xref: /freebsd-src/sys/netinet/tcp_stacks/rack.c (revision 81b22a9892b1047e551fc3f1d6d58031bc59a4c3) 
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_segement
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 
209 static int32_t rack_pkt_delay = 1000;
210 static int32_t rack_send_a_lot_in_prr = 1;
211 static int32_t rack_min_to = 1000;	/* Number of microsecond  min timeout */
212 static int32_t rack_verbose_logging = 0;
213 static int32_t rack_ignore_data_after_close = 1;
214 static int32_t rack_enable_shared_cwnd = 1;
215 static int32_t rack_use_cmp_acks = 1;
216 static int32_t rack_use_fsb = 1;
217 static int32_t rack_use_rfo = 1;
218 static int32_t rack_use_rsm_rfo = 1;
219 static int32_t rack_max_abc_post_recovery = 2;
220 static int32_t rack_client_low_buf = 0;
221 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 */
222 #ifdef TCP_ACCOUNTING
223 static int32_t rack_tcp_accounting = 0;
224 #endif
225 static int32_t rack_limits_scwnd = 1;
226 static int32_t rack_enable_mqueue_for_nonpaced = 0;
227 static int32_t rack_disable_prr = 0;
228 static int32_t use_rack_rr = 1;
229 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
230 static int32_t rack_persist_min = 250000;	/* 250usec */
231 static int32_t rack_persist_max = 2000000;	/* 2 Second in usec's */
232 static int32_t rack_sack_not_required = 1;	/* set to one to allow non-sack to use rack */
233 static int32_t rack_default_init_window = 0;	/* Use system default */
234 static int32_t rack_limit_time_with_srtt = 0;
235 static int32_t rack_autosndbuf_inc = 20;	/* In percentage form */
236 static int32_t rack_enobuf_hw_boost_mult = 2;	/* How many times the hw rate we boost slot using time_between */
237 static int32_t rack_enobuf_hw_max = 12000;	/* 12 ms in usecs */
238 static int32_t rack_enobuf_hw_min = 10000;	/* 10 ms in usecs */
239 static int32_t rack_hw_rwnd_factor = 2;		/* How many max_segs the rwnd must be before we hold off sending */
240 /*
241  * Currently regular tcp has a rto_min of 30ms
242  * the backoff goes 12 times so that ends up
243  * being a total of 122.850 seconds before a
244  * connection is killed.
245  */
246 static uint32_t rack_def_data_window = 20;
247 static uint32_t rack_goal_bdp = 2;
248 static uint32_t rack_min_srtts = 1;
249 static uint32_t rack_min_measure_usec = 0;
250 static int32_t rack_tlp_min = 10000;	/* 10ms */
251 static int32_t rack_rto_min = 30000;	/* 30,000 usec same as main freebsd */
252 static int32_t rack_rto_max = 4000000;	/* 4 seconds in usec's */
253 static const int32_t rack_free_cache = 2;
254 static int32_t rack_hptsi_segments = 40;
255 static int32_t rack_rate_sample_method = USE_RTT_LOW;
256 static int32_t rack_pace_every_seg = 0;
257 static int32_t rack_delayed_ack_time = 40000;	/* 40ms in usecs */
258 static int32_t rack_slot_reduction = 4;
259 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
260 static int32_t rack_cwnd_block_ends_measure = 0;
261 static int32_t rack_rwnd_block_ends_measure = 0;
262 static int32_t rack_def_profile = 0;
263 
264 static int32_t rack_lower_cwnd_at_tlp = 0;
265 static int32_t rack_limited_retran = 0;
266 static int32_t rack_always_send_oldest = 0;
267 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
268 
269 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
270 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
271 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
272 
273 /* Probertt */
274 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
275 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
276 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
277 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
278 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
279 
280 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
281 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
282 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
283 static uint32_t rack_probertt_use_min_rtt_exit = 0;
284 static uint32_t rack_probe_rtt_sets_cwnd = 0;
285 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
286 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in usecs */
287 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
288 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction */
289 static uint32_t rack_min_probertt_hold = 40000;		/* Equal to delayed ack time */
290 static uint32_t rack_probertt_filter_life = 10000000;
291 static uint32_t rack_probertt_lower_within = 10;
292 static uint32_t rack_min_rtt_movement = 250000;	/* Must move at least 250ms (in microseconds)  to count as a lowering */
293 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
294 static int32_t rack_probertt_clear_is = 1;
295 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
296 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
297 
298 /* Part of pacing */
299 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
300 
301 /* Timely information */
302 /* Combine these two gives the range of 'no change' to bw */
303 /* ie the up/down provide the upper and lower bound */
304 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
305 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
306 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
307 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
308 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
309 static int32_t rack_gp_decrease_per = 20;	/* 20% decrease in multipler */
310 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multipler */
311 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
312 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
313 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
314 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
315 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
316 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
317 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
318 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
319 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
320 static int32_t rack_use_max_for_nobackoff = 0;
321 static int32_t rack_timely_int_timely_only = 0;	/* do interim timely's only use the timely algo (no b/w changes)? */
322 static int32_t rack_timely_no_stopping = 0;
323 static int32_t rack_down_raise_thresh = 100;
324 static int32_t rack_req_segs = 1;
325 static uint64_t rack_bw_rate_cap = 0;
326 
327 /* Weird delayed ack mode */
328 static int32_t rack_use_imac_dack = 0;
329 /* Rack specific counters */
330 counter_u64_t rack_badfr;
331 counter_u64_t rack_badfr_bytes;
332 counter_u64_t rack_rtm_prr_retran;
333 counter_u64_t rack_rtm_prr_newdata;
334 counter_u64_t rack_timestamp_mismatch;
335 counter_u64_t rack_reorder_seen;
336 counter_u64_t rack_paced_segments;
337 counter_u64_t rack_unpaced_segments;
338 counter_u64_t rack_calc_zero;
339 counter_u64_t rack_calc_nonzero;
340 counter_u64_t rack_saw_enobuf;
341 counter_u64_t rack_saw_enobuf_hw;
342 counter_u64_t rack_saw_enetunreach;
343 counter_u64_t rack_per_timer_hole;
344 counter_u64_t rack_large_ackcmp;
345 counter_u64_t rack_small_ackcmp;
346 #ifdef INVARIANTS
347 counter_u64_t rack_adjust_map_bw;
348 #endif
349 /* Tail loss probe counters */
350 counter_u64_t rack_tlp_tot;
351 counter_u64_t rack_tlp_newdata;
352 counter_u64_t rack_tlp_retran;
353 counter_u64_t rack_tlp_retran_bytes;
354 counter_u64_t rack_tlp_retran_fail;
355 counter_u64_t rack_to_tot;
356 counter_u64_t rack_to_arm_rack;
357 counter_u64_t rack_to_arm_tlp;
358 counter_u64_t rack_hot_alloc;
359 counter_u64_t rack_to_alloc;
360 counter_u64_t rack_to_alloc_hard;
361 counter_u64_t rack_to_alloc_emerg;
362 counter_u64_t rack_to_alloc_limited;
363 counter_u64_t rack_alloc_limited_conns;
364 counter_u64_t rack_split_limited;
365 
366 #define MAX_NUM_OF_CNTS 13
367 counter_u64_t rack_proc_comp_ack[MAX_NUM_OF_CNTS];
368 counter_u64_t rack_multi_single_eq;
369 counter_u64_t rack_proc_non_comp_ack;
370 
371 counter_u64_t rack_fto_send;
372 counter_u64_t rack_fto_rsm_send;
373 counter_u64_t rack_nfto_resend;
374 counter_u64_t rack_non_fto_send;
375 counter_u64_t rack_extended_rfo;
376 
377 counter_u64_t rack_sack_proc_all;
378 counter_u64_t rack_sack_proc_short;
379 counter_u64_t rack_sack_proc_restart;
380 counter_u64_t rack_sack_attacks_detected;
381 counter_u64_t rack_sack_attacks_reversed;
382 counter_u64_t rack_sack_used_next_merge;
383 counter_u64_t rack_sack_splits;
384 counter_u64_t rack_sack_used_prev_merge;
385 counter_u64_t rack_sack_skipped_acked;
386 counter_u64_t rack_ack_total;
387 counter_u64_t rack_express_sack;
388 counter_u64_t rack_sack_total;
389 counter_u64_t rack_move_none;
390 counter_u64_t rack_move_some;
391 
392 counter_u64_t rack_used_tlpmethod;
393 counter_u64_t rack_used_tlpmethod2;
394 counter_u64_t rack_enter_tlp_calc;
395 counter_u64_t rack_input_idle_reduces;
396 counter_u64_t rack_collapsed_win;
397 counter_u64_t rack_tlp_does_nada;
398 counter_u64_t rack_try_scwnd;
399 counter_u64_t rack_hw_pace_init_fail;
400 counter_u64_t rack_hw_pace_lost;
401 counter_u64_t rack_sbsndptr_right;
402 counter_u64_t rack_sbsndptr_wrong;
403 
404 /* Temp CPU counters */
405 counter_u64_t rack_find_high;
406 
407 counter_u64_t rack_progress_drops;
408 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
409 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
410 
411 
412 #define	RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
413 
414 #define	RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {	\
415 	(tv) = (value) + slop;	 \
416 	if ((u_long)(tv) < (u_long)(tvmin)) \
417 		(tv) = (tvmin); \
418 	if ((u_long)(tv) > (u_long)(tvmax)) \
419 		(tv) = (tvmax); \
420 } while (0)
421 
422 static void
423 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
424 
425 static int
426 rack_process_ack(struct mbuf *m, struct tcphdr *th,
427     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
428     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
429 static int
430 rack_process_data(struct mbuf *m, struct tcphdr *th,
431     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
432     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
433 static void
434 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
435    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
436 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
437 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
438     uint8_t limit_type);
439 static struct rack_sendmap *
440 rack_check_recovery_mode(struct tcpcb *tp,
441     uint32_t tsused);
442 static void
443 rack_cong_signal(struct tcpcb *tp,
444 		 uint32_t type, uint32_t ack);
445 static void rack_counter_destroy(void);
446 static int
447 rack_ctloutput(struct socket *so, struct sockopt *sopt,
448     struct inpcb *inp, struct tcpcb *tp);
449 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
450 static void
451 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
452 static void
453 rack_do_segment(struct mbuf *m, struct tcphdr *th,
454     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
455     uint8_t iptos);
456 static void rack_dtor(void *mem, int32_t size, void *arg);
457 static void
458 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
459     uint32_t flex1, uint32_t flex2,
460     uint32_t flex3, uint32_t flex4,
461     uint32_t flex5, uint32_t flex6,
462     uint16_t flex7, uint8_t mod);
463 
464 static void
465 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
466    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
467    struct rack_sendmap *rsm, uint8_t quality);
468 static struct rack_sendmap *
469 rack_find_high_nonack(struct tcp_rack *rack,
470     struct rack_sendmap *rsm);
471 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
472 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
473 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
474 static int
475 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
476     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
477 static void
478 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
479 			    tcp_seq th_ack, int line, uint8_t quality);
480 static uint32_t
481 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
482 static int32_t rack_handoff_ok(struct tcpcb *tp);
483 static int32_t rack_init(struct tcpcb *tp);
484 static void rack_init_sysctls(void);
485 static void
486 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
487     struct tcphdr *th, int entered_rec, int dup_ack_struck);
488 static void
489 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
490     uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t ts,
491     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls);
492 
493 static void
494 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
495     struct rack_sendmap *rsm);
496 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
497 static int32_t rack_output(struct tcpcb *tp);
498 
499 static uint32_t
500 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
501     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
502     uint32_t cts, int *moved_two);
503 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
504 static void rack_remxt_tmr(struct tcpcb *tp);
505 static int
506 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
507     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
508 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
509 static int32_t rack_stopall(struct tcpcb *tp);
510 static void
511 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type,
512     uint32_t delta);
513 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type);
514 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
515 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type);
516 static uint32_t
517 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
518     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
519 static void
520 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
521     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
522 static int
523 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
524     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
525 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
526 static int
527 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
528     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
529     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
530 static int
531 rack_do_closing(struct mbuf *m, struct tcphdr *th,
532     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
533     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
534 static int
535 rack_do_established(struct mbuf *m, struct tcphdr *th,
536     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
537     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
538 static int
539 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
540     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
541     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
542 static int
543 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
544     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
545     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
546 static int
547 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
548     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
549     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
550 static int
551 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
552     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
553     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
554 static int
555 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
556     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
557     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
558 static int
559 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
560     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
561     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
562 struct rack_sendmap *
563 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
564     uint32_t tsused);
565 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
566     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
567 static void
568      tcp_rack_partialack(struct tcpcb *tp);
569 static int
570 rack_set_profile(struct tcp_rack *rack, int prof);
571 static void
572 rack_apply_deferred_options(struct tcp_rack *rack);
573 
574 int32_t rack_clear_counter=0;
575 
576 static void
577 rack_set_cc_pacing(struct tcp_rack *rack)
578 {
579 	struct sockopt sopt;
580 	struct cc_newreno_opts opt;
581 	struct newreno old, *ptr;
582 	struct tcpcb *tp;
583 	int error;
584 
585 	if (rack->rc_pacing_cc_set)
586 		return;
587 
588 	tp = rack->rc_tp;
589 	if (tp->cc_algo == NULL) {
590 		/* Tcb is leaving */
591 		printf("No cc algorithm?\n");
592 		return;
593 	}
594 	rack->rc_pacing_cc_set = 1;
595 	if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
596 		/* Not new-reno we can't play games with beta! */
597 		goto out;
598 	}
599 	ptr = ((struct newreno *)tp->ccv->cc_data);
600 	if (CC_ALGO(tp)->ctl_output == NULL)  {
601 		/* Huh, why does new_reno no longer have a set function? */
602 		goto out;
603 	}
604 	if (ptr == NULL) {
605 		/* Just the default values */
606 		old.beta = V_newreno_beta_ecn;
607 		old.beta_ecn = V_newreno_beta_ecn;
608 		old.newreno_flags = 0;
609 	} else {
610 		old.beta = ptr->beta;
611 		old.beta_ecn = ptr->beta_ecn;
612 		old.newreno_flags = ptr->newreno_flags;
613 	}
614 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
615 	sopt.sopt_dir = SOPT_SET;
616 	opt.name = CC_NEWRENO_BETA;
617 	opt.val = rack->r_ctl.rc_saved_beta.beta;
618 	error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
619 	if (error)  {
620 		goto out;
621 	}
622 	/*
623 	 * Hack alert we need to set in our newreno_flags
624 	 * so that Abe behavior is also applied.
625 	 */
626 	((struct newreno *)tp->ccv->cc_data)->newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
627 	opt.name = CC_NEWRENO_BETA_ECN;
628 	opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
629 	error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
630 	if (error) {
631 		goto out;
632 	}
633 	/* Save off the original values for restoral */
634 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
635 out:
636 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
637 		union tcp_log_stackspecific log;
638 		struct timeval tv;
639 
640 		ptr = ((struct newreno *)tp->ccv->cc_data);
641 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
642 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
643 		if (ptr) {
644 			log.u_bbr.flex1 = ptr->beta;
645 			log.u_bbr.flex2 = ptr->beta_ecn;
646 			log.u_bbr.flex3 = ptr->newreno_flags;
647 		}
648 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
649 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
650 		log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
651 		log.u_bbr.flex7 = rack->gp_ready;
652 		log.u_bbr.flex7 <<= 1;
653 		log.u_bbr.flex7 |= rack->use_fixed_rate;
654 		log.u_bbr.flex7 <<= 1;
655 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
656 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
657 		log.u_bbr.flex8 = 3;
658 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
659 			       0, &log, false, NULL, NULL, 0, &tv);
660 	}
661 }
662 
663 static void
664 rack_undo_cc_pacing(struct tcp_rack *rack)
665 {
666 	struct newreno old, *ptr;
667 	struct tcpcb *tp;
668 
669 	if (rack->rc_pacing_cc_set == 0)
670 		return;
671 	tp = rack->rc_tp;
672 	rack->rc_pacing_cc_set = 0;
673 	if (tp->cc_algo == NULL)
674 		/* Tcb is leaving */
675 		return;
676 	if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
677 		/* Not new-reno nothing to do! */
678 		return;
679 	}
680 	ptr = ((struct newreno *)tp->ccv->cc_data);
681 	if (ptr == NULL) {
682 		/*
683 		 * This happens at rack_fini() if the
684 		 * cc module gets freed on us. In that
685 		 * case we loose our "new" settings but
686 		 * thats ok, since the tcb is going away anyway.
687 		 */
688 		return;
689 	}
690 	/* Grab out our set values */
691 	memcpy(&old, ptr, sizeof(struct newreno));
692 	/* Copy back in the original values */
693 	memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno));
694 	/* Now save back the values we had set in (for when pacing is restored) */
695 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
696 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
697 		union tcp_log_stackspecific log;
698 		struct timeval tv;
699 
700 		ptr = ((struct newreno *)tp->ccv->cc_data);
701 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
702 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
703 		log.u_bbr.flex1 = ptr->beta;
704 		log.u_bbr.flex2 = ptr->beta_ecn;
705 		log.u_bbr.flex3 = ptr->newreno_flags;
706 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
707 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
708 		log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
709 		log.u_bbr.flex7 = rack->gp_ready;
710 		log.u_bbr.flex7 <<= 1;
711 		log.u_bbr.flex7 |= rack->use_fixed_rate;
712 		log.u_bbr.flex7 <<= 1;
713 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
714 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
715 		log.u_bbr.flex8 = 4;
716 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
717 			       0, &log, false, NULL, NULL, 0, &tv);
718 	}
719 }
720 
721 #ifdef NETFLIX_PEAKRATE
722 static inline void
723 rack_update_peakrate_thr(struct tcpcb *tp)
724 {
725 	/* Keep in mind that t_maxpeakrate is in B/s. */
726 	uint64_t peak;
727 	peak = uqmax((tp->t_maxseg * 2),
728 		     (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
729 	tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
730 }
731 #endif
732 
733 static int
734 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
735 {
736 	uint32_t stat;
737 	int32_t error;
738 	int i;
739 
740 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
741 	if (error || req->newptr == NULL)
742 		return error;
743 
744 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
745 	if (error)
746 		return (error);
747 	if (stat == 1) {
748 #ifdef INVARIANTS
749 		printf("Clearing RACK counters\n");
750 #endif
751 		counter_u64_zero(rack_badfr);
752 		counter_u64_zero(rack_badfr_bytes);
753 		counter_u64_zero(rack_rtm_prr_retran);
754 		counter_u64_zero(rack_rtm_prr_newdata);
755 		counter_u64_zero(rack_timestamp_mismatch);
756 		counter_u64_zero(rack_reorder_seen);
757 		counter_u64_zero(rack_tlp_tot);
758 		counter_u64_zero(rack_tlp_newdata);
759 		counter_u64_zero(rack_tlp_retran);
760 		counter_u64_zero(rack_tlp_retran_bytes);
761 		counter_u64_zero(rack_tlp_retran_fail);
762 		counter_u64_zero(rack_to_tot);
763 		counter_u64_zero(rack_to_arm_rack);
764 		counter_u64_zero(rack_to_arm_tlp);
765 		counter_u64_zero(rack_paced_segments);
766 		counter_u64_zero(rack_calc_zero);
767 		counter_u64_zero(rack_calc_nonzero);
768 		counter_u64_zero(rack_unpaced_segments);
769 		counter_u64_zero(rack_saw_enobuf);
770 		counter_u64_zero(rack_saw_enobuf_hw);
771 		counter_u64_zero(rack_saw_enetunreach);
772 		counter_u64_zero(rack_per_timer_hole);
773 		counter_u64_zero(rack_large_ackcmp);
774 		counter_u64_zero(rack_small_ackcmp);
775 #ifdef INVARIANTS
776 		counter_u64_zero(rack_adjust_map_bw);
777 #endif
778 		counter_u64_zero(rack_to_alloc_hard);
779 		counter_u64_zero(rack_to_alloc_emerg);
780 		counter_u64_zero(rack_sack_proc_all);
781 		counter_u64_zero(rack_fto_send);
782 		counter_u64_zero(rack_fto_rsm_send);
783 		counter_u64_zero(rack_extended_rfo);
784 		counter_u64_zero(rack_hw_pace_init_fail);
785 		counter_u64_zero(rack_hw_pace_lost);
786 		counter_u64_zero(rack_sbsndptr_wrong);
787 		counter_u64_zero(rack_sbsndptr_right);
788 		counter_u64_zero(rack_non_fto_send);
789 		counter_u64_zero(rack_nfto_resend);
790 		counter_u64_zero(rack_sack_proc_short);
791 		counter_u64_zero(rack_sack_proc_restart);
792 		counter_u64_zero(rack_to_alloc);
793 		counter_u64_zero(rack_to_alloc_limited);
794 		counter_u64_zero(rack_alloc_limited_conns);
795 		counter_u64_zero(rack_split_limited);
796 		for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
797 			counter_u64_zero(rack_proc_comp_ack[i]);
798 		}
799 		counter_u64_zero(rack_multi_single_eq);
800 		counter_u64_zero(rack_proc_non_comp_ack);
801 		counter_u64_zero(rack_find_high);
802 		counter_u64_zero(rack_sack_attacks_detected);
803 		counter_u64_zero(rack_sack_attacks_reversed);
804 		counter_u64_zero(rack_sack_used_next_merge);
805 		counter_u64_zero(rack_sack_used_prev_merge);
806 		counter_u64_zero(rack_sack_splits);
807 		counter_u64_zero(rack_sack_skipped_acked);
808 		counter_u64_zero(rack_ack_total);
809 		counter_u64_zero(rack_express_sack);
810 		counter_u64_zero(rack_sack_total);
811 		counter_u64_zero(rack_move_none);
812 		counter_u64_zero(rack_move_some);
813 		counter_u64_zero(rack_used_tlpmethod);
814 		counter_u64_zero(rack_used_tlpmethod2);
815 		counter_u64_zero(rack_enter_tlp_calc);
816 		counter_u64_zero(rack_progress_drops);
817 		counter_u64_zero(rack_tlp_does_nada);
818 		counter_u64_zero(rack_try_scwnd);
819 		counter_u64_zero(rack_collapsed_win);
820 	}
821 	rack_clear_counter = 0;
822 	return (0);
823 }
824 
825 static void
826 rack_init_sysctls(void)
827 {
828 	int i;
829 	struct sysctl_oid *rack_counters;
830 	struct sysctl_oid *rack_attack;
831 	struct sysctl_oid *rack_pacing;
832 	struct sysctl_oid *rack_timely;
833 	struct sysctl_oid *rack_timers;
834 	struct sysctl_oid *rack_tlp;
835 	struct sysctl_oid *rack_misc;
836 	struct sysctl_oid *rack_features;
837 	struct sysctl_oid *rack_measure;
838 	struct sysctl_oid *rack_probertt;
839 	struct sysctl_oid *rack_hw_pacing;
840 
841 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
842 	    SYSCTL_CHILDREN(rack_sysctl_root),
843 	    OID_AUTO,
844 	    "sack_attack",
845 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
846 	    "Rack Sack Attack Counters and Controls");
847 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
848 	    SYSCTL_CHILDREN(rack_sysctl_root),
849 	    OID_AUTO,
850 	    "stats",
851 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
852 	    "Rack Counters");
853 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
854 	    SYSCTL_CHILDREN(rack_sysctl_root),
855 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
856 	    &rack_rate_sample_method , USE_RTT_LOW,
857 	    "What method should we use for rate sampling 0=high, 1=low ");
858 	/* Probe rtt related controls */
859 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
860 	    SYSCTL_CHILDREN(rack_sysctl_root),
861 	    OID_AUTO,
862 	    "probertt",
863 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
864 	    "ProbeRTT related Controls");
865 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
866 	    SYSCTL_CHILDREN(rack_probertt),
867 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
868 	    &rack_atexit_prtt_hbp, 130,
869 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
870 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
871 	    SYSCTL_CHILDREN(rack_probertt),
872 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
873 	    &rack_atexit_prtt, 130,
874 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
875 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
876 	    SYSCTL_CHILDREN(rack_probertt),
877 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
878 	    &rack_per_of_gp_probertt, 60,
879 	    "What percentage of goodput do we pace at in probertt");
880 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
881 	    SYSCTL_CHILDREN(rack_probertt),
882 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
883 	    &rack_per_of_gp_probertt_reduce, 10,
884 	    "What percentage of goodput do we reduce every gp_srtt");
885 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
886 	    SYSCTL_CHILDREN(rack_probertt),
887 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
888 	    &rack_per_of_gp_lowthresh, 40,
889 	    "What percentage of goodput do we allow the multiplier to fall to");
890 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
891 	    SYSCTL_CHILDREN(rack_probertt),
892 	    OID_AUTO, "time_between", CTLFLAG_RW,
893 	    & rack_time_between_probertt, 96000000,
894 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
895 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
896 	    SYSCTL_CHILDREN(rack_probertt),
897 	    OID_AUTO, "safety", CTLFLAG_RW,
898 	    &rack_probe_rtt_safety_val, 2000000,
899 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
900 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
901 	    SYSCTL_CHILDREN(rack_probertt),
902 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
903 	    &rack_probe_rtt_sets_cwnd, 0,
904 	    "Do we set the cwnd too (if always_lower is on)");
905 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
906 	    SYSCTL_CHILDREN(rack_probertt),
907 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
908 	    &rack_max_drain_wait, 2,
909 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
910 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
911 	    SYSCTL_CHILDREN(rack_probertt),
912 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
913 	    &rack_must_drain, 1,
914 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
915 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
916 	    SYSCTL_CHILDREN(rack_probertt),
917 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
918 	    &rack_probertt_use_min_rtt_entry, 1,
919 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
920 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
921 	    SYSCTL_CHILDREN(rack_probertt),
922 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
923 	    &rack_probertt_use_min_rtt_exit, 0,
924 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
925 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
926 	    SYSCTL_CHILDREN(rack_probertt),
927 	    OID_AUTO, "length_div", CTLFLAG_RW,
928 	    &rack_probertt_gpsrtt_cnt_div, 0,
929 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
930 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
931 	    SYSCTL_CHILDREN(rack_probertt),
932 	    OID_AUTO, "length_mul", CTLFLAG_RW,
933 	    &rack_probertt_gpsrtt_cnt_mul, 0,
934 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
935 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
936 	    SYSCTL_CHILDREN(rack_probertt),
937 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
938 	    &rack_min_probertt_hold, 200000,
939 	    "What is the minimum time we hold probertt at target");
940 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
941 	    SYSCTL_CHILDREN(rack_probertt),
942 	    OID_AUTO, "filter_life", CTLFLAG_RW,
943 	    &rack_probertt_filter_life, 10000000,
944 	    "What is the time for the filters life in useconds");
945 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
946 	    SYSCTL_CHILDREN(rack_probertt),
947 	    OID_AUTO, "lower_within", CTLFLAG_RW,
948 	    &rack_probertt_lower_within, 10,
949 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
950 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
951 	    SYSCTL_CHILDREN(rack_probertt),
952 	    OID_AUTO, "must_move", CTLFLAG_RW,
953 	    &rack_min_rtt_movement, 250,
954 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
955 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
956 	    SYSCTL_CHILDREN(rack_probertt),
957 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
958 	    &rack_probertt_clear_is, 1,
959 	    "Do we clear I/S counts on exiting probe-rtt");
960 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
961 	    SYSCTL_CHILDREN(rack_probertt),
962 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
963 	    &rack_max_drain_hbp, 1,
964 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
965 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
966 	    SYSCTL_CHILDREN(rack_probertt),
967 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
968 	    &rack_hbp_thresh, 3,
969 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
970 	/* Pacing related sysctls */
971 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
972 	    SYSCTL_CHILDREN(rack_sysctl_root),
973 	    OID_AUTO,
974 	    "pacing",
975 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
976 	    "Pacing related Controls");
977 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
978 	    SYSCTL_CHILDREN(rack_pacing),
979 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
980 	    &rack_max_per_above, 30,
981 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
982 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
983 	    SYSCTL_CHILDREN(rack_pacing),
984 	    OID_AUTO, "pace_to_one", CTLFLAG_RW,
985 	    &rack_pace_one_seg, 0,
986 	    "Do we allow low b/w pacing of 1MSS instead of two");
987 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
988 	    SYSCTL_CHILDREN(rack_pacing),
989 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
990 	    &rack_limit_time_with_srtt, 0,
991 	    "Do we limit pacing time based on srtt");
992 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
993 	    SYSCTL_CHILDREN(rack_pacing),
994 	    OID_AUTO, "init_win", CTLFLAG_RW,
995 	    &rack_default_init_window, 0,
996 	    "Do we have a rack initial window 0 = system default");
997 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
998 	    SYSCTL_CHILDREN(rack_pacing),
999 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1000 	    &rack_per_of_gp_ss, 250,
1001 	    "If non zero, what percentage of goodput to pace at in slow start");
1002 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1003 	    SYSCTL_CHILDREN(rack_pacing),
1004 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1005 	    &rack_per_of_gp_ca, 150,
1006 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
1007 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1008 	    SYSCTL_CHILDREN(rack_pacing),
1009 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1010 	    &rack_per_of_gp_rec, 200,
1011 	    "If non zero, what percentage of goodput to pace at in recovery");
1012 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1013 	    SYSCTL_CHILDREN(rack_pacing),
1014 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1015 	    &rack_hptsi_segments, 40,
1016 	    "What size is the max for TSO segments in pacing and burst mitigation");
1017 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1018 	    SYSCTL_CHILDREN(rack_pacing),
1019 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
1020 	    &rack_slot_reduction, 4,
1021 	    "When doing only burst mitigation what is the reduce divisor");
1022 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1023 	    SYSCTL_CHILDREN(rack_sysctl_root),
1024 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
1025 	    &rack_pace_every_seg, 0,
1026 	    "If set we use pacing, if clear we use only the original burst mitigation");
1027 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1028 	    SYSCTL_CHILDREN(rack_pacing),
1029 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1030 	    &rack_bw_rate_cap, 0,
1031 	    "If set we apply this value to the absolute rate cap used by pacing");
1032 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1033 	    SYSCTL_CHILDREN(rack_sysctl_root),
1034 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1035 	    &rack_req_measurements, 1,
1036 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1037 	/* Hardware pacing */
1038 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1039 	    SYSCTL_CHILDREN(rack_sysctl_root),
1040 	    OID_AUTO,
1041 	    "hdwr_pacing",
1042 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1043 	    "Pacing related Controls");
1044 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1045 	    SYSCTL_CHILDREN(rack_hw_pacing),
1046 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1047 	    &rack_hw_rwnd_factor, 2,
1048 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1049 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1050 	    SYSCTL_CHILDREN(rack_hw_pacing),
1051 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1052 	    &rack_enobuf_hw_boost_mult, 2,
1053 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1054 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1055 	    SYSCTL_CHILDREN(rack_hw_pacing),
1056 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1057 	    &rack_enobuf_hw_max, 2,
1058 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1059 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1060 	    SYSCTL_CHILDREN(rack_hw_pacing),
1061 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1062 	    &rack_enobuf_hw_min, 2,
1063 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1064 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1065 	    SYSCTL_CHILDREN(rack_hw_pacing),
1066 	    OID_AUTO, "enable", CTLFLAG_RW,
1067 	    &rack_enable_hw_pacing, 0,
1068 	    "Should RACK attempt to use hw pacing?");
1069 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1070 	    SYSCTL_CHILDREN(rack_hw_pacing),
1071 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1072 	    &rack_hw_rate_caps, 1,
1073 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1074 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1075 	    SYSCTL_CHILDREN(rack_hw_pacing),
1076 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1077 	    &rack_hw_rate_min, 0,
1078 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1079 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1080 	    SYSCTL_CHILDREN(rack_hw_pacing),
1081 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1082 	    &rack_hw_rate_to_low, 0,
1083 	    "If we fall below this rate, dis-engage hw pacing?");
1084 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1085 	    SYSCTL_CHILDREN(rack_hw_pacing),
1086 	    OID_AUTO, "up_only", CTLFLAG_RW,
1087 	    &rack_hw_up_only, 1,
1088 	    "Do we allow hw pacing to lower the rate selected?");
1089 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1090 	    SYSCTL_CHILDREN(rack_hw_pacing),
1091 	    OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1092 	    &rack_hw_pace_extra_slots, 2,
1093 	    "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1094 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1095 	    SYSCTL_CHILDREN(rack_sysctl_root),
1096 	    OID_AUTO,
1097 	    "timely",
1098 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1099 	    "Rack Timely RTT Controls");
1100 	/* Timely based GP dynmics */
1101 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1102 	    SYSCTL_CHILDREN(rack_timely),
1103 	    OID_AUTO, "upper", CTLFLAG_RW,
1104 	    &rack_gp_per_bw_mul_up, 2,
1105 	    "Rack timely upper range for equal b/w (in percentage)");
1106 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1107 	    SYSCTL_CHILDREN(rack_timely),
1108 	    OID_AUTO, "lower", CTLFLAG_RW,
1109 	    &rack_gp_per_bw_mul_down, 4,
1110 	    "Rack timely lower range for equal b/w (in percentage)");
1111 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1112 	    SYSCTL_CHILDREN(rack_timely),
1113 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1114 	    &rack_gp_rtt_maxmul, 3,
1115 	    "Rack timely multipler of lowest rtt for rtt_max");
1116 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1117 	    SYSCTL_CHILDREN(rack_timely),
1118 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1119 	    &rack_gp_rtt_mindiv, 4,
1120 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1121 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1122 	    SYSCTL_CHILDREN(rack_timely),
1123 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1124 	    &rack_gp_rtt_minmul, 1,
1125 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1126 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1127 	    SYSCTL_CHILDREN(rack_timely),
1128 	    OID_AUTO, "decrease", CTLFLAG_RW,
1129 	    &rack_gp_decrease_per, 20,
1130 	    "Rack timely decrease percentage of our GP multiplication factor");
1131 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1132 	    SYSCTL_CHILDREN(rack_timely),
1133 	    OID_AUTO, "increase", CTLFLAG_RW,
1134 	    &rack_gp_increase_per, 2,
1135 	    "Rack timely increase perentage of our GP multiplication factor");
1136 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1137 	    SYSCTL_CHILDREN(rack_timely),
1138 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1139 	    &rack_per_lower_bound, 50,
1140 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1141 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1142 	    SYSCTL_CHILDREN(rack_timely),
1143 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1144 	    &rack_per_upper_bound_ss, 0,
1145 	    "Rack timely higest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1146 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1147 	    SYSCTL_CHILDREN(rack_timely),
1148 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1149 	    &rack_per_upper_bound_ca, 0,
1150 	    "Rack timely higest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1151 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1152 	    SYSCTL_CHILDREN(rack_timely),
1153 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1154 	    &rack_do_dyn_mul, 0,
1155 	    "Rack timely do we enable dynmaic timely goodput by default");
1156 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1157 	    SYSCTL_CHILDREN(rack_timely),
1158 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1159 	    &rack_gp_no_rec_chg, 1,
1160 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1161 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1162 	    SYSCTL_CHILDREN(rack_timely),
1163 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1164 	    &rack_timely_dec_clear, 6,
1165 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1166 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1167 	    SYSCTL_CHILDREN(rack_timely),
1168 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1169 	    &rack_timely_max_push_rise, 3,
1170 	    "Rack timely how many times do we push up with b/w increase");
1171 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1172 	    SYSCTL_CHILDREN(rack_timely),
1173 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1174 	    &rack_timely_max_push_drop, 3,
1175 	    "Rack timely how many times do we push back on b/w decent");
1176 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1177 	    SYSCTL_CHILDREN(rack_timely),
1178 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1179 	    &rack_timely_min_segs, 4,
1180 	    "Rack timely when setting the cwnd what is the min num segments");
1181 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1182 	    SYSCTL_CHILDREN(rack_timely),
1183 	    OID_AUTO, "noback_max", CTLFLAG_RW,
1184 	    &rack_use_max_for_nobackoff, 0,
1185 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1186 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1187 	    SYSCTL_CHILDREN(rack_timely),
1188 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1189 	    &rack_timely_int_timely_only, 0,
1190 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1191 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1192 	    SYSCTL_CHILDREN(rack_timely),
1193 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1194 	    &rack_timely_no_stopping, 0,
1195 	    "Rack timely don't stop increase");
1196 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1197 	    SYSCTL_CHILDREN(rack_timely),
1198 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1199 	    &rack_down_raise_thresh, 100,
1200 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1201 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1202 	    SYSCTL_CHILDREN(rack_timely),
1203 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1204 	    &rack_req_segs, 1,
1205 	    "Bottom dragging if not these many segments outstanding and room");
1206 
1207 	/* TLP and Rack related parameters */
1208 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1209 	    SYSCTL_CHILDREN(rack_sysctl_root),
1210 	    OID_AUTO,
1211 	    "tlp",
1212 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1213 	    "TLP and Rack related Controls");
1214 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1215 	    SYSCTL_CHILDREN(rack_tlp),
1216 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1217 	    &use_rack_rr, 1,
1218 	    "Do we use Rack Rapid Recovery");
1219 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1220 	    SYSCTL_CHILDREN(rack_tlp),
1221 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1222 	    &rack_max_abc_post_recovery, 2,
1223 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1224 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1225 	    SYSCTL_CHILDREN(rack_tlp),
1226 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1227 	    &rack_non_rxt_use_cr, 0,
1228 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1229 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1230 	    SYSCTL_CHILDREN(rack_tlp),
1231 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1232 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1233 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1234 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1235 	    SYSCTL_CHILDREN(rack_tlp),
1236 	    OID_AUTO, "limit", CTLFLAG_RW,
1237 	    &rack_tlp_limit, 2,
1238 	    "How many TLP's can be sent without sending new data");
1239 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1240 	    SYSCTL_CHILDREN(rack_tlp),
1241 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1242 	    &rack_tlp_use_greater, 1,
1243 	    "Should we use the rack_rtt time if its greater than srtt");
1244 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1245 	    SYSCTL_CHILDREN(rack_tlp),
1246 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1247 	    &rack_tlp_min, 10000,
1248 	    "TLP minimum timeout per the specification (in microseconds)");
1249 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1250 	    SYSCTL_CHILDREN(rack_tlp),
1251 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1252 	    &rack_always_send_oldest, 0,
1253 	    "Should we always send the oldest TLP and RACK-TLP");
1254 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1255 	    SYSCTL_CHILDREN(rack_tlp),
1256 	    OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1257 	    &rack_limited_retran, 0,
1258 	    "How many times can a rack timeout drive out sends");
1259 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1260 	    SYSCTL_CHILDREN(rack_tlp),
1261 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1262 	    &rack_lower_cwnd_at_tlp, 0,
1263 	    "When a TLP completes a retran should we enter recovery");
1264 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1265 	    SYSCTL_CHILDREN(rack_tlp),
1266 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1267 	    &rack_reorder_thresh, 2,
1268 	    "What factor for rack will be added when seeing reordering (shift right)");
1269 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1270 	    SYSCTL_CHILDREN(rack_tlp),
1271 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1272 	    &rack_tlp_thresh, 1,
1273 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1274 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1275 	    SYSCTL_CHILDREN(rack_tlp),
1276 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1277 	    &rack_reorder_fade, 60000000,
1278 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1279 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1280 	    SYSCTL_CHILDREN(rack_tlp),
1281 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1282 	    &rack_pkt_delay, 1000,
1283 	    "Extra RACK time (in microseconds) besides reordering thresh");
1284 
1285 	/* Timer related controls */
1286 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1287 	    SYSCTL_CHILDREN(rack_sysctl_root),
1288 	    OID_AUTO,
1289 	    "timers",
1290 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1291 	    "Timer related controls");
1292 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1293 	    SYSCTL_CHILDREN(rack_timers),
1294 	    OID_AUTO, "persmin", CTLFLAG_RW,
1295 	    &rack_persist_min, 250000,
1296 	    "What is the minimum time in microseconds between persists");
1297 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1298 	    SYSCTL_CHILDREN(rack_timers),
1299 	    OID_AUTO, "persmax", CTLFLAG_RW,
1300 	    &rack_persist_max, 2000000,
1301 	    "What is the largest delay in microseconds between persists");
1302 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1303 	    SYSCTL_CHILDREN(rack_timers),
1304 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1305 	    &rack_delayed_ack_time, 40000,
1306 	    "Delayed ack time (40ms in microseconds)");
1307 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1308 	    SYSCTL_CHILDREN(rack_timers),
1309 	    OID_AUTO, "minrto", CTLFLAG_RW,
1310 	    &rack_rto_min, 30000,
1311 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1312 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1313 	    SYSCTL_CHILDREN(rack_timers),
1314 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1315 	    &rack_rto_max, 4000000,
1316 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1317 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1318 	    SYSCTL_CHILDREN(rack_timers),
1319 	    OID_AUTO, "minto", CTLFLAG_RW,
1320 	    &rack_min_to, 1000,
1321 	    "Minimum rack timeout in microseconds");
1322 	/* Measure controls */
1323 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1324 	    SYSCTL_CHILDREN(rack_sysctl_root),
1325 	    OID_AUTO,
1326 	    "measure",
1327 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1328 	    "Measure related controls");
1329 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1330 	    SYSCTL_CHILDREN(rack_measure),
1331 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1332 	    &rack_wma_divisor, 8,
1333 	    "When doing b/w calculation what is the  divisor for the WMA");
1334 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1335 	    SYSCTL_CHILDREN(rack_measure),
1336 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1337 	    &rack_cwnd_block_ends_measure, 0,
1338 	    "Does a cwnd just-return end the measurement window (app limited)");
1339 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1340 	    SYSCTL_CHILDREN(rack_measure),
1341 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1342 	    &rack_rwnd_block_ends_measure, 0,
1343 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1344 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1345 	    SYSCTL_CHILDREN(rack_measure),
1346 	    OID_AUTO, "min_target", CTLFLAG_RW,
1347 	    &rack_def_data_window, 20,
1348 	    "What is the minimum target window (in mss) for a GP measurements");
1349 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1350 	    SYSCTL_CHILDREN(rack_measure),
1351 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1352 	    &rack_goal_bdp, 2,
1353 	    "What is the goal BDP to measure");
1354 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1355 	    SYSCTL_CHILDREN(rack_measure),
1356 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1357 	    &rack_min_srtts, 1,
1358 	    "What is the goal BDP to measure");
1359 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1360 	    SYSCTL_CHILDREN(rack_measure),
1361 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1362 	    &rack_min_measure_usec, 0,
1363 	    "What is the Minimum time time for a measurement if 0, this is off");
1364 	/* Features */
1365 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1366 	    SYSCTL_CHILDREN(rack_sysctl_root),
1367 	    OID_AUTO,
1368 	    "features",
1369 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1370 	    "Feature controls");
1371 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1372 	    SYSCTL_CHILDREN(rack_features),
1373 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1374 	    &rack_use_cmp_acks, 1,
1375 	    "Should RACK have LRO send compressed acks");
1376 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1377 	    SYSCTL_CHILDREN(rack_features),
1378 	    OID_AUTO, "fsb", CTLFLAG_RW,
1379 	    &rack_use_fsb, 1,
1380 	    "Should RACK use the fast send block?");
1381 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1382 	    SYSCTL_CHILDREN(rack_features),
1383 	    OID_AUTO, "rfo", CTLFLAG_RW,
1384 	    &rack_use_rfo, 1,
1385 	    "Should RACK use rack_fast_output()?");
1386 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1387 	    SYSCTL_CHILDREN(rack_features),
1388 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1389 	    &rack_use_rsm_rfo, 1,
1390 	    "Should RACK use rack_fast_rsm_output()?");
1391 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1392 	    SYSCTL_CHILDREN(rack_features),
1393 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1394 	    &rack_enable_mqueue_for_nonpaced, 0,
1395 	    "Should RACK use mbuf queuing for non-paced connections");
1396 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1397 	    SYSCTL_CHILDREN(rack_features),
1398 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1399 	    &rack_do_hystart, 0,
1400 	    "Should RACK enable HyStart++ on connections?");
1401 	/* Misc rack controls */
1402 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1403 	    SYSCTL_CHILDREN(rack_sysctl_root),
1404 	    OID_AUTO,
1405 	    "misc",
1406 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1407 	    "Misc related controls");
1408 #ifdef TCP_ACCOUNTING
1409 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1410 	    SYSCTL_CHILDREN(rack_misc),
1411 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1412 	    &rack_tcp_accounting, 0,
1413 	    "Should we turn on TCP accounting for all rack sessions?");
1414 #endif
1415 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1416 	    SYSCTL_CHILDREN(rack_misc),
1417 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1418 	    &rack_dsack_std_based, 3,
1419 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1420 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1421 	    SYSCTL_CHILDREN(rack_misc),
1422 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1423 	    &rack_prr_addbackmax, 2,
1424 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1425 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1426 	    SYSCTL_CHILDREN(rack_misc),
1427 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1428 	    &rack_stats_gets_ms_rtt, 1,
1429 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1430 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1431 	    SYSCTL_CHILDREN(rack_misc),
1432 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1433 	    &rack_client_low_buf, 0,
1434 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1435 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1436 	    SYSCTL_CHILDREN(rack_misc),
1437 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1438 	    &rack_def_profile, 0,
1439 	    "Should RACK use a default profile (0=no, num == profile num)?");
1440 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1441 	    SYSCTL_CHILDREN(rack_misc),
1442 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1443 	    &rack_enable_shared_cwnd, 1,
1444 	    "Should RACK try to use the shared cwnd on connections where allowed");
1445 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1446 	    SYSCTL_CHILDREN(rack_misc),
1447 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1448 	    &rack_limits_scwnd, 1,
1449 	    "Should RACK place low end time limits on the shared cwnd feature");
1450 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1451 	    SYSCTL_CHILDREN(rack_misc),
1452 	    OID_AUTO, "iMac_dack", CTLFLAG_RW,
1453 	    &rack_use_imac_dack, 0,
1454 	    "Should RACK try to emulate iMac delayed ack");
1455 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1456 	    SYSCTL_CHILDREN(rack_misc),
1457 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1458 	    &rack_disable_prr, 0,
1459 	    "Should RACK not use prr and only pace (must have pacing on)");
1460 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1461 	    SYSCTL_CHILDREN(rack_misc),
1462 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1463 	    &rack_verbose_logging, 0,
1464 	    "Should RACK black box logging be verbose");
1465 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1466 	    SYSCTL_CHILDREN(rack_misc),
1467 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1468 	    &rack_ignore_data_after_close, 1,
1469 	    "Do we hold off sending a RST until all pending data is ack'd");
1470 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1471 	    SYSCTL_CHILDREN(rack_misc),
1472 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1473 	    &rack_sack_not_required, 1,
1474 	    "Do we allow rack to run on connections not supporting SACK");
1475 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1476 	    SYSCTL_CHILDREN(rack_misc),
1477 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1478 	    &rack_send_a_lot_in_prr, 1,
1479 	    "Send a lot in prr");
1480 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1481 	    SYSCTL_CHILDREN(rack_misc),
1482 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1483 	    &rack_autosndbuf_inc, 20,
1484 	    "What percentage should rack scale up its snd buffer by?");
1485 	/* Sack Attacker detection stuff */
1486 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1487 	    SYSCTL_CHILDREN(rack_attack),
1488 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1489 	    &rack_highest_sack_thresh_seen, 0,
1490 	    "Highest sack to ack ratio seen");
1491 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1492 	    SYSCTL_CHILDREN(rack_attack),
1493 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1494 	    &rack_highest_move_thresh_seen, 0,
1495 	    "Highest move to non-move ratio seen");
1496 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1497 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1498 	    SYSCTL_CHILDREN(rack_attack),
1499 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1500 	    &rack_ack_total,
1501 	    "Total number of Ack's");
1502 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1503 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1504 	    SYSCTL_CHILDREN(rack_attack),
1505 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1506 	    &rack_express_sack,
1507 	    "Total expresss number of Sack's");
1508 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1509 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1510 	    SYSCTL_CHILDREN(rack_attack),
1511 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1512 	    &rack_sack_total,
1513 	    "Total number of SACKs");
1514 	rack_move_none = counter_u64_alloc(M_WAITOK);
1515 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1516 	    SYSCTL_CHILDREN(rack_attack),
1517 	    OID_AUTO, "move_none", CTLFLAG_RD,
1518 	    &rack_move_none,
1519 	    "Total number of SACK index reuse of postions under threshold");
1520 	rack_move_some = counter_u64_alloc(M_WAITOK);
1521 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1522 	    SYSCTL_CHILDREN(rack_attack),
1523 	    OID_AUTO, "move_some", CTLFLAG_RD,
1524 	    &rack_move_some,
1525 	    "Total number of SACK index reuse of postions over threshold");
1526 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1527 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1528 	    SYSCTL_CHILDREN(rack_attack),
1529 	    OID_AUTO, "attacks", CTLFLAG_RD,
1530 	    &rack_sack_attacks_detected,
1531 	    "Total number of SACK attackers that had sack disabled");
1532 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1533 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1534 	    SYSCTL_CHILDREN(rack_attack),
1535 	    OID_AUTO, "reversed", CTLFLAG_RD,
1536 	    &rack_sack_attacks_reversed,
1537 	    "Total number of SACK attackers that were later determined false positive");
1538 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1539 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1540 	    SYSCTL_CHILDREN(rack_attack),
1541 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1542 	    &rack_sack_used_next_merge,
1543 	    "Total number of times we used the next merge");
1544 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1545 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1546 	    SYSCTL_CHILDREN(rack_attack),
1547 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1548 	    &rack_sack_used_prev_merge,
1549 	    "Total number of times we used the prev merge");
1550 	/* Counters */
1551 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1552 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1553 	    SYSCTL_CHILDREN(rack_counters),
1554 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1555 	    &rack_fto_send, "Total number of rack_fast_output sends");
1556 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1557 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1558 	    SYSCTL_CHILDREN(rack_counters),
1559 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1560 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1561 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1562 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1563 	    SYSCTL_CHILDREN(rack_counters),
1564 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1565 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1566 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1567 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1568 	    SYSCTL_CHILDREN(rack_counters),
1569 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1570 	    &rack_non_fto_send, "Total number of rack_output first sends");
1571 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1572 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1573 	    SYSCTL_CHILDREN(rack_counters),
1574 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1575 	    &rack_extended_rfo, "Total number of times we extended rfo");
1576 
1577 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1578 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1579 	    SYSCTL_CHILDREN(rack_counters),
1580 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1581 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1582 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1583 
1584 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1585 	    SYSCTL_CHILDREN(rack_counters),
1586 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1587 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1588 	rack_badfr = counter_u64_alloc(M_WAITOK);
1589 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1590 	    SYSCTL_CHILDREN(rack_counters),
1591 	    OID_AUTO, "badfr", CTLFLAG_RD,
1592 	    &rack_badfr, "Total number of bad FRs");
1593 	rack_badfr_bytes = counter_u64_alloc(M_WAITOK);
1594 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1595 	    SYSCTL_CHILDREN(rack_counters),
1596 	    OID_AUTO, "badfr_bytes", CTLFLAG_RD,
1597 	    &rack_badfr_bytes, "Total number of bad FRs");
1598 	rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK);
1599 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1600 	    SYSCTL_CHILDREN(rack_counters),
1601 	    OID_AUTO, "prrsndret", CTLFLAG_RD,
1602 	    &rack_rtm_prr_retran,
1603 	    "Total number of prr based retransmits");
1604 	rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK);
1605 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1606 	    SYSCTL_CHILDREN(rack_counters),
1607 	    OID_AUTO, "prrsndnew", CTLFLAG_RD,
1608 	    &rack_rtm_prr_newdata,
1609 	    "Total number of prr based new transmits");
1610 	rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK);
1611 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1612 	    SYSCTL_CHILDREN(rack_counters),
1613 	    OID_AUTO, "tsnf", CTLFLAG_RD,
1614 	    &rack_timestamp_mismatch,
1615 	    "Total number of timestamps that we could not find the reported ts");
1616 	rack_find_high = counter_u64_alloc(M_WAITOK);
1617 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1618 	    SYSCTL_CHILDREN(rack_counters),
1619 	    OID_AUTO, "findhigh", CTLFLAG_RD,
1620 	    &rack_find_high,
1621 	    "Total number of FIN causing find-high");
1622 	rack_reorder_seen = counter_u64_alloc(M_WAITOK);
1623 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1624 	    SYSCTL_CHILDREN(rack_counters),
1625 	    OID_AUTO, "reordering", CTLFLAG_RD,
1626 	    &rack_reorder_seen,
1627 	    "Total number of times we added delay due to reordering");
1628 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1629 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1630 	    SYSCTL_CHILDREN(rack_counters),
1631 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1632 	    &rack_tlp_tot,
1633 	    "Total number of tail loss probe expirations");
1634 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1635 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1636 	    SYSCTL_CHILDREN(rack_counters),
1637 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1638 	    &rack_tlp_newdata,
1639 	    "Total number of tail loss probe sending new data");
1640 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1641 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1642 	    SYSCTL_CHILDREN(rack_counters),
1643 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1644 	    &rack_tlp_retran,
1645 	    "Total number of tail loss probe sending retransmitted data");
1646 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1647 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1648 	    SYSCTL_CHILDREN(rack_counters),
1649 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1650 	    &rack_tlp_retran_bytes,
1651 	    "Total bytes of tail loss probe sending retransmitted data");
1652 	rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK);
1653 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1654 	    SYSCTL_CHILDREN(rack_counters),
1655 	    OID_AUTO, "tlp_retran_fail", CTLFLAG_RD,
1656 	    &rack_tlp_retran_fail,
1657 	    "Total number of tail loss probe sending retransmitted data that failed (wait for t3)");
1658 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1659 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1660 	    SYSCTL_CHILDREN(rack_counters),
1661 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1662 	    &rack_to_tot,
1663 	    "Total number of times the rack to expired");
1664 	rack_to_arm_rack = counter_u64_alloc(M_WAITOK);
1665 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1666 	    SYSCTL_CHILDREN(rack_counters),
1667 	    OID_AUTO, "arm_rack", CTLFLAG_RD,
1668 	    &rack_to_arm_rack,
1669 	    "Total number of times the rack timer armed");
1670 	rack_to_arm_tlp = counter_u64_alloc(M_WAITOK);
1671 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1672 	    SYSCTL_CHILDREN(rack_counters),
1673 	    OID_AUTO, "arm_tlp", CTLFLAG_RD,
1674 	    &rack_to_arm_tlp,
1675 	    "Total number of times the tlp timer armed");
1676 	rack_calc_zero = counter_u64_alloc(M_WAITOK);
1677 	rack_calc_nonzero = counter_u64_alloc(M_WAITOK);
1678 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1679 	    SYSCTL_CHILDREN(rack_counters),
1680 	    OID_AUTO, "calc_zero", CTLFLAG_RD,
1681 	    &rack_calc_zero,
1682 	    "Total number of times pacing time worked out to zero");
1683 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1684 	    SYSCTL_CHILDREN(rack_counters),
1685 	    OID_AUTO, "calc_nonzero", CTLFLAG_RD,
1686 	    &rack_calc_nonzero,
1687 	    "Total number of times pacing time worked out to non-zero");
1688 	rack_paced_segments = counter_u64_alloc(M_WAITOK);
1689 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1690 	    SYSCTL_CHILDREN(rack_counters),
1691 	    OID_AUTO, "paced", CTLFLAG_RD,
1692 	    &rack_paced_segments,
1693 	    "Total number of times a segment send caused hptsi");
1694 	rack_unpaced_segments = counter_u64_alloc(M_WAITOK);
1695 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1696 	    SYSCTL_CHILDREN(rack_counters),
1697 	    OID_AUTO, "unpaced", CTLFLAG_RD,
1698 	    &rack_unpaced_segments,
1699 	    "Total number of times a segment did not cause hptsi");
1700 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1701 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1702 	    SYSCTL_CHILDREN(rack_counters),
1703 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1704 	    &rack_saw_enobuf,
1705 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1706 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1707 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1708 	    SYSCTL_CHILDREN(rack_counters),
1709 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1710 	    &rack_saw_enobuf_hw,
1711 	    "Total number of times a send returned enobuf for hdwr paced connections");
1712 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1713 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1714 	    SYSCTL_CHILDREN(rack_counters),
1715 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1716 	    &rack_saw_enetunreach,
1717 	    "Total number of times a send received a enetunreachable");
1718 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1719 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1720 	    SYSCTL_CHILDREN(rack_counters),
1721 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1722 	    &rack_hot_alloc,
1723 	    "Total allocations from the top of our list");
1724 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1725 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1726 	    SYSCTL_CHILDREN(rack_counters),
1727 	    OID_AUTO, "allocs", CTLFLAG_RD,
1728 	    &rack_to_alloc,
1729 	    "Total allocations of tracking structures");
1730 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1731 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1732 	    SYSCTL_CHILDREN(rack_counters),
1733 	    OID_AUTO, "allochard", CTLFLAG_RD,
1734 	    &rack_to_alloc_hard,
1735 	    "Total allocations done with sleeping the hard way");
1736 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1737 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1738 	    SYSCTL_CHILDREN(rack_counters),
1739 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1740 	    &rack_to_alloc_emerg,
1741 	    "Total allocations done from emergency cache");
1742 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1743 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1744 	    SYSCTL_CHILDREN(rack_counters),
1745 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1746 	    &rack_to_alloc_limited,
1747 	    "Total allocations dropped due to limit");
1748 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1749 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1750 	    SYSCTL_CHILDREN(rack_counters),
1751 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1752 	    &rack_alloc_limited_conns,
1753 	    "Connections with allocations dropped due to limit");
1754 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1755 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1756 	    SYSCTL_CHILDREN(rack_counters),
1757 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1758 	    &rack_split_limited,
1759 	    "Split allocations dropped due to limit");
1760 
1761 	for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
1762 		char name[32];
1763 		sprintf(name, "cmp_ack_cnt_%d", i);
1764 		rack_proc_comp_ack[i] = counter_u64_alloc(M_WAITOK);
1765 		SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1766 				       SYSCTL_CHILDREN(rack_counters),
1767 				       OID_AUTO, name, CTLFLAG_RD,
1768 				       &rack_proc_comp_ack[i],
1769 				       "Number of compressed acks we processed");
1770 	}
1771 	rack_large_ackcmp = counter_u64_alloc(M_WAITOK);
1772 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1773 	    SYSCTL_CHILDREN(rack_counters),
1774 	    OID_AUTO, "cmp_large_mbufs", CTLFLAG_RD,
1775 	    &rack_large_ackcmp,
1776 	    "Number of TCP connections with large mbuf's for compressed acks");
1777 	rack_small_ackcmp = counter_u64_alloc(M_WAITOK);
1778 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1779 	    SYSCTL_CHILDREN(rack_counters),
1780 	    OID_AUTO, "cmp_small_mbufs", CTLFLAG_RD,
1781 	    &rack_small_ackcmp,
1782 	    "Number of TCP connections with small mbuf's for compressed acks");
1783 #ifdef INVARIANTS
1784 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1785 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1786 	    SYSCTL_CHILDREN(rack_counters),
1787 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1788 	    &rack_adjust_map_bw,
1789 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1790 #endif
1791 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1792 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1793 	    SYSCTL_CHILDREN(rack_counters),
1794 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1795 	    &rack_multi_single_eq,
1796 	    "Number of compressed acks total represented");
1797 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1798 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1799 	    SYSCTL_CHILDREN(rack_counters),
1800 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1801 	    &rack_proc_non_comp_ack,
1802 	    "Number of non compresseds acks that we processed");
1803 
1804 
1805 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1806 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1807 	    SYSCTL_CHILDREN(rack_counters),
1808 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1809 	    &rack_sack_proc_all,
1810 	    "Total times we had to walk whole list for sack processing");
1811 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1812 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1813 	    SYSCTL_CHILDREN(rack_counters),
1814 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1815 	    &rack_sack_proc_restart,
1816 	    "Total times we had to walk whole list due to a restart");
1817 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1818 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1819 	    SYSCTL_CHILDREN(rack_counters),
1820 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1821 	    &rack_sack_proc_short,
1822 	    "Total times we took shortcut for sack processing");
1823 	rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK);
1824 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1825 	    SYSCTL_CHILDREN(rack_counters),
1826 	    OID_AUTO, "tlp_calc_entered", CTLFLAG_RD,
1827 	    &rack_enter_tlp_calc,
1828 	    "Total times we called calc-tlp");
1829 	rack_used_tlpmethod = counter_u64_alloc(M_WAITOK);
1830 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1831 	    SYSCTL_CHILDREN(rack_counters),
1832 	    OID_AUTO, "hit_tlp_method", CTLFLAG_RD,
1833 	    &rack_used_tlpmethod,
1834 	    "Total number of runt sacks");
1835 	rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK);
1836 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1837 	    SYSCTL_CHILDREN(rack_counters),
1838 	    OID_AUTO, "hit_tlp_method2", CTLFLAG_RD,
1839 	    &rack_used_tlpmethod2,
1840 	    "Total number of times we hit TLP method 2");
1841 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1842 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1843 	    SYSCTL_CHILDREN(rack_attack),
1844 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1845 	    &rack_sack_skipped_acked,
1846 	    "Total number of times we skipped previously sacked");
1847 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1848 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1849 	    SYSCTL_CHILDREN(rack_attack),
1850 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1851 	    &rack_sack_splits,
1852 	    "Total number of times we did the old fashion tree split");
1853 	rack_progress_drops = counter_u64_alloc(M_WAITOK);
1854 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1855 	    SYSCTL_CHILDREN(rack_counters),
1856 	    OID_AUTO, "prog_drops", CTLFLAG_RD,
1857 	    &rack_progress_drops,
1858 	    "Total number of progress drops");
1859 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1860 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1861 	    SYSCTL_CHILDREN(rack_counters),
1862 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1863 	    &rack_input_idle_reduces,
1864 	    "Total number of idle reductions on input");
1865 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1866 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1867 	    SYSCTL_CHILDREN(rack_counters),
1868 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1869 	    &rack_collapsed_win,
1870 	    "Total number of collapsed windows");
1871 	rack_tlp_does_nada = counter_u64_alloc(M_WAITOK);
1872 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1873 	    SYSCTL_CHILDREN(rack_counters),
1874 	    OID_AUTO, "tlp_nada", CTLFLAG_RD,
1875 	    &rack_tlp_does_nada,
1876 	    "Total number of nada tlp calls");
1877 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1878 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1879 	    SYSCTL_CHILDREN(rack_counters),
1880 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1881 	    &rack_try_scwnd,
1882 	    "Total number of scwnd attempts");
1883 
1884 	rack_per_timer_hole = counter_u64_alloc(M_WAITOK);
1885 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1886 	    SYSCTL_CHILDREN(rack_counters),
1887 	    OID_AUTO, "timer_hole", CTLFLAG_RD,
1888 	    &rack_per_timer_hole,
1889 	    "Total persists start in timer hole");
1890 
1891 	rack_sbsndptr_wrong = counter_u64_alloc(M_WAITOK);
1892 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1893 	    SYSCTL_CHILDREN(rack_counters),
1894 	    OID_AUTO, "sndptr_wrong", CTLFLAG_RD,
1895 	    &rack_sbsndptr_wrong, "Total number of times the saved sbsndptr was incorret");
1896 	rack_sbsndptr_right = counter_u64_alloc(M_WAITOK);
1897 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1898 	    SYSCTL_CHILDREN(rack_counters),
1899 	    OID_AUTO, "sndptr_right", CTLFLAG_RD,
1900 	    &rack_sbsndptr_right, "Total number of times the saved sbsndptr was corret");
1901 
1902 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1903 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1904 	    OID_AUTO, "outsize", CTLFLAG_RD,
1905 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1906 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1907 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1908 	    OID_AUTO, "opts", CTLFLAG_RD,
1909 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1910 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1911 	    SYSCTL_CHILDREN(rack_sysctl_root),
1912 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1913 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1914 }
1915 
1916 static __inline int
1917 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1918 {
1919 	if (SEQ_GEQ(b->r_start, a->r_start) &&
1920 	    SEQ_LT(b->r_start, a->r_end)) {
1921 		/*
1922 		 * The entry b is within the
1923 		 * block a. i.e.:
1924 		 * a --   |-------------|
1925 		 * b --   |----|
1926 		 * <or>
1927 		 * b --       |------|
1928 		 * <or>
1929 		 * b --       |-----------|
1930 		 */
1931 		return (0);
1932 	} else if (SEQ_GEQ(b->r_start, a->r_end)) {
1933 		/*
1934 		 * b falls as either the next
1935 		 * sequence block after a so a
1936 		 * is said to be smaller than b.
