xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision 829f0bcb5fe24bb523c5a9e7bd3bb79412e06906) 
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_ratelimit.h"
34 #include "opt_kern_tls.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 #ifdef TCP_ACCOUNTING
67 #include <sys/sched.h>
68 #include <machine/cpu.h>
69 #endif
70 #include <vm/uma.h>
71 
72 #include <net/route.h>
73 #include <net/route/nhop.h>
74 #include <net/vnet.h>
75 
76 #define TCPSTATES		/* for logging */
77 
78 #include <netinet/in.h>
79 #include <netinet/in_kdtrace.h>
80 #include <netinet/in_pcb.h>
81 #include <netinet/ip.h>
82 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
83 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
84 #include <netinet/ip_var.h>
85 #include <netinet/ip6.h>
86 #include <netinet6/in6_pcb.h>
87 #include <netinet6/ip6_var.h>
88 #include <netinet/tcp.h>
89 #define	TCPOUTFLAGS
90 #include <netinet/tcp_fsm.h>
91 #include <netinet/tcp_log_buf.h>
92 #include <netinet/tcp_seq.h>
93 #include <netinet/tcp_timer.h>
94 #include <netinet/tcp_var.h>
95 #include <netinet/tcp_syncache.h>
96 #include <netinet/tcp_hpts.h>
97 #include <netinet/tcp_ratelimit.h>
98 #include <netinet/tcp_accounting.h>
99 #include <netinet/tcpip.h>
100 #include <netinet/cc/cc.h>
101 #include <netinet/cc/cc_newreno.h>
102 #include <netinet/tcp_fastopen.h>
103 #include <netinet/tcp_lro.h>
104 #ifdef NETFLIX_SHARED_CWND
105 #include <netinet/tcp_shared_cwnd.h>
106 #endif
107 #ifdef TCP_OFFLOAD
108 #include <netinet/tcp_offload.h>
109 #endif
110 #ifdef INET6
111 #include <netinet6/tcp6_var.h>
112 #endif
113 #include <netinet/tcp_ecn.h>
114 
115 #include <netipsec/ipsec_support.h>
116 
117 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
118 #include <netipsec/ipsec.h>
119 #include <netipsec/ipsec6.h>
120 #endif				/* IPSEC */
121 
122 #include <netinet/udp.h>
123 #include <netinet/udp_var.h>
124 #include <machine/in_cksum.h>
125 
126 #ifdef MAC
127 #include <security/mac/mac_framework.h>
128 #endif
129 #include "sack_filter.h"
130 #include "tcp_rack.h"
131 #include "rack_bbr_common.h"
132 
133 uma_zone_t rack_zone;
134 uma_zone_t rack_pcb_zone;
135 
136 #ifndef TICKS2SBT
137 #define	TICKS2SBT(__t)	(tick_sbt * ((sbintime_t)(__t)))
138 #endif
139 
140 VNET_DECLARE(uint32_t, newreno_beta);
141 VNET_DECLARE(uint32_t, newreno_beta_ecn);
142 #define V_newreno_beta VNET(newreno_beta)
143 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
144 
145 
146 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
147 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
148 
149 struct sysctl_ctx_list rack_sysctl_ctx;
150 struct sysctl_oid *rack_sysctl_root;
151 
152 #define CUM_ACKED 1
153 #define SACKED 2
154 
155 /*
156  * The RACK module incorporates a number of
157  * TCP ideas that have been put out into the IETF
158  * over the last few years:
159  * - Matt Mathis's Rate Halving which slowly drops
160  *    the congestion window so that the ack clock can
161  *    be maintained during a recovery.
162  * - Yuchung Cheng's RACK TCP (for which its named) that
163  *    will stop us using the number of dup acks and instead
164  *    use time as the gage of when we retransmit.
165  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
166  *    of Dukkipati et.al.
167  * RACK depends on SACK, so if an endpoint arrives that
168  * cannot do SACK the state machine below will shuttle the
169  * connection back to using the "default" TCP stack that is
170  * in FreeBSD.
171  *
172  * To implement RACK the original TCP stack was first decomposed
173  * into a functional state machine with individual states
174  * for each of the possible TCP connection states. The do_segment
175  * functions role in life is to mandate the connection supports SACK
176  * initially and then assure that the RACK state matches the conenction
177  * state before calling the states do_segment function. Each
178  * state is simplified due to the fact that the original do_segment
179  * has been decomposed and we *know* what state we are in (no
180  * switches on the state) and all tests for SACK are gone. This
181  * greatly simplifies what each state does.
182  *
183  * TCP output is also over-written with a new version since it
184  * must maintain the new rack scoreboard.
185  *
186  */
187 static int32_t rack_tlp_thresh = 1;
188 static int32_t rack_tlp_limit = 2;	/* No more than 2 TLPs w-out new data */
189 static int32_t rack_tlp_use_greater = 1;
190 static int32_t rack_reorder_thresh = 2;
191 static int32_t rack_reorder_fade = 60000000;	/* 0 - never fade, def 60,000,000
192 						 * - 60 seconds */
193 static uint8_t rack_req_measurements = 1;
194 /* Attack threshold detections */
195 static uint32_t rack_highest_sack_thresh_seen = 0;
196 static uint32_t rack_highest_move_thresh_seen = 0;
197 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
198 static int32_t rack_hw_pace_extra_slots = 2;	/* 2 extra MSS time betweens */
199 static int32_t rack_hw_rate_caps = 1; /* 1; */
200 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
201 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
202 static int32_t rack_hw_up_only = 1;
203 static int32_t rack_stats_gets_ms_rtt = 1;
204 static int32_t rack_prr_addbackmax = 2;
205 static int32_t rack_do_hystart = 0;
206 static int32_t rack_apply_rtt_with_reduced_conf = 0;
207 
208 static int32_t rack_pkt_delay = 1000;
209 static int32_t rack_send_a_lot_in_prr = 1;
210 static int32_t rack_min_to = 1000;	/* Number of microsecond  min timeout */
211 static int32_t rack_verbose_logging = 0;
212 static int32_t rack_ignore_data_after_close = 1;
213 static int32_t rack_enable_shared_cwnd = 1;
214 static int32_t rack_use_cmp_acks = 1;
215 static int32_t rack_use_fsb = 1;
216 static int32_t rack_use_rfo = 1;
217 static int32_t rack_use_rsm_rfo = 1;
218 static int32_t rack_max_abc_post_recovery = 2;
219 static int32_t rack_client_low_buf = 0;
220 static int32_t rack_dsack_std_based = 0x3;	/* bit field bit 1 sets rc_rack_tmr_std_based and bit 2 sets rc_rack_use_dsack */
221 #ifdef TCP_ACCOUNTING
222 static int32_t rack_tcp_accounting = 0;
223 #endif
224 static int32_t rack_limits_scwnd = 1;
225 static int32_t rack_enable_mqueue_for_nonpaced = 0;
226 static int32_t rack_disable_prr = 0;
227 static int32_t use_rack_rr = 1;
228 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
229 static int32_t rack_persist_min = 250000;	/* 250usec */
230 static int32_t rack_persist_max = 2000000;	/* 2 Second in usec's */
231 static int32_t rack_sack_not_required = 1;	/* set to one to allow non-sack to use rack */
232 static int32_t rack_default_init_window = 0;	/* Use system default */
233 static int32_t rack_limit_time_with_srtt = 0;
234 static int32_t rack_autosndbuf_inc = 20;	/* In percentage form */
235 static int32_t rack_enobuf_hw_boost_mult = 2;	/* How many times the hw rate we boost slot using time_between */
236 static int32_t rack_enobuf_hw_max = 12000;	/* 12 ms in usecs */
237 static int32_t rack_enobuf_hw_min = 10000;	/* 10 ms in usecs */
238 static int32_t rack_hw_rwnd_factor = 2;		/* How many max_segs the rwnd must be before we hold off sending */
239 
240 /*
241  * Currently regular tcp has a rto_min of 30ms
242  * the backoff goes 12 times so that ends up
243  * being a total of 122.850 seconds before a
244  * connection is killed.
245  */
246 static uint32_t rack_def_data_window = 20;
247 static uint32_t rack_goal_bdp = 2;
248 static uint32_t rack_min_srtts = 1;
249 static uint32_t rack_min_measure_usec = 0;
250 static int32_t rack_tlp_min = 10000;	/* 10ms */
251 static int32_t rack_rto_min = 30000;	/* 30,000 usec same as main freebsd */
252 static int32_t rack_rto_max = 4000000;	/* 4 seconds in usec's */
253 static const int32_t rack_free_cache = 2;
254 static int32_t rack_hptsi_segments = 40;
255 static int32_t rack_rate_sample_method = USE_RTT_LOW;
256 static int32_t rack_pace_every_seg = 0;
257 static int32_t rack_delayed_ack_time = 40000;	/* 40ms in usecs */
258 static int32_t rack_slot_reduction = 4;
259 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
260 static int32_t rack_cwnd_block_ends_measure = 0;
261 static int32_t rack_rwnd_block_ends_measure = 0;
262 static int32_t rack_def_profile = 0;
263 
264 static int32_t rack_lower_cwnd_at_tlp = 0;
265 static int32_t rack_limited_retran = 0;
266 static int32_t rack_always_send_oldest = 0;
267 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
268 
269 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
270 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
271 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
272 
273 /* Probertt */
274 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
275 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
276 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
277 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
278 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
279 
280 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
281 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
282 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
283 static uint32_t rack_probertt_use_min_rtt_exit = 0;
284 static uint32_t rack_probe_rtt_sets_cwnd = 0;
285 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
286 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in usecs */
287 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
288 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction */
289 static uint32_t rack_min_probertt_hold = 40000;		/* Equal to delayed ack time */
290 static uint32_t rack_probertt_filter_life = 10000000;
291 static uint32_t rack_probertt_lower_within = 10;
292 static uint32_t rack_min_rtt_movement = 250000;	/* Must move at least 250ms (in microseconds)  to count as a lowering */
293 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
294 static int32_t rack_probertt_clear_is = 1;
295 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
296 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
297 
298 /* Part of pacing */
299 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
300 
301 /* Timely information */
302 /* Combine these two gives the range of 'no change' to bw */
303 /* ie the up/down provide the upper and lower bound */
304 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
305 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
306 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
307 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
308 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
309 static int32_t rack_gp_decrease_per = 20;	/* 20% decrease in multiplier */
310 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multiplier */
311 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
312 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
313 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
314 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
315 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
316 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
317 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
318 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
319 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
320 static int32_t rack_use_max_for_nobackoff = 0;
321 static int32_t rack_timely_int_timely_only = 0;	/* do interim timely's only use the timely algo (no b/w changes)? */
322 static int32_t rack_timely_no_stopping = 0;
323 static int32_t rack_down_raise_thresh = 100;
324 static int32_t rack_req_segs = 1;
325 static uint64_t rack_bw_rate_cap = 0;
326 static uint32_t rack_trace_point_config = 0;
327 static uint32_t rack_trace_point_bb_mode = 4;
328 static int32_t rack_trace_point_count = 0;
329 
330 
331 /* Weird delayed ack mode */
332 static int32_t rack_use_imac_dack = 0;
333 /* Rack specific counters */
334 counter_u64_t rack_saw_enobuf;
335 counter_u64_t rack_saw_enobuf_hw;
336 counter_u64_t rack_saw_enetunreach;
337 counter_u64_t rack_persists_sends;
338 counter_u64_t rack_persists_acks;
339 counter_u64_t rack_persists_loss;
340 counter_u64_t rack_persists_lost_ends;
341 #ifdef INVARIANTS
342 counter_u64_t rack_adjust_map_bw;
343 #endif
344 /* Tail loss probe counters */
345 counter_u64_t rack_tlp_tot;
346 counter_u64_t rack_tlp_newdata;
347 counter_u64_t rack_tlp_retran;
348 counter_u64_t rack_tlp_retran_bytes;
349 counter_u64_t rack_to_tot;
350 counter_u64_t rack_hot_alloc;
351 counter_u64_t rack_to_alloc;
352 counter_u64_t rack_to_alloc_hard;
353 counter_u64_t rack_to_alloc_emerg;
354 counter_u64_t rack_to_alloc_limited;
355 counter_u64_t rack_alloc_limited_conns;
356 counter_u64_t rack_split_limited;
357 
358 counter_u64_t rack_multi_single_eq;
359 counter_u64_t rack_proc_non_comp_ack;
360 
361 counter_u64_t rack_fto_send;
362 counter_u64_t rack_fto_rsm_send;
363 counter_u64_t rack_nfto_resend;
364 counter_u64_t rack_non_fto_send;
365 counter_u64_t rack_extended_rfo;
366 
367 counter_u64_t rack_sack_proc_all;
368 counter_u64_t rack_sack_proc_short;
369 counter_u64_t rack_sack_proc_restart;
370 counter_u64_t rack_sack_attacks_detected;
371 counter_u64_t rack_sack_attacks_reversed;
372 counter_u64_t rack_sack_used_next_merge;
373 counter_u64_t rack_sack_splits;
374 counter_u64_t rack_sack_used_prev_merge;
375 counter_u64_t rack_sack_skipped_acked;
376 counter_u64_t rack_ack_total;
377 counter_u64_t rack_express_sack;
378 counter_u64_t rack_sack_total;
379 counter_u64_t rack_move_none;
380 counter_u64_t rack_move_some;
381 
382 counter_u64_t rack_input_idle_reduces;
383 counter_u64_t rack_collapsed_win;
384 counter_u64_t rack_collapsed_win_seen;
385 counter_u64_t rack_collapsed_win_rxt;
386 counter_u64_t rack_collapsed_win_rxt_bytes;
387 counter_u64_t rack_try_scwnd;
388 counter_u64_t rack_hw_pace_init_fail;
389 counter_u64_t rack_hw_pace_lost;
390 
391 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
392 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
393 
394 
395 #define	RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
396 
397 #define	RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {	\
398 	(tv) = (value) + slop;	 \
399 	if ((u_long)(tv) < (u_long)(tvmin)) \
400 		(tv) = (tvmin); \
401 	if ((u_long)(tv) > (u_long)(tvmax)) \
402 		(tv) = (tvmax); \
403 } while (0)
404 
405 static void
406 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
407 
408 static int
409 rack_process_ack(struct mbuf *m, struct tcphdr *th,
410     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
411     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
412 static int
413 rack_process_data(struct mbuf *m, struct tcphdr *th,
414     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
415     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
416 static void
417 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
418    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
419 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
420 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
421     uint8_t limit_type);
422 static struct rack_sendmap *
423 rack_check_recovery_mode(struct tcpcb *tp,
424     uint32_t tsused);
425 static void
426 rack_cong_signal(struct tcpcb *tp,
427 		 uint32_t type, uint32_t ack, int );
428 static void rack_counter_destroy(void);
429 static int
430 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt);
431 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
432 static void
433 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
434 static void
435 rack_do_segment(struct mbuf *m, struct tcphdr *th,
436     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
437     uint8_t iptos);
438 static void rack_dtor(void *mem, int32_t size, void *arg);
439 static void
440 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
441     uint32_t flex1, uint32_t flex2,
442     uint32_t flex3, uint32_t flex4,
443     uint32_t flex5, uint32_t flex6,
444     uint16_t flex7, uint8_t mod);
445 
446 static void
447 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
448    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
449    struct rack_sendmap *rsm, uint8_t quality);
450 static struct rack_sendmap *
451 rack_find_high_nonack(struct tcp_rack *rack,
452     struct rack_sendmap *rsm);
453 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
454 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
455 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
456 static int rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt);
457 static void
458 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
459 			    tcp_seq th_ack, int line, uint8_t quality);
460 static uint32_t
461 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
462 static int32_t rack_handoff_ok(struct tcpcb *tp);
463 static int32_t rack_init(struct tcpcb *tp);
464 static void rack_init_sysctls(void);
465 static void
466 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
467     struct tcphdr *th, int entered_rec, int dup_ack_struck);
468 static void
469 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
470     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
471     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls);
472 
473 static void
474 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
475     struct rack_sendmap *rsm);
476 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
477 static int32_t rack_output(struct tcpcb *tp);
478 
479 static uint32_t
480 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
481     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
482     uint32_t cts, int *moved_two);
483 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
484 static void rack_remxt_tmr(struct tcpcb *tp);
485 static int rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt);
486 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
487 static int32_t rack_stopall(struct tcpcb *tp);
488 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
489 static uint32_t
490 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
491     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
492 static void
493 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
494     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
495 static int
496 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
497     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
498 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
499 static int
500 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
501     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
502     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
503 static int
504 rack_do_closing(struct mbuf *m, struct tcphdr *th,
505     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
506     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
507 static int
508 rack_do_established(struct mbuf *m, struct tcphdr *th,
509     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
510     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
511 static int
512 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
513     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
514     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
515 static int
516 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
517     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
518     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
519 static int
520 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
521     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
522     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
523 static int
524 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
525     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
526     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
527 static int
528 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
529     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
530     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
531 static int
532 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
533     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
534     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
535 struct rack_sendmap *
536 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
537     uint32_t tsused);
538 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
539     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
540 static void
541      tcp_rack_partialack(struct tcpcb *tp);
542 static int
543 rack_set_profile(struct tcp_rack *rack, int prof);
544 static void
545 rack_apply_deferred_options(struct tcp_rack *rack);
546 
547 int32_t rack_clear_counter=0;
548 
549 static inline void
550 rack_trace_point(struct tcp_rack *rack, int num)
551 {
552 	if (((rack_trace_point_config == num)  ||
553 	     (rack_trace_point_config = 0xffffffff)) &&
554 	    (rack_trace_point_bb_mode != 0) &&
555 	    (rack_trace_point_count > 0) &&
556 	    (rack->rc_tp->t_logstate == 0)) {
557 		int res;
558 		res = atomic_fetchadd_int(&rack_trace_point_count, -1);
559 		if (res > 0) {
560 			rack->rc_tp->t_logstate = rack_trace_point_bb_mode;
561 		} else {
562 			/* Loss a race assure its zero now */
563 			rack_trace_point_count = 0;
564 		}
565 	}
566 }
567 
568 static void
569 rack_swap_beta_values(struct tcp_rack *rack, uint8_t flex8)
570 {
571 	struct sockopt sopt;
572 	struct cc_newreno_opts opt;
573 	struct newreno old;
574 	struct tcpcb *tp;
575 	int error, failed = 0;
576 
577 	tp = rack->rc_tp;
578 	if (tp->t_cc == NULL) {
579 		/* Tcb is leaving */
580 		return;
581 	}
582 	rack->rc_pacing_cc_set = 1;
583 	if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
584 		/* Not new-reno we can't play games with beta! */
585 		failed = 1;
586 		goto out;
587 
588 	}
589 	if (CC_ALGO(tp)->ctl_output == NULL)  {
590 		/* Huh, not using new-reno so no swaps.? */
591 		failed = 2;
592 		goto out;
593 	}
594 	/* Get the current values out */
595 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
596 	sopt.sopt_dir = SOPT_GET;
597 	opt.name = CC_NEWRENO_BETA;
598 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
599 	if (error)  {
600 		failed = 3;
601 		goto out;
602 	}
603 	old.beta = opt.val;
604 	opt.name = CC_NEWRENO_BETA_ECN;
605 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
606 	if (error)  {
607 		failed = 4;
608 		goto out;
609 	}
610 	old.beta_ecn = opt.val;
611 
612 	/* Now lets set in the values we have stored */
613 	sopt.sopt_dir = SOPT_SET;
614 	opt.name = CC_NEWRENO_BETA;
615 	opt.val = rack->r_ctl.rc_saved_beta.beta;
616 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
617 	if (error)  {
618 		failed = 5;
619 		goto out;
620 	}
621 	opt.name = CC_NEWRENO_BETA_ECN;
622 	opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
623 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
624 	if (error) {
625 		failed = 6;
626 		goto out;
627 	}
628 	/* Save off the values for restoral */
629 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
630 out:
631 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
632 		union tcp_log_stackspecific log;
633 		struct timeval tv;
634 		struct newreno *ptr;
635 
636 		ptr = ((struct newreno *)tp->t_ccv.cc_data);
637 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
638 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
639 		log.u_bbr.flex1 = ptr->beta;
640 		log.u_bbr.flex2 = ptr->beta_ecn;
641 		log.u_bbr.flex3 = ptr->newreno_flags;
642 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
643 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
644 		log.u_bbr.flex6 = failed;
645 		log.u_bbr.flex7 = rack->gp_ready;
646 		log.u_bbr.flex7 <<= 1;
647 		log.u_bbr.flex7 |= rack->use_fixed_rate;
648 		log.u_bbr.flex7 <<= 1;
649 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
650 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
651 		log.u_bbr.flex8 = flex8;
652 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
653 			       0, &log, false, NULL, NULL, 0, &tv);
654 	}
655 }
656 
657 static void
658 rack_set_cc_pacing(struct tcp_rack *rack)
659 {
660 	if (rack->rc_pacing_cc_set)
661 		return;
662 	/*
663 	 * Use the swap utility placing in 3 for flex8 to id a
664 	 * set of a new set of values.
665 	 */
666 	rack->rc_pacing_cc_set = 1;
667 	rack_swap_beta_values(rack, 3);
668 }
669 
670 static void
671 rack_undo_cc_pacing(struct tcp_rack *rack)
672 {
673 	if (rack->rc_pacing_cc_set == 0)
674 		return;
675 	/*
676 	 * Use the swap utility placing in 4 for flex8 to id a
677 	 * restoral of the old values.
678 	 */
679 	rack->rc_pacing_cc_set = 0;
680 	rack_swap_beta_values(rack, 4);
681 }
682 
683 #ifdef NETFLIX_PEAKRATE
684 static inline void
685 rack_update_peakrate_thr(struct tcpcb *tp)
686 {
687 	/* Keep in mind that t_maxpeakrate is in B/s. */
688 	uint64_t peak;
689 	peak = uqmax((tp->t_maxseg * 2),
690 		     (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
691 	tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
692 }
693 #endif
694 
695 static int
696 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
697 {
698 	uint32_t stat;
699 	int32_t error;
700 
701 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
702 	if (error || req->newptr == NULL)
703 		return error;
704 
705 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
706 	if (error)
707 		return (error);
708 	if (stat == 1) {
709 #ifdef INVARIANTS
710 		printf("Clearing RACK counters\n");
711 #endif
712 		counter_u64_zero(rack_tlp_tot);
713 		counter_u64_zero(rack_tlp_newdata);
714 		counter_u64_zero(rack_tlp_retran);
715 		counter_u64_zero(rack_tlp_retran_bytes);
716 		counter_u64_zero(rack_to_tot);
717 		counter_u64_zero(rack_saw_enobuf);
718 		counter_u64_zero(rack_saw_enobuf_hw);
719 		counter_u64_zero(rack_saw_enetunreach);
720 		counter_u64_zero(rack_persists_sends);
721 		counter_u64_zero(rack_persists_acks);
722 		counter_u64_zero(rack_persists_loss);
723 		counter_u64_zero(rack_persists_lost_ends);
724 #ifdef INVARIANTS
725 		counter_u64_zero(rack_adjust_map_bw);
726 #endif
727 		counter_u64_zero(rack_to_alloc_hard);
728 		counter_u64_zero(rack_to_alloc_emerg);
729 		counter_u64_zero(rack_sack_proc_all);
730 		counter_u64_zero(rack_fto_send);
731 		counter_u64_zero(rack_fto_rsm_send);
732 		counter_u64_zero(rack_extended_rfo);
733 		counter_u64_zero(rack_hw_pace_init_fail);
734 		counter_u64_zero(rack_hw_pace_lost);
735 		counter_u64_zero(rack_non_fto_send);
736 		counter_u64_zero(rack_nfto_resend);
737 		counter_u64_zero(rack_sack_proc_short);
738 		counter_u64_zero(rack_sack_proc_restart);
739 		counter_u64_zero(rack_to_alloc);
740 		counter_u64_zero(rack_to_alloc_limited);
741 		counter_u64_zero(rack_alloc_limited_conns);
742 		counter_u64_zero(rack_split_limited);
743 		counter_u64_zero(rack_multi_single_eq);
744 		counter_u64_zero(rack_proc_non_comp_ack);
745 		counter_u64_zero(rack_sack_attacks_detected);
746 		counter_u64_zero(rack_sack_attacks_reversed);
747 		counter_u64_zero(rack_sack_used_next_merge);
748 		counter_u64_zero(rack_sack_used_prev_merge);
749 		counter_u64_zero(rack_sack_splits);
750 		counter_u64_zero(rack_sack_skipped_acked);
751 		counter_u64_zero(rack_ack_total);
752 		counter_u64_zero(rack_express_sack);
753 		counter_u64_zero(rack_sack_total);
754 		counter_u64_zero(rack_move_none);
755 		counter_u64_zero(rack_move_some);
756 		counter_u64_zero(rack_try_scwnd);
757 		counter_u64_zero(rack_collapsed_win);
758 		counter_u64_zero(rack_collapsed_win_rxt);
759 		counter_u64_zero(rack_collapsed_win_seen);
760 		counter_u64_zero(rack_collapsed_win_rxt_bytes);
761 	}
762 	rack_clear_counter = 0;
763 	return (0);
764 }
765 
766 static void
767 rack_init_sysctls(void)
768 {
769 	struct sysctl_oid *rack_counters;
770 	struct sysctl_oid *rack_attack;
771 	struct sysctl_oid *rack_pacing;
772 	struct sysctl_oid *rack_timely;
773 	struct sysctl_oid *rack_timers;
774 	struct sysctl_oid *rack_tlp;
775 	struct sysctl_oid *rack_misc;
776 	struct sysctl_oid *rack_features;
777 	struct sysctl_oid *rack_measure;
778 	struct sysctl_oid *rack_probertt;
779 	struct sysctl_oid *rack_hw_pacing;
780 	struct sysctl_oid *rack_tracepoint;
781 
782 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
783 	    SYSCTL_CHILDREN(rack_sysctl_root),
784 	    OID_AUTO,
785 	    "sack_attack",
786 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
787 	    "Rack Sack Attack Counters and Controls");
788 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
789 	    SYSCTL_CHILDREN(rack_sysctl_root),
790 	    OID_AUTO,
791 	    "stats",
792 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
793 	    "Rack Counters");
794 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
795 	    SYSCTL_CHILDREN(rack_sysctl_root),
796 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
797 	    &rack_rate_sample_method , USE_RTT_LOW,
798 	    "What method should we use for rate sampling 0=high, 1=low ");
799 	/* Probe rtt related controls */
800 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
801 	    SYSCTL_CHILDREN(rack_sysctl_root),
802 	    OID_AUTO,
803 	    "probertt",
804 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
805 	    "ProbeRTT related Controls");
806 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
807 	    SYSCTL_CHILDREN(rack_probertt),
808 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
809 	    &rack_atexit_prtt_hbp, 130,
810 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
811 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
812 	    SYSCTL_CHILDREN(rack_probertt),
813 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
814 	    &rack_atexit_prtt, 130,
815 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
816 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
817 	    SYSCTL_CHILDREN(rack_probertt),
818 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
819 	    &rack_per_of_gp_probertt, 60,
820 	    "What percentage of goodput do we pace at in probertt");
821 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
822 	    SYSCTL_CHILDREN(rack_probertt),
823 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
824 	    &rack_per_of_gp_probertt_reduce, 10,
825 	    "What percentage of goodput do we reduce every gp_srtt");
826 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
827 	    SYSCTL_CHILDREN(rack_probertt),
828 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
829 	    &rack_per_of_gp_lowthresh, 40,
830 	    "What percentage of goodput do we allow the multiplier to fall to");
831 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
832 	    SYSCTL_CHILDREN(rack_probertt),
833 	    OID_AUTO, "time_between", CTLFLAG_RW,
834 	    & rack_time_between_probertt, 96000000,
835 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
836 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
837 	    SYSCTL_CHILDREN(rack_probertt),
838 	    OID_AUTO, "safety", CTLFLAG_RW,
839 	    &rack_probe_rtt_safety_val, 2000000,
840 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
841 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
842 	    SYSCTL_CHILDREN(rack_probertt),
843 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
844 	    &rack_probe_rtt_sets_cwnd, 0,
845 	    "Do we set the cwnd too (if always_lower is on)");
846 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
847 	    SYSCTL_CHILDREN(rack_probertt),
848 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
849 	    &rack_max_drain_wait, 2,
850 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
851 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
852 	    SYSCTL_CHILDREN(rack_probertt),
853 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
854 	    &rack_must_drain, 1,
855 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
856 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
857 	    SYSCTL_CHILDREN(rack_probertt),
858 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
859 	    &rack_probertt_use_min_rtt_entry, 1,
860 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
861 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
862 	    SYSCTL_CHILDREN(rack_probertt),
863 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
864 	    &rack_probertt_use_min_rtt_exit, 0,
865 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
866 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
867 	    SYSCTL_CHILDREN(rack_probertt),
868 	    OID_AUTO, "length_div", CTLFLAG_RW,
869 	    &rack_probertt_gpsrtt_cnt_div, 0,
870 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
871 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
872 	    SYSCTL_CHILDREN(rack_probertt),
873 	    OID_AUTO, "length_mul", CTLFLAG_RW,
874 	    &rack_probertt_gpsrtt_cnt_mul, 0,
875 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
876 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
877 	    SYSCTL_CHILDREN(rack_probertt),
878 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
879 	    &rack_min_probertt_hold, 200000,
880 	    "What is the minimum time we hold probertt at target");
881 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
882 	    SYSCTL_CHILDREN(rack_probertt),
883 	    OID_AUTO, "filter_life", CTLFLAG_RW,
884 	    &rack_probertt_filter_life, 10000000,
885 	    "What is the time for the filters life in useconds");
886 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
887 	    SYSCTL_CHILDREN(rack_probertt),
888 	    OID_AUTO, "lower_within", CTLFLAG_RW,
889 	    &rack_probertt_lower_within, 10,
890 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
891 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
892 	    SYSCTL_CHILDREN(rack_probertt),
893 	    OID_AUTO, "must_move", CTLFLAG_RW,
894 	    &rack_min_rtt_movement, 250,
895 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
896 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
897 	    SYSCTL_CHILDREN(rack_probertt),
898 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
899 	    &rack_probertt_clear_is, 1,
900 	    "Do we clear I/S counts on exiting probe-rtt");
901 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
902 	    SYSCTL_CHILDREN(rack_probertt),
903 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
904 	    &rack_max_drain_hbp, 1,
905 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
906 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
907 	    SYSCTL_CHILDREN(rack_probertt),
908 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
909 	    &rack_hbp_thresh, 3,
910 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
911 
912 	rack_tracepoint = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
913 	    SYSCTL_CHILDREN(rack_sysctl_root),
914 	    OID_AUTO,
915 	    "tp",
916 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
917 	    "Rack tracepoint facility");
918 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
919 	    SYSCTL_CHILDREN(rack_tracepoint),
920 	    OID_AUTO, "number", CTLFLAG_RW,
921 	    &rack_trace_point_config, 0,
922 	    "What is the trace point number to activate (0=none, 0xffffffff = all)?");
923 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
924 	    SYSCTL_CHILDREN(rack_tracepoint),
925 	    OID_AUTO, "bbmode", CTLFLAG_RW,
926 	    &rack_trace_point_bb_mode, 4,
927 	    "What is BB logging mode that is activated?");
928 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
929 	    SYSCTL_CHILDREN(rack_tracepoint),
930 	    OID_AUTO, "count", CTLFLAG_RW,
931 	    &rack_trace_point_count, 0,
932 	    "How many connections will have BB logging turned on that hit the tracepoint?");
933 	/* Pacing related sysctls */
934 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
935 	    SYSCTL_CHILDREN(rack_sysctl_root),
936 	    OID_AUTO,
937 	    "pacing",
938 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
939 	    "Pacing related Controls");
940 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
941 	    SYSCTL_CHILDREN(rack_pacing),
942 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
943 	    &rack_max_per_above, 30,
944 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
945 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
946 	    SYSCTL_CHILDREN(rack_pacing),
947 	    OID_AUTO, "pace_to_one", CTLFLAG_RW,
948 	    &rack_pace_one_seg, 0,
949 	    "Do we allow low b/w pacing of 1MSS instead of two");
950 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
951 	    SYSCTL_CHILDREN(rack_pacing),
952 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
953 	    &rack_limit_time_with_srtt, 0,
954 	    "Do we limit pacing time based on srtt");
955 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
956 	    SYSCTL_CHILDREN(rack_pacing),
957 	    OID_AUTO, "init_win", CTLFLAG_RW,
958 	    &rack_default_init_window, 0,
959 	    "Do we have a rack initial window 0 = system default");
960 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
961 	    SYSCTL_CHILDREN(rack_pacing),
962 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
963 	    &rack_per_of_gp_ss, 250,
964 	    "If non zero, what percentage of goodput to pace at in slow start");
965 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
966 	    SYSCTL_CHILDREN(rack_pacing),
967 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
968 	    &rack_per_of_gp_ca, 150,
969 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
970 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
971 	    SYSCTL_CHILDREN(rack_pacing),
972 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
973 	    &rack_per_of_gp_rec, 200,
974 	    "If non zero, what percentage of goodput to pace at in recovery");
975 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
976 	    SYSCTL_CHILDREN(rack_pacing),
977 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
978 	    &rack_hptsi_segments, 40,
979 	    "What size is the max for TSO segments in pacing and burst mitigation");
980 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
981 	    SYSCTL_CHILDREN(rack_pacing),
982 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
983 	    &rack_slot_reduction, 4,
984 	    "When doing only burst mitigation what is the reduce divisor");
985 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
986 	    SYSCTL_CHILDREN(rack_sysctl_root),
987 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
988 	    &rack_pace_every_seg, 0,
989 	    "If set we use pacing, if clear we use only the original burst mitigation");
990 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
991 	    SYSCTL_CHILDREN(rack_pacing),
992 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
993 	    &rack_bw_rate_cap, 0,
994 	    "If set we apply this value to the absolute rate cap used by pacing");
995 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
996 	    SYSCTL_CHILDREN(rack_sysctl_root),
997 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
998 	    &rack_req_measurements, 1,
999 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1000 	/* Hardware pacing */
1001 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1002 	    SYSCTL_CHILDREN(rack_sysctl_root),
1003 	    OID_AUTO,
1004 	    "hdwr_pacing",
1005 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1006 	    "Pacing related Controls");
1007 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1008 	    SYSCTL_CHILDREN(rack_hw_pacing),
1009 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1010 	    &rack_hw_rwnd_factor, 2,
1011 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1012 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1013 	    SYSCTL_CHILDREN(rack_hw_pacing),
1014 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1015 	    &rack_enobuf_hw_boost_mult, 2,
1016 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1017 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1018 	    SYSCTL_CHILDREN(rack_hw_pacing),
1019 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1020 	    &rack_enobuf_hw_max, 2,
1021 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1022 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1023 	    SYSCTL_CHILDREN(rack_hw_pacing),
1024 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1025 	    &rack_enobuf_hw_min, 2,
1026 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1027 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1028 	    SYSCTL_CHILDREN(rack_hw_pacing),
1029 	    OID_AUTO, "enable", CTLFLAG_RW,
1030 	    &rack_enable_hw_pacing, 0,
1031 	    "Should RACK attempt to use hw pacing?");
1032 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1033 	    SYSCTL_CHILDREN(rack_hw_pacing),
1034 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1035 	    &rack_hw_rate_caps, 1,
1036 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1037 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1038 	    SYSCTL_CHILDREN(rack_hw_pacing),
1039 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1040 	    &rack_hw_rate_min, 0,
1041 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1042 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1043 	    SYSCTL_CHILDREN(rack_hw_pacing),
1044 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1045 	    &rack_hw_rate_to_low, 0,
1046 	    "If we fall below this rate, dis-engage hw pacing?");
1047 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1048 	    SYSCTL_CHILDREN(rack_hw_pacing),
1049 	    OID_AUTO, "up_only", CTLFLAG_RW,
1050 	    &rack_hw_up_only, 1,
1051 	    "Do we allow hw pacing to lower the rate selected?");
1052 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1053 	    SYSCTL_CHILDREN(rack_hw_pacing),
1054 	    OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1055 	    &rack_hw_pace_extra_slots, 2,
1056 	    "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1057 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1058 	    SYSCTL_CHILDREN(rack_sysctl_root),
1059 	    OID_AUTO,
1060 	    "timely",
1061 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1062 	    "Rack Timely RTT Controls");
1063 	/* Timely based GP dynmics */
1064 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1065 	    SYSCTL_CHILDREN(rack_timely),
1066 	    OID_AUTO, "upper", CTLFLAG_RW,
1067 	    &rack_gp_per_bw_mul_up, 2,
1068 	    "Rack timely upper range for equal b/w (in percentage)");
1069 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1070 	    SYSCTL_CHILDREN(rack_timely),
1071 	    OID_AUTO, "lower", CTLFLAG_RW,
1072 	    &rack_gp_per_bw_mul_down, 4,
1073 	    "Rack timely lower range for equal b/w (in percentage)");
1074 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1075 	    SYSCTL_CHILDREN(rack_timely),
1076 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1077 	    &rack_gp_rtt_maxmul, 3,
1078 	    "Rack timely multiplier of lowest rtt for rtt_max");
1079 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1080 	    SYSCTL_CHILDREN(rack_timely),
1081 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1082 	    &rack_gp_rtt_mindiv, 4,
1083 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1084 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1085 	    SYSCTL_CHILDREN(rack_timely),
1086 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1087 	    &rack_gp_rtt_minmul, 1,
1088 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1089 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1090 	    SYSCTL_CHILDREN(rack_timely),
1091 	    OID_AUTO, "decrease", CTLFLAG_RW,
1092 	    &rack_gp_decrease_per, 20,
1093 	    "Rack timely decrease percentage of our GP multiplication factor");
1094 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1095 	    SYSCTL_CHILDREN(rack_timely),
1096 	    OID_AUTO, "increase", CTLFLAG_RW,
1097 	    &rack_gp_increase_per, 2,
1098 	    "Rack timely increase perentage of our GP multiplication factor");
1099 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1100 	    SYSCTL_CHILDREN(rack_timely),
1101 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1102 	    &rack_per_lower_bound, 50,
1103 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1104 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1105 	    SYSCTL_CHILDREN(rack_timely),
1106 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1107 	    &rack_per_upper_bound_ss, 0,
1108 	    "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1109 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1110 	    SYSCTL_CHILDREN(rack_timely),
1111 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1112 	    &rack_per_upper_bound_ca, 0,
1113 	    "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1114 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1115 	    SYSCTL_CHILDREN(rack_timely),
1116 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1117 	    &rack_do_dyn_mul, 0,
1118 	    "Rack timely do we enable dynmaic timely goodput by default");
1119 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1120 	    SYSCTL_CHILDREN(rack_timely),
1121 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1122 	    &rack_gp_no_rec_chg, 1,
1123 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1124 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1125 	    SYSCTL_CHILDREN(rack_timely),
1126 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1127 	    &rack_timely_dec_clear, 6,
1128 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1129 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1130 	    SYSCTL_CHILDREN(rack_timely),
1131 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1132 	    &rack_timely_max_push_rise, 3,
1133 	    "Rack timely how many times do we push up with b/w increase");
1134 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1135 	    SYSCTL_CHILDREN(rack_timely),
1136 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1137 	    &rack_timely_max_push_drop, 3,
1138 	    "Rack timely how many times do we push back on b/w decent");
1139 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1140 	    SYSCTL_CHILDREN(rack_timely),
1141 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1142 	    &rack_timely_min_segs, 4,
1143 	    "Rack timely when setting the cwnd what is the min num segments");
1144 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1145 	    SYSCTL_CHILDREN(rack_timely),
1146 	    OID_AUTO, "noback_max", CTLFLAG_RW,
1147 	    &rack_use_max_for_nobackoff, 0,
1148 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1149 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1150 	    SYSCTL_CHILDREN(rack_timely),
1151 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1152 	    &rack_timely_int_timely_only, 0,
1153 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1154 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1155 	    SYSCTL_CHILDREN(rack_timely),
1156 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1157 	    &rack_timely_no_stopping, 0,
1158 	    "Rack timely don't stop increase");
1159 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1160 	    SYSCTL_CHILDREN(rack_timely),
1161 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1162 	    &rack_down_raise_thresh, 100,
1163 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1164 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1165 	    SYSCTL_CHILDREN(rack_timely),
1166 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1167 	    &rack_req_segs, 1,
1168 	    "Bottom dragging if not these many segments outstanding and room");
1169 
1170 	/* TLP and Rack related parameters */
1171 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1172 	    SYSCTL_CHILDREN(rack_sysctl_root),
1173 	    OID_AUTO,
1174 	    "tlp",
1175 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1176 	    "TLP and Rack related Controls");
1177 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1178 	    SYSCTL_CHILDREN(rack_tlp),
1179 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1180 	    &use_rack_rr, 1,
1181 	    "Do we use Rack Rapid Recovery");
1182 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1183 	    SYSCTL_CHILDREN(rack_tlp),
1184 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1185 	    &rack_max_abc_post_recovery, 2,
1186 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1187 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1188 	    SYSCTL_CHILDREN(rack_tlp),
1189 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1190 	    &rack_non_rxt_use_cr, 0,
1191 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1192 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1193 	    SYSCTL_CHILDREN(rack_tlp),
1194 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1195 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1196 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1197 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1198 	    SYSCTL_CHILDREN(rack_tlp),
1199 	    OID_AUTO, "limit", CTLFLAG_RW,
1200 	    &rack_tlp_limit, 2,
1201 	    "How many TLP's can be sent without sending new data");
1202 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1203 	    SYSCTL_CHILDREN(rack_tlp),
1204 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1205 	    &rack_tlp_use_greater, 1,
1206 	    "Should we use the rack_rtt time if its greater than srtt");
1207 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1208 	    SYSCTL_CHILDREN(rack_tlp),
1209 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1210 	    &rack_tlp_min, 10000,
1211 	    "TLP minimum timeout per the specification (in microseconds)");
1212 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1213 	    SYSCTL_CHILDREN(rack_tlp),
1214 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1215 	    &rack_always_send_oldest, 0,
1216 	    "Should we always send the oldest TLP and RACK-TLP");
1217 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1218 	    SYSCTL_CHILDREN(rack_tlp),
1219 	    OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1220 	    &rack_limited_retran, 0,
1221 	    "How many times can a rack timeout drive out sends");
1222 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1223 	    SYSCTL_CHILDREN(rack_tlp),
1224 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1225 	    &rack_lower_cwnd_at_tlp, 0,
1226 	    "When a TLP completes a retran should we enter recovery");
1227 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1228 	    SYSCTL_CHILDREN(rack_tlp),
1229 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1230 	    &rack_reorder_thresh, 2,
1231 	    "What factor for rack will be added when seeing reordering (shift right)");
1232 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1233 	    SYSCTL_CHILDREN(rack_tlp),
1234 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1235 	    &rack_tlp_thresh, 1,
1236 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1237 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1238 	    SYSCTL_CHILDREN(rack_tlp),
1239 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1240 	    &rack_reorder_fade, 60000000,
1241 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1242 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1243 	    SYSCTL_CHILDREN(rack_tlp),
1244 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1245 	    &rack_pkt_delay, 1000,
1246 	    "Extra RACK time (in microseconds) besides reordering thresh");
1247 
1248 	/* Timer related controls */
1249 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1250 	    SYSCTL_CHILDREN(rack_sysctl_root),
1251 	    OID_AUTO,
1252 	    "timers",
1253 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1254 	    "Timer related controls");
1255 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1256 	    SYSCTL_CHILDREN(rack_timers),
1257 	    OID_AUTO, "persmin", CTLFLAG_RW,
1258 	    &rack_persist_min, 250000,
1259 	    "What is the minimum time in microseconds between persists");
1260 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1261 	    SYSCTL_CHILDREN(rack_timers),
1262 	    OID_AUTO, "persmax", CTLFLAG_RW,
1263 	    &rack_persist_max, 2000000,
1264 	    "What is the largest delay in microseconds between persists");
1265 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1266 	    SYSCTL_CHILDREN(rack_timers),
1267 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1268 	    &rack_delayed_ack_time, 40000,
1269 	    "Delayed ack time (40ms in microseconds)");
1270 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1271 	    SYSCTL_CHILDREN(rack_timers),
1272 	    OID_AUTO, "minrto", CTLFLAG_RW,
1273 	    &rack_rto_min, 30000,
1274 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1275 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1276 	    SYSCTL_CHILDREN(rack_timers),
1277 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1278 	    &rack_rto_max, 4000000,
1279 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1280 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1281 	    SYSCTL_CHILDREN(rack_timers),
1282 	    OID_AUTO, "minto", CTLFLAG_RW,
1283 	    &rack_min_to, 1000,
1284 	    "Minimum rack timeout in microseconds");
1285 	/* Measure controls */
1286 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1287 	    SYSCTL_CHILDREN(rack_sysctl_root),
1288 	    OID_AUTO,
1289 	    "measure",
1290 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1291 	    "Measure related controls");
1292 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1293 	    SYSCTL_CHILDREN(rack_measure),
1294 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1295 	    &rack_wma_divisor, 8,
1296 	    "When doing b/w calculation what is the  divisor for the WMA");
1297 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1298 	    SYSCTL_CHILDREN(rack_measure),
1299 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1300 	    &rack_cwnd_block_ends_measure, 0,
1301 	    "Does a cwnd just-return end the measurement window (app limited)");
1302 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1303 	    SYSCTL_CHILDREN(rack_measure),
1304 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1305 	    &rack_rwnd_block_ends_measure, 0,
1306 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1307 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1308 	    SYSCTL_CHILDREN(rack_measure),
1309 	    OID_AUTO, "min_target", CTLFLAG_RW,
1310 	    &rack_def_data_window, 20,
1311 	    "What is the minimum target window (in mss) for a GP measurements");
1312 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1313 	    SYSCTL_CHILDREN(rack_measure),
1314 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1315 	    &rack_goal_bdp, 2,
1316 	    "What is the goal BDP to measure");
1317 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1318 	    SYSCTL_CHILDREN(rack_measure),
1319 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1320 	    &rack_min_srtts, 1,
1321 	    "What is the goal BDP to measure");
1322 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1323 	    SYSCTL_CHILDREN(rack_measure),
1324 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1325 	    &rack_min_measure_usec, 0,
1326 	    "What is the Minimum time time for a measurement if 0, this is off");
1327 	/* Features */
1328 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1329 	    SYSCTL_CHILDREN(rack_sysctl_root),
1330 	    OID_AUTO,
1331 	    "features",
1332 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1333 	    "Feature controls");
1334 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1335 	    SYSCTL_CHILDREN(rack_features),
1336 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1337 	    &rack_use_cmp_acks, 1,
1338 	    "Should RACK have LRO send compressed acks");
1339 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1340 	    SYSCTL_CHILDREN(rack_features),
1341 	    OID_AUTO, "fsb", CTLFLAG_RW,
1342 	    &rack_use_fsb, 1,
1343 	    "Should RACK use the fast send block?");
1344 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1345 	    SYSCTL_CHILDREN(rack_features),
1346 	    OID_AUTO, "rfo", CTLFLAG_RW,
1347 	    &rack_use_rfo, 1,
1348 	    "Should RACK use rack_fast_output()?");
1349 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1350 	    SYSCTL_CHILDREN(rack_features),
1351 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1352 	    &rack_use_rsm_rfo, 1,
1353 	    "Should RACK use rack_fast_rsm_output()?");
1354 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1355 	    SYSCTL_CHILDREN(rack_features),
1356 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1357 	    &rack_enable_mqueue_for_nonpaced, 0,
1358 	    "Should RACK use mbuf queuing for non-paced connections");
1359 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1360 	    SYSCTL_CHILDREN(rack_features),
1361 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1362 	    &rack_do_hystart, 0,
1363 	    "Should RACK enable HyStart++ on connections?");
1364 	/* Misc rack controls */
1365 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1366 	    SYSCTL_CHILDREN(rack_sysctl_root),
1367 	    OID_AUTO,
1368 	    "misc",
1369 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1370 	    "Misc related controls");
1371 #ifdef TCP_ACCOUNTING
1372 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1373 	    SYSCTL_CHILDREN(rack_misc),
1374 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1375 	    &rack_tcp_accounting, 0,
1376 	    "Should we turn on TCP accounting for all rack sessions?");
1377 #endif
1378 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1379 	    SYSCTL_CHILDREN(rack_misc),
1380 	    OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1381 	    &rack_apply_rtt_with_reduced_conf, 0,
1382 	    "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1383 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1384 	    SYSCTL_CHILDREN(rack_misc),
1385 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1386 	    &rack_dsack_std_based, 3,
1387 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1388 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1389 	    SYSCTL_CHILDREN(rack_misc),
1390 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1391 	    &rack_prr_addbackmax, 2,
1392 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1393 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1394 	    SYSCTL_CHILDREN(rack_misc),
1395 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1396 	    &rack_stats_gets_ms_rtt, 1,
1397 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1398 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1399 	    SYSCTL_CHILDREN(rack_misc),
1400 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1401 	    &rack_client_low_buf, 0,
1402 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1403 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1404 	    SYSCTL_CHILDREN(rack_misc),
1405 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1406 	    &rack_def_profile, 0,
1407 	    "Should RACK use a default profile (0=no, num == profile num)?");
1408 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1409 	    SYSCTL_CHILDREN(rack_misc),
1410 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1411 	    &rack_enable_shared_cwnd, 1,
1412 	    "Should RACK try to use the shared cwnd on connections where allowed");
1413 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1414 	    SYSCTL_CHILDREN(rack_misc),
1415 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1416 	    &rack_limits_scwnd, 1,
1417 	    "Should RACK place low end time limits on the shared cwnd feature");
1418 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1419 	    SYSCTL_CHILDREN(rack_misc),
1420 	    OID_AUTO, "iMac_dack", CTLFLAG_RW,
1421 	    &rack_use_imac_dack, 0,
1422 	    "Should RACK try to emulate iMac delayed ack");
1423 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1424 	    SYSCTL_CHILDREN(rack_misc),
1425 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1426 	    &rack_disable_prr, 0,
1427 	    "Should RACK not use prr and only pace (must have pacing on)");
1428 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1429 	    SYSCTL_CHILDREN(rack_misc),
1430 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1431 	    &rack_verbose_logging, 0,
1432 	    "Should RACK black box logging be verbose");
1433 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1434 	    SYSCTL_CHILDREN(rack_misc),
1435 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1436 	    &rack_ignore_data_after_close, 1,
1437 	    "Do we hold off sending a RST until all pending data is ack'd");
1438 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1439 	    SYSCTL_CHILDREN(rack_misc),
1440 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1441 	    &rack_sack_not_required, 1,
1442 	    "Do we allow rack to run on connections not supporting SACK");
1443 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1444 	    SYSCTL_CHILDREN(rack_misc),
1445 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1446 	    &rack_send_a_lot_in_prr, 1,
1447 	    "Send a lot in prr");
1448 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1449 	    SYSCTL_CHILDREN(rack_misc),
1450 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1451 	    &rack_autosndbuf_inc, 20,
1452 	    "What percentage should rack scale up its snd buffer by?");
1453 	/* Sack Attacker detection stuff */
1454 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1455 	    SYSCTL_CHILDREN(rack_attack),
1456 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1457 	    &rack_highest_sack_thresh_seen, 0,
1458 	    "Highest sack to ack ratio seen");
1459 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1460 	    SYSCTL_CHILDREN(rack_attack),
1461 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1462 	    &rack_highest_move_thresh_seen, 0,
1463 	    "Highest move to non-move ratio seen");
1464 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1465 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1466 	    SYSCTL_CHILDREN(rack_attack),
1467 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1468 	    &rack_ack_total,
1469 	    "Total number of Ack's");
1470 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1471 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1472 	    SYSCTL_CHILDREN(rack_attack),
1473 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1474 	    &rack_express_sack,
1475 	    "Total expresss number of Sack's");
1476 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1477 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1478 	    SYSCTL_CHILDREN(rack_attack),
1479 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1480 	    &rack_sack_total,
1481 	    "Total number of SACKs");
1482 	rack_move_none = counter_u64_alloc(M_WAITOK);
1483 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1484 	    SYSCTL_CHILDREN(rack_attack),
1485 	    OID_AUTO, "move_none", CTLFLAG_RD,
1486 	    &rack_move_none,
1487 	    "Total number of SACK index reuse of positions under threshold");
1488 	rack_move_some = counter_u64_alloc(M_WAITOK);
1489 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1490 	    SYSCTL_CHILDREN(rack_attack),
1491 	    OID_AUTO, "move_some", CTLFLAG_RD,
1492 	    &rack_move_some,
1493 	    "Total number of SACK index reuse of positions over threshold");
1494 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1495 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1496 	    SYSCTL_CHILDREN(rack_attack),
1497 	    OID_AUTO, "attacks", CTLFLAG_RD,
1498 	    &rack_sack_attacks_detected,
1499 	    "Total number of SACK attackers that had sack disabled");
1500 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1501 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1502 	    SYSCTL_CHILDREN(rack_attack),
1503 	    OID_AUTO, "reversed", CTLFLAG_RD,
1504 	    &rack_sack_attacks_reversed,
1505 	    "Total number of SACK attackers that were later determined false positive");
1506 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1507 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1508 	    SYSCTL_CHILDREN(rack_attack),
1509 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1510 	    &rack_sack_used_next_merge,
1511 	    "Total number of times we used the next merge");
1512 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1513 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1514 	    SYSCTL_CHILDREN(rack_attack),
1515 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1516 	    &rack_sack_used_prev_merge,
1517 	    "Total number of times we used the prev merge");
1518 	/* Counters */
1519 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1520 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1521 	    SYSCTL_CHILDREN(rack_counters),
1522 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1523 	    &rack_fto_send, "Total number of rack_fast_output sends");
1524 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1525 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1526 	    SYSCTL_CHILDREN(rack_counters),
1527 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1528 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1529 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1530 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1531 	    SYSCTL_CHILDREN(rack_counters),
1532 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1533 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1534 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1535 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1536 	    SYSCTL_CHILDREN(rack_counters),
1537 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1538 	    &rack_non_fto_send, "Total number of rack_output first sends");
1539 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1540 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1541 	    SYSCTL_CHILDREN(rack_counters),
1542 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1543 	    &rack_extended_rfo, "Total number of times we extended rfo");
1544 
1545 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1546 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1547 	    SYSCTL_CHILDREN(rack_counters),
1548 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1549 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1550 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1551 
1552 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1553 	    SYSCTL_CHILDREN(rack_counters),
1554 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1555 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1556 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1557 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1558 	    SYSCTL_CHILDREN(rack_counters),
1559 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1560 	    &rack_tlp_tot,
1561 	    "Total number of tail loss probe expirations");
1562 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1563 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1564 	    SYSCTL_CHILDREN(rack_counters),
1565 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1566 	    &rack_tlp_newdata,
1567 	    "Total number of tail loss probe sending new data");
1568 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1569 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1570 	    SYSCTL_CHILDREN(rack_counters),
1571 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1572 	    &rack_tlp_retran,
1573 	    "Total number of tail loss probe sending retransmitted data");
1574 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1575 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1576 	    SYSCTL_CHILDREN(rack_counters),
1577 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1578 	    &rack_tlp_retran_bytes,
1579 	    "Total bytes of tail loss probe sending retransmitted data");
1580 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1581 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1582 	    SYSCTL_CHILDREN(rack_counters),
1583 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1584 	    &rack_to_tot,
1585 	    "Total number of times the rack to expired");
1586 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1587 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1588 	    SYSCTL_CHILDREN(rack_counters),
1589 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1590 	    &rack_saw_enobuf,
1591 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1592 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1593 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1594 	    SYSCTL_CHILDREN(rack_counters),
1595 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1596 	    &rack_saw_enobuf_hw,
1597 	    "Total number of times a send returned enobuf for hdwr paced connections");
1598 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1599 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1600 	    SYSCTL_CHILDREN(rack_counters),
1601 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1602 	    &rack_saw_enetunreach,
1603 	    "Total number of times a send received a enetunreachable");
1604 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1605 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1606 	    SYSCTL_CHILDREN(rack_counters),
1607 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1608 	    &rack_hot_alloc,
1609 	    "Total allocations from the top of our list");
1610 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1611 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1612 	    SYSCTL_CHILDREN(rack_counters),
1613 	    OID_AUTO, "allocs", CTLFLAG_RD,
1614 	    &rack_to_alloc,
1615 	    "Total allocations of tracking structures");
1616 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1617 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1618 	    SYSCTL_CHILDREN(rack_counters),
1619 	    OID_AUTO, "allochard", CTLFLAG_RD,
1620 	    &rack_to_alloc_hard,
1621 	    "Total allocations done with sleeping the hard way");
1622 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1623 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1624 	    SYSCTL_CHILDREN(rack_counters),
1625 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1626 	    &rack_to_alloc_emerg,
1627 	    "Total allocations done from emergency cache");
1628 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1629 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1630 	    SYSCTL_CHILDREN(rack_counters),
1631 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1632 	    &rack_to_alloc_limited,
1633 	    "Total allocations dropped due to limit");
1634 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1635 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1636 	    SYSCTL_CHILDREN(rack_counters),
1637 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1638 	    &rack_alloc_limited_conns,
1639 	    "Connections with allocations dropped due to limit");
1640 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1641 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1642 	    SYSCTL_CHILDREN(rack_counters),
1643 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1644 	    &rack_split_limited,
1645 	    "Split allocations dropped due to limit");
1646 	rack_persists_sends = counter_u64_alloc(M_WAITOK);
1647 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1648 	    SYSCTL_CHILDREN(rack_counters),
1649 	    OID_AUTO, "persist_sends", CTLFLAG_RD,
1650 	    &rack_persists_sends,
1651 	    "Number of times we sent a persist probe");
1652 	rack_persists_acks = counter_u64_alloc(M_WAITOK);
1653 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1654 	    SYSCTL_CHILDREN(rack_counters),
1655 	    OID_AUTO, "persist_acks", CTLFLAG_RD,
1656 	    &rack_persists_acks,
1657 	    "Number of times a persist probe was acked");
1658 	rack_persists_loss = counter_u64_alloc(M_WAITOK);
1659 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1660 	    SYSCTL_CHILDREN(rack_counters),
1661 	    OID_AUTO, "persist_loss", CTLFLAG_RD,
1662 	    &rack_persists_loss,
1663 	    "Number of times we detected a lost persist probe (no ack)");
1664 	rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1665 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1666 	    SYSCTL_CHILDREN(rack_counters),
1667 	    OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1668 	    &rack_persists_lost_ends,
1669 	    "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1670 #ifdef INVARIANTS
1671 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1672 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1673 	    SYSCTL_CHILDREN(rack_counters),
1674 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1675 	    &rack_adjust_map_bw,
1676 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1677 #endif
1678 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1679 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1680 	    SYSCTL_CHILDREN(rack_counters),
1681 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1682 	    &rack_multi_single_eq,
1683 	    "Number of compressed acks total represented");
1684 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1685 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1686 	    SYSCTL_CHILDREN(rack_counters),
1687 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1688 	    &rack_proc_non_comp_ack,
1689 	    "Number of non compresseds acks that we processed");
1690 
1691 
1692 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1693 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1694 	    SYSCTL_CHILDREN(rack_counters),
1695 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1696 	    &rack_sack_proc_all,
1697 	    "Total times we had to walk whole list for sack processing");
1698 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1699 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1700 	    SYSCTL_CHILDREN(rack_counters),
1701 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1702 	    &rack_sack_proc_restart,
1703 	    "Total times we had to walk whole list due to a restart");
1704 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1705 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1706 	    SYSCTL_CHILDREN(rack_counters),
1707 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1708 	    &rack_sack_proc_short,
1709 	    "Total times we took shortcut for sack processing");
1710 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1711 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1712 	    SYSCTL_CHILDREN(rack_attack),
1713 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1714 	    &rack_sack_skipped_acked,
1715 	    "Total number of times we skipped previously sacked");
1716 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1717 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1718 	    SYSCTL_CHILDREN(rack_attack),
1719 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1720 	    &rack_sack_splits,
1721 	    "Total number of times we did the old fashion tree split");
1722 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1723 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1724 	    SYSCTL_CHILDREN(rack_counters),
1725 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1726 	    &rack_input_idle_reduces,
1727 	    "Total number of idle reductions on input");
1728 	rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1729 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1730 	    SYSCTL_CHILDREN(rack_counters),
1731 	    OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1732 	    &rack_collapsed_win_seen,
1733 	    "Total number of collapsed window events seen (where our window shrinks)");
1734 
1735 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1736 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1737 	    SYSCTL_CHILDREN(rack_counters),
1738 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1739 	    &rack_collapsed_win,
1740 	    "Total number of collapsed window events where we mark packets");
1741 	rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1742 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1743 	    SYSCTL_CHILDREN(rack_counters),
1744 	    OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1745 	    &rack_collapsed_win_rxt,
1746 	    "Total number of packets that were retransmitted");
1747 	rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1748 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1749 	    SYSCTL_CHILDREN(rack_counters),
1750 	    OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1751 	    &rack_collapsed_win_rxt_bytes,
1752 	    "Total number of bytes that were retransmitted");
1753 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1754 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1755 	    SYSCTL_CHILDREN(rack_counters),
1756 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1757 	    &rack_try_scwnd,
1758 	    "Total number of scwnd attempts");
1759 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1760 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1761 	    OID_AUTO, "outsize", CTLFLAG_RD,
1762 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1763 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1764 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1765 	    OID_AUTO, "opts", CTLFLAG_RD,
1766 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1767 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1768 	    SYSCTL_CHILDREN(rack_sysctl_root),
1769 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1770 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1771 }
1772 
1773 static __inline int
1774 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1775 {
1776 	if (SEQ_GEQ(b->r_start, a->r_start) &&
1777 	    SEQ_LT(b->r_start, a->r_end)) {
1778 		/*
1779 		 * The entry b is within the
1780 		 * block a. i.e.:
1781 		 * a --   |-------------|
1782 		 * b --   |----|
1783 		 * <or>
1784 		 * b --       |------|
1785 		 * <or>
1786 		 * b --       |-----------|
1787 		 */
1788 		return (0);
1789 	} else if (SEQ_GEQ(b->r_start, a->r_end)) {
1790 		/*
1791 		 * b falls as either the next
1792 		 * sequence block after a so a
1793 		 * is said to be smaller than b.
1794 		 * i.e:
1795 		 * a --   |------|
1796 		 * b --          |--------|
1797 		 * or
1798 		 * b --              |-----|
1799 		 */
1800 		return (1);
1801 	}
1802 	/*
1803 	 * Whats left is where a is
1804 	 * larger than b. i.e:
1805 	 * a --         |-------|
1806 	 * b --  |---|
1807 	 * or even possibly
1808 	 * b --   |--------------|
1809 	 */
1810 	return (-1);
1811 }
1812 
1813 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1814 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1815 
1816 static uint32_t
1817 rc_init_window(struct tcp_rack *rack)
1818 {
1819 	uint32_t win;
1820 
1821 	if (rack->rc_init_win == 0) {
1822 		/*
1823 		 * Nothing set by the user, use the system stack
1824 		 * default.
1825 		 */
1826 		return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1827 	}
1828 	win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1829 	return (win);
1830 }
1831 
1832 static uint64_t
1833 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1834 {
1835 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1836 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1837 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1838 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1839 	else
1840 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1841 }
1842 
1843 static uint64_t
1844 rack_get_bw(struct tcp_rack *rack)
1845 {
1846 	if (rack->use_fixed_rate) {
1847 		/* Return the fixed pacing rate */
1848 		return (rack_get_fixed_pacing_bw(rack));
1849 	}
1850 	if (rack->r_ctl.gp_bw == 0) {
1851 		/*
1852 		 * We have yet no b/w measurement,
1853 		 * if we have a user set initial bw
1854 		 * return it. If we don't have that and
1855 		 * we have an srtt, use the tcp IW (10) to
1856 		 * calculate a fictional b/w over the SRTT
1857 		 * which is more or less a guess. Note
1858 		 * we don't use our IW from rack on purpose
1859 		 * so if we have like IW=30, we are not
1860 		 * calculating a "huge" b/w.
1861 		 */
1862 		uint64_t bw, srtt;
1863 		if (rack->r_ctl.init_rate)
1864 			return (rack->r_ctl.init_rate);
1865 
1866 		/* Has the user set a max peak rate? */
1867 #ifdef NETFLIX_PEAKRATE
1868 		if (rack->rc_tp->t_maxpeakrate)
1869 			return (rack->rc_tp->t_maxpeakrate);
1870 #endif
1871 		/* Ok lets come up with the IW guess, if we have a srtt */
1872 		if (rack->rc_tp->t_srtt == 0) {
1873 			/*
1874 			 * Go with old pacing method
1875 			 * i.e. burst mitigation only.
1876 			 */
1877 			return (0);
1878 		}
1879 		/* Ok lets get the initial TCP win (not racks) */
1880 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
1881 		srtt = (uint64_t)rack->rc_tp->t_srtt;
1882 		bw *= (uint64_t)USECS_IN_SECOND;
1883 		bw /= srtt;
1884 		if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1885 			bw = rack->r_ctl.bw_rate_cap;
1886 		return (bw);
1887 	} else {
1888 		uint64_t bw;
1889 
1890 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
1891 			/* Averaging is done, we can return the value */
1892 			bw = rack->r_ctl.gp_bw;
1893 		} else {
1894 			/* Still doing initial average must calculate */
1895 			bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
1896 		}
1897 #ifdef NETFLIX_PEAKRATE
1898 		if ((rack->rc_tp->t_maxpeakrate) &&
1899 		    (bw > rack->rc_tp->t_maxpeakrate)) {
1900 			/* The user has set a peak rate to pace at
1901 			 * don't allow us to pace faster than that.
1902 			 */
1903 			return (rack->rc_tp->t_maxpeakrate);
1904 		}
1905 #endif
1906 		if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1907 			bw = rack->r_ctl.bw_rate_cap;
1908 		return (bw);
1909 	}
1910 }
1911 
1912 static uint16_t
1913 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
1914 {
1915 	if (rack->use_fixed_rate) {
1916 		return (100);
1917 	} else if (rack->in_probe_rtt && (rsm == NULL))
1918 		return (rack->r_ctl.rack_per_of_gp_probertt);
1919 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
1920 		  rack->r_ctl.rack_per_of_gp_rec)) {
1921 		if (rsm) {
1922 			/* a retransmission always use the recovery rate */
1923 			return (rack->r_ctl.rack_per_of_gp_rec);
1924 		} else if (rack->rack_rec_nonrxt_use_cr) {
1925 			/* Directed to use the configured rate */
1926 			goto configured_rate;
1927 		} else if (rack->rack_no_prr &&
1928 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
1929 			/* No PRR, lets just use the b/w estimate only */
1930 			return (100);
1931 		} else {
1932 			/*
1933 			 * Here we may have a non-retransmit but we
1934 			 * have no overrides, so just use the recovery
1935 			 * rate (prr is in effect).
1936 			 */
1937 			return (rack->r_ctl.rack_per_of_gp_rec);
1938 		}
1939 	}
1940 configured_rate:
1941 	/* For the configured rate we look at our cwnd vs the ssthresh */
1942 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1943 		return (rack->r_ctl.rack_per_of_gp_ss);
1944 	else
1945 		return (rack->r_ctl.rack_per_of_gp_ca);
1946 }
1947 
1948 static void
1949 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
1950 {
1951 	/*
1952 	 * Types of logs (mod value)
1953 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
1954 	 * 2 = a dsack round begins, persist is reset to 16.
1955 	 * 3 = a dsack round ends
1956 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
1957 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
1958 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
1959 	 */
1960 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1961 		union tcp_log_stackspecific log;
1962 		struct timeval tv;
1963 
1964 		memset(&log, 0, sizeof(log));
1965 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
1966 		log.u_bbr.flex1 <<= 1;
1967 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
1968 		log.u_bbr.flex1 <<= 1;
1969 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
1970 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
1971 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
1972 		log.u_bbr.flex4 = flex4;
1973 		log.u_bbr.flex5 = flex5;
1974 		log.u_bbr.flex6 = flex6;
1975 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
1976 		log.u_bbr.flex8 = mod;
1977 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1978 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1979 		    &rack->rc_inp->inp_socket->so_rcv,
1980 		    &rack->rc_inp->inp_socket->so_snd,
1981 		    RACK_DSACK_HANDLING, 0,
1982 		    0, &log, false, &tv);
1983 	}
1984 }
1985 
1986 static void
1987 rack_log_hdwr_pacing(struct tcp_rack *rack,
1988 		     uint64_t rate, uint64_t hw_rate, int line,
1989 		     int error, uint16_t mod)
1990 {
1991 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1992 		union tcp_log_stackspecific log;
1993 		struct timeval tv;
1994 		const struct ifnet *ifp;
1995 
1996 		memset(&log, 0, sizeof(log));
1997 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
1998 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
1999 		if (rack->r_ctl.crte) {
2000 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2001 		} else if (rack->rc_inp->inp_route.ro_nh &&
2002 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2003 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2004 		} else
2005 			ifp = NULL;
2006 		if (ifp) {
2007 			log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2008 			log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2009 		}
2010 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2011 		log.u_bbr.bw_inuse = rate;
2012 		log.u_bbr.flex5 = line;
2013 		log.u_bbr.flex6 = error;
2014 		log.u_bbr.flex7 = mod;
2015 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2016 		log.u_bbr.flex8 = rack->use_fixed_rate;
2017 		log.u_bbr.flex8 <<= 1;
2018 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2019 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2020 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2021 		if (rack->r_ctl.crte)
2022 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2023 		else
2024 			log.u_bbr.cur_del_rate = 0;
2025 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2026 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2027 		    &rack->rc_inp->inp_socket->so_rcv,
2028 		    &rack->rc_inp->inp_socket->so_snd,
2029 		    BBR_LOG_HDWR_PACE, 0,
2030 		    0, &log, false, &tv);
2031 	}
2032 }
2033 
2034 static uint64_t
2035 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2036 {
2037 	/*
2038 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2039 	 */
2040 	uint64_t bw_est, high_rate;
2041 	uint64_t gain;
2042 
2043 	gain = (uint64_t)rack_get_output_gain(rack, rsm);
2044 	bw_est = bw * gain;
2045 	bw_est /= (uint64_t)100;
2046 	/* Never fall below the minimum (def 64kbps) */
2047 	if (bw_est < RACK_MIN_BW)
2048 		bw_est = RACK_MIN_BW;
2049 	if (rack->r_rack_hw_rate_caps) {
2050 		/* Rate caps are in place */
2051 		if (rack->r_ctl.crte != NULL) {
2052 			/* We have a hdwr rate already */
2053 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2054 			if (bw_est >= high_rate) {
2055 				/* We are capping bw at the highest rate table entry */
2056 				rack_log_hdwr_pacing(rack,
2057 						     bw_est, high_rate, __LINE__,
2058 						     0, 3);
2059 				bw_est = high_rate;
2060 				if (capped)
2061 					*capped = 1;
2062 			}
2063 		} else if ((rack->rack_hdrw_pacing == 0) &&
2064 			   (rack->rack_hdw_pace_ena) &&
2065 			   (rack->rack_attempt_hdwr_pace == 0) &&
2066 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2067 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2068 			/*
2069 			 * Special case, we have not yet attempted hardware
2070 			 * pacing, and yet we may, when we do, find out if we are
2071 			 * above the highest rate. We need to know the maxbw for the interface
2072 			 * in question (if it supports ratelimiting). We get back
2073 			 * a 0, if the interface is not found in the RL lists.
2074 			 */
2075 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2076 			if (high_rate) {
2077 				/* Yep, we have a rate is it above this rate? */
2078 				if (bw_est > high_rate) {
2079 					bw_est = high_rate;
2080 					if (capped)
2081 						*capped = 1;
2082 				}
2083 			}
2084 		}
2085 	}
2086 	return (bw_est);
2087 }
2088 
2089 static void
2090 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2091 {
2092 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2093 		union tcp_log_stackspecific log;
2094 		struct timeval tv;
2095 
2096 		if ((mod != 1) && (rack_verbose_logging == 0)) {
2097 			/*
2098 			 * We get 3 values currently for mod
2099 			 * 1 - We are retransmitting and this tells the reason.
2100 			 * 2 - We are clearing a dup-ack count.
2101 			 * 3 - We are incrementing a dup-ack count.
2102 			 *
2103 			 * The clear/increment are only logged
2104 			 * if you have BBverbose on.
2105 			 */
2106 			return;
2107 		}
2108 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2109 		log.u_bbr.flex1 = tsused;
2110 		log.u_bbr.flex2 = thresh;
2111 		log.u_bbr.flex3 = rsm->r_flags;
2112 		log.u_bbr.flex4 = rsm->r_dupack;
2113 		log.u_bbr.flex5 = rsm->r_start;
2114 		log.u_bbr.flex6 = rsm->r_end;
2115 		log.u_bbr.flex8 = mod;
2116 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2117 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2118 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2119 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2120 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2121 		log.u_bbr.pacing_gain = rack->r_must_retran;
2122 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2123 		    &rack->rc_inp->inp_socket->so_rcv,
2124 		    &rack->rc_inp->inp_socket->so_snd,
2125 		    BBR_LOG_SETTINGS_CHG, 0,
2126 		    0, &log, false, &tv);
2127 	}
2128 }
2129 
2130 static void
2131 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2132 {
2133 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2134 		union tcp_log_stackspecific log;
2135 		struct timeval tv;
2136 
2137 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2138 		log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2139 		log.u_bbr.flex2 = to;
2140 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2141 		log.u_bbr.flex4 = slot;
2142 		log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2143 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2144 		log.u_bbr.flex7 = rack->rc_in_persist;
2145 		log.u_bbr.flex8 = which;
2146 		if (rack->rack_no_prr)
2147 			log.u_bbr.pkts_out = 0;
2148 		else
2149 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2150 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2151 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2152 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2153 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2154 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2155 		log.u_bbr.pacing_gain = rack->r_must_retran;
2156 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2157 		log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2158 		log.u_bbr.lost = rack_rto_min;
2159 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2160 		    &rack->rc_inp->inp_socket->so_rcv,
2161 		    &rack->rc_inp->inp_socket->so_snd,
2162 		    BBR_LOG_TIMERSTAR, 0,
2163 		    0, &log, false, &tv);
2164 	}
2165 }
2166 
2167 static void
2168 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2169 {
2170 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2171 		union tcp_log_stackspecific log;
2172 		struct timeval tv;
2173 
2174 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2175 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2176 		log.u_bbr.flex8 = to_num;
2177 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2178 		log.u_bbr.flex2 = rack->rc_rack_rtt;
2179 		if (rsm == NULL)
2180 			log.u_bbr.flex3 = 0;
2181 		else
2182 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2183 		if (rack->rack_no_prr)
2184 			log.u_bbr.flex5 = 0;
2185 		else
2186 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2187 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2188 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2189 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2190 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2191 		log.u_bbr.pacing_gain = rack->r_must_retran;
2192 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2193 		    &rack->rc_inp->inp_socket->so_rcv,
2194 		    &rack->rc_inp->inp_socket->so_snd,
2195 		    BBR_LOG_RTO, 0,
2196 		    0, &log, false, &tv);
2197 	}
2198 }
2199 
2200 static void
2201 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2202 		 struct rack_sendmap *prev,
2203 		 struct rack_sendmap *rsm,
2204 		 struct rack_sendmap *next,
2205 		 int flag, uint32_t th_ack, int line)
2206 {
2207 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2208 		union tcp_log_stackspecific log;
2209 		struct timeval tv;
2210 
2211 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2212 		log.u_bbr.flex8 = flag;
2213 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2214 		log.u_bbr.cur_del_rate = (uint64_t)prev;
2215 		log.u_bbr.delRate = (uint64_t)rsm;
2216 		log.u_bbr.rttProp = (uint64_t)next;
2217 		log.u_bbr.flex7 = 0;
2218 		if (prev) {
2219 			log.u_bbr.flex1 = prev->r_start;
2220 			log.u_bbr.flex2 = prev->r_end;
2221 			log.u_bbr.flex7 |= 0x4;
2222 		}
2223 		if (rsm) {
2224 			log.u_bbr.flex3 = rsm->r_start;
2225 			log.u_bbr.flex4 = rsm->r_end;
2226 			log.u_bbr.flex7 |= 0x2;
2227 		}
2228 		if (next) {
2229 			log.u_bbr.flex5 = next->r_start;
2230 			log.u_bbr.flex6 = next->r_end;
2231 			log.u_bbr.flex7 |= 0x1;
2232 		}
2233 		log.u_bbr.applimited = line;
2234 		log.u_bbr.pkts_out = th_ack;
2235 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2236 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2237 		if (rack->rack_no_prr)
2238 			log.u_bbr.lost = 0;
2239 		else
2240 			log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2241 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2242 		    &rack->rc_inp->inp_socket->so_rcv,
2243 		    &rack->rc_inp->inp_socket->so_snd,
2244 		    TCP_LOG_MAPCHG, 0,
2245 		    0, &log, false, &tv);
2246 	}
2247 }
2248 
2249 static void
2250 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2251 		 struct rack_sendmap *rsm, int conf)
2252 {
2253 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2254 		union tcp_log_stackspecific log;
2255 		struct timeval tv;
2256 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2257 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2258 		log.u_bbr.flex1 = t;
2259 		log.u_bbr.flex2 = len;
2260 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2261 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2262 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2263 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2264 		log.u_bbr.flex7 = conf;
2265 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2266 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2267 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2268 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2269 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2270 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2271 		if (rsm) {
2272 			log.u_bbr.pkt_epoch = rsm->r_start;
2273 			log.u_bbr.lost = rsm->r_end;
2274 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2275 			/* We loose any upper of the 24 bits */
2276 			log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2277 		} else {
2278 			/* Its a SYN */
2279 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2280 			log.u_bbr.lost = 0;
2281 			log.u_bbr.cwnd_gain = 0;
2282 			log.u_bbr.pacing_gain = 0;
2283 		}
2284 		/* Write out general bits of interest rrs here */
2285 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2286 		log.u_bbr.use_lt_bw <<= 1;
2287 		log.u_bbr.use_lt_bw |= rack->forced_ack;
2288 		log.u_bbr.use_lt_bw <<= 1;
2289 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2290 		log.u_bbr.use_lt_bw <<= 1;
2291 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2292 		log.u_bbr.use_lt_bw <<= 1;
2293 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2294 		log.u_bbr.use_lt_bw <<= 1;
2295 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2296 		log.u_bbr.use_lt_bw <<= 1;
2297 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2298 		log.u_bbr.use_lt_bw <<= 1;
2299 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2300 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2301 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2302 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2303 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2304 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2305 		log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2306 		log.u_bbr.bw_inuse <<= 32;
2307 		if (rsm)
2308 			log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2309 		TCP_LOG_EVENTP(tp, NULL,
2310 		    &rack->rc_inp->inp_socket->so_rcv,
2311 		    &rack->rc_inp->inp_socket->so_snd,
2312 		    BBR_LOG_BBRRTT, 0,
2313 		    0, &log, false, &tv);
2314 
2315 
2316 	}
2317 }
2318 
2319 static void
2320 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2321 {
2322 	/*
2323 	 * Log the rtt sample we are
2324 	 * applying to the srtt algorithm in
2325 	 * useconds.
2326 	 */
2327 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2328 		union tcp_log_stackspecific log;
2329 		struct timeval tv;
2330 
2331 		/* Convert our ms to a microsecond */
2332 		memset(&log, 0, sizeof(log));
2333 		log.u_bbr.flex1 = rtt;
2334 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2335 		log.u_bbr.flex3 = rack->r_ctl.sack_count;
2336 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2337 		log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2338 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2339 		log.u_bbr.flex7 = 1;
2340 		log.u_bbr.flex8 = rack->sack_attack_disable;
2341 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2342 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2343 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2344 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2345 		log.u_bbr.pacing_gain = rack->r_must_retran;
2346 		/*
2347 		 * We capture in delRate the upper 32 bits as
2348 		 * the confidence level we had declared, and the
2349 		 * lower 32 bits as the actual RTT using the arrival
2350 		 * timestamp.
2351 		 */
2352 		log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2353 		log.u_bbr.delRate <<= 32;
2354 		log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2355 		/* Lets capture all the things that make up t_rtxcur */
2356 		log.u_bbr.applimited = rack_rto_min;
2357 		log.u_bbr.epoch = rack_rto_max;
2358 		log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2359 		log.u_bbr.lost = rack_rto_min;
2360 		log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2361 		log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2362 		log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2363 		log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2364 		log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2365 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2366 		    &rack->rc_inp->inp_socket->so_rcv,
2367 		    &rack->rc_inp->inp_socket->so_snd,
2368 		    TCP_LOG_RTT, 0,
2369 		    0, &log, false, &tv);
2370 	}
2371 }
2372 
2373 static void
2374 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2375 {
2376 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2377 		union tcp_log_stackspecific log;
2378 		struct timeval tv;
2379 
2380 		/* Convert our ms to a microsecond */
2381 		memset(&log, 0, sizeof(log));
2382 		log.u_bbr.flex1 = rtt;
2383 		log.u_bbr.flex2 = send_time;
2384 		log.u_bbr.flex3 = ack_time;
2385 		log.u_bbr.flex4 = where;
2386 		log.u_bbr.flex7 = 2;
2387 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2388 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2389 		    &rack->rc_inp->inp_socket->so_rcv,
2390 		    &rack->rc_inp->inp_socket->so_snd,
2391 		    TCP_LOG_RTT, 0,
2392 		    0, &log, false, &tv);
2393 	}
2394 }
2395 
2396 
2397 
2398 static inline void
2399 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2400 {
2401 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2402 		union tcp_log_stackspecific log;
2403 		struct timeval tv;
2404 
2405 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2406 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2407 		log.u_bbr.flex1 = line;
2408 		log.u_bbr.flex2 = tick;
2409 		log.u_bbr.flex3 = tp->t_maxunacktime;
2410 		log.u_bbr.flex4 = tp->t_acktime;
2411 		log.u_bbr.flex8 = event;
2412 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2413 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2414 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2415 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2416 		log.u_bbr.pacing_gain = rack->r_must_retran;
2417 		TCP_LOG_EVENTP(tp, NULL,
2418 		    &rack->rc_inp->inp_socket->so_rcv,
2419 		    &rack->rc_inp->inp_socket->so_snd,
2420 		    BBR_LOG_PROGRESS, 0,
2421 		    0, &log, false, &tv);
2422 	}
2423 }
2424 
2425 static void
2426 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2427 {
2428 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2429 		union tcp_log_stackspecific log;
2430 
2431 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2432 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2433 		log.u_bbr.flex1 = slot;
2434 		if (rack->rack_no_prr)
2435 			log.u_bbr.flex2 = 0;
2436 		else
2437 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2438 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2439 		log.u_bbr.flex8 = rack->rc_in_persist;
2440 		log.u_bbr.timeStamp = cts;
2441 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2442 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2443 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2444 		log.u_bbr.pacing_gain = rack->r_must_retran;
2445 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2446 		    &rack->rc_inp->inp_socket->so_rcv,
2447 		    &rack->rc_inp->inp_socket->so_snd,
2448 		    BBR_LOG_BBRSND, 0,
2449 		    0, &log, false, tv);
2450 	}
2451 }
2452 
2453 static void
2454 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2455 {
2456 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2457 		union tcp_log_stackspecific log;
2458 		struct timeval tv;
2459 
2460 		memset(&log, 0, sizeof(log));
2461 		log.u_bbr.flex1 = did_out;
2462 		log.u_bbr.flex2 = nxt_pkt;
2463 		log.u_bbr.flex3 = way_out;
2464 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2465 		if (rack->rack_no_prr)
2466 			log.u_bbr.flex5 = 0;
2467 		else
2468 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2469 		log.u_bbr.flex6 = nsegs;
2470 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2471 		log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;	/* Do we have ack-can-send set */
2472 		log.u_bbr.flex7 <<= 1;
2473 		log.u_bbr.flex7 |= rack->r_fast_output;	/* is fast output primed */
2474 		log.u_bbr.flex7 <<= 1;
2475 		log.u_bbr.flex7 |= rack->r_wanted_output;	/* Do we want output */
2476 		log.u_bbr.flex8 = rack->rc_in_persist;
2477 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2478 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2479 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2480 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2481 		log.u_bbr.use_lt_bw <<= 1;
2482 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
2483 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2484 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2485 		log.u_bbr.pacing_gain = rack->r_must_retran;
2486 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2487 		    &rack->rc_inp->inp_socket->so_rcv,
2488 		    &rack->rc_inp->inp_socket->so_snd,
2489 		    BBR_LOG_DOSEG_DONE, 0,
2490 		    0, &log, false, &tv);
2491 	}
2492 }
2493 
2494 static void
2495 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2496 {
2497 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2498 		union tcp_log_stackspecific log;
2499 		struct timeval tv;
2500 
2501 		memset(&log, 0, sizeof(log));
2502 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2503 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2504 		log.u_bbr.flex4 = arg1;
2505 		log.u_bbr.flex5 = arg2;
2506 		log.u_bbr.flex6 = arg3;
2507 		log.u_bbr.flex8 = frm;
2508 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2509 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2510 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2511 		log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2512 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2513 		log.u_bbr.pacing_gain = rack->r_must_retran;
2514 		TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
2515 		    &tptosocket(tp)->so_snd,
2516 		    TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
2517 	}
2518 }
2519 
2520 static void
2521 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2522 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2523 {
2524 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2525 		union tcp_log_stackspecific log;
2526 		struct timeval tv;
2527 
2528 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2529 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2530 		log.u_bbr.flex1 = slot;
2531 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2532 		log.u_bbr.flex4 = reason;
2533 		if (rack->rack_no_prr)
2534 			log.u_bbr.flex5 = 0;
2535 		else
2536 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2537 		log.u_bbr.flex7 = hpts_calling;
2538 		log.u_bbr.flex8 = rack->rc_in_persist;
2539 		log.u_bbr.lt_epoch = cwnd_to_use;
2540 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2541 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2542 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2543 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2544 		log.u_bbr.pacing_gain = rack->r_must_retran;
2545 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2546 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2547 		    &rack->rc_inp->inp_socket->so_rcv,
2548 		    &rack->rc_inp->inp_socket->so_snd,
2549 		    BBR_LOG_JUSTRET, 0,
2550 		    tlen, &log, false, &tv);
2551 	}
2552 }
2553 
2554 static void
2555 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2556 		   struct timeval *tv, uint32_t flags_on_entry)
2557 {
2558 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2559 		union tcp_log_stackspecific log;
2560 
2561 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2562 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2563 		log.u_bbr.flex1 = line;
2564 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2565 		log.u_bbr.flex3 = flags_on_entry;
2566 		log.u_bbr.flex4 = us_cts;
2567 		if (rack->rack_no_prr)
2568 			log.u_bbr.flex5 = 0;
2569 		else
2570 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2571 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2572 		log.u_bbr.flex7 = hpts_removed;
2573 		log.u_bbr.flex8 = 1;
2574 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2575 		log.u_bbr.timeStamp = us_cts;
2576 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2577 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2578 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2579 		log.u_bbr.pacing_gain = rack->r_must_retran;
2580 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2581 		    &rack->rc_inp->inp_socket->so_rcv,
2582 		    &rack->rc_inp->inp_socket->so_snd,
2583 		    BBR_LOG_TIMERCANC, 0,
2584 		    0, &log, false, tv);
2585 	}
2586 }
2587 
2588 static void
2589 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2590 			  uint32_t flex1, uint32_t flex2,
2591 			  uint32_t flex3, uint32_t flex4,
2592 			  uint32_t flex5, uint32_t flex6,
2593 			  uint16_t flex7, uint8_t mod)
2594 {
2595 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2596 		union tcp_log_stackspecific log;
2597 		struct timeval tv;
2598 
2599 		if (mod == 1) {
2600 			/* No you can't use 1, its for the real to cancel */
2601 			return;
2602 		}
2603 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2604 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2605 		log.u_bbr.flex1 = flex1;
2606 		log.u_bbr.flex2 = flex2;
2607 		log.u_bbr.flex3 = flex3;
2608 		log.u_bbr.flex4 = flex4;
2609 		log.u_bbr.flex5 = flex5;
2610 		log.u_bbr.flex6 = flex6;
2611 		log.u_bbr.flex7 = flex7;
2612 		log.u_bbr.flex8 = mod;
2613 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2614 		    &rack->rc_inp->inp_socket->so_rcv,
2615 		    &rack->rc_inp->inp_socket->so_snd,
2616 		    BBR_LOG_TIMERCANC, 0,
2617 		    0, &log, false, &tv);
2618 	}
2619 }
2620 
2621 static void
2622 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2623 {
2624 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2625 		union tcp_log_stackspecific log;
2626 		struct timeval tv;
2627 
2628 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2629 		log.u_bbr.flex1 = timers;
2630 		log.u_bbr.flex2 = ret;
2631 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2632 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2633 		log.u_bbr.flex5 = cts;
2634 		if (rack->rack_no_prr)
2635 			log.u_bbr.flex6 = 0;
2636 		else
2637 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2638 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2639 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2640 		log.u_bbr.pacing_gain = rack->r_must_retran;
2641 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2642 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2643 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2644 		    &rack->rc_inp->inp_socket->so_rcv,
2645 		    &rack->rc_inp->inp_socket->so_snd,
2646 		    BBR_LOG_TO_PROCESS, 0,
2647 		    0, &log, false, &tv);
2648 	}
2649 }
2650 
2651 static void
2652 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
2653 {
2654 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2655 		union tcp_log_stackspecific log;
2656 		struct timeval tv;
2657 
2658 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2659 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2660 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2661 		if (rack->rack_no_prr)
2662 			log.u_bbr.flex3 = 0;
2663 		else
2664 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2665 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2666 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2667 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2668 		log.u_bbr.flex7 = line;
2669 		log.u_bbr.flex8 = frm;
2670 		log.u_bbr.pkts_out = orig_cwnd;
2671 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2672 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2673 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2674 		log.u_bbr.use_lt_bw <<= 1;
2675 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
2676 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2677 		    &rack->rc_inp->inp_socket->so_rcv,
2678 		    &rack->rc_inp->inp_socket->so_snd,
2679 		    BBR_LOG_BBRUPD, 0,
2680 		    0, &log, false, &tv);
2681 	}
2682 }
2683 
2684 #ifdef NETFLIX_EXP_DETECTION
2685 static void
2686 rack_log_sad(struct tcp_rack *rack, int event)
2687 {
2688 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2689 		union tcp_log_stackspecific log;
2690 		struct timeval tv;
2691 
2692 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2693 		log.u_bbr.flex1 = rack->r_ctl.sack_count;
2694 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2695 		log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2696 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2697 		log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2698 		log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2699 		log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2700 		log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2701 		log.u_bbr.lt_epoch |= rack->do_detection;
2702 		log.u_bbr.applimited = tcp_map_minimum;
2703 		log.u_bbr.flex7 = rack->sack_attack_disable;
2704 		log.u_bbr.flex8 = event;
2705 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2706 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2707 		log.u_bbr.delivered = tcp_sad_decay_val;
2708 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2709 		    &rack->rc_inp->inp_socket->so_rcv,
2710 		    &rack->rc_inp->inp_socket->so_snd,
2711 		    TCP_SAD_DETECTION, 0,
2712 		    0, &log, false, &tv);
2713 	}
2714 }
2715 #endif
2716 
2717 static void
2718 rack_counter_destroy(void)
2719 {
2720 	counter_u64_free(rack_fto_send);
2721 	counter_u64_free(rack_fto_rsm_send);
2722 	counter_u64_free(rack_nfto_resend);
2723 	counter_u64_free(rack_hw_pace_init_fail);
2724 	counter_u64_free(rack_hw_pace_lost);
2725 	counter_u64_free(rack_non_fto_send);
2726 	counter_u64_free(rack_extended_rfo);
2727 	counter_u64_free(rack_ack_total);
2728 	counter_u64_free(rack_express_sack);
2729 	counter_u64_free(rack_sack_total);
2730 	counter_u64_free(rack_move_none);
2731 	counter_u64_free(rack_move_some);
2732 	counter_u64_free(rack_sack_attacks_detected);
2733 	counter_u64_free(rack_sack_attacks_reversed);
2734 	counter_u64_free(rack_sack_used_next_merge);
2735 	counter_u64_free(rack_sack_used_prev_merge);
2736 	counter_u64_free(rack_tlp_tot);
2737 	counter_u64_free(rack_tlp_newdata);
2738 	counter_u64_free(rack_tlp_retran);
2739 	counter_u64_free(rack_tlp_retran_bytes);
2740 	counter_u64_free(rack_to_tot);
2741 	counter_u64_free(rack_saw_enobuf);
2742 	counter_u64_free(rack_saw_enobuf_hw);
2743 	counter_u64_free(rack_saw_enetunreach);
2744 	counter_u64_free(rack_hot_alloc);
2745 	counter_u64_free(rack_to_alloc);
2746 	counter_u64_free(rack_to_alloc_hard);
2747 	counter_u64_free(rack_to_alloc_emerg);
2748 	counter_u64_free(rack_to_alloc_limited);
2749 	counter_u64_free(rack_alloc_limited_conns);
2750 	counter_u64_free(rack_split_limited);
2751 	counter_u64_free(rack_multi_single_eq);
2752 	counter_u64_free(rack_proc_non_comp_ack);
2753 	counter_u64_free(rack_sack_proc_all);
2754 	counter_u64_free(rack_sack_proc_restart);
2755 	counter_u64_free(rack_sack_proc_short);
2756 	counter_u64_free(rack_sack_skipped_acked);
2757 	counter_u64_free(rack_sack_splits);
2758 	counter_u64_free(rack_input_idle_reduces);
2759 	counter_u64_free(rack_collapsed_win);
2760 	counter_u64_free(rack_collapsed_win_rxt);
2761 	counter_u64_free(rack_collapsed_win_rxt_bytes);
2762 	counter_u64_free(rack_collapsed_win_seen);
2763 	counter_u64_free(rack_try_scwnd);
2764 	counter_u64_free(rack_persists_sends);
2765 	counter_u64_free(rack_persists_acks);
2766 	counter_u64_free(rack_persists_loss);
2767 	counter_u64_free(rack_persists_lost_ends);
2768 #ifdef INVARIANTS
2769 	counter_u64_free(rack_adjust_map_bw);
2770 #endif
2771 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2772 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2773 }
2774 
2775 static struct rack_sendmap *
2776 rack_alloc(struct tcp_rack *rack)
2777 {
2778 	struct rack_sendmap *rsm;
2779 
2780 	/*
2781 	 * First get the top of the list it in
2782 	 * theory is the "hottest" rsm we have,
2783 	 * possibly just freed by ack processing.
