xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision 963f5dc7a30624e95d72fb7f87b8892651164e46) 
1 /*-
2  * Copyright (c) 2016-2020 Netflix, Inc.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include "opt_inet.h"
31 #include "opt_inet6.h"
32 #include "opt_ipsec.h"
33 #include "opt_tcpdebug.h"
34 #include "opt_ratelimit.h"
35 #include "opt_kern_tls.h"
36 #include <sys/param.h>
37 #include <sys/arb.h>
38 #include <sys/module.h>
39 #include <sys/kernel.h>
40 #ifdef TCP_HHOOK
41 #include <sys/hhook.h>
42 #endif
43 #include <sys/lock.h>
44 #include <sys/malloc.h>
45 #include <sys/lock.h>
46 #include <sys/mutex.h>
47 #include <sys/mbuf.h>
48 #include <sys/proc.h>		/* for proc0 declaration */
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 #include <sys/systm.h>
53 #ifdef STATS
54 #include <sys/qmath.h>
55 #include <sys/tree.h>
56 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
57 #else
58 #include <sys/tree.h>
59 #endif
60 #include <sys/refcount.h>
61 #include <sys/queue.h>
62 #include <sys/tim_filter.h>
63 #include <sys/smp.h>
64 #include <sys/kthread.h>
65 #include <sys/kern_prefetch.h>
66 #include <sys/protosw.h>
67 #ifdef TCP_ACCOUNTING
68 #include <sys/sched.h>
69 #include <machine/cpu.h>
70 #endif
71 #include <vm/uma.h>
72 
73 #include <net/route.h>
74 #include <net/route/nhop.h>
75 #include <net/vnet.h>
76 
77 #define TCPSTATES		/* for logging */
78 
79 #include <netinet/in.h>
80 #include <netinet/in_kdtrace.h>
81 #include <netinet/in_pcb.h>
82 #include <netinet/ip.h>
83 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
84 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
85 #include <netinet/ip_var.h>
86 #include <netinet/ip6.h>
87 #include <netinet6/in6_pcb.h>
88 #include <netinet6/ip6_var.h>
89 #include <netinet/tcp.h>
90 #define	TCPOUTFLAGS
91 #include <netinet/tcp_fsm.h>
92 #include <netinet/tcp_log_buf.h>
93 #include <netinet/tcp_seq.h>
94 #include <netinet/tcp_timer.h>
95 #include <netinet/tcp_var.h>
96 #include <netinet/tcp_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 TCPDEBUG
108 #include <netinet/tcp_debug.h>
109 #endif				/* TCPDEBUG */
110 #ifdef TCP_OFFLOAD
111 #include <netinet/tcp_offload.h>
112 #endif
113 #ifdef INET6
114 #include <netinet6/tcp6_var.h>
115 #endif
116 
117 #include <netipsec/ipsec_support.h>
118 
119 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
120 #include <netipsec/ipsec.h>
121 #include <netipsec/ipsec6.h>
122 #endif				/* IPSEC */
123 
124 #include <netinet/udp.h>
125 #include <netinet/udp_var.h>
126 #include <machine/in_cksum.h>
127 
128 #ifdef MAC
129 #include <security/mac/mac_framework.h>
130 #endif
131 #include "sack_filter.h"
132 #include "tcp_rack.h"
133 #include "rack_bbr_common.h"
134 
135 uma_zone_t rack_zone;
136 uma_zone_t rack_pcb_zone;
137 
138 #ifndef TICKS2SBT
139 #define	TICKS2SBT(__t)	(tick_sbt * ((sbintime_t)(__t)))
140 #endif
141 
142 VNET_DECLARE(uint32_t, newreno_beta);
143 VNET_DECLARE(uint32_t, newreno_beta_ecn);
144 #define V_newreno_beta VNET(newreno_beta)
145 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
146 
147 
148 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
149 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
150 
151 struct sysctl_ctx_list rack_sysctl_ctx;
152 struct sysctl_oid *rack_sysctl_root;
153 
154 #define CUM_ACKED 1
155 #define SACKED 2
156 
157 /*
158  * The RACK module incorporates a number of
159  * TCP ideas that have been put out into the IETF
160  * over the last few years:
161  * - Matt Mathis's Rate Halving which slowly drops
162  *    the congestion window so that the ack clock can
163  *    be maintained during a recovery.
164  * - Yuchung Cheng's RACK TCP (for which its named) that
165  *    will stop us using the number of dup acks and instead
166  *    use time as the gage of when we retransmit.
167  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
168  *    of Dukkipati et.al.
169  * RACK depends on SACK, so if an endpoint arrives that
170  * cannot do SACK the state machine below will shuttle the
171  * connection back to using the "default" TCP stack that is
172  * in FreeBSD.
173  *
174  * To implement RACK the original TCP stack was first decomposed
175  * into a functional state machine with individual states
176  * for each of the possible TCP connection states. The do_segment
177  * functions role in life is to mandate the connection supports SACK
178  * initially and then assure that the RACK state matches the conenction
179  * state before calling the states do_segment function. Each
180  * state is simplified due to the fact that the original do_segment
181  * has been decomposed and we *know* what state we are in (no
182  * switches on the state) and all tests for SACK are gone. This
183  * greatly simplifies what each state does.
184  *
185  * TCP output is also over-written with a new version since it
186  * must maintain the new rack scoreboard.
187  *
188  */
189 static int32_t rack_tlp_thresh = 1;
190 static int32_t rack_tlp_limit = 2;	/* No more than 2 TLPs w-out new data */
191 static int32_t rack_tlp_use_greater = 1;
192 static int32_t rack_reorder_thresh = 2;
193 static int32_t rack_reorder_fade = 60000000;	/* 0 - never fade, def 60,000,000
194 						 * - 60 seconds */
195 static uint8_t rack_req_measurements = 1;
196 /* Attack threshold detections */
197 static uint32_t rack_highest_sack_thresh_seen = 0;
198 static uint32_t rack_highest_move_thresh_seen = 0;
199 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
200 static int32_t rack_hw_pace_extra_slots = 2;	/* 2 extra MSS time betweens */
201 static int32_t rack_hw_rate_caps = 1; /* 1; */
202 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
203 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
204 static int32_t rack_hw_up_only = 1;
205 static int32_t rack_stats_gets_ms_rtt = 1;
206 static int32_t rack_prr_addbackmax = 2;
207 static int32_t rack_do_hystart = 0;
208 static int32_t rack_apply_rtt_with_reduced_conf = 0;
209 
210 static int32_t rack_pkt_delay = 1000;
211 static int32_t rack_send_a_lot_in_prr = 1;
212 static int32_t rack_min_to = 1000;	/* Number of microsecond  min timeout */
213 static int32_t rack_verbose_logging = 0;
214 static int32_t rack_ignore_data_after_close = 1;
215 static int32_t rack_enable_shared_cwnd = 1;
216 static int32_t rack_use_cmp_acks = 1;
217 static int32_t rack_use_fsb = 1;
218 static int32_t rack_use_rfo = 1;
219 static int32_t rack_use_rsm_rfo = 1;
220 static int32_t rack_max_abc_post_recovery = 2;
221 static int32_t rack_client_low_buf = 0;
222 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 */
223 #ifdef TCP_ACCOUNTING
224 static int32_t rack_tcp_accounting = 0;
225 #endif
226 static int32_t rack_limits_scwnd = 1;
227 static int32_t rack_enable_mqueue_for_nonpaced = 0;
228 static int32_t rack_disable_prr = 0;
229 static int32_t use_rack_rr = 1;
230 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
231 static int32_t rack_persist_min = 250000;	/* 250usec */
232 static int32_t rack_persist_max = 2000000;	/* 2 Second in usec's */
233 static int32_t rack_sack_not_required = 1;	/* set to one to allow non-sack to use rack */
234 static int32_t rack_default_init_window = 0;	/* Use system default */
235 static int32_t rack_limit_time_with_srtt = 0;
236 static int32_t rack_autosndbuf_inc = 20;	/* In percentage form */
237 static int32_t rack_enobuf_hw_boost_mult = 2;	/* How many times the hw rate we boost slot using time_between */
238 static int32_t rack_enobuf_hw_max = 12000;	/* 12 ms in usecs */
239 static int32_t rack_enobuf_hw_min = 10000;	/* 10 ms in usecs */
240 static int32_t rack_hw_rwnd_factor = 2;		/* How many max_segs the rwnd must be before we hold off sending */
241 /*
242  * Currently regular tcp has a rto_min of 30ms
243  * the backoff goes 12 times so that ends up
244  * being a total of 122.850 seconds before a
245  * connection is killed.
246  */
247 static uint32_t rack_def_data_window = 20;
248 static uint32_t rack_goal_bdp = 2;
249 static uint32_t rack_min_srtts = 1;
250 static uint32_t rack_min_measure_usec = 0;
251 static int32_t rack_tlp_min = 10000;	/* 10ms */
252 static int32_t rack_rto_min = 30000;	/* 30,000 usec same as main freebsd */
253 static int32_t rack_rto_max = 4000000;	/* 4 seconds in usec's */
254 static const int32_t rack_free_cache = 2;
255 static int32_t rack_hptsi_segments = 40;
256 static int32_t rack_rate_sample_method = USE_RTT_LOW;
257 static int32_t rack_pace_every_seg = 0;
258 static int32_t rack_delayed_ack_time = 40000;	/* 40ms in usecs */
259 static int32_t rack_slot_reduction = 4;
260 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
261 static int32_t rack_cwnd_block_ends_measure = 0;
262 static int32_t rack_rwnd_block_ends_measure = 0;
263 static int32_t rack_def_profile = 0;
264 
265 static int32_t rack_lower_cwnd_at_tlp = 0;
266 static int32_t rack_limited_retran = 0;
267 static int32_t rack_always_send_oldest = 0;
268 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
269 
270 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
271 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
272 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
273 
274 /* Probertt */
275 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
276 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
277 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
278 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
279 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
280 
281 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
282 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
283 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
284 static uint32_t rack_probertt_use_min_rtt_exit = 0;
285 static uint32_t rack_probe_rtt_sets_cwnd = 0;
286 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
287 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in usecs */
288 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
289 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction */
290 static uint32_t rack_min_probertt_hold = 40000;		/* Equal to delayed ack time */
291 static uint32_t rack_probertt_filter_life = 10000000;
292 static uint32_t rack_probertt_lower_within = 10;
293 static uint32_t rack_min_rtt_movement = 250000;	/* Must move at least 250ms (in microseconds)  to count as a lowering */
294 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
295 static int32_t rack_probertt_clear_is = 1;
296 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
297 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
298 
299 /* Part of pacing */
300 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
301 
302 /* Timely information */
303 /* Combine these two gives the range of 'no change' to bw */
304 /* ie the up/down provide the upper and lower bound */
305 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
306 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
307 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
308 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
309 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
310 static int32_t rack_gp_decrease_per = 20;	/* 20% decrease in multipler */
311 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multipler */
312 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
313 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
314 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
315 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
316 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
317 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
318 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
319 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
320 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
321 static int32_t rack_use_max_for_nobackoff = 0;
322 static int32_t rack_timely_int_timely_only = 0;	/* do interim timely's only use the timely algo (no b/w changes)? */
323 static int32_t rack_timely_no_stopping = 0;
324 static int32_t rack_down_raise_thresh = 100;
325 static int32_t rack_req_segs = 1;
326 static uint64_t rack_bw_rate_cap = 0;
327 
328 /* Weird delayed ack mode */
329 static int32_t rack_use_imac_dack = 0;
330 /* Rack specific counters */
331 counter_u64_t rack_badfr;
332 counter_u64_t rack_badfr_bytes;
333 counter_u64_t rack_rtm_prr_retran;
334 counter_u64_t rack_rtm_prr_newdata;
335 counter_u64_t rack_timestamp_mismatch;
336 counter_u64_t rack_reorder_seen;
337 counter_u64_t rack_paced_segments;
338 counter_u64_t rack_unpaced_segments;
339 counter_u64_t rack_calc_zero;
340 counter_u64_t rack_calc_nonzero;
341 counter_u64_t rack_saw_enobuf;
342 counter_u64_t rack_saw_enobuf_hw;
343 counter_u64_t rack_saw_enetunreach;
344 counter_u64_t rack_per_timer_hole;
345 counter_u64_t rack_large_ackcmp;
346 counter_u64_t rack_small_ackcmp;
347 counter_u64_t rack_persists_sends;
348 counter_u64_t rack_persists_acks;
349 counter_u64_t rack_persists_loss;
350 counter_u64_t rack_persists_lost_ends;
351 #ifdef INVARIANTS
352 counter_u64_t rack_adjust_map_bw;
353 #endif
354 /* Tail loss probe counters */
355 counter_u64_t rack_tlp_tot;
356 counter_u64_t rack_tlp_newdata;
357 counter_u64_t rack_tlp_retran;
358 counter_u64_t rack_tlp_retran_bytes;
359 counter_u64_t rack_tlp_retran_fail;
360 counter_u64_t rack_to_tot;
361 counter_u64_t rack_to_arm_rack;
362 counter_u64_t rack_to_arm_tlp;
363 counter_u64_t rack_hot_alloc;
364 counter_u64_t rack_to_alloc;
365 counter_u64_t rack_to_alloc_hard;
366 counter_u64_t rack_to_alloc_emerg;
367 counter_u64_t rack_to_alloc_limited;
368 counter_u64_t rack_alloc_limited_conns;
369 counter_u64_t rack_split_limited;
370 
371 #define MAX_NUM_OF_CNTS 13
372 counter_u64_t rack_proc_comp_ack[MAX_NUM_OF_CNTS];
373 counter_u64_t rack_multi_single_eq;
374 counter_u64_t rack_proc_non_comp_ack;
375 
376 counter_u64_t rack_fto_send;
377 counter_u64_t rack_fto_rsm_send;
378 counter_u64_t rack_nfto_resend;
379 counter_u64_t rack_non_fto_send;
380 counter_u64_t rack_extended_rfo;
381 
382 counter_u64_t rack_sack_proc_all;
383 counter_u64_t rack_sack_proc_short;
384 counter_u64_t rack_sack_proc_restart;
385 counter_u64_t rack_sack_attacks_detected;
386 counter_u64_t rack_sack_attacks_reversed;
387 counter_u64_t rack_sack_used_next_merge;
388 counter_u64_t rack_sack_splits;
389 counter_u64_t rack_sack_used_prev_merge;
390 counter_u64_t rack_sack_skipped_acked;
391 counter_u64_t rack_ack_total;
392 counter_u64_t rack_express_sack;
393 counter_u64_t rack_sack_total;
394 counter_u64_t rack_move_none;
395 counter_u64_t rack_move_some;
396 
397 counter_u64_t rack_used_tlpmethod;
398 counter_u64_t rack_used_tlpmethod2;
399 counter_u64_t rack_enter_tlp_calc;
400 counter_u64_t rack_input_idle_reduces;
401 counter_u64_t rack_collapsed_win;
402 counter_u64_t rack_tlp_does_nada;
403 counter_u64_t rack_try_scwnd;
404 counter_u64_t rack_hw_pace_init_fail;
405 counter_u64_t rack_hw_pace_lost;
406 counter_u64_t rack_sbsndptr_right;
407 counter_u64_t rack_sbsndptr_wrong;
408 
409 /* Temp CPU counters */
410 counter_u64_t rack_find_high;
411 
412 counter_u64_t rack_progress_drops;
413 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
414 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
415 
416 
417 #define	RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
418 
419 #define	RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {	\
420 	(tv) = (value) + slop;	 \
421 	if ((u_long)(tv) < (u_long)(tvmin)) \
422 		(tv) = (tvmin); \
423 	if ((u_long)(tv) > (u_long)(tvmax)) \
424 		(tv) = (tvmax); \
425 } while (0)
426 
427 static void
428 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
429 
430 static int
431 rack_process_ack(struct mbuf *m, struct tcphdr *th,
432     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
433     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
434 static int
435 rack_process_data(struct mbuf *m, struct tcphdr *th,
436     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
437     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
438 static void
439 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
440    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
441 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
442 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
443     uint8_t limit_type);
444 static struct rack_sendmap *
445 rack_check_recovery_mode(struct tcpcb *tp,
446     uint32_t tsused);
447 static void
448 rack_cong_signal(struct tcpcb *tp,
449 		 uint32_t type, uint32_t ack);
450 static void rack_counter_destroy(void);
451 static int
452 rack_ctloutput(struct socket *so, struct sockopt *sopt,
453     struct inpcb *inp, struct tcpcb *tp);
454 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
455 static void
456 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
457 static void
458 rack_do_segment(struct mbuf *m, struct tcphdr *th,
459     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
460     uint8_t iptos);
461 static void rack_dtor(void *mem, int32_t size, void *arg);
462 static void
463 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
464     uint32_t flex1, uint32_t flex2,
465     uint32_t flex3, uint32_t flex4,
466     uint32_t flex5, uint32_t flex6,
467     uint16_t flex7, uint8_t mod);
468 
469 static void
470 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
471    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
472    struct rack_sendmap *rsm, uint8_t quality);
473 static struct rack_sendmap *
474 rack_find_high_nonack(struct tcp_rack *rack,
475     struct rack_sendmap *rsm);
476 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
477 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
478 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
479 static int
480 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
481     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
482 static void
483 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
484 			    tcp_seq th_ack, int line, uint8_t quality);
485 static uint32_t
486 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
487 static int32_t rack_handoff_ok(struct tcpcb *tp);
488 static int32_t rack_init(struct tcpcb *tp);
489 static void rack_init_sysctls(void);
490 static void
491 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
492     struct tcphdr *th, int entered_rec, int dup_ack_struck);
493 static void
494 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
495     uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t ts,
496     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls);
497 
498 static void
499 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
500     struct rack_sendmap *rsm);
501 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
502 static int32_t rack_output(struct tcpcb *tp);
503 
504 static uint32_t
505 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
506     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
507     uint32_t cts, int *moved_two);
508 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
509 static void rack_remxt_tmr(struct tcpcb *tp);
510 static int
511 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
512     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
513 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
514 static int32_t rack_stopall(struct tcpcb *tp);
515 static void
516 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type,
517     uint32_t delta);
518 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type);
519 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
520 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type);
521 static uint32_t
522 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
523     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
524 static void
525 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
526     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
527 static int
528 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
529     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
530 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
531 static int
532 rack_do_close_wait(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 static int
536 rack_do_closing(struct mbuf *m, struct tcphdr *th,
537     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
538     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
539 static int
540 rack_do_established(struct mbuf *m, struct tcphdr *th,
541     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
542     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
543 static int
544 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
545     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
546     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
547 static int
548 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
549     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
550     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
551 static int
552 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
553     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
554     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
555 static int
556 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
557     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
558     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
559 static int
560 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
561     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
562     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
563 static int
564 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
565     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
566     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
567 struct rack_sendmap *
568 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
569     uint32_t tsused);
570 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
571     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
572 static void
573      tcp_rack_partialack(struct tcpcb *tp);
574 static int
575 rack_set_profile(struct tcp_rack *rack, int prof);
576 static void
577 rack_apply_deferred_options(struct tcp_rack *rack);
578 
579 int32_t rack_clear_counter=0;
580 
581 static void
582 rack_set_cc_pacing(struct tcp_rack *rack)
583 {
584 	struct sockopt sopt;
585 	struct cc_newreno_opts opt;
586 	struct newreno old, *ptr;
587 	struct tcpcb *tp;
588 	int error;
589 
590 	if (rack->rc_pacing_cc_set)
591 		return;
592 
593 	tp = rack->rc_tp;
594 	if (tp->cc_algo == NULL) {
595 		/* Tcb is leaving */
596 		printf("No cc algorithm?\n");
597 		return;
598 	}
599 	rack->rc_pacing_cc_set = 1;
600 	if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
601 		/* Not new-reno we can't play games with beta! */
602 		goto out;
603 	}
604 	ptr = ((struct newreno *)tp->ccv->cc_data);
605 	if (CC_ALGO(tp)->ctl_output == NULL)  {
606 		/* Huh, why does new_reno no longer have a set function? */
607 		goto out;
608 	}
609 	if (ptr == NULL) {
610 		/* Just the default values */
611 		old.beta = V_newreno_beta_ecn;
612 		old.beta_ecn = V_newreno_beta_ecn;
613 		old.newreno_flags = 0;
614 	} else {
615 		old.beta = ptr->beta;
616 		old.beta_ecn = ptr->beta_ecn;
617 		old.newreno_flags = ptr->newreno_flags;
618 	}
619 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
620 	sopt.sopt_dir = SOPT_SET;
621 	opt.name = CC_NEWRENO_BETA;
622 	opt.val = rack->r_ctl.rc_saved_beta.beta;
623 	error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
624 	if (error)  {
625 		goto out;
626 	}
627 	/*
628 	 * Hack alert we need to set in our newreno_flags
629 	 * so that Abe behavior is also applied.
630 	 */
631 	((struct newreno *)tp->ccv->cc_data)->newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
632 	opt.name = CC_NEWRENO_BETA_ECN;
633 	opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
634 	error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
635 	if (error) {
636 		goto out;
637 	}
638 	/* Save off the original values for restoral */
639 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
640 out:
641 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
642 		union tcp_log_stackspecific log;
643 		struct timeval tv;
644 
645 		ptr = ((struct newreno *)tp->ccv->cc_data);
646 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
647 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
648 		if (ptr) {
649 			log.u_bbr.flex1 = ptr->beta;
650 			log.u_bbr.flex2 = ptr->beta_ecn;
651 			log.u_bbr.flex3 = ptr->newreno_flags;
652 		}
653 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
654 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
655 		log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
656 		log.u_bbr.flex7 = rack->gp_ready;
657 		log.u_bbr.flex7 <<= 1;
658 		log.u_bbr.flex7 |= rack->use_fixed_rate;
659 		log.u_bbr.flex7 <<= 1;
660 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
661 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
662 		log.u_bbr.flex8 = 3;
663 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
664 			       0, &log, false, NULL, NULL, 0, &tv);
665 	}
666 }
667 
668 static void
669 rack_undo_cc_pacing(struct tcp_rack *rack)
670 {
671 	struct newreno old, *ptr;
672 	struct tcpcb *tp;
673 
674 	if (rack->rc_pacing_cc_set == 0)
675 		return;
676 	tp = rack->rc_tp;
677 	rack->rc_pacing_cc_set = 0;
678 	if (tp->cc_algo == NULL)
679 		/* Tcb is leaving */
680 		return;
681 	if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
682 		/* Not new-reno nothing to do! */
683 		return;
684 	}
685 	ptr = ((struct newreno *)tp->ccv->cc_data);
686 	if (ptr == NULL) {
687 		/*
688 		 * This happens at rack_fini() if the
689 		 * cc module gets freed on us. In that
690 		 * case we loose our "new" settings but
691 		 * thats ok, since the tcb is going away anyway.
692 		 */
693 		return;
694 	}
695 	/* Grab out our set values */
696 	memcpy(&old, ptr, sizeof(struct newreno));
697 	/* Copy back in the original values */
698 	memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno));
699 	/* Now save back the values we had set in (for when pacing is restored) */
700 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
701 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
702 		union tcp_log_stackspecific log;
703 		struct timeval tv;
704 
705 		ptr = ((struct newreno *)tp->ccv->cc_data);
706 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
707 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
708 		log.u_bbr.flex1 = ptr->beta;
709 		log.u_bbr.flex2 = ptr->beta_ecn;
710 		log.u_bbr.flex3 = ptr->newreno_flags;
711 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
712 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
713 		log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
714 		log.u_bbr.flex7 = rack->gp_ready;
715 		log.u_bbr.flex7 <<= 1;
716 		log.u_bbr.flex7 |= rack->use_fixed_rate;
717 		log.u_bbr.flex7 <<= 1;
718 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
719 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
720 		log.u_bbr.flex8 = 4;
721 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
722 			       0, &log, false, NULL, NULL, 0, &tv);
723 	}
724 }
725 
726 #ifdef NETFLIX_PEAKRATE
727 static inline void
728 rack_update_peakrate_thr(struct tcpcb *tp)
729 {
730 	/* Keep in mind that t_maxpeakrate is in B/s. */
731 	uint64_t peak;
732 	peak = uqmax((tp->t_maxseg * 2),
733 		     (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
734 	tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
735 }
736 #endif
737 
738 static int
739 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
740 {
741 	uint32_t stat;
742 	int32_t error;
743 	int i;
744 
745 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
746 	if (error || req->newptr == NULL)
747 		return error;
748 
749 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
750 	if (error)
751 		return (error);
752 	if (stat == 1) {
753 #ifdef INVARIANTS
754 		printf("Clearing RACK counters\n");
755 #endif
756 		counter_u64_zero(rack_badfr);
757 		counter_u64_zero(rack_badfr_bytes);
758 		counter_u64_zero(rack_rtm_prr_retran);
759 		counter_u64_zero(rack_rtm_prr_newdata);
760 		counter_u64_zero(rack_timestamp_mismatch);
761 		counter_u64_zero(rack_reorder_seen);
762 		counter_u64_zero(rack_tlp_tot);
763 		counter_u64_zero(rack_tlp_newdata);
764 		counter_u64_zero(rack_tlp_retran);
765 		counter_u64_zero(rack_tlp_retran_bytes);
766 		counter_u64_zero(rack_tlp_retran_fail);
767 		counter_u64_zero(rack_to_tot);
768 		counter_u64_zero(rack_to_arm_rack);
769 		counter_u64_zero(rack_to_arm_tlp);
770 		counter_u64_zero(rack_paced_segments);
771 		counter_u64_zero(rack_calc_zero);
772 		counter_u64_zero(rack_calc_nonzero);
773 		counter_u64_zero(rack_unpaced_segments);
774 		counter_u64_zero(rack_saw_enobuf);
775 		counter_u64_zero(rack_saw_enobuf_hw);
776 		counter_u64_zero(rack_saw_enetunreach);
777 		counter_u64_zero(rack_per_timer_hole);
778 		counter_u64_zero(rack_large_ackcmp);
779 		counter_u64_zero(rack_small_ackcmp);
780 		counter_u64_zero(rack_persists_sends);
781 		counter_u64_zero(rack_persists_acks);
782 		counter_u64_zero(rack_persists_loss);
783 		counter_u64_zero(rack_persists_lost_ends);
784 #ifdef INVARIANTS
785 		counter_u64_zero(rack_adjust_map_bw);
786 #endif
787 		counter_u64_zero(rack_to_alloc_hard);
788 		counter_u64_zero(rack_to_alloc_emerg);
789 		counter_u64_zero(rack_sack_proc_all);
790 		counter_u64_zero(rack_fto_send);
791 		counter_u64_zero(rack_fto_rsm_send);
792 		counter_u64_zero(rack_extended_rfo);
793 		counter_u64_zero(rack_hw_pace_init_fail);
794 		counter_u64_zero(rack_hw_pace_lost);
795 		counter_u64_zero(rack_sbsndptr_wrong);
796 		counter_u64_zero(rack_sbsndptr_right);
797 		counter_u64_zero(rack_non_fto_send);
798 		counter_u64_zero(rack_nfto_resend);
799 		counter_u64_zero(rack_sack_proc_short);
800 		counter_u64_zero(rack_sack_proc_restart);
801 		counter_u64_zero(rack_to_alloc);
802 		counter_u64_zero(rack_to_alloc_limited);
803 		counter_u64_zero(rack_alloc_limited_conns);
804 		counter_u64_zero(rack_split_limited);
805 		for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
806 			counter_u64_zero(rack_proc_comp_ack[i]);
807 		}
808 		counter_u64_zero(rack_multi_single_eq);
809 		counter_u64_zero(rack_proc_non_comp_ack);
810 		counter_u64_zero(rack_find_high);
811 		counter_u64_zero(rack_sack_attacks_detected);
812 		counter_u64_zero(rack_sack_attacks_reversed);
813 		counter_u64_zero(rack_sack_used_next_merge);
814 		counter_u64_zero(rack_sack_used_prev_merge);
815 		counter_u64_zero(rack_sack_splits);
816 		counter_u64_zero(rack_sack_skipped_acked);
817 		counter_u64_zero(rack_ack_total);
818 		counter_u64_zero(rack_express_sack);
819 		counter_u64_zero(rack_sack_total);
820 		counter_u64_zero(rack_move_none);
821 		counter_u64_zero(rack_move_some);
822 		counter_u64_zero(rack_used_tlpmethod);
823 		counter_u64_zero(rack_used_tlpmethod2);
824 		counter_u64_zero(rack_enter_tlp_calc);
825 		counter_u64_zero(rack_progress_drops);
826 		counter_u64_zero(rack_tlp_does_nada);
827 		counter_u64_zero(rack_try_scwnd);
828 		counter_u64_zero(rack_collapsed_win);
829 	}
830 	rack_clear_counter = 0;
831 	return (0);
832 }
833 
834 static void
835 rack_init_sysctls(void)
836 {
837 	int i;
838 	struct sysctl_oid *rack_counters;
839 	struct sysctl_oid *rack_attack;
840 	struct sysctl_oid *rack_pacing;
841 	struct sysctl_oid *rack_timely;
842 	struct sysctl_oid *rack_timers;
843 	struct sysctl_oid *rack_tlp;
844 	struct sysctl_oid *rack_misc;
845 	struct sysctl_oid *rack_features;
846 	struct sysctl_oid *rack_measure;
847 	struct sysctl_oid *rack_probertt;
848 	struct sysctl_oid *rack_hw_pacing;
849 
850 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
851 	    SYSCTL_CHILDREN(rack_sysctl_root),
852 	    OID_AUTO,
853 	    "sack_attack",
854 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
855 	    "Rack Sack Attack Counters and Controls");
856 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
857 	    SYSCTL_CHILDREN(rack_sysctl_root),
858 	    OID_AUTO,
859 	    "stats",
860 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
861 	    "Rack Counters");
862 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
863 	    SYSCTL_CHILDREN(rack_sysctl_root),
864 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
865 	    &rack_rate_sample_method , USE_RTT_LOW,
866 	    "What method should we use for rate sampling 0=high, 1=low ");
867 	/* Probe rtt related controls */
868 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
869 	    SYSCTL_CHILDREN(rack_sysctl_root),
870 	    OID_AUTO,
871 	    "probertt",
872 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
873 	    "ProbeRTT related Controls");
874 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
875 	    SYSCTL_CHILDREN(rack_probertt),
876 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
877 	    &rack_atexit_prtt_hbp, 130,
878 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
879 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
880 	    SYSCTL_CHILDREN(rack_probertt),
881 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
882 	    &rack_atexit_prtt, 130,
883 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
884 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
885 	    SYSCTL_CHILDREN(rack_probertt),
886 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
887 	    &rack_per_of_gp_probertt, 60,
888 	    "What percentage of goodput do we pace at in probertt");
889 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
890 	    SYSCTL_CHILDREN(rack_probertt),
891 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
892 	    &rack_per_of_gp_probertt_reduce, 10,
893 	    "What percentage of goodput do we reduce every gp_srtt");
894 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
895 	    SYSCTL_CHILDREN(rack_probertt),
896 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
897 	    &rack_per_of_gp_lowthresh, 40,
898 	    "What percentage of goodput do we allow the multiplier to fall to");
899 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
900 	    SYSCTL_CHILDREN(rack_probertt),
901 	    OID_AUTO, "time_between", CTLFLAG_RW,
902 	    & rack_time_between_probertt, 96000000,
903 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
904 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
905 	    SYSCTL_CHILDREN(rack_probertt),
906 	    OID_AUTO, "safety", CTLFLAG_RW,
907 	    &rack_probe_rtt_safety_val, 2000000,
908 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
909 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
910 	    SYSCTL_CHILDREN(rack_probertt),
911 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
912 	    &rack_probe_rtt_sets_cwnd, 0,
913 	    "Do we set the cwnd too (if always_lower is on)");
914 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
915 	    SYSCTL_CHILDREN(rack_probertt),
916 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
917 	    &rack_max_drain_wait, 2,
918 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
919 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
920 	    SYSCTL_CHILDREN(rack_probertt),
921 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
922 	    &rack_must_drain, 1,
923 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
924 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
925 	    SYSCTL_CHILDREN(rack_probertt),
926 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
927 	    &rack_probertt_use_min_rtt_entry, 1,
928 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
929 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
930 	    SYSCTL_CHILDREN(rack_probertt),
931 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
932 	    &rack_probertt_use_min_rtt_exit, 0,
933 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
934 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
935 	    SYSCTL_CHILDREN(rack_probertt),
936 	    OID_AUTO, "length_div", CTLFLAG_RW,
937 	    &rack_probertt_gpsrtt_cnt_div, 0,
938 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
939 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
940 	    SYSCTL_CHILDREN(rack_probertt),
941 	    OID_AUTO, "length_mul", CTLFLAG_RW,
942 	    &rack_probertt_gpsrtt_cnt_mul, 0,
943 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
944 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
945 	    SYSCTL_CHILDREN(rack_probertt),
946 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
947 	    &rack_min_probertt_hold, 200000,
948 	    "What is the minimum time we hold probertt at target");
949 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
950 	    SYSCTL_CHILDREN(rack_probertt),
951 	    OID_AUTO, "filter_life", CTLFLAG_RW,
952 	    &rack_probertt_filter_life, 10000000,
953 	    "What is the time for the filters life in useconds");
954 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
955 	    SYSCTL_CHILDREN(rack_probertt),
956 	    OID_AUTO, "lower_within", CTLFLAG_RW,
957 	    &rack_probertt_lower_within, 10,
958 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
959 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
960 	    SYSCTL_CHILDREN(rack_probertt),
961 	    OID_AUTO, "must_move", CTLFLAG_RW,
962 	    &rack_min_rtt_movement, 250,
963 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
964 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
965 	    SYSCTL_CHILDREN(rack_probertt),
966 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
967 	    &rack_probertt_clear_is, 1,
968 	    "Do we clear I/S counts on exiting probe-rtt");
969 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
970 	    SYSCTL_CHILDREN(rack_probertt),
971 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
972 	    &rack_max_drain_hbp, 1,
973 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
974 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
975 	    SYSCTL_CHILDREN(rack_probertt),
976 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
977 	    &rack_hbp_thresh, 3,
978 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
979 	/* Pacing related sysctls */
980 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
981 	    SYSCTL_CHILDREN(rack_sysctl_root),
982 	    OID_AUTO,
983 	    "pacing",
984 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
985 	    "Pacing related Controls");
986 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
987 	    SYSCTL_CHILDREN(rack_pacing),
988 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
989 	    &rack_max_per_above, 30,
990 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
991 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
992 	    SYSCTL_CHILDREN(rack_pacing),
993 	    OID_AUTO, "pace_to_one", CTLFLAG_RW,
994 	    &rack_pace_one_seg, 0,
995 	    "Do we allow low b/w pacing of 1MSS instead of two");
996 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
997 	    SYSCTL_CHILDREN(rack_pacing),
998 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
999 	    &rack_limit_time_with_srtt, 0,
1000 	    "Do we limit pacing time based on srtt");
1001 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1002 	    SYSCTL_CHILDREN(rack_pacing),
1003 	    OID_AUTO, "init_win", CTLFLAG_RW,
1004 	    &rack_default_init_window, 0,
1005 	    "Do we have a rack initial window 0 = system default");
1006 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1007 	    SYSCTL_CHILDREN(rack_pacing),
1008 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1009 	    &rack_per_of_gp_ss, 250,
1010 	    "If non zero, what percentage of goodput to pace at in slow start");
1011 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1012 	    SYSCTL_CHILDREN(rack_pacing),
1013 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1014 	    &rack_per_of_gp_ca, 150,
1015 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
1016 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1017 	    SYSCTL_CHILDREN(rack_pacing),
1018 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1019 	    &rack_per_of_gp_rec, 200,
1020 	    "If non zero, what percentage of goodput to pace at in recovery");
1021 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1022 	    SYSCTL_CHILDREN(rack_pacing),
1023 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1024 	    &rack_hptsi_segments, 40,
1025 	    "What size is the max for TSO segments in pacing and burst mitigation");
1026 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1027 	    SYSCTL_CHILDREN(rack_pacing),
1028 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
1029 	    &rack_slot_reduction, 4,
1030 	    "When doing only burst mitigation what is the reduce divisor");
1031 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1032 	    SYSCTL_CHILDREN(rack_sysctl_root),
1033 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
1034 	    &rack_pace_every_seg, 0,
1035 	    "If set we use pacing, if clear we use only the original burst mitigation");
1036 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1037 	    SYSCTL_CHILDREN(rack_pacing),
1038 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1039 	    &rack_bw_rate_cap, 0,
1040 	    "If set we apply this value to the absolute rate cap used by pacing");
1041 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1042 	    SYSCTL_CHILDREN(rack_sysctl_root),
1043 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1044 	    &rack_req_measurements, 1,
1045 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1046 	/* Hardware pacing */
1047 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1048 	    SYSCTL_CHILDREN(rack_sysctl_root),
1049 	    OID_AUTO,
1050 	    "hdwr_pacing",
1051 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1052 	    "Pacing related Controls");
1053 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1054 	    SYSCTL_CHILDREN(rack_hw_pacing),
1055 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1056 	    &rack_hw_rwnd_factor, 2,
1057 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1058 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1059 	    SYSCTL_CHILDREN(rack_hw_pacing),
1060 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1061 	    &rack_enobuf_hw_boost_mult, 2,
1062 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1063 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1064 	    SYSCTL_CHILDREN(rack_hw_pacing),
1065 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1066 	    &rack_enobuf_hw_max, 2,
1067 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1068 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1069 	    SYSCTL_CHILDREN(rack_hw_pacing),
1070 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1071 	    &rack_enobuf_hw_min, 2,
1072 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1073 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1074 	    SYSCTL_CHILDREN(rack_hw_pacing),
1075 	    OID_AUTO, "enable", CTLFLAG_RW,
1076 	    &rack_enable_hw_pacing, 0,
1077 	    "Should RACK attempt to use hw pacing?");
1078 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1079 	    SYSCTL_CHILDREN(rack_hw_pacing),
1080 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1081 	    &rack_hw_rate_caps, 1,
1082 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1083 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1084 	    SYSCTL_CHILDREN(rack_hw_pacing),
1085 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1086 	    &rack_hw_rate_min, 0,
1087 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1088 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1089 	    SYSCTL_CHILDREN(rack_hw_pacing),
1090 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1091 	    &rack_hw_rate_to_low, 0,
1092 	    "If we fall below this rate, dis-engage hw pacing?");
1093 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1094 	    SYSCTL_CHILDREN(rack_hw_pacing),
1095 	    OID_AUTO, "up_only", CTLFLAG_RW,
1096 	    &rack_hw_up_only, 1,
1097 	    "Do we allow hw pacing to lower the rate selected?");
1098 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1099 	    SYSCTL_CHILDREN(rack_hw_pacing),
1100 	    OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1101 	    &rack_hw_pace_extra_slots, 2,
1102 	    "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1103 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1104 	    SYSCTL_CHILDREN(rack_sysctl_root),
1105 	    OID_AUTO,
1106 	    "timely",
1107 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1108 	    "Rack Timely RTT Controls");
1109 	/* Timely based GP dynmics */
1110 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1111 	    SYSCTL_CHILDREN(rack_timely),
1112 	    OID_AUTO, "upper", CTLFLAG_RW,
1113 	    &rack_gp_per_bw_mul_up, 2,
1114 	    "Rack timely upper range for equal b/w (in percentage)");
1115 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1116 	    SYSCTL_CHILDREN(rack_timely),
1117 	    OID_AUTO, "lower", CTLFLAG_RW,
1118 	    &rack_gp_per_bw_mul_down, 4,
1119 	    "Rack timely lower range for equal b/w (in percentage)");
1120 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1121 	    SYSCTL_CHILDREN(rack_timely),
1122 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1123 	    &rack_gp_rtt_maxmul, 3,
1124 	    "Rack timely multipler of lowest rtt for rtt_max");
1125 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1126 	    SYSCTL_CHILDREN(rack_timely),
1127 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1128 	    &rack_gp_rtt_mindiv, 4,
1129 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1130 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1131 	    SYSCTL_CHILDREN(rack_timely),
1132 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1133 	    &rack_gp_rtt_minmul, 1,
1134 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1135 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1136 	    SYSCTL_CHILDREN(rack_timely),
1137 	    OID_AUTO, "decrease", CTLFLAG_RW,
1138 	    &rack_gp_decrease_per, 20,
1139 	    "Rack timely decrease percentage of our GP multiplication factor");
1140 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1141 	    SYSCTL_CHILDREN(rack_timely),
1142 	    OID_AUTO, "increase", CTLFLAG_RW,
1143 	    &rack_gp_increase_per, 2,
1144 	    "Rack timely increase perentage of our GP multiplication factor");
1145 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1146 	    SYSCTL_CHILDREN(rack_timely),
1147 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1148 	    &rack_per_lower_bound, 50,
1149 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1150 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1151 	    SYSCTL_CHILDREN(rack_timely),
1152 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1153 	    &rack_per_upper_bound_ss, 0,
1154 	    "Rack timely higest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1155 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1156 	    SYSCTL_CHILDREN(rack_timely),
1157 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1158 	    &rack_per_upper_bound_ca, 0,
1159 	    "Rack timely higest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1160 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1161 	    SYSCTL_CHILDREN(rack_timely),
1162 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1163 	    &rack_do_dyn_mul, 0,
1164 	    "Rack timely do we enable dynmaic timely goodput by default");
1165 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1166 	    SYSCTL_CHILDREN(rack_timely),
1167 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1168 	    &rack_gp_no_rec_chg, 1,
1169 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1170 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1171 	    SYSCTL_CHILDREN(rack_timely),
1172 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1173 	    &rack_timely_dec_clear, 6,
1174 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1175 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1176 	    SYSCTL_CHILDREN(rack_timely),
1177 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1178 	    &rack_timely_max_push_rise, 3,
1179 	    "Rack timely how many times do we push up with b/w increase");
1180 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1181 	    SYSCTL_CHILDREN(rack_timely),
1182 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1183 	    &rack_timely_max_push_drop, 3,
1184 	    "Rack timely how many times do we push back on b/w decent");
1185 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1186 	    SYSCTL_CHILDREN(rack_timely),
1187 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1188 	    &rack_timely_min_segs, 4,
1189 	    "Rack timely when setting the cwnd what is the min num segments");
1190 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1191 	    SYSCTL_CHILDREN(rack_timely),
1192 	    OID_AUTO, "noback_max", CTLFLAG_RW,
1193 	    &rack_use_max_for_nobackoff, 0,
1194 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1195 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1196 	    SYSCTL_CHILDREN(rack_timely),
1197 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1198 	    &rack_timely_int_timely_only, 0,
1199 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1200 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1201 	    SYSCTL_CHILDREN(rack_timely),
1202 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1203 	    &rack_timely_no_stopping, 0,
1204 	    "Rack timely don't stop increase");
1205 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1206 	    SYSCTL_CHILDREN(rack_timely),
1207 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1208 	    &rack_down_raise_thresh, 100,
1209 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1210 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1211 	    SYSCTL_CHILDREN(rack_timely),
1212 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1213 	    &rack_req_segs, 1,
1214 	    "Bottom dragging if not these many segments outstanding and room");
1215 
1216 	/* TLP and Rack related parameters */
1217 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1218 	    SYSCTL_CHILDREN(rack_sysctl_root),
1219 	    OID_AUTO,
1220 	    "tlp",
1221 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1222 	    "TLP and Rack related Controls");
1223 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1224 	    SYSCTL_CHILDREN(rack_tlp),
1225 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1226 	    &use_rack_rr, 1,
1227 	    "Do we use Rack Rapid Recovery");
1228 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1229 	    SYSCTL_CHILDREN(rack_tlp),
1230 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1231 	    &rack_max_abc_post_recovery, 2,
1232 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1233 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1234 	    SYSCTL_CHILDREN(rack_tlp),
1235 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1236 	    &rack_non_rxt_use_cr, 0,
1237 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1238 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1239 	    SYSCTL_CHILDREN(rack_tlp),
1240 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1241 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1242 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1243 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1244 	    SYSCTL_CHILDREN(rack_tlp),
1245 	    OID_AUTO, "limit", CTLFLAG_RW,
1246 	    &rack_tlp_limit, 2,
1247 	    "How many TLP's can be sent without sending new data");
1248 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1249 	    SYSCTL_CHILDREN(rack_tlp),
1250 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1251 	    &rack_tlp_use_greater, 1,
1252 	    "Should we use the rack_rtt time if its greater than srtt");
1253 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1254 	    SYSCTL_CHILDREN(rack_tlp),
1255 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1256 	    &rack_tlp_min, 10000,
1257 	    "TLP minimum timeout per the specification (in microseconds)");
1258 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1259 	    SYSCTL_CHILDREN(rack_tlp),
1260 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1261 	    &rack_always_send_oldest, 0,
1262 	    "Should we always send the oldest TLP and RACK-TLP");
1263 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1264 	    SYSCTL_CHILDREN(rack_tlp),
1265 	    OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1266 	    &rack_limited_retran, 0,
1267 	    "How many times can a rack timeout drive out sends");
1268 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1269 	    SYSCTL_CHILDREN(rack_tlp),
1270 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1271 	    &rack_lower_cwnd_at_tlp, 0,
1272 	    "When a TLP completes a retran should we enter recovery");
1273 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1274 	    SYSCTL_CHILDREN(rack_tlp),
1275 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1276 	    &rack_reorder_thresh, 2,
1277 	    "What factor for rack will be added when seeing reordering (shift right)");
1278 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1279 	    SYSCTL_CHILDREN(rack_tlp),
1280 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1281 	    &rack_tlp_thresh, 1,
1282 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1283 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1284 	    SYSCTL_CHILDREN(rack_tlp),
1285 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1286 	    &rack_reorder_fade, 60000000,
1287 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1288 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1289 	    SYSCTL_CHILDREN(rack_tlp),
1290 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1291 	    &rack_pkt_delay, 1000,
1292 	    "Extra RACK time (in microseconds) besides reordering thresh");
1293 
1294 	/* Timer related controls */
1295 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1296 	    SYSCTL_CHILDREN(rack_sysctl_root),
1297 	    OID_AUTO,
1298 	    "timers",
1299 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1300 	    "Timer related controls");
1301 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1302 	    SYSCTL_CHILDREN(rack_timers),
1303 	    OID_AUTO, "persmin", CTLFLAG_RW,
1304 	    &rack_persist_min, 250000,
1305 	    "What is the minimum time in microseconds between persists");
1306 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1307 	    SYSCTL_CHILDREN(rack_timers),
1308 	    OID_AUTO, "persmax", CTLFLAG_RW,
1309 	    &rack_persist_max, 2000000,
1310 	    "What is the largest delay in microseconds between persists");
1311 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1312 	    SYSCTL_CHILDREN(rack_timers),
1313 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1314 	    &rack_delayed_ack_time, 40000,
1315 	    "Delayed ack time (40ms in microseconds)");
1316 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1317 	    SYSCTL_CHILDREN(rack_timers),
1318 	    OID_AUTO, "minrto", CTLFLAG_RW,
1319 	    &rack_rto_min, 30000,
1320 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1321 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1322 	    SYSCTL_CHILDREN(rack_timers),
1323 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1324 	    &rack_rto_max, 4000000,
1325 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1326 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1327 	    SYSCTL_CHILDREN(rack_timers),
1328 	    OID_AUTO, "minto", CTLFLAG_RW,
1329 	    &rack_min_to, 1000,
1330 	    "Minimum rack timeout in microseconds");
1331 	/* Measure controls */
1332 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1333 	    SYSCTL_CHILDREN(rack_sysctl_root),
1334 	    OID_AUTO,
1335 	    "measure",
1336 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1337 	    "Measure related controls");
1338 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1339 	    SYSCTL_CHILDREN(rack_measure),
1340 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1341 	    &rack_wma_divisor, 8,
1342 	    "When doing b/w calculation what is the  divisor for the WMA");
1343 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1344 	    SYSCTL_CHILDREN(rack_measure),
1345 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1346 	    &rack_cwnd_block_ends_measure, 0,
1347 	    "Does a cwnd just-return end the measurement window (app limited)");
1348 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1349 	    SYSCTL_CHILDREN(rack_measure),
1350 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1351 	    &rack_rwnd_block_ends_measure, 0,
1352 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1353 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1354 	    SYSCTL_CHILDREN(rack_measure),
1355 	    OID_AUTO, "min_target", CTLFLAG_RW,
1356 	    &rack_def_data_window, 20,
1357 	    "What is the minimum target window (in mss) for a GP measurements");
1358 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1359 	    SYSCTL_CHILDREN(rack_measure),
1360 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1361 	    &rack_goal_bdp, 2,
1362 	    "What is the goal BDP to measure");
1363 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1364 	    SYSCTL_CHILDREN(rack_measure),
1365 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1366 	    &rack_min_srtts, 1,
1367 	    "What is the goal BDP to measure");
1368 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1369 	    SYSCTL_CHILDREN(rack_measure),
1370 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1371 	    &rack_min_measure_usec, 0,
1372 	    "What is the Minimum time time for a measurement if 0, this is off");
1373 	/* Features */
1374 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1375 	    SYSCTL_CHILDREN(rack_sysctl_root),
1376 	    OID_AUTO,
1377 	    "features",
1378 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1379 	    "Feature controls");
1380 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1381 	    SYSCTL_CHILDREN(rack_features),
1382 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1383 	    &rack_use_cmp_acks, 1,
1384 	    "Should RACK have LRO send compressed acks");
1385 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1386 	    SYSCTL_CHILDREN(rack_features),
1387 	    OID_AUTO, "fsb", CTLFLAG_RW,
1388 	    &rack_use_fsb, 1,
1389 	    "Should RACK use the fast send block?");
1390 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1391 	    SYSCTL_CHILDREN(rack_features),
1392 	    OID_AUTO, "rfo", CTLFLAG_RW,
1393 	    &rack_use_rfo, 1,
1394 	    "Should RACK use rack_fast_output()?");
1395 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1396 	    SYSCTL_CHILDREN(rack_features),
1397 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1398 	    &rack_use_rsm_rfo, 1,
1399 	    "Should RACK use rack_fast_rsm_output()?");
1400 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1401 	    SYSCTL_CHILDREN(rack_features),
1402 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1403 	    &rack_enable_mqueue_for_nonpaced, 0,
1404 	    "Should RACK use mbuf queuing for non-paced connections");
1405 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1406 	    SYSCTL_CHILDREN(rack_features),
1407 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1408 	    &rack_do_hystart, 0,
1409 	    "Should RACK enable HyStart++ on connections?");
1410 	/* Misc rack controls */
1411 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1412 	    SYSCTL_CHILDREN(rack_sysctl_root),
1413 	    OID_AUTO,
1414 	    "misc",
1415 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1416 	    "Misc related controls");
1417 #ifdef TCP_ACCOUNTING
1418 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1419 	    SYSCTL_CHILDREN(rack_misc),
1420 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1421 	    &rack_tcp_accounting, 0,
1422 	    "Should we turn on TCP accounting for all rack sessions?");
1423 #endif
1424 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1425 	    SYSCTL_CHILDREN(rack_misc),
1426 	    OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1427 	    &rack_apply_rtt_with_reduced_conf, 0,
1428 	    "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1429 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1430 	    SYSCTL_CHILDREN(rack_misc),
1431 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1432 	    &rack_dsack_std_based, 3,
1433 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1434 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1435 	    SYSCTL_CHILDREN(rack_misc),
1436 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1437 	    &rack_prr_addbackmax, 2,
1438 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1439 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1440 	    SYSCTL_CHILDREN(rack_misc),
1441 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1442 	    &rack_stats_gets_ms_rtt, 1,
1443 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1444 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1445 	    SYSCTL_CHILDREN(rack_misc),
1446 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1447 	    &rack_client_low_buf, 0,
1448 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1449 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1450 	    SYSCTL_CHILDREN(rack_misc),
1451 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1452 	    &rack_def_profile, 0,
1453 	    "Should RACK use a default profile (0=no, num == profile num)?");
1454 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1455 	    SYSCTL_CHILDREN(rack_misc),
1456 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1457 	    &rack_enable_shared_cwnd, 1,
1458 	    "Should RACK try to use the shared cwnd on connections where allowed");
1459 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1460 	    SYSCTL_CHILDREN(rack_misc),
1461 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1462 	    &rack_limits_scwnd, 1,
1463 	    "Should RACK place low end time limits on the shared cwnd feature");
1464 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1465 	    SYSCTL_CHILDREN(rack_misc),
1466 	    OID_AUTO, "iMac_dack", CTLFLAG_RW,
1467 	    &rack_use_imac_dack, 0,
1468 	    "Should RACK try to emulate iMac delayed ack");
1469 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1470 	    SYSCTL_CHILDREN(rack_misc),
1471 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1472 	    &rack_disable_prr, 0,
1473 	    "Should RACK not use prr and only pace (must have pacing on)");
1474 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1475 	    SYSCTL_CHILDREN(rack_misc),
1476 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1477 	    &rack_verbose_logging, 0,
1478 	    "Should RACK black box logging be verbose");
1479 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1480 	    SYSCTL_CHILDREN(rack_misc),
1481 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1482 	    &rack_ignore_data_after_close, 1,
1483 	    "Do we hold off sending a RST until all pending data is ack'd");
1484 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1485 	    SYSCTL_CHILDREN(rack_misc),
1486 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1487 	    &rack_sack_not_required, 1,
1488 	    "Do we allow rack to run on connections not supporting SACK");
1489 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1490 	    SYSCTL_CHILDREN(rack_misc),
1491 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1492 	    &rack_send_a_lot_in_prr, 1,
1493 	    "Send a lot in prr");
1494 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1495 	    SYSCTL_CHILDREN(rack_misc),
1496 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1497 	    &rack_autosndbuf_inc, 20,
1498 	    "What percentage should rack scale up its snd buffer by?");
1499 	/* Sack Attacker detection stuff */
1500 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1501 	    SYSCTL_CHILDREN(rack_attack),
1502 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1503 	    &rack_highest_sack_thresh_seen, 0,
1504 	    "Highest sack to ack ratio seen");
1505 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1506 	    SYSCTL_CHILDREN(rack_attack),
1507 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1508 	    &rack_highest_move_thresh_seen, 0,
1509 	    "Highest move to non-move ratio seen");
1510 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1511 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1512 	    SYSCTL_CHILDREN(rack_attack),
1513 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1514 	    &rack_ack_total,
1515 	    "Total number of Ack's");
1516 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1517 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1518 	    SYSCTL_CHILDREN(rack_attack),
1519 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1520 	    &rack_express_sack,
1521 	    "Total expresss number of Sack's");
1522 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1523 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1524 	    SYSCTL_CHILDREN(rack_attack),
1525 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1526 	    &rack_sack_total,
1527 	    "Total number of SACKs");
1528 	rack_move_none = counter_u64_alloc(M_WAITOK);
1529 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1530 	    SYSCTL_CHILDREN(rack_attack),
1531 	    OID_AUTO, "move_none", CTLFLAG_RD,
1532 	    &rack_move_none,
1533 	    "Total number of SACK index reuse of postions under threshold");
1534 	rack_move_some = counter_u64_alloc(M_WAITOK);
1535 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1536 	    SYSCTL_CHILDREN(rack_attack),
1537 	    OID_AUTO, "move_some", CTLFLAG_RD,
1538 	    &rack_move_some,
1539 	    "Total number of SACK index reuse of postions over threshold");
1540 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1541 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1542 	    SYSCTL_CHILDREN(rack_attack),
1543 	    OID_AUTO, "attacks", CTLFLAG_RD,
1544 	    &rack_sack_attacks_detected,
1545 	    "Total number of SACK attackers that had sack disabled");
1546 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1547 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1548 	    SYSCTL_CHILDREN(rack_attack),
1549 	    OID_AUTO, "reversed", CTLFLAG_RD,
1550 	    &rack_sack_attacks_reversed,
1551 	    "Total number of SACK attackers that were later determined false positive");
1552 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1553 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1554 	    SYSCTL_CHILDREN(rack_attack),
1555 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1556 	    &rack_sack_used_next_merge,
1557 	    "Total number of times we used the next merge");
1558 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1559 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1560 	    SYSCTL_CHILDREN(rack_attack),
1561 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1562 	    &rack_sack_used_prev_merge,
1563 	    "Total number of times we used the prev merge");
1564 	/* Counters */
1565 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1566 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1567 	    SYSCTL_CHILDREN(rack_counters),
1568 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1569 	    &rack_fto_send, "Total number of rack_fast_output sends");
1570 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1571 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1572 	    SYSCTL_CHILDREN(rack_counters),
1573 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1574 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1575 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1576 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1577 	    SYSCTL_CHILDREN(rack_counters),
1578 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1579 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1580 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1581 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1582 	    SYSCTL_CHILDREN(rack_counters),
1583 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1584 	    &rack_non_fto_send, "Total number of rack_output first sends");
1585 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1586 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1587 	    SYSCTL_CHILDREN(rack_counters),
1588 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1589 	    &rack_extended_rfo, "Total number of times we extended rfo");
1590 
1591 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1592 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1593 	    SYSCTL_CHILDREN(rack_counters),
1594 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1595 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1596 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1597 
1598 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1599 	    SYSCTL_CHILDREN(rack_counters),
1600 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1601 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1602 	rack_badfr = counter_u64_alloc(M_WAITOK);
1603 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1604 	    SYSCTL_CHILDREN(rack_counters),
1605 	    OID_AUTO, "badfr", CTLFLAG_RD,
1606 	    &rack_badfr, "Total number of bad FRs");
1607 	rack_badfr_bytes = counter_u64_alloc(M_WAITOK);
1608 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1609 	    SYSCTL_CHILDREN(rack_counters),
1610 	    OID_AUTO, "badfr_bytes", CTLFLAG_RD,
1611 	    &rack_badfr_bytes, "Total number of bad FRs");
1612 	rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK);
1613 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1614 	    SYSCTL_CHILDREN(rack_counters),
1615 	    OID_AUTO, "prrsndret", CTLFLAG_RD,
1616 	    &rack_rtm_prr_retran,
1617 	    "Total number of prr based retransmits");
1618 	rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK);
1619 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1620 	    SYSCTL_CHILDREN(rack_counters),
1621 	    OID_AUTO, "prrsndnew", CTLFLAG_RD,
1622 	    &rack_rtm_prr_newdata,
1623 	    "Total number of prr based new transmits");
1624 	rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK);
1625 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1626 	    SYSCTL_CHILDREN(rack_counters),
1627 	    OID_AUTO, "tsnf", CTLFLAG_RD,
1628 	    &rack_timestamp_mismatch,
1629 	    "Total number of timestamps that we could not find the reported ts");
1630 	rack_find_high = counter_u64_alloc(M_WAITOK);
1631 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1632 	    SYSCTL_CHILDREN(rack_counters),
1633 	    OID_AUTO, "findhigh", CTLFLAG_RD,
1634 	    &rack_find_high,
1635 	    "Total number of FIN causing find-high");
1636 	rack_reorder_seen = counter_u64_alloc(M_WAITOK);
1637 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1638 	    SYSCTL_CHILDREN(rack_counters),
1639 	    OID_AUTO, "reordering", CTLFLAG_RD,
1640 	    &rack_reorder_seen,
1641 	    "Total number of times we added delay due to reordering");
1642 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1643 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1644 	    SYSCTL_CHILDREN(rack_counters),
1645 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1646 	    &rack_tlp_tot,
1647 	    "Total number of tail loss probe expirations");
1648 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1649 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1650 	    SYSCTL_CHILDREN(rack_counters),
1651 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1652 	    &rack_tlp_newdata,
1653 	    "Total number of tail loss probe sending new data");
1654 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1655 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1656 	    SYSCTL_CHILDREN(rack_counters),
1657 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1658 	    &rack_tlp_retran,
1659 	    "Total number of tail loss probe sending retransmitted data");
1660 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1661 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1662 	    SYSCTL_CHILDREN(rack_counters),
1663 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1664 	    &rack_tlp_retran_bytes,
1665 	    "Total bytes of tail loss probe sending retransmitted data");
1666 	rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK);
1667 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1668 	    SYSCTL_CHILDREN(rack_counters),
1669 	    OID_AUTO, "tlp_retran_fail", CTLFLAG_RD,
1670 	    &rack_tlp_retran_fail,
1671 	    "Total number of tail loss probe sending retransmitted data that failed (wait for t3)");
1672 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1673 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1674 	    SYSCTL_CHILDREN(rack_counters),
1675 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1676 	    &rack_to_tot,
1677 	    "Total number of times the rack to expired");
1678 	rack_to_arm_rack = counter_u64_alloc(M_WAITOK);
1679 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1680 	    SYSCTL_CHILDREN(rack_counters),
1681 	    OID_AUTO, "arm_rack", CTLFLAG_RD,
1682 	    &rack_to_arm_rack,
1683 	    "Total number of times the rack timer armed");
1684 	rack_to_arm_tlp = counter_u64_alloc(M_WAITOK);
1685 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1686 	    SYSCTL_CHILDREN(rack_counters),
1687 	    OID_AUTO, "arm_tlp", CTLFLAG_RD,
1688 	    &rack_to_arm_tlp,
1689 	    "Total number of times the tlp timer armed");
1690 	rack_calc_zero = counter_u64_alloc(M_WAITOK);
1691 	rack_calc_nonzero = counter_u64_alloc(M_WAITOK);
1692 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1693 	    SYSCTL_CHILDREN(rack_counters),
1694 	    OID_AUTO, "calc_zero", CTLFLAG_RD,
1695 	    &rack_calc_zero,
1696 	    "Total number of times pacing time worked out to zero");
1697 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1698 	    SYSCTL_CHILDREN(rack_counters),
1699 	    OID_AUTO, "calc_nonzero", CTLFLAG_RD,
1700 	    &rack_calc_nonzero,
1701 	    "Total number of times pacing time worked out to non-zero");
1702 	rack_paced_segments = counter_u64_alloc(M_WAITOK);
1703 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1704 	    SYSCTL_CHILDREN(rack_counters),
1705 	    OID_AUTO, "paced", CTLFLAG_RD,
1706 	    &rack_paced_segments,
1707 	    "Total number of times a segment send caused hptsi");
1708 	rack_unpaced_segments = counter_u64_alloc(M_WAITOK);
1709 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1710 	    SYSCTL_CHILDREN(rack_counters),
1711 	    OID_AUTO, "unpaced", CTLFLAG_RD,
1712 	    &rack_unpaced_segments,
1713 	    "Total number of times a segment did not cause hptsi");
1714 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1715 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1716 	    SYSCTL_CHILDREN(rack_counters),
1717 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1718 	    &rack_saw_enobuf,
1719 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1720 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1721 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1722 	    SYSCTL_CHILDREN(rack_counters),
1723 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1724 	    &rack_saw_enobuf_hw,
1725 	    "Total number of times a send returned enobuf for hdwr paced connections");
1726 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1727 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1728 	    SYSCTL_CHILDREN(rack_counters),
1729 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1730 	    &rack_saw_enetunreach,
1731 	    "Total number of times a send received a enetunreachable");
1732 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1733 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1734 	    SYSCTL_CHILDREN(rack_counters),
1735 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1736 	    &rack_hot_alloc,
1737 	    "Total allocations from the top of our list");
1738 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1739 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1740 	    SYSCTL_CHILDREN(rack_counters),
1741 	    OID_AUTO, "allocs", CTLFLAG_RD,
1742 	    &rack_to_alloc,
1743 	    "Total allocations of tracking structures");
1744 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1745 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1746 	    SYSCTL_CHILDREN(rack_counters),
1747 	    OID_AUTO, "allochard", CTLFLAG_RD,
1748 	    &rack_to_alloc_hard,
1749 	    "Total allocations done with sleeping the hard way");
1750 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1751 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1752 	    SYSCTL_CHILDREN(rack_counters),
1753 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1754 	    &rack_to_alloc_emerg,
1755 	    "Total allocations done from emergency cache");
1756 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1757 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1758 	    SYSCTL_CHILDREN(rack_counters),
1759 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1760 	    &rack_to_alloc_limited,
1761 	    "Total allocations dropped due to limit");
1762 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1763 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1764 	    SYSCTL_CHILDREN(rack_counters),
1765 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1766 	    &rack_alloc_limited_conns,
1767 	    "Connections with allocations dropped due to limit");
1768 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1769 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1770 	    SYSCTL_CHILDREN(rack_counters),
1771 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1772 	    &rack_split_limited,
1773 	    "Split allocations dropped due to limit");
1774 
1775 	for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
1776 		char name[32];
1777 		sprintf(name, "cmp_ack_cnt_%d", i);
1778 		rack_proc_comp_ack[i] = counter_u64_alloc(M_WAITOK);
1779 		SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1780 				       SYSCTL_CHILDREN(rack_counters),
1781 				       OID_AUTO, name, CTLFLAG_RD,
1782 				       &rack_proc_comp_ack[i],
1783 				       "Number of compressed acks we processed");
1784 	}
1785 	rack_large_ackcmp = counter_u64_alloc(M_WAITOK);
1786 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1787 	    SYSCTL_CHILDREN(rack_counters),
1788 	    OID_AUTO, "cmp_large_mbufs", CTLFLAG_RD,
1789 	    &rack_large_ackcmp,
1790 	    "Number of TCP connections with large mbuf's for compressed acks");
1791 	rack_persists_sends = counter_u64_alloc(M_WAITOK);
1792 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1793 	    SYSCTL_CHILDREN(rack_counters),
1794 	    OID_AUTO, "persist_sends", CTLFLAG_RD,
1795 	    &rack_persists_sends,
1796 	    "Number of times we sent a persist probe");
1797 	rack_persists_acks = counter_u64_alloc(M_WAITOK);
1798 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1799 	    SYSCTL_CHILDREN(rack_counters),
1800 	    OID_AUTO, "persist_acks", CTLFLAG_RD,
1801 	    &rack_persists_acks,
1802 	    "Number of times a persist probe was acked");
1803 	rack_persists_loss = counter_u64_alloc(M_WAITOK);
1804 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1805 	    SYSCTL_CHILDREN(rack_counters),
1806 	    OID_AUTO, "persist_loss", CTLFLAG_RD,
1807 	    &rack_persists_loss,
1808 	    "Number of times we detected a lost persist probe (no ack)");
1809 	rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1810 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1811 	    SYSCTL_CHILDREN(rack_counters),
1812 	    OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1813 	    &rack_persists_lost_ends,
1814 	    "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1815 	rack_small_ackcmp = counter_u64_alloc(M_WAITOK);
1816 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1817 	    SYSCTL_CHILDREN(rack_counters),
1818 	    OID_AUTO, "cmp_small_mbufs", CTLFLAG_RD,
1819 	    &rack_small_ackcmp,
1820 	    "Number of TCP connections with small mbuf's for compressed acks");
1821 #ifdef INVARIANTS
1822 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1823 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1824 	    SYSCTL_CHILDREN(rack_counters),
1825 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1826 	    &rack_adjust_map_bw,
1827 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1828 #endif
1829 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1830 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1831 	    SYSCTL_CHILDREN(rack_counters),
1832 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1833 	    &rack_multi_single_eq,
1834 	    "Number of compressed acks total represented");
1835 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1836 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1837 	    SYSCTL_CHILDREN(rack_counters),
1838 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1839 	    &rack_proc_non_comp_ack,
1840 	    "Number of non compresseds acks that we processed");
1841 
1842 
1843 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1844 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1845 	    SYSCTL_CHILDREN(rack_counters),
1846 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1847 	    &rack_sack_proc_all,
1848 	    "Total times we had to walk whole list for sack processing");
1849 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1850 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1851 	    SYSCTL_CHILDREN(rack_counters),
1852 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1853 	    &rack_sack_proc_restart,
1854 	    "Total times we had to walk whole list due to a restart");
1855 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1856 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1857 	    SYSCTL_CHILDREN(rack_counters),
1858 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1859 	    &rack_sack_proc_short,
1860 	    "Total times we took shortcut for sack processing");
1861 	rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK);
1862 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1863 	    SYSCTL_CHILDREN(rack_counters),
1864 	    OID_AUTO, "tlp_calc_entered", CTLFLAG_RD,
1865 	    &rack_enter_tlp_calc,
1866 	    "Total times we called calc-tlp");
1867 	rack_used_tlpmethod = counter_u64_alloc(M_WAITOK);
1868 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1869 	    SYSCTL_CHILDREN(rack_counters),
1870 	    OID_AUTO, "hit_tlp_method", CTLFLAG_RD,
1871 	    &rack_used_tlpmethod,
1872 	    "Total number of runt sacks");
1873 	rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK);
1874 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1875 	    SYSCTL_CHILDREN(rack_counters),
1876 	    OID_AUTO, "hit_tlp_method2", CTLFLAG_RD,
1877 	    &rack_used_tlpmethod2,
1878 	    "Total number of times we hit TLP method 2");
1879 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1880 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1881 	    SYSCTL_CHILDREN(rack_attack),
1882 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1883 	    &rack_sack_skipped_acked,
1884 	    "Total number of times we skipped previously sacked");
1885 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1886 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1887 	    SYSCTL_CHILDREN(rack_attack),
1888 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1889 	    &rack_sack_splits,
1890 	    "Total number of times we did the old fashion tree split");
1891 	rack_progress_drops = counter_u64_alloc(M_WAITOK);
1892 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1893 	    SYSCTL_CHILDREN(rack_counters),
1894 	    OID_AUTO, "prog_drops", CTLFLAG_RD,
1895 	    &rack_progress_drops,
1896 	    "Total number of progress drops");
1897 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1898 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1899 	    SYSCTL_CHILDREN(rack_counters),
1900 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1901 	    &rack_input_idle_reduces,
1902 	    "Total number of idle reductions on input");
1903 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1904 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1905 	    SYSCTL_CHILDREN(rack_counters),
1906 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1907 	    &rack_collapsed_win,
1908 	    "Total number of collapsed windows");
1909 	rack_tlp_does_nada = counter_u64_alloc(M_WAITOK);
1910 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1911 	    SYSCTL_CHILDREN(rack_counters),
1912 	    OID_AUTO, "tlp_nada", CTLFLAG_RD,
1913 	    &rack_tlp_does_nada,
1914 	    "Total number of nada tlp calls");
1915 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1916 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1917 	    SYSCTL_CHILDREN(rack_counters),
1918 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1919 	    &rack_try_scwnd,
1920 	    "Total number of scwnd attempts");
1921 
1922 	rack_per_timer_hole = counter_u64_alloc(M_WAITOK);
1923 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1924 	    SYSCTL_CHILDREN(rack_counters),
1925 	    OID_AUTO, "timer_hole", CTLFLAG_RD,
1926 	    &rack_per_timer_hole,
1927 	    "Total persists start in timer hole");
1928 
1929 	rack_sbsndptr_wrong = counter_u64_alloc(M_WAITOK);
1930 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1931 	    SYSCTL_CHILDREN(rack_counters),
1932 	    OID_AUTO, "sndptr_wrong", CTLFLAG_RD,
1933 	    &rack_sbsndptr_wrong, "Total number of times the saved sbsndptr was incorret");
1934 	rack_sbsndptr_right = counter_u64_alloc(M_WAITOK);
1935 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1936 	    SYSCTL_CHILDREN(rack_counters),
1937 	    OID_AUTO, "sndptr_right", CTLFLAG_RD,
1938 	    &rack_sbsndptr_right, "Total number of times the saved sbsndptr was corret");
1939 
1940 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1941 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1942 	    OID_AUTO, "outsize", CTLFLAG_RD,
1943 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1944 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1945 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1946 	    OID_AUTO, "opts", CTLFLAG_RD,
1947 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1948 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1949 	    SYSCTL_CHILDREN(rack_sysctl_root),
1950 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1951 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1952 }
1953 
1954 static __inline int
1955 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1956 {
1957 	if (SEQ_GEQ(b->r_start, a->r_start) &&
1958 	    SEQ_LT(b->r_start, a->r_end)) {
1959 		/*
1960 		 * The entry b is within the
1961 		 * block a. i.e.:
1962 		 * a --   |-------------|
1963 		 * b --   |----|
1964 		 * <or>
1965 		 * b --       |------|
1966 		 * <or>
1967 		 * b --       |-----------|
1968 		 */
1969 		return (0);
1970 	} else if (SEQ_GEQ(b->r_start, a->r_end)) {
1971 		/*
1972 		 * b falls as either the next
1973 		 * sequence block after a so a
1974 		 * is said to be smaller than b.
1975 		 * i.e:
1976 		 * a --   |------|
1977 		 * b --          |--------|
1978 		 * or
1979 		 * b --              |-----|
1980 		 */
1981 		return (1);
1982 	}
1983 	/*
1984 	 * Whats left is where a is
1985 	 * larger than b. i.e:
1986 	 * a --         |-------|
1987 	 * b --  |---|
1988 	 * or even possibly
1989 	 * b --   |--------------|
1990 	 */
1991 	return (-1);
1992 }
1993 
1994 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1995 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1996 
1997 static uint32_t
1998 rc_init_window(struct tcp_rack *rack)
1999 {
2000 	uint32_t win;
2001 
2002 	if (rack->rc_init_win == 0) {
2003 		/*
2004 		 * Nothing set by the user, use the system stack
2005 		 * default.
2006 		 */
2007 		return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
2008 	}
2009 	win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
2010 	return (win);
2011 }
2012 
2013 static uint64_t
2014 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
2015 {
2016 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
2017 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
2018 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2019 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
2020 	else
2021 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
2022 }
2023 
2024 static uint64_t
2025 rack_get_bw(struct tcp_rack *rack)
2026 {
2027 	if (rack->use_fixed_rate) {
2028 		/* Return the fixed pacing rate */
2029 		return (rack_get_fixed_pacing_bw(rack));
2030 	}
2031 	if (rack->r_ctl.gp_bw == 0) {
2032 		/*
2033 		 * We have yet no b/w measurement,
2034 		 * if we have a user set initial bw
2035 		 * return it. If we don't have that and
2036 		 * we have an srtt, use the tcp IW (10) to
2037 		 * calculate a fictional b/w over the SRTT
2038 		 * which is more or less a guess. Note
2039 		 * we don't use our IW from rack on purpose
2040 		 * so if we have like IW=30, we are not
2041 		 * calculating a "huge" b/w.
2042 		 */
2043 		uint64_t bw, srtt;
2044 		if (rack->r_ctl.init_rate)
2045 			return (rack->r_ctl.init_rate);
2046 
2047 		/* Has the user set a max peak rate? */
2048 #ifdef NETFLIX_PEAKRATE
2049 		if (rack->rc_tp->t_maxpeakrate)
2050 			return (rack->rc_tp->t_maxpeakrate);
2051 #endif
2052 		/* Ok lets come up with the IW guess, if we have a srtt */
2053 		if (rack->rc_tp->t_srtt == 0) {
2054 			/*
2055 			 * Go with old pacing method
2056 			 * i.e. burst mitigation only.
2057 			 */
2058 			return (0);
2059 		}
2060 		/* Ok lets get the initial TCP win (not racks) */
2061 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2062 		srtt = (uint64_t)rack->rc_tp->t_srtt;
2063 		bw *= (uint64_t)USECS_IN_SECOND;
2064 		bw /= srtt;
2065 		if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2066 			bw = rack->r_ctl.bw_rate_cap;
2067 		return (bw);
2068 	} else {
2069 		uint64_t bw;
2070 
2071 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2072 			/* Averaging is done, we can return the value */
2073 			bw = rack->r_ctl.gp_bw;
2074 		} else {
2075 			/* Still doing initial average must calculate */
2076 			bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
2077 		}
2078 #ifdef NETFLIX_PEAKRATE
2079 		if ((rack->rc_tp->t_maxpeakrate) &&
2080 		    (bw > rack->rc_tp->t_maxpeakrate)) {
2081 			/* The user has set a peak rate to pace at
2082 			 * don't allow us to pace faster than that.
2083 			 */
2084 			return (rack->rc_tp->t_maxpeakrate);
2085 		}
2086 #endif
2087 		if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2088 			bw = rack->r_ctl.bw_rate_cap;
2089 		return (bw);
2090 	}
2091 }
2092 
2093 static uint16_t
2094 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2095 {
2096 	if (rack->use_fixed_rate) {
2097 		return (100);
2098 	} else if (rack->in_probe_rtt && (rsm == NULL))
2099 		return (rack->r_ctl.rack_per_of_gp_probertt);
2100 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2101 		  rack->r_ctl.rack_per_of_gp_rec)) {
2102 		if (rsm) {
2103 			/* a retransmission always use the recovery rate */
2104 			return (rack->r_ctl.rack_per_of_gp_rec);
2105 		} else if (rack->rack_rec_nonrxt_use_cr) {
2106 			/* Directed to use the configured rate */
2107 			goto configured_rate;
2108 		} else if (rack->rack_no_prr &&
2109 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2110 			/* No PRR, lets just use the b/w estimate only */
2111 			return (100);
2112 		} else {
2113 			/*
2114 			 * Here we may have a non-retransmit but we
2115 			 * have no overrides, so just use the recovery
2116 			 * rate (prr is in effect).
2117 			 */
2118 			return (rack->r_ctl.rack_per_of_gp_rec);
2119 		}
2120 	}
2121 configured_rate:
2122 	/* For the configured rate we look at our cwnd vs the ssthresh */
2123 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2124 		return (rack->r_ctl.rack_per_of_gp_ss);
2125 	else
2126 		return (rack->r_ctl.rack_per_of_gp_ca);
2127 }
2128 
2129 static void
2130 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
2131 {
2132 	/*
2133 	 * Types of logs (mod value)
2134 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
2135 	 * 2 = a dsack round begins, persist is reset to 16.
2136 	 * 3 = a dsack round ends
2137 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
2138 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
2139 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
2140 	 */
2141 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2142 		union tcp_log_stackspecific log;
2143 		struct timeval tv;
2144 
2145 		memset(&log, 0, sizeof(log));
2146 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2147 		log.u_bbr.flex1 <<= 1;
2148 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2149 		log.u_bbr.flex1 <<= 1;
2150 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2151 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2152 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2153 		log.u_bbr.flex4 = flex4;
2154 		log.u_bbr.flex5 = flex5;
2155 		log.u_bbr.flex6 = flex6;
2156 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2157 		log.u_bbr.flex8 = mod;
2158 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
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 		    RACK_DSACK_HANDLING, 0,
2163 		    0, &log, false, &tv);
2164 	}
2165 }
2166 
2167 static void
2168 rack_log_hdwr_pacing(struct tcp_rack *rack,
2169 		     uint64_t rate, uint64_t hw_rate, int line,
2170 		     int error, uint16_t mod)
2171 {
2172 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2173 		union tcp_log_stackspecific log;
2174 		struct timeval tv;
2175 		const struct ifnet *ifp;
2176 
2177 		memset(&log, 0, sizeof(log));
2178 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2179 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2180 		if (rack->r_ctl.crte) {
2181 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2182 		} else if (rack->rc_inp->inp_route.ro_nh &&
2183 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2184 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2185 		} else
2186 			ifp = NULL;
2187 		if (ifp) {
2188 			log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2189 			log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2190 		}
2191 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2192 		log.u_bbr.bw_inuse = rate;
2193 		log.u_bbr.flex5 = line;
2194 		log.u_bbr.flex6 = error;
2195 		log.u_bbr.flex7 = mod;
2196 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2197 		log.u_bbr.flex8 = rack->use_fixed_rate;
2198 		log.u_bbr.flex8 <<= 1;
2199 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2200 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2201 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2202 		if (rack->r_ctl.crte)
2203 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2204 		else
2205 			log.u_bbr.cur_del_rate = 0;
2206 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2207 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2208 		    &rack->rc_inp->inp_socket->so_rcv,
2209 		    &rack->rc_inp->inp_socket->so_snd,
2210 		    BBR_LOG_HDWR_PACE, 0,
2211 		    0, &log, false, &tv);
2212 	}
2213 }
2214 
2215 static uint64_t
2216 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2217 {
2218 	/*
2219 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2220 	 */
2221 	uint64_t bw_est, high_rate;
2222 	uint64_t gain;
2223 
2224 	gain = (uint64_t)rack_get_output_gain(rack, rsm);
2225 	bw_est = bw * gain;
2226 	bw_est /= (uint64_t)100;
2227 	/* Never fall below the minimum (def 64kbps) */
2228 	if (bw_est < RACK_MIN_BW)
2229 		bw_est = RACK_MIN_BW;
2230 	if (rack->r_rack_hw_rate_caps) {
2231 		/* Rate caps are in place */
2232 		if (rack->r_ctl.crte != NULL) {
2233 			/* We have a hdwr rate already */
2234 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2235 			if (bw_est >= high_rate) {
2236 				/* We are capping bw at the highest rate table entry */
2237 				rack_log_hdwr_pacing(rack,
2238 						     bw_est, high_rate, __LINE__,
2239 						     0, 3);
2240 				bw_est = high_rate;
2241 				if (capped)
2242 					*capped = 1;
2243 			}
2244 		} else if ((rack->rack_hdrw_pacing == 0) &&
2245 			   (rack->rack_hdw_pace_ena) &&
2246 			   (rack->rack_attempt_hdwr_pace == 0) &&
2247 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2248 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2249 			/*
2250 			 * Special case, we have not yet attempted hardware
2251 			 * pacing, and yet we may, when we do, find out if we are
2252 			 * above the highest rate. We need to know the maxbw for the interface
2253 			 * in question (if it supports ratelimiting). We get back
2254 			 * a 0, if the interface is not found in the RL lists.
2255 			 */
2256 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2257 			if (high_rate) {
2258 				/* Yep, we have a rate is it above this rate? */
2259 				if (bw_est > high_rate) {
2260 					bw_est = high_rate;
2261 					if (capped)
2262 						*capped = 1;
2263 				}
2264 			}
2265 		}
2266 	}
2267 	return (bw_est);
2268 }
2269 
2270 static void
2271 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2272 {
2273 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2274 		union tcp_log_stackspecific log;
2275 		struct timeval tv;
2276 
2277 		if ((mod != 1) && (rack_verbose_logging == 0)) {
2278 			/*
2279 			 * We get 3 values currently for mod
2280 			 * 1 - We are retransmitting and this tells the reason.
2281 			 * 2 - We are clearing a dup-ack count.
2282 			 * 3 - We are incrementing a dup-ack count.
2283 			 *
2284 			 * The clear/increment are only logged
2285 			 * if you have BBverbose on.
2286 			 */
2287 			return;
2288 		}
2289 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2290 		log.u_bbr.flex1 = tsused;
2291 		log.u_bbr.flex2 = thresh;
2292 		log.u_bbr.flex3 = rsm->r_flags;
2293 		log.u_bbr.flex4 = rsm->r_dupack;
2294 		log.u_bbr.flex5 = rsm->r_start;
2295 		log.u_bbr.flex6 = rsm->r_end;
2296 		log.u_bbr.flex8 = mod;
2297 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2298 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2299 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2300 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2301 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2302 		log.u_bbr.pacing_gain = rack->r_must_retran;
2303 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2304 		    &rack->rc_inp->inp_socket->so_rcv,
2305 		    &rack->rc_inp->inp_socket->so_snd,
2306 		    BBR_LOG_SETTINGS_CHG, 0,
2307 		    0, &log, false, &tv);
2308 	}
2309 }
2310 
2311 static void
2312 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2313 {
2314 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2315 		union tcp_log_stackspecific log;
2316 		struct timeval tv;
2317 
2318 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2319 		log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2320 		log.u_bbr.flex2 = to;
2321 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2322 		log.u_bbr.flex4 = slot;
2323 		log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2324 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2325 		log.u_bbr.flex7 = rack->rc_in_persist;
2326 		log.u_bbr.flex8 = which;
2327 		if (rack->rack_no_prr)
2328 			log.u_bbr.pkts_out = 0;
2329 		else
2330 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2331 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2332 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2333 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2334 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2335 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2336 		log.u_bbr.pacing_gain = rack->r_must_retran;
2337 		log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2338 		log.u_bbr.lost = rack_rto_min;
2339 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2340 		    &rack->rc_inp->inp_socket->so_rcv,
2341 		    &rack->rc_inp->inp_socket->so_snd,
2342 		    BBR_LOG_TIMERSTAR, 0,
2343 		    0, &log, false, &tv);
2344 	}
2345 }
2346 
2347 static void
2348 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2349 {
2350 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2351 		union tcp_log_stackspecific log;
2352 		struct timeval tv;
2353 
2354 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2355 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2356 		log.u_bbr.flex8 = to_num;
2357 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2358 		log.u_bbr.flex2 = rack->rc_rack_rtt;
2359 		if (rsm == NULL)
2360 			log.u_bbr.flex3 = 0;
2361 		else
2362 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2363 		if (rack->rack_no_prr)
2364 			log.u_bbr.flex5 = 0;
2365 		else
2366 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2367 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2368 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2369 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2370 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2371 		log.u_bbr.pacing_gain = rack->r_must_retran;
2372 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2373 		    &rack->rc_inp->inp_socket->so_rcv,
2374 		    &rack->rc_inp->inp_socket->so_snd,
2375 		    BBR_LOG_RTO, 0,
2376 		    0, &log, false, &tv);
2377 	}
2378 }
2379 
2380 static void
2381 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2382 		 struct rack_sendmap *prev,
2383 		 struct rack_sendmap *rsm,
2384 		 struct rack_sendmap *next,
2385 		 int flag, uint32_t th_ack, int line)
2386 {
2387 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2388 		union tcp_log_stackspecific log;
2389 		struct timeval tv;
2390 
2391 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2392 		log.u_bbr.flex8 = flag;
2393 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2394 		log.u_bbr.cur_del_rate = (uint64_t)prev;
2395 		log.u_bbr.delRate = (uint64_t)rsm;
2396 		log.u_bbr.rttProp = (uint64_t)next;
2397 		log.u_bbr.flex7 = 0;
2398 		if (prev) {
2399 			log.u_bbr.flex1 = prev->r_start;
2400 			log.u_bbr.flex2 = prev->r_end;
2401 			log.u_bbr.flex7 |= 0x4;
2402 		}
2403 		if (rsm) {
2404 			log.u_bbr.flex3 = rsm->r_start;
2405 			log.u_bbr.flex4 = rsm->r_end;
2406 			log.u_bbr.flex7 |= 0x2;
2407 		}
2408 		if (next) {
2409 			log.u_bbr.flex5 = next->r_start;
2410 			log.u_bbr.flex6 = next->r_end;
2411 			log.u_bbr.flex7 |= 0x1;
2412 		}
2413 		log.u_bbr.applimited = line;
2414 		log.u_bbr.pkts_out = th_ack;
2415 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2416 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2417 		if (rack->rack_no_prr)
2418 			log.u_bbr.lost = 0;
2419 		else
2420 			log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2421 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2422 		    &rack->rc_inp->inp_socket->so_rcv,
2423 		    &rack->rc_inp->inp_socket->so_snd,
2424 		    TCP_LOG_MAPCHG, 0,
2425 		    0, &log, false, &tv);
2426 	}
2427 }
2428 
2429 static void
2430 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2431 		 struct rack_sendmap *rsm, int conf)
2432 {
2433 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2434 		union tcp_log_stackspecific log;
2435 		struct timeval tv;
2436 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2437 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2438 		log.u_bbr.flex1 = t;
2439 		log.u_bbr.flex2 = len;
2440 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2441 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2442 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2443 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2444 		log.u_bbr.flex7 = conf;
2445 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2446 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2447 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2448 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2449 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2450 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2451 		if (rsm) {
2452 			log.u_bbr.pkt_epoch = rsm->r_start;
2453 			log.u_bbr.lost = rsm->r_end;
2454 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2455 			/* We loose any upper of the 24 bits */
2456 			log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2457 		} else {
2458 			/* Its a SYN */
2459 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2460 			log.u_bbr.lost = 0;
2461 			log.u_bbr.cwnd_gain = 0;
2462 			log.u_bbr.pacing_gain = 0;
2463 		}
2464 		/* Write out general bits of interest rrs here */
2465 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2466 		log.u_bbr.use_lt_bw <<= 1;
2467 		log.u_bbr.use_lt_bw |= rack->forced_ack;
2468 		log.u_bbr.use_lt_bw <<= 1;
2469 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2470 		log.u_bbr.use_lt_bw <<= 1;
2471 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2472 		log.u_bbr.use_lt_bw <<= 1;
2473 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2474 		log.u_bbr.use_lt_bw <<= 1;
2475 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2476 		log.u_bbr.use_lt_bw <<= 1;
2477 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2478 		log.u_bbr.use_lt_bw <<= 1;
2479 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2480 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2481 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2482 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2483 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2484 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2485 		log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2486 		log.u_bbr.bw_inuse <<= 32;
2487 		if (rsm)
2488 			log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2489 		TCP_LOG_EVENTP(tp, NULL,
2490 		    &rack->rc_inp->inp_socket->so_rcv,
2491 		    &rack->rc_inp->inp_socket->so_snd,
2492 		    BBR_LOG_BBRRTT, 0,
2493 		    0, &log, false, &tv);
2494 
2495 
2496 	}
2497 }
2498 
2499 static void
2500 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2501 {
2502 	/*
2503 	 * Log the rtt sample we are
2504 	 * applying to the srtt algorithm in
2505 	 * useconds.
2506 	 */
2507 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2508 		union tcp_log_stackspecific log;
2509 		struct timeval tv;
2510 
2511 		/* Convert our ms to a microsecond */
2512 		memset(&log, 0, sizeof(log));
2513 		log.u_bbr.flex1 = rtt;
2514 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2515 		log.u_bbr.flex3 = rack->r_ctl.sack_count;
2516 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2517 		log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2518 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2519 		log.u_bbr.flex7 = 1;
2520 		log.u_bbr.flex8 = rack->sack_attack_disable;
2521 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2522 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2523 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2524 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2525 		log.u_bbr.pacing_gain = rack->r_must_retran;
2526 		/*
2527 		 * We capture in delRate the upper 32 bits as
2528 		 * the confidence level we had declared, and the
2529 		 * lower 32 bits as the actual RTT using the arrival
2530 		 * timestamp.