1937 		 * i.e:
1938 		 * a --   |------|
1939 		 * b --          |--------|
1940 		 * or
1941 		 * b --              |-----|
1942 		 */
1943 		return (1);
1944 	}
1945 	/*
1946 	 * Whats left is where a is
1947 	 * larger than b. i.e:
1948 	 * a --         |-------|
1949 	 * b --  |---|
1950 	 * or even possibly
1951 	 * b --   |--------------|
1952 	 */
1953 	return (-1);
1954 }
1955 
1956 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1957 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1958 
1959 static uint32_t
1960 rc_init_window(struct tcp_rack *rack)
1961 {
1962 	uint32_t win;
1963 
1964 	if (rack->rc_init_win == 0) {
1965 		/*
1966 		 * Nothing set by the user, use the system stack
1967 		 * default.
1968 		 */
1969 		return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1970 	}
1971 	win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1972 	return (win);
1973 }
1974 
1975 static uint64_t
1976 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1977 {
1978 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1979 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1980 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1981 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1982 	else
1983 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1984 }
1985 
1986 static uint64_t
1987 rack_get_bw(struct tcp_rack *rack)
1988 {
1989 	if (rack->use_fixed_rate) {
1990 		/* Return the fixed pacing rate */
1991 		return (rack_get_fixed_pacing_bw(rack));
1992 	}
1993 	if (rack->r_ctl.gp_bw == 0) {
1994 		/*
1995 		 * We have yet no b/w measurement,
1996 		 * if we have a user set initial bw
1997 		 * return it. If we don't have that and
1998 		 * we have an srtt, use the tcp IW (10) to
1999 		 * calculate a fictional b/w over the SRTT
2000 		 * which is more or less a guess. Note
2001 		 * we don't use our IW from rack on purpose
2002 		 * so if we have like IW=30, we are not
2003 		 * calculating a "huge" b/w.
2004 		 */
2005 		uint64_t bw, srtt;
2006 		if (rack->r_ctl.init_rate)
2007 			return (rack->r_ctl.init_rate);
2008 
2009 		/* Has the user set a max peak rate? */
2010 #ifdef NETFLIX_PEAKRATE
2011 		if (rack->rc_tp->t_maxpeakrate)
2012 			return (rack->rc_tp->t_maxpeakrate);
2013 #endif
2014 		/* Ok lets come up with the IW guess, if we have a srtt */
2015 		if (rack->rc_tp->t_srtt == 0) {
2016 			/*
2017 			 * Go with old pacing method
2018 			 * i.e. burst mitigation only.
2019 			 */
2020 			return (0);
2021 		}
2022 		/* Ok lets get the initial TCP win (not racks) */
2023 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2024 		srtt = (uint64_t)rack->rc_tp->t_srtt;
2025 		bw *= (uint64_t)USECS_IN_SECOND;
2026 		bw /= srtt;
2027 		if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2028 			bw = rack->r_ctl.bw_rate_cap;
2029 		return (bw);
2030 	} else {
2031 		uint64_t bw;
2032 
2033 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2034 			/* Averaging is done, we can return the value */
2035 			bw = rack->r_ctl.gp_bw;
2036 		} else {
2037 			/* Still doing initial average must calculate */
2038 			bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
2039 		}
2040 #ifdef NETFLIX_PEAKRATE
2041 		if ((rack->rc_tp->t_maxpeakrate) &&
2042 		    (bw > rack->rc_tp->t_maxpeakrate)) {
2043 			/* The user has set a peak rate to pace at
2044 			 * don't allow us to pace faster than that.
2045 			 */
2046 			return (rack->rc_tp->t_maxpeakrate);
2047 		}
2048 #endif
2049 		if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
2050 			bw = rack->r_ctl.bw_rate_cap;
2051 		return (bw);
2052 	}
2053 }
2054 
2055 static uint16_t
2056 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2057 {
2058 	if (rack->use_fixed_rate) {
2059 		return (100);
2060 	} else if (rack->in_probe_rtt && (rsm == NULL))
2061 		return (rack->r_ctl.rack_per_of_gp_probertt);
2062 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2063 		  rack->r_ctl.rack_per_of_gp_rec)) {
2064 		if (rsm) {
2065 			/* a retransmission always use the recovery rate */
2066 			return (rack->r_ctl.rack_per_of_gp_rec);
2067 		} else if (rack->rack_rec_nonrxt_use_cr) {
2068 			/* Directed to use the configured rate */
2069 			goto configured_rate;
2070 		} else if (rack->rack_no_prr &&
2071 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2072 			/* No PRR, lets just use the b/w estimate only */
2073 			return (100);
2074 		} else {
2075 			/*
2076 			 * Here we may have a non-retransmit but we
2077 			 * have no overrides, so just use the recovery
2078 			 * rate (prr is in effect).
2079 			 */
2080 			return (rack->r_ctl.rack_per_of_gp_rec);
2081 		}
2082 	}
2083 configured_rate:
2084 	/* For the configured rate we look at our cwnd vs the ssthresh */
2085 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2086 		return (rack->r_ctl.rack_per_of_gp_ss);
2087 	else
2088 		return (rack->r_ctl.rack_per_of_gp_ca);
2089 }
2090 
2091 static void
2092 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
2093 {
2094 	/*
2095 	 * Types of logs (mod value)
2096 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
2097 	 * 2 = a dsack round begins, persist is reset to 16.
2098 	 * 3 = a dsack round ends
2099 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
2100 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
2101 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
2102 	 */
2103 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2104 		union tcp_log_stackspecific log;
2105 		struct timeval tv;
2106 
2107 		memset(&log, 0, sizeof(log));
2108 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2109 		log.u_bbr.flex1 <<= 1;
2110 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2111 		log.u_bbr.flex1 <<= 1;
2112 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2113 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2114 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2115 		log.u_bbr.flex4 = flex4;
2116 		log.u_bbr.flex5 = flex5;
2117 		log.u_bbr.flex6 = flex6;
2118 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2119 		log.u_bbr.flex8 = mod;
2120 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2121 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2122 		    &rack->rc_inp->inp_socket->so_rcv,
2123 		    &rack->rc_inp->inp_socket->so_snd,
2124 		    RACK_DSACK_HANDLING, 0,
2125 		    0, &log, false, &tv);
2126 	}
2127 }
2128 
2129 static void
2130 rack_log_hdwr_pacing(struct tcp_rack *rack,
2131 		     uint64_t rate, uint64_t hw_rate, int line,
2132 		     int error, uint16_t mod)
2133 {
2134 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2135 		union tcp_log_stackspecific log;
2136 		struct timeval tv;
2137 		const struct ifnet *ifp;
2138 
2139 		memset(&log, 0, sizeof(log));
2140 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2141 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2142 		if (rack->r_ctl.crte) {
2143 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2144 		} else if (rack->rc_inp->inp_route.ro_nh &&
2145 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2146 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2147 		} else
2148 			ifp = NULL;
2149 		if (ifp) {
2150 			log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2151 			log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2152 		}
2153 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2154 		log.u_bbr.bw_inuse = rate;
2155 		log.u_bbr.flex5 = line;
2156 		log.u_bbr.flex6 = error;
2157 		log.u_bbr.flex7 = mod;
2158 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2159 		log.u_bbr.flex8 = rack->use_fixed_rate;
2160 		log.u_bbr.flex8 <<= 1;
2161 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2162 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2163 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2164 		if (rack->r_ctl.crte)
2165 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2166 		else
2167 			log.u_bbr.cur_del_rate = 0;
2168 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2169 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2170 		    &rack->rc_inp->inp_socket->so_rcv,
2171 		    &rack->rc_inp->inp_socket->so_snd,
2172 		    BBR_LOG_HDWR_PACE, 0,
2173 		    0, &log, false, &tv);
2174 	}
2175 }
2176 
2177 static uint64_t
2178 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2179 {
2180 	/*
2181 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2182 	 */
2183 	uint64_t bw_est, high_rate;
2184 	uint64_t gain;
2185 
2186 	gain = (uint64_t)rack_get_output_gain(rack, rsm);
2187 	bw_est = bw * gain;
2188 	bw_est /= (uint64_t)100;
2189 	/* Never fall below the minimum (def 64kbps) */
2190 	if (bw_est < RACK_MIN_BW)
2191 		bw_est = RACK_MIN_BW;
2192 	if (rack->r_rack_hw_rate_caps) {
2193 		/* Rate caps are in place */
2194 		if (rack->r_ctl.crte != NULL) {
2195 			/* We have a hdwr rate already */
2196 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2197 			if (bw_est >= high_rate) {
2198 				/* We are capping bw at the highest rate table entry */
2199 				rack_log_hdwr_pacing(rack,
2200 						     bw_est, high_rate, __LINE__,
2201 						     0, 3);
2202 				bw_est = high_rate;
2203 				if (capped)
2204 					*capped = 1;
2205 			}
2206 		} else if ((rack->rack_hdrw_pacing == 0) &&
2207 			   (rack->rack_hdw_pace_ena) &&
2208 			   (rack->rack_attempt_hdwr_pace == 0) &&
2209 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2210 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2211 			/*
2212 			 * Special case, we have not yet attempted hardware
2213 			 * pacing, and yet we may, when we do, find out if we are
2214 			 * above the highest rate. We need to know the maxbw for the interface
2215 			 * in question (if it supports ratelimiting). We get back
2216 			 * a 0, if the interface is not found in the RL lists.
2217 			 */
2218 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2219 			if (high_rate) {
2220 				/* Yep, we have a rate is it above this rate? */
2221 				if (bw_est > high_rate) {
2222 					bw_est = high_rate;
2223 					if (capped)
2224 						*capped = 1;
2225 				}
2226 			}
2227 		}
2228 	}
2229 	return (bw_est);
2230 }
2231 
2232 static void
2233 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2234 {
2235 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2236 		union tcp_log_stackspecific log;
2237 		struct timeval tv;
2238 
2239 		if ((mod != 1) && (rack_verbose_logging == 0)) {
2240 			/*
2241 			 * We get 3 values currently for mod
2242 			 * 1 - We are retransmitting and this tells the reason.
2243 			 * 2 - We are clearing a dup-ack count.
2244 			 * 3 - We are incrementing a dup-ack count.
2245 			 *
2246 			 * The clear/increment are only logged
2247 			 * if you have BBverbose on.
2248 			 */
2249 			return;
2250 		}
2251 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2252 		log.u_bbr.flex1 = tsused;
2253 		log.u_bbr.flex2 = thresh;
2254 		log.u_bbr.flex3 = rsm->r_flags;
2255 		log.u_bbr.flex4 = rsm->r_dupack;
2256 		log.u_bbr.flex5 = rsm->r_start;
2257 		log.u_bbr.flex6 = rsm->r_end;
2258 		log.u_bbr.flex8 = mod;
2259 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2260 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2261 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2262 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2263 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2264 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2265 		log.u_bbr.pacing_gain = rack->r_must_retran;
2266 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2267 		    &rack->rc_inp->inp_socket->so_rcv,
2268 		    &rack->rc_inp->inp_socket->so_snd,
2269 		    BBR_LOG_SETTINGS_CHG, 0,
2270 		    0, &log, false, &tv);
2271 	}
2272 }
2273 
2274 static void
2275 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2276 {
2277 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2278 		union tcp_log_stackspecific log;
2279 		struct timeval tv;
2280 
2281 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2282 		log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2283 		log.u_bbr.flex2 = to;
2284 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2285 		log.u_bbr.flex4 = slot;
2286 		log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2287 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2288 		log.u_bbr.flex7 = rack->rc_in_persist;
2289 		log.u_bbr.flex8 = which;
2290 		if (rack->rack_no_prr)
2291 			log.u_bbr.pkts_out = 0;
2292 		else
2293 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2294 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2295 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2296 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2297 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2298 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2299 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2300 		log.u_bbr.pacing_gain = rack->r_must_retran;
2301 		log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2302 		log.u_bbr.lost = rack_rto_min;
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_TIMERSTAR, 0,
2307 		    0, &log, false, &tv);
2308 	}
2309 }
2310 
2311 static void
2312 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
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.inhpts = rack->rc_inp->inp_in_hpts;
2320 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2321 		log.u_bbr.flex8 = to_num;
2322 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2323 		log.u_bbr.flex2 = rack->rc_rack_rtt;
2324 		if (rsm == NULL)
2325 			log.u_bbr.flex3 = 0;
2326 		else
2327 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2328 		if (rack->rack_no_prr)
2329 			log.u_bbr.flex5 = 0;
2330 		else
2331 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
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 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2338 		    &rack->rc_inp->inp_socket->so_rcv,
2339 		    &rack->rc_inp->inp_socket->so_snd,
2340 		    BBR_LOG_RTO, 0,
2341 		    0, &log, false, &tv);
2342 	}
2343 }
2344 
2345 static void
2346 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2347 		 struct rack_sendmap *prev,
2348 		 struct rack_sendmap *rsm,
2349 		 struct rack_sendmap *next,
2350 		 int flag, uint32_t th_ack, int line)
2351 {
2352 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2353 		union tcp_log_stackspecific log;
2354 		struct timeval tv;
2355 
2356 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2357 		log.u_bbr.flex8 = flag;
2358 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2359 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2360 		log.u_bbr.cur_del_rate = (uint64_t)prev;
2361 		log.u_bbr.delRate = (uint64_t)rsm;
2362 		log.u_bbr.rttProp = (uint64_t)next;
2363 		log.u_bbr.flex7 = 0;
2364 		if (prev) {
2365 			log.u_bbr.flex1 = prev->r_start;
2366 			log.u_bbr.flex2 = prev->r_end;
2367 			log.u_bbr.flex7 |= 0x4;
2368 		}
2369 		if (rsm) {
2370 			log.u_bbr.flex3 = rsm->r_start;
2371 			log.u_bbr.flex4 = rsm->r_end;
2372 			log.u_bbr.flex7 |= 0x2;
2373 		}
2374 		if (next) {
2375 			log.u_bbr.flex5 = next->r_start;
2376 			log.u_bbr.flex6 = next->r_end;
2377 			log.u_bbr.flex7 |= 0x1;
2378 		}
2379 		log.u_bbr.applimited = line;
2380 		log.u_bbr.pkts_out = th_ack;
2381 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2382 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2383 		if (rack->rack_no_prr)
2384 			log.u_bbr.lost = 0;
2385 		else
2386 			log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2387 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2388 		    &rack->rc_inp->inp_socket->so_rcv,
2389 		    &rack->rc_inp->inp_socket->so_snd,
2390 		    TCP_LOG_MAPCHG, 0,
2391 		    0, &log, false, &tv);
2392 	}
2393 }
2394 
2395 static void
2396 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2397 		 struct rack_sendmap *rsm, int conf)
2398 {
2399 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2400 		union tcp_log_stackspecific log;
2401 		struct timeval tv;
2402 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2403 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2404 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2405 		log.u_bbr.flex1 = t;
2406 		log.u_bbr.flex2 = len;
2407 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2408 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2409 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2410 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2411 		log.u_bbr.flex7 = conf;
2412 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2413 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2414 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2415 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2416 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2417 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2418 		if (rsm) {
2419 			log.u_bbr.pkt_epoch = rsm->r_start;
2420 			log.u_bbr.lost = rsm->r_end;
2421 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2422 			/* We loose any upper of the 24 bits */
2423 			log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2424 		} else {
2425 			/* Its a SYN */
2426 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2427 			log.u_bbr.lost = 0;
2428 			log.u_bbr.cwnd_gain = 0;
2429 			log.u_bbr.pacing_gain = 0;
2430 		}
2431 		/* Write out general bits of interest rrs here */
2432 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2433 		log.u_bbr.use_lt_bw <<= 1;
2434 		log.u_bbr.use_lt_bw |= rack->forced_ack;
2435 		log.u_bbr.use_lt_bw <<= 1;
2436 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2437 		log.u_bbr.use_lt_bw <<= 1;
2438 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2439 		log.u_bbr.use_lt_bw <<= 1;
2440 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2441 		log.u_bbr.use_lt_bw <<= 1;
2442 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2443 		log.u_bbr.use_lt_bw <<= 1;
2444 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2445 		log.u_bbr.use_lt_bw <<= 1;
2446 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2447 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2448 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2449 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2450 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2451 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2452 		log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2453 		log.u_bbr.bw_inuse <<= 32;
2454 		if (rsm)
2455 			log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2456 		TCP_LOG_EVENTP(tp, NULL,
2457 		    &rack->rc_inp->inp_socket->so_rcv,
2458 		    &rack->rc_inp->inp_socket->so_snd,
2459 		    BBR_LOG_BBRRTT, 0,
2460 		    0, &log, false, &tv);
2461 
2462 
2463 	}
2464 }
2465 
2466 static void
2467 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2468 {
2469 	/*
2470 	 * Log the rtt sample we are
2471 	 * applying to the srtt algorithm in
2472 	 * useconds.
2473 	 */
2474 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2475 		union tcp_log_stackspecific log;
2476 		struct timeval tv;
2477 
2478 		/* Convert our ms to a microsecond */
2479 		memset(&log, 0, sizeof(log));
2480 		log.u_bbr.flex1 = rtt;
2481 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2482 		log.u_bbr.flex3 = rack->r_ctl.sack_count;
2483 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2484 		log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2485 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2486 		log.u_bbr.flex7 = 1;
2487 		log.u_bbr.flex8 = rack->sack_attack_disable;
2488 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2489 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2490 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2491 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2492 		log.u_bbr.pacing_gain = rack->r_must_retran;
2493 		/*
2494 		 * We capture in delRate the upper 32 bits as
2495 		 * the confidence level we had declared, and the
2496 		 * lower 32 bits as the actual RTT using the arrival
2497 		 * timestamp.
2498 		 */
2499 		log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2500 		log.u_bbr.delRate <<= 32;
2501 		log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2502 		/* Lets capture all the things that make up t_rtxcur */
2503 		log.u_bbr.applimited = rack_rto_min;
2504 		log.u_bbr.epoch = rack_rto_max;
2505 		log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2506 		log.u_bbr.lost = rack_rto_min;
2507 		log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2508 		log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2509 		log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2510 		log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2511 		log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2512 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2513 		    &rack->rc_inp->inp_socket->so_rcv,
2514 		    &rack->rc_inp->inp_socket->so_snd,
2515 		    TCP_LOG_RTT, 0,
2516 		    0, &log, false, &tv);
2517 	}
2518 }
2519 
2520 static void
2521 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2522 {
2523 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2524 		union tcp_log_stackspecific log;
2525 		struct timeval tv;
2526 
2527 		/* Convert our ms to a microsecond */
2528 		memset(&log, 0, sizeof(log));
2529 		log.u_bbr.flex1 = rtt;
2530 		log.u_bbr.flex2 = send_time;
2531 		log.u_bbr.flex3 = ack_time;
2532 		log.u_bbr.flex4 = where;
2533 		log.u_bbr.flex7 = 2;
2534 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2535 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2536 		    &rack->rc_inp->inp_socket->so_rcv,
2537 		    &rack->rc_inp->inp_socket->so_snd,
2538 		    TCP_LOG_RTT, 0,
2539 		    0, &log, false, &tv);
2540 	}
2541 }
2542 
2543 
2544 
2545 static inline void
2546 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2547 {
2548 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2549 		union tcp_log_stackspecific log;
2550 		struct timeval tv;
2551 
2552 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2553 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2554 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2555 		log.u_bbr.flex1 = line;
2556 		log.u_bbr.flex2 = tick;
2557 		log.u_bbr.flex3 = tp->t_maxunacktime;
2558 		log.u_bbr.flex4 = tp->t_acktime;
2559 		log.u_bbr.flex8 = event;
2560 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2561 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2562 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2563 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2564 		log.u_bbr.pacing_gain = rack->r_must_retran;
2565 		TCP_LOG_EVENTP(tp, NULL,
2566 		    &rack->rc_inp->inp_socket->so_rcv,
2567 		    &rack->rc_inp->inp_socket->so_snd,
2568 		    BBR_LOG_PROGRESS, 0,
2569 		    0, &log, false, &tv);
2570 	}
2571 }
2572 
2573 static void
2574 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2575 {
2576 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2577 		union tcp_log_stackspecific log;
2578 
2579 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2580 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2581 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2582 		log.u_bbr.flex1 = slot;
2583 		if (rack->rack_no_prr)
2584 			log.u_bbr.flex2 = 0;
2585 		else
2586 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2587 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2588 		log.u_bbr.flex8 = rack->rc_in_persist;
2589 		log.u_bbr.timeStamp = cts;
2590 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2591 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2592 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2593 		log.u_bbr.pacing_gain = rack->r_must_retran;
2594 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2595 		    &rack->rc_inp->inp_socket->so_rcv,
2596 		    &rack->rc_inp->inp_socket->so_snd,
2597 		    BBR_LOG_BBRSND, 0,
2598 		    0, &log, false, tv);
2599 	}
2600 }
2601 
2602 static void
2603 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2604 {
2605 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2606 		union tcp_log_stackspecific log;
2607 		struct timeval tv;
2608 
2609 		memset(&log, 0, sizeof(log));
2610 		log.u_bbr.flex1 = did_out;
2611 		log.u_bbr.flex2 = nxt_pkt;
2612 		log.u_bbr.flex3 = way_out;
2613 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2614 		if (rack->rack_no_prr)
2615 			log.u_bbr.flex5 = 0;
2616 		else
2617 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2618 		log.u_bbr.flex6 = nsegs;
2619 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2620 		log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;	/* Do we have ack-can-send set */
2621 		log.u_bbr.flex7 <<= 1;
2622 		log.u_bbr.flex7 |= rack->r_fast_output;	/* is fast output primed */
2623 		log.u_bbr.flex7 <<= 1;
2624 		log.u_bbr.flex7 |= rack->r_wanted_output;	/* Do we want output */
2625 		log.u_bbr.flex8 = rack->rc_in_persist;
2626 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2627 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2628 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2629 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2630 		log.u_bbr.use_lt_bw <<= 1;
2631 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
2632 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2633 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2634 		log.u_bbr.pacing_gain = rack->r_must_retran;
2635 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2636 		    &rack->rc_inp->inp_socket->so_rcv,
2637 		    &rack->rc_inp->inp_socket->so_snd,
2638 		    BBR_LOG_DOSEG_DONE, 0,
2639 		    0, &log, false, &tv);
2640 	}
2641 }
2642 
2643 static void
2644 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2645 {
2646 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2647 		union tcp_log_stackspecific log;
2648 		struct timeval tv;
2649 		uint32_t cts;
2650 
2651 		memset(&log, 0, sizeof(log));
2652 		cts = tcp_get_usecs(&tv);
2653 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2654 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2655 		log.u_bbr.flex4 = arg1;
2656 		log.u_bbr.flex5 = arg2;
2657 		log.u_bbr.flex6 = arg3;
2658 		log.u_bbr.flex8 = frm;
2659 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2660 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2661 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2662 		log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2663 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2664 		log.u_bbr.pacing_gain = rack->r_must_retran;
2665 		TCP_LOG_EVENTP(tp, NULL,
2666 		    &tp->t_inpcb->inp_socket->so_rcv,
2667 		    &tp->t_inpcb->inp_socket->so_snd,
2668 		    TCP_HDWR_PACE_SIZE, 0,
2669 		    0, &log, false, &tv);
2670 	}
2671 }
2672 
2673 static void
2674 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2675 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2676 {
2677 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2678 		union tcp_log_stackspecific log;
2679 		struct timeval tv;
2680 
2681 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2682 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2683 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2684 		log.u_bbr.flex1 = slot;
2685 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2686 		log.u_bbr.flex4 = reason;
2687 		if (rack->rack_no_prr)
2688 			log.u_bbr.flex5 = 0;
2689 		else
2690 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2691 		log.u_bbr.flex7 = hpts_calling;
2692 		log.u_bbr.flex8 = rack->rc_in_persist;
2693 		log.u_bbr.lt_epoch = cwnd_to_use;
2694 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2695 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2696 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2697 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2698 		log.u_bbr.pacing_gain = rack->r_must_retran;
2699 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2700 		    &rack->rc_inp->inp_socket->so_rcv,
2701 		    &rack->rc_inp->inp_socket->so_snd,
2702 		    BBR_LOG_JUSTRET, 0,
2703 		    tlen, &log, false, &tv);
2704 	}
2705 }
2706 
2707 static void
2708 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2709 		   struct timeval *tv, uint32_t flags_on_entry)
2710 {
2711 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2712 		union tcp_log_stackspecific log;
2713 
2714 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2715 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
2716 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
2717 		log.u_bbr.flex1 = line;
2718 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2719 		log.u_bbr.flex3 = flags_on_entry;
2720 		log.u_bbr.flex4 = us_cts;
2721 		if (rack->rack_no_prr)
2722 			log.u_bbr.flex5 = 0;
2723 		else
2724 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2725 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2726 		log.u_bbr.flex7 = hpts_removed;
2727 		log.u_bbr.flex8 = 1;
2728 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2729 		log.u_bbr.timeStamp = us_cts;
2730 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2731 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2732 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2733 		log.u_bbr.pacing_gain = rack->r_must_retran;
2734 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2735 		    &rack->rc_inp->inp_socket->so_rcv,
2736 		    &rack->rc_inp->inp_socket->so_snd,
2737 		    BBR_LOG_TIMERCANC, 0,
2738 		    0, &log, false, tv);
2739 	}
2740 }
2741 
2742 static void
2743 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2744 			  uint32_t flex1, uint32_t flex2,
2745 			  uint32_t flex3, uint32_t flex4,
2746 			  uint32_t flex5, uint32_t flex6,
2747 			  uint16_t flex7, uint8_t mod)
2748 {
2749 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2750 		union tcp_log_stackspecific log;
2751 		struct timeval tv;
2752 
2753 		if (mod == 1) {
2754 			/* No you can't use 1, its for the real to cancel */
2755 			return;
2756 		}
2757 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2758 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2759 		log.u_bbr.flex1 = flex1;
2760 		log.u_bbr.flex2 = flex2;
2761 		log.u_bbr.flex3 = flex3;
2762 		log.u_bbr.flex4 = flex4;
2763 		log.u_bbr.flex5 = flex5;
2764 		log.u_bbr.flex6 = flex6;
2765 		log.u_bbr.flex7 = flex7;
2766 		log.u_bbr.flex8 = mod;
2767 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2768 		    &rack->rc_inp->inp_socket->so_rcv,
2769 		    &rack->rc_inp->inp_socket->so_snd,
2770 		    BBR_LOG_TIMERCANC, 0,
2771 		    0, &log, false, &tv);
2772 	}
2773 }
2774 
2775 static void
2776 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2777 {
2778 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2779 		union tcp_log_stackspecific log;
2780 		struct timeval tv;
2781 
2782 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2783 		log.u_bbr.flex1 = timers;
2784 		log.u_bbr.flex2 = ret;
2785 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2786 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2787 		log.u_bbr.flex5 = cts;
2788 		if (rack->rack_no_prr)
2789 			log.u_bbr.flex6 = 0;
2790 		else
2791 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2792 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2793 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2794 		log.u_bbr.pacing_gain = rack->r_must_retran;
2795 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2796 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2797 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2798 		    &rack->rc_inp->inp_socket->so_rcv,
2799 		    &rack->rc_inp->inp_socket->so_snd,
2800 		    BBR_LOG_TO_PROCESS, 0,
2801 		    0, &log, false, &tv);
2802 	}
2803 }
2804 
2805 static void
2806 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd)
2807 {
2808 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2809 		union tcp_log_stackspecific log;
2810 		struct timeval tv;
2811 
2812 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2813 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2814 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2815 		if (rack->rack_no_prr)
2816 			log.u_bbr.flex3 = 0;
2817 		else
2818 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2819 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2820 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2821 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2822 		log.u_bbr.flex8 = frm;
2823 		log.u_bbr.pkts_out = orig_cwnd;
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 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2827 		log.u_bbr.use_lt_bw <<= 1;
2828 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
2829 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2830 		    &rack->rc_inp->inp_socket->so_rcv,
2831 		    &rack->rc_inp->inp_socket->so_snd,
2832 		    BBR_LOG_BBRUPD, 0,
2833 		    0, &log, false, &tv);
2834 	}
2835 }
2836 
2837 #ifdef NETFLIX_EXP_DETECTION
2838 static void
2839 rack_log_sad(struct tcp_rack *rack, int event)
2840 {
2841 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2842 		union tcp_log_stackspecific log;
2843 		struct timeval tv;
2844 
2845 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2846 		log.u_bbr.flex1 = rack->r_ctl.sack_count;
2847 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2848 		log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2849 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2850 		log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2851 		log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2852 		log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2853 		log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2854 		log.u_bbr.lt_epoch |= rack->do_detection;
2855 		log.u_bbr.applimited = tcp_map_minimum;
2856 		log.u_bbr.flex7 = rack->sack_attack_disable;
2857 		log.u_bbr.flex8 = event;
2858 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2859 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2860 		log.u_bbr.delivered = tcp_sad_decay_val;
2861 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2862 		    &rack->rc_inp->inp_socket->so_rcv,
2863 		    &rack->rc_inp->inp_socket->so_snd,
2864 		    TCP_SAD_DETECTION, 0,
2865 		    0, &log, false, &tv);
2866 	}
2867 }
2868 #endif
2869 
2870 static void
2871 rack_counter_destroy(void)
2872 {
2873 	int i;
2874 
2875 	counter_u64_free(rack_fto_send);
2876 	counter_u64_free(rack_fto_rsm_send);
2877 	counter_u64_free(rack_nfto_resend);
2878 	counter_u64_free(rack_hw_pace_init_fail);
2879 	counter_u64_free(rack_hw_pace_lost);
2880 	counter_u64_free(rack_non_fto_send);
2881 	counter_u64_free(rack_extended_rfo);
2882 	counter_u64_free(rack_ack_total);
2883 	counter_u64_free(rack_express_sack);
2884 	counter_u64_free(rack_sack_total);
2885 	counter_u64_free(rack_move_none);
2886 	counter_u64_free(rack_move_some);
2887 	counter_u64_free(rack_sack_attacks_detected);
2888 	counter_u64_free(rack_sack_attacks_reversed);
2889 	counter_u64_free(rack_sack_used_next_merge);
2890 	counter_u64_free(rack_sack_used_prev_merge);
2891 	counter_u64_free(rack_badfr);
2892 	counter_u64_free(rack_badfr_bytes);
2893 	counter_u64_free(rack_rtm_prr_retran);
2894 	counter_u64_free(rack_rtm_prr_newdata);
2895 	counter_u64_free(rack_timestamp_mismatch);
2896 	counter_u64_free(rack_find_high);
2897 	counter_u64_free(rack_reorder_seen);
2898 	counter_u64_free(rack_tlp_tot);
2899 	counter_u64_free(rack_tlp_newdata);
2900 	counter_u64_free(rack_tlp_retran);
2901 	counter_u64_free(rack_tlp_retran_bytes);
2902 	counter_u64_free(rack_tlp_retran_fail);
2903 	counter_u64_free(rack_to_tot);
2904 	counter_u64_free(rack_to_arm_rack);
2905 	counter_u64_free(rack_to_arm_tlp);
2906 	counter_u64_free(rack_calc_zero);
2907 	counter_u64_free(rack_calc_nonzero);
2908 	counter_u64_free(rack_paced_segments);
2909 	counter_u64_free(rack_unpaced_segments);
2910 	counter_u64_free(rack_saw_enobuf);
2911 	counter_u64_free(rack_saw_enobuf_hw);
2912 	counter_u64_free(rack_saw_enetunreach);
2913 	counter_u64_free(rack_hot_alloc);
2914 	counter_u64_free(rack_to_alloc);
2915 	counter_u64_free(rack_to_alloc_hard);
2916 	counter_u64_free(rack_to_alloc_emerg);
2917 	counter_u64_free(rack_to_alloc_limited);
2918 	counter_u64_free(rack_alloc_limited_conns);
2919 	counter_u64_free(rack_split_limited);
2920 	for (i = 0; i < MAX_NUM_OF_CNTS; i++) {
2921 		counter_u64_free(rack_proc_comp_ack[i]);
2922 	}
2923 	counter_u64_free(rack_multi_single_eq);
2924 	counter_u64_free(rack_proc_non_comp_ack);
2925 	counter_u64_free(rack_sack_proc_all);
2926 	counter_u64_free(rack_sack_proc_restart);
2927 	counter_u64_free(rack_sack_proc_short);
2928 	counter_u64_free(rack_enter_tlp_calc);
2929 	counter_u64_free(rack_used_tlpmethod);
2930 	counter_u64_free(rack_used_tlpmethod2);
2931 	counter_u64_free(rack_sack_skipped_acked);
2932 	counter_u64_free(rack_sack_splits);
2933 	counter_u64_free(rack_progress_drops);
2934 	counter_u64_free(rack_input_idle_reduces);
2935 	counter_u64_free(rack_collapsed_win);
2936 	counter_u64_free(rack_tlp_does_nada);
2937 	counter_u64_free(rack_try_scwnd);
2938 	counter_u64_free(rack_per_timer_hole);
2939 	counter_u64_free(rack_large_ackcmp);
2940 	counter_u64_free(rack_small_ackcmp);
2941 #ifdef INVARIANTS
2942 	counter_u64_free(rack_adjust_map_bw);
2943 #endif
2944 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2945 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2946 }
2947 
2948 static struct rack_sendmap *
2949 rack_alloc(struct tcp_rack *rack)
2950 {
2951 	struct rack_sendmap *rsm;
2952 
2953 	/*
2954 	 * First get the top of the list it in
2955 	 * theory is the "hottest" rsm we have,
2956 	 * possibly just freed by ack processing.
2957 	 */
2958 	if (rack->rc_free_cnt > rack_free_cache) {
2959 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2960 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2961 		counter_u64_add(rack_hot_alloc, 1);
2962 		rack->rc_free_cnt--;
2963 		return (rsm);
2964 	}
2965 	/*
2966 	 * Once we get under our free cache we probably
2967 	 * no longer have a "hot" one available. Lets
2968 	 * get one from UMA.
2969 	 */
2970 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
2971 	if (rsm) {
2972 		rack->r_ctl.rc_num_maps_alloced++;
2973 		counter_u64_add(rack_to_alloc, 1);
2974 		return (rsm);
2975 	}
2976 	/*
2977 	 * Dig in to our aux rsm's (the last two) since
2978 	 * UMA failed to get us one.
2979 	 */
2980 	if (rack->rc_free_cnt) {
2981 		counter_u64_add(rack_to_alloc_emerg, 1);
2982 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2983 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2984 		rack->rc_free_cnt--;
2985 		return (rsm);
2986 	}
2987 	return (NULL);
2988 }
2989 
2990 static struct rack_sendmap *
2991 rack_alloc_full_limit(struct tcp_rack *rack)
2992 {
2993 	if ((V_tcp_map_entries_limit > 0) &&
2994 	    (rack->do_detection == 0) &&
2995 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2996 		counter_u64_add(rack_to_alloc_limited, 1);
2997 		if (!rack->alloc_limit_reported) {
2998 			rack->alloc_limit_reported = 1;
2999 			counter_u64_add(rack_alloc_limited_conns, 1);
3000 		}
3001 		return (NULL);
3002 	}
3003 	return (rack_alloc(rack));
3004 }
3005 
3006 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3007 static struct rack_sendmap *
3008 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
3009 {
3010 	struct rack_sendmap *rsm;
3011 
3012 	if (limit_type) {
3013 		/* currently there is only one limit type */
3014 		if (V_tcp_map_split_limit > 0 &&
3015 		    (rack->do_detection == 0) &&
3016 		    rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
3017 			counter_u64_add(rack_split_limited, 1);
3018 			if (!rack->alloc_limit_reported) {
3019 				rack->alloc_limit_reported = 1;
3020 				counter_u64_add(rack_alloc_limited_conns, 1);
3021 			}
3022 			return (NULL);
3023 		}
3024 	}
3025 
3026 	/* allocate and mark in the limit type, if set */
3027 	rsm = rack_alloc(rack);
3028 	if (rsm != NULL && limit_type) {
3029 		rsm->r_limit_type = limit_type;
3030 		rack->r_ctl.rc_num_split_allocs++;
3031 	}
3032 	return (rsm);
3033 }
3034 
3035 static void
3036 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
3037 {
3038 	if (rsm->r_flags & RACK_APP_LIMITED) {
3039 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
3040 			rack->r_ctl.rc_app_limited_cnt--;
3041 		}
3042 	}
3043 	if (rsm->r_limit_type) {
3044 		/* currently there is only one limit type */
3045 		rack->r_ctl.rc_num_split_allocs--;
3046 	}
3047 	if (rsm == rack->r_ctl.rc_first_appl) {
3048 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3049 			rack->r_ctl.rc_first_appl = NULL;
3050 		else {
3051 			/* Follow the next one out */
3052 			struct rack_sendmap fe;
3053 
3054 			fe.r_start = rsm->r_nseq_appl;
3055 			rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
3056 		}
3057 	}
3058 	if (rsm == rack->r_ctl.rc_resend)
3059 		rack->r_ctl.rc_resend = NULL;
3060 	if (rsm == rack->r_ctl.rc_rsm_at_retran)
3061 		rack->r_ctl.rc_rsm_at_retran = NULL;
3062 	if (rsm == rack->r_ctl.rc_end_appl)
3063 		rack->r_ctl.rc_end_appl = NULL;
3064 	if (rack->r_ctl.rc_tlpsend == rsm)
3065 		rack->r_ctl.rc_tlpsend = NULL;
3066 	if (rack->r_ctl.rc_sacklast == rsm)
3067 		rack->r_ctl.rc_sacklast = NULL;
3068 	memset(rsm, 0, sizeof(struct rack_sendmap));
3069 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3070 	rack->rc_free_cnt++;
3071 }
3072 
3073 static void
3074 rack_free_trim(struct tcp_rack *rack)
3075 {
3076 	struct rack_sendmap *rsm;
3077 
3078 	/*
3079 	 * Free up all the tail entries until
3080 	 * we get our list down to the limit.
3081 	 */
3082 	while (rack->rc_free_cnt > rack_free_cache) {
3083 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3084 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3085 		rack->rc_free_cnt--;
3086 		uma_zfree(rack_zone, rsm);
3087 	}
3088 }
3089 
3090 
3091 static uint32_t
3092 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3093 {
3094 	uint64_t srtt, bw, len, tim;
3095 	uint32_t segsiz, def_len, minl;
3096 
3097 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3098 	def_len = rack_def_data_window * segsiz;
3099 	if (rack->rc_gp_filled == 0) {
3100 		/*
3101 		 * We have no measurement (IW is in flight?) so
3102 		 * we can only guess using our data_window sysctl
3103 		 * value (usually 20MSS).
3104 		 */
3105 		return (def_len);
3106 	}
3107 	/*
3108 	 * Now we have a number of factors to consider.
3109 	 *
3110 	 * 1) We have a desired BDP which is usually
3111 	 *    at least 2.
3112 	 * 2) We have a minimum number of rtt's usually 1 SRTT
3113 	 *    but we allow it too to be more.
3114 	 * 3) We want to make sure a measurement last N useconds (if
3115 	 *    we have set rack_min_measure_usec.
3116 	 *
3117 	 * We handle the first concern here by trying to create a data
3118 	 * window of max(rack_def_data_window, DesiredBDP). The
3119 	 * second concern we handle in not letting the measurement
3120 	 * window end normally until at least the required SRTT's
3121 	 * have gone by which is done further below in
3122 	 * rack_enough_for_measurement(). Finally the third concern
3123 	 * we also handle here by calculating how long that time
3124 	 * would take at the current BW and then return the
3125 	 * max of our first calculation and that length. Note
3126 	 * that if rack_min_measure_usec is 0, we don't deal
3127 	 * with concern 3. Also for both Concern 1 and 3 an
3128 	 * application limited period could end the measurement
3129 	 * earlier.
3130 	 *
3131 	 * So lets calculate the BDP with the "known" b/w using
3132 	 * the SRTT has our rtt and then multiply it by the
3133 	 * goal.
3134 	 */
3135 	bw = rack_get_bw(rack);
3136 	srtt = (uint64_t)tp->t_srtt;
3137 	len = bw * srtt;
3138 	len /= (uint64_t)HPTS_USEC_IN_SEC;
3139 	len *= max(1, rack_goal_bdp);
3140 	/* Now we need to round up to the nearest MSS */
3141 	len = roundup(len, segsiz);
3142 	if (rack_min_measure_usec) {
3143 		/* Now calculate our min length for this b/w */
3144 		tim = rack_min_measure_usec;
3145 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3146 		if (minl == 0)
3147 			minl = 1;
3148 		minl = roundup(minl, segsiz);
3149 		if (len < minl)
3150 			len = minl;
3151 	}
3152 	/*
3153 	 * Now if we have a very small window we want
3154 	 * to attempt to get the window that is
3155 	 * as small as possible. This happens on
3156 	 * low b/w connections and we don't want to
3157 	 * span huge numbers of rtt's between measurements.
3158 	 *
3159 	 * We basically include 2 over our "MIN window" so
3160 	 * that the measurement can be shortened (possibly) by
3161 	 * an ack'ed packet.
3162 	 */
3163 	if (len < def_len)
3164 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3165 	else
3166 		return (max((uint32_t)len, def_len));
3167 
3168 }
3169 
3170 static int
3171 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3172 {
3173 	uint32_t tim, srtts, segsiz;
3174 
3175 	/*
3176 	 * Has enough time passed for the GP measurement to be valid?
3177 	 */
3178 	if ((tp->snd_max == tp->snd_una) ||
3179 	    (th_ack == tp->snd_max)){
3180 		/* All is acked */
3181 		*quality = RACK_QUALITY_ALLACKED;
3182 		return (1);
3183 	}
3184 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3185 		/* Not enough bytes yet */
3186 		return (0);
3187 	}
3188 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3189 	if (SEQ_LT(th_ack, tp->gput_ack) &&
3190 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3191 		/* Not enough bytes yet */
3192 		return (0);
3193 	}
3194 	if (rack->r_ctl.rc_first_appl &&
3195 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3196 		/*
3197 		 * We are up to the app limited send point
3198 		 * we have to measure irrespective of the time..
3199 		 */
3200 		*quality = RACK_QUALITY_APPLIMITED;
3201 		return (1);
3202 	}
3203 	/* Now what about time? */
3204 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3205 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3206 	if (tim >= srtts) {
3207 		*quality = RACK_QUALITY_HIGH;
3208 		return (1);
3209 	}
3210 	/* Nope not even a full SRTT has passed */
3211 	return (0);
3212 }
3213 
3214 static void
3215 rack_log_timely(struct tcp_rack *rack,
3216 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3217 		uint64_t up_bnd, int line, uint8_t method)
3218 {
3219 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3220 		union tcp_log_stackspecific log;
3221 		struct timeval tv;
3222 
3223 		memset(&log, 0, sizeof(log));
3224 		log.u_bbr.flex1 = logged;
3225 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3226 		log.u_bbr.flex2 <<= 4;
3227 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3228 		log.u_bbr.flex2 <<= 4;
3229 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3230 		log.u_bbr.flex2 <<= 4;
3231 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3232 		log.u_bbr.flex3 = rack->rc_gp_incr;
3233 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3234 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3235 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3236 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3237 		log.u_bbr.flex8 = method;
3238 		log.u_bbr.cur_del_rate = cur_bw;
3239 		log.u_bbr.delRate = low_bnd;
3240 		log.u_bbr.bw_inuse = up_bnd;
3241 		log.u_bbr.rttProp = rack_get_bw(rack);
3242 		log.u_bbr.pkt_epoch = line;
3243 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3244 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3245 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3246 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3247 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3248 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3249 		log.u_bbr.cwnd_gain <<= 1;
3250 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3251 		log.u_bbr.cwnd_gain <<= 1;
3252 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3253 		log.u_bbr.cwnd_gain <<= 1;
3254 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3255 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3256 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3257 		    &rack->rc_inp->inp_socket->so_rcv,
3258 		    &rack->rc_inp->inp_socket->so_snd,
3259 		    TCP_TIMELY_WORK, 0,
3260 		    0, &log, false, &tv);
3261 	}
3262 }
3263 
3264 static int
3265 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3266 {
3267 	/*
3268 	 * Before we increase we need to know if
3269 	 * the estimate just made was less than
3270 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3271 	 *
3272 	 * If we already are pacing at a fast enough
3273 	 * rate to push us faster there is no sense of
3274 	 * increasing.
3275 	 *
3276 	 * We first caculate our actual pacing rate (ss or ca multipler
3277 	 * times our cur_bw).
3278 	 *
3279 	 * Then we take the last measured rate and multipy by our
3280 	 * maximum pacing overage to give us a max allowable rate.
3281 	 *
3282 	 * If our act_rate is smaller than our max_allowable rate
3283 	 * then we should increase. Else we should hold steady.
3284 	 *
3285 	 */
3286 	uint64_t act_rate, max_allow_rate;
3287 
3288 	if (rack_timely_no_stopping)
3289 		return (1);
3290 
3291 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3292 		/*
3293 		 * Initial startup case or
3294 		 * everything is acked case.
3295 		 */
3296 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3297 				__LINE__, 9);
3298 		return (1);
3299 	}
3300 	if (mult <= 100) {
3301 		/*
3302 		 * We can always pace at or slightly above our rate.
3303 		 */
3304 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3305 				__LINE__, 9);
3306 		return (1);
3307 	}
3308 	act_rate = cur_bw * (uint64_t)mult;
3309 	act_rate /= 100;
3310 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3311 	max_allow_rate /= 100;
3312 	if (act_rate < max_allow_rate) {
3313 		/*
3314 		 * Here the rate we are actually pacing at
3315 		 * is smaller than 10% above our last measurement.
3316 		 * This means we are pacing below what we would
3317 		 * like to try to achieve (plus some wiggle room).
3318 		 */
3319 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3320 				__LINE__, 9);
3321 		return (1);
3322 	} else {
3323 		/*
3324 		 * Here we are already pacing at least rack_max_per_above(10%)
3325 		 * what we are getting back. This indicates most likely
3326 		 * that we are being limited (cwnd/rwnd/app) and can't
3327 		 * get any more b/w. There is no sense of trying to
3328 		 * raise up the pacing rate its not speeding us up
3329 		 * and we already are pacing faster than we are getting.
3330 		 */
3331 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3332 				__LINE__, 8);
3333 		return (0);
3334 	}
3335 }
3336 
3337 static void
3338 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3339 {
3340 	/*
3341 	 * When we drag bottom, we want to assure
3342 	 * that no multiplier is below 1.0, if so
3343 	 * we want to restore it to at least that.
3344 	 */
3345 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3346 		/* This is unlikely we usually do not touch recovery */
3347 		rack->r_ctl.rack_per_of_gp_rec = 100;
3348 	}
3349 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3350 		rack->r_ctl.rack_per_of_gp_ca = 100;
3351 	}
3352 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3353 		rack->r_ctl.rack_per_of_gp_ss = 100;
3354 	}
3355 }
3356 
3357 static void
3358 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3359 {
3360 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3361 		rack->r_ctl.rack_per_of_gp_ca = 100;
3362 	}
3363 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3364 		rack->r_ctl.rack_per_of_gp_ss = 100;
3365 	}
3366 }
3367 
3368 static void
3369 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3370 {
3371 	int32_t  calc, logged, plus;
3372 
3373 	logged = 0;
3374 
3375 	if (override) {
3376 		/*
3377 		 * override is passed when we are
3378 		 * loosing b/w and making one last
3379 		 * gasp at trying to not loose out
3380 		 * to a new-reno flow.
3381 		 */
3382 		goto extra_boost;
3383 	}
3384 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3385 	if (rack->rc_gp_incr &&
3386 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3387 		/*
3388 		 * Reset and get 5 strokes more before the boost. Note
3389 		 * that the count is 0 based so we have to add one.
3390 		 */
3391 extra_boost:
3392 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3393 		rack->rc_gp_timely_inc_cnt = 0;
3394 	} else
3395 		plus = (uint32_t)rack_gp_increase_per;
3396 	/* Must be at least 1% increase for true timely increases */
3397 	if ((plus < 1) &&
3398 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3399 		plus = 1;
3400 	if (rack->rc_gp_saw_rec &&
3401 	    (rack->rc_gp_no_rec_chg == 0) &&
3402 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3403 				  rack->r_ctl.rack_per_of_gp_rec)) {
3404 		/* We have been in recovery ding it too */
3405 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3406 		if (calc > 0xffff)
3407 			calc = 0xffff;
3408 		logged |= 1;
3409 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3410 		if (rack_per_upper_bound_ss &&
3411 		    (rack->rc_dragged_bottom == 0) &&
3412 		    (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3413 			rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3414 	}
3415 	if (rack->rc_gp_saw_ca &&
3416 	    (rack->rc_gp_saw_ss == 0) &&
3417 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3418 				  rack->r_ctl.rack_per_of_gp_ca)) {
3419 		/* In CA */
3420 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3421 		if (calc > 0xffff)
3422 			calc = 0xffff;
3423 		logged |= 2;
3424 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3425 		if (rack_per_upper_bound_ca &&
3426 		    (rack->rc_dragged_bottom == 0) &&
3427 		    (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3428 			rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3429 	}
3430 	if (rack->rc_gp_saw_ss &&
3431 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3432 				  rack->r_ctl.rack_per_of_gp_ss)) {
3433 		/* In SS */
3434 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3435 		if (calc > 0xffff)
3436 			calc = 0xffff;
3437 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3438 		if (rack_per_upper_bound_ss &&
3439 		    (rack->rc_dragged_bottom == 0) &&
3440 		    (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3441 			rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3442 		logged |= 4;
3443 	}
3444 	if (logged &&
3445 	    (rack->rc_gp_incr == 0)){
3446 		/* Go into increment mode */
3447 		rack->rc_gp_incr = 1;
3448 		rack->rc_gp_timely_inc_cnt = 0;
3449 	}
3450 	if (rack->rc_gp_incr &&
3451 	    logged &&
3452 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3453 		rack->rc_gp_timely_inc_cnt++;
3454 	}
3455 	rack_log_timely(rack,  logged, plus, 0, 0,
3456 			__LINE__, 1);
3457 }
3458 
3459 static uint32_t
3460 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3461 {
3462 	/*
3463 	 * norm_grad = rtt_diff / minrtt;
3464 	 * new_per = curper * (1 - B * norm_grad)
3465 	 *
3466 	 * B = rack_gp_decrease_per (default 10%)
3467 	 * rtt_dif = input var current rtt-diff
3468 	 * curper = input var current percentage
3469 	 * minrtt = from rack filter
3470 	 *
3471 	 */
3472 	uint64_t perf;
3473 
3474 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3475 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3476 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
3477 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3478 		     (uint64_t)1000000)) /
3479 		(uint64_t)1000000);
3480 	if (perf > curper) {
3481 		/* TSNH */
3482 		perf = curper - 1;
3483 	}
3484 	return ((uint32_t)perf);
3485 }
3486 
3487 static uint32_t
3488 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3489 {
3490 	/*
3491 	 *                                   highrttthresh
3492 	 * result = curper * (1 - (B * ( 1 -  ------          ))
3493 	 *                                     gp_srtt
3494 	 *
3495 	 * B = rack_gp_decrease_per (default 10%)
3496 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3497 	 */
3498 	uint64_t perf;
3499 	uint32_t highrttthresh;
3500 
3501 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3502 
3503 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3504 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3505 					((uint64_t)highrttthresh * (uint64_t)1000000) /
3506 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3507 	return (perf);
3508 }
3509 
3510 static void
3511 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3512 {
3513 	uint64_t logvar, logvar2, logvar3;
3514 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3515 
3516 	if (rack->rc_gp_incr) {
3517 		/* Turn off increment counting */
3518 		rack->rc_gp_incr = 0;
3519 		rack->rc_gp_timely_inc_cnt = 0;
3520 	}
3521 	ss_red = ca_red = rec_red = 0;
3522 	logged = 0;
3523 	/* Calculate the reduction value */
3524 	if (rtt_diff < 0) {
3525 		rtt_diff *= -1;
3526 	}
3527 	/* Must be at least 1% reduction */
3528 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3529 		/* We have been in recovery ding it too */
3530 		if (timely_says == 2) {
3531 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3532 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3533 			if (alt < new_per)
3534 				val = alt;
3535 			else
3536 				val = new_per;
3537 		} else
3538 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3539 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
3540 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3541 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3542 		} else {
3543 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3544 			rec_red = 0;
3545 		}
3546 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3547 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3548 		logged |= 1;
3549 	}
3550 	if (rack->rc_gp_saw_ss) {
3551 		/* Sent in SS */
3552 		if (timely_says == 2) {
3553 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3554 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3555 			if (alt < new_per)
3556 				val = alt;
3557 			else
3558 				val = new_per;
3559 		} else
3560 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3561 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3562 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3563 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3564 		} else {
3565 			ss_red = new_per;
3566 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3567 			logvar = new_per;
3568 			logvar <<= 32;
3569 			logvar |= alt;
3570 			logvar2 = (uint32_t)rtt;
3571 			logvar2 <<= 32;
3572 			logvar2 |= (uint32_t)rtt_diff;
3573 			logvar3 = rack_gp_rtt_maxmul;
3574 			logvar3 <<= 32;
3575 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3576 			rack_log_timely(rack, timely_says,
3577 					logvar2, logvar3,
3578 					logvar, __LINE__, 10);
3579 		}
3580 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3581 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3582 		logged |= 4;
3583 	} else if (rack->rc_gp_saw_ca) {
3584 		/* Sent in CA */
3585 		if (timely_says == 2) {
3586 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, 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_ca, rtt_diff);
3594 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
3595 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3596 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3597 		} else {
3598 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3599 			ca_red = 0;
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_ca)
3614 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3615 		logged |= 2;
3616 	}
3617 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
3618 		rack->rc_gp_timely_dec_cnt++;
3619 		if (rack_timely_dec_clear &&
3620 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3621 			rack->rc_gp_timely_dec_cnt = 0;
3622 	}
3623 	logvar = ss_red;
3624 	logvar <<= 32;
3625 	logvar |= ca_red;
3626 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3627 			__LINE__, 2);
3628 }
3629 
3630 static void
3631 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3632 		     uint32_t rtt, uint32_t line, uint8_t reas)
3633 {
3634 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3635 		union tcp_log_stackspecific log;
3636 		struct timeval tv;
3637 
3638 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3639 		log.u_bbr.flex1 = line;
3640 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3641 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3642 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3643 		log.u_bbr.flex5 = rtt;
3644 		log.u_bbr.flex6 = rack->rc_highly_buffered;
3645 		log.u_bbr.flex6 <<= 1;
3646 		log.u_bbr.flex6 |= rack->forced_ack;
3647 		log.u_bbr.flex6 <<= 1;
3648 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3649 		log.u_bbr.flex6 <<= 1;
3650 		log.u_bbr.flex6 |= rack->in_probe_rtt;
3651 		log.u_bbr.flex6 <<= 1;
3652 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3653 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3654 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3655 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3656 		log.u_bbr.flex8 = reas;
3657 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3658 		log.u_bbr.delRate = rack_get_bw(rack);
3659 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3660 		log.u_bbr.cur_del_rate <<= 32;
3661 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3662 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3663 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3664 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3665 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3666 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3667 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3668 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3669 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3670 		log.u_bbr.rttProp = us_cts;
3671 		log.u_bbr.rttProp <<= 32;
3672 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3673 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3674 		    &rack->rc_inp->inp_socket->so_rcv,
3675 		    &rack->rc_inp->inp_socket->so_snd,
3676 		    BBR_LOG_RTT_SHRINKS, 0,
3677 		    0, &log, false, &rack->r_ctl.act_rcv_time);
3678 	}
3679 }
3680 
3681 static void
3682 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3683 {
3684 	uint64_t bwdp;
3685 
3686 	bwdp = rack_get_bw(rack);
3687 	bwdp *= (uint64_t)rtt;
3688 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3689 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3690 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3691 		/*
3692 		 * A window protocol must be able to have 4 packets
3693 		 * outstanding as the floor in order to function
3694 		 * (especially considering delayed ack :D).