2784 	 */
2785 	if (rack->rc_free_cnt > rack_free_cache) {
2786 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2787 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2788 		counter_u64_add(rack_hot_alloc, 1);
2789 		rack->rc_free_cnt--;
2790 		return (rsm);
2791 	}
2792 	/*
2793 	 * Once we get under our free cache we probably
2794 	 * no longer have a "hot" one available. Lets
2795 	 * get one from UMA.
2796 	 */
2797 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
2798 	if (rsm) {
2799 		rack->r_ctl.rc_num_maps_alloced++;
2800 		counter_u64_add(rack_to_alloc, 1);
2801 		return (rsm);
2802 	}
2803 	/*
2804 	 * Dig in to our aux rsm's (the last two) since
2805 	 * UMA failed to get us one.
2806 	 */
2807 	if (rack->rc_free_cnt) {
2808 		counter_u64_add(rack_to_alloc_emerg, 1);
2809 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2810 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2811 		rack->rc_free_cnt--;
2812 		return (rsm);
2813 	}
2814 	return (NULL);
2815 }
2816 
2817 static struct rack_sendmap *
2818 rack_alloc_full_limit(struct tcp_rack *rack)
2819 {
2820 	if ((V_tcp_map_entries_limit > 0) &&
2821 	    (rack->do_detection == 0) &&
2822 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2823 		counter_u64_add(rack_to_alloc_limited, 1);
2824 		if (!rack->alloc_limit_reported) {
2825 			rack->alloc_limit_reported = 1;
2826 			counter_u64_add(rack_alloc_limited_conns, 1);
2827 		}
2828 		return (NULL);
2829 	}
2830 	return (rack_alloc(rack));
2831 }
2832 
2833 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2834 static struct rack_sendmap *
2835 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2836 {
2837 	struct rack_sendmap *rsm;
2838 
2839 	if (limit_type) {
2840 		/* currently there is only one limit type */
2841 		if (V_tcp_map_split_limit > 0 &&
2842 		    (rack->do_detection == 0) &&
2843 		    rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2844 			counter_u64_add(rack_split_limited, 1);
2845 			if (!rack->alloc_limit_reported) {
2846 				rack->alloc_limit_reported = 1;
2847 				counter_u64_add(rack_alloc_limited_conns, 1);
2848 			}
2849 			return (NULL);
2850 		}
2851 	}
2852 
2853 	/* allocate and mark in the limit type, if set */
2854 	rsm = rack_alloc(rack);
2855 	if (rsm != NULL && limit_type) {
2856 		rsm->r_limit_type = limit_type;
2857 		rack->r_ctl.rc_num_split_allocs++;
2858 	}
2859 	return (rsm);
2860 }
2861 
2862 static void
2863 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2864 {
2865 	if (rsm->r_flags & RACK_APP_LIMITED) {
2866 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
2867 			rack->r_ctl.rc_app_limited_cnt--;
2868 		}
2869 	}
2870 	if (rsm->r_limit_type) {
2871 		/* currently there is only one limit type */
2872 		rack->r_ctl.rc_num_split_allocs--;
2873 	}
2874 	if (rsm == rack->r_ctl.rc_first_appl) {
2875 		if (rack->r_ctl.rc_app_limited_cnt == 0)
2876 			rack->r_ctl.rc_first_appl = NULL;
2877 		else {
2878 			/* Follow the next one out */
2879 			struct rack_sendmap fe;
2880 
2881 			fe.r_start = rsm->r_nseq_appl;
2882 			rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
2883 		}
2884 	}
2885 	if (rsm == rack->r_ctl.rc_resend)
2886 		rack->r_ctl.rc_resend = NULL;
2887 	if (rsm == rack->r_ctl.rc_end_appl)
2888 		rack->r_ctl.rc_end_appl = NULL;
2889 	if (rack->r_ctl.rc_tlpsend == rsm)
2890 		rack->r_ctl.rc_tlpsend = NULL;
2891 	if (rack->r_ctl.rc_sacklast == rsm)
2892 		rack->r_ctl.rc_sacklast = NULL;
2893 	memset(rsm, 0, sizeof(struct rack_sendmap));
2894 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
2895 	rack->rc_free_cnt++;
2896 }
2897 
2898 static void
2899 rack_free_trim(struct tcp_rack *rack)
2900 {
2901 	struct rack_sendmap *rsm;
2902 
2903 	/*
2904 	 * Free up all the tail entries until
2905 	 * we get our list down to the limit.
2906 	 */
2907 	while (rack->rc_free_cnt > rack_free_cache) {
2908 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
2909 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2910 		rack->rc_free_cnt--;
2911 		uma_zfree(rack_zone, rsm);
2912 	}
2913 }
2914 
2915 
2916 static uint32_t
2917 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
2918 {
2919 	uint64_t srtt, bw, len, tim;
2920 	uint32_t segsiz, def_len, minl;
2921 
2922 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2923 	def_len = rack_def_data_window * segsiz;
2924 	if (rack->rc_gp_filled == 0) {
2925 		/*
2926 		 * We have no measurement (IW is in flight?) so
2927 		 * we can only guess using our data_window sysctl
2928 		 * value (usually 20MSS).
2929 		 */
2930 		return (def_len);
2931 	}
2932 	/*
2933 	 * Now we have a number of factors to consider.
2934 	 *
2935 	 * 1) We have a desired BDP which is usually
2936 	 *    at least 2.
2937 	 * 2) We have a minimum number of rtt's usually 1 SRTT
2938 	 *    but we allow it too to be more.
2939 	 * 3) We want to make sure a measurement last N useconds (if
2940 	 *    we have set rack_min_measure_usec.
2941 	 *
2942 	 * We handle the first concern here by trying to create a data
2943 	 * window of max(rack_def_data_window, DesiredBDP). The
2944 	 * second concern we handle in not letting the measurement
2945 	 * window end normally until at least the required SRTT's
2946 	 * have gone by which is done further below in
2947 	 * rack_enough_for_measurement(). Finally the third concern
2948 	 * we also handle here by calculating how long that time
2949 	 * would take at the current BW and then return the
2950 	 * max of our first calculation and that length. Note
2951 	 * that if rack_min_measure_usec is 0, we don't deal
2952 	 * with concern 3. Also for both Concern 1 and 3 an
2953 	 * application limited period could end the measurement
2954 	 * earlier.
2955 	 *
2956 	 * So lets calculate the BDP with the "known" b/w using
2957 	 * the SRTT has our rtt and then multiply it by the
2958 	 * goal.
2959 	 */
2960 	bw = rack_get_bw(rack);
2961 	srtt = (uint64_t)tp->t_srtt;
2962 	len = bw * srtt;
2963 	len /= (uint64_t)HPTS_USEC_IN_SEC;
2964 	len *= max(1, rack_goal_bdp);
2965 	/* Now we need to round up to the nearest MSS */
2966 	len = roundup(len, segsiz);
2967 	if (rack_min_measure_usec) {
2968 		/* Now calculate our min length for this b/w */
2969 		tim = rack_min_measure_usec;
2970 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
2971 		if (minl == 0)
2972 			minl = 1;
2973 		minl = roundup(minl, segsiz);
2974 		if (len < minl)
2975 			len = minl;
2976 	}
2977 	/*
2978 	 * Now if we have a very small window we want
2979 	 * to attempt to get the window that is
2980 	 * as small as possible. This happens on
2981 	 * low b/w connections and we don't want to
2982 	 * span huge numbers of rtt's between measurements.
2983 	 *
2984 	 * We basically include 2 over our "MIN window" so
2985 	 * that the measurement can be shortened (possibly) by
2986 	 * an ack'ed packet.
2987 	 */
2988 	if (len < def_len)
2989 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
2990 	else
2991 		return (max((uint32_t)len, def_len));
2992 
2993 }
2994 
2995 static int
2996 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
2997 {
2998 	uint32_t tim, srtts, segsiz;
2999 
3000 	/*
3001 	 * Has enough time passed for the GP measurement to be valid?
3002 	 */
3003 	if ((tp->snd_max == tp->snd_una) ||
3004 	    (th_ack == tp->snd_max)){
3005 		/* All is acked */
3006 		*quality = RACK_QUALITY_ALLACKED;
3007 		return (1);
3008 	}
3009 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3010 		/* Not enough bytes yet */
3011 		return (0);
3012 	}
3013 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3014 	if (SEQ_LT(th_ack, tp->gput_ack) &&
3015 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3016 		/* Not enough bytes yet */
3017 		return (0);
3018 	}
3019 	if (rack->r_ctl.rc_first_appl &&
3020 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3021 		/*
3022 		 * We are up to the app limited send point
3023 		 * we have to measure irrespective of the time..
3024 		 */
3025 		*quality = RACK_QUALITY_APPLIMITED;
3026 		return (1);
3027 	}
3028 	/* Now what about time? */
3029 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3030 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3031 	if (tim >= srtts) {
3032 		*quality = RACK_QUALITY_HIGH;
3033 		return (1);
3034 	}
3035 	/* Nope not even a full SRTT has passed */
3036 	return (0);
3037 }
3038 
3039 static void
3040 rack_log_timely(struct tcp_rack *rack,
3041 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3042 		uint64_t up_bnd, int line, uint8_t method)
3043 {
3044 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3045 		union tcp_log_stackspecific log;
3046 		struct timeval tv;
3047 
3048 		memset(&log, 0, sizeof(log));
3049 		log.u_bbr.flex1 = logged;
3050 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3051 		log.u_bbr.flex2 <<= 4;
3052 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3053 		log.u_bbr.flex2 <<= 4;
3054 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3055 		log.u_bbr.flex2 <<= 4;
3056 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3057 		log.u_bbr.flex3 = rack->rc_gp_incr;
3058 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3059 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3060 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3061 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3062 		log.u_bbr.flex8 = method;
3063 		log.u_bbr.cur_del_rate = cur_bw;
3064 		log.u_bbr.delRate = low_bnd;
3065 		log.u_bbr.bw_inuse = up_bnd;
3066 		log.u_bbr.rttProp = rack_get_bw(rack);
3067 		log.u_bbr.pkt_epoch = line;
3068 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3069 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3070 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3071 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3072 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3073 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3074 		log.u_bbr.cwnd_gain <<= 1;
3075 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3076 		log.u_bbr.cwnd_gain <<= 1;
3077 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3078 		log.u_bbr.cwnd_gain <<= 1;
3079 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3080 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3081 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3082 		    &rack->rc_inp->inp_socket->so_rcv,
3083 		    &rack->rc_inp->inp_socket->so_snd,
3084 		    TCP_TIMELY_WORK, 0,
3085 		    0, &log, false, &tv);
3086 	}
3087 }
3088 
3089 static int
3090 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3091 {
3092 	/*
3093 	 * Before we increase we need to know if
3094 	 * the estimate just made was less than
3095 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3096 	 *
3097 	 * If we already are pacing at a fast enough
3098 	 * rate to push us faster there is no sense of
3099 	 * increasing.
3100 	 *
3101 	 * We first caculate our actual pacing rate (ss or ca multiplier
3102 	 * times our cur_bw).
3103 	 *
3104 	 * Then we take the last measured rate and multipy by our
3105 	 * maximum pacing overage to give us a max allowable rate.
3106 	 *
3107 	 * If our act_rate is smaller than our max_allowable rate
3108 	 * then we should increase. Else we should hold steady.
3109 	 *
3110 	 */
3111 	uint64_t act_rate, max_allow_rate;
3112 
3113 	if (rack_timely_no_stopping)
3114 		return (1);
3115 
3116 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3117 		/*
3118 		 * Initial startup case or
3119 		 * everything is acked case.
3120 		 */
3121 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3122 				__LINE__, 9);
3123 		return (1);
3124 	}
3125 	if (mult <= 100) {
3126 		/*
3127 		 * We can always pace at or slightly above our rate.
3128 		 */
3129 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3130 				__LINE__, 9);
3131 		return (1);
3132 	}
3133 	act_rate = cur_bw * (uint64_t)mult;
3134 	act_rate /= 100;
3135 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3136 	max_allow_rate /= 100;
3137 	if (act_rate < max_allow_rate) {
3138 		/*
3139 		 * Here the rate we are actually pacing at
3140 		 * is smaller than 10% above our last measurement.
3141 		 * This means we are pacing below what we would
3142 		 * like to try to achieve (plus some wiggle room).
3143 		 */
3144 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3145 				__LINE__, 9);
3146 		return (1);
3147 	} else {
3148 		/*
3149 		 * Here we are already pacing at least rack_max_per_above(10%)
3150 		 * what we are getting back. This indicates most likely
3151 		 * that we are being limited (cwnd/rwnd/app) and can't
3152 		 * get any more b/w. There is no sense of trying to
3153 		 * raise up the pacing rate its not speeding us up
3154 		 * and we already are pacing faster than we are getting.
3155 		 */
3156 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3157 				__LINE__, 8);
3158 		return (0);
3159 	}
3160 }
3161 
3162 static void
3163 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3164 {
3165 	/*
3166 	 * When we drag bottom, we want to assure
3167 	 * that no multiplier is below 1.0, if so
3168 	 * we want to restore it to at least that.
3169 	 */
3170 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3171 		/* This is unlikely we usually do not touch recovery */
3172 		rack->r_ctl.rack_per_of_gp_rec = 100;
3173 	}
3174 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3175 		rack->r_ctl.rack_per_of_gp_ca = 100;
3176 	}
3177 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3178 		rack->r_ctl.rack_per_of_gp_ss = 100;
3179 	}
3180 }
3181 
3182 static void
3183 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3184 {
3185 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3186 		rack->r_ctl.rack_per_of_gp_ca = 100;
3187 	}
3188 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3189 		rack->r_ctl.rack_per_of_gp_ss = 100;
3190 	}
3191 }
3192 
3193 static void
3194 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3195 {
3196 	int32_t  calc, logged, plus;
3197 
3198 	logged = 0;
3199 
3200 	if (override) {
3201 		/*
3202 		 * override is passed when we are
3203 		 * loosing b/w and making one last
3204 		 * gasp at trying to not loose out
3205 		 * to a new-reno flow.
3206 		 */
3207 		goto extra_boost;
3208 	}
3209 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3210 	if (rack->rc_gp_incr &&
3211 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3212 		/*
3213 		 * Reset and get 5 strokes more before the boost. Note
3214 		 * that the count is 0 based so we have to add one.
3215 		 */
3216 extra_boost:
3217 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3218 		rack->rc_gp_timely_inc_cnt = 0;
3219 	} else
3220 		plus = (uint32_t)rack_gp_increase_per;
3221 	/* Must be at least 1% increase for true timely increases */
3222 	if ((plus < 1) &&
3223 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3224 		plus = 1;
3225 	if (rack->rc_gp_saw_rec &&
3226 	    (rack->rc_gp_no_rec_chg == 0) &&
3227 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3228 				  rack->r_ctl.rack_per_of_gp_rec)) {
3229 		/* We have been in recovery ding it too */
3230 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3231 		if (calc > 0xffff)
3232 			calc = 0xffff;
3233 		logged |= 1;
3234 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3235 		if (rack_per_upper_bound_ss &&
3236 		    (rack->rc_dragged_bottom == 0) &&
3237 		    (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3238 			rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3239 	}
3240 	if (rack->rc_gp_saw_ca &&
3241 	    (rack->rc_gp_saw_ss == 0) &&
3242 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3243 				  rack->r_ctl.rack_per_of_gp_ca)) {
3244 		/* In CA */
3245 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3246 		if (calc > 0xffff)
3247 			calc = 0xffff;
3248 		logged |= 2;
3249 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3250 		if (rack_per_upper_bound_ca &&
3251 		    (rack->rc_dragged_bottom == 0) &&
3252 		    (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3253 			rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3254 	}
3255 	if (rack->rc_gp_saw_ss &&
3256 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3257 				  rack->r_ctl.rack_per_of_gp_ss)) {
3258 		/* In SS */
3259 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3260 		if (calc > 0xffff)
3261 			calc = 0xffff;
3262 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3263 		if (rack_per_upper_bound_ss &&
3264 		    (rack->rc_dragged_bottom == 0) &&
3265 		    (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3266 			rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3267 		logged |= 4;
3268 	}
3269 	if (logged &&
3270 	    (rack->rc_gp_incr == 0)){
3271 		/* Go into increment mode */
3272 		rack->rc_gp_incr = 1;
3273 		rack->rc_gp_timely_inc_cnt = 0;
3274 	}
3275 	if (rack->rc_gp_incr &&
3276 	    logged &&
3277 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3278 		rack->rc_gp_timely_inc_cnt++;
3279 	}
3280 	rack_log_timely(rack,  logged, plus, 0, 0,
3281 			__LINE__, 1);
3282 }
3283 
3284 static uint32_t
3285 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3286 {
3287 	/*
3288 	 * norm_grad = rtt_diff / minrtt;
3289 	 * new_per = curper * (1 - B * norm_grad)
3290 	 *
3291 	 * B = rack_gp_decrease_per (default 10%)
3292 	 * rtt_dif = input var current rtt-diff
3293 	 * curper = input var current percentage
3294 	 * minrtt = from rack filter
3295 	 *
3296 	 */
3297 	uint64_t perf;
3298 
3299 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3300 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3301 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
3302 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3303 		     (uint64_t)1000000)) /
3304 		(uint64_t)1000000);
3305 	if (perf > curper) {
3306 		/* TSNH */
3307 		perf = curper - 1;
3308 	}
3309 	return ((uint32_t)perf);
3310 }
3311 
3312 static uint32_t
3313 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3314 {
3315 	/*
3316 	 *                                   highrttthresh
3317 	 * result = curper * (1 - (B * ( 1 -  ------          ))
3318 	 *                                     gp_srtt
3319 	 *
3320 	 * B = rack_gp_decrease_per (default 10%)
3321 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3322 	 */
3323 	uint64_t perf;
3324 	uint32_t highrttthresh;
3325 
3326 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3327 
3328 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3329 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3330 					((uint64_t)highrttthresh * (uint64_t)1000000) /
3331 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3332 	return (perf);
3333 }
3334 
3335 static void
3336 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3337 {
3338 	uint64_t logvar, logvar2, logvar3;
3339 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3340 
3341 	if (rack->rc_gp_incr) {
3342 		/* Turn off increment counting */
3343 		rack->rc_gp_incr = 0;
3344 		rack->rc_gp_timely_inc_cnt = 0;
3345 	}
3346 	ss_red = ca_red = rec_red = 0;
3347 	logged = 0;
3348 	/* Calculate the reduction value */
3349 	if (rtt_diff < 0) {
3350 		rtt_diff *= -1;
3351 	}
3352 	/* Must be at least 1% reduction */
3353 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3354 		/* We have been in recovery ding it too */
3355 		if (timely_says == 2) {
3356 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3357 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3358 			if (alt < new_per)
3359 				val = alt;
3360 			else
3361 				val = new_per;
3362 		} else
3363 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3364 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
3365 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3366 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3367 		} else {
3368 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3369 			rec_red = 0;
3370 		}
3371 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3372 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3373 		logged |= 1;
3374 	}
3375 	if (rack->rc_gp_saw_ss) {
3376 		/* Sent in SS */
3377 		if (timely_says == 2) {
3378 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3379 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3380 			if (alt < new_per)
3381 				val = alt;
3382 			else
3383 				val = new_per;
3384 		} else
3385 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3386 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3387 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3388 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3389 		} else {
3390 			ss_red = new_per;
3391 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3392 			logvar = new_per;
3393 			logvar <<= 32;
3394 			logvar |= alt;
3395 			logvar2 = (uint32_t)rtt;
3396 			logvar2 <<= 32;
3397 			logvar2 |= (uint32_t)rtt_diff;
3398 			logvar3 = rack_gp_rtt_maxmul;
3399 			logvar3 <<= 32;
3400 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3401 			rack_log_timely(rack, timely_says,
3402 					logvar2, logvar3,
3403 					logvar, __LINE__, 10);
3404 		}
3405 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3406 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3407 		logged |= 4;
3408 	} else if (rack->rc_gp_saw_ca) {
3409 		/* Sent in CA */
3410 		if (timely_says == 2) {
3411 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3412 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3413 			if (alt < new_per)
3414 				val = alt;
3415 			else
3416 				val = new_per;
3417 		} else
3418 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3419 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
3420 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3421 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3422 		} else {
3423 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3424 			ca_red = 0;
3425 			logvar = new_per;
3426 			logvar <<= 32;
3427 			logvar |= alt;
3428 			logvar2 = (uint32_t)rtt;
3429 			logvar2 <<= 32;
3430 			logvar2 |= (uint32_t)rtt_diff;
3431 			logvar3 = rack_gp_rtt_maxmul;
3432 			logvar3 <<= 32;
3433 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3434 			rack_log_timely(rack, timely_says,
3435 					logvar2, logvar3,
3436 					logvar, __LINE__, 10);
3437 		}
3438 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3439 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3440 		logged |= 2;
3441 	}
3442 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
3443 		rack->rc_gp_timely_dec_cnt++;
3444 		if (rack_timely_dec_clear &&
3445 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3446 			rack->rc_gp_timely_dec_cnt = 0;
3447 	}
3448 	logvar = ss_red;
3449 	logvar <<= 32;
3450 	logvar |= ca_red;
3451 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3452 			__LINE__, 2);
3453 }
3454 
3455 static void
3456 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3457 		     uint32_t rtt, uint32_t line, uint8_t reas)
3458 {
3459 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3460 		union tcp_log_stackspecific log;
3461 		struct timeval tv;
3462 
3463 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3464 		log.u_bbr.flex1 = line;
3465 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3466 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3467 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3468 		log.u_bbr.flex5 = rtt;
3469 		log.u_bbr.flex6 = rack->rc_highly_buffered;
3470 		log.u_bbr.flex6 <<= 1;
3471 		log.u_bbr.flex6 |= rack->forced_ack;
3472 		log.u_bbr.flex6 <<= 1;
3473 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3474 		log.u_bbr.flex6 <<= 1;
3475 		log.u_bbr.flex6 |= rack->in_probe_rtt;
3476 		log.u_bbr.flex6 <<= 1;
3477 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3478 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3479 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3480 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3481 		log.u_bbr.flex8 = reas;
3482 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3483 		log.u_bbr.delRate = rack_get_bw(rack);
3484 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3485 		log.u_bbr.cur_del_rate <<= 32;
3486 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3487 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3488 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3489 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3490 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3491 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3492 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3493 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3494 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3495 		log.u_bbr.rttProp = us_cts;
3496 		log.u_bbr.rttProp <<= 32;
3497 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3498 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3499 		    &rack->rc_inp->inp_socket->so_rcv,
3500 		    &rack->rc_inp->inp_socket->so_snd,
3501 		    BBR_LOG_RTT_SHRINKS, 0,
3502 		    0, &log, false, &rack->r_ctl.act_rcv_time);
3503 	}
3504 }
3505 
3506 static void
3507 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3508 {
3509 	uint64_t bwdp;
3510 
3511 	bwdp = rack_get_bw(rack);
3512 	bwdp *= (uint64_t)rtt;
3513 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3514 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3515 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3516 		/*
3517 		 * A window protocol must be able to have 4 packets
3518 		 * outstanding as the floor in order to function
3519 		 * (especially considering delayed ack :D).
3520 		 */
3521 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3522 	}
3523 }
3524 
3525 static void
3526 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3527 {
3528 	/**
3529 	 * ProbeRTT is a bit different in rack_pacing than in
3530 	 * BBR. It is like BBR in that it uses the lowering of
3531 	 * the RTT as a signal that we saw something new and
3532 	 * counts from there for how long between. But it is
3533 	 * different in that its quite simple. It does not
3534 	 * play with the cwnd and wait until we get down
3535 	 * to N segments outstanding and hold that for
3536 	 * 200ms. Instead it just sets the pacing reduction
3537 	 * rate to a set percentage (70 by default) and hold
3538 	 * that for a number of recent GP Srtt's.
3539 	 */
3540 	uint32_t segsiz;
3541 
3542 	if (rack->rc_gp_dyn_mul == 0)
3543 		return;
3544 
3545 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3546 		/* We are idle */
3547 		return;
3548 	}
3549 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3550 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3551 		/*
3552 		 * Stop the goodput now, the idea here is
3553 		 * that future measurements with in_probe_rtt
3554 		 * won't register if they are not greater so
3555 		 * we want to get what info (if any) is available
3556 		 * now.
3557 		 */
3558 		rack_do_goodput_measurement(rack->rc_tp, rack,
3559 					    rack->rc_tp->snd_una, __LINE__,
3560 					    RACK_QUALITY_PROBERTT);
3561 	}
3562 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3563 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3564 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3565 		     rack->r_ctl.rc_pace_min_segs);
3566 	rack->in_probe_rtt = 1;
3567 	rack->measure_saw_probe_rtt = 1;
3568 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3569 	rack->r_ctl.rc_time_probertt_starts = 0;
3570 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3571 	if (rack_probertt_use_min_rtt_entry)
3572 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3573 	else
3574 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3575 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3576 			     __LINE__, RACK_RTTS_ENTERPROBE);
3577 }
3578 
3579 static void
3580 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3581 {
3582 	struct rack_sendmap *rsm;
3583 	uint32_t segsiz;
3584 
3585 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3586 		     rack->r_ctl.rc_pace_min_segs);
3587 	rack->in_probe_rtt = 0;
3588 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3589 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3590 		/*
3591 		 * Stop the goodput now, the idea here is
3592 		 * that future measurements with in_probe_rtt
3593 		 * won't register if they are not greater so
3594 		 * we want to get what info (if any) is available
3595 		 * now.
3596 		 */
3597 		rack_do_goodput_measurement(rack->rc_tp, rack,
3598 					    rack->rc_tp->snd_una, __LINE__,
3599 					    RACK_QUALITY_PROBERTT);
3600 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3601 		/*
3602 		 * We don't have enough data to make a measurement.
3603 		 * So lets just stop and start here after exiting
3604 		 * probe-rtt. We probably are not interested in
3605 		 * the results anyway.
3606 		 */
3607 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3608 	}
3609 	/*
3610 	 * Measurements through the current snd_max are going
3611 	 * to be limited by the slower pacing rate.
3612 	 *
3613 	 * We need to mark these as app-limited so we
3614 	 * don't collapse the b/w.
3615 	 */
3616 	rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3617 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3618 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3619 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3620 		else {
3621 			/*
3622 			 * Go out to the end app limited and mark
3623 			 * this new one as next and move the end_appl up
3624 			 * to this guy.
3625 			 */
3626 			if (rack->r_ctl.rc_end_appl)
3627 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3628 			rack->r_ctl.rc_end_appl = rsm;
3629 		}
3630 		rsm->r_flags |= RACK_APP_LIMITED;
3631 		rack->r_ctl.rc_app_limited_cnt++;
3632 	}
3633 	/*
3634 	 * Now, we need to examine our pacing rate multipliers.
3635 	 * If its under 100%, we need to kick it back up to
3636 	 * 100%. We also don't let it be over our "max" above
3637 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3638 	 * Note setting clamp_atexit_prtt to 0 has the effect
3639 	 * of setting CA/SS to 100% always at exit (which is
3640 	 * the default behavior).
3641 	 */
3642 	if (rack_probertt_clear_is) {
3643 		rack->rc_gp_incr = 0;
3644 		rack->rc_gp_bwred = 0;
3645 		rack->rc_gp_timely_inc_cnt = 0;
3646 		rack->rc_gp_timely_dec_cnt = 0;
3647 	}
3648 	/* Do we do any clamping at exit? */
3649 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3650 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3651 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3652 	}
3653 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3654 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3655 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3656 	}
3657 	/*
3658 	 * Lets set rtt_diff to 0, so that we will get a "boost"
3659 	 * after exiting.
3660 	 */
3661 	rack->r_ctl.rc_rtt_diff = 0;
3662 
3663 	/* Clear all flags so we start fresh */
3664 	rack->rc_tp->t_bytes_acked = 0;
3665 	rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
3666 	/*
3667 	 * If configured to, set the cwnd and ssthresh to
3668 	 * our targets.
3669 	 */
3670 	if (rack_probe_rtt_sets_cwnd) {
3671 		uint64_t ebdp;
3672 		uint32_t setto;
3673 
3674 		/* Set ssthresh so we get into CA once we hit our target */
3675 		if (rack_probertt_use_min_rtt_exit == 1) {
3676 			/* Set to min rtt */
3677 			rack_set_prtt_target(rack, segsiz,
3678 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3679 		} else if (rack_probertt_use_min_rtt_exit == 2) {
3680 			/* Set to current gp rtt */
3681 			rack_set_prtt_target(rack, segsiz,
3682 					     rack->r_ctl.rc_gp_srtt);
3683 		} else if (rack_probertt_use_min_rtt_exit == 3) {
3684 			/* Set to entry gp rtt */
3685 			rack_set_prtt_target(rack, segsiz,
3686 					     rack->r_ctl.rc_entry_gp_rtt);
3687 		} else {
3688 			uint64_t sum;
3689 			uint32_t setval;
3690 
3691 			sum = rack->r_ctl.rc_entry_gp_rtt;
3692 			sum *= 10;
3693 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3694 			if (sum >= 20) {
3695 				/*
3696 				 * A highly buffered path needs
3697 				 * cwnd space for timely to work.
3698 				 * Lets set things up as if
3699 				 * we are heading back here again.
3700 				 */
3701 				setval = rack->r_ctl.rc_entry_gp_rtt;
3702 			} else if (sum >= 15) {
3703 				/*
3704 				 * Lets take the smaller of the
3705 				 * two since we are just somewhat
3706 				 * buffered.
3707 				 */
3708 				setval = rack->r_ctl.rc_gp_srtt;
3709 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
3710 					setval = rack->r_ctl.rc_entry_gp_rtt;
3711 			} else {
3712 				/*
3713 				 * Here we are not highly buffered
3714 				 * and should pick the min we can to
3715 				 * keep from causing loss.
3716 				 */
3717 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3718 			}
3719 			rack_set_prtt_target(rack, segsiz,
3720 					     setval);
3721 		}
3722 		if (rack_probe_rtt_sets_cwnd > 1) {
3723 			/* There is a percentage here to boost */
3724 			ebdp = rack->r_ctl.rc_target_probertt_flight;
3725 			ebdp *= rack_probe_rtt_sets_cwnd;
3726 			ebdp /= 100;
3727 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3728 		} else
3729 			setto = rack->r_ctl.rc_target_probertt_flight;
3730 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3731 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3732 			/* Enforce a min */
3733 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3734 		}
3735 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
3736 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3737 	}
3738 	rack_log_rtt_shrinks(rack,  us_cts,
3739 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3740 			     __LINE__, RACK_RTTS_EXITPROBE);
3741 	/* Clear times last so log has all the info */
3742 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3743 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3744 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3745 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
3746 }
3747 
3748 static void
3749 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3750 {
3751 	/* Check in on probe-rtt */
3752 	if (rack->rc_gp_filled == 0) {
3753 		/* We do not do p-rtt unless we have gp measurements */
3754 		return;
3755 	}
3756 	if (rack->in_probe_rtt) {
3757 		uint64_t no_overflow;
3758 		uint32_t endtime, must_stay;
3759 
3760 		if (rack->r_ctl.rc_went_idle_time &&
3761 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3762 			/*
3763 			 * We went idle during prtt, just exit now.
3764 			 */
3765 			rack_exit_probertt(rack, us_cts);
3766 		} else if (rack_probe_rtt_safety_val &&
3767 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3768 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3769 			/*
3770 			 * Probe RTT safety value triggered!
3771 			 */
3772 			rack_log_rtt_shrinks(rack,  us_cts,
3773 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3774 					     __LINE__, RACK_RTTS_SAFETY);
3775 			rack_exit_probertt(rack, us_cts);
3776 		}
3777 		/* Calculate the max we will wait */
3778 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3779 		if (rack->rc_highly_buffered)
3780 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3781 		/* Calculate the min we must wait */
3782 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3783 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3784 		    TSTMP_LT(us_cts, endtime)) {
3785 			uint32_t calc;
3786 			/* Do we lower more? */
3787 no_exit:
3788 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3789 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3790 			else
3791 				calc = 0;
3792 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3793 			if (calc) {
3794 				/* Maybe */
3795 				calc *= rack_per_of_gp_probertt_reduce;
3796 				rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3797 				/* Limit it too */
3798 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3799 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3800 			}
3801 			/* We must reach target or the time set */
3802 			return;
3803 		}
3804 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
3805 			if ((TSTMP_LT(us_cts, must_stay) &&
3806 			     rack->rc_highly_buffered) ||
3807 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3808 			      rack->r_ctl.rc_target_probertt_flight)) {
3809 				/* We are not past the must_stay time */
3810 				goto no_exit;
3811 			}
3812 			rack_log_rtt_shrinks(rack,  us_cts,
3813 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3814 					     __LINE__, RACK_RTTS_REACHTARGET);
3815 			rack->r_ctl.rc_time_probertt_starts = us_cts;
3816 			if (rack->r_ctl.rc_time_probertt_starts == 0)
3817 				rack->r_ctl.rc_time_probertt_starts = 1;
3818 			/* Restore back to our rate we want to pace at in prtt */
3819 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3820 		}
3821 		/*
3822 		 * Setup our end time, some number of gp_srtts plus 200ms.
3823 		 */
3824 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3825 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3826 		if (rack_probertt_gpsrtt_cnt_div)
3827 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3828 		else
3829 			endtime = 0;
3830 		endtime += rack_min_probertt_hold;
3831 		endtime += rack->r_ctl.rc_time_probertt_starts;
3832 		if (TSTMP_GEQ(us_cts,  endtime)) {
3833 			/* yes, exit probertt */
3834 			rack_exit_probertt(rack, us_cts);
3835 		}
3836 
3837 	} else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3838 		/* Go into probertt, its been too long since we went lower */
3839 		rack_enter_probertt(rack, us_cts);
3840 	}
3841 }
3842 
3843 static void
3844 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3845 		       uint32_t rtt, int32_t rtt_diff)
3846 {
3847 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
3848 	uint32_t losses;
3849 
3850 	if ((rack->rc_gp_dyn_mul == 0) ||
3851 	    (rack->use_fixed_rate) ||
3852 	    (rack->in_probe_rtt) ||
3853 	    (rack->rc_always_pace == 0)) {
3854 		/* No dynamic GP multiplier in play */
3855 		return;
3856 	}
3857 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3858 	cur_bw = rack_get_bw(rack);
3859 	/* Calculate our up and down range */
3860 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3861 	up_bnd /= 100;
3862 	up_bnd += rack->r_ctl.last_gp_comp_bw;
3863 
3864 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3865 	subfr /= 100;
3866 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3867 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3868 		/*
3869 		 * This is the case where our RTT is above
3870 		 * the max target and we have been configured
3871 		 * to just do timely no bonus up stuff in that case.
3872 		 *
3873 		 * There are two configurations, set to 1, and we
3874 		 * just do timely if we are over our max. If its
3875 		 * set above 1 then we slam the multipliers down
3876 		 * to 100 and then decrement per timely.
3877 		 */
3878 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3879 				__LINE__, 3);
3880 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3881 			rack_validate_multipliers_at_or_below_100(rack);
3882 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3883 	} else if ((last_bw_est < low_bnd) && !losses) {
3884 		/*
3885 		 * We are decreasing this is a bit complicated this
3886 		 * means we are loosing ground. This could be
3887 		 * because another flow entered and we are competing
3888 		 * for b/w with it. This will push the RTT up which
3889 		 * makes timely unusable unless we want to get shoved
3890 		 * into a corner and just be backed off (the age
3891 		 * old problem with delay based CC).
3892 		 *
3893 		 * On the other hand if it was a route change we
3894 		 * would like to stay somewhat contained and not
3895 		 * blow out the buffers.
3896 		 */
3897 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3898 				__LINE__, 3);
3899 		rack->r_ctl.last_gp_comp_bw = cur_bw;
3900 		if (rack->rc_gp_bwred == 0) {
3901 			/* Go into reduction counting */
3902 			rack->rc_gp_bwred = 1;
3903 			rack->rc_gp_timely_dec_cnt = 0;
3904 		}
3905 		if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
3906 		    (timely_says == 0)) {
3907 			/*
3908 			 * Push another time with a faster pacing
3909 			 * to try to gain back (we include override to
3910 			 * get a full raise factor).
3911 			 */
3912 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
3913 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
3914 			    (timely_says == 0) ||
3915 			    (rack_down_raise_thresh == 0)) {
3916 				/*
3917 				 * Do an override up in b/w if we were
3918 				 * below the threshold or if the threshold
3919 				 * is zero we always do the raise.
3920 				 */
3921 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
3922 			} else {
3923 				/* Log it stays the same */
3924 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
3925 						__LINE__, 11);
3926 			}
3927 			rack->rc_gp_timely_dec_cnt++;
3928 			/* We are not incrementing really no-count */
3929 			rack->rc_gp_incr = 0;
3930 			rack->rc_gp_timely_inc_cnt = 0;
3931 		} else {
3932 			/*
3933 			 * Lets just use the RTT
3934 			 * information and give up
3935 			 * pushing.
3936 			 */
3937 			goto use_timely;
3938 		}
3939 	} else if ((timely_says != 2) &&
3940 		    !losses &&
3941 		    (last_bw_est > up_bnd)) {
3942 		/*
3943 		 * We are increasing b/w lets keep going, updating
3944 		 * our b/w and ignoring any timely input, unless
3945 		 * of course we are at our max raise (if there is one).
3946 		 */
3947 
3948 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3949 				__LINE__, 3);
3950 		rack->r_ctl.last_gp_comp_bw = cur_bw;
3951 		if (rack->rc_gp_saw_ss &&
3952 		    rack_per_upper_bound_ss &&
3953 		     (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
3954 			    /*
3955 			     * In cases where we can't go higher
3956 			     * we should just use timely.
3957 			     */
3958 			    goto use_timely;
3959 		}
3960 		if (rack->rc_gp_saw_ca &&
3961 		    rack_per_upper_bound_ca &&
3962 		    (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
3963 			    /*
3964 			     * In cases where we can't go higher
3965 			     * we should just use timely.
3966 			     */
3967 			    goto use_timely;
3968 		}
3969 		rack->rc_gp_bwred = 0;
3970 		rack->rc_gp_timely_dec_cnt = 0;
3971 		/* You get a set number of pushes if timely is trying to reduce */
3972 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
3973 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3974 		} else {
3975 			/* Log it stays the same */
3976 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
3977 			    __LINE__, 12);
3978 		}
3979 		return;
3980 	} else {
3981 		/*
3982 		 * We are staying between the lower and upper range bounds
3983 		 * so use timely to decide.
3984 		 */
3985 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3986 				__LINE__, 3);
3987 use_timely:
3988 		if (timely_says) {
3989 			rack->rc_gp_incr = 0;
3990 			rack->rc_gp_timely_inc_cnt = 0;
3991 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
3992 			    !losses &&
3993 			    (last_bw_est < low_bnd)) {
3994 				/* We are loosing ground */
3995 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3996 				rack->rc_gp_timely_dec_cnt++;
3997 				/* We are not incrementing really no-count */
3998 				rack->rc_gp_incr = 0;
3999 				rack->rc_gp_timely_inc_cnt = 0;
4000 			} else
4001 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4002 		} else {
4003 			rack->rc_gp_bwred = 0;
4004 			rack->rc_gp_timely_dec_cnt = 0;
4005 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4006 		}
4007 	}
4008 }
4009 
4010 static int32_t
4011 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4012 {
4013 	int32_t timely_says;
4014 	uint64_t log_mult, log_rtt_a_diff;
4015 
4016 	log_rtt_a_diff = rtt;
4017 	log_rtt_a_diff <<= 32;
4018 	log_rtt_a_diff |= (uint32_t)rtt_diff;
4019 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4020 		    rack_gp_rtt_maxmul)) {
4021 		/* Reduce the b/w multiplier */
4022 		timely_says = 2;
4023 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4024 		log_mult <<= 32;
4025 		log_mult |= prev_rtt;
4026 		rack_log_timely(rack,  timely_says, log_mult,
4027 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4028 				log_rtt_a_diff, __LINE__, 4);
4029 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4030 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4031 			    max(rack_gp_rtt_mindiv , 1)))) {
4032 		/* Increase the b/w multiplier */
4033 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4034 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4035 			 max(rack_gp_rtt_mindiv , 1));
4036 		log_mult <<= 32;
4037 		log_mult |= prev_rtt;
4038 		timely_says = 0;
4039 		rack_log_timely(rack,  timely_says, log_mult ,
4040 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4041 				log_rtt_a_diff, __LINE__, 5);
4042 	} else {
4043 		/*
4044 		 * Use a gradient to find it the timely gradient
4045 		 * is:
4046 		 * grad = rc_rtt_diff / min_rtt;
4047 		 *
4048 		 * anything below or equal to 0 will be
4049 		 * a increase indication. Anything above
4050 		 * zero is a decrease. Note we take care
4051 		 * of the actual gradient calculation
4052 		 * in the reduction (its not needed for
4053 		 * increase).
4054 		 */
4055 		log_mult = prev_rtt;
4056 		if (rtt_diff <= 0) {
4057 			/*
4058 			 * Rttdiff is less than zero, increase the
4059 			 * b/w multiplier (its 0 or negative)
4060 			 */
4061 			timely_says = 0;
4062 			rack_log_timely(rack,  timely_says, log_mult,
4063 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4064 		} else {
4065 			/* Reduce the b/w multiplier */
4066 			timely_says = 1;
4067 			rack_log_timely(rack,  timely_says, log_mult,
4068 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4069 		}
4070 	}
4071 	return (timely_says);
4072 }
4073 
4074 static void
4075 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4076 			    tcp_seq th_ack, int line, uint8_t quality)
4077 {
4078 	uint64_t tim, bytes_ps, ltim, stim, utim;
4079 	uint32_t segsiz, bytes, reqbytes, us_cts;
4080 	int32_t gput, new_rtt_diff, timely_says;
4081 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4082 	int did_add = 0;
4083 
4084 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4085 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4086 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
4087 		tim = us_cts - tp->gput_ts;
4088 	else
4089 		tim = 0;
4090 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4091 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4092 	else
4093 		stim = 0;
4094 	/*
4095 	 * Use the larger of the send time or ack time. This prevents us
4096 	 * from being influenced by ack artifacts to come up with too
4097 	 * high of measurement. Note that since we are spanning over many more
4098 	 * bytes in most of our measurements hopefully that is less likely to
4099 	 * occur.
4100 	 */
4101 	if (tim > stim)
4102 		utim = max(tim, 1);
4103 	else
4104 		utim = max(stim, 1);
4105 	/* Lets get a msec time ltim too for the old stuff */
4106 	ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4107 	gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4108 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4109 	if ((tim == 0) && (stim == 0)) {
4110 		/*
4111 		 * Invalid measurement time, maybe
4112 		 * all on one ack/one send?
4113 		 */
4114 		bytes = 0;
4115 		bytes_ps = 0;
4116 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4117 					   0, 0, 0, 10, __LINE__, NULL, quality);
4118 		goto skip_measurement;
4119 	}
4120 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4121 		/* We never made a us_rtt measurement? */
4122 		bytes = 0;
4123 		bytes_ps = 0;
4124 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4125 					   0, 0, 0, 10, __LINE__, NULL, quality);
4126 		goto skip_measurement;
4127 	}
4128 	/*
4129 	 * Calculate the maximum possible b/w this connection
4130 	 * could have. We base our calculation on the lowest
4131 	 * rtt we have seen during the measurement and the
4132 	 * largest rwnd the client has given us in that time. This
4133 	 * forms a BDP that is the maximum that we could ever
4134 	 * get to the client. Anything larger is not valid.
4135 	 *
4136 	 * I originally had code here that rejected measurements
4137 	 * where the time was less than 1/2 the latest us_rtt.
4138 	 * But after thinking on that I realized its wrong since
4139 	 * say you had a 150Mbps or even 1Gbps link, and you
4140 	 * were a long way away.. example I am in Europe (100ms rtt)
4141 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4142 	 * bytes my time would be 1.2ms, and yet my rtt would say
4143 	 * the measurement was invalid the time was < 50ms. The
4144 	 * same thing is true for 150Mb (8ms of time).
4145 	 *
4146 	 * A better way I realized is to look at what the maximum
4147 	 * the connection could possibly do. This is gated on
4148 	 * the lowest RTT we have seen and the highest rwnd.
4149 	 * We should in theory never exceed that, if we are
4150 	 * then something on the path is storing up packets
4151 	 * and then feeding them all at once to our endpoint
4152 	 * messing up our measurement.
4153 	 */
4154 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4155 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4156 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4157 	if (SEQ_LT(th_ack, tp->gput_seq)) {
4158 		/* No measurement can be made */
4159 		bytes = 0;
4160 		bytes_ps = 0;
4161 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4162 					   0, 0, 0, 10, __LINE__, NULL, quality);
4163 		goto skip_measurement;
4164 	} else
4165 		bytes = (th_ack - tp->gput_seq);
4166 	bytes_ps = (uint64_t)bytes;
4167 	/*
4168 	 * Don't measure a b/w for pacing unless we have gotten at least
4169 	 * an initial windows worth of data in this measurement interval.
4170 	 *
4171 	 * Small numbers of bytes get badly influenced by delayed ack and
4172 	 * other artifacts. Note we take the initial window or our
4173 	 * defined minimum GP (defaulting to 10 which hopefully is the
4174 	 * IW).
4175 	 */
4176 	if (rack->rc_gp_filled == 0) {
4177 		/*
4178 		 * The initial estimate is special. We
4179 		 * have blasted out an IW worth of packets
4180 		 * without a real valid ack ts results. We
4181 		 * then setup the app_limited_needs_set flag,
4182 		 * this should get the first ack in (probably 2
4183 		 * MSS worth) to be recorded as the timestamp.
4184 		 * We thus allow a smaller number of bytes i.e.
4185 		 * IW - 2MSS.
4186 		 */
4187 		reqbytes -= (2 * segsiz);
4188 		/* Also lets fill previous for our first measurement to be neutral */
4189 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4190 	}
4191 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4192 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4193 					   rack->r_ctl.rc_app_limited_cnt,
4194 					   0, 0, 10, __LINE__, NULL, quality);
4195 		goto skip_measurement;
4196 	}
4197 	/*
4198 	 * We now need to calculate the Timely like status so
4199 	 * we can update (possibly) the b/w multipliers.
4200 	 */
4201 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4202 	if (rack->rc_gp_filled == 0) {
4203 		/* No previous reading */
4204 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4205 	} else {
4206 		if (rack->measure_saw_probe_rtt == 0) {
4207 			/*
4208 			 * We don't want a probertt to be counted
4209 			 * since it will be negative incorrectly. We
4210 			 * expect to be reducing the RTT when we
4211 			 * pace at a slower rate.
4212 			 */
4213 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4214 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4215 		}
4216 	}
4217 	timely_says = rack_make_timely_judgement(rack,
4218 		rack->r_ctl.rc_gp_srtt,
4219 		rack->r_ctl.rc_rtt_diff,
4220 	        rack->r_ctl.rc_prev_gp_srtt
4221 		);
4222 	bytes_ps *= HPTS_USEC_IN_SEC;
4223 	bytes_ps /= utim;
4224 	if (bytes_ps > rack->r_ctl.last_max_bw) {
4225 		/*
4226 		 * Something is on path playing
4227 		 * since this b/w is not possible based
4228 		 * on our BDP (highest rwnd and lowest rtt
4229 		 * we saw in the measurement window).
4230 		 *
4231 		 * Another option here would be to
4232 		 * instead skip the measurement.
4233 		 */
4234 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4235 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
4236 					   11, __LINE__, NULL, quality);
4237 		bytes_ps = rack->r_ctl.last_max_bw;
4238 	}
4239 	/* We store gp for b/w in bytes per second */
4240 	if (rack->rc_gp_filled == 0) {
4241 		/* Initial measurement */
4242 		if (bytes_ps) {
4243 			rack->r_ctl.gp_bw = bytes_ps;
4244 			rack->rc_gp_filled = 1;
4245 			rack->r_ctl.num_measurements = 1;
4246 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4247 		} else {
4248 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4249 						   rack->r_ctl.rc_app_limited_cnt,
4250 						   0, 0, 10, __LINE__, NULL, quality);
4251 		}
4252 		if (tcp_in_hpts(rack->rc_inp) &&
4253 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4254 			/*
4255 			 * Ok we can't trust the pacer in this case
4256 			 * where we transition from un-paced to paced.
4257 			 * Or for that matter when the burst mitigation
4258 			 * was making a wild guess and got it wrong.
4259 			 * Stop the pacer and clear up all the aggregate
4260 			 * delays etc.
4261 			 */
4262 			tcp_hpts_remove(rack->rc_inp);
4263 			rack->r_ctl.rc_hpts_flags = 0;
4264 			rack->r_ctl.rc_last_output_to = 0;
4265 		}
4266 		did_add = 2;
4267 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4268 		/* Still a small number run an average */
4269 		rack->r_ctl.gp_bw += bytes_ps;
4270 		addpart = rack->r_ctl.num_measurements;
4271 		rack->r_ctl.num_measurements++;
4272 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4273 			/* We have collected enough to move forward */
4274 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4275 		}
4276 		did_add = 3;
4277 	} else {
4278 		/*
4279 		 * We want to take 1/wma of the goodput and add in to 7/8th
4280 		 * of the old value weighted by the srtt. So if your measurement
4281 		 * period is say 2 SRTT's long you would get 1/4 as the
4282 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
4283 		 *
4284 		 * But we must be careful not to take too much i.e. if the
4285 		 * srtt is say 20ms and the measurement is taken over
4286 		 * 400ms our weight would be 400/20 i.e. 20. On the
4287 		 * other hand if we get a measurement over 1ms with a
4288 		 * 10ms rtt we only want to take a much smaller portion.
4289 		 */
4290 		if (rack->r_ctl.num_measurements < 0xff) {
4291 			rack->r_ctl.num_measurements++;
4292 		}
4293 		srtt = (uint64_t)tp->t_srtt;
4294 		if (srtt == 0) {
4295 			/*
4296 			 * Strange why did t_srtt go back to zero?
4297 			 */
4298 			if (rack->r_ctl.rc_rack_min_rtt)
4299 				srtt = rack->r_ctl.rc_rack_min_rtt;
4300 			else
4301 				srtt = HPTS_USEC_IN_MSEC;
4302 		}
4303 		/*
4304 		 * XXXrrs: Note for reviewers, in playing with
4305 		 * dynamic pacing I discovered this GP calculation
4306 		 * as done originally leads to some undesired results.
4307 		 * Basically you can get longer measurements contributing
4308 		 * too much to the WMA. Thus I changed it if you are doing
4309 		 * dynamic adjustments to only do the aportioned adjustment
4310 		 * if we have a very small (time wise) measurement. Longer
4311 		 * measurements just get there weight (defaulting to 1/8)
4312 		 * add to the WMA. We may want to think about changing
4313 		 * this to always do that for both sides i.e. dynamic
4314 		 * and non-dynamic... but considering lots of folks
4315 		 * were playing with this I did not want to change the
4316 		 * calculation per.se. without your thoughts.. Lawerence?
4317 		 * Peter??
4318 		 */
4319 		if (rack->rc_gp_dyn_mul == 0) {
4320 			subpart = rack->r_ctl.gp_bw * utim;
4321 			subpart /= (srtt * 8);
4322 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
4323 				/*
4324 				 * The b/w update takes no more
4325 				 * away then 1/2 our running total
4326 				 * so factor it in.
4327 				 */
4328 				addpart = bytes_ps * utim;
4329 				addpart /= (srtt * 8);
4330 			} else {
4331 				/*
4332 				 * Don't allow a single measurement
4333 				 * to account for more than 1/2 of the
4334 				 * WMA. This could happen on a retransmission
4335 				 * where utim becomes huge compared to
4336 				 * srtt (multiple retransmissions when using
4337 				 * the sending rate which factors in all the
4338 				 * transmissions from the first one).
4339 				 */
4340 				subpart = rack->r_ctl.gp_bw / 2;
4341 				addpart = bytes_ps / 2;
4342 			}
4343 			resid_bw = rack->r_ctl.gp_bw - subpart;
4344 			rack->r_ctl.gp_bw = resid_bw + addpart;
4345 			did_add = 1;
4346 		} else {
4347 			if ((utim / srtt) <= 1) {
4348 				/*
4349 				 * The b/w update was over a small period
4350 				 * of time. The idea here is to prevent a small
4351 				 * measurement time period from counting
4352 				 * too much. So we scale it based on the
4353 				 * time so it attributes less than 1/rack_wma_divisor
4354 				 * of its measurement.
4355 				 */
4356 				subpart = rack->r_ctl.gp_bw * utim;
4357 				subpart /= (srtt * rack_wma_divisor);
4358 				addpart = bytes_ps * utim;
4359 				addpart /= (srtt * rack_wma_divisor);
4360 			} else {
4361 				/*
4362 				 * The scaled measurement was long
4363 				 * enough so lets just add in the
4364 				 * portion of the measurement i.e. 1/rack_wma_divisor
4365 				 */
4366 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4367 				addpart = bytes_ps / rack_wma_divisor;
4368 			}
4369 			if ((rack->measure_saw_probe_rtt == 0) ||
4370 		            (bytes_ps > rack->r_ctl.gp_bw)) {
4371 				/*
4372 				 * For probe-rtt we only add it in
4373 				 * if its larger, all others we just
4374 				 * add in.
4375 				 */
4376 				did_add = 1;
4377 				resid_bw = rack->r_ctl.gp_bw - subpart;
4378 				rack->r_ctl.gp_bw = resid_bw + addpart;
4379 			}
4380 		}
4381 	}
4382 	if ((rack->gp_ready == 0) &&
4383 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4384 		/* We have enough measurements now */
4385 		rack->gp_ready = 1;
4386 		rack_set_cc_pacing(rack);
4387 		if (rack->defer_options)
4388 			rack_apply_deferred_options(rack);
4389 	}
4390 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4391 				   rack_get_bw(rack), 22, did_add, NULL, quality);
4392 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
4393 	if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4394 		rack_update_multiplier(rack, timely_says, bytes_ps,
4395 				       rack->r_ctl.rc_gp_srtt,
4396 				       rack->r_ctl.rc_rtt_diff);
4397 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4398 				   rack_get_bw(rack), 3, line, NULL, quality);
4399 	/* reset the gp srtt and setup the new prev */
4400 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4401 	/* Record the lost count for the next measurement */
4402 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4403 	/*
4404 	 * We restart our diffs based on the gpsrtt in the
4405 	 * measurement window.
4406 	 */
4407 	rack->rc_gp_rtt_set = 0;
4408 	rack->rc_gp_saw_rec = 0;
4409 	rack->rc_gp_saw_ca = 0;
4410 	rack->rc_gp_saw_ss = 0;
4411 	rack->rc_dragged_bottom = 0;
4412 skip_measurement:
4413 
4414 #ifdef STATS
4415 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4416 				 gput);
4417 	/*
4418 	 * XXXLAS: This is a temporary hack, and should be
4419 	 * chained off VOI_TCP_GPUT when stats(9) grows an
4420 	 * API to deal with chained VOIs.
4421 	 */
4422 	if (tp->t_stats_gput_prev > 0)
4423 		stats_voi_update_abs_s32(tp->t_stats,
4424 					 VOI_TCP_GPUT_ND,
4425 					 ((gput - tp->t_stats_gput_prev) * 100) /
4426 					 tp->t_stats_gput_prev);
4427 #endif
4428 	tp->t_flags &= ~TF_GPUTINPROG;
4429 	tp->t_stats_gput_prev = gput;
4430 	/*
4431 	 * Now are we app limited now and there is space from where we
4432 	 * were to where we want to go?
4433 	 *
4434 	 * We don't do the other case i.e. non-applimited here since
4435 	 * the next send will trigger us picking up the missing data.
4436 	 */
4437 	if (rack->r_ctl.rc_first_appl &&
4438 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
4439 	    rack->r_ctl.rc_app_limited_cnt &&
4440 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4441 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
4442 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4443 		/*
4444 		 * Yep there is enough outstanding to make a measurement here.
4445 		 */
4446 		struct rack_sendmap *rsm, fe;
4447 
4448 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4449 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4450 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4451 		rack->app_limited_needs_set = 0;
4452 		tp->gput_seq = th_ack;
4453 		if (rack->in_probe_rtt)
4454 			rack->measure_saw_probe_rtt = 1;
4455 		else if ((rack->measure_saw_probe_rtt) &&
4456 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4457 			rack->measure_saw_probe_rtt = 0;
4458 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
4459 			/* There is a full window to gain info from */
4460 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4461 		} else {
4462 			/* We can only measure up to the applimited point */
4463 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
4464 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
4465 				/*
4466 				 * We don't have enough to make a measurement.
4467 				 */
4468 				tp->t_flags &= ~TF_GPUTINPROG;
4469 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
4470 							   0, 0, 0, 6, __LINE__, NULL, quality);
4471 				return;
4472 			}
4473 		}
4474 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
4475 			/*
4476 			 * We will get no more data into the SB
4477 			 * this means we need to have the data available
4478 			 * before we start a measurement.
4479 			 */
4480 			if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
4481 				/* Nope not enough data. */
4482 				return;
4483 			}
4484 		}
4485 		tp->t_flags |= TF_GPUTINPROG;
4486 		/*
4487 		 * Now we need to find the timestamp of the send at tp->gput_seq
4488 		 * for the send based measurement.
4489 		 */
4490 		fe.r_start = tp->gput_seq;
4491 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4492 		if (rsm) {
4493 			/* Ok send-based limit is set */
4494 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4495 				/*
4496 				 * Move back to include the earlier part
4497 				 * so our ack time lines up right (this may
4498 				 * make an overlapping measurement but thats
4499 				 * ok).
4500 				 */
4501 				tp->gput_seq = rsm->r_start;
4502 			}
4503 			if (rsm->r_flags & RACK_ACKED)
4504 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4505 			else
4506 				rack->app_limited_needs_set = 1;
4507 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4508 		} else {
4509 			/*
4510 			 * If we don't find the rsm due to some
4511 			 * send-limit set the current time, which
4512 			 * basically disables the send-limit.
4513 			 */
4514 			struct timeval tv;
4515 
4516 			microuptime(&tv);
4517 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4518 		}
4519 		rack_log_pacing_delay_calc(rack,
4520 					   tp->gput_seq,
4521 					   tp->gput_ack,
4522 					   (uint64_t)rsm,
4523 					   tp->gput_ts,
4524 					   rack->r_ctl.rc_app_limited_cnt,
4525 					   9,
4526 					   __LINE__, NULL, quality);
4527 	}
4528 }
4529 
4530 /*
4531  * CC wrapper hook functions
4532  */
4533 static void
4534 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4535     uint16_t type, int32_t recovery)
4536 {
4537 	uint32_t prior_cwnd, acked;
4538 	struct tcp_log_buffer *lgb = NULL;
4539 	uint8_t labc_to_use, quality;
4540 
4541 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4542 	tp->t_ccv.nsegs = nsegs;
4543 	acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
4544 	if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4545 		uint32_t max;
4546 
4547 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4548 		if (tp->t_ccv.bytes_this_ack > max) {
4549 			tp->t_ccv.bytes_this_ack = max;
4550 		}
4551 	}
4552 #ifdef STATS
4553 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4554 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4555 #endif
4556 	quality = RACK_QUALITY_NONE;
4557 	if ((tp->t_flags & TF_GPUTINPROG) &&
4558 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
4559 		/* Measure the Goodput */
4560 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
4561 #ifdef NETFLIX_PEAKRATE
4562 		if ((type == CC_ACK) &&
4563 		    (tp->t_maxpeakrate)) {
4564 			/*
4565 			 * We update t_peakrate_thr. This gives us roughly
4566 			 * one update per round trip time. Note
4567 			 * it will only be used if pace_always is off i.e
4568 			 * we don't do this for paced flows.
4569 			 */
4570 			rack_update_peakrate_thr(tp);
4571 		}
4572 #endif
4573 	}
4574 	/* Which way our we limited, if not cwnd limited no advance in CA */
4575 	if (tp->snd_cwnd <= tp->snd_wnd)
4576 		tp->t_ccv.flags |= CCF_CWND_LIMITED;
4577 	else
4578 		tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
4579 	if (tp->snd_cwnd > tp->snd_ssthresh) {
4580 		tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
4581 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4582 		/* For the setting of a window past use the actual scwnd we are using */
4583 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4584 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4585 			tp->t_ccv.flags |= CCF_ABC_SENTAWND;
4586 		}
4587 	} else {
4588 		tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4589 		tp->t_bytes_acked = 0;
4590 	}
4591 	prior_cwnd = tp->snd_cwnd;
4592 	if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4593 	    (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4594 		labc_to_use = rack->rc_labc;
4595 	else
4596 		labc_to_use = rack_max_abc_post_recovery;
4597 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4598 		union tcp_log_stackspecific log;
4599 		struct timeval tv;
4600 
4601 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4602 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4603 		log.u_bbr.flex1 = th_ack;
4604 		log.u_bbr.flex2 = tp->t_ccv.flags;
4605 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
4606 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
4607 		log.u_bbr.flex5 = labc_to_use;
4608 		log.u_bbr.flex6 = prior_cwnd;
4609 		log.u_bbr.flex7 = V_tcp_do_newsack;
4610 		log.u_bbr.flex8 = 1;
4611 		lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4612 				     0, &log, false, NULL, NULL, 0, &tv);
4613 	}
4614 	if (CC_ALGO(tp)->ack_received != NULL) {
4615 		/* XXXLAS: Find a way to live without this */
4616 		tp->t_ccv.curack = th_ack;
4617 		tp->t_ccv.labc = labc_to_use;
4618 		tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
4619 		CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
4620 	}
4621 	if (lgb) {
4622 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4623 	}
4624 	if (rack->r_must_retran) {
4625 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4626 			/*
4627 			 * We now are beyond the rxt point so lets disable
4628 			 * the flag.
4629 			 */
4630 			rack->r_ctl.rc_out_at_rto = 0;
4631 			rack->r_must_retran = 0;
4632 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4633 			/*
4634 			 * Only decrement the rc_out_at_rto if the cwnd advances
4635 			 * at least a whole segment. Otherwise next time the peer
4636 			 * acks, we won't be able to send this generaly happens
4637 			 * when we are in Congestion Avoidance.
4638 			 */
4639 			if (acked <= rack->r_ctl.rc_out_at_rto){
4640 				rack->r_ctl.rc_out_at_rto -= acked;
4641 			} else {
4642 				rack->r_ctl.rc_out_at_rto = 0;
4643 			}
4644 		}
4645 	}
4646 #ifdef STATS
4647 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4648 #endif
4649 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4650 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4651 	}
4652 #ifdef NETFLIX_PEAKRATE
4653 	/* we enforce max peak rate if it is set and we are not pacing */
4654 	if ((rack->rc_always_pace == 0) &&
4655 	    tp->t_peakrate_thr &&
4656 	    (tp->snd_cwnd > tp->t_peakrate_thr)) {
4657 		tp->snd_cwnd = tp->t_peakrate_thr;
4658 	}
4659 #endif
4660 }
4661 
4662 static void
4663 tcp_rack_partialack(struct tcpcb *tp)
4664 {
4665 	struct tcp_rack *rack;
4666 
4667 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4668 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4669 	/*
4670 	 * If we are doing PRR and have enough
4671 	 * room to send <or> we are pacing and prr
4672 	 * is disabled we will want to see if we
4673 	 * can send data (by setting r_wanted_output to
4674 	 * true).
4675 	 */
4676 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4677 	    rack->rack_no_prr)
4678 		rack->r_wanted_output = 1;
4679 }
4680 
4681 static void
4682 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4683 {
4684 	struct tcp_rack *rack;
4685 	uint32_t orig_cwnd;
4686 
4687 	orig_cwnd = tp->snd_cwnd;
4688 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4689 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4690 	/* only alert CC if we alerted when we entered */
4691 	if (CC_ALGO(tp)->post_recovery != NULL) {
4692 		tp->t_ccv.curack = th_ack;
4693 		CC_ALGO(tp)->post_recovery(&tp->t_ccv);
4694 		if (tp->snd_cwnd < tp->snd_ssthresh) {
4695 			/*
4696 			 * Rack has burst control and pacing
4697 			 * so lets not set this any lower than
4698 			 * snd_ssthresh per RFC-6582 (option 2).
4699 			 */
4700 			tp->snd_cwnd = tp->snd_ssthresh;
4701 		}
4702 	}
4703 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4704 		union tcp_log_stackspecific log;
4705 		struct timeval tv;
4706 
4707 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4708 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4709 		log.u_bbr.flex1 = th_ack;
4710 		log.u_bbr.flex2 = tp->t_ccv.flags;
4711 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
4712 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
4713 		log.u_bbr.flex5 = V_tcp_abc_l_var;
4714 		log.u_bbr.flex6 = orig_cwnd;
4715 		log.u_bbr.flex7 = V_tcp_do_newsack;
4716 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4717 		log.u_bbr.flex8 = 2;
4718 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4719 			       0, &log, false, NULL, NULL, 0, &tv);
4720 	}
4721 	if ((rack->rack_no_prr == 0) &&
4722 	    (rack->no_prr_addback == 0) &&
4723 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
4724 		/*
4725 		 * Suck the next prr cnt back into cwnd, but
4726 		 * only do that if we are not application limited.
4727 		 */
4728 		if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
4729 			/*
4730 			 * We are allowed to add back to the cwnd the amount we did
4731 			 * not get out if:
4732 			 * a) no_prr_addback is off.
4733 			 * b) we are not app limited
4734 			 * c) we are doing prr
4735 			 * <and>
4736 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4737 			 */
4738 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4739 					    rack->r_ctl.rc_prr_sndcnt);
4740 		}
4741 		rack->r_ctl.rc_prr_sndcnt = 0;
4742 		rack_log_to_prr(rack, 1, 0, __LINE__);
4743 	}
4744 	rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
4745 	tp->snd_recover = tp->snd_una;
4746 	if (rack->r_ctl.dsack_persist) {
4747 		rack->r_ctl.dsack_persist--;
4748 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
4749 			rack->r_ctl.num_dsack = 0;
4750 		}
4751 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
4752 	}
4753 	EXIT_RECOVERY(tp->t_flags);
4754 }
4755 
4756 static void
4757 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
4758 {
4759 	struct tcp_rack *rack;
4760 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4761 
4762 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4763 #ifdef STATS
4764 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4765 #endif
4766 	if (IN_RECOVERY(tp->t_flags) == 0) {
4767 		in_rec_at_entry = 0;
4768 		ssthresh_enter = tp->snd_ssthresh;
4769 		cwnd_enter = tp->snd_cwnd;
4770 	} else
4771 		in_rec_at_entry = 1;
4772 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4773 	switch (type) {
4774 	case CC_NDUPACK:
4775 		tp->t_flags &= ~TF_WASFRECOVERY;
4776 		tp->t_flags &= ~TF_WASCRECOVERY;
4777 		if (!IN_FASTRECOVERY(tp->t_flags)) {
4778 			rack->r_ctl.rc_prr_delivered = 0;
4779 			rack->r_ctl.rc_prr_out = 0;
4780 			if (rack->rack_no_prr == 0) {
4781 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4782 				rack_log_to_prr(rack, 2, in_rec_at_entry, line);
4783 			}
4784 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4785 			tp->snd_recover = tp->snd_max;
4786 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4787 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4788 		}
4789 		break;
4790 	case CC_ECN:
4791 		if (!IN_CONGRECOVERY(tp->t_flags) ||
4792 		    /*
4793 		     * Allow ECN reaction on ACK to CWR, if
4794 		     * that data segment was also CE marked.
4795 		     */
4796 		    SEQ_GEQ(ack, tp->snd_recover)) {
4797 			EXIT_CONGRECOVERY(tp->t_flags);
4798 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4799 			tp->snd_recover = tp->snd_max + 1;
4800 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4801 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4802 		}
4803 		break;
4804 	case CC_RTO:
4805 		tp->t_dupacks = 0;
4806 		tp->t_bytes_acked = 0;
4807 		EXIT_RECOVERY(tp->t_flags);
4808 		tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4809 		    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4810 		orig_cwnd = tp->snd_cwnd;
4811 		tp->snd_cwnd = ctf_fixed_maxseg(tp);
4812 		rack_log_to_prr(rack, 16, orig_cwnd, line);
4813 		if (tp->t_flags2 & TF2_ECN_PERMIT)
4814 			tp->t_flags2 |= TF2_ECN_SND_CWR;
4815 		break;
4816 	case CC_RTO_ERR:
4817 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4818 		/* RTO was unnecessary, so reset everything. */
4819 		tp->snd_cwnd = tp->snd_cwnd_prev;
4820 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
4821 		tp->snd_recover = tp->snd_recover_prev;
4822 		if (tp->t_flags & TF_WASFRECOVERY) {
4823 			ENTER_FASTRECOVERY(tp->t_flags);
4824 			tp->t_flags &= ~TF_WASFRECOVERY;
4825 		}
4826 		if (tp->t_flags & TF_WASCRECOVERY) {
4827 			ENTER_CONGRECOVERY(tp->t_flags);
4828 			tp->t_flags &= ~TF_WASCRECOVERY;
4829 		}
4830 		tp->snd_nxt = tp->snd_max;
4831 		tp->t_badrxtwin = 0;
4832 		break;
4833 	}
4834 	if ((CC_ALGO(tp)->cong_signal != NULL)  &&
4835 	    (type != CC_RTO)){
4836 		tp->t_ccv.curack = ack;
4837 		CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
4838 	}
4839 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4840 		rack_log_to_prr(rack, 15, cwnd_enter, line);
4841 		rack->r_ctl.dsack_byte_cnt = 0;
4842 		rack->r_ctl.retran_during_recovery = 0;
4843 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4844 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4845 		rack->r_ent_rec_ns = 1;
4846 	}
4847 }
4848 
4849 static inline void
4850 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4851 {
4852 	uint32_t i_cwnd;
4853 
4854 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4855 
4856 #ifdef NETFLIX_STATS
4857 	KMOD_TCPSTAT_INC(tcps_idle_restarts);
4858 	if (tp->t_state == TCPS_ESTABLISHED)
4859 		KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4860 #endif
4861 	if (CC_ALGO(tp)->after_idle != NULL)
4862 		CC_ALGO(tp)->after_idle(&tp->t_ccv);
4863 
4864 	if (tp->snd_cwnd == 1)
4865 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
4866 	else
4867 		i_cwnd = rc_init_window(rack);
4868 
4869 	/*
4870 	 * Being idle is no different than the initial window. If the cc
4871 	 * clamps it down below the initial window raise it to the initial
4872 	 * window.
4873 	 */
4874 	if (tp->snd_cwnd < i_cwnd) {
4875 		tp->snd_cwnd = i_cwnd;
4876 	}
4877 }
4878 
4879 /*
4880  * Indicate whether this ack should be delayed.  We can delay the ack if
4881  * following conditions are met:
4882  *	- There is no delayed ack timer in progress.
4883  *	- Our last ack wasn't a 0-sized window. We never want to delay
4884  *	  the ack that opens up a 0-sized window.
4885  *	- LRO wasn't used for this segment. We make sure by checking that the
4886  *	  segment size is not larger than the MSS.
4887  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
4888  *	  connection.
4889  */
4890 #define DELAY_ACK(tp, tlen)			 \
4891 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4892 	((tp->t_flags & TF_DELACK) == 0) &&	 \
4893 	(tlen <= tp->t_maxseg) &&		 \
4894 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4895 
4896 static struct rack_sendmap *
4897 rack_find_lowest_rsm(struct tcp_rack *rack)
4898 {
4899 	struct rack_sendmap *rsm;
4900 
4901 	/*
4902 	 * Walk the time-order transmitted list looking for an rsm that is
4903 	 * not acked. This will be the one that was sent the longest time
4904 	 * ago that is still outstanding.
4905 	 */
4906 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4907 		if (rsm->r_flags & RACK_ACKED) {
4908 			continue;
4909 		}
4910 		goto finish;
4911 	}
4912 finish:
4913 	return (rsm);
4914 }
4915 
4916 static struct rack_sendmap *
4917 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
4918 {
4919 	struct rack_sendmap *prsm;
4920 
4921 	/*
4922 	 * Walk the sequence order list backward until we hit and arrive at
4923 	 * the highest seq not acked. In theory when this is called it
4924 	 * should be the last segment (which it was not).
4925 	 */
4926 	prsm = rsm;
4927 	RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
4928 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
4929 			continue;
4930 		}
4931 		return (prsm);
4932 	}
4933 	return (NULL);
4934 }
4935 
4936 static uint32_t
4937 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
4938 {
4939 	int32_t lro;
4940 	uint32_t thresh;
4941 
4942 	/*
4943 	 * lro is the flag we use to determine if we have seen reordering.
4944 	 * If it gets set we have seen reordering. The reorder logic either
4945 	 * works in one of two ways:
4946 	 *
4947 	 * If reorder-fade is configured, then we track the last time we saw
4948 	 * re-ordering occur. If we reach the point where enough time as
4949 	 * passed we no longer consider reordering has occuring.
4950 	 *
4951 	 * Or if reorder-face is 0, then once we see reordering we consider
4952 	 * the connection to alway be subject to reordering and just set lro
4953 	 * to 1.
4954 	 *
4955 	 * In the end if lro is non-zero we add the extra time for
4956 	 * reordering in.
4957 	 */
4958 	if (srtt == 0)
4959 		srtt = 1;
4960 	if (rack->r_ctl.rc_reorder_ts) {
4961 		if (rack->r_ctl.rc_reorder_fade) {
4962 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
4963 				lro = cts - rack->r_ctl.rc_reorder_ts;
4964 				if (lro == 0) {
4965 					/*
4966 					 * No time as passed since the last
4967 					 * reorder, mark it as reordering.
4968 					 */
4969 					lro = 1;
4970 				}
4971 			} else {
4972 				/* Negative time? */
4973 				lro = 0;
4974 			}
4975 			if (lro > rack->r_ctl.rc_reorder_fade) {
4976 				/* Turn off reordering seen too */
4977 				rack->r_ctl.rc_reorder_ts = 0;
4978 				lro = 0;
4979 			}
4980 		} else {
4981 			/* Reodering does not fade */
4982 			lro = 1;
4983 		}
4984 	} else {
4985 		lro = 0;
4986 	}
4987 	if (rack->rc_rack_tmr_std_based == 0) {
4988 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
4989 	} else {
4990 		/* Standards based pkt-delay is 1/4 srtt */
4991 		thresh = srtt +  (srtt >> 2);
4992 	}
4993 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
4994 		/* It must be set, if not you get 1/4 rtt */
4995 		if (rack->r_ctl.rc_reorder_shift)
4996 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
4997 		else
4998 			thresh += (srtt >> 2);
4999 	}
5000 	if (rack->rc_rack_use_dsack &&
5001 	    lro &&
5002 	    (rack->r_ctl.num_dsack > 0)) {
5003 		/*
5004 		 * We only increase the reordering window if we
5005 		 * have seen reordering <and> we have a DSACK count.
5006 		 */
5007 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
5008 		rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
5009 	}
5010 	/* SRTT * 2 is the ceiling */
5011 	if (thresh > (srtt * 2)) {
5012 		thresh = srtt * 2;
5013 	}
5014 	/* And we don't want it above the RTO max either */
5015 	if (thresh > rack_rto_max) {
5016 		thresh = rack_rto_max;
5017 	}
5018 	rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
5019 	return (thresh);
5020 }
5021 
5022 static uint32_t
5023 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5024 		     struct rack_sendmap *rsm, uint32_t srtt)
5025 {
5026 	struct rack_sendmap *prsm;
5027 	uint32_t thresh, len;
5028 	int segsiz;
5029 
5030 	if (srtt == 0)
5031 		srtt = 1;
5032 	if (rack->r_ctl.rc_tlp_threshold)
5033 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5034 	else
5035 		thresh = (srtt * 2);
5036 
5037 	/* Get the previous sent packet, if any */
5038 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5039 	len = rsm->r_end - rsm->r_start;
5040 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5041 		/* Exactly like the ID */
5042 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5043 			uint32_t alt_thresh;
5044 			/*
5045 			 * Compensate for delayed-ack with the d-ack time.
5046 			 */
5047 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5048 			if (alt_thresh > thresh)
5049 				thresh = alt_thresh;
5050 		}
5051 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5052 		/* 2.1 behavior */
5053 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5054 		if (prsm && (len <= segsiz)) {
5055 			/*
5056 			 * Two packets outstanding, thresh should be (2*srtt) +
5057 			 * possible inter-packet delay (if any).
5058 			 */
5059 			uint32_t inter_gap = 0;
5060 			int idx, nidx;
5061 
5062 			idx = rsm->r_rtr_cnt - 1;
5063 			nidx = prsm->r_rtr_cnt - 1;
5064 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5065 				/* Yes it was sent later (or at the same time) */
5066 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5067 			}
5068 			thresh += inter_gap;
5069 		} else if (len <= segsiz) {
5070 			/*
5071 			 * Possibly compensate for delayed-ack.
5072 			 */
5073 			uint32_t alt_thresh;
5074 
5075 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5076 			if (alt_thresh > thresh)
5077 				thresh = alt_thresh;
5078 		}
5079 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5080 		/* 2.2 behavior */
5081 		if (len <= segsiz) {
5082 			uint32_t alt_thresh;
5083 			/*
5084 			 * Compensate for delayed-ack with the d-ack time.
5085 			 */
5086 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5087 			if (alt_thresh > thresh)
5088 				thresh = alt_thresh;
5089 		}
5090 	}
5091 	/* Not above an RTO */
5092 	if (thresh > tp->t_rxtcur) {
5093 		thresh = tp->t_rxtcur;
5094 	}
5095 	/* Not above a RTO max */
5096 	if (thresh > rack_rto_max) {
5097 		thresh = rack_rto_max;
5098 	}
5099 	/* Apply user supplied min TLP */
5100 	if (thresh < rack_tlp_min) {
5101 		thresh = rack_tlp_min;
5102 	}
5103 	return (thresh);
5104 }
5105 
5106 static uint32_t
5107 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5108 {
5109 	/*
5110 	 * We want the rack_rtt which is the
5111 	 * last rtt we measured. However if that
5112 	 * does not exist we fallback to the srtt (which
5113 	 * we probably will never do) and then as a last
5114 	 * resort we use RACK_INITIAL_RTO if no srtt is
5115 	 * yet set.
5116 	 */
5117 	if (rack->rc_rack_rtt)
5118 		return (rack->rc_rack_rtt);
5119 	else if (tp->t_srtt == 0)
5120 		return (RACK_INITIAL_RTO);
5121 	return (tp->t_srtt);
5122 }
5123 
5124 static struct rack_sendmap *
5125 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5126 {
5127 	/*
5128 	 * Check to see that we don't need to fall into recovery. We will
5129 	 * need to do so if our oldest transmit is past the time we should
5130 	 * have had an ack.
5131 	 */
5132 	struct tcp_rack *rack;
5133 	struct rack_sendmap *rsm;
5134 	int32_t idx;
5135 	uint32_t srtt, thresh;
5136 
5137 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5138 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5139 		return (NULL);
5140 	}
5141 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5142 	if (rsm == NULL)
5143 		return (NULL);
5144 
5145 
5146 	if (rsm->r_flags & RACK_ACKED) {
5147 		rsm = rack_find_lowest_rsm(rack);
5148 		if (rsm == NULL)
5149 			return (NULL);
5150 	}
5151 	idx = rsm->r_rtr_cnt - 1;
5152 	srtt = rack_grab_rtt(tp, rack);
5153 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5154 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5155 		return (NULL);
5156 	}
5157 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5158 		return (NULL);
5159 	}
5160 	/* Ok if we reach here we are over-due and this guy can be sent */
5161 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
5162 	return (rsm);
5163 }
5164 
5165 static uint32_t
5166 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5167 {
5168 	int32_t t;
5169 	int32_t tt;
5170 	uint32_t ret_val;
5171 
5172 	t = (tp->t_srtt + (tp->t_rttvar << 2));
5173 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5174  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5175 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5176 	ret_val = (uint32_t)tt;
5177 	return (ret_val);
5178 }
5179 
5180 static uint32_t
5181 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5182 {
5183 	/*
5184 	 * Start the FR timer, we do this based on getting the first one in
5185 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
5186 	 * events we need to stop the running timer (if its running) before
5187 	 * starting the new one.
5188 	 */
5189 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5190 	uint32_t srtt_cur;
5191 	int32_t idx;
5192 	int32_t is_tlp_timer = 0;
5193 	struct rack_sendmap *rsm;
5194 
5195 	if (rack->t_timers_stopped) {
5196 		/* All timers have been stopped none are to run */
5197 		return (0);
5198 	}
5199 	if (rack->rc_in_persist) {
5200 		/* We can't start any timer in persists */
5201 		return (rack_get_persists_timer_val(tp, rack));
5202 	}
5203 	rack->rc_on_min_to = 0;
5204 	if ((tp->t_state < TCPS_ESTABLISHED) ||
5205 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5206 		goto activate_rxt;
5207 	}
5208 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5209 	if ((rsm == NULL) || sup_rack) {
5210 		/* Nothing on the send map or no rack */
5211 activate_rxt:
5212 		time_since_sent = 0;
5213 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5214 		if (rsm) {
5215 			/*
5216 			 * Should we discount the RTX timer any?
5217 			 *
5218 			 * We want to discount it the smallest amount.
5219 			 * If a timer (Rack/TLP or RXT) has gone off more
5220 			 * recently thats the discount we want to use (now - timer time).
5221 			 * If the retransmit of the oldest packet was more recent then
5222 			 * we want to use that (now - oldest-packet-last_transmit_time).
5223 			 *
5224 			 */
5225 			idx = rsm->r_rtr_cnt - 1;
5226 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5227 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5228 			else
5229 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5230 			if (TSTMP_GT(cts, tstmp_touse))
5231 			    time_since_sent = cts - tstmp_touse;
5232 		}
5233 		if (SEQ_LT(tp->snd_una, tp->snd_max) ||
5234 		    sbavail(&tptosocket(tp)->so_snd)) {
5235 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5236 			to = tp->t_rxtcur;
5237 			if (to > time_since_sent)
5238 				to -= time_since_sent;
5239 			else
5240 				to = rack->r_ctl.rc_min_to;
5241 			if (to == 0)
5242 				to = 1;
5243 			/* Special case for KEEPINIT */
5244 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5245 			    (TP_KEEPINIT(tp) != 0) &&
5246 			    rsm) {
5247 				/*
5248 				 * We have to put a ceiling on the rxt timer
5249 				 * of the keep-init timeout.