2531 		 */
2532 		log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2533 		log.u_bbr.delRate <<= 32;
2534 		log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2535 		/* Lets capture all the things that make up t_rtxcur */
2536 		log.u_bbr.applimited = rack_rto_min;
2537 		log.u_bbr.epoch = rack_rto_max;
2538 		log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2539 		log.u_bbr.lost = rack_rto_min;
2540 		log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2541 		log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2542 		log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2543 		log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2544 		log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2545 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2546 		    &rack->rc_inp->inp_socket->so_rcv,
2547 		    &rack->rc_inp->inp_socket->so_snd,
2548 		    TCP_LOG_RTT, 0,
2549 		    0, &log, false, &tv);
2550 	}
2551 }
2552 
2553 static void
2554 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2555 {
2556 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2557 		union tcp_log_stackspecific log;
2558 		struct timeval tv;
2559 
2560 		/* Convert our ms to a microsecond */
2561 		memset(&log, 0, sizeof(log));
2562 		log.u_bbr.flex1 = rtt;
2563 		log.u_bbr.flex2 = send_time;
2564 		log.u_bbr.flex3 = ack_time;
2565 		log.u_bbr.flex4 = where;
2566 		log.u_bbr.flex7 = 2;
2567 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2568 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2569 		    &rack->rc_inp->inp_socket->so_rcv,
2570 		    &rack->rc_inp->inp_socket->so_snd,
2571 		    TCP_LOG_RTT, 0,
2572 		    0, &log, false, &tv);
2573 	}
2574 }
2575 
2576 
2577 
2578 static inline void
2579 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2580 {
2581 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2582 		union tcp_log_stackspecific log;
2583 		struct timeval tv;
2584 
2585 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2586 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2587 		log.u_bbr.flex1 = line;
2588 		log.u_bbr.flex2 = tick;
2589 		log.u_bbr.flex3 = tp->t_maxunacktime;
2590 		log.u_bbr.flex4 = tp->t_acktime;
2591 		log.u_bbr.flex8 = event;
2592 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2593 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2594 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2595 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2596 		log.u_bbr.pacing_gain = rack->r_must_retran;
2597 		TCP_LOG_EVENTP(tp, NULL,
2598 		    &rack->rc_inp->inp_socket->so_rcv,
2599 		    &rack->rc_inp->inp_socket->so_snd,
2600 		    BBR_LOG_PROGRESS, 0,
2601 		    0, &log, false, &tv);
2602 	}
2603 }
2604 
2605 static void
2606 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2607 {
2608 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2609 		union tcp_log_stackspecific log;
2610 
2611 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2612 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2613 		log.u_bbr.flex1 = slot;
2614 		if (rack->rack_no_prr)
2615 			log.u_bbr.flex2 = 0;
2616 		else
2617 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2618 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2619 		log.u_bbr.flex8 = rack->rc_in_persist;
2620 		log.u_bbr.timeStamp = cts;
2621 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2622 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2623 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2624 		log.u_bbr.pacing_gain = rack->r_must_retran;
2625 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2626 		    &rack->rc_inp->inp_socket->so_rcv,
2627 		    &rack->rc_inp->inp_socket->so_snd,
2628 		    BBR_LOG_BBRSND, 0,
2629 		    0, &log, false, tv);
2630 	}
2631 }
2632 
2633 static void
2634 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2635 {
2636 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2637 		union tcp_log_stackspecific log;
2638 		struct timeval tv;
2639 
2640 		memset(&log, 0, sizeof(log));
2641 		log.u_bbr.flex1 = did_out;
2642 		log.u_bbr.flex2 = nxt_pkt;
2643 		log.u_bbr.flex3 = way_out;
2644 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2645 		if (rack->rack_no_prr)
2646 			log.u_bbr.flex5 = 0;
2647 		else
2648 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2649 		log.u_bbr.flex6 = nsegs;
2650 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2651 		log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;	/* Do we have ack-can-send set */
2652 		log.u_bbr.flex7 <<= 1;
2653 		log.u_bbr.flex7 |= rack->r_fast_output;	/* is fast output primed */
2654 		log.u_bbr.flex7 <<= 1;
2655 		log.u_bbr.flex7 |= rack->r_wanted_output;	/* Do we want output */
2656 		log.u_bbr.flex8 = rack->rc_in_persist;
2657 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2658 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2659 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2660 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2661 		log.u_bbr.use_lt_bw <<= 1;
2662 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
2663 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2664 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2665 		log.u_bbr.pacing_gain = rack->r_must_retran;
2666 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2667 		    &rack->rc_inp->inp_socket->so_rcv,
2668 		    &rack->rc_inp->inp_socket->so_snd,
2669 		    BBR_LOG_DOSEG_DONE, 0,
2670 		    0, &log, false, &tv);
2671 	}
2672 }
2673 
2674 static void
2675 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2676 {
2677 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2678 		union tcp_log_stackspecific log;
2679 		struct timeval tv;
2680 		uint32_t cts;
2681 
2682 		memset(&log, 0, sizeof(log));
2683 		cts = tcp_get_usecs(&tv);
2684 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2685 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2686 		log.u_bbr.flex4 = arg1;
2687 		log.u_bbr.flex5 = arg2;
2688 		log.u_bbr.flex6 = arg3;
2689 		log.u_bbr.flex8 = frm;
2690 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2691 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2692 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2693 		log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2694 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2695 		log.u_bbr.pacing_gain = rack->r_must_retran;
2696 		TCP_LOG_EVENTP(tp, NULL,
2697 		    &tp->t_inpcb->inp_socket->so_rcv,
2698 		    &tp->t_inpcb->inp_socket->so_snd,
2699 		    TCP_HDWR_PACE_SIZE, 0,
2700 		    0, &log, false, &tv);
2701 	}
2702 }
2703 
2704 static void
2705 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2706 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2707 {
2708 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2709 		union tcp_log_stackspecific log;
2710 		struct timeval tv;
2711 
2712 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2713 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2714 		log.u_bbr.flex1 = slot;
2715 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2716 		log.u_bbr.flex4 = reason;
2717 		if (rack->rack_no_prr)
2718 			log.u_bbr.flex5 = 0;
2719 		else
2720 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2721 		log.u_bbr.flex7 = hpts_calling;
2722 		log.u_bbr.flex8 = rack->rc_in_persist;
2723 		log.u_bbr.lt_epoch = cwnd_to_use;
2724 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2725 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2726 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2727 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2728 		log.u_bbr.pacing_gain = rack->r_must_retran;
2729 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2730 		    &rack->rc_inp->inp_socket->so_rcv,
2731 		    &rack->rc_inp->inp_socket->so_snd,
2732 		    BBR_LOG_JUSTRET, 0,
2733 		    tlen, &log, false, &tv);
2734 	}
2735 }
2736 
2737 static void
2738 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2739 		   struct timeval *tv, uint32_t flags_on_entry)
2740 {
2741 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2742 		union tcp_log_stackspecific log;
2743 
2744 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2745 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2746 		log.u_bbr.flex1 = line;
2747 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2748 		log.u_bbr.flex3 = flags_on_entry;
2749 		log.u_bbr.flex4 = us_cts;
2750 		if (rack->rack_no_prr)
2751 			log.u_bbr.flex5 = 0;
2752 		else
2753 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2754 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2755 		log.u_bbr.flex7 = hpts_removed;
2756 		log.u_bbr.flex8 = 1;
2757 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2758 		log.u_bbr.timeStamp = us_cts;
2759 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2760 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2761 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2762 		log.u_bbr.pacing_gain = rack->r_must_retran;
2763 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2764 		    &rack->rc_inp->inp_socket->so_rcv,
2765 		    &rack->rc_inp->inp_socket->so_snd,
2766 		    BBR_LOG_TIMERCANC, 0,
2767 		    0, &log, false, tv);
2768 	}
2769 }
2770 
2771 static void
2772 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2773 			  uint32_t flex1, uint32_t flex2,
2774 			  uint32_t flex3, uint32_t flex4,
2775 			  uint32_t flex5, uint32_t flex6,
2776 			  uint16_t flex7, uint8_t mod)
2777 {
2778 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2779 		union tcp_log_stackspecific log;
2780 		struct timeval tv;
2781 
2782 		if (mod == 1) {
2783 			/* No you can't use 1, its for the real to cancel */
2784 			return;
2785 		}
2786 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2787 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2788 		log.u_bbr.flex1 = flex1;
2789 		log.u_bbr.flex2 = flex2;
2790 		log.u_bbr.flex3 = flex3;
2791 		log.u_bbr.flex4 = flex4;
2792 		log.u_bbr.flex5 = flex5;
2793 		log.u_bbr.flex6 = flex6;
2794 		log.u_bbr.flex7 = flex7;
2795 		log.u_bbr.flex8 = mod;
2796 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2797 		    &rack->rc_inp->inp_socket->so_rcv,
2798 		    &rack->rc_inp->inp_socket->so_snd,
2799 		    BBR_LOG_TIMERCANC, 0,
2800 		    0, &log, false, &tv);
2801 	}
2802 }
2803 
2804 static void
2805 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2806 {
2807 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2808 		union tcp_log_stackspecific log;
2809 		struct timeval tv;
2810 
2811 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2812 		log.u_bbr.flex1 = timers;
2813 		log.u_bbr.flex2 = ret;
2814 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2815 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2816 		log.u_bbr.flex5 = cts;
2817 		if (rack->rack_no_prr)
2818 			log.u_bbr.flex6 = 0;
2819 		else
2820 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2821 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2822 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2823 		log.u_bbr.pacing_gain = rack->r_must_retran;
2824 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2825 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2826 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2827 		    &rack->rc_inp->inp_socket->so_rcv,
2828 		    &rack->rc_inp->inp_socket->so_snd,
2829 		    BBR_LOG_TO_PROCESS, 0,
2830 		    0, &log, false, &tv);
2831 	}
2832 }
2833 
2834 static void
2835 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd)
2836 {
2837 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2838 		union tcp_log_stackspecific log;
2839 		struct timeval tv;
2840 
2841 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2842 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2843 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2844 		if (rack->rack_no_prr)
2845 			log.u_bbr.flex3 = 0;
2846 		else
2847 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2848 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2849 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2850 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2851 		log.u_bbr.flex8 = frm;
2852 		log.u_bbr.pkts_out = orig_cwnd;
2853 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2854 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2855 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2856 		log.u_bbr.use_lt_bw <<= 1;
2857 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
2858 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2859 		    &rack->rc_inp->inp_socket->so_rcv,
2860 		    &rack->rc_inp->inp_socket->so_snd,
2861 		    BBR_LOG_BBRUPD, 0,
2862 		    0, &log, false, &tv);
2863 	}
2864 }
2865 
2866 #ifdef NETFLIX_EXP_DETECTION
2867 static void
2868 rack_log_sad(struct tcp_rack *rack, int event)
2869 {
2870 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2871 		union tcp_log_stackspecific log;
2872 		struct timeval tv;
2873 
2874 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2875 		log.u_bbr.flex1 = rack->r_ctl.sack_count;
2876 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2877 		log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2878 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2879 		log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2880 		log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2881 		log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2882 		log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2883 		log.u_bbr.lt_epoch |= rack->do_detection;
2884 		log.u_bbr.applimited = tcp_map_minimum;
2885 		log.u_bbr.flex7 = rack->sack_attack_disable;
2886 		log.u_bbr.flex8 = event;
2887 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2888 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2889 		log.u_bbr.delivered = tcp_sad_decay_val;
2890 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2891 		    &rack->rc_inp->inp_socket->so_rcv,
2892 		    &rack->rc_inp->inp_socket->so_snd,
2893 		    TCP_SAD_DETECTION, 0,
2894 		    0, &log, false, &tv);
2895 	}
2896 }
2897 #endif
2898 
2899 static void
2900 rack_counter_destroy(void)
2901 {
2902 	int i;
2903 
2904 	counter_u64_free(rack_fto_send);
2905 	counter_u64_free(rack_fto_rsm_send);
2906 	counter_u64_free(rack_nfto_resend);
2907 	counter_u64_free(rack_hw_pace_init_fail);
2908 	counter_u64_free(rack_hw_pace_lost);
2909 	counter_u64_free(rack_non_fto_send);
2910 	counter_u64_free(rack_extended_rfo);
2911 	counter_u64_free(rack_ack_total);
2912 	counter_u64_free(rack_express_sack);
2913 	counter_u64_free(rack_sack_total);
2914 	counter_u64_free(rack_move_none);
2915 	counter_u64_free(rack_move_some);
2916 	counter_u64_free(rack_sack_attacks_detected);
2917 	counter_u64_free(rack_sack_attacks_reversed);
2918 	counter_u64_free(rack_sack_used_next_merge);
2919 	counter_u64_free(rack_sack_used_prev_merge);
2920 	counter_u64_free(rack_badfr);
2921 	counter_u64_free(rack_badfr_bytes);
2922 	counter_u64_free(rack_rtm_prr_retran);
2923 	counter_u64_free(rack_rtm_prr_newdata);
2924 	counter_u64_free(rack_timestamp_mismatch);
2925 	counter_u64_free(rack_find_high);
2926 	counter_u64_free(rack_reorder_seen);
2927 	counter_u64_free(rack_tlp_tot);
2928 	counter_u64_free(rack_tlp_newdata);
2929 	counter_u64_free(rack_tlp_retran);
2930 	counter_u64_free(rack_tlp_retran_bytes);
2931 	counter_u64_free(rack_tlp_retran_fail);
2932 	counter_u64_free(rack_to_tot);
2933 	counter_u64_free(rack_to_arm_rack);
2934 	counter_u64_free(rack_to_arm_tlp);
2935 	counter_u64_free(rack_calc_zero);
2936 	counter_u64_free(rack_calc_nonzero);
2937 	counter_u64_free(rack_paced_segments);
2938 	counter_u64_free(rack_unpaced_segments);
2939 	counter_u64_free(rack_saw_enobuf);
2940 	counter_u64_free(rack_saw_enobuf_hw);
2941 	counter_u64_free(rack_saw_enetunreach);
2942 	counter_u64_free(rack_hot_alloc);
2943 	counter_u64_free(rack_to_alloc);
2944 	counter_u64_free(rack_to_alloc_hard);
2945 	counter_u64_free(rack_to_alloc_emerg);
2946 	counter_u64_free(rack_to_alloc_limited);
2947 	counter_u64_free(rack_alloc_limited_conns);
2948 	counter_u64_free(rack_split_limited);
2949 	for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
2950 		counter_u64_free(rack_proc_comp_ack[i]);
2951 	}
2952 	counter_u64_free(rack_multi_single_eq);
2953 	counter_u64_free(rack_proc_non_comp_ack);
2954 	counter_u64_free(rack_sack_proc_all);
2955 	counter_u64_free(rack_sack_proc_restart);
2956 	counter_u64_free(rack_sack_proc_short);
2957 	counter_u64_free(rack_enter_tlp_calc);
2958 	counter_u64_free(rack_used_tlpmethod);
2959 	counter_u64_free(rack_used_tlpmethod2);
2960 	counter_u64_free(rack_sack_skipped_acked);
2961 	counter_u64_free(rack_sack_splits);
2962 	counter_u64_free(rack_progress_drops);
2963 	counter_u64_free(rack_input_idle_reduces);
2964 	counter_u64_free(rack_collapsed_win);
2965 	counter_u64_free(rack_tlp_does_nada);
2966 	counter_u64_free(rack_try_scwnd);
2967 	counter_u64_free(rack_per_timer_hole);
2968 	counter_u64_free(rack_large_ackcmp);
2969 	counter_u64_free(rack_small_ackcmp);
2970 	counter_u64_free(rack_persists_sends);
2971 	counter_u64_free(rack_persists_acks);
2972 	counter_u64_free(rack_persists_loss);
2973 	counter_u64_free(rack_persists_lost_ends);
2974 #ifdef INVARIANTS
2975 	counter_u64_free(rack_adjust_map_bw);
2976 #endif
2977 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2978 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2979 }
2980 
2981 static struct rack_sendmap *
2982 rack_alloc(struct tcp_rack *rack)
2983 {
2984 	struct rack_sendmap *rsm;
2985 
2986 	/*
2987 	 * First get the top of the list it in
2988 	 * theory is the "hottest" rsm we have,
2989 	 * possibly just freed by ack processing.
2990 	 */
2991 	if (rack->rc_free_cnt > rack_free_cache) {
2992 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2993 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2994 		counter_u64_add(rack_hot_alloc, 1);
2995 		rack->rc_free_cnt--;
2996 		return (rsm);
2997 	}
2998 	/*
2999 	 * Once we get under our free cache we probably
3000 	 * no longer have a "hot" one available. Lets
3001 	 * get one from UMA.
3002 	 */
3003 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
3004 	if (rsm) {
3005 		rack->r_ctl.rc_num_maps_alloced++;
3006 		counter_u64_add(rack_to_alloc, 1);
3007 		return (rsm);
3008 	}
3009 	/*
3010 	 * Dig in to our aux rsm's (the last two) since
3011 	 * UMA failed to get us one.
3012 	 */
3013 	if (rack->rc_free_cnt) {
3014 		counter_u64_add(rack_to_alloc_emerg, 1);
3015 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3016 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3017 		rack->rc_free_cnt--;
3018 		return (rsm);
3019 	}
3020 	return (NULL);
3021 }
3022 
3023 static struct rack_sendmap *
3024 rack_alloc_full_limit(struct tcp_rack *rack)
3025 {
3026 	if ((V_tcp_map_entries_limit > 0) &&
3027 	    (rack->do_detection == 0) &&
3028 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3029 		counter_u64_add(rack_to_alloc_limited, 1);
3030 		if (!rack->alloc_limit_reported) {
3031 			rack->alloc_limit_reported = 1;
3032 			counter_u64_add(rack_alloc_limited_conns, 1);
3033 		}
3034 		return (NULL);
3035 	}
3036 	return (rack_alloc(rack));
3037 }
3038 
3039 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3040 static struct rack_sendmap *
3041 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
3042 {
3043 	struct rack_sendmap *rsm;
3044 
3045 	if (limit_type) {
3046 		/* currently there is only one limit type */
3047 		if (V_tcp_map_split_limit > 0 &&
3048 		    (rack->do_detection == 0) &&
3049 		    rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
3050 			counter_u64_add(rack_split_limited, 1);
3051 			if (!rack->alloc_limit_reported) {
3052 				rack->alloc_limit_reported = 1;
3053 				counter_u64_add(rack_alloc_limited_conns, 1);
3054 			}
3055 			return (NULL);
3056 		}
3057 	}
3058 
3059 	/* allocate and mark in the limit type, if set */
3060 	rsm = rack_alloc(rack);
3061 	if (rsm != NULL && limit_type) {
3062 		rsm->r_limit_type = limit_type;
3063 		rack->r_ctl.rc_num_split_allocs++;
3064 	}
3065 	return (rsm);
3066 }
3067 
3068 static void
3069 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
3070 {
3071 	if (rsm->r_flags & RACK_APP_LIMITED) {
3072 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
3073 			rack->r_ctl.rc_app_limited_cnt--;
3074 		}
3075 	}
3076 	if (rsm->r_limit_type) {
3077 		/* currently there is only one limit type */
3078 		rack->r_ctl.rc_num_split_allocs--;
3079 	}
3080 	if (rsm == rack->r_ctl.rc_first_appl) {
3081 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3082 			rack->r_ctl.rc_first_appl = NULL;
3083 		else {
3084 			/* Follow the next one out */
3085 			struct rack_sendmap fe;
3086 
3087 			fe.r_start = rsm->r_nseq_appl;
3088 			rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
3089 		}
3090 	}
3091 	if (rsm == rack->r_ctl.rc_resend)
3092 		rack->r_ctl.rc_resend = NULL;
3093 	if (rsm == rack->r_ctl.rc_rsm_at_retran)
3094 		rack->r_ctl.rc_rsm_at_retran = NULL;
3095 	if (rsm == rack->r_ctl.rc_end_appl)
3096 		rack->r_ctl.rc_end_appl = NULL;
3097 	if (rack->r_ctl.rc_tlpsend == rsm)
3098 		rack->r_ctl.rc_tlpsend = NULL;
3099 	if (rack->r_ctl.rc_sacklast == rsm)
3100 		rack->r_ctl.rc_sacklast = NULL;
3101 	memset(rsm, 0, sizeof(struct rack_sendmap));
3102 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3103 	rack->rc_free_cnt++;
3104 }
3105 
3106 static void
3107 rack_free_trim(struct tcp_rack *rack)
3108 {
3109 	struct rack_sendmap *rsm;
3110 
3111 	/*
3112 	 * Free up all the tail entries until
3113 	 * we get our list down to the limit.
3114 	 */
3115 	while (rack->rc_free_cnt > rack_free_cache) {
3116 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3117 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3118 		rack->rc_free_cnt--;
3119 		uma_zfree(rack_zone, rsm);
3120 	}
3121 }
3122 
3123 
3124 static uint32_t
3125 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3126 {
3127 	uint64_t srtt, bw, len, tim;
3128 	uint32_t segsiz, def_len, minl;
3129 
3130 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3131 	def_len = rack_def_data_window * segsiz;
3132 	if (rack->rc_gp_filled == 0) {
3133 		/*
3134 		 * We have no measurement (IW is in flight?) so
3135 		 * we can only guess using our data_window sysctl
3136 		 * value (usually 20MSS).
3137 		 */
3138 		return (def_len);
3139 	}
3140 	/*
3141 	 * Now we have a number of factors to consider.
3142 	 *
3143 	 * 1) We have a desired BDP which is usually
3144 	 *    at least 2.
3145 	 * 2) We have a minimum number of rtt's usually 1 SRTT
3146 	 *    but we allow it too to be more.
3147 	 * 3) We want to make sure a measurement last N useconds (if
3148 	 *    we have set rack_min_measure_usec.
3149 	 *
3150 	 * We handle the first concern here by trying to create a data
3151 	 * window of max(rack_def_data_window, DesiredBDP). The
3152 	 * second concern we handle in not letting the measurement
3153 	 * window end normally until at least the required SRTT's
3154 	 * have gone by which is done further below in
3155 	 * rack_enough_for_measurement(). Finally the third concern
3156 	 * we also handle here by calculating how long that time
3157 	 * would take at the current BW and then return the
3158 	 * max of our first calculation and that length. Note
3159 	 * that if rack_min_measure_usec is 0, we don't deal
3160 	 * with concern 3. Also for both Concern 1 and 3 an
3161 	 * application limited period could end the measurement
3162 	 * earlier.
3163 	 *
3164 	 * So lets calculate the BDP with the "known" b/w using
3165 	 * the SRTT has our rtt and then multiply it by the
3166 	 * goal.
3167 	 */
3168 	bw = rack_get_bw(rack);
3169 	srtt = (uint64_t)tp->t_srtt;
3170 	len = bw * srtt;
3171 	len /= (uint64_t)HPTS_USEC_IN_SEC;
3172 	len *= max(1, rack_goal_bdp);
3173 	/* Now we need to round up to the nearest MSS */
3174 	len = roundup(len, segsiz);
3175 	if (rack_min_measure_usec) {
3176 		/* Now calculate our min length for this b/w */
3177 		tim = rack_min_measure_usec;
3178 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3179 		if (minl == 0)
3180 			minl = 1;
3181 		minl = roundup(minl, segsiz);
3182 		if (len < minl)
3183 			len = minl;
3184 	}
3185 	/*
3186 	 * Now if we have a very small window we want
3187 	 * to attempt to get the window that is
3188 	 * as small as possible. This happens on
3189 	 * low b/w connections and we don't want to
3190 	 * span huge numbers of rtt's between measurements.
3191 	 *
3192 	 * We basically include 2 over our "MIN window" so
3193 	 * that the measurement can be shortened (possibly) by
3194 	 * an ack'ed packet.
3195 	 */
3196 	if (len < def_len)
3197 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3198 	else
3199 		return (max((uint32_t)len, def_len));
3200 
3201 }
3202 
3203 static int
3204 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3205 {
3206 	uint32_t tim, srtts, segsiz;
3207 
3208 	/*
3209 	 * Has enough time passed for the GP measurement to be valid?
3210 	 */
3211 	if ((tp->snd_max == tp->snd_una) ||
3212 	    (th_ack == tp->snd_max)){
3213 		/* All is acked */
3214 		*quality = RACK_QUALITY_ALLACKED;
3215 		return (1);
3216 	}
3217 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3218 		/* Not enough bytes yet */
3219 		return (0);
3220 	}
3221 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3222 	if (SEQ_LT(th_ack, tp->gput_ack) &&
3223 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3224 		/* Not enough bytes yet */
3225 		return (0);
3226 	}
3227 	if (rack->r_ctl.rc_first_appl &&
3228 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3229 		/*
3230 		 * We are up to the app limited send point
3231 		 * we have to measure irrespective of the time..
3232 		 */
3233 		*quality = RACK_QUALITY_APPLIMITED;
3234 		return (1);
3235 	}
3236 	/* Now what about time? */
3237 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3238 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3239 	if (tim >= srtts) {
3240 		*quality = RACK_QUALITY_HIGH;
3241 		return (1);
3242 	}
3243 	/* Nope not even a full SRTT has passed */
3244 	return (0);
3245 }
3246 
3247 static void
3248 rack_log_timely(struct tcp_rack *rack,
3249 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3250 		uint64_t up_bnd, int line, uint8_t method)
3251 {
3252 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3253 		union tcp_log_stackspecific log;
3254 		struct timeval tv;
3255 
3256 		memset(&log, 0, sizeof(log));
3257 		log.u_bbr.flex1 = logged;
3258 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3259 		log.u_bbr.flex2 <<= 4;
3260 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3261 		log.u_bbr.flex2 <<= 4;
3262 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3263 		log.u_bbr.flex2 <<= 4;
3264 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3265 		log.u_bbr.flex3 = rack->rc_gp_incr;
3266 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3267 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3268 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3269 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3270 		log.u_bbr.flex8 = method;
3271 		log.u_bbr.cur_del_rate = cur_bw;
3272 		log.u_bbr.delRate = low_bnd;
3273 		log.u_bbr.bw_inuse = up_bnd;
3274 		log.u_bbr.rttProp = rack_get_bw(rack);
3275 		log.u_bbr.pkt_epoch = line;
3276 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3277 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3278 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3279 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3280 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3281 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3282 		log.u_bbr.cwnd_gain <<= 1;
3283 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3284 		log.u_bbr.cwnd_gain <<= 1;
3285 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3286 		log.u_bbr.cwnd_gain <<= 1;
3287 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3288 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3289 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3290 		    &rack->rc_inp->inp_socket->so_rcv,
3291 		    &rack->rc_inp->inp_socket->so_snd,
3292 		    TCP_TIMELY_WORK, 0,
3293 		    0, &log, false, &tv);
3294 	}
3295 }
3296 
3297 static int
3298 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3299 {
3300 	/*
3301 	 * Before we increase we need to know if
3302 	 * the estimate just made was less than
3303 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3304 	 *
3305 	 * If we already are pacing at a fast enough
3306 	 * rate to push us faster there is no sense of
3307 	 * increasing.
3308 	 *
3309 	 * We first caculate our actual pacing rate (ss or ca multipler
3310 	 * times our cur_bw).
3311 	 *
3312 	 * Then we take the last measured rate and multipy by our
3313 	 * maximum pacing overage to give us a max allowable rate.
3314 	 *
3315 	 * If our act_rate is smaller than our max_allowable rate
3316 	 * then we should increase. Else we should hold steady.
3317 	 *
3318 	 */
3319 	uint64_t act_rate, max_allow_rate;
3320 
3321 	if (rack_timely_no_stopping)
3322 		return (1);
3323 
3324 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3325 		/*
3326 		 * Initial startup case or
3327 		 * everything is acked case.
3328 		 */
3329 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3330 				__LINE__, 9);
3331 		return (1);
3332 	}
3333 	if (mult <= 100) {
3334 		/*
3335 		 * We can always pace at or slightly above our rate.
3336 		 */
3337 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3338 				__LINE__, 9);
3339 		return (1);
3340 	}
3341 	act_rate = cur_bw * (uint64_t)mult;
3342 	act_rate /= 100;
3343 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3344 	max_allow_rate /= 100;
3345 	if (act_rate < max_allow_rate) {
3346 		/*
3347 		 * Here the rate we are actually pacing at
3348 		 * is smaller than 10% above our last measurement.
3349 		 * This means we are pacing below what we would
3350 		 * like to try to achieve (plus some wiggle room).
3351 		 */
3352 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3353 				__LINE__, 9);
3354 		return (1);
3355 	} else {
3356 		/*
3357 		 * Here we are already pacing at least rack_max_per_above(10%)
3358 		 * what we are getting back. This indicates most likely
3359 		 * that we are being limited (cwnd/rwnd/app) and can't
3360 		 * get any more b/w. There is no sense of trying to
3361 		 * raise up the pacing rate its not speeding us up
3362 		 * and we already are pacing faster than we are getting.
3363 		 */
3364 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3365 				__LINE__, 8);
3366 		return (0);
3367 	}
3368 }
3369 
3370 static void
3371 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3372 {
3373 	/*
3374 	 * When we drag bottom, we want to assure
3375 	 * that no multiplier is below 1.0, if so
3376 	 * we want to restore it to at least that.
3377 	 */
3378 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3379 		/* This is unlikely we usually do not touch recovery */
3380 		rack->r_ctl.rack_per_of_gp_rec = 100;
3381 	}
3382 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3383 		rack->r_ctl.rack_per_of_gp_ca = 100;
3384 	}
3385 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3386 		rack->r_ctl.rack_per_of_gp_ss = 100;
3387 	}
3388 }
3389 
3390 static void
3391 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3392 {
3393 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3394 		rack->r_ctl.rack_per_of_gp_ca = 100;
3395 	}
3396 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3397 		rack->r_ctl.rack_per_of_gp_ss = 100;
3398 	}
3399 }
3400 
3401 static void
3402 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3403 {
3404 	int32_t  calc, logged, plus;
3405 
3406 	logged = 0;
3407 
3408 	if (override) {
3409 		/*
3410 		 * override is passed when we are
3411 		 * loosing b/w and making one last
3412 		 * gasp at trying to not loose out
3413 		 * to a new-reno flow.
3414 		 */
3415 		goto extra_boost;
3416 	}
3417 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3418 	if (rack->rc_gp_incr &&
3419 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3420 		/*
3421 		 * Reset and get 5 strokes more before the boost. Note
3422 		 * that the count is 0 based so we have to add one.
3423 		 */
3424 extra_boost:
3425 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3426 		rack->rc_gp_timely_inc_cnt = 0;
3427 	} else
3428 		plus = (uint32_t)rack_gp_increase_per;
3429 	/* Must be at least 1% increase for true timely increases */
3430 	if ((plus < 1) &&
3431 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3432 		plus = 1;
3433 	if (rack->rc_gp_saw_rec &&
3434 	    (rack->rc_gp_no_rec_chg == 0) &&
3435 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3436 				  rack->r_ctl.rack_per_of_gp_rec)) {
3437 		/* We have been in recovery ding it too */
3438 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3439 		if (calc > 0xffff)
3440 			calc = 0xffff;
3441 		logged |= 1;
3442 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3443 		if (rack_per_upper_bound_ss &&
3444 		    (rack->rc_dragged_bottom == 0) &&
3445 		    (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3446 			rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3447 	}
3448 	if (rack->rc_gp_saw_ca &&
3449 	    (rack->rc_gp_saw_ss == 0) &&
3450 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3451 				  rack->r_ctl.rack_per_of_gp_ca)) {
3452 		/* In CA */
3453 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3454 		if (calc > 0xffff)
3455 			calc = 0xffff;
3456 		logged |= 2;
3457 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3458 		if (rack_per_upper_bound_ca &&
3459 		    (rack->rc_dragged_bottom == 0) &&
3460 		    (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3461 			rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3462 	}
3463 	if (rack->rc_gp_saw_ss &&
3464 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3465 				  rack->r_ctl.rack_per_of_gp_ss)) {
3466 		/* In SS */
3467 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3468 		if (calc > 0xffff)
3469 			calc = 0xffff;
3470 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3471 		if (rack_per_upper_bound_ss &&
3472 		    (rack->rc_dragged_bottom == 0) &&
3473 		    (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3474 			rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3475 		logged |= 4;
3476 	}
3477 	if (logged &&
3478 	    (rack->rc_gp_incr == 0)){
3479 		/* Go into increment mode */
3480 		rack->rc_gp_incr = 1;
3481 		rack->rc_gp_timely_inc_cnt = 0;
3482 	}
3483 	if (rack->rc_gp_incr &&
3484 	    logged &&
3485 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3486 		rack->rc_gp_timely_inc_cnt++;
3487 	}
3488 	rack_log_timely(rack,  logged, plus, 0, 0,
3489 			__LINE__, 1);
3490 }
3491 
3492 static uint32_t
3493 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3494 {
3495 	/*
3496 	 * norm_grad = rtt_diff / minrtt;
3497 	 * new_per = curper * (1 - B * norm_grad)
3498 	 *
3499 	 * B = rack_gp_decrease_per (default 10%)
3500 	 * rtt_dif = input var current rtt-diff
3501 	 * curper = input var current percentage
3502 	 * minrtt = from rack filter
3503 	 *
3504 	 */
3505 	uint64_t perf;
3506 
3507 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3508 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3509 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
3510 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3511 		     (uint64_t)1000000)) /
3512 		(uint64_t)1000000);
3513 	if (perf > curper) {
3514 		/* TSNH */
3515 		perf = curper - 1;
3516 	}
3517 	return ((uint32_t)perf);
3518 }
3519 
3520 static uint32_t
3521 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3522 {
3523 	/*
3524 	 *                                   highrttthresh
3525 	 * result = curper * (1 - (B * ( 1 -  ------          ))
3526 	 *                                     gp_srtt
3527 	 *
3528 	 * B = rack_gp_decrease_per (default 10%)
3529 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3530 	 */
3531 	uint64_t perf;
3532 	uint32_t highrttthresh;
3533 
3534 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3535 
3536 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3537 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3538 					((uint64_t)highrttthresh * (uint64_t)1000000) /
3539 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3540 	return (perf);
3541 }
3542 
3543 static void
3544 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3545 {
3546 	uint64_t logvar, logvar2, logvar3;
3547 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3548 
3549 	if (rack->rc_gp_incr) {
3550 		/* Turn off increment counting */
3551 		rack->rc_gp_incr = 0;
3552 		rack->rc_gp_timely_inc_cnt = 0;
3553 	}
3554 	ss_red = ca_red = rec_red = 0;
3555 	logged = 0;
3556 	/* Calculate the reduction value */
3557 	if (rtt_diff < 0) {
3558 		rtt_diff *= -1;
3559 	}
3560 	/* Must be at least 1% reduction */
3561 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3562 		/* We have been in recovery ding it too */
3563 		if (timely_says == 2) {
3564 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3565 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3566 			if (alt < new_per)
3567 				val = alt;
3568 			else
3569 				val = new_per;
3570 		} else
3571 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3572 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
3573 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3574 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3575 		} else {
3576 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3577 			rec_red = 0;
3578 		}
3579 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3580 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3581 		logged |= 1;
3582 	}
3583 	if (rack->rc_gp_saw_ss) {
3584 		/* Sent in SS */
3585 		if (timely_says == 2) {
3586 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3587 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3588 			if (alt < new_per)
3589 				val = alt;
3590 			else
3591 				val = new_per;
3592 		} else
3593 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3594 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3595 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3596 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3597 		} else {
3598 			ss_red = new_per;
3599 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3600 			logvar = new_per;
3601 			logvar <<= 32;
3602 			logvar |= alt;
3603 			logvar2 = (uint32_t)rtt;
3604 			logvar2 <<= 32;
3605 			logvar2 |= (uint32_t)rtt_diff;
3606 			logvar3 = rack_gp_rtt_maxmul;
3607 			logvar3 <<= 32;
3608 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3609 			rack_log_timely(rack, timely_says,
3610 					logvar2, logvar3,
3611 					logvar, __LINE__, 10);
3612 		}
3613 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3614 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3615 		logged |= 4;
3616 	} else if (rack->rc_gp_saw_ca) {
3617 		/* Sent in CA */
3618 		if (timely_says == 2) {
3619 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3620 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3621 			if (alt < new_per)
3622 				val = alt;
3623 			else
3624 				val = new_per;
3625 		} else
3626 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3627 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
3628 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3629 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3630 		} else {
3631 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3632 			ca_red = 0;
3633 			logvar = new_per;
3634 			logvar <<= 32;
3635 			logvar |= alt;
3636 			logvar2 = (uint32_t)rtt;
3637 			logvar2 <<= 32;
3638 			logvar2 |= (uint32_t)rtt_diff;
3639 			logvar3 = rack_gp_rtt_maxmul;
3640 			logvar3 <<= 32;
3641 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3642 			rack_log_timely(rack, timely_says,
3643 					logvar2, logvar3,
3644 					logvar, __LINE__, 10);
3645 		}
3646 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3647 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3648 		logged |= 2;
3649 	}
3650 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
3651 		rack->rc_gp_timely_dec_cnt++;
3652 		if (rack_timely_dec_clear &&
3653 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3654 			rack->rc_gp_timely_dec_cnt = 0;
3655 	}
3656 	logvar = ss_red;
3657 	logvar <<= 32;
3658 	logvar |= ca_red;
3659 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3660 			__LINE__, 2);
3661 }
3662 
3663 static void
3664 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3665 		     uint32_t rtt, uint32_t line, uint8_t reas)
3666 {
3667 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3668 		union tcp_log_stackspecific log;
3669 		struct timeval tv;
3670 
3671 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3672 		log.u_bbr.flex1 = line;
3673 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3674 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3675 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3676 		log.u_bbr.flex5 = rtt;
3677 		log.u_bbr.flex6 = rack->rc_highly_buffered;
3678 		log.u_bbr.flex6 <<= 1;
3679 		log.u_bbr.flex6 |= rack->forced_ack;
3680 		log.u_bbr.flex6 <<= 1;
3681 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3682 		log.u_bbr.flex6 <<= 1;
3683 		log.u_bbr.flex6 |= rack->in_probe_rtt;
3684 		log.u_bbr.flex6 <<= 1;
3685 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3686 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3687 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3688 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3689 		log.u_bbr.flex8 = reas;
3690 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3691 		log.u_bbr.delRate = rack_get_bw(rack);
3692 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3693 		log.u_bbr.cur_del_rate <<= 32;
3694 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3695 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3696 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3697 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3698 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3699 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3700 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3701 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3702 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3703 		log.u_bbr.rttProp = us_cts;
3704 		log.u_bbr.rttProp <<= 32;
3705 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3706 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3707 		    &rack->rc_inp->inp_socket->so_rcv,
3708 		    &rack->rc_inp->inp_socket->so_snd,
3709 		    BBR_LOG_RTT_SHRINKS, 0,
3710 		    0, &log, false, &rack->r_ctl.act_rcv_time);
3711 	}
3712 }
3713 
3714 static void
3715 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3716 {
3717 	uint64_t bwdp;
3718 
3719 	bwdp = rack_get_bw(rack);
3720 	bwdp *= (uint64_t)rtt;
3721 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3722 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3723 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3724 		/*
3725 		 * A window protocol must be able to have 4 packets
3726 		 * outstanding as the floor in order to function
3727 		 * (especially considering delayed ack :D).
3728 		 */
3729 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3730 	}
3731 }
3732 
3733 static void
3734 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3735 {
3736 	/**
3737 	 * ProbeRTT is a bit different in rack_pacing than in
3738 	 * BBR. It is like BBR in that it uses the lowering of
3739 	 * the RTT as a signal that we saw something new and
3740 	 * counts from there for how long between. But it is
3741 	 * different in that its quite simple. It does not
3742 	 * play with the cwnd and wait until we get down
3743 	 * to N segments outstanding and hold that for
3744 	 * 200ms. Instead it just sets the pacing reduction
3745 	 * rate to a set percentage (70 by default) and hold
3746 	 * that for a number of recent GP Srtt's.
3747 	 */
3748 	uint32_t segsiz;
3749 
3750 	if (rack->rc_gp_dyn_mul == 0)
3751 		return;
3752 
3753 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3754 		/* We are idle */
3755 		return;
3756 	}
3757 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3758 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3759 		/*
3760 		 * Stop the goodput now, the idea here is
3761 		 * that future measurements with in_probe_rtt
3762 		 * won't register if they are not greater so
3763 		 * we want to get what info (if any) is available
3764 		 * now.
3765 		 */
3766 		rack_do_goodput_measurement(rack->rc_tp, rack,
3767 					    rack->rc_tp->snd_una, __LINE__,
3768 					    RACK_QUALITY_PROBERTT);
3769 	}
3770 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3771 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3772 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3773 		     rack->r_ctl.rc_pace_min_segs);
3774 	rack->in_probe_rtt = 1;
3775 	rack->measure_saw_probe_rtt = 1;
3776 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3777 	rack->r_ctl.rc_time_probertt_starts = 0;
3778 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3779 	if (rack_probertt_use_min_rtt_entry)
3780 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3781 	else
3782 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3783 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3784 			     __LINE__, RACK_RTTS_ENTERPROBE);
3785 }
3786 
3787 static void
3788 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3789 {
3790 	struct rack_sendmap *rsm;
3791 	uint32_t segsiz;
3792 
3793 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3794 		     rack->r_ctl.rc_pace_min_segs);
3795 	rack->in_probe_rtt = 0;
3796 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3797 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3798 		/*
3799 		 * Stop the goodput now, the idea here is
3800 		 * that future measurements with in_probe_rtt
3801 		 * won't register if they are not greater so
3802 		 * we want to get what info (if any) is available
3803 		 * now.
3804 		 */
3805 		rack_do_goodput_measurement(rack->rc_tp, rack,
3806 					    rack->rc_tp->snd_una, __LINE__,
3807 					    RACK_QUALITY_PROBERTT);
3808 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3809 		/*
3810 		 * We don't have enough data to make a measurement.
3811 		 * So lets just stop and start here after exiting
3812 		 * probe-rtt. We probably are not interested in
3813 		 * the results anyway.
3814 		 */
3815 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3816 	}
3817 	/*
3818 	 * Measurements through the current snd_max are going
3819 	 * to be limited by the slower pacing rate.
3820 	 *
3821 	 * We need to mark these as app-limited so we
3822 	 * don't collapse the b/w.
3823 	 */
3824 	rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3825 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3826 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3827 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3828 		else {
3829 			/*
3830 			 * Go out to the end app limited and mark
3831 			 * this new one as next and move the end_appl up
3832 			 * to this guy.
3833 			 */
3834 			if (rack->r_ctl.rc_end_appl)
3835 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3836 			rack->r_ctl.rc_end_appl = rsm;
3837 		}
3838 		rsm->r_flags |= RACK_APP_LIMITED;
3839 		rack->r_ctl.rc_app_limited_cnt++;
3840 	}
3841 	/*
3842 	 * Now, we need to examine our pacing rate multipliers.
3843 	 * If its under 100%, we need to kick it back up to
3844 	 * 100%. We also don't let it be over our "max" above
3845 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3846 	 * Note setting clamp_atexit_prtt to 0 has the effect
3847 	 * of setting CA/SS to 100% always at exit (which is
3848 	 * the default behavior).
3849 	 */
3850 	if (rack_probertt_clear_is) {
3851 		rack->rc_gp_incr = 0;
3852 		rack->rc_gp_bwred = 0;
3853 		rack->rc_gp_timely_inc_cnt = 0;
3854 		rack->rc_gp_timely_dec_cnt = 0;
3855 	}
3856 	/* Do we do any clamping at exit? */
3857 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3858 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3859 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3860 	}
3861 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3862 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3863 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3864 	}
3865 	/*
3866 	 * Lets set rtt_diff to 0, so that we will get a "boost"
3867 	 * after exiting.
3868 	 */
3869 	rack->r_ctl.rc_rtt_diff = 0;
3870 
3871 	/* Clear all flags so we start fresh */
3872 	rack->rc_tp->t_bytes_acked = 0;
3873 	rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3874 	/*
3875 	 * If configured to, set the cwnd and ssthresh to
3876 	 * our targets.
3877 	 */
3878 	if (rack_probe_rtt_sets_cwnd) {
3879 		uint64_t ebdp;
3880 		uint32_t setto;
3881 
3882 		/* Set ssthresh so we get into CA once we hit our target */
3883 		if (rack_probertt_use_min_rtt_exit == 1) {
3884 			/* Set to min rtt */
3885 			rack_set_prtt_target(rack, segsiz,
3886 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3887 		} else if (rack_probertt_use_min_rtt_exit == 2) {
3888 			/* Set to current gp rtt */
3889 			rack_set_prtt_target(rack, segsiz,
3890 					     rack->r_ctl.rc_gp_srtt);
3891 		} else if (rack_probertt_use_min_rtt_exit == 3) {
3892 			/* Set to entry gp rtt */
3893 			rack_set_prtt_target(rack, segsiz,
3894 					     rack->r_ctl.rc_entry_gp_rtt);
3895 		} else {
3896 			uint64_t sum;
3897 			uint32_t setval;
3898 
3899 			sum = rack->r_ctl.rc_entry_gp_rtt;
3900 			sum *= 10;
3901 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3902 			if (sum >= 20) {
3903 				/*
3904 				 * A highly buffered path needs
3905 				 * cwnd space for timely to work.
3906 				 * Lets set things up as if
3907 				 * we are heading back here again.
3908 				 */
3909 				setval = rack->r_ctl.rc_entry_gp_rtt;
3910 			} else if (sum >= 15) {
3911 				/*
3912 				 * Lets take the smaller of the
3913 				 * two since we are just somewhat
3914 				 * buffered.
3915 				 */
3916 				setval = rack->r_ctl.rc_gp_srtt;
3917 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
3918 					setval = rack->r_ctl.rc_entry_gp_rtt;
3919 			} else {
3920 				/*
3921 				 * Here we are not highly buffered
3922 				 * and should pick the min we can to
3923 				 * keep from causing loss.
3924 				 */
3925 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3926 			}
3927 			rack_set_prtt_target(rack, segsiz,
3928 					     setval);
3929 		}
3930 		if (rack_probe_rtt_sets_cwnd > 1) {
3931 			/* There is a percentage here to boost */
3932 			ebdp = rack->r_ctl.rc_target_probertt_flight;
3933 			ebdp *= rack_probe_rtt_sets_cwnd;
3934 			ebdp /= 100;
3935 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3936 		} else
3937 			setto = rack->r_ctl.rc_target_probertt_flight;
3938 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3939 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3940 			/* Enforce a min */
3941 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3942 		}
3943 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
3944 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3945 	}
3946 	rack_log_rtt_shrinks(rack,  us_cts,
3947 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3948 			     __LINE__, RACK_RTTS_EXITPROBE);
3949 	/* Clear times last so log has all the info */
3950 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3951 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3952 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3953 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
3954 }
3955 
3956 static void
3957 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3958 {
3959 	/* Check in on probe-rtt */
3960 	if (rack->rc_gp_filled == 0) {
3961 		/* We do not do p-rtt unless we have gp measurements */
3962 		return;
3963 	}
3964 	if (rack->in_probe_rtt) {
3965 		uint64_t no_overflow;
3966 		uint32_t endtime, must_stay;
3967 
3968 		if (rack->r_ctl.rc_went_idle_time &&
3969 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3970 			/*
3971 			 * We went idle during prtt, just exit now.
3972 			 */
3973 			rack_exit_probertt(rack, us_cts);
3974 		} else if (rack_probe_rtt_safety_val &&
3975 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3976 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3977 			/*
3978 			 * Probe RTT safety value triggered!
3979 			 */
3980 			rack_log_rtt_shrinks(rack,  us_cts,
3981 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3982 					     __LINE__, RACK_RTTS_SAFETY);
3983 			rack_exit_probertt(rack, us_cts);
3984 		}
3985 		/* Calculate the max we will wait */
3986 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3987 		if (rack->rc_highly_buffered)
3988 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3989 		/* Calculate the min we must wait */
3990 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3991 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3992 		    TSTMP_LT(us_cts, endtime)) {
3993 			uint32_t calc;
3994 			/* Do we lower more? */
3995 no_exit:
3996 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3997 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3998 			else
3999 				calc = 0;
4000 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
4001 			if (calc) {
4002 				/* Maybe */
4003 				calc *= rack_per_of_gp_probertt_reduce;
4004 				rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
4005 				/* Limit it too */
4006 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
4007 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4008 			}
4009 			/* We must reach target or the time set */
4010 			return;
4011 		}
4012 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
4013 			if ((TSTMP_LT(us_cts, must_stay) &&
4014 			     rack->rc_highly_buffered) ||
4015 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
4016 			      rack->r_ctl.rc_target_probertt_flight)) {
4017 				/* We are not past the must_stay time */
4018 				goto no_exit;
4019 			}
4020 			rack_log_rtt_shrinks(rack,  us_cts,
4021 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4022 					     __LINE__, RACK_RTTS_REACHTARGET);
4023 			rack->r_ctl.rc_time_probertt_starts = us_cts;
4024 			if (rack->r_ctl.rc_time_probertt_starts == 0)
4025 				rack->r_ctl.rc_time_probertt_starts = 1;
4026 			/* Restore back to our rate we want to pace at in prtt */
4027 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4028 		}
4029 		/*
4030 		 * Setup our end time, some number of gp_srtts plus 200ms.
4031 		 */
4032 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
4033 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
4034 		if (rack_probertt_gpsrtt_cnt_div)
4035 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
4036 		else
4037 			endtime = 0;
4038 		endtime += rack_min_probertt_hold;
4039 		endtime += rack->r_ctl.rc_time_probertt_starts;
4040 		if (TSTMP_GEQ(us_cts,  endtime)) {
4041 			/* yes, exit probertt */
4042 			rack_exit_probertt(rack, us_cts);
4043 		}
4044 
4045 	} else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
4046 		/* Go into probertt, its been too long since we went lower */
4047 		rack_enter_probertt(rack, us_cts);
4048 	}
4049 }
4050 
4051 static void
4052 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
4053 		       uint32_t rtt, int32_t rtt_diff)
4054 {
4055 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
4056 	uint32_t losses;
4057 
4058 	if ((rack->rc_gp_dyn_mul == 0) ||
4059 	    (rack->use_fixed_rate) ||
4060 	    (rack->in_probe_rtt) ||
4061 	    (rack->rc_always_pace == 0)) {
4062 		/* No dynamic GP multipler in play */
4063 		return;
4064 	}
4065 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
4066 	cur_bw = rack_get_bw(rack);
4067 	/* Calculate our up and down range */
4068 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
4069 	up_bnd /= 100;
4070 	up_bnd += rack->r_ctl.last_gp_comp_bw;
4071 
4072 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
4073 	subfr /= 100;
4074 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
4075 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
4076 		/*
4077 		 * This is the case where our RTT is above
4078 		 * the max target and we have been configured
4079 		 * to just do timely no bonus up stuff in that case.
4080 		 *
4081 		 * There are two configurations, set to 1, and we
4082 		 * just do timely if we are over our max. If its
4083 		 * set above 1 then we slam the multipliers down
4084 		 * to 100 and then decrement per timely.
4085 		 */
4086 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4087 				__LINE__, 3);
4088 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
4089 			rack_validate_multipliers_at_or_below_100(rack);
4090 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4091 	} else if ((last_bw_est < low_bnd) && !losses) {
4092 		/*
4093 		 * We are decreasing this is a bit complicated this
4094 		 * means we are loosing ground. This could be
4095 		 * because another flow entered and we are competing
4096 		 * for b/w with it. This will push the RTT up which
4097 		 * makes timely unusable unless we want to get shoved
4098 		 * into a corner and just be backed off (the age
4099 		 * old problem with delay based CC).
4100 		 *
4101 		 * On the other hand if it was a route change we
4102 		 * would like to stay somewhat contained and not
4103 		 * blow out the buffers.
4104 		 */
4105 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4106 				__LINE__, 3);
4107 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4108 		if (rack->rc_gp_bwred == 0) {
4109 			/* Go into reduction counting */
4110 			rack->rc_gp_bwred = 1;
4111 			rack->rc_gp_timely_dec_cnt = 0;
4112 		}
4113 		if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
4114 		    (timely_says == 0)) {
4115 			/*
4116 			 * Push another time with a faster pacing
4117 			 * to try to gain back (we include override to
4118 			 * get a full raise factor).
4119 			 */
4120 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4121 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4122 			    (timely_says == 0) ||
4123 			    (rack_down_raise_thresh == 0)) {
4124 				/*
4125 				 * Do an override up in b/w if we were
4126 				 * below the threshold or if the threshold
4127 				 * is zero we always do the raise.
4128 				 */
4129 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4130 			} else {
4131 				/* Log it stays the same */
4132 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4133 						__LINE__, 11);
4134 			}
4135 			rack->rc_gp_timely_dec_cnt++;
4136 			/* We are not incrementing really no-count */
4137 			rack->rc_gp_incr = 0;
4138 			rack->rc_gp_timely_inc_cnt = 0;
4139 		} else {
4140 			/*
4141 			 * Lets just use the RTT
4142 			 * information and give up
4143 			 * pushing.
4144 			 */
4145 			goto use_timely;
4146 		}
4147 	} else if ((timely_says != 2) &&
4148 		    !losses &&
4149 		    (last_bw_est > up_bnd)) {
4150 		/*
4151 		 * We are increasing b/w lets keep going, updating
4152 		 * our b/w and ignoring any timely input, unless
4153 		 * of course we are at our max raise (if there is one).
4154 		 */
4155 
4156 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4157 				__LINE__, 3);
4158 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4159 		if (rack->rc_gp_saw_ss &&
4160 		    rack_per_upper_bound_ss &&
4161 		     (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
4162 			    /*
4163 			     * In cases where we can't go higher
4164 			     * we should just use timely.
4165 			     */
4166 			    goto use_timely;
4167 		}
4168 		if (rack->rc_gp_saw_ca &&
4169 		    rack_per_upper_bound_ca &&
4170 		    (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
4171 			    /*
4172 			     * In cases where we can't go higher
4173 			     * we should just use timely.
4174 			     */
4175 			    goto use_timely;
4176 		}
4177 		rack->rc_gp_bwred = 0;
4178 		rack->rc_gp_timely_dec_cnt = 0;
4179 		/* You get a set number of pushes if timely is trying to reduce */
4180 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4181 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4182 		} else {
4183 			/* Log it stays the same */
4184 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4185 			    __LINE__, 12);
4186 		}
4187 		return;
4188 	} else {
4189 		/*
4190 		 * We are staying between the lower and upper range bounds
4191 		 * so use timely to decide.
4192 		 */
4193 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4194 				__LINE__, 3);
4195 use_timely:
4196 		if (timely_says) {
4197 			rack->rc_gp_incr = 0;
4198 			rack->rc_gp_timely_inc_cnt = 0;
4199 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4200 			    !losses &&
4201 			    (last_bw_est < low_bnd)) {
4202 				/* We are loosing ground */
4203 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4204 				rack->rc_gp_timely_dec_cnt++;
4205 				/* We are not incrementing really no-count */
4206 				rack->rc_gp_incr = 0;
4207 				rack->rc_gp_timely_inc_cnt = 0;
4208 			} else
4209 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4210 		} else {
4211 			rack->rc_gp_bwred = 0;
4212 			rack->rc_gp_timely_dec_cnt = 0;
4213 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4214 		}
4215 	}
4216 }
4217 
4218 static int32_t
4219 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4220 {
4221 	int32_t timely_says;
4222 	uint64_t log_mult, log_rtt_a_diff;
4223 
4224 	log_rtt_a_diff = rtt;
4225 	log_rtt_a_diff <<= 32;
4226 	log_rtt_a_diff |= (uint32_t)rtt_diff;
4227 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4228 		    rack_gp_rtt_maxmul)) {
4229 		/* Reduce the b/w multipler */
4230 		timely_says = 2;
4231 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4232 		log_mult <<= 32;
4233 		log_mult |= prev_rtt;
4234 		rack_log_timely(rack,  timely_says, log_mult,
4235 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4236 				log_rtt_a_diff, __LINE__, 4);
4237 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4238 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4239 			    max(rack_gp_rtt_mindiv , 1)))) {
4240 		/* Increase the b/w multipler */
4241 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4242 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4243 			 max(rack_gp_rtt_mindiv , 1));
4244 		log_mult <<= 32;
4245 		log_mult |= prev_rtt;
4246 		timely_says = 0;
4247 		rack_log_timely(rack,  timely_says, log_mult ,
4248 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4249 				log_rtt_a_diff, __LINE__, 5);
4250 	} else {
4251 		/*
4252 		 * Use a gradient to find it the timely gradient
4253 		 * is:
4254 		 * grad = rc_rtt_diff / min_rtt;
4255 		 *
4256 		 * anything below or equal to 0 will be
4257 		 * a increase indication. Anything above
4258 		 * zero is a decrease. Note we take care
4259 		 * of the actual gradient calculation
4260 		 * in the reduction (its not needed for
4261 		 * increase).
4262 		 */
4263 		log_mult = prev_rtt;
4264 		if (rtt_diff <= 0) {
4265 			/*
4266 			 * Rttdiff is less than zero, increase the
4267 			 * b/w multipler (its 0 or negative)
4268 			 */
4269 			timely_says = 0;
4270 			rack_log_timely(rack,  timely_says, log_mult,
4271 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4272 		} else {
4273 			/* Reduce the b/w multipler */
4274 			timely_says = 1;
4275 			rack_log_timely(rack,  timely_says, log_mult,
4276 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4277 		}
4278 	}
4279 	return (timely_says);
4280 }
4281 
4282 static void
4283 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4284 			    tcp_seq th_ack, int line, uint8_t quality)
4285 {
4286 	uint64_t tim, bytes_ps, ltim, stim, utim;
4287 	uint32_t segsiz, bytes, reqbytes, us_cts;
4288 	int32_t gput, new_rtt_diff, timely_says;
4289 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4290 	int did_add = 0;
4291 
4292 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4293 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4294 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
4295 		tim = us_cts - tp->gput_ts;
4296 	else
4297 		tim = 0;
4298 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4299 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4300 	else
4301 		stim = 0;
4302 	/*
4303 	 * Use the larger of the send time or ack time. This prevents us
4304 	 * from being influenced by ack artifacts to come up with too
4305 	 * high of measurement. Note that since we are spanning over many more
4306 	 * bytes in most of our measurements hopefully that is less likely to
4307 	 * occur.
4308 	 */
4309 	if (tim > stim)
4310 		utim = max(tim, 1);
4311 	else
4312 		utim = max(stim, 1);
4313 	/* Lets get a msec time ltim too for the old stuff */
4314 	ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4315 	gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4316 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4317 	if ((tim == 0) && (stim == 0)) {
4318 		/*
4319 		 * Invalid measurement time, maybe
4320 		 * all on one ack/one send?
4321 		 */
4322 		bytes = 0;
4323 		bytes_ps = 0;
4324 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4325 					   0, 0, 0, 10, __LINE__, NULL, quality);
4326 		goto skip_measurement;
4327 	}
4328 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4329 		/* We never made a us_rtt measurement? */
4330 		bytes = 0;
4331 		bytes_ps = 0;
4332 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4333 					   0, 0, 0, 10, __LINE__, NULL, quality);
4334 		goto skip_measurement;
4335 	}
4336 	/*
4337 	 * Calculate the maximum possible b/w this connection
4338 	 * could have. We base our calculation on the lowest
4339 	 * rtt we have seen during the measurement and the
4340 	 * largest rwnd the client has given us in that time. This
4341 	 * forms a BDP that is the maximum that we could ever
4342 	 * get to the client. Anything larger is not valid.
4343 	 *
4344 	 * I originally had code here that rejected measurements
4345 	 * where the time was less than 1/2 the latest us_rtt.
4346 	 * But after thinking on that I realized its wrong since
4347 	 * say you had a 150Mbps or even 1Gbps link, and you
4348 	 * were a long way away.. example I am in Europe (100ms rtt)
4349 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4350 	 * bytes my time would be 1.2ms, and yet my rtt would say
4351 	 * the measurement was invalid the time was < 50ms. The
4352 	 * same thing is true for 150Mb (8ms of time).
4353 	 *
4354 	 * A better way I realized is to look at what the maximum
4355 	 * the connection could possibly do. This is gated on
4356 	 * the lowest RTT we have seen and the highest rwnd.
4357 	 * We should in theory never exceed that, if we are
4358 	 * then something on the path is storing up packets
4359 	 * and then feeding them all at once to our endpoint
4360 	 * messing up our measurement.
4361 	 */
4362 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4363 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4364 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4365 	if (SEQ_LT(th_ack, tp->gput_seq)) {
4366 		/* No measurement can be made */
4367 		bytes = 0;
4368 		bytes_ps = 0;
4369 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4370 					   0, 0, 0, 10, __LINE__, NULL, quality);
4371 		goto skip_measurement;
4372 	} else
4373 		bytes = (th_ack - tp->gput_seq);
4374 	bytes_ps = (uint64_t)bytes;
4375 	/*
4376 	 * Don't measure a b/w for pacing unless we have gotten at least
4377 	 * an initial windows worth of data in this measurement interval.
4378 	 *
4379 	 * Small numbers of bytes get badly influenced by delayed ack and
4380 	 * other artifacts. Note we take the initial window or our
4381 	 * defined minimum GP (defaulting to 10 which hopefully is the
4382 	 * IW).
4383 	 */
4384 	if (rack->rc_gp_filled == 0) {
4385 		/*
4386 		 * The initial estimate is special. We
4387 		 * have blasted out an IW worth of packets
4388 		 * without a real valid ack ts results. We
4389 		 * then setup the app_limited_needs_set flag,
4390 		 * this should get the first ack in (probably 2
4391 		 * MSS worth) to be recorded as the timestamp.
4392 		 * We thus allow a smaller number of bytes i.e.
4393 		 * IW - 2MSS.
4394 		 */
4395 		reqbytes -= (2 * segsiz);
4396 		/* Also lets fill previous for our first measurement to be neutral */
4397 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4398 	}
4399 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4400 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4401 					   rack->r_ctl.rc_app_limited_cnt,
4402 					   0, 0, 10, __LINE__, NULL, quality);
4403 		goto skip_measurement;
4404 	}
4405 	/*
4406 	 * We now need to calculate the Timely like status so
4407 	 * we can update (possibly) the b/w multipliers.
4408 	 */
4409 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4410 	if (rack->rc_gp_filled == 0) {
4411 		/* No previous reading */
4412 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4413 	} else {
4414 		if (rack->measure_saw_probe_rtt == 0) {
4415 			/*
4416 			 * We don't want a probertt to be counted
4417 			 * since it will be negative incorrectly. We
4418 			 * expect to be reducing the RTT when we
4419 			 * pace at a slower rate.
4420 			 */
4421 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4422 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4423 		}
4424 	}
4425 	timely_says = rack_make_timely_judgement(rack,
4426 		rack->r_ctl.rc_gp_srtt,
4427 		rack->r_ctl.rc_rtt_diff,
4428 	        rack->r_ctl.rc_prev_gp_srtt
4429 		);
4430 	bytes_ps *= HPTS_USEC_IN_SEC;
4431 	bytes_ps /= utim;
4432 	if (bytes_ps > rack->r_ctl.last_max_bw) {
4433 		/*
4434 		 * Something is on path playing
4435 		 * since this b/w is not possible based
4436 		 * on our BDP (highest rwnd and lowest rtt
4437 		 * we saw in the measurement window).
4438 		 *
4439 		 * Another option here would be to
4440 		 * instead skip the measurement.
4441 		 */
4442 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4443 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
4444 					   11, __LINE__, NULL, quality);
4445 		bytes_ps = rack->r_ctl.last_max_bw;
4446 	}
4447 	/* We store gp for b/w in bytes per second */
4448 	if (rack->rc_gp_filled == 0) {
4449 		/* Initial measurment */
4450 		if (bytes_ps) {
4451 			rack->r_ctl.gp_bw = bytes_ps;
4452 			rack->rc_gp_filled = 1;
4453 			rack->r_ctl.num_measurements = 1;
4454 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4455 		} else {
4456 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4457 						   rack->r_ctl.rc_app_limited_cnt,
4458 						   0, 0, 10, __LINE__, NULL, quality);
4459 		}
4460 		if (tcp_in_hpts(rack->rc_inp) &&
4461 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4462 			/*
4463 			 * Ok we can't trust the pacer in this case
4464 			 * where we transition from un-paced to paced.
4465 			 * Or for that matter when the burst mitigation
4466 			 * was making a wild guess and got it wrong.
4467 			 * Stop the pacer and clear up all the aggregate
4468 			 * delays etc.
4469 			 */
4470 			tcp_hpts_remove(rack->rc_inp);
4471 			rack->r_ctl.rc_hpts_flags = 0;
4472 			rack->r_ctl.rc_last_output_to = 0;
4473 		}
4474 		did_add = 2;
4475 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4476 		/* Still a small number run an average */
4477 		rack->r_ctl.gp_bw += bytes_ps;
4478 		addpart = rack->r_ctl.num_measurements;
4479 		rack->r_ctl.num_measurements++;
4480 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4481 			/* We have collected enought to move forward */
4482 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4483 		}
4484 		did_add = 3;
4485 	} else {
4486 		/*
4487 		 * We want to take 1/wma of the goodput and add in to 7/8th
4488 		 * of the old value weighted by the srtt. So if your measurement
4489 		 * period is say 2 SRTT's long you would get 1/4 as the
4490 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
4491 		 *
4492 		 * But we must be careful not to take too much i.e. if the
4493 		 * srtt is say 20ms and the measurement is taken over
4494 		 * 400ms our weight would be 400/20 i.e. 20. On the
4495 		 * other hand if we get a measurement over 1ms with a
4496 		 * 10ms rtt we only want to take a much smaller portion.
4497 		 */
4498 		if (rack->r_ctl.num_measurements < 0xff) {
4499 			rack->r_ctl.num_measurements++;
4500 		}
4501 		srtt = (uint64_t)tp->t_srtt;
4502 		if (srtt == 0) {
4503 			/*
4504 			 * Strange why did t_srtt go back to zero?
4505 			 */
4506 			if (rack->r_ctl.rc_rack_min_rtt)
4507 				srtt = rack->r_ctl.rc_rack_min_rtt;
4508 			else
4509 				srtt = HPTS_USEC_IN_MSEC;
4510 		}
4511 		/*
4512 		 * XXXrrs: Note for reviewers, in playing with
4513 		 * dynamic pacing I discovered this GP calculation
4514 		 * as done originally leads to some undesired results.
4515 		 * Basically you can get longer measurements contributing
4516 		 * too much to the WMA. Thus I changed it if you are doing
4517 		 * dynamic adjustments to only do the aportioned adjustment
4518 		 * if we have a very small (time wise) measurement. Longer
4519 		 * measurements just get there weight (defaulting to 1/8)
4520 		 * add to the WMA. We may want to think about changing
4521 		 * this to always do that for both sides i.e. dynamic
4522 		 * and non-dynamic... but considering lots of folks
4523 		 * were playing with this I did not want to change the
4524 		 * calculation per.se. without your thoughts.. Lawerence?
4525 		 * Peter??
4526 		 */
4527 		if (rack->rc_gp_dyn_mul == 0) {
4528 			subpart = rack->r_ctl.gp_bw * utim;
4529 			subpart /= (srtt * 8);
4530 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
4531 				/*
4532 				 * The b/w update takes no more
4533 				 * away then 1/2 our running total
4534 				 * so factor it in.
4535 				 */
4536 				addpart = bytes_ps * utim;
4537 				addpart /= (srtt * 8);
4538 			} else {
4539 				/*
4540 				 * Don't allow a single measurement
4541 				 * to account for more than 1/2 of the
4542 				 * WMA. This could happen on a retransmission
4543 				 * where utim becomes huge compared to
4544 				 * srtt (multiple retransmissions when using
4545 				 * the sending rate which factors in all the
4546 				 * transmissions from the first one).
4547 				 */
4548 				subpart = rack->r_ctl.gp_bw / 2;
4549 				addpart = bytes_ps / 2;
4550 			}
4551 			resid_bw = rack->r_ctl.gp_bw - subpart;
4552 			rack->r_ctl.gp_bw = resid_bw + addpart;
4553 			did_add = 1;
4554 		} else {
4555 			if ((utim / srtt) <= 1) {
4556 				/*
4557 				 * The b/w update was over a small period
4558 				 * of time. The idea here is to prevent a small
4559 				 * measurement time period from counting
4560 				 * too much. So we scale it based on the
4561 				 * time so it attributes less than 1/rack_wma_divisor
4562 				 * of its measurement.
4563 				 */
4564 				subpart = rack->r_ctl.gp_bw * utim;
4565 				subpart /= (srtt * rack_wma_divisor);
4566 				addpart = bytes_ps * utim;
4567 				addpart /= (srtt * rack_wma_divisor);
4568 			} else {
4569 				/*
4570 				 * The scaled measurement was long
4571 				 * enough so lets just add in the
4572 				 * portion of the measurment i.e. 1/rack_wma_divisor
4573 				 */
4574 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4575 				addpart = bytes_ps / rack_wma_divisor;
4576 			}
4577 			if ((rack->measure_saw_probe_rtt == 0) ||
4578 		            (bytes_ps > rack->r_ctl.gp_bw)) {
4579 				/*
4580 				 * For probe-rtt we only add it in
4581 				 * if its larger, all others we just
4582 				 * add in.
4583 				 */
4584 				did_add = 1;
4585 				resid_bw = rack->r_ctl.gp_bw - subpart;
4586 				rack->r_ctl.gp_bw = resid_bw + addpart;
4587 			}
4588 		}
4589 	}
4590 	if ((rack->gp_ready == 0) &&
4591 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4592 		/* We have enough measurements now */
4593 		rack->gp_ready = 1;
4594 		rack_set_cc_pacing(rack);
4595 		if (rack->defer_options)
4596 			rack_apply_deferred_options(rack);
4597 	}
4598 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4599 				   rack_get_bw(rack), 22, did_add, NULL, quality);
4600 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
4601 	if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4602 		rack_update_multiplier(rack, timely_says, bytes_ps,
4603 				       rack->r_ctl.rc_gp_srtt,
4604 				       rack->r_ctl.rc_rtt_diff);
4605 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4606 				   rack_get_bw(rack), 3, line, NULL, quality);
4607 	/* reset the gp srtt and setup the new prev */
4608 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4609 	/* Record the lost count for the next measurement */
4610 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4611 	/*
4612 	 * We restart our diffs based on the gpsrtt in the
4613 	 * measurement window.
4614 	 */
4615 	rack->rc_gp_rtt_set = 0;
4616 	rack->rc_gp_saw_rec = 0;
4617 	rack->rc_gp_saw_ca = 0;
4618 	rack->rc_gp_saw_ss = 0;
4619 	rack->rc_dragged_bottom = 0;
4620 skip_measurement:
4621 
4622 #ifdef STATS
4623 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4624 				 gput);
4625 	/*
4626 	 * XXXLAS: This is a temporary hack, and should be
4627 	 * chained off VOI_TCP_GPUT when stats(9) grows an
4628 	 * API to deal with chained VOIs.
4629 	 */
4630 	if (tp->t_stats_gput_prev > 0)
4631 		stats_voi_update_abs_s32(tp->t_stats,
4632 					 VOI_TCP_GPUT_ND,
4633 					 ((gput - tp->t_stats_gput_prev) * 100) /
4634 					 tp->t_stats_gput_prev);
4635 #endif
4636 	tp->t_flags &= ~TF_GPUTINPROG;
4637 	tp->t_stats_gput_prev = gput;
4638 	/*
4639 	 * Now are we app limited now and there is space from where we
4640 	 * were to where we want to go?
4641 	 *
4642 	 * We don't do the other case i.e. non-applimited here since
4643 	 * the next send will trigger us picking up the missing data.
4644 	 */
4645 	if (rack->r_ctl.rc_first_appl &&
4646 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
4647 	    rack->r_ctl.rc_app_limited_cnt &&
4648 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4649 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
4650 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4651 		/*
4652 		 * Yep there is enough outstanding to make a measurement here.
4653 		 */
4654 		struct rack_sendmap *rsm, fe;
4655 
4656 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4657 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4658 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4659 		rack->app_limited_needs_set = 0;
4660 		tp->gput_seq = th_ack;
4661 		if (rack->in_probe_rtt)
4662 			rack->measure_saw_probe_rtt = 1;
4663 		else if ((rack->measure_saw_probe_rtt) &&
4664 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4665 			rack->measure_saw_probe_rtt = 0;
4666 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
4667 			/* There is a full window to gain info from */
4668 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4669 		} else {
4670 			/* We can only measure up to the applimited point */
4671 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
4672 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
4673 				/*
4674 				 * We don't have enough to make a measurement.
4675 				 */
4676 				tp->t_flags &= ~TF_GPUTINPROG;
4677 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
4678 							   0, 0, 0, 6, __LINE__, NULL, quality);
4679 				return;
4680 			}
4681 		}
4682 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
4683 			/*
4684 			 * We will get no more data into the SB
4685 			 * this means we need to have the data available
4686 			 * before we start a measurement.
4687 			 */
4688 			if (sbavail(&tp->t_inpcb->inp_socket->so_snd) < (tp->gput_ack - tp->gput_seq)) {
4689 				/* Nope not enough data. */
4690 				return;
4691 			}
4692 		}
4693 		tp->t_flags |= TF_GPUTINPROG;
4694 		/*
4695 		 * Now we need to find the timestamp of the send at tp->gput_seq
4696 		 * for the send based measurement.
4697 		 */
4698 		fe.r_start = tp->gput_seq;
4699 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4700 		if (rsm) {
4701 			/* Ok send-based limit is set */
4702 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4703 				/*
4704 				 * Move back to include the earlier part
4705 				 * so our ack time lines up right (this may
4706 				 * make an overlapping measurement but thats
4707 				 * ok).
4708 				 */
4709 				tp->gput_seq = rsm->r_start;
4710 			}
4711 			if (rsm->r_flags & RACK_ACKED)
4712 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4713 			else
4714 				rack->app_limited_needs_set = 1;
4715 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4716 		} else {
4717 			/*
4718 			 * If we don't find the rsm due to some
4719 			 * send-limit set the current time, which
4720 			 * basically disables the send-limit.
4721 			 */
4722 			struct timeval tv;
4723 
4724 			microuptime(&tv);
4725 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4726 		}
4727 		rack_log_pacing_delay_calc(rack,
4728 					   tp->gput_seq,
4729 					   tp->gput_ack,
4730 					   (uint64_t)rsm,
4731 					   tp->gput_ts,
4732 					   rack->r_ctl.rc_app_limited_cnt,
4733 					   9,
4734 					   __LINE__, NULL, quality);
4735 	}
4736 }
4737 
4738 /*
4739  * CC wrapper hook functions
4740  */
4741 static void
4742 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4743     uint16_t type, int32_t recovery)
4744 {
4745 	uint32_t prior_cwnd, acked;
4746 	struct tcp_log_buffer *lgb = NULL;
4747 	uint8_t labc_to_use, quality;
4748 
4749 	INP_WLOCK_ASSERT(tp->t_inpcb);
4750 	tp->ccv->nsegs = nsegs;
4751 	acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una);
4752 	if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4753 		uint32_t max;
4754 
4755 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4756 		if (tp->ccv->bytes_this_ack > max) {
4757 			tp->ccv->bytes_this_ack = max;
4758 		}
4759 	}
4760 #ifdef STATS
4761 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4762 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4763 #endif
4764 	quality = RACK_QUALITY_NONE;
4765 	if ((tp->t_flags & TF_GPUTINPROG) &&
4766 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
4767 		/* Measure the Goodput */
4768 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
4769 #ifdef NETFLIX_PEAKRATE
4770 		if ((type == CC_ACK) &&
4771 		    (tp->t_maxpeakrate)) {
4772 			/*
4773 			 * We update t_peakrate_thr. This gives us roughly
4774 			 * one update per round trip time. Note
4775 			 * it will only be used if pace_always is off i.e
4776 			 * we don't do this for paced flows.
4777 			 */
4778 			rack_update_peakrate_thr(tp);
4779 		}
4780 #endif
4781 	}
4782 	/* Which way our we limited, if not cwnd limited no advance in CA */
4783 	if (tp->snd_cwnd <= tp->snd_wnd)
4784 		tp->ccv->flags |= CCF_CWND_LIMITED;
4785 	else
4786 		tp->ccv->flags &= ~CCF_CWND_LIMITED;
4787 	if (tp->snd_cwnd > tp->snd_ssthresh) {
4788 		tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
4789 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4790 		/* For the setting of a window past use the actual scwnd we are using */
4791 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4792 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4793 			tp->ccv->flags |= CCF_ABC_SENTAWND;
4794 		}
4795 	} else {
4796 		tp->ccv->flags &= ~CCF_ABC_SENTAWND;
4797 		tp->t_bytes_acked = 0;
4798 	}
4799 	prior_cwnd = tp->snd_cwnd;
4800 	if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4801 	    (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4802 		labc_to_use = rack->rc_labc;
4803 	else
4804 		labc_to_use = rack_max_abc_post_recovery;
4805 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4806 		union tcp_log_stackspecific log;
4807 		struct timeval tv;
4808 
4809 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4810 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4811 		log.u_bbr.flex1 = th_ack;
4812 		log.u_bbr.flex2 = tp->ccv->flags;
4813 		log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4814 		log.u_bbr.flex4 = tp->ccv->nsegs;
4815 		log.u_bbr.flex5 = labc_to_use;
4816 		log.u_bbr.flex6 = prior_cwnd;
4817 		log.u_bbr.flex7 = V_tcp_do_newsack;
4818 		log.u_bbr.flex8 = 1;
4819 		lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4820 				     0, &log, false, NULL, NULL, 0, &tv);
4821 	}
4822 	if (CC_ALGO(tp)->ack_received != NULL) {
4823 		/* XXXLAS: Find a way to live without this */
4824 		tp->ccv->curack = th_ack;
4825 		tp->ccv->labc = labc_to_use;
4826 		tp->ccv->flags |= CCF_USE_LOCAL_ABC;
4827 		CC_ALGO(tp)->ack_received(tp->ccv, type);
4828 	}
4829 	if (lgb) {
4830 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4831 	}
4832 	if (rack->r_must_retran) {
4833 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4834 			/*
4835 			 * We now are beyond the rxt point so lets disable
4836 			 * the flag.
4837 			 */
4838 			rack->r_ctl.rc_out_at_rto = 0;
4839 			rack->r_must_retran = 0;
4840 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4841 			/*
4842 			 * Only decrement the rc_out_at_rto if the cwnd advances
4843 			 * at least a whole segment. Otherwise next time the peer
4844 			 * acks, we won't be able to send this generaly happens
4845 			 * when we are in Congestion Avoidance.
4846 			 */
4847 			if (acked <= rack->r_ctl.rc_out_at_rto){
4848 				rack->r_ctl.rc_out_at_rto -= acked;
4849 			} else {
4850 				rack->r_ctl.rc_out_at_rto = 0;
4851 			}
4852 		}
4853 	}
4854 #ifdef STATS
4855 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4856 #endif
4857 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4858 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4859 	}
4860 #ifdef NETFLIX_PEAKRATE
4861 	/* we enforce max peak rate if it is set and we are not pacing */
4862 	if ((rack->rc_always_pace == 0) &&
4863 	    tp->t_peakrate_thr &&
4864 	    (tp->snd_cwnd > tp->t_peakrate_thr)) {
4865 		tp->snd_cwnd = tp->t_peakrate_thr;
4866 	}
4867 #endif
4868 }
4869 
4870 static void
4871 tcp_rack_partialack(struct tcpcb *tp)
4872 {
4873 	struct tcp_rack *rack;
4874 
4875 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4876 	INP_WLOCK_ASSERT(tp->t_inpcb);
4877 	/*
4878 	 * If we are doing PRR and have enough
4879 	 * room to send <or> we are pacing and prr
4880 	 * is disabled we will want to see if we
4881 	 * can send data (by setting r_wanted_output to
4882 	 * true).
4883 	 */
4884 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4885 	    rack->rack_no_prr)
4886 		rack->r_wanted_output = 1;
4887 }
4888 
4889 static void
4890 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4891 {
4892 	struct tcp_rack *rack;
4893 	uint32_t orig_cwnd;
4894 
4895 	orig_cwnd = tp->snd_cwnd;
4896 	INP_WLOCK_ASSERT(tp->t_inpcb);
4897 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4898 	/* only alert CC if we alerted when we entered */
4899 	if (CC_ALGO(tp)->post_recovery != NULL) {
4900 		tp->ccv->curack = th_ack;
4901 		CC_ALGO(tp)->post_recovery(tp->ccv);
4902 		if (tp->snd_cwnd < tp->snd_ssthresh) {
4903 			/*
4904 			 * Rack has burst control and pacing
4905 			 * so lets not set this any lower than
4906 			 * snd_ssthresh per RFC-6582 (option 2).
4907 			 */
4908 			tp->snd_cwnd = tp->snd_ssthresh;
4909 		}
4910 	}
4911 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4912 		union tcp_log_stackspecific log;
4913 		struct timeval tv;
4914 
4915 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4916 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4917 		log.u_bbr.flex1 = th_ack;
4918 		log.u_bbr.flex2 = tp->ccv->flags;
4919 		log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4920 		log.u_bbr.flex4 = tp->ccv->nsegs;
4921 		log.u_bbr.flex5 = V_tcp_abc_l_var;
4922 		log.u_bbr.flex6 = orig_cwnd;
4923 		log.u_bbr.flex7 = V_tcp_do_newsack;
4924 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4925 		log.u_bbr.flex8 = 2;
4926 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4927 			       0, &log, false, NULL, NULL, 0, &tv);
4928 	}
4929 	if ((rack->rack_no_prr == 0) &&
4930 	    (rack->no_prr_addback == 0) &&
4931 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
4932 		/*
4933 		 * Suck the next prr cnt back into cwnd, but
4934 		 * only do that if we are not application limited.
4935 		 */
4936 		if (ctf_outstanding(tp) <= sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
4937 			/*
4938 			 * We are allowed to add back to the cwnd the amount we did
4939 			 * not get out if:
4940 			 * a) no_prr_addback is off.
4941 			 * b) we are not app limited
4942 			 * c) we are doing prr
4943 			 * <and>
4944 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4945 			 */
4946 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4947 					    rack->r_ctl.rc_prr_sndcnt);
4948 		}
4949 		rack->r_ctl.rc_prr_sndcnt = 0;
4950 		rack_log_to_prr(rack, 1, 0);
4951 	}
4952 	rack_log_to_prr(rack, 14, orig_cwnd);
4953 	tp->snd_recover = tp->snd_una;
4954 	if (rack->r_ctl.dsack_persist) {
4955 		rack->r_ctl.dsack_persist--;
4956 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
4957 			rack->r_ctl.num_dsack = 0;
4958 		}
4959 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
4960 	}
4961 	EXIT_RECOVERY(tp->t_flags);
4962 }
4963 
4964 static void
4965 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack)
4966 {
4967 	struct tcp_rack *rack;
4968 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4969 
4970 	INP_WLOCK_ASSERT(tp->t_inpcb);
4971 #ifdef STATS
4972 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4973 #endif
4974 	if (IN_RECOVERY(tp->t_flags) == 0) {
4975 		in_rec_at_entry = 0;
4976 		ssthresh_enter = tp->snd_ssthresh;
4977 		cwnd_enter = tp->snd_cwnd;
4978 	} else
4979 		in_rec_at_entry = 1;
4980 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4981 	switch (type) {
4982 	case CC_NDUPACK:
4983 		tp->t_flags &= ~TF_WASFRECOVERY;
4984 		tp->t_flags &= ~TF_WASCRECOVERY;
4985 		if (!IN_FASTRECOVERY(tp->t_flags)) {
4986 			rack->r_ctl.rc_prr_delivered = 0;
4987 			rack->r_ctl.rc_prr_out = 0;
4988 			if (rack->rack_no_prr == 0) {
4989 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4990 				rack_log_to_prr(rack, 2, in_rec_at_entry);
4991 			}
4992 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4993 			tp->snd_recover = tp->snd_max;
4994 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4995 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4996 		}
4997 		break;
4998 	case CC_ECN:
4999 		if (!IN_CONGRECOVERY(tp->t_flags) ||
5000 		    /*
5001 		     * Allow ECN reaction on ACK to CWR, if
5002 		     * that data segment was also CE marked.
5003 		     */
5004 		    SEQ_GEQ(ack, tp->snd_recover)) {
5005 			EXIT_CONGRECOVERY(tp->t_flags);
5006 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
5007 			tp->snd_recover = tp->snd_max + 1;
5008 			if (tp->t_flags2 & TF2_ECN_PERMIT)
5009 				tp->t_flags2 |= TF2_ECN_SND_CWR;
5010 		}
5011 		break;
5012 	case CC_RTO:
5013 		tp->t_dupacks = 0;
5014 		tp->t_bytes_acked = 0;
5015 		EXIT_RECOVERY(tp->t_flags);
5016 		tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
5017 		    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
5018 		orig_cwnd = tp->snd_cwnd;
5019 		tp->snd_cwnd = ctf_fixed_maxseg(tp);
5020 		rack_log_to_prr(rack, 16, orig_cwnd);
5021 		if (tp->t_flags2 & TF2_ECN_PERMIT)
5022 			tp->t_flags2 |= TF2_ECN_SND_CWR;
5023 		break;
5024 	case CC_RTO_ERR:
5025 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
5026 		/* RTO was unnecessary, so reset everything. */
5027 		tp->snd_cwnd = tp->snd_cwnd_prev;
5028 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
5029 		tp->snd_recover = tp->snd_recover_prev;
5030 		if (tp->t_flags & TF_WASFRECOVERY) {
5031 			ENTER_FASTRECOVERY(tp->t_flags);
5032 			tp->t_flags &= ~TF_WASFRECOVERY;
5033 		}
5034 		if (tp->t_flags & TF_WASCRECOVERY) {
5035 			ENTER_CONGRECOVERY(tp->t_flags);
5036 			tp->t_flags &= ~TF_WASCRECOVERY;
5037 		}
5038 		tp->snd_nxt = tp->snd_max;
5039 		tp->t_badrxtwin = 0;
5040 		break;
5041 	}
5042 	if ((CC_ALGO(tp)->cong_signal != NULL)  &&
5043 	    (type != CC_RTO)){
5044 		tp->ccv->curack = ack;
5045 		CC_ALGO(tp)->cong_signal(tp->ccv, type);
5046 	}
5047 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
5048 		rack_log_to_prr(rack, 15, cwnd_enter);
5049 		rack->r_ctl.dsack_byte_cnt = 0;
5050 		rack->r_ctl.retran_during_recovery = 0;
5051 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
5052 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
5053 		rack->r_ent_rec_ns = 1;
5054 	}
5055 }
5056 
5057 static inline void
5058 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
5059 {
5060 	uint32_t i_cwnd;
5061 
5062 	INP_WLOCK_ASSERT(tp->t_inpcb);
5063 
5064 #ifdef NETFLIX_STATS
5065 	KMOD_TCPSTAT_INC(tcps_idle_restarts);
5066 	if (tp->t_state == TCPS_ESTABLISHED)
5067 		KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
5068 #endif
5069 	if (CC_ALGO(tp)->after_idle != NULL)
5070 		CC_ALGO(tp)->after_idle(tp->ccv);
5071 
5072 	if (tp->snd_cwnd == 1)
5073 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
5074 	else
5075 		i_cwnd = rc_init_window(rack);
5076 
5077 	/*
5078 	 * Being idle is no differnt than the initial window. If the cc
5079 	 * clamps it down below the initial window raise it to the initial
5080 	 * window.
5081 	 */
5082 	if (tp->snd_cwnd < i_cwnd) {
5083 		tp->snd_cwnd = i_cwnd;
5084 	}
5085 }
5086 
5087 /*
5088  * Indicate whether this ack should be delayed.  We can delay the ack if
5089  * following conditions are met:
5090  *	- There is no delayed ack timer in progress.
5091  *	- Our last ack wasn't a 0-sized window. We never want to delay
5092  *	  the ack that opens up a 0-sized window.
5093  *	- LRO wasn't used for this segment. We make sure by checking that the
5094  *	  segment size is not larger than the MSS.
5095  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
5096  *	  connection.
5097  */
5098 #define DELAY_ACK(tp, tlen)			 \
5099 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
5100 	((tp->t_flags & TF_DELACK) == 0) &&	 \
5101 	(tlen <= tp->t_maxseg) &&		 \
5102 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
5103 
5104 static struct rack_sendmap *
5105 rack_find_lowest_rsm(struct tcp_rack *rack)
5106 {
5107 	struct rack_sendmap *rsm;
5108 
5109 	/*
5110 	 * Walk the time-order transmitted list looking for an rsm that is
5111 	 * not acked. This will be the one that was sent the longest time
5112 	 * ago that is still outstanding.
5113 	 */
5114 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
5115 		if (rsm->r_flags & RACK_ACKED) {
5116 			continue;
5117 		}
5118 		goto finish;
5119 	}
5120 finish:
5121 	return (rsm);
5122 }
5123 
5124 static struct rack_sendmap *
5125 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
5126 {
5127 	struct rack_sendmap *prsm;
5128 
5129 	/*
5130 	 * Walk the sequence order list backward until we hit and arrive at
5131 	 * the highest seq not acked. In theory when this is called it
5132 	 * should be the last segment (which it was not).
5133 	 */
5134 	counter_u64_add(rack_find_high, 1);
5135 	prsm = rsm;
5136 	RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
5137 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
5138 			continue;
5139 		}
5140 		return (prsm);
5141 	}
5142 	return (NULL);
5143 }
5144 
5145 static uint32_t
5146 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
5147 {
5148 	int32_t lro;
5149 	uint32_t thresh;
5150 
5151 	/*
5152 	 * lro is the flag we use to determine if we have seen reordering.
5153 	 * If it gets set we have seen reordering. The reorder logic either
5154 	 * works in one of two ways:
5155 	 *
5156 	 * If reorder-fade is configured, then we track the last time we saw
5157 	 * re-ordering occur. If we reach the point where enough time as
5158 	 * passed we no longer consider reordering has occuring.
5159 	 *
5160 	 * Or if reorder-face is 0, then once we see reordering we consider
5161 	 * the connection to alway be subject to reordering and just set lro
5162 	 * to 1.
5163 	 *
5164 	 * In the end if lro is non-zero we add the extra time for
5165 	 * reordering in.
5166 	 */
5167 	if (srtt == 0)
5168 		srtt = 1;
5169 	if (rack->r_ctl.rc_reorder_ts) {
5170 		if (rack->r_ctl.rc_reorder_fade) {
5171 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5172 				lro = cts - rack->r_ctl.rc_reorder_ts;
5173 				if (lro == 0) {
5174 					/*
5175 					 * No time as passed since the last
5176 					 * reorder, mark it as reordering.