3695 		 */
3696 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3697 	}
3698 }
3699 
3700 static void
3701 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3702 {
3703 	/**
3704 	 * ProbeRTT is a bit different in rack_pacing than in
3705 	 * BBR. It is like BBR in that it uses the lowering of
3706 	 * the RTT as a signal that we saw something new and
3707 	 * counts from there for how long between. But it is
3708 	 * different in that its quite simple. It does not
3709 	 * play with the cwnd and wait until we get down
3710 	 * to N segments outstanding and hold that for
3711 	 * 200ms. Instead it just sets the pacing reduction
3712 	 * rate to a set percentage (70 by default) and hold
3713 	 * that for a number of recent GP Srtt's.
3714 	 */
3715 	uint32_t segsiz;
3716 
3717 	if (rack->rc_gp_dyn_mul == 0)
3718 		return;
3719 
3720 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3721 		/* We are idle */
3722 		return;
3723 	}
3724 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3725 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3726 		/*
3727 		 * Stop the goodput now, the idea here is
3728 		 * that future measurements with in_probe_rtt
3729 		 * won't register if they are not greater so
3730 		 * we want to get what info (if any) is available
3731 		 * now.
3732 		 */
3733 		rack_do_goodput_measurement(rack->rc_tp, rack,
3734 					    rack->rc_tp->snd_una, __LINE__,
3735 					    RACK_QUALITY_PROBERTT);
3736 	}
3737 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3738 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3739 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3740 		     rack->r_ctl.rc_pace_min_segs);
3741 	rack->in_probe_rtt = 1;
3742 	rack->measure_saw_probe_rtt = 1;
3743 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3744 	rack->r_ctl.rc_time_probertt_starts = 0;
3745 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3746 	if (rack_probertt_use_min_rtt_entry)
3747 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3748 	else
3749 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3750 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3751 			     __LINE__, RACK_RTTS_ENTERPROBE);
3752 }
3753 
3754 static void
3755 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3756 {
3757 	struct rack_sendmap *rsm;
3758 	uint32_t segsiz;
3759 
3760 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3761 		     rack->r_ctl.rc_pace_min_segs);
3762 	rack->in_probe_rtt = 0;
3763 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3764 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3765 		/*
3766 		 * Stop the goodput now, the idea here is
3767 		 * that future measurements with in_probe_rtt
3768 		 * won't register if they are not greater so
3769 		 * we want to get what info (if any) is available
3770 		 * now.
3771 		 */
3772 		rack_do_goodput_measurement(rack->rc_tp, rack,
3773 					    rack->rc_tp->snd_una, __LINE__,
3774 					    RACK_QUALITY_PROBERTT);
3775 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3776 		/*
3777 		 * We don't have enough data to make a measurement.
3778 		 * So lets just stop and start here after exiting
3779 		 * probe-rtt. We probably are not interested in
3780 		 * the results anyway.
3781 		 */
3782 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3783 	}
3784 	/*
3785 	 * Measurements through the current snd_max are going
3786 	 * to be limited by the slower pacing rate.
3787 	 *
3788 	 * We need to mark these as app-limited so we
3789 	 * don't collapse the b/w.
3790 	 */
3791 	rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3792 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3793 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3794 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3795 		else {
3796 			/*
3797 			 * Go out to the end app limited and mark
3798 			 * this new one as next and move the end_appl up
3799 			 * to this guy.
3800 			 */
3801 			if (rack->r_ctl.rc_end_appl)
3802 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3803 			rack->r_ctl.rc_end_appl = rsm;
3804 		}
3805 		rsm->r_flags |= RACK_APP_LIMITED;
3806 		rack->r_ctl.rc_app_limited_cnt++;
3807 	}
3808 	/*
3809 	 * Now, we need to examine our pacing rate multipliers.
3810 	 * If its under 100%, we need to kick it back up to
3811 	 * 100%. We also don't let it be over our "max" above
3812 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3813 	 * Note setting clamp_atexit_prtt to 0 has the effect
3814 	 * of setting CA/SS to 100% always at exit (which is
3815 	 * the default behavior).
3816 	 */
3817 	if (rack_probertt_clear_is) {
3818 		rack->rc_gp_incr = 0;
3819 		rack->rc_gp_bwred = 0;
3820 		rack->rc_gp_timely_inc_cnt = 0;
3821 		rack->rc_gp_timely_dec_cnt = 0;
3822 	}
3823 	/* Do we do any clamping at exit? */
3824 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3825 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3826 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3827 	}
3828 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3829 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3830 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3831 	}
3832 	/*
3833 	 * Lets set rtt_diff to 0, so that we will get a "boost"
3834 	 * after exiting.
3835 	 */
3836 	rack->r_ctl.rc_rtt_diff = 0;
3837 
3838 	/* Clear all flags so we start fresh */
3839 	rack->rc_tp->t_bytes_acked = 0;
3840 	rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3841 	/*
3842 	 * If configured to, set the cwnd and ssthresh to
3843 	 * our targets.
3844 	 */
3845 	if (rack_probe_rtt_sets_cwnd) {
3846 		uint64_t ebdp;
3847 		uint32_t setto;
3848 
3849 		/* Set ssthresh so we get into CA once we hit our target */
3850 		if (rack_probertt_use_min_rtt_exit == 1) {
3851 			/* Set to min rtt */
3852 			rack_set_prtt_target(rack, segsiz,
3853 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3854 		} else if (rack_probertt_use_min_rtt_exit == 2) {
3855 			/* Set to current gp rtt */
3856 			rack_set_prtt_target(rack, segsiz,
3857 					     rack->r_ctl.rc_gp_srtt);
3858 		} else if (rack_probertt_use_min_rtt_exit == 3) {
3859 			/* Set to entry gp rtt */
3860 			rack_set_prtt_target(rack, segsiz,
3861 					     rack->r_ctl.rc_entry_gp_rtt);
3862 		} else {
3863 			uint64_t sum;
3864 			uint32_t setval;
3865 
3866 			sum = rack->r_ctl.rc_entry_gp_rtt;
3867 			sum *= 10;
3868 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3869 			if (sum >= 20) {
3870 				/*
3871 				 * A highly buffered path needs
3872 				 * cwnd space for timely to work.
3873 				 * Lets set things up as if
3874 				 * we are heading back here again.
3875 				 */
3876 				setval = rack->r_ctl.rc_entry_gp_rtt;
3877 			} else if (sum >= 15) {
3878 				/*
3879 				 * Lets take the smaller of the
3880 				 * two since we are just somewhat
3881 				 * buffered.
3882 				 */
3883 				setval = rack->r_ctl.rc_gp_srtt;
3884 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
3885 					setval = rack->r_ctl.rc_entry_gp_rtt;
3886 			} else {
3887 				/*
3888 				 * Here we are not highly buffered
3889 				 * and should pick the min we can to
3890 				 * keep from causing loss.
3891 				 */
3892 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3893 			}
3894 			rack_set_prtt_target(rack, segsiz,
3895 					     setval);
3896 		}
3897 		if (rack_probe_rtt_sets_cwnd > 1) {
3898 			/* There is a percentage here to boost */
3899 			ebdp = rack->r_ctl.rc_target_probertt_flight;
3900 			ebdp *= rack_probe_rtt_sets_cwnd;
3901 			ebdp /= 100;
3902 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3903 		} else
3904 			setto = rack->r_ctl.rc_target_probertt_flight;
3905 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3906 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3907 			/* Enforce a min */
3908 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3909 		}
3910 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
3911 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3912 	}
3913 	rack_log_rtt_shrinks(rack,  us_cts,
3914 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3915 			     __LINE__, RACK_RTTS_EXITPROBE);
3916 	/* Clear times last so log has all the info */
3917 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3918 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3919 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3920 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
3921 }
3922 
3923 static void
3924 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3925 {
3926 	/* Check in on probe-rtt */
3927 	if (rack->rc_gp_filled == 0) {
3928 		/* We do not do p-rtt unless we have gp measurements */
3929 		return;
3930 	}
3931 	if (rack->in_probe_rtt) {
3932 		uint64_t no_overflow;
3933 		uint32_t endtime, must_stay;
3934 
3935 		if (rack->r_ctl.rc_went_idle_time &&
3936 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3937 			/*
3938 			 * We went idle during prtt, just exit now.
3939 			 */
3940 			rack_exit_probertt(rack, us_cts);
3941 		} else if (rack_probe_rtt_safety_val &&
3942 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3943 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3944 			/*
3945 			 * Probe RTT safety value triggered!
3946 			 */
3947 			rack_log_rtt_shrinks(rack,  us_cts,
3948 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3949 					     __LINE__, RACK_RTTS_SAFETY);
3950 			rack_exit_probertt(rack, us_cts);
3951 		}
3952 		/* Calculate the max we will wait */
3953 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3954 		if (rack->rc_highly_buffered)
3955 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3956 		/* Calculate the min we must wait */
3957 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3958 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3959 		    TSTMP_LT(us_cts, endtime)) {
3960 			uint32_t calc;
3961 			/* Do we lower more? */
3962 no_exit:
3963 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3964 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3965 			else
3966 				calc = 0;
3967 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3968 			if (calc) {
3969 				/* Maybe */
3970 				calc *= rack_per_of_gp_probertt_reduce;
3971 				rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3972 				/* Limit it too */
3973 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3974 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3975 			}
3976 			/* We must reach target or the time set */
3977 			return;
3978 		}
3979 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
3980 			if ((TSTMP_LT(us_cts, must_stay) &&
3981 			     rack->rc_highly_buffered) ||
3982 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3983 			      rack->r_ctl.rc_target_probertt_flight)) {
3984 				/* We are not past the must_stay time */
3985 				goto no_exit;
3986 			}
3987 			rack_log_rtt_shrinks(rack,  us_cts,
3988 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3989 					     __LINE__, RACK_RTTS_REACHTARGET);
3990 			rack->r_ctl.rc_time_probertt_starts = us_cts;
3991 			if (rack->r_ctl.rc_time_probertt_starts == 0)
3992 				rack->r_ctl.rc_time_probertt_starts = 1;
3993 			/* Restore back to our rate we want to pace at in prtt */
3994 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3995 		}
3996 		/*
3997 		 * Setup our end time, some number of gp_srtts plus 200ms.
3998 		 */
3999 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
4000 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
4001 		if (rack_probertt_gpsrtt_cnt_div)
4002 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
4003 		else
4004 			endtime = 0;
4005 		endtime += rack_min_probertt_hold;
4006 		endtime += rack->r_ctl.rc_time_probertt_starts;
4007 		if (TSTMP_GEQ(us_cts,  endtime)) {
4008 			/* yes, exit probertt */
4009 			rack_exit_probertt(rack, us_cts);
4010 		}
4011 
4012 	} else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
4013 		/* Go into probertt, its been too long since we went lower */
4014 		rack_enter_probertt(rack, us_cts);
4015 	}
4016 }
4017 
4018 static void
4019 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
4020 		       uint32_t rtt, int32_t rtt_diff)
4021 {
4022 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
4023 	uint32_t losses;
4024 
4025 	if ((rack->rc_gp_dyn_mul == 0) ||
4026 	    (rack->use_fixed_rate) ||
4027 	    (rack->in_probe_rtt) ||
4028 	    (rack->rc_always_pace == 0)) {
4029 		/* No dynamic GP multipler in play */
4030 		return;
4031 	}
4032 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
4033 	cur_bw = rack_get_bw(rack);
4034 	/* Calculate our up and down range */
4035 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
4036 	up_bnd /= 100;
4037 	up_bnd += rack->r_ctl.last_gp_comp_bw;
4038 
4039 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
4040 	subfr /= 100;
4041 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
4042 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
4043 		/*
4044 		 * This is the case where our RTT is above
4045 		 * the max target and we have been configured
4046 		 * to just do timely no bonus up stuff in that case.
4047 		 *
4048 		 * There are two configurations, set to 1, and we
4049 		 * just do timely if we are over our max. If its
4050 		 * set above 1 then we slam the multipliers down
4051 		 * to 100 and then decrement per timely.
4052 		 */
4053 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4054 				__LINE__, 3);
4055 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
4056 			rack_validate_multipliers_at_or_below_100(rack);
4057 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4058 	} else if ((last_bw_est < low_bnd) && !losses) {
4059 		/*
4060 		 * We are decreasing this is a bit complicated this
4061 		 * means we are loosing ground. This could be
4062 		 * because another flow entered and we are competing
4063 		 * for b/w with it. This will push the RTT up which
4064 		 * makes timely unusable unless we want to get shoved
4065 		 * into a corner and just be backed off (the age
4066 		 * old problem with delay based CC).
4067 		 *
4068 		 * On the other hand if it was a route change we
4069 		 * would like to stay somewhat contained and not
4070 		 * blow out the buffers.
4071 		 */
4072 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4073 				__LINE__, 3);
4074 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4075 		if (rack->rc_gp_bwred == 0) {
4076 			/* Go into reduction counting */
4077 			rack->rc_gp_bwred = 1;
4078 			rack->rc_gp_timely_dec_cnt = 0;
4079 		}
4080 		if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
4081 		    (timely_says == 0)) {
4082 			/*
4083 			 * Push another time with a faster pacing
4084 			 * to try to gain back (we include override to
4085 			 * get a full raise factor).
4086 			 */
4087 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4088 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4089 			    (timely_says == 0) ||
4090 			    (rack_down_raise_thresh == 0)) {
4091 				/*
4092 				 * Do an override up in b/w if we were
4093 				 * below the threshold or if the threshold
4094 				 * is zero we always do the raise.
4095 				 */
4096 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4097 			} else {
4098 				/* Log it stays the same */
4099 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4100 						__LINE__, 11);
4101 			}
4102 			rack->rc_gp_timely_dec_cnt++;
4103 			/* We are not incrementing really no-count */
4104 			rack->rc_gp_incr = 0;
4105 			rack->rc_gp_timely_inc_cnt = 0;
4106 		} else {
4107 			/*
4108 			 * Lets just use the RTT
4109 			 * information and give up
4110 			 * pushing.
4111 			 */
4112 			goto use_timely;
4113 		}
4114 	} else if ((timely_says != 2) &&
4115 		    !losses &&
4116 		    (last_bw_est > up_bnd)) {
4117 		/*
4118 		 * We are increasing b/w lets keep going, updating
4119 		 * our b/w and ignoring any timely input, unless
4120 		 * of course we are at our max raise (if there is one).
4121 		 */
4122 
4123 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4124 				__LINE__, 3);
4125 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4126 		if (rack->rc_gp_saw_ss &&
4127 		    rack_per_upper_bound_ss &&
4128 		     (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
4129 			    /*
4130 			     * In cases where we can't go higher
4131 			     * we should just use timely.
4132 			     */
4133 			    goto use_timely;
4134 		}
4135 		if (rack->rc_gp_saw_ca &&
4136 		    rack_per_upper_bound_ca &&
4137 		    (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
4138 			    /*
4139 			     * In cases where we can't go higher
4140 			     * we should just use timely.
4141 			     */
4142 			    goto use_timely;
4143 		}
4144 		rack->rc_gp_bwred = 0;
4145 		rack->rc_gp_timely_dec_cnt = 0;
4146 		/* You get a set number of pushes if timely is trying to reduce */
4147 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4148 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4149 		} else {
4150 			/* Log it stays the same */
4151 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4152 			    __LINE__, 12);
4153 		}
4154 		return;
4155 	} else {
4156 		/*
4157 		 * We are staying between the lower and upper range bounds
4158 		 * so use timely to decide.
4159 		 */
4160 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4161 				__LINE__, 3);
4162 use_timely:
4163 		if (timely_says) {
4164 			rack->rc_gp_incr = 0;
4165 			rack->rc_gp_timely_inc_cnt = 0;
4166 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4167 			    !losses &&
4168 			    (last_bw_est < low_bnd)) {
4169 				/* We are loosing ground */
4170 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4171 				rack->rc_gp_timely_dec_cnt++;
4172 				/* We are not incrementing really no-count */
4173 				rack->rc_gp_incr = 0;
4174 				rack->rc_gp_timely_inc_cnt = 0;
4175 			} else
4176 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4177 		} else {
4178 			rack->rc_gp_bwred = 0;
4179 			rack->rc_gp_timely_dec_cnt = 0;
4180 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4181 		}
4182 	}
4183 }
4184 
4185 static int32_t
4186 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4187 {
4188 	int32_t timely_says;
4189 	uint64_t log_mult, log_rtt_a_diff;
4190 
4191 	log_rtt_a_diff = rtt;
4192 	log_rtt_a_diff <<= 32;
4193 	log_rtt_a_diff |= (uint32_t)rtt_diff;
4194 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4195 		    rack_gp_rtt_maxmul)) {
4196 		/* Reduce the b/w multipler */
4197 		timely_says = 2;
4198 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4199 		log_mult <<= 32;
4200 		log_mult |= prev_rtt;
4201 		rack_log_timely(rack,  timely_says, log_mult,
4202 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4203 				log_rtt_a_diff, __LINE__, 4);
4204 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4205 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4206 			    max(rack_gp_rtt_mindiv , 1)))) {
4207 		/* Increase the b/w multipler */
4208 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4209 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4210 			 max(rack_gp_rtt_mindiv , 1));
4211 		log_mult <<= 32;
4212 		log_mult |= prev_rtt;
4213 		timely_says = 0;
4214 		rack_log_timely(rack,  timely_says, log_mult ,
4215 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4216 				log_rtt_a_diff, __LINE__, 5);
4217 	} else {
4218 		/*
4219 		 * Use a gradient to find it the timely gradient
4220 		 * is:
4221 		 * grad = rc_rtt_diff / min_rtt;
4222 		 *
4223 		 * anything below or equal to 0 will be
4224 		 * a increase indication. Anything above
4225 		 * zero is a decrease. Note we take care
4226 		 * of the actual gradient calculation
4227 		 * in the reduction (its not needed for
4228 		 * increase).
4229 		 */
4230 		log_mult = prev_rtt;
4231 		if (rtt_diff <= 0) {
4232 			/*
4233 			 * Rttdiff is less than zero, increase the
4234 			 * b/w multipler (its 0 or negative)
4235 			 */
4236 			timely_says = 0;
4237 			rack_log_timely(rack,  timely_says, log_mult,
4238 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4239 		} else {
4240 			/* Reduce the b/w multipler */
4241 			timely_says = 1;
4242 			rack_log_timely(rack,  timely_says, log_mult,
4243 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4244 		}
4245 	}
4246 	return (timely_says);
4247 }
4248 
4249 static void
4250 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4251 			    tcp_seq th_ack, int line, uint8_t quality)
4252 {
4253 	uint64_t tim, bytes_ps, ltim, stim, utim;
4254 	uint32_t segsiz, bytes, reqbytes, us_cts;
4255 	int32_t gput, new_rtt_diff, timely_says;
4256 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4257 	int did_add = 0;
4258 
4259 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4260 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4261 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
4262 		tim = us_cts - tp->gput_ts;
4263 	else
4264 		tim = 0;
4265 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4266 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4267 	else
4268 		stim = 0;
4269 	/*
4270 	 * Use the larger of the send time or ack time. This prevents us
4271 	 * from being influenced by ack artifacts to come up with too
4272 	 * high of measurement. Note that since we are spanning over many more
4273 	 * bytes in most of our measurements hopefully that is less likely to
4274 	 * occur.
4275 	 */
4276 	if (tim > stim)
4277 		utim = max(tim, 1);
4278 	else
4279 		utim = max(stim, 1);
4280 	/* Lets get a msec time ltim too for the old stuff */
4281 	ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4282 	gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4283 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4284 	if ((tim == 0) && (stim == 0)) {
4285 		/*
4286 		 * Invalid measurement time, maybe
4287 		 * all on one ack/one send?
4288 		 */
4289 		bytes = 0;
4290 		bytes_ps = 0;
4291 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4292 					   0, 0, 0, 10, __LINE__, NULL, quality);
4293 		goto skip_measurement;
4294 	}
4295 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4296 		/* We never made a us_rtt measurement? */
4297 		bytes = 0;
4298 		bytes_ps = 0;
4299 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4300 					   0, 0, 0, 10, __LINE__, NULL, quality);
4301 		goto skip_measurement;
4302 	}
4303 	/*
4304 	 * Calculate the maximum possible b/w this connection
4305 	 * could have. We base our calculation on the lowest
4306 	 * rtt we have seen during the measurement and the
4307 	 * largest rwnd the client has given us in that time. This
4308 	 * forms a BDP that is the maximum that we could ever
4309 	 * get to the client. Anything larger is not valid.
4310 	 *
4311 	 * I originally had code here that rejected measurements
4312 	 * where the time was less than 1/2 the latest us_rtt.
4313 	 * But after thinking on that I realized its wrong since
4314 	 * say you had a 150Mbps or even 1Gbps link, and you
4315 	 * were a long way away.. example I am in Europe (100ms rtt)
4316 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4317 	 * bytes my time would be 1.2ms, and yet my rtt would say
4318 	 * the measurement was invalid the time was < 50ms. The
4319 	 * same thing is true for 150Mb (8ms of time).
4320 	 *
4321 	 * A better way I realized is to look at what the maximum
4322 	 * the connection could possibly do. This is gated on
4323 	 * the lowest RTT we have seen and the highest rwnd.
4324 	 * We should in theory never exceed that, if we are
4325 	 * then something on the path is storing up packets
4326 	 * and then feeding them all at once to our endpoint
4327 	 * messing up our measurement.
4328 	 */
4329 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4330 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4331 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4332 	if (SEQ_LT(th_ack, tp->gput_seq)) {
4333 		/* No measurement can be made */
4334 		bytes = 0;
4335 		bytes_ps = 0;
4336 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4337 					   0, 0, 0, 10, __LINE__, NULL, quality);
4338 		goto skip_measurement;
4339 	} else
4340 		bytes = (th_ack - tp->gput_seq);
4341 	bytes_ps = (uint64_t)bytes;
4342 	/*
4343 	 * Don't measure a b/w for pacing unless we have gotten at least
4344 	 * an initial windows worth of data in this measurement interval.
4345 	 *
4346 	 * Small numbers of bytes get badly influenced by delayed ack and
4347 	 * other artifacts. Note we take the initial window or our
4348 	 * defined minimum GP (defaulting to 10 which hopefully is the
4349 	 * IW).
4350 	 */
4351 	if (rack->rc_gp_filled == 0) {
4352 		/*
4353 		 * The initial estimate is special. We
4354 		 * have blasted out an IW worth of packets
4355 		 * without a real valid ack ts results. We
4356 		 * then setup the app_limited_needs_set flag,
4357 		 * this should get the first ack in (probably 2
4358 		 * MSS worth) to be recorded as the timestamp.
4359 		 * We thus allow a smaller number of bytes i.e.
4360 		 * IW - 2MSS.
4361 		 */
4362 		reqbytes -= (2 * segsiz);
4363 		/* Also lets fill previous for our first measurement to be neutral */
4364 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4365 	}
4366 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4367 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4368 					   rack->r_ctl.rc_app_limited_cnt,
4369 					   0, 0, 10, __LINE__, NULL, quality);
4370 		goto skip_measurement;
4371 	}
4372 	/*
4373 	 * We now need to calculate the Timely like status so
4374 	 * we can update (possibly) the b/w multipliers.
4375 	 */
4376 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4377 	if (rack->rc_gp_filled == 0) {
4378 		/* No previous reading */
4379 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4380 	} else {
4381 		if (rack->measure_saw_probe_rtt == 0) {
4382 			/*
4383 			 * We don't want a probertt to be counted
4384 			 * since it will be negative incorrectly. We
4385 			 * expect to be reducing the RTT when we
4386 			 * pace at a slower rate.
4387 			 */
4388 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4389 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4390 		}
4391 	}
4392 	timely_says = rack_make_timely_judgement(rack,
4393 		rack->r_ctl.rc_gp_srtt,
4394 		rack->r_ctl.rc_rtt_diff,
4395 	        rack->r_ctl.rc_prev_gp_srtt
4396 		);
4397 	bytes_ps *= HPTS_USEC_IN_SEC;
4398 	bytes_ps /= utim;
4399 	if (bytes_ps > rack->r_ctl.last_max_bw) {
4400 		/*
4401 		 * Something is on path playing
4402 		 * since this b/w is not possible based
4403 		 * on our BDP (highest rwnd and lowest rtt
4404 		 * we saw in the measurement window).
4405 		 *
4406 		 * Another option here would be to
4407 		 * instead skip the measurement.
4408 		 */
4409 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4410 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
4411 					   11, __LINE__, NULL, quality);
4412 		bytes_ps = rack->r_ctl.last_max_bw;
4413 	}
4414 	/* We store gp for b/w in bytes per second */
4415 	if (rack->rc_gp_filled == 0) {
4416 		/* Initial measurment */
4417 		if (bytes_ps) {
4418 			rack->r_ctl.gp_bw = bytes_ps;
4419 			rack->rc_gp_filled = 1;
4420 			rack->r_ctl.num_measurements = 1;
4421 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4422 		} else {
4423 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4424 						   rack->r_ctl.rc_app_limited_cnt,
4425 						   0, 0, 10, __LINE__, NULL, quality);
4426 		}
4427 		if (rack->rc_inp->inp_in_hpts &&
4428 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4429 			/*
4430 			 * Ok we can't trust the pacer in this case
4431 			 * where we transition from un-paced to paced.
4432 			 * Or for that matter when the burst mitigation
4433 			 * was making a wild guess and got it wrong.
4434 			 * Stop the pacer and clear up all the aggregate
4435 			 * delays etc.
4436 			 */
4437 			tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
4438 			rack->r_ctl.rc_hpts_flags = 0;
4439 			rack->r_ctl.rc_last_output_to = 0;
4440 		}
4441 		did_add = 2;
4442 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4443 		/* Still a small number run an average */
4444 		rack->r_ctl.gp_bw += bytes_ps;
4445 		addpart = rack->r_ctl.num_measurements;
4446 		rack->r_ctl.num_measurements++;
4447 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4448 			/* We have collected enought to move forward */
4449 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4450 		}
4451 		did_add = 3;
4452 	} else {
4453 		/*
4454 		 * We want to take 1/wma of the goodput and add in to 7/8th
4455 		 * of the old value weighted by the srtt. So if your measurement
4456 		 * period is say 2 SRTT's long you would get 1/4 as the
4457 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
4458 		 *
4459 		 * But we must be careful not to take too much i.e. if the
4460 		 * srtt is say 20ms and the measurement is taken over
4461 		 * 400ms our weight would be 400/20 i.e. 20. On the
4462 		 * other hand if we get a measurement over 1ms with a
4463 		 * 10ms rtt we only want to take a much smaller portion.
4464 		 */
4465 		if (rack->r_ctl.num_measurements < 0xff) {
4466 			rack->r_ctl.num_measurements++;
4467 		}
4468 		srtt = (uint64_t)tp->t_srtt;
4469 		if (srtt == 0) {
4470 			/*
4471 			 * Strange why did t_srtt go back to zero?
4472 			 */
4473 			if (rack->r_ctl.rc_rack_min_rtt)
4474 				srtt = rack->r_ctl.rc_rack_min_rtt;
4475 			else
4476 				srtt = HPTS_USEC_IN_MSEC;
4477 		}
4478 		/*
4479 		 * XXXrrs: Note for reviewers, in playing with
4480 		 * dynamic pacing I discovered this GP calculation
4481 		 * as done originally leads to some undesired results.
4482 		 * Basically you can get longer measurements contributing
4483 		 * too much to the WMA. Thus I changed it if you are doing
4484 		 * dynamic adjustments to only do the aportioned adjustment
4485 		 * if we have a very small (time wise) measurement. Longer
4486 		 * measurements just get there weight (defaulting to 1/8)
4487 		 * add to the WMA. We may want to think about changing
4488 		 * this to always do that for both sides i.e. dynamic
4489 		 * and non-dynamic... but considering lots of folks
4490 		 * were playing with this I did not want to change the
4491 		 * calculation per.se. without your thoughts.. Lawerence?
4492 		 * Peter??
4493 		 */
4494 		if (rack->rc_gp_dyn_mul == 0) {
4495 			subpart = rack->r_ctl.gp_bw * utim;
4496 			subpart /= (srtt * 8);
4497 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
4498 				/*
4499 				 * The b/w update takes no more
4500 				 * away then 1/2 our running total
4501 				 * so factor it in.
4502 				 */
4503 				addpart = bytes_ps * utim;
4504 				addpart /= (srtt * 8);
4505 			} else {
4506 				/*
4507 				 * Don't allow a single measurement
4508 				 * to account for more than 1/2 of the
4509 				 * WMA. This could happen on a retransmission
4510 				 * where utim becomes huge compared to
4511 				 * srtt (multiple retransmissions when using
4512 				 * the sending rate which factors in all the
4513 				 * transmissions from the first one).
4514 				 */
4515 				subpart = rack->r_ctl.gp_bw / 2;
4516 				addpart = bytes_ps / 2;
4517 			}
4518 			resid_bw = rack->r_ctl.gp_bw - subpart;
4519 			rack->r_ctl.gp_bw = resid_bw + addpart;
4520 			did_add = 1;
4521 		} else {
4522 			if ((utim / srtt) <= 1) {
4523 				/*
4524 				 * The b/w update was over a small period
4525 				 * of time. The idea here is to prevent a small
4526 				 * measurement time period from counting
4527 				 * too much. So we scale it based on the
4528 				 * time so it attributes less than 1/rack_wma_divisor
4529 				 * of its measurement.
4530 				 */
4531 				subpart = rack->r_ctl.gp_bw * utim;
4532 				subpart /= (srtt * rack_wma_divisor);
4533 				addpart = bytes_ps * utim;
4534 				addpart /= (srtt * rack_wma_divisor);
4535 			} else {
4536 				/*
4537 				 * The scaled measurement was long
4538 				 * enough so lets just add in the
4539 				 * portion of the measurment i.e. 1/rack_wma_divisor
4540 				 */
4541 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4542 				addpart = bytes_ps / rack_wma_divisor;
4543 			}
4544 			if ((rack->measure_saw_probe_rtt == 0) ||
4545 		            (bytes_ps > rack->r_ctl.gp_bw)) {
4546 				/*
4547 				 * For probe-rtt we only add it in
4548 				 * if its larger, all others we just
4549 				 * add in.
4550 				 */
4551 				did_add = 1;
4552 				resid_bw = rack->r_ctl.gp_bw - subpart;
4553 				rack->r_ctl.gp_bw = resid_bw + addpart;
4554 			}
4555 		}
4556 	}
4557 	if ((rack->gp_ready == 0) &&
4558 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4559 		/* We have enough measurements now */
4560 		rack->gp_ready = 1;
4561 		rack_set_cc_pacing(rack);
4562 		if (rack->defer_options)
4563 			rack_apply_deferred_options(rack);
4564 	}
4565 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4566 				   rack_get_bw(rack), 22, did_add, NULL, quality);
4567 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
4568 	if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4569 		rack_update_multiplier(rack, timely_says, bytes_ps,
4570 				       rack->r_ctl.rc_gp_srtt,
4571 				       rack->r_ctl.rc_rtt_diff);
4572 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4573 				   rack_get_bw(rack), 3, line, NULL, quality);
4574 	/* reset the gp srtt and setup the new prev */
4575 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4576 	/* Record the lost count for the next measurement */
4577 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4578 	/*
4579 	 * We restart our diffs based on the gpsrtt in the
4580 	 * measurement window.
4581 	 */
4582 	rack->rc_gp_rtt_set = 0;
4583 	rack->rc_gp_saw_rec = 0;
4584 	rack->rc_gp_saw_ca = 0;
4585 	rack->rc_gp_saw_ss = 0;
4586 	rack->rc_dragged_bottom = 0;
4587 skip_measurement:
4588 
4589 #ifdef STATS
4590 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4591 				 gput);
4592 	/*
4593 	 * XXXLAS: This is a temporary hack, and should be
4594 	 * chained off VOI_TCP_GPUT when stats(9) grows an
4595 	 * API to deal with chained VOIs.
4596 	 */
4597 	if (tp->t_stats_gput_prev > 0)
4598 		stats_voi_update_abs_s32(tp->t_stats,
4599 					 VOI_TCP_GPUT_ND,
4600 					 ((gput - tp->t_stats_gput_prev) * 100) /
4601 					 tp->t_stats_gput_prev);
4602 #endif
4603 	tp->t_flags &= ~TF_GPUTINPROG;
4604 	tp->t_stats_gput_prev = gput;
4605 	/*
4606 	 * Now are we app limited now and there is space from where we
4607 	 * were to where we want to go?
4608 	 *
4609 	 * We don't do the other case i.e. non-applimited here since
4610 	 * the next send will trigger us picking up the missing data.
4611 	 */
4612 	if (rack->r_ctl.rc_first_appl &&
4613 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
4614 	    rack->r_ctl.rc_app_limited_cnt &&
4615 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4616 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
4617 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4618 		/*
4619 		 * Yep there is enough outstanding to make a measurement here.
4620 		 */
4621 		struct rack_sendmap *rsm, fe;
4622 
4623 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4624 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4625 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4626 		rack->app_limited_needs_set = 0;
4627 		tp->gput_seq = th_ack;
4628 		if (rack->in_probe_rtt)
4629 			rack->measure_saw_probe_rtt = 1;
4630 		else if ((rack->measure_saw_probe_rtt) &&
4631 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4632 			rack->measure_saw_probe_rtt = 0;
4633 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
4634 			/* There is a full window to gain info from */
4635 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4636 		} else {
4637 			/* We can only measure up to the applimited point */
4638 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
4639 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
4640 				/*
4641 				 * We don't have enough to make a measurement.
4642 				 */
4643 				tp->t_flags &= ~TF_GPUTINPROG;
4644 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
4645 							   0, 0, 0, 6, __LINE__, NULL, quality);
4646 				return;
4647 			}
4648 		}
4649 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
4650 			/*
4651 			 * We will get no more data into the SB
4652 			 * this means we need to have the data available
4653 			 * before we start a measurement.
4654 			 */
4655 			if (sbavail(&tp->t_inpcb->inp_socket->so_snd) < (tp->gput_ack - tp->gput_seq)) {
4656 				/* Nope not enough data. */
4657 				return;
4658 			}
4659 		}
4660 		tp->t_flags |= TF_GPUTINPROG;
4661 		/*
4662 		 * Now we need to find the timestamp of the send at tp->gput_seq
4663 		 * for the send based measurement.
4664 		 */
4665 		fe.r_start = tp->gput_seq;
4666 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4667 		if (rsm) {
4668 			/* Ok send-based limit is set */
4669 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4670 				/*
4671 				 * Move back to include the earlier part
4672 				 * so our ack time lines up right (this may
4673 				 * make an overlapping measurement but thats
4674 				 * ok).
4675 				 */
4676 				tp->gput_seq = rsm->r_start;
4677 			}
4678 			if (rsm->r_flags & RACK_ACKED)
4679 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4680 			else
4681 				rack->app_limited_needs_set = 1;
4682 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4683 		} else {
4684 			/*
4685 			 * If we don't find the rsm due to some
4686 			 * send-limit set the current time, which
4687 			 * basically disables the send-limit.
4688 			 */
4689 			struct timeval tv;
4690 
4691 			microuptime(&tv);
4692 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4693 		}
4694 		rack_log_pacing_delay_calc(rack,
4695 					   tp->gput_seq,
4696 					   tp->gput_ack,
4697 					   (uint64_t)rsm,
4698 					   tp->gput_ts,
4699 					   rack->r_ctl.rc_app_limited_cnt,
4700 					   9,
4701 					   __LINE__, NULL, quality);
4702 	}
4703 }
4704 
4705 /*
4706  * CC wrapper hook functions
4707  */
4708 static void
4709 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4710     uint16_t type, int32_t recovery)
4711 {
4712 	uint32_t prior_cwnd, acked;
4713 	struct tcp_log_buffer *lgb = NULL;
4714 	uint8_t labc_to_use, quality;
4715 
4716 	INP_WLOCK_ASSERT(tp->t_inpcb);
4717 	tp->ccv->nsegs = nsegs;
4718 	acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una);
4719 	if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4720 		uint32_t max;
4721 
4722 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4723 		if (tp->ccv->bytes_this_ack > max) {
4724 			tp->ccv->bytes_this_ack = max;
4725 		}
4726 	}
4727 #ifdef STATS
4728 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4729 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4730 #endif
4731 	quality = RACK_QUALITY_NONE;
4732 	if ((tp->t_flags & TF_GPUTINPROG) &&
4733 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
4734 		/* Measure the Goodput */
4735 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
4736 #ifdef NETFLIX_PEAKRATE
4737 		if ((type == CC_ACK) &&
4738 		    (tp->t_maxpeakrate)) {
4739 			/*
4740 			 * We update t_peakrate_thr. This gives us roughly
4741 			 * one update per round trip time. Note
4742 			 * it will only be used if pace_always is off i.e
4743 			 * we don't do this for paced flows.
4744 			 */
4745 			rack_update_peakrate_thr(tp);
4746 		}
4747 #endif
4748 	}
4749 	/* Which way our we limited, if not cwnd limited no advance in CA */
4750 	if (tp->snd_cwnd <= tp->snd_wnd)
4751 		tp->ccv->flags |= CCF_CWND_LIMITED;
4752 	else
4753 		tp->ccv->flags &= ~CCF_CWND_LIMITED;
4754 	if (tp->snd_cwnd > tp->snd_ssthresh) {
4755 		tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
4756 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4757 		/* For the setting of a window past use the actual scwnd we are using */
4758 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4759 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4760 			tp->ccv->flags |= CCF_ABC_SENTAWND;
4761 		}
4762 	} else {
4763 		tp->ccv->flags &= ~CCF_ABC_SENTAWND;
4764 		tp->t_bytes_acked = 0;
4765 	}
4766 	prior_cwnd = tp->snd_cwnd;
4767 	if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4768 	    (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4769 		labc_to_use = rack->rc_labc;
4770 	else
4771 		labc_to_use = rack_max_abc_post_recovery;
4772 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4773 		union tcp_log_stackspecific log;
4774 		struct timeval tv;
4775 
4776 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4777 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4778 		log.u_bbr.flex1 = th_ack;
4779 		log.u_bbr.flex2 = tp->ccv->flags;
4780 		log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4781 		log.u_bbr.flex4 = tp->ccv->nsegs;
4782 		log.u_bbr.flex5 = labc_to_use;
4783 		log.u_bbr.flex6 = prior_cwnd;
4784 		log.u_bbr.flex7 = V_tcp_do_newsack;
4785 		log.u_bbr.flex8 = 1;
4786 		lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4787 				     0, &log, false, NULL, NULL, 0, &tv);
4788 	}
4789 	if (CC_ALGO(tp)->ack_received != NULL) {
4790 		/* XXXLAS: Find a way to live without this */
4791 		tp->ccv->curack = th_ack;
4792 		tp->ccv->labc = labc_to_use;
4793 		tp->ccv->flags |= CCF_USE_LOCAL_ABC;
4794 		CC_ALGO(tp)->ack_received(tp->ccv, type);
4795 	}
4796 	if (lgb) {
4797 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4798 	}
4799 	if (rack->r_must_retran) {
4800 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4801 			/*
4802 			 * We now are beyond the rxt point so lets disable
4803 			 * the flag.
4804 			 */
4805 			rack->r_ctl.rc_out_at_rto = 0;
4806 			rack->r_must_retran = 0;
4807 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4808 			/*
4809 			 * Only decrement the rc_out_at_rto if the cwnd advances
4810 			 * at least a whole segment. Otherwise next time the peer
4811 			 * acks, we won't be able to send this generaly happens
4812 			 * when we are in Congestion Avoidance.
4813 			 */
4814 			if (acked <= rack->r_ctl.rc_out_at_rto){
4815 				rack->r_ctl.rc_out_at_rto -= acked;
4816 			} else {
4817 				rack->r_ctl.rc_out_at_rto = 0;
4818 			}
4819 		}
4820 	}
4821 #ifdef STATS
4822 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4823 #endif
4824 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4825 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4826 	}
4827 #ifdef NETFLIX_PEAKRATE
4828 	/* we enforce max peak rate if it is set and we are not pacing */
4829 	if ((rack->rc_always_pace == 0) &&
4830 	    tp->t_peakrate_thr &&
4831 	    (tp->snd_cwnd > tp->t_peakrate_thr)) {
4832 		tp->snd_cwnd = tp->t_peakrate_thr;
4833 	}
4834 #endif
4835 }
4836 
4837 static void
4838 tcp_rack_partialack(struct tcpcb *tp)
4839 {
4840 	struct tcp_rack *rack;
4841 
4842 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4843 	INP_WLOCK_ASSERT(tp->t_inpcb);
4844 	/*
4845 	 * If we are doing PRR and have enough
4846 	 * room to send <or> we are pacing and prr
4847 	 * is disabled we will want to see if we
4848 	 * can send data (by setting r_wanted_output to
4849 	 * true).
4850 	 */
4851 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4852 	    rack->rack_no_prr)
4853 		rack->r_wanted_output = 1;
4854 }
4855 
4856 static void
4857 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4858 {
4859 	struct tcp_rack *rack;
4860 	uint32_t orig_cwnd;
4861 
4862 	orig_cwnd = tp->snd_cwnd;
4863 	INP_WLOCK_ASSERT(tp->t_inpcb);
4864 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4865 	/* only alert CC if we alerted when we entered */
4866 	if (CC_ALGO(tp)->post_recovery != NULL) {
4867 		tp->ccv->curack = th_ack;
4868 		CC_ALGO(tp)->post_recovery(tp->ccv);
4869 		if (tp->snd_cwnd < tp->snd_ssthresh) {
4870 			/*
4871 			 * Rack has burst control and pacing
4872 			 * so lets not set this any lower than
4873 			 * snd_ssthresh per RFC-6582 (option 2).
4874 			 */
4875 			tp->snd_cwnd = tp->snd_ssthresh;
4876 		}
4877 	}
4878 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4879 		union tcp_log_stackspecific log;
4880 		struct timeval tv;
4881 
4882 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4883 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4884 		log.u_bbr.flex1 = th_ack;
4885 		log.u_bbr.flex2 = tp->ccv->flags;
4886 		log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4887 		log.u_bbr.flex4 = tp->ccv->nsegs;
4888 		log.u_bbr.flex5 = V_tcp_abc_l_var;
4889 		log.u_bbr.flex6 = orig_cwnd;
4890 		log.u_bbr.flex7 = V_tcp_do_newsack;
4891 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4892 		log.u_bbr.flex8 = 2;
4893 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4894 			       0, &log, false, NULL, NULL, 0, &tv);
4895 	}
4896 	if ((rack->rack_no_prr == 0) &&
4897 	    (rack->no_prr_addback == 0) &&
4898 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
4899 		/*
4900 		 * Suck the next prr cnt back into cwnd, but
4901 		 * only do that if we are not application limited.
4902 		 */
4903 		if (ctf_outstanding(tp) <= sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
4904 			/*
4905 			 * We are allowed to add back to the cwnd the amount we did
4906 			 * not get out if:
4907 			 * a) no_prr_addback is off.
4908 			 * b) we are not app limited
4909 			 * c) we are doing prr
4910 			 * <and>
4911 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4912 			 */
4913 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4914 					    rack->r_ctl.rc_prr_sndcnt);
4915 		}
4916 		rack->r_ctl.rc_prr_sndcnt = 0;
4917 		rack_log_to_prr(rack, 1, 0);
4918 	}
4919 	rack_log_to_prr(rack, 14, orig_cwnd);
4920 	tp->snd_recover = tp->snd_una;
4921 	if (rack->r_ctl.dsack_persist) {
4922 		rack->r_ctl.dsack_persist--;
4923 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
4924 			rack->r_ctl.num_dsack = 0;
4925 		}
4926 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
4927 	}
4928 	EXIT_RECOVERY(tp->t_flags);
4929 }
4930 
4931 static void
4932 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack)
4933 {
4934 	struct tcp_rack *rack;
4935 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4936 
4937 	INP_WLOCK_ASSERT(tp->t_inpcb);
4938 #ifdef STATS
4939 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4940 #endif
4941 	if (IN_RECOVERY(tp->t_flags) == 0) {
4942 		in_rec_at_entry = 0;
4943 		ssthresh_enter = tp->snd_ssthresh;
4944 		cwnd_enter = tp->snd_cwnd;
4945 	} else
4946 		in_rec_at_entry = 1;
4947 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4948 	switch (type) {
4949 	case CC_NDUPACK:
4950 		tp->t_flags &= ~TF_WASFRECOVERY;
4951 		tp->t_flags &= ~TF_WASCRECOVERY;
4952 		if (!IN_FASTRECOVERY(tp->t_flags)) {
4953 			rack->r_ctl.rc_prr_delivered = 0;
4954 			rack->r_ctl.rc_prr_out = 0;
4955 			if (rack->rack_no_prr == 0) {
4956 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4957 				rack_log_to_prr(rack, 2, in_rec_at_entry);
4958 			}
4959 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4960 			tp->snd_recover = tp->snd_max;
4961 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4962 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4963 		}
4964 		break;
4965 	case CC_ECN:
4966 		if (!IN_CONGRECOVERY(tp->t_flags) ||
4967 		    /*
4968 		     * Allow ECN reaction on ACK to CWR, if
4969 		     * that data segment was also CE marked.
4970 		     */
4971 		    SEQ_GEQ(ack, tp->snd_recover)) {
4972 			EXIT_CONGRECOVERY(tp->t_flags);
4973 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4974 			tp->snd_recover = tp->snd_max + 1;
4975 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4976 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4977 		}
4978 		break;
4979 	case CC_RTO:
4980 		tp->t_dupacks = 0;
4981 		tp->t_bytes_acked = 0;
4982 		EXIT_RECOVERY(tp->t_flags);
4983 		tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4984 		    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4985 		orig_cwnd = tp->snd_cwnd;
4986 		tp->snd_cwnd = ctf_fixed_maxseg(tp);
4987 		rack_log_to_prr(rack, 16, orig_cwnd);
4988 		if (tp->t_flags2 & TF2_ECN_PERMIT)
4989 			tp->t_flags2 |= TF2_ECN_SND_CWR;
4990 		break;
4991 	case CC_RTO_ERR:
4992 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4993 		/* RTO was unnecessary, so reset everything. */
4994 		tp->snd_cwnd = tp->snd_cwnd_prev;
4995 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
4996 		tp->snd_recover = tp->snd_recover_prev;
4997 		if (tp->t_flags & TF_WASFRECOVERY) {
4998 			ENTER_FASTRECOVERY(tp->t_flags);
4999 			tp->t_flags &= ~TF_WASFRECOVERY;
5000 		}
5001 		if (tp->t_flags & TF_WASCRECOVERY) {
5002 			ENTER_CONGRECOVERY(tp->t_flags);
5003 			tp->t_flags &= ~TF_WASCRECOVERY;
5004 		}
5005 		tp->snd_nxt = tp->snd_max;
5006 		tp->t_badrxtwin = 0;
5007 		break;
5008 	}
5009 	if ((CC_ALGO(tp)->cong_signal != NULL)  &&
5010 	    (type != CC_RTO)){
5011 		tp->ccv->curack = ack;
5012 		CC_ALGO(tp)->cong_signal(tp->ccv, type);
5013 	}
5014 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
5015 		rack_log_to_prr(rack, 15, cwnd_enter);
5016 		rack->r_ctl.dsack_byte_cnt = 0;
5017 		rack->r_ctl.retran_during_recovery = 0;
5018 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
5019 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
5020 		rack->r_ent_rec_ns = 1;
5021 	}
5022 }
5023 
5024 static inline void
5025 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
5026 {
5027 	uint32_t i_cwnd;
5028 
5029 	INP_WLOCK_ASSERT(tp->t_inpcb);
5030 
5031 #ifdef NETFLIX_STATS
5032 	KMOD_TCPSTAT_INC(tcps_idle_restarts);
5033 	if (tp->t_state == TCPS_ESTABLISHED)
5034 		KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
5035 #endif
5036 	if (CC_ALGO(tp)->after_idle != NULL)
5037 		CC_ALGO(tp)->after_idle(tp->ccv);
5038 
5039 	if (tp->snd_cwnd == 1)
5040 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
5041 	else
5042 		i_cwnd = rc_init_window(rack);
5043 
5044 	/*
5045 	 * Being idle is no differnt than the initial window. If the cc
5046 	 * clamps it down below the initial window raise it to the initial
5047 	 * window.
5048 	 */
5049 	if (tp->snd_cwnd < i_cwnd) {
5050 		tp->snd_cwnd = i_cwnd;
5051 	}
5052 }
5053 
5054 /*
5055  * Indicate whether this ack should be delayed.  We can delay the ack if
5056  * following conditions are met:
5057  *	- There is no delayed ack timer in progress.
5058  *	- Our last ack wasn't a 0-sized window. We never want to delay
5059  *	  the ack that opens up a 0-sized window.
5060  *	- LRO wasn't used for this segment. We make sure by checking that the
5061  *	  segment size is not larger than the MSS.
5062  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
5063  *	  connection.
5064  */
5065 #define DELAY_ACK(tp, tlen)			 \
5066 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
5067 	((tp->t_flags & TF_DELACK) == 0) &&	 \
5068 	(tlen <= tp->t_maxseg) &&		 \
5069 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
5070 
5071 static struct rack_sendmap *
5072 rack_find_lowest_rsm(struct tcp_rack *rack)
5073 {
5074 	struct rack_sendmap *rsm;
5075 
5076 	/*
5077 	 * Walk the time-order transmitted list looking for an rsm that is
5078 	 * not acked. This will be the one that was sent the longest time
5079 	 * ago that is still outstanding.
5080 	 */
5081 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
5082 		if (rsm->r_flags & RACK_ACKED) {
5083 			continue;
5084 		}
5085 		goto finish;
5086 	}
5087 finish:
5088 	return (rsm);
5089 }
5090 
5091 static struct rack_sendmap *
5092 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
5093 {
5094 	struct rack_sendmap *prsm;
5095 
5096 	/*
5097 	 * Walk the sequence order list backward until we hit and arrive at
5098 	 * the highest seq not acked. In theory when this is called it
5099 	 * should be the last segment (which it was not).
5100 	 */
5101 	counter_u64_add(rack_find_high, 1);
5102 	prsm = rsm;
5103 	RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
5104 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
5105 			continue;
5106 		}
5107 		return (prsm);
5108 	}
5109 	return (NULL);
5110 }
5111 
5112 static uint32_t
5113 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
5114 {
5115 	int32_t lro;
5116 	uint32_t thresh;
5117 
5118 	/*
5119 	 * lro is the flag we use to determine if we have seen reordering.
5120 	 * If it gets set we have seen reordering. The reorder logic either
5121 	 * works in one of two ways:
5122 	 *
5123 	 * If reorder-fade is configured, then we track the last time we saw
5124 	 * re-ordering occur. If we reach the point where enough time as
5125 	 * passed we no longer consider reordering has occuring.
5126 	 *
5127 	 * Or if reorder-face is 0, then once we see reordering we consider
5128 	 * the connection to alway be subject to reordering and just set lro
5129 	 * to 1.
5130 	 *
5131 	 * In the end if lro is non-zero we add the extra time for
5132 	 * reordering in.
5133 	 */
5134 	if (srtt == 0)
5135 		srtt = 1;
5136 	if (rack->r_ctl.rc_reorder_ts) {
5137 		if (rack->r_ctl.rc_reorder_fade) {
5138 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5139 				lro = cts - rack->r_ctl.rc_reorder_ts;
5140 				if (lro == 0) {
5141 					/*
5142 					 * No time as passed since the last
5143 					 * reorder, mark it as reordering.
5144 					 */
5145 					lro = 1;
5146 				}
5147 			} else {
5148 				/* Negative time? */
5149 				lro = 0;
5150 			}
5151 			if (lro > rack->r_ctl.rc_reorder_fade) {
5152 				/* Turn off reordering seen too */
5153 				rack->r_ctl.rc_reorder_ts = 0;
5154 				lro = 0;
5155 			}
5156 		} else {
5157 			/* Reodering does not fade */
5158 			lro = 1;
5159 		}
5160 	} else {
5161 		lro = 0;
5162 	}
5163 	if (rack->rc_rack_tmr_std_based == 0) {
5164 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
5165 	} else {
5166 		/* Standards based pkt-delay is 1/4 srtt */
5167 		thresh = srtt +  (srtt >> 2);
5168 	}
5169 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
5170 		/* It must be set, if not you get 1/4 rtt */
5171 		if (rack->r_ctl.rc_reorder_shift)
5172 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5173 		else
5174 			thresh += (srtt >> 2);
5175 	}
5176 	if (rack->rc_rack_use_dsack &&
5177 	    lro &&
5178 	    (rack->r_ctl.num_dsack > 0)) {
5179 		/*
5180 		 * We only increase the reordering window if we
5181 		 * have seen reordering <and> we have a DSACK count.
5182 		 */
5183 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
5184 		rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
5185 	}
5186 	/* SRTT * 2 is the ceiling */
5187 	if (thresh > (srtt * 2)) {
5188 		thresh = srtt * 2;
5189 	}
5190 	/* And we don't want it above the RTO max either */
5191 	if (thresh > rack_rto_max) {
5192 		thresh = rack_rto_max;
5193 	}
5194 	rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
5195 	return (thresh);
5196 }
5197 
5198 static uint32_t
5199 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5200 		     struct rack_sendmap *rsm, uint32_t srtt)
5201 {
5202 	struct rack_sendmap *prsm;
5203 	uint32_t thresh, len;
5204 	int segsiz;
5205 
5206 	if (srtt == 0)
5207 		srtt = 1;
5208 	if (rack->r_ctl.rc_tlp_threshold)
5209 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5210 	else
5211 		thresh = (srtt * 2);
5212 
5213 	/* Get the previous sent packet, if any */
5214 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5215 	counter_u64_add(rack_enter_tlp_calc, 1);
5216 	len = rsm->r_end - rsm->r_start;
5217 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5218 		/* Exactly like the ID */
5219 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5220 			uint32_t alt_thresh;
5221 			/*
5222 			 * Compensate for delayed-ack with the d-ack time.
5223 			 */
5224 			counter_u64_add(rack_used_tlpmethod, 1);
5225 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5226 			if (alt_thresh > thresh)
5227 				thresh = alt_thresh;
5228 		}
5229 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5230 		/* 2.1 behavior */
5231 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5232 		if (prsm && (len <= segsiz)) {
5233 			/*
5234 			 * Two packets outstanding, thresh should be (2*srtt) +
5235 			 * possible inter-packet delay (if any).
5236 			 */
5237 			uint32_t inter_gap = 0;
5238 			int idx, nidx;
5239 
5240 			counter_u64_add(rack_used_tlpmethod, 1);
5241 			idx = rsm->r_rtr_cnt - 1;
5242 			nidx = prsm->r_rtr_cnt - 1;
5243 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5244 				/* Yes it was sent later (or at the same time) */
5245 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5246 			}
5247 			thresh += inter_gap;
5248 		} else if (len <= segsiz) {
5249 			/*
5250 			 * Possibly compensate for delayed-ack.
5251 			 */
5252 			uint32_t alt_thresh;
5253 
5254 			counter_u64_add(rack_used_tlpmethod2, 1);
5255 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5256 			if (alt_thresh > thresh)
5257 				thresh = alt_thresh;
5258 		}
5259 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5260 		/* 2.2 behavior */
5261 		if (len <= segsiz) {
5262 			uint32_t alt_thresh;
5263 			/*
5264 			 * Compensate for delayed-ack with the d-ack time.
5265 			 */
5266 			counter_u64_add(rack_used_tlpmethod, 1);
5267 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5268 			if (alt_thresh > thresh)
5269 				thresh = alt_thresh;
5270 		}
5271 	}
5272 	/* Not above an RTO */
5273 	if (thresh > tp->t_rxtcur) {
5274 		thresh = tp->t_rxtcur;
5275 	}
5276 	/* Not above a RTO max */
5277 	if (thresh > rack_rto_max) {
5278 		thresh = rack_rto_max;
5279 	}
5280 	/* Apply user supplied min TLP */
5281 	if (thresh < rack_tlp_min) {
5282 		thresh = rack_tlp_min;
5283 	}
5284 	return (thresh);
5285 }
5286 
5287 static uint32_t
5288 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5289 {
5290 	/*
5291 	 * We want the rack_rtt which is the
5292 	 * last rtt we measured. However if that
5293 	 * does not exist we fallback to the srtt (which
5294 	 * we probably will never do) and then as a last
5295 	 * resort we use RACK_INITIAL_RTO if no srtt is
5296 	 * yet set.
5297 	 */
5298 	if (rack->rc_rack_rtt)
5299 		return (rack->rc_rack_rtt);
5300 	else if (tp->t_srtt == 0)
5301 		return (RACK_INITIAL_RTO);
5302 	return (tp->t_srtt);
5303 }
5304 
5305 static struct rack_sendmap *
5306 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5307 {
5308 	/*
5309 	 * Check to see that we don't need to fall into recovery. We will
5310 	 * need to do so if our oldest transmit is past the time we should
5311 	 * have had an ack.