5250 				 */
5251 				uint32_t max_time, red;
5252 
5253 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5254 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5255 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5256 					if (red < max_time)
5257 						max_time -= red;
5258 					else
5259 						max_time = 1;
5260 				}
5261 				/* Reduce timeout to the keep value if needed */
5262 				if (max_time < to)
5263 					to = max_time;
5264 			}
5265 			return (to);
5266 		}
5267 		return (0);
5268 	}
5269 	if (rsm->r_flags & RACK_ACKED) {
5270 		rsm = rack_find_lowest_rsm(rack);
5271 		if (rsm == NULL) {
5272 			/* No lowest? */
5273 			goto activate_rxt;
5274 		}
5275 	}
5276 	if (rack->sack_attack_disable) {
5277 		/*
5278 		 * We don't want to do
5279 		 * any TLP's if you are an attacker.
5280 		 * Though if you are doing what
5281 		 * is expected you may still have
5282 		 * SACK-PASSED marks.
5283 		 */
5284 		goto activate_rxt;
5285 	}
5286 	/* Convert from ms to usecs */
5287 	if ((rsm->r_flags & RACK_SACK_PASSED) ||
5288 	    (rsm->r_flags & RACK_RWND_COLLAPSED) ||
5289 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5290 		if ((tp->t_flags & TF_SENTFIN) &&
5291 		    ((tp->snd_max - tp->snd_una) == 1) &&
5292 		    (rsm->r_flags & RACK_HAS_FIN)) {
5293 			/*
5294 			 * We don't start a rack timer if all we have is a
5295 			 * FIN outstanding.
5296 			 */
5297 			goto activate_rxt;
5298 		}
5299 		if ((rack->use_rack_rr == 0) &&
5300 		    (IN_FASTRECOVERY(tp->t_flags)) &&
5301 		    (rack->rack_no_prr == 0) &&
5302 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5303 			/*
5304 			 * We are not cheating, in recovery  and
5305 			 * not enough ack's to yet get our next
5306 			 * retransmission out.
5307 			 *
5308 			 * Note that classified attackers do not
5309 			 * get to use the rack-cheat.
5310 			 */
5311 			goto activate_tlp;
5312 		}
5313 		srtt = rack_grab_rtt(tp, rack);
5314 		thresh = rack_calc_thresh_rack(rack, srtt, cts);
5315 		idx = rsm->r_rtr_cnt - 1;
5316 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5317 		if (SEQ_GEQ(exp, cts)) {
5318 			to = exp - cts;
5319 			if (to < rack->r_ctl.rc_min_to) {
5320 				to = rack->r_ctl.rc_min_to;
5321 				if (rack->r_rr_config == 3)
5322 					rack->rc_on_min_to = 1;
5323 			}
5324 		} else {
5325 			to = rack->r_ctl.rc_min_to;
5326 			if (rack->r_rr_config == 3)
5327 				rack->rc_on_min_to = 1;
5328 		}
5329 	} else {
5330 		/* Ok we need to do a TLP not RACK */
5331 activate_tlp:
5332 		if ((rack->rc_tlp_in_progress != 0) &&
5333 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5334 			/*
5335 			 * The previous send was a TLP and we have sent
5336 			 * N TLP's without sending new data.
5337 			 */
5338 			goto activate_rxt;
5339 		}
5340 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5341 		if (rsm == NULL) {
5342 			/* We found no rsm to TLP with. */
5343 			goto activate_rxt;
5344 		}
5345 		if (rsm->r_flags & RACK_HAS_FIN) {
5346 			/* If its a FIN we dont do TLP */
5347 			rsm = NULL;
5348 			goto activate_rxt;
5349 		}
5350 		idx = rsm->r_rtr_cnt - 1;
5351 		time_since_sent = 0;
5352 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5353 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5354 		else
5355 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5356 		if (TSTMP_GT(cts, tstmp_touse))
5357 		    time_since_sent = cts - tstmp_touse;
5358 		is_tlp_timer = 1;
5359 		if (tp->t_srtt) {
5360 			if ((rack->rc_srtt_measure_made == 0) &&
5361 			    (tp->t_srtt == 1)) {
5362 				/*
5363 				 * If another stack as run and set srtt to 1,
5364 				 * then the srtt was 0, so lets use the initial.
5365 				 */
5366 				srtt = RACK_INITIAL_RTO;
5367 			} else {
5368 				srtt_cur = tp->t_srtt;
5369 				srtt = srtt_cur;
5370 			}
5371 		} else
5372 			srtt = RACK_INITIAL_RTO;
5373 		/*
5374 		 * If the SRTT is not keeping up and the
5375 		 * rack RTT has spiked we want to use
5376 		 * the last RTT not the smoothed one.
5377 		 */
5378 		if (rack_tlp_use_greater &&
5379 		    tp->t_srtt &&
5380 		    (srtt < rack_grab_rtt(tp, rack))) {
5381 			srtt = rack_grab_rtt(tp, rack);
5382 		}
5383 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5384 		if (thresh > time_since_sent) {
5385 			to = thresh - time_since_sent;
5386 		} else {
5387 			to = rack->r_ctl.rc_min_to;
5388 			rack_log_alt_to_to_cancel(rack,
5389 						  thresh,		/* flex1 */
5390 						  time_since_sent,	/* flex2 */
5391 						  tstmp_touse,		/* flex3 */
5392 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5393 						  (uint32_t)rsm->r_tim_lastsent[idx],
5394 						  srtt,
5395 						  idx, 99);
5396 		}
5397 		if (to < rack_tlp_min) {
5398 			to = rack_tlp_min;
5399 		}
5400 		if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5401 			/*
5402 			 * If the TLP time works out to larger than the max
5403 			 * RTO lets not do TLP.. just RTO.
5404 			 */
5405 			goto activate_rxt;
5406 		}
5407 	}
5408 	if (is_tlp_timer == 0) {
5409 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5410 	} else {
5411 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5412 	}
5413 	if (to == 0)
5414 		to = 1;
5415 	return (to);
5416 }
5417 
5418 static void
5419 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5420 {
5421 	if (rack->rc_in_persist == 0) {
5422 		if (tp->t_flags & TF_GPUTINPROG) {
5423 			/*
5424 			 * Stop the goodput now, the calling of the
5425 			 * measurement function clears the flag.
5426 			 */
5427 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
5428 						    RACK_QUALITY_PERSIST);
5429 		}
5430 #ifdef NETFLIX_SHARED_CWND
5431 		if (rack->r_ctl.rc_scw) {
5432 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5433 			rack->rack_scwnd_is_idle = 1;
5434 		}
5435 #endif
5436 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5437 		if (rack->r_ctl.rc_went_idle_time == 0)
5438 			rack->r_ctl.rc_went_idle_time = 1;
5439 		rack_timer_cancel(tp, rack, cts, __LINE__);
5440 		rack->r_ctl.persist_lost_ends = 0;
5441 		rack->probe_not_answered = 0;
5442 		rack->forced_ack = 0;
5443 		tp->t_rxtshift = 0;
5444 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5445 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5446 		rack->rc_in_persist = 1;
5447 	}
5448 }
5449 
5450 static void
5451 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5452 {
5453 	if (tcp_in_hpts(rack->rc_inp)) {
5454 		tcp_hpts_remove(rack->rc_inp);
5455 		rack->r_ctl.rc_hpts_flags = 0;
5456 	}
5457 #ifdef NETFLIX_SHARED_CWND
5458 	if (rack->r_ctl.rc_scw) {
5459 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5460 		rack->rack_scwnd_is_idle = 0;
5461 	}
5462 #endif
5463 	if (rack->rc_gp_dyn_mul &&
5464 	    (rack->use_fixed_rate == 0) &&
5465 	    (rack->rc_always_pace)) {
5466 		/*
5467 		 * Do we count this as if a probe-rtt just
5468 		 * finished?
5469 		 */
5470 		uint32_t time_idle, idle_min;
5471 
5472 		time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5473 		idle_min = rack_min_probertt_hold;
5474 		if (rack_probertt_gpsrtt_cnt_div) {
5475 			uint64_t extra;
5476 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5477 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
5478 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5479 			idle_min += (uint32_t)extra;
5480 		}
5481 		if (time_idle >= idle_min) {
5482 			/* Yes, we count it as a probe-rtt. */
5483 			uint32_t us_cts;
5484 
5485 			us_cts = tcp_get_usecs(NULL);
5486 			if (rack->in_probe_rtt == 0) {
5487 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5488 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5489 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5490 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5491 			} else {
5492 				rack_exit_probertt(rack, us_cts);
5493 			}
5494 		}
5495 	}
5496 	rack->rc_in_persist = 0;
5497 	rack->r_ctl.rc_went_idle_time = 0;
5498 	tp->t_rxtshift = 0;
5499 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5500 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5501 	rack->r_ctl.rc_agg_delayed = 0;
5502 	rack->r_early = 0;
5503 	rack->r_late = 0;
5504 	rack->r_ctl.rc_agg_early = 0;
5505 }
5506 
5507 static void
5508 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5509 		   struct hpts_diag *diag, struct timeval *tv)
5510 {
5511 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5512 		union tcp_log_stackspecific log;
5513 
5514 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5515 		log.u_bbr.flex1 = diag->p_nxt_slot;
5516 		log.u_bbr.flex2 = diag->p_cur_slot;
5517 		log.u_bbr.flex3 = diag->slot_req;
5518 		log.u_bbr.flex4 = diag->inp_hptsslot;
5519 		log.u_bbr.flex5 = diag->slot_remaining;
5520 		log.u_bbr.flex6 = diag->need_new_to;
5521 		log.u_bbr.flex7 = diag->p_hpts_active;
5522 		log.u_bbr.flex8 = diag->p_on_min_sleep;
5523 		/* Hijack other fields as needed */
5524 		log.u_bbr.epoch = diag->have_slept;
5525 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
5526 		log.u_bbr.pkts_out = diag->co_ret;
5527 		log.u_bbr.applimited = diag->hpts_sleep_time;
5528 		log.u_bbr.delivered = diag->p_prev_slot;
5529 		log.u_bbr.inflight = diag->p_runningslot;
5530 		log.u_bbr.bw_inuse = diag->wheel_slot;
5531 		log.u_bbr.rttProp = diag->wheel_cts;
5532 		log.u_bbr.timeStamp = cts;
5533 		log.u_bbr.delRate = diag->maxslots;
5534 		log.u_bbr.cur_del_rate = diag->p_curtick;
5535 		log.u_bbr.cur_del_rate <<= 32;
5536 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
5537 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5538 		    &rack->rc_inp->inp_socket->so_rcv,
5539 		    &rack->rc_inp->inp_socket->so_snd,
5540 		    BBR_LOG_HPTSDIAG, 0,
5541 		    0, &log, false, tv);
5542 	}
5543 
5544 }
5545 
5546 static void
5547 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5548 {
5549 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5550 		union tcp_log_stackspecific log;
5551 		struct timeval tv;
5552 
5553 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5554 		log.u_bbr.flex1 = sb->sb_flags;
5555 		log.u_bbr.flex2 = len;
5556 		log.u_bbr.flex3 = sb->sb_state;
5557 		log.u_bbr.flex8 = type;
5558 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5559 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5560 		    &rack->rc_inp->inp_socket->so_rcv,
5561 		    &rack->rc_inp->inp_socket->so_snd,
5562 		    TCP_LOG_SB_WAKE, 0,
5563 		    len, &log, false, &tv);
5564 	}
5565 }
5566 
5567 static void
5568 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5569       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5570 {
5571 	struct hpts_diag diag;
5572 	struct inpcb *inp = tptoinpcb(tp);
5573 	struct timeval tv;
5574 	uint32_t delayed_ack = 0;
5575 	uint32_t hpts_timeout;
5576 	uint32_t entry_slot = slot;
5577 	uint8_t stopped;
5578 	uint32_t left = 0;
5579 	uint32_t us_cts;
5580 
5581 	if ((tp->t_state == TCPS_CLOSED) ||
5582 	    (tp->t_state == TCPS_LISTEN)) {
5583 		return;
5584 	}
5585 	if (tcp_in_hpts(inp)) {
5586 		/* Already on the pacer */
5587 		return;
5588 	}
5589 	stopped = rack->rc_tmr_stopped;
5590 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5591 		left = rack->r_ctl.rc_timer_exp - cts;
5592 	}
5593 	rack->r_ctl.rc_timer_exp = 0;
5594 	rack->r_ctl.rc_hpts_flags = 0;
5595 	us_cts = tcp_get_usecs(&tv);
5596 	/* Now early/late accounting */
5597 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
5598 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5599 		/*
5600 		 * We have a early carry over set,
5601 		 * we can always add more time so we
5602 		 * can always make this compensation.
5603 		 *
5604 		 * Note if ack's are allowed to wake us do not
5605 		 * penalize the next timer for being awoke
5606 		 * by an ack aka the rc_agg_early (non-paced mode).
5607 		 */
5608 		slot += rack->r_ctl.rc_agg_early;
5609 		rack->r_early = 0;
5610 		rack->r_ctl.rc_agg_early = 0;
5611 	}
5612 	if (rack->r_late) {
5613 		/*
5614 		 * This is harder, we can
5615 		 * compensate some but it
5616 		 * really depends on what
5617 		 * the current pacing time is.
5618 		 */
5619 		if (rack->r_ctl.rc_agg_delayed >= slot) {
5620 			/*
5621 			 * We can't compensate for it all.
5622 			 * And we have to have some time
5623 			 * on the clock. We always have a min
5624 			 * 10 slots (10 x 10 i.e. 100 usecs).
5625 			 */
5626 			if (slot <= HPTS_TICKS_PER_SLOT) {
5627 				/* We gain delay */
5628 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
5629 				slot = HPTS_TICKS_PER_SLOT;
5630 			} else {
5631 				/* We take off some */
5632 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
5633 				slot = HPTS_TICKS_PER_SLOT;
5634 			}
5635 		} else {
5636 			slot -= rack->r_ctl.rc_agg_delayed;
5637 			rack->r_ctl.rc_agg_delayed = 0;
5638 			/* Make sure we have 100 useconds at minimum */
5639 			if (slot < HPTS_TICKS_PER_SLOT) {
5640 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
5641 				slot = HPTS_TICKS_PER_SLOT;
5642 			}
5643 			if (rack->r_ctl.rc_agg_delayed == 0)
5644 				rack->r_late = 0;
5645 		}
5646 	}
5647 	if (slot) {
5648 		/* We are pacing too */
5649 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5650 	}
5651 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5652 #ifdef NETFLIX_EXP_DETECTION
5653 	if (rack->sack_attack_disable &&
5654 	    (slot < tcp_sad_pacing_interval)) {
5655 		/*
5656 		 * We have a potential attacker on
5657 		 * the line. We have possibly some
5658 		 * (or now) pacing time set. We want to
5659 		 * slow down the processing of sacks by some
5660 		 * amount (if it is an attacker). Set the default
5661 		 * slot for attackers in place (unless the orginal
5662 		 * interval is longer). Its stored in
5663 		 * micro-seconds, so lets convert to msecs.
5664 		 */
5665 		slot = tcp_sad_pacing_interval;
5666 	}
5667 #endif
5668 	if (tp->t_flags & TF_DELACK) {
5669 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
5670 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5671 	}
5672 	if (delayed_ack && ((hpts_timeout == 0) ||
5673 			    (delayed_ack < hpts_timeout)))
5674 		hpts_timeout = delayed_ack;
5675 	else
5676 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5677 	/*
5678 	 * If no timers are going to run and we will fall off the hptsi
5679 	 * wheel, we resort to a keep-alive timer if its configured.
5680 	 */
5681 	if ((hpts_timeout == 0) &&
5682 	    (slot == 0)) {
5683 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5684 		    (tp->t_state <= TCPS_CLOSING)) {
5685 			/*
5686 			 * Ok we have no timer (persists, rack, tlp, rxt  or
5687 			 * del-ack), we don't have segments being paced. So
5688 			 * all that is left is the keepalive timer.
5689 			 */
5690 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5691 				/* Get the established keep-alive time */
5692 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5693 			} else {
5694 				/*
5695 				 * Get the initial setup keep-alive time,
5696 				 * note that this is probably not going to
5697 				 * happen, since rack will be running a rxt timer
5698 				 * if a SYN of some sort is outstanding. It is
5699 				 * actually handled in rack_timeout_rxt().
5700 				 */
5701 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5702 			}
5703 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5704 			if (rack->in_probe_rtt) {
5705 				/*
5706 				 * We want to instead not wake up a long time from
5707 				 * now but to wake up about the time we would
5708 				 * exit probe-rtt and initiate a keep-alive ack.
5709 				 * This will get us out of probe-rtt and update
5710 				 * our min-rtt.
5711 				 */
5712 				hpts_timeout = rack_min_probertt_hold;
5713 			}
5714 		}
5715 	}
5716 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5717 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5718 		/*
5719 		 * RACK, TLP, persists and RXT timers all are restartable
5720 		 * based on actions input .. i.e we received a packet (ack
5721 		 * or sack) and that changes things (rw, or snd_una etc).
5722 		 * Thus we can restart them with a new value. For
5723 		 * keep-alive, delayed_ack we keep track of what was left
5724 		 * and restart the timer with a smaller value.
5725 		 */
5726 		if (left < hpts_timeout)
5727 			hpts_timeout = left;
5728 	}
5729 	if (hpts_timeout) {
5730 		/*
5731 		 * Hack alert for now we can't time-out over 2,147,483
5732 		 * seconds (a bit more than 596 hours), which is probably ok
5733 		 * :).
5734 		 */
5735 		if (hpts_timeout > 0x7ffffffe)
5736 			hpts_timeout = 0x7ffffffe;
5737 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5738 	}
5739 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
5740 	if ((rack->gp_ready == 0) &&
5741 	    (rack->use_fixed_rate == 0) &&
5742 	    (hpts_timeout < slot) &&
5743 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5744 		/*
5745 		 * We have no good estimate yet for the
5746 		 * old clunky burst mitigation or the
5747 		 * real pacing. And the tlp or rxt is smaller
5748 		 * than the pacing calculation. Lets not
5749 		 * pace that long since we know the calculation
5750 		 * so far is not accurate.
5751 		 */
5752 		slot = hpts_timeout;
5753 	}
5754 	/**
5755 	 * Turn off all the flags for queuing by default. The
5756 	 * flags have important meanings to what happens when
5757 	 * LRO interacts with the transport. Most likely (by default now)
5758 	 * mbuf_queueing and ack compression are on. So the transport
5759 	 * has a couple of flags that control what happens (if those
5760 	 * are not on then these flags won't have any effect since it
5761 	 * won't go through the queuing LRO path).
5762 	 *
5763 	 * INP_MBUF_QUEUE_READY - This flags says that I am busy
5764 	 *                        pacing output, so don't disturb. But
5765 	 *                        it also means LRO can wake me if there
5766 	 *                        is a SACK arrival.
5767 	 *
5768 	 * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5769 	 *                       with the above flag (QUEUE_READY) and
5770 	 *                       when present it says don't even wake me
5771 	 *                       if a SACK arrives.
5772 	 *
5773 	 * The idea behind these flags is that if we are pacing we
5774 	 * set the MBUF_QUEUE_READY and only get woken up if
5775 	 * a SACK arrives (which could change things) or if
5776 	 * our pacing timer expires. If, however, we have a rack
5777 	 * timer running, then we don't even want a sack to wake
5778 	 * us since the rack timer has to expire before we can send.
5779 	 *
5780 	 * Other cases should usually have none of the flags set
5781 	 * so LRO can call into us.
5782 	 */
5783 	inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5784 	if (slot) {
5785 		rack->r_ctl.rc_last_output_to = us_cts + slot;
5786 		/*
5787 		 * A pacing timer (slot) is being set, in
5788 		 * such a case we cannot send (we are blocked by
5789 		 * the timer). So lets tell LRO that it should not
5790 		 * wake us unless there is a SACK. Note this only
5791 		 * will be effective if mbuf queueing is on or
5792 		 * compressed acks are being processed.
5793 		 */
5794 		inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5795 		/*
5796 		 * But wait if we have a Rack timer running
5797 		 * even a SACK should not disturb us (with
5798 		 * the exception of r_rr_config 3).
5799 		 */
5800 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5801 		    (rack->r_rr_config != 3))
5802 			inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5803 		if (rack->rc_ack_can_sendout_data) {
5804 			/*
5805 			 * Ahh but wait, this is that special case
5806 			 * where the pacing timer can be disturbed
5807 			 * backout the changes (used for non-paced
5808 			 * burst limiting).
5809 			 */
5810 			inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5811 		}
5812 		if ((rack->use_rack_rr) &&
5813 		    (rack->r_rr_config < 2) &&
5814 		    ((hpts_timeout) && (hpts_timeout < slot))) {
5815 			/*
5816 			 * Arrange for the hpts to kick back in after the
5817 			 * t-o if the t-o does not cause a send.
5818 			 */
5819 			(void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
5820 						   __LINE__, &diag);
5821 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5822 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5823 		} else {
5824 			(void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(slot),
5825 						   __LINE__, &diag);
5826 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5827 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5828 		}
5829 	} else if (hpts_timeout) {
5830 		/*
5831 		 * With respect to inp_flags2 here, lets let any new acks wake
5832 		 * us up here. Since we are not pacing (no pacing timer), output
5833 		 * can happen so we should let it. If its a Rack timer, then any inbound
5834 		 * packet probably won't change the sending (we will be blocked)
5835 		 * but it may change the prr stats so letting it in (the set defaults
5836 		 * at the start of this block) are good enough.
5837 		 */
5838 		(void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
5839 					   __LINE__, &diag);
5840 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5841 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5842 	} else {
5843 		/* No timer starting */
5844 #ifdef INVARIANTS
5845 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5846 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5847 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5848 		}
5849 #endif
5850 	}
5851 	rack->rc_tmr_stopped = 0;
5852 	if (slot)
5853 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5854 }
5855 
5856 /*
5857  * RACK Timer, here we simply do logging and house keeping.
5858  * the normal rack_output() function will call the
5859  * appropriate thing to check if we need to do a RACK retransmit.
5860  * We return 1, saying don't proceed with rack_output only
5861  * when all timers have been stopped (destroyed PCB?).
5862  */
5863 static int
5864 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5865 {
5866 	/*
5867 	 * This timer simply provides an internal trigger to send out data.
5868 	 * The check_recovery_mode call will see if there are needed
5869 	 * retransmissions, if so we will enter fast-recovery. The output
5870 	 * call may or may not do the same thing depending on sysctl
5871 	 * settings.
5872 	 */
5873 	struct rack_sendmap *rsm;
5874 
5875 	counter_u64_add(rack_to_tot, 1);
5876 	if (rack->r_state && (rack->r_state != tp->t_state))
5877 		rack_set_state(tp, rack);
5878 	rack->rc_on_min_to = 0;
5879 	rsm = rack_check_recovery_mode(tp, cts);
5880 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5881 	if (rsm) {
5882 		rack->r_ctl.rc_resend = rsm;
5883 		rack->r_timer_override = 1;
5884 		if (rack->use_rack_rr) {
5885 			/*
5886 			 * Don't accumulate extra pacing delay
5887 			 * we are allowing the rack timer to
5888 			 * over-ride pacing i.e. rrr takes precedence
5889 			 * if the pacing interval is longer than the rrr
5890 			 * time (in other words we get the min pacing
5891 			 * time versus rrr pacing time).
5892 			 */
5893 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5894 		}
5895 	}
5896 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5897 	if (rsm == NULL) {
5898 		/* restart a timer and return 1 */
5899 		rack_start_hpts_timer(rack, tp, cts,
5900 				      0, 0, 0);
5901 		return (1);
5902 	}
5903 	return (0);
5904 }
5905 
5906 static void
5907 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
5908 {
5909 	if (rsm->m->m_len > rsm->orig_m_len) {
5910 		/*
5911 		 * Mbuf grew, caused by sbcompress, our offset does
5912 		 * not change.
5913 		 */
5914 		rsm->orig_m_len = rsm->m->m_len;
5915 	} else if (rsm->m->m_len < rsm->orig_m_len) {
5916 		/*
5917 		 * Mbuf shrank, trimmed off the top by an ack, our
5918 		 * offset changes.
5919 		 */
5920 		rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
5921 		rsm->orig_m_len = rsm->m->m_len;
5922 	}
5923 }
5924 
5925 static void
5926 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
5927 {
5928 	struct mbuf *m;
5929 	uint32_t soff;
5930 
5931 	if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
5932 		/* Fix up the orig_m_len and possibly the mbuf offset */
5933 		rack_adjust_orig_mlen(src_rsm);
5934 	}
5935 	m = src_rsm->m;
5936 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
5937 	while (soff >= m->m_len) {
5938 		/* Move out past this mbuf */
5939 		soff -= m->m_len;
5940 		m = m->m_next;
5941 		KASSERT((m != NULL),
5942 			("rsm:%p nrsm:%p hit at soff:%u null m",
5943 			 src_rsm, rsm, soff));
5944 	}
5945 	rsm->m = m;
5946 	rsm->soff = soff;
5947 	rsm->orig_m_len = m->m_len;
5948 }
5949 
5950 static __inline void
5951 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
5952 	       struct rack_sendmap *rsm, uint32_t start)
5953 {
5954 	int idx;
5955 
5956 	nrsm->r_start = start;
5957 	nrsm->r_end = rsm->r_end;
5958 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
5959 	nrsm->r_flags = rsm->r_flags;
5960 	nrsm->r_dupack = rsm->r_dupack;
5961 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
5962 	nrsm->r_rtr_bytes = 0;
5963 	nrsm->r_fas = rsm->r_fas;
5964 	rsm->r_end = nrsm->r_start;
5965 	nrsm->r_just_ret = rsm->r_just_ret;
5966 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
5967 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
5968 	}
5969 	/* Now if we have SYN flag we keep it on the left edge */
5970 	if (nrsm->r_flags & RACK_HAS_SYN)
5971 		nrsm->r_flags &= ~RACK_HAS_SYN;
5972 	/* Now if we have a FIN flag we keep it on the right edge */
5973 	if (rsm->r_flags & RACK_HAS_FIN)
5974 		rsm->r_flags &= ~RACK_HAS_FIN;
5975 	/* Push bit must go to the right edge as well */
5976 	if (rsm->r_flags & RACK_HAD_PUSH)
5977 		rsm->r_flags &= ~RACK_HAD_PUSH;
5978 	/* Clone over the state of the hw_tls flag */
5979 	nrsm->r_hw_tls = rsm->r_hw_tls;
5980 	/*
5981 	 * Now we need to find nrsm's new location in the mbuf chain
5982 	 * we basically calculate a new offset, which is soff +
5983 	 * how much is left in original rsm. Then we walk out the mbuf
5984 	 * chain to find the righ position, it may be the same mbuf
5985 	 * or maybe not.
5986 	 */
5987 	KASSERT(((rsm->m != NULL) ||
5988 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
5989 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
5990 	if (rsm->m)
5991 		rack_setup_offset_for_rsm(rsm, nrsm);
5992 }
5993 
5994 static struct rack_sendmap *
5995 rack_merge_rsm(struct tcp_rack *rack,
5996 	       struct rack_sendmap *l_rsm,
5997 	       struct rack_sendmap *r_rsm)
5998 {
5999 	/*
6000 	 * We are merging two ack'd RSM's,
6001 	 * the l_rsm is on the left (lower seq
6002 	 * values) and the r_rsm is on the right
6003 	 * (higher seq value). The simplest way
6004 	 * to merge these is to move the right
6005 	 * one into the left. I don't think there
6006 	 * is any reason we need to try to find
6007 	 * the oldest (or last oldest retransmitted).
6008 	 */
6009 #ifdef INVARIANTS
6010 	struct rack_sendmap *rm;
6011 #endif
6012 	rack_log_map_chg(rack->rc_tp, rack, NULL,
6013 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
6014 	l_rsm->r_end = r_rsm->r_end;
6015 	if (l_rsm->r_dupack < r_rsm->r_dupack)
6016 		l_rsm->r_dupack = r_rsm->r_dupack;
6017 	if (r_rsm->r_rtr_bytes)
6018 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
6019 	if (r_rsm->r_in_tmap) {
6020 		/* This really should not happen */
6021 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
6022 		r_rsm->r_in_tmap = 0;
6023 	}
6024 
6025 	/* Now the flags */
6026 	if (r_rsm->r_flags & RACK_HAS_FIN)
6027 		l_rsm->r_flags |= RACK_HAS_FIN;
6028 	if (r_rsm->r_flags & RACK_TLP)
6029 		l_rsm->r_flags |= RACK_TLP;
6030 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
6031 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
6032 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
6033 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6034 		/*
6035 		 * If both are app-limited then let the
6036 		 * free lower the count. If right is app
6037 		 * limited and left is not, transfer.
6038 		 */
6039 		l_rsm->r_flags |= RACK_APP_LIMITED;
6040 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
6041 		if (r_rsm == rack->r_ctl.rc_first_appl)
6042 			rack->r_ctl.rc_first_appl = l_rsm;
6043 	}
6044 #ifndef INVARIANTS
6045 	(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6046 #else
6047 	rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6048 	if (rm != r_rsm) {
6049 		panic("removing head in rack:%p rsm:%p rm:%p",
6050 		      rack, r_rsm, rm);
6051 	}
6052 #endif
6053 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6054 		/* Transfer the split limit to the map we free */
6055 		r_rsm->r_limit_type = l_rsm->r_limit_type;
6056 		l_rsm->r_limit_type = 0;
6057 	}
6058 	rack_free(rack, r_rsm);
6059 	return (l_rsm);
6060 }
6061 
6062 /*
6063  * TLP Timer, here we simply setup what segment we want to
6064  * have the TLP expire on, the normal rack_output() will then
6065  * send it out.
6066  *
6067  * We return 1, saying don't proceed with rack_output only
6068  * when all timers have been stopped (destroyed PCB?).
6069  */
6070 static int
6071 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
6072 {
6073 	/*
6074 	 * Tail Loss Probe.
6075 	 */
6076 	struct rack_sendmap *rsm = NULL;
6077 #ifdef INVARIANTS
6078 	struct rack_sendmap *insret;
6079 #endif
6080 	struct socket *so = tptosocket(tp);
6081 	uint32_t amm;
6082 	uint32_t out, avail;
6083 	int collapsed_win = 0;
6084 
6085 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6086 		/* Its not time yet */
6087 		return (0);
6088 	}
6089 	if (ctf_progress_timeout_check(tp, true)) {
6090 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6091 		return (-ETIMEDOUT);	/* tcp_drop() */
6092 	}
6093 	/*
6094 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
6095 	 * need to figure out how to force a full MSS segment out.
6096 	 */
6097 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6098 	rack->r_ctl.retran_during_recovery = 0;
6099 	rack->r_ctl.dsack_byte_cnt = 0;
6100 	counter_u64_add(rack_tlp_tot, 1);
6101 	if (rack->r_state && (rack->r_state != tp->t_state))
6102 		rack_set_state(tp, rack);
6103 	avail = sbavail(&so->so_snd);
6104 	out = tp->snd_max - tp->snd_una;
6105 	if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
6106 		/* special case, we need a retransmission */
6107 		collapsed_win = 1;
6108 		goto need_retran;
6109 	}
6110 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
6111 		rack->r_ctl.dsack_persist--;
6112 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6113 			rack->r_ctl.num_dsack = 0;
6114 		}
6115 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6116 	}
6117 	if ((tp->t_flags & TF_GPUTINPROG) &&
6118 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
6119 		/*
6120 		 * If this is the second in a row
6121 		 * TLP and we are doing a measurement
6122 		 * its time to abandon the measurement.
6123 		 * Something is likely broken on
6124 		 * the clients network and measuring a
6125 		 * broken network does us no good.
6126 		 */
6127 		tp->t_flags &= ~TF_GPUTINPROG;
6128 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6129 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
6130 					   tp->gput_seq,
6131 					   0, 0, 18, __LINE__, NULL, 0);
6132 	}
6133 	/*
6134 	 * Check our send oldest always settings, and if
6135 	 * there is an oldest to send jump to the need_retran.
6136 	 */
6137 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6138 		goto need_retran;
6139 
6140 	if (avail > out) {
6141 		/* New data is available */
6142 		amm = avail - out;
6143 		if (amm > ctf_fixed_maxseg(tp)) {
6144 			amm = ctf_fixed_maxseg(tp);
6145 			if ((amm + out) > tp->snd_wnd) {
6146 				/* We are rwnd limited */
6147 				goto need_retran;
6148 			}
6149 		} else if (amm < ctf_fixed_maxseg(tp)) {
6150 			/* not enough to fill a MTU */
6151 			goto need_retran;
6152 		}
6153 		if (IN_FASTRECOVERY(tp->t_flags)) {
6154 			/* Unlikely */
6155 			if (rack->rack_no_prr == 0) {
6156 				if (out + amm <= tp->snd_wnd) {
6157 					rack->r_ctl.rc_prr_sndcnt = amm;
6158 					rack->r_ctl.rc_tlp_new_data = amm;
6159 					rack_log_to_prr(rack, 4, 0, __LINE__);
6160 				}
6161 			} else
6162 				goto need_retran;
6163 		} else {
6164 			/* Set the send-new override */
6165 			if (out + amm <= tp->snd_wnd)
6166 				rack->r_ctl.rc_tlp_new_data = amm;
6167 			else
6168 				goto need_retran;
6169 		}
6170 		rack->r_ctl.rc_tlpsend = NULL;
6171 		counter_u64_add(rack_tlp_newdata, 1);
6172 		goto send;
6173 	}
6174 need_retran:
6175 	/*
6176 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
6177 	 * optionally the first un-acked segment.
6178 	 */
6179 	if (collapsed_win == 0) {
6180 		if (rack_always_send_oldest)
6181 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6182 		else {
6183 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6184 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6185 				rsm = rack_find_high_nonack(rack, rsm);
6186 			}
6187 		}
6188 		if (rsm == NULL) {
6189 #ifdef TCP_BLACKBOX
6190 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6191 #endif
6192 			goto out;
6193 		}
6194 	} else {
6195 		/*
6196 		 * We must find the last segment
6197 		 * that was acceptable by the client.
6198 		 */
6199 		RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6200 			if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6201 				/* Found one */
6202 				break;
6203 			}
6204 		}
6205 		if (rsm == NULL) {
6206 			/* None? if so send the first */
6207 			rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6208 			if (rsm == NULL) {
6209 #ifdef TCP_BLACKBOX
6210 				tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6211 #endif
6212 				goto out;
6213 			}
6214 		}
6215 	}
6216 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6217 		/*
6218 		 * We need to split this the last segment in two.
6219 		 */
6220 		struct rack_sendmap *nrsm;
6221 
6222 		nrsm = rack_alloc_full_limit(rack);
6223 		if (nrsm == NULL) {
6224 			/*
6225 			 * No memory to split, we will just exit and punt
6226 			 * off to the RXT timer.
6227 			 */
6228 			goto out;
6229 		}
6230 		rack_clone_rsm(rack, nrsm, rsm,
6231 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
6232 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6233 #ifndef INVARIANTS
6234 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6235 #else
6236 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6237 		if (insret != NULL) {
6238 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6239 			      nrsm, insret, rack, rsm);
6240 		}
6241 #endif
6242 		if (rsm->r_in_tmap) {
6243 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6244 			nrsm->r_in_tmap = 1;
6245 		}
6246 		rsm = nrsm;
6247 	}
6248 	rack->r_ctl.rc_tlpsend = rsm;
6249 send:
6250 	/* Make sure output path knows we are doing a TLP */
6251 	*doing_tlp = 1;
6252 	rack->r_timer_override = 1;
6253 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6254 	return (0);
6255 out:
6256 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6257 	return (0);
6258 }
6259 
6260 /*
6261  * Delayed ack Timer, here we simply need to setup the
6262  * ACK_NOW flag and remove the DELACK flag. From there
6263  * the output routine will send the ack out.
6264  *
6265  * We only return 1, saying don't proceed, if all timers
6266  * are stopped (destroyed PCB?).
6267  */
6268 static int
6269 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6270 {
6271 
6272 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6273 	tp->t_flags &= ~TF_DELACK;
6274 	tp->t_flags |= TF_ACKNOW;
6275 	KMOD_TCPSTAT_INC(tcps_delack);
6276 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6277 	return (0);
6278 }
6279 
6280 /*
6281  * Persists timer, here we simply send the
6282  * same thing as a keepalive will.
6283  * the one byte send.
6284  *
6285  * We only return 1, saying don't proceed, if all timers
6286  * are stopped (destroyed PCB?).
6287  */
6288 static int
6289 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6290 {
6291 	struct tcptemp *t_template;
6292 	int32_t retval = 1;
6293 
6294 	if (rack->rc_in_persist == 0)
6295 		return (0);
6296 	if (ctf_progress_timeout_check(tp, false)) {
6297 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6298 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6299 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6300 		return (-ETIMEDOUT);	/* tcp_drop() */
6301 	}
6302 	/*
6303 	 * Persistence timer into zero window. Force a byte to be output, if
6304 	 * possible.
6305 	 */
6306 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
6307 	/*
6308 	 * Hack: if the peer is dead/unreachable, we do not time out if the
6309 	 * window is closed.  After a full backoff, drop the connection if
6310 	 * the idle time (no responses to probes) reaches the maximum
6311 	 * backoff that we would use if retransmitting.
6312 	 */
6313 	if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6314 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6315 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6316 		KMOD_TCPSTAT_INC(tcps_persistdrop);
6317 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6318 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6319 		retval = -ETIMEDOUT;	/* tcp_drop() */
6320 		goto out;
6321 	}
6322 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6323 	    tp->snd_una == tp->snd_max)
6324 		rack_exit_persist(tp, rack, cts);
6325 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6326 	/*
6327 	 * If the user has closed the socket then drop a persisting
6328 	 * connection after a much reduced timeout.
6329 	 */
6330 	if (tp->t_state > TCPS_CLOSE_WAIT &&
6331 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6332 		KMOD_TCPSTAT_INC(tcps_persistdrop);
6333 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6334 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6335 		retval = -ETIMEDOUT;	/* tcp_drop() */
6336 		goto out;
6337 	}
6338 	t_template = tcpip_maketemplate(rack->rc_inp);
6339 	if (t_template) {
6340 		/* only set it if we were answered */
6341 		if (rack->forced_ack == 0) {
6342 			rack->forced_ack = 1;
6343 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6344 		} else {
6345 			rack->probe_not_answered = 1;
6346 			counter_u64_add(rack_persists_loss, 1);
6347 			rack->r_ctl.persist_lost_ends++;
6348 		}
6349 		counter_u64_add(rack_persists_sends, 1);
6350 		tcp_respond(tp, t_template->tt_ipgen,
6351 			    &t_template->tt_t, (struct mbuf *)NULL,
6352 			    tp->rcv_nxt, tp->snd_una - 1, 0);
6353 		/* This sends an ack */
6354 		if (tp->t_flags & TF_DELACK)
6355 			tp->t_flags &= ~TF_DELACK;
6356 		free(t_template, M_TEMP);
6357 	}
6358 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6359 		tp->t_rxtshift++;
6360 out:
6361 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6362 	rack_start_hpts_timer(rack, tp, cts,
6363 			      0, 0, 0);
6364 	return (retval);
6365 }
6366 
6367 /*
6368  * If a keepalive goes off, we had no other timers
6369  * happening. We always return 1 here since this
6370  * routine either drops the connection or sends
6371  * out a segment with respond.
6372  */
6373 static int
6374 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6375 {
6376 	struct tcptemp *t_template;
6377 	struct inpcb *inp = tptoinpcb(tp);
6378 
6379 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6380 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6381 	/*
6382 	 * Keep-alive timer went off; send something or drop connection if
6383 	 * idle for too long.
6384 	 */
6385 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
6386 	if (tp->t_state < TCPS_ESTABLISHED)
6387 		goto dropit;
6388 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6389 	    tp->t_state <= TCPS_CLOSING) {
6390 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6391 			goto dropit;
6392 		/*
6393 		 * Send a packet designed to force a response if the peer is
6394 		 * up and reachable: either an ACK if the connection is
6395 		 * still alive, or an RST if the peer has closed the
6396 		 * connection due to timeout or reboot. Using sequence
6397 		 * number tp->snd_una-1 causes the transmitted zero-length
6398 		 * segment to lie outside the receive window; by the
6399 		 * protocol spec, this requires the correspondent TCP to
6400 		 * respond.
6401 		 */
6402 		KMOD_TCPSTAT_INC(tcps_keepprobe);
6403 		t_template = tcpip_maketemplate(inp);
6404 		if (t_template) {
6405 			if (rack->forced_ack == 0) {
6406 				rack->forced_ack = 1;
6407 				rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6408 			} else {
6409 				rack->probe_not_answered = 1;
6410 			}
6411 			tcp_respond(tp, t_template->tt_ipgen,
6412 			    &t_template->tt_t, (struct mbuf *)NULL,
6413 			    tp->rcv_nxt, tp->snd_una - 1, 0);
6414 			free(t_template, M_TEMP);
6415 		}
6416 	}
6417 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6418 	return (1);
6419 dropit:
6420 	KMOD_TCPSTAT_INC(tcps_keepdrops);
6421 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6422 	return (-ETIMEDOUT);	/* tcp_drop() */
6423 }
6424 
6425 /*
6426  * Retransmit helper function, clear up all the ack
6427  * flags and take care of important book keeping.
6428  */
6429 static void
6430 rack_remxt_tmr(struct tcpcb *tp)
6431 {
6432 	/*
6433 	 * The retransmit timer went off, all sack'd blocks must be
6434 	 * un-acked.
6435 	 */
6436 	struct rack_sendmap *rsm, *trsm = NULL;
6437 	struct tcp_rack *rack;
6438 
6439 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6440 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6441 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6442 	if (rack->r_state && (rack->r_state != tp->t_state))
6443 		rack_set_state(tp, rack);
6444 	/*
6445 	 * Ideally we would like to be able to
6446 	 * mark SACK-PASS on anything not acked here.
6447 	 *
6448 	 * However, if we do that we would burst out
6449 	 * all that data 1ms apart. This would be unwise,
6450 	 * so for now we will just let the normal rxt timer
6451 	 * and tlp timer take care of it.
6452 	 *
6453 	 * Also we really need to stick them back in sequence
6454 	 * order. This way we send in the proper order and any
6455 	 * sacks that come floating in will "re-ack" the data.
6456 	 * To do this we zap the tmap with an INIT and then
6457 	 * walk through and place every rsm in the RB tree
6458 	 * back in its seq ordered place.
6459 	 */
6460 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
6461 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6462 		rsm->r_dupack = 0;
6463 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6464 		/* We must re-add it back to the tlist */
6465 		if (trsm == NULL) {
6466 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6467 		} else {
6468 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6469 		}
6470 		rsm->r_in_tmap = 1;
6471 		trsm = rsm;
6472 		if (rsm->r_flags & RACK_ACKED)
6473 			rsm->r_flags |= RACK_WAS_ACKED;
6474 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
6475 		rsm->r_flags |= RACK_MUST_RXT;
6476 	}
6477 	/* Clear the count (we just un-acked them) */
6478 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6479 	rack->r_ctl.rc_sacked = 0;
6480 	rack->r_ctl.rc_sacklast = NULL;
6481 	rack->r_ctl.rc_agg_delayed = 0;
6482 	rack->r_early = 0;
6483 	rack->r_ctl.rc_agg_early = 0;
6484 	rack->r_late = 0;
6485 	/* Clear the tlp rtx mark */
6486 	rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6487 	if (rack->r_ctl.rc_resend != NULL)
6488 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6489 	rack->r_ctl.rc_prr_sndcnt = 0;
6490 	rack_log_to_prr(rack, 6, 0, __LINE__);
6491 	rack->r_timer_override = 1;
6492 	if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6493 #ifdef NETFLIX_EXP_DETECTION
6494 	    || (rack->sack_attack_disable != 0)
6495 #endif
6496 		    ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6497 		/*
6498 		 * For non-sack customers new data
6499 		 * needs to go out as retransmits until
6500 		 * we retransmit up to snd_max.
6501 		 */
6502 		rack->r_must_retran = 1;
6503 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6504 						rack->r_ctl.rc_sacked);
6505 	}
6506 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6507 }
6508 
6509 static void
6510 rack_convert_rtts(struct tcpcb *tp)
6511 {
6512 	if (tp->t_srtt > 1) {
6513 		uint32_t val, frac;
6514 
6515 		val = tp->t_srtt >> TCP_RTT_SHIFT;
6516 		frac = tp->t_srtt & 0x1f;
6517 		tp->t_srtt = TICKS_2_USEC(val);
6518 		/*
6519 		 * frac is the fractional part of the srtt (if any)
6520 		 * but its in ticks and every bit represents
6521 		 * 1/32nd of a hz.
6522 		 */
6523 		if (frac) {
6524 			if (hz == 1000) {
6525 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6526 			} else {
6527 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6528 			}
6529 			tp->t_srtt += frac;
6530 		}
6531 	}
6532 	if (tp->t_rttvar) {
6533 		uint32_t val, frac;
6534 
6535 		val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6536 		frac = tp->t_rttvar & 0x1f;
6537 		tp->t_rttvar = TICKS_2_USEC(val);
6538 		/*
6539 		 * frac is the fractional part of the srtt (if any)
6540 		 * but its in ticks and every bit represents
6541 		 * 1/32nd of a hz.
6542 		 */
6543 		if (frac) {
6544 			if (hz == 1000) {
6545 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6546 			} else {
6547 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6548 			}
6549 			tp->t_rttvar += frac;
6550 		}
6551 	}
6552 	tp->t_rxtcur = RACK_REXMTVAL(tp);
6553 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6554 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6555 	}
6556 	if (tp->t_rxtcur > rack_rto_max) {
6557 		tp->t_rxtcur = rack_rto_max;
6558 	}
6559 }
6560 
6561 static void
6562 rack_cc_conn_init(struct tcpcb *tp)
6563 {
6564 	struct tcp_rack *rack;
6565 	uint32_t srtt;
6566 
6567 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6568 	srtt = tp->t_srtt;
6569 	cc_conn_init(tp);
6570 	/*
6571 	 * Now convert to rack's internal format,
6572 	 * if required.
6573 	 */
6574 	if ((srtt == 0) && (tp->t_srtt != 0))
6575 		rack_convert_rtts(tp);
6576 	/*
6577 	 * We want a chance to stay in slowstart as
6578 	 * we create a connection. TCP spec says that
6579 	 * initially ssthresh is infinite. For our
6580 	 * purposes that is the snd_wnd.
6581 	 */
6582 	if (tp->snd_ssthresh < tp->snd_wnd) {
6583 		tp->snd_ssthresh = tp->snd_wnd;
6584 	}
6585 	/*
6586 	 * We also want to assure a IW worth of
6587 	 * data can get inflight.
6588 	 */
6589 	if (rc_init_window(rack) < tp->snd_cwnd)
6590 		tp->snd_cwnd = rc_init_window(rack);
6591 }
6592 
6593 /*
6594  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6595  * we will setup to retransmit the lowest seq number outstanding.
6596  */
6597 static int
6598 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6599 {
6600 	struct inpcb *inp = tptoinpcb(tp);
6601 	int32_t rexmt;
6602 	int32_t retval = 0;
6603 	bool isipv6;
6604 
6605 	if ((tp->t_flags & TF_GPUTINPROG) &&
6606 	    (tp->t_rxtshift)) {
6607 		/*
6608 		 * We have had a second timeout
6609 		 * measurements on successive rxt's are not profitable.
6610 		 * It is unlikely to be of any use (the network is
6611 		 * broken or the client went away).
6612 		 */
6613 		tp->t_flags &= ~TF_GPUTINPROG;
6614 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6615 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
6616 					   tp->gput_seq,
6617 					   0, 0, 18, __LINE__, NULL, 0);
6618 	}
6619 	if (ctf_progress_timeout_check(tp, false)) {
6620 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6621 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6622 		return (-ETIMEDOUT);	/* tcp_drop() */
6623 	}
6624 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6625 	rack->r_ctl.retran_during_recovery = 0;
6626 	rack->rc_ack_required = 1;
6627 	rack->r_ctl.dsack_byte_cnt = 0;
6628 	if (IN_FASTRECOVERY(tp->t_flags))
6629 		tp->t_flags |= TF_WASFRECOVERY;
6630 	else
6631 		tp->t_flags &= ~TF_WASFRECOVERY;
6632 	if (IN_CONGRECOVERY(tp->t_flags))
6633 		tp->t_flags |= TF_WASCRECOVERY;
6634 	else
6635 		tp->t_flags &= ~TF_WASCRECOVERY;
6636 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6637 	    (tp->snd_una == tp->snd_max)) {
6638 		/* Nothing outstanding .. nothing to do */
6639 		return (0);
6640 	}
6641 	if (rack->r_ctl.dsack_persist) {
6642 		rack->r_ctl.dsack_persist--;
6643 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6644 			rack->r_ctl.num_dsack = 0;
6645 		}
6646 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6647 	}
6648 	/*
6649 	 * Rack can only run one timer  at a time, so we cannot
6650 	 * run a KEEPINIT (gating SYN sending) and a retransmit
6651 	 * timer for the SYN. So if we are in a front state and
6652 	 * have a KEEPINIT timer we need to check the first transmit
6653 	 * against now to see if we have exceeded the KEEPINIT time
6654 	 * (if one is set).
6655 	 */
6656 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6657 	    (TP_KEEPINIT(tp) != 0)) {
6658 		struct rack_sendmap *rsm;
6659 
6660 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6661 		if (rsm) {
6662 			/* Ok we have something outstanding to test keepinit with */
6663 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6664 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6665 				/* We have exceeded the KEEPINIT time */
6666 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6667 				goto drop_it;
6668 			}
6669 		}
6670 	}
6671 	/*
6672 	 * Retransmission timer went off.  Message has not been acked within
6673 	 * retransmit interval.  Back off to a longer retransmit interval
6674 	 * and retransmit one segment.
6675 	 */
6676 	rack_remxt_tmr(tp);
6677 	if ((rack->r_ctl.rc_resend == NULL) ||
6678 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6679 		/*
6680 		 * If the rwnd collapsed on
6681 		 * the one we are retransmitting
6682 		 * it does not count against the
6683 		 * rxt count.
6684 		 */
6685 		tp->t_rxtshift++;
6686 	}
6687 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6688 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6689 drop_it:
6690 		tp->t_rxtshift = TCP_MAXRXTSHIFT;
6691 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6692 		/* XXXGL: previously t_softerror was casted to uint16_t */
6693 		MPASS(tp->t_softerror >= 0);
6694 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
6695 		goto out;	/* tcp_drop() */
6696 	}
6697 	if (tp->t_state == TCPS_SYN_SENT) {
6698 		/*
6699 		 * If the SYN was retransmitted, indicate CWND to be limited
6700 		 * to 1 segment in cc_conn_init().
6701 		 */
6702 		tp->snd_cwnd = 1;
6703 	} else if (tp->t_rxtshift == 1) {
6704 		/*
6705 		 * first retransmit; record ssthresh and cwnd so they can be
6706 		 * recovered if this turns out to be a "bad" retransmit. A
6707 		 * retransmit is considered "bad" if an ACK for this segment
6708 		 * is received within RTT/2 interval; the assumption here is
6709 		 * that the ACK was already in flight.  See "On Estimating
6710 		 * End-to-End Network Path Properties" by Allman and Paxson
6711 		 * for more details.
6712 		 */
6713 		tp->snd_cwnd_prev = tp->snd_cwnd;
6714 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
6715 		tp->snd_recover_prev = tp->snd_recover;
6716 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6717 		tp->t_flags |= TF_PREVVALID;
6718 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6719 		tp->t_flags &= ~TF_PREVVALID;
6720 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6721 	if ((tp->t_state == TCPS_SYN_SENT) ||
6722 	    (tp->t_state == TCPS_SYN_RECEIVED))
6723 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6724 	else
6725 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6726 
6727 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6728 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6729 	/*
6730 	 * We enter the path for PLMTUD if connection is established or, if
6731 	 * connection is FIN_WAIT_1 status, reason for the last is that if
6732 	 * amount of data we send is very small, we could send it in couple
6733 	 * of packets and process straight to FIN. In that case we won't
6734 	 * catch ESTABLISHED state.
6735 	 */
6736 #ifdef INET6
6737 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
6738 #else
6739 	isipv6 = false;
6740 #endif
6741 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
6742 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6743 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6744 	    ((tp->t_state == TCPS_ESTABLISHED) ||
6745 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
6746 		/*
6747 		 * Idea here is that at each stage of mtu probe (usually,
6748 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
6749 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6750 		 * should take care of that.
6751 		 */
6752 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6753 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6754 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6755 		    tp->t_rxtshift % 2 == 0)) {
6756 			/*
6757 			 * Enter Path MTU Black-hole Detection mechanism: -
6758 			 * Disable Path MTU Discovery (IP "DF" bit). -
6759 			 * Reduce MTU to lower value than what we negotiated
6760 			 * with peer.
6761 			 */
6762 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6763 				/* Record that we may have found a black hole. */
6764 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6765 				/* Keep track of previous MSS. */
6766 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6767 			}
6768 
6769 			/*
6770 			 * Reduce the MSS to blackhole value or to the
6771 			 * default in an attempt to retransmit.
6772 			 */
6773 #ifdef INET6
6774 			if (isipv6 &&
6775 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6776 				/* Use the sysctl tuneable blackhole MSS. */
6777 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6778 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6779 			} else if (isipv6) {
6780 				/* Use the default MSS. */
6781 				tp->t_maxseg = V_tcp_v6mssdflt;
6782 				/*
6783 				 * Disable Path MTU Discovery when we switch
6784 				 * to minmss.
6785 				 */
6786 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6787 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6788 			}
6789 #endif
6790 #if defined(INET6) && defined(INET)
6791 			else
6792 #endif
6793 #ifdef INET
6794 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6795 				/* Use the sysctl tuneable blackhole MSS. */
6796 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6797 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6798 			} else {
6799 				/* Use the default MSS. */
6800 				tp->t_maxseg = V_tcp_mssdflt;
6801 				/*
6802 				 * Disable Path MTU Discovery when we switch
6803 				 * to minmss.
6804 				 */
6805 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6806 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6807 			}
6808 #endif
6809 		} else {
6810 			/*
6811 			 * If further retransmissions are still unsuccessful
6812 			 * with a lowered MTU, maybe this isn't a blackhole
6813 			 * and we restore the previous MSS and blackhole
6814 			 * detection flags. The limit '6' is determined by
6815 			 * giving each probe stage (1448, 1188, 524) 2
6816 			 * chances to recover.
6817 			 */
6818 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6819 			    (tp->t_rxtshift >= 6)) {
6820 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6821 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6822 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6823 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6824 			}
6825 		}
6826 	}
6827 	/*
6828 	 * Disable RFC1323 and SACK if we haven't got any response to
6829 	 * our third SYN to work-around some broken terminal servers
6830 	 * (most of which have hopefully been retired) that have bad VJ
6831 	 * header compression code which trashes TCP segments containing
6832 	 * unknown-to-them TCP options.
6833 	 */
6834 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6835 	    (tp->t_rxtshift == 3))
6836 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6837 	/*
6838 	 * If we backed off this far, our srtt estimate is probably bogus.
6839 	 * Clobber it so we'll take the next rtt measurement as our srtt;
6840 	 * move the current srtt into rttvar to keep the current retransmit
6841 	 * times until then.
6842 	 */
6843 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6844 #ifdef INET6
6845 		if ((inp->inp_vflag & INP_IPV6) != 0)
6846 			in6_losing(inp);
6847 		else
6848 #endif
6849 			in_losing(inp);
6850 		tp->t_rttvar += tp->t_srtt;
6851 		tp->t_srtt = 0;
6852 	}
6853 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6854 	tp->snd_recover = tp->snd_max;
6855 	tp->t_flags |= TF_ACKNOW;
6856 	tp->t_rtttime = 0;
6857 	rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
6858 out:
6859 	return (retval);
6860 }
6861 
6862 static int
6863 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
6864 {
6865 	int32_t ret = 0;
6866 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6867 
6868 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
6869 	    (tp->t_flags & TF_GPUTINPROG)) {
6870 		/*
6871 		 * We have a goodput in progress
6872 		 * and we have entered a late state.
6873 		 * Do we have enough data in the sb
6874 		 * to handle the GPUT request?
6875 		 */
6876 		uint32_t bytes;
6877 
6878 		bytes = tp->gput_ack - tp->gput_seq;
6879 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
6880 			bytes += tp->gput_seq - tp->snd_una;
6881 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
6882 			/*
6883 			 * There are not enough bytes in the socket
6884 			 * buffer that have been sent to cover this
6885 			 * measurement. Cancel it.
6886 			 */
6887 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6888 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
6889 						   tp->gput_seq,
6890 						   0, 0, 18, __LINE__, NULL, 0);
6891 			tp->t_flags &= ~TF_GPUTINPROG;
6892 		}
6893 	}
6894 	if (timers == 0) {
6895 		return (0);
6896 	}
6897 	if (tp->t_state == TCPS_LISTEN) {
6898 		/* no timers on listen sockets */
6899 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6900 			return (0);
6901 		return (1);
6902 	}
6903 	if ((timers & PACE_TMR_RACK) &&
6904 	    rack->rc_on_min_to) {
6905 		/*
6906 		 * For the rack timer when we
6907 		 * are on a min-timeout (which means rrr_conf = 3)
6908 		 * we don't want to check the timer. It may
6909 		 * be going off for a pace and thats ok we
6910 		 * want to send the retransmit (if its ready).
6911 		 *
6912 		 * If its on a normal rack timer (non-min) then
6913 		 * we will check if its expired.
6914 		 */
6915 		goto skip_time_check;
6916 	}
6917 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6918 		uint32_t left;
6919 
6920 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6921 			ret = -1;
6922 			rack_log_to_processing(rack, cts, ret, 0);
6923 			return (0);
6924 		}
6925 		if (hpts_calling == 0) {
6926 			/*
6927 			 * A user send or queued mbuf (sack) has called us? We
6928 			 * return 0 and let the pacing guards
6929 			 * deal with it if they should or
6930 			 * should not cause a send.
6931 			 */
6932 			ret = -2;
6933 			rack_log_to_processing(rack, cts, ret, 0);
6934 			return (0);
6935 		}
6936 		/*
6937 		 * Ok our timer went off early and we are not paced false
6938 		 * alarm, go back to sleep.
6939 		 */
6940 		ret = -3;
6941 		left = rack->r_ctl.rc_timer_exp - cts;
6942 		tcp_hpts_insert(tptoinpcb(tp), HPTS_MS_TO_SLOTS(left));
6943 		rack_log_to_processing(rack, cts, ret, left);
6944 		return (1);
6945 	}
6946 skip_time_check:
6947 	rack->rc_tmr_stopped = 0;
6948 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
6949 	if (timers & PACE_TMR_DELACK) {
6950 		ret = rack_timeout_delack(tp, rack, cts);
6951 	} else if (timers & PACE_TMR_RACK) {
6952 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
6953 		rack->r_fast_output = 0;
6954 		ret = rack_timeout_rack(tp, rack, cts);
6955 	} else if (timers & PACE_TMR_TLP) {
6956 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
6957 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
6958 	} else if (timers & PACE_TMR_RXT) {
6959 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
6960 		rack->r_fast_output = 0;
6961 		ret = rack_timeout_rxt(tp, rack, cts);
6962 	} else if (timers & PACE_TMR_PERSIT) {
6963 		ret = rack_timeout_persist(tp, rack, cts);
6964 	} else if (timers & PACE_TMR_KEEP) {
6965 		ret = rack_timeout_keepalive(tp, rack, cts);
6966 	}
6967 	rack_log_to_processing(rack, cts, ret, timers);
6968 	return (ret);
6969 }
6970 
6971 static void
6972 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
6973 {
6974 	struct timeval tv;
6975 	uint32_t us_cts, flags_on_entry;
6976 	uint8_t hpts_removed = 0;
6977 
6978 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
6979 	us_cts = tcp_get_usecs(&tv);
6980 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
6981 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
6982 	     ((tp->snd_max - tp->snd_una) == 0))) {
6983 		tcp_hpts_remove(rack->rc_inp);
6984 		hpts_removed = 1;
6985 		/* If we were not delayed cancel out the flag. */
6986 		if ((tp->snd_max - tp->snd_una) == 0)
6987 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6988 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
6989 	}
6990 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
6991 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
6992 		if (tcp_in_hpts(rack->rc_inp) &&
6993 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
6994 			/*
6995 			 * Canceling timer's when we have no output being
6996 			 * paced. We also must remove ourselves from the
6997 			 * hpts.
6998 			 */
6999 			tcp_hpts_remove(rack->rc_inp);
7000 			hpts_removed = 1;
7001 		}
7002 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7003 	}
7004 	if (hpts_removed == 0)
7005 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7006 }
7007 
7008 static int
7009 rack_stopall(struct tcpcb *tp)
7010 {
7011 	struct tcp_rack *rack;
7012 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7013 	rack->t_timers_stopped = 1;
7014 	return (0);
7015 }
7016 
7017 static void
7018 rack_stop_all_timers(struct tcpcb *tp)
7019 {
7020 	struct tcp_rack *rack;
7021 
7022 	/*
7023 	 * Assure no timers are running.
7024 	 */
7025 	if (tcp_timer_active(tp, TT_PERSIST)) {
7026 		/* We enter in persists, set the flag appropriately */
7027 		rack = (struct tcp_rack *)tp->t_fb_ptr;
7028 		rack->rc_in_persist = 1;
7029 	}
7030 }
7031 
7032 static void
7033 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7034     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7035 {
7036 	int32_t idx;
7037 
7038 	rsm->r_rtr_cnt++;
7039 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7040 	rsm->r_dupack = 0;
7041 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7042 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7043 		rsm->r_flags |= RACK_OVERMAX;
7044 	}
7045 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7046 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7047 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7048 	}
7049 	idx = rsm->r_rtr_cnt - 1;
7050 	rsm->r_tim_lastsent[idx] = ts;
7051 	/*
7052 	 * Here we don't add in the len of send, since its already
7053 	 * in snduna <->snd_max.
7054 	 */
7055 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
7056 				     rack->r_ctl.rc_sacked);
7057 	if (rsm->r_flags & RACK_ACKED) {
7058 		/* Problably MTU discovery messing with us */
7059 		rsm->r_flags &= ~RACK_ACKED;
7060 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7061 	}
7062 	if (rsm->r_in_tmap) {
7063 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7064 		rsm->r_in_tmap = 0;
7065 	}
7066 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7067 	rsm->r_in_tmap = 1;
7068 	/* Take off the must retransmit flag, if its on */
7069 	if (rsm->r_flags & RACK_MUST_RXT) {
7070 		if (rack->r_must_retran)
7071 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
7072 		if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
7073 			/*
7074 			 * We have retransmitted all we need. Clear
7075 			 * any must retransmit flags.
7076 			 */
7077 			rack->r_must_retran = 0;
7078 			rack->r_ctl.rc_out_at_rto = 0;
7079 		}
7080 		rsm->r_flags &= ~RACK_MUST_RXT;
7081 	}
7082 	if (rsm->r_flags & RACK_SACK_PASSED) {
7083 		/* We have retransmitted due to the SACK pass */
7084 		rsm->r_flags &= ~RACK_SACK_PASSED;
7085 		rsm->r_flags |= RACK_WAS_SACKPASS;
7086 	}
7087 }
7088 
7089 static uint32_t
7090 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7091     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7092 {
7093 	/*
7094 	 * We (re-)transmitted starting at rsm->r_start for some length
7095 	 * (possibly less than r_end.
7096 	 */
7097 	struct rack_sendmap *nrsm;
7098 #ifdef INVARIANTS
7099 	struct rack_sendmap *insret;
7100 #endif
7101 	uint32_t c_end;
7102 	int32_t len;
7103 
7104 	len = *lenp;
7105 	c_end = rsm->r_start + len;
7106 	if (SEQ_GEQ(c_end, rsm->r_end)) {
7107 		/*
7108 		 * We retransmitted the whole piece or more than the whole
7109 		 * slopping into the next rsm.
7110 		 */
7111 		rack_update_rsm(tp, rack, rsm, ts, add_flag);
7112 		if (c_end == rsm->r_end) {
7113 			*lenp = 0;
7114 			return (0);
7115 		} else {
7116 			int32_t act_len;
7117 
7118 			/* Hangs over the end return whats left */
7119 			act_len = rsm->r_end - rsm->r_start;
7120 			*lenp = (len - act_len);
7121 			return (rsm->r_end);
7122 		}
7123 		/* We don't get out of this block. */
7124 	}
7125 	/*
7126 	 * Here we retransmitted less than the whole thing which means we
7127 	 * have to split this into what was transmitted and what was not.
7128 	 */
7129 	nrsm = rack_alloc_full_limit(rack);
7130 	if (nrsm == NULL) {
7131 		/*
7132 		 * We can't get memory, so lets not proceed.
7133 		 */
7134 		*lenp = 0;
7135 		return (0);
7136 	}
7137 	/*
7138 	 * So here we are going to take the original rsm and make it what we
7139 	 * retransmitted. nrsm will be the tail portion we did not
7140 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7141 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7142 	 * 1, 6 and the new piece will be 6, 11.
7143 	 */
7144 	rack_clone_rsm(rack, nrsm, rsm, c_end);
7145 	nrsm->r_dupack = 0;
7146 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7147 #ifndef INVARIANTS
7148 	(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7149 #else
7150 	insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7151 	if (insret != NULL) {
7152 		panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7153 		      nrsm, insret, rack, rsm);
7154 	}
7155 #endif
7156 	if (rsm->r_in_tmap) {
7157 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7158 		nrsm->r_in_tmap = 1;
7159 	}
7160 	rsm->r_flags &= (~RACK_HAS_FIN);
7161 	rack_update_rsm(tp, rack, rsm, ts, add_flag);
7162 	/* Log a split of rsm into rsm and nrsm */
7163 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7164 	*lenp = 0;
7165 	return (0);
7166 }
7167 
7168 static void
7169 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7170 		uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
7171 		struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls)
7172 {
7173 	struct tcp_rack *rack;
7174 	struct rack_sendmap *rsm, *nrsm, fe;
7175 #ifdef INVARIANTS
7176 	struct rack_sendmap *insret;
7177 #endif
7178 	register uint32_t snd_max, snd_una;
7179 
7180 	/*
7181 	 * Add to the RACK log of packets in flight or retransmitted. If
7182 	 * there is a TS option we will use the TS echoed, if not we will
7183 	 * grab a TS.
7184 	 *
7185 	 * Retransmissions will increment the count and move the ts to its
7186 	 * proper place. Note that if options do not include TS's then we
7187 	 * won't be able to effectively use the ACK for an RTT on a retran.
7188 	 *
7189 	 * Notes about r_start and r_end. Lets consider a send starting at
7190 	 * sequence 1 for 10 bytes. In such an example the r_start would be
7191 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7192 	 * This means that r_end is actually the first sequence for the next
7193 	 * slot (11).
7194 	 *
7195 	 */
7196 	/*
7197 	 * If err is set what do we do XXXrrs? should we not add the thing?
7198 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
7199 	 * i.e. proceed with add ** do this for now.
7200 	 */
7201 	INP_WLOCK_ASSERT(tptoinpcb(tp));
7202 	if (err)
7203 		/*
7204 		 * We don't log errors -- we could but snd_max does not
7205 		 * advance in this case either.
7206 		 */
7207 		return;
7208 
7209 	if (th_flags & TH_RST) {
7210 		/*
7211 		 * We don't log resets and we return immediately from
7212 		 * sending
7213 		 */
7214 		return;
7215 	}
7216 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7217 	snd_una = tp->snd_una;
7218 	snd_max = tp->snd_max;
7219 	if (th_flags & (TH_SYN | TH_FIN)) {
7220 		/*
7221 		 * The call to rack_log_output is made before bumping
7222 		 * snd_max. This means we can record one extra byte on a SYN
7223 		 * or FIN if seq_out is adding more on and a FIN is present
7224 		 * (and we are not resending).
7225 		 */
7226 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7227 			len++;
7228 		if (th_flags & TH_FIN)
7229 			len++;
7230 		if (SEQ_LT(snd_max, tp->snd_nxt)) {
7231 			/*
7232 			 * The add/update as not been done for the FIN/SYN
7233 			 * yet.
7234 			 */
7235 			snd_max = tp->snd_nxt;
7236 		}
7237 	}
7238 	if (SEQ_LEQ((seq_out + len), snd_una)) {
7239 		/* Are sending an old segment to induce an ack (keep-alive)? */
7240 		return;
7241 	}
7242 	if (SEQ_LT(seq_out, snd_una)) {
7243 		/* huh? should we panic? */
7244 		uint32_t end;
7245 
7246 		end = seq_out + len;
7247 		seq_out = snd_una;
7248 		if (SEQ_GEQ(end, seq_out))
7249 			len = end - seq_out;
7250 		else
7251 			len = 0;
7252 	}
7253 	if (len == 0) {
7254 		/* We don't log zero window probes */
7255 		return;
7256 	}
7257 	if (IN_FASTRECOVERY(tp->t_flags)) {
7258 		rack->r_ctl.rc_prr_out += len;
7259 	}
7260 	/* First question is it a retransmission or new? */
7261 	if (seq_out == snd_max) {
7262 		/* Its new */
7263 again:
7264 		rsm = rack_alloc(rack);
7265 		if (rsm == NULL) {
7266 			/*
7267 			 * Hmm out of memory and the tcb got destroyed while
7268 			 * we tried to wait.
7269 			 */
7270 			return;
7271 		}
7272 		if (th_flags & TH_FIN) {
7273 			rsm->r_flags = RACK_HAS_FIN|add_flag;
7274 		} else {
7275 			rsm->r_flags = add_flag;
7276 		}
7277 		if (hw_tls)
7278 			rsm->r_hw_tls = 1;
7279 		rsm->r_tim_lastsent[0] = cts;
7280 		rsm->r_rtr_cnt = 1;
7281 		rsm->r_rtr_bytes = 0;
7282 		if (th_flags & TH_SYN) {
7283 			/* The data space is one beyond snd_una */
7284 			rsm->r_flags |= RACK_HAS_SYN;
7285 		}
7286 		rsm->r_start = seq_out;
7287 		rsm->r_end = rsm->r_start + len;
7288 		rsm->r_dupack = 0;
7289 		/*
7290 		 * save off the mbuf location that
7291 		 * sndmbuf_noadv returned (which is
7292 		 * where we started copying from)..
7293 		 */
7294 		rsm->m = s_mb;
7295 		rsm->soff = s_moff;
7296 		/*
7297 		 * Here we do add in the len of send, since its not yet
7298 		 * reflected in in snduna <->snd_max
7299 		 */
7300 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
7301 					      rack->r_ctl.rc_sacked) +
7302 			      (rsm->r_end - rsm->r_start));
7303 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7304 		if (rsm->m) {
7305 			if (rsm->m->m_len <= rsm->soff) {
7306 				/*
7307 				 * XXXrrs Question, will this happen?
7308 				 *
7309 				 * If sbsndptr is set at the correct place
7310 				 * then s_moff should always be somewhere
7311 				 * within rsm->m. But if the sbsndptr was
7312 				 * off then that won't be true. If it occurs
7313 				 * we need to walkout to the correct location.
7314 				 */
7315 				struct mbuf *lm;
7316 
7317 				lm = rsm->m;
7318 				while (lm->m_len <= rsm->soff) {
7319 					rsm->soff -= lm->m_len;
7320 					lm = lm->m_next;
7321 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7322 							     __func__, rack, s_moff, s_mb, rsm->soff));
7323 				}
7324 				rsm->m = lm;
7325 			}
7326 			rsm->orig_m_len = rsm->m->m_len;
7327 		} else
7328 			rsm->orig_m_len = 0;
7329 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7330 		/* Log a new rsm */
7331 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7332 #ifndef INVARIANTS
7333 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7334 #else
7335 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7336 		if (insret != NULL) {
7337 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7338 			      nrsm, insret, rack, rsm);
7339 		}
7340 #endif
7341 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7342 		rsm->r_in_tmap = 1;
7343 		/*
7344 		 * Special case detection, is there just a single
7345 		 * packet outstanding when we are not in recovery?
7346 		 *
7347 		 * If this is true mark it so.
7348 		 */
7349 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7350 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7351 			struct rack_sendmap *prsm;
7352 
7353 			prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7354 			if (prsm)
7355 				prsm->r_one_out_nr = 1;
7356 		}
7357 		return;
7358 	}
7359 	/*
7360 	 * If we reach here its a retransmission and we need to find it.
7361 	 */
7362 	memset(&fe, 0, sizeof(fe));
7363 more:
7364 	if (hintrsm && (hintrsm->r_start == seq_out)) {
7365 		rsm = hintrsm;
7366 		hintrsm = NULL;
7367 	} else {
7368 		/* No hints sorry */
7369 		rsm = NULL;
7370 	}
7371 	if ((rsm) && (rsm->r_start == seq_out)) {
7372 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7373 		if (len == 0) {
7374 			return;
7375 		} else {
7376 			goto more;
7377 		}
7378 	}
7379 	/* Ok it was not the last pointer go through it the hard way. */
7380 refind:
7381 	fe.r_start = seq_out;
7382 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7383 	if (rsm) {
7384 		if (rsm->r_start == seq_out) {
7385 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7386 			if (len == 0) {
7387 				return;
7388 			} else {
7389 				goto refind;
7390 			}
7391 		}
7392 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7393 			/* Transmitted within this piece */
7394 			/*
7395 			 * Ok we must split off the front and then let the
7396 			 * update do the rest
7397 			 */
7398 			nrsm = rack_alloc_full_limit(rack);
7399 			if (nrsm == NULL) {
7400 				rack_update_rsm(tp, rack, rsm, cts, add_flag);
7401 				return;
7402 			}
7403 			/*
7404 			 * copy rsm to nrsm and then trim the front of rsm
7405 			 * to not include this part.
7406 			 */
7407 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
7408 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7409 #ifndef INVARIANTS
7410 			(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7411 #else
7412 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7413 			if (insret != NULL) {
7414 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7415 				      nrsm, insret, rack, rsm);
7416 			}
7417 #endif
7418 			if (rsm->r_in_tmap) {
7419 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7420 				nrsm->r_in_tmap = 1;
7421 			}
7422 			rsm->r_flags &= (~RACK_HAS_FIN);
7423 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7424 			if (len == 0) {
7425 				return;
7426 			} else if (len > 0)
7427 				goto refind;
7428 		}
7429 	}
7430 	/*
7431 	 * Hmm not found in map did they retransmit both old and on into the
7432 	 * new?
7433 	 */
7434 	if (seq_out == tp->snd_max) {
7435 		goto again;
7436 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
7437 #ifdef INVARIANTS
7438 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7439 		       seq_out, len, tp->snd_una, tp->snd_max);
7440 		printf("Starting Dump of all rack entries\n");
7441 		RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7442 			printf("rsm:%p start:%u end:%u\n",
7443 			       rsm, rsm->r_start, rsm->r_end);
7444 		}
7445 		printf("Dump complete\n");
7446 		panic("seq_out not found rack:%p tp:%p",
7447 		      rack, tp);
7448 #endif
7449 	} else {
7450 #ifdef INVARIANTS
7451 		/*
7452 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
7453 		 * flag)
7454 		 */
7455 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7456 		      seq_out, len, tp->snd_max, tp);
7457 #endif
7458 	}
7459 }
7460 
7461 /*
7462  * Record one of the RTT updates from an ack into
7463  * our sample structure.
7464  */
7465 
7466 static void
7467 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7468 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7469 {
7470 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7471 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7472 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7473 	}
7474 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7475 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7476 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7477 	}
7478 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7479 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7480 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
7481 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7482 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7483 	}
7484 	if ((confidence == 1) &&
7485 	    ((rsm == NULL) ||
7486 	     (rsm->r_just_ret) ||
7487 	     (rsm->r_one_out_nr &&
7488 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7489 		/*
7490 		 * If the rsm had a just return
7491 		 * hit it then we can't trust the
7492 		 * rtt measurement for buffer deterimination
7493 		 * Note that a confidence of 2, indicates
7494 		 * SACK'd which overrides the r_just_ret or
7495 		 * the r_one_out_nr. If it was a CUM-ACK and
7496 		 * we had only two outstanding, but get an
7497 		 * ack for only 1. Then that also lowers our
7498 		 * confidence.
7499 		 */
7500 		confidence = 0;
7501 	}
7502 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7503 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7504 		if (rack->r_ctl.rack_rs.confidence == 0) {
7505 			/*
7506 			 * We take anything with no current confidence
7507 			 * saved.
7508 			 */
7509 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7510 			rack->r_ctl.rack_rs.confidence = confidence;
7511 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7512 		} else if (confidence || rack->r_ctl.rack_rs.confidence) {
7513 			/*
7514 			 * Once we have a confident number,
7515 			 * we can update it with a smaller
7516 			 * value since this confident number
7517 			 * may include the DSACK time until
7518 			 * the next segment (the second one) arrived.
7519 			 */
7520 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7521 			rack->r_ctl.rack_rs.confidence = confidence;
7522 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7523 		}
7524 	}
7525 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7526 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7527 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7528 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
7529 }
7530 
7531 /*
7532  * Collect new round-trip time estimate
7533  * and update averages and current timeout.
7534  */
7535 static void
7536 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7537 {
7538 	int32_t delta;
7539 	int32_t rtt;
7540 
7541 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7542 		/* No valid sample */
7543 		return;
7544 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7545 		/* We are to use the lowest RTT seen in a single ack */
7546 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7547 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7548 		/* We are to use the highest RTT seen in a single ack */
7549 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7550 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7551 		/* We are to use the average RTT seen in a single ack */
7552 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7553 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7554 	} else {
7555 #ifdef INVARIANTS
7556 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7557 #endif
7558 		return;
7559 	}
7560 	if (rtt == 0)
7561 		rtt = 1;
7562 	if (rack->rc_gp_rtt_set == 0) {
7563 		/*
7564 		 * With no RTT we have to accept
7565 		 * even one we are not confident of.
7566 		 */
7567 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7568 		rack->rc_gp_rtt_set = 1;
7569 	} else if (rack->r_ctl.rack_rs.confidence) {
7570 		/* update the running gp srtt */
7571 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7572 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7573 	}
7574 	if (rack->r_ctl.rack_rs.confidence) {
7575 		/*
7576 		 * record the low and high for highly buffered path computation,
7577 		 * we only do this if we are confident (not a retransmission).
7578 		 */
7579 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7580 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7581 		}
7582 		if (rack->rc_highly_buffered == 0) {
7583 			/*
7584 			 * Currently once we declare a path has
7585 			 * highly buffered there is no going
7586 			 * back, which may be a problem...
7587 			 */
7588 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7589 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7590 						     rack->r_ctl.rc_highest_us_rtt,
7591 						     rack->r_ctl.rc_lowest_us_rtt,
7592 						     RACK_RTTS_SEEHBP);
7593 				rack->rc_highly_buffered = 1;
7594 			}
7595 		}
7596 	}
7597 	if ((rack->r_ctl.rack_rs.confidence) ||
7598 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7599 		/*
7600 		 * If we are highly confident of it <or> it was
7601 		 * never retransmitted we accept it as the last us_rtt.
7602 		 */
7603 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7604 		/* The lowest rtt can be set if its was not retransmited */
7605 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7606 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7607 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
7608 				rack->r_ctl.rc_lowest_us_rtt = 1;
7609 		}
7610 	}
7611 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7612 	if (tp->t_srtt != 0) {
7613 		/*
7614 		 * We keep a simple srtt in microseconds, like our rtt
7615 		 * measurement. We don't need to do any tricks with shifting
7616 		 * etc. Instead we just add in 1/8th of the new measurement
7617 		 * and subtract out 1/8 of the old srtt. We do the same with
7618 		 * the variance after finding the absolute value of the
7619 		 * difference between this sample and the current srtt.
7620 		 */
7621 		delta = tp->t_srtt - rtt;
7622 		/* Take off 1/8th of the current sRTT */
7623 		tp->t_srtt -= (tp->t_srtt >> 3);
7624 		/* Add in 1/8th of the new RTT just measured */
7625 		tp->t_srtt += (rtt >> 3);
7626 		if (tp->t_srtt <= 0)
7627 			tp->t_srtt = 1;
7628 		/* Now lets make the absolute value of the variance */
7629 		if (delta < 0)
7630 			delta = -delta;
7631 		/* Subtract out 1/8th */
7632 		tp->t_rttvar -= (tp->t_rttvar >> 3);
7633 		/* Add in 1/8th of the new variance we just saw */
7634 		tp->t_rttvar += (delta >> 3);
7635 		if (tp->t_rttvar <= 0)
7636 			tp->t_rttvar = 1;
7637 	} else {
7638 		/*
7639 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
7640 		 * variance to half the rtt (so our first retransmit happens
7641 		 * at 3*rtt).
7642 		 */
7643 		tp->t_srtt = rtt;
7644 		tp->t_rttvar = rtt >> 1;
7645 	}
7646 	rack->rc_srtt_measure_made = 1;
7647 	KMOD_TCPSTAT_INC(tcps_rttupdated);
7648 	tp->t_rttupdated++;
7649 #ifdef STATS
7650 	if (rack_stats_gets_ms_rtt == 0) {
7651 		/* Send in the microsecond rtt used for rxt timeout purposes */
7652 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7653 	} else if (rack_stats_gets_ms_rtt == 1) {
7654 		/* Send in the millisecond rtt used for rxt timeout purposes */
7655 		int32_t ms_rtt;
7656 
7657 		/* Round up */
7658 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7659 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7660 	} else if (rack_stats_gets_ms_rtt == 2) {
7661 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
7662 		int32_t ms_rtt;
7663 
7664 		/* Round up */
7665 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7666 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7667 	}  else {
7668 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
7669 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7670 	}
7671 
7672 #endif
7673 	/*
7674 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
7675 	 * way we do the smoothing, srtt and rttvar will each average +1/2
7676 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
7677 	 * tick of rounding and 1 extra tick because of +-1/2 tick
7678 	 * uncertainty in the firing of the timer.  The bias will give us
7679 	 * exactly the 1.5 tick we need.  But, because the bias is
7680 	 * statistical, we have to test that we don't drop below the minimum
7681 	 * feasible timer (which is 2 ticks).
7682 	 */
7683 	tp->t_rxtshift = 0;
7684 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7685 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7686 	rack_log_rtt_sample(rack, rtt);
7687 	tp->t_softerror = 0;
7688 }
7689 
7690 
7691 static void
7692 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7693 {
7694 	/*
7695 	 * Apply to filter the inbound us-rtt at us_cts.
7696 	 */
7697 	uint32_t old_rtt;
7698 
7699 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7700 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7701 			       us_rtt, us_cts);
7702 	if (old_rtt > us_rtt) {
7703 		/* We just hit a new lower rtt time */
7704 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7705 				     __LINE__, RACK_RTTS_NEWRTT);
7706 		/*
7707 		 * Only count it if its lower than what we saw within our
7708 		 * calculated range.
7709 		 */
7710 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7711 			if (rack_probertt_lower_within &&
7712 			    rack->rc_gp_dyn_mul &&
7713 			    (rack->use_fixed_rate == 0) &&
7714 			    (rack->rc_always_pace)) {
7715 				/*
7716 				 * We are seeing a new lower rtt very close
7717 				 * to the time that we would have entered probe-rtt.
7718 				 * This is probably due to the fact that a peer flow
7719 				 * has entered probe-rtt. Lets go in now too.
7720 				 */
7721 				uint32_t val;
7722 
7723 				val = rack_probertt_lower_within * rack_time_between_probertt;
7724 				val /= 100;
7725 				if ((rack->in_probe_rtt == 0)  &&
7726 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
7727 					rack_enter_probertt(rack, us_cts);
7728 				}
7729 			}
7730 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7731 		}
7732 	}
7733 }
7734 
7735 static int
7736 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7737     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7738 {
7739 	uint32_t us_rtt;
7740 	int32_t i, all;
7741 	uint32_t t, len_acked;
7742 
7743 	if ((rsm->r_flags & RACK_ACKED) ||
7744 	    (rsm->r_flags & RACK_WAS_ACKED))
7745 		/* Already done */
7746 		return (0);
7747 	if (rsm->r_no_rtt_allowed) {
7748 		/* Not allowed */
7749 		return (0);
7750 	}
7751 	if (ack_type == CUM_ACKED) {
7752 		if (SEQ_GT(th_ack, rsm->r_end)) {
7753 			len_acked = rsm->r_end - rsm->r_start;
7754 			all = 1;
7755 		} else {
7756 			len_acked = th_ack - rsm->r_start;
7757 			all = 0;
7758 		}
7759 	} else {
7760 		len_acked = rsm->r_end - rsm->r_start;
7761 		all = 0;
7762 	}
7763 	if (rsm->r_rtr_cnt == 1) {
7764 
7765 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7766 		if ((int)t <= 0)
7767 			t = 1;
7768 		if (!tp->t_rttlow || tp->t_rttlow > t)
7769 			tp->t_rttlow = t;
7770 		if (!rack->r_ctl.rc_rack_min_rtt ||
7771 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7772 			rack->r_ctl.rc_rack_min_rtt = t;
7773 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
7774 				rack->r_ctl.rc_rack_min_rtt = 1;
7775 			}
7776 		}
7777 		if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7778 			us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7779 		else
7780 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7781 		if (us_rtt == 0)
7782 			us_rtt = 1;
7783 		if (CC_ALGO(tp)->rttsample != NULL) {
7784 			/* Kick the RTT to the CC */
7785 			CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
7786 		}
7787 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7788 		if (ack_type == SACKED) {
7789 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7790 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7791 		} else {
7792 			/*
7793 			 * We need to setup what our confidence
7794 			 * is in this ack.
7795 			 *
7796 			 * If the rsm was app limited and it is
7797 			 * less than a mss in length (the end
7798 			 * of the send) then we have a gap. If we
7799 			 * were app limited but say we were sending
7800 			 * multiple MSS's then we are more confident
7801 			 * int it.
7802 			 *
7803 			 * When we are not app-limited then we see if
7804 			 * the rsm is being included in the current
7805 			 * measurement, we tell this by the app_limited_needs_set
7806 			 * flag.
7807 			 *
7808 			 * Note that being cwnd blocked is not applimited
7809 			 * as well as the pacing delay between packets which
7810 			 * are sending only 1 or 2 MSS's also will show up
7811 			 * in the RTT. We probably need to examine this algorithm
7812 			 * a bit more and enhance it to account for the delay
7813 			 * between rsm's. We could do that by saving off the
7814 			 * pacing delay of each rsm (in an rsm) and then
7815 			 * factoring that in somehow though for now I am
7816 			 * not sure how :)
7817 			 */
7818 			int calc_conf = 0;
7819 
7820 			if (rsm->r_flags & RACK_APP_LIMITED) {
7821 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7822 					calc_conf = 0;
7823 				else
7824 					calc_conf = 1;
7825 			} else if (rack->app_limited_needs_set == 0) {
7826 				calc_conf = 1;
7827 			} else {
7828 				calc_conf = 0;
7829 			}
7830 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7831 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7832 					    calc_conf, rsm, rsm->r_rtr_cnt);
7833 		}
7834 		if ((rsm->r_flags & RACK_TLP) &&
7835 		    (!IN_FASTRECOVERY(tp->t_flags))) {
7836 			/* Segment was a TLP and our retrans matched */
7837 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7838 				rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
7839 			}
7840 		}
7841 		if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7842 			/* New more recent rack_tmit_time */
7843 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7844 			rack->rc_rack_rtt = t;
7845 		}
7846 		return (1);
7847 	}
7848 	/*
7849 	 * We clear the soft/rxtshift since we got an ack.
7850 	 * There is no assurance we will call the commit() function
7851 	 * so we need to clear these to avoid incorrect handling.
7852 	 */
7853 	tp->t_rxtshift = 0;
7854 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7855 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7856 	tp->t_softerror = 0;
7857 	if (to && (to->to_flags & TOF_TS) &&
7858 	    (ack_type == CUM_ACKED) &&
7859 	    (to->to_tsecr) &&
7860 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7861 		/*
7862 		 * Now which timestamp does it match? In this block the ACK
7863 		 * must be coming from a previous transmission.
7864 		 */
7865 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
7866 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7867 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7868 				if ((int)t <= 0)
7869 					t = 1;
7870 				if (CC_ALGO(tp)->rttsample != NULL) {
7871 					/*
7872 					 * Kick the RTT to the CC, here
7873 					 * we lie a bit in that we know the
7874 					 * retransmission is correct even though
7875 					 * we retransmitted. This is because
7876 					 * we match the timestamps.
7877 					 */
7878 					if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
7879 						us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
7880 					else
7881 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
7882 					CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
7883 				}
7884 				if ((i + 1) < rsm->r_rtr_cnt) {
7885 					/*
7886 					 * The peer ack'd from our previous
7887 					 * transmission. We have a spurious
7888 					 * retransmission and thus we dont
7889 					 * want to update our rack_rtt.
7890 					 *
7891 					 * Hmm should there be a CC revert here?