5177 					 */
5178 					lro = 1;
5179 				}
5180 			} else {
5181 				/* Negative time? */
5182 				lro = 0;
5183 			}
5184 			if (lro > rack->r_ctl.rc_reorder_fade) {
5185 				/* Turn off reordering seen too */
5186 				rack->r_ctl.rc_reorder_ts = 0;
5187 				lro = 0;
5188 			}
5189 		} else {
5190 			/* Reodering does not fade */
5191 			lro = 1;
5192 		}
5193 	} else {
5194 		lro = 0;
5195 	}
5196 	if (rack->rc_rack_tmr_std_based == 0) {
5197 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
5198 	} else {
5199 		/* Standards based pkt-delay is 1/4 srtt */
5200 		thresh = srtt +  (srtt >> 2);
5201 	}
5202 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
5203 		/* It must be set, if not you get 1/4 rtt */
5204 		if (rack->r_ctl.rc_reorder_shift)
5205 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5206 		else
5207 			thresh += (srtt >> 2);
5208 	}
5209 	if (rack->rc_rack_use_dsack &&
5210 	    lro &&
5211 	    (rack->r_ctl.num_dsack > 0)) {
5212 		/*
5213 		 * We only increase the reordering window if we
5214 		 * have seen reordering <and> we have a DSACK count.
5215 		 */
5216 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
5217 		rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
5218 	}
5219 	/* SRTT * 2 is the ceiling */
5220 	if (thresh > (srtt * 2)) {
5221 		thresh = srtt * 2;
5222 	}
5223 	/* And we don't want it above the RTO max either */
5224 	if (thresh > rack_rto_max) {
5225 		thresh = rack_rto_max;
5226 	}
5227 	rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
5228 	return (thresh);
5229 }
5230 
5231 static uint32_t
5232 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5233 		     struct rack_sendmap *rsm, uint32_t srtt)
5234 {
5235 	struct rack_sendmap *prsm;
5236 	uint32_t thresh, len;
5237 	int segsiz;
5238 
5239 	if (srtt == 0)
5240 		srtt = 1;
5241 	if (rack->r_ctl.rc_tlp_threshold)
5242 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5243 	else
5244 		thresh = (srtt * 2);
5245 
5246 	/* Get the previous sent packet, if any */
5247 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5248 	counter_u64_add(rack_enter_tlp_calc, 1);
5249 	len = rsm->r_end - rsm->r_start;
5250 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5251 		/* Exactly like the ID */
5252 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5253 			uint32_t alt_thresh;
5254 			/*
5255 			 * Compensate for delayed-ack with the d-ack time.
5256 			 */
5257 			counter_u64_add(rack_used_tlpmethod, 1);
5258 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5259 			if (alt_thresh > thresh)
5260 				thresh = alt_thresh;
5261 		}
5262 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5263 		/* 2.1 behavior */
5264 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5265 		if (prsm && (len <= segsiz)) {
5266 			/*
5267 			 * Two packets outstanding, thresh should be (2*srtt) +
5268 			 * possible inter-packet delay (if any).
5269 			 */
5270 			uint32_t inter_gap = 0;
5271 			int idx, nidx;
5272 
5273 			counter_u64_add(rack_used_tlpmethod, 1);
5274 			idx = rsm->r_rtr_cnt - 1;
5275 			nidx = prsm->r_rtr_cnt - 1;
5276 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5277 				/* Yes it was sent later (or at the same time) */
5278 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5279 			}
5280 			thresh += inter_gap;
5281 		} else if (len <= segsiz) {
5282 			/*
5283 			 * Possibly compensate for delayed-ack.
5284 			 */
5285 			uint32_t alt_thresh;
5286 
5287 			counter_u64_add(rack_used_tlpmethod2, 1);
5288 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5289 			if (alt_thresh > thresh)
5290 				thresh = alt_thresh;
5291 		}
5292 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5293 		/* 2.2 behavior */
5294 		if (len <= segsiz) {
5295 			uint32_t alt_thresh;
5296 			/*
5297 			 * Compensate for delayed-ack with the d-ack time.
5298 			 */
5299 			counter_u64_add(rack_used_tlpmethod, 1);
5300 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5301 			if (alt_thresh > thresh)
5302 				thresh = alt_thresh;
5303 		}
5304 	}
5305 	/* Not above an RTO */
5306 	if (thresh > tp->t_rxtcur) {
5307 		thresh = tp->t_rxtcur;
5308 	}
5309 	/* Not above a RTO max */
5310 	if (thresh > rack_rto_max) {
5311 		thresh = rack_rto_max;
5312 	}
5313 	/* Apply user supplied min TLP */
5314 	if (thresh < rack_tlp_min) {
5315 		thresh = rack_tlp_min;
5316 	}
5317 	return (thresh);
5318 }
5319 
5320 static uint32_t
5321 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5322 {
5323 	/*
5324 	 * We want the rack_rtt which is the
5325 	 * last rtt we measured. However if that
5326 	 * does not exist we fallback to the srtt (which
5327 	 * we probably will never do) and then as a last
5328 	 * resort we use RACK_INITIAL_RTO if no srtt is
5329 	 * yet set.
5330 	 */
5331 	if (rack->rc_rack_rtt)
5332 		return (rack->rc_rack_rtt);
5333 	else if (tp->t_srtt == 0)
5334 		return (RACK_INITIAL_RTO);
5335 	return (tp->t_srtt);
5336 }
5337 
5338 static struct rack_sendmap *
5339 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5340 {
5341 	/*
5342 	 * Check to see that we don't need to fall into recovery. We will
5343 	 * need to do so if our oldest transmit is past the time we should
5344 	 * have had an ack.
5345 	 */
5346 	struct tcp_rack *rack;
5347 	struct rack_sendmap *rsm;
5348 	int32_t idx;
5349 	uint32_t srtt, thresh;
5350 
5351 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5352 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5353 		return (NULL);
5354 	}
5355 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5356 	if (rsm == NULL)
5357 		return (NULL);
5358 
5359 	if (rsm->r_flags & RACK_ACKED) {
5360 		rsm = rack_find_lowest_rsm(rack);
5361 		if (rsm == NULL)
5362 			return (NULL);
5363 	}
5364 	idx = rsm->r_rtr_cnt - 1;
5365 	srtt = rack_grab_rtt(tp, rack);
5366 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5367 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5368 		return (NULL);
5369 	}
5370 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5371 		return (NULL);
5372 	}
5373 	/* Ok if we reach here we are over-due and this guy can be sent */
5374 	if (IN_RECOVERY(tp->t_flags) == 0) {
5375 		/*
5376 		 * For the one that enters us into recovery record undo
5377 		 * info.
5378 		 */
5379 		rack->r_ctl.rc_rsm_start = rsm->r_start;
5380 		rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
5381 		rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
5382 	}
5383 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
5384 	return (rsm);
5385 }
5386 
5387 static uint32_t
5388 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5389 {
5390 	int32_t t;
5391 	int32_t tt;
5392 	uint32_t ret_val;
5393 
5394 	t = (tp->t_srtt + (tp->t_rttvar << 2));
5395 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5396  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5397 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5398 	ret_val = (uint32_t)tt;
5399 	return (ret_val);
5400 }
5401 
5402 static uint32_t
5403 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5404 {
5405 	/*
5406 	 * Start the FR timer, we do this based on getting the first one in
5407 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
5408 	 * events we need to stop the running timer (if its running) before
5409 	 * starting the new one.
5410 	 */
5411 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5412 	uint32_t srtt_cur;
5413 	int32_t idx;
5414 	int32_t is_tlp_timer = 0;
5415 	struct rack_sendmap *rsm;
5416 
5417 	if (rack->t_timers_stopped) {
5418 		/* All timers have been stopped none are to run */
5419 		return (0);
5420 	}
5421 	if (rack->rc_in_persist) {
5422 		/* We can't start any timer in persists */
5423 		return (rack_get_persists_timer_val(tp, rack));
5424 	}
5425 	rack->rc_on_min_to = 0;
5426 	if ((tp->t_state < TCPS_ESTABLISHED) ||
5427 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5428 		goto activate_rxt;
5429 	}
5430 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5431 	if ((rsm == NULL) || sup_rack) {
5432 		/* Nothing on the send map or no rack */
5433 activate_rxt:
5434 		time_since_sent = 0;
5435 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5436 		if (rsm) {
5437 			/*
5438 			 * Should we discount the RTX timer any?
5439 			 *
5440 			 * We want to discount it the smallest amount.
5441 			 * If a timer (Rack/TLP or RXT) has gone off more
5442 			 * recently thats the discount we want to use (now - timer time).
5443 			 * If the retransmit of the oldest packet was more recent then
5444 			 * we want to use that (now - oldest-packet-last_transmit_time).
5445 			 *
5446 			 */
5447 			idx = rsm->r_rtr_cnt - 1;
5448 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5449 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5450 			else
5451 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5452 			if (TSTMP_GT(cts, tstmp_touse))
5453 			    time_since_sent = cts - tstmp_touse;
5454 		}
5455 		if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
5456 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5457 			to = tp->t_rxtcur;
5458 			if (to > time_since_sent)
5459 				to -= time_since_sent;
5460 			else
5461 				to = rack->r_ctl.rc_min_to;
5462 			if (to == 0)
5463 				to = 1;
5464 			/* Special case for KEEPINIT */
5465 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5466 			    (TP_KEEPINIT(tp) != 0) &&
5467 			    rsm) {
5468 				/*
5469 				 * We have to put a ceiling on the rxt timer
5470 				 * of the keep-init timeout.
5471 				 */
5472 				uint32_t max_time, red;
5473 
5474 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5475 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5476 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5477 					if (red < max_time)
5478 						max_time -= red;
5479 					else
5480 						max_time = 1;
5481 				}
5482 				/* Reduce timeout to the keep value if needed */
5483 				if (max_time < to)
5484 					to = max_time;
5485 			}
5486 			return (to);
5487 		}
5488 		return (0);
5489 	}
5490 	if (rsm->r_flags & RACK_ACKED) {
5491 		rsm = rack_find_lowest_rsm(rack);
5492 		if (rsm == NULL) {
5493 			/* No lowest? */
5494 			goto activate_rxt;
5495 		}
5496 	}
5497 	if (rack->sack_attack_disable) {
5498 		/*
5499 		 * We don't want to do
5500 		 * any TLP's if you are an attacker.
5501 		 * Though if you are doing what
5502 		 * is expected you may still have
5503 		 * SACK-PASSED marks.
5504 		 */
5505 		goto activate_rxt;
5506 	}
5507 	/* Convert from ms to usecs */
5508 	if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5509 		if ((tp->t_flags & TF_SENTFIN) &&
5510 		    ((tp->snd_max - tp->snd_una) == 1) &&
5511 		    (rsm->r_flags & RACK_HAS_FIN)) {
5512 			/*
5513 			 * We don't start a rack timer if all we have is a
5514 			 * FIN outstanding.
5515 			 */
5516 			goto activate_rxt;
5517 		}
5518 		if ((rack->use_rack_rr == 0) &&
5519 		    (IN_FASTRECOVERY(tp->t_flags)) &&
5520 		    (rack->rack_no_prr == 0) &&
5521 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5522 			/*
5523 			 * We are not cheating, in recovery  and
5524 			 * not enough ack's to yet get our next
5525 			 * retransmission out.
5526 			 *
5527 			 * Note that classified attackers do not
5528 			 * get to use the rack-cheat.
5529 			 */
5530 			goto activate_tlp;
5531 		}
5532 		srtt = rack_grab_rtt(tp, rack);
5533 		thresh = rack_calc_thresh_rack(rack, srtt, cts);
5534 		idx = rsm->r_rtr_cnt - 1;
5535 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5536 		if (SEQ_GEQ(exp, cts)) {
5537 			to = exp - cts;
5538 			if (to < rack->r_ctl.rc_min_to) {
5539 				to = rack->r_ctl.rc_min_to;
5540 				if (rack->r_rr_config == 3)
5541 					rack->rc_on_min_to = 1;
5542 			}
5543 		} else {
5544 			to = rack->r_ctl.rc_min_to;
5545 			if (rack->r_rr_config == 3)
5546 				rack->rc_on_min_to = 1;
5547 		}
5548 	} else {
5549 		/* Ok we need to do a TLP not RACK */
5550 activate_tlp:
5551 		if ((rack->rc_tlp_in_progress != 0) &&
5552 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5553 			/*
5554 			 * The previous send was a TLP and we have sent
5555 			 * N TLP's without sending new data.
5556 			 */
5557 			goto activate_rxt;
5558 		}
5559 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5560 		if (rsm == NULL) {
5561 			/* We found no rsm to TLP with. */
5562 			goto activate_rxt;
5563 		}
5564 		if (rsm->r_flags & RACK_HAS_FIN) {
5565 			/* If its a FIN we dont do TLP */
5566 			rsm = NULL;
5567 			goto activate_rxt;
5568 		}
5569 		idx = rsm->r_rtr_cnt - 1;
5570 		time_since_sent = 0;
5571 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5572 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5573 		else
5574 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5575 		if (TSTMP_GT(cts, tstmp_touse))
5576 		    time_since_sent = cts - tstmp_touse;
5577 		is_tlp_timer = 1;
5578 		if (tp->t_srtt) {
5579 			if ((rack->rc_srtt_measure_made == 0) &&
5580 			    (tp->t_srtt == 1)) {
5581 				/*
5582 				 * If another stack as run and set srtt to 1,
5583 				 * then the srtt was 0, so lets use the initial.
5584 				 */
5585 				srtt = RACK_INITIAL_RTO;
5586 			} else {
5587 				srtt_cur = tp->t_srtt;
5588 				srtt = srtt_cur;
5589 			}
5590 		} else
5591 			srtt = RACK_INITIAL_RTO;
5592 		/*
5593 		 * If the SRTT is not keeping up and the
5594 		 * rack RTT has spiked we want to use
5595 		 * the last RTT not the smoothed one.
5596 		 */
5597 		if (rack_tlp_use_greater &&
5598 		    tp->t_srtt &&
5599 		    (srtt < rack_grab_rtt(tp, rack))) {
5600 			srtt = rack_grab_rtt(tp, rack);
5601 		}
5602 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5603 		if (thresh > time_since_sent) {
5604 			to = thresh - time_since_sent;
5605 		} else {
5606 			to = rack->r_ctl.rc_min_to;
5607 			rack_log_alt_to_to_cancel(rack,
5608 						  thresh,		/* flex1 */
5609 						  time_since_sent,	/* flex2 */
5610 						  tstmp_touse,		/* flex3 */
5611 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5612 						  (uint32_t)rsm->r_tim_lastsent[idx],
5613 						  srtt,
5614 						  idx, 99);
5615 		}
5616 		if (to < rack_tlp_min) {
5617 			to = rack_tlp_min;
5618 		}
5619 		if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5620 			/*
5621 			 * If the TLP time works out to larger than the max
5622 			 * RTO lets not do TLP.. just RTO.
5623 			 */
5624 			goto activate_rxt;
5625 		}
5626 	}
5627 	if (is_tlp_timer == 0) {
5628 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5629 	} else {
5630 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5631 	}
5632 	if (to == 0)
5633 		to = 1;
5634 	return (to);
5635 }
5636 
5637 static void
5638 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5639 {
5640 	if (rack->rc_in_persist == 0) {
5641 		if (tp->t_flags & TF_GPUTINPROG) {
5642 			/*
5643 			 * Stop the goodput now, the calling of the
5644 			 * measurement function clears the flag.
5645 			 */
5646 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
5647 						    RACK_QUALITY_PERSIST);
5648 		}
5649 #ifdef NETFLIX_SHARED_CWND
5650 		if (rack->r_ctl.rc_scw) {
5651 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5652 			rack->rack_scwnd_is_idle = 1;
5653 		}
5654 #endif
5655 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5656 		if (rack->r_ctl.rc_went_idle_time == 0)
5657 			rack->r_ctl.rc_went_idle_time = 1;
5658 		rack_timer_cancel(tp, rack, cts, __LINE__);
5659 		rack->r_ctl.persist_lost_ends = 0;
5660 		rack->probe_not_answered = 0;
5661 		rack->forced_ack = 0;
5662 		tp->t_rxtshift = 0;
5663 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5664 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5665 		rack->rc_in_persist = 1;
5666 	}
5667 }
5668 
5669 static void
5670 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5671 {
5672 	if (tcp_in_hpts(rack->rc_inp)) {
5673 		tcp_hpts_remove(rack->rc_inp);
5674 		rack->r_ctl.rc_hpts_flags = 0;
5675 	}
5676 #ifdef NETFLIX_SHARED_CWND
5677 	if (rack->r_ctl.rc_scw) {
5678 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5679 		rack->rack_scwnd_is_idle = 0;
5680 	}
5681 #endif
5682 	if (rack->rc_gp_dyn_mul &&
5683 	    (rack->use_fixed_rate == 0) &&
5684 	    (rack->rc_always_pace)) {
5685 		/*
5686 		 * Do we count this as if a probe-rtt just
5687 		 * finished?
5688 		 */
5689 		uint32_t time_idle, idle_min;
5690 
5691 		time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5692 		idle_min = rack_min_probertt_hold;
5693 		if (rack_probertt_gpsrtt_cnt_div) {
5694 			uint64_t extra;
5695 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5696 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
5697 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5698 			idle_min += (uint32_t)extra;
5699 		}
5700 		if (time_idle >= idle_min) {
5701 			/* Yes, we count it as a probe-rtt. */
5702 			uint32_t us_cts;
5703 
5704 			us_cts = tcp_get_usecs(NULL);
5705 			if (rack->in_probe_rtt == 0) {
5706 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5707 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5708 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5709 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5710 			} else {
5711 				rack_exit_probertt(rack, us_cts);
5712 			}
5713 		}
5714 	}
5715 	rack->rc_in_persist = 0;
5716 	rack->r_ctl.rc_went_idle_time = 0;
5717 	tp->t_rxtshift = 0;
5718 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5719 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5720 	rack->r_ctl.rc_agg_delayed = 0;
5721 	rack->r_early = 0;
5722 	rack->r_late = 0;
5723 	rack->r_ctl.rc_agg_early = 0;
5724 }
5725 
5726 static void
5727 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5728 		   struct hpts_diag *diag, struct timeval *tv)
5729 {
5730 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5731 		union tcp_log_stackspecific log;
5732 
5733 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5734 		log.u_bbr.flex1 = diag->p_nxt_slot;
5735 		log.u_bbr.flex2 = diag->p_cur_slot;
5736 		log.u_bbr.flex3 = diag->slot_req;
5737 		log.u_bbr.flex4 = diag->inp_hptsslot;
5738 		log.u_bbr.flex5 = diag->slot_remaining;
5739 		log.u_bbr.flex6 = diag->need_new_to;
5740 		log.u_bbr.flex7 = diag->p_hpts_active;
5741 		log.u_bbr.flex8 = diag->p_on_min_sleep;
5742 		/* Hijack other fields as needed */
5743 		log.u_bbr.epoch = diag->have_slept;
5744 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
5745 		log.u_bbr.pkts_out = diag->co_ret;
5746 		log.u_bbr.applimited = diag->hpts_sleep_time;
5747 		log.u_bbr.delivered = diag->p_prev_slot;
5748 		log.u_bbr.inflight = diag->p_runningslot;
5749 		log.u_bbr.bw_inuse = diag->wheel_slot;
5750 		log.u_bbr.rttProp = diag->wheel_cts;
5751 		log.u_bbr.timeStamp = cts;
5752 		log.u_bbr.delRate = diag->maxslots;
5753 		log.u_bbr.cur_del_rate = diag->p_curtick;
5754 		log.u_bbr.cur_del_rate <<= 32;
5755 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
5756 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5757 		    &rack->rc_inp->inp_socket->so_rcv,
5758 		    &rack->rc_inp->inp_socket->so_snd,
5759 		    BBR_LOG_HPTSDIAG, 0,
5760 		    0, &log, false, tv);
5761 	}
5762 
5763 }
5764 
5765 static void
5766 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5767 {
5768 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5769 		union tcp_log_stackspecific log;
5770 		struct timeval tv;
5771 
5772 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5773 		log.u_bbr.flex1 = sb->sb_flags;
5774 		log.u_bbr.flex2 = len;
5775 		log.u_bbr.flex3 = sb->sb_state;
5776 		log.u_bbr.flex8 = type;
5777 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5778 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5779 		    &rack->rc_inp->inp_socket->so_rcv,
5780 		    &rack->rc_inp->inp_socket->so_snd,
5781 		    TCP_LOG_SB_WAKE, 0,
5782 		    len, &log, false, &tv);
5783 	}
5784 }
5785 
5786 static void
5787 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5788       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5789 {
5790 	struct hpts_diag diag;
5791 	struct inpcb *inp;
5792 	struct timeval tv;
5793 	uint32_t delayed_ack = 0;
5794 	uint32_t hpts_timeout;
5795 	uint32_t entry_slot = slot;
5796 	uint8_t stopped;
5797 	uint32_t left = 0;
5798 	uint32_t us_cts;
5799 
5800 	inp = tp->t_inpcb;
5801 	if ((tp->t_state == TCPS_CLOSED) ||
5802 	    (tp->t_state == TCPS_LISTEN)) {
5803 		return;
5804 	}
5805 	if (tcp_in_hpts(inp)) {
5806 		/* Already on the pacer */
5807 		return;
5808 	}
5809 	stopped = rack->rc_tmr_stopped;
5810 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5811 		left = rack->r_ctl.rc_timer_exp - cts;
5812 	}
5813 	rack->r_ctl.rc_timer_exp = 0;
5814 	rack->r_ctl.rc_hpts_flags = 0;
5815 	us_cts = tcp_get_usecs(&tv);
5816 	/* Now early/late accounting */
5817 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
5818 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5819 		/*
5820 		 * We have a early carry over set,
5821 		 * we can always add more time so we
5822 		 * can always make this compensation.
5823 		 *
5824 		 * Note if ack's are allowed to wake us do not
5825 		 * penalize the next timer for being awoke
5826 		 * by an ack aka the rc_agg_early (non-paced mode).
5827 		 */
5828 		slot += rack->r_ctl.rc_agg_early;
5829 		rack->r_early = 0;
5830 		rack->r_ctl.rc_agg_early = 0;
5831 	}
5832 	if (rack->r_late) {
5833 		/*
5834 		 * This is harder, we can
5835 		 * compensate some but it
5836 		 * really depends on what
5837 		 * the current pacing time is.
5838 		 */
5839 		if (rack->r_ctl.rc_agg_delayed >= slot) {
5840 			/*
5841 			 * We can't compensate for it all.
5842 			 * And we have to have some time
5843 			 * on the clock. We always have a min
5844 			 * 10 slots (10 x 10 i.e. 100 usecs).
5845 			 */
5846 			if (slot <= HPTS_TICKS_PER_SLOT) {
5847 				/* We gain delay */
5848 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
5849 				slot = HPTS_TICKS_PER_SLOT;
5850 			} else {
5851 				/* We take off some */
5852 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
5853 				slot = HPTS_TICKS_PER_SLOT;
5854 			}
5855 		} else {
5856 			slot -= rack->r_ctl.rc_agg_delayed;
5857 			rack->r_ctl.rc_agg_delayed = 0;
5858 			/* Make sure we have 100 useconds at minimum */
5859 			if (slot < HPTS_TICKS_PER_SLOT) {
5860 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
5861 				slot = HPTS_TICKS_PER_SLOT;
5862 			}
5863 			if (rack->r_ctl.rc_agg_delayed == 0)
5864 				rack->r_late = 0;
5865 		}
5866 	}
5867 	if (slot) {
5868 		/* We are pacing too */
5869 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5870 	}
5871 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5872 #ifdef NETFLIX_EXP_DETECTION
5873 	if (rack->sack_attack_disable &&
5874 	    (slot < tcp_sad_pacing_interval)) {
5875 		/*
5876 		 * We have a potential attacker on
5877 		 * the line. We have possibly some
5878 		 * (or now) pacing time set. We want to
5879 		 * slow down the processing of sacks by some
5880 		 * amount (if it is an attacker). Set the default
5881 		 * slot for attackers in place (unless the orginal
5882 		 * interval is longer). Its stored in
5883 		 * micro-seconds, so lets convert to msecs.
5884 		 */
5885 		slot = tcp_sad_pacing_interval;
5886 	}
5887 #endif
5888 	if (tp->t_flags & TF_DELACK) {
5889 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
5890 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5891 	}
5892 	if (delayed_ack && ((hpts_timeout == 0) ||
5893 			    (delayed_ack < hpts_timeout)))
5894 		hpts_timeout = delayed_ack;
5895 	else
5896 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5897 	/*
5898 	 * If no timers are going to run and we will fall off the hptsi
5899 	 * wheel, we resort to a keep-alive timer if its configured.
5900 	 */
5901 	if ((hpts_timeout == 0) &&
5902 	    (slot == 0)) {
5903 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5904 		    (tp->t_state <= TCPS_CLOSING)) {
5905 			/*
5906 			 * Ok we have no timer (persists, rack, tlp, rxt  or
5907 			 * del-ack), we don't have segments being paced. So
5908 			 * all that is left is the keepalive timer.
5909 			 */
5910 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5911 				/* Get the established keep-alive time */
5912 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5913 			} else {
5914 				/*
5915 				 * Get the initial setup keep-alive time,
5916 				 * note that this is probably not going to
5917 				 * happen, since rack will be running a rxt timer
5918 				 * if a SYN of some sort is outstanding. It is
5919 				 * actually handled in rack_timeout_rxt().
5920 				 */
5921 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5922 			}
5923 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5924 			if (rack->in_probe_rtt) {
5925 				/*
5926 				 * We want to instead not wake up a long time from
5927 				 * now but to wake up about the time we would
5928 				 * exit probe-rtt and initiate a keep-alive ack.
5929 				 * This will get us out of probe-rtt and update
5930 				 * our min-rtt.
5931 				 */
5932 				hpts_timeout = rack_min_probertt_hold;
5933 			}
5934 		}
5935 	}
5936 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5937 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5938 		/*
5939 		 * RACK, TLP, persists and RXT timers all are restartable
5940 		 * based on actions input .. i.e we received a packet (ack
5941 		 * or sack) and that changes things (rw, or snd_una etc).
5942 		 * Thus we can restart them with a new value. For
5943 		 * keep-alive, delayed_ack we keep track of what was left
5944 		 * and restart the timer with a smaller value.
5945 		 */
5946 		if (left < hpts_timeout)
5947 			hpts_timeout = left;
5948 	}
5949 	if (hpts_timeout) {
5950 		/*
5951 		 * Hack alert for now we can't time-out over 2,147,483
5952 		 * seconds (a bit more than 596 hours), which is probably ok
5953 		 * :).
5954 		 */
5955 		if (hpts_timeout > 0x7ffffffe)
5956 			hpts_timeout = 0x7ffffffe;
5957 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5958 	}
5959 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
5960 	if ((rack->gp_ready == 0) &&
5961 	    (rack->use_fixed_rate == 0) &&
5962 	    (hpts_timeout < slot) &&
5963 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5964 		/*
5965 		 * We have no good estimate yet for the
5966 		 * old clunky burst mitigation or the
5967 		 * real pacing. And the tlp or rxt is smaller
5968 		 * than the pacing calculation. Lets not
5969 		 * pace that long since we know the calculation
5970 		 * so far is not accurate.
5971 		 */
5972 		slot = hpts_timeout;
5973 	}
5974 	rack->r_ctl.last_pacing_time = slot;
5975 	/**
5976 	 * Turn off all the flags for queuing by default. The
5977 	 * flags have important meanings to what happens when
5978 	 * LRO interacts with the transport. Most likely (by default now)
5979 	 * mbuf_queueing and ack compression are on. So the transport
5980 	 * has a couple of flags that control what happens (if those
5981 	 * are not on then these flags won't have any effect since it
5982 	 * won't go through the queuing LRO path).
5983 	 *
5984 	 * INP_MBUF_QUEUE_READY - This flags says that I am busy
5985 	 *                        pacing output, so don't disturb. But
5986 	 *                        it also means LRO can wake me if there
5987 	 *                        is a SACK arrival.
5988 	 *
5989 	 * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5990 	 *                       with the above flag (QUEUE_READY) and
5991 	 *                       when present it says don't even wake me
5992 	 *                       if a SACK arrives.
5993 	 *
5994 	 * The idea behind these flags is that if we are pacing we
5995 	 * set the MBUF_QUEUE_READY and only get woken up if
5996 	 * a SACK arrives (which could change things) or if
5997 	 * our pacing timer expires. If, however, we have a rack
5998 	 * timer running, then we don't even want a sack to wake
5999 	 * us since the rack timer has to expire before we can send.
6000 	 *
6001 	 * Other cases should usually have none of the flags set
6002 	 * so LRO can call into us.
6003 	 */
6004 	inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
6005 	if (slot) {
6006 		rack->r_ctl.rc_last_output_to = us_cts + slot;
6007 		/*
6008 		 * A pacing timer (slot) is being set, in
6009 		 * such a case we cannot send (we are blocked by
6010 		 * the timer). So lets tell LRO that it should not
6011 		 * wake us unless there is a SACK. Note this only
6012 		 * will be effective if mbuf queueing is on or
6013 		 * compressed acks are being processed.
6014 		 */
6015 		inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
6016 		/*
6017 		 * But wait if we have a Rack timer running
6018 		 * even a SACK should not disturb us (with
6019 		 * the exception of r_rr_config 3).
6020 		 */
6021 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
6022 		    (rack->r_rr_config != 3))
6023 			inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
6024 		if (rack->rc_ack_can_sendout_data) {
6025 			/*
6026 			 * Ahh but wait, this is that special case
6027 			 * where the pacing timer can be disturbed
6028 			 * backout the changes (used for non-paced
6029 			 * burst limiting).
6030 			 */
6031 			inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
6032 		}
6033 		if ((rack->use_rack_rr) &&
6034 		    (rack->r_rr_config < 2) &&
6035 		    ((hpts_timeout) && (hpts_timeout < slot))) {
6036 			/*
6037 			 * Arrange for the hpts to kick back in after the
6038 			 * t-o if the t-o does not cause a send.
6039 			 */
6040 			(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
6041 						   __LINE__, &diag);
6042 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6043 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6044 		} else {
6045 			(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
6046 						   __LINE__, &diag);
6047 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6048 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
6049 		}
6050 	} else if (hpts_timeout) {
6051 		/*
6052 		 * With respect to inp_flags2 here, lets let any new acks wake
6053 		 * us up here. Since we are not pacing (no pacing timer), output
6054 		 * can happen so we should let it. If its a Rack timer, then any inbound
6055 		 * packet probably won't change the sending (we will be blocked)
6056 		 * but it may change the prr stats so letting it in (the set defaults
6057 		 * at the start of this block) are good enough.
6058 		 */
6059 		(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
6060 					   __LINE__, &diag);
6061 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6062 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6063 	} else {
6064 		/* No timer starting */
6065 #ifdef INVARIANTS
6066 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
6067 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
6068 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
6069 		}
6070 #endif
6071 	}
6072 	rack->rc_tmr_stopped = 0;
6073 	if (slot)
6074 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
6075 }
6076 
6077 /*
6078  * RACK Timer, here we simply do logging and house keeping.
6079  * the normal rack_output() function will call the
6080  * appropriate thing to check if we need to do a RACK retransmit.
6081  * We return 1, saying don't proceed with rack_output only
6082  * when all timers have been stopped (destroyed PCB?).
6083  */
6084 static int
6085 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6086 {
6087 	/*
6088 	 * This timer simply provides an internal trigger to send out data.
6089 	 * The check_recovery_mode call will see if there are needed
6090 	 * retransmissions, if so we will enter fast-recovery. The output
6091 	 * call may or may not do the same thing depending on sysctl
6092 	 * settings.
6093 	 */
6094 	struct rack_sendmap *rsm;
6095 
6096 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6097 		return (1);
6098 	}
6099 	counter_u64_add(rack_to_tot, 1);
6100 	if (rack->r_state && (rack->r_state != tp->t_state))
6101 		rack_set_state(tp, rack);
6102 	rack->rc_on_min_to = 0;
6103 	rsm = rack_check_recovery_mode(tp, cts);
6104 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
6105 	if (rsm) {
6106 		rack->r_ctl.rc_resend = rsm;
6107 		rack->r_timer_override = 1;
6108 		if (rack->use_rack_rr) {
6109 			/*
6110 			 * Don't accumulate extra pacing delay
6111 			 * we are allowing the rack timer to
6112 			 * over-ride pacing i.e. rrr takes precedence
6113 			 * if the pacing interval is longer than the rrr
6114 			 * time (in other words we get the min pacing
6115 			 * time versus rrr pacing time).
6116 			 */
6117 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6118 		}
6119 	}
6120 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
6121 	if (rsm == NULL) {
6122 		/* restart a timer and return 1 */
6123 		rack_start_hpts_timer(rack, tp, cts,
6124 				      0, 0, 0);
6125 		return (1);
6126 	}
6127 	return (0);
6128 }
6129 
6130 static void
6131 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
6132 {
6133 	if (rsm->m->m_len > rsm->orig_m_len) {
6134 		/*
6135 		 * Mbuf grew, caused by sbcompress, our offset does
6136 		 * not change.
6137 		 */
6138 		rsm->orig_m_len = rsm->m->m_len;
6139 	} else if (rsm->m->m_len < rsm->orig_m_len) {
6140 		/*
6141 		 * Mbuf shrank, trimmed off the top by an ack, our
6142 		 * offset changes.
6143 		 */
6144 		rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
6145 		rsm->orig_m_len = rsm->m->m_len;
6146 	}
6147 }
6148 
6149 static void
6150 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
6151 {
6152 	struct mbuf *m;
6153 	uint32_t soff;
6154 
6155 	if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
6156 		/* Fix up the orig_m_len and possibly the mbuf offset */
6157 		rack_adjust_orig_mlen(src_rsm);
6158 	}
6159 	m = src_rsm->m;
6160 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
6161 	while (soff >= m->m_len) {
6162 		/* Move out past this mbuf */
6163 		soff -= m->m_len;
6164 		m = m->m_next;
6165 		KASSERT((m != NULL),
6166 			("rsm:%p nrsm:%p hit at soff:%u null m",
6167 			 src_rsm, rsm, soff));
6168 	}
6169 	rsm->m = m;
6170 	rsm->soff = soff;
6171 	rsm->orig_m_len = m->m_len;
6172 }
6173 
6174 static __inline void
6175 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
6176 	       struct rack_sendmap *rsm, uint32_t start)
6177 {
6178 	int idx;
6179 
6180 	nrsm->r_start = start;
6181 	nrsm->r_end = rsm->r_end;
6182 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
6183 	nrsm->r_flags = rsm->r_flags;
6184 	nrsm->r_dupack = rsm->r_dupack;
6185 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
6186 	nrsm->r_rtr_bytes = 0;
6187 	nrsm->r_fas = rsm->r_fas;
6188 	rsm->r_end = nrsm->r_start;
6189 	nrsm->r_just_ret = rsm->r_just_ret;
6190 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
6191 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
6192 	}
6193 	/* Now if we have SYN flag we keep it on the left edge */
6194 	if (nrsm->r_flags & RACK_HAS_SYN)
6195 		nrsm->r_flags &= ~RACK_HAS_SYN;
6196 	/* Now if we have a FIN flag we keep it on the right edge */
6197 	if (rsm->r_flags & RACK_HAS_FIN)
6198 		rsm->r_flags &= ~RACK_HAS_FIN;
6199 	/* Push bit must go to the right edge as well */
6200 	if (rsm->r_flags & RACK_HAD_PUSH)
6201 		rsm->r_flags &= ~RACK_HAD_PUSH;
6202 	/* Clone over the state of the hw_tls flag */
6203 	nrsm->r_hw_tls = rsm->r_hw_tls;
6204 	/*
6205 	 * Now we need to find nrsm's new location in the mbuf chain
6206 	 * we basically calculate a new offset, which is soff +
6207 	 * how much is left in original rsm. Then we walk out the mbuf
6208 	 * chain to find the righ postion, it may be the same mbuf
6209 	 * or maybe not.
6210 	 */
6211 	KASSERT(((rsm->m != NULL) ||
6212 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
6213 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
6214 	if (rsm->m)
6215 		rack_setup_offset_for_rsm(rsm, nrsm);
6216 }
6217 
6218 static struct rack_sendmap *
6219 rack_merge_rsm(struct tcp_rack *rack,
6220 	       struct rack_sendmap *l_rsm,
6221 	       struct rack_sendmap *r_rsm)
6222 {
6223 	/*
6224 	 * We are merging two ack'd RSM's,
6225 	 * the l_rsm is on the left (lower seq
6226 	 * values) and the r_rsm is on the right
6227 	 * (higher seq value). The simplest way
6228 	 * to merge these is to move the right
6229 	 * one into the left. I don't think there
6230 	 * is any reason we need to try to find
6231 	 * the oldest (or last oldest retransmitted).
6232 	 */
6233 	struct rack_sendmap *rm;
6234 
6235 	rack_log_map_chg(rack->rc_tp, rack, NULL,
6236 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
6237 	l_rsm->r_end = r_rsm->r_end;
6238 	if (l_rsm->r_dupack < r_rsm->r_dupack)
6239 		l_rsm->r_dupack = r_rsm->r_dupack;
6240 	if (r_rsm->r_rtr_bytes)
6241 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
6242 	if (r_rsm->r_in_tmap) {
6243 		/* This really should not happen */
6244 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
6245 		r_rsm->r_in_tmap = 0;
6246 	}
6247 
6248 	/* Now the flags */
6249 	if (r_rsm->r_flags & RACK_HAS_FIN)
6250 		l_rsm->r_flags |= RACK_HAS_FIN;
6251 	if (r_rsm->r_flags & RACK_TLP)
6252 		l_rsm->r_flags |= RACK_TLP;
6253 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
6254 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
6255 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
6256 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6257 		/*
6258 		 * If both are app-limited then let the
6259 		 * free lower the count. If right is app
6260 		 * limited and left is not, transfer.
6261 		 */
6262 		l_rsm->r_flags |= RACK_APP_LIMITED;
6263 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
6264 		if (r_rsm == rack->r_ctl.rc_first_appl)
6265 			rack->r_ctl.rc_first_appl = l_rsm;
6266 	}
6267 	rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6268 #ifdef INVARIANTS
6269 	if (rm != r_rsm) {
6270 		panic("removing head in rack:%p rsm:%p rm:%p",
6271 		      rack, r_rsm, rm);
6272 	}
6273 #endif
6274 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6275 		/* Transfer the split limit to the map we free */
6276 		r_rsm->r_limit_type = l_rsm->r_limit_type;
6277 		l_rsm->r_limit_type = 0;
6278 	}
6279 	rack_free(rack, r_rsm);
6280 	return (l_rsm);
6281 }
6282 
6283 /*
6284  * TLP Timer, here we simply setup what segment we want to
6285  * have the TLP expire on, the normal rack_output() will then
6286  * send it out.
6287  *
6288  * We return 1, saying don't proceed with rack_output only
6289  * when all timers have been stopped (destroyed PCB?).
6290  */
6291 static int
6292 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
6293 {
6294 	/*
6295 	 * Tail Loss Probe.
6296 	 */
6297 	struct rack_sendmap *rsm = NULL;
6298 	struct rack_sendmap *insret;
6299 	struct socket *so;
6300 	uint32_t amm;
6301 	uint32_t out, avail;
6302 	int collapsed_win = 0;
6303 
6304 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6305 		return (1);
6306 	}
6307 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6308 		/* Its not time yet */
6309 		return (0);
6310 	}
6311 	if (ctf_progress_timeout_check(tp, true)) {
6312 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6313 		return (-ETIMEDOUT);	/* tcp_drop() */
6314 	}
6315 	/*
6316 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
6317 	 * need to figure out how to force a full MSS segment out.
6318 	 */
6319 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6320 	rack->r_ctl.retran_during_recovery = 0;
6321 	rack->r_ctl.dsack_byte_cnt = 0;
6322 	counter_u64_add(rack_tlp_tot, 1);
6323 	if (rack->r_state && (rack->r_state != tp->t_state))
6324 		rack_set_state(tp, rack);
6325 	so = tp->t_inpcb->inp_socket;
6326 	avail = sbavail(&so->so_snd);
6327 	out = tp->snd_max - tp->snd_una;
6328 	if (out > tp->snd_wnd) {
6329 		/* special case, we need a retransmission */
6330 		collapsed_win = 1;
6331 		goto need_retran;
6332 	}
6333 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
6334 		rack->r_ctl.dsack_persist--;
6335 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6336 			rack->r_ctl.num_dsack = 0;
6337 		}
6338 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6339 	}
6340 	if ((tp->t_flags & TF_GPUTINPROG) &&
6341 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
6342 		/*
6343 		 * If this is the second in a row
6344 		 * TLP and we are doing a measurement
6345 		 * its time to abandon the measurement.
6346 		 * Something is likely broken on
6347 		 * the clients network and measuring a
6348 		 * broken network does us no good.
6349 		 */
6350 		tp->t_flags &= ~TF_GPUTINPROG;
6351 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6352 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
6353 					   tp->gput_seq,
6354 					   0, 0, 18, __LINE__, NULL, 0);
6355 	}
6356 	/*
6357 	 * Check our send oldest always settings, and if
6358 	 * there is an oldest to send jump to the need_retran.
6359 	 */
6360 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6361 		goto need_retran;
6362 
6363 	if (avail > out) {
6364 		/* New data is available */
6365 		amm = avail - out;
6366 		if (amm > ctf_fixed_maxseg(tp)) {
6367 			amm = ctf_fixed_maxseg(tp);
6368 			if ((amm + out) > tp->snd_wnd) {
6369 				/* We are rwnd limited */
6370 				goto need_retran;
6371 			}
6372 		} else if (amm < ctf_fixed_maxseg(tp)) {
6373 			/* not enough to fill a MTU */
6374 			goto need_retran;
6375 		}
6376 		if (IN_FASTRECOVERY(tp->t_flags)) {
6377 			/* Unlikely */
6378 			if (rack->rack_no_prr == 0) {
6379 				if (out + amm <= tp->snd_wnd) {
6380 					rack->r_ctl.rc_prr_sndcnt = amm;
6381 					rack->r_ctl.rc_tlp_new_data = amm;
6382 					rack_log_to_prr(rack, 4, 0);
6383 				}
6384 			} else
6385 				goto need_retran;
6386 		} else {
6387 			/* Set the send-new override */
6388 			if (out + amm <= tp->snd_wnd)
6389 				rack->r_ctl.rc_tlp_new_data = amm;
6390 			else
6391 				goto need_retran;
6392 		}
6393 		rack->r_ctl.rc_tlpsend = NULL;
6394 		counter_u64_add(rack_tlp_newdata, 1);
6395 		goto send;
6396 	}
6397 need_retran:
6398 	/*
6399 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
6400 	 * optionally the first un-acked segment.
6401 	 */
6402 	if (collapsed_win == 0) {
6403 		if (rack_always_send_oldest)
6404 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6405 		else {
6406 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6407 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6408 				rsm = rack_find_high_nonack(rack, rsm);
6409 			}
6410 		}
6411 		if (rsm == NULL) {
6412 			counter_u64_add(rack_tlp_does_nada, 1);
6413 #ifdef TCP_BLACKBOX
6414 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6415 #endif
6416 			goto out;
6417 		}
6418 	} else {
6419 		/*
6420 		 * We must find the last segment
6421 		 * that was acceptable by the client.
6422 		 */
6423 		RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6424 			if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6425 				/* Found one */
6426 				break;
6427 			}
6428 		}
6429 		if (rsm == NULL) {
6430 			/* None? if so send the first */
6431 			rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6432 			if (rsm == NULL) {
6433 				counter_u64_add(rack_tlp_does_nada, 1);
6434 #ifdef TCP_BLACKBOX
6435 				tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6436 #endif
6437 				goto out;
6438 			}
6439 		}
6440 	}
6441 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6442 		/*
6443 		 * We need to split this the last segment in two.
6444 		 */
6445 		struct rack_sendmap *nrsm;
6446 
6447 		nrsm = rack_alloc_full_limit(rack);
6448 		if (nrsm == NULL) {
6449 			/*
6450 			 * No memory to split, we will just exit and punt
6451 			 * off to the RXT timer.
6452 			 */
6453 			counter_u64_add(rack_tlp_does_nada, 1);
6454 			goto out;
6455 		}
6456 		rack_clone_rsm(rack, nrsm, rsm,
6457 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
6458 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6459 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6460 #ifdef INVARIANTS
6461 		if (insret != NULL) {
6462 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6463 			      nrsm, insret, rack, rsm);
6464 		}
6465 #endif
6466 		if (rsm->r_in_tmap) {
6467 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6468 			nrsm->r_in_tmap = 1;
6469 		}
6470 		rsm = nrsm;
6471 	}
6472 	rack->r_ctl.rc_tlpsend = rsm;
6473 send:
6474 	/* Make sure output path knows we are doing a TLP */
6475 	*doing_tlp = 1;
6476 	rack->r_timer_override = 1;
6477 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6478 	return (0);
6479 out:
6480 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6481 	return (0);
6482 }
6483 
6484 /*
6485  * Delayed ack Timer, here we simply need to setup the
6486  * ACK_NOW flag and remove the DELACK flag. From there
6487  * the output routine will send the ack out.
6488  *
6489  * We only return 1, saying don't proceed, if all timers
6490  * are stopped (destroyed PCB?).
6491  */
6492 static int
6493 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6494 {
6495 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6496 		return (1);
6497 	}
6498 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6499 	tp->t_flags &= ~TF_DELACK;
6500 	tp->t_flags |= TF_ACKNOW;
6501 	KMOD_TCPSTAT_INC(tcps_delack);
6502 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6503 	return (0);
6504 }
6505 
6506 /*
6507  * Persists timer, here we simply send the
6508  * same thing as a keepalive will.
6509  * the one byte send.
6510  *
6511  * We only return 1, saying don't proceed, if all timers
6512  * are stopped (destroyed PCB?).
6513  */
6514 static int
6515 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6516 {
6517 	struct tcptemp *t_template;
6518 	struct inpcb *inp;
6519 	int32_t retval = 1;
6520 
6521 	inp = tp->t_inpcb;
6522 
6523 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6524 		return (1);
6525 	}
6526 	if (rack->rc_in_persist == 0)
6527 		return (0);
6528 	if (ctf_progress_timeout_check(tp, false)) {
6529 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6530 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6531 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6532 		return (-ETIMEDOUT);	/* tcp_drop() */
6533 	}
6534 	KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
6535 	/*
6536 	 * Persistence timer into zero window. Force a byte to be output, if
6537 	 * possible.
6538 	 */
6539 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
6540 	/*
6541 	 * Hack: if the peer is dead/unreachable, we do not time out if the
6542 	 * window is closed.  After a full backoff, drop the connection if
6543 	 * the idle time (no responses to probes) reaches the maximum
6544 	 * backoff that we would use if retransmitting.
6545 	 */
6546 	if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6547 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6548 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6549 		KMOD_TCPSTAT_INC(tcps_persistdrop);
6550 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6551 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6552 		retval = -ETIMEDOUT;	/* tcp_drop() */
6553 		goto out;
6554 	}
6555 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6556 	    tp->snd_una == tp->snd_max)
6557 		rack_exit_persist(tp, rack, cts);
6558 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6559 	/*
6560 	 * If the user has closed the socket then drop a persisting
6561 	 * connection after a much reduced timeout.
6562 	 */
6563 	if (tp->t_state > TCPS_CLOSE_WAIT &&
6564 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6565 		KMOD_TCPSTAT_INC(tcps_persistdrop);
6566 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6567 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6568 		retval = -ETIMEDOUT;	/* tcp_drop() */
6569 		goto out;
6570 	}
6571 	t_template = tcpip_maketemplate(rack->rc_inp);
6572 	if (t_template) {
6573 		/* only set it if we were answered */
6574 		if (rack->forced_ack == 0) {
6575 			rack->forced_ack = 1;
6576 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6577 		} else {
6578 			rack->probe_not_answered = 1;
6579 			counter_u64_add(rack_persists_loss, 1);
6580 			rack->r_ctl.persist_lost_ends++;
6581 		}
6582 		counter_u64_add(rack_persists_sends, 1);
6583 		tcp_respond(tp, t_template->tt_ipgen,
6584 			    &t_template->tt_t, (struct mbuf *)NULL,
6585 			    tp->rcv_nxt, tp->snd_una - 1, 0);
6586 		/* This sends an ack */
6587 		if (tp->t_flags & TF_DELACK)
6588 			tp->t_flags &= ~TF_DELACK;
6589 		free(t_template, M_TEMP);
6590 	}
6591 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6592 		tp->t_rxtshift++;
6593 out:
6594 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6595 	rack_start_hpts_timer(rack, tp, cts,
6596 			      0, 0, 0);
6597 	return (retval);
6598 }
6599 
6600 /*
6601  * If a keepalive goes off, we had no other timers
6602  * happening. We always return 1 here since this
6603  * routine either drops the connection or sends
6604  * out a segment with respond.
6605  */
6606 static int
6607 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6608 {
6609 	struct tcptemp *t_template;
6610 	struct inpcb *inp;
6611 
6612 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6613 		return (1);
6614 	}
6615 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6616 	inp = tp->t_inpcb;
6617 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6618 	/*
6619 	 * Keep-alive timer went off; send something or drop connection if
6620 	 * idle for too long.
6621 	 */
6622 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
6623 	if (tp->t_state < TCPS_ESTABLISHED)
6624 		goto dropit;
6625 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6626 	    tp->t_state <= TCPS_CLOSING) {
6627 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6628 			goto dropit;
6629 		/*
6630 		 * Send a packet designed to force a response if the peer is
6631 		 * up and reachable: either an ACK if the connection is
6632 		 * still alive, or an RST if the peer has closed the
6633 		 * connection due to timeout or reboot. Using sequence
6634 		 * number tp->snd_una-1 causes the transmitted zero-length
6635 		 * segment to lie outside the receive window; by the
6636 		 * protocol spec, this requires the correspondent TCP to
6637 		 * respond.
6638 		 */
6639 		KMOD_TCPSTAT_INC(tcps_keepprobe);
6640 		t_template = tcpip_maketemplate(inp);
6641 		if (t_template) {
6642 			if (rack->forced_ack == 0) {
6643 				rack->forced_ack = 1;
6644 				rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6645 			} else {
6646 				rack->probe_not_answered = 1;
6647 			}
6648 			tcp_respond(tp, t_template->tt_ipgen,
6649 			    &t_template->tt_t, (struct mbuf *)NULL,
6650 			    tp->rcv_nxt, tp->snd_una - 1, 0);
6651 			free(t_template, M_TEMP);
6652 		}
6653 	}
6654 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6655 	return (1);
6656 dropit:
6657 	KMOD_TCPSTAT_INC(tcps_keepdrops);
6658 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6659 	return (-ETIMEDOUT);	/* tcp_drop() */
6660 }
6661 
6662 /*
6663  * Retransmit helper function, clear up all the ack
6664  * flags and take care of important book keeping.
6665  */
6666 static void
6667 rack_remxt_tmr(struct tcpcb *tp)
6668 {
6669 	/*
6670 	 * The retransmit timer went off, all sack'd blocks must be
6671 	 * un-acked.
6672 	 */
6673 	struct rack_sendmap *rsm, *trsm = NULL;
6674 	struct tcp_rack *rack;
6675 
6676 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6677 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6678 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6679 	if (rack->r_state && (rack->r_state != tp->t_state))
6680 		rack_set_state(tp, rack);
6681 	/*
6682 	 * Ideally we would like to be able to
6683 	 * mark SACK-PASS on anything not acked here.
6684 	 *
6685 	 * However, if we do that we would burst out
6686 	 * all that data 1ms apart. This would be unwise,
6687 	 * so for now we will just let the normal rxt timer
6688 	 * and tlp timer take care of it.
6689 	 *
6690 	 * Also we really need to stick them back in sequence
6691 	 * order. This way we send in the proper order and any
6692 	 * sacks that come floating in will "re-ack" the data.
6693 	 * To do this we zap the tmap with an INIT and then
6694 	 * walk through and place every rsm in the RB tree
6695 	 * back in its seq ordered place.
6696 	 */
6697 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
6698 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6699 		rsm->r_dupack = 0;
6700 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6701 		/* We must re-add it back to the tlist */
6702 		if (trsm == NULL) {
6703 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6704 		} else {
6705 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6706 		}
6707 		rsm->r_in_tmap = 1;
6708 		trsm = rsm;
6709 		if (rsm->r_flags & RACK_ACKED)
6710 			rsm->r_flags |= RACK_WAS_ACKED;
6711 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS);
6712 		rsm->r_flags |= RACK_MUST_RXT;
6713 	}
6714 	/* Clear the count (we just un-acked them) */
6715 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6716 	rack->r_ctl.rc_sacked = 0;
6717 	rack->r_ctl.rc_sacklast = NULL;
6718 	rack->r_ctl.rc_agg_delayed = 0;
6719 	rack->r_early = 0;
6720 	rack->r_ctl.rc_agg_early = 0;
6721 	rack->r_late = 0;
6722 	/* Clear the tlp rtx mark */
6723 	rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6724 	if (rack->r_ctl.rc_resend != NULL)
6725 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6726 	rack->r_ctl.rc_prr_sndcnt = 0;
6727 	rack_log_to_prr(rack, 6, 0);
6728 	rack->r_timer_override = 1;
6729 	if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6730 #ifdef NETFLIX_EXP_DETECTION
6731 	    || (rack->sack_attack_disable != 0)
6732 #endif
6733 		    ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6734 		/*
6735 		 * For non-sack customers new data
6736 		 * needs to go out as retransmits until
6737 		 * we retransmit up to snd_max.
6738 		 */
6739 		rack->r_must_retran = 1;
6740 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6741 						rack->r_ctl.rc_sacked);
6742 	}
6743 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6744 }
6745 
6746 static void
6747 rack_convert_rtts(struct tcpcb *tp)
6748 {
6749 	if (tp->t_srtt > 1) {
6750 		uint32_t val, frac;
6751 
6752 		val = tp->t_srtt >> TCP_RTT_SHIFT;
6753 		frac = tp->t_srtt & 0x1f;
6754 		tp->t_srtt = TICKS_2_USEC(val);
6755 		/*
6756 		 * frac is the fractional part of the srtt (if any)
6757 		 * but its in ticks and every bit represents
6758 		 * 1/32nd of a hz.
6759 		 */
6760 		if (frac) {
6761 			if (hz == 1000) {
6762 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6763 			} else {
6764 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6765 			}
6766 			tp->t_srtt += frac;
6767 		}
6768 	}
6769 	if (tp->t_rttvar) {
6770 		uint32_t val, frac;
6771 
6772 		val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6773 		frac = tp->t_rttvar & 0x1f;
6774 		tp->t_rttvar = TICKS_2_USEC(val);
6775 		/*
6776 		 * frac is the fractional part of the srtt (if any)
6777 		 * but its in ticks and every bit represents
6778 		 * 1/32nd of a hz.
6779 		 */
6780 		if (frac) {
6781 			if (hz == 1000) {
6782 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6783 			} else {
6784 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6785 			}
6786 			tp->t_rttvar += frac;
6787 		}
6788 	}
6789 	tp->t_rxtcur = RACK_REXMTVAL(tp);
6790 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6791 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6792 	}
6793 	if (tp->t_rxtcur > rack_rto_max) {
6794 		tp->t_rxtcur = rack_rto_max;
6795 	}
6796 }
6797 
6798 static void
6799 rack_cc_conn_init(struct tcpcb *tp)
6800 {
6801 	struct tcp_rack *rack;
6802 	uint32_t srtt;
6803 
6804 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6805 	srtt = tp->t_srtt;
6806 	cc_conn_init(tp);
6807 	/*
6808 	 * Now convert to rack's internal format,
6809 	 * if required.
6810 	 */
6811 	if ((srtt == 0) && (tp->t_srtt != 0))
6812 		rack_convert_rtts(tp);
6813 	/*
6814 	 * We want a chance to stay in slowstart as
6815 	 * we create a connection. TCP spec says that
6816 	 * initially ssthresh is infinite. For our
6817 	 * purposes that is the snd_wnd.
6818 	 */
6819 	if (tp->snd_ssthresh < tp->snd_wnd) {
6820 		tp->snd_ssthresh = tp->snd_wnd;
6821 	}
6822 	/*
6823 	 * We also want to assure a IW worth of
6824 	 * data can get inflight.
6825 	 */
6826 	if (rc_init_window(rack) < tp->snd_cwnd)
6827 		tp->snd_cwnd = rc_init_window(rack);
6828 }
6829 
6830 /*
6831  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6832  * we will setup to retransmit the lowest seq number outstanding.
6833  */
6834 static int
6835 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6836 {
6837 	int32_t rexmt;
6838 	struct inpcb *inp;
6839 	int32_t retval = 0;
6840 	bool isipv6;
6841 
6842 	inp = tp->t_inpcb;
6843 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6844 		return (1);
6845 	}
6846 	if ((tp->t_flags & TF_GPUTINPROG) &&
6847 	    (tp->t_rxtshift)) {
6848 		/*
6849 		 * We have had a second timeout
6850 		 * measurements on successive rxt's are not profitable.
6851 		 * It is unlikely to be of any use (the network is
6852 		 * broken or the client went away).
6853 		 */
6854 		tp->t_flags &= ~TF_GPUTINPROG;
6855 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6856 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
6857 					   tp->gput_seq,
6858 					   0, 0, 18, __LINE__, NULL, 0);
6859 	}
6860 	if (ctf_progress_timeout_check(tp, false)) {
6861 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6862 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6863 		return (-ETIMEDOUT);	/* tcp_drop() */
6864 	}
6865 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6866 	rack->r_ctl.retran_during_recovery = 0;
6867 	rack->r_ctl.dsack_byte_cnt = 0;
6868 	if (IN_FASTRECOVERY(tp->t_flags))
6869 		tp->t_flags |= TF_WASFRECOVERY;
6870 	else
6871 		tp->t_flags &= ~TF_WASFRECOVERY;
6872 	if (IN_CONGRECOVERY(tp->t_flags))
6873 		tp->t_flags |= TF_WASCRECOVERY;
6874 	else
6875 		tp->t_flags &= ~TF_WASCRECOVERY;
6876 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6877 	    (tp->snd_una == tp->snd_max)) {
6878 		/* Nothing outstanding .. nothing to do */
6879 		return (0);
6880 	}
6881 	if (rack->r_ctl.dsack_persist) {
6882 		rack->r_ctl.dsack_persist--;
6883 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6884 			rack->r_ctl.num_dsack = 0;
6885 		}
6886 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6887 	}
6888 	/*
6889 	 * Rack can only run one timer  at a time, so we cannot
6890 	 * run a KEEPINIT (gating SYN sending) and a retransmit
6891 	 * timer for the SYN. So if we are in a front state and
6892 	 * have a KEEPINIT timer we need to check the first transmit
6893 	 * against now to see if we have exceeded the KEEPINIT time
6894 	 * (if one is set).
6895 	 */
6896 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6897 	    (TP_KEEPINIT(tp) != 0)) {
6898 		struct rack_sendmap *rsm;
6899 
6900 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6901 		if (rsm) {
6902 			/* Ok we have something outstanding to test keepinit with */
6903 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6904 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6905 				/* We have exceeded the KEEPINIT time */
6906 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6907 				goto drop_it;
6908 			}
6909 		}
6910 	}
6911 	/*
6912 	 * Retransmission timer went off.  Message has not been acked within
6913 	 * retransmit interval.  Back off to a longer retransmit interval
6914 	 * and retransmit one segment.
6915 	 */
6916 	rack_remxt_tmr(tp);
6917 	if ((rack->r_ctl.rc_resend == NULL) ||
6918 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6919 		/*
6920 		 * If the rwnd collapsed on
6921 		 * the one we are retransmitting
6922 		 * it does not count against the
6923 		 * rxt count.
6924 		 */
6925 		tp->t_rxtshift++;
6926 	}
6927 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6928 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6929 drop_it:
6930 		tp->t_rxtshift = TCP_MAXRXTSHIFT;
6931 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6932 		/* XXXGL: previously t_softerror was casted to uint16_t */
6933 		MPASS(tp->t_softerror >= 0);
6934 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
6935 		goto out;	/* tcp_drop() */
6936 	}
6937 	if (tp->t_state == TCPS_SYN_SENT) {
6938 		/*
6939 		 * If the SYN was retransmitted, indicate CWND to be limited
6940 		 * to 1 segment in cc_conn_init().
6941 		 */
6942 		tp->snd_cwnd = 1;
6943 	} else if (tp->t_rxtshift == 1) {
6944 		/*
6945 		 * first retransmit; record ssthresh and cwnd so they can be
6946 		 * recovered if this turns out to be a "bad" retransmit. A
6947 		 * retransmit is considered "bad" if an ACK for this segment
6948 		 * is received within RTT/2 interval; the assumption here is
6949 		 * that the ACK was already in flight.  See "On Estimating
6950 		 * End-to-End Network Path Properties" by Allman and Paxson
6951 		 * for more details.
6952 		 */
6953 		tp->snd_cwnd_prev = tp->snd_cwnd;
6954 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
6955 		tp->snd_recover_prev = tp->snd_recover;
6956 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6957 		tp->t_flags |= TF_PREVVALID;
6958 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6959 		tp->t_flags &= ~TF_PREVVALID;
6960 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6961 	if ((tp->t_state == TCPS_SYN_SENT) ||
6962 	    (tp->t_state == TCPS_SYN_RECEIVED))
6963 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6964 	else
6965 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6966 
6967 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6968 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6969 	/*
6970 	 * We enter the path for PLMTUD if connection is established or, if
6971 	 * connection is FIN_WAIT_1 status, reason for the last is that if
6972 	 * amount of data we send is very small, we could send it in couple
6973 	 * of packets and process straight to FIN. In that case we won't
6974 	 * catch ESTABLISHED state.
6975 	 */
6976 #ifdef INET6
6977 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
6978 #else
6979 	isipv6 = false;
6980 #endif
6981 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
6982 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6983 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6984 	    ((tp->t_state == TCPS_ESTABLISHED) ||
6985 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
6986 		/*
6987 		 * Idea here is that at each stage of mtu probe (usually,
6988 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
6989 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6990 		 * should take care of that.
6991 		 */
6992 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6993 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6994 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6995 		    tp->t_rxtshift % 2 == 0)) {
6996 			/*
6997 			 * Enter Path MTU Black-hole Detection mechanism: -
6998 			 * Disable Path MTU Discovery (IP "DF" bit). -
6999 			 * Reduce MTU to lower value than what we negotiated
7000 			 * with peer.
7001 			 */
7002 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
7003 				/* Record that we may have found a black hole. */
7004 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
7005 				/* Keep track of previous MSS. */
7006 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
7007 			}
7008 
7009 			/*
7010 			 * Reduce the MSS to blackhole value or to the
7011 			 * default in an attempt to retransmit.
7012 			 */
7013 #ifdef INET6
7014 			if (isipv6 &&
7015 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
7016 				/* Use the sysctl tuneable blackhole MSS. */
7017 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
7018 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7019 			} else if (isipv6) {
7020 				/* Use the default MSS. */
7021 				tp->t_maxseg = V_tcp_v6mssdflt;
7022 				/*
7023 				 * Disable Path MTU Discovery when we switch
7024 				 * to minmss.
7025 				 */
7026 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7027 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7028 			}
7029 #endif
7030 #if defined(INET6) && defined(INET)
7031 			else
7032 #endif
7033 #ifdef INET
7034 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
7035 				/* Use the sysctl tuneable blackhole MSS. */
7036 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
7037 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7038 			} else {
7039 				/* Use the default MSS. */
7040 				tp->t_maxseg = V_tcp_mssdflt;
7041 				/*
7042 				 * Disable Path MTU Discovery when we switch
7043 				 * to minmss.
7044 				 */
7045 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7046 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7047 			}
7048 #endif
7049 		} else {
7050 			/*
7051 			 * If further retransmissions are still unsuccessful
7052 			 * with a lowered MTU, maybe this isn't a blackhole
7053 			 * and we restore the previous MSS and blackhole
7054 			 * detection flags. The limit '6' is determined by
7055 			 * giving each probe stage (1448, 1188, 524) 2
7056 			 * chances to recover.
7057 			 */
7058 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
7059 			    (tp->t_rxtshift >= 6)) {
7060 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
7061 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
7062 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
7063 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
7064 			}
7065 		}
7066 	}
7067 	/*
7068 	 * Disable RFC1323 and SACK if we haven't got any response to
7069 	 * our third SYN to work-around some broken terminal servers
7070 	 * (most of which have hopefully been retired) that have bad VJ
7071 	 * header compression code which trashes TCP segments containing
7072 	 * unknown-to-them TCP options.
7073 	 */
7074 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
7075 	    (tp->t_rxtshift == 3))
7076 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
7077 	/*
7078 	 * If we backed off this far, our srtt estimate is probably bogus.
7079 	 * Clobber it so we'll take the next rtt measurement as our srtt;
7080 	 * move the current srtt into rttvar to keep the current retransmit
7081 	 * times until then.
7082 	 */
7083 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
7084 #ifdef INET6
7085 		if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
7086 			in6_losing(tp->t_inpcb);
7087 		else
7088 #endif
7089 			in_losing(tp->t_inpcb);
7090 		tp->t_rttvar += tp->t_srtt;
7091 		tp->t_srtt = 0;
7092 	}
7093 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
7094 	tp->snd_recover = tp->snd_max;
7095 	tp->t_flags |= TF_ACKNOW;
7096 	tp->t_rtttime = 0;
7097 	rack_cong_signal(tp, CC_RTO, tp->snd_una);
7098 out:
7099 	return (retval);
7100 }
7101 
7102 static int
7103 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
7104 {
7105 	int32_t ret = 0;
7106 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
7107 
7108 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
7109 	    (tp->t_flags & TF_GPUTINPROG)) {
7110 		/*
7111 		 * We have a goodput in progress
7112 		 * and we have entered a late state.
7113 		 * Do we have enough data in the sb
7114 		 * to handle the GPUT request?
7115 		 */
7116 		uint32_t bytes;
7117 
7118 		bytes = tp->gput_ack - tp->gput_seq;
7119 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
7120 			bytes += tp->gput_seq - tp->snd_una;
7121 		if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
7122 			/*
7123 			 * There are not enough bytes in the socket
7124 			 * buffer that have been sent to cover this
7125 			 * measurement. Cancel it.
7126 			 */
7127 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7128 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
7129 						   tp->gput_seq,
7130 						   0, 0, 18, __LINE__, NULL, 0);
7131 			tp->t_flags &= ~TF_GPUTINPROG;
7132 		}
7133 	}
7134 	if (timers == 0) {
7135 		return (0);
7136 	}
7137 	if (tp->t_state == TCPS_LISTEN) {
7138 		/* no timers on listen sockets */
7139 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
7140 			return (0);
7141 		return (1);
7142 	}
7143 	if ((timers & PACE_TMR_RACK) &&
7144 	    rack->rc_on_min_to) {
7145 		/*
7146 		 * For the rack timer when we
7147 		 * are on a min-timeout (which means rrr_conf = 3)
7148 		 * we don't want to check the timer. It may
7149 		 * be going off for a pace and thats ok we
7150 		 * want to send the retransmit (if its ready).
7151 		 *
7152 		 * If its on a normal rack timer (non-min) then
7153 		 * we will check if its expired.
7154 		 */
7155 		goto skip_time_check;
7156 	}
7157 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7158 		uint32_t left;
7159 
7160 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
7161 			ret = -1;
7162 			rack_log_to_processing(rack, cts, ret, 0);
7163 			return (0);
7164 		}
7165 		if (hpts_calling == 0) {
7166 			/*
7167 			 * A user send or queued mbuf (sack) has called us? We
7168 			 * return 0 and let the pacing guards
7169 			 * deal with it if they should or
7170 			 * should not cause a send.
7171 			 */
7172 			ret = -2;
7173 			rack_log_to_processing(rack, cts, ret, 0);
7174 			return (0);
7175 		}
7176 		/*
7177 		 * Ok our timer went off early and we are not paced false
7178 		 * alarm, go back to sleep.
7179 		 */
7180 		ret = -3;
7181 		left = rack->r_ctl.rc_timer_exp - cts;
7182 		tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
7183 		rack_log_to_processing(rack, cts, ret, left);
7184 		return (1);
7185 	}
7186 skip_time_check:
7187 	rack->rc_tmr_stopped = 0;
7188 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
7189 	if (timers & PACE_TMR_DELACK) {
7190 		ret = rack_timeout_delack(tp, rack, cts);
7191 	} else if (timers & PACE_TMR_RACK) {
7192 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
7193 		rack->r_fast_output = 0;
7194 		ret = rack_timeout_rack(tp, rack, cts);
7195 	} else if (timers & PACE_TMR_TLP) {
7196 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
7197 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
7198 	} else if (timers & PACE_TMR_RXT) {
7199 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
7200 		rack->r_fast_output = 0;
7201 		ret = rack_timeout_rxt(tp, rack, cts);
7202 	} else if (timers & PACE_TMR_PERSIT) {
7203 		ret = rack_timeout_persist(tp, rack, cts);
7204 	} else if (timers & PACE_TMR_KEEP) {
7205 		ret = rack_timeout_keepalive(tp, rack, cts);
7206 	}
7207 	rack_log_to_processing(rack, cts, ret, timers);
7208 	return (ret);
7209 }
7210 
7211 static void
7212 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
7213 {
7214 	struct timeval tv;
7215 	uint32_t us_cts, flags_on_entry;
7216 	uint8_t hpts_removed = 0;
7217 
7218 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
7219 	us_cts = tcp_get_usecs(&tv);
7220 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
7221 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
7222 	     ((tp->snd_max - tp->snd_una) == 0))) {
7223 		tcp_hpts_remove(rack->rc_inp);
7224 		hpts_removed = 1;
7225 		/* If we were not delayed cancel out the flag. */
7226 		if ((tp->snd_max - tp->snd_una) == 0)
7227 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7228 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7229 	}
7230 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
7231 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
7232 		if (tcp_in_hpts(rack->rc_inp) &&
7233 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
7234 			/*
7235 			 * Canceling timer's when we have no output being
7236 			 * paced. We also must remove ourselves from the
7237 			 * hpts.
7238 			 */
7239 			tcp_hpts_remove(rack->rc_inp);
7240 			hpts_removed = 1;
7241 		}
7242 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7243 	}
7244 	if (hpts_removed == 0)
7245 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7246 }
7247 
7248 static void
7249 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
7250 {
7251 	return;
7252 }
7253 
7254 static int
7255 rack_stopall(struct tcpcb *tp)
7256 {
7257 	struct tcp_rack *rack;
7258 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7259 	rack->t_timers_stopped = 1;
7260 	return (0);
7261 }
7262 
7263 static void
7264 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
7265 {
7266 	return;
7267 }
7268 
7269 static int
7270 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
7271 {
7272 	return (0);
7273 }
7274 
7275 static void
7276 rack_stop_all_timers(struct tcpcb *tp)
7277 {
7278 	struct tcp_rack *rack;
7279 
7280 	/*
7281 	 * Assure no timers are running.
7282 	 */
7283 	if (tcp_timer_active(tp, TT_PERSIST)) {
7284 		/* We enter in persists, set the flag appropriately */
7285 		rack = (struct tcp_rack *)tp->t_fb_ptr;
7286 		rack->rc_in_persist = 1;
7287 	}
7288 	tcp_timer_suspend(tp, TT_PERSIST);
7289 	tcp_timer_suspend(tp, TT_REXMT);
7290 	tcp_timer_suspend(tp, TT_KEEP);
7291 	tcp_timer_suspend(tp, TT_DELACK);
7292 }
7293 
7294 static void
7295 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7296     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7297 {
7298 	int32_t idx;
7299 
7300 	rsm->r_rtr_cnt++;
7301 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7302 	rsm->r_dupack = 0;
7303 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7304 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7305 		rsm->r_flags |= RACK_OVERMAX;
7306 	}
7307 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7308 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7309 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7310 	}
7311 	idx = rsm->r_rtr_cnt - 1;
7312 	rsm->r_tim_lastsent[idx] = ts;
7313 	/*
7314 	 * Here we don't add in the len of send, since its already
7315 	 * in snduna <->snd_max.
7316 	 */
7317 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
7318 				     rack->r_ctl.rc_sacked);
7319 	if (rsm->r_flags & RACK_ACKED) {
7320 		/* Problably MTU discovery messing with us */
7321 		rsm->r_flags &= ~RACK_ACKED;
7322 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7323 	}
7324 	if (rsm->r_in_tmap) {
7325 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7326 		rsm->r_in_tmap = 0;
7327 	}
7328 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7329 	rsm->r_in_tmap = 1;
7330 	if (rsm->r_flags & RACK_SACK_PASSED) {
7331 		/* We have retransmitted due to the SACK pass */
7332 		rsm->r_flags &= ~RACK_SACK_PASSED;
7333 		rsm->r_flags |= RACK_WAS_SACKPASS;
7334 	}
7335 }
7336 
7337 static uint32_t
7338 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7339     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7340 {
7341 	/*
7342 	 * We (re-)transmitted starting at rsm->r_start for some length
7343 	 * (possibly less than r_end.
7344 	 */
7345 	struct rack_sendmap *nrsm, *insret;
7346 	uint32_t c_end;
7347 	int32_t len;
7348 
7349 	len = *lenp;
7350 	c_end = rsm->r_start + len;
7351 	if (SEQ_GEQ(c_end, rsm->r_end)) {
7352 		/*
7353 		 * We retransmitted the whole piece or more than the whole
7354 		 * slopping into the next rsm.
7355 		 */
7356 		rack_update_rsm(tp, rack, rsm, ts, add_flag);
7357 		if (c_end == rsm->r_end) {
7358 			*lenp = 0;
7359 			return (0);
7360 		} else {
7361 			int32_t act_len;
7362 
7363 			/* Hangs over the end return whats left */
7364 			act_len = rsm->r_end - rsm->r_start;
7365 			*lenp = (len - act_len);
7366 			return (rsm->r_end);
7367 		}
7368 		/* We don't get out of this block. */
7369 	}
7370 	/*
7371 	 * Here we retransmitted less than the whole thing which means we
7372 	 * have to split this into what was transmitted and what was not.
7373 	 */
7374 	nrsm = rack_alloc_full_limit(rack);
7375 	if (nrsm == NULL) {
7376 		/*
7377 		 * We can't get memory, so lets not proceed.
7378 		 */
7379 		*lenp = 0;
7380 		return (0);
7381 	}
7382 	/*
7383 	 * So here we are going to take the original rsm and make it what we
7384 	 * retransmitted. nrsm will be the tail portion we did not
7385 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7386 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7387 	 * 1, 6 and the new piece will be 6, 11.
7388 	 */
7389 	rack_clone_rsm(rack, nrsm, rsm, c_end);
7390 	nrsm->r_dupack = 0;
7391 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7392 	insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7393 #ifdef INVARIANTS
7394 	if (insret != NULL) {
7395 		panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7396 		      nrsm, insret, rack, rsm);
7397 	}
7398 #endif
7399 	if (rsm->r_in_tmap) {
7400 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7401 		nrsm->r_in_tmap = 1;
7402 	}
7403 	rsm->r_flags &= (~RACK_HAS_FIN);
7404 	rack_update_rsm(tp, rack, rsm, ts, add_flag);
7405 	/* Log a split of rsm into rsm and nrsm */
7406 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7407 	*lenp = 0;
7408 	return (0);
7409 }
7410 
7411 static void
7412 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7413 		uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t cts,
7414 		struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls)
7415 {
7416 	struct tcp_rack *rack;
7417 	struct rack_sendmap *rsm, *nrsm, *insret, fe;
7418 	register uint32_t snd_max, snd_una;
7419 
7420 	/*
7421 	 * Add to the RACK log of packets in flight or retransmitted. If
7422 	 * there is a TS option we will use the TS echoed, if not we will
7423 	 * grab a TS.
7424 	 *
7425 	 * Retransmissions will increment the count and move the ts to its
7426 	 * proper place. Note that if options do not include TS's then we
7427 	 * won't be able to effectively use the ACK for an RTT on a retran.
7428 	 *
7429 	 * Notes about r_start and r_end. Lets consider a send starting at
7430 	 * sequence 1 for 10 bytes. In such an example the r_start would be
7431 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7432 	 * This means that r_end is actually the first sequence for the next
7433 	 * slot (11).
7434 	 *
7435 	 */
7436 	/*
7437 	 * If err is set what do we do XXXrrs? should we not add the thing?
7438 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
7439 	 * i.e. proceed with add ** do this for now.
7440 	 */
7441 	INP_WLOCK_ASSERT(tp->t_inpcb);
7442 	if (err)
7443 		/*
7444 		 * We don't log errors -- we could but snd_max does not
7445 		 * advance in this case either.
7446 		 */
7447 		return;
7448 
7449 	if (th_flags & TH_RST) {
7450 		/*
7451 		 * We don't log resets and we return immediately from
7452 		 * sending
7453 		 */
7454 		return;
7455 	}
7456 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7457 	snd_una = tp->snd_una;
7458 	snd_max = tp->snd_max;
7459 	if (th_flags & (TH_SYN | TH_FIN)) {
7460 		/*
7461 		 * The call to rack_log_output is made before bumping
7462 		 * snd_max. This means we can record one extra byte on a SYN
7463 		 * or FIN if seq_out is adding more on and a FIN is present
7464 		 * (and we are not resending).
7465 		 */
7466 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7467 			len++;
7468 		if (th_flags & TH_FIN)
7469 			len++;
7470 		if (SEQ_LT(snd_max, tp->snd_nxt)) {
7471 			/*
7472 			 * The add/update as not been done for the FIN/SYN
7473 			 * yet.
7474 			 */
7475 			snd_max = tp->snd_nxt;
7476 		}
7477 	}
7478 	if (SEQ_LEQ((seq_out + len), snd_una)) {
7479 		/* Are sending an old segment to induce an ack (keep-alive)? */
7480 		return;
7481 	}
7482 	if (SEQ_LT(seq_out, snd_una)) {
7483 		/* huh? should we panic? */
7484 		uint32_t end;
7485 
7486 		end = seq_out + len;
7487 		seq_out = snd_una;
7488 		if (SEQ_GEQ(end, seq_out))
7489 			len = end - seq_out;
7490 		else
7491 			len = 0;
7492 	}
7493 	if (len == 0) {
7494 		/* We don't log zero window probes */
7495 		return;
7496 	}
7497 	rack->r_ctl.rc_time_last_sent = cts;
7498 	if (IN_FASTRECOVERY(tp->t_flags)) {
7499 		rack->r_ctl.rc_prr_out += len;
7500 	}
7501 	/* First question is it a retransmission or new? */
7502 	if (seq_out == snd_max) {
7503 		/* Its new */
7504 again:
7505 		rsm = rack_alloc(rack);
7506 		if (rsm == NULL) {
7507 			/*
7508 			 * Hmm out of memory and the tcb got destroyed while
7509 			 * we tried to wait.
7510 			 */
7511 			return;
7512 		}
7513 		if (th_flags & TH_FIN) {
7514 			rsm->r_flags = RACK_HAS_FIN|add_flag;
7515 		} else {
7516 			rsm->r_flags = add_flag;
7517 		}
7518 		if (hw_tls)
7519 			rsm->r_hw_tls = 1;
7520 		rsm->r_tim_lastsent[0] = cts;
7521 		rsm->r_rtr_cnt = 1;
7522 		rsm->r_rtr_bytes = 0;
7523 		if (th_flags & TH_SYN) {
7524 			/* The data space is one beyond snd_una */
7525 			rsm->r_flags |= RACK_HAS_SYN;
7526 		}
7527 		rsm->r_start = seq_out;
7528 		rsm->r_end = rsm->r_start + len;
7529 		rsm->r_dupack = 0;
7530 		/*
7531 		 * save off the mbuf location that
7532 		 * sndmbuf_noadv returned (which is
7533 		 * where we started copying from)..
7534 		 */
7535 		rsm->m = s_mb;
7536 		rsm->soff = s_moff;
7537 		/*
7538 		 * Here we do add in the len of send, since its not yet
7539 		 * reflected in in snduna <->snd_max
7540 		 */
7541 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
7542 					      rack->r_ctl.rc_sacked) +
7543 			      (rsm->r_end - rsm->r_start));
7544 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7545 		if (rsm->m) {
7546 			if (rsm->m->m_len <= rsm->soff) {
7547 				/*
7548 				 * XXXrrs Question, will this happen?
7549 				 *
7550 				 * If sbsndptr is set at the correct place
7551 				 * then s_moff should always be somewhere
7552 				 * within rsm->m. But if the sbsndptr was
7553 				 * off then that won't be true. If it occurs
7554 				 * we need to walkout to the correct location.
7555 				 */
7556 				struct mbuf *lm;
7557 
7558 				lm = rsm->m;
7559 				while (lm->m_len <= rsm->soff) {
7560 					rsm->soff -= lm->m_len;
7561 					lm = lm->m_next;
7562 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7563 							     __func__, rack, s_moff, s_mb, rsm->soff));
7564 				}
7565 				rsm->m = lm;
7566 				counter_u64_add(rack_sbsndptr_wrong, 1);
7567 			} else
7568 				counter_u64_add(rack_sbsndptr_right, 1);
7569 			rsm->orig_m_len = rsm->m->m_len;
7570 		} else
7571 			rsm->orig_m_len = 0;
7572 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7573 		/* Log a new rsm */
7574 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7575 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7576 #ifdef INVARIANTS
7577 		if (insret != NULL) {
7578 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7579 			      nrsm, insret, rack, rsm);
7580 		}
7581 #endif
7582 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7583 		rsm->r_in_tmap = 1;
7584 		/*
7585 		 * Special case detection, is there just a single
7586 		 * packet outstanding when we are not in recovery?
7587 		 *
7588 		 * If this is true mark it so.
7589 		 */
7590 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7591 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7592 			struct rack_sendmap *prsm;
7593 
7594 			prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7595 			if (prsm)
7596 				prsm->r_one_out_nr = 1;
7597 		}
7598 		return;
7599 	}
7600 	/*
7601 	 * If we reach here its a retransmission and we need to find it.
7602 	 */
7603 	memset(&fe, 0, sizeof(fe));
7604 more:
7605 	if (hintrsm && (hintrsm->r_start == seq_out)) {
7606 		rsm = hintrsm;
7607 		hintrsm = NULL;
7608 	} else {
7609 		/* No hints sorry */
7610 		rsm = NULL;
7611 	}
7612 	if ((rsm) && (rsm->r_start == seq_out)) {
7613 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7614 		if (len == 0) {
7615 			return;
7616 		} else {
7617 			goto more;
7618 		}
7619 	}
7620 	/* Ok it was not the last pointer go through it the hard way. */
7621 refind:
7622 	fe.r_start = seq_out;
7623 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7624 	if (rsm) {
7625 		if (rsm->r_start == seq_out) {
7626 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7627 			if (len == 0) {
7628 				return;
7629 			} else {
7630 				goto refind;
7631 			}
7632 		}
7633 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7634 			/* Transmitted within this piece */
7635 			/*
7636 			 * Ok we must split off the front and then let the
7637 			 * update do the rest
7638 			 */
7639 			nrsm = rack_alloc_full_limit(rack);
7640 			if (nrsm == NULL) {
7641 				rack_update_rsm(tp, rack, rsm, cts, add_flag);
7642 				return;
7643 			}
7644 			/*
7645 			 * copy rsm to nrsm and then trim the front of rsm
7646 			 * to not include this part.
7647 			 */
7648 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
7649 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7650 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7651 #ifdef INVARIANTS
7652 			if (insret != NULL) {
7653 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7654 				      nrsm, insret, rack, rsm);
7655 			}
7656 #endif
7657 			if (rsm->r_in_tmap) {
7658 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7659 				nrsm->r_in_tmap = 1;
7660 			}
7661 			rsm->r_flags &= (~RACK_HAS_FIN);
7662 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7663 			if (len == 0) {
7664 				return;
7665 			} else if (len > 0)
7666 				goto refind;
7667 		}
7668 	}
7669 	/*
7670 	 * Hmm not found in map did they retransmit both old and on into the
7671 	 * new?
7672 	 */
7673 	if (seq_out == tp->snd_max) {
7674 		goto again;
7675 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
7676 #ifdef INVARIANTS
7677 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7678 		       seq_out, len, tp->snd_una, tp->snd_max);
7679 		printf("Starting Dump of all rack entries\n");
7680 		RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7681 			printf("rsm:%p start:%u end:%u\n",
7682 			       rsm, rsm->r_start, rsm->r_end);
7683 		}
7684 		printf("Dump complete\n");
7685 		panic("seq_out not found rack:%p tp:%p",
7686 		      rack, tp);
7687 #endif
7688 	} else {
7689 #ifdef INVARIANTS
7690 		/*
7691 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
7692 		 * flag)
7693 		 */
7694 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7695 		      seq_out, len, tp->snd_max, tp);
7696 #endif
7697 	}
7698 }
7699 
7700 /*
7701  * Record one of the RTT updates from an ack into
7702  * our sample structure.
7703  */
7704 
7705 static void
7706 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7707 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7708 {
7709 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7710 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7711 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7712 	}
7713 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7714 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7715 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7716 	}
7717 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7718 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7719 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
7720 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7721 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7722 	}
7723 	if ((confidence == 1) &&
7724 	    ((rsm == NULL) ||
7725 	     (rsm->r_just_ret) ||
7726 	     (rsm->r_one_out_nr &&
7727 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7728 		/*
7729 		 * If the rsm had a just return
7730 		 * hit it then we can't trust the
7731 		 * rtt measurement for buffer deterimination
7732 		 * Note that a confidence of 2, indicates
7733 		 * SACK'd which overrides the r_just_ret or
7734 		 * the r_one_out_nr. If it was a CUM-ACK and
7735 		 * we had only two outstanding, but get an
7736 		 * ack for only 1. Then that also lowers our
7737 		 * confidence.
7738 		 */
7739 		confidence = 0;
7740 	}
7741 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7742 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7743 		if (rack->r_ctl.rack_rs.confidence == 0) {
7744 			/*
7745 			 * We take anything with no current confidence
7746 			 * saved.
7747 			 */
7748 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7749 			rack->r_ctl.rack_rs.confidence = confidence;
7750 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7751 		} else if (confidence || rack->r_ctl.rack_rs.confidence) {
7752 			/*
7753 			 * Once we have a confident number,
7754 			 * we can update it with a smaller
7755 			 * value since this confident number
7756 			 * may include the DSACK time until
7757 			 * the next segment (the second one) arrived.
7758 			 */
7759 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7760 			rack->r_ctl.rack_rs.confidence = confidence;
7761 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7762 		}
7763 	}
7764 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7765 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7766 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7767 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
7768 }
7769 
7770 /*
7771  * Collect new round-trip time estimate
7772  * and update averages and current timeout.
7773  */
7774 static void
7775 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7776 {
7777 	int32_t delta;
7778 	uint32_t o_srtt, o_var;
7779 	int32_t hrtt_up = 0;
7780 	int32_t rtt;
7781 
7782 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7783 		/* No valid sample */
7784 		return;
7785 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7786 		/* We are to use the lowest RTT seen in a single ack */
7787 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7788 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7789 		/* We are to use the highest RTT seen in a single ack */
7790 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7791 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7792 		/* We are to use the average RTT seen in a single ack */
7793 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7794 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7795 	} else {
7796 #ifdef INVARIANTS
7797 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7798 #endif
7799 		return;
7800 	}
7801 	if (rtt == 0)
7802 		rtt = 1;
7803 	if (rack->rc_gp_rtt_set == 0) {
7804 		/*
7805 		 * With no RTT we have to accept
7806 		 * even one we are not confident of.
7807 		 */
7808 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7809 		rack->rc_gp_rtt_set = 1;
7810 	} else if (rack->r_ctl.rack_rs.confidence) {
7811 		/* update the running gp srtt */
7812 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7813 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7814 	}
7815 	if (rack->r_ctl.rack_rs.confidence) {
7816 		/*
7817 		 * record the low and high for highly buffered path computation,
7818 		 * we only do this if we are confident (not a retransmission).
7819 		 */
7820 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7821 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7822 			hrtt_up = 1;
7823 		}
7824 		if (rack->rc_highly_buffered == 0) {
7825 			/*
7826 			 * Currently once we declare a path has
7827 			 * highly buffered there is no going
7828 			 * back, which may be a problem...
7829 			 */
7830 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7831 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7832 						     rack->r_ctl.rc_highest_us_rtt,
7833 						     rack->r_ctl.rc_lowest_us_rtt,
7834 						     RACK_RTTS_SEEHBP);
7835 				rack->rc_highly_buffered = 1;
7836 			}
7837 		}
7838 	}
7839 	if ((rack->r_ctl.rack_rs.confidence) ||
7840 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7841 		/*
7842 		 * If we are highly confident of it <or> it was
7843 		 * never retransmitted we accept it as the last us_rtt.
7844 		 */
7845 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7846 		/* The lowest rtt can be set if its was not retransmited */
7847 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7848 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7849 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
7850 				rack->r_ctl.rc_lowest_us_rtt = 1;
7851 		}
7852 	}
7853 	o_srtt = tp->t_srtt;
7854 	o_var = tp->t_rttvar;
7855 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7856 	if (tp->t_srtt != 0) {
7857 		/*
7858 		 * We keep a simple srtt in microseconds, like our rtt
7859 		 * measurement. We don't need to do any tricks with shifting
7860 		 * etc. Instead we just add in 1/8th of the new measurement
7861 		 * and subtract out 1/8 of the old srtt. We do the same with
7862 		 * the variance after finding the absolute value of the
7863 		 * difference between this sample and the current srtt.
7864 		 */
7865 		delta = tp->t_srtt - rtt;
7866 		/* Take off 1/8th of the current sRTT */
7867 		tp->t_srtt -= (tp->t_srtt >> 3);
7868 		/* Add in 1/8th of the new RTT just measured */
7869 		tp->t_srtt += (rtt >> 3);
7870 		if (tp->t_srtt <= 0)
7871 			tp->t_srtt = 1;
7872 		/* Now lets make the absolute value of the variance */
7873 		if (delta < 0)
7874 			delta = -delta;
7875 		/* Subtract out 1/8th */
7876 		tp->t_rttvar -= (tp->t_rttvar >> 3);
7877 		/* Add in 1/8th of the new variance we just saw */
7878 		tp->t_rttvar += (delta >> 3);
7879 		if (tp->t_rttvar <= 0)
7880 			tp->t_rttvar = 1;
7881 		if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
7882 			tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7883 	} else {
7884 		/*
7885 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
7886 		 * variance to half the rtt (so our first retransmit happens
7887 		 * at 3*rtt).