5312 	 */
5313 	struct tcp_rack *rack;
5314 	struct rack_sendmap *rsm;
5315 	int32_t idx;
5316 	uint32_t srtt, thresh;
5317 
5318 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5319 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5320 		return (NULL);
5321 	}
5322 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5323 	if (rsm == NULL)
5324 		return (NULL);
5325 
5326 	if (rsm->r_flags & RACK_ACKED) {
5327 		rsm = rack_find_lowest_rsm(rack);
5328 		if (rsm == NULL)
5329 			return (NULL);
5330 	}
5331 	idx = rsm->r_rtr_cnt - 1;
5332 	srtt = rack_grab_rtt(tp, rack);
5333 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5334 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5335 		return (NULL);
5336 	}
5337 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5338 		return (NULL);
5339 	}
5340 	/* Ok if we reach here we are over-due and this guy can be sent */
5341 	if (IN_RECOVERY(tp->t_flags) == 0) {
5342 		/*
5343 		 * For the one that enters us into recovery record undo
5344 		 * info.
5345 		 */
5346 		rack->r_ctl.rc_rsm_start = rsm->r_start;
5347 		rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
5348 		rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
5349 	}
5350 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
5351 	return (rsm);
5352 }
5353 
5354 static uint32_t
5355 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5356 {
5357 	int32_t t;
5358 	int32_t tt;
5359 	uint32_t ret_val;
5360 
5361 	t = (tp->t_srtt + (tp->t_rttvar << 2));
5362 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5363  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5364 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5365 	ret_val = (uint32_t)tt;
5366 	return (ret_val);
5367 }
5368 
5369 static uint32_t
5370 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5371 {
5372 	/*
5373 	 * Start the FR timer, we do this based on getting the first one in
5374 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
5375 	 * events we need to stop the running timer (if its running) before
5376 	 * starting the new one.
5377 	 */
5378 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5379 	uint32_t srtt_cur;
5380 	int32_t idx;
5381 	int32_t is_tlp_timer = 0;
5382 	struct rack_sendmap *rsm;
5383 
5384 	if (rack->t_timers_stopped) {
5385 		/* All timers have been stopped none are to run */
5386 		return (0);
5387 	}
5388 	if (rack->rc_in_persist) {
5389 		/* We can't start any timer in persists */
5390 		return (rack_get_persists_timer_val(tp, rack));
5391 	}
5392 	rack->rc_on_min_to = 0;
5393 	if ((tp->t_state < TCPS_ESTABLISHED) ||
5394 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5395 		goto activate_rxt;
5396 	}
5397 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5398 	if ((rsm == NULL) || sup_rack) {
5399 		/* Nothing on the send map or no rack */
5400 activate_rxt:
5401 		time_since_sent = 0;
5402 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5403 		if (rsm) {
5404 			/*
5405 			 * Should we discount the RTX timer any?
5406 			 *
5407 			 * We want to discount it the smallest amount.
5408 			 * If a timer (Rack/TLP or RXT) has gone off more
5409 			 * recently thats the discount we want to use (now - timer time).
5410 			 * If the retransmit of the oldest packet was more recent then
5411 			 * we want to use that (now - oldest-packet-last_transmit_time).
5412 			 *
5413 			 */
5414 			idx = rsm->r_rtr_cnt - 1;
5415 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5416 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5417 			else
5418 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5419 			if (TSTMP_GT(cts, tstmp_touse))
5420 			    time_since_sent = cts - tstmp_touse;
5421 		}
5422 		if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
5423 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5424 			to = tp->t_rxtcur;
5425 			if (to > time_since_sent)
5426 				to -= time_since_sent;
5427 			else
5428 				to = rack->r_ctl.rc_min_to;
5429 			if (to == 0)
5430 				to = 1;
5431 			/* Special case for KEEPINIT */
5432 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5433 			    (TP_KEEPINIT(tp) != 0) &&
5434 			    rsm) {
5435 				/*
5436 				 * We have to put a ceiling on the rxt timer
5437 				 * of the keep-init timeout.
5438 				 */
5439 				uint32_t max_time, red;
5440 
5441 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5442 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5443 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5444 					if (red < max_time)
5445 						max_time -= red;
5446 					else
5447 						max_time = 1;
5448 				}
5449 				/* Reduce timeout to the keep value if needed */
5450 				if (max_time < to)
5451 					to = max_time;
5452 			}
5453 			return (to);
5454 		}
5455 		return (0);
5456 	}
5457 	if (rsm->r_flags & RACK_ACKED) {
5458 		rsm = rack_find_lowest_rsm(rack);
5459 		if (rsm == NULL) {
5460 			/* No lowest? */
5461 			goto activate_rxt;
5462 		}
5463 	}
5464 	if (rack->sack_attack_disable) {
5465 		/*
5466 		 * We don't want to do
5467 		 * any TLP's if you are an attacker.
5468 		 * Though if you are doing what
5469 		 * is expected you may still have
5470 		 * SACK-PASSED marks.
5471 		 */
5472 		goto activate_rxt;
5473 	}
5474 	/* Convert from ms to usecs */
5475 	if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5476 		if ((tp->t_flags & TF_SENTFIN) &&
5477 		    ((tp->snd_max - tp->snd_una) == 1) &&
5478 		    (rsm->r_flags & RACK_HAS_FIN)) {
5479 			/*
5480 			 * We don't start a rack timer if all we have is a
5481 			 * FIN outstanding.
5482 			 */
5483 			goto activate_rxt;
5484 		}
5485 		if ((rack->use_rack_rr == 0) &&
5486 		    (IN_FASTRECOVERY(tp->t_flags)) &&
5487 		    (rack->rack_no_prr == 0) &&
5488 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5489 			/*
5490 			 * We are not cheating, in recovery  and
5491 			 * not enough ack's to yet get our next
5492 			 * retransmission out.
5493 			 *
5494 			 * Note that classified attackers do not
5495 			 * get to use the rack-cheat.
5496 			 */
5497 			goto activate_tlp;
5498 		}
5499 		srtt = rack_grab_rtt(tp, rack);
5500 		thresh = rack_calc_thresh_rack(rack, srtt, cts);
5501 		idx = rsm->r_rtr_cnt - 1;
5502 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5503 		if (SEQ_GEQ(exp, cts)) {
5504 			to = exp - cts;
5505 			if (to < rack->r_ctl.rc_min_to) {
5506 				to = rack->r_ctl.rc_min_to;
5507 				if (rack->r_rr_config == 3)
5508 					rack->rc_on_min_to = 1;
5509 			}
5510 		} else {
5511 			to = rack->r_ctl.rc_min_to;
5512 			if (rack->r_rr_config == 3)
5513 				rack->rc_on_min_to = 1;
5514 		}
5515 	} else {
5516 		/* Ok we need to do a TLP not RACK */
5517 activate_tlp:
5518 		if ((rack->rc_tlp_in_progress != 0) &&
5519 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5520 			/*
5521 			 * The previous send was a TLP and we have sent
5522 			 * N TLP's without sending new data.
5523 			 */
5524 			goto activate_rxt;
5525 		}
5526 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5527 		if (rsm == NULL) {
5528 			/* We found no rsm to TLP with. */
5529 			goto activate_rxt;
5530 		}
5531 		if (rsm->r_flags & RACK_HAS_FIN) {
5532 			/* If its a FIN we dont do TLP */
5533 			rsm = NULL;
5534 			goto activate_rxt;
5535 		}
5536 		idx = rsm->r_rtr_cnt - 1;
5537 		time_since_sent = 0;
5538 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5539 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5540 		else
5541 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5542 		if (TSTMP_GT(cts, tstmp_touse))
5543 		    time_since_sent = cts - tstmp_touse;
5544 		is_tlp_timer = 1;
5545 		if (tp->t_srtt) {
5546 			if ((rack->rc_srtt_measure_made == 0) &&
5547 			    (tp->t_srtt == 1)) {
5548 				/*
5549 				 * If another stack as run and set srtt to 1,
5550 				 * then the srtt was 0, so lets use the initial.
5551 				 */
5552 				srtt = RACK_INITIAL_RTO;
5553 			} else {
5554 				srtt_cur = tp->t_srtt;
5555 				srtt = srtt_cur;
5556 			}
5557 		} else
5558 			srtt = RACK_INITIAL_RTO;
5559 		/*
5560 		 * If the SRTT is not keeping up and the
5561 		 * rack RTT has spiked we want to use
5562 		 * the last RTT not the smoothed one.
5563 		 */
5564 		if (rack_tlp_use_greater &&
5565 		    tp->t_srtt &&
5566 		    (srtt < rack_grab_rtt(tp, rack))) {
5567 			srtt = rack_grab_rtt(tp, rack);
5568 		}
5569 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5570 		if (thresh > time_since_sent) {
5571 			to = thresh - time_since_sent;
5572 		} else {
5573 			to = rack->r_ctl.rc_min_to;
5574 			rack_log_alt_to_to_cancel(rack,
5575 						  thresh,		/* flex1 */
5576 						  time_since_sent,	/* flex2 */
5577 						  tstmp_touse,		/* flex3 */
5578 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5579 						  (uint32_t)rsm->r_tim_lastsent[idx],
5580 						  srtt,
5581 						  idx, 99);
5582 		}
5583 		if (to < rack_tlp_min) {
5584 			to = rack_tlp_min;
5585 		}
5586 		if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5587 			/*
5588 			 * If the TLP time works out to larger than the max
5589 			 * RTO lets not do TLP.. just RTO.
5590 			 */
5591 			goto activate_rxt;
5592 		}
5593 	}
5594 	if (is_tlp_timer == 0) {
5595 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5596 	} else {
5597 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5598 	}
5599 	if (to == 0)
5600 		to = 1;
5601 	return (to);
5602 }
5603 
5604 static void
5605 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5606 {
5607 	if (rack->rc_in_persist == 0) {
5608 		if (tp->t_flags & TF_GPUTINPROG) {
5609 			/*
5610 			 * Stop the goodput now, the calling of the
5611 			 * measurement function clears the flag.
5612 			 */
5613 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
5614 						    RACK_QUALITY_PERSIST);
5615 		}
5616 #ifdef NETFLIX_SHARED_CWND
5617 		if (rack->r_ctl.rc_scw) {
5618 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5619 			rack->rack_scwnd_is_idle = 1;
5620 		}
5621 #endif
5622 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5623 		if (rack->r_ctl.rc_went_idle_time == 0)
5624 			rack->r_ctl.rc_went_idle_time = 1;
5625 		rack_timer_cancel(tp, rack, cts, __LINE__);
5626 		tp->t_rxtshift = 0;
5627 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5628 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5629 		rack->rc_in_persist = 1;
5630 	}
5631 }
5632 
5633 static void
5634 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5635 {
5636 	if (rack->rc_inp->inp_in_hpts) {
5637 		tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
5638 		rack->r_ctl.rc_hpts_flags = 0;
5639 	}
5640 #ifdef NETFLIX_SHARED_CWND
5641 	if (rack->r_ctl.rc_scw) {
5642 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5643 		rack->rack_scwnd_is_idle = 0;
5644 	}
5645 #endif
5646 	if (rack->rc_gp_dyn_mul &&
5647 	    (rack->use_fixed_rate == 0) &&
5648 	    (rack->rc_always_pace)) {
5649 		/*
5650 		 * Do we count this as if a probe-rtt just
5651 		 * finished?
5652 		 */
5653 		uint32_t time_idle, idle_min;
5654 
5655 		time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5656 		idle_min = rack_min_probertt_hold;
5657 		if (rack_probertt_gpsrtt_cnt_div) {
5658 			uint64_t extra;
5659 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5660 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
5661 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5662 			idle_min += (uint32_t)extra;
5663 		}
5664 		if (time_idle >= idle_min) {
5665 			/* Yes, we count it as a probe-rtt. */
5666 			uint32_t us_cts;
5667 
5668 			us_cts = tcp_get_usecs(NULL);
5669 			if (rack->in_probe_rtt == 0) {
5670 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5671 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5672 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5673 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5674 			} else {
5675 				rack_exit_probertt(rack, us_cts);
5676 			}
5677 		}
5678 	}
5679 	rack->rc_in_persist = 0;
5680 	rack->r_ctl.rc_went_idle_time = 0;
5681 	tp->t_rxtshift = 0;
5682 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5683 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5684 	rack->r_ctl.rc_agg_delayed = 0;
5685 	rack->r_early = 0;
5686 	rack->r_late = 0;
5687 	rack->r_ctl.rc_agg_early = 0;
5688 }
5689 
5690 static void
5691 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5692 		   struct hpts_diag *diag, struct timeval *tv)
5693 {
5694 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5695 		union tcp_log_stackspecific log;
5696 
5697 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5698 		log.u_bbr.flex1 = diag->p_nxt_slot;
5699 		log.u_bbr.flex2 = diag->p_cur_slot;
5700 		log.u_bbr.flex3 = diag->slot_req;
5701 		log.u_bbr.flex4 = diag->inp_hptsslot;
5702 		log.u_bbr.flex5 = diag->slot_remaining;
5703 		log.u_bbr.flex6 = diag->need_new_to;
5704 		log.u_bbr.flex7 = diag->p_hpts_active;
5705 		log.u_bbr.flex8 = diag->p_on_min_sleep;
5706 		/* Hijack other fields as needed */
5707 		log.u_bbr.epoch = diag->have_slept;
5708 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
5709 		log.u_bbr.pkts_out = diag->co_ret;
5710 		log.u_bbr.applimited = diag->hpts_sleep_time;
5711 		log.u_bbr.delivered = diag->p_prev_slot;
5712 		log.u_bbr.inflight = diag->p_runningslot;
5713 		log.u_bbr.bw_inuse = diag->wheel_slot;
5714 		log.u_bbr.rttProp = diag->wheel_cts;
5715 		log.u_bbr.timeStamp = cts;
5716 		log.u_bbr.delRate = diag->maxslots;
5717 		log.u_bbr.cur_del_rate = diag->p_curtick;
5718 		log.u_bbr.cur_del_rate <<= 32;
5719 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
5720 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5721 		    &rack->rc_inp->inp_socket->so_rcv,
5722 		    &rack->rc_inp->inp_socket->so_snd,
5723 		    BBR_LOG_HPTSDIAG, 0,
5724 		    0, &log, false, tv);
5725 	}
5726 
5727 }
5728 
5729 static void
5730 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5731 {
5732 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5733 		union tcp_log_stackspecific log;
5734 		struct timeval tv;
5735 
5736 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5737 		log.u_bbr.flex1 = sb->sb_flags;
5738 		log.u_bbr.flex2 = len;
5739 		log.u_bbr.flex3 = sb->sb_state;
5740 		log.u_bbr.flex8 = type;
5741 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5742 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5743 		    &rack->rc_inp->inp_socket->so_rcv,
5744 		    &rack->rc_inp->inp_socket->so_snd,
5745 		    TCP_LOG_SB_WAKE, 0,
5746 		    len, &log, false, &tv);
5747 	}
5748 }
5749 
5750 static void
5751 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5752       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5753 {
5754 	struct hpts_diag diag;
5755 	struct inpcb *inp;
5756 	struct timeval tv;
5757 	uint32_t delayed_ack = 0;
5758 	uint32_t hpts_timeout;
5759 	uint32_t entry_slot = slot;
5760 	uint8_t stopped;
5761 	uint32_t left = 0;
5762 	uint32_t us_cts;
5763 
5764 	inp = tp->t_inpcb;
5765 	if ((tp->t_state == TCPS_CLOSED) ||
5766 	    (tp->t_state == TCPS_LISTEN)) {
5767 		return;
5768 	}
5769 	if (inp->inp_in_hpts) {
5770 		/* Already on the pacer */
5771 		return;
5772 	}
5773 	stopped = rack->rc_tmr_stopped;
5774 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5775 		left = rack->r_ctl.rc_timer_exp - cts;
5776 	}
5777 	rack->r_ctl.rc_timer_exp = 0;
5778 	rack->r_ctl.rc_hpts_flags = 0;
5779 	us_cts = tcp_get_usecs(&tv);
5780 	/* Now early/late accounting */
5781 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
5782 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5783 		/*
5784 		 * We have a early carry over set,
5785 		 * we can always add more time so we
5786 		 * can always make this compensation.
5787 		 *
5788 		 * Note if ack's are allowed to wake us do not
5789 		 * penalize the next timer for being awoke
5790 		 * by an ack aka the rc_agg_early (non-paced mode).
5791 		 */
5792 		slot += rack->r_ctl.rc_agg_early;
5793 		rack->r_early = 0;
5794 		rack->r_ctl.rc_agg_early = 0;
5795 	}
5796 	if (rack->r_late) {
5797 		/*
5798 		 * This is harder, we can
5799 		 * compensate some but it
5800 		 * really depends on what
5801 		 * the current pacing time is.
5802 		 */
5803 		if (rack->r_ctl.rc_agg_delayed >= slot) {
5804 			/*
5805 			 * We can't compensate for it all.
5806 			 * And we have to have some time
5807 			 * on the clock. We always have a min
5808 			 * 10 slots (10 x 10 i.e. 100 usecs).
5809 			 */
5810 			if (slot <= HPTS_TICKS_PER_SLOT) {
5811 				/* We gain delay */
5812 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
5813 				slot = HPTS_TICKS_PER_SLOT;
5814 			} else {
5815 				/* We take off some */
5816 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
5817 				slot = HPTS_TICKS_PER_SLOT;
5818 			}
5819 		} else {
5820 			slot -= rack->r_ctl.rc_agg_delayed;
5821 			rack->r_ctl.rc_agg_delayed = 0;
5822 			/* Make sure we have 100 useconds at minimum */
5823 			if (slot < HPTS_TICKS_PER_SLOT) {
5824 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
5825 				slot = HPTS_TICKS_PER_SLOT;
5826 			}
5827 			if (rack->r_ctl.rc_agg_delayed == 0)
5828 				rack->r_late = 0;
5829 		}
5830 	}
5831 	if (slot) {
5832 		/* We are pacing too */
5833 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5834 	}
5835 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5836 #ifdef NETFLIX_EXP_DETECTION
5837 	if (rack->sack_attack_disable &&
5838 	    (slot < tcp_sad_pacing_interval)) {
5839 		/*
5840 		 * We have a potential attacker on
5841 		 * the line. We have possibly some
5842 		 * (or now) pacing time set. We want to
5843 		 * slow down the processing of sacks by some
5844 		 * amount (if it is an attacker). Set the default
5845 		 * slot for attackers in place (unless the orginal
5846 		 * interval is longer). Its stored in
5847 		 * micro-seconds, so lets convert to msecs.
5848 		 */
5849 		slot = tcp_sad_pacing_interval;
5850 	}
5851 #endif
5852 	if (tp->t_flags & TF_DELACK) {
5853 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
5854 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5855 	}
5856 	if (delayed_ack && ((hpts_timeout == 0) ||
5857 			    (delayed_ack < hpts_timeout)))
5858 		hpts_timeout = delayed_ack;
5859 	else
5860 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5861 	/*
5862 	 * If no timers are going to run and we will fall off the hptsi
5863 	 * wheel, we resort to a keep-alive timer if its configured.
5864 	 */
5865 	if ((hpts_timeout == 0) &&
5866 	    (slot == 0)) {
5867 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5868 		    (tp->t_state <= TCPS_CLOSING)) {
5869 			/*
5870 			 * Ok we have no timer (persists, rack, tlp, rxt  or
5871 			 * del-ack), we don't have segments being paced. So
5872 			 * all that is left is the keepalive timer.
5873 			 */
5874 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5875 				/* Get the established keep-alive time */
5876 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5877 			} else {
5878 				/*
5879 				 * Get the initial setup keep-alive time,
5880 				 * note that this is probably not going to
5881 				 * happen, since rack will be running a rxt timer
5882 				 * if a SYN of some sort is outstanding. It is
5883 				 * actually handled in rack_timeout_rxt().
5884 				 */
5885 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5886 			}
5887 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5888 			if (rack->in_probe_rtt) {
5889 				/*
5890 				 * We want to instead not wake up a long time from
5891 				 * now but to wake up about the time we would
5892 				 * exit probe-rtt and initiate a keep-alive ack.
5893 				 * This will get us out of probe-rtt and update
5894 				 * our min-rtt.
5895 				 */
5896 				hpts_timeout = rack_min_probertt_hold;
5897 			}
5898 		}
5899 	}
5900 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5901 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5902 		/*
5903 		 * RACK, TLP, persists and RXT timers all are restartable
5904 		 * based on actions input .. i.e we received a packet (ack
5905 		 * or sack) and that changes things (rw, or snd_una etc).
5906 		 * Thus we can restart them with a new value. For
5907 		 * keep-alive, delayed_ack we keep track of what was left
5908 		 * and restart the timer with a smaller value.
5909 		 */
5910 		if (left < hpts_timeout)
5911 			hpts_timeout = left;
5912 	}
5913 	if (hpts_timeout) {
5914 		/*
5915 		 * Hack alert for now we can't time-out over 2,147,483
5916 		 * seconds (a bit more than 596 hours), which is probably ok
5917 		 * :).
5918 		 */
5919 		if (hpts_timeout > 0x7ffffffe)
5920 			hpts_timeout = 0x7ffffffe;
5921 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5922 	}
5923 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
5924 	if ((rack->gp_ready == 0) &&
5925 	    (rack->use_fixed_rate == 0) &&
5926 	    (hpts_timeout < slot) &&
5927 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5928 		/*
5929 		 * We have no good estimate yet for the
5930 		 * old clunky burst mitigation or the
5931 		 * real pacing. And the tlp or rxt is smaller
5932 		 * than the pacing calculation. Lets not
5933 		 * pace that long since we know the calculation
5934 		 * so far is not accurate.
5935 		 */
5936 		slot = hpts_timeout;
5937 	}
5938 	rack->r_ctl.last_pacing_time = slot;
5939 	/**
5940 	 * Turn off all the flags for queuing by default. The
5941 	 * flags have important meanings to what happens when
5942 	 * LRO interacts with the transport. Most likely (by default now)
5943 	 * mbuf_queueing and ack compression are on. So the transport
5944 	 * has a couple of flags that control what happens (if those
5945 	 * are not on then these flags won't have any effect since it
5946 	 * won't go through the queuing LRO path).
5947 	 *
5948 	 * INP_MBUF_QUEUE_READY - This flags says that I am busy
5949 	 *                        pacing output, so don't disturb. But
5950 	 *                        it also means LRO can wake me if there
5951 	 *                        is a SACK arrival.
5952 	 *
5953 	 * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5954 	 *                       with the above flag (QUEUE_READY) and
5955 	 *                       when present it says don't even wake me
5956 	 *                       if a SACK arrives.
5957 	 *
5958 	 * The idea behind these flags is that if we are pacing we
5959 	 * set the MBUF_QUEUE_READY and only get woken up if
5960 	 * a SACK arrives (which could change things) or if
5961 	 * our pacing timer expires. If, however, we have a rack
5962 	 * timer running, then we don't even want a sack to wake
5963 	 * us since the rack timer has to expire before we can send.
5964 	 *
5965 	 * Other cases should usually have none of the flags set
5966 	 * so LRO can call into us.
5967 	 */
5968 	inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5969 	if (slot) {
5970 		rack->r_ctl.rc_last_output_to = us_cts + slot;
5971 		/*
5972 		 * A pacing timer (slot) is being set, in
5973 		 * such a case we cannot send (we are blocked by
5974 		 * the timer). So lets tell LRO that it should not
5975 		 * wake us unless there is a SACK. Note this only
5976 		 * will be effective if mbuf queueing is on or
5977 		 * compressed acks are being processed.
5978 		 */
5979 		inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5980 		/*
5981 		 * But wait if we have a Rack timer running
5982 		 * even a SACK should not disturb us (with
5983 		 * the exception of r_rr_config 3).
5984 		 */
5985 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5986 		    (rack->r_rr_config != 3))
5987 			inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5988 		if (rack->rc_ack_can_sendout_data) {
5989 			/*
5990 			 * Ahh but wait, this is that special case
5991 			 * where the pacing timer can be disturbed
5992 			 * backout the changes (used for non-paced
5993 			 * burst limiting).
5994 			 */
5995 			inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5996 		}
5997 		if ((rack->use_rack_rr) &&
5998 		    (rack->r_rr_config < 2) &&
5999 		    ((hpts_timeout) && (hpts_timeout < slot))) {
6000 			/*
6001 			 * Arrange for the hpts to kick back in after the
6002 			 * t-o if the t-o does not cause a send.
6003 			 */
6004 			(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
6005 						   __LINE__, &diag);
6006 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6007 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6008 		} else {
6009 			(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot),
6010 						   __LINE__, &diag);
6011 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6012 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
6013 		}
6014 	} else if (hpts_timeout) {
6015 		/*
6016 		 * With respect to inp_flags2 here, lets let any new acks wake
6017 		 * us up here. Since we are not pacing (no pacing timer), output
6018 		 * can happen so we should let it. If its a Rack timer, then any inbound
6019 		 * packet probably won't change the sending (we will be blocked)
6020 		 * but it may change the prr stats so letting it in (the set defaults
6021 		 * at the start of this block) are good enough.
6022 		 */
6023 		(void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout),
6024 					   __LINE__, &diag);
6025 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
6026 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
6027 	} else {
6028 		/* No timer starting */
6029 #ifdef INVARIANTS
6030 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
6031 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
6032 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
6033 		}
6034 #endif
6035 	}
6036 	rack->rc_tmr_stopped = 0;
6037 	if (slot)
6038 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
6039 }
6040 
6041 /*
6042  * RACK Timer, here we simply do logging and house keeping.
6043  * the normal rack_output() function will call the
6044  * appropriate thing to check if we need to do a RACK retransmit.
6045  * We return 1, saying don't proceed with rack_output only
6046  * when all timers have been stopped (destroyed PCB?).
6047  */
6048 static int
6049 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6050 {
6051 	/*
6052 	 * This timer simply provides an internal trigger to send out data.
6053 	 * The check_recovery_mode call will see if there are needed
6054 	 * retransmissions, if so we will enter fast-recovery. The output
6055 	 * call may or may not do the same thing depending on sysctl
6056 	 * settings.
6057 	 */
6058 	struct rack_sendmap *rsm;
6059 
6060 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6061 		return (1);
6062 	}
6063 	counter_u64_add(rack_to_tot, 1);
6064 	if (rack->r_state && (rack->r_state != tp->t_state))
6065 		rack_set_state(tp, rack);
6066 	rack->rc_on_min_to = 0;
6067 	rsm = rack_check_recovery_mode(tp, cts);
6068 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
6069 	if (rsm) {
6070 		rack->r_ctl.rc_resend = rsm;
6071 		rack->r_timer_override = 1;
6072 		if (rack->use_rack_rr) {
6073 			/*
6074 			 * Don't accumulate extra pacing delay
6075 			 * we are allowing the rack timer to
6076 			 * over-ride pacing i.e. rrr takes precedence
6077 			 * if the pacing interval is longer than the rrr
6078 			 * time (in other words we get the min pacing
6079 			 * time versus rrr pacing time).
6080 			 */
6081 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6082 		}
6083 	}
6084 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
6085 	if (rsm == NULL) {
6086 		/* restart a timer and return 1 */
6087 		rack_start_hpts_timer(rack, tp, cts,
6088 				      0, 0, 0);
6089 		return (1);
6090 	}
6091 	return (0);
6092 }
6093 
6094 static void
6095 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
6096 {
6097 	if (rsm->m->m_len > rsm->orig_m_len) {
6098 		/*
6099 		 * Mbuf grew, caused by sbcompress, our offset does
6100 		 * not change.
6101 		 */
6102 		rsm->orig_m_len = rsm->m->m_len;
6103 	} else if (rsm->m->m_len < rsm->orig_m_len) {
6104 		/*
6105 		 * Mbuf shrank, trimmed off the top by an ack, our
6106 		 * offset changes.
6107 		 */
6108 		rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
6109 		rsm->orig_m_len = rsm->m->m_len;
6110 	}
6111 }
6112 
6113 static void
6114 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
6115 {
6116 	struct mbuf *m;
6117 	uint32_t soff;
6118 
6119 	if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
6120 		/* Fix up the orig_m_len and possibly the mbuf offset */
6121 		rack_adjust_orig_mlen(src_rsm);
6122 	}
6123 	m = src_rsm->m;
6124 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
6125 	while (soff >= m->m_len) {
6126 		/* Move out past this mbuf */
6127 		soff -= m->m_len;
6128 		m = m->m_next;
6129 		KASSERT((m != NULL),
6130 			("rsm:%p nrsm:%p hit at soff:%u null m",
6131 			 src_rsm, rsm, soff));
6132 	}
6133 	rsm->m = m;
6134 	rsm->soff = soff;
6135 	rsm->orig_m_len = m->m_len;
6136 }
6137 
6138 static __inline void
6139 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
6140 	       struct rack_sendmap *rsm, uint32_t start)
6141 {
6142 	int idx;
6143 
6144 	nrsm->r_start = start;
6145 	nrsm->r_end = rsm->r_end;
6146 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
6147 	nrsm->r_flags = rsm->r_flags;
6148 	nrsm->r_dupack = rsm->r_dupack;
6149 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
6150 	nrsm->r_rtr_bytes = 0;
6151 	nrsm->r_fas = rsm->r_fas;
6152 	rsm->r_end = nrsm->r_start;
6153 	nrsm->r_just_ret = rsm->r_just_ret;
6154 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
6155 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
6156 	}
6157 	/* Now if we have SYN flag we keep it on the left edge */
6158 	if (nrsm->r_flags & RACK_HAS_SYN)
6159 		nrsm->r_flags &= ~RACK_HAS_SYN;
6160 	/* Now if we have a FIN flag we keep it on the right edge */
6161 	if (rsm->r_flags & RACK_HAS_FIN)
6162 		rsm->r_flags &= ~RACK_HAS_FIN;
6163 	/* Push bit must go to the right edge as well */
6164 	if (rsm->r_flags & RACK_HAD_PUSH)
6165 		rsm->r_flags &= ~RACK_HAD_PUSH;
6166 	/* Clone over the state of the hw_tls flag */
6167 	nrsm->r_hw_tls = rsm->r_hw_tls;
6168 	/*
6169 	 * Now we need to find nrsm's new location in the mbuf chain
6170 	 * we basically calculate a new offset, which is soff +
6171 	 * how much is left in original rsm. Then we walk out the mbuf
6172 	 * chain to find the righ postion, it may be the same mbuf
6173 	 * or maybe not.
6174 	 */
6175 	KASSERT(((rsm->m != NULL) ||
6176 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
6177 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
6178 	if (rsm->m)
6179 		rack_setup_offset_for_rsm(rsm, nrsm);
6180 }
6181 
6182 static struct rack_sendmap *
6183 rack_merge_rsm(struct tcp_rack *rack,
6184 	       struct rack_sendmap *l_rsm,
6185 	       struct rack_sendmap *r_rsm)
6186 {
6187 	/*
6188 	 * We are merging two ack'd RSM's,
6189 	 * the l_rsm is on the left (lower seq
6190 	 * values) and the r_rsm is on the right
6191 	 * (higher seq value). The simplest way
6192 	 * to merge these is to move the right
6193 	 * one into the left. I don't think there
6194 	 * is any reason we need to try to find
6195 	 * the oldest (or last oldest retransmitted).
6196 	 */
6197 	struct rack_sendmap *rm;
6198 
6199 	rack_log_map_chg(rack->rc_tp, rack, NULL,
6200 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
6201 	l_rsm->r_end = r_rsm->r_end;
6202 	if (l_rsm->r_dupack < r_rsm->r_dupack)
6203 		l_rsm->r_dupack = r_rsm->r_dupack;
6204 	if (r_rsm->r_rtr_bytes)
6205 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
6206 	if (r_rsm->r_in_tmap) {
6207 		/* This really should not happen */
6208 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
6209 		r_rsm->r_in_tmap = 0;
6210 	}
6211 
6212 	/* Now the flags */
6213 	if (r_rsm->r_flags & RACK_HAS_FIN)
6214 		l_rsm->r_flags |= RACK_HAS_FIN;
6215 	if (r_rsm->r_flags & RACK_TLP)
6216 		l_rsm->r_flags |= RACK_TLP;
6217 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
6218 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
6219 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
6220 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6221 		/*
6222 		 * If both are app-limited then let the
6223 		 * free lower the count. If right is app
6224 		 * limited and left is not, transfer.
6225 		 */
6226 		l_rsm->r_flags |= RACK_APP_LIMITED;
6227 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
6228 		if (r_rsm == rack->r_ctl.rc_first_appl)
6229 			rack->r_ctl.rc_first_appl = l_rsm;
6230 	}
6231 	rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6232 #ifdef INVARIANTS
6233 	if (rm != r_rsm) {
6234 		panic("removing head in rack:%p rsm:%p rm:%p",
6235 		      rack, r_rsm, rm);
6236 	}
6237 #endif
6238 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6239 		/* Transfer the split limit to the map we free */
6240 		r_rsm->r_limit_type = l_rsm->r_limit_type;
6241 		l_rsm->r_limit_type = 0;
6242 	}
6243 	rack_free(rack, r_rsm);
6244 	return (l_rsm);
6245 }
6246 
6247 /*
6248  * TLP Timer, here we simply setup what segment we want to
6249  * have the TLP expire on, the normal rack_output() will then
6250  * send it out.
6251  *
6252  * We return 1, saying don't proceed with rack_output only
6253  * when all timers have been stopped (destroyed PCB?).
6254  */
6255 static int
6256 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
6257 {
6258 	/*
6259 	 * Tail Loss Probe.
6260 	 */
6261 	struct rack_sendmap *rsm = NULL;
6262 	struct rack_sendmap *insret;
6263 	struct socket *so;
6264 	uint32_t amm;
6265 	uint32_t out, avail;
6266 	int collapsed_win = 0;
6267 
6268 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6269 		return (1);
6270 	}
6271 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6272 		/* Its not time yet */
6273 		return (0);
6274 	}
6275 	if (ctf_progress_timeout_check(tp, true)) {
6276 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6277 		tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
6278 		return (1);
6279 	}
6280 	/*
6281 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
6282 	 * need to figure out how to force a full MSS segment out.
6283 	 */
6284 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6285 	rack->r_ctl.retran_during_recovery = 0;
6286 	rack->r_ctl.dsack_byte_cnt = 0;
6287 	counter_u64_add(rack_tlp_tot, 1);
6288 	if (rack->r_state && (rack->r_state != tp->t_state))
6289 		rack_set_state(tp, rack);
6290 	so = tp->t_inpcb->inp_socket;
6291 	avail = sbavail(&so->so_snd);
6292 	out = tp->snd_max - tp->snd_una;
6293 	if (out > tp->snd_wnd) {
6294 		/* special case, we need a retransmission */
6295 		collapsed_win = 1;
6296 		goto need_retran;
6297 	}
6298 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
6299 		rack->r_ctl.dsack_persist--;
6300 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6301 			rack->r_ctl.num_dsack = 0;
6302 		}
6303 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6304 	}
6305 	if ((tp->t_flags & TF_GPUTINPROG) &&
6306 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
6307 		/*
6308 		 * If this is the second in a row
6309 		 * TLP and we are doing a measurement
6310 		 * its time to abandon the measurement.
6311 		 * Something is likely broken on
6312 		 * the clients network and measuring a
6313 		 * broken network does us no good.
6314 		 */
6315 		tp->t_flags &= ~TF_GPUTINPROG;
6316 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6317 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
6318 					   tp->gput_seq,
6319 					   0, 0, 18, __LINE__, NULL, 0);
6320 	}
6321 	/*
6322 	 * Check our send oldest always settings, and if
6323 	 * there is an oldest to send jump to the need_retran.
6324 	 */
6325 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6326 		goto need_retran;
6327 
6328 	if (avail > out) {
6329 		/* New data is available */
6330 		amm = avail - out;
6331 		if (amm > ctf_fixed_maxseg(tp)) {
6332 			amm = ctf_fixed_maxseg(tp);
6333 			if ((amm + out) > tp->snd_wnd) {
6334 				/* We are rwnd limited */
6335 				goto need_retran;
6336 			}
6337 		} else if (amm < ctf_fixed_maxseg(tp)) {
6338 			/* not enough to fill a MTU */
6339 			goto need_retran;
6340 		}
6341 		if (IN_FASTRECOVERY(tp->t_flags)) {
6342 			/* Unlikely */
6343 			if (rack->rack_no_prr == 0) {
6344 				if (out + amm <= tp->snd_wnd) {
6345 					rack->r_ctl.rc_prr_sndcnt = amm;
6346 					rack->r_ctl.rc_tlp_new_data = amm;
6347 					rack_log_to_prr(rack, 4, 0);
6348 				}
6349 			} else
6350 				goto need_retran;
6351 		} else {
6352 			/* Set the send-new override */
6353 			if (out + amm <= tp->snd_wnd)
6354 				rack->r_ctl.rc_tlp_new_data = amm;
6355 			else
6356 				goto need_retran;
6357 		}
6358 		rack->r_ctl.rc_tlpsend = NULL;
6359 		counter_u64_add(rack_tlp_newdata, 1);
6360 		goto send;
6361 	}
6362 need_retran:
6363 	/*
6364 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
6365 	 * optionally the first un-acked segment.
6366 	 */
6367 	if (collapsed_win == 0) {
6368 		if (rack_always_send_oldest)
6369 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6370 		else {
6371 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6372 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6373 				rsm = rack_find_high_nonack(rack, rsm);
6374 			}
6375 		}
6376 		if (rsm == NULL) {
6377 			counter_u64_add(rack_tlp_does_nada, 1);
6378 #ifdef TCP_BLACKBOX
6379 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6380 #endif
6381 			goto out;
6382 		}
6383 	} else {
6384 		/*
6385 		 * We must find the last segment
6386 		 * that was acceptable by the client.
6387 		 */
6388 		RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6389 			if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6390 				/* Found one */
6391 				break;
6392 			}
6393 		}
6394 		if (rsm == NULL) {
6395 			/* None? if so send the first */
6396 			rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6397 			if (rsm == NULL) {
6398 				counter_u64_add(rack_tlp_does_nada, 1);
6399 #ifdef TCP_BLACKBOX
6400 				tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6401 #endif
6402 				goto out;
6403 			}
6404 		}
6405 	}
6406 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6407 		/*
6408 		 * We need to split this the last segment in two.
6409 		 */
6410 		struct rack_sendmap *nrsm;
6411 
6412 		nrsm = rack_alloc_full_limit(rack);
6413 		if (nrsm == NULL) {
6414 			/*
6415 			 * No memory to split, we will just exit and punt
6416 			 * off to the RXT timer.
6417 			 */
6418 			counter_u64_add(rack_tlp_does_nada, 1);
6419 			goto out;
6420 		}
6421 		rack_clone_rsm(rack, nrsm, rsm,
6422 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
6423 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6424 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6425 #ifdef INVARIANTS
6426 		if (insret != NULL) {
6427 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6428 			      nrsm, insret, rack, rsm);
6429 		}
6430 #endif
6431 		if (rsm->r_in_tmap) {
6432 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6433 			nrsm->r_in_tmap = 1;
6434 		}
6435 		rsm = nrsm;
6436 	}
6437 	rack->r_ctl.rc_tlpsend = rsm;
6438 send:
6439 	/* Make sure output path knows we are doing a TLP */
6440 	*doing_tlp = 1;
6441 	rack->r_timer_override = 1;
6442 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6443 	return (0);
6444 out:
6445 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6446 	return (0);
6447 }
6448 
6449 /*
6450  * Delayed ack Timer, here we simply need to setup the
6451  * ACK_NOW flag and remove the DELACK flag. From there
6452  * the output routine will send the ack out.
6453  *
6454  * We only return 1, saying don't proceed, if all timers
6455  * are stopped (destroyed PCB?).
6456  */
6457 static int
6458 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6459 {
6460 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6461 		return (1);
6462 	}
6463 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6464 	tp->t_flags &= ~TF_DELACK;
6465 	tp->t_flags |= TF_ACKNOW;
6466 	KMOD_TCPSTAT_INC(tcps_delack);
6467 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6468 	return (0);
6469 }
6470 
6471 /*
6472  * Persists timer, here we simply send the
6473  * same thing as a keepalive will.
6474  * the one byte send.
6475  *
6476  * We only return 1, saying don't proceed, if all timers
6477  * are stopped (destroyed PCB?).
6478  */
6479 static int
6480 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6481 {
6482 	struct tcptemp *t_template;
6483 	struct inpcb *inp;
6484 	int32_t retval = 1;
6485 
6486 	inp = tp->t_inpcb;
6487 
6488 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6489 		return (1);
6490 	}
6491 	if (rack->rc_in_persist == 0)
6492 		return (0);
6493 	if (ctf_progress_timeout_check(tp, false)) {
6494 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6495 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6496 		tcp_set_inp_to_drop(inp, ETIMEDOUT);
6497 		return (1);
6498 	}
6499 	KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
6500 	/*
6501 	 * Persistence timer into zero window. Force a byte to be output, if
6502 	 * possible.
6503 	 */
6504 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
6505 	/*
6506 	 * Hack: if the peer is dead/unreachable, we do not time out if the
6507 	 * window is closed.  After a full backoff, drop the connection if
6508 	 * the idle time (no responses to probes) reaches the maximum
6509 	 * backoff that we would use if retransmitting.
6510 	 */
6511 	if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6512 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6513 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6514 		KMOD_TCPSTAT_INC(tcps_persistdrop);
6515 		retval = 1;
6516 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6517 		tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6518 		goto out;
6519 	}
6520 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6521 	    tp->snd_una == tp->snd_max)
6522 		rack_exit_persist(tp, rack, cts);
6523 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6524 	/*
6525 	 * If the user has closed the socket then drop a persisting
6526 	 * connection after a much reduced timeout.
6527 	 */
6528 	if (tp->t_state > TCPS_CLOSE_WAIT &&
6529 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6530 		retval = 1;
6531 		KMOD_TCPSTAT_INC(tcps_persistdrop);
6532 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6533 		tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6534 		goto out;
6535 	}
6536 	t_template = tcpip_maketemplate(rack->rc_inp);
6537 	if (t_template) {
6538 		/* only set it if we were answered */
6539 		if (rack->forced_ack == 0) {
6540 			rack->forced_ack = 1;
6541 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6542 		}
6543 		tcp_respond(tp, t_template->tt_ipgen,
6544 			    &t_template->tt_t, (struct mbuf *)NULL,
6545 			    tp->rcv_nxt, tp->snd_una - 1, 0);
6546 		/* This sends an ack */
6547 		if (tp->t_flags & TF_DELACK)
6548 			tp->t_flags &= ~TF_DELACK;
6549 		free(t_template, M_TEMP);
6550 	}
6551 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6552 		tp->t_rxtshift++;
6553 out:
6554 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6555 	rack_start_hpts_timer(rack, tp, cts,
6556 			      0, 0, 0);
6557 	return (retval);
6558 }
6559 
6560 /*
6561  * If a keepalive goes off, we had no other timers
6562  * happening. We always return 1 here since this
6563  * routine either drops the connection or sends
6564  * out a segment with respond.
6565  */
6566 static int
6567 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6568 {
6569 	struct tcptemp *t_template;
6570 	struct inpcb *inp;
6571 
6572 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6573 		return (1);
6574 	}
6575 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6576 	inp = tp->t_inpcb;
6577 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6578 	/*
6579 	 * Keep-alive timer went off; send something or drop connection if
6580 	 * idle for too long.
6581 	 */
6582 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
6583 	if (tp->t_state < TCPS_ESTABLISHED)
6584 		goto dropit;
6585 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6586 	    tp->t_state <= TCPS_CLOSING) {
6587 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6588 			goto dropit;
6589 		/*
6590 		 * Send a packet designed to force a response if the peer is
6591 		 * up and reachable: either an ACK if the connection is
6592 		 * still alive, or an RST if the peer has closed the
6593 		 * connection due to timeout or reboot. Using sequence
6594 		 * number tp->snd_una-1 causes the transmitted zero-length
6595 		 * segment to lie outside the receive window; by the
6596 		 * protocol spec, this requires the correspondent TCP to
6597 		 * respond.
6598 		 */
6599 		KMOD_TCPSTAT_INC(tcps_keepprobe);
6600 		t_template = tcpip_maketemplate(inp);
6601 		if (t_template) {
6602 			if (rack->forced_ack == 0) {
6603 				rack->forced_ack = 1;
6604 				rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6605 			}
6606 			tcp_respond(tp, t_template->tt_ipgen,
6607 			    &t_template->tt_t, (struct mbuf *)NULL,
6608 			    tp->rcv_nxt, tp->snd_una - 1, 0);
6609 			free(t_template, M_TEMP);
6610 		}
6611 	}
6612 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6613 	return (1);
6614 dropit:
6615 	KMOD_TCPSTAT_INC(tcps_keepdrops);
6616 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6617 	tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
6618 	return (1);
6619 }
6620 
6621 /*
6622  * Retransmit helper function, clear up all the ack
6623  * flags and take care of important book keeping.
6624  */
6625 static void
6626 rack_remxt_tmr(struct tcpcb *tp)
6627 {
6628 	/*
6629 	 * The retransmit timer went off, all sack'd blocks must be
6630 	 * un-acked.
6631 	 */
6632 	struct rack_sendmap *rsm, *trsm = NULL;
6633 	struct tcp_rack *rack;
6634 
6635 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6636 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6637 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6638 	if (rack->r_state && (rack->r_state != tp->t_state))
6639 		rack_set_state(tp, rack);
6640 	/*
6641 	 * Ideally we would like to be able to
6642 	 * mark SACK-PASS on anything not acked here.
6643 	 *
6644 	 * However, if we do that we would burst out
6645 	 * all that data 1ms apart. This would be unwise,
6646 	 * so for now we will just let the normal rxt timer
6647 	 * and tlp timer take care of it.
6648 	 *
6649 	 * Also we really need to stick them back in sequence
6650 	 * order. This way we send in the proper order and any
6651 	 * sacks that come floating in will "re-ack" the data.
6652 	 * To do this we zap the tmap with an INIT and then
6653 	 * walk through and place every rsm in the RB tree
6654 	 * back in its seq ordered place.
6655 	 */
6656 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
6657 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6658 		rsm->r_dupack = 0;
6659 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6660 		/* We must re-add it back to the tlist */
6661 		if (trsm == NULL) {
6662 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6663 		} else {
6664 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6665 		}
6666 		rsm->r_in_tmap = 1;
6667 		trsm = rsm;
6668 		if (rsm->r_flags & RACK_ACKED)
6669 			rsm->r_flags |= RACK_WAS_ACKED;
6670 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS);
6671 		rsm->r_flags |= RACK_MUST_RXT;
6672 	}
6673 	/* Clear the count (we just un-acked them) */
6674 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6675 	rack->r_ctl.rc_sacked = 0;
6676 	rack->r_ctl.rc_sacklast = NULL;
6677 	rack->r_ctl.rc_agg_delayed = 0;
6678 	rack->r_early = 0;
6679 	rack->r_ctl.rc_agg_early = 0;
6680 	rack->r_late = 0;
6681 	/* Clear the tlp rtx mark */
6682 	rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6683 	if (rack->r_ctl.rc_resend != NULL)
6684 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6685 	rack->r_ctl.rc_prr_sndcnt = 0;
6686 	rack_log_to_prr(rack, 6, 0);
6687 	rack->r_timer_override = 1;
6688 	if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6689 #ifdef NETFLIX_EXP_DETECTION
6690 	    || (rack->sack_attack_disable != 0)
6691 #endif
6692 		    ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6693 		/*
6694 		 * For non-sack customers new data
6695 		 * needs to go out as retransmits until
6696 		 * we retransmit up to snd_max.
6697 		 */
6698 		rack->r_must_retran = 1;
6699 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6700 						rack->r_ctl.rc_sacked);
6701 	}
6702 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6703 }
6704 
6705 static void
6706 rack_convert_rtts(struct tcpcb *tp)
6707 {
6708 	if (tp->t_srtt > 1) {
6709 		uint32_t val, frac;
6710 
6711 		val = tp->t_srtt >> TCP_RTT_SHIFT;
6712 		frac = tp->t_srtt & 0x1f;
6713 		tp->t_srtt = TICKS_2_USEC(val);
6714 		/*
6715 		 * frac is the fractional part of the srtt (if any)
6716 		 * but its in ticks and every bit represents
6717 		 * 1/32nd of a hz.
6718 		 */
6719 		if (frac) {
6720 			if (hz == 1000) {
6721 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6722 			} else {
6723 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6724 			}
6725 			tp->t_srtt += frac;
6726 		}
6727 	}
6728 	if (tp->t_rttvar) {
6729 		uint32_t val, frac;
6730 
6731 		val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6732 		frac = tp->t_rttvar & 0x1f;
6733 		tp->t_rttvar = TICKS_2_USEC(val);
6734 		/*
6735 		 * frac is the fractional part of the srtt (if any)
6736 		 * but its in ticks and every bit represents
6737 		 * 1/32nd of a hz.
6738 		 */
6739 		if (frac) {
6740 			if (hz == 1000) {
6741 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6742 			} else {
6743 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6744 			}
6745 			tp->t_rttvar += frac;
6746 		}
6747 	}
6748 	tp->t_rxtcur = RACK_REXMTVAL(tp);
6749 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6750 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6751 	}
6752 	if (tp->t_rxtcur > rack_rto_max) {
6753 		tp->t_rxtcur = rack_rto_max;
6754 	}
6755 }
6756 
6757 static void
6758 rack_cc_conn_init(struct tcpcb *tp)
6759 {
6760 	struct tcp_rack *rack;
6761 	uint32_t srtt;
6762 
6763 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6764 	srtt = tp->t_srtt;
6765 	cc_conn_init(tp);
6766 	/*
6767 	 * Now convert to rack's internal format,
6768 	 * if required.
6769 	 */
6770 	if ((srtt == 0) && (tp->t_srtt != 0))
6771 		rack_convert_rtts(tp);
6772 	/*
6773 	 * We want a chance to stay in slowstart as
6774 	 * we create a connection. TCP spec says that
6775 	 * initially ssthresh is infinite. For our
6776 	 * purposes that is the snd_wnd.
6777 	 */
6778 	if (tp->snd_ssthresh < tp->snd_wnd) {
6779 		tp->snd_ssthresh = tp->snd_wnd;
6780 	}
6781 	/*
6782 	 * We also want to assure a IW worth of
6783 	 * data can get inflight.
6784 	 */
6785 	if (rc_init_window(rack) < tp->snd_cwnd)
6786 		tp->snd_cwnd = rc_init_window(rack);
6787 }
6788 
6789 /*
6790  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6791  * we will setup to retransmit the lowest seq number outstanding.
6792  */
6793 static int
6794 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6795 {
6796 	int32_t rexmt;
6797 	struct inpcb *inp;
6798 	int32_t retval = 0;
6799 	bool isipv6;
6800 
6801 	inp = tp->t_inpcb;
6802 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6803 		return (1);
6804 	}
6805 	if ((tp->t_flags & TF_GPUTINPROG) &&
6806 	    (tp->t_rxtshift)) {
6807 		/*
6808 		 * We have had a second timeout
6809 		 * measurements on successive rxt's are not profitable.
6810 		 * It is unlikely to be of any use (the network is
6811 		 * broken or the client went away).
6812 		 */
6813 		tp->t_flags &= ~TF_GPUTINPROG;
6814 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6815 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
6816 					   tp->gput_seq,
6817 					   0, 0, 18, __LINE__, NULL, 0);
6818 	}
6819 	if (ctf_progress_timeout_check(tp, false)) {
6820 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6821 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6822 		tcp_set_inp_to_drop(inp, ETIMEDOUT);
6823 		return (1);
6824 	}
6825 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6826 	rack->r_ctl.retran_during_recovery = 0;
6827 	rack->r_ctl.dsack_byte_cnt = 0;
6828 	if (IN_FASTRECOVERY(tp->t_flags))
6829 		tp->t_flags |= TF_WASFRECOVERY;
6830 	else
6831 		tp->t_flags &= ~TF_WASFRECOVERY;
6832 	if (IN_CONGRECOVERY(tp->t_flags))
6833 		tp->t_flags |= TF_WASCRECOVERY;
6834 	else
6835 		tp->t_flags &= ~TF_WASCRECOVERY;
6836 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6837 	    (tp->snd_una == tp->snd_max)) {
6838 		/* Nothing outstanding .. nothing to do */
6839 		return (0);
6840 	}
6841 	if (rack->r_ctl.dsack_persist) {
6842 		rack->r_ctl.dsack_persist--;
6843 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6844 			rack->r_ctl.num_dsack = 0;
6845 		}
6846 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6847 	}
6848 	/*
6849 	 * Rack can only run one timer  at a time, so we cannot
6850 	 * run a KEEPINIT (gating SYN sending) and a retransmit
6851 	 * timer for the SYN. So if we are in a front state and
6852 	 * have a KEEPINIT timer we need to check the first transmit
6853 	 * against now to see if we have exceeded the KEEPINIT time
6854 	 * (if one is set).
6855 	 */
6856 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6857 	    (TP_KEEPINIT(tp) != 0)) {
6858 		struct rack_sendmap *rsm;
6859 
6860 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6861 		if (rsm) {
6862 			/* Ok we have something outstanding to test keepinit with */
6863 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6864 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6865 				/* We have exceeded the KEEPINIT time */
6866 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6867 				goto drop_it;
6868 			}
6869 		}
6870 	}
6871 	/*
6872 	 * Retransmission timer went off.  Message has not been acked within
6873 	 * retransmit interval.  Back off to a longer retransmit interval
6874 	 * and retransmit one segment.
6875 	 */
6876 	rack_remxt_tmr(tp);
6877 	if ((rack->r_ctl.rc_resend == NULL) ||
6878 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6879 		/*
6880 		 * If the rwnd collapsed on
6881 		 * the one we are retransmitting
6882 		 * it does not count against the
6883 		 * rxt count.
6884 		 */
6885 		tp->t_rxtshift++;
6886 	}
6887 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6888 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6889 drop_it:
6890 		tp->t_rxtshift = TCP_MAXRXTSHIFT;
6891 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6892 		retval = 1;
6893 		tcp_set_inp_to_drop(rack->rc_inp,
6894 		    (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
6895 		goto out;
6896 	}
6897 	if (tp->t_state == TCPS_SYN_SENT) {
6898 		/*
6899 		 * If the SYN was retransmitted, indicate CWND to be limited
6900 		 * to 1 segment in cc_conn_init().
6901 		 */
6902 		tp->snd_cwnd = 1;
6903 	} else if (tp->t_rxtshift == 1) {
6904 		/*
6905 		 * first retransmit; record ssthresh and cwnd so they can be
6906 		 * recovered if this turns out to be a "bad" retransmit. A
6907 		 * retransmit is considered "bad" if an ACK for this segment
6908 		 * is received within RTT/2 interval; the assumption here is
6909 		 * that the ACK was already in flight.  See "On Estimating
6910 		 * End-to-End Network Path Properties" by Allman and Paxson
6911 		 * for more details.
6912 		 */
6913 		tp->snd_cwnd_prev = tp->snd_cwnd;
6914 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
6915 		tp->snd_recover_prev = tp->snd_recover;
6916 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6917 		tp->t_flags |= TF_PREVVALID;
6918 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6919 		tp->t_flags &= ~TF_PREVVALID;
6920 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6921 	if ((tp->t_state == TCPS_SYN_SENT) ||
6922 	    (tp->t_state == TCPS_SYN_RECEIVED))
6923 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6924 	else
6925 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6926 
6927 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6928 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6929 	/*
6930 	 * We enter the path for PLMTUD if connection is established or, if
6931 	 * connection is FIN_WAIT_1 status, reason for the last is that if
6932 	 * amount of data we send is very small, we could send it in couple
6933 	 * of packets and process straight to FIN. In that case we won't
6934 	 * catch ESTABLISHED state.
6935 	 */
6936 #ifdef INET6
6937 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false;
6938 #else
6939 	isipv6 = false;
6940 #endif
6941 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
6942 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6943 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6944 	    ((tp->t_state == TCPS_ESTABLISHED) ||
6945 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
6946 		/*
6947 		 * Idea here is that at each stage of mtu probe (usually,
6948 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
6949 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6950 		 * should take care of that.
6951 		 */
6952 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6953 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6954 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6955 		    tp->t_rxtshift % 2 == 0)) {
6956 			/*
6957 			 * Enter Path MTU Black-hole Detection mechanism: -
6958 			 * Disable Path MTU Discovery (IP "DF" bit). -
6959 			 * Reduce MTU to lower value than what we negotiated
6960 			 * with peer.
6961 			 */
6962 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6963 				/* Record that we may have found a black hole. */
6964 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6965 				/* Keep track of previous MSS. */
6966 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6967 			}
6968 
6969 			/*
6970 			 * Reduce the MSS to blackhole value or to the
6971 			 * default in an attempt to retransmit.
6972 			 */
6973 #ifdef INET6
6974 			if (isipv6 &&
6975 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6976 				/* Use the sysctl tuneable blackhole MSS. */
6977 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6978 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6979 			} else if (isipv6) {
6980 				/* Use the default MSS. */
6981 				tp->t_maxseg = V_tcp_v6mssdflt;
6982 				/*
6983 				 * Disable Path MTU Discovery when we switch
6984 				 * to minmss.
6985 				 */
6986 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6987 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6988 			}
6989 #endif
6990 #if defined(INET6) && defined(INET)
6991 			else
6992 #endif
6993 #ifdef INET
6994 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6995 				/* Use the sysctl tuneable blackhole MSS. */
6996 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6997 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6998 			} else {
6999 				/* Use the default MSS. */
7000 				tp->t_maxseg = V_tcp_mssdflt;
7001 				/*
7002 				 * Disable Path MTU Discovery when we switch
7003 				 * to minmss.
7004 				 */
7005 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7006 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7007 			}
7008 #endif
7009 		} else {
7010 			/*
7011 			 * If further retransmissions are still unsuccessful
7012 			 * with a lowered MTU, maybe this isn't a blackhole
7013 			 * and we restore the previous MSS and blackhole
7014 			 * detection flags. The limit '6' is determined by
7015 			 * giving each probe stage (1448, 1188, 524) 2
7016 			 * chances to recover.
7017 			 */
7018 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
7019 			    (tp->t_rxtshift >= 6)) {
7020 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
7021 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
7022 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
7023 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
7024 			}
7025 		}
7026 	}
7027 	/*
7028 	 * Disable RFC1323 and SACK if we haven't got any response to
7029 	 * our third SYN to work-around some broken terminal servers
7030 	 * (most of which have hopefully been retired) that have bad VJ
7031 	 * header compression code which trashes TCP segments containing
7032 	 * unknown-to-them TCP options.
7033 	 */
7034 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
7035 	    (tp->t_rxtshift == 3))
7036 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
7037 	/*
7038 	 * If we backed off this far, our srtt estimate is probably bogus.
7039 	 * Clobber it so we'll take the next rtt measurement as our srtt;
7040 	 * move the current srtt into rttvar to keep the current retransmit
7041 	 * times until then.