7892 					 *
7893 					 */
7894 					return (0);
7895 				}
7896 				if (!tp->t_rttlow || tp->t_rttlow > t)
7897 					tp->t_rttlow = t;
7898 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7899 					rack->r_ctl.rc_rack_min_rtt = t;
7900 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
7901 						rack->r_ctl.rc_rack_min_rtt = 1;
7902 					}
7903 				}
7904 				if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7905 					   (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7906 					/* New more recent rack_tmit_time */
7907 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7908 					rack->rc_rack_rtt = t;
7909 				}
7910 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7911 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7912 						    rsm->r_rtr_cnt);
7913 				return (1);
7914 			}
7915 		}
7916 		goto ts_not_found;
7917 	} else {
7918 		/*
7919 		 * Ok its a SACK block that we retransmitted. or a windows
7920 		 * machine without timestamps. We can tell nothing from the
7921 		 * time-stamp since its not there or the time the peer last
7922 		 * recieved a segment that moved forward its cum-ack point.
7923 		 */
7924 ts_not_found:
7925 		i = rsm->r_rtr_cnt - 1;
7926 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7927 		if ((int)t <= 0)
7928 			t = 1;
7929 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7930 			/*
7931 			 * We retransmitted and the ack came back in less
7932 			 * than the smallest rtt we have observed. We most
7933 			 * likely did an improper retransmit as outlined in
7934 			 * 6.2 Step 2 point 2 in the rack-draft so we
7935 			 * don't want to update our rack_rtt. We in
7936 			 * theory (in future) might want to think about reverting our
7937 			 * cwnd state but we won't for now.
7938 			 */
7939 			return (0);
7940 		} else if (rack->r_ctl.rc_rack_min_rtt) {
7941 			/*
7942 			 * We retransmitted it and the retransmit did the
7943 			 * job.
7944 			 */
7945 			if (!rack->r_ctl.rc_rack_min_rtt ||
7946 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7947 				rack->r_ctl.rc_rack_min_rtt = t;
7948 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
7949 					rack->r_ctl.rc_rack_min_rtt = 1;
7950 				}
7951 			}
7952 			if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
7953 				/* New more recent rack_tmit_time */
7954 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
7955 				rack->rc_rack_rtt = t;
7956 			}
7957 			return (1);
7958 		}
7959 	}
7960 	return (0);
7961 }
7962 
7963 /*
7964  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7965  */
7966 static void
7967 rack_log_sack_passed(struct tcpcb *tp,
7968     struct tcp_rack *rack, struct rack_sendmap *rsm)
7969 {
7970 	struct rack_sendmap *nrsm;
7971 
7972 	nrsm = rsm;
7973 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
7974 	    rack_head, r_tnext) {
7975 		if (nrsm == rsm) {
7976 			/* Skip orginal segment he is acked */
7977 			continue;
7978 		}
7979 		if (nrsm->r_flags & RACK_ACKED) {
7980 			/*
7981 			 * Skip ack'd segments, though we
7982 			 * should not see these, since tmap
7983 			 * should not have ack'd segments.
7984 			 */
7985 			continue;
7986 		}
7987 		if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
7988 			/*
7989 			 * If the peer dropped the rwnd on
7990 			 * these then we don't worry about them.
7991 			 */
7992 			continue;
7993 		}
7994 		if (nrsm->r_flags & RACK_SACK_PASSED) {
7995 			/*
7996 			 * We found one that is already marked
7997 			 * passed, we have been here before and
7998 			 * so all others below this are marked.
7999 			 */
8000 			break;
8001 		}
8002 		nrsm->r_flags |= RACK_SACK_PASSED;
8003 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
8004 	}
8005 }
8006 
8007 static void
8008 rack_need_set_test(struct tcpcb *tp,
8009 		   struct tcp_rack *rack,
8010 		   struct rack_sendmap *rsm,
8011 		   tcp_seq th_ack,
8012 		   int line,
8013 		   int use_which)
8014 {
8015 
8016 	if ((tp->t_flags & TF_GPUTINPROG) &&
8017 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8018 		/*
8019 		 * We were app limited, and this ack
8020 		 * butts up or goes beyond the point where we want
8021 		 * to start our next measurement. We need
8022 		 * to record the new gput_ts as here and
8023 		 * possibly update the start sequence.
8024 		 */
8025 		uint32_t seq, ts;
8026 
8027 		if (rsm->r_rtr_cnt > 1) {
8028 			/*
8029 			 * This is a retransmit, can we
8030 			 * really make any assessment at this
8031 			 * point?  We are not really sure of
8032 			 * the timestamp, is it this or the
8033 			 * previous transmission?
8034 			 *
8035 			 * Lets wait for something better that
8036 			 * is not retransmitted.
8037 			 */
8038 			return;
8039 		}
8040 		seq = tp->gput_seq;
8041 		ts = tp->gput_ts;
8042 		rack->app_limited_needs_set = 0;
8043 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8044 		/* Do we start at a new end? */
8045 		if ((use_which == RACK_USE_BEG) &&
8046 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8047 			/*
8048 			 * When we get an ACK that just eats
8049 			 * up some of the rsm, we set RACK_USE_BEG
8050 			 * since whats at r_start (i.e. th_ack)
8051 			 * is left unacked and thats where the
8052 			 * measurement not starts.
8053 			 */
8054 			tp->gput_seq = rsm->r_start;
8055 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8056 		}
8057 		if ((use_which == RACK_USE_END) &&
8058 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8059 			    /*
8060 			     * We use the end when the cumack
8061 			     * is moving forward and completely
8062 			     * deleting the rsm passed so basically
8063 			     * r_end holds th_ack.
8064 			     *
8065 			     * For SACK's we also want to use the end
8066 			     * since this piece just got sacked and
8067 			     * we want to target anything after that
8068 			     * in our measurement.
8069 			     */
8070 			    tp->gput_seq = rsm->r_end;
8071 			    rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8072 		}
8073 		if (use_which == RACK_USE_END_OR_THACK) {
8074 			/*
8075 			 * special case for ack moving forward,
8076 			 * not a sack, we need to move all the
8077 			 * way up to where this ack cum-ack moves
8078 			 * to.
8079 			 */
8080 			if (SEQ_GT(th_ack, rsm->r_end))
8081 				tp->gput_seq = th_ack;
8082 			else
8083 				tp->gput_seq = rsm->r_end;
8084 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8085 		}
8086 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8087 			/*
8088 			 * We moved beyond this guy's range, re-calculate
8089 			 * the new end point.
8090 			 */
8091 			if (rack->rc_gp_filled == 0) {
8092 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8093 			} else {
8094 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8095 			}
8096 		}
8097 		/*
8098 		 * We are moving the goal post, we may be able to clear the
8099 		 * measure_saw_probe_rtt flag.
8100 		 */
8101 		if ((rack->in_probe_rtt == 0) &&
8102 		    (rack->measure_saw_probe_rtt) &&
8103 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8104 			rack->measure_saw_probe_rtt = 0;
8105 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8106 					   seq, tp->gput_seq, 0, 5, line, NULL, 0);
8107 		if (rack->rc_gp_filled &&
8108 		    ((tp->gput_ack - tp->gput_seq) <
8109 		     max(rc_init_window(rack), (MIN_GP_WIN *
8110 						ctf_fixed_maxseg(tp))))) {
8111 			uint32_t ideal_amount;
8112 
8113 			ideal_amount = rack_get_measure_window(tp, rack);
8114 			if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
8115 				/*
8116 				 * There is no sense of continuing this measurement
8117 				 * because its too small to gain us anything we
8118 				 * trust. Skip it and that way we can start a new
8119 				 * measurement quicker.
8120 				 */
8121 				tp->t_flags &= ~TF_GPUTINPROG;
8122 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8123 							   0, 0, 0, 6, __LINE__, NULL, 0);
8124 			} else {
8125 				/*
8126 				 * Reset the window further out.
8127 				 */
8128 				tp->gput_ack = tp->gput_seq + ideal_amount;
8129 			}
8130 		}
8131 	}
8132 }
8133 
8134 static inline int
8135 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
8136 {
8137 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
8138 		/* Behind our TLP definition or right at */
8139 		return (0);
8140 	}
8141 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
8142 		/* The start is beyond or right at our end of TLP definition */
8143 		return (0);
8144 	}
8145 	/* It has to be a sub-part of the original TLP recorded */
8146 	return (1);
8147 }
8148 
8149 
8150 static uint32_t
8151 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8152 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8153 {
8154 	uint32_t start, end, changed = 0;
8155 	struct rack_sendmap stack_map;
8156 	struct rack_sendmap *rsm, *nrsm, fe, *prev, *next;
8157 #ifdef INVARIANTS
8158 	struct rack_sendmap *insret;
8159 #endif
8160 	int32_t used_ref = 1;
8161 	int moved = 0;
8162 
8163 	start = sack->start;
8164 	end = sack->end;
8165 	rsm = *prsm;
8166 	memset(&fe, 0, sizeof(fe));
8167 do_rest_ofb:
8168 	if ((rsm == NULL) ||
8169 	    (SEQ_LT(end, rsm->r_start)) ||
8170 	    (SEQ_GEQ(start, rsm->r_end)) ||
8171 	    (SEQ_LT(start, rsm->r_start))) {
8172 		/*
8173 		 * We are not in the right spot,
8174 		 * find the correct spot in the tree.
8175 		 */
8176 		used_ref = 0;
8177 		fe.r_start = start;
8178 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8179 		moved++;
8180 	}
8181 	if (rsm == NULL) {
8182 		/* TSNH */
8183 		goto out;
8184 	}
8185 	/* Ok we have an ACK for some piece of this rsm */
8186 	if (rsm->r_start != start) {
8187 		if ((rsm->r_flags & RACK_ACKED) == 0) {
8188 			/*
8189 			 * Before any splitting or hookery is
8190 			 * done is it a TLP of interest i.e. rxt?
8191 			 */
8192 			if ((rsm->r_flags & RACK_TLP) &&
8193 			    (rsm->r_rtr_cnt > 1)) {
8194 				/*
8195 				 * We are splitting a rxt TLP, check
8196 				 * if we need to save off the start/end
8197 				 */
8198 				if (rack->rc_last_tlp_acked_set &&
8199 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8200 					/*
8201 					 * We already turned this on since we are inside
8202 					 * the previous one was a partially sack now we
8203 					 * are getting another one (maybe all of it).
8204 					 *
8205 					 */
8206 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8207 					/*
8208 					 * Lets make sure we have all of it though.
8209 					 */
8210 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8211 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8212 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8213 								     rack->r_ctl.last_tlp_acked_end);
8214 					}
8215 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8216 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8217 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8218 								     rack->r_ctl.last_tlp_acked_end);
8219 					}
8220 				} else {
8221 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8222 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8223 					rack->rc_last_tlp_past_cumack = 0;
8224 					rack->rc_last_tlp_acked_set = 1;
8225 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8226 				}
8227 			}
8228 			/**
8229 			 * Need to split this in two pieces the before and after,
8230 			 * the before remains in the map, the after must be
8231 			 * added. In other words we have:
8232 			 * rsm        |--------------|
8233 			 * sackblk        |------->
8234 			 * rsm will become
8235 			 *     rsm    |---|
8236 			 * and nrsm will be  the sacked piece
8237 			 *     nrsm       |----------|
8238 			 *
8239 			 * But before we start down that path lets
8240 			 * see if the sack spans over on top of
8241 			 * the next guy and it is already sacked.
8242 			 *
8243 			 */
8244 			next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8245 			if (next && (next->r_flags & RACK_ACKED) &&
8246 			    SEQ_GEQ(end, next->r_start)) {
8247 				/**
8248 				 * So the next one is already acked, and
8249 				 * we can thus by hookery use our stack_map
8250 				 * to reflect the piece being sacked and
8251 				 * then adjust the two tree entries moving
8252 				 * the start and ends around. So we start like:
8253 				 *  rsm     |------------|             (not-acked)
8254 				 *  next                 |-----------| (acked)
8255 				 *  sackblk        |-------->
8256 				 *  We want to end like so:
8257 				 *  rsm     |------|                   (not-acked)
8258 				 *  next           |-----------------| (acked)
8259 				 *  nrsm           |-----|
8260 				 * Where nrsm is a temporary stack piece we
8261 				 * use to update all the gizmos.
8262 				 */
8263 				/* Copy up our fudge block */
8264 				nrsm = &stack_map;
8265 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8266 				/* Now adjust our tree blocks */
8267 				rsm->r_end = start;
8268 				next->r_start = start;
8269 				/* Now we must adjust back where next->m is */
8270 				rack_setup_offset_for_rsm(rsm, next);
8271 
8272 				/* We don't need to adjust rsm, it did not change */
8273 				/* Clear out the dup ack count of the remainder */
8274 				rsm->r_dupack = 0;
8275 				rsm->r_just_ret = 0;
8276 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8277 				/* Now lets make sure our fudge block is right */
8278 				nrsm->r_start = start;
8279 				/* Now lets update all the stats and such */
8280 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8281 				if (rack->app_limited_needs_set)
8282 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8283 				changed += (nrsm->r_end - nrsm->r_start);
8284 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8285 				if (nrsm->r_flags & RACK_SACK_PASSED) {
8286 					rack->r_ctl.rc_reorder_ts = cts;
8287 				}
8288 				/*
8289 				 * Now we want to go up from rsm (the
8290 				 * one left un-acked) to the next one
8291 				 * in the tmap. We do this so when
8292 				 * we walk backwards we include marking
8293 				 * sack-passed on rsm (The one passed in
8294 				 * is skipped since it is generally called
8295 				 * on something sacked before removing it
8296 				 * from the tmap).
8297 				 */
8298 				if (rsm->r_in_tmap) {
8299 					nrsm = TAILQ_NEXT(rsm, r_tnext);
8300 					/*
8301 					 * Now that we have the next
8302 					 * one walk backwards from there.
8303 					 */
8304 					if (nrsm && nrsm->r_in_tmap)
8305 						rack_log_sack_passed(tp, rack, nrsm);
8306 				}
8307 				/* Now are we done? */
8308 				if (SEQ_LT(end, next->r_end) ||
8309 				    (end == next->r_end)) {
8310 					/* Done with block */
8311 					goto out;
8312 				}
8313 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8314 				counter_u64_add(rack_sack_used_next_merge, 1);
8315 				/* Postion for the next block */
8316 				start = next->r_end;
8317 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8318 				if (rsm == NULL)
8319 					goto out;
8320 			} else {
8321 				/**
8322 				 * We can't use any hookery here, so we
8323 				 * need to split the map. We enter like
8324 				 * so:
8325 				 *  rsm      |--------|
8326 				 *  sackblk       |----->
8327 				 * We will add the new block nrsm and
8328 				 * that will be the new portion, and then
8329 				 * fall through after reseting rsm. So we
8330 				 * split and look like this:
8331 				 *  rsm      |----|
8332 				 *  sackblk       |----->
8333 				 *  nrsm          |---|
8334 				 * We then fall through reseting
8335 				 * rsm to nrsm, so the next block
8336 				 * picks it up.
8337 				 */
8338 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8339 				if (nrsm == NULL) {
8340 					/*
8341 					 * failed XXXrrs what can we do but loose the sack
8342 					 * info?
8343 					 */
8344 					goto out;
8345 				}
8346 				counter_u64_add(rack_sack_splits, 1);
8347 				rack_clone_rsm(rack, nrsm, rsm, start);
8348 				rsm->r_just_ret = 0;
8349 #ifndef INVARIANTS
8350 				(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8351 #else
8352 				insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8353 				if (insret != NULL) {
8354 					panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8355 					      nrsm, insret, rack, rsm);
8356 				}
8357 #endif
8358 				if (rsm->r_in_tmap) {
8359 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8360 					nrsm->r_in_tmap = 1;
8361 				}
8362 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8363 				rsm->r_flags &= (~RACK_HAS_FIN);
8364 				/* Position us to point to the new nrsm that starts the sack blk */
8365 				rsm = nrsm;
8366 			}
8367 		} else {
8368 			/* Already sacked this piece */
8369 			counter_u64_add(rack_sack_skipped_acked, 1);
8370 			moved++;
8371 			if (end == rsm->r_end) {
8372 				/* Done with block */
8373 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8374 				goto out;
8375 			} else if (SEQ_LT(end, rsm->r_end)) {
8376 				/* A partial sack to a already sacked block */
8377 				moved++;
8378 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8379 				goto out;
8380 			} else {
8381 				/*
8382 				 * The end goes beyond this guy
8383 				 * reposition the start to the
8384 				 * next block.
8385 				 */
8386 				start = rsm->r_end;
8387 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8388 				if (rsm == NULL)
8389 					goto out;
8390 			}
8391 		}
8392 	}
8393 	if (SEQ_GEQ(end, rsm->r_end)) {
8394 		/**
8395 		 * The end of this block is either beyond this guy or right
8396 		 * at this guy. I.e.:
8397 		 *  rsm ---                 |-----|
8398 		 *  end                     |-----|
8399 		 *  <or>
8400 		 *  end                     |---------|
8401 		 */
8402 		if ((rsm->r_flags & RACK_ACKED) == 0) {
8403 			/*
8404 			 * Is it a TLP of interest?
8405 			 */
8406 			if ((rsm->r_flags & RACK_TLP) &&
8407 			    (rsm->r_rtr_cnt > 1)) {
8408 				/*
8409 				 * We are splitting a rxt TLP, check
8410 				 * if we need to save off the start/end
8411 				 */
8412 				if (rack->rc_last_tlp_acked_set &&
8413 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8414 					/*
8415 					 * We already turned this on since we are inside
8416 					 * the previous one was a partially sack now we
8417 					 * are getting another one (maybe all of it).
8418 					 */
8419 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8420 					/*
8421 					 * Lets make sure we have all of it though.
8422 					 */
8423 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8424 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8425 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8426 								     rack->r_ctl.last_tlp_acked_end);
8427 					}
8428 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8429 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8430 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8431 								     rack->r_ctl.last_tlp_acked_end);
8432 					}
8433 				} else {
8434 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8435 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8436 					rack->rc_last_tlp_past_cumack = 0;
8437 					rack->rc_last_tlp_acked_set = 1;
8438 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8439 				}
8440 			}
8441 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8442 			changed += (rsm->r_end - rsm->r_start);
8443 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8444 			if (rsm->r_in_tmap) /* should be true */
8445 				rack_log_sack_passed(tp, rack, rsm);
8446 			/* Is Reordering occuring? */
8447 			if (rsm->r_flags & RACK_SACK_PASSED) {
8448 				rsm->r_flags &= ~RACK_SACK_PASSED;
8449 				rack->r_ctl.rc_reorder_ts = cts;
8450 			}
8451 			if (rack->app_limited_needs_set)
8452 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8453 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8454 			rsm->r_flags |= RACK_ACKED;
8455 			if (rsm->r_in_tmap) {
8456 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8457 				rsm->r_in_tmap = 0;
8458 			}
8459 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8460 		} else {
8461 			counter_u64_add(rack_sack_skipped_acked, 1);
8462 			moved++;
8463 		}
8464 		if (end == rsm->r_end) {
8465 			/* This block only - done, setup for next */
8466 			goto out;
8467 		}
8468 		/*
8469 		 * There is more not coverend by this rsm move on
8470 		 * to the next block in the RB tree.
8471 		 */
8472 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8473 		start = rsm->r_end;
8474 		rsm = nrsm;
8475 		if (rsm == NULL)
8476 			goto out;
8477 		goto do_rest_ofb;
8478 	}
8479 	/**
8480 	 * The end of this sack block is smaller than
8481 	 * our rsm i.e.:
8482 	 *  rsm ---                 |-----|
8483 	 *  end                     |--|
8484 	 */
8485 	if ((rsm->r_flags & RACK_ACKED) == 0) {
8486 		/*
8487 		 * Is it a TLP of interest?
8488 		 */
8489 		if ((rsm->r_flags & RACK_TLP) &&
8490 		    (rsm->r_rtr_cnt > 1)) {
8491 			/*
8492 			 * We are splitting a rxt TLP, check
8493 			 * if we need to save off the start/end
8494 			 */
8495 			if (rack->rc_last_tlp_acked_set &&
8496 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8497 				/*
8498 				 * We already turned this on since we are inside
8499 				 * the previous one was a partially sack now we
8500 				 * are getting another one (maybe all of it).
8501 				 */
8502 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8503 				/*
8504 				 * Lets make sure we have all of it though.
8505 				 */
8506 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8507 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8508 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8509 							     rack->r_ctl.last_tlp_acked_end);
8510 				}
8511 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8512 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8513 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8514 							     rack->r_ctl.last_tlp_acked_end);
8515 				}
8516 			} else {
8517 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8518 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8519 				rack->rc_last_tlp_past_cumack = 0;
8520 				rack->rc_last_tlp_acked_set = 1;
8521 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8522 			}
8523 		}
8524 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8525 		if (prev &&
8526 		    (prev->r_flags & RACK_ACKED)) {
8527 			/**
8528 			 * Goal, we want the right remainder of rsm to shrink
8529 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
8530 			 * We want to expand prev to go all the way
8531 			 * to prev->r_end <- end.
8532 			 * so in the tree we have before:
8533 			 *   prev     |--------|         (acked)
8534 			 *   rsm               |-------| (non-acked)
8535 			 *   sackblk           |-|
8536 			 * We churn it so we end up with
8537 			 *   prev     |----------|       (acked)
8538 			 *   rsm                 |-----| (non-acked)
8539 			 *   nrsm              |-| (temporary)
8540 			 *
8541 			 * Note if either prev/rsm is a TLP we don't
8542 			 * do this.
8543 			 */
8544 			nrsm = &stack_map;
8545 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8546 			prev->r_end = end;
8547 			rsm->r_start = end;
8548 			/* Now adjust nrsm (stack copy) to be
8549 			 * the one that is the small
8550 			 * piece that was "sacked".
8551 			 */
8552 			nrsm->r_end = end;
8553 			rsm->r_dupack = 0;
8554 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8555 			/*
8556 			 * Now that the rsm has had its start moved forward
8557 			 * lets go ahead and get its new place in the world.
8558 			 */
8559 			rack_setup_offset_for_rsm(prev, rsm);
8560 			/*
8561 			 * Now nrsm is our new little piece
8562 			 * that is acked (which was merged
8563 			 * to prev). Update the rtt and changed
8564 			 * based on that. Also check for reordering.
8565 			 */
8566 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8567 			if (rack->app_limited_needs_set)
8568 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8569 			changed += (nrsm->r_end - nrsm->r_start);
8570 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8571 			if (nrsm->r_flags & RACK_SACK_PASSED) {
8572 				rack->r_ctl.rc_reorder_ts = cts;
8573 			}
8574 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8575 			rsm = prev;
8576 			counter_u64_add(rack_sack_used_prev_merge, 1);
8577 		} else {
8578 			/**
8579 			 * This is the case where our previous
8580 			 * block is not acked either, so we must
8581 			 * split the block in two.
8582 			 */
8583 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8584 			if (nrsm == NULL) {
8585 				/* failed rrs what can we do but loose the sack info? */
8586 				goto out;
8587 			}
8588 			if ((rsm->r_flags & RACK_TLP) &&
8589 			    (rsm->r_rtr_cnt > 1)) {
8590 				/*
8591 				 * We are splitting a rxt TLP, check
8592 				 * if we need to save off the start/end
8593 				 */
8594 				if (rack->rc_last_tlp_acked_set &&
8595 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8596 					    /*
8597 					     * We already turned this on since this block is inside
8598 					     * the previous one was a partially sack now we
8599 					     * are getting another one (maybe all of it).
8600 					     */
8601 					    rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8602 					    /*
8603 					     * Lets make sure we have all of it though.
8604 					     */
8605 					    if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8606 						    rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8607 						    rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8608 									 rack->r_ctl.last_tlp_acked_end);
8609 					    }
8610 					    if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8611 						    rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8612 						    rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8613 									 rack->r_ctl.last_tlp_acked_end);
8614 					    }
8615 				    } else {
8616 					    rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8617 					    rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8618 					    rack->rc_last_tlp_acked_set = 1;
8619 					    rack->rc_last_tlp_past_cumack = 0;
8620 					    rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8621 				    }
8622 			}
8623 			/**
8624 			 * In this case nrsm becomes
8625 			 * nrsm->r_start = end;
8626 			 * nrsm->r_end = rsm->r_end;
8627 			 * which is un-acked.
8628 			 * <and>
8629 			 * rsm->r_end = nrsm->r_start;
8630 			 * i.e. the remaining un-acked
8631 			 * piece is left on the left
8632 			 * hand side.
8633 			 *
8634 			 * So we start like this
8635 			 * rsm      |----------| (not acked)
8636 			 * sackblk  |---|
8637 			 * build it so we have
8638 			 * rsm      |---|         (acked)
8639 			 * nrsm         |------|  (not acked)
8640 			 */
8641 			counter_u64_add(rack_sack_splits, 1);
8642 			rack_clone_rsm(rack, nrsm, rsm, end);
8643 			rsm->r_flags &= (~RACK_HAS_FIN);
8644 			rsm->r_just_ret = 0;
8645 #ifndef INVARIANTS
8646 			(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8647 #else
8648 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8649 			if (insret != NULL) {
8650 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8651 				      nrsm, insret, rack, rsm);
8652 			}
8653 #endif
8654 			if (rsm->r_in_tmap) {
8655 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8656 				nrsm->r_in_tmap = 1;
8657 			}
8658 			nrsm->r_dupack = 0;
8659 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8660 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8661 			changed += (rsm->r_end - rsm->r_start);
8662 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8663 			if (rsm->r_in_tmap) /* should be true */
8664 				rack_log_sack_passed(tp, rack, rsm);
8665 			/* Is Reordering occuring? */
8666 			if (rsm->r_flags & RACK_SACK_PASSED) {
8667 				rsm->r_flags &= ~RACK_SACK_PASSED;
8668 				rack->r_ctl.rc_reorder_ts = cts;
8669 			}
8670 			if (rack->app_limited_needs_set)
8671 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8672 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8673 			rsm->r_flags |= RACK_ACKED;
8674 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8675 			if (rsm->r_in_tmap) {
8676 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8677 				rsm->r_in_tmap = 0;
8678 			}
8679 		}
8680 	} else if (start != end){
8681 		/*
8682 		 * The block was already acked.
8683 		 */
8684 		counter_u64_add(rack_sack_skipped_acked, 1);
8685 		moved++;
8686 	}
8687 out:
8688 	if (rsm &&
8689 	    ((rsm->r_flags & RACK_TLP) == 0) &&
8690 	    (rsm->r_flags & RACK_ACKED)) {
8691 		/*
8692 		 * Now can we merge where we worked
8693 		 * with either the previous or
8694 		 * next block?
8695 		 */
8696 		next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8697 		while (next) {
8698 			if (next->r_flags & RACK_TLP)
8699 				break;
8700 			if (next->r_flags & RACK_ACKED) {
8701 			/* yep this and next can be merged */
8702 				rsm = rack_merge_rsm(rack, rsm, next);
8703 				next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8704 			} else
8705 				break;
8706 		}
8707 		/* Now what about the previous? */
8708 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8709 		while (prev) {
8710 			if (prev->r_flags & RACK_TLP)
8711 				break;
8712 			if (prev->r_flags & RACK_ACKED) {
8713 				/* yep the previous and this can be merged */
8714 				rsm = rack_merge_rsm(rack, prev, rsm);
8715 				prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8716 			} else
8717 				break;
8718 		}
8719 	}
8720 	if (used_ref == 0) {
8721 		counter_u64_add(rack_sack_proc_all, 1);
8722 	} else {
8723 		counter_u64_add(rack_sack_proc_short, 1);
8724 	}
8725 	/* Save off the next one for quick reference. */
8726 	if (rsm)
8727 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8728 	else
8729 		nrsm = NULL;
8730 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
8731 	/* Pass back the moved. */
8732 	*moved_two = moved;
8733 	return (changed);
8734 }
8735 
8736 static void inline
8737 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8738 {
8739 	struct rack_sendmap *tmap;
8740 
8741 	tmap = NULL;
8742 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
8743 		/* Its no longer sacked, mark it so */
8744 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8745 #ifdef INVARIANTS
8746 		if (rsm->r_in_tmap) {
8747 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
8748 			      rack, rsm, rsm->r_flags);
8749 		}
8750 #endif
8751 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8752 		/* Rebuild it into our tmap */
8753 		if (tmap == NULL) {
8754 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8755 			tmap = rsm;
8756 		} else {
8757 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8758 			tmap = rsm;
8759 		}
8760 		tmap->r_in_tmap = 1;
8761 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8762 	}
8763 	/*
8764 	 * Now lets possibly clear the sack filter so we start
8765 	 * recognizing sacks that cover this area.
8766 	 */
8767 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8768 
8769 }
8770 
8771 static void
8772 rack_do_decay(struct tcp_rack *rack)
8773 {
8774 	struct timeval res;
8775 
8776 #define	timersub(tvp, uvp, vvp)						\
8777 	do {								\
8778 		(vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;		\
8779 		(vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;	\
8780 		if ((vvp)->tv_usec < 0) {				\
8781 			(vvp)->tv_sec--;				\
8782 			(vvp)->tv_usec += 1000000;			\
8783 		}							\
8784 	} while (0)
8785 
8786 	timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8787 #undef timersub
8788 
8789 	rack->r_ctl.input_pkt++;
8790 	if ((rack->rc_in_persist) ||
8791 	    (res.tv_sec >= 1) ||
8792 	    (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8793 		/*
8794 		 * Check for decay of non-SAD,
8795 		 * we want all SAD detection metrics to
8796 		 * decay 1/4 per second (or more) passed.
8797 		 */
8798 #ifdef NETFLIX_EXP_DETECTION
8799 		uint32_t pkt_delta;
8800 
8801 		pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8802 #endif
8803 		/* Update our saved tracking values */
8804 		rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8805 		rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8806 		/* Now do we escape without decay? */
8807 #ifdef NETFLIX_EXP_DETECTION
8808 		if (rack->rc_in_persist ||
8809 		    (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8810 		    (pkt_delta < tcp_sad_low_pps)){
8811 			/*
8812 			 * We don't decay idle connections
8813 			 * or ones that have a low input pps.
8814 			 */
8815 			return;
8816 		}
8817 		/* Decay the counters */
8818 		rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8819 							tcp_sad_decay_val);
8820 		rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8821 							 tcp_sad_decay_val);
8822 		rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8823 							       tcp_sad_decay_val);
8824 		rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8825 								tcp_sad_decay_val);
8826 #endif
8827 	}
8828 }
8829 
8830 static void
8831 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8832 {
8833 	struct rack_sendmap *rsm;
8834 #ifdef INVARIANTS
8835 	struct rack_sendmap *rm;
8836 #endif
8837 
8838 	/*
8839 	 * The ACK point is advancing to th_ack, we must drop off
8840 	 * the packets in the rack log and calculate any eligble
8841 	 * RTT's.
8842 	 */
8843 	rack->r_wanted_output = 1;
8844 
8845 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
8846 	if ((rack->rc_last_tlp_acked_set == 1)&&
8847 	    (rack->rc_last_tlp_past_cumack == 1) &&
8848 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
8849 		/*
8850 		 * We have reached the point where our last rack
8851 		 * tlp retransmit sequence is ahead of the cum-ack.
8852 		 * This can only happen when the cum-ack moves all
8853 		 * the way around (its been a full 2^^31+1 bytes
8854 		 * or more since we sent a retransmitted TLP). Lets
8855 		 * turn off the valid flag since its not really valid.
8856 		 *
8857 		 * Note since sack's also turn on this event we have
8858 		 * a complication, we have to wait to age it out until
8859 		 * the cum-ack is by the TLP before checking which is
8860 		 * what the next else clause does.
8861 		 */
8862 		rack_log_dsack_event(rack, 9, __LINE__,
8863 				     rack->r_ctl.last_tlp_acked_start,
8864 				     rack->r_ctl.last_tlp_acked_end);
8865 		rack->rc_last_tlp_acked_set = 0;
8866 		rack->rc_last_tlp_past_cumack = 0;
8867 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
8868 		   (rack->rc_last_tlp_past_cumack == 0) &&
8869 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
8870 		/*
8871 		 * It is safe to start aging TLP's out.
8872 		 */
8873 		rack->rc_last_tlp_past_cumack = 1;
8874 	}
8875 	/* We do the same for the tlp send seq as well */
8876 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8877 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
8878 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
8879 		rack_log_dsack_event(rack, 9, __LINE__,
8880 				     rack->r_ctl.last_sent_tlp_seq,
8881 				     (rack->r_ctl.last_sent_tlp_seq +
8882 				      rack->r_ctl.last_sent_tlp_len));
8883 		rack->rc_last_sent_tlp_seq_valid = 0;
8884 		rack->rc_last_sent_tlp_past_cumack = 0;
8885 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8886 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
8887 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
8888 		/*
8889 		 * It is safe to start aging TLP's send.
8890 		 */
8891 		rack->rc_last_sent_tlp_past_cumack = 1;
8892 	}
8893 more:
8894 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8895 	if (rsm == NULL) {
8896 		if ((th_ack - 1) == tp->iss) {
8897 			/*
8898 			 * For the SYN incoming case we will not
8899 			 * have called tcp_output for the sending of
8900 			 * the SYN, so there will be no map. All
8901 			 * other cases should probably be a panic.
8902 			 */
8903 			return;
8904 		}
8905 		if (tp->t_flags & TF_SENTFIN) {
8906 			/* if we sent a FIN we often will not have map */
8907 			return;
8908 		}
8909 #ifdef INVARIANTS
8910 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8911 		      tp,
8912 		      tp->t_state, th_ack, rack,
8913 		      tp->snd_una, tp->snd_max, tp->snd_nxt);
8914 #endif
8915 		return;
8916 	}
8917 	if (SEQ_LT(th_ack, rsm->r_start)) {
8918 		/* Huh map is missing this */
8919 #ifdef INVARIANTS
8920 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
8921 		       rsm->r_start,
8922 		       th_ack, tp->t_state, rack->r_state);
8923 #endif
8924 		return;
8925 	}
8926 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
8927 
8928 	/* Now was it a retransmitted TLP? */
8929 	if ((rsm->r_flags & RACK_TLP) &&
8930 	    (rsm->r_rtr_cnt > 1)) {
8931 		/*
8932 		 * Yes, this rsm was a TLP and retransmitted, remember that
8933 		 * since if a DSACK comes back on this we don't want
8934 		 * to think of it as a reordered segment. This may
8935 		 * get updated again with possibly even other TLPs
8936 		 * in flight, but thats ok. Only when we don't send
8937 		 * a retransmitted TLP for 1/2 the sequences space
8938 		 * will it get turned off (above).
8939 		 */
8940 		if (rack->rc_last_tlp_acked_set &&
8941 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8942 			/*
8943 			 * We already turned this on since the end matches,
8944 			 * the previous one was a partially ack now we
8945 			 * are getting another one (maybe all of it).
8946 			 */
8947 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8948 			/*
8949 			 * Lets make sure we have all of it though.
8950 			 */
8951 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8952 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8953 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8954 						     rack->r_ctl.last_tlp_acked_end);
8955 			}
8956 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8957 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8958 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8959 						     rack->r_ctl.last_tlp_acked_end);
8960 			}
8961 		} else {
8962 			rack->rc_last_tlp_past_cumack = 1;
8963 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8964 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8965 			rack->rc_last_tlp_acked_set = 1;
8966 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8967 		}
8968 	}
8969 	/* Now do we consume the whole thing? */
8970 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
8971 		/* Its all consumed. */
8972 		uint32_t left;
8973 		uint8_t newly_acked;
8974 
8975 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
8976 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
8977 		rsm->r_rtr_bytes = 0;
8978 		/* Record the time of highest cumack sent */
8979 		rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8980 #ifndef INVARIANTS
8981 		(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8982 #else
8983 		rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8984 		if (rm != rsm) {
8985 			panic("removing head in rack:%p rsm:%p rm:%p",
8986 			      rack, rsm, rm);
8987 		}
8988 #endif
8989 		if (rsm->r_in_tmap) {
8990 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8991 			rsm->r_in_tmap = 0;
8992 		}
8993 		newly_acked = 1;
8994 		if (rsm->r_flags & RACK_ACKED) {
8995 			/*
8996 			 * It was acked on the scoreboard -- remove
8997 			 * it from total
8998 			 */
8999 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
9000 			newly_acked = 0;
9001 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
9002 			/*
9003 			 * There are segments ACKED on the
9004 			 * scoreboard further up. We are seeing
9005 			 * reordering.
9006 			 */
9007 			rsm->r_flags &= ~RACK_SACK_PASSED;
9008 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9009 			rsm->r_flags |= RACK_ACKED;
9010 			rack->r_ctl.rc_reorder_ts = cts;
9011 			if (rack->r_ent_rec_ns) {
9012 				/*
9013 				 * We have sent no more, and we saw an sack
9014 				 * then ack arrive.
9015 				 */
9016 				rack->r_might_revert = 1;
9017 			}
9018 		}
9019 		if ((rsm->r_flags & RACK_TO_REXT) &&
9020 		    (tp->t_flags & TF_RCVD_TSTMP) &&
9021 		    (to->to_flags & TOF_TS) &&
9022 		    (to->to_tsecr != 0) &&
9023 		    (tp->t_flags & TF_PREVVALID)) {
9024 			/*
9025 			 * We can use the timestamp to see
9026 			 * if this retransmission was from the
9027 			 * first transmit. If so we made a mistake.
9028 			 */
9029 			tp->t_flags &= ~TF_PREVVALID;
9030 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
9031 				/* The first transmit is what this ack is for */
9032 				rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
9033 			}
9034 		}
9035 		left = th_ack - rsm->r_end;
9036 		if (rack->app_limited_needs_set && newly_acked)
9037 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
9038 		/* Free back to zone */
9039 		rack_free(rack, rsm);
9040 		if (left) {
9041 			goto more;
9042 		}
9043 		/* Check for reneging */
9044 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9045 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
9046 			/*
9047 			 * The peer has moved snd_una up to
9048 			 * the edge of this send, i.e. one
9049 			 * that it had previously acked. The only
9050 			 * way that can be true if the peer threw
9051 			 * away data (space issues) that it had
9052 			 * previously sacked (else it would have
9053 			 * given us snd_una up to (rsm->r_end).
9054 			 * We need to undo the acked markings here.
9055 			 *
9056 			 * Note we have to look to make sure th_ack is
9057 			 * our rsm->r_start in case we get an old ack
9058 			 * where th_ack is behind snd_una.
9059 			 */
9060 			rack_peer_reneges(rack, rsm, th_ack);
9061 		}
9062 		return;
9063 	}
9064 	if (rsm->r_flags & RACK_ACKED) {
9065 		/*
9066 		 * It was acked on the scoreboard -- remove it from
9067 		 * total for the part being cum-acked.
9068 		 */
9069 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
9070 	}
9071 	/*
9072 	 * Clear the dup ack count for
9073 	 * the piece that remains.
9074 	 */
9075 	rsm->r_dupack = 0;
9076 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9077 	if (rsm->r_rtr_bytes) {
9078 		/*
9079 		 * It was retransmitted adjust the
9080 		 * sack holes for what was acked.
9081 		 */
9082 		int ack_am;
9083 
9084 		ack_am = (th_ack - rsm->r_start);
9085 		if (ack_am >= rsm->r_rtr_bytes) {
9086 			rack->r_ctl.rc_holes_rxt -= ack_am;
9087 			rsm->r_rtr_bytes -= ack_am;
9088 		}
9089 	}
9090 	/*
9091 	 * Update where the piece starts and record
9092 	 * the time of send of highest cumack sent.
9093 	 */
9094 	rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9095 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
9096 	/* Now we need to move our offset forward too */
9097 	if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
9098 		/* Fix up the orig_m_len and possibly the mbuf offset */
9099 		rack_adjust_orig_mlen(rsm);
9100 	}
9101 	rsm->soff += (th_ack - rsm->r_start);
9102 	rsm->r_start = th_ack;
9103 	/* Now do we need to move the mbuf fwd too? */
9104 	if (rsm->m) {
9105 		while (rsm->soff >= rsm->m->m_len) {
9106 			rsm->soff -= rsm->m->m_len;
9107 			rsm->m = rsm->m->m_next;
9108 			KASSERT((rsm->m != NULL),
9109 				(" nrsm:%p hit at soff:%u null m",
9110 				 rsm, rsm->soff));
9111 		}
9112 		rsm->orig_m_len = rsm->m->m_len;
9113 	}
9114 	if (rack->app_limited_needs_set)
9115 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
9116 }
9117 
9118 static void
9119 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
9120 {
9121 	struct rack_sendmap *rsm;
9122 	int sack_pass_fnd = 0;
9123 
9124 	if (rack->r_might_revert) {
9125 		/*
9126 		 * Ok we have reordering, have not sent anything, we
9127 		 * might want to revert the congestion state if nothing
9128 		 * further has SACK_PASSED on it. Lets check.
9129 		 *
9130 		 * We also get here when we have DSACKs come in for
9131 		 * all the data that we FR'd. Note that a rxt or tlp
9132 		 * timer clears this from happening.
9133 		 */
9134 
9135 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
9136 			if (rsm->r_flags & RACK_SACK_PASSED) {
9137 				sack_pass_fnd = 1;
9138 				break;
9139 			}
9140 		}
9141 		if (sack_pass_fnd == 0) {
9142 			/*
9143 			 * We went into recovery
9144 			 * incorrectly due to reordering!
9145 			 */
9146 			int orig_cwnd;
9147 
9148 			rack->r_ent_rec_ns = 0;
9149 			orig_cwnd = tp->snd_cwnd;
9150 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
9151 			tp->snd_recover = tp->snd_una;
9152 			rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
9153 			EXIT_RECOVERY(tp->t_flags);
9154 		}
9155 		rack->r_might_revert = 0;
9156 	}
9157 }
9158 
9159 #ifdef NETFLIX_EXP_DETECTION
9160 static void
9161 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
9162 {
9163 	if ((rack->do_detection || tcp_force_detection) &&
9164 	    tcp_sack_to_ack_thresh &&
9165 	    tcp_sack_to_move_thresh &&
9166 	    ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
9167 		/*
9168 		 * We have thresholds set to find
9169 		 * possible attackers and disable sack.
9170 		 * Check them.
9171 		 */
9172 		uint64_t ackratio, moveratio, movetotal;
9173 
9174 		/* Log detecting */
9175 		rack_log_sad(rack, 1);
9176 		ackratio = (uint64_t)(rack->r_ctl.sack_count);
9177 		ackratio *= (uint64_t)(1000);
9178 		if (rack->r_ctl.ack_count)
9179 			ackratio /= (uint64_t)(rack->r_ctl.ack_count);
9180 		else {
9181 			/* We really should not hit here */
9182 			ackratio = 1000;
9183 		}
9184 		if ((rack->sack_attack_disable == 0) &&
9185 		    (ackratio > rack_highest_sack_thresh_seen))
9186 			rack_highest_sack_thresh_seen = (uint32_t)ackratio;
9187 		movetotal = rack->r_ctl.sack_moved_extra;
9188 		movetotal += rack->r_ctl.sack_noextra_move;
9189 		moveratio = rack->r_ctl.sack_moved_extra;
9190 		moveratio *= (uint64_t)1000;
9191 		if (movetotal)
9192 			moveratio /= movetotal;
9193 		else {
9194 			/* No moves, thats pretty good */
9195 			moveratio = 0;
9196 		}
9197 		if ((rack->sack_attack_disable == 0) &&
9198 		    (moveratio > rack_highest_move_thresh_seen))
9199 			rack_highest_move_thresh_seen = (uint32_t)moveratio;
9200 		if (rack->sack_attack_disable == 0) {
9201 			if ((ackratio > tcp_sack_to_ack_thresh) &&
9202 			    (moveratio > tcp_sack_to_move_thresh)) {
9203 				/* Disable sack processing */
9204 				rack->sack_attack_disable = 1;
9205 				if (rack->r_rep_attack == 0) {
9206 					rack->r_rep_attack = 1;
9207 					counter_u64_add(rack_sack_attacks_detected, 1);
9208 				}
9209 				if (tcp_attack_on_turns_on_logging) {
9210 					/*
9211 					 * Turn on logging, used for debugging
9212 					 * false positives.
9213 					 */
9214 					rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
9215 				}
9216 				/* Clamp the cwnd at flight size */
9217 				rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
9218 				rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9219 				rack_log_sad(rack, 2);
9220 			}
9221 		} else {
9222 			/* We are sack-disabled check for false positives */
9223 			if ((ackratio <= tcp_restoral_thresh) ||
9224 			    (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
9225 				rack->sack_attack_disable = 0;
9226 				rack_log_sad(rack, 3);
9227 				/* Restart counting */
9228 				rack->r_ctl.sack_count = 0;
9229 				rack->r_ctl.sack_moved_extra = 0;
9230 				rack->r_ctl.sack_noextra_move = 1;
9231 				rack->r_ctl.ack_count = max(1,
9232 				      (bytes_this_ack / segsiz));
9233 
9234 				if (rack->r_rep_reverse == 0) {
9235 					rack->r_rep_reverse = 1;
9236 					counter_u64_add(rack_sack_attacks_reversed, 1);
9237 				}
9238 				/* Restore the cwnd */
9239 				if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
9240 					rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
9241 			}
9242 		}
9243 	}
9244 }
9245 #endif
9246 
9247 static int
9248 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
9249 {
9250 
9251 	uint32_t am, l_end;
9252 	int was_tlp = 0;
9253 
9254 	if (SEQ_GT(end, start))
9255 		am = end - start;
9256 	else
9257 		am = 0;
9258 	if ((rack->rc_last_tlp_acked_set ) &&
9259 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
9260 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
9261 		/*
9262 		 * The DSACK is because of a TLP which we don't
9263 		 * do anything with the reordering window over since
9264 		 * it was not reordering that caused the DSACK but
9265 		 * our previous retransmit TLP.
9266 		 */
9267 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
9268 		was_tlp = 1;
9269 		goto skip_dsack_round;
9270 	}
9271 	if (rack->rc_last_sent_tlp_seq_valid) {
9272 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
9273 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
9274 		    (SEQ_LEQ(end, l_end))) {
9275 			/*
9276 			 * This dsack is from the last sent TLP, ignore it
9277 			 * for reordering purposes.
9278 			 */
9279 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
9280 			was_tlp = 1;
9281 			goto skip_dsack_round;
9282 		}
9283 	}
9284 	if (rack->rc_dsack_round_seen == 0) {
9285 		rack->rc_dsack_round_seen = 1;
9286 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
9287 		rack->r_ctl.num_dsack++;
9288 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
9289 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
9290 	}
9291 skip_dsack_round:
9292 	/*
9293 	 * We keep track of how many DSACK blocks we get
9294 	 * after a recovery incident.
9295 	 */
9296 	rack->r_ctl.dsack_byte_cnt += am;
9297 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
9298 	    rack->r_ctl.retran_during_recovery &&
9299 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
9300 		/*
9301 		 * False recovery most likely culprit is reordering. If
9302 		 * nothing else is missing we need to revert.
9303 		 */
9304 		rack->r_might_revert = 1;
9305 		rack_handle_might_revert(rack->rc_tp, rack);
9306 		rack->r_might_revert = 0;
9307 		rack->r_ctl.retran_during_recovery = 0;
9308 		rack->r_ctl.dsack_byte_cnt = 0;
9309 	}
9310 	return (was_tlp);
9311 }
9312 
9313 static uint32_t
9314 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
9315 {
9316 	return (((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt);
9317 }
9318 
9319 static int32_t
9320 rack_compute_pipe(struct tcpcb *tp)
9321 {
9322 	return ((int32_t)do_rack_compute_pipe(tp,
9323 					      (struct tcp_rack *)tp->t_fb_ptr,
9324 					      tp->snd_una));
9325 }
9326 
9327 static void
9328 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9329 {
9330 	/* Deal with changed and PRR here (in recovery only) */
9331 	uint32_t pipe, snd_una;
9332 
9333 	rack->r_ctl.rc_prr_delivered += changed;
9334 
9335 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9336 		/*
9337 		 * It is all outstanding, we are application limited
9338 		 * and thus we don't need more room to send anything.
9339 		 * Note we use tp->snd_una here and not th_ack because
9340 		 * the data as yet not been cut from the sb.
9341 		 */
9342 		rack->r_ctl.rc_prr_sndcnt = 0;
9343 		return;
9344 	}
9345 	/* Compute prr_sndcnt */
9346 	if (SEQ_GT(tp->snd_una, th_ack)) {
9347 		snd_una = tp->snd_una;
9348 	} else {
9349 		snd_una = th_ack;
9350 	}
9351 	pipe = do_rack_compute_pipe(tp, rack, snd_una);
9352 	if (pipe > tp->snd_ssthresh) {
9353 		long sndcnt;
9354 
9355 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9356 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
9357 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9358 		else {
9359 			rack->r_ctl.rc_prr_sndcnt = 0;
9360 			rack_log_to_prr(rack, 9, 0, __LINE__);
9361 			sndcnt = 0;
9362 		}
9363 		sndcnt++;
9364 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9365 			sndcnt -= rack->r_ctl.rc_prr_out;
9366 		else
9367 			sndcnt = 0;
9368 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
9369 		rack_log_to_prr(rack, 10, 0, __LINE__);
9370 	} else {
9371 		uint32_t limit;
9372 
9373 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9374 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9375 		else
9376 			limit = 0;
9377 		if (changed > limit)
9378 			limit = changed;
9379 		limit += ctf_fixed_maxseg(tp);
9380 		if (tp->snd_ssthresh > pipe) {
9381 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9382 			rack_log_to_prr(rack, 11, 0, __LINE__);
9383 		} else {
9384 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9385 			rack_log_to_prr(rack, 12, 0, __LINE__);
9386 		}
9387 	}
9388 }
9389 
9390 static void
9391 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9392 {
9393 	uint32_t changed;
9394 	struct tcp_rack *rack;
9395 	struct rack_sendmap *rsm;
9396 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9397 	register uint32_t th_ack;
9398 	int32_t i, j, k, num_sack_blks = 0;
9399 	uint32_t cts, acked, ack_point;
9400 	int loop_start = 0, moved_two = 0;
9401 	uint32_t tsused;
9402 
9403 
9404 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9405 	if (tcp_get_flags(th) & TH_RST) {
9406 		/* We don't log resets */
9407 		return;
9408 	}
9409 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9410 	cts = tcp_get_usecs(NULL);
9411 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9412 	changed = 0;
9413 	th_ack = th->th_ack;
9414 	if (rack->sack_attack_disable == 0)
9415 		rack_do_decay(rack);
9416 	if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9417 		/*
9418 		 * You only get credit for
9419 		 * MSS and greater (and you get extra
9420 		 * credit for larger cum-ack moves).
9421 		 */
9422 		int ac;
9423 
9424 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9425 		rack->r_ctl.ack_count += ac;
9426 		counter_u64_add(rack_ack_total, ac);
9427 	}
9428 	if (rack->r_ctl.ack_count > 0xfff00000) {
9429 		/*
9430 		 * reduce the number to keep us under
9431 		 * a uint32_t.
9432 		 */
9433 		rack->r_ctl.ack_count /= 2;
9434 		rack->r_ctl.sack_count /= 2;
9435 	}
9436 	if (SEQ_GT(th_ack, tp->snd_una)) {
9437 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9438 		tp->t_acktime = ticks;
9439 	}
9440 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
9441 		changed = th_ack - rsm->r_start;
9442 	if (changed) {
9443 		rack_process_to_cumack(tp, rack, th_ack, cts, to);
9444 	}
9445 	if ((to->to_flags & TOF_SACK) == 0) {
9446 		/* We are done nothing left and no sack. */
9447 		rack_handle_might_revert(tp, rack);
9448 		/*
9449 		 * For cases where we struck a dup-ack
9450 		 * with no SACK, add to the changes so
9451 		 * PRR will work right.
9452 		 */
9453 		if (dup_ack_struck && (changed == 0)) {
9454 			changed += ctf_fixed_maxseg(rack->rc_tp);
9455 		}
9456 		goto out;
9457 	}
9458 	/* Sack block processing */
9459 	if (SEQ_GT(th_ack, tp->snd_una))
9460 		ack_point = th_ack;
9461 	else
9462 		ack_point = tp->snd_una;
9463 	for (i = 0; i < to->to_nsacks; i++) {
9464 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
9465 		      &sack, sizeof(sack));
9466 		sack.start = ntohl(sack.start);
9467 		sack.end = ntohl(sack.end);
9468 		if (SEQ_GT(sack.end, sack.start) &&
9469 		    SEQ_GT(sack.start, ack_point) &&
9470 		    SEQ_LT(sack.start, tp->snd_max) &&
9471 		    SEQ_GT(sack.end, ack_point) &&
9472 		    SEQ_LEQ(sack.end, tp->snd_max)) {
9473 			sack_blocks[num_sack_blks] = sack;
9474 			num_sack_blks++;
9475 		} else if (SEQ_LEQ(sack.start, th_ack) &&
9476 			   SEQ_LEQ(sack.end, th_ack)) {
9477 			int was_tlp;
9478 
9479 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
9480 			/*
9481 			 * Its a D-SACK block.
9482 			 */
9483 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
9484 		}
9485 	}
9486 	if (rack->rc_dsack_round_seen) {
9487 		/* Is the dsack roound over? */
9488 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
9489 			/* Yes it is */
9490 			rack->rc_dsack_round_seen = 0;
9491 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
9492 		}
9493 	}
9494 	/*
9495 	 * Sort the SACK blocks so we can update the rack scoreboard with
9496 	 * just one pass.
9497 	 */
9498 	num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9499 					 num_sack_blks, th->th_ack);
9500 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9501 	if (num_sack_blks == 0) {
9502 		/* Nothing to sack (DSACKs?) */
9503 		goto out_with_totals;
9504 	}
9505 	if (num_sack_blks < 2) {
9506 		/* Only one, we don't need to sort */
9507 		goto do_sack_work;
9508 	}
9509 	/* Sort the sacks */
9510 	for (i = 0; i < num_sack_blks; i++) {
9511 		for (j = i + 1; j < num_sack_blks; j++) {
9512 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9513 				sack = sack_blocks[i];
9514 				sack_blocks[i] = sack_blocks[j];
9515 				sack_blocks[j] = sack;
9516 			}
9517 		}
9518 	}
9519 	/*
9520 	 * Now are any of the sack block ends the same (yes some
9521 	 * implementations send these)?
9522 	 */
9523 again:
9524 	if (num_sack_blks == 0)
9525 		goto out_with_totals;
9526 	if (num_sack_blks > 1) {
9527 		for (i = 0; i < num_sack_blks; i++) {
9528 			for (j = i + 1; j < num_sack_blks; j++) {
9529 				if (sack_blocks[i].end == sack_blocks[j].end) {
9530 					/*
9531 					 * Ok these two have the same end we
9532 					 * want the smallest end and then
9533 					 * throw away the larger and start
9534 					 * again.
9535 					 */
9536 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9537 						/*
9538 						 * The second block covers
9539 						 * more area use that
9540 						 */
9541 						sack_blocks[i].start = sack_blocks[j].start;
9542 					}
9543 					/*
9544 					 * Now collapse out the dup-sack and
9545 					 * lower the count
9546 					 */
9547 					for (k = (j + 1); k < num_sack_blks; k++) {
9548 						sack_blocks[j].start = sack_blocks[k].start;
9549 						sack_blocks[j].end = sack_blocks[k].end;
9550 						j++;
9551 					}
9552 					num_sack_blks--;
9553 					goto again;
9554 				}
9555 			}
9556 		}
9557 	}
9558 do_sack_work:
9559 	/*
9560 	 * First lets look to see if
9561 	 * we have retransmitted and
9562 	 * can use the transmit next?
9563 	 */
9564 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9565 	if (rsm &&
9566 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9567 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9568 		/*
9569 		 * We probably did the FR and the next
9570 		 * SACK in continues as we would expect.
9571 		 */
9572 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9573 		if (acked) {
9574 			rack->r_wanted_output = 1;
9575 			changed += acked;
9576 		}
9577 		if (num_sack_blks == 1) {
9578 			/*
9579 			 * This is what we would expect from
9580 			 * a normal implementation to happen
9581 			 * after we have retransmitted the FR,
9582 			 * i.e the sack-filter pushes down
9583 			 * to 1 block and the next to be retransmitted
9584 			 * is the sequence in the sack block (has more
9585 			 * are acked). Count this as ACK'd data to boost
9586 			 * up the chances of recovering any false positives.
9587 			 */
9588 			rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9589 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9590 			counter_u64_add(rack_express_sack, 1);
9591 			if (rack->r_ctl.ack_count > 0xfff00000) {
9592 				/*
9593 				 * reduce the number to keep us under
9594 				 * a uint32_t.
9595 				 */
9596 				rack->r_ctl.ack_count /= 2;
9597 				rack->r_ctl.sack_count /= 2;
9598 			}
9599 			goto out_with_totals;
9600 		} else {
9601 			/*
9602 			 * Start the loop through the
9603 			 * rest of blocks, past the first block.
9604 			 */
9605 			moved_two = 0;
9606 			loop_start = 1;
9607 		}
9608 	}
9609 	/* Its a sack of some sort */
9610 	rack->r_ctl.sack_count++;
9611 	if (rack->r_ctl.sack_count > 0xfff00000) {
9612 		/*
9613 		 * reduce the number to keep us under
9614 		 * a uint32_t.
9615 		 */
9616 		rack->r_ctl.ack_count /= 2;
9617 		rack->r_ctl.sack_count /= 2;
9618 	}
9619 	counter_u64_add(rack_sack_total, 1);
9620 	if (rack->sack_attack_disable) {
9621 		/* An attacker disablement is in place */
9622 		if (num_sack_blks > 1) {
9623 			rack->r_ctl.sack_count += (num_sack_blks - 1);
9624 			rack->r_ctl.sack_moved_extra++;
9625 			counter_u64_add(rack_move_some, 1);
9626 			if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9627 				rack->r_ctl.sack_moved_extra /= 2;
9628 				rack->r_ctl.sack_noextra_move /= 2;
9629 			}
9630 		}
9631 		goto out;
9632 	}
9633 	rsm = rack->r_ctl.rc_sacklast;
9634 	for (i = loop_start; i < num_sack_blks; i++) {
9635 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9636 		if (acked) {
9637 			rack->r_wanted_output = 1;
9638 			changed += acked;
9639 		}
9640 		if (moved_two) {
9641 			/*
9642 			 * If we did not get a SACK for at least a MSS and
9643 			 * had to move at all, or if we moved more than our
9644 			 * threshold, it counts against the "extra" move.
9645 			 */
9646 			rack->r_ctl.sack_moved_extra += moved_two;
9647 			counter_u64_add(rack_move_some, 1);
9648 		} else {
9649 			/*
9650 			 * else we did not have to move
9651 			 * any more than we would expect.
9652 			 */
9653 			rack->r_ctl.sack_noextra_move++;
9654 			counter_u64_add(rack_move_none, 1);
9655 		}
9656 		if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9657 			/*
9658 			 * If the SACK was not a full MSS then
9659 			 * we add to sack_count the number of
9660 			 * MSS's (or possibly more than
9661 			 * a MSS if its a TSO send) we had to skip by.
9662 			 */
9663 			rack->r_ctl.sack_count += moved_two;
9664 			counter_u64_add(rack_sack_total, moved_two);
9665 		}
9666 		/*
9667 		 * Now we need to setup for the next
9668 		 * round. First we make sure we won't
9669 		 * exceed the size of our uint32_t on
9670 		 * the various counts, and then clear out
9671 		 * moved_two.
9672 		 */
9673 		if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9674 		    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9675 			rack->r_ctl.sack_moved_extra /= 2;
9676 			rack->r_ctl.sack_noextra_move /= 2;
9677 		}
9678 		if (rack->r_ctl.sack_count > 0xfff00000) {
9679 			rack->r_ctl.ack_count /= 2;
9680 			rack->r_ctl.sack_count /= 2;
9681 		}
9682 		moved_two = 0;
9683 	}
9684 out_with_totals:
9685 	if (num_sack_blks > 1) {
9686 		/*
9687 		 * You get an extra stroke if
9688 		 * you have more than one sack-blk, this
9689 		 * could be where we are skipping forward
9690 		 * and the sack-filter is still working, or
9691 		 * it could be an attacker constantly
9692 		 * moving us.
9693 		 */
9694 		rack->r_ctl.sack_moved_extra++;
9695 		counter_u64_add(rack_move_some, 1);
9696 	}
9697 out:
9698 #ifdef NETFLIX_EXP_DETECTION
9699 	rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9700 #endif
9701 	if (changed) {
9702 		/* Something changed cancel the rack timer */
9703 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9704 	}
9705 	tsused = tcp_get_usecs(NULL);
9706 	rsm = tcp_rack_output(tp, rack, tsused);
9707 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9708 	    rsm &&
9709 	    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
9710 		/* Enter recovery */
9711 		entered_recovery = 1;
9712 		rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
9713 		/*
9714 		 * When we enter recovery we need to assure we send
9715 		 * one packet.
9716 		 */
9717 		if (rack->rack_no_prr == 0) {
9718 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9719 			rack_log_to_prr(rack, 8, 0, __LINE__);
9720 		}
9721 		rack->r_timer_override = 1;
9722 		rack->r_early = 0;
9723 		rack->r_ctl.rc_agg_early = 0;
9724 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
9725 		   rsm &&
9726 		   (rack->r_rr_config == 3)) {
9727 		/*
9728 		 * Assure we can output and we get no
9729 		 * remembered pace time except the retransmit.
9730 		 */
9731 		rack->r_timer_override = 1;
9732 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9733 		rack->r_ctl.rc_resend = rsm;
9734 	}
9735 	if (IN_FASTRECOVERY(tp->t_flags) &&
9736 	    (rack->rack_no_prr == 0) &&
9737 	    (entered_recovery == 0)) {
9738 		rack_update_prr(tp, rack, changed, th_ack);
9739 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9740 		     ((tcp_in_hpts(rack->rc_inp) == 0) &&
9741 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9742 			/*
9743 			 * If you are pacing output you don't want
9744 			 * to override.
9745 			 */
9746 			rack->r_early = 0;
9747 			rack->r_ctl.rc_agg_early = 0;
9748 			rack->r_timer_override = 1;
9749 		}
9750 	}
9751 }
9752 
9753 static void
9754 rack_strike_dupack(struct tcp_rack *rack)
9755 {
9756 	struct rack_sendmap *rsm;
9757 
9758 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9759 	while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9760 		rsm = TAILQ_NEXT(rsm, r_tnext);
9761 		if (rsm->r_flags & RACK_MUST_RXT) {
9762 			/* Sendmap entries that are marked to
9763 			 * be retransmitted do not need dupack's
9764 			 * struck. We get these marks for a number
9765 			 * of reasons (rxt timeout with no sack,
9766 			 * mtu change, or rwnd collapses). When
9767 			 * these events occur, we know we must retransmit
9768 			 * them and mark the sendmap entries. Dupack counting
9769 			 * is not needed since we are already set to retransmit
9770 			 * it as soon as we can.
9771 			 */
9772 			continue;
9773 		}
9774 	}
9775 	if (rsm && (rsm->r_dupack < 0xff)) {
9776 		rsm->r_dupack++;
9777 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9778 			struct timeval tv;
9779 			uint32_t cts;
9780 			/*
9781 			 * Here we see if we need to retransmit. For
9782 			 * a SACK type connection if enough time has passed
9783 			 * we will get a return of the rsm. For a non-sack
9784 			 * connection we will get the rsm returned if the
9785 			 * dupack value is 3 or more.
9786 			 */
9787 			cts = tcp_get_usecs(&tv);
9788 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9789 			if (rack->r_ctl.rc_resend != NULL) {
9790 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9791 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9792 							 rack->rc_tp->snd_una, __LINE__);
9793 				}
9794 				rack->r_wanted_output = 1;
9795 				rack->r_timer_override = 1;
9796 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9797 			}
9798 		} else {
9799 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9800 		}
9801 	}
9802 }
9803 
9804 static void
9805 rack_check_bottom_drag(struct tcpcb *tp,
9806 		       struct tcp_rack *rack,
9807 		       struct socket *so, int32_t acked)
9808 {
9809 	uint32_t segsiz, minseg;
9810 
9811 	segsiz = ctf_fixed_maxseg(tp);
9812 	minseg = segsiz;
9813 
9814 	if (tp->snd_max == tp->snd_una) {
9815 		/*
9816 		 * We are doing dynamic pacing and we are way
9817 		 * under. Basically everything got acked while
9818 		 * we were still waiting on the pacer to expire.
9819 		 *
9820 		 * This means we need to boost the b/w in
9821 		 * addition to any earlier boosting of
9822 		 * the multiplier.
9823 		 */
9824 		rack->rc_dragged_bottom = 1;
9825 		rack_validate_multipliers_at_or_above100(rack);
9826 		/*
9827 		 * Lets use the segment bytes acked plus
9828 		 * the lowest RTT seen as the basis to
9829 		 * form a b/w estimate. This will be off
9830 		 * due to the fact that the true estimate
9831 		 * should be around 1/2 the time of the RTT
9832 		 * but we can settle for that.
9833 		 */
9834 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9835 		    acked) {
9836 			uint64_t bw, calc_bw, rtt;
9837 
9838 			rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9839 			if (rtt == 0) {
9840 				/* no us sample is there a ms one? */
9841 				if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9842 					rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9843 				} else {
9844 					goto no_measurement;
9845 				}
9846 			}
9847 			bw = acked;
9848 			calc_bw = bw * 1000000;
9849 			calc_bw /= rtt;
9850 			if (rack->r_ctl.last_max_bw &&
9851 			    (rack->r_ctl.last_max_bw < calc_bw)) {
9852 				/*
9853 				 * If we have a last calculated max bw
9854 				 * enforce it.
9855 				 */
9856 				calc_bw = rack->r_ctl.last_max_bw;
9857 			}
9858 			/* now plop it in */
9859 			if (rack->rc_gp_filled == 0) {
9860 				if (calc_bw > ONE_POINT_TWO_MEG) {
9861 					/*
9862 					 * If we have no measurement
9863 					 * don't let us set in more than
9864 					 * 1.2Mbps. If we are still too
9865 					 * low after pacing with this we
9866 					 * will hopefully have a max b/w
9867 					 * available to sanity check things.
9868 					 */
9869 					calc_bw = ONE_POINT_TWO_MEG;
9870 				}
9871 				rack->r_ctl.rc_rtt_diff = 0;
9872 				rack->r_ctl.gp_bw = calc_bw;
9873 				rack->rc_gp_filled = 1;
9874 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9875 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
9876 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9877 			} else if (calc_bw > rack->r_ctl.gp_bw) {
9878 				rack->r_ctl.rc_rtt_diff = 0;
9879 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9880 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
9881 				rack->r_ctl.gp_bw = calc_bw;
9882 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9883 			} else
9884 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
9885 			if ((rack->gp_ready == 0) &&
9886 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9887 				/* We have enough measurements now */
9888 				rack->gp_ready = 1;
9889 				rack_set_cc_pacing(rack);
9890 				if (rack->defer_options)
9891 					rack_apply_deferred_options(rack);
9892 			}
9893 			/*
9894 			 * For acks over 1mss we do a extra boost to simulate
9895 			 * where we would get 2 acks (we want 110 for the mul).
9896 			 */
9897 			if (acked > segsiz)
9898 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
9899 		} else {
9900 			/*
9901 			 * zero rtt possibly?, settle for just an old increase.
9902 			 */
9903 no_measurement:
9904 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
9905 		}
9906 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9907 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9908 					       minseg)) &&
9909 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9910 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9911 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9912 		    (segsiz * rack_req_segs))) {
9913 		/*
9914 		 * We are doing dynamic GP pacing and
9915 		 * we have everything except 1MSS or less
9916 		 * bytes left out. We are still pacing away.
9917 		 * And there is data that could be sent, This
9918 		 * means we are inserting delayed ack time in
9919 		 * our measurements because we are pacing too slow.
9920 		 */
9921 		rack_validate_multipliers_at_or_above100(rack);
9922 		rack->rc_dragged_bottom = 1;
9923 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
9924 	}
9925 }
9926 
9927 
9928 
9929 static void
9930 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
9931 {
9932 	/*
9933 	 * The fast output path is enabled and we
9934 	 * have moved the cumack forward. Lets see if
9935 	 * we can expand forward the fast path length by
9936 	 * that amount. What we would ideally like to
9937 	 * do is increase the number of bytes in the
9938 	 * fast path block (left_to_send) by the
9939 	 * acked amount. However we have to gate that
9940 	 * by two factors:
9941 	 * 1) The amount outstanding and the rwnd of the peer
9942 	 *    (i.e. we don't want to exceed the rwnd of the peer).
9943 	 *    <and>
9944 	 * 2) The amount of data left in the socket buffer (i.e.
9945 	 *    we can't send beyond what is in the buffer).
9946 	 *
9947 	 * Note that this does not take into account any increase
9948 	 * in the cwnd. We will only extend the fast path by
9949 	 * what was acked.
9950 	 */
9951 	uint32_t new_total, gating_val;
9952 
9953 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
9954 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
9955 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
9956 	if (new_total <= gating_val) {
9957 		/* We can increase left_to_send by the acked amount */
9958 		counter_u64_add(rack_extended_rfo, 1);
9959 		rack->r_ctl.fsb.left_to_send = new_total;
9960 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
9961 			("rack:%p left_to_send:%u sbavail:%u out:%u",
9962 			 rack, rack->r_ctl.fsb.left_to_send,
9963 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
9964 			 (tp->snd_max - tp->snd_una)));
9965 
9966 	}
9967 }
9968 
9969 static void
9970 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
9971 {
9972 	/*
9973 	 * Here any sendmap entry that points to the
9974 	 * beginning mbuf must be adjusted to the correct
9975 	 * offset. This must be called with:
9976 	 * 1) The socket buffer locked
9977 	 * 2) snd_una adjusted to its new postion.
9978 	 *
9979 	 * Note that (2) implies rack_ack_received has also
9980 	 * been called.
9981 	 *
9982 	 * We grab the first mbuf in the socket buffer and
9983 	 * then go through the front of the sendmap, recalculating
9984 	 * the stored offset for any sendmap entry that has
9985 	 * that mbuf. We must use the sb functions to do this
9986 	 * since its possible an add was done has well as
9987 	 * the subtraction we may have just completed. This should
9988 	 * not be a penalty though, since we just referenced the sb
9989 	 * to go in and trim off the mbufs that we freed (of course
9990 	 * there will be a penalty for the sendmap references though).
9991 	 */
9992 	struct mbuf *m;
9993 	struct rack_sendmap *rsm;
9994 
9995 	SOCKBUF_LOCK_ASSERT(sb);
9996 	m = sb->sb_mb;
9997 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9998 	if ((rsm == NULL) || (m == NULL)) {
9999 		/* Nothing outstanding */
10000 		return;
10001 	}
10002 	while (rsm->m && (rsm->m == m)) {
10003 		/* one to adjust */
10004 #ifdef INVARIANTS
10005 		struct mbuf *tm;
10006 		uint32_t soff;
10007 
10008 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
10009 		if (rsm->orig_m_len != m->m_len) {
10010 			rack_adjust_orig_mlen(rsm);
10011 		}
10012 		if (rsm->soff != soff) {
10013 			/*
10014 			 * This is not a fatal error, we anticipate it
10015 			 * might happen (the else code), so we count it here
10016 			 * so that under invariant we can see that it really
10017 			 * does happen.
10018 			 */
10019 			counter_u64_add(rack_adjust_map_bw, 1);
10020 		}
10021 		rsm->m = tm;
10022 		rsm->soff = soff;
10023 		if (tm)
10024 			rsm->orig_m_len = rsm->m->m_len;
10025 		else
10026 			rsm->orig_m_len = 0;
10027 #else
10028 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
10029 		if (rsm->m)
10030 			rsm->orig_m_len = rsm->m->m_len;
10031 		else
10032 			rsm->orig_m_len = 0;
10033 #endif
10034 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
10035 			      rsm);
10036 		if (rsm == NULL)
10037 			break;
10038 	}
10039 }
10040 
10041 /*
10042  * Return value of 1, we do not need to call rack_process_data().
10043  * return value of 0, rack_process_data can be called.
10044  * For ret_val if its 0 the TCP is locked, if its non-zero
10045  * its unlocked and probably unsafe to touch the TCB.
10046  */
10047 static int
10048 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10049     struct tcpcb *tp, struct tcpopt *to,
10050     uint32_t tiwin, int32_t tlen,
10051     int32_t * ofia, int32_t thflags, int32_t *ret_val)
10052 {
10053 	int32_t ourfinisacked = 0;
10054 	int32_t nsegs, acked_amount;
10055 	int32_t acked;
10056 	struct mbuf *mfree;
10057 	struct tcp_rack *rack;
10058 	int32_t under_pacing = 0;
10059 	int32_t recovery = 0;
10060 
10061 	INP_WLOCK_ASSERT(tptoinpcb(tp));
10062 
10063 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10064 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
10065 		__ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
10066 				      &rack->r_ctl.challenge_ack_ts,
10067 				      &rack->r_ctl.challenge_ack_cnt);
10068 		rack->r_wanted_output = 1;
10069 		return (1);
10070 	}
10071 	if (rack->gp_ready &&
10072 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10073 		under_pacing = 1;
10074 	}
10075 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
10076 		int in_rec, dup_ack_struck = 0;
10077 
10078 		in_rec = IN_FASTRECOVERY(tp->t_flags);
10079 		if (rack->rc_in_persist) {
10080 			tp->t_rxtshift = 0;
10081 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10082 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10083 		}
10084 		if ((th->th_ack == tp->snd_una) &&
10085 		    (tiwin == tp->snd_wnd) &&
10086 		    ((to->to_flags & TOF_SACK) == 0)) {
10087 			rack_strike_dupack(rack);
10088 			dup_ack_struck = 1;
10089 		}
10090 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
10091 	}
10092 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10093 		/*
10094 		 * Old ack, behind (or duplicate to) the last one rcv'd
10095 		 * Note: We mark reordering is occuring if its
10096 		 * less than and we have not closed our window.
10097 		 */
10098 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
10099 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
10100 		}
10101 		return (0);
10102 	}
10103 	/*
10104 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
10105 	 * something we sent.
10106 	 */
10107 	if (tp->t_flags & TF_NEEDSYN) {
10108 		/*
10109 		 * T/TCP: Connection was half-synchronized, and our SYN has
10110 		 * been ACK'd (so connection is now fully synchronized).  Go
10111 		 * to non-starred state, increment snd_una for ACK of SYN,
10112 		 * and check if we can do window scaling.
10113 		 */
10114 		tp->t_flags &= ~TF_NEEDSYN;
10115 		tp->snd_una++;
10116 		/* Do window scaling? */
10117 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10118 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10119 			tp->rcv_scale = tp->request_r_scale;
10120 			/* Send window already scaled. */
10121 		}
10122 	}
10123 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10124 
10125 	acked = BYTES_THIS_ACK(tp, th);
10126 	if (acked) {
10127 		/*
10128 		 * Any time we move the cum-ack forward clear
10129 		 * keep-alive tied probe-not-answered. The
10130 		 * persists clears its own on entry.
10131 		 */
10132 		rack->probe_not_answered = 0;
10133 	}
10134 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10135 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10136 	/*
10137 	 * If we just performed our first retransmit, and the ACK arrives
10138 	 * within our recovery window, then it was a mistake to do the
10139 	 * retransmit in the first place.  Recover our original cwnd and
10140 	 * ssthresh, and proceed to transmit where we left off.
10141 	 */
10142 	if ((tp->t_flags & TF_PREVVALID) &&
10143 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10144 		tp->t_flags &= ~TF_PREVVALID;
10145 		if (tp->t_rxtshift == 1 &&
10146 		    (int)(ticks - tp->t_badrxtwin) < 0)
10147 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
10148 	}
10149 	if (acked) {
10150 		/* assure we are not backed off */
10151 		tp->t_rxtshift = 0;
10152 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10153 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10154 		rack->rc_tlp_in_progress = 0;
10155 		rack->r_ctl.rc_tlp_cnt_out = 0;
10156 		/*
10157 		 * If it is the RXT timer we want to
10158 		 * stop it, so we can restart a TLP.
10159 		 */
10160 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10161 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10162 #ifdef NETFLIX_HTTP_LOGGING
10163 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10164 #endif
10165 	}
10166 	/*
10167 	 * If we have a timestamp reply, update smoothed round trip time. If
10168 	 * no timestamp is present but transmit timer is running and timed
10169 	 * sequence number was acked, update smoothed round trip time. Since
10170 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
10171 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
10172 	 * timer.
10173 	 *
10174 	 * Some boxes send broken timestamp replies during the SYN+ACK
10175 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
10176 	 * and blow up the retransmit timer.
10177 	 */
10178 	/*
10179 	 * If all outstanding data is acked, stop retransmit timer and
10180 	 * remember to restart (more output or persist). If there is more
10181 	 * data to be acked, restart retransmit timer, using current
10182 	 * (possibly backed-off) value.
10183 	 */
10184 	if (acked == 0) {
10185 		if (ofia)
10186 			*ofia = ourfinisacked;
10187 		return (0);
10188 	}
10189 	if (IN_RECOVERY(tp->t_flags)) {
10190 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
10191 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
10192 			tcp_rack_partialack(tp);
10193 		} else {
10194 			rack_post_recovery(tp, th->th_ack);
10195 			recovery = 1;
10196 		}
10197 	}
10198 	/*
10199 	 * Let the congestion control algorithm update congestion control
10200 	 * related information. This typically means increasing the
10201 	 * congestion window.
10202 	 */
10203 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
10204 	SOCKBUF_LOCK(&so->so_snd);
10205 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
10206 	tp->snd_wnd -= acked_amount;
10207 	mfree = sbcut_locked(&so->so_snd, acked_amount);
10208 	if ((sbused(&so->so_snd) == 0) &&
10209 	    (acked > acked_amount) &&
10210 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
10211 	    (tp->t_flags & TF_SENTFIN)) {
10212 		/*
10213 		 * We must be sure our fin
10214 		 * was sent and acked (we can be
10215 		 * in FIN_WAIT_1 without having
10216 		 * sent the fin).
10217 		 */
10218 		ourfinisacked = 1;
10219 	}
10220 	tp->snd_una = th->th_ack;
10221 	if (acked_amount && sbavail(&so->so_snd))
10222 		rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10223 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10224 	/* NB: sowwakeup_locked() does an implicit unlock. */
10225 	sowwakeup_locked(so);
10226 	m_freem(mfree);
10227 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
10228 		tp->snd_recover = tp->snd_una;
10229 
10230 	if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
10231 		tp->snd_nxt = tp->snd_una;
10232 	}
10233 	if (under_pacing &&
10234 	    (rack->use_fixed_rate == 0) &&
10235 	    (rack->in_probe_rtt == 0) &&
10236 	    rack->rc_gp_dyn_mul &&
10237 	    rack->rc_always_pace) {
10238 		/* Check if we are dragging bottom */
10239 		rack_check_bottom_drag(tp, rack, so, acked);
10240 	}
10241 	if (tp->snd_una == tp->snd_max) {
10242 		/* Nothing left outstanding */
10243 		tp->t_flags &= ~TF_PREVVALID;
10244 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10245 		rack->r_ctl.retran_during_recovery = 0;
10246 		rack->r_ctl.dsack_byte_cnt = 0;
10247 		if (rack->r_ctl.rc_went_idle_time == 0)
10248 			rack->r_ctl.rc_went_idle_time = 1;
10249 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10250 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
10251 			tp->t_acktime = 0;
10252 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10253 		/* Set need output so persist might get set */
10254 		rack->r_wanted_output = 1;
10255 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10256 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
10257 		    (sbavail(&so->so_snd) == 0) &&
10258 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
10259 			/*
10260 			 * The socket was gone and the
10261 			 * peer sent data (now or in the past), time to
10262 			 * reset him.
10263 			 */
10264 			*ret_val = 1;
10265 			/* tcp_close will kill the inp pre-log the Reset */
10266 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
10267 			tp = tcp_close(tp);
10268 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
10269 			return (1);
10270 		}
10271 	}
10272 	if (ofia)
10273 		*ofia = ourfinisacked;
10274 	return (0);
10275 }
10276 
10277 
10278 static void
10279 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
10280 		  int dir, uint32_t flags, struct rack_sendmap *rsm)
10281 {
10282 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
10283 		union tcp_log_stackspecific log;
10284 		struct timeval tv;
10285 
10286 		memset(&log, 0, sizeof(log));
10287 		log.u_bbr.flex1 = cnt;
10288 		log.u_bbr.flex2 = split;
10289 		log.u_bbr.flex3 = out;
10290 		log.u_bbr.flex4 = line;
10291 		log.u_bbr.flex5 = rack->r_must_retran;
10292 		log.u_bbr.flex6 = flags;
10293 		log.u_bbr.flex7 = rack->rc_has_collapsed;
10294 		log.u_bbr.flex8 = dir;	/*
10295 					 * 1 is collapsed, 0 is uncollapsed,
10296 					 * 2 is log of a rsm being marked, 3 is a split.
10297 					 */
10298 		if (rsm == NULL)
10299 			log.u_bbr.rttProp = 0;
10300 		else
10301 			log.u_bbr.rttProp = (uint64_t)rsm;
10302 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
10303 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10304 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
10305 		    &rack->rc_inp->inp_socket->so_rcv,
10306 		    &rack->rc_inp->inp_socket->so_snd,
10307 		    TCP_RACK_LOG_COLLAPSE, 0,
10308 		    0, &log, false, &tv);
10309 	}
10310 }
10311 
10312 static void
10313 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, int line)
10314 {
10315 	/*
10316 	 * Here all we do is mark the collapsed point and set the flag.
10317 	 * This may happen again and again, but there is no
10318 	 * sense splitting our map until we know where the
10319 	 * peer finally lands in the collapse.
10320 	 */
10321 	rack_trace_point(rack, RACK_TP_COLLAPSED_WND);
10322 	if ((rack->rc_has_collapsed == 0) ||
10323 	    (rack->r_ctl.last_collapse_point != (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)))
10324 		counter_u64_add(rack_collapsed_win_seen, 1);
10325 	rack->r_ctl.last_collapse_point = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
10326 	rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
10327 	rack->rc_has_collapsed = 1;
10328 	rack->r_collapse_point_valid = 1;
10329 	rack_log_collapse(rack, 0, 0, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
10330 }
10331 
10332 static void
10333 rack_un_collapse_window(struct tcp_rack *rack, int line)
10334 {
10335 	struct rack_sendmap *nrsm, *rsm, fe;
10336 	int cnt = 0, split = 0;
10337 #ifdef INVARIANTS
10338 	struct rack_sendmap *insret;
10339 #endif
10340 
10341 	memset(&fe, 0, sizeof(fe));
10342 	rack->rc_has_collapsed = 0;
10343 	fe.r_start = rack->r_ctl.last_collapse_point;
10344 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
10345 	if (rsm == NULL) {
10346 		/* Nothing to do maybe the peer ack'ed it all */
10347 		rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
10348 		return;
10349 	}
10350 	/* Now do we need to split this one? */
10351 	if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
10352 		rack_log_collapse(rack, rsm->r_start, rsm->r_end,
10353 				  rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
10354 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10355 		if (nrsm == NULL) {
10356 			/* We can't get a rsm, mark all? */
10357 			nrsm = rsm;
10358 			goto no_split;
10359 		}
10360 		/* Clone it */
10361 		split = 1;
10362 		rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
10363 #ifndef INVARIANTS
10364 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10365 #else
10366 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10367 		if (insret != NULL) {
10368 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
10369 			      nrsm, insret, rack, rsm);
10370 		}
10371 #endif
10372 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
10373 				 rack->r_ctl.last_collapse_point, __LINE__);
10374 		if (rsm->r_in_tmap) {
10375 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10376 			nrsm->r_in_tmap = 1;
10377 		}
10378 		/*
10379 		 * Set in the new RSM as the
10380 		 * collapsed starting point
10381 		 */
10382 		rsm = nrsm;
10383 	}
10384 no_split:
10385 	RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
10386 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
10387 		rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
10388 		cnt++;
10389 	}
10390 	if (cnt) {
10391 		counter_u64_add(rack_collapsed_win, 1);
10392 	}
10393 	rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
10394 }
10395 
10396 static void
10397 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10398 			int32_t tlen, int32_t tfo_syn)
10399 {
10400 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
10401 		if (rack->rc_dack_mode &&
10402 		    (tlen > 500) &&
10403 		    (rack->rc_dack_toggle == 1)) {
10404 			goto no_delayed_ack;
10405 		}
10406 		rack_timer_cancel(tp, rack,
10407 				  rack->r_ctl.rc_rcvtime, __LINE__);
10408 		tp->t_flags |= TF_DELACK;
10409 	} else {
10410 no_delayed_ack:
10411 		rack->r_wanted_output = 1;
10412 		tp->t_flags |= TF_ACKNOW;
10413 		if (rack->rc_dack_mode) {
10414 			if (tp->t_flags & TF_DELACK)
10415 				rack->rc_dack_toggle = 1;
10416 			else
10417 				rack->rc_dack_toggle = 0;
10418 		}
10419 	}
10420 }
10421 
10422 static void
10423 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10424 {
10425 	/*
10426 	 * If fast output is in progress, lets validate that
10427 	 * the new window did not shrink on us and make it
10428 	 * so fast output should end.
10429 	 */
10430 	if (rack->r_fast_output) {
10431 		uint32_t out;
10432 
10433 		/*
10434 		 * Calculate what we will send if left as is
10435 		 * and compare that to our send window.
10436 		 */
10437 		out = ctf_outstanding(tp);
10438 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10439 			/* ok we have an issue */
10440 			if (out >= tp->snd_wnd) {
10441 				/* Turn off fast output the window is met or collapsed */
10442 				rack->r_fast_output = 0;
10443 			} else {
10444 				/* we have some room left */
10445 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10446 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10447 					/* If not at least 1 full segment never mind */
10448 					rack->r_fast_output = 0;
10449 				}
10450 			}
10451 		}
10452 	}
10453 }
10454 
10455 
10456 /*
10457  * Return value of 1, the TCB is unlocked and most
10458  * likely gone, return value of 0, the TCP is still
10459  * locked.
10460  */
10461 static int
10462 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10463     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10464     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10465 {
10466 	/*
10467 	 * Update window information. Don't look at window if no ACK: TAC's
10468 	 * send garbage on first SYN.
10469 	 */
10470 	int32_t nsegs;
10471 	int32_t tfo_syn;
10472 	struct tcp_rack *rack;
10473 
10474 	INP_WLOCK_ASSERT(tptoinpcb(tp));
10475 
10476 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10477 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10478 	if ((thflags & TH_ACK) &&
10479 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10480 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10481 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10482 		/* keep track of pure window updates */
10483 		if (tlen == 0 &&
10484 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10485 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10486 		tp->snd_wnd = tiwin;
10487 		rack_validate_fo_sendwin_up(tp, rack);
10488 		tp->snd_wl1 = th->th_seq;
10489 		tp->snd_wl2 = th->th_ack;
10490 		if (tp->snd_wnd > tp->max_sndwnd)
10491 			tp->max_sndwnd = tp->snd_wnd;
10492 		rack->r_wanted_output = 1;
10493 	} else if (thflags & TH_ACK) {
10494 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10495 			tp->snd_wnd = tiwin;
10496 			rack_validate_fo_sendwin_up(tp, rack);
10497 			tp->snd_wl1 = th->th_seq;
10498 			tp->snd_wl2 = th->th_ack;
10499 		}
10500 	}
10501 	if (tp->snd_wnd < ctf_outstanding(tp))
10502 		/* The peer collapsed the window */
10503 		rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__);
10504 	else if (rack->rc_has_collapsed)
10505 		rack_un_collapse_window(rack, __LINE__);
10506 	if ((rack->r_collapse_point_valid) &&
10507 	    (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
10508 		rack->r_collapse_point_valid = 0;
10509 	/* Was persist timer active and now we have window space? */
10510 	if ((rack->rc_in_persist != 0) &&
10511 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10512 				rack->r_ctl.rc_pace_min_segs))) {
10513 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10514 		tp->snd_nxt = tp->snd_max;
10515 		/* Make sure we output to start the timer */
10516 		rack->r_wanted_output = 1;
10517 	}
10518 	/* Do we enter persists? */
10519 	if ((rack->rc_in_persist == 0) &&
10520 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10521 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
10522 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
10523 	    sbavail(&tptosocket(tp)->so_snd) &&
10524 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
10525 		/*
10526 		 * Here the rwnd is less than
10527 		 * the pacing size, we are established,
10528 		 * nothing is outstanding, and there is
10529 		 * data to send. Enter persists.