7888 		 */
7889 		tp->t_srtt = rtt;
7890 		tp->t_rttvar = rtt >> 1;
7891 		tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7892 	}
7893 	rack->rc_srtt_measure_made = 1;
7894 	KMOD_TCPSTAT_INC(tcps_rttupdated);
7895 	tp->t_rttupdated++;
7896 #ifdef STATS
7897 	if (rack_stats_gets_ms_rtt == 0) {
7898 		/* Send in the microsecond rtt used for rxt timeout purposes */
7899 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7900 	} else if (rack_stats_gets_ms_rtt == 1) {
7901 		/* Send in the millisecond rtt used for rxt timeout purposes */
7902 		int32_t ms_rtt;
7903 
7904 		/* Round up */
7905 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7906 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7907 	} else if (rack_stats_gets_ms_rtt == 2) {
7908 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
7909 		int32_t ms_rtt;
7910 
7911 		/* Round up */
7912 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7913 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7914 	}  else {
7915 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
7916 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7917 	}
7918 
7919 #endif
7920 	/*
7921 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
7922 	 * way we do the smoothing, srtt and rttvar will each average +1/2
7923 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
7924 	 * tick of rounding and 1 extra tick because of +-1/2 tick
7925 	 * uncertainty in the firing of the timer.  The bias will give us
7926 	 * exactly the 1.5 tick we need.  But, because the bias is
7927 	 * statistical, we have to test that we don't drop below the minimum
7928 	 * feasible timer (which is 2 ticks).
7929 	 */
7930 	tp->t_rxtshift = 0;
7931 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7932 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7933 	rack_log_rtt_sample(rack, rtt);
7934 	tp->t_softerror = 0;
7935 }
7936 
7937 
7938 static void
7939 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7940 {
7941 	/*
7942 	 * Apply to filter the inbound us-rtt at us_cts.
7943 	 */
7944 	uint32_t old_rtt;
7945 
7946 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7947 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7948 			       us_rtt, us_cts);
7949 	if (rack->r_ctl.last_pacing_time &&
7950 	    rack->rc_gp_dyn_mul &&
7951 	    (rack->r_ctl.last_pacing_time > us_rtt))
7952 		rack->pacing_longer_than_rtt = 1;
7953 	else
7954 		rack->pacing_longer_than_rtt = 0;
7955 	if (old_rtt > us_rtt) {
7956 		/* We just hit a new lower rtt time */
7957 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7958 				     __LINE__, RACK_RTTS_NEWRTT);
7959 		/*
7960 		 * Only count it if its lower than what we saw within our
7961 		 * calculated range.
7962 		 */
7963 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7964 			if (rack_probertt_lower_within &&
7965 			    rack->rc_gp_dyn_mul &&
7966 			    (rack->use_fixed_rate == 0) &&
7967 			    (rack->rc_always_pace)) {
7968 				/*
7969 				 * We are seeing a new lower rtt very close
7970 				 * to the time that we would have entered probe-rtt.
7971 				 * This is probably due to the fact that a peer flow
7972 				 * has entered probe-rtt. Lets go in now too.
7973 				 */
7974 				uint32_t val;
7975 
7976 				val = rack_probertt_lower_within * rack_time_between_probertt;
7977 				val /= 100;
7978 				if ((rack->in_probe_rtt == 0)  &&
7979 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
7980 					rack_enter_probertt(rack, us_cts);
7981 				}
7982 			}
7983 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7984 		}
7985 	}
7986 }
7987 
7988 static int
7989 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7990     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7991 {
7992 	uint32_t us_rtt;
7993 	int32_t i, all;
7994 	uint32_t t, len_acked;
7995 
7996 	if ((rsm->r_flags & RACK_ACKED) ||
7997 	    (rsm->r_flags & RACK_WAS_ACKED))
7998 		/* Already done */
7999 		return (0);
8000 	if (rsm->r_no_rtt_allowed) {
8001 		/* Not allowed */
8002 		return (0);
8003 	}
8004 	if (ack_type == CUM_ACKED) {
8005 		if (SEQ_GT(th_ack, rsm->r_end)) {
8006 			len_acked = rsm->r_end - rsm->r_start;
8007 			all = 1;
8008 		} else {
8009 			len_acked = th_ack - rsm->r_start;
8010 			all = 0;
8011 		}
8012 	} else {
8013 		len_acked = rsm->r_end - rsm->r_start;
8014 		all = 0;
8015 	}
8016 	if (rsm->r_rtr_cnt == 1) {
8017 
8018 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8019 		if ((int)t <= 0)
8020 			t = 1;
8021 		if (!tp->t_rttlow || tp->t_rttlow > t)
8022 			tp->t_rttlow = t;
8023 		if (!rack->r_ctl.rc_rack_min_rtt ||
8024 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8025 			rack->r_ctl.rc_rack_min_rtt = t;
8026 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
8027 				rack->r_ctl.rc_rack_min_rtt = 1;
8028 			}
8029 		}
8030 		if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
8031 			us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8032 		else
8033 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8034 		if (us_rtt == 0)
8035 			us_rtt = 1;
8036 		if (CC_ALGO(tp)->rttsample != NULL) {
8037 			/* Kick the RTT to the CC */
8038 			CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas);
8039 		}
8040 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
8041 		if (ack_type == SACKED) {
8042 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
8043 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
8044 		} else {
8045 			/*
8046 			 * We need to setup what our confidence
8047 			 * is in this ack.
8048 			 *
8049 			 * If the rsm was app limited and it is
8050 			 * less than a mss in length (the end
8051 			 * of the send) then we have a gap. If we
8052 			 * were app limited but say we were sending
8053 			 * multiple MSS's then we are more confident
8054 			 * int it.
8055 			 *
8056 			 * When we are not app-limited then we see if
8057 			 * the rsm is being included in the current
8058 			 * measurement, we tell this by the app_limited_needs_set
8059 			 * flag.
8060 			 *
8061 			 * Note that being cwnd blocked is not applimited
8062 			 * as well as the pacing delay between packets which
8063 			 * are sending only 1 or 2 MSS's also will show up
8064 			 * in the RTT. We probably need to examine this algorithm
8065 			 * a bit more and enhance it to account for the delay
8066 			 * between rsm's. We could do that by saving off the
8067 			 * pacing delay of each rsm (in an rsm) and then
8068 			 * factoring that in somehow though for now I am
8069 			 * not sure how :)
8070 			 */
8071 			int calc_conf = 0;
8072 
8073 			if (rsm->r_flags & RACK_APP_LIMITED) {
8074 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
8075 					calc_conf = 0;
8076 				else
8077 					calc_conf = 1;
8078 			} else if (rack->app_limited_needs_set == 0) {
8079 				calc_conf = 1;
8080 			} else {
8081 				calc_conf = 0;
8082 			}
8083 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
8084 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
8085 					    calc_conf, rsm, rsm->r_rtr_cnt);
8086 		}
8087 		if ((rsm->r_flags & RACK_TLP) &&
8088 		    (!IN_FASTRECOVERY(tp->t_flags))) {
8089 			/* Segment was a TLP and our retrans matched */
8090 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
8091 				rack->r_ctl.rc_rsm_start = tp->snd_max;
8092 				rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
8093 				rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
8094 				rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
8095 			}
8096 		}
8097 		if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
8098 			/* New more recent rack_tmit_time */
8099 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8100 			rack->rc_rack_rtt = t;
8101 		}
8102 		return (1);
8103 	}
8104 	/*
8105 	 * We clear the soft/rxtshift since we got an ack.
8106 	 * There is no assurance we will call the commit() function
8107 	 * so we need to clear these to avoid incorrect handling.
8108 	 */
8109 	tp->t_rxtshift = 0;
8110 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8111 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
8112 	tp->t_softerror = 0;
8113 	if (to && (to->to_flags & TOF_TS) &&
8114 	    (ack_type == CUM_ACKED) &&
8115 	    (to->to_tsecr) &&
8116 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
8117 		/*
8118 		 * Now which timestamp does it match? In this block the ACK
8119 		 * must be coming from a previous transmission.
8120 		 */
8121 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
8122 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
8123 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
8124 				if ((int)t <= 0)
8125 					t = 1;
8126 				if (CC_ALGO(tp)->rttsample != NULL) {
8127 					/*
8128 					 * Kick the RTT to the CC, here
8129 					 * we lie a bit in that we know the
8130 					 * retransmission is correct even though
8131 					 * we retransmitted. This is because
8132 					 * we match the timestamps.
8133 					 */
8134 					if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
8135 						us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
8136 					else
8137 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
8138 					CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas);
8139 				}
8140 				if ((i + 1) < rsm->r_rtr_cnt) {
8141 					/*
8142 					 * The peer ack'd from our previous
8143 					 * transmission. We have a spurious
8144 					 * retransmission and thus we dont
8145 					 * want to update our rack_rtt.
8146 					 *
8147 					 * Hmm should there be a CC revert here?
8148 					 *
8149 					 */
8150 					return (0);
8151 				}
8152 				if (!tp->t_rttlow || tp->t_rttlow > t)
8153 					tp->t_rttlow = t;
8154 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8155 					rack->r_ctl.rc_rack_min_rtt = t;
8156 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
8157 						rack->r_ctl.rc_rack_min_rtt = 1;
8158 					}
8159 				}
8160 				if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
8161 					   (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
8162 					/* New more recent rack_tmit_time */
8163 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8164 					rack->rc_rack_rtt = t;
8165 				}
8166 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
8167 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
8168 						    rsm->r_rtr_cnt);
8169 				return (1);
8170 			}
8171 		}
8172 		goto ts_not_found;
8173 	} else {
8174 		/*
8175 		 * Ok its a SACK block that we retransmitted. or a windows
8176 		 * machine without timestamps. We can tell nothing from the
8177 		 * time-stamp since its not there or the time the peer last
8178 		 * recieved a segment that moved forward its cum-ack point.
8179 		 */
8180 ts_not_found:
8181 		i = rsm->r_rtr_cnt - 1;
8182 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
8183 		if ((int)t <= 0)
8184 			t = 1;
8185 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8186 			/*
8187 			 * We retransmitted and the ack came back in less
8188 			 * than the smallest rtt we have observed. We most
8189 			 * likely did an improper retransmit as outlined in
8190 			 * 6.2 Step 2 point 2 in the rack-draft so we
8191 			 * don't want to update our rack_rtt. We in
8192 			 * theory (in future) might want to think about reverting our
8193 			 * cwnd state but we won't for now.
8194 			 */
8195 			return (0);
8196 		} else if (rack->r_ctl.rc_rack_min_rtt) {
8197 			/*
8198 			 * We retransmitted it and the retransmit did the
8199 			 * job.
8200 			 */
8201 			if (!rack->r_ctl.rc_rack_min_rtt ||
8202 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8203 				rack->r_ctl.rc_rack_min_rtt = t;
8204 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
8205 					rack->r_ctl.rc_rack_min_rtt = 1;
8206 				}
8207 			}
8208 			if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
8209 				/* New more recent rack_tmit_time */
8210 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
8211 				rack->rc_rack_rtt = t;
8212 			}
8213 			return (1);
8214 		}
8215 	}
8216 	return (0);
8217 }
8218 
8219 /*
8220  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
8221  */
8222 static void
8223 rack_log_sack_passed(struct tcpcb *tp,
8224     struct tcp_rack *rack, struct rack_sendmap *rsm)
8225 {
8226 	struct rack_sendmap *nrsm;
8227 
8228 	nrsm = rsm;
8229 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
8230 	    rack_head, r_tnext) {
8231 		if (nrsm == rsm) {
8232 			/* Skip orginal segment he is acked */
8233 			continue;
8234 		}
8235 		if (nrsm->r_flags & RACK_ACKED) {
8236 			/*
8237 			 * Skip ack'd segments, though we
8238 			 * should not see these, since tmap
8239 			 * should not have ack'd segments.
8240 			 */
8241 			continue;
8242 		}
8243 		if (nrsm->r_flags & RACK_SACK_PASSED) {
8244 			/*
8245 			 * We found one that is already marked
8246 			 * passed, we have been here before and
8247 			 * so all others below this are marked.
8248 			 */
8249 			break;
8250 		}
8251 		nrsm->r_flags |= RACK_SACK_PASSED;
8252 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
8253 	}
8254 }
8255 
8256 static void
8257 rack_need_set_test(struct tcpcb *tp,
8258 		   struct tcp_rack *rack,
8259 		   struct rack_sendmap *rsm,
8260 		   tcp_seq th_ack,
8261 		   int line,
8262 		   int use_which)
8263 {
8264 
8265 	if ((tp->t_flags & TF_GPUTINPROG) &&
8266 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8267 		/*
8268 		 * We were app limited, and this ack
8269 		 * butts up or goes beyond the point where we want
8270 		 * to start our next measurement. We need
8271 		 * to record the new gput_ts as here and
8272 		 * possibly update the start sequence.
8273 		 */
8274 		uint32_t seq, ts;
8275 
8276 		if (rsm->r_rtr_cnt > 1) {
8277 			/*
8278 			 * This is a retransmit, can we
8279 			 * really make any assessment at this
8280 			 * point?  We are not really sure of
8281 			 * the timestamp, is it this or the
8282 			 * previous transmission?
8283 			 *
8284 			 * Lets wait for something better that
8285 			 * is not retransmitted.
8286 			 */
8287 			return;
8288 		}
8289 		seq = tp->gput_seq;
8290 		ts = tp->gput_ts;
8291 		rack->app_limited_needs_set = 0;
8292 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8293 		/* Do we start at a new end? */
8294 		if ((use_which == RACK_USE_BEG) &&
8295 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8296 			/*
8297 			 * When we get an ACK that just eats
8298 			 * up some of the rsm, we set RACK_USE_BEG
8299 			 * since whats at r_start (i.e. th_ack)
8300 			 * is left unacked and thats where the
8301 			 * measurement not starts.
8302 			 */
8303 			tp->gput_seq = rsm->r_start;
8304 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8305 		}
8306 		if ((use_which == RACK_USE_END) &&
8307 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8308 			    /*
8309 			     * We use the end when the cumack
8310 			     * is moving forward and completely
8311 			     * deleting the rsm passed so basically
8312 			     * r_end holds th_ack.
8313 			     *
8314 			     * For SACK's we also want to use the end
8315 			     * since this piece just got sacked and
8316 			     * we want to target anything after that
8317 			     * in our measurement.
8318 			     */
8319 			    tp->gput_seq = rsm->r_end;
8320 			    rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8321 		}
8322 		if (use_which == RACK_USE_END_OR_THACK) {
8323 			/*
8324 			 * special case for ack moving forward,
8325 			 * not a sack, we need to move all the
8326 			 * way up to where this ack cum-ack moves
8327 			 * to.
8328 			 */
8329 			if (SEQ_GT(th_ack, rsm->r_end))
8330 				tp->gput_seq = th_ack;
8331 			else
8332 				tp->gput_seq = rsm->r_end;
8333 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8334 		}
8335 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8336 			/*
8337 			 * We moved beyond this guy's range, re-calculate
8338 			 * the new end point.
8339 			 */
8340 			if (rack->rc_gp_filled == 0) {
8341 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8342 			} else {
8343 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8344 			}
8345 		}
8346 		/*
8347 		 * We are moving the goal post, we may be able to clear the
8348 		 * measure_saw_probe_rtt flag.
8349 		 */
8350 		if ((rack->in_probe_rtt == 0) &&
8351 		    (rack->measure_saw_probe_rtt) &&
8352 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8353 			rack->measure_saw_probe_rtt = 0;
8354 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8355 					   seq, tp->gput_seq, 0, 5, line, NULL, 0);
8356 		if (rack->rc_gp_filled &&
8357 		    ((tp->gput_ack - tp->gput_seq) <
8358 		     max(rc_init_window(rack), (MIN_GP_WIN *
8359 						ctf_fixed_maxseg(tp))))) {
8360 			uint32_t ideal_amount;
8361 
8362 			ideal_amount = rack_get_measure_window(tp, rack);
8363 			if (ideal_amount > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
8364 				/*
8365 				 * There is no sense of continuing this measurement
8366 				 * because its too small to gain us anything we
8367 				 * trust. Skip it and that way we can start a new
8368 				 * measurement quicker.
8369 				 */
8370 				tp->t_flags &= ~TF_GPUTINPROG;
8371 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8372 							   0, 0, 0, 6, __LINE__, NULL, 0);
8373 			} else {
8374 				/*
8375 				 * Reset the window further out.
8376 				 */
8377 				tp->gput_ack = tp->gput_seq + ideal_amount;
8378 			}
8379 		}
8380 	}
8381 }
8382 
8383 static inline int
8384 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
8385 {
8386 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
8387 		/* Behind our TLP definition or right at */
8388 		return (0);
8389 	}
8390 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
8391 		/* The start is beyond or right at our end of TLP definition */
8392 		return (0);
8393 	}
8394 	/* It has to be a sub-part of the original TLP recorded */
8395 	return (1);
8396 }
8397 
8398 
8399 static uint32_t
8400 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8401 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8402 {
8403 	uint32_t start, end, changed = 0;
8404 	struct rack_sendmap stack_map;
8405 	struct rack_sendmap *rsm, *nrsm, fe, *insret, *prev, *next;
8406 	int32_t used_ref = 1;
8407 	int moved = 0;
8408 
8409 	start = sack->start;
8410 	end = sack->end;
8411 	rsm = *prsm;
8412 	memset(&fe, 0, sizeof(fe));
8413 do_rest_ofb:
8414 	if ((rsm == NULL) ||
8415 	    (SEQ_LT(end, rsm->r_start)) ||
8416 	    (SEQ_GEQ(start, rsm->r_end)) ||
8417 	    (SEQ_LT(start, rsm->r_start))) {
8418 		/*
8419 		 * We are not in the right spot,
8420 		 * find the correct spot in the tree.
8421 		 */
8422 		used_ref = 0;
8423 		fe.r_start = start;
8424 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8425 		moved++;
8426 	}
8427 	if (rsm == NULL) {
8428 		/* TSNH */
8429 		goto out;
8430 	}
8431 	/* Ok we have an ACK for some piece of this rsm */
8432 	if (rsm->r_start != start) {
8433 		if ((rsm->r_flags & RACK_ACKED) == 0) {
8434 			/*
8435 			 * Before any splitting or hookery is
8436 			 * done is it a TLP of interest i.e. rxt?
8437 			 */
8438 			if ((rsm->r_flags & RACK_TLP) &&
8439 			    (rsm->r_rtr_cnt > 1)) {
8440 				/*
8441 				 * We are splitting a rxt TLP, check
8442 				 * if we need to save off the start/end
8443 				 */
8444 				if (rack->rc_last_tlp_acked_set &&
8445 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8446 					/*
8447 					 * We already turned this on since we are inside
8448 					 * the previous one was a partially sack now we
8449 					 * are getting another one (maybe all of it).
8450 					 *
8451 					 */
8452 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8453 					/*
8454 					 * Lets make sure we have all of it though.
8455 					 */
8456 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8457 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8458 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8459 								     rack->r_ctl.last_tlp_acked_end);
8460 					}
8461 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8462 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8463 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8464 								     rack->r_ctl.last_tlp_acked_end);
8465 					}
8466 				} else {
8467 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8468 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8469 					rack->rc_last_tlp_past_cumack = 0;
8470 					rack->rc_last_tlp_acked_set = 1;
8471 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8472 				}
8473 			}
8474 			/**
8475 			 * Need to split this in two pieces the before and after,
8476 			 * the before remains in the map, the after must be
8477 			 * added. In other words we have:
8478 			 * rsm        |--------------|
8479 			 * sackblk        |------->
8480 			 * rsm will become
8481 			 *     rsm    |---|
8482 			 * and nrsm will be  the sacked piece
8483 			 *     nrsm       |----------|
8484 			 *
8485 			 * But before we start down that path lets
8486 			 * see if the sack spans over on top of
8487 			 * the next guy and it is already sacked.
8488 			 *
8489 			 */
8490 			next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8491 			if (next && (next->r_flags & RACK_ACKED) &&
8492 			    SEQ_GEQ(end, next->r_start)) {
8493 				/**
8494 				 * So the next one is already acked, and
8495 				 * we can thus by hookery use our stack_map
8496 				 * to reflect the piece being sacked and
8497 				 * then adjust the two tree entries moving
8498 				 * the start and ends around. So we start like:
8499 				 *  rsm     |------------|             (not-acked)
8500 				 *  next                 |-----------| (acked)
8501 				 *  sackblk        |-------->
8502 				 *  We want to end like so:
8503 				 *  rsm     |------|                   (not-acked)
8504 				 *  next           |-----------------| (acked)
8505 				 *  nrsm           |-----|
8506 				 * Where nrsm is a temporary stack piece we
8507 				 * use to update all the gizmos.
8508 				 */
8509 				/* Copy up our fudge block */
8510 				nrsm = &stack_map;
8511 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8512 				/* Now adjust our tree blocks */
8513 				rsm->r_end = start;
8514 				next->r_start = start;
8515 				/* Now we must adjust back where next->m is */
8516 				rack_setup_offset_for_rsm(rsm, next);
8517 
8518 				/* We don't need to adjust rsm, it did not change */
8519 				/* Clear out the dup ack count of the remainder */
8520 				rsm->r_dupack = 0;
8521 				rsm->r_just_ret = 0;
8522 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8523 				/* Now lets make sure our fudge block is right */
8524 				nrsm->r_start = start;
8525 				/* Now lets update all the stats and such */
8526 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8527 				if (rack->app_limited_needs_set)
8528 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8529 				changed += (nrsm->r_end - nrsm->r_start);
8530 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8531 				if (nrsm->r_flags & RACK_SACK_PASSED) {
8532 					counter_u64_add(rack_reorder_seen, 1);
8533 					rack->r_ctl.rc_reorder_ts = cts;
8534 				}
8535 				/*
8536 				 * Now we want to go up from rsm (the
8537 				 * one left un-acked) to the next one
8538 				 * in the tmap. We do this so when
8539 				 * we walk backwards we include marking
8540 				 * sack-passed on rsm (The one passed in
8541 				 * is skipped since it is generally called
8542 				 * on something sacked before removing it
8543 				 * from the tmap).
8544 				 */
8545 				if (rsm->r_in_tmap) {
8546 					nrsm = TAILQ_NEXT(rsm, r_tnext);
8547 					/*
8548 					 * Now that we have the next
8549 					 * one walk backwards from there.
8550 					 */
8551 					if (nrsm && nrsm->r_in_tmap)
8552 						rack_log_sack_passed(tp, rack, nrsm);
8553 				}
8554 				/* Now are we done? */
8555 				if (SEQ_LT(end, next->r_end) ||
8556 				    (end == next->r_end)) {
8557 					/* Done with block */
8558 					goto out;
8559 				}
8560 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8561 				counter_u64_add(rack_sack_used_next_merge, 1);
8562 				/* Postion for the next block */
8563 				start = next->r_end;
8564 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8565 				if (rsm == NULL)
8566 					goto out;
8567 			} else {
8568 				/**
8569 				 * We can't use any hookery here, so we
8570 				 * need to split the map. We enter like
8571 				 * so:
8572 				 *  rsm      |--------|
8573 				 *  sackblk       |----->
8574 				 * We will add the new block nrsm and
8575 				 * that will be the new portion, and then
8576 				 * fall through after reseting rsm. So we
8577 				 * split and look like this:
8578 				 *  rsm      |----|
8579 				 *  sackblk       |----->
8580 				 *  nrsm          |---|
8581 				 * We then fall through reseting
8582 				 * rsm to nrsm, so the next block
8583 				 * picks it up.
8584 				 */
8585 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8586 				if (nrsm == NULL) {
8587 					/*
8588 					 * failed XXXrrs what can we do but loose the sack
8589 					 * info?
8590 					 */
8591 					goto out;
8592 				}
8593 				counter_u64_add(rack_sack_splits, 1);
8594 				rack_clone_rsm(rack, nrsm, rsm, start);
8595 				rsm->r_just_ret = 0;
8596 				insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8597 #ifdef INVARIANTS
8598 				if (insret != NULL) {
8599 					panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8600 					      nrsm, insret, rack, rsm);
8601 				}
8602 #endif
8603 				if (rsm->r_in_tmap) {
8604 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8605 					nrsm->r_in_tmap = 1;
8606 				}
8607 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8608 				rsm->r_flags &= (~RACK_HAS_FIN);
8609 				/* Position us to point to the new nrsm that starts the sack blk */
8610 				rsm = nrsm;
8611 			}
8612 		} else {
8613 			/* Already sacked this piece */
8614 			counter_u64_add(rack_sack_skipped_acked, 1);
8615 			moved++;
8616 			if (end == rsm->r_end) {
8617 				/* Done with block */
8618 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8619 				goto out;
8620 			} else if (SEQ_LT(end, rsm->r_end)) {
8621 				/* A partial sack to a already sacked block */
8622 				moved++;
8623 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8624 				goto out;
8625 			} else {
8626 				/*
8627 				 * The end goes beyond this guy
8628 				 * repostion the start to the
8629 				 * next block.
8630 				 */
8631 				start = rsm->r_end;
8632 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8633 				if (rsm == NULL)
8634 					goto out;
8635 			}
8636 		}
8637 	}
8638 	if (SEQ_GEQ(end, rsm->r_end)) {
8639 		/**
8640 		 * The end of this block is either beyond this guy or right
8641 		 * at this guy. I.e.:
8642 		 *  rsm ---                 |-----|
8643 		 *  end                     |-----|
8644 		 *  <or>
8645 		 *  end                     |---------|
8646 		 */
8647 		if ((rsm->r_flags & RACK_ACKED) == 0) {
8648 			/*
8649 			 * Is it a TLP of interest?
8650 			 */
8651 			if ((rsm->r_flags & RACK_TLP) &&
8652 			    (rsm->r_rtr_cnt > 1)) {
8653 				/*
8654 				 * We are splitting a rxt TLP, check
8655 				 * if we need to save off the start/end
8656 				 */
8657 				if (rack->rc_last_tlp_acked_set &&
8658 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8659 					/*
8660 					 * We already turned this on since we are inside
8661 					 * the previous one was a partially sack now we
8662 					 * are getting another one (maybe all of it).
8663 					 */
8664 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8665 					/*
8666 					 * Lets make sure we have all of it though.
8667 					 */
8668 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8669 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8670 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8671 								     rack->r_ctl.last_tlp_acked_end);
8672 					}
8673 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8674 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8675 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8676 								     rack->r_ctl.last_tlp_acked_end);
8677 					}
8678 				} else {
8679 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8680 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8681 					rack->rc_last_tlp_past_cumack = 0;
8682 					rack->rc_last_tlp_acked_set = 1;
8683 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8684 				}
8685 			}
8686 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8687 			changed += (rsm->r_end - rsm->r_start);
8688 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8689 			if (rsm->r_in_tmap) /* should be true */
8690 				rack_log_sack_passed(tp, rack, rsm);
8691 			/* Is Reordering occuring? */
8692 			if (rsm->r_flags & RACK_SACK_PASSED) {
8693 				rsm->r_flags &= ~RACK_SACK_PASSED;
8694 				counter_u64_add(rack_reorder_seen, 1);
8695 				rack->r_ctl.rc_reorder_ts = cts;
8696 			}
8697 			if (rack->app_limited_needs_set)
8698 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8699 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8700 			rsm->r_flags |= RACK_ACKED;
8701 			if (rsm->r_in_tmap) {
8702 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8703 				rsm->r_in_tmap = 0;
8704 			}
8705 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8706 		} else {
8707 			counter_u64_add(rack_sack_skipped_acked, 1);
8708 			moved++;
8709 		}
8710 		if (end == rsm->r_end) {
8711 			/* This block only - done, setup for next */
8712 			goto out;
8713 		}
8714 		/*
8715 		 * There is more not coverend by this rsm move on
8716 		 * to the next block in the RB tree.
8717 		 */
8718 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8719 		start = rsm->r_end;
8720 		rsm = nrsm;
8721 		if (rsm == NULL)
8722 			goto out;
8723 		goto do_rest_ofb;
8724 	}
8725 	/**
8726 	 * The end of this sack block is smaller than
8727 	 * our rsm i.e.:
8728 	 *  rsm ---                 |-----|
8729 	 *  end                     |--|
8730 	 */
8731 	if ((rsm->r_flags & RACK_ACKED) == 0) {
8732 		/*
8733 		 * Is it a TLP of interest?
8734 		 */
8735 		if ((rsm->r_flags & RACK_TLP) &&
8736 		    (rsm->r_rtr_cnt > 1)) {
8737 			/*
8738 			 * We are splitting a rxt TLP, check
8739 			 * if we need to save off the start/end
8740 			 */
8741 			if (rack->rc_last_tlp_acked_set &&
8742 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8743 				/*
8744 				 * We already turned this on since we are inside
8745 				 * the previous one was a partially sack now we
8746 				 * are getting another one (maybe all of it).
8747 				 */
8748 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8749 				/*
8750 				 * Lets make sure we have all of it though.
8751 				 */
8752 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8753 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8754 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8755 							     rack->r_ctl.last_tlp_acked_end);
8756 				}
8757 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8758 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8759 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8760 							     rack->r_ctl.last_tlp_acked_end);
8761 				}
8762 			} else {
8763 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8764 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8765 				rack->rc_last_tlp_past_cumack = 0;
8766 				rack->rc_last_tlp_acked_set = 1;
8767 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8768 			}
8769 		}
8770 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8771 		if (prev &&
8772 		    (prev->r_flags & RACK_ACKED)) {
8773 			/**
8774 			 * Goal, we want the right remainder of rsm to shrink
8775 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
8776 			 * We want to expand prev to go all the way
8777 			 * to prev->r_end <- end.
8778 			 * so in the tree we have before:
8779 			 *   prev     |--------|         (acked)
8780 			 *   rsm               |-------| (non-acked)
8781 			 *   sackblk           |-|
8782 			 * We churn it so we end up with
8783 			 *   prev     |----------|       (acked)
8784 			 *   rsm                 |-----| (non-acked)
8785 			 *   nrsm              |-| (temporary)
8786 			 *
8787 			 * Note if either prev/rsm is a TLP we don't
8788 			 * do this.
8789 			 */
8790 			nrsm = &stack_map;
8791 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8792 			prev->r_end = end;
8793 			rsm->r_start = end;
8794 			/* Now adjust nrsm (stack copy) to be
8795 			 * the one that is the small
8796 			 * piece that was "sacked".
8797 			 */
8798 			nrsm->r_end = end;
8799 			rsm->r_dupack = 0;
8800 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8801 			/*
8802 			 * Now that the rsm has had its start moved forward
8803 			 * lets go ahead and get its new place in the world.
8804 			 */
8805 			rack_setup_offset_for_rsm(prev, rsm);
8806 			/*
8807 			 * Now nrsm is our new little piece
8808 			 * that is acked (which was merged
8809 			 * to prev). Update the rtt and changed
8810 			 * based on that. Also check for reordering.
8811 			 */
8812 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8813 			if (rack->app_limited_needs_set)
8814 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8815 			changed += (nrsm->r_end - nrsm->r_start);
8816 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8817 			if (nrsm->r_flags & RACK_SACK_PASSED) {
8818 				counter_u64_add(rack_reorder_seen, 1);
8819 				rack->r_ctl.rc_reorder_ts = cts;
8820 			}
8821 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8822 			rsm = prev;
8823 			counter_u64_add(rack_sack_used_prev_merge, 1);
8824 		} else {
8825 			/**
8826 			 * This is the case where our previous
8827 			 * block is not acked either, so we must
8828 			 * split the block in two.
8829 			 */
8830 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8831 			if (nrsm == NULL) {
8832 				/* failed rrs what can we do but loose the sack info? */
8833 				goto out;
8834 			}
8835 			if ((rsm->r_flags & RACK_TLP) &&
8836 			    (rsm->r_rtr_cnt > 1)) {
8837 				/*
8838 				 * We are splitting a rxt TLP, check
8839 				 * if we need to save off the start/end
8840 				 */
8841 				if (rack->rc_last_tlp_acked_set &&
8842 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8843 					    /*
8844 					     * We already turned this on since this block is inside
8845 					     * the previous one was a partially sack now we
8846 					     * are getting another one (maybe all of it).
8847 					     */
8848 					    rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8849 					    /*
8850 					     * Lets make sure we have all of it though.
8851 					     */
8852 					    if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8853 						    rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8854 						    rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8855 									 rack->r_ctl.last_tlp_acked_end);
8856 					    }
8857 					    if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8858 						    rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8859 						    rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8860 									 rack->r_ctl.last_tlp_acked_end);
8861 					    }
8862 				    } else {
8863 					    rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8864 					    rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8865 					    rack->rc_last_tlp_acked_set = 1;
8866 					    rack->rc_last_tlp_past_cumack = 0;
8867 					    rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8868 				    }
8869 			}
8870 			/**
8871 			 * In this case nrsm becomes
8872 			 * nrsm->r_start = end;
8873 			 * nrsm->r_end = rsm->r_end;
8874 			 * which is un-acked.
8875 			 * <and>
8876 			 * rsm->r_end = nrsm->r_start;
8877 			 * i.e. the remaining un-acked
8878 			 * piece is left on the left
8879 			 * hand side.
8880 			 *
8881 			 * So we start like this
8882 			 * rsm      |----------| (not acked)
8883 			 * sackblk  |---|
8884 			 * build it so we have
8885 			 * rsm      |---|         (acked)
8886 			 * nrsm         |------|  (not acked)
8887 			 */
8888 			counter_u64_add(rack_sack_splits, 1);
8889 			rack_clone_rsm(rack, nrsm, rsm, end);
8890 			rsm->r_flags &= (~RACK_HAS_FIN);
8891 			rsm->r_just_ret = 0;
8892 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8893 #ifdef INVARIANTS
8894 			if (insret != NULL) {
8895 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8896 				      nrsm, insret, rack, rsm);
8897 			}
8898 #endif
8899 			if (rsm->r_in_tmap) {
8900 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8901 				nrsm->r_in_tmap = 1;
8902 			}
8903 			nrsm->r_dupack = 0;
8904 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8905 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8906 			changed += (rsm->r_end - rsm->r_start);
8907 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8908 			if (rsm->r_in_tmap) /* should be true */
8909 				rack_log_sack_passed(tp, rack, rsm);
8910 			/* Is Reordering occuring? */
8911 			if (rsm->r_flags & RACK_SACK_PASSED) {
8912 				rsm->r_flags &= ~RACK_SACK_PASSED;
8913 				counter_u64_add(rack_reorder_seen, 1);
8914 				rack->r_ctl.rc_reorder_ts = cts;
8915 			}
8916 			if (rack->app_limited_needs_set)
8917 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8918 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8919 			rsm->r_flags |= RACK_ACKED;
8920 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8921 			if (rsm->r_in_tmap) {
8922 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8923 				rsm->r_in_tmap = 0;
8924 			}
8925 		}
8926 	} else if (start != end){
8927 		/*
8928 		 * The block was already acked.
8929 		 */
8930 		counter_u64_add(rack_sack_skipped_acked, 1);
8931 		moved++;
8932 	}
8933 out:
8934 	if (rsm &&
8935 	    ((rsm->r_flags & RACK_TLP) == 0) &&
8936 	    (rsm->r_flags & RACK_ACKED)) {
8937 		/*
8938 		 * Now can we merge where we worked
8939 		 * with either the previous or
8940 		 * next block?
8941 		 */
8942 		next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8943 		while (next) {
8944 			if (next->r_flags & RACK_TLP)
8945 				break;
8946 			if (next->r_flags & RACK_ACKED) {
8947 			/* yep this and next can be merged */
8948 				rsm = rack_merge_rsm(rack, rsm, next);
8949 				next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8950 			} else
8951 				break;
8952 		}
8953 		/* Now what about the previous? */
8954 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8955 		while (prev) {
8956 			if (prev->r_flags & RACK_TLP)
8957 				break;
8958 			if (prev->r_flags & RACK_ACKED) {
8959 				/* yep the previous and this can be merged */
8960 				rsm = rack_merge_rsm(rack, prev, rsm);
8961 				prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8962 			} else
8963 				break;
8964 		}
8965 	}
8966 	if (used_ref == 0) {
8967 		counter_u64_add(rack_sack_proc_all, 1);
8968 	} else {
8969 		counter_u64_add(rack_sack_proc_short, 1);
8970 	}
8971 	/* Save off the next one for quick reference. */
8972 	if (rsm)
8973 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8974 	else
8975 		nrsm = NULL;
8976 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
8977 	/* Pass back the moved. */
8978 	*moved_two = moved;
8979 	return (changed);
8980 }
8981 
8982 static void inline
8983 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8984 {
8985 	struct rack_sendmap *tmap;
8986 
8987 	tmap = NULL;
8988 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
8989 		/* Its no longer sacked, mark it so */
8990 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8991 #ifdef INVARIANTS
8992 		if (rsm->r_in_tmap) {
8993 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
8994 			      rack, rsm, rsm->r_flags);
8995 		}
8996 #endif
8997 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8998 		/* Rebuild it into our tmap */
8999 		if (tmap == NULL) {
9000 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9001 			tmap = rsm;
9002 		} else {
9003 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
9004 			tmap = rsm;
9005 		}
9006 		tmap->r_in_tmap = 1;
9007 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
9008 	}
9009 	/*
9010 	 * Now lets possibly clear the sack filter so we start
9011 	 * recognizing sacks that cover this area.
9012 	 */
9013 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
9014 
9015 }
9016 
9017 static void
9018 rack_do_decay(struct tcp_rack *rack)
9019 {
9020 	struct timeval res;
9021 
9022 #define	timersub(tvp, uvp, vvp)						\
9023 	do {								\
9024 		(vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;		\
9025 		(vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;	\
9026 		if ((vvp)->tv_usec < 0) {				\
9027 			(vvp)->tv_sec--;				\
9028 			(vvp)->tv_usec += 1000000;			\
9029 		}							\
9030 	} while (0)
9031 
9032 	timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
9033 #undef timersub
9034 
9035 	rack->r_ctl.input_pkt++;
9036 	if ((rack->rc_in_persist) ||
9037 	    (res.tv_sec >= 1) ||
9038 	    (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
9039 		/*
9040 		 * Check for decay of non-SAD,
9041 		 * we want all SAD detection metrics to
9042 		 * decay 1/4 per second (or more) passed.
9043 		 */
9044 		uint32_t pkt_delta;
9045 
9046 		pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
9047 		/* Update our saved tracking values */
9048 		rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
9049 		rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
9050 		/* Now do we escape without decay? */
9051 #ifdef NETFLIX_EXP_DETECTION
9052 		if (rack->rc_in_persist ||
9053 		    (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
9054 		    (pkt_delta < tcp_sad_low_pps)){
9055 			/*
9056 			 * We don't decay idle connections
9057 			 * or ones that have a low input pps.
9058 			 */
9059 			return;
9060 		}
9061 		/* Decay the counters */
9062 		rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
9063 							tcp_sad_decay_val);
9064 		rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
9065 							 tcp_sad_decay_val);
9066 		rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
9067 							       tcp_sad_decay_val);
9068 		rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
9069 								tcp_sad_decay_val);
9070 #endif
9071 	}
9072 }
9073 
9074 static void
9075 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
9076 {
9077 	struct rack_sendmap *rsm, *rm;
9078 
9079 	/*
9080 	 * The ACK point is advancing to th_ack, we must drop off
9081 	 * the packets in the rack log and calculate any eligble
9082 	 * RTT's.
9083 	 */
9084 	rack->r_wanted_output = 1;
9085 
9086 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
9087 	if ((rack->rc_last_tlp_acked_set == 1)&&
9088 	    (rack->rc_last_tlp_past_cumack == 1) &&
9089 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
9090 		/*
9091 		 * We have reached the point where our last rack
9092 		 * tlp retransmit sequence is ahead of the cum-ack.
9093 		 * This can only happen when the cum-ack moves all
9094 		 * the way around (its been a full 2^^31+1 bytes
9095 		 * or more since we sent a retransmitted TLP). Lets
9096 		 * turn off the valid flag since its not really valid.
9097 		 *
9098 		 * Note since sack's also turn on this event we have
9099 		 * a complication, we have to wait to age it out until
9100 		 * the cum-ack is by the TLP before checking which is
9101 		 * what the next else clause does.
9102 		 */
9103 		rack_log_dsack_event(rack, 9, __LINE__,
9104 				     rack->r_ctl.last_tlp_acked_start,
9105 				     rack->r_ctl.last_tlp_acked_end);
9106 		rack->rc_last_tlp_acked_set = 0;
9107 		rack->rc_last_tlp_past_cumack = 0;
9108 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
9109 		   (rack->rc_last_tlp_past_cumack == 0) &&
9110 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
9111 		/*
9112 		 * It is safe to start aging TLP's out.
9113 		 */
9114 		rack->rc_last_tlp_past_cumack = 1;
9115 	}
9116 	/* We do the same for the tlp send seq as well */
9117 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
9118 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
9119 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
9120 		rack_log_dsack_event(rack, 9, __LINE__,
9121 				     rack->r_ctl.last_sent_tlp_seq,
9122 				     (rack->r_ctl.last_sent_tlp_seq +
9123 				      rack->r_ctl.last_sent_tlp_len));
9124 		rack->rc_last_sent_tlp_seq_valid = 0;
9125 		rack->rc_last_sent_tlp_past_cumack = 0;
9126 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
9127 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
9128 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
9129 		/*
9130 		 * It is safe to start aging TLP's send.
9131 		 */
9132 		rack->rc_last_sent_tlp_past_cumack = 1;
9133 	}
9134 more:
9135 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9136 	if (rsm == NULL) {
9137 		if ((th_ack - 1) == tp->iss) {
9138 			/*
9139 			 * For the SYN incoming case we will not
9140 			 * have called tcp_output for the sending of
9141 			 * the SYN, so there will be no map. All
9142 			 * other cases should probably be a panic.
9143 			 */
9144 			return;
9145 		}
9146 		if (tp->t_flags & TF_SENTFIN) {
9147 			/* if we sent a FIN we often will not have map */
9148 			return;
9149 		}
9150 #ifdef INVARIANTS
9151 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
9152 		      tp,
9153 		      tp->t_state, th_ack, rack,
9154 		      tp->snd_una, tp->snd_max, tp->snd_nxt);
9155 #endif
9156 		return;
9157 	}
9158 	if (SEQ_LT(th_ack, rsm->r_start)) {
9159 		/* Huh map is missing this */
9160 #ifdef INVARIANTS
9161 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
9162 		       rsm->r_start,
9163 		       th_ack, tp->t_state, rack->r_state);
9164 #endif
9165 		return;
9166 	}
9167 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
9168 
9169 	/* Now was it a retransmitted TLP? */
9170 	if ((rsm->r_flags & RACK_TLP) &&
9171 	    (rsm->r_rtr_cnt > 1)) {
9172 		/*
9173 		 * Yes, this rsm was a TLP and retransmitted, remember that
9174 		 * since if a DSACK comes back on this we don't want
9175 		 * to think of it as a reordered segment. This may
9176 		 * get updated again with possibly even other TLPs
9177 		 * in flight, but thats ok. Only when we don't send
9178 		 * a retransmitted TLP for 1/2 the sequences space
9179 		 * will it get turned off (above).
9180 		 */
9181 		if (rack->rc_last_tlp_acked_set &&
9182 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9183 			/*
9184 			 * We already turned this on since the end matches,
9185 			 * the previous one was a partially ack now we
9186 			 * are getting another one (maybe all of it).
9187 			 */
9188 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9189 			/*
9190 			 * Lets make sure we have all of it though.
9191 			 */
9192 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9193 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9194 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9195 						     rack->r_ctl.last_tlp_acked_end);
9196 			}
9197 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9198 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9199 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9200 						     rack->r_ctl.last_tlp_acked_end);
9201 			}
9202 		} else {
9203 			rack->rc_last_tlp_past_cumack = 1;
9204 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9205 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9206 			rack->rc_last_tlp_acked_set = 1;
9207 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9208 		}
9209 	}
9210 	/* Now do we consume the whole thing? */
9211 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
9212 		/* Its all consumed. */
9213 		uint32_t left;
9214 		uint8_t newly_acked;
9215 
9216 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
9217 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
9218 		rsm->r_rtr_bytes = 0;
9219 		/* Record the time of highest cumack sent */
9220 		rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9221 		rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
9222 #ifdef INVARIANTS
9223 		if (rm != rsm) {
9224 			panic("removing head in rack:%p rsm:%p rm:%p",
9225 			      rack, rsm, rm);
9226 		}
9227 #endif
9228 		if (rsm->r_in_tmap) {
9229 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9230 			rsm->r_in_tmap = 0;
9231 		}
9232 		newly_acked = 1;
9233 		if (rsm->r_flags & RACK_ACKED) {
9234 			/*
9235 			 * It was acked on the scoreboard -- remove
9236 			 * it from total
9237 			 */
9238 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
9239 			newly_acked = 0;
9240 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
9241 			/*
9242 			 * There are segments ACKED on the
9243 			 * scoreboard further up. We are seeing
9244 			 * reordering.
9245 			 */
9246 			rsm->r_flags &= ~RACK_SACK_PASSED;
9247 			counter_u64_add(rack_reorder_seen, 1);
9248 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9249 			rsm->r_flags |= RACK_ACKED;
9250 			rack->r_ctl.rc_reorder_ts = cts;
9251 			if (rack->r_ent_rec_ns) {
9252 				/*
9253 				 * We have sent no more, and we saw an sack
9254 				 * then ack arrive.
9255 				 */
9256 				rack->r_might_revert = 1;
9257 			}
9258 		}
9259 		if ((rsm->r_flags & RACK_TO_REXT) &&
9260 		    (tp->t_flags & TF_RCVD_TSTMP) &&
9261 		    (to->to_flags & TOF_TS) &&
9262 		    (to->to_tsecr != 0) &&
9263 		    (tp->t_flags & TF_PREVVALID)) {
9264 			/*
9265 			 * We can use the timestamp to see
9266 			 * if this retransmission was from the
9267 			 * first transmit. If so we made a mistake.
9268 			 */
9269 			tp->t_flags &= ~TF_PREVVALID;
9270 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
9271 				/* The first transmit is what this ack is for */
9272 				rack_cong_signal(tp, CC_RTO_ERR, th_ack);
9273 			}
9274 		}
9275 		left = th_ack - rsm->r_end;
9276 		if (rack->app_limited_needs_set && newly_acked)
9277 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
9278 		/* Free back to zone */
9279 		rack_free(rack, rsm);
9280 		if (left) {
9281 			goto more;
9282 		}
9283 		/* Check for reneging */
9284 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9285 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
9286 			/*
9287 			 * The peer has moved snd_una up to
9288 			 * the edge of this send, i.e. one
9289 			 * that it had previously acked. The only
9290 			 * way that can be true if the peer threw
9291 			 * away data (space issues) that it had
9292 			 * previously sacked (else it would have
9293 			 * given us snd_una up to (rsm->r_end).
9294 			 * We need to undo the acked markings here.
9295 			 *
9296 			 * Note we have to look to make sure th_ack is
9297 			 * our rsm->r_start in case we get an old ack
9298 			 * where th_ack is behind snd_una.
9299 			 */
9300 			rack_peer_reneges(rack, rsm, th_ack);
9301 		}
9302 		return;
9303 	}
9304 	if (rsm->r_flags & RACK_ACKED) {
9305 		/*
9306 		 * It was acked on the scoreboard -- remove it from
9307 		 * total for the part being cum-acked.
9308 		 */
9309 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
9310 	}
9311 	/*
9312 	 * Clear the dup ack count for
9313 	 * the piece that remains.
9314 	 */
9315 	rsm->r_dupack = 0;
9316 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9317 	if (rsm->r_rtr_bytes) {
9318 		/*
9319 		 * It was retransmitted adjust the
9320 		 * sack holes for what was acked.
9321 		 */
9322 		int ack_am;
9323 
9324 		ack_am = (th_ack - rsm->r_start);
9325 		if (ack_am >= rsm->r_rtr_bytes) {
9326 			rack->r_ctl.rc_holes_rxt -= ack_am;
9327 			rsm->r_rtr_bytes -= ack_am;
9328 		}
9329 	}
9330 	/*
9331 	 * Update where the piece starts and record
9332 	 * the time of send of highest cumack sent.
9333 	 */
9334 	rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9335 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
9336 	/* Now we need to move our offset forward too */
9337 	if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
9338 		/* Fix up the orig_m_len and possibly the mbuf offset */
9339 		rack_adjust_orig_mlen(rsm);
9340 	}
9341 	rsm->soff += (th_ack - rsm->r_start);
9342 	rsm->r_start = th_ack;
9343 	/* Now do we need to move the mbuf fwd too? */
9344 	if (rsm->m) {
9345 		while (rsm->soff >= rsm->m->m_len) {
9346 			rsm->soff -= rsm->m->m_len;
9347 			rsm->m = rsm->m->m_next;
9348 			KASSERT((rsm->m != NULL),
9349 				(" nrsm:%p hit at soff:%u null m",
9350 				 rsm, rsm->soff));
9351 		}
9352 		rsm->orig_m_len = rsm->m->m_len;
9353 	}
9354 	if (rack->app_limited_needs_set)
9355 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
9356 }
9357 
9358 static void
9359 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
9360 {
9361 	struct rack_sendmap *rsm;
9362 	int sack_pass_fnd = 0;
9363 
9364 	if (rack->r_might_revert) {
9365 		/*
9366 		 * Ok we have reordering, have not sent anything, we
9367 		 * might want to revert the congestion state if nothing
9368 		 * further has SACK_PASSED on it. Lets check.
9369 		 *
9370 		 * We also get here when we have DSACKs come in for
9371 		 * all the data that we FR'd. Note that a rxt or tlp
9372 		 * timer clears this from happening.
9373 		 */
9374 
9375 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
9376 			if (rsm->r_flags & RACK_SACK_PASSED) {
9377 				sack_pass_fnd = 1;
9378 				break;
9379 			}
9380 		}
9381 		if (sack_pass_fnd == 0) {
9382 			/*
9383 			 * We went into recovery
9384 			 * incorrectly due to reordering!
9385 			 */
9386 			int orig_cwnd;
9387 
9388 			rack->r_ent_rec_ns = 0;
9389 			orig_cwnd = tp->snd_cwnd;
9390 			tp->snd_cwnd = rack->r_ctl.rc_cwnd_at_erec;
9391 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
9392 			tp->snd_recover = tp->snd_una;
9393 			rack_log_to_prr(rack, 14, orig_cwnd);
9394 			EXIT_RECOVERY(tp->t_flags);
9395 		}
9396 		rack->r_might_revert = 0;
9397 	}
9398 }
9399 
9400 #ifdef NETFLIX_EXP_DETECTION
9401 static void
9402 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
9403 {
9404 	if ((rack->do_detection || tcp_force_detection) &&
9405 	    tcp_sack_to_ack_thresh &&
9406 	    tcp_sack_to_move_thresh &&
9407 	    ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
9408 		/*
9409 		 * We have thresholds set to find
9410 		 * possible attackers and disable sack.
9411 		 * Check them.
9412 		 */
9413 		uint64_t ackratio, moveratio, movetotal;
9414 
9415 		/* Log detecting */
9416 		rack_log_sad(rack, 1);
9417 		ackratio = (uint64_t)(rack->r_ctl.sack_count);
9418 		ackratio *= (uint64_t)(1000);
9419 		if (rack->r_ctl.ack_count)
9420 			ackratio /= (uint64_t)(rack->r_ctl.ack_count);
9421 		else {
9422 			/* We really should not hit here */
9423 			ackratio = 1000;
9424 		}
9425 		if ((rack->sack_attack_disable == 0) &&
9426 		    (ackratio > rack_highest_sack_thresh_seen))
9427 			rack_highest_sack_thresh_seen = (uint32_t)ackratio;
9428 		movetotal = rack->r_ctl.sack_moved_extra;
9429 		movetotal += rack->r_ctl.sack_noextra_move;
9430 		moveratio = rack->r_ctl.sack_moved_extra;
9431 		moveratio *= (uint64_t)1000;
9432 		if (movetotal)
9433 			moveratio /= movetotal;
9434 		else {
9435 			/* No moves, thats pretty good */
9436 			moveratio = 0;
9437 		}
9438 		if ((rack->sack_attack_disable == 0) &&
9439 		    (moveratio > rack_highest_move_thresh_seen))
9440 			rack_highest_move_thresh_seen = (uint32_t)moveratio;
9441 		if (rack->sack_attack_disable == 0) {
9442 			if ((ackratio > tcp_sack_to_ack_thresh) &&
9443 			    (moveratio > tcp_sack_to_move_thresh)) {
9444 				/* Disable sack processing */
9445 				rack->sack_attack_disable = 1;
9446 				if (rack->r_rep_attack == 0) {
9447 					rack->r_rep_attack = 1;
9448 					counter_u64_add(rack_sack_attacks_detected, 1);
9449 				}
9450 				if (tcp_attack_on_turns_on_logging) {
9451 					/*
9452 					 * Turn on logging, used for debugging
9453 					 * false positives.
9454 					 */
9455 					rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
9456 				}
9457 				/* Clamp the cwnd at flight size */
9458 				rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
9459 				rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9460 				rack_log_sad(rack, 2);
9461 			}
9462 		} else {
9463 			/* We are sack-disabled check for false positives */
9464 			if ((ackratio <= tcp_restoral_thresh) ||
9465 			    (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
9466 				rack->sack_attack_disable = 0;
9467 				rack_log_sad(rack, 3);
9468 				/* Restart counting */
9469 				rack->r_ctl.sack_count = 0;
9470 				rack->r_ctl.sack_moved_extra = 0;
9471 				rack->r_ctl.sack_noextra_move = 1;
9472 				rack->r_ctl.ack_count = max(1,
9473 				      (bytes_this_ack / segsiz));
9474 
9475 				if (rack->r_rep_reverse == 0) {
9476 					rack->r_rep_reverse = 1;
9477 					counter_u64_add(rack_sack_attacks_reversed, 1);
9478 				}
9479 				/* Restore the cwnd */
9480 				if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
9481 					rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
9482 			}
9483 		}
9484 	}
9485 }
9486 #endif
9487 
9488 static int
9489 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
9490 {
9491 
9492 	uint32_t am, l_end;
9493 	int was_tlp = 0;
9494 
9495 	if (SEQ_GT(end, start))
9496 		am = end - start;
9497 	else
9498 		am = 0;
9499 	if ((rack->rc_last_tlp_acked_set ) &&
9500 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
9501 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
9502 		/*
9503 		 * The DSACK is because of a TLP which we don't
9504 		 * do anything with the reordering window over since
9505 		 * it was not reordering that caused the DSACK but
9506 		 * our previous retransmit TLP.
9507 		 */
9508 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
9509 		was_tlp = 1;
9510 		goto skip_dsack_round;
9511 	}
9512 	if (rack->rc_last_sent_tlp_seq_valid) {
9513 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
9514 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
9515 		    (SEQ_LEQ(end, l_end))) {
9516 			/*
9517 			 * This dsack is from the last sent TLP, ignore it
9518 			 * for reordering purposes.
9519 			 */
9520 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
9521 			was_tlp = 1;
9522 			goto skip_dsack_round;
9523 		}
9524 	}
9525 	if (rack->rc_dsack_round_seen == 0) {
9526 		rack->rc_dsack_round_seen = 1;
9527 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
9528 		rack->r_ctl.num_dsack++;
9529 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
9530 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
9531 	}
9532 skip_dsack_round:
9533 	/*
9534 	 * We keep track of how many DSACK blocks we get
9535 	 * after a recovery incident.
9536 	 */
9537 	rack->r_ctl.dsack_byte_cnt += am;
9538 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
9539 	    rack->r_ctl.retran_during_recovery &&
9540 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
9541 		/*
9542 		 * False recovery most likely culprit is reordering. If
9543 		 * nothing else is missing we need to revert.
9544 		 */
9545 		rack->r_might_revert = 1;
9546 		rack_handle_might_revert(rack->rc_tp, rack);
9547 		rack->r_might_revert = 0;
9548 		rack->r_ctl.retran_during_recovery = 0;
9549 		rack->r_ctl.dsack_byte_cnt = 0;
9550 	}
9551 	return (was_tlp);
9552 }
9553 
9554 static void
9555 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9556 {
9557 	/* Deal with changed and PRR here (in recovery only) */
9558 	uint32_t pipe, snd_una;
9559 
9560 	rack->r_ctl.rc_prr_delivered += changed;
9561 
9562 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9563 		/*
9564 		 * It is all outstanding, we are application limited
9565 		 * and thus we don't need more room to send anything.
9566 		 * Note we use tp->snd_una here and not th_ack because
9567 		 * the data as yet not been cut from the sb.
9568 		 */
9569 		rack->r_ctl.rc_prr_sndcnt = 0;
9570 		return;
9571 	}
9572 	/* Compute prr_sndcnt */
9573 	if (SEQ_GT(tp->snd_una, th_ack)) {
9574 		snd_una = tp->snd_una;
9575 	} else {
9576 		snd_una = th_ack;
9577 	}
9578 	pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt;
9579 	if (pipe > tp->snd_ssthresh) {
9580 		long sndcnt;
9581 
9582 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9583 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
9584 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9585 		else {
9586 			rack->r_ctl.rc_prr_sndcnt = 0;
9587 			rack_log_to_prr(rack, 9, 0);
9588 			sndcnt = 0;
9589 		}
9590 		sndcnt++;
9591 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9592 			sndcnt -= rack->r_ctl.rc_prr_out;
9593 		else
9594 			sndcnt = 0;
9595 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
9596 		rack_log_to_prr(rack, 10, 0);
9597 	} else {
9598 		uint32_t limit;
9599 
9600 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9601 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9602 		else
9603 			limit = 0;
9604 		if (changed > limit)
9605 			limit = changed;
9606 		limit += ctf_fixed_maxseg(tp);
9607 		if (tp->snd_ssthresh > pipe) {
9608 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9609 			rack_log_to_prr(rack, 11, 0);
9610 		} else {
9611 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9612 			rack_log_to_prr(rack, 12, 0);
9613 		}
9614 	}
9615 }
9616 
9617 static void
9618 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9619 {
9620 	uint32_t changed;
9621 	struct tcp_rack *rack;
9622 	struct rack_sendmap *rsm;
9623 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9624 	register uint32_t th_ack;
9625 	int32_t i, j, k, num_sack_blks = 0;
9626 	uint32_t cts, acked, ack_point, sack_changed = 0;
9627 	int loop_start = 0, moved_two = 0;
9628 	uint32_t tsused;
9629 
9630 
9631 	INP_WLOCK_ASSERT(tp->t_inpcb);
9632 	if (th->th_flags & TH_RST) {
9633 		/* We don't log resets */
9634 		return;
9635 	}
9636 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9637 	cts = tcp_get_usecs(NULL);
9638 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9639 	changed = 0;
9640 	th_ack = th->th_ack;
9641 	if (rack->sack_attack_disable == 0)
9642 		rack_do_decay(rack);
9643 	if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9644 		/*
9645 		 * You only get credit for
9646 		 * MSS and greater (and you get extra
9647 		 * credit for larger cum-ack moves).
9648 		 */
9649 		int ac;
9650 
9651 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9652 		rack->r_ctl.ack_count += ac;
9653 		counter_u64_add(rack_ack_total, ac);
9654 	}
9655 	if (rack->r_ctl.ack_count > 0xfff00000) {
9656 		/*
9657 		 * reduce the number to keep us under
9658 		 * a uint32_t.
9659 		 */
9660 		rack->r_ctl.ack_count /= 2;
9661 		rack->r_ctl.sack_count /= 2;
9662 	}
9663 	if (SEQ_GT(th_ack, tp->snd_una)) {
9664 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9665 		tp->t_acktime = ticks;
9666 	}
9667 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
9668 		changed = th_ack - rsm->r_start;
9669 	if (changed) {
9670 		rack_process_to_cumack(tp, rack, th_ack, cts, to);
9671 	}
9672 	if ((to->to_flags & TOF_SACK) == 0) {
9673 		/* We are done nothing left and no sack. */
9674 		rack_handle_might_revert(tp, rack);
9675 		/*
9676 		 * For cases where we struck a dup-ack
9677 		 * with no SACK, add to the changes so
9678 		 * PRR will work right.
9679 		 */
9680 		if (dup_ack_struck && (changed == 0)) {
9681 			changed += ctf_fixed_maxseg(rack->rc_tp);
9682 		}
9683 		goto out;
9684 	}
9685 	/* Sack block processing */
9686 	if (SEQ_GT(th_ack, tp->snd_una))
9687 		ack_point = th_ack;
9688 	else
9689 		ack_point = tp->snd_una;
9690 	for (i = 0; i < to->to_nsacks; i++) {
9691 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
9692 		      &sack, sizeof(sack));
9693 		sack.start = ntohl(sack.start);
9694 		sack.end = ntohl(sack.end);
9695 		if (SEQ_GT(sack.end, sack.start) &&
9696 		    SEQ_GT(sack.start, ack_point) &&
9697 		    SEQ_LT(sack.start, tp->snd_max) &&
9698 		    SEQ_GT(sack.end, ack_point) &&
9699 		    SEQ_LEQ(sack.end, tp->snd_max)) {
9700 			sack_blocks[num_sack_blks] = sack;
9701 			num_sack_blks++;
9702 		} else if (SEQ_LEQ(sack.start, th_ack) &&
9703 			   SEQ_LEQ(sack.end, th_ack)) {
9704 			int was_tlp;
9705 
9706 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
9707 			/*
9708 			 * Its a D-SACK block.
9709 			 */
9710 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
9711 		}
9712 	}
9713 	if (rack->rc_dsack_round_seen) {
9714 		/* Is the dsack roound over? */
9715 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
9716 			/* Yes it is */
9717 			rack->rc_dsack_round_seen = 0;
9718 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
9719 		}
9720 	}
9721 	/*
9722 	 * Sort the SACK blocks so we can update the rack scoreboard with
9723 	 * just one pass.
9724 	 */
9725 	num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9726 					 num_sack_blks, th->th_ack);
9727 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9728 	if (num_sack_blks == 0) {
9729 		/* Nothing to sack (DSACKs?) */
9730 		goto out_with_totals;
9731 	}
9732 	if (num_sack_blks < 2) {
9733 		/* Only one, we don't need to sort */
9734 		goto do_sack_work;
9735 	}
9736 	/* Sort the sacks */
9737 	for (i = 0; i < num_sack_blks; i++) {
9738 		for (j = i + 1; j < num_sack_blks; j++) {
9739 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9740 				sack = sack_blocks[i];
9741 				sack_blocks[i] = sack_blocks[j];
9742 				sack_blocks[j] = sack;
9743 			}
9744 		}
9745 	}
9746 	/*
9747 	 * Now are any of the sack block ends the same (yes some
9748 	 * implementations send these)?
9749 	 */
9750 again:
9751 	if (num_sack_blks == 0)
9752 		goto out_with_totals;
9753 	if (num_sack_blks > 1) {
9754 		for (i = 0; i < num_sack_blks; i++) {
9755 			for (j = i + 1; j < num_sack_blks; j++) {
9756 				if (sack_blocks[i].end == sack_blocks[j].end) {
9757 					/*
9758 					 * Ok these two have the same end we
9759 					 * want the smallest end and then
9760 					 * throw away the larger and start
9761 					 * again.
9762 					 */
9763 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9764 						/*
9765 						 * The second block covers
9766 						 * more area use that
9767 						 */
9768 						sack_blocks[i].start = sack_blocks[j].start;
9769 					}
9770 					/*
9771 					 * Now collapse out the dup-sack and
9772 					 * lower the count
9773 					 */
9774 					for (k = (j + 1); k < num_sack_blks; k++) {
9775 						sack_blocks[j].start = sack_blocks[k].start;
9776 						sack_blocks[j].end = sack_blocks[k].end;
9777 						j++;
9778 					}
9779 					num_sack_blks--;
9780 					goto again;
9781 				}
9782 			}
9783 		}
9784 	}
9785 do_sack_work:
9786 	/*
9787 	 * First lets look to see if
9788 	 * we have retransmitted and
9789 	 * can use the transmit next?
9790 	 */
9791 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9792 	if (rsm &&
9793 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9794 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9795 		/*
9796 		 * We probably did the FR and the next
9797 		 * SACK in continues as we would expect.
9798 		 */
9799 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9800 		if (acked) {
9801 			rack->r_wanted_output = 1;
9802 			changed += acked;
9803 			sack_changed += acked;
9804 		}
9805 		if (num_sack_blks == 1) {
9806 			/*
9807 			 * This is what we would expect from
9808 			 * a normal implementation to happen
9809 			 * after we have retransmitted the FR,
9810 			 * i.e the sack-filter pushes down
9811 			 * to 1 block and the next to be retransmitted
9812 			 * is the sequence in the sack block (has more
9813 			 * are acked). Count this as ACK'd data to boost
9814 			 * up the chances of recovering any false positives.
9815 			 */
9816 			rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9817 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9818 			counter_u64_add(rack_express_sack, 1);
9819 			if (rack->r_ctl.ack_count > 0xfff00000) {
9820 				/*
9821 				 * reduce the number to keep us under
9822 				 * a uint32_t.
9823 				 */
9824 				rack->r_ctl.ack_count /= 2;
9825 				rack->r_ctl.sack_count /= 2;
9826 			}
9827 			goto out_with_totals;
9828 		} else {
9829 			/*
9830 			 * Start the loop through the
9831 			 * rest of blocks, past the first block.
9832 			 */
9833 			moved_two = 0;
9834 			loop_start = 1;
9835 		}
9836 	}
9837 	/* Its a sack of some sort */
9838 	rack->r_ctl.sack_count++;
9839 	if (rack->r_ctl.sack_count > 0xfff00000) {
9840 		/*
9841 		 * reduce the number to keep us under
9842 		 * a uint32_t.
9843 		 */
9844 		rack->r_ctl.ack_count /= 2;
9845 		rack->r_ctl.sack_count /= 2;
9846 	}
9847 	counter_u64_add(rack_sack_total, 1);
9848 	if (rack->sack_attack_disable) {
9849 		/* An attacker disablement is in place */
9850 		if (num_sack_blks > 1) {
9851 			rack->r_ctl.sack_count += (num_sack_blks - 1);
9852 			rack->r_ctl.sack_moved_extra++;
9853 			counter_u64_add(rack_move_some, 1);
9854 			if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9855 				rack->r_ctl.sack_moved_extra /= 2;
9856 				rack->r_ctl.sack_noextra_move /= 2;
9857 			}
9858 		}
9859 		goto out;
9860 	}
9861 	rsm = rack->r_ctl.rc_sacklast;
9862 	for (i = loop_start; i < num_sack_blks; i++) {
9863 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9864 		if (acked) {
9865 			rack->r_wanted_output = 1;
9866 			changed += acked;
9867 			sack_changed += acked;
9868 		}
9869 		if (moved_two) {
9870 			/*
9871 			 * If we did not get a SACK for at least a MSS and
9872 			 * had to move at all, or if we moved more than our
9873 			 * threshold, it counts against the "extra" move.
9874 			 */
9875 			rack->r_ctl.sack_moved_extra += moved_two;
9876 			counter_u64_add(rack_move_some, 1);
9877 		} else {
9878 			/*
9879 			 * else we did not have to move
9880 			 * any more than we would expect.
9881 			 */
9882 			rack->r_ctl.sack_noextra_move++;
9883 			counter_u64_add(rack_move_none, 1);
9884 		}
9885 		if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9886 			/*
9887 			 * If the SACK was not a full MSS then
9888 			 * we add to sack_count the number of
9889 			 * MSS's (or possibly more than
9890 			 * a MSS if its a TSO send) we had to skip by.
9891 			 */
9892 			rack->r_ctl.sack_count += moved_two;
9893 			counter_u64_add(rack_sack_total, moved_two);
9894 		}
9895 		/*
9896 		 * Now we need to setup for the next
9897 		 * round. First we make sure we won't
9898 		 * exceed the size of our uint32_t on
9899 		 * the various counts, and then clear out
9900 		 * moved_two.
9901 		 */
9902 		if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9903 		    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9904 			rack->r_ctl.sack_moved_extra /= 2;
9905 			rack->r_ctl.sack_noextra_move /= 2;
9906 		}
9907 		if (rack->r_ctl.sack_count > 0xfff00000) {
9908 			rack->r_ctl.ack_count /= 2;
9909 			rack->r_ctl.sack_count /= 2;
9910 		}
9911 		moved_two = 0;
9912 	}
9913 out_with_totals:
9914 	if (num_sack_blks > 1) {
9915 		/*
9916 		 * You get an extra stroke if
9917 		 * you have more than one sack-blk, this
9918 		 * could be where we are skipping forward
9919 		 * and the sack-filter is still working, or
9920 		 * it could be an attacker constantly
9921 		 * moving us.
9922 		 */
9923 		rack->r_ctl.sack_moved_extra++;
9924 		counter_u64_add(rack_move_some, 1);
9925 	}
9926 out:
9927 #ifdef NETFLIX_EXP_DETECTION
9928 	rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9929 #endif
9930 	if (changed) {
9931 		/* Something changed cancel the rack timer */
9932 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9933 	}
9934 	tsused = tcp_get_usecs(NULL);
9935 	rsm = tcp_rack_output(tp, rack, tsused);
9936 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9937 	    rsm) {
9938 		/* Enter recovery */
9939 		rack->r_ctl.rc_rsm_start = rsm->r_start;
9940 		rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
9941 		rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
9942 		entered_recovery = 1;
9943 		rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
9944 		/*
9945 		 * When we enter recovery we need to assure we send
9946 		 * one packet.
9947 		 */
9948 		if (rack->rack_no_prr == 0) {
9949 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9950 			rack_log_to_prr(rack, 8, 0);
9951 		}
9952 		rack->r_timer_override = 1;
9953 		rack->r_early = 0;
9954 		rack->r_ctl.rc_agg_early = 0;
9955 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
9956 		   rsm &&
9957 		   (rack->r_rr_config == 3)) {
9958 		/*
9959 		 * Assure we can output and we get no
9960 		 * remembered pace time except the retransmit.
9961 		 */
9962 		rack->r_timer_override = 1;
9963 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9964 		rack->r_ctl.rc_resend = rsm;
9965 	}
9966 	if (IN_FASTRECOVERY(tp->t_flags) &&
9967 	    (rack->rack_no_prr == 0) &&
9968 	    (entered_recovery == 0)) {
9969 		rack_update_prr(tp, rack, changed, th_ack);
9970 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9971 		     ((tcp_in_hpts(rack->rc_inp) == 0) &&
9972 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9973 			/*
9974 			 * If you are pacing output you don't want
9975 			 * to override.
9976 			 */
9977 			rack->r_early = 0;
9978 			rack->r_ctl.rc_agg_early = 0;
9979 			rack->r_timer_override = 1;
9980 		}
9981 	}
9982 }
9983 
9984 static void
9985 rack_strike_dupack(struct tcp_rack *rack)
9986 {
9987 	struct rack_sendmap *rsm;
9988 
9989 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9990 	while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9991 		rsm = TAILQ_NEXT(rsm, r_tnext);
9992 	}
9993 	if (rsm && (rsm->r_dupack < 0xff)) {
9994 		rsm->r_dupack++;
9995 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9996 			struct timeval tv;
9997 			uint32_t cts;
9998 			/*
9999 			 * Here we see if we need to retransmit. For
10000 			 * a SACK type connection if enough time has passed
10001 			 * we will get a return of the rsm. For a non-sack
10002 			 * connection we will get the rsm returned if the
10003 			 * dupack value is 3 or more.
10004 			 */
10005 			cts = tcp_get_usecs(&tv);
10006 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
10007 			if (rack->r_ctl.rc_resend != NULL) {
10008 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
10009 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
10010 							 rack->rc_tp->snd_una);
10011 				}
10012 				rack->r_wanted_output = 1;
10013 				rack->r_timer_override = 1;
10014 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
10015 			}
10016 		} else {
10017 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
10018 		}
10019 	}
10020 }
10021 
10022 static void
10023 rack_check_bottom_drag(struct tcpcb *tp,
10024 		       struct tcp_rack *rack,
10025 		       struct socket *so, int32_t acked)
10026 {
10027 	uint32_t segsiz, minseg;
10028 
10029 	segsiz = ctf_fixed_maxseg(tp);
10030 	minseg = segsiz;
10031 
10032 	if (tp->snd_max == tp->snd_una) {
10033 		/*
10034 		 * We are doing dynamic pacing and we are way
10035 		 * under. Basically everything got acked while
10036 		 * we were still waiting on the pacer to expire.
10037 		 *
10038 		 * This means we need to boost the b/w in
10039 		 * addition to any earlier boosting of
10040 		 * the multipler.
10041 		 */
10042 		rack->rc_dragged_bottom = 1;
10043 		rack_validate_multipliers_at_or_above100(rack);
10044 		/*
10045 		 * Lets use the segment bytes acked plus
10046 		 * the lowest RTT seen as the basis to
10047 		 * form a b/w estimate. This will be off
10048 		 * due to the fact that the true estimate
10049 		 * should be around 1/2 the time of the RTT
10050 		 * but we can settle for that.
10051 		 */
10052 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
10053 		    acked) {
10054 			uint64_t bw, calc_bw, rtt;
10055 
10056 			rtt = rack->r_ctl.rack_rs.rs_us_rtt;
10057 			if (rtt == 0) {
10058 				/* no us sample is there a ms one? */
10059 				if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
10060 					rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
10061 				} else {
10062 					goto no_measurement;
10063 				}
10064 			}
10065 			bw = acked;
10066 			calc_bw = bw * 1000000;
10067 			calc_bw /= rtt;
10068 			if (rack->r_ctl.last_max_bw &&
10069 			    (rack->r_ctl.last_max_bw < calc_bw)) {
10070 				/*
10071 				 * If we have a last calculated max bw
10072 				 * enforce it.
10073 				 */
10074 				calc_bw = rack->r_ctl.last_max_bw;
10075 			}
10076 			/* now plop it in */
10077 			if (rack->rc_gp_filled == 0) {
10078 				if (calc_bw > ONE_POINT_TWO_MEG) {
10079 					/*
10080 					 * If we have no measurement
10081 					 * don't let us set in more than
10082 					 * 1.2Mbps. If we are still too
10083 					 * low after pacing with this we
10084 					 * will hopefully have a max b/w
10085 					 * available to sanity check things.
10086 					 */
10087 					calc_bw = ONE_POINT_TWO_MEG;
10088 				}
10089 				rack->r_ctl.rc_rtt_diff = 0;
10090 				rack->r_ctl.gp_bw = calc_bw;
10091 				rack->rc_gp_filled = 1;
10092 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
10093 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
10094 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
10095 			} else if (calc_bw > rack->r_ctl.gp_bw) {
10096 				rack->r_ctl.rc_rtt_diff = 0;
10097 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
10098 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
10099 				rack->r_ctl.gp_bw = calc_bw;
10100 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
10101 			} else
10102 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
10103 			if ((rack->gp_ready == 0) &&
10104 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
10105 				/* We have enough measurements now */
10106 				rack->gp_ready = 1;
10107 				rack_set_cc_pacing(rack);
10108 				if (rack->defer_options)
10109 					rack_apply_deferred_options(rack);
10110 			}
10111 			/*
10112 			 * For acks over 1mss we do a extra boost to simulate
10113 			 * where we would get 2 acks (we want 110 for the mul).
10114 			 */
10115 			if (acked > segsiz)
10116 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
10117 		} else {
10118 			/*
10119 			 * zero rtt possibly?, settle for just an old increase.
10120 			 */
10121 no_measurement:
10122 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
10123 		}
10124 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
10125 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
10126 					       minseg)) &&
10127 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
10128 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
10129 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
10130 		    (segsiz * rack_req_segs))) {
10131 		/*
10132 		 * We are doing dynamic GP pacing and
10133 		 * we have everything except 1MSS or less
10134 		 * bytes left out. We are still pacing away.
10135 		 * And there is data that could be sent, This
10136 		 * means we are inserting delayed ack time in
10137 		 * our measurements because we are pacing too slow.
10138 		 */
10139 		rack_validate_multipliers_at_or_above100(rack);
10140 		rack->rc_dragged_bottom = 1;
10141 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
10142 	}
10143 }
10144 
10145 
10146 
10147 static void
10148 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
10149 {
10150 	/*
10151 	 * The fast output path is enabled and we
10152 	 * have moved the cumack forward. Lets see if
10153 	 * we can expand forward the fast path length by
10154 	 * that amount. What we would ideally like to
10155 	 * do is increase the number of bytes in the
10156 	 * fast path block (left_to_send) by the
10157 	 * acked amount. However we have to gate that
10158 	 * by two factors:
10159 	 * 1) The amount outstanding and the rwnd of the peer
10160 	 *    (i.e. we don't want to exceed the rwnd of the peer).
10161 	 *    <and>
10162 	 * 2) The amount of data left in the socket buffer (i.e.
10163 	 *    we can't send beyond what is in the buffer).
10164 	 *
10165 	 * Note that this does not take into account any increase
10166 	 * in the cwnd. We will only extend the fast path by
10167 	 * what was acked.
10168 	 */
10169 	uint32_t new_total, gating_val;
10170 
10171 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
10172 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
10173 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
10174 	if (new_total <= gating_val) {
10175 		/* We can increase left_to_send by the acked amount */
10176 		counter_u64_add(rack_extended_rfo, 1);
10177 		rack->r_ctl.fsb.left_to_send = new_total;
10178 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
10179 			("rack:%p left_to_send:%u sbavail:%u out:%u",
10180 			 rack, rack->r_ctl.fsb.left_to_send,
10181 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
10182 			 (tp->snd_max - tp->snd_una)));
10183 
10184 	}
10185 }
10186 
10187 static void
10188 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
10189 {
10190 	/*
10191 	 * Here any sendmap entry that points to the
10192 	 * beginning mbuf must be adjusted to the correct
10193 	 * offset. This must be called with:
10194 	 * 1) The socket buffer locked
10195 	 * 2) snd_una adjusted to its new postion.
10196 	 *
10197 	 * Note that (2) implies rack_ack_received has also
10198 	 * been called.
10199 	 *
10200 	 * We grab the first mbuf in the socket buffer and
10201 	 * then go through the front of the sendmap, recalculating
10202 	 * the stored offset for any sendmap entry that has
10203 	 * that mbuf. We must use the sb functions to do this
10204 	 * since its possible an add was done has well as
10205 	 * the subtraction we may have just completed. This should
10206 	 * not be a penalty though, since we just referenced the sb
10207 	 * to go in and trim off the mbufs that we freed (of course
10208 	 * there will be a penalty for the sendmap references though).
10209 	 */
10210 	struct mbuf *m;
10211 	struct rack_sendmap *rsm;
10212 
10213 	SOCKBUF_LOCK_ASSERT(sb);
10214 	m = sb->sb_mb;
10215 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
10216 	if ((rsm == NULL) || (m == NULL)) {
10217 		/* Nothing outstanding */
10218 		return;
10219 	}
10220 	while (rsm->m && (rsm->m == m)) {
10221 		/* one to adjust */
10222 #ifdef INVARIANTS
10223 		struct mbuf *tm;
10224 		uint32_t soff;
10225 
10226 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
10227 		if (rsm->orig_m_len != m->m_len) {
10228 			rack_adjust_orig_mlen(rsm);
10229 		}
10230 		if (rsm->soff != soff) {
10231 			/*
10232 			 * This is not a fatal error, we anticipate it
10233 			 * might happen (the else code), so we count it here
10234 			 * so that under invariant we can see that it really
10235 			 * does happen.
10236 			 */
10237 			counter_u64_add(rack_adjust_map_bw, 1);
10238 		}
10239 		rsm->m = tm;
10240 		rsm->soff = soff;
10241 		if (tm)
10242 			rsm->orig_m_len = rsm->m->m_len;
10243 		else
10244 			rsm->orig_m_len = 0;
10245 #else
10246 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
10247 		if (rsm->m)
10248 			rsm->orig_m_len = rsm->m->m_len;
10249 		else
10250 			rsm->orig_m_len = 0;
10251 #endif
10252 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
10253 			      rsm);
10254 		if (rsm == NULL)
10255 			break;
10256 	}
10257 }
10258 
10259 /*
10260  * Return value of 1, we do not need to call rack_process_data().
10261  * return value of 0, rack_process_data can be called.
10262  * For ret_val if its 0 the TCP is locked, if its non-zero
10263  * its unlocked and probably unsafe to touch the TCB.
10264  */
10265 static int
10266 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10267     struct tcpcb *tp, struct tcpopt *to,
10268     uint32_t tiwin, int32_t tlen,
10269     int32_t * ofia, int32_t thflags, int32_t *ret_val)
10270 {
10271 	int32_t ourfinisacked = 0;
10272 	int32_t nsegs, acked_amount;
10273 	int32_t acked;
10274 	struct mbuf *mfree;
10275 	struct tcp_rack *rack;
10276 	int32_t under_pacing = 0;
10277 	int32_t recovery = 0;
10278 
10279 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10280 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
10281 		__ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
10282 				      &rack->r_ctl.challenge_ack_ts,
10283 				      &rack->r_ctl.challenge_ack_cnt);
10284 		rack->r_wanted_output = 1;
10285 		return (1);
10286 	}
10287 	if (rack->gp_ready &&
10288 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10289 		under_pacing = 1;
10290 	}
10291 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
10292 		int in_rec, dup_ack_struck = 0;
10293 
10294 		in_rec = IN_FASTRECOVERY(tp->t_flags);
10295 		if (rack->rc_in_persist) {
10296 			tp->t_rxtshift = 0;
10297 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10298 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10299 		}
10300 		if ((th->th_ack == tp->snd_una) &&
10301 		    (tiwin == tp->snd_wnd) &&
10302 		    ((to->to_flags & TOF_SACK) == 0)) {
10303 			rack_strike_dupack(rack);
10304 			dup_ack_struck = 1;
10305 		}
10306 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
10307 	}
10308 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10309 		/*
10310 		 * Old ack, behind (or duplicate to) the last one rcv'd
10311 		 * Note: We mark reordering is occuring if its
10312 		 * less than and we have not closed our window.
10313 		 */
10314 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
10315 			counter_u64_add(rack_reorder_seen, 1);
10316 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
10317 		}
10318 		return (0);
10319 	}
10320 	/*
10321 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
10322 	 * something we sent.
10323 	 */
10324 	if (tp->t_flags & TF_NEEDSYN) {
10325 		/*
10326 		 * T/TCP: Connection was half-synchronized, and our SYN has
10327 		 * been ACK'd (so connection is now fully synchronized).  Go
10328 		 * to non-starred state, increment snd_una for ACK of SYN,
10329 		 * and check if we can do window scaling.
10330 		 */
10331 		tp->t_flags &= ~TF_NEEDSYN;
10332 		tp->snd_una++;
10333 		/* Do window scaling? */
10334 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10335 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10336 			tp->rcv_scale = tp->request_r_scale;
10337 			/* Send window already scaled. */
10338 		}
10339 	}
10340 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10341 	INP_WLOCK_ASSERT(tp->t_inpcb);
10342 
10343 	acked = BYTES_THIS_ACK(tp, th);
10344 	if (acked) {
10345 		/*
10346 		 * Any time we move the cum-ack forward clear
10347 		 * keep-alive tied probe-not-answered. The
10348 		 * persists clears its own on entry.
10349 		 */
10350 		rack->probe_not_answered = 0;
10351 	}
10352 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10353 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10354 	/*
10355 	 * If we just performed our first retransmit, and the ACK arrives
10356 	 * within our recovery window, then it was a mistake to do the
10357 	 * retransmit in the first place.  Recover our original cwnd and
10358 	 * ssthresh, and proceed to transmit where we left off.
10359 	 */
10360 	if ((tp->t_flags & TF_PREVVALID) &&
10361 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10362 		tp->t_flags &= ~TF_PREVVALID;
10363 		if (tp->t_rxtshift == 1 &&
10364 		    (int)(ticks - tp->t_badrxtwin) < 0)
10365 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
10366 	}
10367 	if (acked) {
10368 		/* assure we are not backed off */
10369 		tp->t_rxtshift = 0;
10370 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10371 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10372 		rack->rc_tlp_in_progress = 0;
10373 		rack->r_ctl.rc_tlp_cnt_out = 0;
10374 		/*
10375 		 * If it is the RXT timer we want to
10376 		 * stop it, so we can restart a TLP.
10377 		 */
10378 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10379 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10380 #ifdef NETFLIX_HTTP_LOGGING
10381 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10382 #endif
10383 	}
10384 	/*
10385 	 * If we have a timestamp reply, update smoothed round trip time. If
10386 	 * no timestamp is present but transmit timer is running and timed
10387 	 * sequence number was acked, update smoothed round trip time. Since
10388 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
10389 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
10390 	 * timer.
10391 	 *
10392 	 * Some boxes send broken timestamp replies during the SYN+ACK
10393 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
10394 	 * and blow up the retransmit timer.
10395 	 */
10396 	/*
10397 	 * If all outstanding data is acked, stop retransmit timer and
10398 	 * remember to restart (more output or persist). If there is more
10399 	 * data to be acked, restart retransmit timer, using current
10400 	 * (possibly backed-off) value.
10401 	 */
10402 	if (acked == 0) {
10403 		if (ofia)
10404 			*ofia = ourfinisacked;
10405 		return (0);
10406 	}
10407 	if (IN_RECOVERY(tp->t_flags)) {
10408 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
10409 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
10410 			tcp_rack_partialack(tp);
10411 		} else {
10412 			rack_post_recovery(tp, th->th_ack);
10413 			recovery = 1;
10414 		}
10415 	}
10416 	/*
10417 	 * Let the congestion control algorithm update congestion control
10418 	 * related information. This typically means increasing the
10419 	 * congestion window.
10420 	 */
10421 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
10422 	SOCKBUF_LOCK(&so->so_snd);
10423 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
10424 	tp->snd_wnd -= acked_amount;
10425 	mfree = sbcut_locked(&so->so_snd, acked_amount);
10426 	if ((sbused(&so->so_snd) == 0) &&
10427 	    (acked > acked_amount) &&
10428 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
10429 	    (tp->t_flags & TF_SENTFIN)) {
10430 		/*
10431 		 * We must be sure our fin
10432 		 * was sent and acked (we can be
10433 		 * in FIN_WAIT_1 without having
10434 		 * sent the fin).
10435 		 */
10436 		ourfinisacked = 1;
10437 	}
10438 	tp->snd_una = th->th_ack;
10439 	if (acked_amount && sbavail(&so->so_snd))
10440 		rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10441 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10442 	/* NB: sowwakeup_locked() does an implicit unlock. */
10443 	sowwakeup_locked(so);
10444 	m_freem(mfree);
10445 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
10446 		tp->snd_recover = tp->snd_una;
10447 
10448 	if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
10449 		tp->snd_nxt = tp->snd_una;
10450 	}
10451 	if (under_pacing &&
10452 	    (rack->use_fixed_rate == 0) &&
10453 	    (rack->in_probe_rtt == 0) &&
10454 	    rack->rc_gp_dyn_mul &&
10455 	    rack->rc_always_pace) {
10456 		/* Check if we are dragging bottom */
10457 		rack_check_bottom_drag(tp, rack, so, acked);
10458 	}
10459 	if (tp->snd_una == tp->snd_max) {
10460 		/* Nothing left outstanding */
10461 		tp->t_flags &= ~TF_PREVVALID;
10462 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10463 		rack->r_ctl.retran_during_recovery = 0;
10464 		rack->r_ctl.dsack_byte_cnt = 0;
10465 		if (rack->r_ctl.rc_went_idle_time == 0)
10466 			rack->r_ctl.rc_went_idle_time = 1;
10467 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10468 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
10469 			tp->t_acktime = 0;
10470 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10471 		/* Set need output so persist might get set */
10472 		rack->r_wanted_output = 1;
10473 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10474 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
10475 		    (sbavail(&so->so_snd) == 0) &&
10476 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
10477 			/*
10478 			 * The socket was gone and the
10479 			 * peer sent data (now or in the past), time to
10480 			 * reset him.
10481 			 */
10482 			*ret_val = 1;
10483 			/* tcp_close will kill the inp pre-log the Reset */
10484 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
10485 			tp = tcp_close(tp);
10486 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
10487 			return (1);
10488 		}
10489 	}
10490 	if (ofia)
10491 		*ofia = ourfinisacked;
10492 	return (0);
10493 }
10494 
10495 static void
10496 rack_collapsed_window(struct tcp_rack *rack)
10497 {
10498 	/*
10499 	 * Now we must walk the
10500 	 * send map and divide the
10501 	 * ones left stranded. These
10502 	 * guys can't cause us to abort
10503 	 * the connection and are really
10504 	 * "unsent". However if a buggy
10505 	 * client actually did keep some
10506 	 * of the data i.e. collapsed the win
10507 	 * and refused to ack and then opened
10508 	 * the win and acked that data. We would
10509 	 * get into an ack war, the simplier
10510 	 * method then of just pretending we
10511 	 * did not send those segments something
10512 	 * won't work.
10513 	 */
10514 	struct rack_sendmap *rsm, *nrsm, fe, *insret;
10515 	tcp_seq max_seq;
10516 
10517 	max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
10518 	memset(&fe, 0, sizeof(fe));
10519 	fe.r_start = max_seq;
10520 	/* Find the first seq past or at maxseq */
10521 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
10522 	if (rsm == NULL) {
10523 		/* Nothing to do strange */
10524 		rack->rc_has_collapsed = 0;
10525 		return;
10526 	}
10527 	/*
10528 	 * Now do we need to split at
10529 	 * the collapse point?