7042 	 */
7043 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
7044 #ifdef INET6
7045 		if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
7046 			in6_losing(tp->t_inpcb);
7047 		else
7048 #endif
7049 			in_losing(tp->t_inpcb);
7050 		tp->t_rttvar += tp->t_srtt;
7051 		tp->t_srtt = 0;
7052 	}
7053 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
7054 	tp->snd_recover = tp->snd_max;
7055 	tp->t_flags |= TF_ACKNOW;
7056 	tp->t_rtttime = 0;
7057 	rack_cong_signal(tp, CC_RTO, tp->snd_una);
7058 out:
7059 	return (retval);
7060 }
7061 
7062 static int
7063 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
7064 {
7065 	int32_t ret = 0;
7066 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
7067 
7068 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
7069 	    (tp->t_flags & TF_GPUTINPROG)) {
7070 		/*
7071 		 * We have a goodput in progress
7072 		 * and we have entered a late state.
7073 		 * Do we have enough data in the sb
7074 		 * to handle the GPUT request?
7075 		 */
7076 		uint32_t bytes;
7077 
7078 		bytes = tp->gput_ack - tp->gput_seq;
7079 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
7080 			bytes += tp->gput_seq - tp->snd_una;
7081 		if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
7082 			/*
7083 			 * There are not enough bytes in the socket
7084 			 * buffer that have been sent to cover this
7085 			 * measurement. Cancel it.
7086 			 */
7087 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7088 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
7089 						   tp->gput_seq,
7090 						   0, 0, 18, __LINE__, NULL, 0);
7091 			tp->t_flags &= ~TF_GPUTINPROG;
7092 		}
7093 	}
7094 	if (timers == 0) {
7095 		return (0);
7096 	}
7097 	if (tp->t_state == TCPS_LISTEN) {
7098 		/* no timers on listen sockets */
7099 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
7100 			return (0);
7101 		return (1);
7102 	}
7103 	if ((timers & PACE_TMR_RACK) &&
7104 	    rack->rc_on_min_to) {
7105 		/*
7106 		 * For the rack timer when we
7107 		 * are on a min-timeout (which means rrr_conf = 3)
7108 		 * we don't want to check the timer. It may
7109 		 * be going off for a pace and thats ok we
7110 		 * want to send the retransmit (if its ready).
7111 		 *
7112 		 * If its on a normal rack timer (non-min) then
7113 		 * we will check if its expired.
7114 		 */
7115 		goto skip_time_check;
7116 	}
7117 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7118 		uint32_t left;
7119 
7120 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
7121 			ret = -1;
7122 			rack_log_to_processing(rack, cts, ret, 0);
7123 			return (0);
7124 		}
7125 		if (hpts_calling == 0) {
7126 			/*
7127 			 * A user send or queued mbuf (sack) has called us? We
7128 			 * return 0 and let the pacing guards
7129 			 * deal with it if they should or
7130 			 * should not cause a send.
7131 			 */
7132 			ret = -2;
7133 			rack_log_to_processing(rack, cts, ret, 0);
7134 			return (0);
7135 		}
7136 		/*
7137 		 * Ok our timer went off early and we are not paced false
7138 		 * alarm, go back to sleep.
7139 		 */
7140 		ret = -3;
7141 		left = rack->r_ctl.rc_timer_exp - cts;
7142 		tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
7143 		rack_log_to_processing(rack, cts, ret, left);
7144 		return (1);
7145 	}
7146 skip_time_check:
7147 	rack->rc_tmr_stopped = 0;
7148 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
7149 	if (timers & PACE_TMR_DELACK) {
7150 		ret = rack_timeout_delack(tp, rack, cts);
7151 	} else if (timers & PACE_TMR_RACK) {
7152 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
7153 		rack->r_fast_output = 0;
7154 		ret = rack_timeout_rack(tp, rack, cts);
7155 	} else if (timers & PACE_TMR_TLP) {
7156 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
7157 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
7158 	} else if (timers & PACE_TMR_RXT) {
7159 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
7160 		rack->r_fast_output = 0;
7161 		ret = rack_timeout_rxt(tp, rack, cts);
7162 	} else if (timers & PACE_TMR_PERSIT) {
7163 		ret = rack_timeout_persist(tp, rack, cts);
7164 	} else if (timers & PACE_TMR_KEEP) {
7165 		ret = rack_timeout_keepalive(tp, rack, cts);
7166 	}
7167 	rack_log_to_processing(rack, cts, ret, timers);
7168 	return (ret);
7169 }
7170 
7171 static void
7172 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
7173 {
7174 	struct timeval tv;
7175 	uint32_t us_cts, flags_on_entry;
7176 	uint8_t hpts_removed = 0;
7177 
7178 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
7179 	us_cts = tcp_get_usecs(&tv);
7180 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
7181 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
7182 	     ((tp->snd_max - tp->snd_una) == 0))) {
7183 		tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
7184 		hpts_removed = 1;
7185 		/* If we were not delayed cancel out the flag. */
7186 		if ((tp->snd_max - tp->snd_una) == 0)
7187 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7188 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7189 	}
7190 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
7191 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
7192 		if (rack->rc_inp->inp_in_hpts &&
7193 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
7194 			/*
7195 			 * Canceling timer's when we have no output being
7196 			 * paced. We also must remove ourselves from the
7197 			 * hpts.
7198 			 */
7199 			tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
7200 			hpts_removed = 1;
7201 		}
7202 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7203 	}
7204 	if (hpts_removed == 0)
7205 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7206 }
7207 
7208 static void
7209 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
7210 {
7211 	return;
7212 }
7213 
7214 static int
7215 rack_stopall(struct tcpcb *tp)
7216 {
7217 	struct tcp_rack *rack;
7218 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7219 	rack->t_timers_stopped = 1;
7220 	return (0);
7221 }
7222 
7223 static void
7224 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
7225 {
7226 	return;
7227 }
7228 
7229 static int
7230 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
7231 {
7232 	return (0);
7233 }
7234 
7235 static void
7236 rack_stop_all_timers(struct tcpcb *tp)
7237 {
7238 	struct tcp_rack *rack;
7239 
7240 	/*
7241 	 * Assure no timers are running.
7242 	 */
7243 	if (tcp_timer_active(tp, TT_PERSIST)) {
7244 		/* We enter in persists, set the flag appropriately */
7245 		rack = (struct tcp_rack *)tp->t_fb_ptr;
7246 		rack->rc_in_persist = 1;
7247 	}
7248 	tcp_timer_suspend(tp, TT_PERSIST);
7249 	tcp_timer_suspend(tp, TT_REXMT);
7250 	tcp_timer_suspend(tp, TT_KEEP);
7251 	tcp_timer_suspend(tp, TT_DELACK);
7252 }
7253 
7254 static void
7255 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7256     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7257 {
7258 	int32_t idx;
7259 
7260 	rsm->r_rtr_cnt++;
7261 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7262 	rsm->r_dupack = 0;
7263 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7264 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7265 		rsm->r_flags |= RACK_OVERMAX;
7266 	}
7267 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7268 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7269 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7270 	}
7271 	idx = rsm->r_rtr_cnt - 1;
7272 	rsm->r_tim_lastsent[idx] = ts;
7273 	/*
7274 	 * Here we don't add in the len of send, since its already
7275 	 * in snduna <->snd_max.
7276 	 */
7277 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
7278 				     rack->r_ctl.rc_sacked);
7279 	if (rsm->r_flags & RACK_ACKED) {
7280 		/* Problably MTU discovery messing with us */
7281 		rsm->r_flags &= ~RACK_ACKED;
7282 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7283 	}
7284 	if (rsm->r_in_tmap) {
7285 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7286 		rsm->r_in_tmap = 0;
7287 	}
7288 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7289 	rsm->r_in_tmap = 1;
7290 	if (rsm->r_flags & RACK_SACK_PASSED) {
7291 		/* We have retransmitted due to the SACK pass */
7292 		rsm->r_flags &= ~RACK_SACK_PASSED;
7293 		rsm->r_flags |= RACK_WAS_SACKPASS;
7294 	}
7295 }
7296 
7297 static uint32_t
7298 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7299     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7300 {
7301 	/*
7302 	 * We (re-)transmitted starting at rsm->r_start for some length
7303 	 * (possibly less than r_end.
7304 	 */
7305 	struct rack_sendmap *nrsm, *insret;
7306 	uint32_t c_end;
7307 	int32_t len;
7308 
7309 	len = *lenp;
7310 	c_end = rsm->r_start + len;
7311 	if (SEQ_GEQ(c_end, rsm->r_end)) {
7312 		/*
7313 		 * We retransmitted the whole piece or more than the whole
7314 		 * slopping into the next rsm.
7315 		 */
7316 		rack_update_rsm(tp, rack, rsm, ts, add_flag);
7317 		if (c_end == rsm->r_end) {
7318 			*lenp = 0;
7319 			return (0);
7320 		} else {
7321 			int32_t act_len;
7322 
7323 			/* Hangs over the end return whats left */
7324 			act_len = rsm->r_end - rsm->r_start;
7325 			*lenp = (len - act_len);
7326 			return (rsm->r_end);
7327 		}
7328 		/* We don't get out of this block. */
7329 	}
7330 	/*
7331 	 * Here we retransmitted less than the whole thing which means we
7332 	 * have to split this into what was transmitted and what was not.
7333 	 */
7334 	nrsm = rack_alloc_full_limit(rack);
7335 	if (nrsm == NULL) {
7336 		/*
7337 		 * We can't get memory, so lets not proceed.
7338 		 */
7339 		*lenp = 0;
7340 		return (0);
7341 	}
7342 	/*
7343 	 * So here we are going to take the original rsm and make it what we
7344 	 * retransmitted. nrsm will be the tail portion we did not
7345 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7346 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7347 	 * 1, 6 and the new piece will be 6, 11.
7348 	 */
7349 	rack_clone_rsm(rack, nrsm, rsm, c_end);
7350 	nrsm->r_dupack = 0;
7351 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7352 	insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7353 #ifdef INVARIANTS
7354 	if (insret != NULL) {
7355 		panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7356 		      nrsm, insret, rack, rsm);
7357 	}
7358 #endif
7359 	if (rsm->r_in_tmap) {
7360 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7361 		nrsm->r_in_tmap = 1;
7362 	}
7363 	rsm->r_flags &= (~RACK_HAS_FIN);
7364 	rack_update_rsm(tp, rack, rsm, ts, add_flag);
7365 	/* Log a split of rsm into rsm and nrsm */
7366 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7367 	*lenp = 0;
7368 	return (0);
7369 }
7370 
7371 static void
7372 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7373 		uint32_t seq_out, uint8_t th_flags, int32_t err, uint64_t cts,
7374 		struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls)
7375 {
7376 	struct tcp_rack *rack;
7377 	struct rack_sendmap *rsm, *nrsm, *insret, fe;
7378 	register uint32_t snd_max, snd_una;
7379 
7380 	/*
7381 	 * Add to the RACK log of packets in flight or retransmitted. If
7382 	 * there is a TS option we will use the TS echoed, if not we will
7383 	 * grab a TS.
7384 	 *
7385 	 * Retransmissions will increment the count and move the ts to its
7386 	 * proper place. Note that if options do not include TS's then we
7387 	 * won't be able to effectively use the ACK for an RTT on a retran.
7388 	 *
7389 	 * Notes about r_start and r_end. Lets consider a send starting at
7390 	 * sequence 1 for 10 bytes. In such an example the r_start would be
7391 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7392 	 * This means that r_end is actually the first sequence for the next
7393 	 * slot (11).
7394 	 *
7395 	 */
7396 	/*
7397 	 * If err is set what do we do XXXrrs? should we not add the thing?
7398 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
7399 	 * i.e. proceed with add ** do this for now.
7400 	 */
7401 	INP_WLOCK_ASSERT(tp->t_inpcb);
7402 	if (err)
7403 		/*
7404 		 * We don't log errors -- we could but snd_max does not
7405 		 * advance in this case either.
7406 		 */
7407 		return;
7408 
7409 	if (th_flags & TH_RST) {
7410 		/*
7411 		 * We don't log resets and we return immediately from
7412 		 * sending
7413 		 */
7414 		return;
7415 	}
7416 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7417 	snd_una = tp->snd_una;
7418 	snd_max = tp->snd_max;
7419 	if (th_flags & (TH_SYN | TH_FIN)) {
7420 		/*
7421 		 * The call to rack_log_output is made before bumping
7422 		 * snd_max. This means we can record one extra byte on a SYN
7423 		 * or FIN if seq_out is adding more on and a FIN is present
7424 		 * (and we are not resending).
7425 		 */
7426 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7427 			len++;
7428 		if (th_flags & TH_FIN)
7429 			len++;
7430 		if (SEQ_LT(snd_max, tp->snd_nxt)) {
7431 			/*
7432 			 * The add/update as not been done for the FIN/SYN
7433 			 * yet.
7434 			 */
7435 			snd_max = tp->snd_nxt;
7436 		}
7437 	}
7438 	if (SEQ_LEQ((seq_out + len), snd_una)) {
7439 		/* Are sending an old segment to induce an ack (keep-alive)? */
7440 		return;
7441 	}
7442 	if (SEQ_LT(seq_out, snd_una)) {
7443 		/* huh? should we panic? */
7444 		uint32_t end;
7445 
7446 		end = seq_out + len;
7447 		seq_out = snd_una;
7448 		if (SEQ_GEQ(end, seq_out))
7449 			len = end - seq_out;
7450 		else
7451 			len = 0;
7452 	}
7453 	if (len == 0) {
7454 		/* We don't log zero window probes */
7455 		return;
7456 	}
7457 	rack->r_ctl.rc_time_last_sent = cts;
7458 	if (IN_FASTRECOVERY(tp->t_flags)) {
7459 		rack->r_ctl.rc_prr_out += len;
7460 	}
7461 	/* First question is it a retransmission or new? */
7462 	if (seq_out == snd_max) {
7463 		/* Its new */
7464 again:
7465 		rsm = rack_alloc(rack);
7466 		if (rsm == NULL) {
7467 			/*
7468 			 * Hmm out of memory and the tcb got destroyed while
7469 			 * we tried to wait.
7470 			 */
7471 			return;
7472 		}
7473 		if (th_flags & TH_FIN) {
7474 			rsm->r_flags = RACK_HAS_FIN|add_flag;
7475 		} else {
7476 			rsm->r_flags = add_flag;
7477 		}
7478 		if (hw_tls)
7479 			rsm->r_hw_tls = 1;
7480 		rsm->r_tim_lastsent[0] = cts;
7481 		rsm->r_rtr_cnt = 1;
7482 		rsm->r_rtr_bytes = 0;
7483 		if (th_flags & TH_SYN) {
7484 			/* The data space is one beyond snd_una */
7485 			rsm->r_flags |= RACK_HAS_SYN;
7486 		}
7487 		rsm->r_start = seq_out;
7488 		rsm->r_end = rsm->r_start + len;
7489 		rsm->r_dupack = 0;
7490 		/*
7491 		 * save off the mbuf location that
7492 		 * sndmbuf_noadv returned (which is
7493 		 * where we started copying from)..
7494 		 */
7495 		rsm->m = s_mb;
7496 		rsm->soff = s_moff;
7497 		/*
7498 		 * Here we do add in the len of send, since its not yet
7499 		 * reflected in in snduna <->snd_max
7500 		 */
7501 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
7502 					      rack->r_ctl.rc_sacked) +
7503 			      (rsm->r_end - rsm->r_start));
7504 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7505 		if (rsm->m) {
7506 			if (rsm->m->m_len <= rsm->soff) {
7507 				/*
7508 				 * XXXrrs Question, will this happen?
7509 				 *
7510 				 * If sbsndptr is set at the correct place
7511 				 * then s_moff should always be somewhere
7512 				 * within rsm->m. But if the sbsndptr was
7513 				 * off then that won't be true. If it occurs
7514 				 * we need to walkout to the correct location.
7515 				 */
7516 				struct mbuf *lm;
7517 
7518 				lm = rsm->m;
7519 				while (lm->m_len <= rsm->soff) {
7520 					rsm->soff -= lm->m_len;
7521 					lm = lm->m_next;
7522 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7523 							     __func__, rack, s_moff, s_mb, rsm->soff));
7524 				}
7525 				rsm->m = lm;
7526 				counter_u64_add(rack_sbsndptr_wrong, 1);
7527 			} else
7528 				counter_u64_add(rack_sbsndptr_right, 1);
7529 			rsm->orig_m_len = rsm->m->m_len;
7530 		} else
7531 			rsm->orig_m_len = 0;
7532 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7533 		/* Log a new rsm */
7534 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7535 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7536 #ifdef INVARIANTS
7537 		if (insret != NULL) {
7538 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7539 			      nrsm, insret, rack, rsm);
7540 		}
7541 #endif
7542 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7543 		rsm->r_in_tmap = 1;
7544 		/*
7545 		 * Special case detection, is there just a single
7546 		 * packet outstanding when we are not in recovery?
7547 		 *
7548 		 * If this is true mark it so.
7549 		 */
7550 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7551 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7552 			struct rack_sendmap *prsm;
7553 
7554 			prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7555 			if (prsm)
7556 				prsm->r_one_out_nr = 1;
7557 		}
7558 		return;
7559 	}
7560 	/*
7561 	 * If we reach here its a retransmission and we need to find it.
7562 	 */
7563 	memset(&fe, 0, sizeof(fe));
7564 more:
7565 	if (hintrsm && (hintrsm->r_start == seq_out)) {
7566 		rsm = hintrsm;
7567 		hintrsm = NULL;
7568 	} else {
7569 		/* No hints sorry */
7570 		rsm = NULL;
7571 	}
7572 	if ((rsm) && (rsm->r_start == seq_out)) {
7573 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7574 		if (len == 0) {
7575 			return;
7576 		} else {
7577 			goto more;
7578 		}
7579 	}
7580 	/* Ok it was not the last pointer go through it the hard way. */
7581 refind:
7582 	fe.r_start = seq_out;
7583 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7584 	if (rsm) {
7585 		if (rsm->r_start == seq_out) {
7586 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7587 			if (len == 0) {
7588 				return;
7589 			} else {
7590 				goto refind;
7591 			}
7592 		}
7593 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7594 			/* Transmitted within this piece */
7595 			/*
7596 			 * Ok we must split off the front and then let the
7597 			 * update do the rest
7598 			 */
7599 			nrsm = rack_alloc_full_limit(rack);
7600 			if (nrsm == NULL) {
7601 				rack_update_rsm(tp, rack, rsm, cts, add_flag);
7602 				return;
7603 			}
7604 			/*
7605 			 * copy rsm to nrsm and then trim the front of rsm
7606 			 * to not include this part.
7607 			 */
7608 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
7609 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7610 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7611 #ifdef INVARIANTS
7612 			if (insret != NULL) {
7613 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7614 				      nrsm, insret, rack, rsm);
7615 			}
7616 #endif
7617 			if (rsm->r_in_tmap) {
7618 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7619 				nrsm->r_in_tmap = 1;
7620 			}
7621 			rsm->r_flags &= (~RACK_HAS_FIN);
7622 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7623 			if (len == 0) {
7624 				return;
7625 			} else if (len > 0)
7626 				goto refind;
7627 		}
7628 	}
7629 	/*
7630 	 * Hmm not found in map did they retransmit both old and on into the
7631 	 * new?
7632 	 */
7633 	if (seq_out == tp->snd_max) {
7634 		goto again;
7635 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
7636 #ifdef INVARIANTS
7637 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7638 		       seq_out, len, tp->snd_una, tp->snd_max);
7639 		printf("Starting Dump of all rack entries\n");
7640 		RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7641 			printf("rsm:%p start:%u end:%u\n",
7642 			       rsm, rsm->r_start, rsm->r_end);
7643 		}
7644 		printf("Dump complete\n");
7645 		panic("seq_out not found rack:%p tp:%p",
7646 		      rack, tp);
7647 #endif
7648 	} else {
7649 #ifdef INVARIANTS
7650 		/*
7651 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
7652 		 * flag)
7653 		 */
7654 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7655 		      seq_out, len, tp->snd_max, tp);
7656 #endif
7657 	}
7658 }
7659 
7660 /*
7661  * Record one of the RTT updates from an ack into
7662  * our sample structure.
7663  */
7664 
7665 static void
7666 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7667 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7668 {
7669 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7670 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7671 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7672 	}
7673 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7674 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7675 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7676 	}
7677 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7678 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7679 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
7680 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7681 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7682 	}
7683 	if ((confidence == 1) &&
7684 	    ((rsm == NULL) ||
7685 	     (rsm->r_just_ret) ||
7686 	     (rsm->r_one_out_nr &&
7687 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7688 		/*
7689 		 * If the rsm had a just return
7690 		 * hit it then we can't trust the
7691 		 * rtt measurement for buffer deterimination
7692 		 * Note that a confidence of 2, indicates
7693 		 * SACK'd which overrides the r_just_ret or
7694 		 * the r_one_out_nr. If it was a CUM-ACK and
7695 		 * we had only two outstanding, but get an
7696 		 * ack for only 1. Then that also lowers our
7697 		 * confidence.
7698 		 */
7699 		confidence = 0;
7700 	}
7701 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7702 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7703 		if (rack->r_ctl.rack_rs.confidence == 0) {
7704 			/*
7705 			 * We take anything with no current confidence
7706 			 * saved.
7707 			 */
7708 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7709 			rack->r_ctl.rack_rs.confidence = confidence;
7710 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7711 		} else if (confidence || rack->r_ctl.rack_rs.confidence) {
7712 			/*
7713 			 * Once we have a confident number,
7714 			 * we can update it with a smaller
7715 			 * value since this confident number
7716 			 * may include the DSACK time until
7717 			 * the next segment (the second one) arrived.
7718 			 */
7719 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7720 			rack->r_ctl.rack_rs.confidence = confidence;
7721 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7722 		}
7723 	}
7724 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7725 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7726 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7727 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
7728 }
7729 
7730 /*
7731  * Collect new round-trip time estimate
7732  * and update averages and current timeout.
7733  */
7734 static void
7735 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7736 {
7737 	int32_t delta;
7738 	uint32_t o_srtt, o_var;
7739 	int32_t hrtt_up = 0;
7740 	int32_t rtt;
7741 
7742 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7743 		/* No valid sample */
7744 		return;
7745 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7746 		/* We are to use the lowest RTT seen in a single ack */
7747 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7748 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7749 		/* We are to use the highest RTT seen in a single ack */
7750 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7751 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7752 		/* We are to use the average RTT seen in a single ack */
7753 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7754 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7755 	} else {
7756 #ifdef INVARIANTS
7757 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7758 #endif
7759 		return;
7760 	}
7761 	if (rtt == 0)
7762 		rtt = 1;
7763 	if (rack->rc_gp_rtt_set == 0) {
7764 		/*
7765 		 * With no RTT we have to accept
7766 		 * even one we are not confident of.
7767 		 */
7768 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7769 		rack->rc_gp_rtt_set = 1;
7770 	} else if (rack->r_ctl.rack_rs.confidence) {
7771 		/* update the running gp srtt */
7772 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7773 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7774 	}
7775 	if (rack->r_ctl.rack_rs.confidence) {
7776 		/*
7777 		 * record the low and high for highly buffered path computation,
7778 		 * we only do this if we are confident (not a retransmission).
7779 		 */
7780 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7781 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7782 			hrtt_up = 1;
7783 		}
7784 		if (rack->rc_highly_buffered == 0) {
7785 			/*
7786 			 * Currently once we declare a path has
7787 			 * highly buffered there is no going
7788 			 * back, which may be a problem...
7789 			 */
7790 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7791 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7792 						     rack->r_ctl.rc_highest_us_rtt,
7793 						     rack->r_ctl.rc_lowest_us_rtt,
7794 						     RACK_RTTS_SEEHBP);
7795 				rack->rc_highly_buffered = 1;
7796 			}
7797 		}
7798 	}
7799 	if ((rack->r_ctl.rack_rs.confidence) ||
7800 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7801 		/*
7802 		 * If we are highly confident of it <or> it was
7803 		 * never retransmitted we accept it as the last us_rtt.
7804 		 */
7805 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7806 		/* The lowest rtt can be set if its was not retransmited */
7807 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7808 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7809 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
7810 				rack->r_ctl.rc_lowest_us_rtt = 1;
7811 		}
7812 	}
7813 	o_srtt = tp->t_srtt;
7814 	o_var = tp->t_rttvar;
7815 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7816 	if (tp->t_srtt != 0) {
7817 		/*
7818 		 * We keep a simple srtt in microseconds, like our rtt
7819 		 * measurement. We don't need to do any tricks with shifting
7820 		 * etc. Instead we just add in 1/8th of the new measurement
7821 		 * and subtract out 1/8 of the old srtt. We do the same with
7822 		 * the variance after finding the absolute value of the
7823 		 * difference between this sample and the current srtt.
7824 		 */
7825 		delta = tp->t_srtt - rtt;
7826 		/* Take off 1/8th of the current sRTT */
7827 		tp->t_srtt -= (tp->t_srtt >> 3);
7828 		/* Add in 1/8th of the new RTT just measured */
7829 		tp->t_srtt += (rtt >> 3);
7830 		if (tp->t_srtt <= 0)
7831 			tp->t_srtt = 1;
7832 		/* Now lets make the absolute value of the variance */
7833 		if (delta < 0)
7834 			delta = -delta;
7835 		/* Subtract out 1/8th */
7836 		tp->t_rttvar -= (tp->t_rttvar >> 3);
7837 		/* Add in 1/8th of the new variance we just saw */
7838 		tp->t_rttvar += (delta >> 3);
7839 		if (tp->t_rttvar <= 0)
7840 			tp->t_rttvar = 1;
7841 		if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
7842 			tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7843 	} else {
7844 		/*
7845 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
7846 		 * variance to half the rtt (so our first retransmit happens
7847 		 * at 3*rtt).
7848 		 */
7849 		tp->t_srtt = rtt;
7850 		tp->t_rttvar = rtt >> 1;
7851 		tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7852 	}
7853 	rack->rc_srtt_measure_made = 1;
7854 	KMOD_TCPSTAT_INC(tcps_rttupdated);
7855 	tp->t_rttupdated++;
7856 #ifdef STATS
7857 	if (rack_stats_gets_ms_rtt == 0) {
7858 		/* Send in the microsecond rtt used for rxt timeout purposes */
7859 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7860 	} else if (rack_stats_gets_ms_rtt == 1) {
7861 		/* Send in the millisecond rtt used for rxt timeout purposes */
7862 		int32_t ms_rtt;
7863 
7864 		/* Round up */
7865 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7866 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7867 	} else if (rack_stats_gets_ms_rtt == 2) {
7868 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
7869 		int32_t ms_rtt;
7870 
7871 		/* Round up */
7872 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7873 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7874 	}  else {
7875 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
7876 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7877 	}
7878 
7879 #endif
7880 	/*
7881 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
7882 	 * way we do the smoothing, srtt and rttvar will each average +1/2
7883 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
7884 	 * tick of rounding and 1 extra tick because of +-1/2 tick
7885 	 * uncertainty in the firing of the timer.  The bias will give us
7886 	 * exactly the 1.5 tick we need.  But, because the bias is
7887 	 * statistical, we have to test that we don't drop below the minimum
7888 	 * feasible timer (which is 2 ticks).
7889 	 */
7890 	tp->t_rxtshift = 0;
7891 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7892 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7893 	rack_log_rtt_sample(rack, rtt);
7894 	tp->t_softerror = 0;
7895 }
7896 
7897 
7898 static void
7899 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7900 {
7901 	/*
7902 	 * Apply to filter the inbound us-rtt at us_cts.
7903 	 */
7904 	uint32_t old_rtt;
7905 
7906 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7907 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7908 			       us_rtt, us_cts);
7909 	if (rack->r_ctl.last_pacing_time &&
7910 	    rack->rc_gp_dyn_mul &&
7911 	    (rack->r_ctl.last_pacing_time > us_rtt))
7912 		rack->pacing_longer_than_rtt = 1;
7913 	else
7914 		rack->pacing_longer_than_rtt = 0;
7915 	if (old_rtt > us_rtt) {
7916 		/* We just hit a new lower rtt time */
7917 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7918 				     __LINE__, RACK_RTTS_NEWRTT);
7919 		/*
7920 		 * Only count it if its lower than what we saw within our
7921 		 * calculated range.
7922 		 */
7923 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7924 			if (rack_probertt_lower_within &&
7925 			    rack->rc_gp_dyn_mul &&
7926 			    (rack->use_fixed_rate == 0) &&
7927 			    (rack->rc_always_pace)) {
7928 				/*
7929 				 * We are seeing a new lower rtt very close
7930 				 * to the time that we would have entered probe-rtt.
7931 				 * This is probably due to the fact that a peer flow
7932 				 * has entered probe-rtt. Lets go in now too.
7933 				 */
7934 				uint32_t val;
7935 
7936 				val = rack_probertt_lower_within * rack_time_between_probertt;
7937 				val /= 100;
7938 				if ((rack->in_probe_rtt == 0)  &&
7939 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
7940 					rack_enter_probertt(rack, us_cts);
7941 				}
7942 			}
7943 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7944 		}
7945 	}
7946 }
7947 
7948 static int
7949 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7950     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7951 {
7952 	uint32_t us_rtt;
7953 	int32_t i, all;
7954 	uint32_t t, len_acked;
7955 
7956 	if ((rsm->r_flags & RACK_ACKED) ||
7957 	    (rsm->r_flags & RACK_WAS_ACKED))
7958 		/* Already done */
7959 		return (0);
7960 	if (rsm->r_no_rtt_allowed) {
7961 		/* Not allowed */
7962 		return (0);
7963 	}
7964 	if (ack_type == CUM_ACKED) {
7965 		if (SEQ_GT(th_ack, rsm->r_end)) {
7966 			len_acked = rsm->r_end - rsm->r_start;
7967 			all = 1;
7968 		} else {
7969 			len_acked = th_ack - rsm->r_start;
7970 			all = 0;
7971 		}
7972 	} else {
7973 		len_acked = rsm->r_end - rsm->r_start;
7974 		all = 0;
7975 	}
7976 	if (rsm->r_rtr_cnt == 1) {
7977 
7978 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7979 		if ((int)t <= 0)
7980 			t = 1;
7981 		if (!tp->t_rttlow || tp->t_rttlow > t)
7982 			tp->t_rttlow = t;
7983 		if (!rack->r_ctl.rc_rack_min_rtt ||
7984 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7985 			rack->r_ctl.rc_rack_min_rtt = t;
7986 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
7987 				rack->r_ctl.rc_rack_min_rtt = 1;
7988 			}
7989 		}
7990 		if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7991 			us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7992 		else
7993 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7994 		if (us_rtt == 0)
7995 			us_rtt = 1;
7996 		if (CC_ALGO(tp)->rttsample != NULL) {
7997 			/* Kick the RTT to the CC */
7998 			CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas);
7999 		}
8000 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
8001 		if (ack_type == SACKED) {
8002 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
8003 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
8004 		} else {
8005 			/*
8006 			 * We need to setup what our confidence
8007 			 * is in this ack.
8008 			 *
8009 			 * If the rsm was app limited and it is
8010 			 * less than a mss in length (the end
8011 			 * of the send) then we have a gap. If we
8012 			 * were app limited but say we were sending
8013 			 * multiple MSS's then we are more confident
8014 			 * int it.
8015 			 *
8016 			 * When we are not app-limited then we see if
8017 			 * the rsm is being included in the current
8018 			 * measurement, we tell this by the app_limited_needs_set
8019 			 * flag.
8020 			 *
8021 			 * Note that being cwnd blocked is not applimited
8022 			 * as well as the pacing delay between packets which
8023 			 * are sending only 1 or 2 MSS's also will show up
8024 			 * in the RTT. We probably need to examine this algorithm
8025 			 * a bit more and enhance it to account for the delay
8026 			 * between rsm's. We could do that by saving off the
8027 			 * pacing delay of each rsm (in an rsm) and then
8028 			 * factoring that in somehow though for now I am
8029 			 * not sure how :)
8030 			 */
8031 			int calc_conf = 0;
8032 
8033 			if (rsm->r_flags & RACK_APP_LIMITED) {
8034 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
8035 					calc_conf = 0;
8036 				else
8037 					calc_conf = 1;
8038 			} else if (rack->app_limited_needs_set == 0) {
8039 				calc_conf = 1;
8040 			} else {
8041 				calc_conf = 0;
8042 			}
8043 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
8044 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
8045 					    calc_conf, rsm, rsm->r_rtr_cnt);
8046 		}
8047 		if ((rsm->r_flags & RACK_TLP) &&
8048 		    (!IN_FASTRECOVERY(tp->t_flags))) {
8049 			/* Segment was a TLP and our retrans matched */
8050 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
8051 				rack->r_ctl.rc_rsm_start = tp->snd_max;
8052 				rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
8053 				rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
8054 				rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
8055 			}
8056 		}
8057 		if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
8058 			/* New more recent rack_tmit_time */
8059 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8060 			rack->rc_rack_rtt = t;
8061 		}
8062 		return (1);
8063 	}
8064 	/*
8065 	 * We clear the soft/rxtshift since we got an ack.
8066 	 * There is no assurance we will call the commit() function
8067 	 * so we need to clear these to avoid incorrect handling.
8068 	 */
8069 	tp->t_rxtshift = 0;
8070 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8071 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
8072 	tp->t_softerror = 0;
8073 	if (to && (to->to_flags & TOF_TS) &&
8074 	    (ack_type == CUM_ACKED) &&
8075 	    (to->to_tsecr) &&
8076 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
8077 		/*
8078 		 * Now which timestamp does it match? In this block the ACK
8079 		 * must be coming from a previous transmission.
8080 		 */
8081 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
8082 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
8083 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
8084 				if ((int)t <= 0)
8085 					t = 1;
8086 				if (CC_ALGO(tp)->rttsample != NULL) {
8087 					/*
8088 					 * Kick the RTT to the CC, here
8089 					 * we lie a bit in that we know the
8090 					 * retransmission is correct even though
8091 					 * we retransmitted. This is because
8092 					 * we match the timestamps.
8093 					 */
8094 					if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
8095 						us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
8096 					else
8097 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
8098 					CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas);
8099 				}
8100 				if ((i + 1) < rsm->r_rtr_cnt) {
8101 					/*
8102 					 * The peer ack'd from our previous
8103 					 * transmission. We have a spurious
8104 					 * retransmission and thus we dont
8105 					 * want to update our rack_rtt.
8106 					 *
8107 					 * Hmm should there be a CC revert here?
8108 					 *
8109 					 */
8110 					return (0);
8111 				}
8112 				if (!tp->t_rttlow || tp->t_rttlow > t)
8113 					tp->t_rttlow = t;
8114 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8115 					rack->r_ctl.rc_rack_min_rtt = t;
8116 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
8117 						rack->r_ctl.rc_rack_min_rtt = 1;
8118 					}
8119 				}
8120 				if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
8121 					   (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
8122 					/* New more recent rack_tmit_time */
8123 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8124 					rack->rc_rack_rtt = t;
8125 				}
8126 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
8127 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
8128 						    rsm->r_rtr_cnt);
8129 				return (1);
8130 			}
8131 		}
8132 		goto ts_not_found;
8133 	} else {
8134 		/*
8135 		 * Ok its a SACK block that we retransmitted. or a windows
8136 		 * machine without timestamps. We can tell nothing from the
8137 		 * time-stamp since its not there or the time the peer last
8138 		 * recieved a segment that moved forward its cum-ack point.
8139 		 */
8140 ts_not_found:
8141 		i = rsm->r_rtr_cnt - 1;
8142 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
8143 		if ((int)t <= 0)
8144 			t = 1;
8145 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8146 			/*
8147 			 * We retransmitted and the ack came back in less
8148 			 * than the smallest rtt we have observed. We most
8149 			 * likely did an improper retransmit as outlined in
8150 			 * 6.2 Step 2 point 2 in the rack-draft so we
8151 			 * don't want to update our rack_rtt. We in
8152 			 * theory (in future) might want to think about reverting our
8153 			 * cwnd state but we won't for now.
8154 			 */
8155 			return (0);
8156 		} else if (rack->r_ctl.rc_rack_min_rtt) {
8157 			/*
8158 			 * We retransmitted it and the retransmit did the
8159 			 * job.
8160 			 */
8161 			if (!rack->r_ctl.rc_rack_min_rtt ||
8162 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8163 				rack->r_ctl.rc_rack_min_rtt = t;
8164 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
8165 					rack->r_ctl.rc_rack_min_rtt = 1;
8166 				}
8167 			}
8168 			if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
8169 				/* New more recent rack_tmit_time */
8170 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
8171 				rack->rc_rack_rtt = t;
8172 			}
8173 			return (1);
8174 		}
8175 	}
8176 	return (0);
8177 }
8178 
8179 /*
8180  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
8181  */
8182 static void
8183 rack_log_sack_passed(struct tcpcb *tp,
8184     struct tcp_rack *rack, struct rack_sendmap *rsm)
8185 {
8186 	struct rack_sendmap *nrsm;
8187 
8188 	nrsm = rsm;
8189 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
8190 	    rack_head, r_tnext) {
8191 		if (nrsm == rsm) {
8192 			/* Skip orginal segment he is acked */
8193 			continue;
8194 		}
8195 		if (nrsm->r_flags & RACK_ACKED) {
8196 			/*
8197 			 * Skip ack'd segments, though we
8198 			 * should not see these, since tmap
8199 			 * should not have ack'd segments.
8200 			 */
8201 			continue;
8202 		}
8203 		if (nrsm->r_flags & RACK_SACK_PASSED) {
8204 			/*
8205 			 * We found one that is already marked
8206 			 * passed, we have been here before and
8207 			 * so all others below this are marked.
8208 			 */
8209 			break;
8210 		}
8211 		nrsm->r_flags |= RACK_SACK_PASSED;
8212 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
8213 	}
8214 }
8215 
8216 static void
8217 rack_need_set_test(struct tcpcb *tp,
8218 		   struct tcp_rack *rack,
8219 		   struct rack_sendmap *rsm,
8220 		   tcp_seq th_ack,
8221 		   int line,
8222 		   int use_which)
8223 {
8224 
8225 	if ((tp->t_flags & TF_GPUTINPROG) &&
8226 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8227 		/*
8228 		 * We were app limited, and this ack
8229 		 * butts up or goes beyond the point where we want
8230 		 * to start our next measurement. We need
8231 		 * to record the new gput_ts as here and
8232 		 * possibly update the start sequence.
8233 		 */
8234 		uint32_t seq, ts;
8235 
8236 		if (rsm->r_rtr_cnt > 1) {
8237 			/*
8238 			 * This is a retransmit, can we
8239 			 * really make any assessment at this
8240 			 * point?  We are not really sure of
8241 			 * the timestamp, is it this or the
8242 			 * previous transmission?
8243 			 *
8244 			 * Lets wait for something better that
8245 			 * is not retransmitted.
8246 			 */
8247 			return;
8248 		}
8249 		seq = tp->gput_seq;
8250 		ts = tp->gput_ts;
8251 		rack->app_limited_needs_set = 0;
8252 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8253 		/* Do we start at a new end? */
8254 		if ((use_which == RACK_USE_BEG) &&
8255 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8256 			/*
8257 			 * When we get an ACK that just eats
8258 			 * up some of the rsm, we set RACK_USE_BEG
8259 			 * since whats at r_start (i.e. th_ack)
8260 			 * is left unacked and thats where the
8261 			 * measurement not starts.
8262 			 */
8263 			tp->gput_seq = rsm->r_start;
8264 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8265 		}
8266 		if ((use_which == RACK_USE_END) &&
8267 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8268 			    /*
8269 			     * We use the end when the cumack
8270 			     * is moving forward and completely
8271 			     * deleting the rsm passed so basically
8272 			     * r_end holds th_ack.
8273 			     *
8274 			     * For SACK's we also want to use the end
8275 			     * since this piece just got sacked and
8276 			     * we want to target anything after that
8277 			     * in our measurement.
8278 			     */
8279 			    tp->gput_seq = rsm->r_end;
8280 			    rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8281 		}
8282 		if (use_which == RACK_USE_END_OR_THACK) {
8283 			/*
8284 			 * special case for ack moving forward,
8285 			 * not a sack, we need to move all the
8286 			 * way up to where this ack cum-ack moves
8287 			 * to.
8288 			 */
8289 			if (SEQ_GT(th_ack, rsm->r_end))
8290 				tp->gput_seq = th_ack;
8291 			else
8292 				tp->gput_seq = rsm->r_end;
8293 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8294 		}
8295 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8296 			/*
8297 			 * We moved beyond this guy's range, re-calculate
8298 			 * the new end point.
8299 			 */
8300 			if (rack->rc_gp_filled == 0) {
8301 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8302 			} else {
8303 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8304 			}
8305 		}
8306 		/*
8307 		 * We are moving the goal post, we may be able to clear the
8308 		 * measure_saw_probe_rtt flag.
8309 		 */
8310 		if ((rack->in_probe_rtt == 0) &&
8311 		    (rack->measure_saw_probe_rtt) &&
8312 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8313 			rack->measure_saw_probe_rtt = 0;
8314 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8315 					   seq, tp->gput_seq, 0, 5, line, NULL, 0);
8316 		if (rack->rc_gp_filled &&
8317 		    ((tp->gput_ack - tp->gput_seq) <
8318 		     max(rc_init_window(rack), (MIN_GP_WIN *
8319 						ctf_fixed_maxseg(tp))))) {
8320 			uint32_t ideal_amount;
8321 
8322 			ideal_amount = rack_get_measure_window(tp, rack);
8323 			if (ideal_amount > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
8324 				/*
8325 				 * There is no sense of continuing this measurement
8326 				 * because its too small to gain us anything we
8327 				 * trust. Skip it and that way we can start a new
8328 				 * measurement quicker.
8329 				 */
8330 				tp->t_flags &= ~TF_GPUTINPROG;
8331 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8332 							   0, 0, 0, 6, __LINE__, NULL, 0);
8333 			} else {
8334 				/*
8335 				 * Reset the window further out.
8336 				 */
8337 				tp->gput_ack = tp->gput_seq + ideal_amount;
8338 			}
8339 		}
8340 	}
8341 }
8342 
8343 static inline int
8344 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
8345 {
8346 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
8347 		/* Behind our TLP definition or right at */
8348 		return (0);
8349 	}
8350 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
8351 		/* The start is beyond or right at our end of TLP definition */
8352 		return (0);
8353 	}
8354 	/* It has to be a sub-part of the original TLP recorded */
8355 	return (1);
8356 }
8357 
8358 
8359 static uint32_t
8360 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8361 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8362 {
8363 	uint32_t start, end, changed = 0;
8364 	struct rack_sendmap stack_map;
8365 	struct rack_sendmap *rsm, *nrsm, fe, *insret, *prev, *next;
8366 	int32_t used_ref = 1;
8367 	int moved = 0;
8368 
8369 	start = sack->start;
8370 	end = sack->end;
8371 	rsm = *prsm;
8372 	memset(&fe, 0, sizeof(fe));
8373 do_rest_ofb:
8374 	if ((rsm == NULL) ||
8375 	    (SEQ_LT(end, rsm->r_start)) ||
8376 	    (SEQ_GEQ(start, rsm->r_end)) ||
8377 	    (SEQ_LT(start, rsm->r_start))) {
8378 		/*
8379 		 * We are not in the right spot,
8380 		 * find the correct spot in the tree.
8381 		 */
8382 		used_ref = 0;
8383 		fe.r_start = start;
8384 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8385 		moved++;
8386 	}
8387 	if (rsm == NULL) {
8388 		/* TSNH */
8389 		goto out;
8390 	}
8391 	/* Ok we have an ACK for some piece of this rsm */
8392 	if (rsm->r_start != start) {
8393 		if ((rsm->r_flags & RACK_ACKED) == 0) {
8394 			/*
8395 			 * Before any splitting or hookery is
8396 			 * done is it a TLP of interest i.e. rxt?
8397 			 */
8398 			if ((rsm->r_flags & RACK_TLP) &&
8399 			    (rsm->r_rtr_cnt > 1)) {
8400 				/*
8401 				 * We are splitting a rxt TLP, check
8402 				 * if we need to save off the start/end
8403 				 */
8404 				if (rack->rc_last_tlp_acked_set &&
8405 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8406 					/*
8407 					 * We already turned this on since we are inside
8408 					 * the previous one was a partially sack now we
8409 					 * are getting another one (maybe all of it).
8410 					 *
8411 					 */
8412 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8413 					/*
8414 					 * Lets make sure we have all of it though.
8415 					 */
8416 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8417 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8418 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8419 								     rack->r_ctl.last_tlp_acked_end);
8420 					}
8421 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8422 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8423 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8424 								     rack->r_ctl.last_tlp_acked_end);
8425 					}
8426 				} else {
8427 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8428 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8429 					rack->rc_last_tlp_past_cumack = 0;
8430 					rack->rc_last_tlp_acked_set = 1;
8431 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8432 				}
8433 			}
8434 			/**
8435 			 * Need to split this in two pieces the before and after,
8436 			 * the before remains in the map, the after must be
8437 			 * added. In other words we have:
8438 			 * rsm        |--------------|
8439 			 * sackblk        |------->
8440 			 * rsm will become
8441 			 *     rsm    |---|
8442 			 * and nrsm will be  the sacked piece
8443 			 *     nrsm       |----------|
8444 			 *
8445 			 * But before we start down that path lets
8446 			 * see if the sack spans over on top of
8447 			 * the next guy and it is already sacked.
8448 			 *
8449 			 */
8450 			next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8451 			if (next && (next->r_flags & RACK_ACKED) &&
8452 			    SEQ_GEQ(end, next->r_start)) {
8453 				/**
8454 				 * So the next one is already acked, and
8455 				 * we can thus by hookery use our stack_map
8456 				 * to reflect the piece being sacked and
8457 				 * then adjust the two tree entries moving
8458 				 * the start and ends around. So we start like:
8459 				 *  rsm     |------------|             (not-acked)
8460 				 *  next                 |-----------| (acked)
8461 				 *  sackblk        |-------->
8462 				 *  We want to end like so:
8463 				 *  rsm     |------|                   (not-acked)
8464 				 *  next           |-----------------| (acked)
8465 				 *  nrsm           |-----|
8466 				 * Where nrsm is a temporary stack piece we
8467 				 * use to update all the gizmos.
8468 				 */
8469 				/* Copy up our fudge block */
8470 				nrsm = &stack_map;
8471 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8472 				/* Now adjust our tree blocks */
8473 				rsm->r_end = start;
8474 				next->r_start = start;
8475 				/* Now we must adjust back where next->m is */
8476 				rack_setup_offset_for_rsm(rsm, next);
8477 
8478 				/* We don't need to adjust rsm, it did not change */
8479 				/* Clear out the dup ack count of the remainder */
8480 				rsm->r_dupack = 0;
8481 				rsm->r_just_ret = 0;
8482 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8483 				/* Now lets make sure our fudge block is right */
8484 				nrsm->r_start = start;
8485 				/* Now lets update all the stats and such */
8486 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8487 				if (rack->app_limited_needs_set)
8488 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8489 				changed += (nrsm->r_end - nrsm->r_start);
8490 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8491 				if (nrsm->r_flags & RACK_SACK_PASSED) {
8492 					counter_u64_add(rack_reorder_seen, 1);
8493 					rack->r_ctl.rc_reorder_ts = cts;
8494 				}
8495 				/*
8496 				 * Now we want to go up from rsm (the
8497 				 * one left un-acked) to the next one
8498 				 * in the tmap. We do this so when
8499 				 * we walk backwards we include marking
8500 				 * sack-passed on rsm (The one passed in
8501 				 * is skipped since it is generally called
8502 				 * on something sacked before removing it
8503 				 * from the tmap).
8504 				 */
8505 				if (rsm->r_in_tmap) {
8506 					nrsm = TAILQ_NEXT(rsm, r_tnext);
8507 					/*
8508 					 * Now that we have the next
8509 					 * one walk backwards from there.
8510 					 */
8511 					if (nrsm && nrsm->r_in_tmap)
8512 						rack_log_sack_passed(tp, rack, nrsm);
8513 				}
8514 				/* Now are we done? */
8515 				if (SEQ_LT(end, next->r_end) ||
8516 				    (end == next->r_end)) {
8517 					/* Done with block */
8518 					goto out;
8519 				}
8520 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8521 				counter_u64_add(rack_sack_used_next_merge, 1);
8522 				/* Postion for the next block */
8523 				start = next->r_end;
8524 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8525 				if (rsm == NULL)
8526 					goto out;
8527 			} else {
8528 				/**
8529 				 * We can't use any hookery here, so we
8530 				 * need to split the map. We enter like
8531 				 * so:
8532 				 *  rsm      |--------|
8533 				 *  sackblk       |----->
8534 				 * We will add the new block nrsm and
8535 				 * that will be the new portion, and then
8536 				 * fall through after reseting rsm. So we
8537 				 * split and look like this:
8538 				 *  rsm      |----|
8539 				 *  sackblk       |----->
8540 				 *  nrsm          |---|
8541 				 * We then fall through reseting
8542 				 * rsm to nrsm, so the next block
8543 				 * picks it up.
8544 				 */
8545 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8546 				if (nrsm == NULL) {
8547 					/*
8548 					 * failed XXXrrs what can we do but loose the sack
8549 					 * info?
8550 					 */
8551 					goto out;
8552 				}
8553 				counter_u64_add(rack_sack_splits, 1);
8554 				rack_clone_rsm(rack, nrsm, rsm, start);
8555 				rsm->r_just_ret = 0;
8556 				insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8557 #ifdef INVARIANTS
8558 				if (insret != NULL) {
8559 					panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8560 					      nrsm, insret, rack, rsm);
8561 				}
8562 #endif
8563 				if (rsm->r_in_tmap) {
8564 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8565 					nrsm->r_in_tmap = 1;
8566 				}
8567 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8568 				rsm->r_flags &= (~RACK_HAS_FIN);
8569 				/* Position us to point to the new nrsm that starts the sack blk */
8570 				rsm = nrsm;
8571 			}
8572 		} else {
8573 			/* Already sacked this piece */
8574 			counter_u64_add(rack_sack_skipped_acked, 1);
8575 			moved++;
8576 			if (end == rsm->r_end) {
8577 				/* Done with block */
8578 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8579 				goto out;
8580 			} else if (SEQ_LT(end, rsm->r_end)) {
8581 				/* A partial sack to a already sacked block */
8582 				moved++;
8583 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8584 				goto out;
8585 			} else {
8586 				/*
8587 				 * The end goes beyond this guy
8588 				 * repostion the start to the
8589 				 * next block.
8590 				 */
8591 				start = rsm->r_end;
8592 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8593 				if (rsm == NULL)
8594 					goto out;
8595 			}
8596 		}
8597 	}
8598 	if (SEQ_GEQ(end, rsm->r_end)) {
8599 		/**
8600 		 * The end of this block is either beyond this guy or right
8601 		 * at this guy. I.e.:
8602 		 *  rsm ---                 |-----|
8603 		 *  end                     |-----|
8604 		 *  <or>
8605 		 *  end                     |---------|
8606 		 */
8607 		if ((rsm->r_flags & RACK_ACKED) == 0) {
8608 			/*
8609 			 * Is it a TLP of interest?
8610 			 */
8611 			if ((rsm->r_flags & RACK_TLP) &&
8612 			    (rsm->r_rtr_cnt > 1)) {
8613 				/*
8614 				 * We are splitting a rxt TLP, check
8615 				 * if we need to save off the start/end
8616 				 */
8617 				if (rack->rc_last_tlp_acked_set &&
8618 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8619 					/*
8620 					 * We already turned this on since we are inside
8621 					 * the previous one was a partially sack now we
8622 					 * are getting another one (maybe all of it).
8623 					 */
8624 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8625 					/*
8626 					 * Lets make sure we have all of it though.
8627 					 */
8628 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8629 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8630 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8631 								     rack->r_ctl.last_tlp_acked_end);
8632 					}
8633 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8634 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8635 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8636 								     rack->r_ctl.last_tlp_acked_end);
8637 					}
8638 				} else {
8639 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8640 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8641 					rack->rc_last_tlp_past_cumack = 0;
8642 					rack->rc_last_tlp_acked_set = 1;
8643 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8644 				}
8645 			}
8646 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8647 			changed += (rsm->r_end - rsm->r_start);
8648 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8649 			if (rsm->r_in_tmap) /* should be true */
8650 				rack_log_sack_passed(tp, rack, rsm);
8651 			/* Is Reordering occuring? */
8652 			if (rsm->r_flags & RACK_SACK_PASSED) {
8653 				rsm->r_flags &= ~RACK_SACK_PASSED;
8654 				counter_u64_add(rack_reorder_seen, 1);
8655 				rack->r_ctl.rc_reorder_ts = cts;
8656 			}
8657 			if (rack->app_limited_needs_set)
8658 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8659 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8660 			rsm->r_flags |= RACK_ACKED;
8661 			if (rsm->r_in_tmap) {
8662 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8663 				rsm->r_in_tmap = 0;
8664 			}
8665 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8666 		} else {
8667 			counter_u64_add(rack_sack_skipped_acked, 1);
8668 			moved++;
8669 		}
8670 		if (end == rsm->r_end) {
8671 			/* This block only - done, setup for next */
8672 			goto out;
8673 		}
8674 		/*
8675 		 * There is more not coverend by this rsm move on
8676 		 * to the next block in the RB tree.
8677 		 */
8678 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8679 		start = rsm->r_end;
8680 		rsm = nrsm;
8681 		if (rsm == NULL)
8682 			goto out;
8683 		goto do_rest_ofb;
8684 	}
8685 	/**
8686 	 * The end of this sack block is smaller than
8687 	 * our rsm i.e.:
8688 	 *  rsm ---                 |-----|
8689 	 *  end                     |--|
8690 	 */
8691 	if ((rsm->r_flags & RACK_ACKED) == 0) {
8692 		/*
8693 		 * Is it a TLP of interest?
8694 		 */
8695 		if ((rsm->r_flags & RACK_TLP) &&
8696 		    (rsm->r_rtr_cnt > 1)) {
8697 			/*
8698 			 * We are splitting a rxt TLP, check
8699 			 * if we need to save off the start/end
8700 			 */
8701 			if (rack->rc_last_tlp_acked_set &&
8702 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8703 				/*
8704 				 * We already turned this on since we are inside
8705 				 * the previous one was a partially sack now we
8706 				 * are getting another one (maybe all of it).
8707 				 */
8708 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8709 				/*
8710 				 * Lets make sure we have all of it though.
8711 				 */
8712 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8713 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8714 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8715 							     rack->r_ctl.last_tlp_acked_end);
8716 				}
8717 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8718 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8719 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8720 							     rack->r_ctl.last_tlp_acked_end);
8721 				}
8722 			} else {
8723 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8724 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8725 				rack->rc_last_tlp_past_cumack = 0;
8726 				rack->rc_last_tlp_acked_set = 1;
8727 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8728 			}
8729 		}
8730 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8731 		if (prev &&
8732 		    (prev->r_flags & RACK_ACKED)) {
8733 			/**
8734 			 * Goal, we want the right remainder of rsm to shrink
8735 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
8736 			 * We want to expand prev to go all the way
8737 			 * to prev->r_end <- end.
8738 			 * so in the tree we have before:
8739 			 *   prev     |--------|         (acked)
8740 			 *   rsm               |-------| (non-acked)
8741 			 *   sackblk           |-|
8742 			 * We churn it so we end up with
8743 			 *   prev     |----------|       (acked)
8744 			 *   rsm                 |-----| (non-acked)
8745 			 *   nrsm              |-| (temporary)
8746 			 *
8747 			 * Note if either prev/rsm is a TLP we don't
8748 			 * do this.
8749 			 */
8750 			nrsm = &stack_map;
8751 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8752 			prev->r_end = end;
8753 			rsm->r_start = end;
8754 			/* Now adjust nrsm (stack copy) to be
8755 			 * the one that is the small
8756 			 * piece that was "sacked".
8757 			 */
8758 			nrsm->r_end = end;
8759 			rsm->r_dupack = 0;
8760 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8761 			/*
8762 			 * Now that the rsm has had its start moved forward
8763 			 * lets go ahead and get its new place in the world.
8764 			 */
8765 			rack_setup_offset_for_rsm(prev, rsm);
8766 			/*
8767 			 * Now nrsm is our new little piece
8768 			 * that is acked (which was merged
8769 			 * to prev). Update the rtt and changed
8770 			 * based on that. Also check for reordering.
8771 			 */
8772 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8773 			if (rack->app_limited_needs_set)
8774 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8775 			changed += (nrsm->r_end - nrsm->r_start);
8776 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8777 			if (nrsm->r_flags & RACK_SACK_PASSED) {
8778 				counter_u64_add(rack_reorder_seen, 1);
8779 				rack->r_ctl.rc_reorder_ts = cts;
8780 			}
8781 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8782 			rsm = prev;
8783 			counter_u64_add(rack_sack_used_prev_merge, 1);
8784 		} else {
8785 			/**
8786 			 * This is the case where our previous
8787 			 * block is not acked either, so we must
8788 			 * split the block in two.
8789 			 */
8790 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8791 			if (nrsm == NULL) {
8792 				/* failed rrs what can we do but loose the sack info? */
8793 				goto out;
8794 			}
8795 			if ((rsm->r_flags & RACK_TLP) &&
8796 			    (rsm->r_rtr_cnt > 1)) {
8797 				/*
8798 				 * We are splitting a rxt TLP, check
8799 				 * if we need to save off the start/end
8800 				 */
8801 				if (rack->rc_last_tlp_acked_set &&
8802 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8803 					    /*
8804 					     * We already turned this on since this block is inside
8805 					     * the previous one was a partially sack now we
8806 					     * are getting another one (maybe all of it).
8807 					     */
8808 					    rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8809 					    /*
8810 					     * Lets make sure we have all of it though.
8811 					     */
8812 					    if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8813 						    rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8814 						    rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8815 									 rack->r_ctl.last_tlp_acked_end);
8816 					    }
8817 					    if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8818 						    rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8819 						    rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8820 									 rack->r_ctl.last_tlp_acked_end);
8821 					    }
8822 				    } else {
8823 					    rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8824 					    rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8825 					    rack->rc_last_tlp_acked_set = 1;
8826 					    rack->rc_last_tlp_past_cumack = 0;
8827 					    rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8828 				    }
8829 			}
8830 			/**
8831 			 * In this case nrsm becomes
8832 			 * nrsm->r_start = end;
8833 			 * nrsm->r_end = rsm->r_end;
8834 			 * which is un-acked.
8835 			 * <and>
8836 			 * rsm->r_end = nrsm->r_start;
8837 			 * i.e. the remaining un-acked
8838 			 * piece is left on the left
8839 			 * hand side.