10530 		 */
10531 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10532 	}
10533 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10534 		m_freem(m);
10535 		return (0);
10536 	}
10537 	/*
10538 	 * don't process the URG bit, ignore them drag
10539 	 * along the up.
10540 	 */
10541 	tp->rcv_up = tp->rcv_nxt;
10542 
10543 	/*
10544 	 * Process the segment text, merging it into the TCP sequencing
10545 	 * queue, and arranging for acknowledgment of receipt if necessary.
10546 	 * This process logically involves adjusting tp->rcv_wnd as data is
10547 	 * presented to the user (this happens in tcp_usrreq.c, case
10548 	 * PRU_RCVD).  If a FIN has already been received on this connection
10549 	 * then we just ignore the text.
10550 	 */
10551 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10552 		   IS_FASTOPEN(tp->t_flags));
10553 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10554 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10555 		tcp_seq save_start = th->th_seq;
10556 		tcp_seq save_rnxt  = tp->rcv_nxt;
10557 		int     save_tlen  = tlen;
10558 
10559 		m_adj(m, drop_hdrlen);	/* delayed header drop */
10560 		/*
10561 		 * Insert segment which includes th into TCP reassembly
10562 		 * queue with control block tp.  Set thflags to whether
10563 		 * reassembly now includes a segment with FIN.  This handles
10564 		 * the common case inline (segment is the next to be
10565 		 * received on an established connection, and the queue is
10566 		 * empty), avoiding linkage into and removal from the queue
10567 		 * and repetition of various conversions. Set DELACK for
10568 		 * segments received in order, but ack immediately when
10569 		 * segments are out of order (so fast retransmit can work).
10570 		 */
10571 		if (th->th_seq == tp->rcv_nxt &&
10572 		    SEGQ_EMPTY(tp) &&
10573 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
10574 		    tfo_syn)) {
10575 #ifdef NETFLIX_SB_LIMITS
10576 			u_int mcnt, appended;
10577 
10578 			if (so->so_rcv.sb_shlim) {
10579 				mcnt = m_memcnt(m);
10580 				appended = 0;
10581 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10582 				    CFO_NOSLEEP, NULL) == false) {
10583 					counter_u64_add(tcp_sb_shlim_fails, 1);
10584 					m_freem(m);
10585 					return (0);
10586 				}
10587 			}
10588 #endif
10589 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10590 			tp->rcv_nxt += tlen;
10591 			if (tlen &&
10592 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10593 			    (tp->t_fbyte_in == 0)) {
10594 				tp->t_fbyte_in = ticks;
10595 				if (tp->t_fbyte_in == 0)
10596 					tp->t_fbyte_in = 1;
10597 				if (tp->t_fbyte_out && tp->t_fbyte_in)
10598 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10599 			}
10600 			thflags = tcp_get_flags(th) & TH_FIN;
10601 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10602 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10603 			SOCKBUF_LOCK(&so->so_rcv);
10604 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10605 				m_freem(m);
10606 			} else
10607 #ifdef NETFLIX_SB_LIMITS
10608 				appended =
10609 #endif
10610 					sbappendstream_locked(&so->so_rcv, m, 0);
10611 
10612 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10613 			/* NB: sorwakeup_locked() does an implicit unlock. */
10614 			sorwakeup_locked(so);
10615 #ifdef NETFLIX_SB_LIMITS
10616 			if (so->so_rcv.sb_shlim && appended != mcnt)
10617 				counter_fo_release(so->so_rcv.sb_shlim,
10618 				    mcnt - appended);
10619 #endif
10620 		} else {
10621 			/*
10622 			 * XXX: Due to the header drop above "th" is
10623 			 * theoretically invalid by now.  Fortunately
10624 			 * m_adj() doesn't actually frees any mbufs when
10625 			 * trimming from the head.
10626 			 */
10627 			tcp_seq temp = save_start;
10628 
10629 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
10630 			tp->t_flags |= TF_ACKNOW;
10631 			if (tp->t_flags & TF_WAKESOR) {
10632 				tp->t_flags &= ~TF_WAKESOR;
10633 				/* NB: sorwakeup_locked() does an implicit unlock. */
10634 				sorwakeup_locked(so);
10635 			}
10636 		}
10637 		if ((tp->t_flags & TF_SACK_PERMIT) &&
10638 		    (save_tlen > 0) &&
10639 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
10640 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10641 				/*
10642 				 * DSACK actually handled in the fastpath
10643 				 * above.
10644 				 */
10645 				RACK_OPTS_INC(tcp_sack_path_1);
10646 				tcp_update_sack_list(tp, save_start,
10647 				    save_start + save_tlen);
10648 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10649 				if ((tp->rcv_numsacks >= 1) &&
10650 				    (tp->sackblks[0].end == save_start)) {
10651 					/*
10652 					 * Partial overlap, recorded at todrop
10653 					 * above.
10654 					 */
10655 					RACK_OPTS_INC(tcp_sack_path_2a);
10656 					tcp_update_sack_list(tp,
10657 					    tp->sackblks[0].start,
10658 					    tp->sackblks[0].end);
10659 				} else {
10660 					RACK_OPTS_INC(tcp_sack_path_2b);
10661 					tcp_update_dsack_list(tp, save_start,
10662 					    save_start + save_tlen);
10663 				}
10664 			} else if (tlen >= save_tlen) {
10665 				/* Update of sackblks. */
10666 				RACK_OPTS_INC(tcp_sack_path_3);
10667 				tcp_update_dsack_list(tp, save_start,
10668 				    save_start + save_tlen);
10669 			} else if (tlen > 0) {
10670 				RACK_OPTS_INC(tcp_sack_path_4);
10671 				tcp_update_dsack_list(tp, save_start,
10672 				    save_start + tlen);
10673 			}
10674 		}
10675 	} else {
10676 		m_freem(m);
10677 		thflags &= ~TH_FIN;
10678 	}
10679 
10680 	/*
10681 	 * If FIN is received ACK the FIN and let the user know that the
10682 	 * connection is closing.
10683 	 */
10684 	if (thflags & TH_FIN) {
10685 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10686 			/* The socket upcall is handled by socantrcvmore. */
10687 			socantrcvmore(so);
10688 			/*
10689 			 * If connection is half-synchronized (ie NEEDSYN
10690 			 * flag on) then delay ACK, so it may be piggybacked
10691 			 * when SYN is sent. Otherwise, since we received a
10692 			 * FIN then no more input can be expected, send ACK
10693 			 * now.
10694 			 */
10695 			if (tp->t_flags & TF_NEEDSYN) {
10696 				rack_timer_cancel(tp, rack,
10697 				    rack->r_ctl.rc_rcvtime, __LINE__);
10698 				tp->t_flags |= TF_DELACK;
10699 			} else {
10700 				tp->t_flags |= TF_ACKNOW;
10701 			}
10702 			tp->rcv_nxt++;
10703 		}
10704 		switch (tp->t_state) {
10705 			/*
10706 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
10707 			 * CLOSE_WAIT state.
10708 			 */
10709 		case TCPS_SYN_RECEIVED:
10710 			tp->t_starttime = ticks;
10711 			/* FALLTHROUGH */
10712 		case TCPS_ESTABLISHED:
10713 			rack_timer_cancel(tp, rack,
10714 			    rack->r_ctl.rc_rcvtime, __LINE__);
10715 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
10716 			break;
10717 
10718 			/*
10719 			 * If still in FIN_WAIT_1 STATE FIN has not been
10720 			 * acked so enter the CLOSING state.
10721 			 */
10722 		case TCPS_FIN_WAIT_1:
10723 			rack_timer_cancel(tp, rack,
10724 			    rack->r_ctl.rc_rcvtime, __LINE__);
10725 			tcp_state_change(tp, TCPS_CLOSING);
10726 			break;
10727 
10728 			/*
10729 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
10730 			 * starting the time-wait timer, turning off the
10731 			 * other standard timers.
10732 			 */
10733 		case TCPS_FIN_WAIT_2:
10734 			rack_timer_cancel(tp, rack,
10735 			    rack->r_ctl.rc_rcvtime, __LINE__);
10736 			tcp_twstart(tp);
10737 			return (1);
10738 		}
10739 	}
10740 	/*
10741 	 * Return any desired output.
10742 	 */
10743 	if ((tp->t_flags & TF_ACKNOW) ||
10744 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10745 		rack->r_wanted_output = 1;
10746 	}
10747 	return (0);
10748 }
10749 
10750 /*
10751  * Here nothing is really faster, its just that we
10752  * have broken out the fast-data path also just like
10753  * the fast-ack.
10754  */
10755 static int
10756 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10757     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10758     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10759 {
10760 	int32_t nsegs;
10761 	int32_t newsize = 0;	/* automatic sockbuf scaling */
10762 	struct tcp_rack *rack;
10763 #ifdef NETFLIX_SB_LIMITS
10764 	u_int mcnt, appended;
10765 #endif
10766 
10767 	/*
10768 	 * If last ACK falls within this segment's sequence numbers, record
10769 	 * the timestamp. NOTE that the test is modified according to the
10770 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10771 	 */
10772 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10773 		return (0);
10774 	}
10775 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10776 		return (0);
10777 	}
10778 	if (tiwin && tiwin != tp->snd_wnd) {
10779 		return (0);
10780 	}
10781 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10782 		return (0);
10783 	}
10784 	if (__predict_false((to->to_flags & TOF_TS) &&
10785 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10786 		return (0);
10787 	}
10788 	if (__predict_false((th->th_ack != tp->snd_una))) {
10789 		return (0);
10790 	}
10791 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10792 		return (0);
10793 	}
10794 	if ((to->to_flags & TOF_TS) != 0 &&
10795 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10796 		tp->ts_recent_age = tcp_ts_getticks();
10797 		tp->ts_recent = to->to_tsval;
10798 	}
10799 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10800 	/*
10801 	 * This is a pure, in-sequence data packet with nothing on the
10802 	 * reassembly queue and we have enough buffer space to take it.
10803 	 */
10804 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10805 
10806 #ifdef NETFLIX_SB_LIMITS
10807 	if (so->so_rcv.sb_shlim) {
10808 		mcnt = m_memcnt(m);
10809 		appended = 0;
10810 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10811 		    CFO_NOSLEEP, NULL) == false) {
10812 			counter_u64_add(tcp_sb_shlim_fails, 1);
10813 			m_freem(m);
10814 			return (1);
10815 		}
10816 	}
10817 #endif
10818 	/* Clean receiver SACK report if present */
10819 	if (tp->rcv_numsacks)
10820 		tcp_clean_sackreport(tp);
10821 	KMOD_TCPSTAT_INC(tcps_preddat);
10822 	tp->rcv_nxt += tlen;
10823 	if (tlen &&
10824 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10825 	    (tp->t_fbyte_in == 0)) {
10826 		tp->t_fbyte_in = ticks;
10827 		if (tp->t_fbyte_in == 0)
10828 			tp->t_fbyte_in = 1;
10829 		if (tp->t_fbyte_out && tp->t_fbyte_in)
10830 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10831 	}
10832 	/*
10833 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10834 	 */
10835 	tp->snd_wl1 = th->th_seq;
10836 	/*
10837 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10838 	 */
10839 	tp->rcv_up = tp->rcv_nxt;
10840 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10841 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10842 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10843 
10844 	/* Add data to socket buffer. */
10845 	SOCKBUF_LOCK(&so->so_rcv);
10846 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10847 		m_freem(m);
10848 	} else {
10849 		/*
10850 		 * Set new socket buffer size. Give up when limit is
10851 		 * reached.
10852 		 */
10853 		if (newsize)
10854 			if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
10855 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10856 		m_adj(m, drop_hdrlen);	/* delayed header drop */
10857 #ifdef NETFLIX_SB_LIMITS
10858 		appended =
10859 #endif
10860 			sbappendstream_locked(&so->so_rcv, m, 0);
10861 		ctf_calc_rwin(so, tp);
10862 	}
10863 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10864 	/* NB: sorwakeup_locked() does an implicit unlock. */
10865 	sorwakeup_locked(so);
10866 #ifdef NETFLIX_SB_LIMITS
10867 	if (so->so_rcv.sb_shlim && mcnt != appended)
10868 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10869 #endif
10870 	rack_handle_delayed_ack(tp, rack, tlen, 0);
10871 	if (tp->snd_una == tp->snd_max)
10872 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10873 	return (1);
10874 }
10875 
10876 /*
10877  * This subfunction is used to try to highly optimize the
10878  * fast path. We again allow window updates that are
10879  * in sequence to remain in the fast-path. We also add
10880  * in the __predict's to attempt to help the compiler.
10881  * Note that if we return a 0, then we can *not* process
10882  * it and the caller should push the packet into the
10883  * slow-path.
10884  */
10885 static int
10886 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10887     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10888     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10889 {
10890 	int32_t acked;
10891 	int32_t nsegs;
10892 	int32_t under_pacing = 0;
10893 	struct tcp_rack *rack;
10894 
10895 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10896 		/* Old ack, behind (or duplicate to) the last one rcv'd */
10897 		return (0);
10898 	}
10899 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
10900 		/* Above what we have sent? */
10901 		return (0);
10902 	}
10903 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10904 		/* We are retransmitting */
10905 		return (0);
10906 	}
10907 	if (__predict_false(tiwin == 0)) {
10908 		/* zero window */
10909 		return (0);
10910 	}
10911 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
10912 		/* We need a SYN or a FIN, unlikely.. */
10913 		return (0);
10914 	}
10915 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
10916 		/* Timestamp is behind .. old ack with seq wrap? */
10917 		return (0);
10918 	}
10919 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
10920 		/* Still recovering */
10921 		return (0);
10922 	}
10923 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10924 	if (rack->r_ctl.rc_sacked) {
10925 		/* We have sack holes on our scoreboard */
10926 		return (0);
10927 	}
10928 	/* Ok if we reach here, we can process a fast-ack */
10929 	if (rack->gp_ready &&
10930 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10931 		under_pacing = 1;
10932 	}
10933 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10934 	rack_log_ack(tp, to, th, 0, 0);
10935 	/* Did the window get updated? */
10936 	if (tiwin != tp->snd_wnd) {
10937 		tp->snd_wnd = tiwin;
10938 		rack_validate_fo_sendwin_up(tp, rack);
10939 		tp->snd_wl1 = th->th_seq;
10940 		if (tp->snd_wnd > tp->max_sndwnd)
10941 			tp->max_sndwnd = tp->snd_wnd;
10942 	}
10943 	/* Do we exit persists? */
10944 	if ((rack->rc_in_persist != 0) &&
10945 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10946 			       rack->r_ctl.rc_pace_min_segs))) {
10947 		rack_exit_persist(tp, rack, cts);
10948 	}
10949 	/* Do we enter persists? */
10950 	if ((rack->rc_in_persist == 0) &&
10951 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10952 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
10953 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
10954 	    sbavail(&tptosocket(tp)->so_snd) &&
10955 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
10956 		/*
10957 		 * Here the rwnd is less than
10958 		 * the pacing size, we are established,
10959 		 * nothing is outstanding, and there is
10960 		 * data to send. Enter persists.
10961 		 */
10962 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10963 	}
10964 	/*
10965 	 * If last ACK falls within this segment's sequence numbers, record
10966 	 * the timestamp. NOTE that the test is modified according to the
10967 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10968 	 */
10969 	if ((to->to_flags & TOF_TS) != 0 &&
10970 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10971 		tp->ts_recent_age = tcp_ts_getticks();
10972 		tp->ts_recent = to->to_tsval;
10973 	}
10974 	/*
10975 	 * This is a pure ack for outstanding data.
10976 	 */
10977 	KMOD_TCPSTAT_INC(tcps_predack);
10978 
10979 	/*
10980 	 * "bad retransmit" recovery.
10981 	 */
10982 	if ((tp->t_flags & TF_PREVVALID) &&
10983 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10984 		tp->t_flags &= ~TF_PREVVALID;
10985 		if (tp->t_rxtshift == 1 &&
10986 		    (int)(ticks - tp->t_badrxtwin) < 0)
10987 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
10988 	}
10989 	/*
10990 	 * Recalculate the transmit timer / rtt.
10991 	 *
10992 	 * Some boxes send broken timestamp replies during the SYN+ACK
10993 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
10994 	 * and blow up the retransmit timer.
10995 	 */
10996 	acked = BYTES_THIS_ACK(tp, th);
10997 
10998 #ifdef TCP_HHOOK
10999 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
11000 	hhook_run_tcp_est_in(tp, th, to);
11001 #endif
11002 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
11003 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
11004 	if (acked) {
11005 		struct mbuf *mfree;
11006 
11007 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
11008 		SOCKBUF_LOCK(&so->so_snd);
11009 		mfree = sbcut_locked(&so->so_snd, acked);
11010 		tp->snd_una = th->th_ack;
11011 		/* Note we want to hold the sb lock through the sendmap adjust */
11012 		rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
11013 		/* Wake up the socket if we have room to write more */
11014 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
11015 		sowwakeup_locked(so);
11016 		m_freem(mfree);
11017 		tp->t_rxtshift = 0;
11018 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11019 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11020 		rack->rc_tlp_in_progress = 0;
11021 		rack->r_ctl.rc_tlp_cnt_out = 0;
11022 		/*
11023 		 * If it is the RXT timer we want to
11024 		 * stop it, so we can restart a TLP.
11025 		 */
11026 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
11027 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11028 #ifdef NETFLIX_HTTP_LOGGING
11029 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
11030 #endif
11031 	}
11032 	/*
11033 	 * Let the congestion control algorithm update congestion control
11034 	 * related information. This typically means increasing the
11035 	 * congestion window.
11036 	 */
11037 	if (tp->snd_wnd < ctf_outstanding(tp)) {
11038 		/* The peer collapsed the window */
11039 		rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__);
11040 	} else if (rack->rc_has_collapsed)
11041 		rack_un_collapse_window(rack, __LINE__);
11042 	if ((rack->r_collapse_point_valid) &&
11043 	    (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
11044 		rack->r_collapse_point_valid = 0;
11045 	/*
11046 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
11047 	 */
11048 	tp->snd_wl2 = th->th_ack;
11049 	tp->t_dupacks = 0;
11050 	m_freem(m);
11051 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
11052 
11053 	/*
11054 	 * If all outstanding data are acked, stop retransmit timer,
11055 	 * otherwise restart timer using current (possibly backed-off)
11056 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
11057 	 * If data are ready to send, let tcp_output decide between more
11058 	 * output or persist.
11059 	 */
11060 	if (under_pacing &&
11061 	    (rack->use_fixed_rate == 0) &&
11062 	    (rack->in_probe_rtt == 0) &&
11063 	    rack->rc_gp_dyn_mul &&
11064 	    rack->rc_always_pace) {
11065 		/* Check if we are dragging bottom */
11066 		rack_check_bottom_drag(tp, rack, so, acked);
11067 	}
11068 	if (tp->snd_una == tp->snd_max) {
11069 		tp->t_flags &= ~TF_PREVVALID;
11070 		rack->r_ctl.retran_during_recovery = 0;
11071 		rack->r_ctl.dsack_byte_cnt = 0;
11072 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
11073 		if (rack->r_ctl.rc_went_idle_time == 0)
11074 			rack->r_ctl.rc_went_idle_time = 1;
11075 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
11076 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
11077 			tp->t_acktime = 0;
11078 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11079 	}
11080 	if (acked && rack->r_fast_output)
11081 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
11082 	if (sbavail(&so->so_snd)) {
11083 		rack->r_wanted_output = 1;
11084 	}
11085 	return (1);
11086 }
11087 
11088 /*
11089  * Return value of 1, the TCB is unlocked and most
11090  * likely gone, return value of 0, the TCP is still
11091  * locked.
11092  */
11093 static int
11094 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
11095     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11096     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11097 {
11098 	int32_t ret_val = 0;
11099 	int32_t todrop;
11100 	int32_t ourfinisacked = 0;
11101 	struct tcp_rack *rack;
11102 
11103 	INP_WLOCK_ASSERT(tptoinpcb(tp));
11104 
11105 	ctf_calc_rwin(so, tp);
11106 	/*
11107 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
11108 	 * SYN, drop the input. if seg contains a RST, then drop the
11109 	 * connection. if seg does not contain SYN, then drop it. Otherwise
11110 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
11111 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
11112 	 * contains an ECE and ECN support is enabled, the stream is ECN
11113 	 * capable. if SYN has been acked change to ESTABLISHED else
11114 	 * SYN_RCVD state arrange for segment to be acked (eventually)
11115 	 * continue processing rest of data/controls.
11116 	 */
11117 	if ((thflags & TH_ACK) &&
11118 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
11119 	    SEQ_GT(th->th_ack, tp->snd_max))) {
11120 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11121 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11122 		return (1);
11123 	}
11124 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
11125 		TCP_PROBE5(connect__refused, NULL, tp,
11126 		    mtod(m, const char *), tp, th);
11127 		tp = tcp_drop(tp, ECONNREFUSED);
11128 		ctf_do_drop(m, tp);
11129 		return (1);
11130 	}
11131 	if (thflags & TH_RST) {
11132 		ctf_do_drop(m, tp);
11133 		return (1);
11134 	}
11135 	if (!(thflags & TH_SYN)) {
11136 		ctf_do_drop(m, tp);
11137 		return (1);
11138 	}
11139 	tp->irs = th->th_seq;
11140 	tcp_rcvseqinit(tp);
11141 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11142 	if (thflags & TH_ACK) {
11143 		int tfo_partial = 0;
11144 
11145 		KMOD_TCPSTAT_INC(tcps_connects);
11146 		soisconnected(so);
11147 #ifdef MAC
11148 		mac_socketpeer_set_from_mbuf(m, so);
11149 #endif
11150 		/* Do window scaling on this connection? */
11151 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11152 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11153 			tp->rcv_scale = tp->request_r_scale;
11154 		}
11155 		tp->rcv_adv += min(tp->rcv_wnd,
11156 		    TCP_MAXWIN << tp->rcv_scale);
11157 		/*
11158 		 * If not all the data that was sent in the TFO SYN
11159 		 * has been acked, resend the remainder right away.
11160 		 */
11161 		if (IS_FASTOPEN(tp->t_flags) &&
11162 		    (tp->snd_una != tp->snd_max)) {
11163 			tp->snd_nxt = th->th_ack;
11164 			tfo_partial = 1;
11165 		}
11166 		/*
11167 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
11168 		 * will be turned on later.
11169 		 */
11170 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
11171 			rack_timer_cancel(tp, rack,
11172 					  rack->r_ctl.rc_rcvtime, __LINE__);
11173 			tp->t_flags |= TF_DELACK;
11174 		} else {
11175 			rack->r_wanted_output = 1;
11176 			tp->t_flags |= TF_ACKNOW;
11177 			rack->rc_dack_toggle = 0;
11178 		}
11179 
11180 		tcp_ecn_input_syn_sent(tp, thflags, iptos);
11181 
11182 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
11183 			/*
11184 			 * We advance snd_una for the
11185 			 * fast open case. If th_ack is
11186 			 * acknowledging data beyond
11187 			 * snd_una we can't just call
11188 			 * ack-processing since the
11189 			 * data stream in our send-map
11190 			 * will start at snd_una + 1 (one
11191 			 * beyond the SYN). If its just
11192 			 * equal we don't need to do that
11193 			 * and there is no send_map.
11194 			 */
11195 			tp->snd_una++;
11196 		}
11197 		/*
11198 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
11199 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
11200 		 */
11201 		tp->t_starttime = ticks;
11202 		if (tp->t_flags & TF_NEEDFIN) {
11203 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
11204 			tp->t_flags &= ~TF_NEEDFIN;
11205 			thflags &= ~TH_SYN;
11206 		} else {
11207 			tcp_state_change(tp, TCPS_ESTABLISHED);
11208 			TCP_PROBE5(connect__established, NULL, tp,
11209 			    mtod(m, const char *), tp, th);
11210 			rack_cc_conn_init(tp);
11211 		}
11212 	} else {
11213 		/*
11214 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
11215 		 * open.  If segment contains CC option and there is a
11216 		 * cached CC, apply TAO test. If it succeeds, connection is *
11217 		 * half-synchronized. Otherwise, do 3-way handshake:
11218 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
11219 		 * there was no CC option, clear cached CC value.
11220 		 */
11221 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
11222 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
11223 	}
11224 	/*
11225 	 * Advance th->th_seq to correspond to first data byte. If data,
11226 	 * trim to stay within window, dropping FIN if necessary.
11227 	 */
11228 	th->th_seq++;
11229 	if (tlen > tp->rcv_wnd) {
11230 		todrop = tlen - tp->rcv_wnd;
11231 		m_adj(m, -todrop);
11232 		tlen = tp->rcv_wnd;
11233 		thflags &= ~TH_FIN;
11234 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
11235 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
11236 	}
11237 	tp->snd_wl1 = th->th_seq - 1;
11238 	tp->rcv_up = th->th_seq;
11239 	/*
11240 	 * Client side of transaction: already sent SYN and data. If the
11241 	 * remote host used T/TCP to validate the SYN, our data will be
11242 	 * ACK'd; if so, enter normal data segment processing in the middle
11243 	 * of step 5, ack processing. Otherwise, goto step 6.
11244 	 */
11245 	if (thflags & TH_ACK) {
11246 		/* For syn-sent we need to possibly update the rtt */
11247 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11248 			uint32_t t, mcts;
11249 
11250 			mcts = tcp_ts_getticks();
11251 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11252 			if (!tp->t_rttlow || tp->t_rttlow > t)
11253 				tp->t_rttlow = t;
11254 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
11255 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11256 			tcp_rack_xmit_timer_commit(rack, tp);
11257 		}
11258 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
11259 			return (ret_val);
11260 		/* We may have changed to FIN_WAIT_1 above */
11261 		if (tp->t_state == TCPS_FIN_WAIT_1) {
11262 			/*
11263 			 * In FIN_WAIT_1 STATE in addition to the processing
11264 			 * for the ESTABLISHED state if our FIN is now
11265 			 * acknowledged then enter FIN_WAIT_2.
11266 			 */
11267 			if (ourfinisacked) {
11268 				/*
11269 				 * If we can't receive any more data, then
11270 				 * closing user can proceed. Starting the
11271 				 * timer is contrary to the specification,
11272 				 * but if we don't get a FIN we'll hang
11273 				 * forever.
11274 				 *
11275 				 * XXXjl: we should release the tp also, and
11276 				 * use a compressed state.
11277 				 */
11278 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11279 					soisdisconnected(so);
11280 					tcp_timer_activate(tp, TT_2MSL,
11281 					    (tcp_fast_finwait2_recycle ?
11282 					    tcp_finwait2_timeout :
11283 					    TP_MAXIDLE(tp)));
11284 				}
11285 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
11286 			}
11287 		}
11288 	}
11289 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11290 	   tiwin, thflags, nxt_pkt));
11291 }
11292 
11293 /*
11294  * Return value of 1, the TCB is unlocked and most
11295  * likely gone, return value of 0, the TCP is still
11296  * locked.
11297  */
11298 static int
11299 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
11300     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11301     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11302 {
11303 	struct tcp_rack *rack;
11304 	int32_t ret_val = 0;
11305 	int32_t ourfinisacked = 0;
11306 
11307 	ctf_calc_rwin(so, tp);
11308 	if ((thflags & TH_ACK) &&
11309 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
11310 	    SEQ_GT(th->th_ack, tp->snd_max))) {
11311 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11312 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11313 		return (1);
11314 	}
11315 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11316 	if (IS_FASTOPEN(tp->t_flags)) {
11317 		/*
11318 		 * When a TFO connection is in SYN_RECEIVED, the
11319 		 * only valid packets are the initial SYN, a
11320 		 * retransmit/copy of the initial SYN (possibly with
11321 		 * a subset of the original data), a valid ACK, a
11322 		 * FIN, or a RST.
11323 		 */
11324 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
11325 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11326 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11327 			return (1);
11328 		} else if (thflags & TH_SYN) {
11329 			/* non-initial SYN is ignored */
11330 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
11331 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
11332 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
11333 				ctf_do_drop(m, NULL);
11334 				return (0);
11335 			}
11336 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
11337 			ctf_do_drop(m, NULL);
11338 			return (0);
11339 		}
11340 	}
11341 
11342 	if ((thflags & TH_RST) ||
11343 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11344 		return (__ctf_process_rst(m, th, so, tp,
11345 					  &rack->r_ctl.challenge_ack_ts,
11346 					  &rack->r_ctl.challenge_ack_cnt));
11347 	/*
11348 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11349 	 * it's less than ts_recent, drop it.
11350 	 */
11351 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11352 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11353 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11354 			return (ret_val);
11355 	}
11356 	/*
11357 	 * In the SYN-RECEIVED state, validate that the packet belongs to
11358 	 * this connection before trimming the data to fit the receive
11359 	 * window.  Check the sequence number versus IRS since we know the
11360 	 * sequence numbers haven't wrapped.  This is a partial fix for the
11361 	 * "LAND" DoS attack.
11362 	 */
11363 	if (SEQ_LT(th->th_seq, tp->irs)) {
11364 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11365 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11366 		return (1);
11367 	}
11368 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11369 			      &rack->r_ctl.challenge_ack_ts,
11370 			      &rack->r_ctl.challenge_ack_cnt)) {
11371 		return (ret_val);
11372 	}
11373 	/*
11374 	 * If last ACK falls within this segment's sequence numbers, record
11375 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11376 	 * from the latest proposal of the tcplw@cray.com list (Braden
11377 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11378 	 * with our earlier PAWS tests, so this check should be solely
11379 	 * predicated on the sequence space of this segment. 3) That we
11380 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11381 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11382 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11383 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11384 	 * p.869. In such cases, we can still calculate the RTT correctly
11385 	 * when RCV.NXT == Last.ACK.Sent.
11386 	 */
11387 	if ((to->to_flags & TOF_TS) != 0 &&
11388 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11389 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11390 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11391 		tp->ts_recent_age = tcp_ts_getticks();
11392 		tp->ts_recent = to->to_tsval;
11393 	}
11394 	tp->snd_wnd = tiwin;
11395 	rack_validate_fo_sendwin_up(tp, rack);
11396 	/*
11397 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11398 	 * is on (half-synchronized state), then queue data for later
11399 	 * processing; else drop segment and return.
11400 	 */
11401 	if ((thflags & TH_ACK) == 0) {
11402 		if (IS_FASTOPEN(tp->t_flags)) {
11403 			rack_cc_conn_init(tp);
11404 		}
11405 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11406 		    tiwin, thflags, nxt_pkt));
11407 	}
11408 	KMOD_TCPSTAT_INC(tcps_connects);
11409 	if (tp->t_flags & TF_SONOTCONN) {
11410 		tp->t_flags &= ~TF_SONOTCONN;
11411 		soisconnected(so);
11412 	}
11413 	/* Do window scaling? */
11414 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11415 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11416 		tp->rcv_scale = tp->request_r_scale;
11417 	}
11418 	/*
11419 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11420 	 * FIN-WAIT-1
11421 	 */
11422 	tp->t_starttime = ticks;
11423 	if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11424 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11425 		tp->t_tfo_pending = NULL;
11426 	}
11427 	if (tp->t_flags & TF_NEEDFIN) {
11428 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
11429 		tp->t_flags &= ~TF_NEEDFIN;
11430 	} else {
11431 		tcp_state_change(tp, TCPS_ESTABLISHED);
11432 		TCP_PROBE5(accept__established, NULL, tp,
11433 		    mtod(m, const char *), tp, th);
11434 		/*
11435 		 * TFO connections call cc_conn_init() during SYN
11436 		 * processing.  Calling it again here for such connections
11437 		 * is not harmless as it would undo the snd_cwnd reduction
11438 		 * that occurs when a TFO SYN|ACK is retransmitted.
11439 		 */
11440 		if (!IS_FASTOPEN(tp->t_flags))
11441 			rack_cc_conn_init(tp);
11442 	}
11443 	/*
11444 	 * Account for the ACK of our SYN prior to
11445 	 * regular ACK processing below, except for
11446 	 * simultaneous SYN, which is handled later.
11447 	 */
11448 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11449 		tp->snd_una++;
11450 	/*
11451 	 * If segment contains data or ACK, will call tcp_reass() later; if
11452 	 * not, do so now to pass queued data to user.
11453 	 */
11454 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
11455 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11456 		    (struct mbuf *)0);
11457 		if (tp->t_flags & TF_WAKESOR) {
11458 			tp->t_flags &= ~TF_WAKESOR;
11459 			/* NB: sorwakeup_locked() does an implicit unlock. */
11460 			sorwakeup_locked(so);
11461 		}
11462 	}
11463 	tp->snd_wl1 = th->th_seq - 1;
11464 	/* For syn-recv we need to possibly update the rtt */
11465 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11466 		uint32_t t, mcts;
11467 
11468 		mcts = tcp_ts_getticks();
11469 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11470 		if (!tp->t_rttlow || tp->t_rttlow > t)
11471 			tp->t_rttlow = t;
11472 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11473 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11474 		tcp_rack_xmit_timer_commit(rack, tp);
11475 	}
11476 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11477 		return (ret_val);
11478 	}
11479 	if (tp->t_state == TCPS_FIN_WAIT_1) {
11480 		/* We could have went to FIN_WAIT_1 (or EST) above */
11481 		/*
11482 		 * In FIN_WAIT_1 STATE in addition to the processing for the
11483 		 * ESTABLISHED state if our FIN is now acknowledged then
11484 		 * enter FIN_WAIT_2.
11485 		 */
11486 		if (ourfinisacked) {
11487 			/*
11488 			 * If we can't receive any more data, then closing
11489 			 * user can proceed. Starting the timer is contrary
11490 			 * to the specification, but if we don't get a FIN
11491 			 * we'll hang forever.
11492 			 *
11493 			 * XXXjl: we should release the tp also, and use a
11494 			 * compressed state.
11495 			 */
11496 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11497 				soisdisconnected(so);
11498 				tcp_timer_activate(tp, TT_2MSL,
11499 				    (tcp_fast_finwait2_recycle ?
11500 				    tcp_finwait2_timeout :
11501 				    TP_MAXIDLE(tp)));
11502 			}
11503 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
11504 		}
11505 	}
11506 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11507 	    tiwin, thflags, nxt_pkt));
11508 }
11509 
11510 /*
11511  * Return value of 1, the TCB is unlocked and most
11512  * likely gone, return value of 0, the TCP is still
11513  * locked.
11514  */
11515 static int
11516 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11517     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11518     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11519 {
11520 	int32_t ret_val = 0;
11521 	struct tcp_rack *rack;
11522 
11523 	/*
11524 	 * Header prediction: check for the two common cases of a
11525 	 * uni-directional data xfer.  If the packet has no control flags,
11526 	 * is in-sequence, the window didn't change and we're not
11527 	 * retransmitting, it's a candidate.  If the length is zero and the
11528 	 * ack moved forward, we're the sender side of the xfer.  Just free
11529 	 * the data acked & wake any higher level process that was blocked
11530 	 * waiting for space.  If the length is non-zero and the ack didn't
11531 	 * move, we're the receiver side.  If we're getting packets in-order
11532 	 * (the reassembly queue is empty), add the data toc The socket
11533 	 * buffer and note that we need a delayed ack. Make sure that the
11534 	 * hidden state-flags are also off. Since we check for
11535 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11536 	 */
11537 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11538 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11539 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11540 	    __predict_true(SEGQ_EMPTY(tp)) &&
11541 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
11542 		if (tlen == 0) {
11543 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11544 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11545 				return (0);
11546 			}
11547 		} else {
11548 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11549 			    tiwin, nxt_pkt, iptos)) {
11550 				return (0);
11551 			}
11552 		}
11553 	}
11554 	ctf_calc_rwin(so, tp);
11555 
11556 	if ((thflags & TH_RST) ||
11557 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11558 		return (__ctf_process_rst(m, th, so, tp,
11559 					  &rack->r_ctl.challenge_ack_ts,
11560 					  &rack->r_ctl.challenge_ack_cnt));
11561 
11562 	/*
11563 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11564 	 * synchronized state.
11565 	 */
11566 	if (thflags & TH_SYN) {
11567 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11568 		return (ret_val);
11569 	}
11570 	/*
11571 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11572 	 * it's less than ts_recent, drop it.
11573 	 */
11574 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11575 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11576 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11577 			return (ret_val);
11578 	}
11579 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11580 			      &rack->r_ctl.challenge_ack_ts,
11581 			      &rack->r_ctl.challenge_ack_cnt)) {
11582 		return (ret_val);
11583 	}
11584 	/*
11585 	 * If last ACK falls within this segment's sequence numbers, record
11586 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11587 	 * from the latest proposal of the tcplw@cray.com list (Braden
11588 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11589 	 * with our earlier PAWS tests, so this check should be solely
11590 	 * predicated on the sequence space of this segment. 3) That we
11591 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11592 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11593 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11594 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11595 	 * p.869. In such cases, we can still calculate the RTT correctly
11596 	 * when RCV.NXT == Last.ACK.Sent.
11597 	 */
11598 	if ((to->to_flags & TOF_TS) != 0 &&
11599 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11600 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11601 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11602 		tp->ts_recent_age = tcp_ts_getticks();
11603 		tp->ts_recent = to->to_tsval;
11604 	}
11605 	/*
11606 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11607 	 * is on (half-synchronized state), then queue data for later
11608 	 * processing; else drop segment and return.
11609 	 */
11610 	if ((thflags & TH_ACK) == 0) {
11611 		if (tp->t_flags & TF_NEEDSYN) {
11612 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11613 			    tiwin, thflags, nxt_pkt));
11614 
11615 		} else if (tp->t_flags & TF_ACKNOW) {
11616 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11617 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11618 			return (ret_val);
11619 		} else {
11620 			ctf_do_drop(m, NULL);
11621 			return (0);
11622 		}
11623 	}
11624 	/*
11625 	 * Ack processing.
11626 	 */
11627 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11628 		return (ret_val);
11629 	}
11630 	if (sbavail(&so->so_snd)) {
11631 		if (ctf_progress_timeout_check(tp, true)) {
11632 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11633 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11634 			return (1);
11635 		}
11636 	}
11637 	/* State changes only happen in rack_process_data() */
11638 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11639 	    tiwin, thflags, nxt_pkt));
11640 }
11641 
11642 /*
11643  * Return value of 1, the TCB is unlocked and most
11644  * likely gone, return value of 0, the TCP is still
11645  * locked.
11646  */
11647 static int
11648 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11649     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11650     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11651 {
11652 	int32_t ret_val = 0;
11653 	struct tcp_rack *rack;
11654 
11655 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11656 	ctf_calc_rwin(so, tp);
11657 	if ((thflags & TH_RST) ||
11658 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11659 		return (__ctf_process_rst(m, th, so, tp,
11660 					  &rack->r_ctl.challenge_ack_ts,
11661 					  &rack->r_ctl.challenge_ack_cnt));
11662 	/*
11663 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11664 	 * synchronized state.
11665 	 */
11666 	if (thflags & TH_SYN) {
11667 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11668 		return (ret_val);
11669 	}
11670 	/*
11671 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11672 	 * it's less than ts_recent, drop it.
11673 	 */
11674 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11675 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11676 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11677 			return (ret_val);
11678 	}
11679 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11680 			      &rack->r_ctl.challenge_ack_ts,
11681 			      &rack->r_ctl.challenge_ack_cnt)) {
11682 		return (ret_val);
11683 	}
11684 	/*
11685 	 * If last ACK falls within this segment's sequence numbers, record
11686 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11687 	 * from the latest proposal of the tcplw@cray.com list (Braden
11688 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11689 	 * with our earlier PAWS tests, so this check should be solely
11690 	 * predicated on the sequence space of this segment. 3) That we
11691 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11692 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11693 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11694 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11695 	 * p.869. In such cases, we can still calculate the RTT correctly
11696 	 * when RCV.NXT == Last.ACK.Sent.
11697 	 */
11698 	if ((to->to_flags & TOF_TS) != 0 &&
11699 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11700 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11701 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11702 		tp->ts_recent_age = tcp_ts_getticks();
11703 		tp->ts_recent = to->to_tsval;
11704 	}
11705 	/*
11706 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11707 	 * is on (half-synchronized state), then queue data for later
11708 	 * processing; else drop segment and return.
11709 	 */
11710 	if ((thflags & TH_ACK) == 0) {
11711 		if (tp->t_flags & TF_NEEDSYN) {
11712 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11713 			    tiwin, thflags, nxt_pkt));
11714 
11715 		} else if (tp->t_flags & TF_ACKNOW) {
11716 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11717 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11718 			return (ret_val);
11719 		} else {
11720 			ctf_do_drop(m, NULL);
11721 			return (0);
11722 		}
11723 	}
11724 	/*
11725 	 * Ack processing.
11726 	 */
11727 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11728 		return (ret_val);
11729 	}
11730 	if (sbavail(&so->so_snd)) {
11731 		if (ctf_progress_timeout_check(tp, true)) {
11732 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11733 						tp, tick, PROGRESS_DROP, __LINE__);
11734 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11735 			return (1);
11736 		}
11737 	}
11738 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11739 	    tiwin, thflags, nxt_pkt));
11740 }
11741 
11742 static int
11743 rack_check_data_after_close(struct mbuf *m,
11744     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11745 {
11746 	struct tcp_rack *rack;
11747 
11748 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11749 	if (rack->rc_allow_data_af_clo == 0) {
11750 	close_now:
11751 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11752 		/* tcp_close will kill the inp pre-log the Reset */
11753 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11754 		tp = tcp_close(tp);
11755 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11756 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11757 		return (1);
11758 	}
11759 	if (sbavail(&so->so_snd) == 0)
11760 		goto close_now;
11761 	/* Ok we allow data that is ignored and a followup reset */
11762 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11763 	tp->rcv_nxt = th->th_seq + *tlen;
11764 	tp->t_flags2 |= TF2_DROP_AF_DATA;
11765 	rack->r_wanted_output = 1;
11766 	*tlen = 0;
11767 	return (0);
11768 }
11769 
11770 /*
11771  * Return value of 1, the TCB is unlocked and most
11772  * likely gone, return value of 0, the TCP is still
11773  * locked.
11774  */
11775 static int
11776 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11777     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11778     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11779 {
11780 	int32_t ret_val = 0;
11781 	int32_t ourfinisacked = 0;
11782 	struct tcp_rack *rack;
11783 
11784 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11785 	ctf_calc_rwin(so, tp);
11786 
11787 	if ((thflags & TH_RST) ||
11788 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11789 		return (__ctf_process_rst(m, th, so, tp,
11790 					  &rack->r_ctl.challenge_ack_ts,
11791 					  &rack->r_ctl.challenge_ack_cnt));
11792 	/*
11793 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11794 	 * synchronized state.
11795 	 */
11796 	if (thflags & TH_SYN) {
11797 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11798 		return (ret_val);
11799 	}
11800 	/*
11801 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11802 	 * it's less than ts_recent, drop it.
11803 	 */
11804 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11805 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11806 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11807 			return (ret_val);
11808 	}
11809 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11810 			      &rack->r_ctl.challenge_ack_ts,
11811 			      &rack->r_ctl.challenge_ack_cnt)) {
11812 		return (ret_val);
11813 	}
11814 	/*
11815 	 * If new data are received on a connection after the user processes
11816 	 * are gone, then RST the other end.
11817 	 */
11818 	if ((tp->t_flags & TF_CLOSED) && tlen &&
11819 	    rack_check_data_after_close(m, tp, &tlen, th, so))
11820 		return (1);
11821 	/*
11822 	 * If last ACK falls within this segment's sequence numbers, record
11823 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11824 	 * from the latest proposal of the tcplw@cray.com list (Braden
11825 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11826 	 * with our earlier PAWS tests, so this check should be solely
11827 	 * predicated on the sequence space of this segment. 3) That we
11828 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11829 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11830 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11831 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11832 	 * p.869. In such cases, we can still calculate the RTT correctly
11833 	 * when RCV.NXT == Last.ACK.Sent.
11834 	 */
11835 	if ((to->to_flags & TOF_TS) != 0 &&
11836 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11837 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11838 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11839 		tp->ts_recent_age = tcp_ts_getticks();
11840 		tp->ts_recent = to->to_tsval;
11841 	}
11842 	/*
11843 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11844 	 * is on (half-synchronized state), then queue data for later
11845 	 * processing; else drop segment and return.
11846 	 */
11847 	if ((thflags & TH_ACK) == 0) {
11848 		if (tp->t_flags & TF_NEEDSYN) {
11849 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11850 			    tiwin, thflags, nxt_pkt));
11851 		} else if (tp->t_flags & TF_ACKNOW) {
11852 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11853 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11854 			return (ret_val);
11855 		} else {
11856 			ctf_do_drop(m, NULL);
11857 			return (0);
11858 		}
11859 	}
11860 	/*
11861 	 * Ack processing.
11862 	 */
11863 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11864 		return (ret_val);
11865 	}
11866 	if (ourfinisacked) {
11867 		/*
11868 		 * If we can't receive any more data, then closing user can
11869 		 * proceed. Starting the timer is contrary to the
11870 		 * specification, but if we don't get a FIN we'll hang
11871 		 * forever.
11872 		 *
11873 		 * XXXjl: we should release the tp also, and use a
11874 		 * compressed state.
11875 		 */
11876 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11877 			soisdisconnected(so);
11878 			tcp_timer_activate(tp, TT_2MSL,
11879 			    (tcp_fast_finwait2_recycle ?
11880 			    tcp_finwait2_timeout :
11881 			    TP_MAXIDLE(tp)));
11882 		}
11883 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
11884 	}
11885 	if (sbavail(&so->so_snd)) {
11886 		if (ctf_progress_timeout_check(tp, true)) {
11887 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11888 						tp, tick, PROGRESS_DROP, __LINE__);
11889 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11890 			return (1);
11891 		}
11892 	}
11893 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11894 	    tiwin, thflags, nxt_pkt));
11895 }
11896 
11897 /*
11898  * Return value of 1, the TCB is unlocked and most
11899  * likely gone, return value of 0, the TCP is still
11900  * locked.
11901  */
11902 static int
11903 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
11904     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11905     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11906 {
11907 	int32_t ret_val = 0;
11908 	int32_t ourfinisacked = 0;
11909 	struct tcp_rack *rack;
11910 
11911 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11912 	ctf_calc_rwin(so, tp);
11913 
11914 	if ((thflags & TH_RST) ||
11915 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11916 		return (__ctf_process_rst(m, th, so, tp,
11917 					  &rack->r_ctl.challenge_ack_ts,
11918 					  &rack->r_ctl.challenge_ack_cnt));
11919 	/*
11920 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11921 	 * synchronized state.
11922 	 */
11923 	if (thflags & TH_SYN) {
11924 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11925 		return (ret_val);
11926 	}
11927 	/*
11928 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11929 	 * it's less than ts_recent, drop it.
11930 	 */
11931 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11932 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11933 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11934 			return (ret_val);
11935 	}
11936 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11937 			      &rack->r_ctl.challenge_ack_ts,
11938 			      &rack->r_ctl.challenge_ack_cnt)) {
11939 		return (ret_val);
11940 	}
11941 	/*
11942 	 * If new data are received on a connection after the user processes
11943 	 * are gone, then RST the other end.
11944 	 */
11945 	if ((tp->t_flags & TF_CLOSED) && tlen &&
11946 	    rack_check_data_after_close(m, tp, &tlen, th, so))
11947 		return (1);
11948 	/*
11949 	 * If last ACK falls within this segment's sequence numbers, record
11950 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11951 	 * from the latest proposal of the tcplw@cray.com list (Braden
11952 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11953 	 * with our earlier PAWS tests, so this check should be solely
11954 	 * predicated on the sequence space of this segment. 3) That we
11955 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11956 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11957 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11958 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11959 	 * p.869. In such cases, we can still calculate the RTT correctly
11960 	 * when RCV.NXT == Last.ACK.Sent.
11961 	 */
11962 	if ((to->to_flags & TOF_TS) != 0 &&
11963 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11964 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11965 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11966 		tp->ts_recent_age = tcp_ts_getticks();
11967 		tp->ts_recent = to->to_tsval;
11968 	}
11969 	/*
11970 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11971 	 * is on (half-synchronized state), then queue data for later
11972 	 * processing; else drop segment and return.
11973 	 */
11974 	if ((thflags & TH_ACK) == 0) {
11975 		if (tp->t_flags & TF_NEEDSYN) {
11976 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11977 			    tiwin, thflags, nxt_pkt));
11978 		} else if (tp->t_flags & TF_ACKNOW) {
11979 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11980 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11981 			return (ret_val);
11982 		} else {
11983 			ctf_do_drop(m, NULL);
11984 			return (0);
11985 		}
11986 	}
11987 	/*
11988 	 * Ack processing.
11989 	 */
11990 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11991 		return (ret_val);
11992 	}
11993 	if (ourfinisacked) {
11994 		tcp_twstart(tp);
11995 		m_freem(m);
11996 		return (1);
11997 	}
11998 	if (sbavail(&so->so_snd)) {
11999 		if (ctf_progress_timeout_check(tp, true)) {
12000 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12001 						tp, tick, PROGRESS_DROP, __LINE__);
12002 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12003 			return (1);
12004 		}
12005 	}
12006 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12007 	    tiwin, thflags, nxt_pkt));
12008 }
12009 
12010 /*
12011  * Return value of 1, the TCB is unlocked and most
12012  * likely gone, return value of 0, the TCP is still
12013  * locked.
12014  */
12015 static int
12016 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12017     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12018     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12019 {
12020 	int32_t ret_val = 0;
12021 	int32_t ourfinisacked = 0;
12022 	struct tcp_rack *rack;
12023 
12024 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12025 	ctf_calc_rwin(so, tp);
12026 
12027 	if ((thflags & TH_RST) ||
12028 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12029 		return (__ctf_process_rst(m, th, so, tp,
12030 					  &rack->r_ctl.challenge_ack_ts,
12031 					  &rack->r_ctl.challenge_ack_cnt));
12032 	/*
12033 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12034 	 * synchronized state.
12035 	 */
12036 	if (thflags & TH_SYN) {
12037 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
12038 		return (ret_val);
12039 	}
12040 	/*
12041 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12042 	 * it's less than ts_recent, drop it.
12043 	 */
12044 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12045 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12046 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12047 			return (ret_val);
12048 	}
12049 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12050 			      &rack->r_ctl.challenge_ack_ts,
12051 			      &rack->r_ctl.challenge_ack_cnt)) {
12052 		return (ret_val);
12053 	}
12054 	/*
12055 	 * If new data are received on a connection after the user processes
12056 	 * are gone, then RST the other end.
12057 	 */
12058 	if ((tp->t_flags & TF_CLOSED) && tlen &&
12059 	    rack_check_data_after_close(m, tp, &tlen, th, so))
12060 		return (1);
12061 	/*
12062 	 * If last ACK falls within this segment's sequence numbers, record
12063 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12064 	 * from the latest proposal of the tcplw@cray.com list (Braden
12065 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12066 	 * with our earlier PAWS tests, so this check should be solely
12067 	 * predicated on the sequence space of this segment. 3) That we
12068 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12069 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12070 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12071 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12072 	 * p.869. In such cases, we can still calculate the RTT correctly
12073 	 * when RCV.NXT == Last.ACK.Sent.
12074 	 */
12075 	if ((to->to_flags & TOF_TS) != 0 &&
12076 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12077 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12078 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12079 		tp->ts_recent_age = tcp_ts_getticks();
12080 		tp->ts_recent = to->to_tsval;
12081 	}
12082 	/*
12083 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12084 	 * is on (half-synchronized state), then queue data for later
12085 	 * processing; else drop segment and return.
12086 	 */
12087 	if ((thflags & TH_ACK) == 0) {
12088 		if (tp->t_flags & TF_NEEDSYN) {
12089 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12090 			    tiwin, thflags, nxt_pkt));
12091 		} else if (tp->t_flags & TF_ACKNOW) {
12092 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12093 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12094 			return (ret_val);
12095 		} else {
12096 			ctf_do_drop(m, NULL);
12097 			return (0);
12098 		}
12099 	}
12100 	/*
12101 	 * case TCPS_LAST_ACK: Ack processing.
12102 	 */
12103 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12104 		return (ret_val);
12105 	}
12106 	if (ourfinisacked) {
12107 		tp = tcp_close(tp);
12108 		ctf_do_drop(m, tp);
12109 		return (1);
12110 	}
12111 	if (sbavail(&so->so_snd)) {
12112 		if (ctf_progress_timeout_check(tp, true)) {
12113 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12114 						tp, tick, PROGRESS_DROP, __LINE__);
12115 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12116 			return (1);
12117 		}
12118 	}
12119 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12120 	    tiwin, thflags, nxt_pkt));
12121 }
12122 
12123 /*
12124  * Return value of 1, the TCB is unlocked and most
12125  * likely gone, return value of 0, the TCP is still
12126  * locked.
12127  */
12128 static int
12129 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
12130     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12131     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12132 {
12133 	int32_t ret_val = 0;
12134 	int32_t ourfinisacked = 0;
12135 	struct tcp_rack *rack;
12136 
12137 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12138 	ctf_calc_rwin(so, tp);
12139 
12140 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
12141 	if ((thflags & TH_RST) ||
12142 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12143 		return (__ctf_process_rst(m, th, so, tp,
12144 					  &rack->r_ctl.challenge_ack_ts,
12145 					  &rack->r_ctl.challenge_ack_cnt));
12146 	/*
12147 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12148 	 * synchronized state.
12149 	 */
12150 	if (thflags & TH_SYN) {
12151 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
12152 		return (ret_val);
12153 	}
12154 	/*
12155 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12156 	 * it's less than ts_recent, drop it.
12157 	 */
12158 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12159 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12160 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12161 			return (ret_val);
12162 	}
12163 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12164 			      &rack->r_ctl.challenge_ack_ts,
12165 			      &rack->r_ctl.challenge_ack_cnt)) {
12166 		return (ret_val);
12167 	}
12168 	/*
12169 	 * If new data are received on a connection after the user processes
12170 	 * are gone, then RST the other end.
12171 	 */
12172 	if ((tp->t_flags & TF_CLOSED) && tlen &&
12173 	    rack_check_data_after_close(m, tp, &tlen, th, so))
12174 		return (1);
12175 	/*
12176 	 * If last ACK falls within this segment's sequence numbers, record
12177 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12178 	 * from the latest proposal of the tcplw@cray.com list (Braden
12179 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12180 	 * with our earlier PAWS tests, so this check should be solely
12181 	 * predicated on the sequence space of this segment. 3) That we
12182 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12183 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12184 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12185 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12186 	 * p.869. In such cases, we can still calculate the RTT correctly
12187 	 * when RCV.NXT == Last.ACK.Sent.
12188 	 */
12189 	if ((to->to_flags & TOF_TS) != 0 &&
12190 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12191 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12192 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12193 		tp->ts_recent_age = tcp_ts_getticks();
12194 		tp->ts_recent = to->to_tsval;
12195 	}
12196 	/*
12197 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12198 	 * is on (half-synchronized state), then queue data for later
12199 	 * processing; else drop segment and return.
12200 	 */
12201 	if ((thflags & TH_ACK) == 0) {
12202 		if (tp->t_flags & TF_NEEDSYN) {
12203 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12204 			    tiwin, thflags, nxt_pkt));
12205 		} else if (tp->t_flags & TF_ACKNOW) {
12206 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12207 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12208 			return (ret_val);
12209 		} else {
12210 			ctf_do_drop(m, NULL);
12211 			return (0);
12212 		}
12213 	}
12214 	/*
12215 	 * Ack processing.
12216 	 */
12217 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12218 		return (ret_val);
12219 	}
12220 	if (sbavail(&so->so_snd)) {
12221 		if (ctf_progress_timeout_check(tp, true)) {
12222 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12223 						tp, tick, PROGRESS_DROP, __LINE__);
12224 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12225 			return (1);
12226 		}
12227 	}
12228 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12229 	    tiwin, thflags, nxt_pkt));
12230 }
12231 
12232 static void inline
12233 rack_clear_rate_sample(struct tcp_rack *rack)
12234 {
12235 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
12236 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
12237 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
12238 }
12239 
12240 static void
12241 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
12242 {
12243 	uint64_t bw_est, rate_wanted;
12244 	int chged = 0;
12245 	uint32_t user_max, orig_min, orig_max;
12246 
12247 	orig_min = rack->r_ctl.rc_pace_min_segs;
12248 	orig_max = rack->r_ctl.rc_pace_max_segs;
12249 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
12250 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
12251 		chged = 1;
12252 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
12253 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
12254 		if (user_max != rack->r_ctl.rc_pace_max_segs)
12255 			chged = 1;
12256 	}
12257 	if (rack->rc_force_max_seg) {
12258 		rack->r_ctl.rc_pace_max_segs = user_max;
12259 	} else if (rack->use_fixed_rate) {
12260 		bw_est = rack_get_bw(rack);
12261 		if ((rack->r_ctl.crte == NULL) ||
12262 		    (bw_est != rack->r_ctl.crte->rate)) {
12263 			rack->r_ctl.rc_pace_max_segs = user_max;
12264 		} else {
12265 			/* We are pacing right at the hardware rate */
12266 			uint32_t segsiz;
12267 
12268 			segsiz = min(ctf_fixed_maxseg(tp),
12269 				     rack->r_ctl.rc_pace_min_segs);
12270 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
12271 				                           tp, bw_est, segsiz, 0,
12272 							   rack->r_ctl.crte, NULL);
12273 		}
12274 	} else if (rack->rc_always_pace) {
12275 		if (rack->r_ctl.gp_bw ||
12276 #ifdef NETFLIX_PEAKRATE
12277 		    rack->rc_tp->t_maxpeakrate ||
12278 #endif
12279 		    rack->r_ctl.init_rate) {
12280 			/* We have a rate of some sort set */
12281 			uint32_t  orig;
12282 
12283 			bw_est = rack_get_bw(rack);
12284 			orig = rack->r_ctl.rc_pace_max_segs;
12285 			if (fill_override)
12286 				rate_wanted = *fill_override;
12287 			else
12288 				rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
12289 			if (rate_wanted) {
12290 				/* We have something */
12291 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
12292 										   rate_wanted,
12293 										   ctf_fixed_maxseg(rack->rc_tp));
12294 			} else
12295 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
12296 			if (orig != rack->r_ctl.rc_pace_max_segs)
12297 				chged = 1;
12298 		} else if ((rack->r_ctl.gp_bw == 0) &&
12299 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
12300 			/*
12301 			 * If we have nothing limit us to bursting
12302 			 * out IW sized pieces.
12303 			 */
12304 			chged = 1;
12305 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
12306 		}
12307 	}
12308 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
12309 		chged = 1;
12310 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
12311 	}
12312 	if (chged)
12313 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
12314 }
12315 
12316 
12317 static void
12318 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
12319 {
12320 #ifdef INET6
12321 	struct ip6_hdr *ip6 = NULL;
12322 #endif
12323 #ifdef INET
12324 	struct ip *ip = NULL;
12325 #endif
12326 	struct udphdr *udp = NULL;
12327 
12328 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
12329 #ifdef INET6
12330 	if (rack->r_is_v6) {
12331 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12332 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
12333 		if (tp->t_port) {
12334 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12335 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
12336 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12337 			udp->uh_dport = tp->t_port;
12338 			rack->r_ctl.fsb.udp = udp;
12339 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12340 		} else
12341 		{
12342 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
12343 			rack->r_ctl.fsb.udp = NULL;
12344 		}
12345 		tcpip_fillheaders(rack->rc_inp,
12346 				  tp->t_port,
12347 				  ip6, rack->r_ctl.fsb.th);
12348 	} else
12349 #endif				/* INET6 */
12350 	{
12351 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
12352 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
12353 		if (tp->t_port) {
12354 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12355 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
12356 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12357 			udp->uh_dport = tp->t_port;
12358 			rack->r_ctl.fsb.udp = udp;
12359 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12360 		} else
12361 		{
12362 			rack->r_ctl.fsb.udp = NULL;
12363 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
12364 		}
12365 		tcpip_fillheaders(rack->rc_inp,
12366 				  tp->t_port,
12367 				  ip, rack->r_ctl.fsb.th);
12368 	}
12369 	rack->r_fsb_inited = 1;
12370 }
12371 
12372 static int
12373 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12374 {
12375 	/*
12376 	 * Allocate the larger of spaces V6 if available else just
12377 	 * V4 and include udphdr (overbook)
12378 	 */
12379 #ifdef INET6
12380 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12381 #else
12382 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12383 #endif
12384 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12385 					    M_TCPFSB, M_NOWAIT|M_ZERO);
12386 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12387 		return (ENOMEM);
12388 	}
12389 	rack->r_fsb_inited = 0;
12390 	return (0);
12391 }
12392 
12393 static int
12394 rack_init(struct tcpcb *tp)
12395 {
12396 	struct inpcb *inp = tptoinpcb(tp);
12397 	struct tcp_rack *rack = NULL;
12398 #ifdef INVARIANTS
12399 	struct rack_sendmap *insret;
12400 #endif
12401 	uint32_t iwin, snt, us_cts;
12402 	int err;
12403 
12404 	tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12405 	if (tp->t_fb_ptr == NULL) {
12406 		/*
12407 		 * We need to allocate memory but cant. The INP and INP_INFO
12408 		 * locks and they are recursive (happens during setup. So a
12409 		 * scheme to drop the locks fails :(
12410 		 *
12411 		 */
12412 		return (ENOMEM);
12413 	}
12414 	memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12415 
12416 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12417 	RB_INIT(&rack->r_ctl.rc_mtree);
12418 	TAILQ_INIT(&rack->r_ctl.rc_free);
12419 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
12420 	rack->rc_tp = tp;
12421 	rack->rc_inp = inp;
12422 	/* Set the flag */
12423 	rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
12424 	/* Probably not needed but lets be sure */
12425 	rack_clear_rate_sample(rack);
12426 	/*
12427 	 * Save off the default values, socket options will poke
12428 	 * at these if pacing is not on or we have not yet
12429 	 * reached where pacing is on (gp_ready/fixed enabled).
12430 	 * When they get set into the CC module (when gp_ready
12431 	 * is enabled or we enable fixed) then we will set these
12432 	 * values into the CC and place in here the old values
12433 	 * so we have a restoral. Then we will set the flag
12434 	 * rc_pacing_cc_set. That way whenever we turn off pacing
12435 	 * or switch off this stack, we will know to go restore
12436 	 * the saved values.
12437 	 */
12438 	rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12439 	rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12440 	/* We want abe like behavior as well */
12441 	rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
12442 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12443 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12444 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12445 	rack->r_ctl.roundends = tp->snd_max;
12446 	if (use_rack_rr)
12447 		rack->use_rack_rr = 1;
12448 	if (V_tcp_delack_enabled)
12449 		tp->t_delayed_ack = 1;
12450 	else
12451 		tp->t_delayed_ack = 0;
12452 #ifdef TCP_ACCOUNTING
12453 	if (rack_tcp_accounting) {
12454 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12455 	}
12456 #endif
12457 	if (rack_enable_shared_cwnd)
12458 		rack->rack_enable_scwnd = 1;
12459 	rack->rc_user_set_max_segs = rack_hptsi_segments;
12460 	rack->rc_force_max_seg = 0;
12461 	if (rack_use_imac_dack)
12462 		rack->rc_dack_mode = 1;
12463 	TAILQ_INIT(&rack->r_ctl.opt_list);
12464 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12465 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12466 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12467 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12468 	rack->r_ctl.rc_highest_us_rtt = 0;
12469 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12470 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12471 	if (rack_use_cmp_acks)
12472 		rack->r_use_cmp_ack = 1;
12473 	if (rack_disable_prr)
12474 		rack->rack_no_prr = 1;
12475 	if (rack_gp_no_rec_chg)
12476 		rack->rc_gp_no_rec_chg = 1;
12477 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12478 		rack->rc_always_pace = 1;
12479 		if (rack->use_fixed_rate || rack->gp_ready)
12480 			rack_set_cc_pacing(rack);
12481 	} else
12482 		rack->rc_always_pace = 0;
12483 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12484 		rack->r_mbuf_queue = 1;
12485 	else
12486 		rack->r_mbuf_queue = 0;
12487 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12488 		inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12489 	else
12490 		inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12491 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
12492 	if (rack_limits_scwnd)
12493 		rack->r_limit_scw = 1;
12494 	else
12495 		rack->r_limit_scw = 0;
12496 	rack->rc_labc = V_tcp_abc_l_var;
12497 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12498 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12499 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12500 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12501 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12502 	rack->r_ctl.rc_min_to = rack_min_to;
12503 	microuptime(&rack->r_ctl.act_rcv_time);
12504 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12505 	rack->rc_init_win = rack_default_init_window;
12506 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12507 	if (rack_hw_up_only)
12508 		rack->r_up_only = 1;
12509 	if (rack_do_dyn_mul) {
12510 		/* When dynamic adjustment is on CA needs to start at 100% */
12511 		rack->rc_gp_dyn_mul = 1;
12512 		if (rack_do_dyn_mul >= 100)
12513 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12514 	} else
12515 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12516 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12517 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12518 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12519 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12520 				rack_probertt_filter_life);
12521 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12522 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12523 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
12524 	rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12525 	rack->r_ctl.rc_time_probertt_starts = 0;
12526 	if (rack_dsack_std_based & 0x1) {
12527 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
12528 		rack->rc_rack_tmr_std_based = 1;
12529 	}
12530 	if (rack_dsack_std_based & 0x2) {
12531 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
12532 		rack->rc_rack_use_dsack = 1;
12533 	}
12534 	/* We require at least one measurement, even if the sysctl is 0 */
12535 	if (rack_req_measurements)
12536 		rack->r_ctl.req_measurements = rack_req_measurements;
12537 	else
12538 		rack->r_ctl.req_measurements = 1;
12539 	if (rack_enable_hw_pacing)
12540 		rack->rack_hdw_pace_ena = 1;
12541 	if (rack_hw_rate_caps)
12542 		rack->r_rack_hw_rate_caps = 1;
12543 	/* Do we force on detection? */
12544 #ifdef NETFLIX_EXP_DETECTION
12545 	if (tcp_force_detection)
12546 		rack->do_detection = 1;
12547 	else
12548 #endif
12549 		rack->do_detection = 0;
12550 	if (rack_non_rxt_use_cr)
12551 		rack->rack_rec_nonrxt_use_cr = 1;
12552 	err = rack_init_fsb(tp, rack);
12553 	if (err) {
12554 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12555 		tp->t_fb_ptr = NULL;
12556 		return (err);
12557 	}
12558 	if (tp->snd_una != tp->snd_max) {
12559 		/* Create a send map for the current outstanding data */
12560 		struct rack_sendmap *rsm;
12561 
12562 		rsm = rack_alloc(rack);
12563 		if (rsm == NULL) {
12564 			uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12565 			tp->t_fb_ptr = NULL;
12566 			return (ENOMEM);
12567 		}
12568 		rsm->r_no_rtt_allowed = 1;
12569 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12570 		rsm->r_rtr_cnt = 1;
12571 		rsm->r_rtr_bytes = 0;
12572 		if (tp->t_flags & TF_SENTFIN)
12573 			rsm->r_flags |= RACK_HAS_FIN;
12574 		if ((tp->snd_una == tp->iss) &&
12575 		    !TCPS_HAVEESTABLISHED(tp->t_state))
12576 			rsm->r_flags |= RACK_HAS_SYN;
12577 		rsm->r_start = tp->snd_una;
12578 		rsm->r_end = tp->snd_max;
12579 		rsm->r_dupack = 0;
12580 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12581 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12582 			if (rsm->m)
12583 				rsm->orig_m_len = rsm->m->m_len;
12584 			else
12585 				rsm->orig_m_len = 0;
12586 		} else {
12587 			/*
12588 			 * This can happen if we have a stand-alone FIN or
12589 			 *  SYN.
12590 			 */
12591 			rsm->m = NULL;
12592 			rsm->orig_m_len = 0;
12593 			rsm->soff = 0;
12594 		}
12595 #ifndef INVARIANTS
12596 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12597 #else
12598 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12599 		if (insret != NULL) {
12600 			panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12601 			      insret, rack, rsm);
12602 		}
12603 #endif
12604 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12605 		rsm->r_in_tmap = 1;
12606 	}
12607 	/*
12608 	 * Timers in Rack are kept in microseconds so lets
12609 	 * convert any initial incoming variables
12610 	 * from ticks into usecs. Note that we
12611 	 * also change the values of t_srtt and t_rttvar, if
12612 	 * they are non-zero. They are kept with a 5
12613 	 * bit decimal so we have to carefully convert
12614 	 * these to get the full precision.
12615 	 */
12616 	rack_convert_rtts(tp);
12617 	tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12618 	if (rack_do_hystart) {
12619 		tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
12620 		if (rack_do_hystart > 1)
12621 			tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
12622 		if (rack_do_hystart > 2)
12623 			tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
12624 	}
12625 	if (rack_def_profile)
12626 		rack_set_profile(rack, rack_def_profile);
12627 	/* Cancel the GP measurement in progress */
12628 	tp->t_flags &= ~TF_GPUTINPROG;
12629 	if (SEQ_GT(tp->snd_max, tp->iss))
12630 		snt = tp->snd_max - tp->iss;
12631 	else
12632 		snt = 0;
12633 	iwin = rc_init_window(rack);
12634 	if (snt < iwin) {
12635 		/* We are not past the initial window
12636 		 * so we need to make sure cwnd is
12637 		 * correct.
12638 		 */
12639 		if (tp->snd_cwnd < iwin)
12640 			tp->snd_cwnd = iwin;
12641 		/*
12642 		 * If we are within the initial window
12643 		 * we want ssthresh to be unlimited. Setting
12644 		 * it to the rwnd (which the default stack does
12645 		 * and older racks) is not really a good idea
12646 		 * since we want to be in SS and grow both the
12647 		 * cwnd and the rwnd (via dynamic rwnd growth). If
12648 		 * we set it to the rwnd then as the peer grows its
12649 		 * rwnd we will be stuck in CA and never hit SS.
12650 		 *
12651 		 * Its far better to raise it up high (this takes the
12652 		 * risk that there as been a loss already, probably
12653 		 * we should have an indicator in all stacks of loss
12654 		 * but we don't), but considering the normal use this
12655 		 * is a risk worth taking. The consequences of not
12656 		 * hitting SS are far worse than going one more time
12657 		 * into it early on (before we have sent even a IW).
12658 		 * It is highly unlikely that we will have had a loss
12659 		 * before getting the IW out.
12660 		 */
12661 		tp->snd_ssthresh = 0xffffffff;
12662 	}
12663 	rack_stop_all_timers(tp);
12664 	/* Lets setup the fsb block */
12665 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12666 	rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12667 			     __LINE__, RACK_RTTS_INIT);
12668 	return (0);
12669 }
12670 
12671 static int
12672 rack_handoff_ok(struct tcpcb *tp)
12673 {
12674 	if ((tp->t_state == TCPS_CLOSED) ||
12675 	    (tp->t_state == TCPS_LISTEN)) {
12676 		/* Sure no problem though it may not stick */
12677 		return (0);
12678 	}
12679 	if ((tp->t_state == TCPS_SYN_SENT) ||
12680 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
12681 		/*
12682 		 * We really don't know if you support sack,
12683 		 * you have to get to ESTAB or beyond to tell.
12684 		 */
12685 		return (EAGAIN);
12686 	}
12687 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12688 		/*
12689 		 * Rack will only send a FIN after all data is acknowledged.
12690 		 * So in this case we have more data outstanding. We can't
12691 		 * switch stacks until either all data and only the FIN
12692 		 * is left (in which case rack_init() now knows how
12693 		 * to deal with that) <or> all is acknowledged and we
12694 		 * are only left with incoming data, though why you
12695 		 * would want to switch to rack after all data is acknowledged
12696 		 * I have no idea (rrs)!
12697 		 */
12698 		return (EAGAIN);
12699 	}
12700 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12701 		return (0);
12702 	}
12703 	/*
12704 	 * If we reach here we don't do SACK on this connection so we can
12705 	 * never do rack.
12706 	 */
12707 	return (EINVAL);
12708 }
12709 
12710 
12711 static void
12712 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12713 {
12714 	struct inpcb *inp = tptoinpcb(tp);
12715 
12716 	if (tp->t_fb_ptr) {
12717 		struct tcp_rack *rack;
12718 		struct rack_sendmap *rsm, *nrsm;
12719 #ifdef INVARIANTS
12720 		struct rack_sendmap *rm;
12721 #endif
12722 
12723 		rack = (struct tcp_rack *)tp->t_fb_ptr;
12724 		if (tp->t_in_pkt) {
12725 			/*
12726 			 * It is unsafe to process the packets since a
12727 			 * reset may be lurking in them (its rare but it
12728 			 * can occur). If we were to find a RST, then we
12729 			 * would end up dropping the connection and the
12730 			 * INP lock, so when we return the caller (tcp_usrreq)
12731 			 * will blow up when it trys to unlock the inp.
12732 			 */
12733 			struct mbuf *save, *m;
12734 
12735 			m = tp->t_in_pkt;
12736 			tp->t_in_pkt = NULL;
12737 			tp->t_tail_pkt = NULL;
12738 			while (m) {
12739 				save = m->m_nextpkt;
12740 				m->m_nextpkt = NULL;
12741 				m_freem(m);
12742 				m = save;
12743 			}
12744 		}
12745 		tp->t_flags &= ~TF_FORCEDATA;
12746 #ifdef NETFLIX_SHARED_CWND
12747 		if (rack->r_ctl.rc_scw) {
12748 			uint32_t limit;
12749 
12750 			if (rack->r_limit_scw)
12751 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12752 			else
12753 				limit = 0;
12754 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12755 						  rack->r_ctl.rc_scw_index,
12756 						  limit);
12757 			rack->r_ctl.rc_scw = NULL;
12758 		}
12759 #endif
12760 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
12761 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12762 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12763 			rack->r_ctl.fsb.th = NULL;
12764 		}
12765 		/* Convert back to ticks, with  */
12766 		if (tp->t_srtt > 1) {
12767 			uint32_t val, frac;
12768 
12769 			val = USEC_2_TICKS(tp->t_srtt);
12770 			frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12771 			tp->t_srtt = val << TCP_RTT_SHIFT;
12772 			/*
12773 			 * frac is the fractional part here is left
12774 			 * over from converting to hz and shifting.
12775 			 * We need to convert this to the 5 bit
12776 			 * remainder.
12777 			 */
12778 			if (frac) {
12779 				if (hz == 1000) {
12780 					frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12781 				} else {
12782 					frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12783 				}
12784 				tp->t_srtt += frac;
12785 			}
12786 		}
12787 		if (tp->t_rttvar) {
12788 			uint32_t val, frac;
12789 
12790 			val = USEC_2_TICKS(tp->t_rttvar);
12791 			frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12792 			tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12793 			/*
12794 			 * frac is the fractional part here is left
12795 			 * over from converting to hz and shifting.
12796 			 * We need to convert this to the 5 bit
12797 			 * remainder.
12798 			 */
12799 			if (frac) {
12800 				if (hz == 1000) {
12801 					frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12802 				} else {
12803 					frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12804 				}
12805 				tp->t_rttvar += frac;
12806 			}
12807 		}
12808 		tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12809 		tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12810 		if (rack->rc_always_pace) {
12811 			tcp_decrement_paced_conn();
12812 			rack_undo_cc_pacing(rack);
12813 			rack->rc_always_pace = 0;
12814 		}
12815 		/* Clean up any options if they were not applied */
12816 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12817 			struct deferred_opt_list *dol;
12818 
12819 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12820 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12821 			free(dol, M_TCPDO);
12822 		}
12823 		/* rack does not use force data but other stacks may clear it */
12824 		if (rack->r_ctl.crte != NULL) {
12825 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12826 			rack->rack_hdrw_pacing = 0;
12827 			rack->r_ctl.crte = NULL;
12828 		}
12829 #ifdef TCP_BLACKBOX
12830 		tcp_log_flowend(tp);
12831 #endif
12832 		RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12833 #ifndef INVARIANTS
12834 			(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12835 #else
12836 			rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12837 			if (rm != rsm) {
12838 				panic("At fini, rack:%p rsm:%p rm:%p",
12839 				      rack, rsm, rm);
12840 			}
12841 #endif
12842 			uma_zfree(rack_zone, rsm);
12843 		}
12844 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12845 		while (rsm) {
12846 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12847 			uma_zfree(rack_zone, rsm);
12848 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12849 		}
12850 		rack->rc_free_cnt = 0;
12851 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12852 		tp->t_fb_ptr = NULL;
12853 	}
12854 	inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12855 	inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12856 	inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12857 	inp->inp_flags2 &= ~INP_MBUF_ACKCMP;
12858 	/* Cancel the GP measurement in progress */
12859 	tp->t_flags &= ~TF_GPUTINPROG;
12860 	inp->inp_flags2 &= ~INP_MBUF_L_ACKS;
12861 	/* Make sure snd_nxt is correctly set */
12862 	tp->snd_nxt = tp->snd_max;
12863 }
12864 
12865 static void
12866 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12867 {
12868 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
12869 		rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
12870 	}
12871 	switch (tp->t_state) {
12872 	case TCPS_SYN_SENT:
12873 		rack->r_state = TCPS_SYN_SENT;
12874 		rack->r_substate = rack_do_syn_sent;
12875 		break;
12876 	case TCPS_SYN_RECEIVED:
12877 		rack->r_state = TCPS_SYN_RECEIVED;
12878 		rack->r_substate = rack_do_syn_recv;
12879 		break;
12880 	case TCPS_ESTABLISHED:
12881 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12882 		rack->r_state = TCPS_ESTABLISHED;
12883 		rack->r_substate = rack_do_established;
12884 		break;
12885 	case TCPS_CLOSE_WAIT:
12886 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12887 		rack->r_state = TCPS_CLOSE_WAIT;
12888 		rack->r_substate = rack_do_close_wait;
12889 		break;
12890 	case TCPS_FIN_WAIT_1:
12891 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12892 		rack->r_state = TCPS_FIN_WAIT_1;
12893 		rack->r_substate = rack_do_fin_wait_1;
12894 		break;
12895 	case TCPS_CLOSING:
12896 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12897 		rack->r_state = TCPS_CLOSING;
12898 		rack->r_substate = rack_do_closing;
12899 		break;
12900 	case TCPS_LAST_ACK:
12901 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12902 		rack->r_state = TCPS_LAST_ACK;
12903 		rack->r_substate = rack_do_lastack;
12904 		break;
12905 	case TCPS_FIN_WAIT_2:
12906 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12907 		rack->r_state = TCPS_FIN_WAIT_2;
12908 		rack->r_substate = rack_do_fin_wait_2;
12909 		break;
12910 	case TCPS_LISTEN:
12911 	case TCPS_CLOSED:
12912 	case TCPS_TIME_WAIT:
12913 	default:
12914 		break;
12915 	};
12916 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
12917 		rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
12918 
12919 }
12920 
12921 static void
12922 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
12923 {
12924 	/*
12925 	 * We received an ack, and then did not
12926 	 * call send or were bounced out due to the
12927 	 * hpts was running. Now a timer is up as well, is
12928 	 * it the right timer?
12929 	 */
12930 	struct rack_sendmap *rsm;
12931 	int tmr_up;
12932 
12933 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
12934 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
12935 		return;
12936 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
12937 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
12938 	    (tmr_up == PACE_TMR_RXT)) {
12939 		/* Should be an RXT */
12940 		return;
12941 	}
12942 	if (rsm == NULL) {
12943 		/* Nothing outstanding? */
12944 		if (tp->t_flags & TF_DELACK) {
12945 			if (tmr_up == PACE_TMR_DELACK)
12946 				/* We are supposed to have delayed ack up and we do */
12947 				return;
12948 		} else if (sbavail(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) {
12949 			/*
12950 			 * if we hit enobufs then we would expect the possibility
12951 			 * of nothing outstanding and the RXT up (and the hptsi timer).
12952 			 */
12953 			return;
12954 		} else if (((V_tcp_always_keepalive ||
12955 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
12956 			    (tp->t_state <= TCPS_CLOSING)) &&
12957 			   (tmr_up == PACE_TMR_KEEP) &&
12958 			   (tp->snd_max == tp->snd_una)) {
12959 			/* We should have keep alive up and we do */
12960 			return;
12961 		}
12962 	}
12963 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
12964 		   ((tmr_up == PACE_TMR_TLP) ||
12965 		    (tmr_up == PACE_TMR_RACK) ||
12966 		    (tmr_up == PACE_TMR_RXT))) {
12967 		/*
12968 		 * Either a Rack, TLP or RXT is fine if  we
12969 		 * have outstanding data.
12970 		 */
12971 		return;
12972 	} else if (tmr_up == PACE_TMR_DELACK) {
12973 		/*
12974 		 * If the delayed ack was going to go off
12975 		 * before the rtx/tlp/rack timer were going to
12976 		 * expire, then that would be the timer in control.
12977 		 * Note we don't check the time here trusting the
12978 		 * code is correct.
12979 		 */
12980 		return;
12981 	}
12982 	/*
12983 	 * Ok the timer originally started is not what we want now.
12984 	 * We will force the hpts to be stopped if any, and restart
12985 	 * with the slot set to what was in the saved slot.
12986 	 */
12987 	if (tcp_in_hpts(rack->rc_inp)) {
12988 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
12989 			uint32_t us_cts;
12990 
12991 			us_cts = tcp_get_usecs(NULL);
12992 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
12993 				rack->r_early = 1;
12994 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
12995 			}
12996 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
12997 		}
12998 		tcp_hpts_remove(rack->rc_inp);
12999 	}
13000 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13001 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13002 }
13003 
13004 
13005 static void
13006 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts, uint32_t high_seq)
13007 {
13008 	if ((SEQ_LT(tp->snd_wl1, seq) ||
13009 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
13010 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
13011 		/* keep track of pure window updates */
13012 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
13013 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
13014 		tp->snd_wnd = tiwin;
13015 		rack_validate_fo_sendwin_up(tp, rack);
13016 		tp->snd_wl1 = seq;
13017 		tp->snd_wl2 = ack;
13018 		if (tp->snd_wnd > tp->max_sndwnd)
13019 			tp->max_sndwnd = tp->snd_wnd;
13020 	    rack->r_wanted_output = 1;
13021 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
13022 		tp->snd_wnd = tiwin;
13023 		rack_validate_fo_sendwin_up(tp, rack);
13024 		tp->snd_wl1 = seq;
13025 		tp->snd_wl2 = ack;
13026 	} else {
13027 		/* Not a valid win update */
13028 		return;
13029 	}
13030 	/* Do we exit persists? */
13031 	if ((rack->rc_in_persist != 0) &&
13032 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13033 				rack->r_ctl.rc_pace_min_segs))) {
13034 		rack_exit_persist(tp, rack, cts);
13035 	}
13036 	/* Do we enter persists? */
13037 	if ((rack->rc_in_persist == 0) &&
13038 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13039 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
13040 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
13041 	    sbavail(&tptosocket(tp)->so_snd) &&
13042 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
13043 		/*
13044 		 * Here the rwnd is less than
13045 		 * the pacing size, we are established,
13046 		 * nothing is outstanding, and there is
13047 		 * data to send. Enter persists.
13048 		 */
13049 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
13050 	}
13051 }
13052 
13053 static void
13054 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
13055 {
13056 
13057 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13058 		struct inpcb *inp = tptoinpcb(tp);
13059 		union tcp_log_stackspecific log;
13060 		struct timeval ltv;
13061 		char tcp_hdr_buf[60];
13062 		struct tcphdr *th;
13063 		struct timespec ts;
13064 		uint32_t orig_snd_una;
13065 		uint8_t xx = 0;
13066 
13067 #ifdef NETFLIX_HTTP_LOGGING
13068 		struct http_sendfile_track *http_req;
13069 
13070 		if (SEQ_GT(ae->ack, tp->snd_una)) {
13071 			http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
13072 		} else {
13073 			http_req = tcp_http_find_req_for_seq(tp, ae->ack);
13074 		}
13075 #endif
13076 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13077 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
13078 		if (rack->rack_no_prr == 0)
13079 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13080 		else
13081 			log.u_bbr.flex1 = 0;
13082 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13083 		log.u_bbr.use_lt_bw <<= 1;
13084 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
13085 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13086 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
13087 		log.u_bbr.pkts_out = tp->t_maxseg;
13088 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13089 		log.u_bbr.flex7 = 1;
13090 		log.u_bbr.lost = ae->flags;
13091 		log.u_bbr.cwnd_gain = ackval;
13092 		log.u_bbr.pacing_gain = 0x2;
13093 		if (ae->flags & TSTMP_HDWR) {
13094 			/* Record the hardware timestamp if present */
13095 			log.u_bbr.flex3 = M_TSTMP;
13096 			ts.tv_sec = ae->timestamp / 1000000000;
13097 			ts.tv_nsec = ae->timestamp % 1000000000;
13098 			ltv.tv_sec = ts.tv_sec;
13099 			ltv.tv_usec = ts.tv_nsec / 1000;
13100 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13101 		} else if (ae->flags & TSTMP_LRO) {
13102 			/* Record the LRO the arrival timestamp */
13103 			log.u_bbr.flex3 = M_TSTMP_LRO;
13104 			ts.tv_sec = ae->timestamp / 1000000000;
13105 			ts.tv_nsec = ae->timestamp % 1000000000;
13106 			ltv.tv_sec = ts.tv_sec;
13107 			ltv.tv_usec = ts.tv_nsec / 1000;
13108 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13109 		}
13110 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13111 		/* Log the rcv time */
13112 		log.u_bbr.delRate = ae->timestamp;
13113 #ifdef NETFLIX_HTTP_LOGGING
13114 		log.u_bbr.applimited = tp->t_http_closed;
13115 		log.u_bbr.applimited <<= 8;
13116 		log.u_bbr.applimited |= tp->t_http_open;
13117 		log.u_bbr.applimited <<= 8;
13118 		log.u_bbr.applimited |= tp->t_http_req;
13119 		if (http_req) {
13120 			/* Copy out any client req info */
13121 			/* seconds */
13122 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13123 			/* useconds */
13124 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13125 			log.u_bbr.rttProp = http_req->timestamp;
13126 			log.u_bbr.cur_del_rate = http_req->start;
13127 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13128 				log.u_bbr.flex8 |= 1;
13129 			} else {
13130 				log.u_bbr.flex8 |= 2;
13131 				log.u_bbr.bw_inuse = http_req->end;
13132 			}
13133 			log.u_bbr.flex6 = http_req->start_seq;
13134 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13135 				log.u_bbr.flex8 |= 4;
13136 				log.u_bbr.epoch = http_req->end_seq;
13137 			}
13138 		}
13139 #endif
13140 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
13141 		th = (struct tcphdr *)tcp_hdr_buf;
13142 		th->th_seq = ae->seq;
13143 		th->th_ack = ae->ack;
13144 		th->th_win = ae->win;
13145 		/* Now fill in the ports */
13146 		th->th_sport = inp->inp_fport;
13147 		th->th_dport = inp->inp_lport;
13148 		tcp_set_flags(th, ae->flags);
13149 		/* Now do we have a timestamp option? */
13150 		if (ae->flags & HAS_TSTMP) {
13151 			u_char *cp;
13152 			uint32_t val;
13153 
13154 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
13155 			cp = (u_char *)(th + 1);
13156 			*cp = TCPOPT_NOP;
13157 			cp++;
13158 			*cp = TCPOPT_NOP;
13159 			cp++;
13160 			*cp = TCPOPT_TIMESTAMP;
13161 			cp++;
13162 			*cp = TCPOLEN_TIMESTAMP;
13163 			cp++;
13164 			val = htonl(ae->ts_value);
13165 			bcopy((char *)&val,
13166 			      (char *)cp, sizeof(uint32_t));
13167 			val = htonl(ae->ts_echo);
13168 			bcopy((char *)&val,
13169 			      (char *)(cp + 4), sizeof(uint32_t));
13170 		} else
13171 			th->th_off = (sizeof(struct tcphdr) >> 2);
13172 
13173 		/*
13174 		 * For sane logging we need to play a little trick.
13175 		 * If the ack were fully processed we would have moved
13176 		 * snd_una to high_seq, but since compressed acks are
13177 		 * processed in two phases, at this point (logging) snd_una
13178 		 * won't be advanced. So we would see multiple acks showing
13179 		 * the advancement. We can prevent that by "pretending" that
13180 		 * snd_una was advanced and then un-advancing it so that the
13181 		 * logging code has the right value for tlb_snd_una.
13182 		 */
13183 		if (tp->snd_una != high_seq) {
13184 			orig_snd_una = tp->snd_una;
13185 			tp->snd_una = high_seq;
13186 			xx = 1;
13187 		} else
13188 			xx = 0;
13189 		TCP_LOG_EVENTP(tp, th,
13190 			       &tptosocket(tp)->so_rcv,
13191 			       &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
13192 			       0, &log, true, &ltv);
13193 		if (xx) {
13194 			tp->snd_una = orig_snd_una;
13195 		}
13196 	}
13197 
13198 }
13199 
13200 static void
13201 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
13202 {
13203 	uint32_t us_rtt;
13204 	/*
13205 	 * A persist or keep-alive was forced out, update our
13206 	 * min rtt time. Note now worry about lost responses.