10530 	 */
10531 	if (SEQ_GT(max_seq, rsm->r_start)) {
10532 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10533 		if (nrsm == NULL) {
10534 			/* We can't get a rsm, mark all? */
10535 			nrsm = rsm;
10536 			goto no_split;
10537 		}
10538 		/* Clone it */
10539 		rack_clone_rsm(rack, nrsm, rsm, max_seq);
10540 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10541 #ifdef INVARIANTS
10542 		if (insret != NULL) {
10543 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
10544 			      nrsm, insret, rack, rsm);
10545 		}
10546 #endif
10547 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT, max_seq, __LINE__);
10548 		if (rsm->r_in_tmap) {
10549 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10550 			nrsm->r_in_tmap = 1;
10551 		}
10552 		/*
10553 		 * Set in the new RSM as the
10554 		 * collapsed starting point
10555 		 */
10556 		rsm = nrsm;
10557 	}
10558 no_split:
10559 	counter_u64_add(rack_collapsed_win, 1);
10560 	RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
10561 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
10562 	}
10563 	rack->rc_has_collapsed = 1;
10564 }
10565 
10566 static void
10567 rack_un_collapse_window(struct tcp_rack *rack)
10568 {
10569 	struct rack_sendmap *rsm;
10570 
10571 	RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
10572 		if (rsm->r_flags & RACK_RWND_COLLAPSED)
10573 			rsm->r_flags &= ~RACK_RWND_COLLAPSED;
10574 		else
10575 			break;
10576 	}
10577 	rack->rc_has_collapsed = 0;
10578 }
10579 
10580 static void
10581 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10582 			int32_t tlen, int32_t tfo_syn)
10583 {
10584 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
10585 		if (rack->rc_dack_mode &&
10586 		    (tlen > 500) &&
10587 		    (rack->rc_dack_toggle == 1)) {
10588 			goto no_delayed_ack;
10589 		}
10590 		rack_timer_cancel(tp, rack,
10591 				  rack->r_ctl.rc_rcvtime, __LINE__);
10592 		tp->t_flags |= TF_DELACK;
10593 	} else {
10594 no_delayed_ack:
10595 		rack->r_wanted_output = 1;
10596 		tp->t_flags |= TF_ACKNOW;
10597 		if (rack->rc_dack_mode) {
10598 			if (tp->t_flags & TF_DELACK)
10599 				rack->rc_dack_toggle = 1;
10600 			else
10601 				rack->rc_dack_toggle = 0;
10602 		}
10603 	}
10604 }
10605 
10606 static void
10607 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10608 {
10609 	/*
10610 	 * If fast output is in progress, lets validate that
10611 	 * the new window did not shrink on us and make it
10612 	 * so fast output should end.
10613 	 */
10614 	if (rack->r_fast_output) {
10615 		uint32_t out;
10616 
10617 		/*
10618 		 * Calculate what we will send if left as is
10619 		 * and compare that to our send window.
10620 		 */
10621 		out = ctf_outstanding(tp);
10622 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10623 			/* ok we have an issue */
10624 			if (out >= tp->snd_wnd) {
10625 				/* Turn off fast output the window is met or collapsed */
10626 				rack->r_fast_output = 0;
10627 			} else {
10628 				/* we have some room left */
10629 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10630 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10631 					/* If not at least 1 full segment never mind */
10632 					rack->r_fast_output = 0;
10633 				}
10634 			}
10635 		}
10636 	}
10637 }
10638 
10639 
10640 /*
10641  * Return value of 1, the TCB is unlocked and most
10642  * likely gone, return value of 0, the TCP is still
10643  * locked.
10644  */
10645 static int
10646 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10647     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10648     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10649 {
10650 	/*
10651 	 * Update window information. Don't look at window if no ACK: TAC's
10652 	 * send garbage on first SYN.
10653 	 */
10654 	int32_t nsegs;
10655 	int32_t tfo_syn;
10656 	struct tcp_rack *rack;
10657 
10658 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10659 	INP_WLOCK_ASSERT(tp->t_inpcb);
10660 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10661 	if ((thflags & TH_ACK) &&
10662 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10663 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10664 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10665 		/* keep track of pure window updates */
10666 		if (tlen == 0 &&
10667 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10668 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10669 		tp->snd_wnd = tiwin;
10670 		rack_validate_fo_sendwin_up(tp, rack);
10671 		tp->snd_wl1 = th->th_seq;
10672 		tp->snd_wl2 = th->th_ack;
10673 		if (tp->snd_wnd > tp->max_sndwnd)
10674 			tp->max_sndwnd = tp->snd_wnd;
10675 		rack->r_wanted_output = 1;
10676 	} else if (thflags & TH_ACK) {
10677 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10678 			tp->snd_wnd = tiwin;
10679 			rack_validate_fo_sendwin_up(tp, rack);
10680 			tp->snd_wl1 = th->th_seq;
10681 			tp->snd_wl2 = th->th_ack;
10682 		}
10683 	}
10684 	if (tp->snd_wnd < ctf_outstanding(tp))
10685 		/* The peer collapsed the window */
10686 		rack_collapsed_window(rack);
10687 	else if (rack->rc_has_collapsed)
10688 		rack_un_collapse_window(rack);
10689 	/* Was persist timer active and now we have window space? */
10690 	if ((rack->rc_in_persist != 0) &&
10691 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10692 				rack->r_ctl.rc_pace_min_segs))) {
10693 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10694 		tp->snd_nxt = tp->snd_max;
10695 		/* Make sure we output to start the timer */
10696 		rack->r_wanted_output = 1;
10697 	}
10698 	/* Do we enter persists? */
10699 	if ((rack->rc_in_persist == 0) &&
10700 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10701 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
10702 	    (tp->snd_max == tp->snd_una) &&
10703 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10704 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10705 		/*
10706 		 * Here the rwnd is less than
10707 		 * the pacing size, we are established,
10708 		 * nothing is outstanding, and there is
10709 		 * data to send. Enter persists.
10710 		 */
10711 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10712 	}
10713 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10714 		m_freem(m);
10715 		return (0);
10716 	}
10717 	/*
10718 	 * don't process the URG bit, ignore them drag
10719 	 * along the up.
10720 	 */
10721 	tp->rcv_up = tp->rcv_nxt;
10722 	INP_WLOCK_ASSERT(tp->t_inpcb);
10723 
10724 	/*
10725 	 * Process the segment text, merging it into the TCP sequencing
10726 	 * queue, and arranging for acknowledgment of receipt if necessary.
10727 	 * This process logically involves adjusting tp->rcv_wnd as data is
10728 	 * presented to the user (this happens in tcp_usrreq.c, case
10729 	 * PRU_RCVD).  If a FIN has already been received on this connection
10730 	 * then we just ignore the text.
10731 	 */
10732 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10733 		   IS_FASTOPEN(tp->t_flags));
10734 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10735 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10736 		tcp_seq save_start = th->th_seq;
10737 		tcp_seq save_rnxt  = tp->rcv_nxt;
10738 		int     save_tlen  = tlen;
10739 
10740 		m_adj(m, drop_hdrlen);	/* delayed header drop */
10741 		/*
10742 		 * Insert segment which includes th into TCP reassembly
10743 		 * queue with control block tp.  Set thflags to whether
10744 		 * reassembly now includes a segment with FIN.  This handles
10745 		 * the common case inline (segment is the next to be
10746 		 * received on an established connection, and the queue is
10747 		 * empty), avoiding linkage into and removal from the queue
10748 		 * and repetition of various conversions. Set DELACK for
10749 		 * segments received in order, but ack immediately when
10750 		 * segments are out of order (so fast retransmit can work).
10751 		 */
10752 		if (th->th_seq == tp->rcv_nxt &&
10753 		    SEGQ_EMPTY(tp) &&
10754 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
10755 		    tfo_syn)) {
10756 #ifdef NETFLIX_SB_LIMITS
10757 			u_int mcnt, appended;
10758 
10759 			if (so->so_rcv.sb_shlim) {
10760 				mcnt = m_memcnt(m);
10761 				appended = 0;
10762 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10763 				    CFO_NOSLEEP, NULL) == false) {
10764 					counter_u64_add(tcp_sb_shlim_fails, 1);
10765 					m_freem(m);
10766 					return (0);
10767 				}
10768 			}
10769 #endif
10770 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10771 			tp->rcv_nxt += tlen;
10772 			if (tlen &&
10773 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10774 			    (tp->t_fbyte_in == 0)) {
10775 				tp->t_fbyte_in = ticks;
10776 				if (tp->t_fbyte_in == 0)
10777 					tp->t_fbyte_in = 1;
10778 				if (tp->t_fbyte_out && tp->t_fbyte_in)
10779 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10780 			}
10781 			thflags = th->th_flags & TH_FIN;
10782 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10783 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10784 			SOCKBUF_LOCK(&so->so_rcv);
10785 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10786 				m_freem(m);
10787 			} else
10788 #ifdef NETFLIX_SB_LIMITS
10789 				appended =
10790 #endif
10791 					sbappendstream_locked(&so->so_rcv, m, 0);
10792 
10793 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10794 			/* NB: sorwakeup_locked() does an implicit unlock. */
10795 			sorwakeup_locked(so);
10796 #ifdef NETFLIX_SB_LIMITS
10797 			if (so->so_rcv.sb_shlim && appended != mcnt)
10798 				counter_fo_release(so->so_rcv.sb_shlim,
10799 				    mcnt - appended);
10800 #endif
10801 		} else {
10802 			/*
10803 			 * XXX: Due to the header drop above "th" is
10804 			 * theoretically invalid by now.  Fortunately
10805 			 * m_adj() doesn't actually frees any mbufs when
10806 			 * trimming from the head.
10807 			 */
10808 			tcp_seq temp = save_start;
10809 
10810 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
10811 			tp->t_flags |= TF_ACKNOW;
10812 			if (tp->t_flags & TF_WAKESOR) {
10813 				tp->t_flags &= ~TF_WAKESOR;
10814 				/* NB: sorwakeup_locked() does an implicit unlock. */
10815 				sorwakeup_locked(so);
10816 			}
10817 		}
10818 		if ((tp->t_flags & TF_SACK_PERMIT) &&
10819 		    (save_tlen > 0) &&
10820 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
10821 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10822 				/*
10823 				 * DSACK actually handled in the fastpath
10824 				 * above.
10825 				 */
10826 				RACK_OPTS_INC(tcp_sack_path_1);
10827 				tcp_update_sack_list(tp, save_start,
10828 				    save_start + save_tlen);
10829 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10830 				if ((tp->rcv_numsacks >= 1) &&
10831 				    (tp->sackblks[0].end == save_start)) {
10832 					/*
10833 					 * Partial overlap, recorded at todrop
10834 					 * above.
10835 					 */
10836 					RACK_OPTS_INC(tcp_sack_path_2a);
10837 					tcp_update_sack_list(tp,
10838 					    tp->sackblks[0].start,
10839 					    tp->sackblks[0].end);
10840 				} else {
10841 					RACK_OPTS_INC(tcp_sack_path_2b);
10842 					tcp_update_dsack_list(tp, save_start,
10843 					    save_start + save_tlen);
10844 				}
10845 			} else if (tlen >= save_tlen) {
10846 				/* Update of sackblks. */
10847 				RACK_OPTS_INC(tcp_sack_path_3);
10848 				tcp_update_dsack_list(tp, save_start,
10849 				    save_start + save_tlen);
10850 			} else if (tlen > 0) {
10851 				RACK_OPTS_INC(tcp_sack_path_4);
10852 				tcp_update_dsack_list(tp, save_start,
10853 				    save_start + tlen);
10854 			}
10855 		}
10856 	} else {
10857 		m_freem(m);
10858 		thflags &= ~TH_FIN;
10859 	}
10860 
10861 	/*
10862 	 * If FIN is received ACK the FIN and let the user know that the
10863 	 * connection is closing.
10864 	 */
10865 	if (thflags & TH_FIN) {
10866 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10867 			/* The socket upcall is handled by socantrcvmore. */
10868 			socantrcvmore(so);
10869 			/*
10870 			 * If connection is half-synchronized (ie NEEDSYN
10871 			 * flag on) then delay ACK, so it may be piggybacked
10872 			 * when SYN is sent. Otherwise, since we received a
10873 			 * FIN then no more input can be expected, send ACK
10874 			 * now.
10875 			 */
10876 			if (tp->t_flags & TF_NEEDSYN) {
10877 				rack_timer_cancel(tp, rack,
10878 				    rack->r_ctl.rc_rcvtime, __LINE__);
10879 				tp->t_flags |= TF_DELACK;
10880 			} else {
10881 				tp->t_flags |= TF_ACKNOW;
10882 			}
10883 			tp->rcv_nxt++;
10884 		}
10885 		switch (tp->t_state) {
10886 			/*
10887 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
10888 			 * CLOSE_WAIT state.
10889 			 */
10890 		case TCPS_SYN_RECEIVED:
10891 			tp->t_starttime = ticks;
10892 			/* FALLTHROUGH */
10893 		case TCPS_ESTABLISHED:
10894 			rack_timer_cancel(tp, rack,
10895 			    rack->r_ctl.rc_rcvtime, __LINE__);
10896 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
10897 			break;
10898 
10899 			/*
10900 			 * If still in FIN_WAIT_1 STATE FIN has not been
10901 			 * acked so enter the CLOSING state.
10902 			 */
10903 		case TCPS_FIN_WAIT_1:
10904 			rack_timer_cancel(tp, rack,
10905 			    rack->r_ctl.rc_rcvtime, __LINE__);
10906 			tcp_state_change(tp, TCPS_CLOSING);
10907 			break;
10908 
10909 			/*
10910 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
10911 			 * starting the time-wait timer, turning off the
10912 			 * other standard timers.
10913 			 */
10914 		case TCPS_FIN_WAIT_2:
10915 			rack_timer_cancel(tp, rack,
10916 			    rack->r_ctl.rc_rcvtime, __LINE__);
10917 			tcp_twstart(tp);
10918 			return (1);
10919 		}
10920 	}
10921 	/*
10922 	 * Return any desired output.
10923 	 */
10924 	if ((tp->t_flags & TF_ACKNOW) ||
10925 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10926 		rack->r_wanted_output = 1;
10927 	}
10928 	INP_WLOCK_ASSERT(tp->t_inpcb);
10929 	return (0);
10930 }
10931 
10932 /*
10933  * Here nothing is really faster, its just that we
10934  * have broken out the fast-data path also just like
10935  * the fast-ack.
10936  */
10937 static int
10938 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10939     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10940     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10941 {
10942 	int32_t nsegs;
10943 	int32_t newsize = 0;	/* automatic sockbuf scaling */
10944 	struct tcp_rack *rack;
10945 #ifdef NETFLIX_SB_LIMITS
10946 	u_int mcnt, appended;
10947 #endif
10948 #ifdef TCPDEBUG
10949 	/*
10950 	 * The size of tcp_saveipgen must be the size of the max ip header,
10951 	 * now IPv6.
10952 	 */
10953 	u_char tcp_saveipgen[IP6_HDR_LEN];
10954 	struct tcphdr tcp_savetcp;
10955 	short ostate = 0;
10956 
10957 #endif
10958 	/*
10959 	 * If last ACK falls within this segment's sequence numbers, record
10960 	 * the timestamp. NOTE that the test is modified according to the
10961 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10962 	 */
10963 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10964 		return (0);
10965 	}
10966 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10967 		return (0);
10968 	}
10969 	if (tiwin && tiwin != tp->snd_wnd) {
10970 		return (0);
10971 	}
10972 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10973 		return (0);
10974 	}
10975 	if (__predict_false((to->to_flags & TOF_TS) &&
10976 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10977 		return (0);
10978 	}
10979 	if (__predict_false((th->th_ack != tp->snd_una))) {
10980 		return (0);
10981 	}
10982 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10983 		return (0);
10984 	}
10985 	if ((to->to_flags & TOF_TS) != 0 &&
10986 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10987 		tp->ts_recent_age = tcp_ts_getticks();
10988 		tp->ts_recent = to->to_tsval;
10989 	}
10990 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10991 	/*
10992 	 * This is a pure, in-sequence data packet with nothing on the
10993 	 * reassembly queue and we have enough buffer space to take it.
10994 	 */
10995 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10996 
10997 #ifdef NETFLIX_SB_LIMITS
10998 	if (so->so_rcv.sb_shlim) {
10999 		mcnt = m_memcnt(m);
11000 		appended = 0;
11001 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
11002 		    CFO_NOSLEEP, NULL) == false) {
11003 			counter_u64_add(tcp_sb_shlim_fails, 1);
11004 			m_freem(m);
11005 			return (1);
11006 		}
11007 	}
11008 #endif
11009 	/* Clean receiver SACK report if present */
11010 	if (tp->rcv_numsacks)
11011 		tcp_clean_sackreport(tp);
11012 	KMOD_TCPSTAT_INC(tcps_preddat);
11013 	tp->rcv_nxt += tlen;
11014 	if (tlen &&
11015 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
11016 	    (tp->t_fbyte_in == 0)) {
11017 		tp->t_fbyte_in = ticks;
11018 		if (tp->t_fbyte_in == 0)
11019 			tp->t_fbyte_in = 1;
11020 		if (tp->t_fbyte_out && tp->t_fbyte_in)
11021 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
11022 	}
11023 	/*
11024 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
11025 	 */
11026 	tp->snd_wl1 = th->th_seq;
11027 	/*
11028 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
11029 	 */
11030 	tp->rcv_up = tp->rcv_nxt;
11031 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
11032 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
11033 #ifdef TCPDEBUG
11034 	if (so->so_options & SO_DEBUG)
11035 		tcp_trace(TA_INPUT, ostate, tp,
11036 		    (void *)tcp_saveipgen, &tcp_savetcp, 0);
11037 #endif
11038 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
11039 
11040 	/* Add data to socket buffer. */
11041 	SOCKBUF_LOCK(&so->so_rcv);
11042 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11043 		m_freem(m);
11044 	} else {
11045 		/*
11046 		 * Set new socket buffer size. Give up when limit is
11047 		 * reached.
11048 		 */
11049 		if (newsize)
11050 			if (!sbreserve_locked(&so->so_rcv,
11051 			    newsize, so, NULL))
11052 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
11053 		m_adj(m, drop_hdrlen);	/* delayed header drop */
11054 #ifdef NETFLIX_SB_LIMITS
11055 		appended =
11056 #endif
11057 			sbappendstream_locked(&so->so_rcv, m, 0);
11058 		ctf_calc_rwin(so, tp);
11059 	}
11060 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
11061 	/* NB: sorwakeup_locked() does an implicit unlock. */
11062 	sorwakeup_locked(so);
11063 #ifdef NETFLIX_SB_LIMITS
11064 	if (so->so_rcv.sb_shlim && mcnt != appended)
11065 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
11066 #endif
11067 	rack_handle_delayed_ack(tp, rack, tlen, 0);
11068 	if (tp->snd_una == tp->snd_max)
11069 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
11070 	return (1);
11071 }
11072 
11073 /*
11074  * This subfunction is used to try to highly optimize the
11075  * fast path. We again allow window updates that are
11076  * in sequence to remain in the fast-path. We also add
11077  * in the __predict's to attempt to help the compiler.
11078  * Note that if we return a 0, then we can *not* process
11079  * it and the caller should push the packet into the
11080  * slow-path.
11081  */
11082 static int
11083 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11084     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11085     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
11086 {
11087 	int32_t acked;
11088 	int32_t nsegs;
11089 #ifdef TCPDEBUG
11090 	/*
11091 	 * The size of tcp_saveipgen must be the size of the max ip header,
11092 	 * now IPv6.
11093 	 */
11094 	u_char tcp_saveipgen[IP6_HDR_LEN];
11095 	struct tcphdr tcp_savetcp;
11096 	short ostate = 0;
11097 #endif
11098 	int32_t under_pacing = 0;
11099 	struct tcp_rack *rack;
11100 
11101 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
11102 		/* Old ack, behind (or duplicate to) the last one rcv'd */
11103 		return (0);
11104 	}
11105 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
11106 		/* Above what we have sent? */
11107 		return (0);
11108 	}
11109 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
11110 		/* We are retransmitting */
11111 		return (0);
11112 	}
11113 	if (__predict_false(tiwin == 0)) {
11114 		/* zero window */
11115 		return (0);
11116 	}
11117 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
11118 		/* We need a SYN or a FIN, unlikely.. */
11119 		return (0);
11120 	}
11121 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
11122 		/* Timestamp is behind .. old ack with seq wrap? */
11123 		return (0);
11124 	}
11125 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
11126 		/* Still recovering */
11127 		return (0);
11128 	}
11129 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11130 	if (rack->r_ctl.rc_sacked) {
11131 		/* We have sack holes on our scoreboard */
11132 		return (0);
11133 	}
11134 	/* Ok if we reach here, we can process a fast-ack */
11135 	if (rack->gp_ready &&
11136 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11137 		under_pacing = 1;
11138 	}
11139 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
11140 	rack_log_ack(tp, to, th, 0, 0);
11141 	/* Did the window get updated? */
11142 	if (tiwin != tp->snd_wnd) {
11143 		tp->snd_wnd = tiwin;
11144 		rack_validate_fo_sendwin_up(tp, rack);
11145 		tp->snd_wl1 = th->th_seq;
11146 		if (tp->snd_wnd > tp->max_sndwnd)
11147 			tp->max_sndwnd = tp->snd_wnd;
11148 	}
11149 	/* Do we exit persists? */
11150 	if ((rack->rc_in_persist != 0) &&
11151 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
11152 			       rack->r_ctl.rc_pace_min_segs))) {
11153 		rack_exit_persist(tp, rack, cts);
11154 	}
11155 	/* Do we enter persists? */
11156 	if ((rack->rc_in_persist == 0) &&
11157 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
11158 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
11159 	    (tp->snd_max == tp->snd_una) &&
11160 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
11161 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
11162 		/*
11163 		 * Here the rwnd is less than
11164 		 * the pacing size, we are established,
11165 		 * nothing is outstanding, and there is
11166 		 * data to send. Enter persists.
11167 		 */
11168 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
11169 	}
11170 	/*
11171 	 * If last ACK falls within this segment's sequence numbers, record
11172 	 * the timestamp. NOTE that the test is modified according to the
11173 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
11174 	 */
11175 	if ((to->to_flags & TOF_TS) != 0 &&
11176 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
11177 		tp->ts_recent_age = tcp_ts_getticks();
11178 		tp->ts_recent = to->to_tsval;
11179 	}
11180 	/*
11181 	 * This is a pure ack for outstanding data.
11182 	 */
11183 	KMOD_TCPSTAT_INC(tcps_predack);
11184 
11185 	/*
11186 	 * "bad retransmit" recovery.
11187 	 */
11188 	if ((tp->t_flags & TF_PREVVALID) &&
11189 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
11190 		tp->t_flags &= ~TF_PREVVALID;
11191 		if (tp->t_rxtshift == 1 &&
11192 		    (int)(ticks - tp->t_badrxtwin) < 0)
11193 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
11194 	}
11195 	/*
11196 	 * Recalculate the transmit timer / rtt.
11197 	 *
11198 	 * Some boxes send broken timestamp replies during the SYN+ACK
11199 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
11200 	 * and blow up the retransmit timer.
11201 	 */
11202 	acked = BYTES_THIS_ACK(tp, th);
11203 
11204 #ifdef TCP_HHOOK
11205 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
11206 	hhook_run_tcp_est_in(tp, th, to);
11207 #endif
11208 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
11209 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
11210 	if (acked) {
11211 		struct mbuf *mfree;
11212 
11213 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
11214 		SOCKBUF_LOCK(&so->so_snd);
11215 		mfree = sbcut_locked(&so->so_snd, acked);
11216 		tp->snd_una = th->th_ack;
11217 		/* Note we want to hold the sb lock through the sendmap adjust */
11218 		rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
11219 		/* Wake up the socket if we have room to write more */
11220 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
11221 		sowwakeup_locked(so);
11222 		m_freem(mfree);
11223 		tp->t_rxtshift = 0;
11224 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11225 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11226 		rack->rc_tlp_in_progress = 0;
11227 		rack->r_ctl.rc_tlp_cnt_out = 0;
11228 		/*
11229 		 * If it is the RXT timer we want to
11230 		 * stop it, so we can restart a TLP.
11231 		 */
11232 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
11233 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11234 #ifdef NETFLIX_HTTP_LOGGING
11235 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
11236 #endif
11237 	}
11238 	/*
11239 	 * Let the congestion control algorithm update congestion control
11240 	 * related information. This typically means increasing the
11241 	 * congestion window.
11242 	 */
11243 	if (tp->snd_wnd < ctf_outstanding(tp)) {
11244 		/* The peer collapsed the window */
11245 		rack_collapsed_window(rack);
11246 	} else if (rack->rc_has_collapsed)
11247 		rack_un_collapse_window(rack);
11248 
11249 	/*
11250 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
11251 	 */
11252 	tp->snd_wl2 = th->th_ack;
11253 	tp->t_dupacks = 0;
11254 	m_freem(m);
11255 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
11256 
11257 	/*
11258 	 * If all outstanding data are acked, stop retransmit timer,
11259 	 * otherwise restart timer using current (possibly backed-off)
11260 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
11261 	 * If data are ready to send, let tcp_output decide between more
11262 	 * output or persist.
11263 	 */
11264 #ifdef TCPDEBUG
11265 	if (so->so_options & SO_DEBUG)
11266 		tcp_trace(TA_INPUT, ostate, tp,
11267 		    (void *)tcp_saveipgen,
11268 		    &tcp_savetcp, 0);
11269 #endif
11270 	if (under_pacing &&
11271 	    (rack->use_fixed_rate == 0) &&
11272 	    (rack->in_probe_rtt == 0) &&
11273 	    rack->rc_gp_dyn_mul &&
11274 	    rack->rc_always_pace) {
11275 		/* Check if we are dragging bottom */
11276 		rack_check_bottom_drag(tp, rack, so, acked);
11277 	}
11278 	if (tp->snd_una == tp->snd_max) {
11279 		tp->t_flags &= ~TF_PREVVALID;
11280 		rack->r_ctl.retran_during_recovery = 0;
11281 		rack->r_ctl.dsack_byte_cnt = 0;
11282 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
11283 		if (rack->r_ctl.rc_went_idle_time == 0)
11284 			rack->r_ctl.rc_went_idle_time = 1;
11285 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
11286 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
11287 			tp->t_acktime = 0;
11288 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11289 	}
11290 	if (acked && rack->r_fast_output)
11291 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
11292 	if (sbavail(&so->so_snd)) {
11293 		rack->r_wanted_output = 1;
11294 	}
11295 	return (1);
11296 }
11297 
11298 /*
11299  * Return value of 1, the TCB is unlocked and most
11300  * likely gone, return value of 0, the TCP is still
11301  * locked.
11302  */
11303 static int
11304 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
11305     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11306     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11307 {
11308 	int32_t ret_val = 0;
11309 	int32_t todrop;
11310 	int32_t ourfinisacked = 0;
11311 	struct tcp_rack *rack;
11312 
11313 	ctf_calc_rwin(so, tp);
11314 	/*
11315 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
11316 	 * SYN, drop the input. if seg contains a RST, then drop the
11317 	 * connection. if seg does not contain SYN, then drop it. Otherwise
11318 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
11319 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
11320 	 * contains an ECE and ECN support is enabled, the stream is ECN
11321 	 * capable. if SYN has been acked change to ESTABLISHED else
11322 	 * SYN_RCVD state arrange for segment to be acked (eventually)
11323 	 * continue processing rest of data/controls.
11324 	 */
11325 	if ((thflags & TH_ACK) &&
11326 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
11327 	    SEQ_GT(th->th_ack, tp->snd_max))) {
11328 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11329 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11330 		return (1);
11331 	}
11332 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
11333 		TCP_PROBE5(connect__refused, NULL, tp,
11334 		    mtod(m, const char *), tp, th);
11335 		tp = tcp_drop(tp, ECONNREFUSED);
11336 		ctf_do_drop(m, tp);
11337 		return (1);
11338 	}
11339 	if (thflags & TH_RST) {
11340 		ctf_do_drop(m, tp);
11341 		return (1);
11342 	}
11343 	if (!(thflags & TH_SYN)) {
11344 		ctf_do_drop(m, tp);
11345 		return (1);
11346 	}
11347 	tp->irs = th->th_seq;
11348 	tcp_rcvseqinit(tp);
11349 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11350 	if (thflags & TH_ACK) {
11351 		int tfo_partial = 0;
11352 
11353 		KMOD_TCPSTAT_INC(tcps_connects);
11354 		soisconnected(so);
11355 #ifdef MAC
11356 		mac_socketpeer_set_from_mbuf(m, so);
11357 #endif
11358 		/* Do window scaling on this connection? */
11359 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11360 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11361 			tp->rcv_scale = tp->request_r_scale;
11362 		}
11363 		tp->rcv_adv += min(tp->rcv_wnd,
11364 		    TCP_MAXWIN << tp->rcv_scale);
11365 		/*
11366 		 * If not all the data that was sent in the TFO SYN
11367 		 * has been acked, resend the remainder right away.
11368 		 */
11369 		if (IS_FASTOPEN(tp->t_flags) &&
11370 		    (tp->snd_una != tp->snd_max)) {
11371 			tp->snd_nxt = th->th_ack;
11372 			tfo_partial = 1;
11373 		}
11374 		/*
11375 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
11376 		 * will be turned on later.
11377 		 */
11378 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
11379 			rack_timer_cancel(tp, rack,
11380 					  rack->r_ctl.rc_rcvtime, __LINE__);
11381 			tp->t_flags |= TF_DELACK;
11382 		} else {
11383 			rack->r_wanted_output = 1;
11384 			tp->t_flags |= TF_ACKNOW;
11385 			rack->rc_dack_toggle = 0;
11386 		}
11387 		if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) &&
11388 		    (V_tcp_do_ecn == 1)) {
11389 			tp->t_flags2 |= TF2_ECN_PERMIT;
11390 			KMOD_TCPSTAT_INC(tcps_ecn_shs);
11391 		}
11392 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
11393 			/*
11394 			 * We advance snd_una for the
11395 			 * fast open case. If th_ack is
11396 			 * acknowledging data beyond
11397 			 * snd_una we can't just call
11398 			 * ack-processing since the
11399 			 * data stream in our send-map
11400 			 * will start at snd_una + 1 (one
11401 			 * beyond the SYN). If its just
11402 			 * equal we don't need to do that
11403 			 * and there is no send_map.
11404 			 */
11405 			tp->snd_una++;
11406 		}
11407 		/*
11408 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
11409 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
11410 		 */
11411 		tp->t_starttime = ticks;
11412 		if (tp->t_flags & TF_NEEDFIN) {
11413 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
11414 			tp->t_flags &= ~TF_NEEDFIN;
11415 			thflags &= ~TH_SYN;
11416 		} else {
11417 			tcp_state_change(tp, TCPS_ESTABLISHED);
11418 			TCP_PROBE5(connect__established, NULL, tp,
11419 			    mtod(m, const char *), tp, th);
11420 			rack_cc_conn_init(tp);
11421 		}
11422 	} else {
11423 		/*
11424 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
11425 		 * open.  If segment contains CC option and there is a
11426 		 * cached CC, apply TAO test. If it succeeds, connection is *
11427 		 * half-synchronized. Otherwise, do 3-way handshake:
11428 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
11429 		 * there was no CC option, clear cached CC value.
11430 		 */
11431 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
11432 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
11433 	}
11434 	INP_WLOCK_ASSERT(tp->t_inpcb);
11435 	/*
11436 	 * Advance th->th_seq to correspond to first data byte. If data,
11437 	 * trim to stay within window, dropping FIN if necessary.
11438 	 */
11439 	th->th_seq++;
11440 	if (tlen > tp->rcv_wnd) {
11441 		todrop = tlen - tp->rcv_wnd;
11442 		m_adj(m, -todrop);
11443 		tlen = tp->rcv_wnd;
11444 		thflags &= ~TH_FIN;
11445 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
11446 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
11447 	}
11448 	tp->snd_wl1 = th->th_seq - 1;
11449 	tp->rcv_up = th->th_seq;
11450 	/*
11451 	 * Client side of transaction: already sent SYN and data. If the
11452 	 * remote host used T/TCP to validate the SYN, our data will be
11453 	 * ACK'd; if so, enter normal data segment processing in the middle
11454 	 * of step 5, ack processing. Otherwise, goto step 6.
11455 	 */
11456 	if (thflags & TH_ACK) {
11457 		/* For syn-sent we need to possibly update the rtt */
11458 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11459 			uint32_t t, mcts;
11460 
11461 			mcts = tcp_ts_getticks();
11462 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11463 			if (!tp->t_rttlow || tp->t_rttlow > t)
11464 				tp->t_rttlow = t;
11465 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
11466 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11467 			tcp_rack_xmit_timer_commit(rack, tp);
11468 		}
11469 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
11470 			return (ret_val);
11471 		/* We may have changed to FIN_WAIT_1 above */
11472 		if (tp->t_state == TCPS_FIN_WAIT_1) {
11473 			/*
11474 			 * In FIN_WAIT_1 STATE in addition to the processing
11475 			 * for the ESTABLISHED state if our FIN is now
11476 			 * acknowledged then enter FIN_WAIT_2.
11477 			 */
11478 			if (ourfinisacked) {
11479 				/*
11480 				 * If we can't receive any more data, then
11481 				 * closing user can proceed. Starting the
11482 				 * timer is contrary to the specification,
11483 				 * but if we don't get a FIN we'll hang
11484 				 * forever.
11485 				 *
11486 				 * XXXjl: we should release the tp also, and
11487 				 * use a compressed state.
11488 				 */
11489 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11490 					soisdisconnected(so);
11491 					tcp_timer_activate(tp, TT_2MSL,
11492 					    (tcp_fast_finwait2_recycle ?
11493 					    tcp_finwait2_timeout :
11494 					    TP_MAXIDLE(tp)));
11495 				}
11496 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
11497 			}
11498 		}
11499 	}
11500 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11501 	   tiwin, thflags, nxt_pkt));
11502 }
11503 
11504 /*
11505  * Return value of 1, the TCB is unlocked and most
11506  * likely gone, return value of 0, the TCP is still
11507  * locked.
11508  */
11509 static int
11510 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
11511     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11512     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11513 {
11514 	struct tcp_rack *rack;
11515 	int32_t ret_val = 0;
11516 	int32_t ourfinisacked = 0;
11517 
11518 	ctf_calc_rwin(so, tp);
11519 	if ((thflags & TH_ACK) &&
11520 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
11521 	    SEQ_GT(th->th_ack, tp->snd_max))) {
11522 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11523 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11524 		return (1);
11525 	}
11526 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11527 	if (IS_FASTOPEN(tp->t_flags)) {
11528 		/*
11529 		 * When a TFO connection is in SYN_RECEIVED, the
11530 		 * only valid packets are the initial SYN, a
11531 		 * retransmit/copy of the initial SYN (possibly with
11532 		 * a subset of the original data), a valid ACK, a
11533 		 * FIN, or a RST.
11534 		 */
11535 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
11536 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11537 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11538 			return (1);
11539 		} else if (thflags & TH_SYN) {
11540 			/* non-initial SYN is ignored */
11541 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
11542 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
11543 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
11544 				ctf_do_drop(m, NULL);
11545 				return (0);
11546 			}
11547 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
11548 			ctf_do_drop(m, NULL);
11549 			return (0);
11550 		}
11551 	}
11552 
11553 	if ((thflags & TH_RST) ||
11554 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11555 		return (__ctf_process_rst(m, th, so, tp,
11556 					  &rack->r_ctl.challenge_ack_ts,
11557 					  &rack->r_ctl.challenge_ack_cnt));
11558 	/*
11559 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11560 	 * it's less than ts_recent, drop it.
11561 	 */
11562 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11563 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11564 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11565 			return (ret_val);
11566 	}
11567 	/*
11568 	 * In the SYN-RECEIVED state, validate that the packet belongs to
11569 	 * this connection before trimming the data to fit the receive
11570 	 * window.  Check the sequence number versus IRS since we know the
11571 	 * sequence numbers haven't wrapped.  This is a partial fix for the
11572 	 * "LAND" DoS attack.
11573 	 */
11574 	if (SEQ_LT(th->th_seq, tp->irs)) {
11575 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11576 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11577 		return (1);
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 	tp->snd_wnd = tiwin;
11606 	rack_validate_fo_sendwin_up(tp, rack);
11607 	/*
11608 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11609 	 * is on (half-synchronized state), then queue data for later
11610 	 * processing; else drop segment and return.
11611 	 */
11612 	if ((thflags & TH_ACK) == 0) {
11613 		if (IS_FASTOPEN(tp->t_flags)) {
11614 			rack_cc_conn_init(tp);
11615 		}
11616 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11617 		    tiwin, thflags, nxt_pkt));
11618 	}
11619 	KMOD_TCPSTAT_INC(tcps_connects);
11620 	soisconnected(so);
11621 	/* Do window scaling? */
11622 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11623 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11624 		tp->rcv_scale = tp->request_r_scale;
11625 	}
11626 	/*
11627 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11628 	 * FIN-WAIT-1
11629 	 */
11630 	tp->t_starttime = ticks;
11631 	if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11632 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11633 		tp->t_tfo_pending = NULL;
11634 	}
11635 	if (tp->t_flags & TF_NEEDFIN) {
11636 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
11637 		tp->t_flags &= ~TF_NEEDFIN;
11638 	} else {
11639 		tcp_state_change(tp, TCPS_ESTABLISHED);
11640 		TCP_PROBE5(accept__established, NULL, tp,
11641 		    mtod(m, const char *), tp, th);
11642 		/*
11643 		 * TFO connections call cc_conn_init() during SYN
11644 		 * processing.  Calling it again here for such connections
11645 		 * is not harmless as it would undo the snd_cwnd reduction
11646 		 * that occurs when a TFO SYN|ACK is retransmitted.
11647 		 */
11648 		if (!IS_FASTOPEN(tp->t_flags))
11649 			rack_cc_conn_init(tp);
11650 	}
11651 	/*
11652 	 * Account for the ACK of our SYN prior to
11653 	 * regular ACK processing below, except for
11654 	 * simultaneous SYN, which is handled later.
11655 	 */
11656 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11657 		tp->snd_una++;
11658 	/*
11659 	 * If segment contains data or ACK, will call tcp_reass() later; if
11660 	 * not, do so now to pass queued data to user.
11661 	 */
11662 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
11663 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11664 		    (struct mbuf *)0);
11665 		if (tp->t_flags & TF_WAKESOR) {
11666 			tp->t_flags &= ~TF_WAKESOR;
11667 			/* NB: sorwakeup_locked() does an implicit unlock. */
11668 			sorwakeup_locked(so);
11669 		}
11670 	}
11671 	tp->snd_wl1 = th->th_seq - 1;
11672 	/* For syn-recv we need to possibly update the rtt */
11673 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11674 		uint32_t t, mcts;
11675 
11676 		mcts = tcp_ts_getticks();
11677 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11678 		if (!tp->t_rttlow || tp->t_rttlow > t)
11679 			tp->t_rttlow = t;
11680 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11681 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11682 		tcp_rack_xmit_timer_commit(rack, tp);
11683 	}
11684 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11685 		return (ret_val);
11686 	}
11687 	if (tp->t_state == TCPS_FIN_WAIT_1) {
11688 		/* We could have went to FIN_WAIT_1 (or EST) above */
11689 		/*
11690 		 * In FIN_WAIT_1 STATE in addition to the processing for the
11691 		 * ESTABLISHED state if our FIN is now acknowledged then
11692 		 * enter FIN_WAIT_2.
11693 		 */
11694 		if (ourfinisacked) {
11695 			/*
11696 			 * If we can't receive any more data, then closing
11697 			 * user can proceed. Starting the timer is contrary
11698 			 * to the specification, but if we don't get a FIN
11699 			 * we'll hang forever.
11700 			 *
11701 			 * XXXjl: we should release the tp also, and use a
11702 			 * compressed state.
11703 			 */
11704 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11705 				soisdisconnected(so);
11706 				tcp_timer_activate(tp, TT_2MSL,
11707 				    (tcp_fast_finwait2_recycle ?
11708 				    tcp_finwait2_timeout :
11709 				    TP_MAXIDLE(tp)));
11710 			}
11711 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
11712 		}
11713 	}
11714 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11715 	    tiwin, thflags, nxt_pkt));
11716 }
11717 
11718 /*
11719  * Return value of 1, the TCB is unlocked and most
11720  * likely gone, return value of 0, the TCP is still
11721  * locked.
11722  */
11723 static int
11724 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11725     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11726     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11727 {
11728 	int32_t ret_val = 0;
11729 	struct tcp_rack *rack;
11730 
11731 	/*
11732 	 * Header prediction: check for the two common cases of a
11733 	 * uni-directional data xfer.  If the packet has no control flags,
11734 	 * is in-sequence, the window didn't change and we're not
11735 	 * retransmitting, it's a candidate.  If the length is zero and the
11736 	 * ack moved forward, we're the sender side of the xfer.  Just free
11737 	 * the data acked & wake any higher level process that was blocked
11738 	 * waiting for space.  If the length is non-zero and the ack didn't
11739 	 * move, we're the receiver side.  If we're getting packets in-order
11740 	 * (the reassembly queue is empty), add the data toc The socket
11741 	 * buffer and note that we need a delayed ack. Make sure that the
11742 	 * hidden state-flags are also off. Since we check for
11743 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11744 	 */
11745 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11746 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11747 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11748 	    __predict_true(SEGQ_EMPTY(tp)) &&
11749 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
11750 		if (tlen == 0) {
11751 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11752 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11753 				return (0);
11754 			}
11755 		} else {
11756 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11757 			    tiwin, nxt_pkt, iptos)) {
11758 				return (0);
11759 			}
11760 		}
11761 	}
11762 	ctf_calc_rwin(so, tp);
11763 
11764 	if ((thflags & TH_RST) ||
11765 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11766 		return (__ctf_process_rst(m, th, so, tp,
11767 					  &rack->r_ctl.challenge_ack_ts,
11768 					  &rack->r_ctl.challenge_ack_cnt));
11769 
11770 	/*
11771 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11772 	 * synchronized state.
11773 	 */
11774 	if (thflags & TH_SYN) {
11775 		ctf_challenge_ack(m, th, tp, &ret_val);
11776 		return (ret_val);
11777 	}
11778 	/*
11779 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11780 	 * it's less than ts_recent, drop it.
11781 	 */
11782 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11783 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11784 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11785 			return (ret_val);
11786 	}
11787 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11788 			      &rack->r_ctl.challenge_ack_ts,
11789 			      &rack->r_ctl.challenge_ack_cnt)) {
11790 		return (ret_val);
11791 	}
11792 	/*
11793 	 * If last ACK falls within this segment's sequence numbers, record
11794 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11795 	 * from the latest proposal of the tcplw@cray.com list (Braden
11796 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11797 	 * with our earlier PAWS tests, so this check should be solely
11798 	 * predicated on the sequence space of this segment. 3) That we
11799 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11800 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11801 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11802 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11803 	 * p.869. In such cases, we can still calculate the RTT correctly
11804 	 * when RCV.NXT == Last.ACK.Sent.
11805 	 */
11806 	if ((to->to_flags & TOF_TS) != 0 &&
11807 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11808 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11809 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11810 		tp->ts_recent_age = tcp_ts_getticks();
11811 		tp->ts_recent = to->to_tsval;
11812 	}
11813 	/*
11814 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11815 	 * is on (half-synchronized state), then queue data for later
11816 	 * processing; else drop segment and return.
11817 	 */
11818 	if ((thflags & TH_ACK) == 0) {
11819 		if (tp->t_flags & TF_NEEDSYN) {
11820 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11821 			    tiwin, thflags, nxt_pkt));
11822 
11823 		} else if (tp->t_flags & TF_ACKNOW) {
11824 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11825 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11826 			return (ret_val);
11827 		} else {
11828 			ctf_do_drop(m, NULL);
11829 			return (0);
11830 		}
11831 	}
11832 	/*
11833 	 * Ack processing.
11834 	 */
11835 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11836 		return (ret_val);
11837 	}
11838 	if (sbavail(&so->so_snd)) {
11839 		if (ctf_progress_timeout_check(tp, true)) {
11840 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11841 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11842 			return (1);
11843 		}
11844 	}
11845 	/* State changes only happen in rack_process_data() */
11846 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11847 	    tiwin, thflags, nxt_pkt));
11848 }
11849 
11850 /*
11851  * Return value of 1, the TCB is unlocked and most
11852  * likely gone, return value of 0, the TCP is still
11853  * locked.
11854  */
11855 static int
11856 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11857     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11858     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11859 {
11860 	int32_t ret_val = 0;
11861 	struct tcp_rack *rack;
11862 
11863 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11864 	ctf_calc_rwin(so, tp);
11865 	if ((thflags & TH_RST) ||
11866 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11867 		return (__ctf_process_rst(m, th, so, tp,
11868 					  &rack->r_ctl.challenge_ack_ts,
11869 					  &rack->r_ctl.challenge_ack_cnt));
11870 	/*
11871 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11872 	 * synchronized state.
11873 	 */
11874 	if (thflags & TH_SYN) {
11875 		ctf_challenge_ack(m, th, tp, &ret_val);
11876 		return (ret_val);
11877 	}
11878 	/*
11879 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11880 	 * it's less than ts_recent, drop it.
11881 	 */
11882 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11883 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11884 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11885 			return (ret_val);
11886 	}
11887 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11888 			      &rack->r_ctl.challenge_ack_ts,
11889 			      &rack->r_ctl.challenge_ack_cnt)) {
11890 		return (ret_val);
11891 	}
11892 	/*
11893 	 * If last ACK falls within this segment's sequence numbers, record
11894 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11895 	 * from the latest proposal of the tcplw@cray.com list (Braden
11896 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11897 	 * with our earlier PAWS tests, so this check should be solely
11898 	 * predicated on the sequence space of this segment. 3) That we
11899 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11900 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11901 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11902 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11903 	 * p.869. In such cases, we can still calculate the RTT correctly
11904 	 * when RCV.NXT == Last.ACK.Sent.
11905 	 */
11906 	if ((to->to_flags & TOF_TS) != 0 &&
11907 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11908 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11909 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11910 		tp->ts_recent_age = tcp_ts_getticks();
11911 		tp->ts_recent = to->to_tsval;
11912 	}
11913 	/*
11914 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11915 	 * is on (half-synchronized state), then queue data for later
11916 	 * processing; else drop segment and return.
11917 	 */
11918 	if ((thflags & TH_ACK) == 0) {
11919 		if (tp->t_flags & TF_NEEDSYN) {
11920 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11921 			    tiwin, thflags, nxt_pkt));
11922 
11923 		} else if (tp->t_flags & TF_ACKNOW) {
11924 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11925 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11926 			return (ret_val);
11927 		} else {
11928 			ctf_do_drop(m, NULL);
11929 			return (0);
11930 		}
11931 	}
11932 	/*
11933 	 * Ack processing.
11934 	 */
11935 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11936 		return (ret_val);
11937 	}
11938 	if (sbavail(&so->so_snd)) {
11939 		if (ctf_progress_timeout_check(tp, true)) {
11940 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11941 						tp, tick, PROGRESS_DROP, __LINE__);
11942 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11943 			return (1);
11944 		}
11945 	}
11946 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11947 	    tiwin, thflags, nxt_pkt));
11948 }
11949 
11950 static int
11951 rack_check_data_after_close(struct mbuf *m,
11952     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11953 {
11954 	struct tcp_rack *rack;
11955 
11956 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11957 	if (rack->rc_allow_data_af_clo == 0) {
11958 	close_now:
11959 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11960 		/* tcp_close will kill the inp pre-log the Reset */
11961 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11962 		tp = tcp_close(tp);
11963 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11964 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11965 		return (1);
11966 	}
11967 	if (sbavail(&so->so_snd) == 0)
11968 		goto close_now;
11969 	/* Ok we allow data that is ignored and a followup reset */
11970 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11971 	tp->rcv_nxt = th->th_seq + *tlen;
11972 	tp->t_flags2 |= TF2_DROP_AF_DATA;
11973 	rack->r_wanted_output = 1;
11974 	*tlen = 0;
11975 	return (0);
11976 }
11977 
11978 /*
11979  * Return value of 1, the TCB is unlocked and most
11980  * likely gone, return value of 0, the TCP is still
11981  * locked.
11982  */
11983 static int
11984 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11985     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11986     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11987 {
11988 	int32_t ret_val = 0;
11989 	int32_t ourfinisacked = 0;
11990 	struct tcp_rack *rack;
11991 
11992 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11993 	ctf_calc_rwin(so, tp);
11994 
11995 	if ((thflags & TH_RST) ||
11996 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11997 		return (__ctf_process_rst(m, th, so, tp,
11998 					  &rack->r_ctl.challenge_ack_ts,
11999 					  &rack->r_ctl.challenge_ack_cnt));
12000 	/*
12001 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12002 	 * synchronized state.
12003 	 */
12004 	if (thflags & TH_SYN) {
12005 		ctf_challenge_ack(m, th, tp, &ret_val);
12006 		return (ret_val);
12007 	}
12008 	/*
12009 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12010 	 * it's less than ts_recent, drop it.
12011 	 */
12012 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12013 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12014 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12015 			return (ret_val);
12016 	}
12017 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12018 			      &rack->r_ctl.challenge_ack_ts,
12019 			      &rack->r_ctl.challenge_ack_cnt)) {
12020 		return (ret_val);
12021 	}
12022 	/*
12023 	 * If new data are received on a connection after the user processes
12024 	 * are gone, then RST the other end.
12025 	 */
12026 	if ((so->so_state & SS_NOFDREF) && tlen) {
12027 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
12028 			return (1);
12029 	}
12030 	/*
12031 	 * If last ACK falls within this segment's sequence numbers, record
12032 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12033 	 * from the latest proposal of the tcplw@cray.com list (Braden
12034 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12035 	 * with our earlier PAWS tests, so this check should be solely
12036 	 * predicated on the sequence space of this segment. 3) That we
12037 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12038 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12039 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12040 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12041 	 * p.869. In such cases, we can still calculate the RTT correctly
12042 	 * when RCV.NXT == Last.ACK.Sent.
12043 	 */
12044 	if ((to->to_flags & TOF_TS) != 0 &&
12045 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12046 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12047 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12048 		tp->ts_recent_age = tcp_ts_getticks();
12049 		tp->ts_recent = to->to_tsval;
12050 	}
12051 	/*
12052 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12053 	 * is on (half-synchronized state), then queue data for later
12054 	 * processing; else drop segment and return.
12055 	 */
12056 	if ((thflags & TH_ACK) == 0) {
12057 		if (tp->t_flags & TF_NEEDSYN) {
12058 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12059 			    tiwin, thflags, nxt_pkt));
12060 		} else if (tp->t_flags & TF_ACKNOW) {
12061 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12062 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12063 			return (ret_val);
12064 		} else {
12065 			ctf_do_drop(m, NULL);
12066 			return (0);
12067 		}
12068 	}
12069 	/*
12070 	 * Ack processing.
12071 	 */
12072 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12073 		return (ret_val);
12074 	}
12075 	if (ourfinisacked) {
12076 		/*
12077 		 * If we can't receive any more data, then closing user can
12078 		 * proceed. Starting the timer is contrary to the
12079 		 * specification, but if we don't get a FIN we'll hang
12080 		 * forever.
12081 		 *
12082 		 * XXXjl: we should release the tp also, and use a
12083 		 * compressed state.
12084 		 */
12085 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12086 			soisdisconnected(so);
12087 			tcp_timer_activate(tp, TT_2MSL,
12088 			    (tcp_fast_finwait2_recycle ?
12089 			    tcp_finwait2_timeout :
12090 			    TP_MAXIDLE(tp)));
12091 		}
12092 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
12093 	}
12094 	if (sbavail(&so->so_snd)) {
12095 		if (ctf_progress_timeout_check(tp, true)) {
12096 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12097 						tp, tick, PROGRESS_DROP, __LINE__);
12098 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12099 			return (1);
12100 		}
12101 	}
12102 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12103 	    tiwin, thflags, nxt_pkt));
12104 }
12105 
12106 /*
12107  * Return value of 1, the TCB is unlocked and most
12108  * likely gone, return value of 0, the TCP is still
12109  * locked.
12110  */
12111 static int
12112 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
12113     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12114     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12115 {
12116 	int32_t ret_val = 0;
12117 	int32_t ourfinisacked = 0;
12118 	struct tcp_rack *rack;
12119 
12120 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12121 	ctf_calc_rwin(so, tp);
12122 
12123 	if ((thflags & TH_RST) ||
12124 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12125 		return (__ctf_process_rst(m, th, so, tp,
12126 					  &rack->r_ctl.challenge_ack_ts,
12127 					  &rack->r_ctl.challenge_ack_cnt));
12128 	/*
12129 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12130 	 * synchronized state.
12131 	 */
12132 	if (thflags & TH_SYN) {
12133 		ctf_challenge_ack(m, th, tp, &ret_val);
12134 		return (ret_val);
12135 	}
12136 	/*
12137 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12138 	 * it's less than ts_recent, drop it.
12139 	 */
12140 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12141 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12142 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12143 			return (ret_val);
12144 	}
12145 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12146 			      &rack->r_ctl.challenge_ack_ts,
12147 			      &rack->r_ctl.challenge_ack_cnt)) {
12148 		return (ret_val);
12149 	}
12150 	/*
12151 	 * If new data are received on a connection after the user processes
12152 	 * are gone, then RST the other end.
12153 	 */
12154 	if ((so->so_state & SS_NOFDREF) && tlen) {
12155 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
12156 			return (1);
12157 	}
12158 	/*
12159 	 * If last ACK falls within this segment's sequence numbers, record
12160 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12161 	 * from the latest proposal of the tcplw@cray.com list (Braden
12162 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12163 	 * with our earlier PAWS tests, so this check should be solely
12164 	 * predicated on the sequence space of this segment. 3) That we
12165 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12166 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12167 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12168 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12169 	 * p.869. In such cases, we can still calculate the RTT correctly
12170 	 * when RCV.NXT == Last.ACK.Sent.
12171 	 */
12172 	if ((to->to_flags & TOF_TS) != 0 &&
12173 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12174 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12175 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12176 		tp->ts_recent_age = tcp_ts_getticks();
12177 		tp->ts_recent = to->to_tsval;
12178 	}
12179 	/*
12180 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12181 	 * is on (half-synchronized state), then queue data for later
12182 	 * processing; else drop segment and return.
12183 	 */
12184 	if ((thflags & TH_ACK) == 0) {
12185 		if (tp->t_flags & TF_NEEDSYN) {
12186 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12187 			    tiwin, thflags, nxt_pkt));
12188 		} else if (tp->t_flags & TF_ACKNOW) {
12189 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12190 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12191 			return (ret_val);
12192 		} else {
12193 			ctf_do_drop(m, NULL);
12194 			return (0);
12195 		}
12196 	}
12197 	/*
12198 	 * Ack processing.
12199 	 */
12200 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12201 		return (ret_val);
12202 	}
12203 	if (ourfinisacked) {
12204 		tcp_twstart(tp);
12205 		m_freem(m);
12206 		return (1);
12207 	}
12208 	if (sbavail(&so->so_snd)) {
12209 		if (ctf_progress_timeout_check(tp, true)) {
12210 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12211 						tp, tick, PROGRESS_DROP, __LINE__);
12212 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12213 			return (1);
12214 		}
12215 	}
12216 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12217 	    tiwin, thflags, nxt_pkt));
12218 }
12219 
12220 /*
12221  * Return value of 1, the TCB is unlocked and most
12222  * likely gone, return value of 0, the TCP is still
12223  * locked.
12224  */
12225 static int
12226 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12227     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12228     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12229 {
12230 	int32_t ret_val = 0;
12231 	int32_t ourfinisacked = 0;
12232 	struct tcp_rack *rack;
12233 
12234 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12235 	ctf_calc_rwin(so, tp);
12236 
12237 	if ((thflags & TH_RST) ||
12238 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12239 		return (__ctf_process_rst(m, th, so, tp,
12240 					  &rack->r_ctl.challenge_ack_ts,
12241 					  &rack->r_ctl.challenge_ack_cnt));
12242 	/*
12243 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12244 	 * synchronized state.
12245 	 */
12246 	if (thflags & TH_SYN) {
12247 		ctf_challenge_ack(m, th, tp, &ret_val);
12248 		return (ret_val);
12249 	}
12250 	/*
12251 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12252 	 * it's less than ts_recent, drop it.
12253 	 */
12254 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12255 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12256 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12257 			return (ret_val);
12258 	}
12259 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12260 			      &rack->r_ctl.challenge_ack_ts,
12261 			      &rack->r_ctl.challenge_ack_cnt)) {
12262 		return (ret_val);
12263 	}
12264 	/*
12265 	 * If new data are received on a connection after the user processes
12266 	 * are gone, then RST the other end.
12267 	 */
12268 	if ((so->so_state & SS_NOFDREF) && tlen) {
12269 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
12270 			return (1);
12271 	}
12272 	/*
12273 	 * If last ACK falls within this segment's sequence numbers, record
12274 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12275 	 * from the latest proposal of the tcplw@cray.com list (Braden
12276 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12277 	 * with our earlier PAWS tests, so this check should be solely
12278 	 * predicated on the sequence space of this segment. 3) That we
12279 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12280 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12281 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12282 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12283 	 * p.869. In such cases, we can still calculate the RTT correctly
12284 	 * when RCV.NXT == Last.ACK.Sent.
12285 	 */
12286 	if ((to->to_flags & TOF_TS) != 0 &&
12287 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12288 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12289 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12290 		tp->ts_recent_age = tcp_ts_getticks();
12291 		tp->ts_recent = to->to_tsval;
12292 	}
12293 	/*
12294 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12295 	 * is on (half-synchronized state), then queue data for later
12296 	 * processing; else drop segment and return.
12297 	 */
12298 	if ((thflags & TH_ACK) == 0) {
12299 		if (tp->t_flags & TF_NEEDSYN) {
12300 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12301 			    tiwin, thflags, nxt_pkt));
12302 		} else if (tp->t_flags & TF_ACKNOW) {
12303 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12304 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12305 			return (ret_val);
12306 		} else {
12307 			ctf_do_drop(m, NULL);
12308 			return (0);
12309 		}
12310 	}
12311 	/*
12312 	 * case TCPS_LAST_ACK: Ack processing.
12313 	 */
12314 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12315 		return (ret_val);
12316 	}
12317 	if (ourfinisacked) {
12318 		tp = tcp_close(tp);
12319 		ctf_do_drop(m, tp);
12320 		return (1);
12321 	}
12322 	if (sbavail(&so->so_snd)) {
12323 		if (ctf_progress_timeout_check(tp, true)) {
12324 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12325 						tp, tick, PROGRESS_DROP, __LINE__);
12326 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12327 			return (1);
12328 		}
12329 	}
12330 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12331 	    tiwin, thflags, nxt_pkt));
12332 }
12333 
12334 /*
12335  * Return value of 1, the TCB is unlocked and most
12336  * likely gone, return value of 0, the TCP is still
12337  * locked.
12338  */
12339 static int
12340 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
12341     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12342     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12343 {
12344 	int32_t ret_val = 0;
12345 	int32_t ourfinisacked = 0;
12346 	struct tcp_rack *rack;
12347 
12348 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12349 	ctf_calc_rwin(so, tp);
12350 
12351 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
12352 	if ((thflags & TH_RST) ||
12353 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12354 		return (__ctf_process_rst(m, th, so, tp,
12355 					  &rack->r_ctl.challenge_ack_ts,
12356 					  &rack->r_ctl.challenge_ack_cnt));
12357 	/*
12358 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12359 	 * synchronized state.
12360 	 */
12361 	if (thflags & TH_SYN) {
12362 		ctf_challenge_ack(m, th, tp, &ret_val);
12363 		return (ret_val);
12364 	}
12365 	/*
12366 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12367 	 * it's less than ts_recent, drop it.
12368 	 */
12369 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12370 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12371 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12372 			return (ret_val);
12373 	}
12374 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12375 			      &rack->r_ctl.challenge_ack_ts,
12376 			      &rack->r_ctl.challenge_ack_cnt)) {
12377 		return (ret_val);
12378 	}
12379 	/*
12380 	 * If new data are received on a connection after the user processes
12381 	 * are gone, then RST the other end.
12382 	 */
12383 	if ((so->so_state & SS_NOFDREF) &&
12384 	    tlen) {
12385 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
12386 			return (1);
12387 	}
12388 	/*
12389 	 * If last ACK falls within this segment's sequence numbers, record
12390 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12391 	 * from the latest proposal of the tcplw@cray.com list (Braden
12392 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12393 	 * with our earlier PAWS tests, so this check should be solely
12394 	 * predicated on the sequence space of this segment. 3) That we
12395 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12396 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12397 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12398 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12399 	 * p.869. In such cases, we can still calculate the RTT correctly
12400 	 * when RCV.NXT == Last.ACK.Sent.
12401 	 */
12402 	if ((to->to_flags & TOF_TS) != 0 &&
12403 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12404 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12405 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12406 		tp->ts_recent_age = tcp_ts_getticks();
12407 		tp->ts_recent = to->to_tsval;
12408 	}
12409 	/*
12410 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12411 	 * is on (half-synchronized state), then queue data for later
12412 	 * processing; else drop segment and return.
12413 	 */
12414 	if ((thflags & TH_ACK) == 0) {
12415 		if (tp->t_flags & TF_NEEDSYN) {
12416 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12417 			    tiwin, thflags, nxt_pkt));
12418 		} else if (tp->t_flags & TF_ACKNOW) {
12419 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12420 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12421 			return (ret_val);
12422 		} else {
12423 			ctf_do_drop(m, NULL);
12424 			return (0);
12425 		}
12426 	}
12427 	/*
12428 	 * Ack processing.
12429 	 */
12430 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12431 		return (ret_val);
12432 	}
12433 	if (sbavail(&so->so_snd)) {
12434 		if (ctf_progress_timeout_check(tp, true)) {
12435 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12436 						tp, tick, PROGRESS_DROP, __LINE__);
12437 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12438 			return (1);
12439 		}
12440 	}
12441 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12442 	    tiwin, thflags, nxt_pkt));
12443 }
12444 
12445 static void inline
12446 rack_clear_rate_sample(struct tcp_rack *rack)
12447 {
12448 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
12449 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
12450 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
12451 }
12452 
12453 static void
12454 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
12455 {
12456 	uint64_t bw_est, rate_wanted;
12457 	int chged = 0;
12458 	uint32_t user_max, orig_min, orig_max;
12459 
12460 	orig_min = rack->r_ctl.rc_pace_min_segs;
12461 	orig_max = rack->r_ctl.rc_pace_max_segs;
12462 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
12463 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
12464 		chged = 1;
12465 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
12466 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
12467 		if (user_max != rack->r_ctl.rc_pace_max_segs)
12468 			chged = 1;
12469 	}
12470 	if (rack->rc_force_max_seg) {
12471 		rack->r_ctl.rc_pace_max_segs = user_max;
12472 	} else if (rack->use_fixed_rate) {
12473 		bw_est = rack_get_bw(rack);
12474 		if ((rack->r_ctl.crte == NULL) ||
12475 		    (bw_est != rack->r_ctl.crte->rate)) {
12476 			rack->r_ctl.rc_pace_max_segs = user_max;
12477 		} else {
12478 			/* We are pacing right at the hardware rate */
12479 			uint32_t segsiz;
12480 
12481 			segsiz = min(ctf_fixed_maxseg(tp),
12482 				     rack->r_ctl.rc_pace_min_segs);
12483 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
12484 				                           tp, bw_est, segsiz, 0,
12485 							   rack->r_ctl.crte, NULL);
12486 		}
12487 	} else if (rack->rc_always_pace) {
12488 		if (rack->r_ctl.gp_bw ||
12489 #ifdef NETFLIX_PEAKRATE
12490 		    rack->rc_tp->t_maxpeakrate ||
12491 #endif
12492 		    rack->r_ctl.init_rate) {
12493 			/* We have a rate of some sort set */
12494 			uint32_t  orig;
12495 
12496 			bw_est = rack_get_bw(rack);
12497 			orig = rack->r_ctl.rc_pace_max_segs;
12498 			if (fill_override)
12499 				rate_wanted = *fill_override;
12500 			else
12501 				rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
12502 			if (rate_wanted) {
12503 				/* We have something */
12504 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
12505 										   rate_wanted,
12506 										   ctf_fixed_maxseg(rack->rc_tp));
12507 			} else
12508 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
12509 			if (orig != rack->r_ctl.rc_pace_max_segs)
12510 				chged = 1;
12511 		} else if ((rack->r_ctl.gp_bw == 0) &&
12512 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
12513 			/*
12514 			 * If we have nothing limit us to bursting
12515 			 * out IW sized pieces.
12516 			 */
12517 			chged = 1;
12518 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
12519 		}
12520 	}
12521 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
12522 		chged = 1;
12523 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
12524 	}
12525 	if (chged)
12526 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
12527 }
12528 
12529 
12530 static void
12531 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
12532 {
12533 #ifdef INET6
12534 	struct ip6_hdr *ip6 = NULL;
12535 #endif
12536 #ifdef INET
12537 	struct ip *ip = NULL;
12538 #endif
12539 	struct udphdr *udp = NULL;
12540 
12541 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
12542 #ifdef INET6
12543 	if (rack->r_is_v6) {
12544 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12545 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
12546 		if (tp->t_port) {
12547 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12548 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
12549 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12550 			udp->uh_dport = tp->t_port;
12551 			rack->r_ctl.fsb.udp = udp;
12552 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12553 		} else
12554 		{
12555 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
12556 			rack->r_ctl.fsb.udp = NULL;
12557 		}
12558 		tcpip_fillheaders(rack->rc_inp,
12559 				  tp->t_port,
12560 				  ip6, rack->r_ctl.fsb.th);
12561 	} else
12562 #endif				/* INET6 */
12563 	{
12564 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
12565 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
12566 		if (tp->t_port) {
12567 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12568 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
12569 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12570 			udp->uh_dport = tp->t_port;
12571 			rack->r_ctl.fsb.udp = udp;
12572 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12573 		} else
12574 		{
12575 			rack->r_ctl.fsb.udp = NULL;
12576 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
12577 		}
12578 		tcpip_fillheaders(rack->rc_inp,
12579 				  tp->t_port,
12580 				  ip, rack->r_ctl.fsb.th);
12581 	}
12582 	rack->r_fsb_inited = 1;
12583 }
12584 
12585 static int
12586 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12587 {
12588 	/*
12589 	 * Allocate the larger of spaces V6 if available else just
12590 	 * V4 and include udphdr (overbook)
12591 	 */
12592 #ifdef INET6
12593 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12594 #else
12595 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12596 #endif
12597 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12598 					    M_TCPFSB, M_NOWAIT|M_ZERO);
12599 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12600 		return (ENOMEM);
12601 	}
12602 	rack->r_fsb_inited = 0;
12603 	return (0);
12604 }
12605 
12606 static int
12607 rack_init(struct tcpcb *tp)
12608 {
12609 	struct tcp_rack *rack = NULL;
12610 	struct rack_sendmap *insret;
12611 	uint32_t iwin, snt, us_cts;
12612 	int err;
12613 
12614 	tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12615 	if (tp->t_fb_ptr == NULL) {
12616 		/*
12617 		 * We need to allocate memory but cant. The INP and INP_INFO
12618 		 * locks and they are recusive (happens during setup. So a
12619 		 * scheme to drop the locks fails :(
12620 		 *
12621 		 */
12622 		return (ENOMEM);
12623 	}
12624 	memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12625 
12626 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12627 	RB_INIT(&rack->r_ctl.rc_mtree);
12628 	TAILQ_INIT(&rack->r_ctl.rc_free);
12629 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
12630 	rack->rc_tp = tp;
12631 	rack->rc_inp = tp->t_inpcb;
12632 	/* Set the flag */
12633 	rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12634 	/* Probably not needed but lets be sure */
12635 	rack_clear_rate_sample(rack);
12636 	/*
12637 	 * Save off the default values, socket options will poke
12638 	 * at these if pacing is not on or we have not yet
12639 	 * reached where pacing is on (gp_ready/fixed enabled).
12640 	 * When they get set into the CC module (when gp_ready
12641 	 * is enabled or we enable fixed) then we will set these
12642 	 * values into the CC and place in here the old values
12643 	 * so we have a restoral. Then we will set the flag
12644 	 * rc_pacing_cc_set. That way whenever we turn off pacing
12645 	 * or switch off this stack, we will know to go restore
12646 	 * the saved values.
12647 	 */
12648 	rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12649 	rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12650 	/* We want abe like behavior as well */
12651 	rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
12652 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12653 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12654 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12655 	rack->r_ctl.roundends = tp->snd_max;
12656 	if (use_rack_rr)
12657 		rack->use_rack_rr = 1;
12658 	if (V_tcp_delack_enabled)
12659 		tp->t_delayed_ack = 1;
12660 	else
12661 		tp->t_delayed_ack = 0;
12662 #ifdef TCP_ACCOUNTING
12663 	if (rack_tcp_accounting) {
12664 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12665 	}
12666 #endif
12667 	if (rack_enable_shared_cwnd)
12668 		rack->rack_enable_scwnd = 1;
12669 	rack->rc_user_set_max_segs = rack_hptsi_segments;
12670 	rack->rc_force_max_seg = 0;
12671 	if (rack_use_imac_dack)
12672 		rack->rc_dack_mode = 1;
12673 	TAILQ_INIT(&rack->r_ctl.opt_list);
12674 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12675 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12676 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12677 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12678 	rack->r_ctl.rc_highest_us_rtt = 0;
12679 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12680 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12681 	if (rack_use_cmp_acks)
12682 		rack->r_use_cmp_ack = 1;
12683 	if (rack_disable_prr)
12684 		rack->rack_no_prr = 1;
12685 	if (rack_gp_no_rec_chg)
12686 		rack->rc_gp_no_rec_chg = 1;
12687 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12688 		rack->rc_always_pace = 1;
12689 		if (rack->use_fixed_rate || rack->gp_ready)
12690 			rack_set_cc_pacing(rack);
12691 	} else
12692 		rack->rc_always_pace = 0;
12693 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12694 		rack->r_mbuf_queue = 1;
12695 	else
12696 		rack->r_mbuf_queue = 0;
12697 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12698 		tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12699 	else
12700 		tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12701 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
12702 	if (rack_limits_scwnd)
12703 		rack->r_limit_scw = 1;
12704 	else
12705 		rack->r_limit_scw = 0;
12706 	rack->rc_labc = V_tcp_abc_l_var;
12707 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12708 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12709 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12710 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12711 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12712 	rack->r_ctl.rc_min_to = rack_min_to;
12713 	microuptime(&rack->r_ctl.act_rcv_time);
12714 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12715 	rack->r_running_late = 0;
12716 	rack->r_running_early = 0;
12717 	rack->rc_init_win = rack_default_init_window;
12718 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12719 	if (rack_hw_up_only)
12720 		rack->r_up_only = 1;
12721 	if (rack_do_dyn_mul) {
12722 		/* When dynamic adjustment is on CA needs to start at 100% */
12723 		rack->rc_gp_dyn_mul = 1;
12724 		if (rack_do_dyn_mul >= 100)
12725 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12726 	} else
12727 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12728 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12729 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12730 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12731 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12732 				rack_probertt_filter_life);
12733 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12734 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12735 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
12736 	rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12737 	rack->r_ctl.rc_time_probertt_starts = 0;
12738 	if (rack_dsack_std_based & 0x1) {
12739 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
12740 		rack->rc_rack_tmr_std_based = 1;
12741 	}
12742 	if (rack_dsack_std_based & 0x2) {
12743 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
12744 		rack->rc_rack_use_dsack = 1;
12745 	}
12746 	/* We require at least one measurement, even if the sysctl is 0 */
12747 	if (rack_req_measurements)
12748 		rack->r_ctl.req_measurements = rack_req_measurements;
12749 	else
12750 		rack->r_ctl.req_measurements = 1;
12751 	if (rack_enable_hw_pacing)
12752 		rack->rack_hdw_pace_ena = 1;
12753 	if (rack_hw_rate_caps)
12754 		rack->r_rack_hw_rate_caps = 1;
12755 	/* Do we force on detection? */
12756 #ifdef NETFLIX_EXP_DETECTION
12757 	if (tcp_force_detection)
12758 		rack->do_detection = 1;
12759 	else
12760 #endif
12761 		rack->do_detection = 0;
12762 	if (rack_non_rxt_use_cr)
12763 		rack->rack_rec_nonrxt_use_cr = 1;
12764 	err = rack_init_fsb(tp, rack);
12765 	if (err) {
12766 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12767 		tp->t_fb_ptr = NULL;
12768 		return (err);
12769 	}
12770 	if (tp->snd_una != tp->snd_max) {
12771 		/* Create a send map for the current outstanding data */
12772 		struct rack_sendmap *rsm;
12773 
12774 		rsm = rack_alloc(rack);
12775 		if (rsm == NULL) {
12776 			uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12777 			tp->t_fb_ptr = NULL;
12778 			return (ENOMEM);
12779 		}
12780 		rsm->r_no_rtt_allowed = 1;
12781 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12782 		rsm->r_rtr_cnt = 1;
12783 		rsm->r_rtr_bytes = 0;
12784 		if (tp->t_flags & TF_SENTFIN) {
12785 			rsm->r_end = tp->snd_max - 1;
12786 			rsm->r_flags |= RACK_HAS_FIN;
12787 		} else {
12788 			rsm->r_end = tp->snd_max;
12789 		}
12790 		if (tp->snd_una == tp->iss) {
12791 			/* The data space is one beyond snd_una */
12792 			rsm->r_flags |= RACK_HAS_SYN;
12793 			rsm->r_start = tp->iss;
12794 			rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
12795 		} else
12796 			rsm->r_start = tp->snd_una;
12797 		rsm->r_dupack = 0;
12798 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12799 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12800 			if (rsm->m)
12801 				rsm->orig_m_len = rsm->m->m_len;
12802 			else
12803 				rsm->orig_m_len = 0;
12804 		} else {
12805 			/*
12806 			 * This can happen if we have a stand-alone FIN or
12807 			 *  SYN.
12808 			 */
12809 			rsm->m = NULL;
12810 			rsm->orig_m_len = 0;
12811 			rsm->soff = 0;
12812 		}
12813 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12814 #ifdef INVARIANTS
12815 		if (insret != NULL) {
12816 			panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12817 			      insret, rack, rsm);
12818 		}
12819 #endif
12820 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12821 		rsm->r_in_tmap = 1;
12822 	}
12823 	/*
12824 	 * Timers in Rack are kept in microseconds so lets
12825 	 * convert any initial incoming variables
12826 	 * from ticks into usecs. Note that we
12827 	 * also change the values of t_srtt and t_rttvar, if
12828 	 * they are non-zero. They are kept with a 5
12829 	 * bit decimal so we have to carefully convert
12830 	 * these to get the full precision.
12831 	 */
12832 	rack_convert_rtts(tp);
12833 	tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12834 	if (rack_do_hystart) {
12835 		struct sockopt sopt;
12836 		struct cc_newreno_opts opt;
12837 
12838 		sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
12839 		sopt.sopt_dir = SOPT_SET;
12840 		opt.name = CC_NEWRENO_ENABLE_HYSTART;
12841 		opt.val = rack_do_hystart;
12842 		if (CC_ALGO(tp)->ctl_output != NULL)
12843 			(void)CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
12844 	}
12845 	if (rack_def_profile)
12846 		rack_set_profile(rack, rack_def_profile);
12847 	/* Cancel the GP measurement in progress */
12848 	tp->t_flags &= ~TF_GPUTINPROG;
12849 	if (SEQ_GT(tp->snd_max, tp->iss))
12850 		snt = tp->snd_max - tp->iss;
12851 	else
12852 		snt = 0;
12853 	iwin = rc_init_window(rack);
12854 	if (snt < iwin) {
12855 		/* We are not past the initial window
12856 		 * so we need to make sure cwnd is
12857 		 * correct.
12858 		 */
12859 		if (tp->snd_cwnd < iwin)
12860 			tp->snd_cwnd = iwin;
12861 		/*
12862 		 * If we are within the initial window
12863 		 * we want ssthresh to be unlimited. Setting
12864 		 * it to the rwnd (which the default stack does
12865 		 * and older racks) is not really a good idea
12866 		 * since we want to be in SS and grow both the
12867 		 * cwnd and the rwnd (via dynamic rwnd growth). If
12868 		 * we set it to the rwnd then as the peer grows its
12869 		 * rwnd we will be stuck in CA and never hit SS.
12870 		 *
12871 		 * Its far better to raise it up high (this takes the
12872 		 * risk that there as been a loss already, probably
12873 		 * we should have an indicator in all stacks of loss
12874 		 * but we don't), but considering the normal use this
12875 		 * is a risk worth taking. The consequences of not
12876 		 * hitting SS are far worse than going one more time
12877 		 * into it early on (before we have sent even a IW).
12878 		 * It is highly unlikely that we will have had a loss
12879 		 * before getting the IW out.
12880 		 */
12881 		tp->snd_ssthresh = 0xffffffff;
12882 	}
12883 	rack_stop_all_timers(tp);
12884 	/* Lets setup the fsb block */
12885 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12886 	rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12887 			     __LINE__, RACK_RTTS_INIT);
12888 	return (0);
12889 }
12890 
12891 static int
12892 rack_handoff_ok(struct tcpcb *tp)
12893 {
12894 	if ((tp->t_state == TCPS_CLOSED) ||
12895 	    (tp->t_state == TCPS_LISTEN)) {
12896 		/* Sure no problem though it may not stick */
12897 		return (0);
12898 	}
12899 	if ((tp->t_state == TCPS_SYN_SENT) ||
12900 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
12901 		/*
12902 		 * We really don't know if you support sack,
12903 		 * you have to get to ESTAB or beyond to tell.
12904 		 */
12905 		return (EAGAIN);
12906 	}
12907 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12908 		/*
12909 		 * Rack will only send a FIN after all data is acknowledged.
12910 		 * So in this case we have more data outstanding. We can't
12911 		 * switch stacks until either all data and only the FIN
12912 		 * is left (in which case rack_init() now knows how
12913 		 * to deal with that) <or> all is acknowledged and we
12914 		 * are only left with incoming data, though why you
12915 		 * would want to switch to rack after all data is acknowledged
12916 		 * I have no idea (rrs)!
12917 		 */
12918 		return (EAGAIN);
12919 	}
12920 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12921 		return (0);
12922 	}
12923 	/*
12924 	 * If we reach here we don't do SACK on this connection so we can
12925 	 * never do rack.
12926 	 */
12927 	return (EINVAL);
12928 }
12929 
12930 
12931 static void
12932 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12933 {
12934 	int ack_cmp = 0;
12935 
12936 	if (tp->t_fb_ptr) {
12937 		struct tcp_rack *rack;
12938 		struct rack_sendmap *rsm, *nrsm, *rm;
12939 
12940 		rack = (struct tcp_rack *)tp->t_fb_ptr;
12941 		if (tp->t_in_pkt) {
12942 			/*
12943 			 * It is unsafe to process the packets since a
12944 			 * reset may be lurking in them (its rare but it
12945 			 * can occur). If we were to find a RST, then we
12946 			 * would end up dropping the connection and the
12947 			 * INP lock, so when we return the caller (tcp_usrreq)
12948 			 * will blow up when it trys to unlock the inp.
12949 			 */
12950 			struct mbuf *save, *m;
12951 
12952 			m = tp->t_in_pkt;
12953 			tp->t_in_pkt = NULL;
12954 			tp->t_tail_pkt = NULL;
12955 			while (m) {
12956 				save = m->m_nextpkt;
12957 				m->m_nextpkt = NULL;
12958 				m_freem(m);
12959 				m = save;
12960 			}
12961 			if ((tp->t_inpcb) &&
12962 			    (tp->t_inpcb->inp_flags2 & INP_MBUF_ACKCMP))
12963 				ack_cmp = 1;
12964 			if (ack_cmp) {
12965 				/* Total if we used large or small (if ack-cmp was used). */
12966 				if (rack->rc_inp->inp_flags2 & INP_MBUF_L_ACKS)
12967 					counter_u64_add(rack_large_ackcmp, 1);
12968 				else
12969 					counter_u64_add(rack_small_ackcmp, 1);
12970 			}
12971 		}
12972 		tp->t_flags &= ~TF_FORCEDATA;
12973 #ifdef NETFLIX_SHARED_CWND
12974 		if (rack->r_ctl.rc_scw) {
12975 			uint32_t limit;
12976 
12977 			if (rack->r_limit_scw)
12978 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12979 			else
12980 				limit = 0;
12981 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12982 						  rack->r_ctl.rc_scw_index,
12983 						  limit);
12984 			rack->r_ctl.rc_scw = NULL;
12985 		}
12986 #endif
12987 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
12988 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12989 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12990 			rack->r_ctl.fsb.th = NULL;
12991 		}
12992 		/* Convert back to ticks, with  */
12993 		if (tp->t_srtt > 1) {
12994 			uint32_t val, frac;
12995 
12996 			val = USEC_2_TICKS(tp->t_srtt);
12997 			frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12998 			tp->t_srtt = val << TCP_RTT_SHIFT;
12999 			/*
13000 			 * frac is the fractional part here is left
13001 			 * over from converting to hz and shifting.
13002 			 * We need to convert this to the 5 bit
13003 			 * remainder.
13004 			 */
13005 			if (frac) {
13006 				if (hz == 1000) {
13007 					frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
13008 				} else {
13009 					frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
13010 				}
13011 				tp->t_srtt += frac;
13012 			}
13013 		}
13014 		if (tp->t_rttvar) {
13015 			uint32_t val, frac;
13016 
13017 			val = USEC_2_TICKS(tp->t_rttvar);
13018 			frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
13019 			tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
13020 			/*
13021 			 * frac is the fractional part here is left
13022 			 * over from converting to hz and shifting.
13023 			 * We need to convert this to the 5 bit
13024 			 * remainder.
13025 			 */
13026 			if (frac) {
13027 				if (hz == 1000) {
13028 					frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
13029 				} else {
13030 					frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
13031 				}
13032 				tp->t_rttvar += frac;
13033 			}
13034 		}
13035 		tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
13036 		tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
13037 		if (rack->rc_always_pace) {
13038 			tcp_decrement_paced_conn();
13039 			rack_undo_cc_pacing(rack);
13040 			rack->rc_always_pace = 0;
13041 		}
13042 		/* Clean up any options if they were not applied */
13043 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
13044 			struct deferred_opt_list *dol;
13045 
13046 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
13047 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
13048 			free(dol, M_TCPDO);
13049 		}
13050 		/* rack does not use force data but other stacks may clear it */
13051 		if (rack->r_ctl.crte != NULL) {
13052 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
13053 			rack->rack_hdrw_pacing = 0;
13054 			rack->r_ctl.crte = NULL;
13055 		}
13056 #ifdef TCP_BLACKBOX
13057 		tcp_log_flowend(tp);
13058 #endif
13059 		RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
13060 			rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
13061 #ifdef INVARIANTS
13062 			if (rm != rsm) {
13063 				panic("At fini, rack:%p rsm:%p rm:%p",
13064 				      rack, rsm, rm);
13065 			}
13066 #endif
13067 			uma_zfree(rack_zone, rsm);
13068 		}
13069 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
13070 		while (rsm) {
13071 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
13072 			uma_zfree(rack_zone, rsm);
13073 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
13074 		}
13075 		rack->rc_free_cnt = 0;
13076 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
13077 		tp->t_fb_ptr = NULL;
13078 	}
13079 	if (tp->t_inpcb) {
13080 		tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
13081 		tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
13082 		tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
13083 		tp->t_inpcb->inp_flags2 &= ~INP_MBUF_ACKCMP;
13084 		/* Cancel the GP measurement in progress */
13085 		tp->t_flags &= ~TF_GPUTINPROG;
13086 		tp->t_inpcb->inp_flags2 &= ~INP_MBUF_L_ACKS;
13087 	}
13088 	/* Make sure snd_nxt is correctly set */
13089 	tp->snd_nxt = tp->snd_max;
13090 }
13091 
13092 static void
13093 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
13094 {
13095 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
13096 		rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
13097 	}
13098 	switch (tp->t_state) {
13099 	case TCPS_SYN_SENT:
13100 		rack->r_state = TCPS_SYN_SENT;
13101 		rack->r_substate = rack_do_syn_sent;
13102 		break;
13103 	case TCPS_SYN_RECEIVED:
13104 		rack->r_state = TCPS_SYN_RECEIVED;
13105 		rack->r_substate = rack_do_syn_recv;
13106 		break;
13107 	case TCPS_ESTABLISHED:
13108 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13109 		rack->r_state = TCPS_ESTABLISHED;
13110 		rack->r_substate = rack_do_established;
13111 		break;
13112 	case TCPS_CLOSE_WAIT:
13113 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13114 		rack->r_state = TCPS_CLOSE_WAIT;
13115 		rack->r_substate = rack_do_close_wait;
13116 		break;
13117 	case TCPS_FIN_WAIT_1:
13118 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13119 		rack->r_state = TCPS_FIN_WAIT_1;
13120 		rack->r_substate = rack_do_fin_wait_1;
13121 		break;
13122 	case TCPS_CLOSING:
13123 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13124 		rack->r_state = TCPS_CLOSING;
13125 		rack->r_substate = rack_do_closing;
13126 		break;
13127 	case TCPS_LAST_ACK:
13128 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13129 		rack->r_state = TCPS_LAST_ACK;
13130 		rack->r_substate = rack_do_lastack;
13131 		break;
13132 	case TCPS_FIN_WAIT_2:
13133 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13134 		rack->r_state = TCPS_FIN_WAIT_2;
13135 		rack->r_substate = rack_do_fin_wait_2;
13136 		break;
13137 	case TCPS_LISTEN:
13138 	case TCPS_CLOSED:
13139 	case TCPS_TIME_WAIT:
13140 	default:
13141 		break;
13142 	};
13143 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
13144 		rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
13145 
13146 }
13147 
13148 static void
13149 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
13150 {
13151 	/*
13152 	 * We received an ack, and then did not
13153 	 * call send or were bounced out due to the
13154 	 * hpts was running. Now a timer is up as well, is
13155 	 * it the right timer?
13156 	 */
13157 	struct rack_sendmap *rsm;
13158 	int tmr_up;
13159 
13160 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
13161 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
13162 		return;
13163 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
13164 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
13165 	    (tmr_up == PACE_TMR_RXT)) {
13166 		/* Should be an RXT */
13167 		return;
13168 	}
13169 	if (rsm == NULL) {
13170 		/* Nothing outstanding? */
13171 		if (tp->t_flags & TF_DELACK) {
13172 			if (tmr_up == PACE_TMR_DELACK)
13173 				/* We are supposed to have delayed ack up and we do */
13174 				return;
13175 		} else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) {
13176 			/*
13177 			 * if we hit enobufs then we would expect the possiblity
13178 			 * of nothing outstanding and the RXT up (and the hptsi timer).
13179 			 */
13180 			return;
13181 		} else if (((V_tcp_always_keepalive ||
13182 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
13183 			    (tp->t_state <= TCPS_CLOSING)) &&
13184 			   (tmr_up == PACE_TMR_KEEP) &&
13185 			   (tp->snd_max == tp->snd_una)) {
13186 			/* We should have keep alive up and we do */
13187 			return;
13188 		}
13189 	}
13190 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
13191 		   ((tmr_up == PACE_TMR_TLP) ||
13192 		    (tmr_up == PACE_TMR_RACK) ||
13193 		    (tmr_up == PACE_TMR_RXT))) {
13194 		/*
13195 		 * Either a Rack, TLP or RXT is fine if  we
13196 		 * have outstanding data.
13197 		 */
13198 		return;
13199 	} else if (tmr_up == PACE_TMR_DELACK) {
13200 		/*
13201 		 * If the delayed ack was going to go off
13202 		 * before the rtx/tlp/rack timer were going to
13203 		 * expire, then that would be the timer in control.
13204 		 * Note we don't check the time here trusting the
13205 		 * code is correct.
13206 		 */
13207 		return;
13208 	}
13209 	/*
13210 	 * Ok the timer originally started is not what we want now.
13211 	 * We will force the hpts to be stopped if any, and restart
13212 	 * with the slot set to what was in the saved slot.
13213 	 */
13214 	if (tcp_in_hpts(rack->rc_inp)) {
13215 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
13216 			uint32_t us_cts;
13217 
13218 			us_cts = tcp_get_usecs(NULL);
13219 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
13220 				rack->r_early = 1;
13221 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
13222 			}
13223 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
13224 		}
13225 		tcp_hpts_remove(tp->t_inpcb);
13226 	}
13227 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13228 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13229 }
13230 
13231 
13232 static void
13233 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)
13234 {
13235 	if ((SEQ_LT(tp->snd_wl1, seq) ||
13236 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
13237 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
13238 		/* keep track of pure window updates */
13239 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
13240 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
13241 		tp->snd_wnd = tiwin;
13242 		rack_validate_fo_sendwin_up(tp, rack);
13243 		tp->snd_wl1 = seq;
13244 		tp->snd_wl2 = ack;
13245 		if (tp->snd_wnd > tp->max_sndwnd)
13246 			tp->max_sndwnd = tp->snd_wnd;
13247 	    rack->r_wanted_output = 1;
13248 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
13249 		tp->snd_wnd = tiwin;
13250 		rack_validate_fo_sendwin_up(tp, rack);
13251 		tp->snd_wl1 = seq;
13252 		tp->snd_wl2 = ack;
13253 	} else {
13254 		/* Not a valid win update */
13255 		return;
13256 	}
13257 	if (tp->snd_wnd > tp->max_sndwnd)
13258 		tp->max_sndwnd = tp->snd_wnd;
13259 	if (tp->snd_wnd < (tp->snd_max - high_seq)) {
13260 		/* The peer collapsed the window */
13261 		rack_collapsed_window(rack);
13262 	} else if (rack->rc_has_collapsed)
13263 		rack_un_collapse_window(rack);
13264 	/* Do we exit persists? */
13265 	if ((rack->rc_in_persist != 0) &&
13266 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13267 				rack->r_ctl.rc_pace_min_segs))) {
13268 		rack_exit_persist(tp, rack, cts);
13269 	}
13270 	/* Do we enter persists? */
13271 	if ((rack->rc_in_persist == 0) &&
13272 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13273 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
13274 	    (tp->snd_max == tp->snd_una) &&
13275 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
13276 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
13277 		/*
13278 		 * Here the rwnd is less than
13279 		 * the pacing size, we are established,
13280 		 * nothing is outstanding, and there is
13281 		 * data to send. Enter persists.