8840 			 *
8841 			 * So we start like this
8842 			 * rsm      |----------| (not acked)
8843 			 * sackblk  |---|
8844 			 * build it so we have
8845 			 * rsm      |---|         (acked)
8846 			 * nrsm         |------|  (not acked)
8847 			 */
8848 			counter_u64_add(rack_sack_splits, 1);
8849 			rack_clone_rsm(rack, nrsm, rsm, end);
8850 			rsm->r_flags &= (~RACK_HAS_FIN);
8851 			rsm->r_just_ret = 0;
8852 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8853 #ifdef INVARIANTS
8854 			if (insret != NULL) {
8855 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8856 				      nrsm, insret, rack, rsm);
8857 			}
8858 #endif
8859 			if (rsm->r_in_tmap) {
8860 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8861 				nrsm->r_in_tmap = 1;
8862 			}
8863 			nrsm->r_dupack = 0;
8864 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8865 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8866 			changed += (rsm->r_end - rsm->r_start);
8867 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8868 			if (rsm->r_in_tmap) /* should be true */
8869 				rack_log_sack_passed(tp, rack, rsm);
8870 			/* Is Reordering occuring? */
8871 			if (rsm->r_flags & RACK_SACK_PASSED) {
8872 				rsm->r_flags &= ~RACK_SACK_PASSED;
8873 				counter_u64_add(rack_reorder_seen, 1);
8874 				rack->r_ctl.rc_reorder_ts = cts;
8875 			}
8876 			if (rack->app_limited_needs_set)
8877 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8878 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8879 			rsm->r_flags |= RACK_ACKED;
8880 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8881 			if (rsm->r_in_tmap) {
8882 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8883 				rsm->r_in_tmap = 0;
8884 			}
8885 		}
8886 	} else if (start != end){
8887 		/*
8888 		 * The block was already acked.
8889 		 */
8890 		counter_u64_add(rack_sack_skipped_acked, 1);
8891 		moved++;
8892 	}
8893 out:
8894 	if (rsm &&
8895 	    ((rsm->r_flags & RACK_TLP) == 0) &&
8896 	    (rsm->r_flags & RACK_ACKED)) {
8897 		/*
8898 		 * Now can we merge where we worked
8899 		 * with either the previous or
8900 		 * next block?
8901 		 */
8902 		next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8903 		while (next) {
8904 			if (next->r_flags & RACK_TLP)
8905 				break;
8906 			if (next->r_flags & RACK_ACKED) {
8907 			/* yep this and next can be merged */
8908 				rsm = rack_merge_rsm(rack, rsm, next);
8909 				next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8910 			} else
8911 				break;
8912 		}
8913 		/* Now what about the previous? */
8914 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8915 		while (prev) {
8916 			if (prev->r_flags & RACK_TLP)
8917 				break;
8918 			if (prev->r_flags & RACK_ACKED) {
8919 				/* yep the previous and this can be merged */
8920 				rsm = rack_merge_rsm(rack, prev, rsm);
8921 				prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8922 			} else
8923 				break;
8924 		}
8925 	}
8926 	if (used_ref == 0) {
8927 		counter_u64_add(rack_sack_proc_all, 1);
8928 	} else {
8929 		counter_u64_add(rack_sack_proc_short, 1);
8930 	}
8931 	/* Save off the next one for quick reference. */
8932 	if (rsm)
8933 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8934 	else
8935 		nrsm = NULL;
8936 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
8937 	/* Pass back the moved. */
8938 	*moved_two = moved;
8939 	return (changed);
8940 }
8941 
8942 static void inline
8943 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8944 {
8945 	struct rack_sendmap *tmap;
8946 
8947 	tmap = NULL;
8948 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
8949 		/* Its no longer sacked, mark it so */
8950 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8951 #ifdef INVARIANTS
8952 		if (rsm->r_in_tmap) {
8953 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
8954 			      rack, rsm, rsm->r_flags);
8955 		}
8956 #endif
8957 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8958 		/* Rebuild it into our tmap */
8959 		if (tmap == NULL) {
8960 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8961 			tmap = rsm;
8962 		} else {
8963 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8964 			tmap = rsm;
8965 		}
8966 		tmap->r_in_tmap = 1;
8967 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8968 	}
8969 	/*
8970 	 * Now lets possibly clear the sack filter so we start
8971 	 * recognizing sacks that cover this area.
8972 	 */
8973 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8974 
8975 }
8976 
8977 static void
8978 rack_do_decay(struct tcp_rack *rack)
8979 {
8980 	struct timeval res;
8981 
8982 #define	timersub(tvp, uvp, vvp)						\
8983 	do {								\
8984 		(vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;		\
8985 		(vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;	\
8986 		if ((vvp)->tv_usec < 0) {				\
8987 			(vvp)->tv_sec--;				\
8988 			(vvp)->tv_usec += 1000000;			\
8989 		}							\
8990 	} while (0)
8991 
8992 	timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8993 #undef timersub
8994 
8995 	rack->r_ctl.input_pkt++;
8996 	if ((rack->rc_in_persist) ||
8997 	    (res.tv_sec >= 1) ||
8998 	    (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8999 		/*
9000 		 * Check for decay of non-SAD,
9001 		 * we want all SAD detection metrics to
9002 		 * decay 1/4 per second (or more) passed.
9003 		 */
9004 		uint32_t pkt_delta;
9005 
9006 		pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
9007 		/* Update our saved tracking values */
9008 		rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
9009 		rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
9010 		/* Now do we escape without decay? */
9011 #ifdef NETFLIX_EXP_DETECTION
9012 		if (rack->rc_in_persist ||
9013 		    (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
9014 		    (pkt_delta < tcp_sad_low_pps)){
9015 			/*
9016 			 * We don't decay idle connections
9017 			 * or ones that have a low input pps.
9018 			 */
9019 			return;
9020 		}
9021 		/* Decay the counters */
9022 		rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
9023 							tcp_sad_decay_val);
9024 		rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
9025 							 tcp_sad_decay_val);
9026 		rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
9027 							       tcp_sad_decay_val);
9028 		rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
9029 								tcp_sad_decay_val);
9030 #endif
9031 	}
9032 }
9033 
9034 static void
9035 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
9036 {
9037 	struct rack_sendmap *rsm, *rm;
9038 
9039 	/*
9040 	 * The ACK point is advancing to th_ack, we must drop off
9041 	 * the packets in the rack log and calculate any eligble
9042 	 * RTT's.
9043 	 */
9044 	rack->r_wanted_output = 1;
9045 
9046 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
9047 	if ((rack->rc_last_tlp_acked_set == 1)&&
9048 	    (rack->rc_last_tlp_past_cumack == 1) &&
9049 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
9050 		/*
9051 		 * We have reached the point where our last rack
9052 		 * tlp retransmit sequence is ahead of the cum-ack.
9053 		 * This can only happen when the cum-ack moves all
9054 		 * the way around (its been a full 2^^31+1 bytes
9055 		 * or more since we sent a retransmitted TLP). Lets
9056 		 * turn off the valid flag since its not really valid.
9057 		 *
9058 		 * Note since sack's also turn on this event we have
9059 		 * a complication, we have to wait to age it out until
9060 		 * the cum-ack is by the TLP before checking which is
9061 		 * what the next else clause does.
9062 		 */
9063 		rack_log_dsack_event(rack, 9, __LINE__,
9064 				     rack->r_ctl.last_tlp_acked_start,
9065 				     rack->r_ctl.last_tlp_acked_end);
9066 		rack->rc_last_tlp_acked_set = 0;
9067 		rack->rc_last_tlp_past_cumack = 0;
9068 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
9069 		   (rack->rc_last_tlp_past_cumack == 0) &&
9070 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
9071 		/*
9072 		 * It is safe to start aging TLP's out.
9073 		 */
9074 		rack->rc_last_tlp_past_cumack = 1;
9075 	}
9076 	/* We do the same for the tlp send seq as well */
9077 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
9078 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
9079 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
9080 		rack_log_dsack_event(rack, 9, __LINE__,
9081 				     rack->r_ctl.last_sent_tlp_seq,
9082 				     (rack->r_ctl.last_sent_tlp_seq +
9083 				      rack->r_ctl.last_sent_tlp_len));
9084 		rack->rc_last_sent_tlp_seq_valid = 0;
9085 		rack->rc_last_sent_tlp_past_cumack = 0;
9086 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
9087 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
9088 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
9089 		/*
9090 		 * It is safe to start aging TLP's send.
9091 		 */
9092 		rack->rc_last_sent_tlp_past_cumack = 1;
9093 	}
9094 more:
9095 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9096 	if (rsm == NULL) {
9097 		if ((th_ack - 1) == tp->iss) {
9098 			/*
9099 			 * For the SYN incoming case we will not
9100 			 * have called tcp_output for the sending of
9101 			 * the SYN, so there will be no map. All
9102 			 * other cases should probably be a panic.
9103 			 */
9104 			return;
9105 		}
9106 		if (tp->t_flags & TF_SENTFIN) {
9107 			/* if we sent a FIN we often will not have map */
9108 			return;
9109 		}
9110 #ifdef INVARIANTS
9111 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
9112 		      tp,
9113 		      tp->t_state, th_ack, rack,
9114 		      tp->snd_una, tp->snd_max, tp->snd_nxt);
9115 #endif
9116 		return;
9117 	}
9118 	if (SEQ_LT(th_ack, rsm->r_start)) {
9119 		/* Huh map is missing this */
9120 #ifdef INVARIANTS
9121 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
9122 		       rsm->r_start,
9123 		       th_ack, tp->t_state, rack->r_state);
9124 #endif
9125 		return;
9126 	}
9127 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
9128 
9129 	/* Now was it a retransmitted TLP? */
9130 	if ((rsm->r_flags & RACK_TLP) &&
9131 	    (rsm->r_rtr_cnt > 1)) {
9132 		/*
9133 		 * Yes, this rsm was a TLP and retransmitted, remember that
9134 		 * since if a DSACK comes back on this we don't want
9135 		 * to think of it as a reordered segment. This may
9136 		 * get updated again with possibly even other TLPs
9137 		 * in flight, but thats ok. Only when we don't send
9138 		 * a retransmitted TLP for 1/2 the sequences space
9139 		 * will it get turned off (above).
9140 		 */
9141 		if (rack->rc_last_tlp_acked_set &&
9142 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9143 			/*
9144 			 * We already turned this on since the end matches,
9145 			 * the previous one was a partially ack now we
9146 			 * are getting another one (maybe all of it).
9147 			 */
9148 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9149 			/*
9150 			 * Lets make sure we have all of it though.
9151 			 */
9152 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9153 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9154 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9155 						     rack->r_ctl.last_tlp_acked_end);
9156 			}
9157 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9158 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9159 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9160 						     rack->r_ctl.last_tlp_acked_end);
9161 			}
9162 		} else {
9163 			rack->rc_last_tlp_past_cumack = 1;
9164 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9165 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9166 			rack->rc_last_tlp_acked_set = 1;
9167 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9168 		}
9169 	}
9170 	/* Now do we consume the whole thing? */
9171 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
9172 		/* Its all consumed. */
9173 		uint32_t left;
9174 		uint8_t newly_acked;
9175 
9176 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
9177 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
9178 		rsm->r_rtr_bytes = 0;
9179 		/* Record the time of highest cumack sent */
9180 		rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9181 		rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
9182 #ifdef INVARIANTS
9183 		if (rm != rsm) {
9184 			panic("removing head in rack:%p rsm:%p rm:%p",
9185 			      rack, rsm, rm);
9186 		}
9187 #endif
9188 		if (rsm->r_in_tmap) {
9189 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9190 			rsm->r_in_tmap = 0;
9191 		}
9192 		newly_acked = 1;
9193 		if (rsm->r_flags & RACK_ACKED) {
9194 			/*
9195 			 * It was acked on the scoreboard -- remove
9196 			 * it from total
9197 			 */
9198 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
9199 			newly_acked = 0;
9200 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
9201 			/*
9202 			 * There are segments ACKED on the
9203 			 * scoreboard further up. We are seeing
9204 			 * reordering.
9205 			 */
9206 			rsm->r_flags &= ~RACK_SACK_PASSED;
9207 			counter_u64_add(rack_reorder_seen, 1);
9208 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9209 			rsm->r_flags |= RACK_ACKED;
9210 			rack->r_ctl.rc_reorder_ts = cts;
9211 			if (rack->r_ent_rec_ns) {
9212 				/*
9213 				 * We have sent no more, and we saw an sack
9214 				 * then ack arrive.
9215 				 */
9216 				rack->r_might_revert = 1;
9217 			}
9218 		}
9219 		if ((rsm->r_flags & RACK_TO_REXT) &&
9220 		    (tp->t_flags & TF_RCVD_TSTMP) &&
9221 		    (to->to_flags & TOF_TS) &&
9222 		    (to->to_tsecr != 0) &&
9223 		    (tp->t_flags & TF_PREVVALID)) {
9224 			/*
9225 			 * We can use the timestamp to see
9226 			 * if this retransmission was from the
9227 			 * first transmit. If so we made a mistake.
9228 			 */
9229 			tp->t_flags &= ~TF_PREVVALID;
9230 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
9231 				/* The first transmit is what this ack is for */
9232 				rack_cong_signal(tp, CC_RTO_ERR, th_ack);
9233 			}
9234 		}
9235 		left = th_ack - rsm->r_end;
9236 		if (rack->app_limited_needs_set && newly_acked)
9237 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
9238 		/* Free back to zone */
9239 		rack_free(rack, rsm);
9240 		if (left) {
9241 			goto more;
9242 		}
9243 		/* Check for reneging */
9244 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9245 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
9246 			/*
9247 			 * The peer has moved snd_una up to
9248 			 * the edge of this send, i.e. one
9249 			 * that it had previously acked. The only
9250 			 * way that can be true if the peer threw
9251 			 * away data (space issues) that it had
9252 			 * previously sacked (else it would have
9253 			 * given us snd_una up to (rsm->r_end).
9254 			 * We need to undo the acked markings here.
9255 			 *
9256 			 * Note we have to look to make sure th_ack is
9257 			 * our rsm->r_start in case we get an old ack
9258 			 * where th_ack is behind snd_una.
9259 			 */
9260 			rack_peer_reneges(rack, rsm, th_ack);
9261 		}
9262 		return;
9263 	}
9264 	if (rsm->r_flags & RACK_ACKED) {
9265 		/*
9266 		 * It was acked on the scoreboard -- remove it from
9267 		 * total for the part being cum-acked.
9268 		 */
9269 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
9270 	}
9271 	/*
9272 	 * Clear the dup ack count for
9273 	 * the piece that remains.
9274 	 */
9275 	rsm->r_dupack = 0;
9276 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9277 	if (rsm->r_rtr_bytes) {
9278 		/*
9279 		 * It was retransmitted adjust the
9280 		 * sack holes for what was acked.
9281 		 */
9282 		int ack_am;
9283 
9284 		ack_am = (th_ack - rsm->r_start);
9285 		if (ack_am >= rsm->r_rtr_bytes) {
9286 			rack->r_ctl.rc_holes_rxt -= ack_am;
9287 			rsm->r_rtr_bytes -= ack_am;
9288 		}
9289 	}
9290 	/*
9291 	 * Update where the piece starts and record
9292 	 * the time of send of highest cumack sent.
9293 	 */
9294 	rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9295 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
9296 	/* Now we need to move our offset forward too */
9297 	if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
9298 		/* Fix up the orig_m_len and possibly the mbuf offset */
9299 		rack_adjust_orig_mlen(rsm);
9300 	}
9301 	rsm->soff += (th_ack - rsm->r_start);
9302 	rsm->r_start = th_ack;
9303 	/* Now do we need to move the mbuf fwd too? */
9304 	if (rsm->m) {
9305 		while (rsm->soff >= rsm->m->m_len) {
9306 			rsm->soff -= rsm->m->m_len;
9307 			rsm->m = rsm->m->m_next;
9308 			KASSERT((rsm->m != NULL),
9309 				(" nrsm:%p hit at soff:%u null m",
9310 				 rsm, rsm->soff));
9311 		}
9312 		rsm->orig_m_len = rsm->m->m_len;
9313 	}
9314 	if (rack->app_limited_needs_set)
9315 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
9316 }
9317 
9318 static void
9319 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
9320 {
9321 	struct rack_sendmap *rsm;
9322 	int sack_pass_fnd = 0;
9323 
9324 	if (rack->r_might_revert) {
9325 		/*
9326 		 * Ok we have reordering, have not sent anything, we
9327 		 * might want to revert the congestion state if nothing
9328 		 * further has SACK_PASSED on it. Lets check.
9329 		 *
9330 		 * We also get here when we have DSACKs come in for
9331 		 * all the data that we FR'd. Note that a rxt or tlp
9332 		 * timer clears this from happening.
9333 		 */
9334 
9335 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
9336 			if (rsm->r_flags & RACK_SACK_PASSED) {
9337 				sack_pass_fnd = 1;
9338 				break;
9339 			}
9340 		}
9341 		if (sack_pass_fnd == 0) {
9342 			/*
9343 			 * We went into recovery
9344 			 * incorrectly due to reordering!
9345 			 */
9346 			int orig_cwnd;
9347 
9348 			rack->r_ent_rec_ns = 0;
9349 			orig_cwnd = tp->snd_cwnd;
9350 			tp->snd_cwnd = rack->r_ctl.rc_cwnd_at_erec;
9351 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
9352 			tp->snd_recover = tp->snd_una;
9353 			rack_log_to_prr(rack, 14, orig_cwnd);
9354 			EXIT_RECOVERY(tp->t_flags);
9355 		}
9356 		rack->r_might_revert = 0;
9357 	}
9358 }
9359 
9360 #ifdef NETFLIX_EXP_DETECTION
9361 static void
9362 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
9363 {
9364 	if ((rack->do_detection || tcp_force_detection) &&
9365 	    tcp_sack_to_ack_thresh &&
9366 	    tcp_sack_to_move_thresh &&
9367 	    ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
9368 		/*
9369 		 * We have thresholds set to find
9370 		 * possible attackers and disable sack.
9371 		 * Check them.
9372 		 */
9373 		uint64_t ackratio, moveratio, movetotal;
9374 
9375 		/* Log detecting */
9376 		rack_log_sad(rack, 1);
9377 		ackratio = (uint64_t)(rack->r_ctl.sack_count);
9378 		ackratio *= (uint64_t)(1000);
9379 		if (rack->r_ctl.ack_count)
9380 			ackratio /= (uint64_t)(rack->r_ctl.ack_count);
9381 		else {
9382 			/* We really should not hit here */
9383 			ackratio = 1000;
9384 		}
9385 		if ((rack->sack_attack_disable == 0) &&
9386 		    (ackratio > rack_highest_sack_thresh_seen))
9387 			rack_highest_sack_thresh_seen = (uint32_t)ackratio;
9388 		movetotal = rack->r_ctl.sack_moved_extra;
9389 		movetotal += rack->r_ctl.sack_noextra_move;
9390 		moveratio = rack->r_ctl.sack_moved_extra;
9391 		moveratio *= (uint64_t)1000;
9392 		if (movetotal)
9393 			moveratio /= movetotal;
9394 		else {
9395 			/* No moves, thats pretty good */
9396 			moveratio = 0;
9397 		}
9398 		if ((rack->sack_attack_disable == 0) &&
9399 		    (moveratio > rack_highest_move_thresh_seen))
9400 			rack_highest_move_thresh_seen = (uint32_t)moveratio;
9401 		if (rack->sack_attack_disable == 0) {
9402 			if ((ackratio > tcp_sack_to_ack_thresh) &&
9403 			    (moveratio > tcp_sack_to_move_thresh)) {
9404 				/* Disable sack processing */
9405 				rack->sack_attack_disable = 1;
9406 				if (rack->r_rep_attack == 0) {
9407 					rack->r_rep_attack = 1;
9408 					counter_u64_add(rack_sack_attacks_detected, 1);
9409 				}
9410 				if (tcp_attack_on_turns_on_logging) {
9411 					/*
9412 					 * Turn on logging, used for debugging
9413 					 * false positives.
9414 					 */
9415 					rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
9416 				}
9417 				/* Clamp the cwnd at flight size */
9418 				rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
9419 				rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9420 				rack_log_sad(rack, 2);
9421 			}
9422 		} else {
9423 			/* We are sack-disabled check for false positives */
9424 			if ((ackratio <= tcp_restoral_thresh) ||
9425 			    (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
9426 				rack->sack_attack_disable = 0;
9427 				rack_log_sad(rack, 3);
9428 				/* Restart counting */
9429 				rack->r_ctl.sack_count = 0;
9430 				rack->r_ctl.sack_moved_extra = 0;
9431 				rack->r_ctl.sack_noextra_move = 1;
9432 				rack->r_ctl.ack_count = max(1,
9433 				      (bytes_this_ack / segsiz));
9434 
9435 				if (rack->r_rep_reverse == 0) {
9436 					rack->r_rep_reverse = 1;
9437 					counter_u64_add(rack_sack_attacks_reversed, 1);
9438 				}
9439 				/* Restore the cwnd */
9440 				if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
9441 					rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
9442 			}
9443 		}
9444 	}
9445 }
9446 #endif
9447 
9448 static int
9449 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
9450 {
9451 
9452 	uint32_t am, l_end;
9453 	int was_tlp = 0;
9454 
9455 	if (SEQ_GT(end, start))
9456 		am = end - start;
9457 	else
9458 		am = 0;
9459 	if ((rack->rc_last_tlp_acked_set ) &&
9460 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
9461 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
9462 		/*
9463 		 * The DSACK is because of a TLP which we don't
9464 		 * do anything with the reordering window over since
9465 		 * it was not reordering that caused the DSACK but
9466 		 * our previous retransmit TLP.
9467 		 */
9468 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
9469 		was_tlp = 1;
9470 		goto skip_dsack_round;
9471 	}
9472 	if (rack->rc_last_sent_tlp_seq_valid) {
9473 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
9474 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
9475 		    (SEQ_LEQ(end, l_end))) {
9476 			/*
9477 			 * This dsack is from the last sent TLP, ignore it
9478 			 * for reordering purposes.
9479 			 */
9480 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
9481 			was_tlp = 1;
9482 			goto skip_dsack_round;
9483 		}
9484 	}
9485 	if (rack->rc_dsack_round_seen == 0) {
9486 		rack->rc_dsack_round_seen = 1;
9487 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
9488 		rack->r_ctl.num_dsack++;
9489 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
9490 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
9491 	}
9492 skip_dsack_round:
9493 	/*
9494 	 * We keep track of how many DSACK blocks we get
9495 	 * after a recovery incident.
9496 	 */
9497 	rack->r_ctl.dsack_byte_cnt += am;
9498 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
9499 	    rack->r_ctl.retran_during_recovery &&
9500 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
9501 		/*
9502 		 * False recovery most likely culprit is reordering. If
9503 		 * nothing else is missing we need to revert.
9504 		 */
9505 		rack->r_might_revert = 1;
9506 		rack_handle_might_revert(rack->rc_tp, rack);
9507 		rack->r_might_revert = 0;
9508 		rack->r_ctl.retran_during_recovery = 0;
9509 		rack->r_ctl.dsack_byte_cnt = 0;
9510 	}
9511 	return (was_tlp);
9512 }
9513 
9514 static void
9515 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9516 {
9517 	/* Deal with changed and PRR here (in recovery only) */
9518 	uint32_t pipe, snd_una;
9519 
9520 	rack->r_ctl.rc_prr_delivered += changed;
9521 
9522 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9523 		/*
9524 		 * It is all outstanding, we are application limited
9525 		 * and thus we don't need more room to send anything.
9526 		 * Note we use tp->snd_una here and not th_ack because
9527 		 * the data as yet not been cut from the sb.
9528 		 */
9529 		rack->r_ctl.rc_prr_sndcnt = 0;
9530 		return;
9531 	}
9532 	/* Compute prr_sndcnt */
9533 	if (SEQ_GT(tp->snd_una, th_ack)) {
9534 		snd_una = tp->snd_una;
9535 	} else {
9536 		snd_una = th_ack;
9537 	}
9538 	pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt;
9539 	if (pipe > tp->snd_ssthresh) {
9540 		long sndcnt;
9541 
9542 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9543 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
9544 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9545 		else {
9546 			rack->r_ctl.rc_prr_sndcnt = 0;
9547 			rack_log_to_prr(rack, 9, 0);
9548 			sndcnt = 0;
9549 		}
9550 		sndcnt++;
9551 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9552 			sndcnt -= rack->r_ctl.rc_prr_out;
9553 		else
9554 			sndcnt = 0;
9555 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
9556 		rack_log_to_prr(rack, 10, 0);
9557 	} else {
9558 		uint32_t limit;
9559 
9560 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9561 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9562 		else
9563 			limit = 0;
9564 		if (changed > limit)
9565 			limit = changed;
9566 		limit += ctf_fixed_maxseg(tp);
9567 		if (tp->snd_ssthresh > pipe) {
9568 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9569 			rack_log_to_prr(rack, 11, 0);
9570 		} else {
9571 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9572 			rack_log_to_prr(rack, 12, 0);
9573 		}
9574 	}
9575 }
9576 
9577 static void
9578 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9579 {
9580 	uint32_t changed;
9581 	struct tcp_rack *rack;
9582 	struct rack_sendmap *rsm;
9583 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9584 	register uint32_t th_ack;
9585 	int32_t i, j, k, num_sack_blks = 0;
9586 	uint32_t cts, acked, ack_point, sack_changed = 0;
9587 	int loop_start = 0, moved_two = 0;
9588 	uint32_t tsused;
9589 
9590 
9591 	INP_WLOCK_ASSERT(tp->t_inpcb);
9592 	if (th->th_flags & TH_RST) {
9593 		/* We don't log resets */
9594 		return;
9595 	}
9596 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9597 	cts = tcp_get_usecs(NULL);
9598 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9599 	changed = 0;
9600 	th_ack = th->th_ack;
9601 	if (rack->sack_attack_disable == 0)
9602 		rack_do_decay(rack);
9603 	if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9604 		/*
9605 		 * You only get credit for
9606 		 * MSS and greater (and you get extra
9607 		 * credit for larger cum-ack moves).
9608 		 */
9609 		int ac;
9610 
9611 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9612 		rack->r_ctl.ack_count += ac;
9613 		counter_u64_add(rack_ack_total, ac);
9614 	}
9615 	if (rack->r_ctl.ack_count > 0xfff00000) {
9616 		/*
9617 		 * reduce the number to keep us under
9618 		 * a uint32_t.
9619 		 */
9620 		rack->r_ctl.ack_count /= 2;
9621 		rack->r_ctl.sack_count /= 2;
9622 	}
9623 	if (SEQ_GT(th_ack, tp->snd_una)) {
9624 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9625 		tp->t_acktime = ticks;
9626 	}
9627 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
9628 		changed = th_ack - rsm->r_start;
9629 	if (changed) {
9630 		rack_process_to_cumack(tp, rack, th_ack, cts, to);
9631 	}
9632 	if ((to->to_flags & TOF_SACK) == 0) {
9633 		/* We are done nothing left and no sack. */
9634 		rack_handle_might_revert(tp, rack);
9635 		/*
9636 		 * For cases where we struck a dup-ack
9637 		 * with no SACK, add to the changes so
9638 		 * PRR will work right.
9639 		 */
9640 		if (dup_ack_struck && (changed == 0)) {
9641 			changed += ctf_fixed_maxseg(rack->rc_tp);
9642 		}
9643 		goto out;
9644 	}
9645 	/* Sack block processing */
9646 	if (SEQ_GT(th_ack, tp->snd_una))
9647 		ack_point = th_ack;
9648 	else
9649 		ack_point = tp->snd_una;
9650 	for (i = 0; i < to->to_nsacks; i++) {
9651 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
9652 		      &sack, sizeof(sack));
9653 		sack.start = ntohl(sack.start);
9654 		sack.end = ntohl(sack.end);
9655 		if (SEQ_GT(sack.end, sack.start) &&
9656 		    SEQ_GT(sack.start, ack_point) &&
9657 		    SEQ_LT(sack.start, tp->snd_max) &&
9658 		    SEQ_GT(sack.end, ack_point) &&
9659 		    SEQ_LEQ(sack.end, tp->snd_max)) {
9660 			sack_blocks[num_sack_blks] = sack;
9661 			num_sack_blks++;
9662 		} else if (SEQ_LEQ(sack.start, th_ack) &&
9663 			   SEQ_LEQ(sack.end, th_ack)) {
9664 			int was_tlp;
9665 
9666 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
9667 			/*
9668 			 * Its a D-SACK block.
9669 			 */
9670 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
9671 		}
9672 	}
9673 	if (rack->rc_dsack_round_seen) {
9674 		/* Is the dsack roound over? */
9675 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
9676 			/* Yes it is */
9677 			rack->rc_dsack_round_seen = 0;
9678 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
9679 		}
9680 	}
9681 	/*
9682 	 * Sort the SACK blocks so we can update the rack scoreboard with
9683 	 * just one pass.
9684 	 */
9685 	num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9686 					 num_sack_blks, th->th_ack);
9687 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9688 	if (num_sack_blks == 0) {
9689 		/* Nothing to sack (DSACKs?) */
9690 		goto out_with_totals;
9691 	}
9692 	if (num_sack_blks < 2) {
9693 		/* Only one, we don't need to sort */
9694 		goto do_sack_work;
9695 	}
9696 	/* Sort the sacks */
9697 	for (i = 0; i < num_sack_blks; i++) {
9698 		for (j = i + 1; j < num_sack_blks; j++) {
9699 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9700 				sack = sack_blocks[i];
9701 				sack_blocks[i] = sack_blocks[j];
9702 				sack_blocks[j] = sack;
9703 			}
9704 		}
9705 	}
9706 	/*
9707 	 * Now are any of the sack block ends the same (yes some
9708 	 * implementations send these)?
9709 	 */
9710 again:
9711 	if (num_sack_blks == 0)
9712 		goto out_with_totals;
9713 	if (num_sack_blks > 1) {
9714 		for (i = 0; i < num_sack_blks; i++) {
9715 			for (j = i + 1; j < num_sack_blks; j++) {
9716 				if (sack_blocks[i].end == sack_blocks[j].end) {
9717 					/*
9718 					 * Ok these two have the same end we
9719 					 * want the smallest end and then
9720 					 * throw away the larger and start
9721 					 * again.
9722 					 */
9723 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9724 						/*
9725 						 * The second block covers
9726 						 * more area use that
9727 						 */
9728 						sack_blocks[i].start = sack_blocks[j].start;
9729 					}
9730 					/*
9731 					 * Now collapse out the dup-sack and
9732 					 * lower the count
9733 					 */
9734 					for (k = (j + 1); k < num_sack_blks; k++) {
9735 						sack_blocks[j].start = sack_blocks[k].start;
9736 						sack_blocks[j].end = sack_blocks[k].end;
9737 						j++;
9738 					}
9739 					num_sack_blks--;
9740 					goto again;
9741 				}
9742 			}
9743 		}
9744 	}
9745 do_sack_work:
9746 	/*
9747 	 * First lets look to see if
9748 	 * we have retransmitted and
9749 	 * can use the transmit next?
9750 	 */
9751 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9752 	if (rsm &&
9753 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9754 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9755 		/*
9756 		 * We probably did the FR and the next
9757 		 * SACK in continues as we would expect.
9758 		 */
9759 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9760 		if (acked) {
9761 			rack->r_wanted_output = 1;
9762 			changed += acked;
9763 			sack_changed += acked;
9764 		}
9765 		if (num_sack_blks == 1) {
9766 			/*
9767 			 * This is what we would expect from
9768 			 * a normal implementation to happen
9769 			 * after we have retransmitted the FR,
9770 			 * i.e the sack-filter pushes down
9771 			 * to 1 block and the next to be retransmitted
9772 			 * is the sequence in the sack block (has more
9773 			 * are acked). Count this as ACK'd data to boost
9774 			 * up the chances of recovering any false positives.
9775 			 */
9776 			rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9777 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9778 			counter_u64_add(rack_express_sack, 1);
9779 			if (rack->r_ctl.ack_count > 0xfff00000) {
9780 				/*
9781 				 * reduce the number to keep us under
9782 				 * a uint32_t.
9783 				 */
9784 				rack->r_ctl.ack_count /= 2;
9785 				rack->r_ctl.sack_count /= 2;
9786 			}
9787 			goto out_with_totals;
9788 		} else {
9789 			/*
9790 			 * Start the loop through the
9791 			 * rest of blocks, past the first block.
9792 			 */
9793 			moved_two = 0;
9794 			loop_start = 1;
9795 		}
9796 	}
9797 	/* Its a sack of some sort */
9798 	rack->r_ctl.sack_count++;
9799 	if (rack->r_ctl.sack_count > 0xfff00000) {
9800 		/*
9801 		 * reduce the number to keep us under
9802 		 * a uint32_t.
9803 		 */
9804 		rack->r_ctl.ack_count /= 2;
9805 		rack->r_ctl.sack_count /= 2;
9806 	}
9807 	counter_u64_add(rack_sack_total, 1);
9808 	if (rack->sack_attack_disable) {
9809 		/* An attacker disablement is in place */
9810 		if (num_sack_blks > 1) {
9811 			rack->r_ctl.sack_count += (num_sack_blks - 1);
9812 			rack->r_ctl.sack_moved_extra++;
9813 			counter_u64_add(rack_move_some, 1);
9814 			if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9815 				rack->r_ctl.sack_moved_extra /= 2;
9816 				rack->r_ctl.sack_noextra_move /= 2;
9817 			}
9818 		}
9819 		goto out;
9820 	}
9821 	rsm = rack->r_ctl.rc_sacklast;
9822 	for (i = loop_start; i < num_sack_blks; i++) {
9823 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9824 		if (acked) {
9825 			rack->r_wanted_output = 1;
9826 			changed += acked;
9827 			sack_changed += acked;
9828 		}
9829 		if (moved_two) {
9830 			/*
9831 			 * If we did not get a SACK for at least a MSS and
9832 			 * had to move at all, or if we moved more than our
9833 			 * threshold, it counts against the "extra" move.
9834 			 */
9835 			rack->r_ctl.sack_moved_extra += moved_two;
9836 			counter_u64_add(rack_move_some, 1);
9837 		} else {
9838 			/*
9839 			 * else we did not have to move
9840 			 * any more than we would expect.
9841 			 */
9842 			rack->r_ctl.sack_noextra_move++;
9843 			counter_u64_add(rack_move_none, 1);
9844 		}
9845 		if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9846 			/*
9847 			 * If the SACK was not a full MSS then
9848 			 * we add to sack_count the number of
9849 			 * MSS's (or possibly more than
9850 			 * a MSS if its a TSO send) we had to skip by.
9851 			 */
9852 			rack->r_ctl.sack_count += moved_two;
9853 			counter_u64_add(rack_sack_total, moved_two);
9854 		}
9855 		/*
9856 		 * Now we need to setup for the next
9857 		 * round. First we make sure we won't
9858 		 * exceed the size of our uint32_t on
9859 		 * the various counts, and then clear out
9860 		 * moved_two.
9861 		 */
9862 		if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9863 		    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9864 			rack->r_ctl.sack_moved_extra /= 2;
9865 			rack->r_ctl.sack_noextra_move /= 2;
9866 		}
9867 		if (rack->r_ctl.sack_count > 0xfff00000) {
9868 			rack->r_ctl.ack_count /= 2;
9869 			rack->r_ctl.sack_count /= 2;
9870 		}
9871 		moved_two = 0;
9872 	}
9873 out_with_totals:
9874 	if (num_sack_blks > 1) {
9875 		/*
9876 		 * You get an extra stroke if
9877 		 * you have more than one sack-blk, this
9878 		 * could be where we are skipping forward
9879 		 * and the sack-filter is still working, or
9880 		 * it could be an attacker constantly
9881 		 * moving us.
9882 		 */
9883 		rack->r_ctl.sack_moved_extra++;
9884 		counter_u64_add(rack_move_some, 1);
9885 	}
9886 out:
9887 #ifdef NETFLIX_EXP_DETECTION
9888 	rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9889 #endif
9890 	if (changed) {
9891 		/* Something changed cancel the rack timer */
9892 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9893 	}
9894 	tsused = tcp_get_usecs(NULL);
9895 	rsm = tcp_rack_output(tp, rack, tsused);
9896 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9897 	    rsm) {
9898 		/* Enter recovery */
9899 		rack->r_ctl.rc_rsm_start = rsm->r_start;
9900 		rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
9901 		rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
9902 		entered_recovery = 1;
9903 		rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
9904 		/*
9905 		 * When we enter recovery we need to assure we send
9906 		 * one packet.
9907 		 */
9908 		if (rack->rack_no_prr == 0) {
9909 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9910 			rack_log_to_prr(rack, 8, 0);
9911 		}
9912 		rack->r_timer_override = 1;
9913 		rack->r_early = 0;
9914 		rack->r_ctl.rc_agg_early = 0;
9915 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
9916 		   rsm &&
9917 		   (rack->r_rr_config == 3)) {
9918 		/*
9919 		 * Assure we can output and we get no
9920 		 * remembered pace time except the retransmit.
9921 		 */
9922 		rack->r_timer_override = 1;
9923 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9924 		rack->r_ctl.rc_resend = rsm;
9925 	}
9926 	if (IN_FASTRECOVERY(tp->t_flags) &&
9927 	    (rack->rack_no_prr == 0) &&
9928 	    (entered_recovery == 0)) {
9929 		rack_update_prr(tp, rack, changed, th_ack);
9930 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9931 		     ((rack->rc_inp->inp_in_hpts == 0) &&
9932 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9933 			/*
9934 			 * If you are pacing output you don't want
9935 			 * to override.
9936 			 */
9937 			rack->r_early = 0;
9938 			rack->r_ctl.rc_agg_early = 0;
9939 			rack->r_timer_override = 1;
9940 		}
9941 	}
9942 }
9943 
9944 static void
9945 rack_strike_dupack(struct tcp_rack *rack)
9946 {
9947 	struct rack_sendmap *rsm;
9948 
9949 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9950 	while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9951 		rsm = TAILQ_NEXT(rsm, r_tnext);
9952 	}
9953 	if (rsm && (rsm->r_dupack < 0xff)) {
9954 		rsm->r_dupack++;
9955 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9956 			struct timeval tv;
9957 			uint32_t cts;
9958 			/*
9959 			 * Here we see if we need to retransmit. For
9960 			 * a SACK type connection if enough time has passed
9961 			 * we will get a return of the rsm. For a non-sack
9962 			 * connection we will get the rsm returned if the
9963 			 * dupack value is 3 or more.
9964 			 */
9965 			cts = tcp_get_usecs(&tv);
9966 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9967 			if (rack->r_ctl.rc_resend != NULL) {
9968 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9969 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9970 							 rack->rc_tp->snd_una);
9971 				}
9972 				rack->r_wanted_output = 1;
9973 				rack->r_timer_override = 1;
9974 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9975 			}
9976 		} else {
9977 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9978 		}
9979 	}
9980 }
9981 
9982 static void
9983 rack_check_bottom_drag(struct tcpcb *tp,
9984 		       struct tcp_rack *rack,
9985 		       struct socket *so, int32_t acked)
9986 {
9987 	uint32_t segsiz, minseg;
9988 
9989 	segsiz = ctf_fixed_maxseg(tp);
9990 	minseg = segsiz;
9991 
9992 	if (tp->snd_max == tp->snd_una) {
9993 		/*
9994 		 * We are doing dynamic pacing and we are way
9995 		 * under. Basically everything got acked while
9996 		 * we were still waiting on the pacer to expire.
9997 		 *
9998 		 * This means we need to boost the b/w in
9999 		 * addition to any earlier boosting of
10000 		 * the multipler.
10001 		 */
10002 		rack->rc_dragged_bottom = 1;
10003 		rack_validate_multipliers_at_or_above100(rack);
10004 		/*
10005 		 * Lets use the segment bytes acked plus
10006 		 * the lowest RTT seen as the basis to
10007 		 * form a b/w estimate. This will be off
10008 		 * due to the fact that the true estimate
10009 		 * should be around 1/2 the time of the RTT
10010 		 * but we can settle for that.
10011 		 */
10012 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
10013 		    acked) {
10014 			uint64_t bw, calc_bw, rtt;
10015 
10016 			rtt = rack->r_ctl.rack_rs.rs_us_rtt;
10017 			if (rtt == 0) {
10018 				/* no us sample is there a ms one? */
10019 				if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
10020 					rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
10021 				} else {
10022 					goto no_measurement;
10023 				}
10024 			}
10025 			bw = acked;
10026 			calc_bw = bw * 1000000;
10027 			calc_bw /= rtt;
10028 			if (rack->r_ctl.last_max_bw &&
10029 			    (rack->r_ctl.last_max_bw < calc_bw)) {
10030 				/*
10031 				 * If we have a last calculated max bw
10032 				 * enforce it.
10033 				 */
10034 				calc_bw = rack->r_ctl.last_max_bw;
10035 			}
10036 			/* now plop it in */
10037 			if (rack->rc_gp_filled == 0) {
10038 				if (calc_bw > ONE_POINT_TWO_MEG) {
10039 					/*
10040 					 * If we have no measurement
10041 					 * don't let us set in more than
10042 					 * 1.2Mbps. If we are still too
10043 					 * low after pacing with this we
10044 					 * will hopefully have a max b/w
10045 					 * available to sanity check things.
10046 					 */
10047 					calc_bw = ONE_POINT_TWO_MEG;
10048 				}
10049 				rack->r_ctl.rc_rtt_diff = 0;
10050 				rack->r_ctl.gp_bw = calc_bw;
10051 				rack->rc_gp_filled = 1;
10052 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
10053 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
10054 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
10055 			} else if (calc_bw > rack->r_ctl.gp_bw) {
10056 				rack->r_ctl.rc_rtt_diff = 0;
10057 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
10058 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
10059 				rack->r_ctl.gp_bw = calc_bw;
10060 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
10061 			} else
10062 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
10063 			if ((rack->gp_ready == 0) &&
10064 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
10065 				/* We have enough measurements now */
10066 				rack->gp_ready = 1;
10067 				rack_set_cc_pacing(rack);
10068 				if (rack->defer_options)
10069 					rack_apply_deferred_options(rack);
10070 			}
10071 			/*
10072 			 * For acks over 1mss we do a extra boost to simulate
10073 			 * where we would get 2 acks (we want 110 for the mul).
10074 			 */
10075 			if (acked > segsiz)
10076 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
10077 		} else {
10078 			/*
10079 			 * zero rtt possibly?, settle for just an old increase.
10080 			 */
10081 no_measurement:
10082 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
10083 		}
10084 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
10085 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
10086 					       minseg)) &&
10087 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
10088 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
10089 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
10090 		    (segsiz * rack_req_segs))) {
10091 		/*
10092 		 * We are doing dynamic GP pacing and
10093 		 * we have everything except 1MSS or less
10094 		 * bytes left out. We are still pacing away.
10095 		 * And there is data that could be sent, This
10096 		 * means we are inserting delayed ack time in
10097 		 * our measurements because we are pacing too slow.
10098 		 */
10099 		rack_validate_multipliers_at_or_above100(rack);
10100 		rack->rc_dragged_bottom = 1;
10101 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
10102 	}
10103 }
10104 
10105 
10106 
10107 static void
10108 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
10109 {
10110 	/*
10111 	 * The fast output path is enabled and we
10112 	 * have moved the cumack forward. Lets see if
10113 	 * we can expand forward the fast path length by
10114 	 * that amount. What we would ideally like to
10115 	 * do is increase the number of bytes in the
10116 	 * fast path block (left_to_send) by the
10117 	 * acked amount. However we have to gate that
10118 	 * by two factors:
10119 	 * 1) The amount outstanding and the rwnd of the peer
10120 	 *    (i.e. we don't want to exceed the rwnd of the peer).
10121 	 *    <and>
10122 	 * 2) The amount of data left in the socket buffer (i.e.
10123 	 *    we can't send beyond what is in the buffer).
10124 	 *
10125 	 * Note that this does not take into account any increase
10126 	 * in the cwnd. We will only extend the fast path by
10127 	 * what was acked.
10128 	 */
10129 	uint32_t new_total, gating_val;
10130 
10131 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
10132 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
10133 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
10134 	if (new_total <= gating_val) {
10135 		/* We can increase left_to_send by the acked amount */
10136 		counter_u64_add(rack_extended_rfo, 1);
10137 		rack->r_ctl.fsb.left_to_send = new_total;
10138 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
10139 			("rack:%p left_to_send:%u sbavail:%u out:%u",
10140 			 rack, rack->r_ctl.fsb.left_to_send,
10141 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
10142 			 (tp->snd_max - tp->snd_una)));
10143 
10144 	}
10145 }
10146 
10147 static void
10148 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
10149 {
10150 	/*
10151 	 * Here any sendmap entry that points to the
10152 	 * beginning mbuf must be adjusted to the correct
10153 	 * offset. This must be called with:
10154 	 * 1) The socket buffer locked
10155 	 * 2) snd_una adjusted to its new postion.
10156 	 *
10157 	 * Note that (2) implies rack_ack_received has also
10158 	 * been called.
10159 	 *
10160 	 * We grab the first mbuf in the socket buffer and
10161 	 * then go through the front of the sendmap, recalculating
10162 	 * the stored offset for any sendmap entry that has
10163 	 * that mbuf. We must use the sb functions to do this
10164 	 * since its possible an add was done has well as
10165 	 * the subtraction we may have just completed. This should
10166 	 * not be a penalty though, since we just referenced the sb
10167 	 * to go in and trim off the mbufs that we freed (of course
10168 	 * there will be a penalty for the sendmap references though).
10169 	 */
10170 	struct mbuf *m;
10171 	struct rack_sendmap *rsm;
10172 
10173 	SOCKBUF_LOCK_ASSERT(sb);
10174 	m = sb->sb_mb;
10175 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
10176 	if ((rsm == NULL) || (m == NULL)) {
10177 		/* Nothing outstanding */
10178 		return;
10179 	}
10180 	while (rsm->m && (rsm->m == m)) {
10181 		/* one to adjust */
10182 #ifdef INVARIANTS
10183 		struct mbuf *tm;
10184 		uint32_t soff;
10185 
10186 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
10187 		if (rsm->orig_m_len != m->m_len) {
10188 			rack_adjust_orig_mlen(rsm);
10189 		}
10190 		if (rsm->soff != soff) {
10191 			/*
10192 			 * This is not a fatal error, we anticipate it
10193 			 * might happen (the else code), so we count it here
10194 			 * so that under invariant we can see that it really
10195 			 * does happen.
10196 			 */
10197 			counter_u64_add(rack_adjust_map_bw, 1);
10198 		}
10199 		rsm->m = tm;
10200 		rsm->soff = soff;
10201 		if (tm)
10202 			rsm->orig_m_len = rsm->m->m_len;
10203 		else
10204 			rsm->orig_m_len = 0;
10205 #else
10206 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
10207 		if (rsm->m)
10208 			rsm->orig_m_len = rsm->m->m_len;
10209 		else
10210 			rsm->orig_m_len = 0;
10211 #endif
10212 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
10213 			      rsm);
10214 		if (rsm == NULL)
10215 			break;
10216 	}
10217 }
10218 
10219 /*
10220  * Return value of 1, we do not need to call rack_process_data().
10221  * return value of 0, rack_process_data can be called.
10222  * For ret_val if its 0 the TCP is locked, if its non-zero
10223  * its unlocked and probably unsafe to touch the TCB.
10224  */
10225 static int
10226 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10227     struct tcpcb *tp, struct tcpopt *to,
10228     uint32_t tiwin, int32_t tlen,
10229     int32_t * ofia, int32_t thflags, int32_t *ret_val)
10230 {
10231 	int32_t ourfinisacked = 0;
10232 	int32_t nsegs, acked_amount;
10233 	int32_t acked;
10234 	struct mbuf *mfree;
10235 	struct tcp_rack *rack;
10236 	int32_t under_pacing = 0;
10237 	int32_t recovery = 0;
10238 
10239 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10240 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
10241 		__ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
10242 				      &rack->r_ctl.challenge_ack_ts,
10243 				      &rack->r_ctl.challenge_ack_cnt);
10244 		rack->r_wanted_output = 1;
10245 		return (1);
10246 	}
10247 	if (rack->gp_ready &&
10248 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10249 		under_pacing = 1;
10250 	}
10251 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
10252 		int in_rec, dup_ack_struck = 0;
10253 
10254 		in_rec = IN_FASTRECOVERY(tp->t_flags);
10255 		if (rack->rc_in_persist) {
10256 			tp->t_rxtshift = 0;
10257 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10258 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10259 		}
10260 		if ((th->th_ack == tp->snd_una) &&
10261 		    (tiwin == tp->snd_wnd) &&
10262 		    ((to->to_flags & TOF_SACK) == 0)) {
10263 			rack_strike_dupack(rack);
10264 			dup_ack_struck = 1;
10265 		}
10266 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
10267 	}
10268 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10269 		/*
10270 		 * Old ack, behind (or duplicate to) the last one rcv'd
10271 		 * Note: We mark reordering is occuring if its
10272 		 * less than and we have not closed our window.
10273 		 */
10274 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
10275 			counter_u64_add(rack_reorder_seen, 1);
10276 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
10277 		}
10278 		return (0);
10279 	}
10280 	/*
10281 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
10282 	 * something we sent.
10283 	 */
10284 	if (tp->t_flags & TF_NEEDSYN) {
10285 		/*
10286 		 * T/TCP: Connection was half-synchronized, and our SYN has
10287 		 * been ACK'd (so connection is now fully synchronized).  Go
10288 		 * to non-starred state, increment snd_una for ACK of SYN,
10289 		 * and check if we can do window scaling.
10290 		 */
10291 		tp->t_flags &= ~TF_NEEDSYN;
10292 		tp->snd_una++;
10293 		/* Do window scaling? */
10294 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10295 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10296 			tp->rcv_scale = tp->request_r_scale;
10297 			/* Send window already scaled. */
10298 		}
10299 	}
10300 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10301 	INP_WLOCK_ASSERT(tp->t_inpcb);
10302 
10303 	acked = BYTES_THIS_ACK(tp, th);
10304 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10305 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10306 	/*
10307 	 * If we just performed our first retransmit, and the ACK arrives
10308 	 * within our recovery window, then it was a mistake to do the
10309 	 * retransmit in the first place.  Recover our original cwnd and
10310 	 * ssthresh, and proceed to transmit where we left off.
10311 	 */
10312 	if ((tp->t_flags & TF_PREVVALID) &&
10313 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10314 		tp->t_flags &= ~TF_PREVVALID;
10315 		if (tp->t_rxtshift == 1 &&
10316 		    (int)(ticks - tp->t_badrxtwin) < 0)
10317 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
10318 	}
10319 	if (acked) {
10320 		/* assure we are not backed off */
10321 		tp->t_rxtshift = 0;
10322 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10323 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10324 		rack->rc_tlp_in_progress = 0;
10325 		rack->r_ctl.rc_tlp_cnt_out = 0;
10326 		/*
10327 		 * If it is the RXT timer we want to
10328 		 * stop it, so we can restart a TLP.
10329 		 */
10330 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10331 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10332 #ifdef NETFLIX_HTTP_LOGGING
10333 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10334 #endif
10335 	}
10336 	/*
10337 	 * If we have a timestamp reply, update smoothed round trip time. If
10338 	 * no timestamp is present but transmit timer is running and timed
10339 	 * sequence number was acked, update smoothed round trip time. Since
10340 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
10341 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
10342 	 * timer.
10343 	 *
10344 	 * Some boxes send broken timestamp replies during the SYN+ACK
10345 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
10346 	 * and blow up the retransmit timer.
10347 	 */
10348 	/*
10349 	 * If all outstanding data is acked, stop retransmit timer and
10350 	 * remember to restart (more output or persist). If there is more
10351 	 * data to be acked, restart retransmit timer, using current
10352 	 * (possibly backed-off) value.
10353 	 */
10354 	if (acked == 0) {
10355 		if (ofia)
10356 			*ofia = ourfinisacked;
10357 		return (0);
10358 	}
10359 	if (IN_RECOVERY(tp->t_flags)) {
10360 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
10361 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
10362 			tcp_rack_partialack(tp);
10363 		} else {
10364 			rack_post_recovery(tp, th->th_ack);
10365 			recovery = 1;
10366 		}
10367 	}
10368 	/*
10369 	 * Let the congestion control algorithm update congestion control
10370 	 * related information. This typically means increasing the
10371 	 * congestion window.
10372 	 */
10373 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
10374 	SOCKBUF_LOCK(&so->so_snd);
10375 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
10376 	tp->snd_wnd -= acked_amount;
10377 	mfree = sbcut_locked(&so->so_snd, acked_amount);
10378 	if ((sbused(&so->so_snd) == 0) &&
10379 	    (acked > acked_amount) &&
10380 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
10381 	    (tp->t_flags & TF_SENTFIN)) {
10382 		/*
10383 		 * We must be sure our fin
10384 		 * was sent and acked (we can be
10385 		 * in FIN_WAIT_1 without having
10386 		 * sent the fin).
10387 		 */
10388 		ourfinisacked = 1;
10389 	}
10390 	tp->snd_una = th->th_ack;
10391 	if (acked_amount && sbavail(&so->so_snd))
10392 		rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10393 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10394 	/* NB: sowwakeup_locked() does an implicit unlock. */
10395 	sowwakeup_locked(so);
10396 	m_freem(mfree);
10397 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
10398 		tp->snd_recover = tp->snd_una;
10399 
10400 	if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
10401 		tp->snd_nxt = tp->snd_una;
10402 	}
10403 	if (under_pacing &&
10404 	    (rack->use_fixed_rate == 0) &&
10405 	    (rack->in_probe_rtt == 0) &&
10406 	    rack->rc_gp_dyn_mul &&
10407 	    rack->rc_always_pace) {
10408 		/* Check if we are dragging bottom */
10409 		rack_check_bottom_drag(tp, rack, so, acked);
10410 	}
10411 	if (tp->snd_una == tp->snd_max) {
10412 		/* Nothing left outstanding */
10413 		tp->t_flags &= ~TF_PREVVALID;
10414 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10415 		rack->r_ctl.retran_during_recovery = 0;
10416 		rack->r_ctl.dsack_byte_cnt = 0;
10417 		if (rack->r_ctl.rc_went_idle_time == 0)
10418 			rack->r_ctl.rc_went_idle_time = 1;
10419 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10420 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
10421 			tp->t_acktime = 0;
10422 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10423 		/* Set need output so persist might get set */
10424 		rack->r_wanted_output = 1;
10425 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10426 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
10427 		    (sbavail(&so->so_snd) == 0) &&
10428 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
10429 			/*
10430 			 * The socket was gone and the
10431 			 * peer sent data (now or in the past), time to
10432 			 * reset him.
10433 			 */
10434 			*ret_val = 1;
10435 			/* tcp_close will kill the inp pre-log the Reset */
10436 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
10437 			tp = tcp_close(tp);
10438 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
10439 			return (1);
10440 		}
10441 	}
10442 	if (ofia)
10443 		*ofia = ourfinisacked;
10444 	return (0);
10445 }
10446 
10447 static void
10448 rack_collapsed_window(struct tcp_rack *rack)
10449 {
10450 	/*
10451 	 * Now we must walk the
10452 	 * send map and divide the
10453 	 * ones left stranded. These
10454 	 * guys can't cause us to abort
10455 	 * the connection and are really
10456 	 * "unsent". However if a buggy
10457 	 * client actually did keep some
10458 	 * of the data i.e. collapsed the win
10459 	 * and refused to ack and then opened
10460 	 * the win and acked that data. We would
10461 	 * get into an ack war, the simplier
10462 	 * method then of just pretending we
10463 	 * did not send those segments something
10464 	 * won't work.
10465 	 */
10466 	struct rack_sendmap *rsm, *nrsm, fe, *insret;
10467 	tcp_seq max_seq;
10468 
10469 	max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
10470 	memset(&fe, 0, sizeof(fe));
10471 	fe.r_start = max_seq;
10472 	/* Find the first seq past or at maxseq */
10473 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
10474 	if (rsm == NULL) {
10475 		/* Nothing to do strange */
10476 		rack->rc_has_collapsed = 0;
10477 		return;
10478 	}
10479 	/*
10480 	 * Now do we need to split at
10481 	 * the collapse point?
10482 	 */
10483 	if (SEQ_GT(max_seq, rsm->r_start)) {
10484 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10485 		if (nrsm == NULL) {
10486 			/* We can't get a rsm, mark all? */
10487 			nrsm = rsm;
10488 			goto no_split;
10489 		}
10490 		/* Clone it */
10491 		rack_clone_rsm(rack, nrsm, rsm, max_seq);
10492 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10493 #ifdef INVARIANTS
10494 		if (insret != NULL) {
10495 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
10496 			      nrsm, insret, rack, rsm);
10497 		}
10498 #endif
10499 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT, max_seq, __LINE__);
10500 		if (rsm->r_in_tmap) {
10501 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10502 			nrsm->r_in_tmap = 1;
10503 		}
10504 		/*
10505 		 * Set in the new RSM as the
10506 		 * collapsed starting point
10507 		 */
10508 		rsm = nrsm;
10509 	}
10510 no_split:
10511 	counter_u64_add(rack_collapsed_win, 1);
10512 	RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
10513 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
10514 	}
10515 	rack->rc_has_collapsed = 1;
10516 }
10517 
10518 static void
10519 rack_un_collapse_window(struct tcp_rack *rack)
10520 {
10521 	struct rack_sendmap *rsm;
10522 
10523 	RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
10524 		if (rsm->r_flags & RACK_RWND_COLLAPSED)
10525 			rsm->r_flags &= ~RACK_RWND_COLLAPSED;
10526 		else
10527 			break;
10528 	}
10529 	rack->rc_has_collapsed = 0;
10530 }
10531 
10532 static void
10533 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10534 			int32_t tlen, int32_t tfo_syn)
10535 {
10536 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
10537 		if (rack->rc_dack_mode &&
10538 		    (tlen > 500) &&
10539 		    (rack->rc_dack_toggle == 1)) {
10540 			goto no_delayed_ack;
10541 		}
10542 		rack_timer_cancel(tp, rack,
10543 				  rack->r_ctl.rc_rcvtime, __LINE__);
10544 		tp->t_flags |= TF_DELACK;
10545 	} else {
10546 no_delayed_ack:
10547 		rack->r_wanted_output = 1;
10548 		tp->t_flags |= TF_ACKNOW;
10549 		if (rack->rc_dack_mode) {
10550 			if (tp->t_flags & TF_DELACK)
10551 				rack->rc_dack_toggle = 1;
10552 			else
10553 				rack->rc_dack_toggle = 0;
10554 		}
10555 	}
10556 }
10557 
10558 static void
10559 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10560 {
10561 	/*
10562 	 * If fast output is in progress, lets validate that
10563 	 * the new window did not shrink on us and make it
10564 	 * so fast output should end.
10565 	 */
10566 	if (rack->r_fast_output) {
10567 		uint32_t out;
10568 
10569 		/*
10570 		 * Calculate what we will send if left as is
10571 		 * and compare that to our send window.
10572 		 */
10573 		out = ctf_outstanding(tp);
10574 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10575 			/* ok we have an issue */
10576 			if (out >= tp->snd_wnd) {
10577 				/* Turn off fast output the window is met or collapsed */
10578 				rack->r_fast_output = 0;
10579 			} else {
10580 				/* we have some room left */
10581 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10582 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10583 					/* If not at least 1 full segment never mind */
10584 					rack->r_fast_output = 0;
10585 				}
10586 			}
10587 		}
10588 	}
10589 }
10590 
10591 
10592 /*
10593  * Return value of 1, the TCB is unlocked and most
10594  * likely gone, return value of 0, the TCP is still
10595  * locked.