13207 	 * When a subsequent keep-alive or persist times out
13208 	 * and forced_ack is still on, then the last probe
13209 	 * was not responded to. In such cases we have a
13210 	 * sysctl that controls the behavior. Either we apply
13211 	 * the rtt but with reduced confidence (0). Or we just
13212 	 * plain don't apply the rtt estimate. Having data flow
13213 	 * will clear the probe_not_answered flag i.e. cum-ack
13214 	 * move forward <or> exiting and reentering persists.
13215 	 */
13216 
13217 	rack->forced_ack = 0;
13218 	rack->rc_tp->t_rxtshift = 0;
13219 	if ((rack->rc_in_persist &&
13220 	     (tiwin == rack->rc_tp->snd_wnd)) ||
13221 	    (rack->rc_in_persist == 0)) {
13222 		/*
13223 		 * In persists only apply the RTT update if this is
13224 		 * a response to our window probe. And that
13225 		 * means the rwnd sent must match the current
13226 		 * snd_wnd. If it does not, then we got a
13227 		 * window update ack instead. For keepalive
13228 		 * we allow the answer no matter what the window.
13229 		 *
13230 		 * Note that if the probe_not_answered is set then
13231 		 * the forced_ack_ts is the oldest one i.e. the first
13232 		 * probe sent that might have been lost. This assures
13233 		 * us that if we do calculate an RTT it is longer not
13234 		 * some short thing.
13235 		 */
13236 		if (rack->rc_in_persist)
13237 			counter_u64_add(rack_persists_acks, 1);
13238 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13239 		if (us_rtt == 0)
13240 			us_rtt = 1;
13241 		if (rack->probe_not_answered == 0) {
13242 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13243 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
13244 		} else {
13245 			/* We have a retransmitted probe here too */
13246 			if (rack_apply_rtt_with_reduced_conf) {
13247 				rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13248 				tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
13249 			}
13250 		}
13251 	}
13252 }
13253 
13254 static int
13255 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
13256 {
13257 	/*
13258 	 * Handle a "special" compressed ack mbuf. Each incoming
13259 	 * ack has only four possible dispositions:
13260 	 *
13261 	 * A) It moves the cum-ack forward
13262 	 * B) It is behind the cum-ack.
13263 	 * C) It is a window-update ack.
13264 	 * D) It is a dup-ack.
13265 	 *
13266 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
13267 	 * in the incoming mbuf. We also need to still pay attention
13268 	 * to nxt_pkt since there may be another packet after this
13269 	 * one.
13270 	 */
13271 #ifdef TCP_ACCOUNTING
13272 	uint64_t ts_val;
13273 	uint64_t rdstc;
13274 #endif
13275 	int segsiz;
13276 	struct timespec ts;
13277 	struct tcp_rack *rack;
13278 	struct tcp_ackent *ae;
13279 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
13280 	int cnt, i, did_out, ourfinisacked = 0;
13281 	struct tcpopt to_holder, *to = NULL;
13282 #ifdef TCP_ACCOUNTING
13283 	int win_up_req = 0;
13284 #endif
13285 	int nsegs = 0;
13286 	int under_pacing = 1;
13287 	int recovery = 0;
13288 #ifdef TCP_ACCOUNTING
13289 	sched_pin();
13290 #endif
13291 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13292 	if (rack->gp_ready &&
13293 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
13294 		under_pacing = 0;
13295 	else
13296 		under_pacing = 1;
13297 
13298 	if (rack->r_state != tp->t_state)
13299 		rack_set_state(tp, rack);
13300 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13301 	    (tp->t_flags & TF_GPUTINPROG)) {
13302 		/*
13303 		 * We have a goodput in progress
13304 		 * and we have entered a late state.
13305 		 * Do we have enough data in the sb
13306 		 * to handle the GPUT request?
13307 		 */
13308 		uint32_t bytes;
13309 
13310 		bytes = tp->gput_ack - tp->gput_seq;
13311 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
13312 			bytes += tp->gput_seq - tp->snd_una;
13313 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
13314 			/*
13315 			 * There are not enough bytes in the socket
13316 			 * buffer that have been sent to cover this
13317 			 * measurement. Cancel it.
13318 			 */
13319 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
13320 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
13321 						   tp->gput_seq,
13322 						   0, 0, 18, __LINE__, NULL, 0);
13323 			tp->t_flags &= ~TF_GPUTINPROG;
13324 		}
13325 	}
13326 	to = &to_holder;
13327 	to->to_flags = 0;
13328 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
13329 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
13330 	cnt = m->m_len / sizeof(struct tcp_ackent);
13331 	counter_u64_add(rack_multi_single_eq, cnt);
13332 	high_seq = tp->snd_una;
13333 	the_win = tp->snd_wnd;
13334 	win_seq = tp->snd_wl1;
13335 	win_upd_ack = tp->snd_wl2;
13336 	cts = tcp_tv_to_usectick(tv);
13337 	ms_cts = tcp_tv_to_mssectick(tv);
13338 	rack->r_ctl.rc_rcvtime = cts;
13339 	segsiz = ctf_fixed_maxseg(tp);
13340 	if ((rack->rc_gp_dyn_mul) &&
13341 	    (rack->use_fixed_rate == 0) &&
13342 	    (rack->rc_always_pace)) {
13343 		/* Check in on probertt */
13344 		rack_check_probe_rtt(rack, cts);
13345 	}
13346 	for (i = 0; i < cnt; i++) {
13347 #ifdef TCP_ACCOUNTING
13348 		ts_val = get_cyclecount();
13349 #endif
13350 		rack_clear_rate_sample(rack);
13351 		ae = ((mtod(m, struct tcp_ackent *)) + i);
13352 		/* Setup the window */
13353 		tiwin = ae->win << tp->snd_scale;
13354 		if (tiwin > rack->r_ctl.rc_high_rwnd)
13355 			rack->r_ctl.rc_high_rwnd = tiwin;
13356 		/* figure out the type of ack */
13357 		if (SEQ_LT(ae->ack, high_seq)) {
13358 			/* Case B*/
13359 			ae->ack_val_set = ACK_BEHIND;
13360 		} else if (SEQ_GT(ae->ack, high_seq)) {
13361 			/* Case A */
13362 			ae->ack_val_set = ACK_CUMACK;
13363 		} else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
13364 			/* Case D */
13365 			ae->ack_val_set = ACK_DUPACK;
13366 		} else {
13367 			/* Case C */
13368 			ae->ack_val_set = ACK_RWND;
13369 		}
13370 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
13371 		/* Validate timestamp */
13372 		if (ae->flags & HAS_TSTMP) {
13373 			/* Setup for a timestamp */
13374 			to->to_flags = TOF_TS;
13375 			ae->ts_echo -= tp->ts_offset;
13376 			to->to_tsecr = ae->ts_echo;
13377 			to->to_tsval = ae->ts_value;
13378 			/*
13379 			 * If echoed timestamp is later than the current time, fall back to
13380 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13381 			 * were used when this connection was established.
13382 			 */
13383 			if (TSTMP_GT(ae->ts_echo, ms_cts))
13384 				to->to_tsecr = 0;
13385 			if (tp->ts_recent &&
13386 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
13387 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
13388 #ifdef TCP_ACCOUNTING
13389 					rdstc = get_cyclecount();
13390 					if (rdstc > ts_val) {
13391 						counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
13392 								(rdstc - ts_val));
13393 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13394 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13395 						}
13396 					}
13397 #endif
13398 					continue;
13399 				}
13400 			}
13401 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
13402 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
13403 				tp->ts_recent_age = tcp_ts_getticks();
13404 				tp->ts_recent = ae->ts_value;
13405 			}
13406 		} else {
13407 			/* Setup for a no options */
13408 			to->to_flags = 0;
13409 		}
13410 		/* Update the rcv time and perform idle reduction possibly */
13411 		if  (tp->t_idle_reduce &&
13412 		     (tp->snd_max == tp->snd_una) &&
13413 		     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13414 			counter_u64_add(rack_input_idle_reduces, 1);
13415 			rack_cc_after_idle(rack, tp);
13416 		}
13417 		tp->t_rcvtime = ticks;
13418 		/* Now what about ECN of a chain of pure ACKs? */
13419 		if (tcp_ecn_input_segment(tp, ae->flags, 0,
13420 			tcp_packets_this_ack(tp, ae->ack),
13421 			ae->codepoint))
13422 			rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
13423 #ifdef TCP_ACCOUNTING
13424 		/* Count for the specific type of ack in */
13425 		counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
13426 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13427 			tp->tcp_cnt_counters[ae->ack_val_set]++;
13428 		}
13429 #endif
13430 		/*
13431 		 * Note how we could move up these in the determination
13432 		 * above, but we don't so that way the timestamp checks (and ECN)
13433 		 * is done first before we do any processing on the ACK.
13434 		 * The non-compressed path through the code has this
13435 		 * weakness (noted by @jtl) that it actually does some
13436 		 * processing before verifying the timestamp information.
13437 		 * We don't take that path here which is why we set
13438 		 * the ack_val_set first, do the timestamp and ecn
13439 		 * processing, and then look at what we have setup.
13440 		 */
13441 		if (ae->ack_val_set == ACK_BEHIND) {
13442 			/*
13443 			 * Case B flag reordering, if window is not closed
13444 			 * or it could be a keep-alive or persists
13445 			 */
13446 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
13447 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13448 			}
13449 		} else if (ae->ack_val_set == ACK_DUPACK) {
13450 			/* Case D */
13451 			rack_strike_dupack(rack);
13452 		} else if (ae->ack_val_set == ACK_RWND) {
13453 			/* Case C */
13454 			if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13455 				ts.tv_sec = ae->timestamp / 1000000000;
13456 				ts.tv_nsec = ae->timestamp % 1000000000;
13457 				rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13458 				rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13459 			} else {
13460 				rack->r_ctl.act_rcv_time = *tv;
13461 			}
13462 			if (rack->forced_ack) {
13463 				rack_handle_probe_response(rack, tiwin,
13464 							   tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
13465 			}
13466 #ifdef TCP_ACCOUNTING
13467 			win_up_req = 1;
13468 #endif
13469 			win_upd_ack = ae->ack;
13470 			win_seq = ae->seq;
13471 			the_win = tiwin;
13472 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13473 		} else {
13474 			/* Case A */
13475 			if (SEQ_GT(ae->ack, tp->snd_max)) {
13476 				/*
13477 				 * We just send an ack since the incoming
13478 				 * ack is beyond the largest seq we sent.
13479 				 */
13480 				if ((tp->t_flags & TF_ACKNOW) == 0) {
13481 					ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13482 					if (tp->t_flags && TF_ACKNOW)
13483 						rack->r_wanted_output = 1;
13484 				}
13485 			} else {
13486 				nsegs++;
13487 				/* If the window changed setup to update */
13488 				if (tiwin != tp->snd_wnd) {
13489 					win_upd_ack = ae->ack;
13490 					win_seq = ae->seq;
13491 					the_win = tiwin;
13492 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13493 				}
13494 #ifdef TCP_ACCOUNTING
13495 				/* Account for the acks */
13496 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13497 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13498 				}
13499 				counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13500 						(((ae->ack - high_seq) + segsiz - 1) / segsiz));
13501 #endif
13502 				high_seq = ae->ack;
13503 				if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
13504 					union tcp_log_stackspecific log;
13505 					struct timeval tv;
13506 
13507 					memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13508 					log.u_bbr.timeStamp = tcp_get_usecs(&tv);
13509 					log.u_bbr.flex1 = high_seq;
13510 					log.u_bbr.flex2 = rack->r_ctl.roundends;
13511 					log.u_bbr.flex3 = rack->r_ctl.current_round;
13512 					log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
13513 					log.u_bbr.flex8 = 8;
13514 					tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
13515 						       0, &log, false, NULL, NULL, 0, &tv);
13516 				}
13517 				/*
13518 				 * The draft (v3) calls for us to use SEQ_GEQ, but that
13519 				 * causes issues when we are just going app limited. Lets
13520 				 * instead use SEQ_GT <or> where its equal but more data
13521 				 * is outstanding.
13522 				 */
13523 				if ((SEQ_GT(high_seq, rack->r_ctl.roundends)) ||
13524 				    ((high_seq == rack->r_ctl.roundends) &&
13525 				     SEQ_GT(tp->snd_max, tp->snd_una))) {
13526 					rack->r_ctl.current_round++;
13527 					rack->r_ctl.roundends = tp->snd_max;
13528 					if (CC_ALGO(tp)->newround != NULL) {
13529 						CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
13530 					}
13531 				}
13532 				/* Setup our act_rcv_time */
13533 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13534 					ts.tv_sec = ae->timestamp / 1000000000;
13535 					ts.tv_nsec = ae->timestamp % 1000000000;
13536 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13537 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13538 				} else {
13539 					rack->r_ctl.act_rcv_time = *tv;
13540 				}
13541 				rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13542 				if (rack->rc_dsack_round_seen) {
13543 					/* Is the dsack round over? */
13544 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
13545 						/* Yes it is */
13546 						rack->rc_dsack_round_seen = 0;
13547 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
13548 					}
13549 				}
13550 			}
13551 		}
13552 		/* And lets be sure to commit the rtt measurements for this ack */
13553 		tcp_rack_xmit_timer_commit(rack, tp);
13554 #ifdef TCP_ACCOUNTING
13555 		rdstc = get_cyclecount();
13556 		if (rdstc > ts_val) {
13557 			counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13558 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13559 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13560 				if (ae->ack_val_set == ACK_CUMACK)
13561 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13562 			}
13563 		}
13564 #endif
13565 	}
13566 #ifdef TCP_ACCOUNTING
13567 	ts_val = get_cyclecount();
13568 #endif
13569 	/* Tend to any collapsed window */
13570 	if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
13571 		/* The peer collapsed the window */
13572 		rack_collapsed_window(rack, (tp->snd_max - high_seq), __LINE__);
13573 	} else if (rack->rc_has_collapsed)
13574 		rack_un_collapse_window(rack, __LINE__);
13575 	if ((rack->r_collapse_point_valid) &&
13576 	    (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
13577 		rack->r_collapse_point_valid = 0;
13578 	acked_amount = acked = (high_seq - tp->snd_una);
13579 	if (acked) {
13580 		/*
13581 		 * Clear the probe not answered flag
13582 		 * since cum-ack moved forward.
13583 		 */
13584 		rack->probe_not_answered = 0;
13585 		if (rack->sack_attack_disable == 0)
13586 			rack_do_decay(rack);
13587 		if (acked >= segsiz) {
13588 			/*
13589 			 * You only get credit for
13590 			 * MSS and greater (and you get extra
13591 			 * credit for larger cum-ack moves).
13592 			 */
13593 			int ac;
13594 
13595 			ac = acked / segsiz;
13596 			rack->r_ctl.ack_count += ac;
13597 			counter_u64_add(rack_ack_total, ac);
13598 		}
13599 		if (rack->r_ctl.ack_count > 0xfff00000) {
13600 			/*
13601 			 * reduce the number to keep us under
13602 			 * a uint32_t.
13603 			 */
13604 			rack->r_ctl.ack_count /= 2;
13605 			rack->r_ctl.sack_count /= 2;
13606 		}
13607 		if (tp->t_flags & TF_NEEDSYN) {
13608 			/*
13609 			 * T/TCP: Connection was half-synchronized, and our SYN has
13610 			 * been ACK'd (so connection is now fully synchronized).  Go
13611 			 * to non-starred state, increment snd_una for ACK of SYN,
13612 			 * and check if we can do window scaling.
13613 			 */
13614 			tp->t_flags &= ~TF_NEEDSYN;
13615 			tp->snd_una++;
13616 			acked_amount = acked = (high_seq - tp->snd_una);
13617 		}
13618 		if (acked > sbavail(&so->so_snd))
13619 			acked_amount = sbavail(&so->so_snd);
13620 #ifdef NETFLIX_EXP_DETECTION
13621 		/*
13622 		 * We only care on a cum-ack move if we are in a sack-disabled
13623 		 * state. We have already added in to the ack_count, and we never
13624 		 * would disable on a cum-ack move, so we only care to do the
13625 		 * detection if it may "undo" it, i.e. we were in disabled already.
13626 		 */
13627 		if (rack->sack_attack_disable)
13628 			rack_do_detection(tp, rack, acked_amount, segsiz);
13629 #endif
13630 		if (IN_FASTRECOVERY(tp->t_flags) &&
13631 		    (rack->rack_no_prr == 0))
13632 			rack_update_prr(tp, rack, acked_amount, high_seq);
13633 		if (IN_RECOVERY(tp->t_flags)) {
13634 			if (SEQ_LT(high_seq, tp->snd_recover) &&
13635 			    (SEQ_LT(high_seq, tp->snd_max))) {
13636 				tcp_rack_partialack(tp);
13637 			} else {
13638 				rack_post_recovery(tp, high_seq);
13639 				recovery = 1;
13640 			}
13641 		}
13642 		/* Handle the rack-log-ack part (sendmap) */
13643 		if ((sbused(&so->so_snd) == 0) &&
13644 		    (acked > acked_amount) &&
13645 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
13646 		    (tp->t_flags & TF_SENTFIN)) {
13647 			/*
13648 			 * We must be sure our fin
13649 			 * was sent and acked (we can be
13650 			 * in FIN_WAIT_1 without having
13651 			 * sent the fin).
13652 			 */
13653 			ourfinisacked = 1;
13654 			/*
13655 			 * Lets make sure snd_una is updated
13656 			 * since most likely acked_amount = 0 (it
13657 			 * should be).
13658 			 */
13659 			tp->snd_una = high_seq;
13660 		}
13661 		/* Did we make a RTO error? */
13662 		if ((tp->t_flags & TF_PREVVALID) &&
13663 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13664 			tp->t_flags &= ~TF_PREVVALID;
13665 			if (tp->t_rxtshift == 1 &&
13666 			    (int)(ticks - tp->t_badrxtwin) < 0)
13667 				rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
13668 		}
13669 		/* Handle the data in the socket buffer */
13670 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13671 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13672 		if (acked_amount > 0) {
13673 			struct mbuf *mfree;
13674 
13675 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13676 			SOCKBUF_LOCK(&so->so_snd);
13677 			mfree = sbcut_locked(&so->so_snd, acked_amount);
13678 			tp->snd_una = high_seq;
13679 			/* Note we want to hold the sb lock through the sendmap adjust */
13680 			rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13681 			/* Wake up the socket if we have room to write more */
13682 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13683 			sowwakeup_locked(so);
13684 			m_freem(mfree);
13685 		}
13686 		/* update progress */
13687 		tp->t_acktime = ticks;
13688 		rack_log_progress_event(rack, tp, tp->t_acktime,
13689 					PROGRESS_UPDATE, __LINE__);
13690 		/* Clear out shifts and such */
13691 		tp->t_rxtshift = 0;
13692 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13693 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13694 		rack->rc_tlp_in_progress = 0;
13695 		rack->r_ctl.rc_tlp_cnt_out = 0;
13696 		/* Send recover and snd_nxt must be dragged along */
13697 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
13698 			tp->snd_recover = tp->snd_una;
13699 		if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13700 			tp->snd_nxt = tp->snd_una;
13701 		/*
13702 		 * If the RXT timer is running we want to
13703 		 * stop it, so we can restart a TLP (or new RXT).
13704 		 */
13705 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13706 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13707 #ifdef NETFLIX_HTTP_LOGGING
13708 		tcp_http_check_for_comp(rack->rc_tp, high_seq);
13709 #endif
13710 		tp->snd_wl2 = high_seq;
13711 		tp->t_dupacks = 0;
13712 		if (under_pacing &&
13713 		    (rack->use_fixed_rate == 0) &&
13714 		    (rack->in_probe_rtt == 0) &&
13715 		    rack->rc_gp_dyn_mul &&
13716 		    rack->rc_always_pace) {
13717 			/* Check if we are dragging bottom */
13718 			rack_check_bottom_drag(tp, rack, so, acked);
13719 		}
13720 		if (tp->snd_una == tp->snd_max) {
13721 			tp->t_flags &= ~TF_PREVVALID;
13722 			rack->r_ctl.retran_during_recovery = 0;
13723 			rack->r_ctl.dsack_byte_cnt = 0;
13724 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13725 			if (rack->r_ctl.rc_went_idle_time == 0)
13726 				rack->r_ctl.rc_went_idle_time = 1;
13727 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13728 			if (sbavail(&tptosocket(tp)->so_snd) == 0)
13729 				tp->t_acktime = 0;
13730 			/* Set so we might enter persists... */
13731 			rack->r_wanted_output = 1;
13732 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13733 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13734 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13735 			    (sbavail(&so->so_snd) == 0) &&
13736 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13737 				/*
13738 				 * The socket was gone and the
13739 				 * peer sent data (not now in the past), time to
13740 				 * reset him.
13741 				 */
13742 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13743 				/* tcp_close will kill the inp pre-log the Reset */
13744 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13745 #ifdef TCP_ACCOUNTING
13746 				rdstc = get_cyclecount();
13747 				if (rdstc > ts_val) {
13748 					counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13749 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13750 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13751 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13752 					}
13753 				}
13754 #endif
13755 				m_freem(m);
13756 				tp = tcp_close(tp);
13757 				if (tp == NULL) {
13758 #ifdef TCP_ACCOUNTING
13759 					sched_unpin();
13760 #endif
13761 					return (1);
13762 				}
13763 				/*
13764 				 * We would normally do drop-with-reset which would
13765 				 * send back a reset. We can't since we don't have
13766 				 * all the needed bits. Instead lets arrange for
13767 				 * a call to tcp_output(). That way since we
13768 				 * are in the closed state we will generate a reset.
13769 				 *
13770 				 * Note if tcp_accounting is on we don't unpin since
13771 				 * we do that after the goto label.
13772 				 */
13773 				goto send_out_a_rst;
13774 			}
13775 			if ((sbused(&so->so_snd) == 0) &&
13776 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
13777 			    (tp->t_flags & TF_SENTFIN)) {
13778 				/*
13779 				 * If we can't receive any more data, then closing user can
13780 				 * proceed. Starting the timer is contrary to the
13781 				 * specification, but if we don't get a FIN we'll hang
13782 				 * forever.
13783 				 *
13784 				 */
13785 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13786 					soisdisconnected(so);
13787 					tcp_timer_activate(tp, TT_2MSL,
13788 							   (tcp_fast_finwait2_recycle ?
13789 							    tcp_finwait2_timeout :
13790 							    TP_MAXIDLE(tp)));
13791 				}
13792 				if (ourfinisacked == 0) {
13793 					/*
13794 					 * We don't change to fin-wait-2 if we have our fin acked
13795 					 * which means we are probably in TCPS_CLOSING.
13796 					 */
13797 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
13798 				}
13799 			}
13800 		}
13801 		/* Wake up the socket if we have room to write more */
13802 		if (sbavail(&so->so_snd)) {
13803 			rack->r_wanted_output = 1;
13804 			if (ctf_progress_timeout_check(tp, true)) {
13805 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13806 							tp, tick, PROGRESS_DROP, __LINE__);
13807 				/*
13808 				 * We cheat here and don't send a RST, we should send one
13809 				 * when the pacer drops the connection.
13810 				 */
13811 #ifdef TCP_ACCOUNTING
13812 				rdstc = get_cyclecount();
13813 				if (rdstc > ts_val) {
13814 					counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13815 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13816 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13817 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13818 					}
13819 				}
13820 				sched_unpin();
13821 #endif
13822 				(void)tcp_drop(tp, ETIMEDOUT);
13823 				m_freem(m);
13824 				return (1);
13825 			}
13826 		}
13827 		if (ourfinisacked) {
13828 			switch(tp->t_state) {
13829 			case TCPS_CLOSING:
13830 #ifdef TCP_ACCOUNTING
13831 				rdstc = get_cyclecount();
13832 				if (rdstc > ts_val) {
13833 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13834 							(rdstc - ts_val));
13835 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13836 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13837 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13838 					}
13839 				}
13840 				sched_unpin();
13841 #endif
13842 				tcp_twstart(tp);
13843 				m_freem(m);
13844 				return (1);
13845 				break;
13846 			case TCPS_LAST_ACK:
13847 #ifdef TCP_ACCOUNTING
13848 				rdstc = get_cyclecount();
13849 				if (rdstc > ts_val) {
13850 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13851 							(rdstc - ts_val));
13852 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13853 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13854 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13855 					}
13856 				}
13857 				sched_unpin();
13858 #endif
13859 				tp = tcp_close(tp);
13860 				ctf_do_drop(m, tp);
13861 				return (1);
13862 				break;
13863 			case TCPS_FIN_WAIT_1:
13864 #ifdef TCP_ACCOUNTING
13865 				rdstc = get_cyclecount();
13866 				if (rdstc > ts_val) {
13867 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13868 							(rdstc - ts_val));
13869 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13870 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13871 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13872 					}
13873 				}
13874 #endif
13875 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13876 					soisdisconnected(so);
13877 					tcp_timer_activate(tp, TT_2MSL,
13878 							   (tcp_fast_finwait2_recycle ?
13879 							    tcp_finwait2_timeout :
13880 							    TP_MAXIDLE(tp)));
13881 				}
13882 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
13883 				break;
13884 			default:
13885 				break;
13886 			}
13887 		}
13888 		if (rack->r_fast_output) {
13889 			/*
13890 			 * We re doing fast output.. can we expand that?
13891 			 */
13892 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
13893 		}
13894 #ifdef TCP_ACCOUNTING
13895 		rdstc = get_cyclecount();
13896 		if (rdstc > ts_val) {
13897 			counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13898 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13899 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13900 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13901 			}
13902 		}
13903 
13904 	} else if (win_up_req) {
13905 		rdstc = get_cyclecount();
13906 		if (rdstc > ts_val) {
13907 			counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
13908 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13909 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
13910 			}
13911 		}
13912 #endif
13913 	}
13914 	/* Now is there a next packet, if so we are done */
13915 	m_freem(m);
13916 	did_out = 0;
13917 	if (nxt_pkt) {
13918 #ifdef TCP_ACCOUNTING
13919 		sched_unpin();
13920 #endif
13921 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
13922 		return (0);
13923 	}
13924 	rack_handle_might_revert(tp, rack);
13925 	ctf_calc_rwin(so, tp);
13926 	if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13927 	send_out_a_rst:
13928 		if (tcp_output(tp) < 0) {
13929 #ifdef TCP_ACCOUNTING
13930 			sched_unpin();
13931 #endif
13932 			return (1);
13933 		}
13934 		did_out = 1;
13935 	}
13936 	rack_free_trim(rack);
13937 #ifdef TCP_ACCOUNTING
13938 	sched_unpin();
13939 #endif
13940 	rack_timer_audit(tp, rack, &so->so_snd);
13941 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
13942 	return (0);
13943 }
13944 
13945 
13946 static int
13947 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
13948     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
13949     int32_t nxt_pkt, struct timeval *tv)
13950 {
13951 	struct inpcb *inp = tptoinpcb(tp);
13952 #ifdef TCP_ACCOUNTING
13953 	uint64_t ts_val;
13954 #endif
13955 	int32_t thflags, retval, did_out = 0;
13956 	int32_t way_out = 0;
13957 	/*
13958 	 * cts - is the current time from tv (caller gets ts) in microseconds.
13959 	 * ms_cts - is the current time from tv in milliseconds.
13960 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
13961 	 */
13962 	uint32_t cts, us_cts, ms_cts;
13963 	uint32_t tiwin, high_seq;
13964 	struct timespec ts;
13965 	struct tcpopt to;
13966 	struct tcp_rack *rack;
13967 	struct rack_sendmap *rsm;
13968 	int32_t prev_state = 0;
13969 #ifdef TCP_ACCOUNTING
13970 	int ack_val_set = 0xf;
13971 #endif
13972 	int nsegs;
13973 
13974 	NET_EPOCH_ASSERT();
13975 	INP_WLOCK_ASSERT(inp);
13976 
13977 	/*
13978 	 * tv passed from common code is from either M_TSTMP_LRO or
13979 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
13980 	 */
13981 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13982 	if (m->m_flags & M_ACKCMP) {
13983 		/*
13984 		 * All compressed ack's are ack's by definition so
13985 		 * remove any ack required flag and then do the processing.
13986 		 */
13987 		rack->rc_ack_required = 0;
13988 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
13989 	}
13990 	if (m->m_flags & M_ACKCMP) {
13991 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
13992 	}
13993 	cts = tcp_tv_to_usectick(tv);
13994 	ms_cts =  tcp_tv_to_mssectick(tv);
13995 	nsegs = m->m_pkthdr.lro_nsegs;
13996 	counter_u64_add(rack_proc_non_comp_ack, 1);
13997 	thflags = tcp_get_flags(th);
13998 #ifdef TCP_ACCOUNTING
13999 	sched_pin();
14000 	if (thflags & TH_ACK)
14001 		ts_val = get_cyclecount();
14002 #endif
14003 	if ((m->m_flags & M_TSTMP) ||
14004 	    (m->m_flags & M_TSTMP_LRO)) {
14005 		mbuf_tstmp2timespec(m, &ts);
14006 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
14007 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
14008 	} else
14009 		rack->r_ctl.act_rcv_time = *tv;
14010 	kern_prefetch(rack, &prev_state);
14011 	prev_state = 0;
14012 	/*
14013 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
14014 	 * the scale is zero.
14015 	 */
14016 	tiwin = th->th_win << tp->snd_scale;
14017 #ifdef TCP_ACCOUNTING
14018 	if (thflags & TH_ACK) {
14019 		/*
14020 		 * We have a tradeoff here. We can either do what we are
14021 		 * doing i.e. pinning to this CPU and then doing the accounting
14022 		 * <or> we could do a critical enter, setup the rdtsc and cpu
14023 		 * as in below, and then validate we are on the same CPU on
14024 		 * exit. I have choosen to not do the critical enter since
14025 		 * that often will gain you a context switch, and instead lock
14026 		 * us (line above this if) to the same CPU with sched_pin(). This
14027 		 * means we may be context switched out for a higher priority
14028 		 * interupt but we won't be moved to another CPU.
14029 		 *
14030 		 * If this occurs (which it won't very often since we most likely
14031 		 * are running this code in interupt context and only a higher
14032 		 * priority will bump us ... clock?) we will falsely add in
14033 		 * to the time the interupt processing time plus the ack processing
14034 		 * time. This is ok since its a rare event.
14035 		 */
14036 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
14037 						    ctf_fixed_maxseg(tp));
14038 	}
14039 #endif
14040 	/*
14041 	 * Parse options on any incoming segment.
14042 	 */
14043 	memset(&to, 0, sizeof(to));
14044 	tcp_dooptions(&to, (u_char *)(th + 1),
14045 	    (th->th_off << 2) - sizeof(struct tcphdr),
14046 	    (thflags & TH_SYN) ? TO_SYN : 0);
14047 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
14048 	    __func__));
14049 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
14050 	    __func__));
14051 
14052 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
14053 	    (tp->t_flags & TF_GPUTINPROG)) {
14054 		/*
14055 		 * We have a goodput in progress
14056 		 * and we have entered a late state.
14057 		 * Do we have enough data in the sb
14058 		 * to handle the GPUT request?
14059 		 */
14060 		uint32_t bytes;
14061 
14062 		bytes = tp->gput_ack - tp->gput_seq;
14063 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
14064 			bytes += tp->gput_seq - tp->snd_una;
14065 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
14066 			/*
14067 			 * There are not enough bytes in the socket
14068 			 * buffer that have been sent to cover this
14069 			 * measurement. Cancel it.
14070 			 */
14071 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
14072 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
14073 						   tp->gput_seq,
14074 						   0, 0, 18, __LINE__, NULL, 0);
14075 			tp->t_flags &= ~TF_GPUTINPROG;
14076 		}
14077 	}
14078 	high_seq = th->th_ack;
14079 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14080 		union tcp_log_stackspecific log;
14081 		struct timeval ltv;
14082 #ifdef NETFLIX_HTTP_LOGGING
14083 		struct http_sendfile_track *http_req;
14084 
14085 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
14086 			http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
14087 		} else {
14088 			http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
14089 		}
14090 #endif
14091 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14092 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
14093 		if (rack->rack_no_prr == 0)
14094 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
14095 		else
14096 			log.u_bbr.flex1 = 0;
14097 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
14098 		log.u_bbr.use_lt_bw <<= 1;
14099 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
14100 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
14101 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14102 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
14103 		log.u_bbr.flex3 = m->m_flags;
14104 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
14105 		log.u_bbr.lost = thflags;
14106 		log.u_bbr.pacing_gain = 0x1;
14107 #ifdef TCP_ACCOUNTING
14108 		log.u_bbr.cwnd_gain = ack_val_set;
14109 #endif
14110 		log.u_bbr.flex7 = 2;
14111 		if (m->m_flags & M_TSTMP) {
14112 			/* Record the hardware timestamp if present */
14113 			mbuf_tstmp2timespec(m, &ts);
14114 			ltv.tv_sec = ts.tv_sec;
14115 			ltv.tv_usec = ts.tv_nsec / 1000;
14116 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
14117 		} else if (m->m_flags & M_TSTMP_LRO) {
14118 			/* Record the LRO the arrival timestamp */
14119 			mbuf_tstmp2timespec(m, &ts);
14120 			ltv.tv_sec = ts.tv_sec;
14121 			ltv.tv_usec = ts.tv_nsec / 1000;
14122 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
14123 		}
14124 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
14125 		/* Log the rcv time */
14126 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
14127 #ifdef NETFLIX_HTTP_LOGGING
14128 		log.u_bbr.applimited = tp->t_http_closed;
14129 		log.u_bbr.applimited <<= 8;
14130 		log.u_bbr.applimited |= tp->t_http_open;
14131 		log.u_bbr.applimited <<= 8;
14132 		log.u_bbr.applimited |= tp->t_http_req;
14133 		if (http_req) {
14134 			/* Copy out any client req info */
14135 			/* seconds */
14136 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
14137 			/* useconds */
14138 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
14139 			log.u_bbr.rttProp = http_req->timestamp;
14140 			log.u_bbr.cur_del_rate = http_req->start;
14141 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
14142 				log.u_bbr.flex8 |= 1;
14143 			} else {
14144 				log.u_bbr.flex8 |= 2;
14145 				log.u_bbr.bw_inuse = http_req->end;
14146 			}
14147 			log.u_bbr.flex6 = http_req->start_seq;
14148 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
14149 				log.u_bbr.flex8 |= 4;
14150 				log.u_bbr.epoch = http_req->end_seq;
14151 			}
14152 		}
14153 #endif
14154 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
14155 		    tlen, &log, true, &ltv);
14156 	}
14157 	/* Remove ack required flag if set, we have one  */
14158 	if (thflags & TH_ACK)
14159 		rack->rc_ack_required = 0;
14160 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
14161 		way_out = 4;
14162 		retval = 0;
14163 		m_freem(m);
14164 		goto done_with_input;
14165 	}
14166 	/*
14167 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
14168 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
14169 	 */
14170 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
14171 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
14172 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14173 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14174 #ifdef TCP_ACCOUNTING
14175 		sched_unpin();
14176 #endif
14177 		return (1);
14178 	}
14179 	/*
14180 	 * If timestamps were negotiated during SYN/ACK and a
14181 	 * segment without a timestamp is received, silently drop
14182 	 * the segment, unless it is a RST segment or missing timestamps are
14183 	 * tolerated.
14184 	 * See section 3.2 of RFC 7323.
14185 	 */
14186 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
14187 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
14188 		way_out = 5;
14189 		retval = 0;
14190 		m_freem(m);
14191 		goto done_with_input;
14192 	}
14193 
14194 	/*
14195 	 * Segment received on connection. Reset idle time and keep-alive
14196 	 * timer. XXX: This should be done after segment validation to
14197 	 * ignore broken/spoofed segs.
14198 	 */
14199 	if  (tp->t_idle_reduce &&
14200 	     (tp->snd_max == tp->snd_una) &&
14201 	     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
14202 		counter_u64_add(rack_input_idle_reduces, 1);
14203 		rack_cc_after_idle(rack, tp);
14204 	}
14205 	tp->t_rcvtime = ticks;
14206 #ifdef STATS
14207 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
14208 #endif
14209 	if (tiwin > rack->r_ctl.rc_high_rwnd)
14210 		rack->r_ctl.rc_high_rwnd = tiwin;
14211 	/*
14212 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
14213 	 * this to occur after we've validated the segment.
14214 	 */
14215 	if (tcp_ecn_input_segment(tp, thflags, tlen,
14216 	    tcp_packets_this_ack(tp, th->th_ack),
14217 	    iptos))
14218 		rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
14219 
14220 	/*
14221 	 * If echoed timestamp is later than the current time, fall back to
14222 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
14223 	 * were used when this connection was established.
14224 	 */
14225 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
14226 		to.to_tsecr -= tp->ts_offset;
14227 		if (TSTMP_GT(to.to_tsecr, ms_cts))
14228 			to.to_tsecr = 0;
14229 	}
14230 
14231 	/*
14232 	 * If its the first time in we need to take care of options and
14233 	 * verify we can do SACK for rack!
14234 	 */
14235 	if (rack->r_state == 0) {
14236 		/* Should be init'd by rack_init() */
14237 		KASSERT(rack->rc_inp != NULL,
14238 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
14239 		if (rack->rc_inp == NULL) {
14240 			rack->rc_inp = inp;
14241 		}
14242 
14243 		/*
14244 		 * Process options only when we get SYN/ACK back. The SYN
14245 		 * case for incoming connections is handled in tcp_syncache.
14246 		 * According to RFC1323 the window field in a SYN (i.e., a
14247 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
14248 		 * this is traditional behavior, may need to be cleaned up.
14249 		 */
14250 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
14251 			/* Handle parallel SYN for ECN */
14252 			tcp_ecn_input_parallel_syn(tp, thflags, iptos);
14253 			if ((to.to_flags & TOF_SCALE) &&
14254 			    (tp->t_flags & TF_REQ_SCALE)) {
14255 				tp->t_flags |= TF_RCVD_SCALE;
14256 				tp->snd_scale = to.to_wscale;
14257 			} else
14258 				tp->t_flags &= ~TF_REQ_SCALE;
14259 			/*
14260 			 * Initial send window.  It will be updated with the
14261 			 * next incoming segment to the scaled value.
14262 			 */
14263 			tp->snd_wnd = th->th_win;
14264 			rack_validate_fo_sendwin_up(tp, rack);
14265 			if ((to.to_flags & TOF_TS) &&
14266 			    (tp->t_flags & TF_REQ_TSTMP)) {
14267 				tp->t_flags |= TF_RCVD_TSTMP;
14268 				tp->ts_recent = to.to_tsval;
14269 				tp->ts_recent_age = cts;
14270 			} else
14271 				tp->t_flags &= ~TF_REQ_TSTMP;
14272 			if (to.to_flags & TOF_MSS) {
14273 				tcp_mss(tp, to.to_mss);
14274 			}
14275 			if ((tp->t_flags & TF_SACK_PERMIT) &&
14276 			    (to.to_flags & TOF_SACKPERM) == 0)
14277 				tp->t_flags &= ~TF_SACK_PERMIT;
14278 			if (IS_FASTOPEN(tp->t_flags)) {
14279 				if (to.to_flags & TOF_FASTOPEN) {
14280 					uint16_t mss;
14281 
14282 					if (to.to_flags & TOF_MSS)
14283 						mss = to.to_mss;
14284 					else
14285 						if ((inp->inp_vflag & INP_IPV6) != 0)
14286 							mss = TCP6_MSS;
14287 						else
14288 							mss = TCP_MSS;
14289 					tcp_fastopen_update_cache(tp, mss,
14290 					    to.to_tfo_len, to.to_tfo_cookie);
14291 				} else
14292 					tcp_fastopen_disable_path(tp);
14293 			}
14294 		}
14295 		/*
14296 		 * At this point we are at the initial call. Here we decide
14297 		 * if we are doing RACK or not. We do this by seeing if
14298 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
14299 		 * The code now does do dup-ack counting so if you don't
14300 		 * switch back you won't get rack & TLP, but you will still
14301 		 * get this stack.
14302 		 */
14303 
14304 		if ((rack_sack_not_required == 0) &&
14305 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
14306 			tcp_switch_back_to_default(tp);
14307 			(*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
14308 			    tlen, iptos);
14309 #ifdef TCP_ACCOUNTING
14310 			sched_unpin();
14311 #endif
14312 			return (1);
14313 		}
14314 		tcp_set_hpts(inp);
14315 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
14316 	}
14317 	if (thflags & TH_FIN)
14318 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
14319 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14320 	if ((rack->rc_gp_dyn_mul) &&
14321 	    (rack->use_fixed_rate == 0) &&
14322 	    (rack->rc_always_pace)) {
14323 		/* Check in on probertt */
14324 		rack_check_probe_rtt(rack, us_cts);
14325 	}
14326 	rack_clear_rate_sample(rack);
14327 	if ((rack->forced_ack) &&
14328 	    ((tcp_get_flags(th) & TH_RST) == 0)) {
14329 		rack_handle_probe_response(rack, tiwin, us_cts);
14330 	}
14331 	/*
14332 	 * This is the one exception case where we set the rack state
14333 	 * always. All other times (timers etc) we must have a rack-state
14334 	 * set (so we assure we have done the checks above for SACK).
14335 	 */
14336 	rack->r_ctl.rc_rcvtime = cts;
14337 	if (rack->r_state != tp->t_state)
14338 		rack_set_state(tp, rack);
14339 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
14340 	    (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
14341 		kern_prefetch(rsm, &prev_state);
14342 	prev_state = rack->r_state;
14343 	retval = (*rack->r_substate) (m, th, so,
14344 	    tp, &to, drop_hdrlen,
14345 	    tlen, tiwin, thflags, nxt_pkt, iptos);
14346 	if (retval == 0) {
14347 		/*
14348 		 * If retval is 1 the tcb is unlocked and most likely the tp
14349 		 * is gone.
14350 		 */
14351 		INP_WLOCK_ASSERT(inp);
14352 		if ((rack->rc_gp_dyn_mul) &&
14353 		    (rack->rc_always_pace) &&
14354 		    (rack->use_fixed_rate == 0) &&
14355 		    rack->in_probe_rtt &&
14356 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
14357 			/*
14358 			 * If we are going for target, lets recheck before
14359 			 * we output.
14360 			 */
14361 			rack_check_probe_rtt(rack, us_cts);
14362 		}
14363 		if (rack->set_pacing_done_a_iw == 0) {
14364 			/* How much has been acked? */
14365 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
14366 				/* We have enough to set in the pacing segment size */
14367 				rack->set_pacing_done_a_iw = 1;
14368 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
14369 			}
14370 		}
14371 		tcp_rack_xmit_timer_commit(rack, tp);
14372 #ifdef TCP_ACCOUNTING
14373 		/*
14374 		 * If we set the ack_val_se to what ack processing we are doing
14375 		 * we also want to track how many cycles we burned. Note
14376 		 * the bits after tcp_output we let be "free". This is because
14377 		 * we are also tracking the tcp_output times as well. Note the
14378 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
14379 		 * 0xf cannot be returned and is what we initialize it too to
14380 		 * indicate we are not doing the tabulations.
14381 		 */
14382 		if (ack_val_set != 0xf) {
14383 			uint64_t crtsc;
14384 
14385 			crtsc = get_cyclecount();
14386 			counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14387 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14388 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
14389 			}
14390 		}
14391 #endif
14392 		if (nxt_pkt == 0) {
14393 			if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14394 do_output_now:
14395 				if (tcp_output(tp) < 0)
14396 					return (1);
14397 				did_out = 1;
14398 			}
14399 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
14400 			rack_free_trim(rack);
14401 		}
14402 		/* Update any rounds needed */
14403 		if (rack_verbose_logging &&  (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
14404 			union tcp_log_stackspecific log;
14405 			struct timeval tv;
14406 
14407 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14408 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14409 			log.u_bbr.flex1 = high_seq;
14410 			log.u_bbr.flex2 = rack->r_ctl.roundends;
14411 			log.u_bbr.flex3 = rack->r_ctl.current_round;
14412 			log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
14413 			log.u_bbr.flex8 = 9;
14414 			tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
14415 				       0, &log, false, NULL, NULL, 0, &tv);
14416 		}
14417 		/*
14418 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
14419 		 * causes issues when we are just going app limited. Lets
14420 		 * instead use SEQ_GT <or> where its equal but more data
14421 		 * is outstanding.
14422 		 */
14423 		if ((SEQ_GT(tp->snd_una, rack->r_ctl.roundends)) ||
14424 		    ((tp->snd_una == rack->r_ctl.roundends) && SEQ_GT(tp->snd_max, tp->snd_una))) {
14425 			rack->r_ctl.current_round++;
14426 			rack->r_ctl.roundends = tp->snd_max;
14427 			if (CC_ALGO(tp)->newround != NULL) {
14428 				CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
14429 			}
14430 		}
14431 		if ((nxt_pkt == 0) &&
14432 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
14433 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
14434 		     (tp->t_flags & TF_DELACK) ||
14435 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
14436 		      (tp->t_state <= TCPS_CLOSING)))) {
14437 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
14438 			if ((tp->snd_max == tp->snd_una) &&
14439 			    ((tp->t_flags & TF_DELACK) == 0) &&
14440 			    (tcp_in_hpts(rack->rc_inp)) &&
14441 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
14442 				/* keep alive not needed if we are hptsi output yet */
14443 				;
14444 			} else {
14445 				int late = 0;
14446 				if (tcp_in_hpts(inp)) {
14447 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14448 						us_cts = tcp_get_usecs(NULL);
14449 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
14450 							rack->r_early = 1;
14451 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
14452 						} else
14453 							late = 1;
14454 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
14455 					}
14456 					tcp_hpts_remove(inp);
14457 				}
14458 				if (late && (did_out == 0)) {
14459 					/*
14460 					 * We are late in the sending
14461 					 * and we did not call the output
14462 					 * (this probably should not happen).
14463 					 */
14464 					goto do_output_now;
14465 				}
14466 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14467 			}
14468 			way_out = 1;
14469 		} else if (nxt_pkt == 0) {
14470 			/* Do we have the correct timer running? */
14471 			rack_timer_audit(tp, rack, &so->so_snd);
14472 			way_out = 2;
14473 		}
14474 	done_with_input:
14475 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
14476 		if (did_out)
14477 			rack->r_wanted_output = 0;
14478 #ifdef TCP_ACCOUNTING
14479 	} else {
14480 		/*
14481 		 * Track the time (see above).
14482 		 */
14483 		if (ack_val_set != 0xf) {
14484 			uint64_t crtsc;
14485 
14486 			crtsc = get_cyclecount();
14487 			counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14488 			/*
14489 			 * Note we *DO NOT* increment the per-tcb counters since
14490 			 * in the else the TP may be gone!!
14491 			 */
14492 		}
14493 #endif
14494 	}
14495 #ifdef TCP_ACCOUNTING
14496 	sched_unpin();
14497 #endif
14498 	return (retval);
14499 }
14500 
14501 void
14502 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
14503     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
14504 {
14505 	struct timeval tv;
14506 
14507 	/* First lets see if we have old packets */
14508 	if (tp->t_in_pkt) {
14509 		if (ctf_do_queued_segments(so, tp, 1)) {
14510 			m_freem(m);
14511 			return;
14512 		}
14513 	}
14514 	if (m->m_flags & M_TSTMP_LRO) {
14515 		mbuf_tstmp2timeval(m, &tv);
14516 	} else {
14517 		/* Should not be should we kassert instead? */
14518 		tcp_get_usecs(&tv);
14519 	}
14520 	if (rack_do_segment_nounlock(m, th, so, tp,
14521 				     drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14522 		INP_WUNLOCK(tptoinpcb(tp));
14523 	}
14524 }
14525 
14526 struct rack_sendmap *
14527 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14528 {
14529 	struct rack_sendmap *rsm = NULL;
14530 	int32_t idx;
14531 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
14532 
14533 	/* Return the next guy to be re-transmitted */
14534 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14535 		return (NULL);
14536 	}
14537 	if (tp->t_flags & TF_SENTFIN) {
14538 		/* retran the end FIN? */
14539 		return (NULL);
14540 	}
14541 	/* ok lets look at this one */
14542 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14543 	if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
14544 		return (rsm);
14545 	}
14546 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14547 		goto check_it;
14548 	}
14549 	rsm = rack_find_lowest_rsm(rack);
14550 	if (rsm == NULL) {
14551 		return (NULL);
14552 	}
14553 check_it:
14554 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14555 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14556 		/*
14557 		 * No sack so we automatically do the 3 strikes and
14558 		 * retransmit (no rack timer would be started).
14559 		 */
14560 
14561 		return (rsm);
14562 	}
14563 	if (rsm->r_flags & RACK_ACKED) {
14564 		return (NULL);
14565 	}
14566 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14567 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14568 		/* Its not yet ready */
14569 		return (NULL);
14570 	}
14571 	srtt = rack_grab_rtt(tp, rack);
14572 	idx = rsm->r_rtr_cnt - 1;
14573 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14574 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14575 	if ((tsused == ts_low) ||
14576 	    (TSTMP_LT(tsused, ts_low))) {
14577 		/* No time since sending */
14578 		return (NULL);
14579 	}
14580 	if ((tsused - ts_low) < thresh) {
14581 		/* It has not been long enough yet */
14582 		return (NULL);
14583 	}
14584 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14585 	    ((rsm->r_flags & RACK_SACK_PASSED) &&
14586 	     (rack->sack_attack_disable == 0))) {
14587 		/*
14588 		 * We have passed the dup-ack threshold <or>
14589 		 * a SACK has indicated this is missing.
14590 		 * Note that if you are a declared attacker
14591 		 * it is only the dup-ack threshold that
14592 		 * will cause retransmits.
14593 		 */
14594 		/* log retransmit reason */
14595 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14596 		rack->r_fast_output = 0;
14597 		return (rsm);
14598 	}
14599 	return (NULL);
14600 }
14601 
14602 static void
14603 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14604 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14605 			   int line, struct rack_sendmap *rsm, uint8_t quality)
14606 {
14607 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14608 		union tcp_log_stackspecific log;
14609 		struct timeval tv;
14610 
14611 		memset(&log, 0, sizeof(log));
14612 		log.u_bbr.flex1 = slot;
14613 		log.u_bbr.flex2 = len;
14614 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14615 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14616 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14617 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14618 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14619 		log.u_bbr.use_lt_bw <<= 1;
14620 		log.u_bbr.use_lt_bw |= rack->r_late;
14621 		log.u_bbr.use_lt_bw <<= 1;
14622 		log.u_bbr.use_lt_bw |= rack->r_early;
14623 		log.u_bbr.use_lt_bw <<= 1;
14624 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14625 		log.u_bbr.use_lt_bw <<= 1;
14626 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14627 		log.u_bbr.use_lt_bw <<= 1;
14628 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14629 		log.u_bbr.use_lt_bw <<= 1;
14630 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14631 		log.u_bbr.use_lt_bw <<= 1;
14632 		log.u_bbr.use_lt_bw |= rack->gp_ready;
14633 		log.u_bbr.pkt_epoch = line;
14634 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14635 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14636 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14637 		log.u_bbr.bw_inuse = bw_est;
14638 		log.u_bbr.delRate = bw;
14639 		if (rack->r_ctl.gp_bw == 0)
14640 			log.u_bbr.cur_del_rate = 0;
14641 		else
14642 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
14643 		log.u_bbr.rttProp = len_time;
14644 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14645 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14646 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14647 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14648 			/* We are in slow start */
14649 			log.u_bbr.flex7 = 1;
14650 		} else {
14651 			/* we are on congestion avoidance */
14652 			log.u_bbr.flex7 = 0;
14653 		}
14654 		log.u_bbr.flex8 = method;
14655 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14656 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14657 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14658 		log.u_bbr.cwnd_gain <<= 1;
14659 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14660 		log.u_bbr.cwnd_gain <<= 1;
14661 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14662 		log.u_bbr.bbr_substate = quality;
14663 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
14664 		    &rack->rc_inp->inp_socket->so_rcv,
14665 		    &rack->rc_inp->inp_socket->so_snd,
14666 		    BBR_LOG_HPTSI_CALC, 0,
14667 		    0, &log, false, &tv);
14668 	}
14669 }
14670 
14671 static uint32_t
14672 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14673 {
14674 	uint32_t new_tso, user_max;
14675 
14676 	user_max = rack->rc_user_set_max_segs * mss;
14677 	if (rack->rc_force_max_seg) {
14678 		return (user_max);
14679 	}
14680 	if (rack->use_fixed_rate &&
14681 	    ((rack->r_ctl.crte == NULL) ||
14682 	     (bw != rack->r_ctl.crte->rate))) {
14683 		/* Use the user mss since we are not exactly matched */
14684 		return (user_max);
14685 	}
14686 	new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14687 	if (new_tso > user_max)
14688 		new_tso = user_max;
14689 	return (new_tso);
14690 }
14691 
14692 static int32_t
14693 pace_to_fill_cwnd(struct tcp_rack *rack, int32_t slot, uint32_t len, uint32_t segsiz, int *capped, uint64_t *rate_wanted, uint8_t non_paced)
14694 {
14695 	uint64_t lentim, fill_bw;
14696 
14697 	/* Lets first see if we are full, if so continue with normal rate */
14698 	rack->r_via_fill_cw = 0;
14699 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14700 		return (slot);
14701 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14702 		return (slot);
14703 	if (rack->r_ctl.rc_last_us_rtt == 0)
14704 		return (slot);
14705 	if (rack->rc_pace_fill_if_rttin_range &&
14706 	    (rack->r_ctl.rc_last_us_rtt >=
14707 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14708 		/* The rtt is huge, N * smallest, lets not fill */
14709 		return (slot);
14710 	}
14711 	/*
14712 	 * first lets calculate the b/w based on the last us-rtt
14713 	 * and the sndwnd.
14714 	 */
14715 	fill_bw = rack->r_ctl.cwnd_to_use;
14716 	/* Take the rwnd if its smaller */
14717 	if (fill_bw > rack->rc_tp->snd_wnd)
14718 		fill_bw = rack->rc_tp->snd_wnd;
14719 	if (rack->r_fill_less_agg) {
14720 		/*
14721 		 * Now take away the inflight (this will reduce our
14722 		 * aggressiveness and yeah, if we get that much out in 1RTT
14723 		 * we will have had acks come back and still be behind).
14724 		 */
14725 		fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14726 	}
14727 	/* Now lets make it into a b/w */
14728 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14729 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14730 	/* We are below the min b/w */
14731 	if (non_paced)
14732 		*rate_wanted = fill_bw;
14733 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14734 		return (slot);
14735 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14736 		fill_bw = rack->r_ctl.bw_rate_cap;
14737 	rack->r_via_fill_cw = 1;
14738 	if (rack->r_rack_hw_rate_caps &&
14739 	    (rack->r_ctl.crte != NULL)) {
14740 		uint64_t high_rate;
14741 
14742 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14743 		if (fill_bw > high_rate) {
14744 			/* We are capping bw at the highest rate table entry */
14745 			if (*rate_wanted > high_rate) {
14746 				/* The original rate was also capped */
14747 				rack->r_via_fill_cw = 0;
14748 			}
14749 			rack_log_hdwr_pacing(rack,
14750 					     fill_bw, high_rate, __LINE__,
14751 					     0, 3);
14752 			fill_bw = high_rate;
14753 			if (capped)
14754 				*capped = 1;
14755 		}
14756 	} else if ((rack->r_ctl.crte == NULL) &&
14757 		   (rack->rack_hdrw_pacing == 0) &&
14758 		   (rack->rack_hdw_pace_ena) &&
14759 		   rack->r_rack_hw_rate_caps &&
14760 		   (rack->rack_attempt_hdwr_pace == 0) &&
14761 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
14762 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14763 		/*
14764 		 * Ok we may have a first attempt that is greater than our top rate
14765 		 * lets check.
14766 		 */
14767 		uint64_t high_rate;
14768 
14769 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14770 		if (high_rate) {
14771 			if (fill_bw > high_rate) {
14772 				fill_bw = high_rate;
14773 				if (capped)
14774 					*capped = 1;
14775 			}
14776 		}
14777 	}
14778 	/*
14779 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
14780 	 * in a rtt, what does that time wise equate too?
14781 	 */
14782 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14783 	lentim /= fill_bw;
14784 	*rate_wanted = fill_bw;
14785 	if (non_paced || (lentim < slot)) {
14786 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14787 					   0, lentim, 12, __LINE__, NULL, 0);
14788 		return ((int32_t)lentim);
14789 	} else
14790 		return (slot);
14791 }
14792 
14793 static int32_t
14794 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14795 {
14796 	uint64_t srtt;
14797 	int32_t slot = 0;
14798 	int can_start_hw_pacing = 1;
14799 	int err;
14800 
14801 	if (rack->rc_always_pace == 0) {
14802 		/*
14803 		 * We use the most optimistic possible cwnd/srtt for
14804 		 * sending calculations. This will make our
14805 		 * calculation anticipate getting more through
14806 		 * quicker then possible. But thats ok we don't want
14807 		 * the peer to have a gap in data sending.
14808 		 */
14809 		uint64_t cwnd, tr_perms = 0;
14810 		int32_t reduce = 0;
14811 
14812 	old_method:
14813 		/*
14814 		 * We keep no precise pacing with the old method
14815 		 * instead we use the pacer to mitigate bursts.
14816 		 */
14817 		if (rack->r_ctl.rc_rack_min_rtt)
14818 			srtt = rack->r_ctl.rc_rack_min_rtt;
14819 		else
14820 			srtt = max(tp->t_srtt, 1);
14821 		if (rack->r_ctl.rc_rack_largest_cwnd)
14822 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14823 		else
14824 			cwnd = rack->r_ctl.cwnd_to_use;
14825 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
14826 		tr_perms = (cwnd * 1000) / srtt;
14827 		if (tr_perms == 0) {
14828 			tr_perms = ctf_fixed_maxseg(tp);
14829 		}
14830 		/*
14831 		 * Calculate how long this will take to drain, if
14832 		 * the calculation comes out to zero, thats ok we
14833 		 * will use send_a_lot to possibly spin around for
14834 		 * more increasing tot_len_this_send to the point
14835 		 * that its going to require a pace, or we hit the
14836 		 * cwnd. Which in that case we are just waiting for
14837 		 * a ACK.
14838 		 */
14839 		slot = len / tr_perms;
14840 		/* Now do we reduce the time so we don't run dry? */
14841 		if (slot && rack_slot_reduction) {
14842 			reduce = (slot / rack_slot_reduction);
14843 			if (reduce < slot) {
14844 				slot -= reduce;
14845 			} else
14846 				slot = 0;
14847 		}
14848 		slot *= HPTS_USEC_IN_MSEC;
14849 		if (rack->rc_pace_to_cwnd) {
14850 			uint64_t rate_wanted = 0;
14851 
14852 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14853 			rack->rc_ack_can_sendout_data = 1;
14854 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
14855 		} else
14856 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
14857 	} else {
14858 		uint64_t bw_est, res, lentim, rate_wanted;
14859 		uint32_t orig_val, segs, oh;
14860 		int capped = 0;
14861 		int prev_fill;
14862 
14863 		if ((rack->r_rr_config == 1) && rsm) {
14864 			return (rack->r_ctl.rc_min_to);
14865 		}
14866 		if (rack->use_fixed_rate) {
14867 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14868 		} else if ((rack->r_ctl.init_rate == 0) &&
14869 #ifdef NETFLIX_PEAKRATE
14870 			   (rack->rc_tp->t_maxpeakrate == 0) &&
14871 #endif
14872 			   (rack->r_ctl.gp_bw == 0)) {
14873 			/* no way to yet do an estimate */
14874 			bw_est = rate_wanted = 0;
14875 		} else {
14876 			bw_est = rack_get_bw(rack);
14877 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14878 		}
14879 		if ((bw_est == 0) || (rate_wanted == 0) ||
14880 		    ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14881 			/*
14882 			 * No way yet to make a b/w estimate or
14883 			 * our raise is set incorrectly.
14884 			 */
14885 			goto old_method;
14886 		}
14887 		/* We need to account for all the overheads */
14888 		segs = (len + segsiz - 1) / segsiz;
14889 		/*
14890 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
14891 		 * and how much data we put in each packet. Yes this
14892 		 * means we may be off if we are larger than 1500 bytes
14893 		 * or smaller. But this just makes us more conservative.
14894 		 */
14895 		if (rack_hw_rate_min &&
14896 		    (bw_est < rack_hw_rate_min))
14897 			can_start_hw_pacing = 0;
14898 		if (ETHERNET_SEGMENT_SIZE > segsiz)
14899 			oh = ETHERNET_SEGMENT_SIZE - segsiz;
14900 		else
14901 			oh = 0;
14902 		segs *= oh;
14903 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
14904 		res = lentim / rate_wanted;
14905 		slot = (uint32_t)res;
14906 		orig_val = rack->r_ctl.rc_pace_max_segs;
14907 		if (rack->r_ctl.crte == NULL) {
14908 			/*
14909 			 * Only do this if we are not hardware pacing
14910 			 * since if we are doing hw-pacing below we will
14911 			 * set make a call after setting up or changing
14912 			 * the rate.
14913 			 */
14914 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
14915 		} else if (rack->rc_inp->inp_snd_tag == NULL) {
14916 			/*
14917 			 * We lost our rate somehow, this can happen
14918 			 * if the interface changed underneath us.
14919 			 */
14920 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14921 			rack->r_ctl.crte = NULL;
14922 			/* Lets re-allow attempting to setup pacing */
14923 			rack->rack_hdrw_pacing = 0;
14924 			rack->rack_attempt_hdwr_pace = 0;
14925 			rack_log_hdwr_pacing(rack,
14926 					     rate_wanted, bw_est, __LINE__,
14927 					     0, 6);
14928 		}
14929 		/* Did we change the TSO size, if so log it */
14930 		if (rack->r_ctl.rc_pace_max_segs != orig_val)
14931 			rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL, 0);
14932 		prev_fill = rack->r_via_fill_cw;
14933 		if ((rack->rc_pace_to_cwnd) &&
14934 		    (capped == 0) &&
14935 		    (rack->use_fixed_rate == 0) &&
14936 		    (rack->in_probe_rtt == 0) &&
14937 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
14938 			/*
14939 			 * We want to pace at our rate *or* faster to
14940 			 * fill the cwnd to the max if its not full.
14941 			 */
14942 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
14943 		}
14944 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
14945 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14946 			if ((rack->rack_hdw_pace_ena) &&
14947 			    (can_start_hw_pacing > 0) &&
14948 			    (rack->rack_hdrw_pacing == 0) &&
14949 			    (rack->rack_attempt_hdwr_pace == 0)) {
14950 				/*
14951 				 * Lets attempt to turn on hardware pacing
14952 				 * if we can.
14953 				 */
14954 				rack->rack_attempt_hdwr_pace = 1;
14955 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
14956 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
14957 								       rate_wanted,
14958 								       RS_PACING_GEQ,
14959 								       &err, &rack->r_ctl.crte_prev_rate);
14960 				if (rack->r_ctl.crte) {
14961 					rack->rack_hdrw_pacing = 1;
14962 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
14963 												 0, rack->r_ctl.crte,
14964 												 NULL);
14965 					rack_log_hdwr_pacing(rack,
14966 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14967 							     err, 0);
14968 					rack->r_ctl.last_hw_bw_req = rate_wanted;
14969 				} else {
14970 					counter_u64_add(rack_hw_pace_init_fail, 1);
14971 				}
14972 			} else if (rack->rack_hdrw_pacing &&
14973 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
14974 				/* Do we need to adjust our rate? */
14975 				const struct tcp_hwrate_limit_table *nrte;
14976 
14977 				if (rack->r_up_only &&
14978 				    (rate_wanted < rack->r_ctl.crte->rate)) {
14979 					/**
14980 					 * We have four possible states here
14981 					 * having to do with the previous time
14982 					 * and this time.
14983 					 *   previous  |  this-time
14984 					 * A)     0      |     0   -- fill_cw not in the picture
14985 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
14986 					 * C)     1      |     1   -- all rates from fill_cw
14987 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
14988 					 *
14989 					 * For case A, C and D we don't allow a drop. But for
14990 					 * case B where we now our on our steady rate we do
14991 					 * allow a drop.
14992 					 *
14993 					 */
14994 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
14995 						goto done_w_hdwr;
14996 				}
14997 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
14998 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
14999 					if (rack_hw_rate_to_low &&
15000 					    (bw_est < rack_hw_rate_to_low)) {
15001 						/*
15002 						 * The pacing rate is too low for hardware, but
15003 						 * do allow hardware pacing to be restarted.
15004 						 */
15005 						rack_log_hdwr_pacing(rack,
15006 							     bw_est, rack->r_ctl.crte->rate, __LINE__,
15007 							     0, 5);
15008 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
15009 						rack->r_ctl.crte = NULL;
15010 						rack->rack_attempt_hdwr_pace = 0;
15011 						rack->rack_hdrw_pacing = 0;
15012 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15013 						goto done_w_hdwr;
15014 					}
15015 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
15016 								   rack->rc_tp,
15017 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
15018 								   rate_wanted,
15019 								   RS_PACING_GEQ,
15020 								   &err, &rack->r_ctl.crte_prev_rate);
15021 					if (nrte == NULL) {
15022 						/* Lost the rate */
15023 						rack->rack_hdrw_pacing = 0;
15024 						rack->r_ctl.crte = NULL;
15025 						rack_log_hdwr_pacing(rack,
15026 								     rate_wanted, 0, __LINE__,
15027 								     err, 1);
15028 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15029 						counter_u64_add(rack_hw_pace_lost, 1);
15030 					} else if (nrte != rack->r_ctl.crte) {
15031 						rack->r_ctl.crte = nrte;
15032 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
15033 													 segsiz, 0,
15034 													 rack->r_ctl.crte,
15035 													 NULL);
15036 						rack_log_hdwr_pacing(rack,
15037 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15038 								     err, 2);
15039 						rack->r_ctl.last_hw_bw_req = rate_wanted;
15040 					}
15041 				} else {
15042 					/* We just need to adjust the segment size */
15043 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15044 					rack_log_hdwr_pacing(rack,
15045 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15046 							     0, 4);
15047 					rack->r_ctl.last_hw_bw_req = rate_wanted;
15048 				}
15049 			}
15050 		}
15051 		if ((rack->r_ctl.crte != NULL) &&
15052 		    (rack->r_ctl.crte->rate == rate_wanted)) {
15053 			/*
15054 			 * We need to add a extra if the rates
15055 			 * are exactly matched. The idea is
15056 			 * we want the software to make sure the
15057 			 * queue is empty before adding more, this
15058 			 * gives us N MSS extra pace times where
15059 			 * N is our sysctl
15060 			 */
15061 			slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
15062 		}
15063 done_w_hdwr:
15064 		if (rack_limit_time_with_srtt &&
15065 		    (rack->use_fixed_rate == 0) &&
15066 #ifdef NETFLIX_PEAKRATE
15067 		    (rack->rc_tp->t_maxpeakrate == 0) &&
15068 #endif
15069 		    (rack->rack_hdrw_pacing == 0)) {
15070 			/*
15071 			 * Sanity check, we do not allow the pacing delay
15072 			 * to be longer than the SRTT of the path. If it is
15073 			 * a slow path, then adding a packet should increase
15074 			 * the RTT and compensate for this i.e. the srtt will
15075 			 * be greater so the allowed pacing time will be greater.
15076 			 *
15077 			 * Note this restriction is not for where a peak rate
15078 			 * is set, we are doing fixed pacing or hardware pacing.
15079 			 */
15080 			if (rack->rc_tp->t_srtt)
15081 				srtt = rack->rc_tp->t_srtt;
15082 			else
15083 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
15084 			if (srtt < (uint64_t)slot) {
15085 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
15086 				slot = srtt;
15087 			}
15088 		}
15089 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
15090 	}
15091 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
15092 		/*
15093 		 * If this rate is seeing enobufs when it
15094 		 * goes to send then either the nic is out
15095 		 * of gas or we are mis-estimating the time
15096 		 * somehow and not letting the queue empty
15097 		 * completely. Lets add to the pacing time.
15098 		 */
15099 		int hw_boost_delay;
15100 
15101 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
15102 		if (hw_boost_delay > rack_enobuf_hw_max)
15103 			hw_boost_delay = rack_enobuf_hw_max;
15104 		else if (hw_boost_delay < rack_enobuf_hw_min)
15105 			hw_boost_delay = rack_enobuf_hw_min;
15106 		slot += hw_boost_delay;
15107 	}
15108 	return (slot);
15109 }
15110 
15111 static void
15112 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
15113     tcp_seq startseq, uint32_t sb_offset)
15114 {
15115 	struct rack_sendmap *my_rsm = NULL;
15116 	struct rack_sendmap fe;
15117 
15118 	if (tp->t_state < TCPS_ESTABLISHED) {
15119 		/*
15120 		 * We don't start any measurements if we are
15121 		 * not at least established.
15122 		 */
15123 		return;
15124 	}
15125 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
15126 		/*
15127 		 * We will get no more data into the SB
15128 		 * this means we need to have the data available
15129 		 * before we start a measurement.
15130 		 */
15131 
15132 		if (sbavail(&tptosocket(tp)->so_snd) <
15133 		    max(rc_init_window(rack),
15134 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
15135 			/* Nope not enough data */
15136 			return;
15137 		}
15138 	}
15139 	tp->t_flags |= TF_GPUTINPROG;
15140 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
15141 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
15142 	tp->gput_seq = startseq;
15143 	rack->app_limited_needs_set = 0;
15144 	if (rack->in_probe_rtt)
15145 		rack->measure_saw_probe_rtt = 1;
15146 	else if ((rack->measure_saw_probe_rtt) &&
15147 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
15148 		rack->measure_saw_probe_rtt = 0;
15149 	if (rack->rc_gp_filled)
15150 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15151 	else {
15152 		/* Special case initial measurement */
15153 		struct timeval tv;
15154 
15155 		tp->gput_ts = tcp_get_usecs(&tv);
15156 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15157 	}
15158 	/*
15159 	 * We take a guess out into the future,
15160 	 * if we have no measurement and no
15161 	 * initial rate, we measure the first
15162 	 * initial-windows worth of data to
15163 	 * speed up getting some GP measurement and
15164 	 * thus start pacing.
15165 	 */
15166 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
15167 		rack->app_limited_needs_set = 1;
15168 		tp->gput_ack = startseq + max(rc_init_window(rack),
15169 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
15170 		rack_log_pacing_delay_calc(rack,
15171 					   tp->gput_seq,
15172 					   tp->gput_ack,
15173 					   0,
15174 					   tp->gput_ts,
15175 					   rack->r_ctl.rc_app_limited_cnt,
15176 					   9,
15177 					   __LINE__, NULL, 0);
15178 		return;
15179 	}
15180 	if (sb_offset) {
15181 		/*
15182 		 * We are out somewhere in the sb
15183 		 * can we use the already outstanding data?
15184 		 */
15185 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
15186 			/*
15187 			 * Yes first one is good and in this case
15188 			 * the tp->gput_ts is correctly set based on
15189 			 * the last ack that arrived (no need to
15190 			 * set things up when an ack comes in).
15191 			 */
15192 			my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
15193 			if ((my_rsm == NULL) ||
15194 			    (my_rsm->r_rtr_cnt != 1)) {
15195 				/* retransmission? */
15196 				goto use_latest;
15197 			}
15198 		} else {
15199 			if (rack->r_ctl.rc_first_appl == NULL) {
15200 				/*
15201 				 * If rc_first_appl is NULL
15202 				 * then the cnt should be 0.
15203 				 * This is probably an error, maybe
15204 				 * a KASSERT would be approprate.
15205 				 */
15206 				goto use_latest;
15207 			}
15208 			/*
15209 			 * If we have a marker pointer to the last one that is
15210 			 * app limited we can use that, but we need to set
15211 			 * things up so that when it gets ack'ed we record
15212 			 * the ack time (if its not already acked).
15213 			 */
15214 			rack->app_limited_needs_set = 1;
15215 			/*
15216 			 * We want to get to the rsm that is either
15217 			 * next with space i.e. over 1 MSS or the one
15218 			 * after that (after the app-limited).
15219 			 */
15220 			my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15221 					 rack->r_ctl.rc_first_appl);
15222 			if (my_rsm) {
15223 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
15224 					/* Have to use the next one */
15225 					my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15226 							 my_rsm);
15227 				else {
15228 					/* Use after the first MSS of it is acked */
15229 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
15230 					goto start_set;
15231 				}
15232 			}
15233 			if ((my_rsm == NULL) ||
15234 			    (my_rsm->r_rtr_cnt != 1)) {
15235 				/*
15236 				 * Either its a retransmit or
15237 				 * the last is the app-limited one.
15238 				 */
15239 				goto use_latest;
15240 			}
15241 		}
15242 		tp->gput_seq = my_rsm->r_start;
15243 start_set:
15244 		if (my_rsm->r_flags & RACK_ACKED) {
15245 			/*
15246 			 * This one has been acked use the arrival ack time
15247 			 */
15248 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15249 			rack->app_limited_needs_set = 0;
15250 		}
15251 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15252 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
15253 		rack_log_pacing_delay_calc(rack,
15254 					   tp->gput_seq,
15255 					   tp->gput_ack,
15256 					   (uint64_t)my_rsm,
15257 					   tp->gput_ts,
15258 					   rack->r_ctl.rc_app_limited_cnt,
15259 					   9,
15260 					   __LINE__, NULL, 0);
15261 		return;
15262 	}
15263 
15264 use_latest:
15265 	/*
15266 	 * We don't know how long we may have been
15267 	 * idle or if this is the first-send. Lets
15268 	 * setup the flag so we will trim off
15269 	 * the first ack'd data so we get a true
15270 	 * measurement.
15271 	 */
15272 	rack->app_limited_needs_set = 1;
15273 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
15274 	/* Find this guy so we can pull the send time */
15275 	fe.r_start = startseq;
15276 	my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
15277 	if (my_rsm) {
15278 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15279 		if (my_rsm->r_flags & RACK_ACKED) {
15280 			/*
15281 			 * Unlikely since its probably what was
15282 			 * just transmitted (but I am paranoid).
15283 			 */
15284 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15285 			rack->app_limited_needs_set = 0;
15286 		}
15287 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
15288 			/* This also is unlikely */
15289 			tp->gput_seq = my_rsm->r_start;
15290 		}
15291 	} else {
15292 		/*
15293 		 * TSNH unless we have some send-map limit,
15294 		 * and even at that it should not be hitting
15295 		 * that limit (we should have stopped sending).
15296 		 */
15297 		struct timeval tv;
15298 
15299 		microuptime(&tv);
15300 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15301 	}
15302 	rack_log_pacing_delay_calc(rack,
15303 				   tp->gput_seq,
15304 				   tp->gput_ack,
15305 				   (uint64_t)my_rsm,
15306 				   tp->gput_ts,
15307 				   rack->r_ctl.rc_app_limited_cnt,
15308 				   9, __LINE__, NULL, 0);
15309 }
15310 
15311 static inline uint32_t
15312 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
15313     uint32_t avail, int32_t sb_offset)
15314 {
15315 	uint32_t len;
15316 	uint32_t sendwin;
15317 
15318 	if (tp->snd_wnd > cwnd_to_use)
15319 		sendwin = cwnd_to_use;
15320 	else
15321 		sendwin = tp->snd_wnd;
15322 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
15323 		/* We never want to go over our peers rcv-window */
15324 		len = 0;
15325 	} else {
15326 		uint32_t flight;
15327 
15328 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15329 		if (flight >= sendwin) {
15330 			/*
15331 			 * We have in flight what we are allowed by cwnd (if
15332 			 * it was rwnd blocking it would have hit above out
15333 			 * >= tp->snd_wnd).
15334 			 */
15335 			return (0);
15336 		}
15337 		len = sendwin - flight;
15338 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
15339 			/* We would send too much (beyond the rwnd) */
15340 			len = tp->snd_wnd - ctf_outstanding(tp);
15341 		}
15342 		if ((len + sb_offset) > avail) {
15343 			/*
15344 			 * We don't have that much in the SB, how much is
15345 			 * there?
15346 			 */
15347 			len = avail - sb_offset;
15348 		}
15349 	}
15350 	return (len);
15351 }
15352 
15353 static void
15354 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
15355 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
15356 	     int rsm_is_null, int optlen, int line, uint16_t mode)
15357 {
15358 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15359 		union tcp_log_stackspecific log;
15360 		struct timeval tv;
15361 
15362 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15363 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15364 		log.u_bbr.flex1 = error;
15365 		log.u_bbr.flex2 = flags;
15366 		log.u_bbr.flex3 = rsm_is_null;
15367 		log.u_bbr.flex4 = ipoptlen;
15368 		log.u_bbr.flex5 = tp->rcv_numsacks;
15369 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15370 		log.u_bbr.flex7 = optlen;
15371 		log.u_bbr.flex8 = rack->r_fsb_inited;
15372 		log.u_bbr.applimited = rack->r_fast_output;
15373 		log.u_bbr.bw_inuse = rack_get_bw(rack);
15374 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15375 		log.u_bbr.cwnd_gain = mode;
15376 		log.u_bbr.pkts_out = orig_len;
15377 		log.u_bbr.lt_epoch = len;
15378 		log.u_bbr.delivered = line;
15379 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15380 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15381 		tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
15382 			       len, &log, false, NULL, NULL, 0, &tv);
15383 	}
15384 }
15385 
15386 
15387 static struct mbuf *
15388 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
15389 		   struct rack_fast_send_blk *fsb,
15390 		   int32_t seglimit, int32_t segsize, int hw_tls)
15391 {
15392 #ifdef KERN_TLS
15393 	struct ktls_session *tls, *ntls;
15394 #ifdef INVARIANTS
15395 	struct mbuf *start;
15396 #endif
15397 #endif
15398 	struct mbuf *m, *n, **np, *smb;
15399 	struct mbuf *top;
15400 	int32_t off, soff;
15401 	int32_t len = *plen;
15402 	int32_t fragsize;
15403 	int32_t len_cp = 0;
15404 	uint32_t mlen, frags;
15405 
15406 	soff = off = the_off;
15407 	smb = m = the_m;
15408 	np = &top;
15409 	top = NULL;
15410 #ifdef KERN_TLS
15411 	if (hw_tls && (m->m_flags & M_EXTPG))
15412 		tls = m->m_epg_tls;
15413 	else
15414 		tls = NULL;
15415 #ifdef INVARIANTS
15416 	start = m;
15417 #endif
15418 #endif
15419 	while (len > 0) {
15420 		if (m == NULL) {
15421 			*plen = len_cp;
15422 			break;
15423 		}
15424 #ifdef KERN_TLS
15425 		if (hw_tls) {
15426 			if (m->m_flags & M_EXTPG)
15427 				ntls = m->m_epg_tls;
15428 			else
15429 				ntls = NULL;
15430 
15431 			/*
15432 			 * Avoid mixing TLS records with handshake
15433 			 * data or TLS records from different
15434 			 * sessions.
15435 			 */
15436 			if (tls != ntls) {
15437 				MPASS(m != start);
15438 				*plen = len_cp;
15439 				break;
15440 			}
15441 		}
15442 #endif
15443 		mlen = min(len, m->m_len - off);
15444 		if (seglimit) {
15445 			/*
15446 			 * For M_EXTPG mbufs, add 3 segments
15447 			 * + 1 in case we are crossing page boundaries
15448 			 * + 2 in case the TLS hdr/trailer are used
15449 			 * It is cheaper to just add the segments
15450 			 * than it is to take the cache miss to look
15451 			 * at the mbuf ext_pgs state in detail.
15452 			 */
15453 			if (m->m_flags & M_EXTPG) {
15454 				fragsize = min(segsize, PAGE_SIZE);
15455 				frags = 3;
15456 			} else {
15457 				fragsize = segsize;
15458 				frags = 0;
15459 			}
15460 
15461 			/* Break if we really can't fit anymore. */
15462 			if ((frags + 1) >= seglimit) {
15463 				*plen =	len_cp;
15464 				break;
15465 			}
15466 
15467 			/*
15468 			 * Reduce size if you can't copy the whole
15469 			 * mbuf. If we can't copy the whole mbuf, also
15470 			 * adjust len so the loop will end after this
15471 			 * mbuf.
15472 			 */
15473 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
15474 				mlen = (seglimit - frags - 1) * fragsize;
15475 				len = mlen;
15476 				*plen = len_cp + len;
15477 			}
15478 			frags += howmany(mlen, fragsize);
15479 			if (frags == 0)
15480 				frags++;
15481 			seglimit -= frags;
15482 			KASSERT(seglimit > 0,
15483 			    ("%s: seglimit went too low", __func__));
15484 		}
15485 		n = m_get(M_NOWAIT, m->m_type);
15486 		*np = n;
15487 		if (n == NULL)
15488 			goto nospace;
15489 		n->m_len = mlen;
15490 		soff += mlen;
15491 		len_cp += n->m_len;
15492 		if (m->m_flags & (M_EXT|M_EXTPG)) {
15493 			n->m_data = m->m_data + off;
15494 			mb_dupcl(n, m);
15495 		} else {
15496 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
15497 			    (u_int)n->m_len);
15498 		}
15499 		len -= n->m_len;
15500 		off = 0;
15501 		m = m->m_next;
15502 		np = &n->m_next;
15503 		if (len || (soff == smb->m_len)) {
15504 			/*
15505 			 * We have more so we move forward  or
15506 			 * we have consumed the entire mbuf and
15507 			 * len has fell to 0.
15508 			 */
15509 			soff = 0;
15510 			smb = m;
15511 		}
15512 
15513 	}
15514 	if (fsb != NULL) {
15515 		fsb->m = smb;
15516 		fsb->off = soff;
15517 		if (smb) {
15518 			/*
15519 			 * Save off the size of the mbuf. We do
15520 			 * this so that we can recognize when it
15521 			 * has been trimmed by sbcut() as acks
15522 			 * come in.
15523 			 */
15524 			fsb->o_m_len = smb->m_len;
15525 		} else {
15526 			/*
15527 			 * This is the case where the next mbuf went to NULL. This
15528 			 * means with this copy we have sent everything in the sb.
15529 			 * In theory we could clear the fast_output flag, but lets
15530 			 * not since its possible that we could get more added
15531 			 * and acks that call the extend function which would let
15532 			 * us send more.
15533 			 */
15534 			fsb->o_m_len = 0;
15535 		}
15536 	}
15537 	return (top);
15538 nospace:
15539 	if (top)
15540 		m_freem(top);
15541 	return (NULL);
15542 
15543 }
15544 
15545 /*
15546  * This is a copy of m_copym(), taking the TSO segment size/limit
15547  * constraints into account, and advancing the sndptr as it goes.
15548  */
15549 static struct mbuf *
15550 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15551 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15552 {
15553 	struct mbuf *m, *n;
15554 	int32_t soff;
15555 
15556 	soff = rack->r_ctl.fsb.off;
15557 	m = rack->r_ctl.fsb.m;
15558 	if (rack->r_ctl.fsb.o_m_len > m->m_len) {
15559 		/*
15560 		 * The mbuf had the front of it chopped off by an ack
15561 		 * we need to adjust the soff/off by that difference.
15562 		 */
15563 		uint32_t delta;
15564 
15565 		delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15566 		soff -= delta;
15567 	} else if (rack->r_ctl.fsb.o_m_len < m->m_len) {
15568 		/*
15569 		 * The mbuf was expanded probably by
15570 		 * a m_compress. Just update o_m_len.
15571 		 */
15572 		rack->r_ctl.fsb.o_m_len = m->m_len;
15573 	}
15574 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15575 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15576 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15577 				 __FUNCTION__,
15578 				 rack, *plen, m, m->m_len));
15579 	/* Save off the right location before we copy and advance */
15580 	*s_soff = soff;
15581 	*s_mb = rack->r_ctl.fsb.m;
15582 	n = rack_fo_base_copym(m, soff, plen,
15583 			       &rack->r_ctl.fsb,
15584 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
15585 	return (n);
15586 }
15587 
15588 static int
15589 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15590 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
15591 {
15592 	/*
15593 	 * Enter the fast retransmit path. We are given that a sched_pin is
15594 	 * in place (if accounting is compliled in) and the cycle count taken
15595 	 * at the entry is in the ts_val. The concept her is that the rsm
15596 	 * now holds the mbuf offsets and such so we can directly transmit
15597 	 * without a lot of overhead, the len field is already set for
15598 	 * us to prohibit us from sending too much (usually its 1MSS).
15599 	 */
15600 	struct ip *ip = NULL;
15601 	struct udphdr *udp = NULL;
15602 	struct tcphdr *th = NULL;
15603 	struct mbuf *m = NULL;
15604 	struct inpcb *inp;
15605 	uint8_t *cpto;
15606 	struct tcp_log_buffer *lgb;
15607 #ifdef TCP_ACCOUNTING
15608 	uint64_t crtsc;
15609 	int cnt_thru = 1;
15610 #endif
15611 	struct tcpopt to;
15612 	u_char opt[TCP_MAXOLEN];
15613 	uint32_t hdrlen, optlen;
15614 	int32_t slot, segsiz, max_val, tso = 0, error, ulen = 0;
15615 	uint16_t flags;
15616 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
15617 	uint32_t if_hw_tsomaxsegsize;
15618 
15619 #ifdef INET6
15620 	struct ip6_hdr *ip6 = NULL;
15621 
15622 	if (rack->r_is_v6) {
15623 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15624 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15625 	} else
15626 #endif				/* INET6 */
15627 	{
15628 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15629 		hdrlen = sizeof(struct tcpiphdr);
15630 	}
15631 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15632 		goto failed;
15633 	}
15634 	if (doing_tlp) {
15635 		/* Its a TLP add the flag, it may already be there but be sure */
15636 		rsm->r_flags |= RACK_TLP;
15637 	} else {
15638 		/* If it was a TLP it is not not on this retransmit */
15639 		rsm->r_flags &= ~RACK_TLP;
15640 	}
15641 	startseq = rsm->r_start;
15642 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15643 	inp = rack->rc_inp;
15644 	to.to_flags = 0;
15645 	flags = tcp_outflags[tp->t_state];
15646 	if (flags & (TH_SYN|TH_RST)) {
15647 		goto failed;
15648 	}
15649 	if (rsm->r_flags & RACK_HAS_FIN) {
15650 		/* We can't send a FIN here */
15651 		goto failed;
15652 	}
15653 	if (flags & TH_FIN) {
15654 		/* We never send a FIN */
15655 		flags &= ~TH_FIN;
15656 	}
15657 	if (tp->t_flags & TF_RCVD_TSTMP) {
15658 		to.to_tsval = ms_cts + tp->ts_offset;
15659 		to.to_tsecr = tp->ts_recent;
15660 		to.to_flags = TOF_TS;
15661 	}
15662 	optlen = tcp_addoptions(&to, opt);
15663 	hdrlen += optlen;
15664 	udp = rack->r_ctl.fsb.udp;
15665 	if (udp)
15666 		hdrlen += sizeof(struct udphdr);
15667 	if (rack->r_ctl.rc_pace_max_segs)
15668 		max_val = rack->r_ctl.rc_pace_max_segs;
15669 	else if (rack->rc_user_set_max_segs)
15670 		max_val = rack->rc_user_set_max_segs * segsiz;
15671 	else
15672 		max_val = len;
15673 	if ((tp->t_flags & TF_TSO) &&
15674 	    V_tcp_do_tso &&
15675 	    (len > segsiz) &&
15676 	    (tp->t_port == 0))
15677 		tso = 1;
15678 #ifdef INET6
15679 	if (MHLEN < hdrlen + max_linkhdr)
15680 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15681 	else
15682 #endif
15683 		m = m_gethdr(M_NOWAIT, MT_DATA);
15684 	if (m == NULL)
15685 		goto failed;
15686 	m->m_data += max_linkhdr;
15687 	m->m_len = hdrlen;
15688 	th = rack->r_ctl.fsb.th;
15689 	/* Establish the len to send */
15690 	if (len > max_val)
15691 		len = max_val;
15692 	if ((tso) && (len + optlen > tp->t_maxseg)) {
15693 		uint32_t if_hw_tsomax;
15694 		int32_t max_len;
15695 
15696 		/* extract TSO information */
15697 		if_hw_tsomax = tp->t_tsomax;
15698 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15699 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15700 		/*
15701 		 * Check if we should limit by maximum payload
15702 		 * length:
15703 		 */
15704 		if (if_hw_tsomax != 0) {
15705 			/* compute maximum TSO length */
15706 			max_len = (if_hw_tsomax - hdrlen -
15707 				   max_linkhdr);
15708 			if (max_len <= 0) {
15709 				goto failed;
15710 			} else if (len > max_len) {
15711 				len = max_len;
15712 			}
15713 		}
15714 		if (len <= segsiz) {
15715 			/*
15716 			 * In case there are too many small fragments don't
15717 			 * use TSO:
15718 			 */
15719 			tso = 0;
15720 		}
15721 	} else {
15722 		tso = 0;
15723 	}
15724 	if ((tso == 0) && (len > segsiz))
15725 		len = segsiz;
15726 	if ((len == 0) ||
15727 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
15728 		goto failed;
15729 	}
15730 	th->th_seq = htonl(rsm->r_start);
15731 	th->th_ack = htonl(tp->rcv_nxt);
15732 	/*
15733 	 * The PUSH bit should only be applied
15734 	 * if the full retransmission is made. If
15735 	 * we are sending less than this is the
15736 	 * left hand edge and should not have
15737 	 * the PUSH bit.