13282 		 */
13283 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
13284 	}
13285 }
13286 
13287 static void
13288 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
13289 {
13290 
13291 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13292 		union tcp_log_stackspecific log;
13293 		struct timeval ltv;
13294 		char tcp_hdr_buf[60];
13295 		struct tcphdr *th;
13296 		struct timespec ts;
13297 		uint32_t orig_snd_una;
13298 		uint8_t xx = 0;
13299 
13300 #ifdef NETFLIX_HTTP_LOGGING
13301 		struct http_sendfile_track *http_req;
13302 
13303 		if (SEQ_GT(ae->ack, tp->snd_una)) {
13304 			http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
13305 		} else {
13306 			http_req = tcp_http_find_req_for_seq(tp, ae->ack);
13307 		}
13308 #endif
13309 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13310 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
13311 		if (rack->rack_no_prr == 0)
13312 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13313 		else
13314 			log.u_bbr.flex1 = 0;
13315 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13316 		log.u_bbr.use_lt_bw <<= 1;
13317 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
13318 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13319 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
13320 		log.u_bbr.pkts_out = tp->t_maxseg;
13321 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13322 		log.u_bbr.flex7 = 1;
13323 		log.u_bbr.lost = ae->flags;
13324 		log.u_bbr.cwnd_gain = ackval;
13325 		log.u_bbr.pacing_gain = 0x2;
13326 		if (ae->flags & TSTMP_HDWR) {
13327 			/* Record the hardware timestamp if present */
13328 			log.u_bbr.flex3 = M_TSTMP;
13329 			ts.tv_sec = ae->timestamp / 1000000000;
13330 			ts.tv_nsec = ae->timestamp % 1000000000;
13331 			ltv.tv_sec = ts.tv_sec;
13332 			ltv.tv_usec = ts.tv_nsec / 1000;
13333 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13334 		} else if (ae->flags & TSTMP_LRO) {
13335 			/* Record the LRO the arrival timestamp */
13336 			log.u_bbr.flex3 = M_TSTMP_LRO;
13337 			ts.tv_sec = ae->timestamp / 1000000000;
13338 			ts.tv_nsec = ae->timestamp % 1000000000;
13339 			ltv.tv_sec = ts.tv_sec;
13340 			ltv.tv_usec = ts.tv_nsec / 1000;
13341 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13342 		}
13343 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13344 		/* Log the rcv time */
13345 		log.u_bbr.delRate = ae->timestamp;
13346 #ifdef NETFLIX_HTTP_LOGGING
13347 		log.u_bbr.applimited = tp->t_http_closed;
13348 		log.u_bbr.applimited <<= 8;
13349 		log.u_bbr.applimited |= tp->t_http_open;
13350 		log.u_bbr.applimited <<= 8;
13351 		log.u_bbr.applimited |= tp->t_http_req;
13352 		if (http_req) {
13353 			/* Copy out any client req info */
13354 			/* seconds */
13355 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13356 			/* useconds */
13357 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13358 			log.u_bbr.rttProp = http_req->timestamp;
13359 			log.u_bbr.cur_del_rate = http_req->start;
13360 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13361 				log.u_bbr.flex8 |= 1;
13362 			} else {
13363 				log.u_bbr.flex8 |= 2;
13364 				log.u_bbr.bw_inuse = http_req->end;
13365 			}
13366 			log.u_bbr.flex6 = http_req->start_seq;
13367 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13368 				log.u_bbr.flex8 |= 4;
13369 				log.u_bbr.epoch = http_req->end_seq;
13370 			}
13371 		}
13372 #endif
13373 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
13374 		th = (struct tcphdr *)tcp_hdr_buf;
13375 		th->th_seq = ae->seq;
13376 		th->th_ack = ae->ack;
13377 		th->th_win = ae->win;
13378 		/* Now fill in the ports */
13379 		th->th_sport = tp->t_inpcb->inp_fport;
13380 		th->th_dport = tp->t_inpcb->inp_lport;
13381 		th->th_flags = ae->flags & 0xff;
13382 		/* Now do we have a timestamp option? */
13383 		if (ae->flags & HAS_TSTMP) {
13384 			u_char *cp;
13385 			uint32_t val;
13386 
13387 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
13388 			cp = (u_char *)(th + 1);
13389 			*cp = TCPOPT_NOP;
13390 			cp++;
13391 			*cp = TCPOPT_NOP;
13392 			cp++;
13393 			*cp = TCPOPT_TIMESTAMP;
13394 			cp++;
13395 			*cp = TCPOLEN_TIMESTAMP;
13396 			cp++;
13397 			val = htonl(ae->ts_value);
13398 			bcopy((char *)&val,
13399 			      (char *)cp, sizeof(uint32_t));
13400 			val = htonl(ae->ts_echo);
13401 			bcopy((char *)&val,
13402 			      (char *)(cp + 4), sizeof(uint32_t));
13403 		} else
13404 			th->th_off = (sizeof(struct tcphdr) >> 2);
13405 
13406 		/*
13407 		 * For sane logging we need to play a little trick.
13408 		 * If the ack were fully processed we would have moved
13409 		 * snd_una to high_seq, but since compressed acks are
13410 		 * processed in two phases, at this point (logging) snd_una
13411 		 * won't be advanced. So we would see multiple acks showing
13412 		 * the advancement. We can prevent that by "pretending" that
13413 		 * snd_una was advanced and then un-advancing it so that the
13414 		 * logging code has the right value for tlb_snd_una.
13415 		 */
13416 		if (tp->snd_una != high_seq) {
13417 			orig_snd_una = tp->snd_una;
13418 			tp->snd_una = high_seq;
13419 			xx = 1;
13420 		} else
13421 			xx = 0;
13422 		TCP_LOG_EVENTP(tp, th,
13423 			       &tp->t_inpcb->inp_socket->so_rcv,
13424 			       &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_IN, 0,
13425 			       0, &log, true, &ltv);
13426 		if (xx) {
13427 			tp->snd_una = orig_snd_una;
13428 		}
13429 	}
13430 
13431 }
13432 
13433 static void
13434 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
13435 {
13436 	uint32_t us_rtt;
13437 	/*
13438 	 * A persist or keep-alive was forced out, update our
13439 	 * min rtt time. Note now worry about lost responses.
13440 	 * When a subsequent keep-alive or persist times out
13441 	 * and forced_ack is still on, then the last probe
13442 	 * was not responded to. In such cases we have a
13443 	 * sysctl that controls the behavior. Either we apply
13444 	 * the rtt but with reduced confidence (0). Or we just
13445 	 * plain don't apply the rtt estimate. Having data flow
13446 	 * will clear the probe_not_answered flag i.e. cum-ack
13447 	 * move forward <or> exiting and reentering persists.
13448 	 */
13449 
13450 	rack->forced_ack = 0;
13451 	rack->rc_tp->t_rxtshift = 0;
13452 	if ((rack->rc_in_persist &&
13453 	     (tiwin == rack->rc_tp->snd_wnd)) ||
13454 	    (rack->rc_in_persist == 0)) {
13455 		/*
13456 		 * In persists only apply the RTT update if this is
13457 		 * a response to our window probe. And that
13458 		 * means the rwnd sent must match the current
13459 		 * snd_wnd. If it does not, then we got a
13460 		 * window update ack instead. For keepalive
13461 		 * we allow the answer no matter what the window.
13462 		 *
13463 		 * Note that if the probe_not_answered is set then
13464 		 * the forced_ack_ts is the oldest one i.e. the first
13465 		 * probe sent that might have been lost. This assures
13466 		 * us that if we do calculate an RTT it is longer not
13467 		 * some short thing.
13468 		 */
13469 		if (rack->rc_in_persist)
13470 			counter_u64_add(rack_persists_acks, 1);
13471 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13472 		if (us_rtt == 0)
13473 			us_rtt = 1;
13474 		if (rack->probe_not_answered == 0) {
13475 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13476 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
13477 		} else {
13478 			/* We have a retransmitted probe here too */
13479 			if (rack_apply_rtt_with_reduced_conf) {
13480 				rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13481 				tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
13482 			}
13483 		}
13484 	}
13485 }
13486 
13487 
13488 static int
13489 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
13490 {
13491 	/*
13492 	 * Handle a "special" compressed ack mbuf. Each incoming
13493 	 * ack has only four possible dispositions:
13494 	 *
13495 	 * A) It moves the cum-ack forward
13496 	 * B) It is behind the cum-ack.
13497 	 * C) It is a window-update ack.
13498 	 * D) It is a dup-ack.
13499 	 *
13500 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
13501 	 * in the incoming mbuf. We also need to still pay attention
13502 	 * to nxt_pkt since there may be another packet after this
13503 	 * one.
13504 	 */
13505 #ifdef TCP_ACCOUNTING
13506 	uint64_t ts_val;
13507 	uint64_t rdstc;
13508 #endif
13509 	int segsiz;
13510 	struct timespec ts;
13511 	struct tcp_rack *rack;
13512 	struct tcp_ackent *ae;
13513 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
13514 	int cnt, i, did_out, ourfinisacked = 0;
13515 	struct tcpopt to_holder, *to = NULL;
13516 	int win_up_req = 0;
13517 	int nsegs = 0;
13518 	int under_pacing = 1;
13519 	int recovery = 0;
13520 	int idx;
13521 #ifdef TCP_ACCOUNTING
13522 	sched_pin();
13523 #endif
13524 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13525 	if (rack->gp_ready &&
13526 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
13527 		under_pacing = 0;
13528 	else
13529 		under_pacing = 1;
13530 
13531 	if (rack->r_state != tp->t_state)
13532 		rack_set_state(tp, rack);
13533 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13534 	    (tp->t_flags & TF_GPUTINPROG)) {
13535 		/*
13536 		 * We have a goodput in progress
13537 		 * and we have entered a late state.
13538 		 * Do we have enough data in the sb
13539 		 * to handle the GPUT request?
13540 		 */
13541 		uint32_t bytes;
13542 
13543 		bytes = tp->gput_ack - tp->gput_seq;
13544 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
13545 			bytes += tp->gput_seq - tp->snd_una;
13546 		if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
13547 			/*
13548 			 * There are not enough bytes in the socket
13549 			 * buffer that have been sent to cover this
13550 			 * measurement. Cancel it.
13551 			 */
13552 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
13553 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
13554 						   tp->gput_seq,
13555 						   0, 0, 18, __LINE__, NULL, 0);
13556 			tp->t_flags &= ~TF_GPUTINPROG;
13557 		}
13558 	}
13559 	to = &to_holder;
13560 	to->to_flags = 0;
13561 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
13562 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
13563 	cnt = m->m_len / sizeof(struct tcp_ackent);
13564 	idx = cnt / 5;
13565 	if (idx >= MAX_NUM_OF_CNTS)
13566 		idx = MAX_NUM_OF_CNTS - 1;
13567 	counter_u64_add(rack_proc_comp_ack[idx], 1);
13568 	counter_u64_add(rack_multi_single_eq, cnt);
13569 	high_seq = tp->snd_una;
13570 	the_win = tp->snd_wnd;
13571 	win_seq = tp->snd_wl1;
13572 	win_upd_ack = tp->snd_wl2;
13573 	cts = tcp_tv_to_usectick(tv);
13574 	ms_cts = tcp_tv_to_mssectick(tv);
13575 	segsiz = ctf_fixed_maxseg(tp);
13576 	if ((rack->rc_gp_dyn_mul) &&
13577 	    (rack->use_fixed_rate == 0) &&
13578 	    (rack->rc_always_pace)) {
13579 		/* Check in on probertt */
13580 		rack_check_probe_rtt(rack, cts);
13581 	}
13582 	for (i = 0; i < cnt; i++) {
13583 #ifdef TCP_ACCOUNTING
13584 		ts_val = get_cyclecount();
13585 #endif
13586 		rack_clear_rate_sample(rack);
13587 		ae = ((mtod(m, struct tcp_ackent *)) + i);
13588 		/* Setup the window */
13589 		tiwin = ae->win << tp->snd_scale;
13590 		/* figure out the type of ack */
13591 		if (SEQ_LT(ae->ack, high_seq)) {
13592 			/* Case B*/
13593 			ae->ack_val_set = ACK_BEHIND;
13594 		} else if (SEQ_GT(ae->ack, high_seq)) {
13595 			/* Case A */
13596 			ae->ack_val_set = ACK_CUMACK;
13597 		} else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
13598 			/* Case D */
13599 			ae->ack_val_set = ACK_DUPACK;
13600 		} else {
13601 			/* Case C */
13602 			ae->ack_val_set = ACK_RWND;
13603 		}
13604 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
13605 		/* Validate timestamp */
13606 		if (ae->flags & HAS_TSTMP) {
13607 			/* Setup for a timestamp */
13608 			to->to_flags = TOF_TS;
13609 			ae->ts_echo -= tp->ts_offset;
13610 			to->to_tsecr = ae->ts_echo;
13611 			to->to_tsval = ae->ts_value;
13612 			/*
13613 			 * If echoed timestamp is later than the current time, fall back to
13614 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13615 			 * were used when this connection was established.
13616 			 */
13617 			if (TSTMP_GT(ae->ts_echo, ms_cts))
13618 				to->to_tsecr = 0;
13619 			if (tp->ts_recent &&
13620 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
13621 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
13622 #ifdef TCP_ACCOUNTING
13623 					rdstc = get_cyclecount();
13624 					if (rdstc > ts_val) {
13625 						counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
13626 								(rdstc - ts_val));
13627 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13628 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13629 						}
13630 					}
13631 #endif
13632 					continue;
13633 				}
13634 			}
13635 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
13636 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
13637 				tp->ts_recent_age = tcp_ts_getticks();
13638 				tp->ts_recent = ae->ts_value;
13639 			}
13640 		} else {
13641 			/* Setup for a no options */
13642 			to->to_flags = 0;
13643 		}
13644 		/* Update the rcv time and perform idle reduction possibly */
13645 		if  (tp->t_idle_reduce &&
13646 		     (tp->snd_max == tp->snd_una) &&
13647 		     ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13648 			counter_u64_add(rack_input_idle_reduces, 1);
13649 			rack_cc_after_idle(rack, tp);
13650 		}
13651 		tp->t_rcvtime = ticks;
13652 		/* Now what about ECN? */
13653 		if (tp->t_flags2 & TF2_ECN_PERMIT) {
13654 			if (ae->flags & TH_CWR) {
13655 				tp->t_flags2 &= ~TF2_ECN_SND_ECE;
13656 				tp->t_flags |= TF_ACKNOW;
13657 			}
13658 			switch (ae->codepoint & IPTOS_ECN_MASK) {
13659 			case IPTOS_ECN_CE:
13660 				tp->t_flags2 |= TF2_ECN_SND_ECE;
13661 				KMOD_TCPSTAT_INC(tcps_ecn_ce);
13662 				break;
13663 			case IPTOS_ECN_ECT0:
13664 				KMOD_TCPSTAT_INC(tcps_ecn_ect0);
13665 				break;
13666 			case IPTOS_ECN_ECT1:
13667 				KMOD_TCPSTAT_INC(tcps_ecn_ect1);
13668 				break;
13669 			}
13670 
13671 			/* Process a packet differently from RFC3168. */
13672 			cc_ecnpkt_handler_flags(tp, ae->flags, ae->codepoint);
13673 			/* Congestion experienced. */
13674 			if (ae->flags & TH_ECE) {
13675 				rack_cong_signal(tp,  CC_ECN, ae->ack);
13676 			}
13677 		}
13678 #ifdef TCP_ACCOUNTING
13679 		/* Count for the specific type of ack in */
13680 		counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
13681 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13682 			tp->tcp_cnt_counters[ae->ack_val_set]++;
13683 		}
13684 #endif
13685 		/*
13686 		 * Note how we could move up these in the determination
13687 		 * above, but we don't so that way the timestamp checks (and ECN)
13688 		 * is done first before we do any processing on the ACK.
13689 		 * The non-compressed path through the code has this
13690 		 * weakness (noted by @jtl) that it actually does some
13691 		 * processing before verifying the timestamp information.
13692 		 * We don't take that path here which is why we set
13693 		 * the ack_val_set first, do the timestamp and ecn
13694 		 * processing, and then look at what we have setup.
13695 		 */
13696 		if (ae->ack_val_set == ACK_BEHIND) {
13697 			/*
13698 			 * Case B flag reordering, if window is not closed
13699 			 * or it could be a keep-alive or persists
13700 			 */
13701 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
13702 				counter_u64_add(rack_reorder_seen, 1);
13703 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13704 			}
13705 		} else if (ae->ack_val_set == ACK_DUPACK) {
13706 			/* Case D */
13707 			rack_strike_dupack(rack);
13708 		} else if (ae->ack_val_set == ACK_RWND) {
13709 			/* Case C */
13710 			if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13711 				ts.tv_sec = ae->timestamp / 1000000000;
13712 				ts.tv_nsec = ae->timestamp % 1000000000;
13713 				rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13714 				rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13715 			} else {
13716 				rack->r_ctl.act_rcv_time = *tv;
13717 			}
13718 			if (rack->forced_ack) {
13719 				rack_handle_probe_response(rack, tiwin,
13720 							   tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
13721 			}
13722 			win_up_req = 1;
13723 			win_upd_ack = ae->ack;
13724 			win_seq = ae->seq;
13725 			the_win = tiwin;
13726 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13727 		} else {
13728 			/* Case A */
13729 			if (SEQ_GT(ae->ack, tp->snd_max)) {
13730 				/*
13731 				 * We just send an ack since the incoming
13732 				 * ack is beyond the largest seq we sent.
13733 				 */
13734 				if ((tp->t_flags & TF_ACKNOW) == 0) {
13735 					ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13736 					if (tp->t_flags && TF_ACKNOW)
13737 						rack->r_wanted_output = 1;
13738 				}
13739 			} else {
13740 				nsegs++;
13741 				/* If the window changed setup to update */
13742 				if (tiwin != tp->snd_wnd) {
13743 					win_upd_ack = ae->ack;
13744 					win_seq = ae->seq;
13745 					the_win = tiwin;
13746 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13747 				}
13748 #ifdef TCP_ACCOUNTING
13749 				/* Account for the acks */
13750 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13751 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13752 				}
13753 				counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13754 						(((ae->ack - high_seq) + segsiz - 1) / segsiz));
13755 #endif
13756 				high_seq = ae->ack;
13757 				if (SEQ_GEQ(high_seq, rack->r_ctl.roundends)) {
13758 					rack->r_ctl.current_round++;
13759 					rack->r_ctl.roundends = tp->snd_max;
13760 					if (CC_ALGO(tp)->newround != NULL) {
13761 						CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round);
13762 					}
13763 				}
13764 				/* Setup our act_rcv_time */
13765 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13766 					ts.tv_sec = ae->timestamp / 1000000000;
13767 					ts.tv_nsec = ae->timestamp % 1000000000;
13768 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13769 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13770 				} else {
13771 					rack->r_ctl.act_rcv_time = *tv;
13772 				}
13773 				rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13774 				if (rack->rc_dsack_round_seen) {
13775 					/* Is the dsack round over? */
13776 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
13777 						/* Yes it is */
13778 						rack->rc_dsack_round_seen = 0;
13779 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
13780 					}
13781 				}
13782 			}
13783 		}
13784 		/* And lets be sure to commit the rtt measurements for this ack */
13785 		tcp_rack_xmit_timer_commit(rack, tp);
13786 #ifdef TCP_ACCOUNTING
13787 		rdstc = get_cyclecount();
13788 		if (rdstc > ts_val) {
13789 			counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13790 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13791 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13792 				if (ae->ack_val_set == ACK_CUMACK)
13793 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13794 			}
13795 		}
13796 #endif
13797 	}
13798 #ifdef TCP_ACCOUNTING
13799 	ts_val = get_cyclecount();
13800 #endif
13801 	acked_amount = acked = (high_seq - tp->snd_una);
13802 	if (acked) {
13803 		/*
13804 		 * Clear the probe not answered flag
13805 		 * since cum-ack moved forward.
13806 		 */
13807 		rack->probe_not_answered = 0;
13808 		if (rack->sack_attack_disable == 0)
13809 			rack_do_decay(rack);
13810 		if (acked >= segsiz) {
13811 			/*
13812 			 * You only get credit for
13813 			 * MSS and greater (and you get extra
13814 			 * credit for larger cum-ack moves).
13815 			 */
13816 			int ac;
13817 
13818 			ac = acked / segsiz;
13819 			rack->r_ctl.ack_count += ac;
13820 			counter_u64_add(rack_ack_total, ac);
13821 		}
13822 		if (rack->r_ctl.ack_count > 0xfff00000) {
13823 			/*
13824 			 * reduce the number to keep us under
13825 			 * a uint32_t.
13826 			 */
13827 			rack->r_ctl.ack_count /= 2;
13828 			rack->r_ctl.sack_count /= 2;
13829 		}
13830 		if (tp->t_flags & TF_NEEDSYN) {
13831 			/*
13832 			 * T/TCP: Connection was half-synchronized, and our SYN has
13833 			 * been ACK'd (so connection is now fully synchronized).  Go
13834 			 * to non-starred state, increment snd_una for ACK of SYN,
13835 			 * and check if we can do window scaling.
13836 			 */
13837 			tp->t_flags &= ~TF_NEEDSYN;
13838 			tp->snd_una++;
13839 			acked_amount = acked = (high_seq - tp->snd_una);
13840 		}
13841 		if (acked > sbavail(&so->so_snd))
13842 			acked_amount = sbavail(&so->so_snd);
13843 #ifdef NETFLIX_EXP_DETECTION
13844 		/*
13845 		 * We only care on a cum-ack move if we are in a sack-disabled
13846 		 * state. We have already added in to the ack_count, and we never
13847 		 * would disable on a cum-ack move, so we only care to do the
13848 		 * detection if it may "undo" it, i.e. we were in disabled already.
13849 		 */
13850 		if (rack->sack_attack_disable)
13851 			rack_do_detection(tp, rack, acked_amount, segsiz);
13852 #endif
13853 		if (IN_FASTRECOVERY(tp->t_flags) &&
13854 		    (rack->rack_no_prr == 0))
13855 			rack_update_prr(tp, rack, acked_amount, high_seq);
13856 		if (IN_RECOVERY(tp->t_flags)) {
13857 			if (SEQ_LT(high_seq, tp->snd_recover) &&
13858 			    (SEQ_LT(high_seq, tp->snd_max))) {
13859 				tcp_rack_partialack(tp);
13860 			} else {
13861 				rack_post_recovery(tp, high_seq);
13862 				recovery = 1;
13863 			}
13864 		}
13865 		/* Handle the rack-log-ack part (sendmap) */
13866 		if ((sbused(&so->so_snd) == 0) &&
13867 		    (acked > acked_amount) &&
13868 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
13869 		    (tp->t_flags & TF_SENTFIN)) {
13870 			/*
13871 			 * We must be sure our fin
13872 			 * was sent and acked (we can be
13873 			 * in FIN_WAIT_1 without having
13874 			 * sent the fin).
13875 			 */
13876 			ourfinisacked = 1;
13877 			/*
13878 			 * Lets make sure snd_una is updated
13879 			 * since most likely acked_amount = 0 (it
13880 			 * should be).
13881 			 */
13882 			tp->snd_una = high_seq;
13883 		}
13884 		/* Did we make a RTO error? */
13885 		if ((tp->t_flags & TF_PREVVALID) &&
13886 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13887 			tp->t_flags &= ~TF_PREVVALID;
13888 			if (tp->t_rxtshift == 1 &&
13889 			    (int)(ticks - tp->t_badrxtwin) < 0)
13890 				rack_cong_signal(tp, CC_RTO_ERR, high_seq);
13891 		}
13892 		/* Handle the data in the socket buffer */
13893 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13894 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13895 		if (acked_amount > 0) {
13896 			struct mbuf *mfree;
13897 
13898 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13899 			SOCKBUF_LOCK(&so->so_snd);
13900 			mfree = sbcut_locked(&so->so_snd, acked_amount);
13901 			tp->snd_una = high_seq;
13902 			/* Note we want to hold the sb lock through the sendmap adjust */
13903 			rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13904 			/* Wake up the socket if we have room to write more */
13905 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13906 			sowwakeup_locked(so);
13907 			m_freem(mfree);
13908 		}
13909 		/* update progress */
13910 		tp->t_acktime = ticks;
13911 		rack_log_progress_event(rack, tp, tp->t_acktime,
13912 					PROGRESS_UPDATE, __LINE__);
13913 		/* Clear out shifts and such */
13914 		tp->t_rxtshift = 0;
13915 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13916 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13917 		rack->rc_tlp_in_progress = 0;
13918 		rack->r_ctl.rc_tlp_cnt_out = 0;
13919 		/* Send recover and snd_nxt must be dragged along */
13920 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
13921 			tp->snd_recover = tp->snd_una;
13922 		if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13923 			tp->snd_nxt = tp->snd_una;
13924 		/*
13925 		 * If the RXT timer is running we want to
13926 		 * stop it, so we can restart a TLP (or new RXT).
13927 		 */
13928 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13929 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13930 #ifdef NETFLIX_HTTP_LOGGING
13931 		tcp_http_check_for_comp(rack->rc_tp, high_seq);
13932 #endif
13933 		tp->snd_wl2 = high_seq;
13934 		tp->t_dupacks = 0;
13935 		if (under_pacing &&
13936 		    (rack->use_fixed_rate == 0) &&
13937 		    (rack->in_probe_rtt == 0) &&
13938 		    rack->rc_gp_dyn_mul &&
13939 		    rack->rc_always_pace) {
13940 			/* Check if we are dragging bottom */
13941 			rack_check_bottom_drag(tp, rack, so, acked);
13942 		}
13943 		if (tp->snd_una == tp->snd_max) {
13944 			tp->t_flags &= ~TF_PREVVALID;
13945 			rack->r_ctl.retran_during_recovery = 0;
13946 			rack->r_ctl.dsack_byte_cnt = 0;
13947 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13948 			if (rack->r_ctl.rc_went_idle_time == 0)
13949 				rack->r_ctl.rc_went_idle_time = 1;
13950 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13951 			if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
13952 				tp->t_acktime = 0;
13953 			/* Set so we might enter persists... */
13954 			rack->r_wanted_output = 1;
13955 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13956 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13957 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13958 			    (sbavail(&so->so_snd) == 0) &&
13959 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13960 				/*
13961 				 * The socket was gone and the
13962 				 * peer sent data (not now in the past), time to
13963 				 * reset him.
13964 				 */
13965 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13966 				/* tcp_close will kill the inp pre-log the Reset */
13967 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13968 #ifdef TCP_ACCOUNTING
13969 				rdstc = get_cyclecount();
13970 				if (rdstc > ts_val) {
13971 					counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13972 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13973 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13974 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13975 					}
13976 				}
13977 #endif
13978 				m_freem(m);
13979 				tp = tcp_close(tp);
13980 				if (tp == NULL) {
13981 #ifdef TCP_ACCOUNTING
13982 					sched_unpin();
13983 #endif
13984 					return (1);
13985 				}
13986 				/*
13987 				 * We would normally do drop-with-reset which would
13988 				 * send back a reset. We can't since we don't have
13989 				 * all the needed bits. Instead lets arrange for
13990 				 * a call to tcp_output(). That way since we
13991 				 * are in the closed state we will generate a reset.
13992 				 *
13993 				 * Note if tcp_accounting is on we don't unpin since
13994 				 * we do that after the goto label.
13995 				 */
13996 				goto send_out_a_rst;
13997 			}
13998 			if ((sbused(&so->so_snd) == 0) &&
13999 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
14000 			    (tp->t_flags & TF_SENTFIN)) {
14001 				/*
14002 				 * If we can't receive any more data, then closing user can
14003 				 * proceed. Starting the timer is contrary to the
14004 				 * specification, but if we don't get a FIN we'll hang
14005 				 * forever.
14006 				 *
14007 				 */
14008 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
14009 					soisdisconnected(so);
14010 					tcp_timer_activate(tp, TT_2MSL,
14011 							   (tcp_fast_finwait2_recycle ?
14012 							    tcp_finwait2_timeout :
14013 							    TP_MAXIDLE(tp)));
14014 				}
14015 				if (ourfinisacked == 0) {
14016 					/*
14017 					 * We don't change to fin-wait-2 if we have our fin acked
14018 					 * which means we are probably in TCPS_CLOSING.
14019 					 */
14020 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
14021 				}
14022 			}
14023 		}
14024 		/* Wake up the socket if we have room to write more */
14025 		if (sbavail(&so->so_snd)) {
14026 			rack->r_wanted_output = 1;
14027 			if (ctf_progress_timeout_check(tp, true)) {
14028 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14029 							tp, tick, PROGRESS_DROP, __LINE__);
14030 				/*
14031 				 * We cheat here and don't send a RST, we should send one
14032 				 * when the pacer drops the connection.
14033 				 */
14034 #ifdef TCP_ACCOUNTING
14035 				rdstc = get_cyclecount();
14036 				if (rdstc > ts_val) {
14037 					counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
14038 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14039 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
14040 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
14041 					}
14042 				}
14043 				sched_unpin();
14044 #endif
14045 				(void)tcp_drop(tp, ETIMEDOUT);
14046 				m_freem(m);
14047 				return (1);
14048 			}
14049 		}
14050 		if (ourfinisacked) {
14051 			switch(tp->t_state) {
14052 			case TCPS_CLOSING:
14053 #ifdef TCP_ACCOUNTING
14054 				rdstc = get_cyclecount();
14055 				if (rdstc > ts_val) {
14056 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
14057 							(rdstc - ts_val));
14058 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14059 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
14060 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
14061 					}
14062 				}
14063 				sched_unpin();
14064 #endif
14065 				tcp_twstart(tp);
14066 				m_freem(m);
14067 				return (1);
14068 				break;
14069 			case TCPS_LAST_ACK:
14070 #ifdef TCP_ACCOUNTING
14071 				rdstc = get_cyclecount();
14072 				if (rdstc > ts_val) {
14073 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
14074 							(rdstc - ts_val));
14075 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14076 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
14077 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
14078 					}
14079 				}
14080 				sched_unpin();
14081 #endif
14082 				tp = tcp_close(tp);
14083 				ctf_do_drop(m, tp);
14084 				return (1);
14085 				break;
14086 			case TCPS_FIN_WAIT_1:
14087 #ifdef TCP_ACCOUNTING
14088 				rdstc = get_cyclecount();
14089 				if (rdstc > ts_val) {
14090 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
14091 							(rdstc - ts_val));
14092 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14093 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
14094 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
14095 					}
14096 				}
14097 #endif
14098 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
14099 					soisdisconnected(so);
14100 					tcp_timer_activate(tp, TT_2MSL,
14101 							   (tcp_fast_finwait2_recycle ?
14102 							    tcp_finwait2_timeout :
14103 							    TP_MAXIDLE(tp)));
14104 				}
14105 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
14106 				break;
14107 			default:
14108 				break;
14109 			}
14110 		}
14111 		if (rack->r_fast_output) {
14112 			/*
14113 			 * We re doing fast output.. can we expand that?
14114 			 */
14115 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
14116 		}
14117 #ifdef TCP_ACCOUNTING
14118 		rdstc = get_cyclecount();
14119 		if (rdstc > ts_val) {
14120 			counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
14121 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14122 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
14123 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
14124 			}
14125 		}
14126 
14127 	} else if (win_up_req) {
14128 		rdstc = get_cyclecount();
14129 		if (rdstc > ts_val) {
14130 			counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
14131 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14132 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
14133 			}
14134 		}
14135 #endif
14136 	}
14137 	/* Now is there a next packet, if so we are done */
14138 	m_freem(m);
14139 	did_out = 0;
14140 	if (nxt_pkt) {
14141 #ifdef TCP_ACCOUNTING
14142 		sched_unpin();
14143 #endif
14144 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
14145 		return (0);
14146 	}
14147 	rack_handle_might_revert(tp, rack);
14148 	ctf_calc_rwin(so, tp);
14149 	if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14150 	send_out_a_rst:
14151 		if (tcp_output(tp) < 0) {
14152 #ifdef TCP_ACCOUNTING
14153 			sched_unpin();
14154 #endif
14155 			return (1);
14156 		}
14157 		did_out = 1;
14158 	}
14159 	rack_free_trim(rack);
14160 #ifdef TCP_ACCOUNTING
14161 	sched_unpin();
14162 #endif
14163 	rack_timer_audit(tp, rack, &so->so_snd);
14164 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
14165 	return (0);
14166 }
14167 
14168 
14169 static int
14170 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
14171     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
14172     int32_t nxt_pkt, struct timeval *tv)
14173 {
14174 #ifdef TCP_ACCOUNTING
14175 	uint64_t ts_val;
14176 #endif
14177 	int32_t thflags, retval, did_out = 0;
14178 	int32_t way_out = 0;
14179 	/*
14180 	 * cts - is the current time from tv (caller gets ts) in microseconds.
14181 	 * ms_cts - is the current time from tv in milliseconds.
14182 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
14183 	 */
14184 	uint32_t cts, us_cts, ms_cts;
14185 	uint32_t tiwin;
14186 	struct timespec ts;
14187 	struct tcpopt to;
14188 	struct tcp_rack *rack;
14189 	struct rack_sendmap *rsm;
14190 	int32_t prev_state = 0;
14191 #ifdef TCP_ACCOUNTING
14192 	int ack_val_set = 0xf;
14193 #endif
14194 	int nsegs;
14195 	/*
14196 	 * tv passed from common code is from either M_TSTMP_LRO or
14197 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
14198 	 */
14199 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14200 	if (m->m_flags & M_ACKCMP) {
14201 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
14202 	}
14203 	if (m->m_flags & M_ACKCMP) {
14204 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
14205 	}
14206 	cts = tcp_tv_to_usectick(tv);
14207 	ms_cts =  tcp_tv_to_mssectick(tv);
14208 	nsegs = m->m_pkthdr.lro_nsegs;
14209 	counter_u64_add(rack_proc_non_comp_ack, 1);
14210 	thflags = th->th_flags;
14211 #ifdef TCP_ACCOUNTING
14212 	sched_pin();
14213 	if (thflags & TH_ACK)
14214 		ts_val = get_cyclecount();
14215 #endif
14216 	if ((m->m_flags & M_TSTMP) ||
14217 	    (m->m_flags & M_TSTMP_LRO)) {
14218 		mbuf_tstmp2timespec(m, &ts);
14219 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
14220 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
14221 	} else
14222 		rack->r_ctl.act_rcv_time = *tv;
14223 	kern_prefetch(rack, &prev_state);
14224 	prev_state = 0;
14225 	/*
14226 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
14227 	 * the scale is zero.
14228 	 */
14229 	tiwin = th->th_win << tp->snd_scale;
14230 #ifdef TCP_ACCOUNTING
14231 	if (thflags & TH_ACK) {
14232 		/*
14233 		 * We have a tradeoff here. We can either do what we are
14234 		 * doing i.e. pinning to this CPU and then doing the accounting
14235 		 * <or> we could do a critical enter, setup the rdtsc and cpu
14236 		 * as in below, and then validate we are on the same CPU on
14237 		 * exit. I have choosen to not do the critical enter since
14238 		 * that often will gain you a context switch, and instead lock
14239 		 * us (line above this if) to the same CPU with sched_pin(). This
14240 		 * means we may be context switched out for a higher priority
14241 		 * interupt but we won't be moved to another CPU.
14242 		 *
14243 		 * If this occurs (which it won't very often since we most likely
14244 		 * are running this code in interupt context and only a higher
14245 		 * priority will bump us ... clock?) we will falsely add in
14246 		 * to the time the interupt processing time plus the ack processing
14247 		 * time. This is ok since its a rare event.
14248 		 */
14249 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
14250 						    ctf_fixed_maxseg(tp));
14251 	}
14252 #endif
14253 	/*
14254 	 * Parse options on any incoming segment.
14255 	 */
14256 	memset(&to, 0, sizeof(to));
14257 	tcp_dooptions(&to, (u_char *)(th + 1),
14258 	    (th->th_off << 2) - sizeof(struct tcphdr),
14259 	    (thflags & TH_SYN) ? TO_SYN : 0);
14260 	NET_EPOCH_ASSERT();
14261 	INP_WLOCK_ASSERT(tp->t_inpcb);
14262 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
14263 	    __func__));
14264 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
14265 	    __func__));
14266 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
14267 	    (tp->t_flags & TF_GPUTINPROG)) {
14268 		/*
14269 		 * We have a goodput in progress
14270 		 * and we have entered a late state.
14271 		 * Do we have enough data in the sb
14272 		 * to handle the GPUT request?
14273 		 */
14274 		uint32_t bytes;
14275 
14276 		bytes = tp->gput_ack - tp->gput_seq;
14277 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
14278 			bytes += tp->gput_seq - tp->snd_una;
14279 		if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
14280 			/*
14281 			 * There are not enough bytes in the socket
14282 			 * buffer that have been sent to cover this
14283 			 * measurement. Cancel it.
14284 			 */
14285 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
14286 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
14287 						   tp->gput_seq,
14288 						   0, 0, 18, __LINE__, NULL, 0);
14289 			tp->t_flags &= ~TF_GPUTINPROG;
14290 		}
14291 	}
14292 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14293 		union tcp_log_stackspecific log;
14294 		struct timeval ltv;
14295 #ifdef NETFLIX_HTTP_LOGGING
14296 		struct http_sendfile_track *http_req;
14297 
14298 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
14299 			http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
14300 		} else {
14301 			http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
14302 		}
14303 #endif
14304 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14305 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
14306 		if (rack->rack_no_prr == 0)
14307 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
14308 		else
14309 			log.u_bbr.flex1 = 0;
14310 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
14311 		log.u_bbr.use_lt_bw <<= 1;
14312 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
14313 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
14314 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14315 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
14316 		log.u_bbr.flex3 = m->m_flags;
14317 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
14318 		log.u_bbr.lost = thflags;
14319 		log.u_bbr.pacing_gain = 0x1;
14320 #ifdef TCP_ACCOUNTING
14321 		log.u_bbr.cwnd_gain = ack_val_set;
14322 #endif
14323 		log.u_bbr.flex7 = 2;
14324 		if (m->m_flags & M_TSTMP) {
14325 			/* Record the hardware timestamp if present */
14326 			mbuf_tstmp2timespec(m, &ts);
14327 			ltv.tv_sec = ts.tv_sec;
14328 			ltv.tv_usec = ts.tv_nsec / 1000;
14329 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
14330 		} else if (m->m_flags & M_TSTMP_LRO) {
14331 			/* Record the LRO the arrival timestamp */
14332 			mbuf_tstmp2timespec(m, &ts);
14333 			ltv.tv_sec = ts.tv_sec;
14334 			ltv.tv_usec = ts.tv_nsec / 1000;
14335 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
14336 		}
14337 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
14338 		/* Log the rcv time */
14339 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
14340 #ifdef NETFLIX_HTTP_LOGGING
14341 		log.u_bbr.applimited = tp->t_http_closed;
14342 		log.u_bbr.applimited <<= 8;
14343 		log.u_bbr.applimited |= tp->t_http_open;
14344 		log.u_bbr.applimited <<= 8;
14345 		log.u_bbr.applimited |= tp->t_http_req;
14346 		if (http_req) {
14347 			/* Copy out any client req info */
14348 			/* seconds */
14349 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
14350 			/* useconds */
14351 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
14352 			log.u_bbr.rttProp = http_req->timestamp;
14353 			log.u_bbr.cur_del_rate = http_req->start;
14354 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
14355 				log.u_bbr.flex8 |= 1;
14356 			} else {
14357 				log.u_bbr.flex8 |= 2;
14358 				log.u_bbr.bw_inuse = http_req->end;
14359 			}
14360 			log.u_bbr.flex6 = http_req->start_seq;
14361 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
14362 				log.u_bbr.flex8 |= 4;
14363 				log.u_bbr.epoch = http_req->end_seq;
14364 			}
14365 		}
14366 #endif
14367 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
14368 		    tlen, &log, true, &ltv);
14369 	}
14370 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
14371 		way_out = 4;
14372 		retval = 0;
14373 		m_freem(m);
14374 		goto done_with_input;
14375 	}
14376 	/*
14377 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
14378 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
14379 	 */
14380 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
14381 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
14382 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14383 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14384 #ifdef TCP_ACCOUNTING
14385 		sched_unpin();
14386 #endif
14387 		return (1);
14388 	}
14389 	/*
14390 	 * If timestamps were negotiated during SYN/ACK and a
14391 	 * segment without a timestamp is received, silently drop
14392 	 * the segment, unless it is a RST segment or missing timestamps are
14393 	 * tolerated.
14394 	 * See section 3.2 of RFC 7323.
14395 	 */
14396 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
14397 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
14398 		way_out = 5;
14399 		retval = 0;
14400 		m_freem(m);
14401 		goto done_with_input;
14402 	}
14403 
14404 	/*
14405 	 * Segment received on connection. Reset idle time and keep-alive
14406 	 * timer. XXX: This should be done after segment validation to
14407 	 * ignore broken/spoofed segs.
14408 	 */
14409 	if  (tp->t_idle_reduce &&
14410 	     (tp->snd_max == tp->snd_una) &&
14411 	     ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
14412 		counter_u64_add(rack_input_idle_reduces, 1);
14413 		rack_cc_after_idle(rack, tp);
14414 	}
14415 	tp->t_rcvtime = ticks;
14416 #ifdef STATS
14417 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
14418 #endif
14419 	if (tiwin > rack->r_ctl.rc_high_rwnd)
14420 		rack->r_ctl.rc_high_rwnd = tiwin;
14421 	/*
14422 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
14423 	 * this to occur after we've validated the segment.
14424 	 */
14425 	if (tp->t_flags2 & TF2_ECN_PERMIT) {
14426 		if (thflags & TH_CWR) {
14427 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
14428 			tp->t_flags |= TF_ACKNOW;
14429 		}
14430 		switch (iptos & IPTOS_ECN_MASK) {
14431 		case IPTOS_ECN_CE:
14432 			tp->t_flags2 |= TF2_ECN_SND_ECE;
14433 			KMOD_TCPSTAT_INC(tcps_ecn_ce);
14434 			break;
14435 		case IPTOS_ECN_ECT0:
14436 			KMOD_TCPSTAT_INC(tcps_ecn_ect0);
14437 			break;
14438 		case IPTOS_ECN_ECT1:
14439 			KMOD_TCPSTAT_INC(tcps_ecn_ect1);
14440 			break;
14441 		}
14442 
14443 		/* Process a packet differently from RFC3168. */
14444 		cc_ecnpkt_handler(tp, th, iptos);
14445 
14446 		/* Congestion experienced. */
14447 		if (thflags & TH_ECE) {
14448 			rack_cong_signal(tp, CC_ECN, th->th_ack);
14449 		}
14450 	}
14451 
14452 	/*
14453 	 * If echoed timestamp is later than the current time, fall back to
14454 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
14455 	 * were used when this connection was established.
14456 	 */
14457 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
14458 		to.to_tsecr -= tp->ts_offset;
14459 		if (TSTMP_GT(to.to_tsecr, ms_cts))
14460 			to.to_tsecr = 0;
14461 	}
14462 
14463 	/*
14464 	 * If its the first time in we need to take care of options and
14465 	 * verify we can do SACK for rack!
14466 	 */
14467 	if (rack->r_state == 0) {
14468 		/* Should be init'd by rack_init() */
14469 		KASSERT(rack->rc_inp != NULL,
14470 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
14471 		if (rack->rc_inp == NULL) {
14472 			rack->rc_inp = tp->t_inpcb;
14473 		}
14474 
14475 		/*
14476 		 * Process options only when we get SYN/ACK back. The SYN
14477 		 * case for incoming connections is handled in tcp_syncache.
14478 		 * According to RFC1323 the window field in a SYN (i.e., a
14479 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
14480 		 * this is traditional behavior, may need to be cleaned up.
14481 		 */
14482 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
14483 			/* Handle parallel SYN for ECN */
14484 			if (!(thflags & TH_ACK) &&
14485 			    ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) &&
14486 			    ((V_tcp_do_ecn == 1) || (V_tcp_do_ecn == 2))) {
14487 				tp->t_flags2 |= TF2_ECN_PERMIT;
14488 				tp->t_flags2 |= TF2_ECN_SND_ECE;
14489 				TCPSTAT_INC(tcps_ecn_shs);
14490 			}
14491 			if ((to.to_flags & TOF_SCALE) &&
14492 			    (tp->t_flags & TF_REQ_SCALE)) {
14493 				tp->t_flags |= TF_RCVD_SCALE;
14494 				tp->snd_scale = to.to_wscale;
14495 			} else
14496 				tp->t_flags &= ~TF_REQ_SCALE;
14497 			/*
14498 			 * Initial send window.  It will be updated with the
14499 			 * next incoming segment to the scaled value.
14500 			 */
14501 			tp->snd_wnd = th->th_win;
14502 			rack_validate_fo_sendwin_up(tp, rack);
14503 			if ((to.to_flags & TOF_TS) &&
14504 			    (tp->t_flags & TF_REQ_TSTMP)) {
14505 				tp->t_flags |= TF_RCVD_TSTMP;
14506 				tp->ts_recent = to.to_tsval;
14507 				tp->ts_recent_age = cts;
14508 			} else
14509 				tp->t_flags &= ~TF_REQ_TSTMP;
14510 			if (to.to_flags & TOF_MSS) {
14511 				tcp_mss(tp, to.to_mss);
14512 			}
14513 			if ((tp->t_flags & TF_SACK_PERMIT) &&
14514 			    (to.to_flags & TOF_SACKPERM) == 0)
14515 				tp->t_flags &= ~TF_SACK_PERMIT;
14516 			if (IS_FASTOPEN(tp->t_flags)) {
14517 				if (to.to_flags & TOF_FASTOPEN) {
14518 					uint16_t mss;
14519 
14520 					if (to.to_flags & TOF_MSS)
14521 						mss = to.to_mss;
14522 					else
14523 						if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
14524 							mss = TCP6_MSS;
14525 						else
14526 							mss = TCP_MSS;
14527 					tcp_fastopen_update_cache(tp, mss,
14528 					    to.to_tfo_len, to.to_tfo_cookie);
14529 				} else
14530 					tcp_fastopen_disable_path(tp);
14531 			}
14532 		}
14533 		/*
14534 		 * At this point we are at the initial call. Here we decide
14535 		 * if we are doing RACK or not. We do this by seeing if
14536 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
14537 		 * The code now does do dup-ack counting so if you don't
14538 		 * switch back you won't get rack & TLP, but you will still
14539 		 * get this stack.
14540 		 */
14541 
14542 		if ((rack_sack_not_required == 0) &&
14543 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
14544 			tcp_switch_back_to_default(tp);
14545 			(*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
14546 			    tlen, iptos);
14547 #ifdef TCP_ACCOUNTING
14548 			sched_unpin();
14549 #endif
14550 			return (1);
14551 		}
14552 		tcp_set_hpts(tp->t_inpcb);
14553 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
14554 	}
14555 	if (thflags & TH_FIN)
14556 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
14557 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14558 	if ((rack->rc_gp_dyn_mul) &&
14559 	    (rack->use_fixed_rate == 0) &&
14560 	    (rack->rc_always_pace)) {
14561 		/* Check in on probertt */
14562 		rack_check_probe_rtt(rack, us_cts);
14563 	}
14564 	rack_clear_rate_sample(rack);
14565 	if ((rack->forced_ack) &&
14566 	    ((th->th_flags & TH_RST) == 0)) {
14567 		rack_handle_probe_response(rack, tiwin, us_cts);
14568 	}
14569 	/*
14570 	 * This is the one exception case where we set the rack state
14571 	 * always. All other times (timers etc) we must have a rack-state
14572 	 * set (so we assure we have done the checks above for SACK).
14573 	 */
14574 	rack->r_ctl.rc_rcvtime = cts;
14575 	if (rack->r_state != tp->t_state)
14576 		rack_set_state(tp, rack);
14577 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
14578 	    (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
14579 		kern_prefetch(rsm, &prev_state);
14580 	prev_state = rack->r_state;
14581 	retval = (*rack->r_substate) (m, th, so,
14582 	    tp, &to, drop_hdrlen,
14583 	    tlen, tiwin, thflags, nxt_pkt, iptos);
14584 #ifdef INVARIANTS
14585 	if ((retval == 0) &&
14586 	    (tp->t_inpcb == NULL)) {
14587 		panic("retval:%d tp:%p t_inpcb:NULL state:%d",
14588 		    retval, tp, prev_state);
14589 	}
14590 #endif
14591 	if (retval == 0) {
14592 		/*
14593 		 * If retval is 1 the tcb is unlocked and most likely the tp
14594 		 * is gone.
14595 		 */
14596 		INP_WLOCK_ASSERT(tp->t_inpcb);
14597 		if ((rack->rc_gp_dyn_mul) &&
14598 		    (rack->rc_always_pace) &&
14599 		    (rack->use_fixed_rate == 0) &&
14600 		    rack->in_probe_rtt &&
14601 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
14602 			/*
14603 			 * If we are going for target, lets recheck before
14604 			 * we output.
14605 			 */
14606 			rack_check_probe_rtt(rack, us_cts);
14607 		}
14608 		if (rack->set_pacing_done_a_iw == 0) {
14609 			/* How much has been acked? */
14610 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
14611 				/* We have enough to set in the pacing segment size */
14612 				rack->set_pacing_done_a_iw = 1;
14613 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
14614 			}
14615 		}
14616 		tcp_rack_xmit_timer_commit(rack, tp);
14617 #ifdef TCP_ACCOUNTING
14618 		/*
14619 		 * If we set the ack_val_se to what ack processing we are doing
14620 		 * we also want to track how many cycles we burned. Note
14621 		 * the bits after tcp_output we let be "free". This is because
14622 		 * we are also tracking the tcp_output times as well. Note the
14623 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
14624 		 * 0xf cannot be returned and is what we initialize it too to
14625 		 * indicate we are not doing the tabulations.
14626 		 */
14627 		if (ack_val_set != 0xf) {
14628 			uint64_t crtsc;
14629 
14630 			crtsc = get_cyclecount();
14631 			counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14632 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14633 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
14634 			}
14635 		}
14636 #endif
14637 		if (nxt_pkt == 0) {
14638 			if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14639 do_output_now:
14640 				if (tcp_output(tp) < 0)
14641 					return (1);
14642 				did_out = 1;
14643 			}
14644 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
14645 			rack_free_trim(rack);
14646 		}
14647 		/* Update any rounds needed */
14648 		if (SEQ_GEQ(tp->snd_una, rack->r_ctl.roundends)) {
14649 			rack->r_ctl.current_round++;
14650 			rack->r_ctl.roundends = tp->snd_max;
14651 			if (CC_ALGO(tp)->newround != NULL) {
14652 				CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round);
14653 			}
14654 		}
14655 		if ((nxt_pkt == 0) &&
14656 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
14657 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
14658 		     (tp->t_flags & TF_DELACK) ||
14659 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
14660 		      (tp->t_state <= TCPS_CLOSING)))) {
14661 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
14662 			if ((tp->snd_max == tp->snd_una) &&
14663 			    ((tp->t_flags & TF_DELACK) == 0) &&
14664 			    (tcp_in_hpts(rack->rc_inp)) &&
14665 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
14666 				/* keep alive not needed if we are hptsi output yet */
14667 				;
14668 			} else {
14669 				int late = 0;
14670 				if (tcp_in_hpts(rack->rc_inp)) {
14671 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14672 						us_cts = tcp_get_usecs(NULL);
14673 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
14674 							rack->r_early = 1;
14675 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
14676 						} else
14677 							late = 1;
14678 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
14679 					}
14680 					tcp_hpts_remove(tp->t_inpcb);
14681 				}
14682 				if (late && (did_out == 0)) {
14683 					/*
14684 					 * We are late in the sending
14685 					 * and we did not call the output
14686 					 * (this probably should not happen).
14687 					 */
14688 					goto do_output_now;
14689 				}
14690 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14691 			}
14692 			way_out = 1;
14693 		} else if (nxt_pkt == 0) {
14694 			/* Do we have the correct timer running? */
14695 			rack_timer_audit(tp, rack, &so->so_snd);
14696 			way_out = 2;
14697 		}
14698 	done_with_input:
14699 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
14700 		if (did_out)
14701 			rack->r_wanted_output = 0;
14702 #ifdef INVARIANTS
14703 		if (tp->t_inpcb == NULL) {
14704 			panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
14705 			      did_out,
14706 			      retval, tp, prev_state);
14707 		}
14708 #endif
14709 #ifdef TCP_ACCOUNTING
14710 	} else {
14711 		/*
14712 		 * Track the time (see above).
14713 		 */
14714 		if (ack_val_set != 0xf) {
14715 			uint64_t crtsc;
14716 
14717 			crtsc = get_cyclecount();
14718 			counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14719 			/*
14720 			 * Note we *DO NOT* increment the per-tcb counters since
14721 			 * in the else the TP may be gone!!
14722 			 */
14723 		}
14724 #endif
14725 	}
14726 #ifdef TCP_ACCOUNTING
14727 	sched_unpin();
14728 #endif
14729 	return (retval);
14730 }
14731 
14732 void
14733 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
14734     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
14735 {
14736 	struct timeval tv;
14737 
14738 	/* First lets see if we have old packets */
14739 	if (tp->t_in_pkt) {
14740 		if (ctf_do_queued_segments(so, tp, 1)) {
14741 			m_freem(m);
14742 			return;
14743 		}
14744 	}
14745 	if (m->m_flags & M_TSTMP_LRO) {
14746 		tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
14747 		tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
14748 	} else {
14749 		/* Should not be should we kassert instead? */
14750 		tcp_get_usecs(&tv);
14751 	}
14752 	if (rack_do_segment_nounlock(m, th, so, tp,
14753 				     drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14754 		INP_WUNLOCK(tp->t_inpcb);
14755 	}
14756 }
14757 
14758 struct rack_sendmap *
14759 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14760 {
14761 	struct rack_sendmap *rsm = NULL;
14762 	int32_t idx;
14763 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
14764 
14765 	/* Return the next guy to be re-transmitted */
14766 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14767 		return (NULL);
14768 	}
14769 	if (tp->t_flags & TF_SENTFIN) {
14770 		/* retran the end FIN? */
14771 		return (NULL);
14772 	}
14773 	/* ok lets look at this one */
14774 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14775 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14776 		goto check_it;
14777 	}
14778 	rsm = rack_find_lowest_rsm(rack);
14779 	if (rsm == NULL) {
14780 		return (NULL);
14781 	}
14782 check_it:
14783 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14784 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14785 		/*
14786 		 * No sack so we automatically do the 3 strikes and
14787 		 * retransmit (no rack timer would be started).
14788 		 */
14789 
14790 		return (rsm);
14791 	}
14792 	if (rsm->r_flags & RACK_ACKED) {
14793 		return (NULL);
14794 	}
14795 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14796 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14797 		/* Its not yet ready */
14798 		return (NULL);
14799 	}
14800 	srtt = rack_grab_rtt(tp, rack);
14801 	idx = rsm->r_rtr_cnt - 1;
14802 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14803 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14804 	if ((tsused == ts_low) ||
14805 	    (TSTMP_LT(tsused, ts_low))) {
14806 		/* No time since sending */
14807 		return (NULL);
14808 	}
14809 	if ((tsused - ts_low) < thresh) {
14810 		/* It has not been long enough yet */
14811 		return (NULL);
14812 	}
14813 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14814 	    ((rsm->r_flags & RACK_SACK_PASSED) &&
14815 	     (rack->sack_attack_disable == 0))) {
14816 		/*
14817 		 * We have passed the dup-ack threshold <or>
14818 		 * a SACK has indicated this is missing.
14819 		 * Note that if you are a declared attacker
14820 		 * it is only the dup-ack threshold that
14821 		 * will cause retransmits.
14822 		 */
14823 		/* log retransmit reason */
14824 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14825 		rack->r_fast_output = 0;
14826 		return (rsm);
14827 	}
14828 	return (NULL);
14829 }
14830 
14831 static void
14832 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14833 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14834 			   int line, struct rack_sendmap *rsm, uint8_t quality)
14835 {
14836 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14837 		union tcp_log_stackspecific log;
14838 		struct timeval tv;
14839 
14840 		memset(&log, 0, sizeof(log));
14841 		log.u_bbr.flex1 = slot;
14842 		log.u_bbr.flex2 = len;
14843 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14844 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14845 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14846 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14847 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14848 		log.u_bbr.use_lt_bw <<= 1;
14849 		log.u_bbr.use_lt_bw |= rack->r_late;
14850 		log.u_bbr.use_lt_bw <<= 1;
14851 		log.u_bbr.use_lt_bw |= rack->r_early;
14852 		log.u_bbr.use_lt_bw <<= 1;
14853 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14854 		log.u_bbr.use_lt_bw <<= 1;
14855 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14856 		log.u_bbr.use_lt_bw <<= 1;
14857 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14858 		log.u_bbr.use_lt_bw <<= 1;
14859 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14860 		log.u_bbr.use_lt_bw <<= 1;
14861 		log.u_bbr.use_lt_bw |= rack->gp_ready;
14862 		log.u_bbr.pkt_epoch = line;
14863 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14864 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14865 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14866 		log.u_bbr.bw_inuse = bw_est;
14867 		log.u_bbr.delRate = bw;
14868 		if (rack->r_ctl.gp_bw == 0)
14869 			log.u_bbr.cur_del_rate = 0;
14870 		else
14871 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
14872 		log.u_bbr.rttProp = len_time;
14873 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14874 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14875 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14876 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14877 			/* We are in slow start */
14878 			log.u_bbr.flex7 = 1;
14879 		} else {
14880 			/* we are on congestion avoidance */
14881 			log.u_bbr.flex7 = 0;
14882 		}
14883 		log.u_bbr.flex8 = method;
14884 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14885 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14886 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14887 		log.u_bbr.cwnd_gain <<= 1;
14888 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14889 		log.u_bbr.cwnd_gain <<= 1;
14890 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14891 		log.u_bbr.bbr_substate = quality;
14892 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
14893 		    &rack->rc_inp->inp_socket->so_rcv,
14894 		    &rack->rc_inp->inp_socket->so_snd,
14895 		    BBR_LOG_HPTSI_CALC, 0,
14896 		    0, &log, false, &tv);
14897 	}
14898 }
14899 
14900 static uint32_t
14901 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14902 {
14903 	uint32_t new_tso, user_max;
14904 
14905 	user_max = rack->rc_user_set_max_segs * mss;
14906 	if (rack->rc_force_max_seg) {
14907 		return (user_max);
14908 	}
14909 	if (rack->use_fixed_rate &&
14910 	    ((rack->r_ctl.crte == NULL) ||
14911 	     (bw != rack->r_ctl.crte->rate))) {
14912 		/* Use the user mss since we are not exactly matched */
14913 		return (user_max);
14914 	}
14915 	new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14916 	if (new_tso > user_max)
14917 		new_tso = user_max;
14918 	return (new_tso);
14919 }
14920 
14921 static int32_t
14922 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)
14923 {
14924 	uint64_t lentim, fill_bw;
14925 
14926 	/* Lets first see if we are full, if so continue with normal rate */
14927 	rack->r_via_fill_cw = 0;
14928 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14929 		return (slot);
14930 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14931 		return (slot);
14932 	if (rack->r_ctl.rc_last_us_rtt == 0)
14933 		return (slot);
14934 	if (rack->rc_pace_fill_if_rttin_range &&
14935 	    (rack->r_ctl.rc_last_us_rtt >=
14936 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14937 		/* The rtt is huge, N * smallest, lets not fill */
14938 		return (slot);
14939 	}
14940 	/*
14941 	 * first lets calculate the b/w based on the last us-rtt
14942 	 * and the sndwnd.
14943 	 */
14944 	fill_bw = rack->r_ctl.cwnd_to_use;
14945 	/* Take the rwnd if its smaller */
14946 	if (fill_bw > rack->rc_tp->snd_wnd)
14947 		fill_bw = rack->rc_tp->snd_wnd;
14948 	if (rack->r_fill_less_agg) {
14949 		/*
14950 		 * Now take away the inflight (this will reduce our
14951 		 * aggressiveness and yeah, if we get that much out in 1RTT
14952 		 * we will have had acks come back and still be behind).
14953 		 */
14954 		fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14955 	}
14956 	/* Now lets make it into a b/w */
14957 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14958 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14959 	/* We are below the min b/w */
14960 	if (non_paced)
14961 		*rate_wanted = fill_bw;
14962 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14963 		return (slot);
14964 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14965 		fill_bw = rack->r_ctl.bw_rate_cap;
14966 	rack->r_via_fill_cw = 1;
14967 	if (rack->r_rack_hw_rate_caps &&
14968 	    (rack->r_ctl.crte != NULL)) {
14969 		uint64_t high_rate;
14970 
14971 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14972 		if (fill_bw > high_rate) {
14973 			/* We are capping bw at the highest rate table entry */
14974 			if (*rate_wanted > high_rate) {
14975 				/* The original rate was also capped */
14976 				rack->r_via_fill_cw = 0;
14977 			}
14978 			rack_log_hdwr_pacing(rack,
14979 					     fill_bw, high_rate, __LINE__,
14980 					     0, 3);
14981 			fill_bw = high_rate;
14982 			if (capped)
14983 				*capped = 1;
14984 		}
14985 	} else if ((rack->r_ctl.crte == NULL) &&
14986 		   (rack->rack_hdrw_pacing == 0) &&
14987 		   (rack->rack_hdw_pace_ena) &&
14988 		   rack->r_rack_hw_rate_caps &&
14989 		   (rack->rack_attempt_hdwr_pace == 0) &&
14990 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
14991 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14992 		/*
14993 		 * Ok we may have a first attempt that is greater than our top rate
14994 		 * lets check.
14995 		 */
14996 		uint64_t high_rate;
14997 
14998 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14999 		if (high_rate) {
15000 			if (fill_bw > high_rate) {
15001 				fill_bw = high_rate;
15002 				if (capped)
15003 					*capped = 1;
15004 			}
15005 		}
15006 	}
15007 	/*
15008 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
15009 	 * in a rtt, what does that time wise equate too?
15010 	 */
15011 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
15012 	lentim /= fill_bw;
15013 	*rate_wanted = fill_bw;
15014 	if (non_paced || (lentim < slot)) {
15015 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
15016 					   0, lentim, 12, __LINE__, NULL, 0);
15017 		return ((int32_t)lentim);
15018 	} else
15019 		return (slot);
15020 }
15021 
15022 static int32_t
15023 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
15024 {
15025 	uint64_t srtt;
15026 	int32_t slot = 0;
15027 	int can_start_hw_pacing = 1;
15028 	int err;
15029 
15030 	if (rack->rc_always_pace == 0) {
15031 		/*
15032 		 * We use the most optimistic possible cwnd/srtt for
15033 		 * sending calculations. This will make our
15034 		 * calculation anticipate getting more through
15035 		 * quicker then possible. But thats ok we don't want
15036 		 * the peer to have a gap in data sending.
15037 		 */
15038 		uint64_t cwnd, tr_perms = 0;
15039 		int32_t reduce = 0;
15040 
15041 	old_method:
15042 		/*
15043 		 * We keep no precise pacing with the old method
15044 		 * instead we use the pacer to mitigate bursts.
15045 		 */
15046 		if (rack->r_ctl.rc_rack_min_rtt)
15047 			srtt = rack->r_ctl.rc_rack_min_rtt;
15048 		else
15049 			srtt = max(tp->t_srtt, 1);
15050 		if (rack->r_ctl.rc_rack_largest_cwnd)
15051 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
15052 		else
15053 			cwnd = rack->r_ctl.cwnd_to_use;
15054 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
15055 		tr_perms = (cwnd * 1000) / srtt;
15056 		if (tr_perms == 0) {
15057 			tr_perms = ctf_fixed_maxseg(tp);
15058 		}
15059 		/*
15060 		 * Calculate how long this will take to drain, if
15061 		 * the calculation comes out to zero, thats ok we
15062 		 * will use send_a_lot to possibly spin around for
15063 		 * more increasing tot_len_this_send to the point
15064 		 * that its going to require a pace, or we hit the
15065 		 * cwnd. Which in that case we are just waiting for
15066 		 * a ACK.
15067 		 */
15068 		slot = len / tr_perms;
15069 		/* Now do we reduce the time so we don't run dry? */
15070 		if (slot && rack_slot_reduction) {
15071 			reduce = (slot / rack_slot_reduction);
15072 			if (reduce < slot) {
15073 				slot -= reduce;
15074 			} else
15075 				slot = 0;
15076 		}
15077 		slot *= HPTS_USEC_IN_MSEC;
15078 		if (rack->rc_pace_to_cwnd) {
15079 			uint64_t rate_wanted = 0;
15080 
15081 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
15082 			rack->rc_ack_can_sendout_data = 1;
15083 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
15084 		} else
15085 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
15086 	} else {
15087 		uint64_t bw_est, res, lentim, rate_wanted;
15088 		uint32_t orig_val, segs, oh;
15089 		int capped = 0;
15090 		int prev_fill;
15091 
15092 		if ((rack->r_rr_config == 1) && rsm) {
15093 			return (rack->r_ctl.rc_min_to);
15094 		}
15095 		if (rack->use_fixed_rate) {
15096 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
15097 		} else if ((rack->r_ctl.init_rate == 0) &&
15098 #ifdef NETFLIX_PEAKRATE
15099 			   (rack->rc_tp->t_maxpeakrate == 0) &&
15100 #endif
15101 			   (rack->r_ctl.gp_bw == 0)) {
15102 			/* no way to yet do an estimate */
15103 			bw_est = rate_wanted = 0;
15104 		} else {
15105 			bw_est = rack_get_bw(rack);
15106 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
15107 		}
15108 		if ((bw_est == 0) || (rate_wanted == 0) ||
15109 		    ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
15110 			/*
15111 			 * No way yet to make a b/w estimate or
15112 			 * our raise is set incorrectly.
15113 			 */
15114 			goto old_method;
15115 		}
15116 		/* We need to account for all the overheads */
15117 		segs = (len + segsiz - 1) / segsiz;
15118 		/*
15119 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
15120 		 * and how much data we put in each packet. Yes this
15121 		 * means we may be off if we are larger than 1500 bytes
15122 		 * or smaller. But this just makes us more conservative.
15123 		 */
15124 		if (rack_hw_rate_min &&
15125 		    (bw_est < rack_hw_rate_min))
15126 			can_start_hw_pacing = 0;
15127 		if (ETHERNET_SEGMENT_SIZE > segsiz)
15128 			oh = ETHERNET_SEGMENT_SIZE - segsiz;
15129 		else
15130 			oh = 0;
15131 		segs *= oh;
15132 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
15133 		res = lentim / rate_wanted;
15134 		slot = (uint32_t)res;
15135 		orig_val = rack->r_ctl.rc_pace_max_segs;
15136 		if (rack->r_ctl.crte == NULL) {
15137 			/*
15138 			 * Only do this if we are not hardware pacing
15139 			 * since if we are doing hw-pacing below we will
15140 			 * set make a call after setting up or changing
15141 			 * the rate.
15142 			 */
15143 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
15144 		} else if (rack->rc_inp->inp_snd_tag == NULL) {
15145 			/*
15146 			 * We lost our rate somehow, this can happen
15147 			 * if the interface changed underneath us.
15148 			 */
15149 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
15150 			rack->r_ctl.crte = NULL;
15151 			/* Lets re-allow attempting to setup pacing */
15152 			rack->rack_hdrw_pacing = 0;
15153 			rack->rack_attempt_hdwr_pace = 0;
15154 			rack_log_hdwr_pacing(rack,
15155 					     rate_wanted, bw_est, __LINE__,
15156 					     0, 6);
15157 		}
15158 		/* Did we change the TSO size, if so log it */
15159 		if (rack->r_ctl.rc_pace_max_segs != orig_val)
15160 			rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL, 0);
15161 		prev_fill = rack->r_via_fill_cw;
15162 		if ((rack->rc_pace_to_cwnd) &&
15163 		    (capped == 0) &&
15164 		    (rack->use_fixed_rate == 0) &&
15165 		    (rack->in_probe_rtt == 0) &&
15166 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
15167 			/*
15168 			 * We want to pace at our rate *or* faster to
15169 			 * fill the cwnd to the max if its not full.
15170 			 */
15171 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
15172 		}
15173 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
15174 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
15175 			if ((rack->rack_hdw_pace_ena) &&
15176 			    (can_start_hw_pacing > 0) &&
15177 			    (rack->rack_hdrw_pacing == 0) &&
15178 			    (rack->rack_attempt_hdwr_pace == 0)) {
15179 				/*
15180 				 * Lets attempt to turn on hardware pacing
15181 				 * if we can.
15182 				 */
15183 				rack->rack_attempt_hdwr_pace = 1;
15184 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
15185 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
15186 								       rate_wanted,
15187 								       RS_PACING_GEQ,
15188 								       &err, &rack->r_ctl.crte_prev_rate);
15189 				if (rack->r_ctl.crte) {
15190 					rack->rack_hdrw_pacing = 1;
15191 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
15192 												 0, rack->r_ctl.crte,
15193 												 NULL);
15194 					rack_log_hdwr_pacing(rack,
15195 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15196 							     err, 0);
15197 					rack->r_ctl.last_hw_bw_req = rate_wanted;
15198 				} else {
15199 					counter_u64_add(rack_hw_pace_init_fail, 1);
15200 				}
15201 			} else if (rack->rack_hdrw_pacing &&
15202 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
15203 				/* Do we need to adjust our rate? */
15204 				const struct tcp_hwrate_limit_table *nrte;
15205 
15206 				if (rack->r_up_only &&
15207 				    (rate_wanted < rack->r_ctl.crte->rate)) {
15208 					/**
15209 					 * We have four possible states here
15210 					 * having to do with the previous time
15211 					 * and this time.
15212 					 *   previous  |  this-time
15213 					 * A)     0      |     0   -- fill_cw not in the picture
15214 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
15215 					 * C)     1      |     1   -- all rates from fill_cw
15216 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
15217 					 *
15218 					 * For case A, C and D we don't allow a drop. But for
15219 					 * case B where we now our on our steady rate we do
15220 					 * allow a drop.
15221 					 *
15222 					 */
15223 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
15224 						goto done_w_hdwr;
15225 				}
15226 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
15227 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
15228 					if (rack_hw_rate_to_low &&
15229 					    (bw_est < rack_hw_rate_to_low)) {
15230 						/*
15231 						 * The pacing rate is too low for hardware, but
15232 						 * do allow hardware pacing to be restarted.
15233 						 */
15234 						rack_log_hdwr_pacing(rack,
15235 							     bw_est, rack->r_ctl.crte->rate, __LINE__,
15236 							     0, 5);
15237 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
15238 						rack->r_ctl.crte = NULL;
15239 						rack->rack_attempt_hdwr_pace = 0;
15240 						rack->rack_hdrw_pacing = 0;
15241 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15242 						goto done_w_hdwr;
15243 					}
15244 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
15245 								   rack->rc_tp,
15246 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
15247 								   rate_wanted,
15248 								   RS_PACING_GEQ,
15249 								   &err, &rack->r_ctl.crte_prev_rate);
15250 					if (nrte == NULL) {
15251 						/* Lost the rate */
15252 						rack->rack_hdrw_pacing = 0;
15253 						rack->r_ctl.crte = NULL;
15254 						rack_log_hdwr_pacing(rack,
15255 								     rate_wanted, 0, __LINE__,
15256 								     err, 1);
15257 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15258 						counter_u64_add(rack_hw_pace_lost, 1);
15259 					} else if (nrte != rack->r_ctl.crte) {
15260 						rack->r_ctl.crte = nrte;
15261 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
15262 													 segsiz, 0,
15263 													 rack->r_ctl.crte,
15264 													 NULL);
15265 						rack_log_hdwr_pacing(rack,
15266 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15267 								     err, 2);
15268 						rack->r_ctl.last_hw_bw_req = rate_wanted;
15269 					}
15270 				} else {
15271 					/* We just need to adjust the segment size */
15272 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15273 					rack_log_hdwr_pacing(rack,
15274 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15275 							     0, 4);
15276 					rack->r_ctl.last_hw_bw_req = rate_wanted;
15277 				}
15278 			}
15279 		}
15280 		if ((rack->r_ctl.crte != NULL) &&
15281 		    (rack->r_ctl.crte->rate == rate_wanted)) {
15282 			/*
15283 			 * We need to add a extra if the rates
15284 			 * are exactly matched. The idea is
15285 			 * we want the software to make sure the
15286 			 * queue is empty before adding more, this
15287 			 * gives us N MSS extra pace times where
15288 			 * N is our sysctl
15289 			 */
15290 			slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
15291 		}
15292 done_w_hdwr:
15293 		if (rack_limit_time_with_srtt &&
15294 		    (rack->use_fixed_rate == 0) &&
15295 #ifdef NETFLIX_PEAKRATE
15296 		    (rack->rc_tp->t_maxpeakrate == 0) &&
15297 #endif
15298 		    (rack->rack_hdrw_pacing == 0)) {
15299 			/*
15300 			 * Sanity check, we do not allow the pacing delay
15301 			 * to be longer than the SRTT of the path. If it is
15302 			 * a slow path, then adding a packet should increase
15303 			 * the RTT and compensate for this i.e. the srtt will
15304 			 * be greater so the allowed pacing time will be greater.
15305 			 *
15306 			 * Note this restriction is not for where a peak rate
15307 			 * is set, we are doing fixed pacing or hardware pacing.
15308 			 */
15309 			if (rack->rc_tp->t_srtt)
15310 				srtt = rack->rc_tp->t_srtt;
15311 			else
15312 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
15313 			if (srtt < (uint64_t)slot) {
15314 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
15315 				slot = srtt;
15316 			}
15317 		}
15318 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
15319 	}
15320 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
15321 		/*
15322 		 * If this rate is seeing enobufs when it
15323 		 * goes to send then either the nic is out
15324 		 * of gas or we are mis-estimating the time
15325 		 * somehow and not letting the queue empty
15326 		 * completely. Lets add to the pacing time.
15327 		 */
15328 		int hw_boost_delay;
15329 
15330 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
15331 		if (hw_boost_delay > rack_enobuf_hw_max)
15332 			hw_boost_delay = rack_enobuf_hw_max;
15333 		else if (hw_boost_delay < rack_enobuf_hw_min)
15334 			hw_boost_delay = rack_enobuf_hw_min;
15335 		slot += hw_boost_delay;
15336 	}
15337 	if (slot)
15338 		counter_u64_add(rack_calc_nonzero, 1);
15339 	else
15340 		counter_u64_add(rack_calc_zero, 1);
15341 	return (slot);
15342 }
15343 
15344 static void
15345 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
15346     tcp_seq startseq, uint32_t sb_offset)
15347 {
15348 	struct rack_sendmap *my_rsm = NULL;
15349 	struct rack_sendmap fe;
15350 
15351 	if (tp->t_state < TCPS_ESTABLISHED) {
15352 		/*
15353 		 * We don't start any measurements if we are
15354 		 * not at least established.
15355 		 */
15356 		return;
15357 	}
15358 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
15359 		/*
15360 		 * We will get no more data into the SB
15361 		 * this means we need to have the data available
15362 		 * before we start a measurement.
15363 		 */
15364 
15365 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) <
15366 		    max(rc_init_window(rack),
15367 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
15368 			/* Nope not enough data */
15369 			return;
15370 		}
15371 	}
15372 	tp->t_flags |= TF_GPUTINPROG;
15373 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
15374 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
15375 	tp->gput_seq = startseq;
15376 	rack->app_limited_needs_set = 0;
15377 	if (rack->in_probe_rtt)
15378 		rack->measure_saw_probe_rtt = 1;
15379 	else if ((rack->measure_saw_probe_rtt) &&
15380 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
15381 		rack->measure_saw_probe_rtt = 0;
15382 	if (rack->rc_gp_filled)
15383 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15384 	else {
15385 		/* Special case initial measurement */
15386 		struct timeval tv;
15387 
15388 		tp->gput_ts = tcp_get_usecs(&tv);
15389 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15390 	}
15391 	/*
15392 	 * We take a guess out into the future,
15393 	 * if we have no measurement and no
15394 	 * initial rate, we measure the first
15395 	 * initial-windows worth of data to
15396 	 * speed up getting some GP measurement and
15397 	 * thus start pacing.
15398 	 */
15399 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
15400 		rack->app_limited_needs_set = 1;
15401 		tp->gput_ack = startseq + max(rc_init_window(rack),
15402 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
15403 		rack_log_pacing_delay_calc(rack,
15404 					   tp->gput_seq,
15405 					   tp->gput_ack,
15406 					   0,
15407 					   tp->gput_ts,
15408 					   rack->r_ctl.rc_app_limited_cnt,
15409 					   9,
15410 					   __LINE__, NULL, 0);
15411 		return;
15412 	}
15413 	if (sb_offset) {
15414 		/*
15415 		 * We are out somewhere in the sb
15416 		 * can we use the already outstanding data?
15417 		 */
15418 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
15419 			/*
15420 			 * Yes first one is good and in this case
15421 			 * the tp->gput_ts is correctly set based on
15422 			 * the last ack that arrived (no need to
15423 			 * set things up when an ack comes in).
15424 			 */
15425 			my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
15426 			if ((my_rsm == NULL) ||
15427 			    (my_rsm->r_rtr_cnt != 1)) {
15428 				/* retransmission? */
15429 				goto use_latest;
15430 			}
15431 		} else {
15432 			if (rack->r_ctl.rc_first_appl == NULL) {
15433 				/*
15434 				 * If rc_first_appl is NULL
15435 				 * then the cnt should be 0.
15436 				 * This is probably an error, maybe
15437 				 * a KASSERT would be approprate.
15438 				 */
15439 				goto use_latest;
15440 			}
15441 			/*
15442 			 * If we have a marker pointer to the last one that is
15443 			 * app limited we can use that, but we need to set
15444 			 * things up so that when it gets ack'ed we record
15445 			 * the ack time (if its not already acked).
15446 			 */
15447 			rack->app_limited_needs_set = 1;
15448 			/*
15449 			 * We want to get to the rsm that is either
15450 			 * next with space i.e. over 1 MSS or the one
15451 			 * after that (after the app-limited).
15452 			 */
15453 			my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15454 					 rack->r_ctl.rc_first_appl);
15455 			if (my_rsm) {
15456 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
15457 					/* Have to use the next one */
15458 					my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15459 							 my_rsm);
15460 				else {
15461 					/* Use after the first MSS of it is acked */
15462 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
15463 					goto start_set;
15464 				}
15465 			}
15466 			if ((my_rsm == NULL) ||
15467 			    (my_rsm->r_rtr_cnt != 1)) {
15468 				/*
15469 				 * Either its a retransmit or
15470 				 * the last is the app-limited one.
15471 				 */
15472 				goto use_latest;
15473 			}
15474 		}
15475 		tp->gput_seq = my_rsm->r_start;
15476 start_set:
15477 		if (my_rsm->r_flags & RACK_ACKED) {
15478 			/*
15479 			 * This one has been acked use the arrival ack time
15480 			 */
15481 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15482 			rack->app_limited_needs_set = 0;
15483 		}
15484 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15485 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
15486 		rack_log_pacing_delay_calc(rack,
15487 					   tp->gput_seq,
15488 					   tp->gput_ack,
15489 					   (uint64_t)my_rsm,
15490 					   tp->gput_ts,
15491 					   rack->r_ctl.rc_app_limited_cnt,
15492 					   9,
15493 					   __LINE__, NULL, 0);
15494 		return;
15495 	}
15496 
15497 use_latest:
15498 	/*
15499 	 * We don't know how long we may have been
15500 	 * idle or if this is the first-send. Lets
15501 	 * setup the flag so we will trim off
15502 	 * the first ack'd data so we get a true
15503 	 * measurement.
15504 	 */
15505 	rack->app_limited_needs_set = 1;
15506 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
15507 	/* Find this guy so we can pull the send time */
15508 	fe.r_start = startseq;
15509 	my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
15510 	if (my_rsm) {
15511 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15512 		if (my_rsm->r_flags & RACK_ACKED) {
15513 			/*
15514 			 * Unlikely since its probably what was
15515 			 * just transmitted (but I am paranoid).
15516 			 */
15517 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15518 			rack->app_limited_needs_set = 0;
15519 		}
15520 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
15521 			/* This also is unlikely */
15522 			tp->gput_seq = my_rsm->r_start;
15523 		}
15524 	} else {
15525 		/*
15526 		 * TSNH unless we have some send-map limit,
15527 		 * and even at that it should not be hitting
15528 		 * that limit (we should have stopped sending).
15529 		 */
15530 		struct timeval tv;
15531 
15532 		microuptime(&tv);
15533 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15534 	}
15535 	rack_log_pacing_delay_calc(rack,
15536 				   tp->gput_seq,
15537 				   tp->gput_ack,
15538 				   (uint64_t)my_rsm,
15539 				   tp->gput_ts,
15540 				   rack->r_ctl.rc_app_limited_cnt,
15541 				   9, __LINE__, NULL, 0);
15542 }
15543 
15544 static inline uint32_t
15545 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
15546     uint32_t avail, int32_t sb_offset)
15547 {
15548 	uint32_t len;
15549 	uint32_t sendwin;
15550 
15551 	if (tp->snd_wnd > cwnd_to_use)
15552 		sendwin = cwnd_to_use;
15553 	else
15554 		sendwin = tp->snd_wnd;
15555 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
15556 		/* We never want to go over our peers rcv-window */
15557 		len = 0;
15558 	} else {
15559 		uint32_t flight;
15560 
15561 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15562 		if (flight >= sendwin) {
15563 			/*
15564 			 * We have in flight what we are allowed by cwnd (if
15565 			 * it was rwnd blocking it would have hit above out
15566 			 * >= tp->snd_wnd).
15567 			 */
15568 			return (0);
15569 		}
15570 		len = sendwin - flight;
15571 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
15572 			/* We would send too much (beyond the rwnd) */
15573 			len = tp->snd_wnd - ctf_outstanding(tp);
15574 		}
15575 		if ((len + sb_offset) > avail) {
15576 			/*
15577 			 * We don't have that much in the SB, how much is
15578 			 * there?
15579 			 */
15580 			len = avail - sb_offset;
15581 		}
15582 	}
15583 	return (len);
15584 }
15585 
15586 static void
15587 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
15588 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
15589 	     int rsm_is_null, int optlen, int line, uint16_t mode)
15590 {
15591 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15592 		union tcp_log_stackspecific log;
15593 		struct timeval tv;
15594 
15595 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15596 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15597 		log.u_bbr.flex1 = error;
15598 		log.u_bbr.flex2 = flags;
15599 		log.u_bbr.flex3 = rsm_is_null;
15600 		log.u_bbr.flex4 = ipoptlen;
15601 		log.u_bbr.flex5 = tp->rcv_numsacks;
15602 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15603 		log.u_bbr.flex7 = optlen;
15604 		log.u_bbr.flex8 = rack->r_fsb_inited;
15605 		log.u_bbr.applimited = rack->r_fast_output;
15606 		log.u_bbr.bw_inuse = rack_get_bw(rack);
15607 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15608 		log.u_bbr.cwnd_gain = mode;
15609 		log.u_bbr.pkts_out = orig_len;
15610 		log.u_bbr.lt_epoch = len;
15611 		log.u_bbr.delivered = line;
15612 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15613 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15614 		tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
15615 			       len, &log, false, NULL, NULL, 0, &tv);
15616 	}
15617 }
15618 
15619 
15620 static struct mbuf *
15621 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
15622 		   struct rack_fast_send_blk *fsb,
15623 		   int32_t seglimit, int32_t segsize, int hw_tls)
15624 {
15625 #ifdef KERN_TLS
15626 	struct ktls_session *tls, *ntls;
15627 	struct mbuf *start;
15628 #endif
15629 	struct mbuf *m, *n, **np, *smb;
15630 	struct mbuf *top;
15631 	int32_t off, soff;
15632 	int32_t len = *plen;
15633 	int32_t fragsize;
15634 	int32_t len_cp = 0;
15635 	uint32_t mlen, frags;
15636 
15637 	soff = off = the_off;
15638 	smb = m = the_m;
15639 	np = &top;
15640 	top = NULL;
15641 #ifdef KERN_TLS
15642 	if (hw_tls && (m->m_flags & M_EXTPG))
15643 		tls = m->m_epg_tls;
15644 	else
15645 		tls = NULL;
15646 	start = m;
15647 #endif
15648 	while (len > 0) {
15649 		if (m == NULL) {
15650 			*plen = len_cp;
15651 			break;
15652 		}
15653 #ifdef KERN_TLS
15654 		if (hw_tls) {
15655 			if (m->m_flags & M_EXTPG)
15656 				ntls = m->m_epg_tls;
15657 			else
15658 				ntls = NULL;
15659 
15660 			/*
15661 			 * Avoid mixing TLS records with handshake
15662 			 * data or TLS records from different
15663 			 * sessions.
15664 			 */
15665 			if (tls != ntls) {
15666 				MPASS(m != start);
15667 				*plen = len_cp;
15668 				break;
15669 			}
15670 		}
15671 #endif
15672 		mlen = min(len, m->m_len - off);
15673 		if (seglimit) {
15674 			/*
15675 			 * For M_EXTPG mbufs, add 3 segments
15676 			 * + 1 in case we are crossing page boundaries
15677 			 * + 2 in case the TLS hdr/trailer are used
15678 			 * It is cheaper to just add the segments
15679 			 * than it is to take the cache miss to look
15680 			 * at the mbuf ext_pgs state in detail.
15681 			 */
15682 			if (m->m_flags & M_EXTPG) {
15683 				fragsize = min(segsize, PAGE_SIZE);
15684 				frags = 3;
15685 			} else {
15686 				fragsize = segsize;
15687 				frags = 0;
15688 			}
15689 
15690 			/* Break if we really can't fit anymore. */
15691 			if ((frags + 1) >= seglimit) {
15692 				*plen =	len_cp;
15693 				break;
15694 			}
15695 
15696 			/*
15697 			 * Reduce size if you can't copy the whole
15698 			 * mbuf. If we can't copy the whole mbuf, also
15699 			 * adjust len so the loop will end after this
15700 			 * mbuf.
15701 			 */
15702 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
15703 				mlen = (seglimit - frags - 1) * fragsize;
15704 				len = mlen;
15705 				*plen = len_cp + len;
15706 			}
15707 			frags += howmany(mlen, fragsize);
15708 			if (frags == 0)
15709 				frags++;
15710 			seglimit -= frags;
15711 			KASSERT(seglimit > 0,
15712 			    ("%s: seglimit went too low", __func__));
15713 		}
15714 		n = m_get(M_NOWAIT, m->m_type);
15715 		*np = n;
15716 		if (n == NULL)
15717 			goto nospace;
15718 		n->m_len = mlen;
15719 		soff += mlen;
15720 		len_cp += n->m_len;
15721 		if (m->m_flags & (M_EXT|M_EXTPG)) {
15722 			n->m_data = m->m_data + off;
15723 			mb_dupcl(n, m);
15724 		} else {
15725 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
15726 			    (u_int)n->m_len);
15727 		}
15728 		len -= n->m_len;
15729 		off = 0;
15730 		m = m->m_next;
15731 		np = &n->m_next;
15732 		if (len || (soff == smb->m_len)) {
15733 			/*
15734 			 * We have more so we move forward  or
15735 			 * we have consumed the entire mbuf and
15736 			 * len has fell to 0.
15737 			 */
15738 			soff = 0;
15739 			smb = m;
15740 		}
15741 
15742 	}
15743 	if (fsb != NULL) {
15744 		fsb->m = smb;
15745 		fsb->off = soff;
15746 		if (smb) {
15747 			/*
15748 			 * Save off the size of the mbuf. We do
15749 			 * this so that we can recognize when it
15750 			 * has been trimmed by sbcut() as acks
15751 			 * come in.
15752 			 */
15753 			fsb->o_m_len = smb->m_len;
15754 		} else {
15755 			/*
15756 			 * This is the case where the next mbuf went to NULL. This
15757 			 * means with this copy we have sent everything in the sb.
15758 			 * In theory we could clear the fast_output flag, but lets
15759 			 * not since its possible that we could get more added
15760 			 * and acks that call the extend function which would let
15761 			 * us send more.
15762 			 */
15763 			fsb->o_m_len = 0;
15764 		}
15765 	}
15766 	return (top);
15767 nospace:
15768 	if (top)
15769 		m_freem(top);
15770 	return (NULL);
15771 
15772 }
15773 
15774 /*
15775  * This is a copy of m_copym(), taking the TSO segment size/limit
15776  * constraints into account, and advancing the sndptr as it goes.
15777  */
15778 static struct mbuf *
15779 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15780 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15781 {
15782 	struct mbuf *m, *n;
15783 	int32_t soff;
15784 
15785 	soff = rack->r_ctl.fsb.off;
15786 	m = rack->r_ctl.fsb.m;
15787 	if (rack->r_ctl.fsb.o_m_len > m->m_len) {
15788 		/*
15789 		 * The mbuf had the front of it chopped off by an ack
15790 		 * we need to adjust the soff/off by that difference.
15791 		 */
15792 		uint32_t delta;
15793 
15794 		delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15795 		soff -= delta;
15796 	} else if (rack->r_ctl.fsb.o_m_len < m->m_len) {
15797 		/*
15798 		 * The mbuf was expanded probably by
15799 		 * a m_compress. Just update o_m_len.
15800 		 */
15801 		rack->r_ctl.fsb.o_m_len = m->m_len;
15802 	}
15803 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15804 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15805 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15806 				 __FUNCTION__,
15807 				 rack, *plen, m, m->m_len));
15808 	/* Save off the right location before we copy and advance */
15809 	*s_soff = soff;
15810 	*s_mb = rack->r_ctl.fsb.m;
15811 	n = rack_fo_base_copym(m, soff, plen,
15812 			       &rack->r_ctl.fsb,
15813 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
15814 	return (n);
15815 }
15816 
15817 static int
15818 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15819 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
15820 {
15821 	/*
15822 	 * Enter the fast retransmit path. We are given that a sched_pin is
15823 	 * in place (if accounting is compliled in) and the cycle count taken
15824 	 * at the entry is in the ts_val. The concept her is that the rsm
15825 	 * now holds the mbuf offsets and such so we can directly transmit
15826 	 * without a lot of overhead, the len field is already set for
15827 	 * us to prohibit us from sending too much (usually its 1MSS).