10596  */
10597 static int
10598 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10599     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10600     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10601 {
10602 	/*
10603 	 * Update window information. Don't look at window if no ACK: TAC's
10604 	 * send garbage on first SYN.
10605 	 */
10606 	int32_t nsegs;
10607 	int32_t tfo_syn;
10608 	struct tcp_rack *rack;
10609 
10610 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10611 	INP_WLOCK_ASSERT(tp->t_inpcb);
10612 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10613 	if ((thflags & TH_ACK) &&
10614 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10615 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10616 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10617 		/* keep track of pure window updates */
10618 		if (tlen == 0 &&
10619 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10620 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10621 		tp->snd_wnd = tiwin;
10622 		rack_validate_fo_sendwin_up(tp, rack);
10623 		tp->snd_wl1 = th->th_seq;
10624 		tp->snd_wl2 = th->th_ack;
10625 		if (tp->snd_wnd > tp->max_sndwnd)
10626 			tp->max_sndwnd = tp->snd_wnd;
10627 		rack->r_wanted_output = 1;
10628 	} else if (thflags & TH_ACK) {
10629 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10630 			tp->snd_wnd = tiwin;
10631 			rack_validate_fo_sendwin_up(tp, rack);
10632 			tp->snd_wl1 = th->th_seq;
10633 			tp->snd_wl2 = th->th_ack;
10634 		}
10635 	}
10636 	if (tp->snd_wnd < ctf_outstanding(tp))
10637 		/* The peer collapsed the window */
10638 		rack_collapsed_window(rack);
10639 	else if (rack->rc_has_collapsed)
10640 		rack_un_collapse_window(rack);
10641 	/* Was persist timer active and now we have window space? */
10642 	if ((rack->rc_in_persist != 0) &&
10643 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10644 				rack->r_ctl.rc_pace_min_segs))) {
10645 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10646 		tp->snd_nxt = tp->snd_max;
10647 		/* Make sure we output to start the timer */
10648 		rack->r_wanted_output = 1;
10649 	}
10650 	/* Do we enter persists? */
10651 	if ((rack->rc_in_persist == 0) &&
10652 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10653 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
10654 	    (tp->snd_max == tp->snd_una) &&
10655 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
10656 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
10657 		/*
10658 		 * Here the rwnd is less than
10659 		 * the pacing size, we are established,
10660 		 * nothing is outstanding, and there is
10661 		 * data to send. Enter persists.
10662 		 */
10663 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10664 	}
10665 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10666 		m_freem(m);
10667 		return (0);
10668 	}
10669 	/*
10670 	 * don't process the URG bit, ignore them drag
10671 	 * along the up.
10672 	 */
10673 	tp->rcv_up = tp->rcv_nxt;
10674 	INP_WLOCK_ASSERT(tp->t_inpcb);
10675 
10676 	/*
10677 	 * Process the segment text, merging it into the TCP sequencing
10678 	 * queue, and arranging for acknowledgment of receipt if necessary.
10679 	 * This process logically involves adjusting tp->rcv_wnd as data is
10680 	 * presented to the user (this happens in tcp_usrreq.c, case
10681 	 * PRU_RCVD).  If a FIN has already been received on this connection
10682 	 * then we just ignore the text.
10683 	 */
10684 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10685 		   IS_FASTOPEN(tp->t_flags));
10686 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10687 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10688 		tcp_seq save_start = th->th_seq;
10689 		tcp_seq save_rnxt  = tp->rcv_nxt;
10690 		int     save_tlen  = tlen;
10691 
10692 		m_adj(m, drop_hdrlen);	/* delayed header drop */
10693 		/*
10694 		 * Insert segment which includes th into TCP reassembly
10695 		 * queue with control block tp.  Set thflags to whether
10696 		 * reassembly now includes a segment with FIN.  This handles
10697 		 * the common case inline (segment is the next to be
10698 		 * received on an established connection, and the queue is
10699 		 * empty), avoiding linkage into and removal from the queue
10700 		 * and repetition of various conversions. Set DELACK for
10701 		 * segments received in order, but ack immediately when
10702 		 * segments are out of order (so fast retransmit can work).
10703 		 */
10704 		if (th->th_seq == tp->rcv_nxt &&
10705 		    SEGQ_EMPTY(tp) &&
10706 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
10707 		    tfo_syn)) {
10708 #ifdef NETFLIX_SB_LIMITS
10709 			u_int mcnt, appended;
10710 
10711 			if (so->so_rcv.sb_shlim) {
10712 				mcnt = m_memcnt(m);
10713 				appended = 0;
10714 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10715 				    CFO_NOSLEEP, NULL) == false) {
10716 					counter_u64_add(tcp_sb_shlim_fails, 1);
10717 					m_freem(m);
10718 					return (0);
10719 				}
10720 			}
10721 #endif
10722 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10723 			tp->rcv_nxt += tlen;
10724 			if (tlen &&
10725 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10726 			    (tp->t_fbyte_in == 0)) {
10727 				tp->t_fbyte_in = ticks;
10728 				if (tp->t_fbyte_in == 0)
10729 					tp->t_fbyte_in = 1;
10730 				if (tp->t_fbyte_out && tp->t_fbyte_in)
10731 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10732 			}
10733 			thflags = th->th_flags & TH_FIN;
10734 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10735 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10736 			SOCKBUF_LOCK(&so->so_rcv);
10737 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10738 				m_freem(m);
10739 			} else
10740 #ifdef NETFLIX_SB_LIMITS
10741 				appended =
10742 #endif
10743 					sbappendstream_locked(&so->so_rcv, m, 0);
10744 
10745 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10746 			/* NB: sorwakeup_locked() does an implicit unlock. */
10747 			sorwakeup_locked(so);
10748 #ifdef NETFLIX_SB_LIMITS
10749 			if (so->so_rcv.sb_shlim && appended != mcnt)
10750 				counter_fo_release(so->so_rcv.sb_shlim,
10751 				    mcnt - appended);
10752 #endif
10753 		} else {
10754 			/*
10755 			 * XXX: Due to the header drop above "th" is
10756 			 * theoretically invalid by now.  Fortunately
10757 			 * m_adj() doesn't actually frees any mbufs when
10758 			 * trimming from the head.
10759 			 */
10760 			tcp_seq temp = save_start;
10761 
10762 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
10763 			tp->t_flags |= TF_ACKNOW;
10764 			if (tp->t_flags & TF_WAKESOR) {
10765 				tp->t_flags &= ~TF_WAKESOR;
10766 				/* NB: sorwakeup_locked() does an implicit unlock. */
10767 				sorwakeup_locked(so);
10768 			}
10769 		}
10770 		if ((tp->t_flags & TF_SACK_PERMIT) &&
10771 		    (save_tlen > 0) &&
10772 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
10773 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10774 				/*
10775 				 * DSACK actually handled in the fastpath
10776 				 * above.
10777 				 */
10778 				RACK_OPTS_INC(tcp_sack_path_1);
10779 				tcp_update_sack_list(tp, save_start,
10780 				    save_start + save_tlen);
10781 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10782 				if ((tp->rcv_numsacks >= 1) &&
10783 				    (tp->sackblks[0].end == save_start)) {
10784 					/*
10785 					 * Partial overlap, recorded at todrop
10786 					 * above.
10787 					 */
10788 					RACK_OPTS_INC(tcp_sack_path_2a);
10789 					tcp_update_sack_list(tp,
10790 					    tp->sackblks[0].start,
10791 					    tp->sackblks[0].end);
10792 				} else {
10793 					RACK_OPTS_INC(tcp_sack_path_2b);
10794 					tcp_update_dsack_list(tp, save_start,
10795 					    save_start + save_tlen);
10796 				}
10797 			} else if (tlen >= save_tlen) {
10798 				/* Update of sackblks. */
10799 				RACK_OPTS_INC(tcp_sack_path_3);
10800 				tcp_update_dsack_list(tp, save_start,
10801 				    save_start + save_tlen);
10802 			} else if (tlen > 0) {
10803 				RACK_OPTS_INC(tcp_sack_path_4);
10804 				tcp_update_dsack_list(tp, save_start,
10805 				    save_start + tlen);
10806 			}
10807 		}
10808 	} else {
10809 		m_freem(m);
10810 		thflags &= ~TH_FIN;
10811 	}
10812 
10813 	/*
10814 	 * If FIN is received ACK the FIN and let the user know that the
10815 	 * connection is closing.
10816 	 */
10817 	if (thflags & TH_FIN) {
10818 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10819 			/* The socket upcall is handled by socantrcvmore. */
10820 			socantrcvmore(so);
10821 			/*
10822 			 * If connection is half-synchronized (ie NEEDSYN
10823 			 * flag on) then delay ACK, so it may be piggybacked
10824 			 * when SYN is sent. Otherwise, since we received a
10825 			 * FIN then no more input can be expected, send ACK
10826 			 * now.
10827 			 */
10828 			if (tp->t_flags & TF_NEEDSYN) {
10829 				rack_timer_cancel(tp, rack,
10830 				    rack->r_ctl.rc_rcvtime, __LINE__);
10831 				tp->t_flags |= TF_DELACK;
10832 			} else {
10833 				tp->t_flags |= TF_ACKNOW;
10834 			}
10835 			tp->rcv_nxt++;
10836 		}
10837 		switch (tp->t_state) {
10838 			/*
10839 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
10840 			 * CLOSE_WAIT state.
10841 			 */
10842 		case TCPS_SYN_RECEIVED:
10843 			tp->t_starttime = ticks;
10844 			/* FALLTHROUGH */
10845 		case TCPS_ESTABLISHED:
10846 			rack_timer_cancel(tp, rack,
10847 			    rack->r_ctl.rc_rcvtime, __LINE__);
10848 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
10849 			break;
10850 
10851 			/*
10852 			 * If still in FIN_WAIT_1 STATE FIN has not been
10853 			 * acked so enter the CLOSING state.
10854 			 */
10855 		case TCPS_FIN_WAIT_1:
10856 			rack_timer_cancel(tp, rack,
10857 			    rack->r_ctl.rc_rcvtime, __LINE__);
10858 			tcp_state_change(tp, TCPS_CLOSING);
10859 			break;
10860 
10861 			/*
10862 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
10863 			 * starting the time-wait timer, turning off the
10864 			 * other standard timers.
10865 			 */
10866 		case TCPS_FIN_WAIT_2:
10867 			rack_timer_cancel(tp, rack,
10868 			    rack->r_ctl.rc_rcvtime, __LINE__);
10869 			tcp_twstart(tp);
10870 			return (1);
10871 		}
10872 	}
10873 	/*
10874 	 * Return any desired output.
10875 	 */
10876 	if ((tp->t_flags & TF_ACKNOW) ||
10877 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10878 		rack->r_wanted_output = 1;
10879 	}
10880 	INP_WLOCK_ASSERT(tp->t_inpcb);
10881 	return (0);
10882 }
10883 
10884 /*
10885  * Here nothing is really faster, its just that we
10886  * have broken out the fast-data path also just like
10887  * the fast-ack.
10888  */
10889 static int
10890 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10891     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10892     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10893 {
10894 	int32_t nsegs;
10895 	int32_t newsize = 0;	/* automatic sockbuf scaling */
10896 	struct tcp_rack *rack;
10897 #ifdef NETFLIX_SB_LIMITS
10898 	u_int mcnt, appended;
10899 #endif
10900 #ifdef TCPDEBUG
10901 	/*
10902 	 * The size of tcp_saveipgen must be the size of the max ip header,
10903 	 * now IPv6.
10904 	 */
10905 	u_char tcp_saveipgen[IP6_HDR_LEN];
10906 	struct tcphdr tcp_savetcp;
10907 	short ostate = 0;
10908 
10909 #endif
10910 	/*
10911 	 * If last ACK falls within this segment's sequence numbers, record
10912 	 * the timestamp. NOTE that the test is modified according to the
10913 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10914 	 */
10915 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10916 		return (0);
10917 	}
10918 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10919 		return (0);
10920 	}
10921 	if (tiwin && tiwin != tp->snd_wnd) {
10922 		return (0);
10923 	}
10924 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10925 		return (0);
10926 	}
10927 	if (__predict_false((to->to_flags & TOF_TS) &&
10928 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10929 		return (0);
10930 	}
10931 	if (__predict_false((th->th_ack != tp->snd_una))) {
10932 		return (0);
10933 	}
10934 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10935 		return (0);
10936 	}
10937 	if ((to->to_flags & TOF_TS) != 0 &&
10938 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10939 		tp->ts_recent_age = tcp_ts_getticks();
10940 		tp->ts_recent = to->to_tsval;
10941 	}
10942 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10943 	/*
10944 	 * This is a pure, in-sequence data packet with nothing on the
10945 	 * reassembly queue and we have enough buffer space to take it.
10946 	 */
10947 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10948 
10949 #ifdef NETFLIX_SB_LIMITS
10950 	if (so->so_rcv.sb_shlim) {
10951 		mcnt = m_memcnt(m);
10952 		appended = 0;
10953 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10954 		    CFO_NOSLEEP, NULL) == false) {
10955 			counter_u64_add(tcp_sb_shlim_fails, 1);
10956 			m_freem(m);
10957 			return (1);
10958 		}
10959 	}
10960 #endif
10961 	/* Clean receiver SACK report if present */
10962 	if (tp->rcv_numsacks)
10963 		tcp_clean_sackreport(tp);
10964 	KMOD_TCPSTAT_INC(tcps_preddat);
10965 	tp->rcv_nxt += tlen;
10966 	if (tlen &&
10967 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10968 	    (tp->t_fbyte_in == 0)) {
10969 		tp->t_fbyte_in = ticks;
10970 		if (tp->t_fbyte_in == 0)
10971 			tp->t_fbyte_in = 1;
10972 		if (tp->t_fbyte_out && tp->t_fbyte_in)
10973 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10974 	}
10975 	/*
10976 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10977 	 */
10978 	tp->snd_wl1 = th->th_seq;
10979 	/*
10980 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10981 	 */
10982 	tp->rcv_up = tp->rcv_nxt;
10983 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10984 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10985 #ifdef TCPDEBUG
10986 	if (so->so_options & SO_DEBUG)
10987 		tcp_trace(TA_INPUT, ostate, tp,
10988 		    (void *)tcp_saveipgen, &tcp_savetcp, 0);
10989 #endif
10990 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10991 
10992 	/* Add data to socket buffer. */
10993 	SOCKBUF_LOCK(&so->so_rcv);
10994 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10995 		m_freem(m);
10996 	} else {
10997 		/*
10998 		 * Set new socket buffer size. Give up when limit is
10999 		 * reached.
11000 		 */
11001 		if (newsize)
11002 			if (!sbreserve_locked(&so->so_rcv,
11003 			    newsize, so, NULL))
11004 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
11005 		m_adj(m, drop_hdrlen);	/* delayed header drop */
11006 #ifdef NETFLIX_SB_LIMITS
11007 		appended =
11008 #endif
11009 			sbappendstream_locked(&so->so_rcv, m, 0);
11010 		ctf_calc_rwin(so, tp);
11011 	}
11012 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
11013 	/* NB: sorwakeup_locked() does an implicit unlock. */
11014 	sorwakeup_locked(so);
11015 #ifdef NETFLIX_SB_LIMITS
11016 	if (so->so_rcv.sb_shlim && mcnt != appended)
11017 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
11018 #endif
11019 	rack_handle_delayed_ack(tp, rack, tlen, 0);
11020 	if (tp->snd_una == tp->snd_max)
11021 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
11022 	return (1);
11023 }
11024 
11025 /*
11026  * This subfunction is used to try to highly optimize the
11027  * fast path. We again allow window updates that are
11028  * in sequence to remain in the fast-path. We also add
11029  * in the __predict's to attempt to help the compiler.
11030  * Note that if we return a 0, then we can *not* process
11031  * it and the caller should push the packet into the
11032  * slow-path.
11033  */
11034 static int
11035 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
11036     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11037     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
11038 {
11039 	int32_t acked;
11040 	int32_t nsegs;
11041 #ifdef TCPDEBUG
11042 	/*
11043 	 * The size of tcp_saveipgen must be the size of the max ip header,
11044 	 * now IPv6.
11045 	 */
11046 	u_char tcp_saveipgen[IP6_HDR_LEN];
11047 	struct tcphdr tcp_savetcp;
11048 	short ostate = 0;
11049 #endif
11050 	int32_t under_pacing = 0;
11051 	struct tcp_rack *rack;
11052 
11053 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
11054 		/* Old ack, behind (or duplicate to) the last one rcv'd */
11055 		return (0);
11056 	}
11057 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
11058 		/* Above what we have sent? */
11059 		return (0);
11060 	}
11061 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
11062 		/* We are retransmitting */
11063 		return (0);
11064 	}
11065 	if (__predict_false(tiwin == 0)) {
11066 		/* zero window */
11067 		return (0);
11068 	}
11069 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
11070 		/* We need a SYN or a FIN, unlikely.. */
11071 		return (0);
11072 	}
11073 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
11074 		/* Timestamp is behind .. old ack with seq wrap? */
11075 		return (0);
11076 	}
11077 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
11078 		/* Still recovering */
11079 		return (0);
11080 	}
11081 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11082 	if (rack->r_ctl.rc_sacked) {
11083 		/* We have sack holes on our scoreboard */
11084 		return (0);
11085 	}
11086 	/* Ok if we reach here, we can process a fast-ack */
11087 	if (rack->gp_ready &&
11088 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11089 		under_pacing = 1;
11090 	}
11091 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
11092 	rack_log_ack(tp, to, th, 0, 0);
11093 	/* Did the window get updated? */
11094 	if (tiwin != tp->snd_wnd) {
11095 		tp->snd_wnd = tiwin;
11096 		rack_validate_fo_sendwin_up(tp, rack);
11097 		tp->snd_wl1 = th->th_seq;
11098 		if (tp->snd_wnd > tp->max_sndwnd)
11099 			tp->max_sndwnd = tp->snd_wnd;
11100 	}
11101 	/* Do we exit persists? */
11102 	if ((rack->rc_in_persist != 0) &&
11103 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
11104 			       rack->r_ctl.rc_pace_min_segs))) {
11105 		rack_exit_persist(tp, rack, cts);
11106 	}
11107 	/* Do we enter persists? */
11108 	if ((rack->rc_in_persist == 0) &&
11109 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
11110 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
11111 	    (tp->snd_max == tp->snd_una) &&
11112 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
11113 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
11114 		/*
11115 		 * Here the rwnd is less than
11116 		 * the pacing size, we are established,
11117 		 * nothing is outstanding, and there is
11118 		 * data to send. Enter persists.
11119 		 */
11120 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
11121 	}
11122 	/*
11123 	 * If last ACK falls within this segment's sequence numbers, record
11124 	 * the timestamp. NOTE that the test is modified according to the
11125 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
11126 	 */
11127 	if ((to->to_flags & TOF_TS) != 0 &&
11128 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
11129 		tp->ts_recent_age = tcp_ts_getticks();
11130 		tp->ts_recent = to->to_tsval;
11131 	}
11132 	/*
11133 	 * This is a pure ack for outstanding data.
11134 	 */
11135 	KMOD_TCPSTAT_INC(tcps_predack);
11136 
11137 	/*
11138 	 * "bad retransmit" recovery.
11139 	 */
11140 	if ((tp->t_flags & TF_PREVVALID) &&
11141 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
11142 		tp->t_flags &= ~TF_PREVVALID;
11143 		if (tp->t_rxtshift == 1 &&
11144 		    (int)(ticks - tp->t_badrxtwin) < 0)
11145 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack);
11146 	}
11147 	/*
11148 	 * Recalculate the transmit timer / rtt.
11149 	 *
11150 	 * Some boxes send broken timestamp replies during the SYN+ACK
11151 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
11152 	 * and blow up the retransmit timer.
11153 	 */
11154 	acked = BYTES_THIS_ACK(tp, th);
11155 
11156 #ifdef TCP_HHOOK
11157 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
11158 	hhook_run_tcp_est_in(tp, th, to);
11159 #endif
11160 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
11161 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
11162 	if (acked) {
11163 		struct mbuf *mfree;
11164 
11165 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
11166 		SOCKBUF_LOCK(&so->so_snd);
11167 		mfree = sbcut_locked(&so->so_snd, acked);
11168 		tp->snd_una = th->th_ack;
11169 		/* Note we want to hold the sb lock through the sendmap adjust */
11170 		rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
11171 		/* Wake up the socket if we have room to write more */
11172 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
11173 		sowwakeup_locked(so);
11174 		m_freem(mfree);
11175 		tp->t_rxtshift = 0;
11176 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11177 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11178 		rack->rc_tlp_in_progress = 0;
11179 		rack->r_ctl.rc_tlp_cnt_out = 0;
11180 		/*
11181 		 * If it is the RXT timer we want to
11182 		 * stop it, so we can restart a TLP.
11183 		 */
11184 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
11185 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11186 #ifdef NETFLIX_HTTP_LOGGING
11187 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
11188 #endif
11189 	}
11190 	/*
11191 	 * Let the congestion control algorithm update congestion control
11192 	 * related information. This typically means increasing the
11193 	 * congestion window.
11194 	 */
11195 	if (tp->snd_wnd < ctf_outstanding(tp)) {
11196 		/* The peer collapsed the window */
11197 		rack_collapsed_window(rack);
11198 	} else if (rack->rc_has_collapsed)
11199 		rack_un_collapse_window(rack);
11200 
11201 	/*
11202 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
11203 	 */
11204 	tp->snd_wl2 = th->th_ack;
11205 	tp->t_dupacks = 0;
11206 	m_freem(m);
11207 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
11208 
11209 	/*
11210 	 * If all outstanding data are acked, stop retransmit timer,
11211 	 * otherwise restart timer using current (possibly backed-off)
11212 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
11213 	 * If data are ready to send, let tcp_output decide between more
11214 	 * output or persist.
11215 	 */
11216 #ifdef TCPDEBUG
11217 	if (so->so_options & SO_DEBUG)
11218 		tcp_trace(TA_INPUT, ostate, tp,
11219 		    (void *)tcp_saveipgen,
11220 		    &tcp_savetcp, 0);
11221 #endif
11222 	if (under_pacing &&
11223 	    (rack->use_fixed_rate == 0) &&
11224 	    (rack->in_probe_rtt == 0) &&
11225 	    rack->rc_gp_dyn_mul &&
11226 	    rack->rc_always_pace) {
11227 		/* Check if we are dragging bottom */
11228 		rack_check_bottom_drag(tp, rack, so, acked);
11229 	}
11230 	if (tp->snd_una == tp->snd_max) {
11231 		tp->t_flags &= ~TF_PREVVALID;
11232 		rack->r_ctl.retran_during_recovery = 0;
11233 		rack->r_ctl.dsack_byte_cnt = 0;
11234 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
11235 		if (rack->r_ctl.rc_went_idle_time == 0)
11236 			rack->r_ctl.rc_went_idle_time = 1;
11237 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
11238 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
11239 			tp->t_acktime = 0;
11240 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11241 	}
11242 	if (acked && rack->r_fast_output)
11243 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
11244 	if (sbavail(&so->so_snd)) {
11245 		rack->r_wanted_output = 1;
11246 	}
11247 	return (1);
11248 }
11249 
11250 /*
11251  * Return value of 1, the TCB is unlocked and most
11252  * likely gone, return value of 0, the TCP is still
11253  * locked.
11254  */
11255 static int
11256 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
11257     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11258     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11259 {
11260 	int32_t ret_val = 0;
11261 	int32_t todrop;
11262 	int32_t ourfinisacked = 0;
11263 	struct tcp_rack *rack;
11264 
11265 	ctf_calc_rwin(so, tp);
11266 	/*
11267 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
11268 	 * SYN, drop the input. if seg contains a RST, then drop the
11269 	 * connection. if seg does not contain SYN, then drop it. Otherwise
11270 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
11271 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
11272 	 * contains an ECE and ECN support is enabled, the stream is ECN
11273 	 * capable. if SYN has been acked change to ESTABLISHED else
11274 	 * SYN_RCVD state arrange for segment to be acked (eventually)
11275 	 * continue processing rest of data/controls.
11276 	 */
11277 	if ((thflags & TH_ACK) &&
11278 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
11279 	    SEQ_GT(th->th_ack, tp->snd_max))) {
11280 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11281 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11282 		return (1);
11283 	}
11284 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
11285 		TCP_PROBE5(connect__refused, NULL, tp,
11286 		    mtod(m, const char *), tp, th);
11287 		tp = tcp_drop(tp, ECONNREFUSED);
11288 		ctf_do_drop(m, tp);
11289 		return (1);
11290 	}
11291 	if (thflags & TH_RST) {
11292 		ctf_do_drop(m, tp);
11293 		return (1);
11294 	}
11295 	if (!(thflags & TH_SYN)) {
11296 		ctf_do_drop(m, tp);
11297 		return (1);
11298 	}
11299 	tp->irs = th->th_seq;
11300 	tcp_rcvseqinit(tp);
11301 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11302 	if (thflags & TH_ACK) {
11303 		int tfo_partial = 0;
11304 
11305 		KMOD_TCPSTAT_INC(tcps_connects);
11306 		soisconnected(so);
11307 #ifdef MAC
11308 		mac_socketpeer_set_from_mbuf(m, so);
11309 #endif
11310 		/* Do window scaling on this connection? */
11311 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11312 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11313 			tp->rcv_scale = tp->request_r_scale;
11314 		}
11315 		tp->rcv_adv += min(tp->rcv_wnd,
11316 		    TCP_MAXWIN << tp->rcv_scale);
11317 		/*
11318 		 * If not all the data that was sent in the TFO SYN
11319 		 * has been acked, resend the remainder right away.
11320 		 */
11321 		if (IS_FASTOPEN(tp->t_flags) &&
11322 		    (tp->snd_una != tp->snd_max)) {
11323 			tp->snd_nxt = th->th_ack;
11324 			tfo_partial = 1;
11325 		}
11326 		/*
11327 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
11328 		 * will be turned on later.
11329 		 */
11330 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
11331 			rack_timer_cancel(tp, rack,
11332 					  rack->r_ctl.rc_rcvtime, __LINE__);
11333 			tp->t_flags |= TF_DELACK;
11334 		} else {
11335 			rack->r_wanted_output = 1;
11336 			tp->t_flags |= TF_ACKNOW;
11337 			rack->rc_dack_toggle = 0;
11338 		}
11339 		if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) &&
11340 		    (V_tcp_do_ecn == 1)) {
11341 			tp->t_flags2 |= TF2_ECN_PERMIT;
11342 			KMOD_TCPSTAT_INC(tcps_ecn_shs);
11343 		}
11344 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
11345 			/*
11346 			 * We advance snd_una for the
11347 			 * fast open case. If th_ack is
11348 			 * acknowledging data beyond
11349 			 * snd_una we can't just call
11350 			 * ack-processing since the
11351 			 * data stream in our send-map
11352 			 * will start at snd_una + 1 (one
11353 			 * beyond the SYN). If its just
11354 			 * equal we don't need to do that
11355 			 * and there is no send_map.
11356 			 */
11357 			tp->snd_una++;
11358 		}
11359 		/*
11360 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
11361 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
11362 		 */
11363 		tp->t_starttime = ticks;
11364 		if (tp->t_flags & TF_NEEDFIN) {
11365 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
11366 			tp->t_flags &= ~TF_NEEDFIN;
11367 			thflags &= ~TH_SYN;
11368 		} else {
11369 			tcp_state_change(tp, TCPS_ESTABLISHED);
11370 			TCP_PROBE5(connect__established, NULL, tp,
11371 			    mtod(m, const char *), tp, th);
11372 			rack_cc_conn_init(tp);
11373 		}
11374 	} else {
11375 		/*
11376 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
11377 		 * open.  If segment contains CC option and there is a
11378 		 * cached CC, apply TAO test. If it succeeds, connection is *
11379 		 * half-synchronized. Otherwise, do 3-way handshake:
11380 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
11381 		 * there was no CC option, clear cached CC value.
11382 		 */
11383 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
11384 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
11385 	}
11386 	INP_WLOCK_ASSERT(tp->t_inpcb);
11387 	/*
11388 	 * Advance th->th_seq to correspond to first data byte. If data,
11389 	 * trim to stay within window, dropping FIN if necessary.
11390 	 */
11391 	th->th_seq++;
11392 	if (tlen > tp->rcv_wnd) {
11393 		todrop = tlen - tp->rcv_wnd;
11394 		m_adj(m, -todrop);
11395 		tlen = tp->rcv_wnd;
11396 		thflags &= ~TH_FIN;
11397 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
11398 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
11399 	}
11400 	tp->snd_wl1 = th->th_seq - 1;
11401 	tp->rcv_up = th->th_seq;
11402 	/*
11403 	 * Client side of transaction: already sent SYN and data. If the
11404 	 * remote host used T/TCP to validate the SYN, our data will be
11405 	 * ACK'd; if so, enter normal data segment processing in the middle
11406 	 * of step 5, ack processing. Otherwise, goto step 6.
11407 	 */
11408 	if (thflags & TH_ACK) {
11409 		/* For syn-sent we need to possibly update the rtt */
11410 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11411 			uint32_t t, mcts;
11412 
11413 			mcts = tcp_ts_getticks();
11414 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11415 			if (!tp->t_rttlow || tp->t_rttlow > t)
11416 				tp->t_rttlow = t;
11417 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
11418 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11419 			tcp_rack_xmit_timer_commit(rack, tp);
11420 		}
11421 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
11422 			return (ret_val);
11423 		/* We may have changed to FIN_WAIT_1 above */
11424 		if (tp->t_state == TCPS_FIN_WAIT_1) {
11425 			/*
11426 			 * In FIN_WAIT_1 STATE in addition to the processing
11427 			 * for the ESTABLISHED state if our FIN is now
11428 			 * acknowledged then enter FIN_WAIT_2.
11429 			 */
11430 			if (ourfinisacked) {
11431 				/*
11432 				 * If we can't receive any more data, then
11433 				 * closing user can proceed. Starting the
11434 				 * timer is contrary to the specification,
11435 				 * but if we don't get a FIN we'll hang
11436 				 * forever.
11437 				 *
11438 				 * XXXjl: we should release the tp also, and
11439 				 * use a compressed state.
11440 				 */
11441 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11442 					soisdisconnected(so);
11443 					tcp_timer_activate(tp, TT_2MSL,
11444 					    (tcp_fast_finwait2_recycle ?
11445 					    tcp_finwait2_timeout :
11446 					    TP_MAXIDLE(tp)));
11447 				}
11448 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
11449 			}
11450 		}
11451 	}
11452 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11453 	   tiwin, thflags, nxt_pkt));
11454 }
11455 
11456 /*
11457  * Return value of 1, the TCB is unlocked and most
11458  * likely gone, return value of 0, the TCP is still
11459  * locked.
11460  */
11461 static int
11462 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
11463     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11464     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11465 {
11466 	struct tcp_rack *rack;
11467 	int32_t ret_val = 0;
11468 	int32_t ourfinisacked = 0;
11469 
11470 	ctf_calc_rwin(so, tp);
11471 	if ((thflags & TH_ACK) &&
11472 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
11473 	    SEQ_GT(th->th_ack, tp->snd_max))) {
11474 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11475 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11476 		return (1);
11477 	}
11478 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11479 	if (IS_FASTOPEN(tp->t_flags)) {
11480 		/*
11481 		 * When a TFO connection is in SYN_RECEIVED, the
11482 		 * only valid packets are the initial SYN, a
11483 		 * retransmit/copy of the initial SYN (possibly with
11484 		 * a subset of the original data), a valid ACK, a
11485 		 * FIN, or a RST.
11486 		 */
11487 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
11488 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11489 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11490 			return (1);
11491 		} else if (thflags & TH_SYN) {
11492 			/* non-initial SYN is ignored */
11493 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
11494 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
11495 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
11496 				ctf_do_drop(m, NULL);
11497 				return (0);
11498 			}
11499 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
11500 			ctf_do_drop(m, NULL);
11501 			return (0);
11502 		}
11503 	}
11504 	if ((thflags & TH_RST) ||
11505 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11506 		return (ctf_process_rst(m, th, so, tp));
11507 	/*
11508 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11509 	 * it's less than ts_recent, drop it.
11510 	 */
11511 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11512 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11513 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11514 			return (ret_val);
11515 	}
11516 	/*
11517 	 * In the SYN-RECEIVED state, validate that the packet belongs to
11518 	 * this connection before trimming the data to fit the receive
11519 	 * window.  Check the sequence number versus IRS since we know the
11520 	 * sequence numbers haven't wrapped.  This is a partial fix for the
11521 	 * "LAND" DoS attack.
11522 	 */
11523 	if (SEQ_LT(th->th_seq, tp->irs)) {
11524 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11525 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11526 		return (1);
11527 	}
11528 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11529 			      &rack->r_ctl.challenge_ack_ts,
11530 			      &rack->r_ctl.challenge_ack_cnt)) {
11531 		return (ret_val);
11532 	}
11533 	/*
11534 	 * If last ACK falls within this segment's sequence numbers, record
11535 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11536 	 * from the latest proposal of the tcplw@cray.com list (Braden
11537 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11538 	 * with our earlier PAWS tests, so this check should be solely
11539 	 * predicated on the sequence space of this segment. 3) That we
11540 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11541 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11542 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11543 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11544 	 * p.869. In such cases, we can still calculate the RTT correctly
11545 	 * when RCV.NXT == Last.ACK.Sent.
11546 	 */
11547 	if ((to->to_flags & TOF_TS) != 0 &&
11548 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11549 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11550 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11551 		tp->ts_recent_age = tcp_ts_getticks();
11552 		tp->ts_recent = to->to_tsval;
11553 	}
11554 	tp->snd_wnd = tiwin;
11555 	rack_validate_fo_sendwin_up(tp, rack);
11556 	/*
11557 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11558 	 * is on (half-synchronized state), then queue data for later
11559 	 * processing; else drop segment and return.
11560 	 */
11561 	if ((thflags & TH_ACK) == 0) {
11562 		if (IS_FASTOPEN(tp->t_flags)) {
11563 			rack_cc_conn_init(tp);
11564 		}
11565 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11566 		    tiwin, thflags, nxt_pkt));
11567 	}
11568 	KMOD_TCPSTAT_INC(tcps_connects);
11569 	soisconnected(so);
11570 	/* Do window scaling? */
11571 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11572 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11573 		tp->rcv_scale = tp->request_r_scale;
11574 	}
11575 	/*
11576 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11577 	 * FIN-WAIT-1
11578 	 */
11579 	tp->t_starttime = ticks;
11580 	if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11581 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11582 		tp->t_tfo_pending = NULL;
11583 	}
11584 	if (tp->t_flags & TF_NEEDFIN) {
11585 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
11586 		tp->t_flags &= ~TF_NEEDFIN;
11587 	} else {
11588 		tcp_state_change(tp, TCPS_ESTABLISHED);
11589 		TCP_PROBE5(accept__established, NULL, tp,
11590 		    mtod(m, const char *), tp, th);
11591 		/*
11592 		 * TFO connections call cc_conn_init() during SYN
11593 		 * processing.  Calling it again here for such connections
11594 		 * is not harmless as it would undo the snd_cwnd reduction
11595 		 * that occurs when a TFO SYN|ACK is retransmitted.
11596 		 */
11597 		if (!IS_FASTOPEN(tp->t_flags))
11598 			rack_cc_conn_init(tp);
11599 	}
11600 	/*
11601 	 * Account for the ACK of our SYN prior to
11602 	 * regular ACK processing below, except for
11603 	 * simultaneous SYN, which is handled later.
11604 	 */
11605 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11606 		tp->snd_una++;
11607 	/*
11608 	 * If segment contains data or ACK, will call tcp_reass() later; if
11609 	 * not, do so now to pass queued data to user.
11610 	 */
11611 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
11612 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11613 		    (struct mbuf *)0);
11614 		if (tp->t_flags & TF_WAKESOR) {
11615 			tp->t_flags &= ~TF_WAKESOR;
11616 			/* NB: sorwakeup_locked() does an implicit unlock. */
11617 			sorwakeup_locked(so);
11618 		}
11619 	}
11620 	tp->snd_wl1 = th->th_seq - 1;
11621 	/* For syn-recv we need to possibly update the rtt */
11622 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11623 		uint32_t t, mcts;
11624 
11625 		mcts = tcp_ts_getticks();
11626 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11627 		if (!tp->t_rttlow || tp->t_rttlow > t)
11628 			tp->t_rttlow = t;
11629 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11630 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11631 		tcp_rack_xmit_timer_commit(rack, tp);
11632 	}
11633 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11634 		return (ret_val);
11635 	}
11636 	if (tp->t_state == TCPS_FIN_WAIT_1) {
11637 		/* We could have went to FIN_WAIT_1 (or EST) above */
11638 		/*
11639 		 * In FIN_WAIT_1 STATE in addition to the processing for the
11640 		 * ESTABLISHED state if our FIN is now acknowledged then
11641 		 * enter FIN_WAIT_2.
11642 		 */
11643 		if (ourfinisacked) {
11644 			/*
11645 			 * If we can't receive any more data, then closing
11646 			 * user can proceed. Starting the timer is contrary
11647 			 * to the specification, but if we don't get a FIN
11648 			 * we'll hang forever.
11649 			 *
11650 			 * XXXjl: we should release the tp also, and use a
11651 			 * compressed state.
11652 			 */
11653 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11654 				soisdisconnected(so);
11655 				tcp_timer_activate(tp, TT_2MSL,
11656 				    (tcp_fast_finwait2_recycle ?
11657 				    tcp_finwait2_timeout :
11658 				    TP_MAXIDLE(tp)));
11659 			}
11660 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
11661 		}
11662 	}
11663 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11664 	    tiwin, thflags, nxt_pkt));
11665 }
11666 
11667 /*
11668  * Return value of 1, the TCB is unlocked and most
11669  * likely gone, return value of 0, the TCP is still
11670  * locked.
11671  */
11672 static int
11673 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11674     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11675     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11676 {
11677 	int32_t ret_val = 0;
11678 	struct tcp_rack *rack;
11679 
11680 	/*
11681 	 * Header prediction: check for the two common cases of a
11682 	 * uni-directional data xfer.  If the packet has no control flags,
11683 	 * is in-sequence, the window didn't change and we're not
11684 	 * retransmitting, it's a candidate.  If the length is zero and the
11685 	 * ack moved forward, we're the sender side of the xfer.  Just free
11686 	 * the data acked & wake any higher level process that was blocked
11687 	 * waiting for space.  If the length is non-zero and the ack didn't
11688 	 * move, we're the receiver side.  If we're getting packets in-order
11689 	 * (the reassembly queue is empty), add the data toc The socket
11690 	 * buffer and note that we need a delayed ack. Make sure that the
11691 	 * hidden state-flags are also off. Since we check for
11692 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11693 	 */
11694 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11695 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11696 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11697 	    __predict_true(SEGQ_EMPTY(tp)) &&
11698 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
11699 		if (tlen == 0) {
11700 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11701 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11702 				return (0);
11703 			}
11704 		} else {
11705 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11706 			    tiwin, nxt_pkt, iptos)) {
11707 				return (0);
11708 			}
11709 		}
11710 	}
11711 	ctf_calc_rwin(so, tp);
11712 
11713 	if ((thflags & TH_RST) ||
11714 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11715 		return (ctf_process_rst(m, th, so, tp));
11716 
11717 	/*
11718 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11719 	 * synchronized state.
11720 	 */
11721 	if (thflags & TH_SYN) {
11722 		ctf_challenge_ack(m, th, tp, &ret_val);
11723 		return (ret_val);
11724 	}
11725 	/*
11726 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11727 	 * it's less than ts_recent, drop it.
11728 	 */
11729 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11730 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11731 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11732 			return (ret_val);
11733 	}
11734 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11735 			      &rack->r_ctl.challenge_ack_ts,
11736 			      &rack->r_ctl.challenge_ack_cnt)) {
11737 		return (ret_val);
11738 	}
11739 	/*
11740 	 * If last ACK falls within this segment's sequence numbers, record
11741 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11742 	 * from the latest proposal of the tcplw@cray.com list (Braden
11743 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11744 	 * with our earlier PAWS tests, so this check should be solely
11745 	 * predicated on the sequence space of this segment. 3) That we
11746 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11747 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11748 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11749 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11750 	 * p.869. In such cases, we can still calculate the RTT correctly
11751 	 * when RCV.NXT == Last.ACK.Sent.
11752 	 */
11753 	if ((to->to_flags & TOF_TS) != 0 &&
11754 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11755 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11756 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11757 		tp->ts_recent_age = tcp_ts_getticks();
11758 		tp->ts_recent = to->to_tsval;
11759 	}
11760 	/*
11761 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11762 	 * is on (half-synchronized state), then queue data for later
11763 	 * processing; else drop segment and return.
11764 	 */
11765 	if ((thflags & TH_ACK) == 0) {
11766 		if (tp->t_flags & TF_NEEDSYN) {
11767 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11768 			    tiwin, thflags, nxt_pkt));
11769 
11770 		} else if (tp->t_flags & TF_ACKNOW) {
11771 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11772 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11773 			return (ret_val);
11774 		} else {
11775 			ctf_do_drop(m, NULL);
11776 			return (0);
11777 		}
11778 	}
11779 	/*
11780 	 * Ack processing.
11781 	 */
11782 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11783 		return (ret_val);
11784 	}
11785 	if (sbavail(&so->so_snd)) {
11786 		if (ctf_progress_timeout_check(tp, true)) {
11787 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11788 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11789 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11790 			return (1);
11791 		}
11792 	}
11793 	/* State changes only happen in rack_process_data() */
11794 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11795 	    tiwin, thflags, nxt_pkt));
11796 }
11797 
11798 /*
11799  * Return value of 1, the TCB is unlocked and most
11800  * likely gone, return value of 0, the TCP is still
11801  * locked.
11802  */
11803 static int
11804 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11805     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11806     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11807 {
11808 	int32_t ret_val = 0;
11809 	struct tcp_rack *rack;
11810 
11811 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11812 	ctf_calc_rwin(so, tp);
11813 	if ((thflags & TH_RST) ||
11814 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11815 		return (ctf_process_rst(m, th, so, tp));
11816 	/*
11817 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11818 	 * synchronized state.
11819 	 */
11820 	if (thflags & TH_SYN) {
11821 		ctf_challenge_ack(m, th, tp, &ret_val);
11822 		return (ret_val);
11823 	}
11824 	/*
11825 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11826 	 * it's less than ts_recent, drop it.
11827 	 */
11828 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11829 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11830 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11831 			return (ret_val);
11832 	}
11833 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11834 			      &rack->r_ctl.challenge_ack_ts,
11835 			      &rack->r_ctl.challenge_ack_cnt)) {
11836 		return (ret_val);
11837 	}
11838 	/*
11839 	 * If last ACK falls within this segment's sequence numbers, record
11840 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11841 	 * from the latest proposal of the tcplw@cray.com list (Braden
11842 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11843 	 * with our earlier PAWS tests, so this check should be solely
11844 	 * predicated on the sequence space of this segment. 3) That we
11845 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11846 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11847 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11848 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11849 	 * p.869. In such cases, we can still calculate the RTT correctly
11850 	 * when RCV.NXT == Last.ACK.Sent.
11851 	 */
11852 	if ((to->to_flags & TOF_TS) != 0 &&
11853 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11854 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11855 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11856 		tp->ts_recent_age = tcp_ts_getticks();
11857 		tp->ts_recent = to->to_tsval;
11858 	}
11859 	/*
11860 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11861 	 * is on (half-synchronized state), then queue data for later
11862 	 * processing; else drop segment and return.
11863 	 */
11864 	if ((thflags & TH_ACK) == 0) {
11865 		if (tp->t_flags & TF_NEEDSYN) {
11866 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11867 			    tiwin, thflags, nxt_pkt));
11868 
11869 		} else if (tp->t_flags & TF_ACKNOW) {
11870 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11871 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11872 			return (ret_val);
11873 		} else {
11874 			ctf_do_drop(m, NULL);
11875 			return (0);
11876 		}
11877 	}
11878 	/*
11879 	 * Ack processing.
11880 	 */
11881 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11882 		return (ret_val);
11883 	}
11884 	if (sbavail(&so->so_snd)) {
11885 		if (ctf_progress_timeout_check(tp, true)) {
11886 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11887 						tp, tick, PROGRESS_DROP, __LINE__);
11888 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
11889 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11890 			return (1);
11891 		}
11892 	}
11893 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11894 	    tiwin, thflags, nxt_pkt));
11895 }
11896 
11897 static int
11898 rack_check_data_after_close(struct mbuf *m,
11899     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11900 {
11901 	struct tcp_rack *rack;
11902 
11903 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11904 	if (rack->rc_allow_data_af_clo == 0) {
11905 	close_now:
11906 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11907 		/* tcp_close will kill the inp pre-log the Reset */
11908 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11909 		tp = tcp_close(tp);
11910 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11911 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11912 		return (1);
11913 	}
11914 	if (sbavail(&so->so_snd) == 0)
11915 		goto close_now;
11916 	/* Ok we allow data that is ignored and a followup reset */
11917 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11918 	tp->rcv_nxt = th->th_seq + *tlen;
11919 	tp->t_flags2 |= TF2_DROP_AF_DATA;
11920 	rack->r_wanted_output = 1;
11921 	*tlen = 0;
11922 	return (0);
11923 }
11924 
11925 /*
11926  * Return value of 1, the TCB is unlocked and most
11927  * likely gone, return value of 0, the TCP is still
11928  * locked.
11929  */
11930 static int
11931 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11932     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11933     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11934 {
11935 	int32_t ret_val = 0;
11936 	int32_t ourfinisacked = 0;
11937 	struct tcp_rack *rack;
11938 
11939 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11940 	ctf_calc_rwin(so, tp);
11941 
11942 	if ((thflags & TH_RST) ||
11943 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11944 		return (ctf_process_rst(m, th, so, tp));
11945 	/*
11946 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11947 	 * synchronized state.
11948 	 */
11949 	if (thflags & TH_SYN) {
11950 		ctf_challenge_ack(m, th, tp, &ret_val);
11951 		return (ret_val);
11952 	}
11953 	/*
11954 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11955 	 * it's less than ts_recent, drop it.
11956 	 */
11957 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11958 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11959 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11960 			return (ret_val);
11961 	}
11962 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11963 			      &rack->r_ctl.challenge_ack_ts,
11964 			      &rack->r_ctl.challenge_ack_cnt)) {
11965 		return (ret_val);
11966 	}
11967 	/*
11968 	 * If new data are received on a connection after the user processes
11969 	 * are gone, then RST the other end.
11970 	 */
11971 	if ((so->so_state & SS_NOFDREF) && tlen) {
11972 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
11973 			return (1);
11974 	}
11975 	/*
11976 	 * If last ACK falls within this segment's sequence numbers, record
11977 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11978 	 * from the latest proposal of the tcplw@cray.com list (Braden
11979 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11980 	 * with our earlier PAWS tests, so this check should be solely
11981 	 * predicated on the sequence space of this segment. 3) That we
11982 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11983 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11984 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11985 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11986 	 * p.869. In such cases, we can still calculate the RTT correctly
11987 	 * when RCV.NXT == Last.ACK.Sent.
11988 	 */
11989 	if ((to->to_flags & TOF_TS) != 0 &&
11990 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11991 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11992 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11993 		tp->ts_recent_age = tcp_ts_getticks();
11994 		tp->ts_recent = to->to_tsval;
11995 	}
11996 	/*
11997 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11998 	 * is on (half-synchronized state), then queue data for later
11999 	 * processing; else drop segment and return.
12000 	 */
12001 	if ((thflags & TH_ACK) == 0) {
12002 		if (tp->t_flags & TF_NEEDSYN) {
12003 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12004 			    tiwin, thflags, nxt_pkt));
12005 		} else if (tp->t_flags & TF_ACKNOW) {
12006 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12007 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12008 			return (ret_val);
12009 		} else {
12010 			ctf_do_drop(m, NULL);
12011 			return (0);
12012 		}
12013 	}
12014 	/*
12015 	 * Ack processing.
12016 	 */
12017 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12018 		return (ret_val);
12019 	}
12020 	if (ourfinisacked) {
12021 		/*
12022 		 * If we can't receive any more data, then closing user can
12023 		 * proceed. Starting the timer is contrary to the
12024 		 * specification, but if we don't get a FIN we'll hang
12025 		 * forever.
12026 		 *
12027 		 * XXXjl: we should release the tp also, and use a
12028 		 * compressed state.
12029 		 */
12030 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12031 			soisdisconnected(so);
12032 			tcp_timer_activate(tp, TT_2MSL,
12033 			    (tcp_fast_finwait2_recycle ?
12034 			    tcp_finwait2_timeout :
12035 			    TP_MAXIDLE(tp)));
12036 		}
12037 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
12038 	}
12039 	if (sbavail(&so->so_snd)) {
12040 		if (ctf_progress_timeout_check(tp, true)) {
12041 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12042 						tp, tick, PROGRESS_DROP, __LINE__);
12043 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
12044 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12045 			return (1);
12046 		}
12047 	}
12048 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12049 	    tiwin, thflags, nxt_pkt));
12050 }
12051 
12052 /*
12053  * Return value of 1, the TCB is unlocked and most
12054  * likely gone, return value of 0, the TCP is still
12055  * locked.
12056  */
12057 static int
12058 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
12059     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12060     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12061 {
12062 	int32_t ret_val = 0;
12063 	int32_t ourfinisacked = 0;
12064 	struct tcp_rack *rack;
12065 
12066 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12067 	ctf_calc_rwin(so, tp);
12068 
12069 	if ((thflags & TH_RST) ||
12070 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12071 		return (ctf_process_rst(m, th, so, tp));
12072 	/*
12073 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12074 	 * synchronized state.
12075 	 */
12076 	if (thflags & TH_SYN) {
12077 		ctf_challenge_ack(m, th, tp, &ret_val);
12078 		return (ret_val);
12079 	}
12080 	/*
12081 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12082 	 * it's less than ts_recent, drop it.
12083 	 */
12084 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12085 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12086 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12087 			return (ret_val);
12088 	}
12089 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12090 			      &rack->r_ctl.challenge_ack_ts,
12091 			      &rack->r_ctl.challenge_ack_cnt)) {
12092 		return (ret_val);
12093 	}
12094 	/*
12095 	 * If new data are received on a connection after the user processes
12096 	 * are gone, then RST the other end.
12097 	 */
12098 	if ((so->so_state & SS_NOFDREF) && tlen) {
12099 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
12100 			return (1);
12101 	}
12102 	/*
12103 	 * If last ACK falls within this segment's sequence numbers, record
12104 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12105 	 * from the latest proposal of the tcplw@cray.com list (Braden
12106 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12107 	 * with our earlier PAWS tests, so this check should be solely
12108 	 * predicated on the sequence space of this segment. 3) That we
12109 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12110 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12111 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12112 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12113 	 * p.869. In such cases, we can still calculate the RTT correctly
12114 	 * when RCV.NXT == Last.ACK.Sent.
12115 	 */
12116 	if ((to->to_flags & TOF_TS) != 0 &&
12117 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12118 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12119 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12120 		tp->ts_recent_age = tcp_ts_getticks();
12121 		tp->ts_recent = to->to_tsval;
12122 	}
12123 	/*
12124 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12125 	 * is on (half-synchronized state), then queue data for later
12126 	 * processing; else drop segment and return.
12127 	 */
12128 	if ((thflags & TH_ACK) == 0) {
12129 		if (tp->t_flags & TF_NEEDSYN) {
12130 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12131 			    tiwin, thflags, nxt_pkt));
12132 		} else if (tp->t_flags & TF_ACKNOW) {
12133 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12134 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12135 			return (ret_val);
12136 		} else {
12137 			ctf_do_drop(m, NULL);
12138 			return (0);
12139 		}
12140 	}
12141 	/*
12142 	 * Ack processing.
12143 	 */
12144 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12145 		return (ret_val);
12146 	}
12147 	if (ourfinisacked) {
12148 		tcp_twstart(tp);
12149 		m_freem(m);
12150 		return (1);
12151 	}
12152 	if (sbavail(&so->so_snd)) {
12153 		if (ctf_progress_timeout_check(tp, true)) {
12154 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12155 						tp, tick, PROGRESS_DROP, __LINE__);
12156 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
12157 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12158 			return (1);
12159 		}
12160 	}
12161 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12162 	    tiwin, thflags, nxt_pkt));
12163 }
12164 
12165 /*
12166  * Return value of 1, the TCB is unlocked and most
12167  * likely gone, return value of 0, the TCP is still
12168  * locked.
12169  */
12170 static int
12171 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12172     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12173     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12174 {
12175 	int32_t ret_val = 0;
12176 	int32_t ourfinisacked = 0;
12177 	struct tcp_rack *rack;
12178 
12179 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12180 	ctf_calc_rwin(so, tp);
12181 
12182 	if ((thflags & TH_RST) ||
12183 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12184 		return (ctf_process_rst(m, th, so, tp));
12185 	/*
12186 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12187 	 * synchronized state.
12188 	 */
12189 	if (thflags & TH_SYN) {
12190 		ctf_challenge_ack(m, th, tp, &ret_val);
12191 		return (ret_val);
12192 	}
12193 	/*
12194 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12195 	 * it's less than ts_recent, drop it.
12196 	 */
12197 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12198 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12199 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12200 			return (ret_val);
12201 	}
12202 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12203 			      &rack->r_ctl.challenge_ack_ts,
12204 			      &rack->r_ctl.challenge_ack_cnt)) {
12205 		return (ret_val);
12206 	}
12207 	/*
12208 	 * If new data are received on a connection after the user processes
12209 	 * are gone, then RST the other end.
12210 	 */
12211 	if ((so->so_state & SS_NOFDREF) && tlen) {
12212 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
12213 			return (1);
12214 	}
12215 	/*
12216 	 * If last ACK falls within this segment's sequence numbers, record
12217 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12218 	 * from the latest proposal of the tcplw@cray.com list (Braden
12219 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12220 	 * with our earlier PAWS tests, so this check should be solely
12221 	 * predicated on the sequence space of this segment. 3) That we
12222 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12223 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12224 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12225 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12226 	 * p.869. In such cases, we can still calculate the RTT correctly
12227 	 * when RCV.NXT == Last.ACK.Sent.
12228 	 */
12229 	if ((to->to_flags & TOF_TS) != 0 &&
12230 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12231 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12232 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12233 		tp->ts_recent_age = tcp_ts_getticks();
12234 		tp->ts_recent = to->to_tsval;
12235 	}
12236 	/*
12237 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12238 	 * is on (half-synchronized state), then queue data for later
12239 	 * processing; else drop segment and return.
12240 	 */
12241 	if ((thflags & TH_ACK) == 0) {
12242 		if (tp->t_flags & TF_NEEDSYN) {
12243 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12244 			    tiwin, thflags, nxt_pkt));
12245 		} else if (tp->t_flags & TF_ACKNOW) {
12246 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12247 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12248 			return (ret_val);
12249 		} else {
12250 			ctf_do_drop(m, NULL);
12251 			return (0);
12252 		}
12253 	}
12254 	/*
12255 	 * case TCPS_LAST_ACK: Ack processing.
12256 	 */
12257 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12258 		return (ret_val);
12259 	}
12260 	if (ourfinisacked) {
12261 		tp = tcp_close(tp);
12262 		ctf_do_drop(m, tp);
12263 		return (1);
12264 	}
12265 	if (sbavail(&so->so_snd)) {
12266 		if (ctf_progress_timeout_check(tp, true)) {
12267 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12268 						tp, tick, PROGRESS_DROP, __LINE__);
12269 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
12270 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12271 			return (1);
12272 		}
12273 	}
12274 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12275 	    tiwin, thflags, nxt_pkt));
12276 }
12277 
12278 /*
12279  * Return value of 1, the TCB is unlocked and most
12280  * likely gone, return value of 0, the TCP is still
12281  * locked.
12282  */
12283 static int
12284 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
12285     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12286     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12287 {
12288 	int32_t ret_val = 0;
12289 	int32_t ourfinisacked = 0;
12290 	struct tcp_rack *rack;
12291 
12292 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12293 	ctf_calc_rwin(so, tp);
12294 
12295 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
12296 	if ((thflags & TH_RST) ||
12297 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12298 		return (ctf_process_rst(m, th, so, tp));
12299 	/*
12300 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12301 	 * synchronized state.
12302 	 */
12303 	if (thflags & TH_SYN) {
12304 		ctf_challenge_ack(m, th, tp, &ret_val);
12305 		return (ret_val);
12306 	}
12307 	/*
12308 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12309 	 * it's less than ts_recent, drop it.
12310 	 */
12311 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12312 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12313 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12314 			return (ret_val);
12315 	}
12316 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12317 			      &rack->r_ctl.challenge_ack_ts,
12318 			      &rack->r_ctl.challenge_ack_cnt)) {
12319 		return (ret_val);
12320 	}
12321 	/*
12322 	 * If new data are received on a connection after the user processes
12323 	 * are gone, then RST the other end.
12324 	 */
12325 	if ((so->so_state & SS_NOFDREF) &&
12326 	    tlen) {
12327 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
12328 			return (1);
12329 	}
12330 	/*
12331 	 * If last ACK falls within this segment's sequence numbers, record
12332 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12333 	 * from the latest proposal of the tcplw@cray.com list (Braden
12334 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12335 	 * with our earlier PAWS tests, so this check should be solely
12336 	 * predicated on the sequence space of this segment. 3) That we
12337 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12338 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12339 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12340 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12341 	 * p.869. In such cases, we can still calculate the RTT correctly
12342 	 * when RCV.NXT == Last.ACK.Sent.
12343 	 */
12344 	if ((to->to_flags & TOF_TS) != 0 &&
12345 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12346 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12347 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12348 		tp->ts_recent_age = tcp_ts_getticks();
12349 		tp->ts_recent = to->to_tsval;
12350 	}
12351 	/*
12352 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12353 	 * is on (half-synchronized state), then queue data for later
12354 	 * processing; else drop segment and return.
12355 	 */
12356 	if ((thflags & TH_ACK) == 0) {
12357 		if (tp->t_flags & TF_NEEDSYN) {
12358 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12359 			    tiwin, thflags, nxt_pkt));
12360 		} else if (tp->t_flags & TF_ACKNOW) {
12361 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12362 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12363 			return (ret_val);
12364 		} else {
12365 			ctf_do_drop(m, NULL);
12366 			return (0);
12367 		}
12368 	}
12369 	/*
12370 	 * Ack processing.