15738 	 */
15739 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
15740 	    (len == (rsm->r_end - rsm->r_start)))
15741 		flags |= TH_PUSH;
15742 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15743 	if (th->th_win == 0) {
15744 		tp->t_sndzerowin++;
15745 		tp->t_flags |= TF_RXWIN0SENT;
15746 	} else
15747 		tp->t_flags &= ~TF_RXWIN0SENT;
15748 	if (rsm->r_flags & RACK_TLP) {
15749 		/*
15750 		 * TLP should not count in retran count, but
15751 		 * in its own bin
15752 		 */
15753 		counter_u64_add(rack_tlp_retran, 1);
15754 		counter_u64_add(rack_tlp_retran_bytes, len);
15755 	} else {
15756 		tp->t_sndrexmitpack++;
15757 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15758 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15759 	}
15760 #ifdef STATS
15761 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15762 				 len);
15763 #endif
15764 	if (rsm->m == NULL)
15765 		goto failed;
15766 	if (rsm->orig_m_len != rsm->m->m_len) {
15767 		/* Fix up the orig_m_len and possibly the mbuf offset */
15768 		rack_adjust_orig_mlen(rsm);
15769 	}
15770 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
15771 	if (len <= segsiz) {
15772 		/*
15773 		 * Must have ran out of mbufs for the copy
15774 		 * shorten it to no longer need tso. Lets
15775 		 * not put on sendalot since we are low on
15776 		 * mbufs.
15777 		 */
15778 		tso = 0;
15779 	}
15780 	if ((m->m_next == NULL) || (len <= 0)){
15781 		goto failed;
15782 	}
15783 	if (udp) {
15784 		if (rack->r_is_v6)
15785 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
15786 		else
15787 			ulen = hdrlen + len - sizeof(struct ip);
15788 		udp->uh_ulen = htons(ulen);
15789 	}
15790 	m->m_pkthdr.rcvif = (struct ifnet *)0;
15791 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
15792 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
15793 		int ect = tcp_ecn_output_established(tp, &flags, len, true);
15794 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
15795 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
15796 		    tp->t_flags2 &= ~TF2_ECN_SND_ECE;
15797 #ifdef INET6
15798 		if (rack->r_is_v6) {
15799 		    ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
15800 		    ip6->ip6_flow |= htonl(ect << 20);
15801 		}
15802 		else
15803 #endif
15804 		{
15805 		    ip->ip_tos &= ~IPTOS_ECN_MASK;
15806 		    ip->ip_tos |= ect;
15807 		}
15808 	}
15809 	tcp_set_flags(th, flags);
15810 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
15811 #ifdef INET6
15812 	if (rack->r_is_v6) {
15813 		if (tp->t_port) {
15814 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15815 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15816 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15817 			th->th_sum = htons(0);
15818 			UDPSTAT_INC(udps_opackets);
15819 		} else {
15820 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15821 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15822 			th->th_sum = in6_cksum_pseudo(ip6,
15823 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15824 						      0);
15825 		}
15826 	}
15827 #endif
15828 #if defined(INET6) && defined(INET)
15829 	else
15830 #endif
15831 #ifdef INET
15832 	{
15833 		if (tp->t_port) {
15834 			m->m_pkthdr.csum_flags = CSUM_UDP;
15835 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15836 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15837 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15838 			th->th_sum = htons(0);
15839 			UDPSTAT_INC(udps_opackets);
15840 		} else {
15841 			m->m_pkthdr.csum_flags = CSUM_TCP;
15842 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15843 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
15844 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15845 									IPPROTO_TCP + len + optlen));
15846 		}
15847 		/* IP version must be set here for ipv4/ipv6 checking later */
15848 		KASSERT(ip->ip_v == IPVERSION,
15849 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
15850 	}
15851 #endif
15852 	if (tso) {
15853 		KASSERT(len > tp->t_maxseg - optlen,
15854 			("%s: len <= tso_segsz tp:%p", __func__, tp));
15855 		m->m_pkthdr.csum_flags |= CSUM_TSO;
15856 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15857 	}
15858 #ifdef INET6
15859 	if (rack->r_is_v6) {
15860 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15861 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15862 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15863 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15864 		else
15865 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15866 	}
15867 #endif
15868 #if defined(INET) && defined(INET6)
15869 	else
15870 #endif
15871 #ifdef INET
15872 	{
15873 		ip->ip_len = htons(m->m_pkthdr.len);
15874 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15875 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15876 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15877 			if (tp->t_port == 0 || len < V_tcp_minmss) {
15878 				ip->ip_off |= htons(IP_DF);
15879 			}
15880 		} else {
15881 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15882 		}
15883 	}
15884 #endif
15885 	/* Time to copy in our header */
15886 	cpto = mtod(m, uint8_t *);
15887 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15888 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15889 	if (optlen) {
15890 		bcopy(opt, th + 1, optlen);
15891 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15892 	} else {
15893 		th->th_off = sizeof(struct tcphdr) >> 2;
15894 	}
15895 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15896 		union tcp_log_stackspecific log;
15897 
15898 		if (rsm->r_flags & RACK_RWND_COLLAPSED) {
15899 			rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
15900 			counter_u64_add(rack_collapsed_win_rxt, 1);
15901 			counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
15902 		}
15903 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15904 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15905 		if (rack->rack_no_prr)
15906 			log.u_bbr.flex1 = 0;
15907 		else
15908 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15909 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15910 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15911 		log.u_bbr.flex4 = max_val;
15912 		log.u_bbr.flex5 = 0;
15913 		/* Save off the early/late values */
15914 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15915 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15916 		log.u_bbr.bw_inuse = rack_get_bw(rack);
15917 		if (doing_tlp == 0)
15918 			log.u_bbr.flex8 = 1;
15919 		else
15920 			log.u_bbr.flex8 = 2;
15921 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15922 		log.u_bbr.flex7 = 55;
15923 		log.u_bbr.pkts_out = tp->t_maxseg;
15924 		log.u_bbr.timeStamp = cts;
15925 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15926 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15927 		log.u_bbr.delivered = 0;
15928 		lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15929 				     len, &log, false, NULL, NULL, 0, tv);
15930 	} else
15931 		lgb = NULL;
15932 #ifdef INET6
15933 	if (rack->r_is_v6) {
15934 		error = ip6_output(m, NULL,
15935 				   &inp->inp_route6,
15936 				   0, NULL, NULL, inp);
15937 	}
15938 #endif
15939 #if defined(INET) && defined(INET6)
15940 	else
15941 #endif
15942 #ifdef INET
15943 	{
15944 		error = ip_output(m, NULL,
15945 				  &inp->inp_route,
15946 				  0, 0, inp);
15947 	}
15948 #endif
15949 	m = NULL;
15950 	if (lgb) {
15951 		lgb->tlb_errno = error;
15952 		lgb = NULL;
15953 	}
15954 	if (error) {
15955 		goto failed;
15956 	}
15957 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
15958 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls);
15959 	if (doing_tlp && (rack->fast_rsm_hack == 0)) {
15960 		rack->rc_tlp_in_progress = 1;
15961 		rack->r_ctl.rc_tlp_cnt_out++;
15962 	}
15963 	if (error == 0) {
15964 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
15965 		if (doing_tlp) {
15966 			rack->rc_last_sent_tlp_past_cumack = 0;
15967 			rack->rc_last_sent_tlp_seq_valid = 1;
15968 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
15969 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
15970 		}
15971 	}
15972 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15973 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
15974 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
15975 		rack->r_ctl.retran_during_recovery += len;
15976 	{
15977 		int idx;
15978 
15979 		idx = (len / segsiz) + 3;
15980 		if (idx >= TCP_MSS_ACCT_ATIMER)
15981 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15982 		else
15983 			counter_u64_add(rack_out_size[idx], 1);
15984 	}
15985 	if (tp->t_rtttime == 0) {
15986 		tp->t_rtttime = ticks;
15987 		tp->t_rtseq = startseq;
15988 		KMOD_TCPSTAT_INC(tcps_segstimed);
15989 	}
15990 	counter_u64_add(rack_fto_rsm_send, 1);
15991 	if (error && (error == ENOBUFS)) {
15992 		if (rack->r_ctl.crte != NULL) {
15993 			rack_trace_point(rack, RACK_TP_HWENOBUF);
15994 		} else
15995 			rack_trace_point(rack, RACK_TP_ENOBUF);
15996 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
15997 		if (rack->rc_enobuf < 0x7f)
15998 			rack->rc_enobuf++;
15999 		if (slot < (10 * HPTS_USEC_IN_MSEC))
16000 			slot = 10 * HPTS_USEC_IN_MSEC;
16001 	} else
16002 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
16003 	if ((slot == 0) ||
16004 	    (rack->rc_always_pace == 0) ||
16005 	    (rack->r_rr_config == 1)) {
16006 		/*
16007 		 * We have no pacing set or we
16008 		 * are using old-style rack or
16009 		 * we are overridden to use the old 1ms pacing.
16010 		 */
16011 		slot = rack->r_ctl.rc_min_to;
16012 	}
16013 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
16014 #ifdef TCP_ACCOUNTING
16015 	crtsc = get_cyclecount();
16016 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16017 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16018 	}
16019 	counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16020 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16021 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16022 	}
16023 	counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16024 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16025 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
16026 	}
16027 	counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
16028 	sched_unpin();
16029 #endif
16030 	return (0);
16031 failed:
16032 	if (m)
16033 		m_free(m);
16034 	return (-1);
16035 }
16036 
16037 static void
16038 rack_sndbuf_autoscale(struct tcp_rack *rack)
16039 {
16040 	/*
16041 	 * Automatic sizing of send socket buffer.  Often the send buffer
16042 	 * size is not optimally adjusted to the actual network conditions
16043 	 * at hand (delay bandwidth product).  Setting the buffer size too
16044 	 * small limits throughput on links with high bandwidth and high
16045 	 * delay (eg. trans-continental/oceanic links).  Setting the
16046 	 * buffer size too big consumes too much real kernel memory,
16047 	 * especially with many connections on busy servers.
16048 	 *
16049 	 * The criteria to step up the send buffer one notch are:
16050 	 *  1. receive window of remote host is larger than send buffer
16051 	 *     (with a fudge factor of 5/4th);
16052 	 *  2. send buffer is filled to 7/8th with data (so we actually
16053 	 *     have data to make use of it);
16054 	 *  3. send buffer fill has not hit maximal automatic size;
16055 	 *  4. our send window (slow start and cogestion controlled) is
16056 	 *     larger than sent but unacknowledged data in send buffer.
16057 	 *
16058 	 * Note that the rack version moves things much faster since
16059 	 * we want to avoid hitting cache lines in the rack_fast_output()
16060 	 * path so this is called much less often and thus moves
16061 	 * the SB forward by a percentage.
16062 	 */
16063 	struct socket *so;
16064 	struct tcpcb *tp;
16065 	uint32_t sendwin, scaleup;
16066 
16067 	tp = rack->rc_tp;
16068 	so = rack->rc_inp->inp_socket;
16069 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
16070 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
16071 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
16072 		    sbused(&so->so_snd) >=
16073 		    (so->so_snd.sb_hiwat / 8 * 7) &&
16074 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
16075 		    sendwin >= (sbused(&so->so_snd) -
16076 		    (tp->snd_nxt - tp->snd_una))) {
16077 			if (rack_autosndbuf_inc)
16078 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
16079 			else
16080 				scaleup = V_tcp_autosndbuf_inc;
16081 			if (scaleup < V_tcp_autosndbuf_inc)
16082 				scaleup = V_tcp_autosndbuf_inc;
16083 			scaleup += so->so_snd.sb_hiwat;
16084 			if (scaleup > V_tcp_autosndbuf_max)
16085 				scaleup = V_tcp_autosndbuf_max;
16086 			if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
16087 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
16088 		}
16089 	}
16090 }
16091 
16092 static int
16093 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
16094 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
16095 {
16096 	/*
16097 	 * Enter to do fast output. We are given that the sched_pin is
16098 	 * in place (if accounting is compiled in) and the cycle count taken
16099 	 * at entry is in place in ts_val. The idea here is that
16100 	 * we know how many more bytes needs to be sent (presumably either
16101 	 * during pacing or to fill the cwnd and that was greater than
16102 	 * the max-burst). We have how much to send and all the info we
16103 	 * need to just send.
16104 	 */
16105 	struct ip *ip = NULL;
16106 	struct udphdr *udp = NULL;
16107 	struct tcphdr *th = NULL;
16108 	struct mbuf *m, *s_mb;
16109 	struct inpcb *inp;
16110 	uint8_t *cpto;
16111 	struct tcp_log_buffer *lgb;
16112 #ifdef TCP_ACCOUNTING
16113 	uint64_t crtsc;
16114 #endif
16115 	struct tcpopt to;
16116 	u_char opt[TCP_MAXOLEN];
16117 	uint32_t hdrlen, optlen;
16118 #ifdef TCP_ACCOUNTING
16119 	int cnt_thru = 1;
16120 #endif
16121 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
16122 	uint16_t flags;
16123 	uint32_t s_soff;
16124 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
16125 	uint32_t if_hw_tsomaxsegsize;
16126 	uint16_t add_flag = RACK_SENT_FP;
16127 #ifdef INET6
16128 	struct ip6_hdr *ip6 = NULL;
16129 
16130 	if (rack->r_is_v6) {
16131 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
16132 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
16133 	} else
16134 #endif				/* INET6 */
16135 	{
16136 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
16137 		hdrlen = sizeof(struct tcpiphdr);
16138 	}
16139 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
16140 		m = NULL;
16141 		goto failed;
16142 	}
16143 	startseq = tp->snd_max;
16144 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16145 	inp = rack->rc_inp;
16146 	len = rack->r_ctl.fsb.left_to_send;
16147 	to.to_flags = 0;
16148 	flags = rack->r_ctl.fsb.tcp_flags;
16149 	if (tp->t_flags & TF_RCVD_TSTMP) {
16150 		to.to_tsval = ms_cts + tp->ts_offset;
16151 		to.to_tsecr = tp->ts_recent;
16152 		to.to_flags = TOF_TS;
16153 	}
16154 	optlen = tcp_addoptions(&to, opt);
16155 	hdrlen += optlen;
16156 	udp = rack->r_ctl.fsb.udp;
16157 	if (udp)
16158 		hdrlen += sizeof(struct udphdr);
16159 	if (rack->r_ctl.rc_pace_max_segs)
16160 		max_val = rack->r_ctl.rc_pace_max_segs;
16161 	else if (rack->rc_user_set_max_segs)
16162 		max_val = rack->rc_user_set_max_segs * segsiz;
16163 	else
16164 		max_val = len;
16165 	if ((tp->t_flags & TF_TSO) &&
16166 	    V_tcp_do_tso &&
16167 	    (len > segsiz) &&
16168 	    (tp->t_port == 0))
16169 		tso = 1;
16170 again:
16171 #ifdef INET6
16172 	if (MHLEN < hdrlen + max_linkhdr)
16173 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
16174 	else
16175 #endif
16176 		m = m_gethdr(M_NOWAIT, MT_DATA);
16177 	if (m == NULL)
16178 		goto failed;
16179 	m->m_data += max_linkhdr;
16180 	m->m_len = hdrlen;
16181 	th = rack->r_ctl.fsb.th;
16182 	/* Establish the len to send */
16183 	if (len > max_val)
16184 		len = max_val;
16185 	if ((tso) && (len + optlen > tp->t_maxseg)) {
16186 		uint32_t if_hw_tsomax;
16187 		int32_t max_len;
16188 
16189 		/* extract TSO information */
16190 		if_hw_tsomax = tp->t_tsomax;
16191 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
16192 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
16193 		/*
16194 		 * Check if we should limit by maximum payload
16195 		 * length:
16196 		 */
16197 		if (if_hw_tsomax != 0) {
16198 			/* compute maximum TSO length */
16199 			max_len = (if_hw_tsomax - hdrlen -
16200 				   max_linkhdr);
16201 			if (max_len <= 0) {
16202 				goto failed;
16203 			} else if (len > max_len) {
16204 				len = max_len;
16205 			}
16206 		}
16207 		if (len <= segsiz) {
16208 			/*
16209 			 * In case there are too many small fragments don't
16210 			 * use TSO:
16211 			 */
16212 			tso = 0;
16213 		}
16214 	} else {
16215 		tso = 0;
16216 	}
16217 	if ((tso == 0) && (len > segsiz))
16218 		len = segsiz;
16219 	if ((len == 0) ||
16220 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
16221 		goto failed;
16222 	}
16223 	sb_offset = tp->snd_max - tp->snd_una;
16224 	th->th_seq = htonl(tp->snd_max);
16225 	th->th_ack = htonl(tp->rcv_nxt);
16226 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
16227 	if (th->th_win == 0) {
16228 		tp->t_sndzerowin++;
16229 		tp->t_flags |= TF_RXWIN0SENT;
16230 	} else
16231 		tp->t_flags &= ~TF_RXWIN0SENT;
16232 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
16233 	KMOD_TCPSTAT_INC(tcps_sndpack);
16234 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
16235 #ifdef STATS
16236 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
16237 				 len);
16238 #endif
16239 	if (rack->r_ctl.fsb.m == NULL)
16240 		goto failed;
16241 
16242 	/* s_mb and s_soff are saved for rack_log_output */
16243 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
16244 				    &s_mb, &s_soff);
16245 	if (len <= segsiz) {
16246 		/*
16247 		 * Must have ran out of mbufs for the copy
16248 		 * shorten it to no longer need tso. Lets
16249 		 * not put on sendalot since we are low on
16250 		 * mbufs.
16251 		 */
16252 		tso = 0;
16253 	}
16254 	if (rack->r_ctl.fsb.rfo_apply_push &&
16255 	    (len == rack->r_ctl.fsb.left_to_send)) {
16256 		flags |= TH_PUSH;
16257 		add_flag |= RACK_HAD_PUSH;
16258 	}
16259 	if ((m->m_next == NULL) || (len <= 0)){
16260 		goto failed;
16261 	}
16262 	if (udp) {
16263 		if (rack->r_is_v6)
16264 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
16265 		else
16266 			ulen = hdrlen + len - sizeof(struct ip);
16267 		udp->uh_ulen = htons(ulen);
16268 	}
16269 	m->m_pkthdr.rcvif = (struct ifnet *)0;
16270 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
16271 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
16272 		int ect = tcp_ecn_output_established(tp, &flags, len, false);
16273 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
16274 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
16275 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
16276 #ifdef INET6
16277 		if (rack->r_is_v6) {
16278 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
16279 			ip6->ip6_flow |= htonl(ect << 20);
16280 		}
16281 		else
16282 #endif
16283 		{
16284 			ip->ip_tos &= ~IPTOS_ECN_MASK;
16285 			ip->ip_tos |= ect;
16286 		}
16287 	}
16288 	tcp_set_flags(th, flags);
16289 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
16290 #ifdef INET6
16291 	if (rack->r_is_v6) {
16292 		if (tp->t_port) {
16293 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
16294 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16295 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
16296 			th->th_sum = htons(0);
16297 			UDPSTAT_INC(udps_opackets);
16298 		} else {
16299 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
16300 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16301 			th->th_sum = in6_cksum_pseudo(ip6,
16302 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
16303 						      0);
16304 		}
16305 	}
16306 #endif
16307 #if defined(INET6) && defined(INET)
16308 	else
16309 #endif
16310 #ifdef INET
16311 	{
16312 		if (tp->t_port) {
16313 			m->m_pkthdr.csum_flags = CSUM_UDP;
16314 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16315 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
16316 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
16317 			th->th_sum = htons(0);
16318 			UDPSTAT_INC(udps_opackets);
16319 		} else {
16320 			m->m_pkthdr.csum_flags = CSUM_TCP;
16321 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16322 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
16323 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
16324 									IPPROTO_TCP + len + optlen));
16325 		}
16326 		/* IP version must be set here for ipv4/ipv6 checking later */
16327 		KASSERT(ip->ip_v == IPVERSION,
16328 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
16329 	}
16330 #endif
16331 	if (tso) {
16332 		KASSERT(len > tp->t_maxseg - optlen,
16333 			("%s: len <= tso_segsz tp:%p", __func__, tp));
16334 		m->m_pkthdr.csum_flags |= CSUM_TSO;
16335 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
16336 	}
16337 #ifdef INET6
16338 	if (rack->r_is_v6) {
16339 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
16340 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
16341 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
16342 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16343 		else
16344 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16345 	}
16346 #endif
16347 #if defined(INET) && defined(INET6)
16348 	else
16349 #endif
16350 #ifdef INET
16351 	{
16352 		ip->ip_len = htons(m->m_pkthdr.len);
16353 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
16354 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
16355 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16356 			if (tp->t_port == 0 || len < V_tcp_minmss) {
16357 				ip->ip_off |= htons(IP_DF);
16358 			}
16359 		} else {
16360 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16361 		}
16362 	}
16363 #endif
16364 	/* Time to copy in our header */
16365 	cpto = mtod(m, uint8_t *);
16366 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
16367 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
16368 	if (optlen) {
16369 		bcopy(opt, th + 1, optlen);
16370 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
16371 	} else {
16372 		th->th_off = sizeof(struct tcphdr) >> 2;
16373 	}
16374 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
16375 		union tcp_log_stackspecific log;
16376 
16377 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16378 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
16379 		if (rack->rack_no_prr)
16380 			log.u_bbr.flex1 = 0;
16381 		else
16382 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16383 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16384 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16385 		log.u_bbr.flex4 = max_val;
16386 		log.u_bbr.flex5 = 0;
16387 		/* Save off the early/late values */
16388 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16389 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16390 		log.u_bbr.bw_inuse = rack_get_bw(rack);
16391 		log.u_bbr.flex8 = 0;
16392 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16393 		log.u_bbr.flex7 = 44;
16394 		log.u_bbr.pkts_out = tp->t_maxseg;
16395 		log.u_bbr.timeStamp = cts;
16396 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16397 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16398 		log.u_bbr.delivered = 0;
16399 		lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16400 				     len, &log, false, NULL, NULL, 0, tv);
16401 	} else
16402 		lgb = NULL;
16403 #ifdef INET6
16404 	if (rack->r_is_v6) {
16405 		error = ip6_output(m, NULL,
16406 				   &inp->inp_route6,
16407 				   0, NULL, NULL, inp);
16408 	}
16409 #endif
16410 #if defined(INET) && defined(INET6)
16411 	else
16412 #endif
16413 #ifdef INET
16414 	{
16415 		error = ip_output(m, NULL,
16416 				  &inp->inp_route,
16417 				  0, 0, inp);
16418 	}
16419 #endif
16420 	if (lgb) {
16421 		lgb->tlb_errno = error;
16422 		lgb = NULL;
16423 	}
16424 	if (error) {
16425 		*send_err = error;
16426 		m = NULL;
16427 		goto failed;
16428 	}
16429 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
16430 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls);
16431 	m = NULL;
16432 	if (tp->snd_una == tp->snd_max) {
16433 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
16434 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
16435 		tp->t_acktime = ticks;
16436 	}
16437 	if (error == 0)
16438 		tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
16439 
16440 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
16441 	tot_len += len;
16442 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
16443 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
16444 	tp->snd_max += len;
16445 	tp->snd_nxt = tp->snd_max;
16446 	{
16447 		int idx;
16448 
16449 		idx = (len / segsiz) + 3;
16450 		if (idx >= TCP_MSS_ACCT_ATIMER)
16451 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16452 		else
16453 			counter_u64_add(rack_out_size[idx], 1);
16454 	}
16455 	if (len <= rack->r_ctl.fsb.left_to_send)
16456 		rack->r_ctl.fsb.left_to_send -= len;
16457 	else
16458 		rack->r_ctl.fsb.left_to_send = 0;
16459 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
16460 		rack->r_fast_output = 0;
16461 		rack->r_ctl.fsb.left_to_send = 0;
16462 		/* At the end of fast_output scale up the sb */
16463 		SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
16464 		rack_sndbuf_autoscale(rack);
16465 		SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
16466 	}
16467 	if (tp->t_rtttime == 0) {
16468 		tp->t_rtttime = ticks;
16469 		tp->t_rtseq = startseq;
16470 		KMOD_TCPSTAT_INC(tcps_segstimed);
16471 	}
16472 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
16473 	    (max_val > len) &&
16474 	    (tso == 0)) {
16475 		max_val -= len;
16476 		len = segsiz;
16477 		th = rack->r_ctl.fsb.th;
16478 #ifdef TCP_ACCOUNTING
16479 		cnt_thru++;
16480 #endif
16481 		goto again;
16482 	}
16483 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16484 	counter_u64_add(rack_fto_send, 1);
16485 	slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
16486 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
16487 #ifdef TCP_ACCOUNTING
16488 	crtsc = get_cyclecount();
16489 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16490 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16491 	}
16492 	counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16493 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16494 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16495 	}
16496 	counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16497 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16498 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
16499 	}
16500 	counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
16501 	sched_unpin();
16502 #endif
16503 	return (0);
16504 failed:
16505 	if (m)
16506 		m_free(m);
16507 	rack->r_fast_output = 0;
16508 	return (-1);
16509 }
16510 
16511 static struct rack_sendmap *
16512 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
16513 {
16514 	struct rack_sendmap *rsm = NULL;
16515 	struct rack_sendmap fe;
16516 	int thresh;
16517 
16518 restart:
16519 	fe.r_start = rack->r_ctl.last_collapse_point;
16520 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
16521 	if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
16522 		/* Nothing, strange turn off validity  */
16523 		rack->r_collapse_point_valid = 0;
16524 		return (NULL);
16525 	}
16526 	/* Can we send it yet? */
16527 	if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
16528 		/*
16529 		 * Receiver window has not grown enough for
16530 		 * the segment to be put on the wire.
16531 		 */
16532 		return (NULL);
16533 	}
16534 	if (rsm->r_flags & RACK_ACKED) {
16535 		/*
16536 		 * It has been sacked, lets move to the
16537 		 * next one if possible.
16538 		 */
16539 		rack->r_ctl.last_collapse_point = rsm->r_end;
16540 		/* Are we done? */
16541 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
16542 			    rack->r_ctl.high_collapse_point)) {
16543 			rack->r_collapse_point_valid = 0;
16544 			return (NULL);
16545 		}
16546 		goto restart;
16547 	}
16548 	/* Now has it been long enough ? */
16549 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts);
16550 	if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
16551 		rack_log_collapse(rack, rsm->r_start,
16552 				  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
16553 				  thresh, __LINE__, 6, rsm->r_flags, rsm);
16554 		return (rsm);
16555 	}
16556 	/* Not enough time */
16557 	rack_log_collapse(rack, rsm->r_start,
16558 			  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
16559 			  thresh, __LINE__, 7, rsm->r_flags, rsm);
16560 	return (NULL);
16561 }
16562 
16563 static int
16564 rack_output(struct tcpcb *tp)
16565 {
16566 	struct socket *so;
16567 	uint32_t recwin;
16568 	uint32_t sb_offset, s_moff = 0;
16569 	int32_t len, error = 0;
16570 	uint16_t flags;
16571 	struct mbuf *m, *s_mb = NULL;
16572 	struct mbuf *mb;
16573 	uint32_t if_hw_tsomaxsegcount = 0;
16574 	uint32_t if_hw_tsomaxsegsize;
16575 	int32_t segsiz, minseg;
16576 	long tot_len_this_send = 0;
16577 #ifdef INET
16578 	struct ip *ip = NULL;
16579 #endif
16580 	struct udphdr *udp = NULL;
16581 	struct tcp_rack *rack;
16582 	struct tcphdr *th;
16583 	uint8_t pass = 0;
16584 	uint8_t mark = 0;
16585 	uint8_t wanted_cookie = 0;
16586 	u_char opt[TCP_MAXOLEN];
16587 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
16588 	uint32_t rack_seq;
16589 
16590 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16591 	unsigned ipsec_optlen = 0;
16592 
16593 #endif
16594 	int32_t idle, sendalot;
16595 	int32_t sub_from_prr = 0;
16596 	volatile int32_t sack_rxmit;
16597 	struct rack_sendmap *rsm = NULL;
16598 	int32_t tso, mtu;
16599 	struct tcpopt to;
16600 	int32_t slot = 0;
16601 	int32_t sup_rack = 0;
16602 	uint32_t cts, ms_cts, delayed, early;
16603 	uint16_t add_flag = RACK_SENT_SP;
16604 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
16605 	uint8_t hpts_calling,  doing_tlp = 0;
16606 	uint32_t cwnd_to_use, pace_max_seg;
16607 	int32_t do_a_prefetch = 0;
16608 	int32_t prefetch_rsm = 0;
16609 	int32_t orig_len = 0;
16610 	struct timeval tv;
16611 	int32_t prefetch_so_done = 0;
16612 	struct tcp_log_buffer *lgb;
16613 	struct inpcb *inp = tptoinpcb(tp);
16614 	struct sockbuf *sb;
16615 	uint64_t ts_val = 0;
16616 #ifdef TCP_ACCOUNTING
16617 	uint64_t crtsc;
16618 #endif
16619 #ifdef INET6
16620 	struct ip6_hdr *ip6 = NULL;
16621 	int32_t isipv6;
16622 #endif
16623 	bool hw_tls = false;
16624 
16625 	NET_EPOCH_ASSERT();
16626 	INP_WLOCK_ASSERT(inp);
16627 
16628 	/* setup and take the cache hits here */
16629 	rack = (struct tcp_rack *)tp->t_fb_ptr;
16630 #ifdef TCP_ACCOUNTING
16631 	sched_pin();
16632 	ts_val = get_cyclecount();
16633 #endif
16634 	hpts_calling = inp->inp_hpts_calls;
16635 #ifdef TCP_OFFLOAD
16636 	if (tp->t_flags & TF_TOE) {
16637 #ifdef TCP_ACCOUNTING
16638 		sched_unpin();
16639 #endif
16640 		return (tcp_offload_output(tp));
16641 	}
16642 #endif
16643 	/*
16644 	 * For TFO connections in SYN_RECEIVED, only allow the initial
16645 	 * SYN|ACK and those sent by the retransmit timer.
16646 	 */
16647 	if (IS_FASTOPEN(tp->t_flags) &&
16648 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
16649 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16650 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16651 #ifdef TCP_ACCOUNTING
16652 		sched_unpin();
16653 #endif
16654 		return (0);
16655 	}
16656 #ifdef INET6
16657 	if (rack->r_state) {
16658 		/* Use the cache line loaded if possible */
16659 		isipv6 = rack->r_is_v6;
16660 	} else {
16661 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16662 	}
16663 #endif
16664 	early = 0;
16665 	cts = tcp_get_usecs(&tv);
16666 	ms_cts = tcp_tv_to_mssectick(&tv);
16667 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16668 	    tcp_in_hpts(rack->rc_inp)) {
16669 		/*
16670 		 * We are on the hpts for some timer but not hptsi output.
16671 		 * Remove from the hpts unconditionally.
16672 		 */
16673 		rack_timer_cancel(tp, rack, cts, __LINE__);
16674 	}
16675 	/* Are we pacing and late? */
16676 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16677 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16678 		/* We are delayed */
16679 		delayed = cts - rack->r_ctl.rc_last_output_to;
16680 	} else {
16681 		delayed = 0;
16682 	}
16683 	/* Do the timers, which may override the pacer */
16684 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16685 		int retval;
16686 
16687 		retval = rack_process_timers(tp, rack, cts, hpts_calling,
16688 		    &doing_tlp);
16689 		if (retval != 0) {
16690 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16691 #ifdef TCP_ACCOUNTING
16692 			sched_unpin();
16693 #endif
16694 			/*
16695 			 * If timers want tcp_drop(), then pass error out,
16696 			 * otherwise suppress it.
16697 			 */
16698 			return (retval < 0 ? retval : 0);
16699 		}
16700 	}
16701 	if (rack->rc_in_persist) {
16702 		if (tcp_in_hpts(rack->rc_inp) == 0) {
16703 			/* Timer is not running */
16704 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16705 		}
16706 #ifdef TCP_ACCOUNTING
16707 		sched_unpin();
16708 #endif
16709 		return (0);
16710 	}
16711 	if ((rack->rc_ack_required == 1) &&
16712 	    (rack->r_timer_override == 0)){
16713 		/* A timeout occurred and no ack has arrived */
16714 		if (tcp_in_hpts(rack->rc_inp) == 0) {
16715 			/* Timer is not running */
16716 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16717 		}
16718 #ifdef TCP_ACCOUNTING
16719 		sched_unpin();
16720 #endif
16721 		return (0);
16722 	}
16723 	if ((rack->r_timer_override) ||
16724 	    (rack->rc_ack_can_sendout_data) ||
16725 	    (delayed) ||
16726 	    (tp->t_state < TCPS_ESTABLISHED)) {
16727 		rack->rc_ack_can_sendout_data = 0;
16728 		if (tcp_in_hpts(rack->rc_inp))
16729 			tcp_hpts_remove(rack->rc_inp);
16730 	} else if (tcp_in_hpts(rack->rc_inp)) {
16731 		/*
16732 		 * On the hpts you can't pass even if ACKNOW is on, we will
16733 		 * when the hpts fires.
16734 		 */
16735 #ifdef TCP_ACCOUNTING
16736 		crtsc = get_cyclecount();
16737 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16738 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16739 		}
16740 		counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16741 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16742 			tp->tcp_cnt_counters[SND_BLOCKED]++;
16743 		}
16744 		counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16745 		sched_unpin();
16746 #endif
16747 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16748 		return (0);
16749 	}
16750 	rack->rc_inp->inp_hpts_calls = 0;
16751 	/* Finish out both pacing early and late accounting */
16752 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16753 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16754 		early = rack->r_ctl.rc_last_output_to - cts;
16755 	} else
16756 		early = 0;
16757 	if (delayed) {
16758 		rack->r_ctl.rc_agg_delayed += delayed;
16759 		rack->r_late = 1;
16760 	} else if (early) {
16761 		rack->r_ctl.rc_agg_early += early;
16762 		rack->r_early = 1;
16763 	}
16764 	/* Now that early/late accounting is done turn off the flag */
16765 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16766 	rack->r_wanted_output = 0;
16767 	rack->r_timer_override = 0;
16768 	if ((tp->t_state != rack->r_state) &&
16769 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
16770 		rack_set_state(tp, rack);
16771 	}
16772 	if ((rack->r_fast_output) &&
16773 	    (doing_tlp == 0) &&
16774 	    (tp->rcv_numsacks == 0)) {
16775 		int ret;
16776 
16777 		error = 0;
16778 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16779 		if (ret >= 0)
16780 			return(ret);
16781 		else if (error) {
16782 			inp = rack->rc_inp;
16783 			so = inp->inp_socket;
16784 			sb = &so->so_snd;
16785 			goto nomore;
16786 		}
16787 	}
16788 	inp = rack->rc_inp;
16789 	/*
16790 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
16791 	 * only allow the initial SYN or SYN|ACK and those sent
16792 	 * by the retransmit timer.
16793 	 */
16794 	if (IS_FASTOPEN(tp->t_flags) &&
16795 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
16796 	     (tp->t_state == TCPS_SYN_SENT)) &&
16797 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16798 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
16799 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16800 		so = inp->inp_socket;
16801 		sb = &so->so_snd;
16802 		goto just_return_nolock;
16803 	}
16804 	/*
16805 	 * Determine length of data that should be transmitted, and flags
16806 	 * that will be used. If there is some data or critical controls
16807 	 * (SYN, RST) to send, then transmit; otherwise, investigate
16808 	 * further.
16809 	 */
16810 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16811 	if (tp->t_idle_reduce) {
16812 		if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16813 			rack_cc_after_idle(rack, tp);
16814 	}
16815 	tp->t_flags &= ~TF_LASTIDLE;
16816 	if (idle) {
16817 		if (tp->t_flags & TF_MORETOCOME) {
16818 			tp->t_flags |= TF_LASTIDLE;
16819 			idle = 0;
16820 		}
16821 	}
16822 	if ((tp->snd_una == tp->snd_max) &&
16823 	    rack->r_ctl.rc_went_idle_time &&
16824 	    TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16825 		idle = cts - rack->r_ctl.rc_went_idle_time;
16826 		if (idle > rack_min_probertt_hold) {
16827 			/* Count as a probe rtt */
16828 			if (rack->in_probe_rtt == 0) {
16829 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
16830 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16831 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16832 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16833 			} else {
16834 				rack_exit_probertt(rack, cts);
16835 			}
16836 		}
16837 		idle = 0;
16838 	}
16839 	if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED))
16840 		rack_init_fsb_block(tp, rack);
16841 again:
16842 	/*
16843 	 * If we've recently taken a timeout, snd_max will be greater than
16844 	 * snd_nxt.  There may be SACK information that allows us to avoid
16845 	 * resending already delivered data.  Adjust snd_nxt accordingly.
16846 	 */
16847 	sendalot = 0;
16848 	cts = tcp_get_usecs(&tv);
16849 	ms_cts = tcp_tv_to_mssectick(&tv);
16850 	tso = 0;
16851 	mtu = 0;
16852 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16853 	minseg = segsiz;
16854 	if (rack->r_ctl.rc_pace_max_segs == 0)
16855 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16856 	else
16857 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16858 	sb_offset = tp->snd_max - tp->snd_una;
16859 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16860 	flags = tcp_outflags[tp->t_state];
16861 	while (rack->rc_free_cnt < rack_free_cache) {
16862 		rsm = rack_alloc(rack);
16863 		if (rsm == NULL) {
16864 			if (inp->inp_hpts_calls)
16865 				/* Retry in a ms */
16866 				slot = (1 * HPTS_USEC_IN_MSEC);
16867 			so = inp->inp_socket;
16868 			sb = &so->so_snd;
16869 			goto just_return_nolock;
16870 		}
16871 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16872 		rack->rc_free_cnt++;
16873 		rsm = NULL;
16874 	}
16875 	if (inp->inp_hpts_calls)
16876 		inp->inp_hpts_calls = 0;
16877 	sack_rxmit = 0;
16878 	len = 0;
16879 	rsm = NULL;
16880 	if (flags & TH_RST) {
16881 		SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16882 		so = inp->inp_socket;
16883 		sb = &so->so_snd;
16884 		goto send;
16885 	}
16886 	if (rack->r_ctl.rc_resend) {
16887 		/* Retransmit timer */
16888 		rsm = rack->r_ctl.rc_resend;
16889 		rack->r_ctl.rc_resend = NULL;
16890 		len = rsm->r_end - rsm->r_start;
16891 		sack_rxmit = 1;
16892 		sendalot = 0;
16893 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16894 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16895 			 __func__, __LINE__,
16896 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
16897 		sb_offset = rsm->r_start - tp->snd_una;
16898 		if (len >= segsiz)
16899 			len = segsiz;
16900 	} else if (rack->r_collapse_point_valid &&
16901 		   ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
16902 		/*
16903 		 * If an RSM is returned then enough time has passed
16904 		 * for us to retransmit it. Move up the collapse point,
16905 		 * since this rsm has its chance to retransmit now.
16906 		 */
16907 		rack_trace_point(rack, RACK_TP_COLLAPSED_RXT);
16908 		rack->r_ctl.last_collapse_point = rsm->r_end;
16909 		/* Are we done? */
16910 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
16911 			    rack->r_ctl.high_collapse_point))
16912 			rack->r_collapse_point_valid = 0;
16913 		sack_rxmit = 1;
16914 		/* We are not doing a TLP */
16915 		doing_tlp = 0;
16916 		len = rsm->r_end - rsm->r_start;
16917 		sb_offset = rsm->r_start - tp->snd_una;
16918 		sendalot = 0;
16919 		if ((rack->full_size_rxt == 0) &&
16920 		    (rack->shape_rxt_to_pacing_min == 0) &&
16921 		    (len >= segsiz))
16922 			len = segsiz;
16923 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16924 		/* We have a retransmit that takes precedence */
16925 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16926 		    ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
16927 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16928 			/* Enter recovery if not induced by a time-out */
16929 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
16930 		}
16931 #ifdef INVARIANTS
16932 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16933 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16934 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
16935 		}
16936 #endif
16937 		len = rsm->r_end - rsm->r_start;
16938 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16939 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16940 			 __func__, __LINE__,
16941 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
16942 		sb_offset = rsm->r_start - tp->snd_una;
16943 		sendalot = 0;
16944 		if (len >= segsiz)
16945 			len = segsiz;
16946 		if (len > 0) {
16947 			sack_rxmit = 1;
16948 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
16949 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
16950 			    min(len, segsiz));
16951 		}
16952 	} else if (rack->r_ctl.rc_tlpsend) {
16953 		/* Tail loss probe */
16954 		long cwin;
16955 		long tlen;
16956 
16957 		/*
16958 		 * Check if we can do a TLP with a RACK'd packet
16959 		 * this can happen if we are not doing the rack
16960 		 * cheat and we skipped to a TLP and it
16961 		 * went off.
16962 		 */
16963 		rsm = rack->r_ctl.rc_tlpsend;
16964 		/* We are doing a TLP make sure the flag is preent */
16965 		rsm->r_flags |= RACK_TLP;
16966 		rack->r_ctl.rc_tlpsend = NULL;
16967 		sack_rxmit = 1;
16968 		tlen = rsm->r_end - rsm->r_start;
16969 		if (tlen > segsiz)
16970 			tlen = segsiz;
16971 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16972 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16973 			 __func__, __LINE__,
16974 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
16975 		sb_offset = rsm->r_start - tp->snd_una;
16976 		cwin = min(tp->snd_wnd, tlen);
16977 		len = cwin;
16978 	}
16979 	if (rack->r_must_retran &&
16980 	    (doing_tlp == 0) &&
16981 	    (SEQ_GT(tp->snd_max, tp->snd_una)) &&
16982 	    (rsm == NULL)) {
16983 		/*
16984 		 * There are two different ways that we
16985 		 * can get into this block:
16986 		 * a) This is a non-sack connection, we had a time-out
16987 		 *    and thus r_must_retran was set and everything
16988 		 *    left outstanding as been marked for retransmit.
16989 		 * b) The MTU of the path shrank, so that everything
16990 		 *    was marked to be retransmitted with the smaller
16991 		 *    mtu and r_must_retran was set.
16992 		 *
16993 		 * This means that we expect the sendmap (outstanding)
16994 		 * to all be marked must. We can use the tmap to
16995 		 * look at them.
16996 		 *
16997 		 */
16998 		int sendwin, flight;
16999 
17000 		sendwin = min(tp->snd_wnd, tp->snd_cwnd);
17001 		flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
17002 		if (flight >= sendwin) {
17003 			/*
17004 			 * We can't send yet.
17005 			 */
17006 			so = inp->inp_socket;
17007 			sb = &so->so_snd;
17008 			goto just_return_nolock;
17009 		}
17010 		/*
17011 		 * This is the case a/b mentioned above. All
17012 		 * outstanding/not-acked should be marked.
17013 		 * We can use the tmap to find them.
17014 		 */
17015 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17016 		if (rsm == NULL) {
17017 			/* TSNH */
17018 			rack->r_must_retran = 0;
17019 			rack->r_ctl.rc_out_at_rto = 0;
17020 			so = inp->inp_socket;
17021 			sb = &so->so_snd;
17022 			goto just_return_nolock;
17023 		}
17024 		if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
17025 			/*
17026 			 * The first one does not have the flag, did we collapse
17027 			 * further up in our list?
17028 			 */
17029 			rack->r_must_retran = 0;
17030 			rack->r_ctl.rc_out_at_rto = 0;
17031 			rsm = NULL;
17032 			sack_rxmit = 0;
17033 		} else {
17034 			sack_rxmit = 1;
17035 			len = rsm->r_end - rsm->r_start;
17036 			sb_offset = rsm->r_start - tp->snd_una;
17037 			sendalot = 0;
17038 			if ((rack->full_size_rxt == 0) &&
17039 			    (rack->shape_rxt_to_pacing_min == 0) &&
17040 			    (len >= segsiz))
17041 				len = segsiz;
17042 			/*
17043 			 * Delay removing the flag RACK_MUST_RXT so
17044 			 * that the fastpath for retransmit will
17045 			 * work with this rsm.
17046 			 */
17047 		}
17048 	}
17049 	/*
17050 	 * Enforce a connection sendmap count limit if set
17051 	 * as long as we are not retransmiting.
17052 	 */
17053 	if ((rsm == NULL) &&
17054 	    (rack->do_detection == 0) &&
17055 	    (V_tcp_map_entries_limit > 0) &&
17056 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
17057 		counter_u64_add(rack_to_alloc_limited, 1);
17058 		if (!rack->alloc_limit_reported) {
17059 			rack->alloc_limit_reported = 1;
17060 			counter_u64_add(rack_alloc_limited_conns, 1);
17061 		}
17062 		so = inp->inp_socket;
17063 		sb = &so->so_snd;
17064 		goto just_return_nolock;
17065 	}
17066 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
17067 		/* we are retransmitting the fin */
17068 		len--;
17069 		if (len) {
17070 			/*
17071 			 * When retransmitting data do *not* include the
17072 			 * FIN. This could happen from a TLP probe.
17073 			 */
17074 			flags &= ~TH_FIN;
17075 		}
17076 	}
17077 	if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
17078 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
17079 		int ret;
17080 
17081 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
17082 		if (ret == 0)
17083 			return (0);
17084 	}
17085 	so = inp->inp_socket;
17086 	sb = &so->so_snd;
17087 	if (do_a_prefetch == 0) {
17088 		kern_prefetch(sb, &do_a_prefetch);
17089 		do_a_prefetch = 1;
17090 	}
17091 #ifdef NETFLIX_SHARED_CWND
17092 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
17093 	    rack->rack_enable_scwnd) {
17094 		/* We are doing cwnd sharing */
17095 		if (rack->gp_ready &&
17096 		    (rack->rack_attempted_scwnd == 0) &&
17097 		    (rack->r_ctl.rc_scw == NULL) &&
17098 		    tp->t_lib) {
17099 			/* The pcbid is in, lets make an attempt */
17100 			counter_u64_add(rack_try_scwnd, 1);
17101 			rack->rack_attempted_scwnd = 1;
17102 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
17103 								   &rack->r_ctl.rc_scw_index,
17104 								   segsiz);
17105 		}
17106 		if (rack->r_ctl.rc_scw &&
17107 		    (rack->rack_scwnd_is_idle == 1) &&
17108 		    sbavail(&so->so_snd)) {
17109 			/* we are no longer out of data */
17110 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17111 			rack->rack_scwnd_is_idle = 0;
17112 		}
17113 		if (rack->r_ctl.rc_scw) {
17114 			/* First lets update and get the cwnd */
17115 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
17116 								    rack->r_ctl.rc_scw_index,
17117 								    tp->snd_cwnd, tp->snd_wnd, segsiz);
17118 		}
17119 	}
17120 #endif
17121 	/*
17122 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
17123 	 * state flags.
17124 	 */
17125 	if (tp->t_flags & TF_NEEDFIN)
17126 		flags |= TH_FIN;
17127 	if (tp->t_flags & TF_NEEDSYN)
17128 		flags |= TH_SYN;
17129 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
17130 		void *end_rsm;
17131 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
17132 		if (end_rsm)
17133 			kern_prefetch(end_rsm, &prefetch_rsm);
17134 		prefetch_rsm = 1;
17135 	}
17136 	SOCKBUF_LOCK(sb);
17137 	/*
17138 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
17139 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
17140 	 * negative length.  This can also occur when TCP opens up its
17141 	 * congestion window while receiving additional duplicate acks after
17142 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
17143 	 * the fast-retransmit.
17144 	 *
17145 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
17146 	 * set to snd_una, the sb_offset will be 0, and the length may wind
17147 	 * up 0.
17148 	 *
17149 	 * If sack_rxmit is true we are retransmitting from the scoreboard
17150 	 * in which case len is already set.
17151 	 */
17152 	if ((sack_rxmit == 0) &&
17153 	    (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
17154 		uint32_t avail;
17155 
17156 		avail = sbavail(sb);
17157 		if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
17158 			sb_offset = tp->snd_nxt - tp->snd_una;
17159 		else
17160 			sb_offset = 0;
17161 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
17162 			if (rack->r_ctl.rc_tlp_new_data) {
17163 				/* TLP is forcing out new data */
17164 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
17165 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
17166 				}
17167 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
17168 					if (tp->snd_wnd > sb_offset)
17169 						len = tp->snd_wnd - sb_offset;
17170 					else
17171 						len = 0;
17172 				} else {
17173 					len = rack->r_ctl.rc_tlp_new_data;
17174 				}
17175 				rack->r_ctl.rc_tlp_new_data = 0;
17176 			}  else {
17177 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
17178 			}
17179 			if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
17180 				/*
17181 				 * For prr=off, we need to send only 1 MSS
17182 				 * at a time. We do this because another sack could
17183 				 * be arriving that causes us to send retransmits and
17184 				 * we don't want to be on a long pace due to a larger send
17185 				 * that keeps us from sending out the retransmit.
17186 				 */
17187 				len = segsiz;
17188 			}
17189 		} else {
17190 			uint32_t outstanding;
17191 			/*
17192 			 * We are inside of a Fast recovery episode, this
17193 			 * is caused by a SACK or 3 dup acks. At this point
17194 			 * we have sent all the retransmissions and we rely
17195 			 * on PRR to dictate what we will send in the form of
17196 			 * new data.
17197 			 */
17198 
17199 			outstanding = tp->snd_max - tp->snd_una;
17200 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
17201 				if (tp->snd_wnd > outstanding) {
17202 					len = tp->snd_wnd - outstanding;
17203 					/* Check to see if we have the data */
17204 					if ((sb_offset + len) > avail) {
17205 						/* It does not all fit */
17206 						if (avail > sb_offset)
17207 							len = avail - sb_offset;
17208 						else
17209 							len = 0;
17210 					}
17211 				} else {
17212 					len = 0;
17213 				}
17214 			} else if (avail > sb_offset) {
17215 				len = avail - sb_offset;
17216 			} else {
17217 				len = 0;
17218 			}
17219 			if (len > 0) {
17220 				if (len > rack->r_ctl.rc_prr_sndcnt) {
17221 					len = rack->r_ctl.rc_prr_sndcnt;
17222 				}
17223 				if (len > 0) {
17224 					sub_from_prr = 1;
17225 				}
17226 			}
17227 			if (len > segsiz) {
17228 				/*
17229 				 * We should never send more than a MSS when
17230 				 * retransmitting or sending new data in prr
17231 				 * mode unless the override flag is on. Most
17232 				 * likely the PRR algorithm is not going to
17233 				 * let us send a lot as well :-)
17234 				 */
17235 				if (rack->r_ctl.rc_prr_sendalot == 0) {
17236 					len = segsiz;
17237 				}
17238 			} else if (len < segsiz) {
17239 				/*
17240 				 * Do we send any? The idea here is if the
17241 				 * send empty's the socket buffer we want to
17242 				 * do it. However if not then lets just wait
17243 				 * for our prr_sndcnt to get bigger.
17244 				 */
17245 				long leftinsb;
17246 
17247 				leftinsb = sbavail(sb) - sb_offset;
17248 				if (leftinsb > len) {
17249 					/* This send does not empty the sb */
17250 					len = 0;
17251 				}
17252 			}
17253 		}
17254 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
17255 		/*
17256 		 * If you have not established
17257 		 * and are not doing FAST OPEN
17258 		 * no data please.
17259 		 */
17260 		if ((sack_rxmit == 0) &&
17261 		    (!IS_FASTOPEN(tp->t_flags))){
17262 			len = 0;
17263 			sb_offset = 0;
17264 		}
17265 	}
17266 	if (prefetch_so_done == 0) {
17267 		kern_prefetch(so, &prefetch_so_done);
17268 		prefetch_so_done = 1;
17269 	}
17270 	/*
17271 	 * Lop off SYN bit if it has already been sent.  However, if this is
17272 	 * SYN-SENT state and if segment contains data and if we don't know
17273 	 * that foreign host supports TAO, suppress sending segment.
17274 	 */
17275 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
17276 	    ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
17277 		/*
17278 		 * When sending additional segments following a TFO SYN|ACK,
17279 		 * do not include the SYN bit.
17280 		 */
17281 		if (IS_FASTOPEN(tp->t_flags) &&
17282 		    (tp->t_state == TCPS_SYN_RECEIVED))
17283 			flags &= ~TH_SYN;
17284 	}
17285 	/*
17286 	 * Be careful not to send data and/or FIN on SYN segments. This
17287 	 * measure is needed to prevent interoperability problems with not
17288 	 * fully conformant TCP implementations.
17289 	 */
17290 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
17291 		len = 0;
17292 		flags &= ~TH_FIN;
17293 	}
17294 	/*
17295 	 * On TFO sockets, ensure no data is sent in the following cases:
17296 	 *
17297 	 *  - When retransmitting SYN|ACK on a passively-created socket
17298 	 *
17299 	 *  - When retransmitting SYN on an actively created socket
17300 	 *
17301 	 *  - When sending a zero-length cookie (cookie request) on an
17302 	 *    actively created socket
17303 	 *
17304 	 *  - When the socket is in the CLOSED state (RST is being sent)
17305 	 */
17306 	if (IS_FASTOPEN(tp->t_flags) &&
17307 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
17308 	     ((tp->t_state == TCPS_SYN_SENT) &&
17309 	      (tp->t_tfo_client_cookie_len == 0)) ||
17310 	     (flags & TH_RST))) {
17311 		sack_rxmit = 0;
17312 		len = 0;
17313 	}
17314 	/* Without fast-open there should never be data sent on a SYN */
17315 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
17316 		tp->snd_nxt = tp->iss;
17317 		len = 0;
17318 	}
17319 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
17320 		/* We only send 1 MSS if we have a DSACK block */
17321 		add_flag |= RACK_SENT_W_DSACK;
17322 		len = segsiz;
17323 	}
17324 	orig_len = len;
17325 	if (len <= 0) {
17326 		/*
17327 		 * If FIN has been sent but not acked, but we haven't been
17328 		 * called to retransmit, len will be < 0.  Otherwise, window
17329 		 * shrank after we sent into it.  If window shrank to 0,
17330 		 * cancel pending retransmit, pull snd_nxt back to (closed)
17331 		 * window, and set the persist timer if it isn't already
17332 		 * going.  If the window didn't close completely, just wait
17333 		 * for an ACK.
17334 		 *
17335 		 * We also do a general check here to ensure that we will
17336 		 * set the persist timer when we have data to send, but a
17337 		 * 0-byte window. This makes sure the persist timer is set
17338 		 * even if the packet hits one of the "goto send" lines
17339 		 * below.
17340 		 */
17341 		len = 0;
17342 		if ((tp->snd_wnd == 0) &&
17343 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17344 		    (tp->snd_una == tp->snd_max) &&
17345 		    (sb_offset < (int)sbavail(sb))) {
17346 			rack_enter_persist(tp, rack, cts);
17347 		}
17348 	} else if ((rsm == NULL) &&
17349 		   (doing_tlp == 0) &&
17350 		   (len < pace_max_seg)) {
17351 		/*
17352 		 * We are not sending a maximum sized segment for
17353 		 * some reason. Should we not send anything (think
17354 		 * sws or persists)?
17355 		 */
17356 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17357 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17358 		    (len < minseg) &&
17359 		    (len < (int)(sbavail(sb) - sb_offset))) {
17360 			/*
17361 			 * Here the rwnd is less than
17362 			 * the minimum pacing size, this is not a retransmit,
17363 			 * we are established and
17364 			 * the send is not the last in the socket buffer
17365 			 * we send nothing, and we may enter persists
17366 			 * if nothing is outstanding.
17367 			 */
17368 			len = 0;
17369 			if (tp->snd_max == tp->snd_una) {
17370 				/*
17371 				 * Nothing out we can
17372 				 * go into persists.
17373 				 */
17374 				rack_enter_persist(tp, rack, cts);
17375 			}
17376 		     } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
17377 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17378 			   (len < (int)(sbavail(sb) - sb_offset)) &&
17379 			   (len < minseg)) {
17380 			/*
17381 			 * Here we are not retransmitting, and
17382 			 * the cwnd is not so small that we could
17383 			 * not send at least a min size (rxt timer
17384 			 * not having gone off), We have 2 segments or
17385 			 * more already in flight, its not the tail end
17386 			 * of the socket buffer  and the cwnd is blocking
17387 			 * us from sending out a minimum pacing segment size.
17388 			 * Lets not send anything.
17389 			 */
17390 			len = 0;
17391 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
17392 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17393 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17394 			   (len < (int)(sbavail(sb) - sb_offset)) &&
17395 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
17396 			/*
17397 			 * Here we have a send window but we have
17398 			 * filled it up and we can't send another pacing segment.
17399 			 * We also have in flight more than 2 segments
17400 			 * and we are not completing the sb i.e. we allow
17401 			 * the last bytes of the sb to go out even if
17402 			 * its not a full pacing segment.
17403 			 */
17404 			len = 0;
17405 		} else if ((rack->r_ctl.crte != NULL) &&
17406 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
17407 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
17408 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
17409 			   (len < (int)(sbavail(sb) - sb_offset))) {
17410 			/*
17411 			 * Here we are doing hardware pacing, this is not a TLP,
17412 			 * we are not sending a pace max segment size, there is rwnd
17413 			 * room to send at least N pace_max_seg, the cwnd is greater
17414 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
17415 			 * more segments in flight and its not the tail of the socket buffer.
17416 			 *
17417 			 * We don't want to send instead we need to get more ack's in to
17418 			 * allow us to send a full pacing segment. Normally, if we are pacing
17419 			 * about the right speed, we should have finished our pacing
17420 			 * send as most of the acks have come back if we are at the
17421 			 * right rate. This is a bit fuzzy since return path delay
17422 			 * can delay the acks, which is why we want to make sure we
17423 			 * have cwnd space to have a bit more than a max pace segments in flight.
17424 			 *
17425 			 * If we have not gotten our acks back we are pacing at too high a
17426 			 * rate delaying will not hurt and will bring our GP estimate down by
17427 			 * injecting the delay. If we don't do this we will send
17428 			 * 2 MSS out in response to the acks being clocked in which
17429 			 * defeats the point of hw-pacing (i.e. to help us get
17430 			 * larger TSO's out).
17431 			 */
17432 			len = 0;
17433 
17434 		}
17435 
17436 	}
17437 	/* len will be >= 0 after this point. */
17438 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17439 	rack_sndbuf_autoscale(rack);
17440 	/*
17441 	 * Decide if we can use TCP Segmentation Offloading (if supported by
17442 	 * hardware).
17443 	 *
17444 	 * TSO may only be used if we are in a pure bulk sending state.  The
17445 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
17446 	 * options prevent using TSO.  With TSO the TCP header is the same
17447 	 * (except for the sequence number) for all generated packets.  This
17448 	 * makes it impossible to transmit any options which vary per
17449 	 * generated segment or packet.
17450 	 *
17451 	 * IPv4 handling has a clear separation of ip options and ip header
17452 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
17453 	 * the right thing below to provide length of just ip options and thus
17454 	 * checking for ipoptlen is enough to decide if ip options are present.
17455 	 */
17456 	ipoptlen = 0;
17457 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17458 	/*
17459 	 * Pre-calculate here as we save another lookup into the darknesses
17460 	 * of IPsec that way and can actually decide if TSO is ok.
17461 	 */
17462 #ifdef INET6
17463 	if (isipv6 && IPSEC_ENABLED(ipv6))
17464 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
17465 #ifdef INET
17466 	else
17467 #endif
17468 #endif				/* INET6 */
17469 #ifdef INET
17470 		if (IPSEC_ENABLED(ipv4))
17471 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
17472 #endif				/* INET */
17473 #endif
17474 
17475 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17476 	ipoptlen += ipsec_optlen;
17477 #endif
17478 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
17479 	    (tp->t_port == 0) &&
17480 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
17481 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
17482 	    ipoptlen == 0)
17483 		tso = 1;
17484 	{
17485 		uint32_t outstanding __unused;
17486 
17487 		outstanding = tp->snd_max - tp->snd_una;
17488 		if (tp->t_flags & TF_SENTFIN) {
17489 			/*
17490 			 * If we sent a fin, snd_max is 1 higher than
17491 			 * snd_una
17492 			 */
17493 			outstanding--;
17494 		}
17495 		if (sack_rxmit) {
17496 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
17497 				flags &= ~TH_FIN;
17498 		} else {
17499 			if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
17500 				   sbused(sb)))
17501 				flags &= ~TH_FIN;
17502 		}
17503 	}
17504 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
17505 	    (long)TCP_MAXWIN << tp->rcv_scale);
17506 
17507 	/*
17508 	 * Sender silly window avoidance.   We transmit under the following
17509 	 * conditions when len is non-zero:
17510 	 *
17511 	 * - We have a full segment (or more with TSO) - This is the last
17512 	 * buffer in a write()/send() and we are either idle or running
17513 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
17514 	 * then 1/2 the maximum send window's worth of data (receiver may be
17515 	 * limited the window size) - we need to retransmit
17516 	 */
17517 	if (len) {
17518 		if (len >= segsiz) {
17519 			goto send;
17520 		}
17521 		/*
17522 		 * NOTE! on localhost connections an 'ack' from the remote
17523 		 * end may occur synchronously with the output and cause us
17524 		 * to flush a buffer queued with moretocome.  XXX
17525 		 *
17526 		 */
17527 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
17528 		    (idle || (tp->t_flags & TF_NODELAY)) &&
17529 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17530 		    (tp->t_flags & TF_NOPUSH) == 0) {
17531 			pass = 2;
17532 			goto send;
17533 		}
17534 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
17535 			pass = 22;
17536 			goto send;
17537 		}
17538 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
17539 			pass = 4;
17540 			goto send;
17541 		}
17542 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {	/* retransmit case */
17543 			pass = 5;
17544 			goto send;
17545 		}
17546 		if (sack_rxmit) {
17547 			pass = 6;
17548 			goto send;
17549 		}
17550 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
17551 		    (ctf_outstanding(tp) < (segsiz * 2))) {
17552 			/*
17553 			 * We have less than two MSS outstanding (delayed ack)
17554 			 * and our rwnd will not let us send a full sized
17555 			 * MSS. Lets go ahead and let this small segment
17556 			 * out because we want to try to have at least two
17557 			 * packets inflight to not be caught by delayed ack.
17558 			 */
17559 			pass = 12;
17560 			goto send;
17561 		}
17562 	}
17563 	/*
17564 	 * Sending of standalone window updates.
17565 	 *
17566 	 * Window updates are important when we close our window due to a
17567 	 * full socket buffer and are opening it again after the application
17568 	 * reads data from it.  Once the window has opened again and the
17569 	 * remote end starts to send again the ACK clock takes over and
17570 	 * provides the most current window information.
17571 	 *
17572 	 * We must avoid the silly window syndrome whereas every read from
17573 	 * the receive buffer, no matter how small, causes a window update
17574 	 * to be sent.  We also should avoid sending a flurry of window
17575 	 * updates when the socket buffer had queued a lot of data and the
17576 	 * application is doing small reads.
17577 	 *
17578 	 * Prevent a flurry of pointless window updates by only sending an
17579 	 * update when we can increase the advertized window by more than
17580 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
17581 	 * full or is very small be more aggressive and send an update
17582 	 * whenever we can increase by two mss sized segments. In all other
17583 	 * situations the ACK's to new incoming data will carry further
17584 	 * window increases.
17585 	 *
17586 	 * Don't send an independent window update if a delayed ACK is
17587 	 * pending (it will get piggy-backed on it) or the remote side
17588 	 * already has done a half-close and won't send more data.  Skip
17589 	 * this if the connection is in T/TCP half-open state.
17590 	 */
17591 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
17592 	    !(tp->t_flags & TF_DELACK) &&
17593 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
17594 		/*
17595 		 * "adv" is the amount we could increase the window, taking
17596 		 * into account that we are limited by TCP_MAXWIN <<
17597 		 * tp->rcv_scale.
17598 		 */
17599 		int32_t adv;
17600 		int oldwin;
17601 
17602 		adv = recwin;
17603 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
17604 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
17605 			if (adv > oldwin)
17606 			    adv -= oldwin;
17607 			else {
17608 				/* We can't increase the window */
17609 				adv = 0;
17610 			}
17611 		} else
17612 			oldwin = 0;
17613 
17614 		/*
17615 		 * If the new window size ends up being the same as or less
17616 		 * than the old size when it is scaled, then don't force
17617 		 * a window update.
17618 		 */
17619 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
17620 			goto dontupdate;
17621 
17622 		if (adv >= (int32_t)(2 * segsiz) &&
17623 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
17624 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
17625 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
17626 			pass = 7;
17627 			goto send;
17628 		}
17629 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
17630 			pass = 23;
17631 			goto send;
17632 		}
17633 	}
17634 dontupdate:
17635 
17636 	/*
17637 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
17638 	 * is also a catch-all for the retransmit timer timeout case.
17639 	 */
17640 	if (tp->t_flags & TF_ACKNOW) {
17641 		pass = 8;
17642 		goto send;
17643 	}
17644 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
17645 		pass = 9;
17646 		goto send;
17647 	}
17648 	/*
17649 	 * If our state indicates that FIN should be sent and we have not
17650 	 * yet done so, then we need to send.
17651 	 */
17652 	if ((flags & TH_FIN) &&
17653 	    (tp->snd_nxt == tp->snd_una)) {
17654 		pass = 11;
17655 		goto send;
17656 	}
17657 	/*
17658 	 * No reason to send a segment, just return.
17659 	 */
17660 just_return:
17661 	SOCKBUF_UNLOCK(sb);
17662 just_return_nolock:
17663 	{
17664 		int app_limited = CTF_JR_SENT_DATA;
17665 
17666 		if (tot_len_this_send > 0) {
17667 			/* Make sure snd_nxt is up to max */
17668 			rack->r_ctl.fsb.recwin = recwin;
17669 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17670 			if ((error == 0) &&
17671 			    rack_use_rfo &&
17672 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
17673 			    (ipoptlen == 0) &&
17674 			    (tp->snd_nxt == tp->snd_max) &&
17675 			    (tp->rcv_numsacks == 0) &&
17676 			    rack->r_fsb_inited &&
17677 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
17678 			    (rack->r_must_retran == 0) &&
17679 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
17680 			    (len > 0) && (orig_len > 0) &&
17681 			    (orig_len > len) &&
17682 			    ((orig_len - len) >= segsiz) &&
17683 			    ((optlen == 0) ||
17684 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17685 				/* We can send at least one more MSS using our fsb */
17686 
17687 				rack->r_fast_output = 1;
17688 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17689 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17690 				rack->r_ctl.fsb.tcp_flags = flags;
17691 				rack->r_ctl.fsb.left_to_send = orig_len - len;
17692 				if (hw_tls)
17693 					rack->r_ctl.fsb.hw_tls = 1;
17694 				else
17695 					rack->r_ctl.fsb.hw_tls = 0;
17696 				KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17697 					("rack:%p left_to_send:%u sbavail:%u out:%u",
17698 					rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17699 					 (tp->snd_max - tp->snd_una)));
17700 				if (rack->r_ctl.fsb.left_to_send < segsiz)
17701 					rack->r_fast_output = 0;
17702 				else {
17703 					if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17704 						rack->r_ctl.fsb.rfo_apply_push = 1;
17705 					else
17706 						rack->r_ctl.fsb.rfo_apply_push = 0;
17707 				}
17708 			} else
17709 				rack->r_fast_output = 0;
17710 
17711 
17712 			rack_log_fsb(rack, tp, so, flags,
17713 				     ipoptlen, orig_len, len, 0,
17714 				     1, optlen, __LINE__, 1);
17715 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17716 				tp->snd_nxt = tp->snd_max;
17717 		} else {
17718 			int end_window = 0;
17719 			uint32_t seq = tp->gput_ack;
17720 
17721 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17722 			if (rsm) {
17723 				/*
17724 				 * Mark the last sent that we just-returned (hinting
17725 				 * that delayed ack may play a role in any rtt measurement).
17726 				 */
17727 				rsm->r_just_ret = 1;
17728 			}
17729 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17730 			rack->r_ctl.rc_agg_delayed = 0;
17731 			rack->r_early = 0;
17732 			rack->r_late = 0;
17733 			rack->r_ctl.rc_agg_early = 0;
17734 			if ((ctf_outstanding(tp) +
17735 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17736 				 minseg)) >= tp->snd_wnd) {
17737 				/* We are limited by the rwnd */
17738 				app_limited = CTF_JR_RWND_LIMITED;
17739 				if (IN_FASTRECOVERY(tp->t_flags))
17740 				    rack->r_ctl.rc_prr_sndcnt = 0;
17741 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
17742 				/* We are limited by whats available -- app limited */
17743 				app_limited = CTF_JR_APP_LIMITED;
17744 				if (IN_FASTRECOVERY(tp->t_flags))
17745 				    rack->r_ctl.rc_prr_sndcnt = 0;
17746 			} else if ((idle == 0) &&
17747 				   ((tp->t_flags & TF_NODELAY) == 0) &&
17748 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17749 				   (len < segsiz)) {
17750 				/*
17751 				 * No delay is not on and the
17752 				 * user is sending less than 1MSS. This
17753 				 * brings out SWS avoidance so we
17754 				 * don't send. Another app-limited case.
17755 				 */
17756 				app_limited = CTF_JR_APP_LIMITED;
17757 			} else if (tp->t_flags & TF_NOPUSH) {
17758 				/*
17759 				 * The user has requested no push of
17760 				 * the last segment and we are
17761 				 * at the last segment. Another app
17762 				 * limited case.
17763 				 */
17764 				app_limited = CTF_JR_APP_LIMITED;
17765 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17766 				/* Its the cwnd */
17767 				app_limited = CTF_JR_CWND_LIMITED;
17768 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
17769 				   (rack->rack_no_prr == 0) &&
17770 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17771 				app_limited = CTF_JR_PRR;
17772 			} else {
17773 				/* Now why here are we not sending? */
17774 #ifdef NOW
17775 #ifdef INVARIANTS
17776 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17777 #endif
17778 #endif
17779 				app_limited = CTF_JR_ASSESSING;
17780 			}
17781 			/*
17782 			 * App limited in some fashion, for our pacing GP
17783 			 * measurements we don't want any gap (even cwnd).
17784 			 * Close  down the measurement window.
17785 			 */
17786 			if (rack_cwnd_block_ends_measure &&
17787 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
17788 			     (app_limited == CTF_JR_PRR))) {
17789 				/*
17790 				 * The reason we are not sending is
17791 				 * the cwnd (or prr). We have been configured
17792 				 * to end the measurement window in
17793 				 * this case.
17794 				 */
17795 				end_window = 1;
17796 			} else if (rack_rwnd_block_ends_measure &&
17797 				   (app_limited == CTF_JR_RWND_LIMITED)) {
17798 				/*
17799 				 * We are rwnd limited and have been
17800 				 * configured to end the measurement
17801 				 * window in this case.
17802 				 */
17803 				end_window = 1;
17804 			} else if (app_limited == CTF_JR_APP_LIMITED) {
17805 				/*
17806 				 * A true application limited period, we have
17807 				 * ran out of data.
17808 				 */
17809 				end_window = 1;
17810 			} else if (app_limited == CTF_JR_ASSESSING) {
17811 				/*
17812 				 * In the assessing case we hit the end of
17813 				 * the if/else and had no known reason
17814 				 * This will panic us under invariants..
17815 				 *
17816 				 * If we get this out in logs we need to
17817 				 * investagate which reason we missed.
17818 				 */
17819 				end_window = 1;
17820 			}
17821 			if (end_window) {
17822 				uint8_t log = 0;
17823 
17824 				/* Adjust the Gput measurement */
17825 				if ((tp->t_flags & TF_GPUTINPROG) &&
17826 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
17827 					tp->gput_ack = tp->snd_max;
17828 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
17829 						/*
17830 						 * There is not enough to measure.
17831 						 */
17832 						tp->t_flags &= ~TF_GPUTINPROG;
17833 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17834 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
17835 									   tp->gput_seq,
17836 									   0, 0, 18, __LINE__, NULL, 0);
17837 					} else
17838 						log = 1;
17839 				}
17840 				/* Mark the last packet has app limited */
17841 				rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17842 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17843 					if (rack->r_ctl.rc_app_limited_cnt == 0)
17844 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17845 					else {
17846 						/*
17847 						 * Go out to the end app limited and mark
17848 						 * this new one as next and move the end_appl up
17849 						 * to this guy.
17850 						 */
17851 						if (rack->r_ctl.rc_end_appl)
17852 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17853 						rack->r_ctl.rc_end_appl = rsm;
17854 					}
17855 					rsm->r_flags |= RACK_APP_LIMITED;
17856 					rack->r_ctl.rc_app_limited_cnt++;
17857 				}
17858 				if (log)
17859 					rack_log_pacing_delay_calc(rack,
17860 								   rack->r_ctl.rc_app_limited_cnt, seq,
17861 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
17862 			}
17863 		}
17864 		/* Check if we need to go into persists or not */
17865 		if ((tp->snd_max == tp->snd_una) &&
17866 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
17867 		    sbavail(sb) &&
17868 		    (sbavail(sb) > tp->snd_wnd) &&
17869 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17870 			/* Yes lets make sure to move to persist before timer-start */
17871 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17872 		}
17873 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17874 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17875 	}
17876 #ifdef NETFLIX_SHARED_CWND
17877 	if ((sbavail(sb) == 0) &&
17878 	    rack->r_ctl.rc_scw) {
17879 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17880 		rack->rack_scwnd_is_idle = 1;
17881 	}
17882 #endif
17883 #ifdef TCP_ACCOUNTING
17884 	if (tot_len_this_send > 0) {
17885 		crtsc = get_cyclecount();
17886 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17887 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
17888 		}
17889 		counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17890 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17891 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17892 		}
17893 		counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
17894 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17895 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17896 		}
17897 		counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
17898 	} else {
17899 		crtsc = get_cyclecount();
17900 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17901 			tp->tcp_cnt_counters[SND_LIMITED]++;
17902 		}
17903 		counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
17904 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17905 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17906 		}
17907 		counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
17908 	}
17909 	sched_unpin();
17910 #endif
17911 	return (0);
17912 
17913 send:
17914 	if (rsm || sack_rxmit)
17915 		counter_u64_add(rack_nfto_resend, 1);
17916 	else
17917 		counter_u64_add(rack_non_fto_send, 1);
17918 	if ((flags & TH_FIN) &&
17919 	    sbavail(sb)) {
17920 		/*
17921 		 * We do not transmit a FIN
17922 		 * with data outstanding. We
17923 		 * need to make it so all data
17924 		 * is acked first.
17925 		 */
17926 		flags &= ~TH_FIN;
17927 	}
17928 	/* Enforce stack imposed max seg size if we have one */
17929 	if (rack->r_ctl.rc_pace_max_segs &&
17930 	    (len > rack->r_ctl.rc_pace_max_segs)) {
17931 		mark = 1;
17932 		len = rack->r_ctl.rc_pace_max_segs;
17933 	}
17934 	SOCKBUF_LOCK_ASSERT(sb);
17935 	if (len > 0) {
17936 		if (len >= segsiz)
17937 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
17938 		else
17939 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
17940 	}
17941 	/*
17942 	 * Before ESTABLISHED, force sending of initial options unless TCP
17943 	 * set not to do any options. NOTE: we assume that the IP/TCP header
17944 	 * plus TCP options always fit in a single mbuf, leaving room for a
17945 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
17946 	 * + optlen <= MCLBYTES
17947 	 */
17948 	optlen = 0;
17949 #ifdef INET6
17950 	if (isipv6)
17951 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
17952 	else
17953 #endif
17954 		hdrlen = sizeof(struct tcpiphdr);
17955 
17956 	/*
17957 	 * Compute options for segment. We only have to care about SYN and
17958 	 * established connection segments.  Options for SYN-ACK segments
17959 	 * are handled in TCP syncache.
17960 	 */
17961 	to.to_flags = 0;
17962 	if ((tp->t_flags & TF_NOOPT) == 0) {
17963 		/* Maximum segment size. */
17964 		if (flags & TH_SYN) {
17965 			tp->snd_nxt = tp->iss;
17966 			to.to_mss = tcp_mssopt(&inp->inp_inc);
17967 			if (tp->t_port)
17968 				to.to_mss -= V_tcp_udp_tunneling_overhead;
17969 			to.to_flags |= TOF_MSS;
17970 
17971 			/*
17972 			 * On SYN or SYN|ACK transmits on TFO connections,
17973 			 * only include the TFO option if it is not a
17974 			 * retransmit, as the presence of the TFO option may
17975 			 * have caused the original SYN or SYN|ACK to have
17976 			 * been dropped by a middlebox.
17977 			 */
17978 			if (IS_FASTOPEN(tp->t_flags) &&
17979 			    (tp->t_rxtshift == 0)) {
17980 				if (tp->t_state == TCPS_SYN_RECEIVED) {
17981 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
17982 					to.to_tfo_cookie =
17983 						(u_int8_t *)&tp->t_tfo_cookie.server;
17984 					to.to_flags |= TOF_FASTOPEN;
17985 					wanted_cookie = 1;
17986 				} else if (tp->t_state == TCPS_SYN_SENT) {
17987 					to.to_tfo_len =
17988 						tp->t_tfo_client_cookie_len;
17989 					to.to_tfo_cookie =
17990 						tp->t_tfo_cookie.client;
17991 					to.to_flags |= TOF_FASTOPEN;
17992 					wanted_cookie = 1;
17993 					/*
17994 					 * If we wind up having more data to
17995 					 * send with the SYN than can fit in
17996 					 * one segment, don't send any more
17997 					 * until the SYN|ACK comes back from
17998 					 * the other end.
17999 					 */
18000 					sendalot = 0;
18001 				}
18002 			}
18003 		}
18004 		/* Window scaling. */
18005 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
18006 			to.to_wscale = tp->request_r_scale;
18007 			to.to_flags |= TOF_SCALE;
18008 		}
18009 		/* Timestamps. */
18010 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
18011 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
18012 			to.to_tsval = ms_cts + tp->ts_offset;
18013 			to.to_tsecr = tp->ts_recent;
18014 			to.to_flags |= TOF_TS;
18015 		}
18016 		/* Set receive buffer autosizing timestamp. */
18017 		if (tp->rfbuf_ts == 0 &&
18018 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
18019 			tp->rfbuf_ts = tcp_ts_getticks();
18020 		/* Selective ACK's. */
18021 		if (tp->t_flags & TF_SACK_PERMIT) {
18022 			if (flags & TH_SYN)
18023 				to.to_flags |= TOF_SACKPERM;
18024 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18025 				 tp->rcv_numsacks > 0) {
18026 				to.to_flags |= TOF_SACK;
18027 				to.to_nsacks = tp->rcv_numsacks;
18028 				to.to_sacks = (u_char *)tp->sackblks;
18029 			}
18030 		}
18031 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18032 		/* TCP-MD5 (RFC2385). */
18033 		if (tp->t_flags & TF_SIGNATURE)
18034 			to.to_flags |= TOF_SIGNATURE;
18035 #endif				/* TCP_SIGNATURE */
18036 
18037 		/* Processing the options. */
18038 		hdrlen += optlen = tcp_addoptions(&to, opt);
18039 		/*
18040 		 * If we wanted a TFO option to be added, but it was unable
18041 		 * to fit, ensure no data is sent.
18042 		 */
18043 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
18044 		    !(to.to_flags & TOF_FASTOPEN))
18045 			len = 0;
18046 	}
18047 	if (tp->t_port) {
18048 		if (V_tcp_udp_tunneling_port == 0) {
18049 			/* The port was removed?? */
18050 			SOCKBUF_UNLOCK(&so->so_snd);
18051 #ifdef TCP_ACCOUNTING
18052 			crtsc = get_cyclecount();
18053 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18054 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18055 			}
18056 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18057 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18058 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18059 			}
18060 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18061 			sched_unpin();
18062 #endif
18063 			return (EHOSTUNREACH);
18064 		}
18065 		hdrlen += sizeof(struct udphdr);
18066 	}
18067 #ifdef INET6
18068 	if (isipv6)
18069 		ipoptlen = ip6_optlen(inp);
18070 	else
18071 #endif
18072 		if (inp->inp_options)
18073 			ipoptlen = inp->inp_options->m_len -
18074 				offsetof(struct ipoption, ipopt_list);
18075 		else
18076 			ipoptlen = 0;
18077 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18078 	ipoptlen += ipsec_optlen;
18079 #endif
18080 
18081 	/*
18082 	 * Adjust data length if insertion of options will bump the packet
18083 	 * length beyond the t_maxseg length. Clear the FIN bit because we
18084 	 * cut off the tail of the segment.
18085 	 */
18086 	if (len + optlen + ipoptlen > tp->t_maxseg) {
18087 		if (tso) {
18088 			uint32_t if_hw_tsomax;
18089 			uint32_t moff;
18090 			int32_t max_len;
18091 
18092 			/* extract TSO information */
18093 			if_hw_tsomax = tp->t_tsomax;
18094 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18095 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18096 			KASSERT(ipoptlen == 0,
18097 				("%s: TSO can't do IP options", __func__));
18098 
18099 			/*
18100 			 * Check if we should limit by maximum payload
18101 			 * length:
18102 			 */
18103 			if (if_hw_tsomax != 0) {
18104 				/* compute maximum TSO length */
18105 				max_len = (if_hw_tsomax - hdrlen -
18106 					   max_linkhdr);
18107 				if (max_len <= 0) {
18108 					len = 0;
18109 				} else if (len > max_len) {
18110 					sendalot = 1;
18111 					len = max_len;
18112 					mark = 2;
18113 				}
18114 			}
18115 			/*
18116 			 * Prevent the last segment from being fractional
18117 			 * unless the send sockbuf can be emptied:
18118 			 */
18119 			max_len = (tp->t_maxseg - optlen);
18120 			if ((sb_offset + len) < sbavail(sb)) {
18121 				moff = len % (u_int)max_len;
18122 				if (moff != 0) {
18123 					mark = 3;
18124 					len -= moff;
18125 				}
18126 			}
18127 			/*
18128 			 * In case there are too many small fragments don't
18129 			 * use TSO:
18130 			 */
18131 			if (len <= segsiz) {
18132 				mark = 4;
18133 				tso = 0;
18134 			}
18135 			/*
18136 			 * Send the FIN in a separate segment after the bulk
18137 			 * sending is done. We don't trust the TSO
18138 			 * implementations to clear the FIN flag on all but
18139 			 * the last segment.
18140 			 */
18141 			if (tp->t_flags & TF_NEEDFIN) {
18142 				sendalot = 4;
18143 			}
18144 		} else {
18145 			mark = 5;
18146 			if (optlen + ipoptlen >= tp->t_maxseg) {
18147 				/*
18148 				 * Since we don't have enough space to put
18149 				 * the IP header chain and the TCP header in
18150 				 * one packet as required by RFC 7112, don't
18151 				 * send it. Also ensure that at least one
18152 				 * byte of the payload can be put into the
18153 				 * TCP segment.
18154 				 */
18155 				SOCKBUF_UNLOCK(&so->so_snd);
18156 				error = EMSGSIZE;
18157 				sack_rxmit = 0;
18158 				goto out;
18159 			}
18160 			len = tp->t_maxseg - optlen - ipoptlen;
18161 			sendalot = 5;
18162 		}
18163 	} else {
18164 		tso = 0;
18165 		mark = 6;
18166 	}
18167 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
18168 		("%s: len > IP_MAXPACKET", __func__));
18169 #ifdef DIAGNOSTIC
18170 #ifdef INET6
18171 	if (max_linkhdr + hdrlen > MCLBYTES)
18172 #else
18173 		if (max_linkhdr + hdrlen > MHLEN)
18174 #endif
18175 			panic("tcphdr too big");
18176 #endif
18177 
18178 	/*
18179 	 * This KASSERT is here to catch edge cases at a well defined place.
18180 	 * Before, those had triggered (random) panic conditions further
18181 	 * down.
18182 	 */
18183 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
18184 	if ((len == 0) &&
18185 	    (flags & TH_FIN) &&
18186 	    (sbused(sb))) {
18187 		/*
18188 		 * We have outstanding data, don't send a fin by itself!.
18189 		 */
18190 		goto just_return;
18191 	}
18192 	/*
18193 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
18194 	 * and initialize the header from the template for sends on this
18195 	 * connection.
18196 	 */
18197 	hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
18198 	if (len) {
18199 		uint32_t max_val;
18200 		uint32_t moff;
18201 
18202 		if (rack->r_ctl.rc_pace_max_segs)
18203 			max_val = rack->r_ctl.rc_pace_max_segs;
18204 		else if (rack->rc_user_set_max_segs)
18205 			max_val = rack->rc_user_set_max_segs * segsiz;
18206 		else
18207 			max_val = len;
18208 		/*
18209 		 * We allow a limit on sending with hptsi.
18210 		 */
18211 		if (len > max_val) {
18212 			mark = 7;
18213 			len = max_val;
18214 		}
18215 #ifdef INET6
18216 		if (MHLEN < hdrlen + max_linkhdr)
18217 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18218 		else
18219 #endif
18220 			m = m_gethdr(M_NOWAIT, MT_DATA);
18221 
18222 		if (m == NULL) {
18223 			SOCKBUF_UNLOCK(sb);
18224 			error = ENOBUFS;
18225 			sack_rxmit = 0;
18226 			goto out;
18227 		}
18228 		m->m_data += max_linkhdr;
18229 		m->m_len = hdrlen;
18230 
18231 		/*
18232 		 * Start the m_copy functions from the closest mbuf to the
18233 		 * sb_offset in the socket buffer chain.
18234 		 */
18235 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
18236 		s_mb = mb;
18237 		s_moff = moff;
18238 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
18239 			m_copydata(mb, moff, (int)len,
18240 				   mtod(m, caddr_t)+hdrlen);
18241 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18242 				sbsndptr_adv(sb, mb, len);
18243 			m->m_len += len;
18244 		} else {
18245 			struct sockbuf *msb;
18246 
18247 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18248 				msb = NULL;
18249 			else
18250 				msb = sb;
18251 			m->m_next = tcp_m_copym(
18252 				mb, moff, &len,
18253 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
18254 				((rsm == NULL) ? hw_tls : 0)
18255 #ifdef NETFLIX_COPY_ARGS
18256 				, &s_mb, &s_moff
18257 #endif
18258 				);
18259 			if (len <= (tp->t_maxseg - optlen)) {
18260 				/*
18261 				 * Must have ran out of mbufs for the copy
18262 				 * shorten it to no longer need tso. Lets
18263 				 * not put on sendalot since we are low on
18264 				 * mbufs.
18265 				 */
18266 				tso = 0;
18267 			}
18268 			if (m->m_next == NULL) {
18269 				SOCKBUF_UNLOCK(sb);
18270 				(void)m_free(m);
18271 				error = ENOBUFS;
18272 				sack_rxmit = 0;
18273 				goto out;
18274 			}
18275 		}
18276 		if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
18277 			if (rsm && (rsm->r_flags & RACK_TLP)) {
18278 				/*
18279 				 * TLP should not count in retran count, but
18280 				 * in its own bin
18281 				 */
18282 				counter_u64_add(rack_tlp_retran, 1);
18283 				counter_u64_add(rack_tlp_retran_bytes, len);
18284 			} else {
18285 				tp->t_sndrexmitpack++;
18286 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18287 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18288 			}
18289 #ifdef STATS
18290 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18291 						 len);
18292 #endif
18293 		} else {
18294 			KMOD_TCPSTAT_INC(tcps_sndpack);
18295 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
18296 #ifdef STATS
18297 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
18298 						 len);
18299 #endif
18300 		}
18301 		/*
18302 		 * If we're sending everything we've got, set PUSH. (This
18303 		 * will keep happy those implementations which only give
18304 		 * data to the user when a buffer fills or a PUSH comes in.)