15828 	 */
15829 	struct ip *ip = NULL;
15830 	struct udphdr *udp = NULL;
15831 	struct tcphdr *th = NULL;
15832 	struct mbuf *m = NULL;
15833 	struct inpcb *inp;
15834 	uint8_t *cpto;
15835 	struct tcp_log_buffer *lgb;
15836 #ifdef TCP_ACCOUNTING
15837 	uint64_t crtsc;
15838 	int cnt_thru = 1;
15839 #endif
15840 	struct tcpopt to;
15841 	u_char opt[TCP_MAXOLEN];
15842 	uint32_t hdrlen, optlen;
15843 	int32_t slot, segsiz, max_val, tso = 0, error, flags, ulen = 0;
15844 	uint32_t us_cts;
15845 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
15846 	uint32_t if_hw_tsomaxsegsize;
15847 
15848 #ifdef INET6
15849 	struct ip6_hdr *ip6 = NULL;
15850 
15851 	if (rack->r_is_v6) {
15852 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15853 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15854 	} else
15855 #endif				/* INET6 */
15856 	{
15857 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15858 		hdrlen = sizeof(struct tcpiphdr);
15859 	}
15860 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15861 		goto failed;
15862 	}
15863 	if (doing_tlp) {
15864 		/* Its a TLP add the flag, it may already be there but be sure */
15865 		rsm->r_flags |= RACK_TLP;
15866 	} else {
15867 		/* If it was a TLP it is not not on this retransmit */
15868 		rsm->r_flags &= ~RACK_TLP;
15869 	}
15870 	startseq = rsm->r_start;
15871 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15872 	inp = rack->rc_inp;
15873 	to.to_flags = 0;
15874 	flags = tcp_outflags[tp->t_state];
15875 	if (flags & (TH_SYN|TH_RST)) {
15876 		goto failed;
15877 	}
15878 	if (rsm->r_flags & RACK_HAS_FIN) {
15879 		/* We can't send a FIN here */
15880 		goto failed;
15881 	}
15882 	if (flags & TH_FIN) {
15883 		/* We never send a FIN */
15884 		flags &= ~TH_FIN;
15885 	}
15886 	if (tp->t_flags & TF_RCVD_TSTMP) {
15887 		to.to_tsval = ms_cts + tp->ts_offset;
15888 		to.to_tsecr = tp->ts_recent;
15889 		to.to_flags = TOF_TS;
15890 	}
15891 	optlen = tcp_addoptions(&to, opt);
15892 	hdrlen += optlen;
15893 	udp = rack->r_ctl.fsb.udp;
15894 	if (udp)
15895 		hdrlen += sizeof(struct udphdr);
15896 	if (rack->r_ctl.rc_pace_max_segs)
15897 		max_val = rack->r_ctl.rc_pace_max_segs;
15898 	else if (rack->rc_user_set_max_segs)
15899 		max_val = rack->rc_user_set_max_segs * segsiz;
15900 	else
15901 		max_val = len;
15902 	if ((tp->t_flags & TF_TSO) &&
15903 	    V_tcp_do_tso &&
15904 	    (len > segsiz) &&
15905 	    (tp->t_port == 0))
15906 		tso = 1;
15907 #ifdef INET6
15908 	if (MHLEN < hdrlen + max_linkhdr)
15909 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15910 	else
15911 #endif
15912 		m = m_gethdr(M_NOWAIT, MT_DATA);
15913 	if (m == NULL)
15914 		goto failed;
15915 	m->m_data += max_linkhdr;
15916 	m->m_len = hdrlen;
15917 	th = rack->r_ctl.fsb.th;
15918 	/* Establish the len to send */
15919 	if (len > max_val)
15920 		len = max_val;
15921 	if ((tso) && (len + optlen > tp->t_maxseg)) {
15922 		uint32_t if_hw_tsomax;
15923 		int32_t max_len;
15924 
15925 		/* extract TSO information */
15926 		if_hw_tsomax = tp->t_tsomax;
15927 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15928 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15929 		/*
15930 		 * Check if we should limit by maximum payload
15931 		 * length:
15932 		 */
15933 		if (if_hw_tsomax != 0) {
15934 			/* compute maximum TSO length */
15935 			max_len = (if_hw_tsomax - hdrlen -
15936 				   max_linkhdr);
15937 			if (max_len <= 0) {
15938 				goto failed;
15939 			} else if (len > max_len) {
15940 				len = max_len;
15941 			}
15942 		}
15943 		if (len <= segsiz) {
15944 			/*
15945 			 * In case there are too many small fragments don't
15946 			 * use TSO:
15947 			 */
15948 			tso = 0;
15949 		}
15950 	} else {
15951 		tso = 0;
15952 	}
15953 	if ((tso == 0) && (len > segsiz))
15954 		len = segsiz;
15955 	us_cts = tcp_get_usecs(tv);
15956 	if ((len == 0) ||
15957 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
15958 		goto failed;
15959 	}
15960 	th->th_seq = htonl(rsm->r_start);
15961 	th->th_ack = htonl(tp->rcv_nxt);
15962 	/*
15963 	 * The PUSH bit should only be applied
15964 	 * if the full retransmission is made. If
15965 	 * we are sending less than this is the
15966 	 * left hand edge and should not have
15967 	 * the PUSH bit.
15968 	 */
15969 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
15970 	    (len == (rsm->r_end - rsm->r_start)))
15971 		flags |= TH_PUSH;
15972 	th->th_flags = flags;
15973 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15974 	if (th->th_win == 0) {
15975 		tp->t_sndzerowin++;
15976 		tp->t_flags |= TF_RXWIN0SENT;
15977 	} else
15978 		tp->t_flags &= ~TF_RXWIN0SENT;
15979 	if (rsm->r_flags & RACK_TLP) {
15980 		/*
15981 		 * TLP should not count in retran count, but
15982 		 * in its own bin
15983 		 */
15984 		counter_u64_add(rack_tlp_retran, 1);
15985 		counter_u64_add(rack_tlp_retran_bytes, len);
15986 	} else {
15987 		tp->t_sndrexmitpack++;
15988 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15989 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15990 	}
15991 #ifdef STATS
15992 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15993 				 len);
15994 #endif
15995 	if (rsm->m == NULL)
15996 		goto failed;
15997 	if (rsm->orig_m_len != rsm->m->m_len) {
15998 		/* Fix up the orig_m_len and possibly the mbuf offset */
15999 		rack_adjust_orig_mlen(rsm);
16000 	}
16001 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
16002 	if (len <= segsiz) {
16003 		/*
16004 		 * Must have ran out of mbufs for the copy
16005 		 * shorten it to no longer need tso. Lets
16006 		 * not put on sendalot since we are low on
16007 		 * mbufs.
16008 		 */
16009 		tso = 0;
16010 	}
16011 	if ((m->m_next == NULL) || (len <= 0)){
16012 		goto failed;
16013 	}
16014 	if (udp) {
16015 		if (rack->r_is_v6)
16016 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
16017 		else
16018 			ulen = hdrlen + len - sizeof(struct ip);
16019 		udp->uh_ulen = htons(ulen);
16020 	}
16021 	m->m_pkthdr.rcvif = (struct ifnet *)0;
16022 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
16023 #ifdef INET6
16024 	if (rack->r_is_v6) {
16025 		if (tp->t_port) {
16026 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
16027 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16028 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
16029 			th->th_sum = htons(0);
16030 			UDPSTAT_INC(udps_opackets);
16031 		} else {
16032 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
16033 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16034 			th->th_sum = in6_cksum_pseudo(ip6,
16035 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
16036 						      0);
16037 		}
16038 	}
16039 #endif
16040 #if defined(INET6) && defined(INET)
16041 	else
16042 #endif
16043 #ifdef INET
16044 	{
16045 		if (tp->t_port) {
16046 			m->m_pkthdr.csum_flags = CSUM_UDP;
16047 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16048 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
16049 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
16050 			th->th_sum = htons(0);
16051 			UDPSTAT_INC(udps_opackets);
16052 		} else {
16053 			m->m_pkthdr.csum_flags = CSUM_TCP;
16054 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16055 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
16056 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
16057 									IPPROTO_TCP + len + optlen));
16058 		}
16059 		/* IP version must be set here for ipv4/ipv6 checking later */
16060 		KASSERT(ip->ip_v == IPVERSION,
16061 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
16062 	}
16063 #endif
16064 	if (tso) {
16065 		KASSERT(len > tp->t_maxseg - optlen,
16066 			("%s: len <= tso_segsz tp:%p", __func__, tp));
16067 		m->m_pkthdr.csum_flags |= CSUM_TSO;
16068 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
16069 	}
16070 #ifdef INET6
16071 	if (rack->r_is_v6) {
16072 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
16073 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
16074 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
16075 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16076 		else
16077 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16078 	}
16079 #endif
16080 #if defined(INET) && defined(INET6)
16081 	else
16082 #endif
16083 #ifdef INET
16084 	{
16085 		ip->ip_len = htons(m->m_pkthdr.len);
16086 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
16087 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
16088 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16089 			if (tp->t_port == 0 || len < V_tcp_minmss) {
16090 				ip->ip_off |= htons(IP_DF);
16091 			}
16092 		} else {
16093 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16094 		}
16095 	}
16096 #endif
16097 	/* Time to copy in our header */
16098 	cpto = mtod(m, uint8_t *);
16099 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
16100 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
16101 	if (optlen) {
16102 		bcopy(opt, th + 1, optlen);
16103 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
16104 	} else {
16105 		th->th_off = sizeof(struct tcphdr) >> 2;
16106 	}
16107 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
16108 		union tcp_log_stackspecific log;
16109 
16110 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16111 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
16112 		if (rack->rack_no_prr)
16113 			log.u_bbr.flex1 = 0;
16114 		else
16115 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16116 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16117 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16118 		log.u_bbr.flex4 = max_val;
16119 		log.u_bbr.flex5 = 0;
16120 		/* Save off the early/late values */
16121 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16122 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16123 		log.u_bbr.bw_inuse = rack_get_bw(rack);
16124 		if (doing_tlp == 0)
16125 			log.u_bbr.flex8 = 1;
16126 		else
16127 			log.u_bbr.flex8 = 2;
16128 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16129 		log.u_bbr.flex7 = 55;
16130 		log.u_bbr.pkts_out = tp->t_maxseg;
16131 		log.u_bbr.timeStamp = cts;
16132 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16133 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16134 		log.u_bbr.delivered = 0;
16135 		lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16136 				     len, &log, false, NULL, NULL, 0, tv);
16137 	} else
16138 		lgb = NULL;
16139 #ifdef INET6
16140 	if (rack->r_is_v6) {
16141 		error = ip6_output(m, NULL,
16142 				   &inp->inp_route6,
16143 				   0, NULL, NULL, inp);
16144 	}
16145 #endif
16146 #if defined(INET) && defined(INET6)
16147 	else
16148 #endif
16149 #ifdef INET
16150 	{
16151 		error = ip_output(m, NULL,
16152 				  &inp->inp_route,
16153 				  0, 0, inp);
16154 	}
16155 #endif
16156 	m = NULL;
16157 	if (lgb) {
16158 		lgb->tlb_errno = error;
16159 		lgb = NULL;
16160 	}
16161 	if (error) {
16162 		goto failed;
16163 	}
16164 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
16165 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls);
16166 	if (doing_tlp && (rack->fast_rsm_hack == 0)) {
16167 		rack->rc_tlp_in_progress = 1;
16168 		rack->r_ctl.rc_tlp_cnt_out++;
16169 	}
16170 	if (error == 0) {
16171 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
16172 		if (doing_tlp) {
16173 			rack->rc_last_sent_tlp_past_cumack = 0;
16174 			rack->rc_last_sent_tlp_seq_valid = 1;
16175 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
16176 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
16177 		}
16178 	}
16179 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16180 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
16181 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
16182 		rack->r_ctl.retran_during_recovery += len;
16183 	{
16184 		int idx;
16185 
16186 		idx = (len / segsiz) + 3;
16187 		if (idx >= TCP_MSS_ACCT_ATIMER)
16188 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16189 		else
16190 			counter_u64_add(rack_out_size[idx], 1);
16191 	}
16192 	if (tp->t_rtttime == 0) {
16193 		tp->t_rtttime = ticks;
16194 		tp->t_rtseq = startseq;
16195 		KMOD_TCPSTAT_INC(tcps_segstimed);
16196 	}
16197 	counter_u64_add(rack_fto_rsm_send, 1);
16198 	if (error && (error == ENOBUFS)) {
16199 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
16200 		if (rack->rc_enobuf < 0x7f)
16201 			rack->rc_enobuf++;
16202 		if (slot < (10 * HPTS_USEC_IN_MSEC))
16203 			slot = 10 * HPTS_USEC_IN_MSEC;
16204 	} else
16205 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
16206 	if ((slot == 0) ||
16207 	    (rack->rc_always_pace == 0) ||
16208 	    (rack->r_rr_config == 1)) {
16209 		/*
16210 		 * We have no pacing set or we
16211 		 * are using old-style rack or
16212 		 * we are overriden to use the old 1ms pacing.
16213 		 */
16214 		slot = rack->r_ctl.rc_min_to;
16215 	}
16216 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
16217 	if (rack->r_must_retran) {
16218 		rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
16219 		if ((SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) ||
16220 		    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
16221 			/*
16222 			 * We have retransmitted all we need. If
16223 			 * RACK_MUST_RXT is not set then we need to
16224 			 * not retransmit this guy.
16225 			 */
16226 			rack->r_must_retran = 0;
16227 			rack->r_ctl.rc_out_at_rto = 0;
16228 			if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
16229 				/* Not one we should rxt */
16230 				goto failed;
16231 			} else {
16232 				/* Clear the flag */
16233 				rsm->r_flags &= ~RACK_MUST_RXT;
16234 			}
16235 		} else {
16236 			/* Remove  the flag */
16237 			rsm->r_flags &= ~RACK_MUST_RXT;
16238 		}
16239 	}
16240 #ifdef TCP_ACCOUNTING
16241 	crtsc = get_cyclecount();
16242 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16243 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16244 	}
16245 	counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16246 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16247 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16248 	}
16249 	counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16250 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16251 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
16252 	}
16253 	counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
16254 	sched_unpin();
16255 #endif
16256 	return (0);
16257 failed:
16258 	if (m)
16259 		m_free(m);
16260 	return (-1);
16261 }
16262 
16263 static void
16264 rack_sndbuf_autoscale(struct tcp_rack *rack)
16265 {
16266 	/*
16267 	 * Automatic sizing of send socket buffer.  Often the send buffer
16268 	 * size is not optimally adjusted to the actual network conditions
16269 	 * at hand (delay bandwidth product).  Setting the buffer size too
16270 	 * small limits throughput on links with high bandwidth and high
16271 	 * delay (eg. trans-continental/oceanic links).  Setting the
16272 	 * buffer size too big consumes too much real kernel memory,
16273 	 * especially with many connections on busy servers.
16274 	 *
16275 	 * The criteria to step up the send buffer one notch are:
16276 	 *  1. receive window of remote host is larger than send buffer
16277 	 *     (with a fudge factor of 5/4th);
16278 	 *  2. send buffer is filled to 7/8th with data (so we actually
16279 	 *     have data to make use of it);
16280 	 *  3. send buffer fill has not hit maximal automatic size;
16281 	 *  4. our send window (slow start and cogestion controlled) is
16282 	 *     larger than sent but unacknowledged data in send buffer.
16283 	 *
16284 	 * Note that the rack version moves things much faster since
16285 	 * we want to avoid hitting cache lines in the rack_fast_output()
16286 	 * path so this is called much less often and thus moves
16287 	 * the SB forward by a percentage.
16288 	 */
16289 	struct socket *so;
16290 	struct tcpcb *tp;
16291 	uint32_t sendwin, scaleup;
16292 
16293 	tp = rack->rc_tp;
16294 	so = rack->rc_inp->inp_socket;
16295 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
16296 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
16297 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
16298 		    sbused(&so->so_snd) >=
16299 		    (so->so_snd.sb_hiwat / 8 * 7) &&
16300 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
16301 		    sendwin >= (sbused(&so->so_snd) -
16302 		    (tp->snd_nxt - tp->snd_una))) {
16303 			if (rack_autosndbuf_inc)
16304 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
16305 			else
16306 				scaleup = V_tcp_autosndbuf_inc;
16307 			if (scaleup < V_tcp_autosndbuf_inc)
16308 				scaleup = V_tcp_autosndbuf_inc;
16309 			scaleup += so->so_snd.sb_hiwat;
16310 			if (scaleup > V_tcp_autosndbuf_max)
16311 				scaleup = V_tcp_autosndbuf_max;
16312 			if (!sbreserve_locked(&so->so_snd, scaleup, so, curthread))
16313 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
16314 		}
16315 	}
16316 }
16317 
16318 static int
16319 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
16320 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
16321 {
16322 	/*
16323 	 * Enter to do fast output. We are given that the sched_pin is
16324 	 * in place (if accounting is compiled in) and the cycle count taken
16325 	 * at entry is in place in ts_val. The idea here is that
16326 	 * we know how many more bytes needs to be sent (presumably either
16327 	 * during pacing or to fill the cwnd and that was greater than
16328 	 * the max-burst). We have how much to send and all the info we
16329 	 * need to just send.
16330 	 */
16331 	struct ip *ip = NULL;
16332 	struct udphdr *udp = NULL;
16333 	struct tcphdr *th = NULL;
16334 	struct mbuf *m, *s_mb;
16335 	struct inpcb *inp;
16336 	uint8_t *cpto;
16337 	struct tcp_log_buffer *lgb;
16338 #ifdef TCP_ACCOUNTING
16339 	uint64_t crtsc;
16340 #endif
16341 	struct tcpopt to;
16342 	u_char opt[TCP_MAXOLEN];
16343 	uint32_t hdrlen, optlen;
16344 	int cnt_thru = 1;
16345 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, flags, ulen = 0;
16346 	uint32_t us_cts, s_soff;
16347 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
16348 	uint32_t if_hw_tsomaxsegsize;
16349 	uint16_t add_flag = RACK_SENT_FP;
16350 #ifdef INET6
16351 	struct ip6_hdr *ip6 = NULL;
16352 
16353 	if (rack->r_is_v6) {
16354 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
16355 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
16356 	} else
16357 #endif				/* INET6 */
16358 	{
16359 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
16360 		hdrlen = sizeof(struct tcpiphdr);
16361 	}
16362 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
16363 		m = NULL;
16364 		goto failed;
16365 	}
16366 	startseq = tp->snd_max;
16367 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16368 	inp = rack->rc_inp;
16369 	len = rack->r_ctl.fsb.left_to_send;
16370 	to.to_flags = 0;
16371 	flags = rack->r_ctl.fsb.tcp_flags;
16372 	if (tp->t_flags & TF_RCVD_TSTMP) {
16373 		to.to_tsval = ms_cts + tp->ts_offset;
16374 		to.to_tsecr = tp->ts_recent;
16375 		to.to_flags = TOF_TS;
16376 	}
16377 	optlen = tcp_addoptions(&to, opt);
16378 	hdrlen += optlen;
16379 	udp = rack->r_ctl.fsb.udp;
16380 	if (udp)
16381 		hdrlen += sizeof(struct udphdr);
16382 	if (rack->r_ctl.rc_pace_max_segs)
16383 		max_val = rack->r_ctl.rc_pace_max_segs;
16384 	else if (rack->rc_user_set_max_segs)
16385 		max_val = rack->rc_user_set_max_segs * segsiz;
16386 	else
16387 		max_val = len;
16388 	if ((tp->t_flags & TF_TSO) &&
16389 	    V_tcp_do_tso &&
16390 	    (len > segsiz) &&
16391 	    (tp->t_port == 0))
16392 		tso = 1;
16393 again:
16394 #ifdef INET6
16395 	if (MHLEN < hdrlen + max_linkhdr)
16396 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
16397 	else
16398 #endif
16399 		m = m_gethdr(M_NOWAIT, MT_DATA);
16400 	if (m == NULL)
16401 		goto failed;
16402 	m->m_data += max_linkhdr;
16403 	m->m_len = hdrlen;
16404 	th = rack->r_ctl.fsb.th;
16405 	/* Establish the len to send */
16406 	if (len > max_val)
16407 		len = max_val;
16408 	if ((tso) && (len + optlen > tp->t_maxseg)) {
16409 		uint32_t if_hw_tsomax;
16410 		int32_t max_len;
16411 
16412 		/* extract TSO information */
16413 		if_hw_tsomax = tp->t_tsomax;
16414 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
16415 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
16416 		/*
16417 		 * Check if we should limit by maximum payload
16418 		 * length:
16419 		 */
16420 		if (if_hw_tsomax != 0) {
16421 			/* compute maximum TSO length */
16422 			max_len = (if_hw_tsomax - hdrlen -
16423 				   max_linkhdr);
16424 			if (max_len <= 0) {
16425 				goto failed;
16426 			} else if (len > max_len) {
16427 				len = max_len;
16428 			}
16429 		}
16430 		if (len <= segsiz) {
16431 			/*
16432 			 * In case there are too many small fragments don't
16433 			 * use TSO:
16434 			 */
16435 			tso = 0;
16436 		}
16437 	} else {
16438 		tso = 0;
16439 	}
16440 	if ((tso == 0) && (len > segsiz))
16441 		len = segsiz;
16442 	us_cts = tcp_get_usecs(tv);
16443 	if ((len == 0) ||
16444 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
16445 		goto failed;
16446 	}
16447 	sb_offset = tp->snd_max - tp->snd_una;
16448 	th->th_seq = htonl(tp->snd_max);
16449 	th->th_ack = htonl(tp->rcv_nxt);
16450 	th->th_flags = flags;
16451 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
16452 	if (th->th_win == 0) {
16453 		tp->t_sndzerowin++;
16454 		tp->t_flags |= TF_RXWIN0SENT;
16455 	} else
16456 		tp->t_flags &= ~TF_RXWIN0SENT;
16457 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
16458 	KMOD_TCPSTAT_INC(tcps_sndpack);
16459 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
16460 #ifdef STATS
16461 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
16462 				 len);
16463 #endif
16464 	if (rack->r_ctl.fsb.m == NULL)
16465 		goto failed;
16466 
16467 	/* s_mb and s_soff are saved for rack_log_output */
16468 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
16469 				    &s_mb, &s_soff);
16470 	if (len <= segsiz) {
16471 		/*
16472 		 * Must have ran out of mbufs for the copy
16473 		 * shorten it to no longer need tso. Lets
16474 		 * not put on sendalot since we are low on
16475 		 * mbufs.
16476 		 */
16477 		tso = 0;
16478 	}
16479 	if (rack->r_ctl.fsb.rfo_apply_push &&
16480 	    (len == rack->r_ctl.fsb.left_to_send)) {
16481 		th->th_flags |= TH_PUSH;
16482 		add_flag |= RACK_HAD_PUSH;
16483 	}
16484 	if ((m->m_next == NULL) || (len <= 0)){
16485 		goto failed;
16486 	}
16487 	if (udp) {
16488 		if (rack->r_is_v6)
16489 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
16490 		else
16491 			ulen = hdrlen + len - sizeof(struct ip);
16492 		udp->uh_ulen = htons(ulen);
16493 	}
16494 	m->m_pkthdr.rcvif = (struct ifnet *)0;
16495 	if (tp->t_state == TCPS_ESTABLISHED &&
16496 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
16497 		/*
16498 		 * If the peer has ECN, mark data packets with ECN capable
16499 		 * transmission (ECT). Ignore pure ack packets,
16500 		 * retransmissions.
16501 		 */
16502 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max)) {
16503 #ifdef INET6
16504 			if (rack->r_is_v6)
16505 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
16506 			else
16507 #endif
16508 				ip->ip_tos |= IPTOS_ECN_ECT0;
16509 			KMOD_TCPSTAT_INC(tcps_ecn_ect0);
16510 			/*
16511 			 * Reply with proper ECN notifications.
16512 			 * Only set CWR on new data segments.
16513 			 */
16514 			if (tp->t_flags2 & TF2_ECN_SND_CWR) {
16515 				flags |= TH_CWR;
16516 				tp->t_flags2 &= ~TF2_ECN_SND_CWR;
16517 			}
16518 		}
16519 		if (tp->t_flags2 & TF2_ECN_SND_ECE)
16520 			flags |= TH_ECE;
16521 	}
16522 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
16523 #ifdef INET6
16524 	if (rack->r_is_v6) {
16525 		if (tp->t_port) {
16526 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
16527 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16528 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
16529 			th->th_sum = htons(0);
16530 			UDPSTAT_INC(udps_opackets);
16531 		} else {
16532 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
16533 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16534 			th->th_sum = in6_cksum_pseudo(ip6,
16535 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
16536 						      0);
16537 		}
16538 	}
16539 #endif
16540 #if defined(INET6) && defined(INET)
16541 	else
16542 #endif
16543 #ifdef INET
16544 	{
16545 		if (tp->t_port) {
16546 			m->m_pkthdr.csum_flags = CSUM_UDP;
16547 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16548 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
16549 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
16550 			th->th_sum = htons(0);
16551 			UDPSTAT_INC(udps_opackets);
16552 		} else {
16553 			m->m_pkthdr.csum_flags = CSUM_TCP;
16554 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16555 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
16556 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
16557 									IPPROTO_TCP + len + optlen));
16558 		}
16559 		/* IP version must be set here for ipv4/ipv6 checking later */
16560 		KASSERT(ip->ip_v == IPVERSION,
16561 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
16562 	}
16563 #endif
16564 	if (tso) {
16565 		KASSERT(len > tp->t_maxseg - optlen,
16566 			("%s: len <= tso_segsz tp:%p", __func__, tp));
16567 		m->m_pkthdr.csum_flags |= CSUM_TSO;
16568 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
16569 	}
16570 #ifdef INET6
16571 	if (rack->r_is_v6) {
16572 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
16573 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
16574 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
16575 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16576 		else
16577 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16578 	}
16579 #endif
16580 #if defined(INET) && defined(INET6)
16581 	else
16582 #endif
16583 #ifdef INET
16584 	{
16585 		ip->ip_len = htons(m->m_pkthdr.len);
16586 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
16587 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
16588 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16589 			if (tp->t_port == 0 || len < V_tcp_minmss) {
16590 				ip->ip_off |= htons(IP_DF);
16591 			}
16592 		} else {
16593 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16594 		}
16595 	}
16596 #endif
16597 	/* Time to copy in our header */
16598 	cpto = mtod(m, uint8_t *);
16599 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
16600 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
16601 	if (optlen) {
16602 		bcopy(opt, th + 1, optlen);
16603 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
16604 	} else {
16605 		th->th_off = sizeof(struct tcphdr) >> 2;
16606 	}
16607 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
16608 		union tcp_log_stackspecific log;
16609 
16610 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16611 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
16612 		if (rack->rack_no_prr)
16613 			log.u_bbr.flex1 = 0;
16614 		else
16615 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16616 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16617 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16618 		log.u_bbr.flex4 = max_val;
16619 		log.u_bbr.flex5 = 0;
16620 		/* Save off the early/late values */
16621 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16622 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16623 		log.u_bbr.bw_inuse = rack_get_bw(rack);
16624 		log.u_bbr.flex8 = 0;
16625 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16626 		log.u_bbr.flex7 = 44;
16627 		log.u_bbr.pkts_out = tp->t_maxseg;
16628 		log.u_bbr.timeStamp = cts;
16629 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16630 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16631 		log.u_bbr.delivered = 0;
16632 		lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16633 				     len, &log, false, NULL, NULL, 0, tv);
16634 	} else
16635 		lgb = NULL;
16636 #ifdef INET6
16637 	if (rack->r_is_v6) {
16638 		error = ip6_output(m, NULL,
16639 				   &inp->inp_route6,
16640 				   0, NULL, NULL, inp);
16641 	}
16642 #endif
16643 #if defined(INET) && defined(INET6)
16644 	else
16645 #endif
16646 #ifdef INET
16647 	{
16648 		error = ip_output(m, NULL,
16649 				  &inp->inp_route,
16650 				  0, 0, inp);
16651 	}
16652 #endif
16653 	if (lgb) {
16654 		lgb->tlb_errno = error;
16655 		lgb = NULL;
16656 	}
16657 	if (error) {
16658 		*send_err = error;
16659 		m = NULL;
16660 		goto failed;
16661 	}
16662 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
16663 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls);
16664 	m = NULL;
16665 	if (tp->snd_una == tp->snd_max) {
16666 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
16667 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
16668 		tp->t_acktime = ticks;
16669 	}
16670 	if (error == 0)
16671 		tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
16672 
16673 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
16674 	tot_len += len;
16675 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
16676 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
16677 	tp->snd_max += len;
16678 	tp->snd_nxt = tp->snd_max;
16679 	{
16680 		int idx;
16681 
16682 		idx = (len / segsiz) + 3;
16683 		if (idx >= TCP_MSS_ACCT_ATIMER)
16684 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16685 		else
16686 			counter_u64_add(rack_out_size[idx], 1);
16687 	}
16688 	if (len <= rack->r_ctl.fsb.left_to_send)
16689 		rack->r_ctl.fsb.left_to_send -= len;
16690 	else
16691 		rack->r_ctl.fsb.left_to_send = 0;
16692 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
16693 		rack->r_fast_output = 0;
16694 		rack->r_ctl.fsb.left_to_send = 0;
16695 		/* At the end of fast_output scale up the sb */
16696 		SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
16697 		rack_sndbuf_autoscale(rack);
16698 		SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
16699 	}
16700 	if (tp->t_rtttime == 0) {
16701 		tp->t_rtttime = ticks;
16702 		tp->t_rtseq = startseq;
16703 		KMOD_TCPSTAT_INC(tcps_segstimed);
16704 	}
16705 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
16706 	    (max_val > len) &&
16707 	    (tso == 0)) {
16708 		max_val -= len;
16709 		len = segsiz;
16710 		th = rack->r_ctl.fsb.th;
16711 		cnt_thru++;
16712 		goto again;
16713 	}
16714 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16715 	counter_u64_add(rack_fto_send, 1);
16716 	slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
16717 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
16718 #ifdef TCP_ACCOUNTING
16719 	crtsc = get_cyclecount();
16720 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16721 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16722 	}
16723 	counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16724 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16725 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16726 	}
16727 	counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16728 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16729 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
16730 	}
16731 	counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
16732 	sched_unpin();
16733 #endif
16734 	return (0);
16735 failed:
16736 	if (m)
16737 		m_free(m);
16738 	rack->r_fast_output = 0;
16739 	return (-1);
16740 }
16741 
16742 static int
16743 rack_output(struct tcpcb *tp)
16744 {
16745 	struct socket *so;
16746 	uint32_t recwin;
16747 	uint32_t sb_offset, s_moff = 0;
16748 	int32_t len, flags, error = 0;
16749 	struct mbuf *m, *s_mb = NULL;
16750 	struct mbuf *mb;
16751 	uint32_t if_hw_tsomaxsegcount = 0;
16752 	uint32_t if_hw_tsomaxsegsize;
16753 	int32_t segsiz, minseg;
16754 	long tot_len_this_send = 0;
16755 #ifdef INET
16756 	struct ip *ip = NULL;
16757 #endif
16758 #ifdef TCPDEBUG
16759 	struct ipovly *ipov = NULL;
16760 #endif
16761 	struct udphdr *udp = NULL;
16762 	struct tcp_rack *rack;
16763 	struct tcphdr *th;
16764 	uint8_t pass = 0;
16765 	uint8_t mark = 0;
16766 	uint8_t wanted_cookie = 0;
16767 	u_char opt[TCP_MAXOLEN];
16768 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
16769 	uint32_t rack_seq;
16770 
16771 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16772 	unsigned ipsec_optlen = 0;
16773 
16774 #endif
16775 	int32_t idle, sendalot;
16776 	int32_t sub_from_prr = 0;
16777 	volatile int32_t sack_rxmit;
16778 	struct rack_sendmap *rsm = NULL;
16779 	int32_t tso, mtu;
16780 	struct tcpopt to;
16781 	int32_t slot = 0;
16782 	int32_t sup_rack = 0;
16783 	uint32_t cts, ms_cts, delayed, early;
16784 	uint16_t add_flag = RACK_SENT_SP;
16785 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
16786 	uint8_t hpts_calling,  doing_tlp = 0;
16787 	uint32_t cwnd_to_use, pace_max_seg;
16788 	int32_t do_a_prefetch = 0;
16789 	int32_t prefetch_rsm = 0;
16790 	int32_t orig_len = 0;
16791 	struct timeval tv;
16792 	int32_t prefetch_so_done = 0;
16793 	struct tcp_log_buffer *lgb;
16794 	struct inpcb *inp;
16795 	struct sockbuf *sb;
16796 	uint64_t ts_val = 0;
16797 #ifdef TCP_ACCOUNTING
16798 	uint64_t crtsc;
16799 #endif
16800 #ifdef INET6
16801 	struct ip6_hdr *ip6 = NULL;
16802 	int32_t isipv6;
16803 #endif
16804 	uint8_t filled_all = 0;
16805 	bool hw_tls = false;
16806 
16807 	/* setup and take the cache hits here */
16808 	rack = (struct tcp_rack *)tp->t_fb_ptr;
16809 #ifdef TCP_ACCOUNTING
16810 	sched_pin();
16811 	ts_val = get_cyclecount();
16812 #endif
16813 	hpts_calling = rack->rc_inp->inp_hpts_calls;
16814 	NET_EPOCH_ASSERT();
16815 	INP_WLOCK_ASSERT(rack->rc_inp);
16816 #ifdef TCP_OFFLOAD
16817 	if (tp->t_flags & TF_TOE) {
16818 #ifdef TCP_ACCOUNTING
16819 		sched_unpin();
16820 #endif
16821 		return (tcp_offload_output(tp));
16822 	}
16823 #endif
16824 	/*
16825 	 * For TFO connections in SYN_RECEIVED, only allow the initial
16826 	 * SYN|ACK and those sent by the retransmit timer.
16827 	 */
16828 	if (IS_FASTOPEN(tp->t_flags) &&
16829 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
16830 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16831 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16832 #ifdef TCP_ACCOUNTING
16833 		sched_unpin();
16834 #endif
16835 		return (0);
16836 	}
16837 #ifdef INET6
16838 	if (rack->r_state) {
16839 		/* Use the cache line loaded if possible */
16840 		isipv6 = rack->r_is_v6;
16841 	} else {
16842 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16843 	}
16844 #endif
16845 	early = 0;
16846 	cts = tcp_get_usecs(&tv);
16847 	ms_cts = tcp_tv_to_mssectick(&tv);
16848 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16849 	    tcp_in_hpts(rack->rc_inp)) {
16850 		/*
16851 		 * We are on the hpts for some timer but not hptsi output.
16852 		 * Remove from the hpts unconditionally.
16853 		 */
16854 		rack_timer_cancel(tp, rack, cts, __LINE__);
16855 	}
16856 	/* Are we pacing and late? */
16857 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16858 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16859 		/* We are delayed */
16860 		delayed = cts - rack->r_ctl.rc_last_output_to;
16861 	} else {
16862 		delayed = 0;
16863 	}
16864 	/* Do the timers, which may override the pacer */
16865 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16866 		int retval;
16867 
16868 		retval = rack_process_timers(tp, rack, cts, hpts_calling,
16869 		    &doing_tlp);
16870 		if (retval != 0) {
16871 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16872 #ifdef TCP_ACCOUNTING
16873 			sched_unpin();
16874 #endif
16875 			/*
16876 			 * If timers want tcp_drop(), then pass error out,
16877 			 * otherwise suppress it.
16878 			 */
16879 			return (retval < 0 ? retval : 0);
16880 		}
16881 	}
16882 	if (rack->rc_in_persist) {
16883 		if (tcp_in_hpts(rack->rc_inp) == 0) {
16884 			/* Timer is not running */
16885 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16886 		}
16887 #ifdef TCP_ACCOUNTING
16888 		sched_unpin();
16889 #endif
16890 		return (0);
16891 	}
16892 	if ((rack->r_timer_override) ||
16893 	    (rack->rc_ack_can_sendout_data) ||
16894 	    (delayed) ||
16895 	    (tp->t_state < TCPS_ESTABLISHED)) {
16896 		rack->rc_ack_can_sendout_data = 0;
16897 		if (tcp_in_hpts(rack->rc_inp))
16898 			tcp_hpts_remove(rack->rc_inp);
16899 	} else if (tcp_in_hpts(rack->rc_inp)) {
16900 		/*
16901 		 * On the hpts you can't pass even if ACKNOW is on, we will
16902 		 * when the hpts fires.
16903 		 */
16904 #ifdef TCP_ACCOUNTING
16905 		crtsc = get_cyclecount();
16906 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16907 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16908 		}
16909 		counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16910 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16911 			tp->tcp_cnt_counters[SND_BLOCKED]++;
16912 		}
16913 		counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16914 		sched_unpin();
16915 #endif
16916 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16917 		return (0);
16918 	}
16919 	rack->rc_inp->inp_hpts_calls = 0;
16920 	/* Finish out both pacing early and late accounting */
16921 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16922 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16923 		early = rack->r_ctl.rc_last_output_to - cts;
16924 	} else
16925 		early = 0;
16926 	if (delayed) {
16927 		rack->r_ctl.rc_agg_delayed += delayed;
16928 		rack->r_late = 1;
16929 	} else if (early) {
16930 		rack->r_ctl.rc_agg_early += early;
16931 		rack->r_early = 1;
16932 	}
16933 	/* Now that early/late accounting is done turn off the flag */
16934 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16935 	rack->r_wanted_output = 0;
16936 	rack->r_timer_override = 0;
16937 	if ((tp->t_state != rack->r_state) &&
16938 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
16939 		rack_set_state(tp, rack);
16940 	}
16941 	if ((rack->r_fast_output) &&
16942 	    (doing_tlp == 0) &&
16943 	    (tp->rcv_numsacks == 0)) {
16944 		int ret;
16945 
16946 		error = 0;
16947 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16948 		if (ret >= 0)
16949 			return(ret);
16950 		else if (error) {
16951 			inp = rack->rc_inp;
16952 			so = inp->inp_socket;
16953 			sb = &so->so_snd;
16954 			goto nomore;
16955 		}
16956 	}
16957 	inp = rack->rc_inp;
16958 	/*
16959 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
16960 	 * only allow the initial SYN or SYN|ACK and those sent
16961 	 * by the retransmit timer.
16962 	 */
16963 	if (IS_FASTOPEN(tp->t_flags) &&
16964 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
16965 	     (tp->t_state == TCPS_SYN_SENT)) &&
16966 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16967 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
16968 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16969 		so = inp->inp_socket;
16970 		sb = &so->so_snd;
16971 		goto just_return_nolock;
16972 	}
16973 	/*
16974 	 * Determine length of data that should be transmitted, and flags
16975 	 * that will be used. If there is some data or critical controls
16976 	 * (SYN, RST) to send, then transmit; otherwise, investigate
16977 	 * further.
16978 	 */
16979 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16980 	if (tp->t_idle_reduce) {
16981 		if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16982 			rack_cc_after_idle(rack, tp);
16983 	}
16984 	tp->t_flags &= ~TF_LASTIDLE;
16985 	if (idle) {
16986 		if (tp->t_flags & TF_MORETOCOME) {
16987 			tp->t_flags |= TF_LASTIDLE;
16988 			idle = 0;
16989 		}
16990 	}
16991 	if ((tp->snd_una == tp->snd_max) &&
16992 	    rack->r_ctl.rc_went_idle_time &&
16993 	    TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16994 		idle = cts - rack->r_ctl.rc_went_idle_time;
16995 		if (idle > rack_min_probertt_hold) {
16996 			/* Count as a probe rtt */
16997 			if (rack->in_probe_rtt == 0) {
16998 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
16999 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
17000 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
17001 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
17002 			} else {
17003 				rack_exit_probertt(rack, cts);
17004 			}
17005 		}
17006 		idle = 0;
17007 	}
17008 	if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED))
17009 		rack_init_fsb_block(tp, rack);
17010 again:
17011 	/*
17012 	 * If we've recently taken a timeout, snd_max will be greater than
17013 	 * snd_nxt.  There may be SACK information that allows us to avoid
17014 	 * resending already delivered data.  Adjust snd_nxt accordingly.
17015 	 */
17016 	sendalot = 0;
17017 	cts = tcp_get_usecs(&tv);
17018 	ms_cts = tcp_tv_to_mssectick(&tv);
17019 	tso = 0;
17020 	mtu = 0;
17021 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
17022 	minseg = segsiz;
17023 	if (rack->r_ctl.rc_pace_max_segs == 0)
17024 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
17025 	else
17026 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
17027 	sb_offset = tp->snd_max - tp->snd_una;
17028 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
17029 	flags = tcp_outflags[tp->t_state];
17030 	while (rack->rc_free_cnt < rack_free_cache) {
17031 		rsm = rack_alloc(rack);
17032 		if (rsm == NULL) {
17033 			if (inp->inp_hpts_calls)
17034 				/* Retry in a ms */
17035 				slot = (1 * HPTS_USEC_IN_MSEC);
17036 			so = inp->inp_socket;
17037 			sb = &so->so_snd;
17038 			goto just_return_nolock;
17039 		}
17040 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
17041 		rack->rc_free_cnt++;
17042 		rsm = NULL;
17043 	}
17044 	if (inp->inp_hpts_calls)
17045 		inp->inp_hpts_calls = 0;
17046 	sack_rxmit = 0;
17047 	len = 0;
17048 	rsm = NULL;
17049 	if (flags & TH_RST) {
17050 		SOCKBUF_LOCK(&inp->inp_socket->so_snd);
17051 		so = inp->inp_socket;
17052 		sb = &so->so_snd;
17053 		goto send;
17054 	}
17055 	if (rack->r_ctl.rc_resend) {
17056 		/* Retransmit timer */
17057 		rsm = rack->r_ctl.rc_resend;
17058 		rack->r_ctl.rc_resend = NULL;
17059 		len = rsm->r_end - rsm->r_start;
17060 		sack_rxmit = 1;
17061 		sendalot = 0;
17062 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
17063 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
17064 			 __func__, __LINE__,
17065 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
17066 		sb_offset = rsm->r_start - tp->snd_una;
17067 		if (len >= segsiz)
17068 			len = segsiz;
17069 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
17070 		/* We have a retransmit that takes precedence */
17071 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
17072 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
17073 			/* Enter recovery if not induced by a time-out */
17074 			rack->r_ctl.rc_rsm_start = rsm->r_start;
17075 			rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
17076 			rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
17077 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
17078 		}
17079 #ifdef INVARIANTS
17080 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
17081 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
17082 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
17083 		}
17084 #endif
17085 		len = rsm->r_end - rsm->r_start;
17086 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
17087 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
17088 			 __func__, __LINE__,
17089 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
17090 		sb_offset = rsm->r_start - tp->snd_una;
17091 		sendalot = 0;
17092 		if (len >= segsiz)
17093 			len = segsiz;
17094 		if (len > 0) {
17095 			sack_rxmit = 1;
17096 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
17097 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
17098 			    min(len, segsiz));
17099 			counter_u64_add(rack_rtm_prr_retran, 1);
17100 		}
17101 	} else if (rack->r_ctl.rc_tlpsend) {
17102 		/* Tail loss probe */
17103 		long cwin;
17104 		long tlen;
17105 
17106 		/*
17107 		 * Check if we can do a TLP with a RACK'd packet
17108 		 * this can happen if we are not doing the rack
17109 		 * cheat and we skipped to a TLP and it
17110 		 * went off.
17111 		 */
17112 		rsm = rack->r_ctl.rc_tlpsend;
17113 		/* We are doing a TLP make sure the flag is preent */
17114 		rsm->r_flags |= RACK_TLP;
17115 		rack->r_ctl.rc_tlpsend = NULL;
17116 		sack_rxmit = 1;
17117 		tlen = rsm->r_end - rsm->r_start;
17118 		if (tlen > segsiz)
17119 			tlen = segsiz;
17120 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
17121 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
17122 			 __func__, __LINE__,
17123 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
17124 		sb_offset = rsm->r_start - tp->snd_una;
17125 		cwin = min(tp->snd_wnd, tlen);
17126 		len = cwin;
17127 	}
17128 	if (rack->r_must_retran &&
17129 	    (doing_tlp == 0) &&
17130 	    (rsm == NULL)) {
17131 		/*
17132 		 * Non-Sack and we had a RTO or Sack/non-Sack and a
17133 		 * MTU change, we need to retransmit until we reach
17134 		 * the former snd_max (rack->r_ctl.rc_snd_max_at_rto).
17135 		 */
17136 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
17137 			int sendwin, flight;
17138 
17139 			sendwin = min(tp->snd_wnd, tp->snd_cwnd);
17140 			flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
17141 			if (flight >= sendwin) {
17142 				so = inp->inp_socket;
17143 				sb = &so->so_snd;
17144 				goto just_return_nolock;
17145 			}
17146 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17147 			if (rsm == NULL) {
17148 				/* TSNH */
17149 				rack->r_must_retran = 0;
17150 				rack->r_ctl.rc_out_at_rto = 0;
17151 				rack->r_must_retran = 0;
17152 				so = inp->inp_socket;
17153 				sb = &so->so_snd;
17154 				goto just_return_nolock;
17155 			}
17156 			if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
17157 				/* It does not have the flag, we are done */
17158 				rack->r_must_retran = 0;
17159 				rack->r_ctl.rc_out_at_rto = 0;
17160 			} else {
17161 				sack_rxmit = 1;
17162 				len = rsm->r_end - rsm->r_start;
17163 				sendalot = 0;
17164 				sb_offset = rsm->r_start - tp->snd_una;
17165 				if (len >= segsiz)
17166 					len = segsiz;
17167 				/*
17168 				 * Delay removing the flag RACK_MUST_RXT so
17169 				 * that the fastpath for retransmit will
17170 				 * work with this rsm.
17171 				 */
17172 
17173 			}
17174 		} else {
17175 			/* We must be done if there is nothing outstanding */
17176 			rack->r_must_retran = 0;
17177 			rack->r_ctl.rc_out_at_rto = 0;
17178 		}
17179 	}
17180 	/*
17181 	 * Enforce a connection sendmap count limit if set
17182 	 * as long as we are not retransmiting.
17183 	 */
17184 	if ((rsm == NULL) &&
17185 	    (rack->do_detection == 0) &&
17186 	    (V_tcp_map_entries_limit > 0) &&
17187 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
17188 		counter_u64_add(rack_to_alloc_limited, 1);
17189 		if (!rack->alloc_limit_reported) {
17190 			rack->alloc_limit_reported = 1;
17191 			counter_u64_add(rack_alloc_limited_conns, 1);
17192 		}
17193 		so = inp->inp_socket;
17194 		sb = &so->so_snd;
17195 		goto just_return_nolock;
17196 	}
17197 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
17198 		/* we are retransmitting the fin */
17199 		len--;
17200 		if (len) {
17201 			/*
17202 			 * When retransmitting data do *not* include the
17203 			 * FIN. This could happen from a TLP probe.
17204 			 */
17205 			flags &= ~TH_FIN;
17206 		}
17207 	}
17208 #ifdef INVARIANTS
17209 	/* For debugging */
17210 	rack->r_ctl.rc_rsm_at_retran = rsm;
17211 #endif
17212 	if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
17213 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
17214 		int ret;
17215 
17216 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
17217 		if (ret == 0)
17218 			return (0);
17219 	}
17220 	if (rsm && (rsm->r_flags & RACK_MUST_RXT)) {
17221 		/*
17222 		 * Clear the flag in prep for the send
17223 		 * note that if we can't get an mbuf
17224 		 * and fail, we won't retransmit this
17225 		 * rsm but that should be ok (its rare).
17226 		 */
17227 		rsm->r_flags &= ~RACK_MUST_RXT;
17228 	}
17229 	so = inp->inp_socket;
17230 	sb = &so->so_snd;
17231 	if (do_a_prefetch == 0) {
17232 		kern_prefetch(sb, &do_a_prefetch);
17233 		do_a_prefetch = 1;
17234 	}
17235 #ifdef NETFLIX_SHARED_CWND
17236 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
17237 	    rack->rack_enable_scwnd) {
17238 		/* We are doing cwnd sharing */
17239 		if (rack->gp_ready &&
17240 		    (rack->rack_attempted_scwnd == 0) &&
17241 		    (rack->r_ctl.rc_scw == NULL) &&
17242 		    tp->t_lib) {
17243 			/* The pcbid is in, lets make an attempt */
17244 			counter_u64_add(rack_try_scwnd, 1);
17245 			rack->rack_attempted_scwnd = 1;
17246 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
17247 								   &rack->r_ctl.rc_scw_index,
17248 								   segsiz);
17249 		}
17250 		if (rack->r_ctl.rc_scw &&
17251 		    (rack->rack_scwnd_is_idle == 1) &&
17252 		    sbavail(&so->so_snd)) {
17253 			/* we are no longer out of data */
17254 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17255 			rack->rack_scwnd_is_idle = 0;
17256 		}
17257 		if (rack->r_ctl.rc_scw) {
17258 			/* First lets update and get the cwnd */
17259 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
17260 								    rack->r_ctl.rc_scw_index,
17261 								    tp->snd_cwnd, tp->snd_wnd, segsiz);
17262 		}
17263 	}
17264 #endif
17265 	/*
17266 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
17267 	 * state flags.
17268 	 */
17269 	if (tp->t_flags & TF_NEEDFIN)
17270 		flags |= TH_FIN;
17271 	if (tp->t_flags & TF_NEEDSYN)
17272 		flags |= TH_SYN;
17273 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
17274 		void *end_rsm;
17275 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
17276 		if (end_rsm)
17277 			kern_prefetch(end_rsm, &prefetch_rsm);
17278 		prefetch_rsm = 1;
17279 	}
17280 	SOCKBUF_LOCK(sb);
17281 	/*
17282 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
17283 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
17284 	 * negative length.  This can also occur when TCP opens up its
17285 	 * congestion window while receiving additional duplicate acks after
17286 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
17287 	 * the fast-retransmit.
17288 	 *
17289 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
17290 	 * set to snd_una, the sb_offset will be 0, and the length may wind
17291 	 * up 0.
17292 	 *
17293 	 * If sack_rxmit is true we are retransmitting from the scoreboard
17294 	 * in which case len is already set.
17295 	 */
17296 	if ((sack_rxmit == 0) &&
17297 	    (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
17298 		uint32_t avail;
17299 
17300 		avail = sbavail(sb);
17301 		if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
17302 			sb_offset = tp->snd_nxt - tp->snd_una;
17303 		else
17304 			sb_offset = 0;
17305 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
17306 			if (rack->r_ctl.rc_tlp_new_data) {
17307 				/* TLP is forcing out new data */
17308 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
17309 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
17310 				}
17311 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
17312 					if (tp->snd_wnd > sb_offset)
17313 						len = tp->snd_wnd - sb_offset;
17314 					else
17315 						len = 0;
17316 				} else {
17317 					len = rack->r_ctl.rc_tlp_new_data;
17318 				}
17319 				rack->r_ctl.rc_tlp_new_data = 0;
17320 			}  else {
17321 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
17322 			}
17323 			if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
17324 				/*
17325 				 * For prr=off, we need to send only 1 MSS
17326 				 * at a time. We do this because another sack could
17327 				 * be arriving that causes us to send retransmits and
17328 				 * we don't want to be on a long pace due to a larger send
17329 				 * that keeps us from sending out the retransmit.
17330 				 */
17331 				len = segsiz;
17332 			}
17333 		} else {
17334 			uint32_t outstanding;
17335 			/*
17336 			 * We are inside of a Fast recovery episode, this
17337 			 * is caused by a SACK or 3 dup acks. At this point
17338 			 * we have sent all the retransmissions and we rely
17339 			 * on PRR to dictate what we will send in the form of
17340 			 * new data.
17341 			 */
17342 
17343 			outstanding = tp->snd_max - tp->snd_una;
17344 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
17345 				if (tp->snd_wnd > outstanding) {
17346 					len = tp->snd_wnd - outstanding;
17347 					/* Check to see if we have the data */
17348 					if ((sb_offset + len) > avail) {
17349 						/* It does not all fit */
17350 						if (avail > sb_offset)
17351 							len = avail - sb_offset;
17352 						else
17353 							len = 0;
17354 					}
17355 				} else {
17356 					len = 0;
17357 				}
17358 			} else if (avail > sb_offset) {
17359 				len = avail - sb_offset;
17360 			} else {
17361 				len = 0;
17362 			}
17363 			if (len > 0) {
17364 				if (len > rack->r_ctl.rc_prr_sndcnt) {
17365 					len = rack->r_ctl.rc_prr_sndcnt;
17366 				}
17367 				if (len > 0) {
17368 					sub_from_prr = 1;
17369 					counter_u64_add(rack_rtm_prr_newdata, 1);
17370 				}
17371 			}
17372 			if (len > segsiz) {
17373 				/*
17374 				 * We should never send more than a MSS when
17375 				 * retransmitting or sending new data in prr
17376 				 * mode unless the override flag is on. Most
17377 				 * likely the PRR algorithm is not going to
17378 				 * let us send a lot as well :-)
17379 				 */
17380 				if (rack->r_ctl.rc_prr_sendalot == 0) {
17381 					len = segsiz;
17382 				}
17383 			} else if (len < segsiz) {
17384 				/*
17385 				 * Do we send any? The idea here is if the
17386 				 * send empty's the socket buffer we want to
17387 				 * do it. However if not then lets just wait
17388 				 * for our prr_sndcnt to get bigger.
17389 				 */
17390 				long leftinsb;
17391 
17392 				leftinsb = sbavail(sb) - sb_offset;
17393 				if (leftinsb > len) {
17394 					/* This send does not empty the sb */
17395 					len = 0;
17396 				}
17397 			}
17398 		}
17399 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
17400 		/*
17401 		 * If you have not established
17402 		 * and are not doing FAST OPEN
17403 		 * no data please.
17404 		 */
17405 		if ((sack_rxmit == 0) &&
17406 		    (!IS_FASTOPEN(tp->t_flags))){
17407 			len = 0;
17408 			sb_offset = 0;
17409 		}
17410 	}
17411 	if (prefetch_so_done == 0) {
17412 		kern_prefetch(so, &prefetch_so_done);
17413 		prefetch_so_done = 1;
17414 	}
17415 	/*
17416 	 * Lop off SYN bit if it has already been sent.  However, if this is
17417 	 * SYN-SENT state and if segment contains data and if we don't know
17418 	 * that foreign host supports TAO, suppress sending segment.
17419 	 */
17420 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
17421 	    ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
17422 		/*
17423 		 * When sending additional segments following a TFO SYN|ACK,
17424 		 * do not include the SYN bit.
17425 		 */
17426 		if (IS_FASTOPEN(tp->t_flags) &&
17427 		    (tp->t_state == TCPS_SYN_RECEIVED))
17428 			flags &= ~TH_SYN;
17429 	}
17430 	/*
17431 	 * Be careful not to send data and/or FIN on SYN segments. This
17432 	 * measure is needed to prevent interoperability problems with not
17433 	 * fully conformant TCP implementations.
17434 	 */
17435 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
17436 		len = 0;
17437 		flags &= ~TH_FIN;
17438 	}
17439 	/*
17440 	 * On TFO sockets, ensure no data is sent in the following cases:
17441 	 *
17442 	 *  - When retransmitting SYN|ACK on a passively-created socket
17443 	 *
17444 	 *  - When retransmitting SYN on an actively created socket
17445 	 *
17446 	 *  - When sending a zero-length cookie (cookie request) on an
17447 	 *    actively created socket
17448 	 *
17449 	 *  - When the socket is in the CLOSED state (RST is being sent)
17450 	 */
17451 	if (IS_FASTOPEN(tp->t_flags) &&
17452 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
17453 	     ((tp->t_state == TCPS_SYN_SENT) &&
17454 	      (tp->t_tfo_client_cookie_len == 0)) ||
17455 	     (flags & TH_RST))) {
17456 		sack_rxmit = 0;
17457 		len = 0;
17458 	}
17459 	/* Without fast-open there should never be data sent on a SYN */
17460 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
17461 		tp->snd_nxt = tp->iss;
17462 		len = 0;
17463 	}
17464 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
17465 		/* We only send 1 MSS if we have a DSACK block */
17466 		add_flag |= RACK_SENT_W_DSACK;
17467 		len = segsiz;
17468 	}
17469 	orig_len = len;
17470 	if (len <= 0) {
17471 		/*
17472 		 * If FIN has been sent but not acked, but we haven't been
17473 		 * called to retransmit, len will be < 0.  Otherwise, window
17474 		 * shrank after we sent into it.  If window shrank to 0,
17475 		 * cancel pending retransmit, pull snd_nxt back to (closed)
17476 		 * window, and set the persist timer if it isn't already
17477 		 * going.  If the window didn't close completely, just wait
17478 		 * for an ACK.
17479 		 *
17480 		 * We also do a general check here to ensure that we will
17481 		 * set the persist timer when we have data to send, but a
17482 		 * 0-byte window. This makes sure the persist timer is set
17483 		 * even if the packet hits one of the "goto send" lines
17484 		 * below.
17485 		 */
17486 		len = 0;
17487 		if ((tp->snd_wnd == 0) &&
17488 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17489 		    (tp->snd_una == tp->snd_max) &&
17490 		    (sb_offset < (int)sbavail(sb))) {
17491 			rack_enter_persist(tp, rack, cts);
17492 		}
17493 	} else if ((rsm == NULL) &&
17494 		   (doing_tlp == 0) &&
17495 		   (len < pace_max_seg)) {
17496 		/*
17497 		 * We are not sending a maximum sized segment for
17498 		 * some reason. Should we not send anything (think
17499 		 * sws or persists)?
17500 		 */
17501 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17502 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17503 		    (len < minseg) &&
17504 		    (len < (int)(sbavail(sb) - sb_offset))) {
17505 			/*
17506 			 * Here the rwnd is less than
17507 			 * the minimum pacing size, this is not a retransmit,
17508 			 * we are established and
17509 			 * the send is not the last in the socket buffer
17510 			 * we send nothing, and we may enter persists
17511 			 * if nothing is outstanding.
17512 			 */
17513 			len = 0;
17514 			if (tp->snd_max == tp->snd_una) {
17515 				/*
17516 				 * Nothing out we can
17517 				 * go into persists.
17518 				 */
17519 				rack_enter_persist(tp, rack, cts);
17520 			}
17521 		     } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
17522 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17523 			   (len < (int)(sbavail(sb) - sb_offset)) &&
17524 			   (len < minseg)) {
17525 			/*
17526 			 * Here we are not retransmitting, and
17527 			 * the cwnd is not so small that we could
17528 			 * not send at least a min size (rxt timer
17529 			 * not having gone off), We have 2 segments or
17530 			 * more already in flight, its not the tail end
17531 			 * of the socket buffer  and the cwnd is blocking
17532 			 * us from sending out a minimum pacing segment size.
17533 			 * Lets not send anything.
17534 			 */
17535 			len = 0;
17536 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
17537 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17538 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17539 			   (len < (int)(sbavail(sb) - sb_offset)) &&
17540 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
17541 			/*
17542 			 * Here we have a send window but we have
17543 			 * filled it up and we can't send another pacing segment.
17544 			 * We also have in flight more than 2 segments
17545 			 * and we are not completing the sb i.e. we allow
17546 			 * the last bytes of the sb to go out even if
17547 			 * its not a full pacing segment.
17548 			 */
17549 			len = 0;
17550 		} else if ((rack->r_ctl.crte != NULL) &&
17551 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
17552 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
17553 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
17554 			   (len < (int)(sbavail(sb) - sb_offset))) {
17555 			/*
17556 			 * Here we are doing hardware pacing, this is not a TLP,
17557 			 * we are not sending a pace max segment size, there is rwnd
17558 			 * room to send at least N pace_max_seg, the cwnd is greater
17559 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
17560 			 * more segments in flight and its not the tail of the socket buffer.
17561 			 *
17562 			 * We don't want to send instead we need to get more ack's in to
17563 			 * allow us to send a full pacing segment. Normally, if we are pacing
17564 			 * about the right speed, we should have finished our pacing
17565 			 * send as most of the acks have come back if we are at the
17566 			 * right rate. This is a bit fuzzy since return path delay
17567 			 * can delay the acks, which is why we want to make sure we
17568 			 * have cwnd space to have a bit more than a max pace segments in flight.
17569 			 *
17570 			 * If we have not gotten our acks back we are pacing at too high a
17571 			 * rate delaying will not hurt and will bring our GP estimate down by
17572 			 * injecting the delay. If we don't do this we will send
17573 			 * 2 MSS out in response to the acks being clocked in which
17574 			 * defeats the point of hw-pacing (i.e. to help us get
17575 			 * larger TSO's out).
17576 			 */
17577 			len = 0;
17578 
17579 		}
17580 
17581 	}
17582 	/* len will be >= 0 after this point. */
17583 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17584 	rack_sndbuf_autoscale(rack);
17585 	/*
17586 	 * Decide if we can use TCP Segmentation Offloading (if supported by
17587 	 * hardware).
17588 	 *
17589 	 * TSO may only be used if we are in a pure bulk sending state.  The
17590 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
17591 	 * options prevent using TSO.  With TSO the TCP header is the same
17592 	 * (except for the sequence number) for all generated packets.  This
17593 	 * makes it impossible to transmit any options which vary per
17594 	 * generated segment or packet.
17595 	 *
17596 	 * IPv4 handling has a clear separation of ip options and ip header
17597 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
17598 	 * the right thing below to provide length of just ip options and thus
17599 	 * checking for ipoptlen is enough to decide if ip options are present.
17600 	 */
17601 	ipoptlen = 0;
17602 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17603 	/*
17604 	 * Pre-calculate here as we save another lookup into the darknesses
17605 	 * of IPsec that way and can actually decide if TSO is ok.
17606 	 */
17607 #ifdef INET6
17608 	if (isipv6 && IPSEC_ENABLED(ipv6))
17609 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
17610 #ifdef INET
17611 	else
17612 #endif
17613 #endif				/* INET6 */
17614 #ifdef INET
17615 		if (IPSEC_ENABLED(ipv4))
17616 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
17617 #endif				/* INET */
17618 #endif
17619 
17620 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17621 	ipoptlen += ipsec_optlen;
17622 #endif
17623 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
17624 	    (tp->t_port == 0) &&
17625 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
17626 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
17627 	    ipoptlen == 0)
17628 		tso = 1;
17629 	{
17630 		uint32_t outstanding;
17631 
17632 		outstanding = tp->snd_max - tp->snd_una;
17633 		if (tp->t_flags & TF_SENTFIN) {
17634 			/*
17635 			 * If we sent a fin, snd_max is 1 higher than
17636 			 * snd_una
17637 			 */
17638 			outstanding--;
17639 		}
17640 		if (sack_rxmit) {
17641 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
17642 				flags &= ~TH_FIN;
17643 		} else {
17644 			if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
17645 				   sbused(sb)))
17646 				flags &= ~TH_FIN;
17647 		}
17648 	}
17649 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
17650 	    (long)TCP_MAXWIN << tp->rcv_scale);
17651 
17652 	/*
17653 	 * Sender silly window avoidance.   We transmit under the following
17654 	 * conditions when len is non-zero:
17655 	 *
17656 	 * - We have a full segment (or more with TSO) - This is the last
17657 	 * buffer in a write()/send() and we are either idle or running
17658 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
17659 	 * then 1/2 the maximum send window's worth of data (receiver may be
17660 	 * limited the window size) - we need to retransmit
17661 	 */
17662 	if (len) {
17663 		if (len >= segsiz) {
17664 			goto send;
17665 		}
17666 		/*
17667 		 * NOTE! on localhost connections an 'ack' from the remote
17668 		 * end may occur synchronously with the output and cause us
17669 		 * to flush a buffer queued with moretocome.  XXX
17670 		 *
17671 		 */
17672 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
17673 		    (idle || (tp->t_flags & TF_NODELAY)) &&
17674 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17675 		    (tp->t_flags & TF_NOPUSH) == 0) {
17676 			pass = 2;
17677 			goto send;
17678 		}
17679 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
17680 			pass = 22;
17681 			goto send;
17682 		}
17683 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
17684 			pass = 4;
17685 			goto send;
17686 		}
17687 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {	/* retransmit case */
17688 			pass = 5;
17689 			goto send;
17690 		}
17691 		if (sack_rxmit) {
17692 			pass = 6;
17693 			goto send;
17694 		}
17695 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
17696 		    (ctf_outstanding(tp) < (segsiz * 2))) {
17697 			/*
17698 			 * We have less than two MSS outstanding (delayed ack)
17699 			 * and our rwnd will not let us send a full sized
17700 			 * MSS. Lets go ahead and let this small segment
17701 			 * out because we want to try to have at least two
17702 			 * packets inflight to not be caught by delayed ack.
17703 			 */
17704 			pass = 12;
17705 			goto send;
17706 		}
17707 	}
17708 	/*
17709 	 * Sending of standalone window updates.
17710 	 *
17711 	 * Window updates are important when we close our window due to a
17712 	 * full socket buffer and are opening it again after the application
17713 	 * reads data from it.  Once the window has opened again and the
17714 	 * remote end starts to send again the ACK clock takes over and
17715 	 * provides the most current window information.
17716 	 *
17717 	 * We must avoid the silly window syndrome whereas every read from
17718 	 * the receive buffer, no matter how small, causes a window update
17719 	 * to be sent.  We also should avoid sending a flurry of window
17720 	 * updates when the socket buffer had queued a lot of data and the
17721 	 * application is doing small reads.
17722 	 *
17723 	 * Prevent a flurry of pointless window updates by only sending an
17724 	 * update when we can increase the advertized window by more than
17725 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
17726 	 * full or is very small be more aggressive and send an update
17727 	 * whenever we can increase by two mss sized segments. In all other
17728 	 * situations the ACK's to new incoming data will carry further
17729 	 * window increases.
17730 	 *
17731 	 * Don't send an independent window update if a delayed ACK is
17732 	 * pending (it will get piggy-backed on it) or the remote side
17733 	 * already has done a half-close and won't send more data.  Skip
17734 	 * this if the connection is in T/TCP half-open state.
17735 	 */
17736 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
17737 	    !(tp->t_flags & TF_DELACK) &&
17738 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
17739 		/*
17740 		 * "adv" is the amount we could increase the window, taking
17741 		 * into account that we are limited by TCP_MAXWIN <<
17742 		 * tp->rcv_scale.
17743 		 */
17744 		int32_t adv;
17745 		int oldwin;
17746 
17747 		adv = recwin;
17748 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
17749 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
17750 			if (adv > oldwin)
17751 			    adv -= oldwin;
17752 			else {
17753 				/* We can't increase the window */
17754 				adv = 0;
17755 			}
17756 		} else
17757 			oldwin = 0;
17758 
17759 		/*
17760 		 * If the new window size ends up being the same as or less
17761 		 * than the old size when it is scaled, then don't force
17762 		 * a window update.
17763 		 */
17764 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
17765 			goto dontupdate;
17766 
17767 		if (adv >= (int32_t)(2 * segsiz) &&
17768 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
17769 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
17770 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
17771 			pass = 7;
17772 			goto send;
17773 		}
17774 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
17775 			pass = 23;
17776 			goto send;
17777 		}
17778 	}
17779 dontupdate:
17780 
17781 	/*
17782 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
17783 	 * is also a catch-all for the retransmit timer timeout case.
17784 	 */
17785 	if (tp->t_flags & TF_ACKNOW) {
17786 		pass = 8;
17787 		goto send;
17788 	}
17789 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
17790 		pass = 9;
17791 		goto send;
17792 	}
17793 	/*
17794 	 * If our state indicates that FIN should be sent and we have not
17795 	 * yet done so, then we need to send.
17796 	 */
17797 	if ((flags & TH_FIN) &&
17798 	    (tp->snd_nxt == tp->snd_una)) {
17799 		pass = 11;
17800 		goto send;
17801 	}
17802 	/*
17803 	 * No reason to send a segment, just return.
17804 	 */
17805 just_return:
17806 	SOCKBUF_UNLOCK(sb);
17807 just_return_nolock:
17808 	{
17809 		int app_limited = CTF_JR_SENT_DATA;
17810 
17811 		if (tot_len_this_send > 0) {
17812 			/* Make sure snd_nxt is up to max */
17813 			rack->r_ctl.fsb.recwin = recwin;
17814 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17815 			if ((error == 0) &&
17816 			    rack_use_rfo &&
17817 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
17818 			    (ipoptlen == 0) &&
17819 			    (tp->snd_nxt == tp->snd_max) &&
17820 			    (tp->rcv_numsacks == 0) &&
17821 			    rack->r_fsb_inited &&
17822 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
17823 			    (rack->r_must_retran == 0) &&
17824 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
17825 			    (len > 0) && (orig_len > 0) &&
17826 			    (orig_len > len) &&
17827 			    ((orig_len - len) >= segsiz) &&
17828 			    ((optlen == 0) ||
17829 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17830 				/* We can send at least one more MSS using our fsb */
17831 
17832 				rack->r_fast_output = 1;
17833 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17834 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17835 				rack->r_ctl.fsb.tcp_flags = flags;
17836 				rack->r_ctl.fsb.left_to_send = orig_len - len;
17837 				if (hw_tls)
17838 					rack->r_ctl.fsb.hw_tls = 1;
17839 				else
17840 					rack->r_ctl.fsb.hw_tls = 0;
17841 				KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17842 					("rack:%p left_to_send:%u sbavail:%u out:%u",
17843 					rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17844 					 (tp->snd_max - tp->snd_una)));
17845 				if (rack->r_ctl.fsb.left_to_send < segsiz)
17846 					rack->r_fast_output = 0;
17847 				else {
17848 					if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17849 						rack->r_ctl.fsb.rfo_apply_push = 1;
17850 					else
17851 						rack->r_ctl.fsb.rfo_apply_push = 0;
17852 				}
17853 			} else
17854 				rack->r_fast_output = 0;
17855 
17856 
17857 			rack_log_fsb(rack, tp, so, flags,
17858 				     ipoptlen, orig_len, len, 0,
17859 				     1, optlen, __LINE__, 1);
17860 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17861 				tp->snd_nxt = tp->snd_max;
17862 		} else {
17863 			int end_window = 0;
17864 			uint32_t seq = tp->gput_ack;
17865 
17866 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17867 			if (rsm) {
17868 				/*
17869 				 * Mark the last sent that we just-returned (hinting
17870 				 * that delayed ack may play a role in any rtt measurement).
17871 				 */
17872 				rsm->r_just_ret = 1;
17873 			}
17874 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17875 			rack->r_ctl.rc_agg_delayed = 0;
17876 			rack->r_early = 0;
17877 			rack->r_late = 0;
17878 			rack->r_ctl.rc_agg_early = 0;
17879 			if ((ctf_outstanding(tp) +
17880 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17881 				 minseg)) >= tp->snd_wnd) {
17882 				/* We are limited by the rwnd */
17883 				app_limited = CTF_JR_RWND_LIMITED;
17884 				if (IN_FASTRECOVERY(tp->t_flags))
17885 				    rack->r_ctl.rc_prr_sndcnt = 0;
17886 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
17887 				/* We are limited by whats available -- app limited */
17888 				app_limited = CTF_JR_APP_LIMITED;
17889 				if (IN_FASTRECOVERY(tp->t_flags))
17890 				    rack->r_ctl.rc_prr_sndcnt = 0;
17891 			} else if ((idle == 0) &&
17892 				   ((tp->t_flags & TF_NODELAY) == 0) &&
17893 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17894 				   (len < segsiz)) {
17895 				/*
17896 				 * No delay is not on and the
17897 				 * user is sending less than 1MSS. This
17898 				 * brings out SWS avoidance so we
17899 				 * don't send. Another app-limited case.
17900 				 */
17901 				app_limited = CTF_JR_APP_LIMITED;
17902 			} else if (tp->t_flags & TF_NOPUSH) {
17903 				/*
17904 				 * The user has requested no push of
17905 				 * the last segment and we are
17906 				 * at the last segment. Another app
17907 				 * limited case.
17908 				 */
17909 				app_limited = CTF_JR_APP_LIMITED;
17910 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17911 				/* Its the cwnd */
17912 				app_limited = CTF_JR_CWND_LIMITED;
17913 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
17914 				   (rack->rack_no_prr == 0) &&
17915 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17916 				app_limited = CTF_JR_PRR;
17917 			} else {
17918 				/* Now why here are we not sending? */
17919 #ifdef NOW
17920 #ifdef INVARIANTS
17921 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17922 #endif
17923 #endif
17924 				app_limited = CTF_JR_ASSESSING;
17925 			}
17926 			/*
17927 			 * App limited in some fashion, for our pacing GP
17928 			 * measurements we don't want any gap (even cwnd).
17929 			 * Close  down the measurement window.
17930 			 */
17931 			if (rack_cwnd_block_ends_measure &&
17932 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
17933 			     (app_limited == CTF_JR_PRR))) {
17934 				/*
17935 				 * The reason we are not sending is
17936 				 * the cwnd (or prr). We have been configured
17937 				 * to end the measurement window in
17938 				 * this case.
17939 				 */
17940 				end_window = 1;
17941 			} else if (rack_rwnd_block_ends_measure &&
17942 				   (app_limited == CTF_JR_RWND_LIMITED)) {
17943 				/*
17944 				 * We are rwnd limited and have been
17945 				 * configured to end the measurement
17946 				 * window in this case.
17947 				 */
17948 				end_window = 1;
17949 			} else if (app_limited == CTF_JR_APP_LIMITED) {
17950 				/*
17951 				 * A true application limited period, we have
17952 				 * ran out of data.
17953 				 */
17954 				end_window = 1;
17955 			} else if (app_limited == CTF_JR_ASSESSING) {
17956 				/*
17957 				 * In the assessing case we hit the end of
17958 				 * the if/else and had no known reason
17959 				 * This will panic us under invariants..
17960 				 *
17961 				 * If we get this out in logs we need to
17962 				 * investagate which reason we missed.
17963 				 */
17964 				end_window = 1;
17965 			}
17966 			if (end_window) {
17967 				uint8_t log = 0;
17968 
17969 				/* Adjust the Gput measurement */
17970 				if ((tp->t_flags & TF_GPUTINPROG) &&
17971 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
17972 					tp->gput_ack = tp->snd_max;
17973 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
17974 						/*
17975 						 * There is not enough to measure.
17976 						 */
17977 						tp->t_flags &= ~TF_GPUTINPROG;
17978 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17979 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
17980 									   tp->gput_seq,
17981 									   0, 0, 18, __LINE__, NULL, 0);
17982 					} else
17983 						log = 1;
17984 				}
17985 				/* Mark the last packet has app limited */
17986 				rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17987 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17988 					if (rack->r_ctl.rc_app_limited_cnt == 0)
17989 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17990 					else {
17991 						/*
17992 						 * Go out to the end app limited and mark
17993 						 * this new one as next and move the end_appl up
17994 						 * to this guy.
17995 						 */
17996 						if (rack->r_ctl.rc_end_appl)
17997 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17998 						rack->r_ctl.rc_end_appl = rsm;
17999 					}
18000 					rsm->r_flags |= RACK_APP_LIMITED;
18001 					rack->r_ctl.rc_app_limited_cnt++;
18002 				}
18003 				if (log)
18004 					rack_log_pacing_delay_calc(rack,
18005 								   rack->r_ctl.rc_app_limited_cnt, seq,
18006 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
18007 			}
18008 		}
18009 		if (slot) {
18010 			/* set the rack tcb into the slot N */
18011 			counter_u64_add(rack_paced_segments, 1);
18012 		} else if (tot_len_this_send) {
18013 			counter_u64_add(rack_unpaced_segments, 1);
18014 		}
18015 		/* Check if we need to go into persists or not */
18016 		if ((tp->snd_max == tp->snd_una) &&
18017 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
18018 		    sbavail(sb) &&
18019 		    (sbavail(sb) > tp->snd_wnd) &&
18020 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
18021 			/* Yes lets make sure to move to persist before timer-start */
18022 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
18023 		}
18024 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
18025 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
18026 	}
18027 #ifdef NETFLIX_SHARED_CWND
18028 	if ((sbavail(sb) == 0) &&
18029 	    rack->r_ctl.rc_scw) {
18030 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
18031 		rack->rack_scwnd_is_idle = 1;
18032 	}
18033 #endif
18034 #ifdef TCP_ACCOUNTING
18035 	if (tot_len_this_send > 0) {
18036 		crtsc = get_cyclecount();
18037 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18038 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
18039 		}
18040 		counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
18041 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18042 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
18043 		}
18044 		counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
18045 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18046 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
18047 		}
18048 		counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
18049 	} else {
18050 		crtsc = get_cyclecount();
18051 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18052 			tp->tcp_cnt_counters[SND_LIMITED]++;
18053 		}
18054 		counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
18055 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18056 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
18057 		}
18058 		counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
18059 	}
18060 	sched_unpin();
18061 #endif
18062 	return (0);
18063 
18064 send:
18065 	if (rsm || sack_rxmit)
18066 		counter_u64_add(rack_nfto_resend, 1);
18067 	else
18068 		counter_u64_add(rack_non_fto_send, 1);
18069 	if ((flags & TH_FIN) &&
18070 	    sbavail(sb)) {
18071 		/*
18072 		 * We do not transmit a FIN
18073 		 * with data outstanding. We
18074 		 * need to make it so all data
18075 		 * is acked first.
18076 		 */
18077 		flags &= ~TH_FIN;
18078 	}
18079 	/* Enforce stack imposed max seg size if we have one */
18080 	if (rack->r_ctl.rc_pace_max_segs &&
18081 	    (len > rack->r_ctl.rc_pace_max_segs)) {
18082 		mark = 1;
18083 		len = rack->r_ctl.rc_pace_max_segs;
18084 	}
18085 	SOCKBUF_LOCK_ASSERT(sb);
18086 	if (len > 0) {
18087 		if (len >= segsiz)
18088 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
18089 		else
18090 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
18091 	}
18092 	/*
18093 	 * Before ESTABLISHED, force sending of initial options unless TCP
18094 	 * set not to do any options. NOTE: we assume that the IP/TCP header
18095 	 * plus TCP options always fit in a single mbuf, leaving room for a
18096 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
18097 	 * + optlen <= MCLBYTES
18098 	 */
18099 	optlen = 0;
18100 #ifdef INET6
18101 	if (isipv6)
18102 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
18103 	else
18104 #endif
18105 		hdrlen = sizeof(struct tcpiphdr);
18106 
18107 	/*
18108 	 * Compute options for segment. We only have to care about SYN and
18109 	 * established connection segments.  Options for SYN-ACK segments
18110 	 * are handled in TCP syncache.
18111 	 */
18112 	to.to_flags = 0;
18113 	if ((tp->t_flags & TF_NOOPT) == 0) {
18114 		/* Maximum segment size. */
18115 		if (flags & TH_SYN) {
18116 			tp->snd_nxt = tp->iss;
18117 			to.to_mss = tcp_mssopt(&inp->inp_inc);
18118 			if (tp->t_port)
18119 				to.to_mss -= V_tcp_udp_tunneling_overhead;
18120 			to.to_flags |= TOF_MSS;
18121 
18122 			/*
18123 			 * On SYN or SYN|ACK transmits on TFO connections,
18124 			 * only include the TFO option if it is not a
18125 			 * retransmit, as the presence of the TFO option may
18126 			 * have caused the original SYN or SYN|ACK to have
18127 			 * been dropped by a middlebox.
18128 			 */
18129 			if (IS_FASTOPEN(tp->t_flags) &&
18130 			    (tp->t_rxtshift == 0)) {
18131 				if (tp->t_state == TCPS_SYN_RECEIVED) {
18132 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
18133 					to.to_tfo_cookie =
18134 						(u_int8_t *)&tp->t_tfo_cookie.server;
18135 					to.to_flags |= TOF_FASTOPEN;
18136 					wanted_cookie = 1;
18137 				} else if (tp->t_state == TCPS_SYN_SENT) {
18138 					to.to_tfo_len =
18139 						tp->t_tfo_client_cookie_len;
18140 					to.to_tfo_cookie =
18141 						tp->t_tfo_cookie.client;
18142 					to.to_flags |= TOF_FASTOPEN;
18143 					wanted_cookie = 1;
18144 					/*
18145 					 * If we wind up having more data to
18146 					 * send with the SYN than can fit in
18147 					 * one segment, don't send any more
18148 					 * until the SYN|ACK comes back from
18149 					 * the other end.
18150 					 */
18151 					sendalot = 0;
18152 				}
18153 			}
18154 		}
18155 		/* Window scaling. */
18156 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
18157 			to.to_wscale = tp->request_r_scale;
18158 			to.to_flags |= TOF_SCALE;
18159 		}
18160 		/* Timestamps. */
18161 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
18162 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
18163 			to.to_tsval = ms_cts + tp->ts_offset;
18164 			to.to_tsecr = tp->ts_recent;
18165 			to.to_flags |= TOF_TS;
18166 		}
18167 		/* Set receive buffer autosizing timestamp. */
18168 		if (tp->rfbuf_ts == 0 &&
18169 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
18170 			tp->rfbuf_ts = tcp_ts_getticks();
18171 		/* Selective ACK's. */
18172 		if (tp->t_flags & TF_SACK_PERMIT) {
18173 			if (flags & TH_SYN)
18174 				to.to_flags |= TOF_SACKPERM;
18175 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18176 				 tp->rcv_numsacks > 0) {
18177 				to.to_flags |= TOF_SACK;
18178 				to.to_nsacks = tp->rcv_numsacks;
18179 				to.to_sacks = (u_char *)tp->sackblks;
18180 			}
18181 		}
18182 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18183 		/* TCP-MD5 (RFC2385). */
18184 		if (tp->t_flags & TF_SIGNATURE)
18185 			to.to_flags |= TOF_SIGNATURE;
18186 #endif				/* TCP_SIGNATURE */
18187 
18188 		/* Processing the options. */
18189 		hdrlen += optlen = tcp_addoptions(&to, opt);
18190 		/*
18191 		 * If we wanted a TFO option to be added, but it was unable
18192 		 * to fit, ensure no data is sent.
18193 		 */
18194 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
18195 		    !(to.to_flags & TOF_FASTOPEN))
18196 			len = 0;
18197 	}
18198 	if (tp->t_port) {
18199 		if (V_tcp_udp_tunneling_port == 0) {
18200 			/* The port was removed?? */
18201 			SOCKBUF_UNLOCK(&so->so_snd);
18202 #ifdef TCP_ACCOUNTING
18203 			crtsc = get_cyclecount();
18204 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18205 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18206 			}
18207 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18208 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18209 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18210 			}
18211 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18212 			sched_unpin();
18213 #endif
18214 			return (EHOSTUNREACH);
18215 		}
18216 		hdrlen += sizeof(struct udphdr);
18217 	}
18218 #ifdef INET6
18219 	if (isipv6)
18220 		ipoptlen = ip6_optlen(tp->t_inpcb);
18221 	else
18222 #endif
18223 		if (tp->t_inpcb->inp_options)
18224 			ipoptlen = tp->t_inpcb->inp_options->m_len -
18225 				offsetof(struct ipoption, ipopt_list);
18226 		else
18227 			ipoptlen = 0;
18228 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18229 	ipoptlen += ipsec_optlen;
18230 #endif
18231 
18232 	/*
18233 	 * Adjust data length if insertion of options will bump the packet
18234 	 * length beyond the t_maxseg length. Clear the FIN bit because we
18235 	 * cut off the tail of the segment.
18236 	 */
18237 	if (len + optlen + ipoptlen > tp->t_maxseg) {
18238 		if (tso) {
18239 			uint32_t if_hw_tsomax;
18240 			uint32_t moff;
18241 			int32_t max_len;
18242 
18243 			/* extract TSO information */
18244 			if_hw_tsomax = tp->t_tsomax;
18245 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18246 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18247 			KASSERT(ipoptlen == 0,
18248 				("%s: TSO can't do IP options", __func__));
18249 
18250 			/*
18251 			 * Check if we should limit by maximum payload
18252 			 * length:
18253 			 */
18254 			if (if_hw_tsomax != 0) {
18255 				/* compute maximum TSO length */
18256 				max_len = (if_hw_tsomax - hdrlen -
18257 					   max_linkhdr);
18258 				if (max_len <= 0) {
18259 					len = 0;
18260 				} else if (len > max_len) {
18261 					sendalot = 1;
18262 					len = max_len;
18263 					mark = 2;
18264 				}
18265 			}
18266 			/*
18267 			 * Prevent the last segment from being fractional
18268 			 * unless the send sockbuf can be emptied:
18269 			 */
18270 			max_len = (tp->t_maxseg - optlen);
18271 			if ((sb_offset + len) < sbavail(sb)) {
18272 				moff = len % (u_int)max_len;
18273 				if (moff != 0) {
18274 					mark = 3;
18275 					len -= moff;
18276 				}
18277 			}
18278 			/*
18279 			 * In case there are too many small fragments don't
18280 			 * use TSO:
18281 			 */
18282 			if (len <= segsiz) {
18283 				mark = 4;
18284 				tso = 0;
18285 			}
18286 			/*
18287 			 * Send the FIN in a separate segment after the bulk
18288 			 * sending is done. We don't trust the TSO
18289 			 * implementations to clear the FIN flag on all but
18290 			 * the last segment.
18291 			 */
18292 			if (tp->t_flags & TF_NEEDFIN) {
18293 				sendalot = 4;
18294 			}
18295 		} else {
18296 			mark = 5;
18297 			if (optlen + ipoptlen >= tp->t_maxseg) {
18298 				/*
18299 				 * Since we don't have enough space to put
18300 				 * the IP header chain and the TCP header in
18301 				 * one packet as required by RFC 7112, don't
18302 				 * send it. Also ensure that at least one
18303 				 * byte of the payload can be put into the
18304 				 * TCP segment.
18305 				 */
18306 				SOCKBUF_UNLOCK(&so->so_snd);
18307 				error = EMSGSIZE;
18308 				sack_rxmit = 0;
18309 				goto out;
18310 			}
18311 			len = tp->t_maxseg - optlen - ipoptlen;
18312 			sendalot = 5;
18313 		}
18314 	} else {
18315 		tso = 0;
18316 		mark = 6;
18317 	}
18318 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
18319 		("%s: len > IP_MAXPACKET", __func__));
18320 #ifdef DIAGNOSTIC
18321 #ifdef INET6
18322 	if (max_linkhdr + hdrlen > MCLBYTES)
18323 #else
18324 		if (max_linkhdr + hdrlen > MHLEN)
18325 #endif
18326 			panic("tcphdr too big");
18327 #endif
18328 
18329 	/*
18330 	 * This KASSERT is here to catch edge cases at a well defined place.
18331 	 * Before, those had triggered (random) panic conditions further
18332 	 * down.
18333 	 */
18334 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
18335 	if ((len == 0) &&
18336 	    (flags & TH_FIN) &&
18337 	    (sbused(sb))) {
18338 		/*
18339 		 * We have outstanding data, don't send a fin by itself!.
18340 		 */
18341 		goto just_return;
18342 	}
18343 	/*
18344 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
18345 	 * and initialize the header from the template for sends on this
18346 	 * connection.
18347 	 */
18348 	hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
18349 	if (len) {
18350 		uint32_t max_val;
18351 		uint32_t moff;
18352 
18353 		if (rack->r_ctl.rc_pace_max_segs)
18354 			max_val = rack->r_ctl.rc_pace_max_segs;
18355 		else if (rack->rc_user_set_max_segs)
18356 			max_val = rack->rc_user_set_max_segs * segsiz;
18357 		else
18358 			max_val = len;
18359 		/*
18360 		 * We allow a limit on sending with hptsi.
18361 		 */
18362 		if (len > max_val) {
18363 			mark = 7;
18364 			len = max_val;
18365 		}
18366 #ifdef INET6
18367 		if (MHLEN < hdrlen + max_linkhdr)
18368 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18369 		else
18370 #endif
18371 			m = m_gethdr(M_NOWAIT, MT_DATA);
18372 
18373 		if (m == NULL) {
18374 			SOCKBUF_UNLOCK(sb);
18375 			error = ENOBUFS;
18376 			sack_rxmit = 0;
18377 			goto out;
18378 		}
18379 		m->m_data += max_linkhdr;
18380 		m->m_len = hdrlen;
18381 
18382 		/*
18383 		 * Start the m_copy functions from the closest mbuf to the
18384 		 * sb_offset in the socket buffer chain.
18385 		 */
18386 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
18387 		s_mb = mb;
18388 		s_moff = moff;
18389 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
18390 			m_copydata(mb, moff, (int)len,
18391 				   mtod(m, caddr_t)+hdrlen);
18392 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18393 				sbsndptr_adv(sb, mb, len);
18394 			m->m_len += len;
18395 		} else {
18396 			struct sockbuf *msb;
18397 
18398 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18399 				msb = NULL;
18400 			else
18401 				msb = sb;
18402 			m->m_next = tcp_m_copym(
18403 				mb, moff, &len,
18404 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
18405 				((rsm == NULL) ? hw_tls : 0)
18406 #ifdef NETFLIX_COPY_ARGS
18407 				, &filled_all
18408 #endif
18409 				);
18410 			if (len <= (tp->t_maxseg - optlen)) {
18411 				/*
18412 				 * Must have ran out of mbufs for the copy
18413 				 * shorten it to no longer need tso. Lets
18414 				 * not put on sendalot since we are low on
18415 				 * mbufs.
18416 				 */
18417 				tso = 0;
18418 			}
18419 			if (m->m_next == NULL) {
18420 				SOCKBUF_UNLOCK(sb);
18421 				(void)m_free(m);
18422 				error = ENOBUFS;
18423 				sack_rxmit = 0;
18424 				goto out;
18425 			}
18426 		}
18427 		if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
18428 			if (rsm && (rsm->r_flags & RACK_TLP)) {
18429 				/*
18430 				 * TLP should not count in retran count, but
18431 				 * in its own bin
18432 				 */
18433 				counter_u64_add(rack_tlp_retran, 1);
18434 				counter_u64_add(rack_tlp_retran_bytes, len);
18435 			} else {
18436 				tp->t_sndrexmitpack++;
18437 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18438 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18439 			}
18440 #ifdef STATS
18441 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18442 						 len);
18443 #endif
18444 		} else {
18445 			KMOD_TCPSTAT_INC(tcps_sndpack);
18446 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
18447 #ifdef STATS
18448 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
18449 						 len);
18450 #endif
18451 		}
18452 		/*
18453 		 * If we're sending everything we've got, set PUSH. (This
18454 		 * will keep happy those implementations which only give
18455 		 * data to the user when a buffer fills or a PUSH comes in.)