18305 		 */
18306 		if (sb_offset + len == sbused(sb) &&
18307 		    sbused(sb) &&
18308 		    !(flags & TH_SYN)) {
18309 			flags |= TH_PUSH;
18310 			add_flag |= RACK_HAD_PUSH;
18311 		}
18312 
18313 		SOCKBUF_UNLOCK(sb);
18314 	} else {
18315 		SOCKBUF_UNLOCK(sb);
18316 		if (tp->t_flags & TF_ACKNOW)
18317 			KMOD_TCPSTAT_INC(tcps_sndacks);
18318 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
18319 			KMOD_TCPSTAT_INC(tcps_sndctrl);
18320 		else
18321 			KMOD_TCPSTAT_INC(tcps_sndwinup);
18322 
18323 		m = m_gethdr(M_NOWAIT, MT_DATA);
18324 		if (m == NULL) {
18325 			error = ENOBUFS;
18326 			sack_rxmit = 0;
18327 			goto out;
18328 		}
18329 #ifdef INET6
18330 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
18331 		    MHLEN >= hdrlen) {
18332 			M_ALIGN(m, hdrlen);
18333 		} else
18334 #endif
18335 			m->m_data += max_linkhdr;
18336 		m->m_len = hdrlen;
18337 	}
18338 	SOCKBUF_UNLOCK_ASSERT(sb);
18339 	m->m_pkthdr.rcvif = (struct ifnet *)0;
18340 #ifdef MAC
18341 	mac_inpcb_create_mbuf(inp, m);
18342 #endif
18343 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
18344 #ifdef INET6
18345 		if (isipv6)
18346 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18347 		else
18348 #endif				/* INET6 */
18349 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18350 		th = rack->r_ctl.fsb.th;
18351 		udp = rack->r_ctl.fsb.udp;
18352 		if (udp) {
18353 #ifdef INET6
18354 			if (isipv6)
18355 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
18356 			else
18357 #endif				/* INET6 */
18358 				ulen = hdrlen + len - sizeof(struct ip);
18359 			udp->uh_ulen = htons(ulen);
18360 		}
18361 	} else {
18362 #ifdef INET6
18363 		if (isipv6) {
18364 			ip6 = mtod(m, struct ip6_hdr *);
18365 			if (tp->t_port) {
18366 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
18367 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18368 				udp->uh_dport = tp->t_port;
18369 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
18370 				udp->uh_ulen = htons(ulen);
18371 				th = (struct tcphdr *)(udp + 1);
18372 			} else
18373 				th = (struct tcphdr *)(ip6 + 1);
18374 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
18375 		} else
18376 #endif				/* INET6 */
18377 		{
18378 			ip = mtod(m, struct ip *);
18379 			if (tp->t_port) {
18380 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
18381 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18382 				udp->uh_dport = tp->t_port;
18383 				ulen = hdrlen + len - sizeof(struct ip);
18384 				udp->uh_ulen = htons(ulen);
18385 				th = (struct tcphdr *)(udp + 1);
18386 			} else
18387 				th = (struct tcphdr *)(ip + 1);
18388 			tcpip_fillheaders(inp, tp->t_port, ip, th);
18389 		}
18390 	}
18391 	/*
18392 	 * Fill in fields, remembering maximum advertised window for use in
18393 	 * delaying messages about window sizes. If resending a FIN, be sure
18394 	 * not to use a new sequence number.
18395 	 */
18396 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
18397 	    tp->snd_nxt == tp->snd_max)
18398 		tp->snd_nxt--;
18399 	/*
18400 	 * If we are starting a connection, send ECN setup SYN packet. If we
18401 	 * are on a retransmit, we may resend those bits a number of times
18402 	 * as per RFC 3168.
18403 	 */
18404 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
18405 		flags |= tcp_ecn_output_syn_sent(tp);
18406 	}
18407 	/* Also handle parallel SYN for ECN */
18408 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
18409 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18410 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
18411 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18412 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
18413 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18414 #ifdef INET6
18415 		if (isipv6) {
18416 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18417 			ip6->ip6_flow |= htonl(ect << 20);
18418 		}
18419 		else
18420 #endif
18421 		{
18422 			ip->ip_tos &= ~IPTOS_ECN_MASK;
18423 			ip->ip_tos |= ect;
18424 		}
18425 	}
18426 	/*
18427 	 * If we are doing retransmissions, then snd_nxt will not reflect
18428 	 * the first unsent octet.  For ACK only packets, we do not want the
18429 	 * sequence number of the retransmitted packet, we want the sequence
18430 	 * number of the next unsent octet.  So, if there is no data (and no
18431 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
18432 	 * ti_seq.  But if we are in persist state, snd_max might reflect
18433 	 * one byte beyond the right edge of the window, so use snd_nxt in
18434 	 * that case, since we know we aren't doing a retransmission.
18435 	 * (retransmit and persist are mutually exclusive...)
18436 	 */
18437 	if (sack_rxmit == 0) {
18438 		if (len || (flags & (TH_SYN | TH_FIN))) {
18439 			th->th_seq = htonl(tp->snd_nxt);
18440 			rack_seq = tp->snd_nxt;
18441 		} else {
18442 			th->th_seq = htonl(tp->snd_max);
18443 			rack_seq = tp->snd_max;
18444 		}
18445 	} else {
18446 		th->th_seq = htonl(rsm->r_start);
18447 		rack_seq = rsm->r_start;
18448 	}
18449 	th->th_ack = htonl(tp->rcv_nxt);
18450 	tcp_set_flags(th, flags);
18451 	/*
18452 	 * Calculate receive window.  Don't shrink window, but avoid silly
18453 	 * window syndrome.
18454 	 * If a RST segment is sent, advertise a window of zero.
18455 	 */
18456 	if (flags & TH_RST) {
18457 		recwin = 0;
18458 	} else {
18459 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
18460 		    recwin < (long)segsiz) {
18461 			recwin = 0;
18462 		}
18463 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
18464 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
18465 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
18466 	}
18467 
18468 	/*
18469 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
18470 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
18471 	 * handled in syncache.
18472 	 */
18473 	if (flags & TH_SYN)
18474 		th->th_win = htons((u_short)
18475 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
18476 	else {
18477 		/* Avoid shrinking window with window scaling. */
18478 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
18479 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
18480 	}
18481 	/*
18482 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
18483 	 * window.  This may cause the remote transmitter to stall.  This
18484 	 * flag tells soreceive() to disable delayed acknowledgements when
18485 	 * draining the buffer.  This can occur if the receiver is
18486 	 * attempting to read more data than can be buffered prior to
18487 	 * transmitting on the connection.
18488 	 */
18489 	if (th->th_win == 0) {
18490 		tp->t_sndzerowin++;
18491 		tp->t_flags |= TF_RXWIN0SENT;
18492 	} else
18493 		tp->t_flags &= ~TF_RXWIN0SENT;
18494 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
18495 	/* Now are we using fsb?, if so copy the template data to the mbuf */
18496 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
18497 		uint8_t *cpto;
18498 
18499 		cpto = mtod(m, uint8_t *);
18500 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18501 		/*
18502 		 * We have just copied in:
18503 		 * IP/IP6
18504 		 * <optional udphdr>
18505 		 * tcphdr (no options)
18506 		 *
18507 		 * We need to grab the correct pointers into the mbuf
18508 		 * for both the tcp header, and possibly the udp header (if tunneling).
18509 		 * We do this by using the offset in the copy buffer and adding it
18510 		 * to the mbuf base pointer (cpto).
18511 		 */
18512 #ifdef INET6
18513 		if (isipv6)
18514 			ip6 = mtod(m, struct ip6_hdr *);
18515 		else
18516 #endif				/* INET6 */
18517 			ip = mtod(m, struct ip *);
18518 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18519 		/* If we have a udp header lets set it into the mbuf as well */
18520 		if (udp)
18521 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
18522 	}
18523 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18524 	if (to.to_flags & TOF_SIGNATURE) {
18525 		/*
18526 		 * Calculate MD5 signature and put it into the place
18527 		 * determined before.
18528 		 * NOTE: since TCP options buffer doesn't point into
18529 		 * mbuf's data, calculate offset and use it.
18530 		 */
18531 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18532 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18533 			/*
18534 			 * Do not send segment if the calculation of MD5
18535 			 * digest has failed.
18536 			 */
18537 			goto out;
18538 		}
18539 	}
18540 #endif
18541 	if (optlen) {
18542 		bcopy(opt, th + 1, optlen);
18543 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18544 	}
18545 	/*
18546 	 * Put TCP length in extended header, and then checksum extended
18547 	 * header and data.
18548 	 */
18549 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
18550 #ifdef INET6
18551 	if (isipv6) {
18552 		/*
18553 		 * ip6_plen is not need to be filled now, and will be filled
18554 		 * in ip6_output.
18555 		 */
18556 		if (tp->t_port) {
18557 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18558 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18559 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18560 			th->th_sum = htons(0);
18561 			UDPSTAT_INC(udps_opackets);
18562 		} else {
18563 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18564 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18565 			th->th_sum = in6_cksum_pseudo(ip6,
18566 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18567 						      0);
18568 		}
18569 	}
18570 #endif
18571 #if defined(INET6) && defined(INET)
18572 	else
18573 #endif
18574 #ifdef INET
18575 	{
18576 		if (tp->t_port) {
18577 			m->m_pkthdr.csum_flags = CSUM_UDP;
18578 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18579 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18580 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18581 			th->th_sum = htons(0);
18582 			UDPSTAT_INC(udps_opackets);
18583 		} else {
18584 			m->m_pkthdr.csum_flags = CSUM_TCP;
18585 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18586 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
18587 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18588 									IPPROTO_TCP + len + optlen));
18589 		}
18590 		/* IP version must be set here for ipv4/ipv6 checking later */
18591 		KASSERT(ip->ip_v == IPVERSION,
18592 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
18593 	}
18594 #endif
18595 	/*
18596 	 * Enable TSO and specify the size of the segments. The TCP pseudo
18597 	 * header checksum is always provided. XXX: Fixme: This is currently
18598 	 * not the case for IPv6.
18599 	 */
18600 	if (tso) {
18601 		KASSERT(len > tp->t_maxseg - optlen,
18602 			("%s: len <= tso_segsz", __func__));
18603 		m->m_pkthdr.csum_flags |= CSUM_TSO;
18604 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
18605 	}
18606 	KASSERT(len + hdrlen == m_length(m, NULL),
18607 		("%s: mbuf chain different than expected: %d + %u != %u",
18608 		 __func__, len, hdrlen, m_length(m, NULL)));
18609 
18610 #ifdef TCP_HHOOK
18611 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
18612 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
18613 #endif
18614 	/* We're getting ready to send; log now. */
18615 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
18616 		union tcp_log_stackspecific log;
18617 
18618 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18619 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
18620 		if (rack->rack_no_prr)
18621 			log.u_bbr.flex1 = 0;
18622 		else
18623 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18624 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18625 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18626 		log.u_bbr.flex4 = orig_len;
18627 		/* Save off the early/late values */
18628 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18629 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18630 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18631 		log.u_bbr.flex8 = 0;
18632 		if (rsm) {
18633 			if (rsm->r_flags & RACK_RWND_COLLAPSED) {
18634 				rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
18635 				counter_u64_add(rack_collapsed_win_rxt, 1);
18636 				counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
18637 			}
18638 			if (doing_tlp)
18639 				log.u_bbr.flex8 = 2;
18640 			else
18641 				log.u_bbr.flex8 = 1;
18642 		} else {
18643 			if (doing_tlp)
18644 				log.u_bbr.flex8 = 3;
18645 			else
18646 				log.u_bbr.flex8 = 0;
18647 		}
18648 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
18649 		log.u_bbr.flex7 = mark;
18650 		log.u_bbr.flex7 <<= 8;
18651 		log.u_bbr.flex7 |= pass;
18652 		log.u_bbr.pkts_out = tp->t_maxseg;
18653 		log.u_bbr.timeStamp = cts;
18654 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18655 		log.u_bbr.lt_epoch = cwnd_to_use;
18656 		log.u_bbr.delivered = sendalot;
18657 		lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
18658 				     len, &log, false, NULL, NULL, 0, &tv);
18659 	} else
18660 		lgb = NULL;
18661 
18662 	/*
18663 	 * Fill in IP length and desired time to live and send to IP level.
18664 	 * There should be a better way to handle ttl and tos; we could keep
18665 	 * them in the template, but need a way to checksum without them.
18666 	 */
18667 	/*
18668 	 * m->m_pkthdr.len should have been set before cksum calcuration,
18669 	 * because in6_cksum() need it.
18670 	 */
18671 #ifdef INET6
18672 	if (isipv6) {
18673 		/*
18674 		 * we separately set hoplimit for every segment, since the
18675 		 * user might want to change the value via setsockopt. Also,
18676 		 * desired default hop limit might be changed via Neighbor
18677 		 * Discovery.
18678 		 */
18679 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18680 
18681 		/*
18682 		 * Set the packet size here for the benefit of DTrace
18683 		 * probes. ip6_output() will set it properly; it's supposed
18684 		 * to include the option header lengths as well.
18685 		 */
18686 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18687 
18688 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18689 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18690 		else
18691 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18692 
18693 		if (tp->t_state == TCPS_SYN_SENT)
18694 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18695 
18696 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18697 		/* TODO: IPv6 IP6TOS_ECT bit on */
18698 		error = ip6_output(m,
18699 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18700 				   inp->in6p_outputopts,
18701 #else
18702 				   NULL,
18703 #endif
18704 				   &inp->inp_route6,
18705 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18706 				   NULL, NULL, inp);
18707 
18708 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18709 			mtu = inp->inp_route6.ro_nh->nh_mtu;
18710 	}
18711 #endif				/* INET6 */
18712 #if defined(INET) && defined(INET6)
18713 	else
18714 #endif
18715 #ifdef INET
18716 	{
18717 		ip->ip_len = htons(m->m_pkthdr.len);
18718 #ifdef INET6
18719 		if (inp->inp_vflag & INP_IPV6PROTO)
18720 			ip->ip_ttl = in6_selecthlim(inp, NULL);
18721 #endif				/* INET6 */
18722 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18723 		/*
18724 		 * If we do path MTU discovery, then we set DF on every
18725 		 * packet. This might not be the best thing to do according
18726 		 * to RFC3390 Section 2. However the tcp hostcache migitates
18727 		 * the problem so it affects only the first tcp connection
18728 		 * with a host.
18729 		 *
18730 		 * NB: Don't set DF on small MTU/MSS to have a safe
18731 		 * fallback.
18732 		 */
18733 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18734 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18735 			if (tp->t_port == 0 || len < V_tcp_minmss) {
18736 				ip->ip_off |= htons(IP_DF);
18737 			}
18738 		} else {
18739 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18740 		}
18741 
18742 		if (tp->t_state == TCPS_SYN_SENT)
18743 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18744 
18745 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
18746 
18747 		error = ip_output(m,
18748 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18749 				  inp->inp_options,
18750 #else
18751 				  NULL,
18752 #endif
18753 				  &inp->inp_route,
18754 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18755 				  inp);
18756 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18757 			mtu = inp->inp_route.ro_nh->nh_mtu;
18758 	}
18759 #endif				/* INET */
18760 
18761 out:
18762 	if (lgb) {
18763 		lgb->tlb_errno = error;
18764 		lgb = NULL;
18765 	}
18766 	/*
18767 	 * In transmit state, time the transmission and arrange for the
18768 	 * retransmit.  In persist state, just set snd_max.
18769 	 */
18770 	if (error == 0) {
18771 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
18772 		if (rsm && doing_tlp) {
18773 			rack->rc_last_sent_tlp_past_cumack = 0;
18774 			rack->rc_last_sent_tlp_seq_valid = 1;
18775 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18776 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18777 		}
18778 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
18779 		if (rsm && (doing_tlp == 0)) {
18780 			/* Set we retransmitted */
18781 			rack->rc_gp_saw_rec = 1;
18782 		} else {
18783 			if (cwnd_to_use > tp->snd_ssthresh) {
18784 				/* Set we sent in CA */
18785 				rack->rc_gp_saw_ca = 1;
18786 			} else {
18787 				/* Set we sent in SS */
18788 				rack->rc_gp_saw_ss = 1;
18789 			}
18790 		}
18791 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18792 		    (tp->t_flags & TF_SACK_PERMIT) &&
18793 		    tp->rcv_numsacks > 0)
18794 			tcp_clean_dsack_blocks(tp);
18795 		tot_len_this_send += len;
18796 		if (len == 0)
18797 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18798 		else if (len == 1) {
18799 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18800 		} else if (len > 1) {
18801 			int idx;
18802 
18803 			idx = (len / segsiz) + 3;
18804 			if (idx >= TCP_MSS_ACCT_ATIMER)
18805 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18806 			else
18807 				counter_u64_add(rack_out_size[idx], 1);
18808 		}
18809 	}
18810 	if ((rack->rack_no_prr == 0) &&
18811 	    sub_from_prr &&
18812 	    (error == 0)) {
18813 		if (rack->r_ctl.rc_prr_sndcnt >= len)
18814 			rack->r_ctl.rc_prr_sndcnt -= len;
18815 		else
18816 			rack->r_ctl.rc_prr_sndcnt = 0;
18817 	}
18818 	sub_from_prr = 0;
18819 	if (doing_tlp) {
18820 		/* Make sure the TLP is added */
18821 		add_flag |= RACK_TLP;
18822 	} else if (rsm) {
18823 		/* If its a resend without TLP then it must not have the flag */
18824 		rsm->r_flags &= ~RACK_TLP;
18825 	}
18826 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18827 			rack_to_usec_ts(&tv),
18828 			rsm, add_flag, s_mb, s_moff, hw_tls);
18829 
18830 
18831 	if ((error == 0) &&
18832 	    (len > 0) &&
18833 	    (tp->snd_una == tp->snd_max))
18834 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
18835 	{
18836 		tcp_seq startseq = tp->snd_nxt;
18837 
18838 		/* Track our lost count */
18839 		if (rsm && (doing_tlp == 0))
18840 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18841 		/*
18842 		 * Advance snd_nxt over sequence space of this segment.
18843 		 */
18844 		if (error)
18845 			/* We don't log or do anything with errors */
18846 			goto nomore;
18847 		if (doing_tlp == 0) {
18848 			if (rsm == NULL) {
18849 				/*
18850 				 * Not a retransmission of some
18851 				 * sort, new data is going out so
18852 				 * clear our TLP count and flag.
18853 				 */
18854 				rack->rc_tlp_in_progress = 0;
18855 				rack->r_ctl.rc_tlp_cnt_out = 0;
18856 			}
18857 		} else {
18858 			/*
18859 			 * We have just sent a TLP, mark that it is true
18860 			 * and make sure our in progress is set so we
18861 			 * continue to check the count.
18862 			 */
18863 			rack->rc_tlp_in_progress = 1;
18864 			rack->r_ctl.rc_tlp_cnt_out++;
18865 		}
18866 		if (flags & (TH_SYN | TH_FIN)) {
18867 			if (flags & TH_SYN)
18868 				tp->snd_nxt++;
18869 			if (flags & TH_FIN) {
18870 				tp->snd_nxt++;
18871 				tp->t_flags |= TF_SENTFIN;
18872 			}
18873 		}
18874 		/* In the ENOBUFS case we do *not* update snd_max */
18875 		if (sack_rxmit)
18876 			goto nomore;
18877 
18878 		tp->snd_nxt += len;
18879 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18880 			if (tp->snd_una == tp->snd_max) {
18881 				/*
18882 				 * Update the time we just added data since
18883 				 * none was outstanding.
18884 				 */
18885 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18886 				tp->t_acktime = ticks;
18887 			}
18888 			tp->snd_max = tp->snd_nxt;
18889 			/*
18890 			 * Time this transmission if not a retransmission and
18891 			 * not currently timing anything.
18892 			 * This is only relevant in case of switching back to
18893 			 * the base stack.
18894 			 */
18895 			if (tp->t_rtttime == 0) {
18896 				tp->t_rtttime = ticks;
18897 				tp->t_rtseq = startseq;
18898 				KMOD_TCPSTAT_INC(tcps_segstimed);
18899 			}
18900 			if (len &&
18901 			    ((tp->t_flags & TF_GPUTINPROG) == 0))
18902 				rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18903 		}
18904 		/*
18905 		 * If we are doing FO we need to update the mbuf position and subtract
18906 		 * this happens when the peer sends us duplicate information and
18907 		 * we thus want to send a DSACK.
18908 		 *
18909 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18910 		 * turned off? If not then we are going to echo multiple DSACK blocks
18911 		 * out (with the TSO), which we should not be doing.
18912 		 */
18913 		if (rack->r_fast_output && len) {
18914 			if (rack->r_ctl.fsb.left_to_send > len)
18915 				rack->r_ctl.fsb.left_to_send -= len;
18916 			else
18917 				rack->r_ctl.fsb.left_to_send = 0;
18918 			if (rack->r_ctl.fsb.left_to_send < segsiz)
18919 				rack->r_fast_output = 0;
18920 			if (rack->r_fast_output) {
18921 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18922 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18923 			}
18924 		}
18925 	}
18926 nomore:
18927 	if (error) {
18928 		rack->r_ctl.rc_agg_delayed = 0;
18929 		rack->r_early = 0;
18930 		rack->r_late = 0;
18931 		rack->r_ctl.rc_agg_early = 0;
18932 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
18933 		/*
18934 		 * Failures do not advance the seq counter above. For the
18935 		 * case of ENOBUFS we will fall out and retry in 1ms with
18936 		 * the hpts. Everything else will just have to retransmit
18937 		 * with the timer.
18938 		 *
18939 		 * In any case, we do not want to loop around for another
18940 		 * send without a good reason.
18941 		 */
18942 		sendalot = 0;
18943 		switch (error) {
18944 		case EPERM:
18945 			tp->t_softerror = error;
18946 #ifdef TCP_ACCOUNTING
18947 			crtsc = get_cyclecount();
18948 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18949 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18950 			}
18951 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18952 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18953 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18954 			}
18955 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18956 			sched_unpin();
18957 #endif
18958 			return (error);
18959 		case ENOBUFS:
18960 			/*
18961 			 * Pace us right away to retry in a some
18962 			 * time
18963 			 */
18964 			if (rack->r_ctl.crte != NULL) {
18965 				rack_trace_point(rack, RACK_TP_HWENOBUF);
18966 			} else
18967 				rack_trace_point(rack, RACK_TP_ENOBUF);
18968 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18969 			if (rack->rc_enobuf < 0x7f)
18970 				rack->rc_enobuf++;
18971 			if (slot < (10 * HPTS_USEC_IN_MSEC))
18972 				slot = 10 * HPTS_USEC_IN_MSEC;
18973 			if (rack->r_ctl.crte != NULL) {
18974 				counter_u64_add(rack_saw_enobuf_hw, 1);
18975 				tcp_rl_log_enobuf(rack->r_ctl.crte);
18976 			}
18977 			counter_u64_add(rack_saw_enobuf, 1);
18978 			goto enobufs;
18979 		case EMSGSIZE:
18980 			/*
18981 			 * For some reason the interface we used initially
18982 			 * to send segments changed to another or lowered
18983 			 * its MTU. If TSO was active we either got an
18984 			 * interface without TSO capabilits or TSO was
18985 			 * turned off. If we obtained mtu from ip_output()
18986 			 * then update it and try again.
18987 			 */
18988 			if (tso)
18989 				tp->t_flags &= ~TF_TSO;
18990 			if (mtu != 0) {
18991 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
18992 				goto again;
18993 			}
18994 			slot = 10 * HPTS_USEC_IN_MSEC;
18995 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
18996 #ifdef TCP_ACCOUNTING
18997 			crtsc = get_cyclecount();
18998 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18999 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19000 			}
19001 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19002 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19003 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19004 			}
19005 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19006 			sched_unpin();
19007 #endif
19008 			return (error);
19009 		case ENETUNREACH:
19010 			counter_u64_add(rack_saw_enetunreach, 1);
19011 		case EHOSTDOWN:
19012 		case EHOSTUNREACH:
19013 		case ENETDOWN:
19014 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
19015 				tp->t_softerror = error;
19016 			}
19017 			/* FALLTHROUGH */
19018 		default:
19019 			slot = 10 * HPTS_USEC_IN_MSEC;
19020 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19021 #ifdef TCP_ACCOUNTING
19022 			crtsc = get_cyclecount();
19023 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19024 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19025 			}
19026 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19027 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19028 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19029 			}
19030 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19031 			sched_unpin();
19032 #endif
19033 			return (error);
19034 		}
19035 	} else {
19036 		rack->rc_enobuf = 0;
19037 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
19038 			rack->r_ctl.retran_during_recovery += len;
19039 	}
19040 	KMOD_TCPSTAT_INC(tcps_sndtotal);
19041 
19042 	/*
19043 	 * Data sent (as far as we can tell). If this advertises a larger
19044 	 * window than any other segment, then remember the size of the
19045 	 * advertised window. Any pending ACK has now been sent.
19046 	 */
19047 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
19048 		tp->rcv_adv = tp->rcv_nxt + recwin;
19049 
19050 	tp->last_ack_sent = tp->rcv_nxt;
19051 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19052 enobufs:
19053 	if (sendalot) {
19054 		/* Do we need to turn off sendalot? */
19055 		if (rack->r_ctl.rc_pace_max_segs &&
19056 		    (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
19057 			/* We hit our max. */
19058 			sendalot = 0;
19059 		} else if ((rack->rc_user_set_max_segs) &&
19060 			   (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
19061 			/* We hit the user defined max */
19062 			sendalot = 0;
19063 		}
19064 	}
19065 	if ((error == 0) && (flags & TH_FIN))
19066 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
19067 	if (flags & TH_RST) {
19068 		/*
19069 		 * We don't send again after sending a RST.
19070 		 */
19071 		slot = 0;
19072 		sendalot = 0;
19073 		if (error == 0)
19074 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
19075 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
19076 		/*
19077 		 * Get our pacing rate, if an error
19078 		 * occurred in sending (ENOBUF) we would
19079 		 * hit the else if with slot preset. Other
19080 		 * errors return.
19081 		 */
19082 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
19083 	}
19084 	if (rsm &&
19085 	    (rsm->r_flags & RACK_HAS_SYN) == 0 &&
19086 	    rack->use_rack_rr) {
19087 		/* Its a retransmit and we use the rack cheat? */
19088 		if ((slot == 0) ||
19089 		    (rack->rc_always_pace == 0) ||
19090 		    (rack->r_rr_config == 1)) {
19091 			/*
19092 			 * We have no pacing set or we
19093 			 * are using old-style rack or
19094 			 * we are overridden to use the old 1ms pacing.
19095 			 */
19096 			slot = rack->r_ctl.rc_min_to;
19097 		}
19098 	}
19099 	/* We have sent clear the flag */
19100 	rack->r_ent_rec_ns = 0;
19101 	if (rack->r_must_retran) {
19102 		if (rsm) {
19103 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
19104 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
19105 				/*
19106 				 * We have retransmitted all.
19107 				 */
19108 				rack->r_must_retran = 0;
19109 				rack->r_ctl.rc_out_at_rto = 0;
19110 			}
19111 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19112 			/*
19113 			 * Sending new data will also kill
19114 			 * the loop.
19115 			 */
19116 			rack->r_must_retran = 0;
19117 			rack->r_ctl.rc_out_at_rto = 0;
19118 		}
19119 	}
19120 	rack->r_ctl.fsb.recwin = recwin;
19121 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
19122 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19123 		/*
19124 		 * We hit an RTO and now have past snd_max at the RTO
19125 		 * clear all the WAS flags.
19126 		 */
19127 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
19128 	}
19129 	if (slot) {
19130 		/* set the rack tcb into the slot N */
19131 		if ((error == 0) &&
19132 		    rack_use_rfo &&
19133 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
19134 		    (rsm == NULL) &&
19135 		    (tp->snd_nxt == tp->snd_max) &&
19136 		    (ipoptlen == 0) &&
19137 		    (tp->rcv_numsacks == 0) &&
19138 		    rack->r_fsb_inited &&
19139 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
19140 		    (rack->r_must_retran == 0) &&
19141 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
19142 		    (len > 0) && (orig_len > 0) &&
19143 		    (orig_len > len) &&
19144 		    ((orig_len - len) >= segsiz) &&
19145 		    ((optlen == 0) ||
19146 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19147 			/* We can send at least one more MSS using our fsb */
19148 
19149 			rack->r_fast_output = 1;
19150 			rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19151 			rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19152 			rack->r_ctl.fsb.tcp_flags = flags;
19153 			rack->r_ctl.fsb.left_to_send = orig_len - len;
19154 			if (hw_tls)
19155 				rack->r_ctl.fsb.hw_tls = 1;
19156 			else
19157 				rack->r_ctl.fsb.hw_tls = 0;
19158 			KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19159 				("rack:%p left_to_send:%u sbavail:%u out:%u",
19160 				 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19161 				 (tp->snd_max - tp->snd_una)));
19162 			if (rack->r_ctl.fsb.left_to_send < segsiz)
19163 				rack->r_fast_output = 0;
19164 			else {
19165 				if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19166 					rack->r_ctl.fsb.rfo_apply_push = 1;
19167 				else
19168 					rack->r_ctl.fsb.rfo_apply_push = 0;
19169 			}
19170 		} else
19171 			rack->r_fast_output = 0;
19172 		rack_log_fsb(rack, tp, so, flags,
19173 			     ipoptlen, orig_len, len, error,
19174 			     (rsm == NULL), optlen, __LINE__, 2);
19175 	} else if (sendalot) {
19176 		int ret;
19177 
19178 		sack_rxmit = 0;
19179 		if ((error == 0) &&
19180 		    rack_use_rfo &&
19181 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
19182 		    (rsm == NULL) &&
19183 		    (ipoptlen == 0) &&
19184 		    (tp->rcv_numsacks == 0) &&
19185 		    (tp->snd_nxt == tp->snd_max) &&
19186 		    (rack->r_must_retran == 0) &&
19187 		    rack->r_fsb_inited &&
19188 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
19189 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
19190 		    (len > 0) && (orig_len > 0) &&
19191 		    (orig_len > len) &&
19192 		    ((orig_len - len) >= segsiz) &&
19193 		    ((optlen == 0) ||
19194 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19195 			/* we can use fast_output for more */
19196 
19197 			rack->r_fast_output = 1;
19198 			rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19199 			rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19200 			rack->r_ctl.fsb.tcp_flags = flags;
19201 			rack->r_ctl.fsb.left_to_send = orig_len - len;
19202 			if (hw_tls)
19203 				rack->r_ctl.fsb.hw_tls = 1;
19204 			else
19205 				rack->r_ctl.fsb.hw_tls = 0;
19206 			KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19207 				("rack:%p left_to_send:%u sbavail:%u out:%u",
19208 				 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19209 				 (tp->snd_max - tp->snd_una)));
19210 			if (rack->r_ctl.fsb.left_to_send < segsiz) {
19211 				rack->r_fast_output = 0;
19212 			}
19213 			if (rack->r_fast_output) {
19214 				if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19215 					rack->r_ctl.fsb.rfo_apply_push = 1;
19216 				else
19217 					rack->r_ctl.fsb.rfo_apply_push = 0;
19218 				rack_log_fsb(rack, tp, so, flags,
19219 					     ipoptlen, orig_len, len, error,
19220 					     (rsm == NULL), optlen, __LINE__, 3);
19221 				error = 0;
19222 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19223 				if (ret >= 0)
19224 					return (ret);
19225 			        else if (error)
19226 					goto nomore;
19227 
19228 			}
19229 		}
19230 		goto again;
19231 	}
19232 	/* Assure when we leave that snd_nxt will point to top */
19233 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
19234 		tp->snd_nxt = tp->snd_max;
19235 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
19236 #ifdef TCP_ACCOUNTING
19237 	crtsc = get_cyclecount() - ts_val;
19238 	if (tot_len_this_send) {
19239 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19240 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
19241 		}
19242 		counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
19243 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19244 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
19245 		}
19246 		counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
19247 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19248 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
19249 		}
19250 		counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
19251 	} else {
19252 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19253 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
19254 		}
19255 		counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
19256 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19257 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
19258 		}
19259 		counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
19260 	}
19261 	sched_unpin();
19262 #endif
19263 	if (error == ENOBUFS)
19264 		error = 0;
19265 	return (error);
19266 }
19267 
19268 static void
19269 rack_update_seg(struct tcp_rack *rack)
19270 {
19271 	uint32_t orig_val;
19272 
19273 	orig_val = rack->r_ctl.rc_pace_max_segs;
19274 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
19275 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
19276 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
19277 }
19278 
19279 static void
19280 rack_mtu_change(struct tcpcb *tp)
19281 {
19282 	/*
19283 	 * The MSS may have changed
19284 	 */
19285 	struct tcp_rack *rack;
19286 	struct rack_sendmap *rsm;
19287 
19288 	rack = (struct tcp_rack *)tp->t_fb_ptr;
19289 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
19290 		/*
19291 		 * The MTU has changed we need to resend everything
19292 		 * since all we have sent is lost. We first fix
19293 		 * up the mtu though.
19294 		 */
19295 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
19296 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
19297 		rack_remxt_tmr(tp);
19298 		rack->r_fast_output = 0;
19299 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
19300 						rack->r_ctl.rc_sacked);
19301 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
19302 		rack->r_must_retran = 1;
19303 		/* Mark all inflight to needing to be rxt'd */
19304 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
19305 			rsm->r_flags |= RACK_MUST_RXT;
19306 		}
19307 	}
19308 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
19309 	/* We don't use snd_nxt to retransmit */
19310 	tp->snd_nxt = tp->snd_max;
19311 }
19312 
19313 static int
19314 rack_set_profile(struct tcp_rack *rack, int prof)
19315 {
19316 	int err = EINVAL;
19317 	if (prof == 1) {
19318 		/* pace_always=1 */
19319 		if (rack->rc_always_pace == 0) {
19320 			if (tcp_can_enable_pacing() == 0)
19321 				return (EBUSY);
19322 		}
19323 		rack->rc_always_pace = 1;
19324 		if (rack->use_fixed_rate || rack->gp_ready)
19325 			rack_set_cc_pacing(rack);
19326 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19327 		rack->rack_attempt_hdwr_pace = 0;
19328 		/* cmpack=1 */
19329 		if (rack_use_cmp_acks)
19330 			rack->r_use_cmp_ack = 1;
19331 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19332 		    rack->r_use_cmp_ack)
19333 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19334 		/* scwnd=1 */
19335 		rack->rack_enable_scwnd = 1;
19336 		/* dynamic=100 */
19337 		rack->rc_gp_dyn_mul = 1;
19338 		/* gp_inc_ca */
19339 		rack->r_ctl.rack_per_of_gp_ca = 100;
19340 		/* rrr_conf=3 */
19341 		rack->r_rr_config = 3;
19342 		/* npush=2 */
19343 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19344 		/* fillcw=1 */
19345 		rack->rc_pace_to_cwnd = 1;
19346 		rack->rc_pace_fill_if_rttin_range = 0;
19347 		rack->rtt_limit_mul = 0;
19348 		/* noprr=1 */
19349 		rack->rack_no_prr = 1;
19350 		/* lscwnd=1 */
19351 		rack->r_limit_scw = 1;
19352 		/* gp_inc_rec */
19353 		rack->r_ctl.rack_per_of_gp_rec = 90;
19354 		err = 0;
19355 
19356 	} else if (prof == 3) {
19357 		/* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
19358 		/* pace_always=1 */
19359 		if (rack->rc_always_pace == 0) {
19360 			if (tcp_can_enable_pacing() == 0)
19361 				return (EBUSY);
19362 		}
19363 		rack->rc_always_pace = 1;
19364 		if (rack->use_fixed_rate || rack->gp_ready)
19365 			rack_set_cc_pacing(rack);
19366 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19367 		rack->rack_attempt_hdwr_pace = 0;
19368 		/* cmpack=1 */
19369 		if (rack_use_cmp_acks)
19370 			rack->r_use_cmp_ack = 1;
19371 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19372 		    rack->r_use_cmp_ack)
19373 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19374 		/* scwnd=1 */
19375 		rack->rack_enable_scwnd = 1;
19376 		/* dynamic=100 */
19377 		rack->rc_gp_dyn_mul = 1;
19378 		/* gp_inc_ca */
19379 		rack->r_ctl.rack_per_of_gp_ca = 100;
19380 		/* rrr_conf=3 */
19381 		rack->r_rr_config = 3;
19382 		/* npush=2 */
19383 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19384 		/* fillcw=2 */
19385 		rack->rc_pace_to_cwnd = 1;
19386 		rack->r_fill_less_agg = 1;
19387 		rack->rc_pace_fill_if_rttin_range = 0;
19388 		rack->rtt_limit_mul = 0;
19389 		/* noprr=1 */
19390 		rack->rack_no_prr = 1;
19391 		/* lscwnd=1 */
19392 		rack->r_limit_scw = 1;
19393 		/* gp_inc_rec */
19394 		rack->r_ctl.rack_per_of_gp_rec = 90;
19395 		err = 0;
19396 
19397 
19398 	} else if (prof == 2) {
19399 		/* cmpack=1 */
19400 		if (rack->rc_always_pace == 0) {
19401 			if (tcp_can_enable_pacing() == 0)
19402 				return (EBUSY);
19403 		}
19404 		rack->rc_always_pace = 1;
19405 		if (rack->use_fixed_rate || rack->gp_ready)
19406 			rack_set_cc_pacing(rack);
19407 		rack->r_use_cmp_ack = 1;
19408 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19409 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19410 		/* pace_always=1 */
19411 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19412 		/* scwnd=1 */
19413 		rack->rack_enable_scwnd = 1;
19414 		/* dynamic=100 */
19415 		rack->rc_gp_dyn_mul = 1;
19416 		rack->r_ctl.rack_per_of_gp_ca = 100;
19417 		/* rrr_conf=3 */
19418 		rack->r_rr_config = 3;
19419 		/* npush=2 */
19420 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19421 		/* fillcw=1 */
19422 		rack->rc_pace_to_cwnd = 1;
19423 		rack->rc_pace_fill_if_rttin_range = 0;
19424 		rack->rtt_limit_mul = 0;
19425 		/* noprr=1 */
19426 		rack->rack_no_prr = 1;
19427 		/* lscwnd=0 */
19428 		rack->r_limit_scw = 0;
19429 		err = 0;
19430 	} else if (prof == 0) {
19431 		/* This changes things back to the default settings */
19432 		err = 0;
19433 		if (rack->rc_always_pace) {
19434 			tcp_decrement_paced_conn();
19435 			rack_undo_cc_pacing(rack);
19436 			rack->rc_always_pace = 0;
19437 		}
19438 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
19439 			rack->rc_always_pace = 1;
19440 			if (rack->use_fixed_rate || rack->gp_ready)
19441 				rack_set_cc_pacing(rack);
19442 		} else
19443 			rack->rc_always_pace = 0;
19444 		if (rack_dsack_std_based & 0x1) {
19445 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
19446 			rack->rc_rack_tmr_std_based = 1;
19447 		}
19448 		if (rack_dsack_std_based & 0x2) {
19449 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
19450 			rack->rc_rack_use_dsack = 1;
19451 		}
19452 		if (rack_use_cmp_acks)
19453 			rack->r_use_cmp_ack = 1;
19454 		else
19455 			rack->r_use_cmp_ack = 0;
19456 		if (rack_disable_prr)
19457 			rack->rack_no_prr = 1;
19458 		else
19459 			rack->rack_no_prr = 0;
19460 		if (rack_gp_no_rec_chg)
19461 			rack->rc_gp_no_rec_chg = 1;
19462 		else
19463 			rack->rc_gp_no_rec_chg = 0;
19464 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
19465 			rack->r_mbuf_queue = 1;
19466 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19467 				rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19468 			rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19469 		} else {
19470 			rack->r_mbuf_queue = 0;
19471 			rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19472 		}
19473 		if (rack_enable_shared_cwnd)
19474 			rack->rack_enable_scwnd = 1;
19475 		else
19476 			rack->rack_enable_scwnd = 0;
19477 		if (rack_do_dyn_mul) {
19478 			/* When dynamic adjustment is on CA needs to start at 100% */
19479 			rack->rc_gp_dyn_mul = 1;
19480 			if (rack_do_dyn_mul >= 100)
19481 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
19482 		} else {
19483 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
19484 			rack->rc_gp_dyn_mul = 0;
19485 		}
19486 		rack->r_rr_config = 0;
19487 		rack->r_ctl.rc_no_push_at_mrtt = 0;
19488 		rack->rc_pace_to_cwnd = 0;
19489 		rack->rc_pace_fill_if_rttin_range = 0;
19490 		rack->rtt_limit_mul = 0;
19491 
19492 		if (rack_enable_hw_pacing)
19493 			rack->rack_hdw_pace_ena = 1;
19494 		else
19495 			rack->rack_hdw_pace_ena = 0;
19496 		if (rack_disable_prr)
19497 			rack->rack_no_prr = 1;
19498 		else
19499 			rack->rack_no_prr = 0;
19500 		if (rack_limits_scwnd)
19501 			rack->r_limit_scw  = 1;
19502 		else
19503 			rack->r_limit_scw  = 0;
19504 		err = 0;
19505 	}
19506 	return (err);
19507 }
19508 
19509 static int
19510 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
19511 {
19512 	struct deferred_opt_list *dol;
19513 
19514 	dol = malloc(sizeof(struct deferred_opt_list),
19515 		     M_TCPFSB, M_NOWAIT|M_ZERO);
19516 	if (dol == NULL) {
19517 		/*
19518 		 * No space yikes -- fail out..
19519 		 */
19520 		return (0);
19521 	}
19522 	dol->optname = sopt_name;
19523 	dol->optval = loptval;
19524 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
19525 	return (1);
19526 }
19527 
19528 static int
19529 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
19530 		    uint32_t optval, uint64_t loptval)
19531 {
19532 	struct epoch_tracker et;
19533 	struct sockopt sopt;
19534 	struct cc_newreno_opts opt;
19535 	struct inpcb *inp = tptoinpcb(tp);
19536 	uint64_t val;
19537 	int error = 0;
19538 	uint16_t ca, ss;
19539 
19540 	switch (sopt_name) {
19541 
19542 	case TCP_RACK_DSACK_OPT:
19543 		RACK_OPTS_INC(tcp_rack_dsack_opt);
19544 		if (optval & 0x1) {
19545 			rack->rc_rack_tmr_std_based = 1;
19546 		} else {
19547 			rack->rc_rack_tmr_std_based = 0;
19548 		}
19549 		if (optval & 0x2) {
19550 			rack->rc_rack_use_dsack = 1;
19551 		} else {
19552 			rack->rc_rack_use_dsack = 0;
19553 		}
19554 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
19555 		break;
19556 	case TCP_RACK_PACING_BETA:
19557 		RACK_OPTS_INC(tcp_rack_beta);
19558 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
19559 			/* This only works for newreno. */
19560 			error = EINVAL;
19561 			break;
19562 		}
19563 		if (rack->rc_pacing_cc_set) {
19564 			/*
19565 			 * Set them into the real CC module
19566 			 * whats in the rack pcb is the old values
19567 			 * to be used on restoral/
19568 			 */
19569 			sopt.sopt_dir = SOPT_SET;
19570 			opt.name = CC_NEWRENO_BETA;
19571 			opt.val = optval;
19572 			if (CC_ALGO(tp)->ctl_output != NULL)
19573 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
19574 			else {
19575 				error = ENOENT;
19576 				break;
19577 			}
19578 		} else {
19579 			/*
19580 			 * Not pacing yet so set it into our local
19581 			 * rack pcb storage.
19582 			 */
19583 			rack->r_ctl.rc_saved_beta.beta = optval;
19584 		}
19585 		break;
19586 	case TCP_RACK_TIMER_SLOP:
19587 		RACK_OPTS_INC(tcp_rack_timer_slop);
19588 		rack->r_ctl.timer_slop = optval;
19589 		if (rack->rc_tp->t_srtt) {
19590 			/*
19591 			 * If we have an SRTT lets update t_rxtcur
19592 			 * to have the new slop.
19593 			 */
19594 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
19595 					   rack_rto_min, rack_rto_max,
19596 					   rack->r_ctl.timer_slop);
19597 		}
19598 		break;
19599 	case TCP_RACK_PACING_BETA_ECN:
19600 		RACK_OPTS_INC(tcp_rack_beta_ecn);
19601 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
19602 			/* This only works for newreno. */
19603 			error = EINVAL;
19604 			break;
19605 		}
19606 		if (rack->rc_pacing_cc_set) {
19607 			/*
19608 			 * Set them into the real CC module
19609 			 * whats in the rack pcb is the old values
19610 			 * to be used on restoral/
19611 			 */
19612 			sopt.sopt_dir = SOPT_SET;
19613 			opt.name = CC_NEWRENO_BETA_ECN;
19614 			opt.val = optval;
19615 			if (CC_ALGO(tp)->ctl_output != NULL)
19616 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
19617 			else
19618 				error = ENOENT;
19619 		} else {
19620 			/*
19621 			 * Not pacing yet so set it into our local
19622 			 * rack pcb storage.
19623 			 */
19624 			rack->r_ctl.rc_saved_beta.beta_ecn = optval;
19625 			rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
19626 		}
19627 		break;
19628 	case TCP_DEFER_OPTIONS:
19629 		RACK_OPTS_INC(tcp_defer_opt);
19630 		if (optval) {
19631 			if (rack->gp_ready) {
19632 				/* Too late */
19633 				error = EINVAL;
19634 				break;
19635 			}
19636 			rack->defer_options = 1;
19637 		} else
19638 			rack->defer_options = 0;
19639 		break;
19640 	case TCP_RACK_MEASURE_CNT:
19641 		RACK_OPTS_INC(tcp_rack_measure_cnt);
19642 		if (optval && (optval <= 0xff)) {
19643 			rack->r_ctl.req_measurements = optval;
19644 		} else
19645 			error = EINVAL;
19646 		break;
19647 	case TCP_REC_ABC_VAL:
19648 		RACK_OPTS_INC(tcp_rec_abc_val);
19649 		if (optval > 0)
19650 			rack->r_use_labc_for_rec = 1;
19651 		else
19652 			rack->r_use_labc_for_rec = 0;
19653 		break;
19654 	case TCP_RACK_ABC_VAL:
19655 		RACK_OPTS_INC(tcp_rack_abc_val);
19656 		if ((optval > 0) && (optval < 255))
19657 			rack->rc_labc = optval;
19658 		else
19659 			error = EINVAL;
19660 		break;
19661 	case TCP_HDWR_UP_ONLY:
19662 		RACK_OPTS_INC(tcp_pacing_up_only);
19663 		if (optval)
19664 			rack->r_up_only = 1;
19665 		else
19666 			rack->r_up_only = 0;
19667 		break;
19668 	case TCP_PACING_RATE_CAP:
19669 		RACK_OPTS_INC(tcp_pacing_rate_cap);
19670 		rack->r_ctl.bw_rate_cap = loptval;
19671 		break;
19672 	case TCP_RACK_PROFILE:
19673 		RACK_OPTS_INC(tcp_profile);
19674 		error = rack_set_profile(rack, optval);
19675 		break;
19676 	case TCP_USE_CMP_ACKS:
19677 		RACK_OPTS_INC(tcp_use_cmp_acks);
19678 		if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
19679 			/* You can't turn it off once its on! */
19680 			error = EINVAL;
19681 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
19682 			rack->r_use_cmp_ack = 1;
19683 			rack->r_mbuf_queue = 1;
19684 			inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19685 		}
19686 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
19687 			inp->inp_flags2 |= INP_MBUF_ACKCMP;
19688 		break;
19689 	case TCP_SHARED_CWND_TIME_LIMIT:
19690 		RACK_OPTS_INC(tcp_lscwnd);
19691 		if (optval)
19692 			rack->r_limit_scw = 1;
19693 		else
19694 			rack->r_limit_scw = 0;
19695 		break;
19696  	case TCP_RACK_PACE_TO_FILL:
19697 		RACK_OPTS_INC(tcp_fillcw);
19698 		if (optval == 0)
19699 			rack->rc_pace_to_cwnd = 0;
19700 		else {
19701 			rack->rc_pace_to_cwnd = 1;
19702 			if (optval > 1)
19703 				rack->r_fill_less_agg = 1;
19704 		}
19705 		if ((optval >= rack_gp_rtt_maxmul) &&
19706 		    rack_gp_rtt_maxmul &&
19707 		    (optval < 0xf)) {
19708 			rack->rc_pace_fill_if_rttin_range = 1;
19709 			rack->rtt_limit_mul = optval;
19710 		} else {
19711 			rack->rc_pace_fill_if_rttin_range = 0;
19712 			rack->rtt_limit_mul = 0;
19713 		}
19714 		break;
19715 	case TCP_RACK_NO_PUSH_AT_MAX:
19716 		RACK_OPTS_INC(tcp_npush);
19717 		if (optval == 0)
19718 			rack->r_ctl.rc_no_push_at_mrtt = 0;
19719 		else if (optval < 0xff)
19720 			rack->r_ctl.rc_no_push_at_mrtt = optval;
19721 		else
19722 			error = EINVAL;
19723 		break;
19724 	case TCP_SHARED_CWND_ENABLE:
19725 		RACK_OPTS_INC(tcp_rack_scwnd);
19726 		if (optval == 0)
19727 			rack->rack_enable_scwnd = 0;
19728 		else
19729 			rack->rack_enable_scwnd = 1;
19730 		break;
19731 	case TCP_RACK_MBUF_QUEUE:
19732 		/* Now do we use the LRO mbuf-queue feature */
19733 		RACK_OPTS_INC(tcp_rack_mbufq);
19734 		if (optval || rack->r_use_cmp_ack)
19735 			rack->r_mbuf_queue = 1;
19736 		else
19737 			rack->r_mbuf_queue = 0;
19738 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19739 			inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19740 		else
19741 			inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19742 		break;
19743 	case TCP_RACK_NONRXT_CFG_RATE:
19744 		RACK_OPTS_INC(tcp_rack_cfg_rate);
19745 		if (optval == 0)
19746 			rack->rack_rec_nonrxt_use_cr = 0;
19747 		else
19748 			rack->rack_rec_nonrxt_use_cr = 1;
19749 		break;
19750 	case TCP_NO_PRR:
19751 		RACK_OPTS_INC(tcp_rack_noprr);
19752 		if (optval == 0)
19753 			rack->rack_no_prr = 0;
19754 		else if (optval == 1)
19755 			rack->rack_no_prr = 1;
19756 		else if (optval == 2)
19757 			rack->no_prr_addback = 1;
19758 		else
19759 			error = EINVAL;
19760 		break;
19761 	case TCP_TIMELY_DYN_ADJ:
19762 		RACK_OPTS_INC(tcp_timely_dyn);
19763 		if (optval == 0)
19764 			rack->rc_gp_dyn_mul = 0;
19765 		else {
19766 			rack->rc_gp_dyn_mul = 1;
19767 			if (optval >= 100) {
19768 				/*
19769 				 * If the user sets something 100 or more
19770 				 * its the gp_ca value.
19771 				 */
19772 				rack->r_ctl.rack_per_of_gp_ca  = optval;
19773 			}
19774 		}
19775 		break;
19776 	case TCP_RACK_DO_DETECTION:
19777 		RACK_OPTS_INC(tcp_rack_do_detection);
19778 		if (optval == 0)
19779 			rack->do_detection = 0;
19780 		else
19781 			rack->do_detection = 1;
19782 		break;
19783 	case TCP_RACK_TLP_USE:
19784 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19785 			error = EINVAL;
19786 			break;
19787 		}
19788 		RACK_OPTS_INC(tcp_tlp_use);
19789 		rack->rack_tlp_threshold_use = optval;
19790 		break;
19791 	case TCP_RACK_TLP_REDUCE:
19792 		/* RACK TLP cwnd reduction (bool) */
19793 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
19794 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19795 		break;
19796 	/*  Pacing related ones */
19797 	case TCP_RACK_PACE_ALWAYS:
19798 		/*
19799 		 * zero is old rack method, 1 is new
19800 		 * method using a pacing rate.
19801 		 */
19802 		RACK_OPTS_INC(tcp_rack_pace_always);
19803 		if (optval > 0) {
19804 			if (rack->rc_always_pace) {
19805 				error = EALREADY;
19806 				break;
19807 			} else if (tcp_can_enable_pacing()) {
19808 				rack->rc_always_pace = 1;
19809 				if (rack->use_fixed_rate || rack->gp_ready)
19810 					rack_set_cc_pacing(rack);
19811 			}
19812 			else {
19813 				error = ENOSPC;
19814 				break;
19815 			}
19816 		} else {
19817 			if (rack->rc_always_pace) {
19818 				tcp_decrement_paced_conn();
19819 				rack->rc_always_pace = 0;
19820 				rack_undo_cc_pacing(rack);
19821 			}
19822 		}
19823 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19824 			inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19825 		else
19826 			inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19827 		/* A rate may be set irate or other, if so set seg size */
19828 		rack_update_seg(rack);
19829 		break;
19830 	case TCP_BBR_RACK_INIT_RATE:
19831 		RACK_OPTS_INC(tcp_initial_rate);
19832 		val = optval;
19833 		/* Change from kbits per second to bytes per second */
19834 		val *= 1000;
19835 		val /= 8;
19836 		rack->r_ctl.init_rate = val;
19837 		if (rack->rc_init_win != rack_default_init_window) {
19838 			uint32_t win, snt;
19839 
19840 			/*
19841 			 * Options don't always get applied
19842 			 * in the order you think. So in order
19843 			 * to assure we update a cwnd we need
19844 			 * to check and see if we are still
19845 			 * where we should raise the cwnd.
19846 			 */
19847 			win = rc_init_window(rack);
19848 			if (SEQ_GT(tp->snd_max, tp->iss))
19849 				snt = tp->snd_max - tp->iss;
19850 			else
19851 				snt = 0;
19852 			if ((snt < win) &&
19853 			    (tp->snd_cwnd < win))
19854 				tp->snd_cwnd = win;
19855 		}
19856 		if (rack->rc_always_pace)
19857 			rack_update_seg(rack);
19858 		break;
19859 	case TCP_BBR_IWINTSO:
19860 		RACK_OPTS_INC(tcp_initial_win);
19861 		if (optval && (optval <= 0xff)) {
19862 			uint32_t win, snt;
19863 
19864 			rack->rc_init_win = optval;
19865 			win = rc_init_window(rack);
19866 			if (SEQ_GT(tp->snd_max, tp->iss))
19867 				snt = tp->snd_max - tp->iss;
19868 			else
19869 				snt = 0;
19870 			if ((snt < win) &&
19871 			    (tp->t_srtt |
19872 #ifdef NETFLIX_PEAKRATE
19873 			     tp->t_maxpeakrate |
19874 #endif
19875 			     rack->r_ctl.init_rate)) {
19876 				/*
19877 				 * We are not past the initial window
19878 				 * and we have some bases for pacing,
19879 				 * so we need to possibly adjust up
19880 				 * the cwnd. Note even if we don't set
19881 				 * the cwnd, its still ok to raise the rc_init_win
19882 				 * which can be used coming out of idle when we
19883 				 * would have a rate.
19884 				 */
19885 				if (tp->snd_cwnd < win)
19886 					tp->snd_cwnd = win;
19887 			}
19888 			if (rack->rc_always_pace)
19889 				rack_update_seg(rack);
19890 		} else
19891 			error = EINVAL;
19892 		break;
19893 	case TCP_RACK_FORCE_MSEG:
19894 		RACK_OPTS_INC(tcp_rack_force_max_seg);
19895 		if (optval)
19896 			rack->rc_force_max_seg = 1;
19897 		else
19898 			rack->rc_force_max_seg = 0;
19899 		break;
19900 	case TCP_RACK_PACE_MAX_SEG:
19901 		/* Max segments size in a pace in bytes */
19902 		RACK_OPTS_INC(tcp_rack_max_seg);
19903 		rack->rc_user_set_max_segs = optval;
19904 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
19905 		break;
19906 	case TCP_RACK_PACE_RATE_REC:
19907 		/* Set the fixed pacing rate in Bytes per second ca */
19908 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19909 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19910 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19911 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19912 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19913 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19914 		rack->use_fixed_rate = 1;
19915 		if (rack->rc_always_pace)
19916 			rack_set_cc_pacing(rack);
19917 		rack_log_pacing_delay_calc(rack,
19918 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
19919 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
19920 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19921 					   __LINE__, NULL,0);
19922 		break;
19923 
19924 	case TCP_RACK_PACE_RATE_SS:
19925 		/* Set the fixed pacing rate in Bytes per second ca */
19926 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19927 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19928 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19929 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19930 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19931 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19932 		rack->use_fixed_rate = 1;
19933 		if (rack->rc_always_pace)
19934 			rack_set_cc_pacing(rack);
19935 		rack_log_pacing_delay_calc(rack,
19936 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
19937 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
19938 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19939 					   __LINE__, NULL, 0);
19940 		break;
19941 
19942 	case TCP_RACK_PACE_RATE_CA:
19943 		/* Set the fixed pacing rate in Bytes per second ca */
19944 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
19945 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19946 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19947 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19948 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19949 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19950 		rack->use_fixed_rate = 1;
19951 		if (rack->rc_always_pace)
19952 			rack_set_cc_pacing(rack);
19953 		rack_log_pacing_delay_calc(rack,
19954 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
19955 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
19956 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19957 					   __LINE__, NULL, 0);
19958 		break;
19959 	case TCP_RACK_GP_INCREASE_REC:
19960 		RACK_OPTS_INC(tcp_gp_inc_rec);
19961 		rack->r_ctl.rack_per_of_gp_rec = optval;
19962 		rack_log_pacing_delay_calc(rack,
19963 					   rack->r_ctl.rack_per_of_gp_ss,
19964 					   rack->r_ctl.rack_per_of_gp_ca,
19965 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19966 					   __LINE__, NULL, 0);
19967 		break;
19968 	case TCP_RACK_GP_INCREASE_CA:
19969 		RACK_OPTS_INC(tcp_gp_inc_ca);
19970 		ca = optval;
19971 		if (ca < 100) {
19972 			/*
19973 			 * We don't allow any reduction
19974 			 * over the GP b/w.
19975 			 */
19976 			error = EINVAL;
19977 			break;
19978 		}
19979 		rack->r_ctl.rack_per_of_gp_ca = ca;
19980 		rack_log_pacing_delay_calc(rack,
19981 					   rack->r_ctl.rack_per_of_gp_ss,
19982 					   rack->r_ctl.rack_per_of_gp_ca,
19983 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19984 					   __LINE__, NULL, 0);
19985 		break;
19986 	case TCP_RACK_GP_INCREASE_SS:
19987 		RACK_OPTS_INC(tcp_gp_inc_ss);
19988 		ss = optval;
19989 		if (ss < 100) {
19990 			/*
19991 			 * We don't allow any reduction
19992 			 * over the GP b/w.
19993 			 */
19994 			error = EINVAL;
19995 			break;
19996 		}
19997 		rack->r_ctl.rack_per_of_gp_ss = ss;
19998 		rack_log_pacing_delay_calc(rack,
19999 					   rack->r_ctl.rack_per_of_gp_ss,
20000 					   rack->r_ctl.rack_per_of_gp_ca,
20001 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20002 					   __LINE__, NULL, 0);
20003 		break;
20004 	case TCP_RACK_RR_CONF:
20005 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
20006 		if (optval && optval <= 3)
20007 			rack->r_rr_config = optval;
20008 		else
20009 			rack->r_rr_config = 0;
20010 		break;
20011 	case TCP_HDWR_RATE_CAP:
20012 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
20013 		if (optval) {
20014 			if (rack->r_rack_hw_rate_caps == 0)
20015 				rack->r_rack_hw_rate_caps = 1;
20016 			else
20017 				error = EALREADY;
20018 		} else {
20019 			rack->r_rack_hw_rate_caps = 0;
20020 		}
20021 		break;
20022 	case TCP_BBR_HDWR_PACE:
20023 		RACK_OPTS_INC(tcp_hdwr_pacing);
20024 		if (optval){
20025 			if (rack->rack_hdrw_pacing == 0) {
20026 				rack->rack_hdw_pace_ena = 1;
20027 				rack->rack_attempt_hdwr_pace = 0;
20028 			} else
20029 				error = EALREADY;
20030 		} else {
20031 			rack->rack_hdw_pace_ena = 0;
20032 #ifdef RATELIMIT
20033 			if (rack->r_ctl.crte != NULL) {
20034 				rack->rack_hdrw_pacing = 0;
20035 				rack->rack_attempt_hdwr_pace = 0;
20036 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
20037 				rack->r_ctl.crte = NULL;
20038 			}
20039 #endif
20040 		}
20041 		break;
20042 	/*  End Pacing related ones */
20043 	case TCP_RACK_PRR_SENDALOT:
20044 		/* Allow PRR to send more than one seg */
20045 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
20046 		rack->r_ctl.rc_prr_sendalot = optval;
20047 		break;
20048 	case TCP_RACK_MIN_TO:
20049 		/* Minimum time between rack t-o's in ms */
20050 		RACK_OPTS_INC(tcp_rack_min_to);
20051 		rack->r_ctl.rc_min_to = optval;
20052 		break;
20053 	case TCP_RACK_EARLY_SEG:
20054 		/* If early recovery max segments */
20055 		RACK_OPTS_INC(tcp_rack_early_seg);
20056 		rack->r_ctl.rc_early_recovery_segs = optval;
20057 		break;
20058 	case TCP_RACK_ENABLE_HYSTART:
20059 	{
20060 		if (optval) {
20061 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
20062 			if (rack_do_hystart > RACK_HYSTART_ON)
20063 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
20064 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
20065 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
20066 		} else {
20067 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
20068 		}
20069 	}
20070 	break;
20071 	case TCP_RACK_REORD_THRESH:
20072 		/* RACK reorder threshold (shift amount) */
20073 		RACK_OPTS_INC(tcp_rack_reord_thresh);
20074 		if ((optval > 0) && (optval < 31))
20075 			rack->r_ctl.rc_reorder_shift = optval;
20076 		else
20077 			error = EINVAL;
20078 		break;
20079 	case TCP_RACK_REORD_FADE:
20080 		/* Does reordering fade after ms time */
20081 		RACK_OPTS_INC(tcp_rack_reord_fade);
20082 		rack->r_ctl.rc_reorder_fade = optval;
20083 		break;
20084 	case TCP_RACK_TLP_THRESH:
20085 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
20086 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
20087 		if (optval)
20088 			rack->r_ctl.rc_tlp_threshold = optval;
20089 		else
20090 			error = EINVAL;
20091 		break;
20092 	case TCP_BBR_USE_RACK_RR:
20093 		RACK_OPTS_INC(tcp_rack_rr);
20094 		if (optval)
20095 			rack->use_rack_rr = 1;
20096 		else
20097 			rack->use_rack_rr = 0;
20098 		break;
20099 	case TCP_FAST_RSM_HACK:
20100 		RACK_OPTS_INC(tcp_rack_fastrsm_hack);
20101 		if (optval)
20102 			rack->fast_rsm_hack = 1;
20103 		else
20104 			rack->fast_rsm_hack = 0;
20105 		break;
20106 	case TCP_RACK_PKT_DELAY:
20107 		/* RACK added ms i.e. rack-rtt + reord + N */
20108 		RACK_OPTS_INC(tcp_rack_pkt_delay);
20109 		rack->r_ctl.rc_pkt_delay = optval;
20110 		break;
20111 	case TCP_DELACK:
20112 		RACK_OPTS_INC(tcp_rack_delayed_ack);
20113 		if (optval == 0)
20114 			tp->t_delayed_ack = 0;
20115 		else
20116 			tp->t_delayed_ack = 1;
20117 		if (tp->t_flags & TF_DELACK) {
20118 			tp->t_flags &= ~TF_DELACK;
20119 			tp->t_flags |= TF_ACKNOW;
20120 			NET_EPOCH_ENTER(et);
20121 			rack_output(tp);
20122 			NET_EPOCH_EXIT(et);
20123 		}
20124 		break;
20125 
20126 	case TCP_BBR_RACK_RTT_USE:
20127 		RACK_OPTS_INC(tcp_rack_rtt_use);
20128 		if ((optval != USE_RTT_HIGH) &&
20129 		    (optval != USE_RTT_LOW) &&
20130 		    (optval != USE_RTT_AVG))
20131 			error = EINVAL;
20132 		else
20133 			rack->r_ctl.rc_rate_sample_method = optval;
20134 		break;
20135 	case TCP_DATA_AFTER_CLOSE:
20136 		RACK_OPTS_INC(tcp_data_after_close);
20137 		if (optval)
20138 			rack->rc_allow_data_af_clo = 1;
20139 		else
20140 			rack->rc_allow_data_af_clo = 0;
20141 		break;
20142 	default:
20143 		break;
20144 	}
20145 #ifdef NETFLIX_STATS
20146 	tcp_log_socket_option(tp, sopt_name, optval, error);
20147 #endif
20148 	return (error);
20149 }
20150 
20151 
20152 static void
20153 rack_apply_deferred_options(struct tcp_rack *rack)
20154 {
20155 	struct deferred_opt_list *dol, *sdol;
20156 	uint32_t s_optval;
20157 
20158 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
20159 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
20160 		/* Disadvantage of deferal is you loose the error return */
20161 		s_optval = (uint32_t)dol->optval;
20162 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
20163 		free(dol, M_TCPDO);
20164 	}
20165 }
20166 
20167 static void
20168 rack_hw_tls_change(struct tcpcb *tp, int chg)
20169 {
20170 	/*
20171 	 * HW tls state has changed.. fix all
20172 	 * rsm's in flight.
20173 	 */
20174 	struct tcp_rack *rack;
20175 	struct rack_sendmap *rsm;
20176 
20177 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20178 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
20179 		if (chg)
20180 			rsm->r_hw_tls = 1;
20181 		else
20182 			rsm->r_hw_tls = 0;
20183 	}
20184 	if (chg)
20185 		rack->r_ctl.fsb.hw_tls = 1;
20186 	else
20187 		rack->r_ctl.fsb.hw_tls = 0;
20188 }
20189 
20190 static int
20191 rack_pru_options(struct tcpcb *tp, int flags)
20192 {
20193 	if (flags & PRUS_OOB)
20194 		return (EOPNOTSUPP);
20195 	return (0);
20196 }
20197 
20198 static struct tcp_function_block __tcp_rack = {
20199 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
20200 	.tfb_tcp_output = rack_output,
20201 	.tfb_do_queued_segments = ctf_do_queued_segments,
20202 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
20203 	.tfb_tcp_do_segment = rack_do_segment,
20204 	.tfb_tcp_ctloutput = rack_ctloutput,
20205 	.tfb_tcp_fb_init = rack_init,
20206 	.tfb_tcp_fb_fini = rack_fini,
20207 	.tfb_tcp_timer_stop_all = rack_stopall,
20208 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
20209 	.tfb_tcp_handoff_ok = rack_handoff_ok,
20210 	.tfb_tcp_mtu_chg = rack_mtu_change,
20211 	.tfb_pru_options = rack_pru_options,
20212 	.tfb_hwtls_change = rack_hw_tls_change,
20213 	.tfb_compute_pipe = rack_compute_pipe,
20214 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
20215 };
20216 
20217 /*
20218  * rack_ctloutput() must drop the inpcb lock before performing copyin on
20219  * socket option arguments.  When it re-acquires the lock after the copy, it
20220  * has to revalidate that the connection is still valid for the socket
20221  * option.
20222  */
20223 static int
20224 rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt)
20225 {
20226 #ifdef INET6
20227 	struct ip6_hdr *ip6;
20228 #endif
20229 #ifdef INET
20230 	struct ip *ip;
20231 #endif
20232 	struct tcpcb *tp;
20233 	struct tcp_rack *rack;
20234 	uint64_t loptval;
20235 	int32_t error = 0, optval;
20236 
20237 	tp = intotcpcb(inp);
20238 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20239 	if (rack == NULL) {
20240 		INP_WUNLOCK(inp);
20241 		return (EINVAL);
20242 	}
20243 #ifdef INET6
20244 	ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
20245 #endif
20246 #ifdef INET
20247 	ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
20248 #endif
20249 
20250 	switch (sopt->sopt_level) {
20251 #ifdef INET6
20252 	case IPPROTO_IPV6:
20253 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
20254 		switch (sopt->sopt_name) {
20255 		case IPV6_USE_MIN_MTU:
20256 			tcp6_use_min_mtu(tp);
20257 			break;
20258 		case IPV6_TCLASS:
20259 			/*
20260 			 * The DSCP codepoint has changed, update the fsb.
20261 			 */
20262 			ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
20263 			    (rack->rc_inp->inp_flow & IPV6_FLOWINFO_MASK);
20264 			break;
20265 		}
20266 		INP_WUNLOCK(inp);
20267 		return (0);
20268 #endif
20269 #ifdef INET
20270 	case IPPROTO_IP:
20271 		switch (sopt->sopt_name) {
20272 		case IP_TOS:
20273 			/*
20274 			 * The DSCP codepoint has changed, update the fsb.
20275 			 */
20276 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
20277 			break;
20278 		case IP_TTL:
20279 			/*
20280 			 * The TTL has changed, update the fsb.
20281 			 */
20282 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
20283 			break;
20284 		}
20285 		INP_WUNLOCK(inp);
20286 		return (0);
20287 #endif
20288 	}
20289 
20290 	switch (sopt->sopt_name) {
20291 	case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
20292 	/*  Pacing related ones */
20293 	case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
20294 	case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
20295 	case TCP_BBR_IWINTSO:			/*  URL:tso_iwin */
20296 	case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
20297 	case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
20298 	case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
20299 	case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
20300 	case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
20301 	case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
20302 	case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
20303 	case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
20304 	case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
20305 	case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
20306 	case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
20307 	case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
20308 	case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
20309        /* End pacing related */
20310 	case TCP_FAST_RSM_HACK:			/*  URL:frsm_hack */
20311 	case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
20312 	case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
20313 	case TCP_RACK_MIN_TO:			/*  URL:min_to */
20314 	case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
20315 	case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
20316 	case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
20317 	case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
20318 	case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
20319 	case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
20320 	case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
20321 	case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
20322 	case TCP_RACK_DO_DETECTION:		/*  URL:detect */
20323 	case TCP_NO_PRR:			/*  URL:noprr */
20324 	case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
20325 	case TCP_DATA_AFTER_CLOSE:		/*  no URL */
20326 	case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
20327 	case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
20328 	case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
20329 	case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
20330 	case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
20331 	case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
20332 	case TCP_RACK_PROFILE:			/*  URL:profile */
20333 	case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
20334 	case TCP_RACK_ABC_VAL:			/*  URL:labc */
20335 	case TCP_REC_ABC_VAL:			/*  URL:reclabc */
20336 	case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
20337 	case TCP_DEFER_OPTIONS:			/*  URL:defer */
20338 	case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
20339 	case TCP_RACK_PACING_BETA:		/*  URL:pacing_beta */
20340 	case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
20341 	case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
20342 	case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
20343 		break;
20344 	default:
20345 		/* Filter off all unknown options to the base stack */
20346 		return (tcp_default_ctloutput(inp, sopt));
20347 		break;
20348 	}
20349 	INP_WUNLOCK(inp);
20350 	if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
20351 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
20352 		/*
20353 		 * We truncate it down to 32 bits for the socket-option trace this
20354 		 * means rates > 34Gbps won't show right, but thats probably ok.
20355 		 */
20356 		optval = (uint32_t)loptval;
20357 	} else {
20358 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
20359 		/* Save it in 64 bit form too */
20360 		loptval = optval;
20361 	}
20362 	if (error)
20363 		return (error);
20364 	INP_WLOCK(inp);
20365 	if (inp->inp_flags & INP_DROPPED) {
20366 		INP_WUNLOCK(inp);
20367 		return (ECONNRESET);
20368 	}
20369 	if (tp->t_fb != &__tcp_rack) {
20370 		INP_WUNLOCK(inp);
20371 		return (ENOPROTOOPT);
20372 	}
20373 	if (rack->defer_options && (rack->gp_ready == 0) &&
20374 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
20375 	    (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
20376 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
20377 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
20378 		/* Options are beind deferred */
20379 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
20380 			INP_WUNLOCK(inp);
20381 			return (0);
20382 		} else {
20383 			/* No memory to defer, fail */
20384 			INP_WUNLOCK(inp);
20385 			return (ENOMEM);
20386 		}
20387 	}
20388 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
20389 	INP_WUNLOCK(inp);
20390 	return (error);
20391 }
20392 
20393 static void
20394 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
20395 {
20396 
20397 	INP_WLOCK_ASSERT(tptoinpcb(tp));
20398 	bzero(ti, sizeof(*ti));
20399 
20400 	ti->tcpi_state = tp->t_state;
20401 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
20402 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
20403 	if (tp->t_flags & TF_SACK_PERMIT)
20404 		ti->tcpi_options |= TCPI_OPT_SACK;
20405 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
20406 		ti->tcpi_options |= TCPI_OPT_WSCALE;
20407 		ti->tcpi_snd_wscale = tp->snd_scale;
20408 		ti->tcpi_rcv_wscale = tp->rcv_scale;
20409 	}
20410 	if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
20411 		ti->tcpi_options |= TCPI_OPT_ECN;
20412 	if (tp->t_flags & TF_FASTOPEN)
20413 		ti->tcpi_options |= TCPI_OPT_TFO;
20414 	/* still kept in ticks is t_rcvtime */
20415 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
20416 	/* Since we hold everything in precise useconds this is easy */
20417 	ti->tcpi_rtt = tp->t_srtt;
20418 	ti->tcpi_rttvar = tp->t_rttvar;
20419 	ti->tcpi_rto = tp->t_rxtcur;
20420 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
20421 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
20422 	/*
20423 	 * FreeBSD-specific extension fields for tcp_info.
20424 	 */
20425 	ti->tcpi_rcv_space = tp->rcv_wnd;
20426 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
20427 	ti->tcpi_snd_wnd = tp->snd_wnd;
20428 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
20429 	ti->tcpi_snd_nxt = tp->snd_nxt;
20430 	ti->tcpi_snd_mss = tp->t_maxseg;
20431 	ti->tcpi_rcv_mss = tp->t_maxseg;
20432 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
20433 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
20434 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
20435 #ifdef NETFLIX_STATS
20436 	ti->tcpi_total_tlp = tp->t_sndtlppack;
20437 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
20438 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
20439 #endif
20440 #ifdef TCP_OFFLOAD
20441 	if (tp->t_flags & TF_TOE) {
20442 		ti->tcpi_options |= TCPI_OPT_TOE;
20443 		tcp_offload_tcp_info(tp, ti);
20444 	}
20445 #endif
20446 }
20447 
20448 static int
20449 rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt)
20450 {
20451 	struct tcpcb *tp;
20452 	struct tcp_rack *rack;
20453 	int32_t error, optval;
20454 	uint64_t val, loptval;
20455 	struct	tcp_info ti;
20456 	/*
20457 	 * Because all our options are either boolean or an int, we can just
20458 	 * pull everything into optval and then unlock and copy. If we ever
20459 	 * add a option that is not a int, then this will have quite an
20460 	 * impact to this routine.
20461 	 */
20462 	error = 0;
20463 	tp = intotcpcb(inp);
20464 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20465 	if (rack == NULL) {
20466 		INP_WUNLOCK(inp);
20467 		return (EINVAL);
20468 	}
20469 	switch (sopt->sopt_name) {
20470 	case TCP_INFO:
20471 		/* First get the info filled */
20472 		rack_fill_info(tp, &ti);
20473 		/* Fix up the rtt related fields if needed */
20474 		INP_WUNLOCK(inp);
20475 		error = sooptcopyout(sopt, &ti, sizeof ti);
20476 		return (error);
20477 	/*
20478 	 * Beta is the congestion control value for NewReno that influences how
20479 	 * much of a backoff happens when loss is detected. It is normally set
20480 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
20481 	 * when you exit recovery.
20482 	 */
20483 	case TCP_RACK_PACING_BETA:
20484 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
20485 			error = EINVAL;
20486 		else if (rack->rc_pacing_cc_set == 0)
20487 			optval = rack->r_ctl.rc_saved_beta.beta;
20488 		else {
20489 			/*
20490 			 * Reach out into the CC data and report back what
20491 			 * I have previously set. Yeah it looks hackish but
20492 			 * we don't want to report the saved values.
20493 			 */
20494 			if (tp->t_ccv.cc_data)
20495 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta;
20496 			else
20497 				error = EINVAL;
20498 		}
20499 		break;
20500 		/*
20501 		 * Beta_ecn is the congestion control value for NewReno that influences how
20502 		 * much of a backoff happens when a ECN mark is detected. It is normally set
20503 		 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
20504 		 * you exit recovery. Note that classic ECN has a beta of 50, it is only
20505 		 * ABE Ecn that uses this "less" value, but we do too with pacing :)
20506 		 */
20507 
20508 	case TCP_RACK_PACING_BETA_ECN:
20509 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
20510 			error = EINVAL;
20511 		else if (rack->rc_pacing_cc_set == 0)
20512 			optval = rack->r_ctl.rc_saved_beta.beta_ecn;
20513 		else {
20514 			/*
20515 			 * Reach out into the CC data and report back what
20516 			 * I have previously set. Yeah it looks hackish but
20517 			 * we don't want to report the saved values.
20518 			 */
20519 			if (tp->t_ccv.cc_data)
20520 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
20521 			else
20522 				error = EINVAL;
20523 		}
20524 		break;
20525 	case TCP_RACK_DSACK_OPT:
20526 		optval = 0;
20527 		if (rack->rc_rack_tmr_std_based) {
20528 			optval |= 1;
20529 		}
20530 		if (rack->rc_rack_use_dsack) {
20531 			optval |= 2;
20532 		}
20533 		break;
20534  	case TCP_RACK_ENABLE_HYSTART:
20535 	{
20536 		if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
20537 			optval = RACK_HYSTART_ON;
20538 			if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
20539 				optval = RACK_HYSTART_ON_W_SC;
20540 			if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
20541 				optval = RACK_HYSTART_ON_W_SC_C;
20542 		} else {
20543 			optval = RACK_HYSTART_OFF;
20544 		}
20545 	}
20546 	break;
20547 	case TCP_FAST_RSM_HACK:
20548 		optval = rack->fast_rsm_hack;
20549 		break;
20550 	case TCP_DEFER_OPTIONS:
20551 		optval = rack->defer_options;
20552 		break;
20553 	case TCP_RACK_MEASURE_CNT:
20554 		optval = rack->r_ctl.req_measurements;
20555 		break;
20556 	case TCP_REC_ABC_VAL:
20557 		optval = rack->r_use_labc_for_rec;
20558 		break;
20559 	case TCP_RACK_ABC_VAL:
20560 		optval = rack->rc_labc;
20561 		break;
20562 	case TCP_HDWR_UP_ONLY:
20563 		optval= rack->r_up_only;
20564 		break;
20565 	case TCP_PACING_RATE_CAP:
20566 		loptval = rack->r_ctl.bw_rate_cap;
20567 		break;
20568 	case TCP_RACK_PROFILE:
20569 		/* You cannot retrieve a profile, its write only */
20570 		error = EINVAL;
20571 		break;
20572 	case TCP_USE_CMP_ACKS:
20573 		optval = rack->r_use_cmp_ack;
20574 		break;
20575 	case TCP_RACK_PACE_TO_FILL:
20576 		optval = rack->rc_pace_to_cwnd;
20577 		if (optval && rack->r_fill_less_agg)
20578 			optval++;
20579 		break;
20580 	case TCP_RACK_NO_PUSH_AT_MAX:
20581 		optval = rack->r_ctl.rc_no_push_at_mrtt;
20582 		break;
20583 	case TCP_SHARED_CWND_ENABLE:
20584 		optval = rack->rack_enable_scwnd;
20585 		break;
20586 	case TCP_RACK_NONRXT_CFG_RATE:
20587 		optval = rack->rack_rec_nonrxt_use_cr;
20588 		break;
20589 	case TCP_NO_PRR:
20590 		if (rack->rack_no_prr  == 1)
20591 			optval = 1;
20592 		else if (rack->no_prr_addback == 1)
20593 			optval = 2;
20594 		else
20595 			optval = 0;
20596 		break;
20597 	case TCP_RACK_DO_DETECTION:
20598 		optval = rack->do_detection;
20599 		break;
20600 	case TCP_RACK_MBUF_QUEUE:
20601 		/* Now do we use the LRO mbuf-queue feature */
20602 		optval = rack->r_mbuf_queue;
20603 		break;
20604 	case TCP_TIMELY_DYN_ADJ:
20605 		optval = rack->rc_gp_dyn_mul;
20606 		break;
20607 	case TCP_BBR_IWINTSO:
20608 		optval = rack->rc_init_win;
20609 		break;
20610 	case TCP_RACK_TLP_REDUCE:
20611 		/* RACK TLP cwnd reduction (bool) */
20612 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
20613 		break;
20614 	case TCP_BBR_RACK_INIT_RATE:
20615 		val = rack->r_ctl.init_rate;
20616 		/* convert to kbits per sec */
20617 		val *= 8;
20618 		val /= 1000;
20619 		optval = (uint32_t)val;
20620 		break;
20621 	case TCP_RACK_FORCE_MSEG:
20622 		optval = rack->rc_force_max_seg;
20623 		break;
20624 	case TCP_RACK_PACE_MAX_SEG:
20625 		/* Max segments in a pace */
20626 		optval = rack->rc_user_set_max_segs;
20627 		break;
20628 	case TCP_RACK_PACE_ALWAYS:
20629 		/* Use the always pace method */
20630 		optval = rack->rc_always_pace;
20631 		break;
20632 	case TCP_RACK_PRR_SENDALOT:
20633 		/* Allow PRR to send more than one seg */
20634 		optval = rack->r_ctl.rc_prr_sendalot;
20635 		break;
20636 	case TCP_RACK_MIN_TO:
20637 		/* Minimum time between rack t-o's in ms */
20638 		optval = rack->r_ctl.rc_min_to;
20639 		break;
20640 	case TCP_RACK_EARLY_SEG:
20641 		/* If early recovery max segments */
20642 		optval = rack->r_ctl.rc_early_recovery_segs;
20643 		break;
20644 	case TCP_RACK_REORD_THRESH:
20645 		/* RACK reorder threshold (shift amount) */
20646 		optval = rack->r_ctl.rc_reorder_shift;
20647 		break;
20648 	case TCP_RACK_REORD_FADE:
20649 		/* Does reordering fade after ms time */
20650 		optval = rack->r_ctl.rc_reorder_fade;
20651 		break;
20652 	case TCP_BBR_USE_RACK_RR:
20653 		/* Do we use the rack cheat for rxt */
20654 		optval = rack->use_rack_rr;
20655 		break;
20656 	case TCP_RACK_RR_CONF:
20657 		optval = rack->r_rr_config;
20658 		break;
20659 	case TCP_HDWR_RATE_CAP:
20660 		optval = rack->r_rack_hw_rate_caps;
20661 		break;
20662 	case TCP_BBR_HDWR_PACE:
20663 		optval = rack->rack_hdw_pace_ena;
20664 		break;
20665 	case TCP_RACK_TLP_THRESH:
20666 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
20667 		optval = rack->r_ctl.rc_tlp_threshold;
20668 		break;
20669 	case TCP_RACK_PKT_DELAY:
20670 		/* RACK added ms i.e. rack-rtt + reord + N */
20671 		optval = rack->r_ctl.rc_pkt_delay;
20672 		break;
20673 	case TCP_RACK_TLP_USE:
20674 		optval = rack->rack_tlp_threshold_use;
20675 		break;
20676 	case TCP_RACK_PACE_RATE_CA:
20677 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
20678 		break;
20679 	case TCP_RACK_PACE_RATE_SS:
20680 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
20681 		break;
20682 	case TCP_RACK_PACE_RATE_REC:
20683 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
20684 		break;
20685 	case TCP_RACK_GP_INCREASE_SS:
20686 		optval = rack->r_ctl.rack_per_of_gp_ca;
20687 		break;
20688 	case TCP_RACK_GP_INCREASE_CA:
20689 		optval = rack->r_ctl.rack_per_of_gp_ss;
20690 		break;
20691 	case TCP_BBR_RACK_RTT_USE:
20692 		optval = rack->r_ctl.rc_rate_sample_method;
20693 		break;
20694 	case TCP_DELACK:
20695 		optval = tp->t_delayed_ack;
20696 		break;
20697 	case TCP_DATA_AFTER_CLOSE:
20698 		optval = rack->rc_allow_data_af_clo;
20699 		break;
20700 	case TCP_SHARED_CWND_TIME_LIMIT:
20701 		optval = rack->r_limit_scw;
20702 		break;
20703 	case TCP_RACK_TIMER_SLOP:
20704 		optval = rack->r_ctl.timer_slop;
20705 		break;
20706 	default:
20707 		return (tcp_default_ctloutput(inp, sopt));
20708 		break;
20709 	}
20710 	INP_WUNLOCK(inp);
20711 	if (error == 0) {
20712 		if (TCP_PACING_RATE_CAP)
20713 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
20714 		else
20715 			error = sooptcopyout(sopt, &optval, sizeof optval);
20716 	}
20717 	return (error);
20718 }
20719 
20720 static int
20721 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt)
20722 {
20723 	if (sopt->sopt_dir == SOPT_SET) {
20724 		return (rack_set_sockopt(inp, sopt));
20725 	} else if (sopt->sopt_dir == SOPT_GET) {
20726 		return (rack_get_sockopt(inp, sopt));
20727 	} else {
20728 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
20729 	}
20730 }
20731 
20732 static const char *rack_stack_names[] = {
20733 	__XSTRING(STACKNAME),
20734 #ifdef STACKALIAS
20735 	__XSTRING(STACKALIAS),
20736 #endif
20737 };
20738 
20739 static int
20740 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
20741 {
20742 	memset(mem, 0, size);
20743 	return (0);
20744 }
20745 
20746 static void
20747 rack_dtor(void *mem, int32_t size, void *arg)
20748 {
20749 
20750 }
20751 
20752 static bool rack_mod_inited = false;
20753 
20754 static int
20755 tcp_addrack(module_t mod, int32_t type, void *data)
20756 {
20757 	int32_t err = 0;
20758 	int num_stacks;
20759 
20760 	switch (type) {
20761 	case MOD_LOAD:
20762 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
20763 		    sizeof(struct rack_sendmap),
20764 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
20765 
20766 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
20767 		    sizeof(struct tcp_rack),
20768 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
20769 
20770 		sysctl_ctx_init(&rack_sysctl_ctx);
20771 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
20772 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
20773 		    OID_AUTO,
20774 #ifdef STACKALIAS
20775 		    __XSTRING(STACKALIAS),
20776 #else
20777 		    __XSTRING(STACKNAME),
20778 #endif
20779 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
20780 		    "");
20781 		if (rack_sysctl_root == NULL) {
20782 			printf("Failed to add sysctl node\n");
20783 			err = EFAULT;
20784 			goto free_uma;
20785 		}
20786 		rack_init_sysctls();
20787 		num_stacks = nitems(rack_stack_names);
20788 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
20789 		    rack_stack_names, &num_stacks);
20790 		if (err) {
20791 			printf("Failed to register %s stack name for "
20792 			    "%s module\n", rack_stack_names[num_stacks],
20793 			    __XSTRING(MODNAME));
20794 			sysctl_ctx_free(&rack_sysctl_ctx);
20795 free_uma:
20796 			uma_zdestroy(rack_zone);
20797 			uma_zdestroy(rack_pcb_zone);
20798 			rack_counter_destroy();
20799 			printf("Failed to register rack module -- err:%d\n", err);
20800 			return (err);
20801 		}
20802 		tcp_lro_reg_mbufq();
20803 		rack_mod_inited = true;
20804 		break;
20805 	case MOD_QUIESCE:
20806 		err = deregister_tcp_functions(&__tcp_rack, true, false);
20807 		break;
20808 	case MOD_UNLOAD:
20809 		err = deregister_tcp_functions(&__tcp_rack, false, true);
20810 		if (err == EBUSY)
20811 			break;
20812 		if (rack_mod_inited) {
20813 			uma_zdestroy(rack_zone);
20814 			uma_zdestroy(rack_pcb_zone);
20815 			sysctl_ctx_free(&rack_sysctl_ctx);
20816 			rack_counter_destroy();
20817 			rack_mod_inited = false;
20818 		}
20819 		tcp_lro_dereg_mbufq();
20820 		err = 0;
20821 		break;
20822 	default:
20823 		return (EOPNOTSUPP);
20824 	}
20825 	return (err);
20826 }
20827 
20828 static moduledata_t tcp_rack = {
20829 	.name = __XSTRING(MODNAME),
20830 	.evhand = tcp_addrack,
20831 	.priv = 0
20832 };
20833 
20834 MODULE_VERSION(MODNAME, 1);
20835 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
20836 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
20837