18456 		 */
18457 		if (sb_offset + len == sbused(sb) &&
18458 		    sbused(sb) &&
18459 		    !(flags & TH_SYN)) {
18460 			flags |= TH_PUSH;
18461 			add_flag |= RACK_HAD_PUSH;
18462 		}
18463 
18464 		SOCKBUF_UNLOCK(sb);
18465 	} else {
18466 		SOCKBUF_UNLOCK(sb);
18467 		if (tp->t_flags & TF_ACKNOW)
18468 			KMOD_TCPSTAT_INC(tcps_sndacks);
18469 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
18470 			KMOD_TCPSTAT_INC(tcps_sndctrl);
18471 		else
18472 			KMOD_TCPSTAT_INC(tcps_sndwinup);
18473 
18474 		m = m_gethdr(M_NOWAIT, MT_DATA);
18475 		if (m == NULL) {
18476 			error = ENOBUFS;
18477 			sack_rxmit = 0;
18478 			goto out;
18479 		}
18480 #ifdef INET6
18481 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
18482 		    MHLEN >= hdrlen) {
18483 			M_ALIGN(m, hdrlen);
18484 		} else
18485 #endif
18486 			m->m_data += max_linkhdr;
18487 		m->m_len = hdrlen;
18488 	}
18489 	SOCKBUF_UNLOCK_ASSERT(sb);
18490 	m->m_pkthdr.rcvif = (struct ifnet *)0;
18491 #ifdef MAC
18492 	mac_inpcb_create_mbuf(inp, m);
18493 #endif
18494 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
18495 #ifdef INET6
18496 		if (isipv6)
18497 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18498 		else
18499 #endif				/* INET6 */
18500 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18501 		th = rack->r_ctl.fsb.th;
18502 		udp = rack->r_ctl.fsb.udp;
18503 		if (udp) {
18504 #ifdef INET6
18505 			if (isipv6)
18506 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
18507 			else
18508 #endif				/* INET6 */
18509 				ulen = hdrlen + len - sizeof(struct ip);
18510 			udp->uh_ulen = htons(ulen);
18511 		}
18512 	} else {
18513 #ifdef INET6
18514 		if (isipv6) {
18515 			ip6 = mtod(m, struct ip6_hdr *);
18516 			if (tp->t_port) {
18517 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
18518 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18519 				udp->uh_dport = tp->t_port;
18520 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
18521 				udp->uh_ulen = htons(ulen);
18522 				th = (struct tcphdr *)(udp + 1);
18523 			} else
18524 				th = (struct tcphdr *)(ip6 + 1);
18525 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
18526 		} else
18527 #endif				/* INET6 */
18528 		{
18529 			ip = mtod(m, struct ip *);
18530 #ifdef TCPDEBUG
18531 			ipov = (struct ipovly *)ip;
18532 #endif
18533 			if (tp->t_port) {
18534 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
18535 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18536 				udp->uh_dport = tp->t_port;
18537 				ulen = hdrlen + len - sizeof(struct ip);
18538 				udp->uh_ulen = htons(ulen);
18539 				th = (struct tcphdr *)(udp + 1);
18540 			} else
18541 				th = (struct tcphdr *)(ip + 1);
18542 			tcpip_fillheaders(inp, tp->t_port, ip, th);
18543 		}
18544 	}
18545 	/*
18546 	 * Fill in fields, remembering maximum advertised window for use in
18547 	 * delaying messages about window sizes. If resending a FIN, be sure
18548 	 * not to use a new sequence number.
18549 	 */
18550 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
18551 	    tp->snd_nxt == tp->snd_max)
18552 		tp->snd_nxt--;
18553 	/*
18554 	 * If we are starting a connection, send ECN setup SYN packet. If we
18555 	 * are on a retransmit, we may resend those bits a number of times
18556 	 * as per RFC 3168.
18557 	 */
18558 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
18559 		if (tp->t_rxtshift >= 1) {
18560 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
18561 				flags |= TH_ECE | TH_CWR;
18562 		} else
18563 			flags |= TH_ECE | TH_CWR;
18564 	}
18565 	/* Handle parallel SYN for ECN */
18566 	if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18567 	    (tp->t_flags2 & TF2_ECN_SND_ECE)) {
18568 		flags |= TH_ECE;
18569 		tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18570 	}
18571 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18572 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
18573 		/*
18574 		 * If the peer has ECN, mark data packets with ECN capable
18575 		 * transmission (ECT). Ignore pure ack packets,
18576 		 * retransmissions.
18577 		 */
18578 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
18579 		    (sack_rxmit == 0)) {
18580 #ifdef INET6
18581 			if (isipv6)
18582 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
18583 			else
18584 #endif
18585 				ip->ip_tos |= IPTOS_ECN_ECT0;
18586 			KMOD_TCPSTAT_INC(tcps_ecn_ect0);
18587 			/*
18588 			 * Reply with proper ECN notifications.
18589 			 * Only set CWR on new data segments.
18590 			 */
18591 			if (tp->t_flags2 & TF2_ECN_SND_CWR) {
18592 				flags |= TH_CWR;
18593 				tp->t_flags2 &= ~TF2_ECN_SND_CWR;
18594 			}
18595 		}
18596 		if (tp->t_flags2 & TF2_ECN_SND_ECE)
18597 			flags |= TH_ECE;
18598 	}
18599 	/*
18600 	 * If we are doing retransmissions, then snd_nxt will not reflect
18601 	 * the first unsent octet.  For ACK only packets, we do not want the
18602 	 * sequence number of the retransmitted packet, we want the sequence
18603 	 * number of the next unsent octet.  So, if there is no data (and no
18604 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
18605 	 * ti_seq.  But if we are in persist state, snd_max might reflect
18606 	 * one byte beyond the right edge of the window, so use snd_nxt in
18607 	 * that case, since we know we aren't doing a retransmission.
18608 	 * (retransmit and persist are mutually exclusive...)
18609 	 */
18610 	if (sack_rxmit == 0) {
18611 		if (len || (flags & (TH_SYN | TH_FIN))) {
18612 			th->th_seq = htonl(tp->snd_nxt);
18613 			rack_seq = tp->snd_nxt;
18614 		} else {
18615 			th->th_seq = htonl(tp->snd_max);
18616 			rack_seq = tp->snd_max;
18617 		}
18618 	} else {
18619 		th->th_seq = htonl(rsm->r_start);
18620 		rack_seq = rsm->r_start;
18621 	}
18622 	th->th_ack = htonl(tp->rcv_nxt);
18623 	th->th_flags = flags;
18624 	/*
18625 	 * Calculate receive window.  Don't shrink window, but avoid silly
18626 	 * window syndrome.
18627 	 * If a RST segment is sent, advertise a window of zero.
18628 	 */
18629 	if (flags & TH_RST) {
18630 		recwin = 0;
18631 	} else {
18632 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
18633 		    recwin < (long)segsiz) {
18634 			recwin = 0;
18635 		}
18636 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
18637 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
18638 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
18639 	}
18640 
18641 	/*
18642 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
18643 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
18644 	 * handled in syncache.
18645 	 */
18646 	if (flags & TH_SYN)
18647 		th->th_win = htons((u_short)
18648 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
18649 	else {
18650 		/* Avoid shrinking window with window scaling. */
18651 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
18652 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
18653 	}
18654 	/*
18655 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
18656 	 * window.  This may cause the remote transmitter to stall.  This
18657 	 * flag tells soreceive() to disable delayed acknowledgements when
18658 	 * draining the buffer.  This can occur if the receiver is
18659 	 * attempting to read more data than can be buffered prior to
18660 	 * transmitting on the connection.
18661 	 */
18662 	if (th->th_win == 0) {
18663 		tp->t_sndzerowin++;
18664 		tp->t_flags |= TF_RXWIN0SENT;
18665 	} else
18666 		tp->t_flags &= ~TF_RXWIN0SENT;
18667 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
18668 	/* Now are we using fsb?, if so copy the template data to the mbuf */
18669 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
18670 		uint8_t *cpto;
18671 
18672 		cpto = mtod(m, uint8_t *);
18673 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18674 		/*
18675 		 * We have just copied in:
18676 		 * IP/IP6
18677 		 * <optional udphdr>
18678 		 * tcphdr (no options)
18679 		 *
18680 		 * We need to grab the correct pointers into the mbuf
18681 		 * for both the tcp header, and possibly the udp header (if tunneling).
18682 		 * We do this by using the offset in the copy buffer and adding it
18683 		 * to the mbuf base pointer (cpto).
18684 		 */
18685 #ifdef INET6
18686 		if (isipv6)
18687 			ip6 = mtod(m, struct ip6_hdr *);
18688 		else
18689 #endif				/* INET6 */
18690 			ip = mtod(m, struct ip *);
18691 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18692 		/* If we have a udp header lets set it into the mbuf as well */
18693 		if (udp)
18694 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
18695 	}
18696 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18697 	if (to.to_flags & TOF_SIGNATURE) {
18698 		/*
18699 		 * Calculate MD5 signature and put it into the place
18700 		 * determined before.
18701 		 * NOTE: since TCP options buffer doesn't point into
18702 		 * mbuf's data, calculate offset and use it.
18703 		 */
18704 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18705 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18706 			/*
18707 			 * Do not send segment if the calculation of MD5
18708 			 * digest has failed.
18709 			 */
18710 			goto out;
18711 		}
18712 	}
18713 #endif
18714 	if (optlen) {
18715 		bcopy(opt, th + 1, optlen);
18716 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18717 	}
18718 	/*
18719 	 * Put TCP length in extended header, and then checksum extended
18720 	 * header and data.
18721 	 */
18722 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
18723 #ifdef INET6
18724 	if (isipv6) {
18725 		/*
18726 		 * ip6_plen is not need to be filled now, and will be filled
18727 		 * in ip6_output.
18728 		 */
18729 		if (tp->t_port) {
18730 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18731 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18732 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18733 			th->th_sum = htons(0);
18734 			UDPSTAT_INC(udps_opackets);
18735 		} else {
18736 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18737 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18738 			th->th_sum = in6_cksum_pseudo(ip6,
18739 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18740 						      0);
18741 		}
18742 	}
18743 #endif
18744 #if defined(INET6) && defined(INET)
18745 	else
18746 #endif
18747 #ifdef INET
18748 	{
18749 		if (tp->t_port) {
18750 			m->m_pkthdr.csum_flags = CSUM_UDP;
18751 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18752 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18753 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18754 			th->th_sum = htons(0);
18755 			UDPSTAT_INC(udps_opackets);
18756 		} else {
18757 			m->m_pkthdr.csum_flags = CSUM_TCP;
18758 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18759 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
18760 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18761 									IPPROTO_TCP + len + optlen));
18762 		}
18763 		/* IP version must be set here for ipv4/ipv6 checking later */
18764 		KASSERT(ip->ip_v == IPVERSION,
18765 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
18766 	}
18767 #endif
18768 	/*
18769 	 * Enable TSO and specify the size of the segments. The TCP pseudo
18770 	 * header checksum is always provided. XXX: Fixme: This is currently
18771 	 * not the case for IPv6.
18772 	 */
18773 	if (tso) {
18774 		KASSERT(len > tp->t_maxseg - optlen,
18775 			("%s: len <= tso_segsz", __func__));
18776 		m->m_pkthdr.csum_flags |= CSUM_TSO;
18777 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
18778 	}
18779 	KASSERT(len + hdrlen == m_length(m, NULL),
18780 		("%s: mbuf chain different than expected: %d + %u != %u",
18781 		 __func__, len, hdrlen, m_length(m, NULL)));
18782 
18783 #ifdef TCP_HHOOK
18784 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
18785 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
18786 #endif
18787 	/* We're getting ready to send; log now. */
18788 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
18789 		union tcp_log_stackspecific log;
18790 
18791 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18792 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
18793 		if (rack->rack_no_prr)
18794 			log.u_bbr.flex1 = 0;
18795 		else
18796 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18797 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18798 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18799 		log.u_bbr.flex4 = orig_len;
18800 		if (filled_all)
18801 			log.u_bbr.flex5 = 0x80000000;
18802 		else
18803 			log.u_bbr.flex5 = 0;
18804 		/* Save off the early/late values */
18805 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18806 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18807 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18808 		if (rsm || sack_rxmit) {
18809 			if (doing_tlp)
18810 				log.u_bbr.flex8 = 2;
18811 			else
18812 				log.u_bbr.flex8 = 1;
18813 		} else {
18814 			if (doing_tlp)
18815 				log.u_bbr.flex8 = 3;
18816 			else
18817 				log.u_bbr.flex8 = 0;
18818 		}
18819 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
18820 		log.u_bbr.flex7 = mark;
18821 		log.u_bbr.flex7 <<= 8;
18822 		log.u_bbr.flex7 |= pass;
18823 		log.u_bbr.pkts_out = tp->t_maxseg;
18824 		log.u_bbr.timeStamp = cts;
18825 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18826 		log.u_bbr.lt_epoch = cwnd_to_use;
18827 		log.u_bbr.delivered = sendalot;
18828 		lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
18829 				     len, &log, false, NULL, NULL, 0, &tv);
18830 	} else
18831 		lgb = NULL;
18832 
18833 	/*
18834 	 * Fill in IP length and desired time to live and send to IP level.
18835 	 * There should be a better way to handle ttl and tos; we could keep
18836 	 * them in the template, but need a way to checksum without them.
18837 	 */
18838 	/*
18839 	 * m->m_pkthdr.len should have been set before cksum calcuration,
18840 	 * because in6_cksum() need it.
18841 	 */
18842 #ifdef INET6
18843 	if (isipv6) {
18844 		/*
18845 		 * we separately set hoplimit for every segment, since the
18846 		 * user might want to change the value via setsockopt. Also,
18847 		 * desired default hop limit might be changed via Neighbor
18848 		 * Discovery.
18849 		 */
18850 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18851 
18852 		/*
18853 		 * Set the packet size here for the benefit of DTrace
18854 		 * probes. ip6_output() will set it properly; it's supposed
18855 		 * to include the option header lengths as well.
18856 		 */
18857 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18858 
18859 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18860 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18861 		else
18862 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18863 
18864 		if (tp->t_state == TCPS_SYN_SENT)
18865 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18866 
18867 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18868 		/* TODO: IPv6 IP6TOS_ECT bit on */
18869 		error = ip6_output(m,
18870 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18871 				   inp->in6p_outputopts,
18872 #else
18873 				   NULL,
18874 #endif
18875 				   &inp->inp_route6,
18876 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18877 				   NULL, NULL, inp);
18878 
18879 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18880 			mtu = inp->inp_route6.ro_nh->nh_mtu;
18881 	}
18882 #endif				/* INET6 */
18883 #if defined(INET) && defined(INET6)
18884 	else
18885 #endif
18886 #ifdef INET
18887 	{
18888 		ip->ip_len = htons(m->m_pkthdr.len);
18889 #ifdef INET6
18890 		if (inp->inp_vflag & INP_IPV6PROTO)
18891 			ip->ip_ttl = in6_selecthlim(inp, NULL);
18892 #endif				/* INET6 */
18893 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18894 		/*
18895 		 * If we do path MTU discovery, then we set DF on every
18896 		 * packet. This might not be the best thing to do according
18897 		 * to RFC3390 Section 2. However the tcp hostcache migitates
18898 		 * the problem so it affects only the first tcp connection
18899 		 * with a host.
18900 		 *
18901 		 * NB: Don't set DF on small MTU/MSS to have a safe
18902 		 * fallback.
18903 		 */
18904 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18905 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18906 			if (tp->t_port == 0 || len < V_tcp_minmss) {
18907 				ip->ip_off |= htons(IP_DF);
18908 			}
18909 		} else {
18910 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18911 		}
18912 
18913 		if (tp->t_state == TCPS_SYN_SENT)
18914 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18915 
18916 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
18917 
18918 		error = ip_output(m,
18919 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18920 				  inp->inp_options,
18921 #else
18922 				  NULL,
18923 #endif
18924 				  &inp->inp_route,
18925 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18926 				  inp);
18927 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18928 			mtu = inp->inp_route.ro_nh->nh_mtu;
18929 	}
18930 #endif				/* INET */
18931 
18932 out:
18933 	if (lgb) {
18934 		lgb->tlb_errno = error;
18935 		lgb = NULL;
18936 	}
18937 	/*
18938 	 * In transmit state, time the transmission and arrange for the
18939 	 * retransmit.  In persist state, just set snd_max.
18940 	 */
18941 	if (error == 0) {
18942 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
18943 		if (rsm && doing_tlp) {
18944 			rack->rc_last_sent_tlp_past_cumack = 0;
18945 			rack->rc_last_sent_tlp_seq_valid = 1;
18946 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18947 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18948 		}
18949 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
18950 		if (rsm && (doing_tlp == 0)) {
18951 			/* Set we retransmitted */
18952 			rack->rc_gp_saw_rec = 1;
18953 		} else {
18954 			if (cwnd_to_use > tp->snd_ssthresh) {
18955 				/* Set we sent in CA */
18956 				rack->rc_gp_saw_ca = 1;
18957 			} else {
18958 				/* Set we sent in SS */
18959 				rack->rc_gp_saw_ss = 1;
18960 			}
18961 		}
18962 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18963 		    (tp->t_flags & TF_SACK_PERMIT) &&
18964 		    tp->rcv_numsacks > 0)
18965 			tcp_clean_dsack_blocks(tp);
18966 		tot_len_this_send += len;
18967 		if (len == 0)
18968 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18969 		else if (len == 1) {
18970 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18971 		} else if (len > 1) {
18972 			int idx;
18973 
18974 			idx = (len / segsiz) + 3;
18975 			if (idx >= TCP_MSS_ACCT_ATIMER)
18976 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18977 			else
18978 				counter_u64_add(rack_out_size[idx], 1);
18979 		}
18980 	}
18981 	if ((rack->rack_no_prr == 0) &&
18982 	    sub_from_prr &&
18983 	    (error == 0)) {
18984 		if (rack->r_ctl.rc_prr_sndcnt >= len)
18985 			rack->r_ctl.rc_prr_sndcnt -= len;
18986 		else
18987 			rack->r_ctl.rc_prr_sndcnt = 0;
18988 	}
18989 	sub_from_prr = 0;
18990 	if (doing_tlp) {
18991 		/* Make sure the TLP is added */
18992 		add_flag |= RACK_TLP;
18993 	} else if (rsm) {
18994 		/* If its a resend without TLP then it must not have the flag */
18995 		rsm->r_flags &= ~RACK_TLP;
18996 	}
18997 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18998 			rack_to_usec_ts(&tv),
18999 			rsm, add_flag, s_mb, s_moff, hw_tls);
19000 
19001 
19002 	if ((error == 0) &&
19003 	    (len > 0) &&
19004 	    (tp->snd_una == tp->snd_max))
19005 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
19006 	{
19007 		tcp_seq startseq = tp->snd_nxt;
19008 
19009 		/* Track our lost count */
19010 		if (rsm && (doing_tlp == 0))
19011 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
19012 		/*
19013 		 * Advance snd_nxt over sequence space of this segment.
19014 		 */
19015 		if (error)
19016 			/* We don't log or do anything with errors */
19017 			goto nomore;
19018 		if (doing_tlp == 0) {
19019 			if (rsm == NULL) {
19020 				/*
19021 				 * Not a retransmission of some
19022 				 * sort, new data is going out so
19023 				 * clear our TLP count and flag.
19024 				 */
19025 				rack->rc_tlp_in_progress = 0;
19026 				rack->r_ctl.rc_tlp_cnt_out = 0;
19027 			}
19028 		} else {
19029 			/*
19030 			 * We have just sent a TLP, mark that it is true
19031 			 * and make sure our in progress is set so we
19032 			 * continue to check the count.
19033 			 */
19034 			rack->rc_tlp_in_progress = 1;
19035 			rack->r_ctl.rc_tlp_cnt_out++;
19036 		}
19037 		if (flags & (TH_SYN | TH_FIN)) {
19038 			if (flags & TH_SYN)
19039 				tp->snd_nxt++;
19040 			if (flags & TH_FIN) {
19041 				tp->snd_nxt++;
19042 				tp->t_flags |= TF_SENTFIN;
19043 			}
19044 		}
19045 		/* In the ENOBUFS case we do *not* update snd_max */
19046 		if (sack_rxmit)
19047 			goto nomore;
19048 
19049 		tp->snd_nxt += len;
19050 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
19051 			if (tp->snd_una == tp->snd_max) {
19052 				/*
19053 				 * Update the time we just added data since
19054 				 * none was outstanding.
19055 				 */
19056 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
19057 				tp->t_acktime = ticks;
19058 			}
19059 			tp->snd_max = tp->snd_nxt;
19060 			/*
19061 			 * Time this transmission if not a retransmission and
19062 			 * not currently timing anything.
19063 			 * This is only relevant in case of switching back to
19064 			 * the base stack.
19065 			 */
19066 			if (tp->t_rtttime == 0) {
19067 				tp->t_rtttime = ticks;
19068 				tp->t_rtseq = startseq;
19069 				KMOD_TCPSTAT_INC(tcps_segstimed);
19070 			}
19071 			if (len &&
19072 			    ((tp->t_flags & TF_GPUTINPROG) == 0))
19073 				rack_start_gp_measurement(tp, rack, startseq, sb_offset);
19074 		}
19075 		/*
19076 		 * If we are doing FO we need to update the mbuf position and subtract
19077 		 * this happens when the peer sends us duplicate information and
19078 		 * we thus want to send a DSACK.
19079 		 *
19080 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
19081 		 * turned off? If not then we are going to echo multiple DSACK blocks
19082 		 * out (with the TSO), which we should not be doing.
19083 		 */
19084 		if (rack->r_fast_output && len) {
19085 			if (rack->r_ctl.fsb.left_to_send > len)
19086 				rack->r_ctl.fsb.left_to_send -= len;
19087 			else
19088 				rack->r_ctl.fsb.left_to_send = 0;
19089 			if (rack->r_ctl.fsb.left_to_send < segsiz)
19090 				rack->r_fast_output = 0;
19091 			if (rack->r_fast_output) {
19092 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19093 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19094 			}
19095 		}
19096 	}
19097 nomore:
19098 	if (error) {
19099 		rack->r_ctl.rc_agg_delayed = 0;
19100 		rack->r_early = 0;
19101 		rack->r_late = 0;
19102 		rack->r_ctl.rc_agg_early = 0;
19103 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
19104 		/*
19105 		 * Failures do not advance the seq counter above. For the
19106 		 * case of ENOBUFS we will fall out and retry in 1ms with
19107 		 * the hpts. Everything else will just have to retransmit
19108 		 * with the timer.
19109 		 *
19110 		 * In any case, we do not want to loop around for another
19111 		 * send without a good reason.
19112 		 */
19113 		sendalot = 0;
19114 		switch (error) {
19115 		case EPERM:
19116 			tp->t_softerror = error;
19117 #ifdef TCP_ACCOUNTING
19118 			crtsc = get_cyclecount();
19119 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19120 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19121 			}
19122 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19123 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19124 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19125 			}
19126 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19127 			sched_unpin();
19128 #endif
19129 			return (error);
19130 		case ENOBUFS:
19131 			/*
19132 			 * Pace us right away to retry in a some
19133 			 * time
19134 			 */
19135 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
19136 			if (rack->rc_enobuf < 0x7f)
19137 				rack->rc_enobuf++;
19138 			if (slot < (10 * HPTS_USEC_IN_MSEC))
19139 				slot = 10 * HPTS_USEC_IN_MSEC;
19140 			if (rack->r_ctl.crte != NULL) {
19141 				counter_u64_add(rack_saw_enobuf_hw, 1);
19142 				tcp_rl_log_enobuf(rack->r_ctl.crte);
19143 			}
19144 			counter_u64_add(rack_saw_enobuf, 1);
19145 			goto enobufs;
19146 		case EMSGSIZE:
19147 			/*
19148 			 * For some reason the interface we used initially
19149 			 * to send segments changed to another or lowered
19150 			 * its MTU. If TSO was active we either got an
19151 			 * interface without TSO capabilits or TSO was
19152 			 * turned off. If we obtained mtu from ip_output()
19153 			 * then update it and try again.
19154 			 */
19155 			if (tso)
19156 				tp->t_flags &= ~TF_TSO;
19157 			if (mtu != 0) {
19158 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
19159 				goto again;
19160 			}
19161 			slot = 10 * HPTS_USEC_IN_MSEC;
19162 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19163 #ifdef TCP_ACCOUNTING
19164 			crtsc = get_cyclecount();
19165 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19166 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19167 			}
19168 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19169 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19170 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19171 			}
19172 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19173 			sched_unpin();
19174 #endif
19175 			return (error);
19176 		case ENETUNREACH:
19177 			counter_u64_add(rack_saw_enetunreach, 1);
19178 		case EHOSTDOWN:
19179 		case EHOSTUNREACH:
19180 		case ENETDOWN:
19181 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
19182 				tp->t_softerror = error;
19183 			}
19184 			/* FALLTHROUGH */
19185 		default:
19186 			slot = 10 * HPTS_USEC_IN_MSEC;
19187 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19188 #ifdef TCP_ACCOUNTING
19189 			crtsc = get_cyclecount();
19190 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19191 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19192 			}
19193 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19194 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19195 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19196 			}
19197 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19198 			sched_unpin();
19199 #endif
19200 			return (error);
19201 		}
19202 	} else {
19203 		rack->rc_enobuf = 0;
19204 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
19205 			rack->r_ctl.retran_during_recovery += len;
19206 	}
19207 	KMOD_TCPSTAT_INC(tcps_sndtotal);
19208 
19209 	/*
19210 	 * Data sent (as far as we can tell). If this advertises a larger
19211 	 * window than any other segment, then remember the size of the
19212 	 * advertised window. Any pending ACK has now been sent.
19213 	 */
19214 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
19215 		tp->rcv_adv = tp->rcv_nxt + recwin;
19216 
19217 	tp->last_ack_sent = tp->rcv_nxt;
19218 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19219 enobufs:
19220 	if (sendalot) {
19221 		/* Do we need to turn off sendalot? */
19222 		if (rack->r_ctl.rc_pace_max_segs &&
19223 		    (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
19224 			/* We hit our max. */
19225 			sendalot = 0;
19226 		} else if ((rack->rc_user_set_max_segs) &&
19227 			   (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
19228 			/* We hit the user defined max */
19229 			sendalot = 0;
19230 		}
19231 	}
19232 	if ((error == 0) && (flags & TH_FIN))
19233 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
19234 	if (flags & TH_RST) {
19235 		/*
19236 		 * We don't send again after sending a RST.
19237 		 */
19238 		slot = 0;
19239 		sendalot = 0;
19240 		if (error == 0)
19241 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
19242 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
19243 		/*
19244 		 * Get our pacing rate, if an error
19245 		 * occurred in sending (ENOBUF) we would
19246 		 * hit the else if with slot preset. Other
19247 		 * errors return.
19248 		 */
19249 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
19250 	}
19251 	if (rsm &&
19252 	    (rsm->r_flags & RACK_HAS_SYN) == 0 &&
19253 	    rack->use_rack_rr) {
19254 		/* Its a retransmit and we use the rack cheat? */
19255 		if ((slot == 0) ||
19256 		    (rack->rc_always_pace == 0) ||
19257 		    (rack->r_rr_config == 1)) {
19258 			/*
19259 			 * We have no pacing set or we
19260 			 * are using old-style rack or
19261 			 * we are overriden to use the old 1ms pacing.
19262 			 */
19263 			slot = rack->r_ctl.rc_min_to;
19264 		}
19265 	}
19266 	/* We have sent clear the flag */
19267 	rack->r_ent_rec_ns = 0;
19268 	if (rack->r_must_retran) {
19269 		if (rsm) {
19270 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
19271 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
19272 				/*
19273 				 * We have retransmitted all.
19274 				 */
19275 				rack->r_must_retran = 0;
19276 				rack->r_ctl.rc_out_at_rto = 0;
19277 			}
19278 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19279 			/*
19280 			 * Sending new data will also kill
19281 			 * the loop.
19282 			 */
19283 			rack->r_must_retran = 0;
19284 			rack->r_ctl.rc_out_at_rto = 0;
19285 		}
19286 	}
19287 	rack->r_ctl.fsb.recwin = recwin;
19288 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
19289 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19290 		/*
19291 		 * We hit an RTO and now have past snd_max at the RTO
19292 		 * clear all the WAS flags.
19293 		 */
19294 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
19295 	}
19296 	if (slot) {
19297 		/* set the rack tcb into the slot N */
19298 		counter_u64_add(rack_paced_segments, 1);
19299 		if ((error == 0) &&
19300 		    rack_use_rfo &&
19301 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
19302 		    (rsm == NULL) &&
19303 		    (tp->snd_nxt == tp->snd_max) &&
19304 		    (ipoptlen == 0) &&
19305 		    (tp->rcv_numsacks == 0) &&
19306 		    rack->r_fsb_inited &&
19307 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
19308 		    (rack->r_must_retran == 0) &&
19309 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
19310 		    (len > 0) && (orig_len > 0) &&
19311 		    (orig_len > len) &&
19312 		    ((orig_len - len) >= segsiz) &&
19313 		    ((optlen == 0) ||
19314 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19315 			/* We can send at least one more MSS using our fsb */
19316 
19317 			rack->r_fast_output = 1;
19318 			rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19319 			rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19320 			rack->r_ctl.fsb.tcp_flags = flags;
19321 			rack->r_ctl.fsb.left_to_send = orig_len - len;
19322 			if (hw_tls)
19323 				rack->r_ctl.fsb.hw_tls = 1;
19324 			else
19325 				rack->r_ctl.fsb.hw_tls = 0;
19326 			KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19327 				("rack:%p left_to_send:%u sbavail:%u out:%u",
19328 				 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19329 				 (tp->snd_max - tp->snd_una)));
19330 			if (rack->r_ctl.fsb.left_to_send < segsiz)
19331 				rack->r_fast_output = 0;
19332 			else {
19333 				if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19334 					rack->r_ctl.fsb.rfo_apply_push = 1;
19335 				else
19336 					rack->r_ctl.fsb.rfo_apply_push = 0;
19337 			}
19338 		} else
19339 			rack->r_fast_output = 0;
19340 		rack_log_fsb(rack, tp, so, flags,
19341 			     ipoptlen, orig_len, len, error,
19342 			     (rsm == NULL), optlen, __LINE__, 2);
19343 	} else if (sendalot) {
19344 		int ret;
19345 
19346 		if (len)
19347 			counter_u64_add(rack_unpaced_segments, 1);
19348 		sack_rxmit = 0;
19349 		if ((error == 0) &&
19350 		    rack_use_rfo &&
19351 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
19352 		    (rsm == NULL) &&
19353 		    (ipoptlen == 0) &&
19354 		    (tp->rcv_numsacks == 0) &&
19355 		    (tp->snd_nxt == tp->snd_max) &&
19356 		    (rack->r_must_retran == 0) &&
19357 		    rack->r_fsb_inited &&
19358 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
19359 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
19360 		    (len > 0) && (orig_len > 0) &&
19361 		    (orig_len > len) &&
19362 		    ((orig_len - len) >= segsiz) &&
19363 		    ((optlen == 0) ||
19364 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19365 			/* we can use fast_output for more */
19366 
19367 			rack->r_fast_output = 1;
19368 			rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19369 			rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19370 			rack->r_ctl.fsb.tcp_flags = flags;
19371 			rack->r_ctl.fsb.left_to_send = orig_len - len;
19372 			if (hw_tls)
19373 				rack->r_ctl.fsb.hw_tls = 1;
19374 			else
19375 				rack->r_ctl.fsb.hw_tls = 0;
19376 			KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19377 				("rack:%p left_to_send:%u sbavail:%u out:%u",
19378 				 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19379 				 (tp->snd_max - tp->snd_una)));
19380 			if (rack->r_ctl.fsb.left_to_send < segsiz) {
19381 				rack->r_fast_output = 0;
19382 			}
19383 			if (rack->r_fast_output) {
19384 				if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19385 					rack->r_ctl.fsb.rfo_apply_push = 1;
19386 				else
19387 					rack->r_ctl.fsb.rfo_apply_push = 0;
19388 				rack_log_fsb(rack, tp, so, flags,
19389 					     ipoptlen, orig_len, len, error,
19390 					     (rsm == NULL), optlen, __LINE__, 3);
19391 				error = 0;
19392 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19393 				if (ret >= 0)
19394 					return (ret);
19395 			        else if (error)
19396 					goto nomore;
19397 
19398 			}
19399 		}
19400 		goto again;
19401 	} else if (len) {
19402 		counter_u64_add(rack_unpaced_segments, 1);
19403 	}
19404 	/* Assure when we leave that snd_nxt will point to top */
19405 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
19406 		tp->snd_nxt = tp->snd_max;
19407 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
19408 #ifdef TCP_ACCOUNTING
19409 	crtsc = get_cyclecount() - ts_val;
19410 	if (tot_len_this_send) {
19411 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19412 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
19413 		}
19414 		counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
19415 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19416 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
19417 		}
19418 		counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
19419 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19420 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
19421 		}
19422 		counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
19423 	} else {
19424 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19425 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
19426 		}
19427 		counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
19428 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19429 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
19430 		}
19431 		counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
19432 	}
19433 	sched_unpin();
19434 #endif
19435 	if (error == ENOBUFS)
19436 		error = 0;
19437 	return (error);
19438 }
19439 
19440 static void
19441 rack_update_seg(struct tcp_rack *rack)
19442 {
19443 	uint32_t orig_val;
19444 
19445 	orig_val = rack->r_ctl.rc_pace_max_segs;
19446 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
19447 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
19448 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
19449 }
19450 
19451 static void
19452 rack_mtu_change(struct tcpcb *tp)
19453 {
19454 	/*
19455 	 * The MSS may have changed
19456 	 */
19457 	struct tcp_rack *rack;
19458 	struct rack_sendmap *rsm;
19459 
19460 	rack = (struct tcp_rack *)tp->t_fb_ptr;
19461 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
19462 		/*
19463 		 * The MTU has changed we need to resend everything
19464 		 * since all we have sent is lost. We first fix
19465 		 * up the mtu though.
19466 		 */
19467 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
19468 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
19469 		rack_remxt_tmr(tp);
19470 		rack->r_fast_output = 0;
19471 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
19472 						rack->r_ctl.rc_sacked);
19473 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
19474 		rack->r_must_retran = 1;
19475 		/* Mark all inflight to needing to be rxt'd */
19476 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
19477 			rsm->r_flags |= RACK_MUST_RXT;
19478 		}
19479 	}
19480 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
19481 	/* We don't use snd_nxt to retransmit */
19482 	tp->snd_nxt = tp->snd_max;
19483 }
19484 
19485 static int
19486 rack_set_profile(struct tcp_rack *rack, int prof)
19487 {
19488 	int err = EINVAL;
19489 	if (prof == 1) {
19490 		/* pace_always=1 */
19491 		if (rack->rc_always_pace == 0) {
19492 			if (tcp_can_enable_pacing() == 0)
19493 				return (EBUSY);
19494 		}
19495 		rack->rc_always_pace = 1;
19496 		if (rack->use_fixed_rate || rack->gp_ready)
19497 			rack_set_cc_pacing(rack);
19498 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19499 		rack->rack_attempt_hdwr_pace = 0;
19500 		/* cmpack=1 */
19501 		if (rack_use_cmp_acks)
19502 			rack->r_use_cmp_ack = 1;
19503 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19504 		    rack->r_use_cmp_ack)
19505 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19506 		/* scwnd=1 */
19507 		rack->rack_enable_scwnd = 1;
19508 		/* dynamic=100 */
19509 		rack->rc_gp_dyn_mul = 1;
19510 		/* gp_inc_ca */
19511 		rack->r_ctl.rack_per_of_gp_ca = 100;
19512 		/* rrr_conf=3 */
19513 		rack->r_rr_config = 3;
19514 		/* npush=2 */
19515 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19516 		/* fillcw=1 */
19517 		rack->rc_pace_to_cwnd = 1;
19518 		rack->rc_pace_fill_if_rttin_range = 0;
19519 		rack->rtt_limit_mul = 0;
19520 		/* noprr=1 */
19521 		rack->rack_no_prr = 1;
19522 		/* lscwnd=1 */
19523 		rack->r_limit_scw = 1;
19524 		/* gp_inc_rec */
19525 		rack->r_ctl.rack_per_of_gp_rec = 90;
19526 		err = 0;
19527 
19528 	} else if (prof == 3) {
19529 		/* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
19530 		/* pace_always=1 */
19531 		if (rack->rc_always_pace == 0) {
19532 			if (tcp_can_enable_pacing() == 0)
19533 				return (EBUSY);
19534 		}
19535 		rack->rc_always_pace = 1;
19536 		if (rack->use_fixed_rate || rack->gp_ready)
19537 			rack_set_cc_pacing(rack);
19538 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19539 		rack->rack_attempt_hdwr_pace = 0;
19540 		/* cmpack=1 */
19541 		if (rack_use_cmp_acks)
19542 			rack->r_use_cmp_ack = 1;
19543 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19544 		    rack->r_use_cmp_ack)
19545 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19546 		/* scwnd=1 */
19547 		rack->rack_enable_scwnd = 1;
19548 		/* dynamic=100 */
19549 		rack->rc_gp_dyn_mul = 1;
19550 		/* gp_inc_ca */
19551 		rack->r_ctl.rack_per_of_gp_ca = 100;
19552 		/* rrr_conf=3 */
19553 		rack->r_rr_config = 3;
19554 		/* npush=2 */
19555 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19556 		/* fillcw=2 */
19557 		rack->rc_pace_to_cwnd = 1;
19558 		rack->r_fill_less_agg = 1;
19559 		rack->rc_pace_fill_if_rttin_range = 0;
19560 		rack->rtt_limit_mul = 0;
19561 		/* noprr=1 */
19562 		rack->rack_no_prr = 1;
19563 		/* lscwnd=1 */
19564 		rack->r_limit_scw = 1;
19565 		/* gp_inc_rec */
19566 		rack->r_ctl.rack_per_of_gp_rec = 90;
19567 		err = 0;
19568 
19569 
19570 	} else if (prof == 2) {
19571 		/* cmpack=1 */
19572 		if (rack->rc_always_pace == 0) {
19573 			if (tcp_can_enable_pacing() == 0)
19574 				return (EBUSY);
19575 		}
19576 		rack->rc_always_pace = 1;
19577 		if (rack->use_fixed_rate || rack->gp_ready)
19578 			rack_set_cc_pacing(rack);
19579 		rack->r_use_cmp_ack = 1;
19580 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19581 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19582 		/* pace_always=1 */
19583 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19584 		/* scwnd=1 */
19585 		rack->rack_enable_scwnd = 1;
19586 		/* dynamic=100 */
19587 		rack->rc_gp_dyn_mul = 1;
19588 		rack->r_ctl.rack_per_of_gp_ca = 100;
19589 		/* rrr_conf=3 */
19590 		rack->r_rr_config = 3;
19591 		/* npush=2 */
19592 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19593 		/* fillcw=1 */
19594 		rack->rc_pace_to_cwnd = 1;
19595 		rack->rc_pace_fill_if_rttin_range = 0;
19596 		rack->rtt_limit_mul = 0;
19597 		/* noprr=1 */
19598 		rack->rack_no_prr = 1;
19599 		/* lscwnd=0 */
19600 		rack->r_limit_scw = 0;
19601 		err = 0;
19602 	} else if (prof == 0) {
19603 		/* This changes things back to the default settings */
19604 		err = 0;
19605 		if (rack->rc_always_pace) {
19606 			tcp_decrement_paced_conn();
19607 			rack_undo_cc_pacing(rack);
19608 			rack->rc_always_pace = 0;
19609 		}
19610 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
19611 			rack->rc_always_pace = 1;
19612 			if (rack->use_fixed_rate || rack->gp_ready)
19613 				rack_set_cc_pacing(rack);
19614 		} else
19615 			rack->rc_always_pace = 0;
19616 		if (rack_dsack_std_based & 0x1) {
19617 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
19618 			rack->rc_rack_tmr_std_based = 1;
19619 		}
19620 		if (rack_dsack_std_based & 0x2) {
19621 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
19622 			rack->rc_rack_use_dsack = 1;
19623 		}
19624 		if (rack_use_cmp_acks)
19625 			rack->r_use_cmp_ack = 1;
19626 		else
19627 			rack->r_use_cmp_ack = 0;
19628 		if (rack_disable_prr)
19629 			rack->rack_no_prr = 1;
19630 		else
19631 			rack->rack_no_prr = 0;
19632 		if (rack_gp_no_rec_chg)
19633 			rack->rc_gp_no_rec_chg = 1;
19634 		else
19635 			rack->rc_gp_no_rec_chg = 0;
19636 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
19637 			rack->r_mbuf_queue = 1;
19638 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19639 				rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19640 			rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19641 		} else {
19642 			rack->r_mbuf_queue = 0;
19643 			rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19644 		}
19645 		if (rack_enable_shared_cwnd)
19646 			rack->rack_enable_scwnd = 1;
19647 		else
19648 			rack->rack_enable_scwnd = 0;
19649 		if (rack_do_dyn_mul) {
19650 			/* When dynamic adjustment is on CA needs to start at 100% */
19651 			rack->rc_gp_dyn_mul = 1;
19652 			if (rack_do_dyn_mul >= 100)
19653 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
19654 		} else {
19655 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
19656 			rack->rc_gp_dyn_mul = 0;
19657 		}
19658 		rack->r_rr_config = 0;
19659 		rack->r_ctl.rc_no_push_at_mrtt = 0;
19660 		rack->rc_pace_to_cwnd = 0;
19661 		rack->rc_pace_fill_if_rttin_range = 0;
19662 		rack->rtt_limit_mul = 0;
19663 
19664 		if (rack_enable_hw_pacing)
19665 			rack->rack_hdw_pace_ena = 1;
19666 		else
19667 			rack->rack_hdw_pace_ena = 0;
19668 		if (rack_disable_prr)
19669 			rack->rack_no_prr = 1;
19670 		else
19671 			rack->rack_no_prr = 0;
19672 		if (rack_limits_scwnd)
19673 			rack->r_limit_scw  = 1;
19674 		else
19675 			rack->r_limit_scw  = 0;
19676 		err = 0;
19677 	}
19678 	return (err);
19679 }
19680 
19681 static int
19682 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
19683 {
19684 	struct deferred_opt_list *dol;
19685 
19686 	dol = malloc(sizeof(struct deferred_opt_list),
19687 		     M_TCPFSB, M_NOWAIT|M_ZERO);
19688 	if (dol == NULL) {
19689 		/*
19690 		 * No space yikes -- fail out..
19691 		 */
19692 		return (0);
19693 	}
19694 	dol->optname = sopt_name;
19695 	dol->optval = loptval;
19696 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
19697 	return (1);
19698 }
19699 
19700 static int
19701 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
19702 		    uint32_t optval, uint64_t loptval)
19703 {
19704 	struct epoch_tracker et;
19705 	struct sockopt sopt;
19706 	struct cc_newreno_opts opt;
19707 	uint64_t val;
19708 	int error = 0;
19709 	uint16_t ca, ss;
19710 
19711 	switch (sopt_name) {
19712 
19713 	case TCP_RACK_DSACK_OPT:
19714 		RACK_OPTS_INC(tcp_rack_dsack_opt);
19715 		if (optval & 0x1) {
19716 			rack->rc_rack_tmr_std_based = 1;
19717 		} else {
19718 			rack->rc_rack_tmr_std_based = 0;
19719 		}
19720 		if (optval & 0x2) {
19721 			rack->rc_rack_use_dsack = 1;
19722 		} else {
19723 			rack->rc_rack_use_dsack = 0;
19724 		}
19725 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
19726 		break;
19727 	case TCP_RACK_PACING_BETA:
19728 		RACK_OPTS_INC(tcp_rack_beta);
19729 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
19730 			/* This only works for newreno. */
19731 			error = EINVAL;
19732 			break;
19733 		}
19734 		if (rack->rc_pacing_cc_set) {
19735 			/*
19736 			 * Set them into the real CC module
19737 			 * whats in the rack pcb is the old values
19738 			 * to be used on restoral/
19739 			 */
19740 			sopt.sopt_dir = SOPT_SET;
19741 			opt.name = CC_NEWRENO_BETA;
19742 			opt.val = optval;
19743 			if (CC_ALGO(tp)->ctl_output != NULL)
19744 				error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
19745 			else {
19746 				error = ENOENT;
19747 				break;
19748 			}
19749 		} else {
19750 			/*
19751 			 * Not pacing yet so set it into our local
19752 			 * rack pcb storage.
19753 			 */
19754 			rack->r_ctl.rc_saved_beta.beta = optval;
19755 		}
19756 		break;
19757 	case TCP_RACK_TIMER_SLOP:
19758 		RACK_OPTS_INC(tcp_rack_timer_slop);
19759 		rack->r_ctl.timer_slop = optval;
19760 		if (rack->rc_tp->t_srtt) {
19761 			/*
19762 			 * If we have an SRTT lets update t_rxtcur
19763 			 * to have the new slop.
19764 			 */
19765 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
19766 					   rack_rto_min, rack_rto_max,
19767 					   rack->r_ctl.timer_slop);
19768 		}
19769 		break;
19770 	case TCP_RACK_PACING_BETA_ECN:
19771 		RACK_OPTS_INC(tcp_rack_beta_ecn);
19772 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
19773 			/* This only works for newreno. */
19774 			error = EINVAL;
19775 			break;
19776 		}
19777 		if (rack->rc_pacing_cc_set) {
19778 			/*
19779 			 * Set them into the real CC module
19780 			 * whats in the rack pcb is the old values
19781 			 * to be used on restoral/
19782 			 */
19783 			sopt.sopt_dir = SOPT_SET;
19784 			opt.name = CC_NEWRENO_BETA_ECN;
19785 			opt.val = optval;
19786 			if (CC_ALGO(tp)->ctl_output != NULL)
19787 				error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
19788 			else
19789 				error = ENOENT;
19790 		} else {
19791 			/*
19792 			 * Not pacing yet so set it into our local
19793 			 * rack pcb storage.
19794 			 */
19795 			rack->r_ctl.rc_saved_beta.beta_ecn = optval;
19796 			rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
19797 		}
19798 		break;
19799 	case TCP_DEFER_OPTIONS:
19800 		RACK_OPTS_INC(tcp_defer_opt);
19801 		if (optval) {
19802 			if (rack->gp_ready) {
19803 				/* Too late */
19804 				error = EINVAL;
19805 				break;
19806 			}
19807 			rack->defer_options = 1;
19808 		} else
19809 			rack->defer_options = 0;
19810 		break;
19811 	case TCP_RACK_MEASURE_CNT:
19812 		RACK_OPTS_INC(tcp_rack_measure_cnt);
19813 		if (optval && (optval <= 0xff)) {
19814 			rack->r_ctl.req_measurements = optval;
19815 		} else
19816 			error = EINVAL;
19817 		break;
19818 	case TCP_REC_ABC_VAL:
19819 		RACK_OPTS_INC(tcp_rec_abc_val);
19820 		if (optval > 0)
19821 			rack->r_use_labc_for_rec = 1;
19822 		else
19823 			rack->r_use_labc_for_rec = 0;
19824 		break;
19825 	case TCP_RACK_ABC_VAL:
19826 		RACK_OPTS_INC(tcp_rack_abc_val);
19827 		if ((optval > 0) && (optval < 255))
19828 			rack->rc_labc = optval;
19829 		else
19830 			error = EINVAL;
19831 		break;
19832 	case TCP_HDWR_UP_ONLY:
19833 		RACK_OPTS_INC(tcp_pacing_up_only);
19834 		if (optval)
19835 			rack->r_up_only = 1;
19836 		else
19837 			rack->r_up_only = 0;
19838 		break;
19839 	case TCP_PACING_RATE_CAP:
19840 		RACK_OPTS_INC(tcp_pacing_rate_cap);
19841 		rack->r_ctl.bw_rate_cap = loptval;
19842 		break;
19843 	case TCP_RACK_PROFILE:
19844 		RACK_OPTS_INC(tcp_profile);
19845 		error = rack_set_profile(rack, optval);
19846 		break;
19847 	case TCP_USE_CMP_ACKS:
19848 		RACK_OPTS_INC(tcp_use_cmp_acks);
19849 		if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
19850 			/* You can't turn it off once its on! */
19851 			error = EINVAL;
19852 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
19853 			rack->r_use_cmp_ack = 1;
19854 			rack->r_mbuf_queue = 1;
19855 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19856 		}
19857 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
19858 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19859 		break;
19860 	case TCP_SHARED_CWND_TIME_LIMIT:
19861 		RACK_OPTS_INC(tcp_lscwnd);
19862 		if (optval)
19863 			rack->r_limit_scw = 1;
19864 		else
19865 			rack->r_limit_scw = 0;
19866 		break;
19867  	case TCP_RACK_PACE_TO_FILL:
19868 		RACK_OPTS_INC(tcp_fillcw);
19869 		if (optval == 0)
19870 			rack->rc_pace_to_cwnd = 0;
19871 		else {
19872 			rack->rc_pace_to_cwnd = 1;
19873 			if (optval > 1)
19874 				rack->r_fill_less_agg = 1;
19875 		}
19876 		if ((optval >= rack_gp_rtt_maxmul) &&
19877 		    rack_gp_rtt_maxmul &&
19878 		    (optval < 0xf)) {
19879 			rack->rc_pace_fill_if_rttin_range = 1;
19880 			rack->rtt_limit_mul = optval;
19881 		} else {
19882 			rack->rc_pace_fill_if_rttin_range = 0;
19883 			rack->rtt_limit_mul = 0;
19884 		}
19885 		break;
19886 	case TCP_RACK_NO_PUSH_AT_MAX:
19887 		RACK_OPTS_INC(tcp_npush);
19888 		if (optval == 0)
19889 			rack->r_ctl.rc_no_push_at_mrtt = 0;
19890 		else if (optval < 0xff)
19891 			rack->r_ctl.rc_no_push_at_mrtt = optval;
19892 		else
19893 			error = EINVAL;
19894 		break;
19895 	case TCP_SHARED_CWND_ENABLE:
19896 		RACK_OPTS_INC(tcp_rack_scwnd);
19897 		if (optval == 0)
19898 			rack->rack_enable_scwnd = 0;
19899 		else
19900 			rack->rack_enable_scwnd = 1;
19901 		break;
19902 	case TCP_RACK_MBUF_QUEUE:
19903 		/* Now do we use the LRO mbuf-queue feature */
19904 		RACK_OPTS_INC(tcp_rack_mbufq);
19905 		if (optval || rack->r_use_cmp_ack)
19906 			rack->r_mbuf_queue = 1;
19907 		else
19908 			rack->r_mbuf_queue = 0;
19909 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19910 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19911 		else
19912 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19913 		break;
19914 	case TCP_RACK_NONRXT_CFG_RATE:
19915 		RACK_OPTS_INC(tcp_rack_cfg_rate);
19916 		if (optval == 0)
19917 			rack->rack_rec_nonrxt_use_cr = 0;
19918 		else
19919 			rack->rack_rec_nonrxt_use_cr = 1;
19920 		break;
19921 	case TCP_NO_PRR:
19922 		RACK_OPTS_INC(tcp_rack_noprr);
19923 		if (optval == 0)
19924 			rack->rack_no_prr = 0;
19925 		else if (optval == 1)
19926 			rack->rack_no_prr = 1;
19927 		else if (optval == 2)
19928 			rack->no_prr_addback = 1;
19929 		else
19930 			error = EINVAL;
19931 		break;
19932 	case TCP_TIMELY_DYN_ADJ:
19933 		RACK_OPTS_INC(tcp_timely_dyn);
19934 		if (optval == 0)
19935 			rack->rc_gp_dyn_mul = 0;
19936 		else {
19937 			rack->rc_gp_dyn_mul = 1;
19938 			if (optval >= 100) {
19939 				/*
19940 				 * If the user sets something 100 or more
19941 				 * its the gp_ca value.
19942 				 */
19943 				rack->r_ctl.rack_per_of_gp_ca  = optval;
19944 			}
19945 		}
19946 		break;
19947 	case TCP_RACK_DO_DETECTION:
19948 		RACK_OPTS_INC(tcp_rack_do_detection);
19949 		if (optval == 0)
19950 			rack->do_detection = 0;
19951 		else
19952 			rack->do_detection = 1;
19953 		break;
19954 	case TCP_RACK_TLP_USE:
19955 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19956 			error = EINVAL;
19957 			break;
19958 		}
19959 		RACK_OPTS_INC(tcp_tlp_use);
19960 		rack->rack_tlp_threshold_use = optval;
19961 		break;
19962 	case TCP_RACK_TLP_REDUCE:
19963 		/* RACK TLP cwnd reduction (bool) */
19964 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
19965 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19966 		break;
19967 	/*  Pacing related ones */
19968 	case TCP_RACK_PACE_ALWAYS:
19969 		/*
19970 		 * zero is old rack method, 1 is new
19971 		 * method using a pacing rate.
19972 		 */
19973 		RACK_OPTS_INC(tcp_rack_pace_always);
19974 		if (optval > 0) {
19975 			if (rack->rc_always_pace) {
19976 				error = EALREADY;
19977 				break;
19978 			} else if (tcp_can_enable_pacing()) {
19979 				rack->rc_always_pace = 1;
19980 				if (rack->use_fixed_rate || rack->gp_ready)
19981 					rack_set_cc_pacing(rack);
19982 			}
19983 			else {
19984 				error = ENOSPC;
19985 				break;
19986 			}
19987 		} else {
19988 			if (rack->rc_always_pace) {
19989 				tcp_decrement_paced_conn();
19990 				rack->rc_always_pace = 0;
19991 				rack_undo_cc_pacing(rack);
19992 			}
19993 		}
19994 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19995 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19996 		else
19997 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19998 		/* A rate may be set irate or other, if so set seg size */
19999 		rack_update_seg(rack);
20000 		break;
20001 	case TCP_BBR_RACK_INIT_RATE:
20002 		RACK_OPTS_INC(tcp_initial_rate);
20003 		val = optval;
20004 		/* Change from kbits per second to bytes per second */
20005 		val *= 1000;
20006 		val /= 8;
20007 		rack->r_ctl.init_rate = val;
20008 		if (rack->rc_init_win != rack_default_init_window) {
20009 			uint32_t win, snt;
20010 
20011 			/*
20012 			 * Options don't always get applied
20013 			 * in the order you think. So in order
20014 			 * to assure we update a cwnd we need
20015 			 * to check and see if we are still
20016 			 * where we should raise the cwnd.
20017 			 */
20018 			win = rc_init_window(rack);
20019 			if (SEQ_GT(tp->snd_max, tp->iss))
20020 				snt = tp->snd_max - tp->iss;
20021 			else
20022 				snt = 0;
20023 			if ((snt < win) &&
20024 			    (tp->snd_cwnd < win))
20025 				tp->snd_cwnd = win;
20026 		}
20027 		if (rack->rc_always_pace)
20028 			rack_update_seg(rack);
20029 		break;
20030 	case TCP_BBR_IWINTSO:
20031 		RACK_OPTS_INC(tcp_initial_win);
20032 		if (optval && (optval <= 0xff)) {
20033 			uint32_t win, snt;
20034 
20035 			rack->rc_init_win = optval;
20036 			win = rc_init_window(rack);
20037 			if (SEQ_GT(tp->snd_max, tp->iss))
20038 				snt = tp->snd_max - tp->iss;
20039 			else
20040 				snt = 0;
20041 			if ((snt < win) &&
20042 			    (tp->t_srtt |
20043 #ifdef NETFLIX_PEAKRATE
20044 			     tp->t_maxpeakrate |
20045 #endif
20046 			     rack->r_ctl.init_rate)) {
20047 				/*
20048 				 * We are not past the initial window
20049 				 * and we have some bases for pacing,
20050 				 * so we need to possibly adjust up
20051 				 * the cwnd. Note even if we don't set
20052 				 * the cwnd, its still ok to raise the rc_init_win
20053 				 * which can be used coming out of idle when we
20054 				 * would have a rate.
20055 				 */
20056 				if (tp->snd_cwnd < win)
20057 					tp->snd_cwnd = win;
20058 			}
20059 			if (rack->rc_always_pace)
20060 				rack_update_seg(rack);
20061 		} else
20062 			error = EINVAL;
20063 		break;
20064 	case TCP_RACK_FORCE_MSEG:
20065 		RACK_OPTS_INC(tcp_rack_force_max_seg);
20066 		if (optval)
20067 			rack->rc_force_max_seg = 1;
20068 		else
20069 			rack->rc_force_max_seg = 0;
20070 		break;
20071 	case TCP_RACK_PACE_MAX_SEG:
20072 		/* Max segments size in a pace in bytes */
20073 		RACK_OPTS_INC(tcp_rack_max_seg);
20074 		rack->rc_user_set_max_segs = optval;
20075 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
20076 		break;
20077 	case TCP_RACK_PACE_RATE_REC:
20078 		/* Set the fixed pacing rate in Bytes per second ca */
20079 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
20080 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
20081 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
20082 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
20083 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
20084 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
20085 		rack->use_fixed_rate = 1;
20086 		if (rack->rc_always_pace)
20087 			rack_set_cc_pacing(rack);
20088 		rack_log_pacing_delay_calc(rack,
20089 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
20090 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
20091 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
20092 					   __LINE__, NULL,0);
20093 		break;
20094 
20095 	case TCP_RACK_PACE_RATE_SS:
20096 		/* Set the fixed pacing rate in Bytes per second ca */
20097 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
20098 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
20099 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
20100 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
20101 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
20102 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
20103 		rack->use_fixed_rate = 1;
20104 		if (rack->rc_always_pace)
20105 			rack_set_cc_pacing(rack);
20106 		rack_log_pacing_delay_calc(rack,
20107 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
20108 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
20109 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
20110 					   __LINE__, NULL, 0);
20111 		break;
20112 
20113 	case TCP_RACK_PACE_RATE_CA:
20114 		/* Set the fixed pacing rate in Bytes per second ca */
20115 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
20116 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
20117 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
20118 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
20119 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
20120 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
20121 		rack->use_fixed_rate = 1;
20122 		if (rack->rc_always_pace)
20123 			rack_set_cc_pacing(rack);
20124 		rack_log_pacing_delay_calc(rack,
20125 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
20126 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
20127 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
20128 					   __LINE__, NULL, 0);
20129 		break;
20130 	case TCP_RACK_GP_INCREASE_REC:
20131 		RACK_OPTS_INC(tcp_gp_inc_rec);
20132 		rack->r_ctl.rack_per_of_gp_rec = optval;
20133 		rack_log_pacing_delay_calc(rack,
20134 					   rack->r_ctl.rack_per_of_gp_ss,
20135 					   rack->r_ctl.rack_per_of_gp_ca,
20136 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20137 					   __LINE__, NULL, 0);
20138 		break;
20139 	case TCP_RACK_GP_INCREASE_CA:
20140 		RACK_OPTS_INC(tcp_gp_inc_ca);
20141 		ca = optval;
20142 		if (ca < 100) {
20143 			/*
20144 			 * We don't allow any reduction
20145 			 * over the GP b/w.
20146 			 */
20147 			error = EINVAL;
20148 			break;
20149 		}
20150 		rack->r_ctl.rack_per_of_gp_ca = ca;
20151 		rack_log_pacing_delay_calc(rack,
20152 					   rack->r_ctl.rack_per_of_gp_ss,
20153 					   rack->r_ctl.rack_per_of_gp_ca,
20154 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20155 					   __LINE__, NULL, 0);
20156 		break;
20157 	case TCP_RACK_GP_INCREASE_SS:
20158 		RACK_OPTS_INC(tcp_gp_inc_ss);
20159 		ss = optval;
20160 		if (ss < 100) {
20161 			/*
20162 			 * We don't allow any reduction
20163 			 * over the GP b/w.
20164 			 */
20165 			error = EINVAL;
20166 			break;
20167 		}
20168 		rack->r_ctl.rack_per_of_gp_ss = ss;
20169 		rack_log_pacing_delay_calc(rack,
20170 					   rack->r_ctl.rack_per_of_gp_ss,
20171 					   rack->r_ctl.rack_per_of_gp_ca,
20172 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20173 					   __LINE__, NULL, 0);
20174 		break;
20175 	case TCP_RACK_RR_CONF:
20176 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
20177 		if (optval && optval <= 3)
20178 			rack->r_rr_config = optval;
20179 		else
20180 			rack->r_rr_config = 0;
20181 		break;
20182 	case TCP_HDWR_RATE_CAP:
20183 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
20184 		if (optval) {
20185 			if (rack->r_rack_hw_rate_caps == 0)
20186 				rack->r_rack_hw_rate_caps = 1;
20187 			else
20188 				error = EALREADY;
20189 		} else {
20190 			rack->r_rack_hw_rate_caps = 0;
20191 		}
20192 		break;
20193 	case TCP_BBR_HDWR_PACE:
20194 		RACK_OPTS_INC(tcp_hdwr_pacing);
20195 		if (optval){
20196 			if (rack->rack_hdrw_pacing == 0) {
20197 				rack->rack_hdw_pace_ena = 1;
20198 				rack->rack_attempt_hdwr_pace = 0;
20199 			} else
20200 				error = EALREADY;
20201 		} else {
20202 			rack->rack_hdw_pace_ena = 0;
20203 #ifdef RATELIMIT
20204 			if (rack->r_ctl.crte != NULL) {
20205 				rack->rack_hdrw_pacing = 0;
20206 				rack->rack_attempt_hdwr_pace = 0;
20207 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
20208 				rack->r_ctl.crte = NULL;
20209 			}
20210 #endif
20211 		}
20212 		break;
20213 	/*  End Pacing related ones */
20214 	case TCP_RACK_PRR_SENDALOT:
20215 		/* Allow PRR to send more than one seg */
20216 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
20217 		rack->r_ctl.rc_prr_sendalot = optval;
20218 		break;
20219 	case TCP_RACK_MIN_TO:
20220 		/* Minimum time between rack t-o's in ms */
20221 		RACK_OPTS_INC(tcp_rack_min_to);
20222 		rack->r_ctl.rc_min_to = optval;
20223 		break;
20224 	case TCP_RACK_EARLY_SEG:
20225 		/* If early recovery max segments */
20226 		RACK_OPTS_INC(tcp_rack_early_seg);
20227 		rack->r_ctl.rc_early_recovery_segs = optval;
20228 		break;
20229 	case TCP_RACK_ENABLE_HYSTART:
20230 	{
20231 		struct sockopt sopt;
20232 		struct cc_newreno_opts opt;
20233 
20234 		sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
20235 		sopt.sopt_dir = SOPT_SET;
20236 		opt.name = CC_NEWRENO_ENABLE_HYSTART;
20237 		opt.val = optval;
20238 		if (CC_ALGO(tp)->ctl_output != NULL)
20239 			error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
20240 		else
20241 			error = EINVAL;
20242 	}
20243 	break;
20244 	case TCP_RACK_REORD_THRESH:
20245 		/* RACK reorder threshold (shift amount) */
20246 		RACK_OPTS_INC(tcp_rack_reord_thresh);
20247 		if ((optval > 0) && (optval < 31))
20248 			rack->r_ctl.rc_reorder_shift = optval;
20249 		else
20250 			error = EINVAL;
20251 		break;
20252 	case TCP_RACK_REORD_FADE:
20253 		/* Does reordering fade after ms time */
20254 		RACK_OPTS_INC(tcp_rack_reord_fade);
20255 		rack->r_ctl.rc_reorder_fade = optval;
20256 		break;
20257 	case TCP_RACK_TLP_THRESH:
20258 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
20259 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
20260 		if (optval)
20261 			rack->r_ctl.rc_tlp_threshold = optval;
20262 		else
20263 			error = EINVAL;
20264 		break;
20265 	case TCP_BBR_USE_RACK_RR:
20266 		RACK_OPTS_INC(tcp_rack_rr);
20267 		if (optval)
20268 			rack->use_rack_rr = 1;
20269 		else
20270 			rack->use_rack_rr = 0;
20271 		break;
20272 	case TCP_FAST_RSM_HACK:
20273 		RACK_OPTS_INC(tcp_rack_fastrsm_hack);
20274 		if (optval)
20275 			rack->fast_rsm_hack = 1;
20276 		else
20277 			rack->fast_rsm_hack = 0;
20278 		break;
20279 	case TCP_RACK_PKT_DELAY:
20280 		/* RACK added ms i.e. rack-rtt + reord + N */
20281 		RACK_OPTS_INC(tcp_rack_pkt_delay);
20282 		rack->r_ctl.rc_pkt_delay = optval;
20283 		break;
20284 	case TCP_DELACK:
20285 		RACK_OPTS_INC(tcp_rack_delayed_ack);
20286 		if (optval == 0)
20287 			tp->t_delayed_ack = 0;
20288 		else
20289 			tp->t_delayed_ack = 1;
20290 		if (tp->t_flags & TF_DELACK) {
20291 			tp->t_flags &= ~TF_DELACK;
20292 			tp->t_flags |= TF_ACKNOW;
20293 			NET_EPOCH_ENTER(et);
20294 			rack_output(tp);
20295 			NET_EPOCH_EXIT(et);
20296 		}
20297 		break;
20298 
20299 	case TCP_BBR_RACK_RTT_USE:
20300 		RACK_OPTS_INC(tcp_rack_rtt_use);
20301 		if ((optval != USE_RTT_HIGH) &&
20302 		    (optval != USE_RTT_LOW) &&
20303 		    (optval != USE_RTT_AVG))
20304 			error = EINVAL;
20305 		else
20306 			rack->r_ctl.rc_rate_sample_method = optval;
20307 		break;
20308 	case TCP_DATA_AFTER_CLOSE:
20309 		RACK_OPTS_INC(tcp_data_after_close);
20310 		if (optval)
20311 			rack->rc_allow_data_af_clo = 1;
20312 		else
20313 			rack->rc_allow_data_af_clo = 0;
20314 		break;
20315 	default:
20316 		break;
20317 	}
20318 #ifdef NETFLIX_STATS
20319 	tcp_log_socket_option(tp, sopt_name, optval, error);
20320 #endif
20321 	return (error);
20322 }
20323 
20324 
20325 static void
20326 rack_apply_deferred_options(struct tcp_rack *rack)
20327 {
20328 	struct deferred_opt_list *dol, *sdol;
20329 	uint32_t s_optval;
20330 
20331 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
20332 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
20333 		/* Disadvantage of deferal is you loose the error return */
20334 		s_optval = (uint32_t)dol->optval;
20335 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
20336 		free(dol, M_TCPDO);
20337 	}
20338 }
20339 
20340 static void
20341 rack_hw_tls_change(struct tcpcb *tp, int chg)
20342 {
20343 	/*
20344 	 * HW tls state has changed.. fix all
20345 	 * rsm's in flight.
20346 	 */
20347 	struct tcp_rack *rack;
20348 	struct rack_sendmap *rsm;
20349 
20350 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20351 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
20352 		if (chg)
20353 			rsm->r_hw_tls = 1;
20354 		else
20355 			rsm->r_hw_tls = 0;
20356 	}
20357 	if (chg)
20358 		rack->r_ctl.fsb.hw_tls = 1;
20359 	else
20360 		rack->r_ctl.fsb.hw_tls = 0;
20361 }
20362 
20363 static int
20364 rack_pru_options(struct tcpcb *tp, int flags)
20365 {
20366 	if (flags & PRUS_OOB)
20367 		return (EOPNOTSUPP);
20368 	return (0);
20369 }
20370 
20371 static struct tcp_function_block __tcp_rack = {
20372 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
20373 	.tfb_tcp_output = rack_output,
20374 	.tfb_do_queued_segments = ctf_do_queued_segments,
20375 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
20376 	.tfb_tcp_do_segment = rack_do_segment,
20377 	.tfb_tcp_ctloutput = rack_ctloutput,
20378 	.tfb_tcp_fb_init = rack_init,
20379 	.tfb_tcp_fb_fini = rack_fini,
20380 	.tfb_tcp_timer_stop_all = rack_stopall,
20381 	.tfb_tcp_timer_activate = rack_timer_activate,
20382 	.tfb_tcp_timer_active = rack_timer_active,
20383 	.tfb_tcp_timer_stop = rack_timer_stop,
20384 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
20385 	.tfb_tcp_handoff_ok = rack_handoff_ok,
20386 	.tfb_tcp_mtu_chg = rack_mtu_change,
20387 	.tfb_pru_options = rack_pru_options,
20388 	.tfb_hwtls_change = rack_hw_tls_change,
20389 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
20390 };
20391 
20392 /*
20393  * rack_ctloutput() must drop the inpcb lock before performing copyin on
20394  * socket option arguments.  When it re-acquires the lock after the copy, it
20395  * has to revalidate that the connection is still valid for the socket
20396  * option.
20397  */
20398 static int
20399 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
20400     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
20401 {
20402 #ifdef INET6
20403 	struct ip6_hdr *ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
20404 #endif
20405 #ifdef INET
20406 	struct ip *ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
20407 #endif
20408 	uint64_t loptval;
20409 	int32_t error = 0, optval;
20410 
20411 	switch (sopt->sopt_level) {
20412 #ifdef INET6
20413 	case IPPROTO_IPV6:
20414 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
20415 		switch (sopt->sopt_name) {
20416 		case IPV6_USE_MIN_MTU:
20417 			tcp6_use_min_mtu(tp);
20418 			break;
20419 		case IPV6_TCLASS:
20420 			/*
20421 			 * The DSCP codepoint has changed, update the fsb.
20422 			 */
20423 			ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
20424 			    (rack->rc_inp->inp_flow & IPV6_FLOWINFO_MASK);
20425 			break;
20426 		}
20427 		INP_WUNLOCK(inp);
20428 		return (0);
20429 #endif
20430 #ifdef INET
20431 	case IPPROTO_IP:
20432 		switch (sopt->sopt_name) {
20433 		case IP_TOS:
20434 			/*
20435 			 * The DSCP codepoint has changed, update the fsb.
20436 			 */
20437 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
20438 			break;
20439 		case IP_TTL:
20440 			/*
20441 			 * The TTL has changed, update the fsb.
20442 			 */
20443 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
20444 			break;
20445 		}
20446 		INP_WUNLOCK(inp);
20447 		return (0);
20448 #endif
20449 	}
20450 
20451 	switch (sopt->sopt_name) {
20452 	case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
20453 	/*  Pacing related ones */
20454 	case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
20455 	case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
20456 	case TCP_BBR_IWINTSO:			/*  URL:tso_iwin */
20457 	case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
20458 	case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
20459 	case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
20460 	case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
20461 	case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
20462 	case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
20463 	case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
20464 	case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
20465 	case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
20466 	case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
20467 	case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
20468 	case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
20469 	case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
20470        /* End pacing related */
20471 	case TCP_FAST_RSM_HACK:			/*  URL:frsm_hack */
20472 	case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
20473 	case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
20474 	case TCP_RACK_MIN_TO:			/*  URL:min_to */
20475 	case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
20476 	case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
20477 	case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
20478 	case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
20479 	case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
20480 	case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
20481 	case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
20482 	case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
20483 	case TCP_RACK_DO_DETECTION:		/*  URL:detect */
20484 	case TCP_NO_PRR:			/*  URL:noprr */
20485 	case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
20486 	case TCP_DATA_AFTER_CLOSE:		/*  no URL */
20487 	case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
20488 	case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
20489 	case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
20490 	case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
20491 	case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
20492 	case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
20493 	case TCP_RACK_PROFILE:			/*  URL:profile */
20494 	case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
20495 	case TCP_RACK_ABC_VAL:			/*  URL:labc */
20496 	case TCP_REC_ABC_VAL:			/*  URL:reclabc */
20497 	case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
20498 	case TCP_DEFER_OPTIONS:			/*  URL:defer */
20499 	case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
20500 	case TCP_RACK_PACING_BETA:		/*  URL:pacing_beta */
20501 	case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
20502 	case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
20503 	case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
20504 		break;
20505 	default:
20506 		/* Filter off all unknown options to the base stack */
20507 		return (tcp_default_ctloutput(so, sopt, inp, tp));
20508 		break;
20509 	}
20510 	INP_WUNLOCK(inp);
20511 	if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
20512 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
20513 		/*
20514 		 * We truncate it down to 32 bits for the socket-option trace this
20515 		 * means rates > 34Gbps won't show right, but thats probably ok.
20516 		 */
20517 		optval = (uint32_t)loptval;
20518 	} else {
20519 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
20520 		/* Save it in 64 bit form too */
20521 		loptval = optval;
20522 	}
20523 	if (error)
20524 		return (error);
20525 	INP_WLOCK(inp);
20526 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
20527 		INP_WUNLOCK(inp);
20528 		return (ECONNRESET);
20529 	}
20530 	if (tp->t_fb != &__tcp_rack) {
20531 		INP_WUNLOCK(inp);
20532 		return (ENOPROTOOPT);
20533 	}
20534 	if (rack->defer_options && (rack->gp_ready == 0) &&
20535 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
20536 	    (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
20537 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
20538 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
20539 		/* Options are beind deferred */
20540 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
20541 			INP_WUNLOCK(inp);
20542 			return (0);
20543 		} else {
20544 			/* No memory to defer, fail */
20545 			INP_WUNLOCK(inp);
20546 			return (ENOMEM);
20547 		}
20548 	}
20549 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
20550 	INP_WUNLOCK(inp);
20551 	return (error);
20552 }
20553 
20554 static void
20555 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
20556 {
20557 
20558 	INP_WLOCK_ASSERT(tp->t_inpcb);
20559 	bzero(ti, sizeof(*ti));
20560 
20561 	ti->tcpi_state = tp->t_state;
20562 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
20563 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
20564 	if (tp->t_flags & TF_SACK_PERMIT)
20565 		ti->tcpi_options |= TCPI_OPT_SACK;
20566 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
20567 		ti->tcpi_options |= TCPI_OPT_WSCALE;
20568 		ti->tcpi_snd_wscale = tp->snd_scale;
20569 		ti->tcpi_rcv_wscale = tp->rcv_scale;
20570 	}
20571 	if (tp->t_flags2 & TF2_ECN_PERMIT)
20572 		ti->tcpi_options |= TCPI_OPT_ECN;
20573 	if (tp->t_flags & TF_FASTOPEN)
20574 		ti->tcpi_options |= TCPI_OPT_TFO;
20575 	/* still kept in ticks is t_rcvtime */
20576 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
20577 	/* Since we hold everything in precise useconds this is easy */
20578 	ti->tcpi_rtt = tp->t_srtt;
20579 	ti->tcpi_rttvar = tp->t_rttvar;
20580 	ti->tcpi_rto = tp->t_rxtcur;
20581 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
20582 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
20583 	/*
20584 	 * FreeBSD-specific extension fields for tcp_info.
20585 	 */
20586 	ti->tcpi_rcv_space = tp->rcv_wnd;
20587 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
20588 	ti->tcpi_snd_wnd = tp->snd_wnd;
20589 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
20590 	ti->tcpi_snd_nxt = tp->snd_nxt;
20591 	ti->tcpi_snd_mss = tp->t_maxseg;
20592 	ti->tcpi_rcv_mss = tp->t_maxseg;
20593 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
20594 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
20595 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
20596 #ifdef NETFLIX_STATS
20597 	ti->tcpi_total_tlp = tp->t_sndtlppack;
20598 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
20599 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
20600 #endif
20601 #ifdef TCP_OFFLOAD
20602 	if (tp->t_flags & TF_TOE) {
20603 		ti->tcpi_options |= TCPI_OPT_TOE;
20604 		tcp_offload_tcp_info(tp, ti);
20605 	}
20606 #endif
20607 }
20608 
20609 static int
20610 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
20611     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
20612 {
20613 	int32_t error, optval;
20614 	uint64_t val, loptval;
20615 	struct	tcp_info ti;
20616 	/*
20617 	 * Because all our options are either boolean or an int, we can just
20618 	 * pull everything into optval and then unlock and copy. If we ever
20619 	 * add a option that is not a int, then this will have quite an
20620 	 * impact to this routine.
20621 	 */
20622 	error = 0;
20623 	switch (sopt->sopt_name) {
20624 	case TCP_INFO:
20625 		/* First get the info filled */
20626 		rack_fill_info(tp, &ti);
20627 		/* Fix up the rtt related fields if needed */
20628 		INP_WUNLOCK(inp);
20629 		error = sooptcopyout(sopt, &ti, sizeof ti);
20630 		return (error);
20631 	/*
20632 	 * Beta is the congestion control value for NewReno that influences how
20633 	 * much of a backoff happens when loss is detected. It is normally set
20634 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
20635 	 * when you exit recovery.
20636 	 */
20637 	case TCP_RACK_PACING_BETA:
20638 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
20639 			error = EINVAL;
20640 		else if (rack->rc_pacing_cc_set == 0)
20641 			optval = rack->r_ctl.rc_saved_beta.beta;
20642 		else {
20643 			/*
20644 			 * Reach out into the CC data and report back what
20645 			 * I have previously set. Yeah it looks hackish but
20646 			 * we don't want to report the saved values.
20647 			 */
20648 			if (tp->ccv->cc_data)
20649 				optval = ((struct newreno *)tp->ccv->cc_data)->beta;
20650 			else
20651 				error = EINVAL;
20652 		}
20653 		break;
20654 		/*
20655 		 * Beta_ecn is the congestion control value for NewReno that influences how
20656 		 * much of a backoff happens when a ECN mark is detected. It is normally set
20657 		 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
20658 		 * you exit recovery. Note that classic ECN has a beta of 50, it is only
20659 		 * ABE Ecn that uses this "less" value, but we do too with pacing :)
20660 		 */
20661 
20662 	case TCP_RACK_PACING_BETA_ECN:
20663 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
20664 			error = EINVAL;
20665 		else if (rack->rc_pacing_cc_set == 0)
20666 			optval = rack->r_ctl.rc_saved_beta.beta_ecn;
20667 		else {
20668 			/*
20669 			 * Reach out into the CC data and report back what
20670 			 * I have previously set. Yeah it looks hackish but
20671 			 * we don't want to report the saved values.
20672 			 */
20673 			if (tp->ccv->cc_data)
20674 				optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn;
20675 			else
20676 				error = EINVAL;
20677 		}
20678 		break;
20679 	case TCP_RACK_DSACK_OPT:
20680 		optval = 0;
20681 		if (rack->rc_rack_tmr_std_based) {
20682 			optval |= 1;
20683 		}
20684 		if (rack->rc_rack_use_dsack) {
20685 			optval |= 2;
20686 		}
20687 		break;
20688  	case TCP_RACK_ENABLE_HYSTART:
20689 	{
20690 		struct sockopt sopt;
20691 		struct cc_newreno_opts opt;
20692 
20693 		sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
20694 		sopt.sopt_dir = SOPT_GET;
20695 		opt.name = CC_NEWRENO_ENABLE_HYSTART;
20696 		if (CC_ALGO(tp)->ctl_output != NULL)
20697 			error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
20698 		else
20699 			error = EINVAL;
20700 		optval = opt.val;
20701 	}
20702 	break;
20703 	case TCP_FAST_RSM_HACK:
20704 		optval = rack->fast_rsm_hack;
20705 		break;
20706 	case TCP_DEFER_OPTIONS:
20707 		optval = rack->defer_options;
20708 		break;
20709 	case TCP_RACK_MEASURE_CNT:
20710 		optval = rack->r_ctl.req_measurements;
20711 		break;
20712 	case TCP_REC_ABC_VAL:
20713 		optval = rack->r_use_labc_for_rec;
20714 		break;
20715 	case TCP_RACK_ABC_VAL:
20716 		optval = rack->rc_labc;
20717 		break;
20718 	case TCP_HDWR_UP_ONLY:
20719 		optval= rack->r_up_only;
20720 		break;
20721 	case TCP_PACING_RATE_CAP:
20722 		loptval = rack->r_ctl.bw_rate_cap;
20723 		break;
20724 	case TCP_RACK_PROFILE:
20725 		/* You cannot retrieve a profile, its write only */
20726 		error = EINVAL;
20727 		break;
20728 	case TCP_USE_CMP_ACKS:
20729 		optval = rack->r_use_cmp_ack;
20730 		break;
20731 	case TCP_RACK_PACE_TO_FILL:
20732 		optval = rack->rc_pace_to_cwnd;
20733 		if (optval && rack->r_fill_less_agg)
20734 			optval++;
20735 		break;
20736 	case TCP_RACK_NO_PUSH_AT_MAX:
20737 		optval = rack->r_ctl.rc_no_push_at_mrtt;
20738 		break;
20739 	case TCP_SHARED_CWND_ENABLE:
20740 		optval = rack->rack_enable_scwnd;
20741 		break;
20742 	case TCP_RACK_NONRXT_CFG_RATE:
20743 		optval = rack->rack_rec_nonrxt_use_cr;
20744 		break;
20745 	case TCP_NO_PRR:
20746 		if (rack->rack_no_prr  == 1)
20747 			optval = 1;
20748 		else if (rack->no_prr_addback == 1)
20749 			optval = 2;
20750 		else
20751 			optval = 0;
20752 		break;
20753 	case TCP_RACK_DO_DETECTION:
20754 		optval = rack->do_detection;
20755 		break;
20756 	case TCP_RACK_MBUF_QUEUE:
20757 		/* Now do we use the LRO mbuf-queue feature */
20758 		optval = rack->r_mbuf_queue;
20759 		break;
20760 	case TCP_TIMELY_DYN_ADJ:
20761 		optval = rack->rc_gp_dyn_mul;
20762 		break;
20763 	case TCP_BBR_IWINTSO:
20764 		optval = rack->rc_init_win;
20765 		break;
20766 	case TCP_RACK_TLP_REDUCE:
20767 		/* RACK TLP cwnd reduction (bool) */
20768 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
20769 		break;
20770 	case TCP_BBR_RACK_INIT_RATE:
20771 		val = rack->r_ctl.init_rate;
20772 		/* convert to kbits per sec */
20773 		val *= 8;
20774 		val /= 1000;
20775 		optval = (uint32_t)val;
20776 		break;
20777 	case TCP_RACK_FORCE_MSEG:
20778 		optval = rack->rc_force_max_seg;
20779 		break;
20780 	case TCP_RACK_PACE_MAX_SEG:
20781 		/* Max segments in a pace */
20782 		optval = rack->rc_user_set_max_segs;
20783 		break;
20784 	case TCP_RACK_PACE_ALWAYS:
20785 		/* Use the always pace method */
20786 		optval = rack->rc_always_pace;
20787 		break;
20788 	case TCP_RACK_PRR_SENDALOT:
20789 		/* Allow PRR to send more than one seg */
20790 		optval = rack->r_ctl.rc_prr_sendalot;
20791 		break;
20792 	case TCP_RACK_MIN_TO:
20793 		/* Minimum time between rack t-o's in ms */
20794 		optval = rack->r_ctl.rc_min_to;
20795 		break;
20796 	case TCP_RACK_EARLY_SEG:
20797 		/* If early recovery max segments */
20798 		optval = rack->r_ctl.rc_early_recovery_segs;
20799 		break;
20800 	case TCP_RACK_REORD_THRESH:
20801 		/* RACK reorder threshold (shift amount) */
20802 		optval = rack->r_ctl.rc_reorder_shift;
20803 		break;
20804 	case TCP_RACK_REORD_FADE:
20805 		/* Does reordering fade after ms time */
20806 		optval = rack->r_ctl.rc_reorder_fade;
20807 		break;
20808 	case TCP_BBR_USE_RACK_RR:
20809 		/* Do we use the rack cheat for rxt */
20810 		optval = rack->use_rack_rr;
20811 		break;
20812 	case TCP_RACK_RR_CONF:
20813 		optval = rack->r_rr_config;
20814 		break;
20815 	case TCP_HDWR_RATE_CAP:
20816 		optval = rack->r_rack_hw_rate_caps;
20817 		break;
20818 	case TCP_BBR_HDWR_PACE:
20819 		optval = rack->rack_hdw_pace_ena;
20820 		break;
20821 	case TCP_RACK_TLP_THRESH:
20822 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
20823 		optval = rack->r_ctl.rc_tlp_threshold;
20824 		break;
20825 	case TCP_RACK_PKT_DELAY:
20826 		/* RACK added ms i.e. rack-rtt + reord + N */
20827 		optval = rack->r_ctl.rc_pkt_delay;
20828 		break;
20829 	case TCP_RACK_TLP_USE:
20830 		optval = rack->rack_tlp_threshold_use;
20831 		break;
20832 	case TCP_RACK_PACE_RATE_CA:
20833 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
20834 		break;
20835 	case TCP_RACK_PACE_RATE_SS:
20836 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
20837 		break;
20838 	case TCP_RACK_PACE_RATE_REC:
20839 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
20840 		break;
20841 	case TCP_RACK_GP_INCREASE_SS:
20842 		optval = rack->r_ctl.rack_per_of_gp_ca;
20843 		break;
20844 	case TCP_RACK_GP_INCREASE_CA:
20845 		optval = rack->r_ctl.rack_per_of_gp_ss;
20846 		break;
20847 	case TCP_BBR_RACK_RTT_USE:
20848 		optval = rack->r_ctl.rc_rate_sample_method;
20849 		break;
20850 	case TCP_DELACK:
20851 		optval = tp->t_delayed_ack;
20852 		break;
20853 	case TCP_DATA_AFTER_CLOSE:
20854 		optval = rack->rc_allow_data_af_clo;
20855 		break;
20856 	case TCP_SHARED_CWND_TIME_LIMIT:
20857 		optval = rack->r_limit_scw;
20858 		break;
20859 	case TCP_RACK_TIMER_SLOP:
20860 		optval = rack->r_ctl.timer_slop;
20861 		break;
20862 	default:
20863 		return (tcp_default_ctloutput(so, sopt, inp, tp));
20864 		break;
20865 	}
20866 	INP_WUNLOCK(inp);
20867 	if (error == 0) {
20868 		if (TCP_PACING_RATE_CAP)
20869 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
20870 		else
20871 			error = sooptcopyout(sopt, &optval, sizeof optval);
20872 	}
20873 	return (error);
20874 }
20875 
20876 static int
20877 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
20878 {
20879 	int32_t error = EINVAL;
20880 	struct tcp_rack *rack;
20881 
20882 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20883 	if (rack == NULL) {
20884 		/* Huh? */
20885 		goto out;
20886 	}
20887 	if (sopt->sopt_dir == SOPT_SET) {
20888 		return (rack_set_sockopt(so, sopt, inp, tp, rack));
20889 	} else if (sopt->sopt_dir == SOPT_GET) {
20890 		return (rack_get_sockopt(so, sopt, inp, tp, rack));
20891 	}
20892 out:
20893 	INP_WUNLOCK(inp);
20894 	return (error);
20895 }
20896 
20897 static const char *rack_stack_names[] = {
20898 	__XSTRING(STACKNAME),
20899 #ifdef STACKALIAS
20900 	__XSTRING(STACKALIAS),
20901 #endif
20902 };
20903 
20904 static int
20905 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
20906 {
20907 	memset(mem, 0, size);
20908 	return (0);
20909 }
20910 
20911 static void
20912 rack_dtor(void *mem, int32_t size, void *arg)
20913 {
20914 
20915 }
20916 
20917 static bool rack_mod_inited = false;
20918 
20919 static int
20920 tcp_addrack(module_t mod, int32_t type, void *data)
20921 {
20922 	int32_t err = 0;
20923 	int num_stacks;
20924 
20925 	switch (type) {
20926 	case MOD_LOAD:
20927 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
20928 		    sizeof(struct rack_sendmap),
20929 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
20930 
20931 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
20932 		    sizeof(struct tcp_rack),
20933 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
20934 
20935 		sysctl_ctx_init(&rack_sysctl_ctx);
20936 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
20937 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
20938 		    OID_AUTO,
20939 #ifdef STACKALIAS
20940 		    __XSTRING(STACKALIAS),
20941 #else
20942 		    __XSTRING(STACKNAME),
20943 #endif
20944 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
20945 		    "");
20946 		if (rack_sysctl_root == NULL) {
20947 			printf("Failed to add sysctl node\n");
20948 			err = EFAULT;
20949 			goto free_uma;
20950 		}
20951 		rack_init_sysctls();
20952 		num_stacks = nitems(rack_stack_names);
20953 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
20954 		    rack_stack_names, &num_stacks);
20955 		if (err) {
20956 			printf("Failed to register %s stack name for "
20957 			    "%s module\n", rack_stack_names[num_stacks],
20958 			    __XSTRING(MODNAME));
20959 			sysctl_ctx_free(&rack_sysctl_ctx);
20960 free_uma:
20961 			uma_zdestroy(rack_zone);
20962 			uma_zdestroy(rack_pcb_zone);
20963 			rack_counter_destroy();
20964 			printf("Failed to register rack module -- err:%d\n", err);
20965 			return (err);
20966 		}
20967 		tcp_lro_reg_mbufq();
20968 		rack_mod_inited = true;
20969 		break;
20970 	case MOD_QUIESCE:
20971 		err = deregister_tcp_functions(&__tcp_rack, true, false);
20972 		break;
20973 	case MOD_UNLOAD:
20974 		err = deregister_tcp_functions(&__tcp_rack, false, true);
20975 		if (err == EBUSY)
20976 			break;
20977 		if (rack_mod_inited) {
20978 			uma_zdestroy(rack_zone);
20979 			uma_zdestroy(rack_pcb_zone);
20980 			sysctl_ctx_free(&rack_sysctl_ctx);
20981 			rack_counter_destroy();
20982 			rack_mod_inited = false;
20983 		}
20984 		tcp_lro_dereg_mbufq();
20985 		err = 0;
20986 		break;
20987 	default:
20988 		return (EOPNOTSUPP);
20989 	}
20990 	return (err);
20991 }
20992 
20993 static moduledata_t tcp_rack = {
20994 	.name = __XSTRING(MODNAME),
20995 	.evhand = tcp_addrack,
20996 	.priv = 0
20997 };
20998 
20999 MODULE_VERSION(MODNAME, 1);
21000 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
21001 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
21002