12371 	 */
12372 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12373 		return (ret_val);
12374 	}
12375 	if (sbavail(&so->so_snd)) {
12376 		if (ctf_progress_timeout_check(tp, true)) {
12377 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12378 						tp, tick, PROGRESS_DROP, __LINE__);
12379 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
12380 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12381 			return (1);
12382 		}
12383 	}
12384 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12385 	    tiwin, thflags, nxt_pkt));
12386 }
12387 
12388 static void inline
12389 rack_clear_rate_sample(struct tcp_rack *rack)
12390 {
12391 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
12392 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
12393 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
12394 }
12395 
12396 static void
12397 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
12398 {
12399 	uint64_t bw_est, rate_wanted;
12400 	int chged = 0;
12401 	uint32_t user_max, orig_min, orig_max;
12402 
12403 	orig_min = rack->r_ctl.rc_pace_min_segs;
12404 	orig_max = rack->r_ctl.rc_pace_max_segs;
12405 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
12406 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
12407 		chged = 1;
12408 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
12409 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
12410 		if (user_max != rack->r_ctl.rc_pace_max_segs)
12411 			chged = 1;
12412 	}
12413 	if (rack->rc_force_max_seg) {
12414 		rack->r_ctl.rc_pace_max_segs = user_max;
12415 	} else if (rack->use_fixed_rate) {
12416 		bw_est = rack_get_bw(rack);
12417 		if ((rack->r_ctl.crte == NULL) ||
12418 		    (bw_est != rack->r_ctl.crte->rate)) {
12419 			rack->r_ctl.rc_pace_max_segs = user_max;
12420 		} else {
12421 			/* We are pacing right at the hardware rate */
12422 			uint32_t segsiz;
12423 
12424 			segsiz = min(ctf_fixed_maxseg(tp),
12425 				     rack->r_ctl.rc_pace_min_segs);
12426 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
12427 				                           tp, bw_est, segsiz, 0,
12428 							   rack->r_ctl.crte, NULL);
12429 		}
12430 	} else if (rack->rc_always_pace) {
12431 		if (rack->r_ctl.gp_bw ||
12432 #ifdef NETFLIX_PEAKRATE
12433 		    rack->rc_tp->t_maxpeakrate ||
12434 #endif
12435 		    rack->r_ctl.init_rate) {
12436 			/* We have a rate of some sort set */
12437 			uint32_t  orig;
12438 
12439 			bw_est = rack_get_bw(rack);
12440 			orig = rack->r_ctl.rc_pace_max_segs;
12441 			if (fill_override)
12442 				rate_wanted = *fill_override;
12443 			else
12444 				rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
12445 			if (rate_wanted) {
12446 				/* We have something */
12447 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
12448 										   rate_wanted,
12449 										   ctf_fixed_maxseg(rack->rc_tp));
12450 			} else
12451 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
12452 			if (orig != rack->r_ctl.rc_pace_max_segs)
12453 				chged = 1;
12454 		} else if ((rack->r_ctl.gp_bw == 0) &&
12455 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
12456 			/*
12457 			 * If we have nothing limit us to bursting
12458 			 * out IW sized pieces.
12459 			 */
12460 			chged = 1;
12461 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
12462 		}
12463 	}
12464 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
12465 		chged = 1;
12466 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
12467 	}
12468 	if (chged)
12469 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
12470 }
12471 
12472 
12473 static void
12474 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
12475 {
12476 #ifdef INET6
12477 	struct ip6_hdr *ip6 = NULL;
12478 #endif
12479 #ifdef INET
12480 	struct ip *ip = NULL;
12481 #endif
12482 	struct udphdr *udp = NULL;
12483 
12484 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
12485 #ifdef INET6
12486 	if (rack->r_is_v6) {
12487 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12488 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
12489 		if (tp->t_port) {
12490 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12491 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
12492 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12493 			udp->uh_dport = tp->t_port;
12494 			rack->r_ctl.fsb.udp = udp;
12495 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12496 		} else
12497 		{
12498 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
12499 			rack->r_ctl.fsb.udp = NULL;
12500 		}
12501 		tcpip_fillheaders(rack->rc_inp,
12502 				  tp->t_port,
12503 				  ip6, rack->r_ctl.fsb.th);
12504 	} else
12505 #endif				/* INET6 */
12506 	{
12507 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
12508 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
12509 		if (tp->t_port) {
12510 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12511 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
12512 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12513 			udp->uh_dport = tp->t_port;
12514 			rack->r_ctl.fsb.udp = udp;
12515 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12516 		} else
12517 		{
12518 			rack->r_ctl.fsb.udp = NULL;
12519 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
12520 		}
12521 		tcpip_fillheaders(rack->rc_inp,
12522 				  tp->t_port,
12523 				  ip, rack->r_ctl.fsb.th);
12524 	}
12525 	rack->r_fsb_inited = 1;
12526 }
12527 
12528 static int
12529 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12530 {
12531 	/*
12532 	 * Allocate the larger of spaces V6 if available else just
12533 	 * V4 and include udphdr (overbook)
12534 	 */
12535 #ifdef INET6
12536 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12537 #else
12538 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12539 #endif
12540 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12541 					    M_TCPFSB, M_NOWAIT|M_ZERO);
12542 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12543 		return (ENOMEM);
12544 	}
12545 	rack->r_fsb_inited = 0;
12546 	return (0);
12547 }
12548 
12549 static int
12550 rack_init(struct tcpcb *tp)
12551 {
12552 	struct tcp_rack *rack = NULL;
12553 	struct rack_sendmap *insret;
12554 	uint32_t iwin, snt, us_cts;
12555 	int err;
12556 
12557 	tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12558 	if (tp->t_fb_ptr == NULL) {
12559 		/*
12560 		 * We need to allocate memory but cant. The INP and INP_INFO
12561 		 * locks and they are recusive (happens during setup. So a
12562 		 * scheme to drop the locks fails :(
12563 		 *
12564 		 */
12565 		return (ENOMEM);
12566 	}
12567 	memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12568 
12569 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12570 	RB_INIT(&rack->r_ctl.rc_mtree);
12571 	TAILQ_INIT(&rack->r_ctl.rc_free);
12572 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
12573 	rack->rc_tp = tp;
12574 	rack->rc_inp = tp->t_inpcb;
12575 	/* Set the flag */
12576 	rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
12577 	/* Probably not needed but lets be sure */
12578 	rack_clear_rate_sample(rack);
12579 	/*
12580 	 * Save off the default values, socket options will poke
12581 	 * at these if pacing is not on or we have not yet
12582 	 * reached where pacing is on (gp_ready/fixed enabled).
12583 	 * When they get set into the CC module (when gp_ready
12584 	 * is enabled or we enable fixed) then we will set these
12585 	 * values into the CC and place in here the old values
12586 	 * so we have a restoral. Then we will set the flag
12587 	 * rc_pacing_cc_set. That way whenever we turn off pacing
12588 	 * or switch off this stack, we will know to go restore
12589 	 * the saved values.
12590 	 */
12591 	rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12592 	rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12593 	/* We want abe like behavior as well */
12594 	rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
12595 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12596 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12597 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12598 	rack->r_ctl.roundends = tp->snd_max;
12599 	if (use_rack_rr)
12600 		rack->use_rack_rr = 1;
12601 	if (V_tcp_delack_enabled)
12602 		tp->t_delayed_ack = 1;
12603 	else
12604 		tp->t_delayed_ack = 0;
12605 #ifdef TCP_ACCOUNTING
12606 	if (rack_tcp_accounting) {
12607 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12608 	}
12609 #endif
12610 	if (rack_enable_shared_cwnd)
12611 		rack->rack_enable_scwnd = 1;
12612 	rack->rc_user_set_max_segs = rack_hptsi_segments;
12613 	rack->rc_force_max_seg = 0;
12614 	if (rack_use_imac_dack)
12615 		rack->rc_dack_mode = 1;
12616 	TAILQ_INIT(&rack->r_ctl.opt_list);
12617 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12618 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12619 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12620 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12621 	rack->r_ctl.rc_highest_us_rtt = 0;
12622 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12623 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12624 	if (rack_use_cmp_acks)
12625 		rack->r_use_cmp_ack = 1;
12626 	if (rack_disable_prr)
12627 		rack->rack_no_prr = 1;
12628 	if (rack_gp_no_rec_chg)
12629 		rack->rc_gp_no_rec_chg = 1;
12630 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12631 		rack->rc_always_pace = 1;
12632 		if (rack->use_fixed_rate || rack->gp_ready)
12633 			rack_set_cc_pacing(rack);
12634 	} else
12635 		rack->rc_always_pace = 0;
12636 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12637 		rack->r_mbuf_queue = 1;
12638 	else
12639 		rack->r_mbuf_queue = 0;
12640 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12641 		tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12642 	else
12643 		tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12644 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
12645 	if (rack_limits_scwnd)
12646 		rack->r_limit_scw = 1;
12647 	else
12648 		rack->r_limit_scw = 0;
12649 	rack->rc_labc = V_tcp_abc_l_var;
12650 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12651 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12652 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12653 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12654 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12655 	rack->r_ctl.rc_min_to = rack_min_to;
12656 	microuptime(&rack->r_ctl.act_rcv_time);
12657 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12658 	rack->r_running_late = 0;
12659 	rack->r_running_early = 0;
12660 	rack->rc_init_win = rack_default_init_window;
12661 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12662 	if (rack_hw_up_only)
12663 		rack->r_up_only = 1;
12664 	if (rack_do_dyn_mul) {
12665 		/* When dynamic adjustment is on CA needs to start at 100% */
12666 		rack->rc_gp_dyn_mul = 1;
12667 		if (rack_do_dyn_mul >= 100)
12668 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12669 	} else
12670 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12671 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12672 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12673 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12674 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12675 				rack_probertt_filter_life);
12676 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12677 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12678 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
12679 	rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12680 	rack->r_ctl.rc_time_probertt_starts = 0;
12681 	if (rack_dsack_std_based & 0x1) {
12682 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
12683 		rack->rc_rack_tmr_std_based = 1;
12684 	}
12685 	if (rack_dsack_std_based & 0x2) {
12686 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
12687 		rack->rc_rack_use_dsack = 1;
12688 	}
12689 	/* We require at least one measurement, even if the sysctl is 0 */
12690 	if (rack_req_measurements)
12691 		rack->r_ctl.req_measurements = rack_req_measurements;
12692 	else
12693 		rack->r_ctl.req_measurements = 1;
12694 	if (rack_enable_hw_pacing)
12695 		rack->rack_hdw_pace_ena = 1;
12696 	if (rack_hw_rate_caps)
12697 		rack->r_rack_hw_rate_caps = 1;
12698 	/* Do we force on detection? */
12699 #ifdef NETFLIX_EXP_DETECTION
12700 	if (tcp_force_detection)
12701 		rack->do_detection = 1;
12702 	else
12703 #endif
12704 		rack->do_detection = 0;
12705 	if (rack_non_rxt_use_cr)
12706 		rack->rack_rec_nonrxt_use_cr = 1;
12707 	err = rack_init_fsb(tp, rack);
12708 	if (err) {
12709 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12710 		tp->t_fb_ptr = NULL;
12711 		return (err);
12712 	}
12713 	if (tp->snd_una != tp->snd_max) {
12714 		/* Create a send map for the current outstanding data */
12715 		struct rack_sendmap *rsm;
12716 
12717 		rsm = rack_alloc(rack);
12718 		if (rsm == NULL) {
12719 			uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12720 			tp->t_fb_ptr = NULL;
12721 			return (ENOMEM);
12722 		}
12723 		rsm->r_no_rtt_allowed = 1;
12724 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12725 		rsm->r_rtr_cnt = 1;
12726 		rsm->r_rtr_bytes = 0;
12727 		if (tp->t_flags & TF_SENTFIN) {
12728 			rsm->r_end = tp->snd_max - 1;
12729 			rsm->r_flags |= RACK_HAS_FIN;
12730 		} else {
12731 			rsm->r_end = tp->snd_max;
12732 		}
12733 		if (tp->snd_una == tp->iss) {
12734 			/* The data space is one beyond snd_una */
12735 			rsm->r_flags |= RACK_HAS_SYN;
12736 			rsm->r_start = tp->iss;
12737 			rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
12738 		} else
12739 			rsm->r_start = tp->snd_una;
12740 		rsm->r_dupack = 0;
12741 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12742 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12743 			if (rsm->m)
12744 				rsm->orig_m_len = rsm->m->m_len;
12745 			else
12746 				rsm->orig_m_len = 0;
12747 		} else {
12748 			/*
12749 			 * This can happen if we have a stand-alone FIN or
12750 			 *  SYN.
12751 			 */
12752 			rsm->m = NULL;
12753 			rsm->orig_m_len = 0;
12754 			rsm->soff = 0;
12755 		}
12756 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12757 #ifdef INVARIANTS
12758 		if (insret != NULL) {
12759 			panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12760 			      insret, rack, rsm);
12761 		}
12762 #endif
12763 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12764 		rsm->r_in_tmap = 1;
12765 	}
12766 	/*
12767 	 * Timers in Rack are kept in microseconds so lets
12768 	 * convert any initial incoming variables
12769 	 * from ticks into usecs. Note that we
12770 	 * also change the values of t_srtt and t_rttvar, if
12771 	 * they are non-zero. They are kept with a 5
12772 	 * bit decimal so we have to carefully convert
12773 	 * these to get the full precision.
12774 	 */
12775 	rack_convert_rtts(tp);
12776 	tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12777 	if (rack_do_hystart) {
12778 		struct sockopt sopt;
12779 		struct cc_newreno_opts opt;
12780 
12781 		sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
12782 		sopt.sopt_dir = SOPT_SET;
12783 		opt.name = CC_NEWRENO_ENABLE_HYSTART;
12784 		opt.val = rack_do_hystart;
12785 		if (CC_ALGO(tp)->ctl_output != NULL)
12786 			(void)CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
12787 	}
12788 	if (rack_def_profile)
12789 		rack_set_profile(rack, rack_def_profile);
12790 	/* Cancel the GP measurement in progress */
12791 	tp->t_flags &= ~TF_GPUTINPROG;
12792 	if (SEQ_GT(tp->snd_max, tp->iss))
12793 		snt = tp->snd_max - tp->iss;
12794 	else
12795 		snt = 0;
12796 	iwin = rc_init_window(rack);
12797 	if (snt < iwin) {
12798 		/* We are not past the initial window
12799 		 * so we need to make sure cwnd is
12800 		 * correct.
12801 		 */
12802 		if (tp->snd_cwnd < iwin)
12803 			tp->snd_cwnd = iwin;
12804 		/*
12805 		 * If we are within the initial window
12806 		 * we want ssthresh to be unlimited. Setting
12807 		 * it to the rwnd (which the default stack does
12808 		 * and older racks) is not really a good idea
12809 		 * since we want to be in SS and grow both the
12810 		 * cwnd and the rwnd (via dynamic rwnd growth). If
12811 		 * we set it to the rwnd then as the peer grows its
12812 		 * rwnd we will be stuck in CA and never hit SS.
12813 		 *
12814 		 * Its far better to raise it up high (this takes the
12815 		 * risk that there as been a loss already, probably
12816 		 * we should have an indicator in all stacks of loss
12817 		 * but we don't), but considering the normal use this
12818 		 * is a risk worth taking. The consequences of not
12819 		 * hitting SS are far worse than going one more time
12820 		 * into it early on (before we have sent even a IW).
12821 		 * It is highly unlikely that we will have had a loss
12822 		 * before getting the IW out.
12823 		 */
12824 		tp->snd_ssthresh = 0xffffffff;
12825 	}
12826 	rack_stop_all_timers(tp);
12827 	/* Lets setup the fsb block */
12828 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12829 	rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12830 			     __LINE__, RACK_RTTS_INIT);
12831 	return (0);
12832 }
12833 
12834 static int
12835 rack_handoff_ok(struct tcpcb *tp)
12836 {
12837 	if ((tp->t_state == TCPS_CLOSED) ||
12838 	    (tp->t_state == TCPS_LISTEN)) {
12839 		/* Sure no problem though it may not stick */
12840 		return (0);
12841 	}
12842 	if ((tp->t_state == TCPS_SYN_SENT) ||
12843 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
12844 		/*
12845 		 * We really don't know if you support sack,
12846 		 * you have to get to ESTAB or beyond to tell.
12847 		 */
12848 		return (EAGAIN);
12849 	}
12850 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12851 		/*
12852 		 * Rack will only send a FIN after all data is acknowledged.
12853 		 * So in this case we have more data outstanding. We can't
12854 		 * switch stacks until either all data and only the FIN
12855 		 * is left (in which case rack_init() now knows how
12856 		 * to deal with that) <or> all is acknowledged and we
12857 		 * are only left with incoming data, though why you
12858 		 * would want to switch to rack after all data is acknowledged
12859 		 * I have no idea (rrs)!
12860 		 */
12861 		return (EAGAIN);
12862 	}
12863 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12864 		return (0);
12865 	}
12866 	/*
12867 	 * If we reach here we don't do SACK on this connection so we can
12868 	 * never do rack.
12869 	 */
12870 	return (EINVAL);
12871 }
12872 
12873 
12874 static void
12875 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12876 {
12877 	int ack_cmp = 0;
12878 
12879 	if (tp->t_fb_ptr) {
12880 		struct tcp_rack *rack;
12881 		struct rack_sendmap *rsm, *nrsm, *rm;
12882 
12883 		rack = (struct tcp_rack *)tp->t_fb_ptr;
12884 		if (tp->t_in_pkt) {
12885 			/*
12886 			 * It is unsafe to process the packets since a
12887 			 * reset may be lurking in them (its rare but it
12888 			 * can occur). If we were to find a RST, then we
12889 			 * would end up dropping the connection and the
12890 			 * INP lock, so when we return the caller (tcp_usrreq)
12891 			 * will blow up when it trys to unlock the inp.
12892 			 */
12893 			struct mbuf *save, *m;
12894 
12895 			m = tp->t_in_pkt;
12896 			tp->t_in_pkt = NULL;
12897 			tp->t_tail_pkt = NULL;
12898 			while (m) {
12899 				save = m->m_nextpkt;
12900 				m->m_nextpkt = NULL;
12901 				m_freem(m);
12902 				m = save;
12903 			}
12904 			if ((tp->t_inpcb) &&
12905 			    (tp->t_inpcb->inp_flags2 & INP_MBUF_ACKCMP))
12906 				ack_cmp = 1;
12907 			if (ack_cmp) {
12908 				/* Total if we used large or small (if ack-cmp was used). */
12909 				if (rack->rc_inp->inp_flags2 & INP_MBUF_L_ACKS)
12910 					counter_u64_add(rack_large_ackcmp, 1);
12911 				else
12912 					counter_u64_add(rack_small_ackcmp, 1);
12913 			}
12914 		}
12915 		tp->t_flags &= ~TF_FORCEDATA;
12916 #ifdef NETFLIX_SHARED_CWND
12917 		if (rack->r_ctl.rc_scw) {
12918 			uint32_t limit;
12919 
12920 			if (rack->r_limit_scw)
12921 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12922 			else
12923 				limit = 0;
12924 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12925 						  rack->r_ctl.rc_scw_index,
12926 						  limit);
12927 			rack->r_ctl.rc_scw = NULL;
12928 		}
12929 #endif
12930 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
12931 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12932 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12933 			rack->r_ctl.fsb.th = NULL;
12934 		}
12935 		/* Convert back to ticks, with  */
12936 		if (tp->t_srtt > 1) {
12937 			uint32_t val, frac;
12938 
12939 			val = USEC_2_TICKS(tp->t_srtt);
12940 			frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12941 			tp->t_srtt = val << TCP_RTT_SHIFT;
12942 			/*
12943 			 * frac is the fractional part here is left
12944 			 * over from converting to hz and shifting.
12945 			 * We need to convert this to the 5 bit
12946 			 * remainder.
12947 			 */
12948 			if (frac) {
12949 				if (hz == 1000) {
12950 					frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12951 				} else {
12952 					frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12953 				}
12954 				tp->t_srtt += frac;
12955 			}
12956 		}
12957 		if (tp->t_rttvar) {
12958 			uint32_t val, frac;
12959 
12960 			val = USEC_2_TICKS(tp->t_rttvar);
12961 			frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12962 			tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12963 			/*
12964 			 * frac is the fractional part here is left
12965 			 * over from converting to hz and shifting.
12966 			 * We need to convert this to the 5 bit
12967 			 * remainder.
12968 			 */
12969 			if (frac) {
12970 				if (hz == 1000) {
12971 					frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12972 				} else {
12973 					frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12974 				}
12975 				tp->t_rttvar += frac;
12976 			}
12977 		}
12978 		tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12979 		tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12980 		if (rack->rc_always_pace) {
12981 			tcp_decrement_paced_conn();
12982 			rack_undo_cc_pacing(rack);
12983 			rack->rc_always_pace = 0;
12984 		}
12985 		/* Clean up any options if they were not applied */
12986 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12987 			struct deferred_opt_list *dol;
12988 
12989 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12990 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12991 			free(dol, M_TCPDO);
12992 		}
12993 		/* rack does not use force data but other stacks may clear it */
12994 		if (rack->r_ctl.crte != NULL) {
12995 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12996 			rack->rack_hdrw_pacing = 0;
12997 			rack->r_ctl.crte = NULL;
12998 		}
12999 #ifdef TCP_BLACKBOX
13000 		tcp_log_flowend(tp);
13001 #endif
13002 		RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
13003 			rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
13004 #ifdef INVARIANTS
13005 			if (rm != rsm) {
13006 				panic("At fini, rack:%p rsm:%p rm:%p",
13007 				      rack, rsm, rm);
13008 			}
13009 #endif
13010 			uma_zfree(rack_zone, rsm);
13011 		}
13012 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
13013 		while (rsm) {
13014 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
13015 			uma_zfree(rack_zone, rsm);
13016 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
13017 		}
13018 		rack->rc_free_cnt = 0;
13019 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
13020 		tp->t_fb_ptr = NULL;
13021 	}
13022 	if (tp->t_inpcb) {
13023 		tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
13024 		tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
13025 		tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
13026 		tp->t_inpcb->inp_flags2 &= ~INP_MBUF_ACKCMP;
13027 		/* Cancel the GP measurement in progress */
13028 		tp->t_flags &= ~TF_GPUTINPROG;
13029 		tp->t_inpcb->inp_flags2 &= ~INP_MBUF_L_ACKS;
13030 	}
13031 	/* Make sure snd_nxt is correctly set */
13032 	tp->snd_nxt = tp->snd_max;
13033 }
13034 
13035 static void
13036 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
13037 {
13038 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
13039 		rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
13040 	}
13041 	switch (tp->t_state) {
13042 	case TCPS_SYN_SENT:
13043 		rack->r_state = TCPS_SYN_SENT;
13044 		rack->r_substate = rack_do_syn_sent;
13045 		break;
13046 	case TCPS_SYN_RECEIVED:
13047 		rack->r_state = TCPS_SYN_RECEIVED;
13048 		rack->r_substate = rack_do_syn_recv;
13049 		break;
13050 	case TCPS_ESTABLISHED:
13051 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13052 		rack->r_state = TCPS_ESTABLISHED;
13053 		rack->r_substate = rack_do_established;
13054 		break;
13055 	case TCPS_CLOSE_WAIT:
13056 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13057 		rack->r_state = TCPS_CLOSE_WAIT;
13058 		rack->r_substate = rack_do_close_wait;
13059 		break;
13060 	case TCPS_FIN_WAIT_1:
13061 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13062 		rack->r_state = TCPS_FIN_WAIT_1;
13063 		rack->r_substate = rack_do_fin_wait_1;
13064 		break;
13065 	case TCPS_CLOSING:
13066 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13067 		rack->r_state = TCPS_CLOSING;
13068 		rack->r_substate = rack_do_closing;
13069 		break;
13070 	case TCPS_LAST_ACK:
13071 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13072 		rack->r_state = TCPS_LAST_ACK;
13073 		rack->r_substate = rack_do_lastack;
13074 		break;
13075 	case TCPS_FIN_WAIT_2:
13076 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13077 		rack->r_state = TCPS_FIN_WAIT_2;
13078 		rack->r_substate = rack_do_fin_wait_2;
13079 		break;
13080 	case TCPS_LISTEN:
13081 	case TCPS_CLOSED:
13082 	case TCPS_TIME_WAIT:
13083 	default:
13084 		break;
13085 	};
13086 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
13087 		rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
13088 
13089 }
13090 
13091 static void
13092 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
13093 {
13094 	/*
13095 	 * We received an ack, and then did not
13096 	 * call send or were bounced out due to the
13097 	 * hpts was running. Now a timer is up as well, is
13098 	 * it the right timer?
13099 	 */
13100 	struct rack_sendmap *rsm;
13101 	int tmr_up;
13102 
13103 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
13104 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
13105 		return;
13106 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
13107 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
13108 	    (tmr_up == PACE_TMR_RXT)) {
13109 		/* Should be an RXT */
13110 		return;
13111 	}
13112 	if (rsm == NULL) {
13113 		/* Nothing outstanding? */
13114 		if (tp->t_flags & TF_DELACK) {
13115 			if (tmr_up == PACE_TMR_DELACK)
13116 				/* We are supposed to have delayed ack up and we do */
13117 				return;
13118 		} else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) {
13119 			/*
13120 			 * if we hit enobufs then we would expect the possiblity
13121 			 * of nothing outstanding and the RXT up (and the hptsi timer).
13122 			 */
13123 			return;
13124 		} else if (((V_tcp_always_keepalive ||
13125 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
13126 			    (tp->t_state <= TCPS_CLOSING)) &&
13127 			   (tmr_up == PACE_TMR_KEEP) &&
13128 			   (tp->snd_max == tp->snd_una)) {
13129 			/* We should have keep alive up and we do */
13130 			return;
13131 		}
13132 	}
13133 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
13134 		   ((tmr_up == PACE_TMR_TLP) ||
13135 		    (tmr_up == PACE_TMR_RACK) ||
13136 		    (tmr_up == PACE_TMR_RXT))) {
13137 		/*
13138 		 * Either a Rack, TLP or RXT is fine if  we
13139 		 * have outstanding data.
13140 		 */
13141 		return;
13142 	} else if (tmr_up == PACE_TMR_DELACK) {
13143 		/*
13144 		 * If the delayed ack was going to go off
13145 		 * before the rtx/tlp/rack timer were going to
13146 		 * expire, then that would be the timer in control.
13147 		 * Note we don't check the time here trusting the
13148 		 * code is correct.
13149 		 */
13150 		return;
13151 	}
13152 	/*
13153 	 * Ok the timer originally started is not what we want now.
13154 	 * We will force the hpts to be stopped if any, and restart
13155 	 * with the slot set to what was in the saved slot.
13156 	 */
13157 	if (rack->rc_inp->inp_in_hpts) {
13158 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
13159 			uint32_t us_cts;
13160 
13161 			us_cts = tcp_get_usecs(NULL);
13162 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
13163 				rack->r_early = 1;
13164 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
13165 			}
13166 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
13167 		}
13168 		tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
13169 	}
13170 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13171 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13172 }
13173 
13174 
13175 static void
13176 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)
13177 {
13178 	if ((SEQ_LT(tp->snd_wl1, seq) ||
13179 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
13180 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
13181 		/* keep track of pure window updates */
13182 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
13183 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
13184 		tp->snd_wnd = tiwin;
13185 		rack_validate_fo_sendwin_up(tp, rack);
13186 		tp->snd_wl1 = seq;
13187 		tp->snd_wl2 = ack;
13188 		if (tp->snd_wnd > tp->max_sndwnd)
13189 			tp->max_sndwnd = tp->snd_wnd;
13190 	    rack->r_wanted_output = 1;
13191 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
13192 		tp->snd_wnd = tiwin;
13193 		rack_validate_fo_sendwin_up(tp, rack);
13194 		tp->snd_wl1 = seq;
13195 		tp->snd_wl2 = ack;
13196 	} else {
13197 		/* Not a valid win update */
13198 		return;
13199 	}
13200 	if (tp->snd_wnd > tp->max_sndwnd)
13201 		tp->max_sndwnd = tp->snd_wnd;
13202 	if (tp->snd_wnd < (tp->snd_max - high_seq)) {
13203 		/* The peer collapsed the window */
13204 		rack_collapsed_window(rack);
13205 	} else if (rack->rc_has_collapsed)
13206 		rack_un_collapse_window(rack);
13207 	/* Do we exit persists? */
13208 	if ((rack->rc_in_persist != 0) &&
13209 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13210 				rack->r_ctl.rc_pace_min_segs))) {
13211 		rack_exit_persist(tp, rack, cts);
13212 	}
13213 	/* Do we enter persists? */
13214 	if ((rack->rc_in_persist == 0) &&
13215 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13216 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
13217 	    (tp->snd_max == tp->snd_una) &&
13218 	    sbavail(&tp->t_inpcb->inp_socket->so_snd) &&
13219 	    (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) {
13220 		/*
13221 		 * Here the rwnd is less than
13222 		 * the pacing size, we are established,
13223 		 * nothing is outstanding, and there is
13224 		 * data to send. Enter persists.
13225 		 */
13226 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
13227 	}
13228 }
13229 
13230 static void
13231 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
13232 {
13233 
13234 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13235 		union tcp_log_stackspecific log;
13236 		struct timeval ltv;
13237 		char tcp_hdr_buf[60];
13238 		struct tcphdr *th;
13239 		struct timespec ts;
13240 		uint32_t orig_snd_una;
13241 		uint8_t xx = 0;
13242 
13243 #ifdef NETFLIX_HTTP_LOGGING
13244 		struct http_sendfile_track *http_req;
13245 
13246 		if (SEQ_GT(ae->ack, tp->snd_una)) {
13247 			http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
13248 		} else {
13249 			http_req = tcp_http_find_req_for_seq(tp, ae->ack);
13250 		}
13251 #endif
13252 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13253 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
13254 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
13255 		if (rack->rack_no_prr == 0)
13256 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13257 		else
13258 			log.u_bbr.flex1 = 0;
13259 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13260 		log.u_bbr.use_lt_bw <<= 1;
13261 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
13262 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13263 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
13264 		log.u_bbr.pkts_out = tp->t_maxseg;
13265 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13266 		log.u_bbr.flex7 = 1;
13267 		log.u_bbr.lost = ae->flags;
13268 		log.u_bbr.cwnd_gain = ackval;
13269 		log.u_bbr.pacing_gain = 0x2;
13270 		if (ae->flags & TSTMP_HDWR) {
13271 			/* Record the hardware timestamp if present */
13272 			log.u_bbr.flex3 = M_TSTMP;
13273 			ts.tv_sec = ae->timestamp / 1000000000;
13274 			ts.tv_nsec = ae->timestamp % 1000000000;
13275 			ltv.tv_sec = ts.tv_sec;
13276 			ltv.tv_usec = ts.tv_nsec / 1000;
13277 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13278 		} else if (ae->flags & TSTMP_LRO) {
13279 			/* Record the LRO the arrival timestamp */
13280 			log.u_bbr.flex3 = M_TSTMP_LRO;
13281 			ts.tv_sec = ae->timestamp / 1000000000;
13282 			ts.tv_nsec = ae->timestamp % 1000000000;
13283 			ltv.tv_sec = ts.tv_sec;
13284 			ltv.tv_usec = ts.tv_nsec / 1000;
13285 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13286 		}
13287 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13288 		/* Log the rcv time */
13289 		log.u_bbr.delRate = ae->timestamp;
13290 #ifdef NETFLIX_HTTP_LOGGING
13291 		log.u_bbr.applimited = tp->t_http_closed;
13292 		log.u_bbr.applimited <<= 8;
13293 		log.u_bbr.applimited |= tp->t_http_open;
13294 		log.u_bbr.applimited <<= 8;
13295 		log.u_bbr.applimited |= tp->t_http_req;
13296 		if (http_req) {
13297 			/* Copy out any client req info */
13298 			/* seconds */
13299 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13300 			/* useconds */
13301 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13302 			log.u_bbr.rttProp = http_req->timestamp;
13303 			log.u_bbr.cur_del_rate = http_req->start;
13304 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13305 				log.u_bbr.flex8 |= 1;
13306 			} else {
13307 				log.u_bbr.flex8 |= 2;
13308 				log.u_bbr.bw_inuse = http_req->end;
13309 			}
13310 			log.u_bbr.flex6 = http_req->start_seq;
13311 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13312 				log.u_bbr.flex8 |= 4;
13313 				log.u_bbr.epoch = http_req->end_seq;
13314 			}
13315 		}
13316 #endif
13317 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
13318 		th = (struct tcphdr *)tcp_hdr_buf;
13319 		th->th_seq = ae->seq;
13320 		th->th_ack = ae->ack;
13321 		th->th_win = ae->win;
13322 		/* Now fill in the ports */
13323 		th->th_sport = tp->t_inpcb->inp_fport;
13324 		th->th_dport = tp->t_inpcb->inp_lport;
13325 		th->th_flags = ae->flags & 0xff;
13326 		/* Now do we have a timestamp option? */
13327 		if (ae->flags & HAS_TSTMP) {
13328 			u_char *cp;
13329 			uint32_t val;
13330 
13331 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
13332 			cp = (u_char *)(th + 1);
13333 			*cp = TCPOPT_NOP;
13334 			cp++;
13335 			*cp = TCPOPT_NOP;
13336 			cp++;
13337 			*cp = TCPOPT_TIMESTAMP;
13338 			cp++;
13339 			*cp = TCPOLEN_TIMESTAMP;
13340 			cp++;
13341 			val = htonl(ae->ts_value);
13342 			bcopy((char *)&val,
13343 			      (char *)cp, sizeof(uint32_t));
13344 			val = htonl(ae->ts_echo);
13345 			bcopy((char *)&val,
13346 			      (char *)(cp + 4), sizeof(uint32_t));
13347 		} else
13348 			th->th_off = (sizeof(struct tcphdr) >> 2);
13349 
13350 		/*
13351 		 * For sane logging we need to play a little trick.
13352 		 * If the ack were fully processed we would have moved
13353 		 * snd_una to high_seq, but since compressed acks are
13354 		 * processed in two phases, at this point (logging) snd_una
13355 		 * won't be advanced. So we would see multiple acks showing
13356 		 * the advancement. We can prevent that by "pretending" that
13357 		 * snd_una was advanced and then un-advancing it so that the
13358 		 * logging code has the right value for tlb_snd_una.
13359 		 */
13360 		if (tp->snd_una != high_seq) {
13361 			orig_snd_una = tp->snd_una;
13362 			tp->snd_una = high_seq;
13363 			xx = 1;
13364 		} else
13365 			xx = 0;
13366 		TCP_LOG_EVENTP(tp, th,
13367 			       &tp->t_inpcb->inp_socket->so_rcv,
13368 			       &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_IN, 0,
13369 			       0, &log, true, &ltv);
13370 		if (xx) {
13371 			tp->snd_una = orig_snd_una;
13372 		}
13373 	}
13374 
13375 }
13376 
13377 static int
13378 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
13379 {
13380 	/*
13381 	 * Handle a "special" compressed ack mbuf. Each incoming
13382 	 * ack has only four possible dispositions:
13383 	 *
13384 	 * A) It moves the cum-ack forward
13385 	 * B) It is behind the cum-ack.
13386 	 * C) It is a window-update ack.
13387 	 * D) It is a dup-ack.
13388 	 *
13389 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
13390 	 * in the incoming mbuf. We also need to still pay attention
13391 	 * to nxt_pkt since there may be another packet after this
13392 	 * one.
13393 	 */
13394 #ifdef TCP_ACCOUNTING
13395 	uint64_t ts_val;
13396 	uint64_t rdstc;
13397 #endif
13398 	int segsiz;
13399 	struct timespec ts;
13400 	struct tcp_rack *rack;
13401 	struct tcp_ackent *ae;
13402 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
13403 	int cnt, i, did_out, ourfinisacked = 0;
13404 	struct tcpopt to_holder, *to = NULL;
13405 	int win_up_req = 0;
13406 	int nsegs = 0;
13407 	int under_pacing = 1;
13408 	int recovery = 0;
13409 	int idx;
13410 #ifdef TCP_ACCOUNTING
13411 	sched_pin();
13412 #endif
13413 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13414 	if (rack->gp_ready &&
13415 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
13416 		under_pacing = 0;
13417 	else
13418 		under_pacing = 1;
13419 
13420 	if (rack->r_state != tp->t_state)
13421 		rack_set_state(tp, rack);
13422 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13423 	    (tp->t_flags & TF_GPUTINPROG)) {
13424 		/*
13425 		 * We have a goodput in progress
13426 		 * and we have entered a late state.
13427 		 * Do we have enough data in the sb
13428 		 * to handle the GPUT request?
13429 		 */
13430 		uint32_t bytes;
13431 
13432 		bytes = tp->gput_ack - tp->gput_seq;
13433 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
13434 			bytes += tp->gput_seq - tp->snd_una;
13435 		if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
13436 			/*
13437 			 * There are not enough bytes in the socket
13438 			 * buffer that have been sent to cover this
13439 			 * measurement. Cancel it.
13440 			 */
13441 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
13442 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
13443 						   tp->gput_seq,
13444 						   0, 0, 18, __LINE__, NULL, 0);
13445 			tp->t_flags &= ~TF_GPUTINPROG;
13446 		}
13447 	}
13448 	to = &to_holder;
13449 	to->to_flags = 0;
13450 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
13451 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
13452 	cnt = m->m_len / sizeof(struct tcp_ackent);
13453 	idx = cnt / 5;
13454 	if (idx >= MAX_NUM_OF_CNTS)
13455 		idx = MAX_NUM_OF_CNTS - 1;
13456 	counter_u64_add(rack_proc_comp_ack[idx], 1);
13457 	counter_u64_add(rack_multi_single_eq, cnt);
13458 	high_seq = tp->snd_una;
13459 	the_win = tp->snd_wnd;
13460 	win_seq = tp->snd_wl1;
13461 	win_upd_ack = tp->snd_wl2;
13462 	cts = tcp_tv_to_usectick(tv);
13463 	ms_cts = tcp_tv_to_mssectick(tv);
13464 	segsiz = ctf_fixed_maxseg(tp);
13465 	if ((rack->rc_gp_dyn_mul) &&
13466 	    (rack->use_fixed_rate == 0) &&
13467 	    (rack->rc_always_pace)) {
13468 		/* Check in on probertt */
13469 		rack_check_probe_rtt(rack, cts);
13470 	}
13471 	for (i = 0; i < cnt; i++) {
13472 #ifdef TCP_ACCOUNTING
13473 		ts_val = get_cyclecount();
13474 #endif
13475 		rack_clear_rate_sample(rack);
13476 		ae = ((mtod(m, struct tcp_ackent *)) + i);
13477 		/* Setup the window */
13478 		tiwin = ae->win << tp->snd_scale;
13479 		/* figure out the type of ack */
13480 		if (SEQ_LT(ae->ack, high_seq)) {
13481 			/* Case B*/
13482 			ae->ack_val_set = ACK_BEHIND;
13483 		} else if (SEQ_GT(ae->ack, high_seq)) {
13484 			/* Case A */
13485 			ae->ack_val_set = ACK_CUMACK;
13486 		} else if (tiwin == the_win) {
13487 			/* Case D */
13488 			ae->ack_val_set = ACK_DUPACK;
13489 		} else {
13490 			/* Case C */
13491 			ae->ack_val_set = ACK_RWND;
13492 		}
13493 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
13494 		/* Validate timestamp */
13495 		if (ae->flags & HAS_TSTMP) {
13496 			/* Setup for a timestamp */
13497 			to->to_flags = TOF_TS;
13498 			ae->ts_echo -= tp->ts_offset;
13499 			to->to_tsecr = ae->ts_echo;
13500 			to->to_tsval = ae->ts_value;
13501 			/*
13502 			 * If echoed timestamp is later than the current time, fall back to
13503 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13504 			 * were used when this connection was established.
13505 			 */
13506 			if (TSTMP_GT(ae->ts_echo, ms_cts))
13507 				to->to_tsecr = 0;
13508 			if (tp->ts_recent &&
13509 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
13510 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
13511 #ifdef TCP_ACCOUNTING
13512 					rdstc = get_cyclecount();
13513 					if (rdstc > ts_val) {
13514 						counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
13515 								(rdstc - ts_val));
13516 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13517 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13518 						}
13519 					}
13520 #endif
13521 					continue;
13522 				}
13523 			}
13524 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
13525 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
13526 				tp->ts_recent_age = tcp_ts_getticks();
13527 				tp->ts_recent = ae->ts_value;
13528 			}
13529 		} else {
13530 			/* Setup for a no options */
13531 			to->to_flags = 0;
13532 		}
13533 		/* Update the rcv time and perform idle reduction possibly */
13534 		if  (tp->t_idle_reduce &&
13535 		     (tp->snd_max == tp->snd_una) &&
13536 		     ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13537 			counter_u64_add(rack_input_idle_reduces, 1);
13538 			rack_cc_after_idle(rack, tp);
13539 		}
13540 		tp->t_rcvtime = ticks;
13541 		/* Now what about ECN? */
13542 		if (tp->t_flags2 & TF2_ECN_PERMIT) {
13543 			if (ae->flags & TH_CWR) {
13544 				tp->t_flags2 &= ~TF2_ECN_SND_ECE;
13545 				tp->t_flags |= TF_ACKNOW;
13546 			}
13547 			switch (ae->codepoint & IPTOS_ECN_MASK) {
13548 			case IPTOS_ECN_CE:
13549 				tp->t_flags2 |= TF2_ECN_SND_ECE;
13550 				KMOD_TCPSTAT_INC(tcps_ecn_ce);
13551 				break;
13552 			case IPTOS_ECN_ECT0:
13553 				KMOD_TCPSTAT_INC(tcps_ecn_ect0);
13554 				break;
13555 			case IPTOS_ECN_ECT1:
13556 				KMOD_TCPSTAT_INC(tcps_ecn_ect1);
13557 				break;
13558 			}
13559 
13560 			/* Process a packet differently from RFC3168. */
13561 			cc_ecnpkt_handler_flags(tp, ae->flags, ae->codepoint);
13562 			/* Congestion experienced. */
13563 			if (ae->flags & TH_ECE) {
13564 				rack_cong_signal(tp,  CC_ECN, ae->ack);
13565 			}
13566 		}
13567 #ifdef TCP_ACCOUNTING
13568 		/* Count for the specific type of ack in */
13569 		counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
13570 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13571 			tp->tcp_cnt_counters[ae->ack_val_set]++;
13572 		}
13573 #endif
13574 		/*
13575 		 * Note how we could move up these in the determination
13576 		 * above, but we don't so that way the timestamp checks (and ECN)
13577 		 * is done first before we do any processing on the ACK.
13578 		 * The non-compressed path through the code has this
13579 		 * weakness (noted by @jtl) that it actually does some
13580 		 * processing before verifying the timestamp information.
13581 		 * We don't take that path here which is why we set
13582 		 * the ack_val_set first, do the timestamp and ecn
13583 		 * processing, and then look at what we have setup.
13584 		 */
13585 		if (ae->ack_val_set == ACK_BEHIND) {
13586 			/*
13587 			 * Case B flag reordering, if window is not closed
13588 			 * or it could be a keep-alive or persists
13589 			 */
13590 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
13591 				counter_u64_add(rack_reorder_seen, 1);
13592 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13593 			}
13594 		} else if (ae->ack_val_set == ACK_DUPACK) {
13595 			/* Case D */
13596 			rack_strike_dupack(rack);
13597 		} else if (ae->ack_val_set == ACK_RWND) {
13598 			/* Case C */
13599 			win_up_req = 1;
13600 			win_upd_ack = ae->ack;
13601 			win_seq = ae->seq;
13602 			the_win = tiwin;
13603 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13604 		} else {
13605 			/* Case A */
13606 			if (SEQ_GT(ae->ack, tp->snd_max)) {
13607 				/*
13608 				 * We just send an ack since the incoming
13609 				 * ack is beyond the largest seq we sent.
13610 				 */
13611 				if ((tp->t_flags & TF_ACKNOW) == 0) {
13612 					ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13613 					if (tp->t_flags && TF_ACKNOW)
13614 						rack->r_wanted_output = 1;
13615 				}
13616 			} else {
13617 				nsegs++;
13618 				/* If the window changed setup to update */
13619 				if (tiwin != tp->snd_wnd) {
13620 					win_upd_ack = ae->ack;
13621 					win_seq = ae->seq;
13622 					the_win = tiwin;
13623 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13624 				}
13625 #ifdef TCP_ACCOUNTING
13626 				/* Account for the acks */
13627 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13628 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13629 				}
13630 				counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13631 						(((ae->ack - high_seq) + segsiz - 1) / segsiz));
13632 #endif
13633 				high_seq = ae->ack;
13634 				if (SEQ_GEQ(high_seq, rack->r_ctl.roundends)) {
13635 					rack->r_ctl.current_round++;
13636 					rack->r_ctl.roundends = tp->snd_max;
13637 					if (CC_ALGO(tp)->newround != NULL) {
13638 						CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round);
13639 					}
13640 				}
13641 				/* Setup our act_rcv_time */
13642 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13643 					ts.tv_sec = ae->timestamp / 1000000000;
13644 					ts.tv_nsec = ae->timestamp % 1000000000;
13645 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13646 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13647 				} else {
13648 					rack->r_ctl.act_rcv_time = *tv;
13649 				}
13650 				rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13651 				if (rack->rc_dsack_round_seen) {
13652 					/* Is the dsack round over? */
13653 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
13654 						/* Yes it is */
13655 						rack->rc_dsack_round_seen = 0;
13656 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
13657 					}
13658 				}
13659 			}
13660 		}
13661 		/* And lets be sure to commit the rtt measurements for this ack */
13662 		tcp_rack_xmit_timer_commit(rack, tp);
13663 #ifdef TCP_ACCOUNTING
13664 		rdstc = get_cyclecount();
13665 		if (rdstc > ts_val) {
13666 			counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13667 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13668 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13669 				if (ae->ack_val_set == ACK_CUMACK)
13670 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13671 			}
13672 		}
13673 #endif
13674 	}
13675 #ifdef TCP_ACCOUNTING
13676 	ts_val = get_cyclecount();
13677 #endif
13678 	acked_amount = acked = (high_seq - tp->snd_una);
13679 	if (acked) {
13680 		if (rack->sack_attack_disable == 0)
13681 			rack_do_decay(rack);
13682 		if (acked >= segsiz) {
13683 			/*
13684 			 * You only get credit for
13685 			 * MSS and greater (and you get extra
13686 			 * credit for larger cum-ack moves).
13687 			 */
13688 			int ac;
13689 
13690 			ac = acked / segsiz;
13691 			rack->r_ctl.ack_count += ac;
13692 			counter_u64_add(rack_ack_total, ac);
13693 		}
13694 		if (rack->r_ctl.ack_count > 0xfff00000) {
13695 			/*
13696 			 * reduce the number to keep us under
13697 			 * a uint32_t.
13698 			 */
13699 			rack->r_ctl.ack_count /= 2;
13700 			rack->r_ctl.sack_count /= 2;
13701 		}
13702 		if (tp->t_flags & TF_NEEDSYN) {
13703 			/*
13704 			 * T/TCP: Connection was half-synchronized, and our SYN has
13705 			 * been ACK'd (so connection is now fully synchronized).  Go
13706 			 * to non-starred state, increment snd_una for ACK of SYN,
13707 			 * and check if we can do window scaling.
13708 			 */
13709 			tp->t_flags &= ~TF_NEEDSYN;
13710 			tp->snd_una++;
13711 			acked_amount = acked = (high_seq - tp->snd_una);
13712 		}
13713 		if (acked > sbavail(&so->so_snd))
13714 			acked_amount = sbavail(&so->so_snd);
13715 #ifdef NETFLIX_EXP_DETECTION
13716 		/*
13717 		 * We only care on a cum-ack move if we are in a sack-disabled
13718 		 * state. We have already added in to the ack_count, and we never
13719 		 * would disable on a cum-ack move, so we only care to do the
13720 		 * detection if it may "undo" it, i.e. we were in disabled already.
13721 		 */
13722 		if (rack->sack_attack_disable)
13723 			rack_do_detection(tp, rack, acked_amount, segsiz);
13724 #endif
13725 		if (IN_FASTRECOVERY(tp->t_flags) &&
13726 		    (rack->rack_no_prr == 0))
13727 			rack_update_prr(tp, rack, acked_amount, high_seq);
13728 		if (IN_RECOVERY(tp->t_flags)) {
13729 			if (SEQ_LT(high_seq, tp->snd_recover) &&
13730 			    (SEQ_LT(high_seq, tp->snd_max))) {
13731 				tcp_rack_partialack(tp);
13732 			} else {
13733 				rack_post_recovery(tp, high_seq);
13734 				recovery = 1;
13735 			}
13736 		}
13737 		/* Handle the rack-log-ack part (sendmap) */
13738 		if ((sbused(&so->so_snd) == 0) &&
13739 		    (acked > acked_amount) &&
13740 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
13741 		    (tp->t_flags & TF_SENTFIN)) {
13742 			/*
13743 			 * We must be sure our fin
13744 			 * was sent and acked (we can be
13745 			 * in FIN_WAIT_1 without having
13746 			 * sent the fin).
13747 			 */
13748 			ourfinisacked = 1;
13749 			/*
13750 			 * Lets make sure snd_una is updated
13751 			 * since most likely acked_amount = 0 (it
13752 			 * should be).
13753 			 */
13754 			tp->snd_una = high_seq;
13755 		}
13756 		/* Did we make a RTO error? */
13757 		if ((tp->t_flags & TF_PREVVALID) &&
13758 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13759 			tp->t_flags &= ~TF_PREVVALID;
13760 			if (tp->t_rxtshift == 1 &&
13761 			    (int)(ticks - tp->t_badrxtwin) < 0)
13762 				rack_cong_signal(tp, CC_RTO_ERR, high_seq);
13763 		}
13764 		/* Handle the data in the socket buffer */
13765 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13766 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13767 		if (acked_amount > 0) {
13768 			struct mbuf *mfree;
13769 
13770 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13771 			SOCKBUF_LOCK(&so->so_snd);
13772 			mfree = sbcut_locked(&so->so_snd, acked_amount);
13773 			tp->snd_una = high_seq;
13774 			/* Note we want to hold the sb lock through the sendmap adjust */
13775 			rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13776 			/* Wake up the socket if we have room to write more */
13777 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13778 			sowwakeup_locked(so);
13779 			m_freem(mfree);
13780 		}
13781 		/* update progress */
13782 		tp->t_acktime = ticks;
13783 		rack_log_progress_event(rack, tp, tp->t_acktime,
13784 					PROGRESS_UPDATE, __LINE__);
13785 		/* Clear out shifts and such */
13786 		tp->t_rxtshift = 0;
13787 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13788 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13789 		rack->rc_tlp_in_progress = 0;
13790 		rack->r_ctl.rc_tlp_cnt_out = 0;
13791 		/* Send recover and snd_nxt must be dragged along */
13792 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
13793 			tp->snd_recover = tp->snd_una;
13794 		if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13795 			tp->snd_nxt = tp->snd_una;
13796 		/*
13797 		 * If the RXT timer is running we want to
13798 		 * stop it, so we can restart a TLP (or new RXT).
13799 		 */
13800 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13801 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13802 #ifdef NETFLIX_HTTP_LOGGING
13803 		tcp_http_check_for_comp(rack->rc_tp, high_seq);
13804 #endif
13805 		tp->snd_wl2 = high_seq;
13806 		tp->t_dupacks = 0;
13807 		if (under_pacing &&
13808 		    (rack->use_fixed_rate == 0) &&
13809 		    (rack->in_probe_rtt == 0) &&
13810 		    rack->rc_gp_dyn_mul &&
13811 		    rack->rc_always_pace) {
13812 			/* Check if we are dragging bottom */
13813 			rack_check_bottom_drag(tp, rack, so, acked);
13814 		}
13815 		if (tp->snd_una == tp->snd_max) {
13816 			tp->t_flags &= ~TF_PREVVALID;
13817 			rack->r_ctl.retran_during_recovery = 0;
13818 			rack->r_ctl.dsack_byte_cnt = 0;
13819 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13820 			if (rack->r_ctl.rc_went_idle_time == 0)
13821 				rack->r_ctl.rc_went_idle_time = 1;
13822 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13823 			if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0)
13824 				tp->t_acktime = 0;
13825 			/* Set so we might enter persists... */
13826 			rack->r_wanted_output = 1;
13827 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13828 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13829 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13830 			    (sbavail(&so->so_snd) == 0) &&
13831 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13832 				/*
13833 				 * The socket was gone and the
13834 				 * peer sent data (not now in the past), time to
13835 				 * reset him.
13836 				 */
13837 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13838 				/* tcp_close will kill the inp pre-log the Reset */
13839 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13840 #ifdef TCP_ACCOUNTING
13841 				rdstc = get_cyclecount();
13842 				if (rdstc > ts_val) {
13843 					counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13844 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13845 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13846 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13847 					}
13848 				}
13849 #endif
13850 				m_freem(m);
13851 				tp = tcp_close(tp);
13852 				if (tp == NULL) {
13853 #ifdef TCP_ACCOUNTING
13854 					sched_unpin();
13855 #endif
13856 					return (1);
13857 				}
13858 				/*
13859 				 * We would normally do drop-with-reset which would
13860 				 * send back a reset. We can't since we don't have
13861 				 * all the needed bits. Instead lets arrange for
13862 				 * a call to tcp_output(). That way since we
13863 				 * are in the closed state we will generate a reset.
13864 				 *
13865 				 * Note if tcp_accounting is on we don't unpin since
13866 				 * we do that after the goto label.
13867 				 */
13868 				goto send_out_a_rst;
13869 			}
13870 			if ((sbused(&so->so_snd) == 0) &&
13871 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
13872 			    (tp->t_flags & TF_SENTFIN)) {
13873 				/*
13874 				 * If we can't receive any more data, then closing user can
13875 				 * proceed. Starting the timer is contrary to the
13876 				 * specification, but if we don't get a FIN we'll hang
13877 				 * forever.
13878 				 *
13879 				 */
13880 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13881 					soisdisconnected(so);
13882 					tcp_timer_activate(tp, TT_2MSL,
13883 							   (tcp_fast_finwait2_recycle ?
13884 							    tcp_finwait2_timeout :
13885 							    TP_MAXIDLE(tp)));
13886 				}
13887 				if (ourfinisacked == 0) {
13888 					/*
13889 					 * We don't change to fin-wait-2 if we have our fin acked
13890 					 * which means we are probably in TCPS_CLOSING.
13891 					 */
13892 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
13893 				}
13894 			}
13895 		}
13896 		/* Wake up the socket if we have room to write more */
13897 		if (sbavail(&so->so_snd)) {
13898 			rack->r_wanted_output = 1;
13899 			if (ctf_progress_timeout_check(tp, true)) {
13900 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13901 							tp, tick, PROGRESS_DROP, __LINE__);
13902 				tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
13903 				/*
13904 				 * We cheat here and don't send a RST, we should send one
13905 				 * when the pacer drops the connection.
13906 				 */
13907 #ifdef TCP_ACCOUNTING
13908 				rdstc = get_cyclecount();
13909 				if (rdstc > ts_val) {
13910 					counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13911 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13912 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13913 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13914 					}
13915 				}
13916 				sched_unpin();
13917 #endif
13918 				INP_WUNLOCK(rack->rc_inp);
13919 				m_freem(m);
13920 				return (1);
13921 			}
13922 		}
13923 		if (ourfinisacked) {
13924 			switch(tp->t_state) {
13925 			case TCPS_CLOSING:
13926 #ifdef TCP_ACCOUNTING
13927 				rdstc = get_cyclecount();
13928 				if (rdstc > ts_val) {
13929 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13930 							(rdstc - ts_val));
13931 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13932 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13933 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13934 					}
13935 				}
13936 				sched_unpin();
13937 #endif
13938 				tcp_twstart(tp);
13939 				m_freem(m);
13940 				return (1);
13941 				break;
13942 			case TCPS_LAST_ACK:
13943 #ifdef TCP_ACCOUNTING
13944 				rdstc = get_cyclecount();
13945 				if (rdstc > ts_val) {
13946 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13947 							(rdstc - ts_val));
13948 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13949 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13950 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13951 					}
13952 				}
13953 				sched_unpin();
13954 #endif
13955 				tp = tcp_close(tp);
13956 				ctf_do_drop(m, tp);
13957 				return (1);
13958 				break;
13959 			case TCPS_FIN_WAIT_1:
13960 #ifdef TCP_ACCOUNTING
13961 				rdstc = get_cyclecount();
13962 				if (rdstc > ts_val) {
13963 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13964 							(rdstc - ts_val));
13965 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13966 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13967 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13968 					}
13969 				}
13970 #endif
13971 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13972 					soisdisconnected(so);
13973 					tcp_timer_activate(tp, TT_2MSL,
13974 							   (tcp_fast_finwait2_recycle ?
13975 							    tcp_finwait2_timeout :
13976 							    TP_MAXIDLE(tp)));
13977 				}
13978 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
13979 				break;
13980 			default:
13981 				break;
13982 			}
13983 		}
13984 		if (rack->r_fast_output) {
13985 			/*
13986 			 * We re doing fast output.. can we expand that?
13987 			 */
13988 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
13989 		}
13990 #ifdef TCP_ACCOUNTING
13991 		rdstc = get_cyclecount();
13992 		if (rdstc > ts_val) {
13993 			counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13994 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13995 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13996 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13997 			}
13998 		}
13999 
14000 	} else if (win_up_req) {
14001 		rdstc = get_cyclecount();
14002 		if (rdstc > ts_val) {
14003 			counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
14004 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14005 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
14006 			}
14007 		}
14008 #endif
14009 	}
14010 	/* Now is there a next packet, if so we are done */
14011 	m_freem(m);
14012 	did_out = 0;
14013 	if (nxt_pkt) {
14014 #ifdef TCP_ACCOUNTING
14015 		sched_unpin();
14016 #endif
14017 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
14018 		return (0);
14019 	}
14020 	rack_handle_might_revert(tp, rack);
14021 	ctf_calc_rwin(so, tp);
14022 	if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14023 	send_out_a_rst:
14024 		(void)tp->t_fb->tfb_tcp_output(tp);
14025 		did_out = 1;
14026 	}
14027 	rack_free_trim(rack);
14028 #ifdef TCP_ACCOUNTING
14029 	sched_unpin();
14030 #endif
14031 	rack_timer_audit(tp, rack, &so->so_snd);
14032 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
14033 	return (0);
14034 }
14035 
14036 
14037 static int
14038 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
14039     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
14040     int32_t nxt_pkt, struct timeval *tv)
14041 {
14042 #ifdef TCP_ACCOUNTING
14043 	uint64_t ts_val;
14044 #endif
14045 	int32_t thflags, retval, did_out = 0;
14046 	int32_t way_out = 0;
14047 	/*
14048 	 * cts - is the current time from tv (caller gets ts) in microseconds.
14049 	 * ms_cts - is the current time from tv in milliseconds.
14050 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
14051 	 */
14052 	uint32_t cts, us_cts, ms_cts;
14053 	uint32_t tiwin;
14054 	struct timespec ts;
14055 	struct tcpopt to;
14056 	struct tcp_rack *rack;
14057 	struct rack_sendmap *rsm;
14058 	int32_t prev_state = 0;
14059 #ifdef TCP_ACCOUNTING
14060 	int ack_val_set = 0xf;
14061 #endif
14062 	int nsegs;
14063 	/*
14064 	 * tv passed from common code is from either M_TSTMP_LRO or
14065 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
14066 	 */
14067 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14068 	if (m->m_flags & M_ACKCMP) {
14069 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
14070 	}
14071 	if (m->m_flags & M_ACKCMP) {
14072 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
14073 	}
14074 	cts = tcp_tv_to_usectick(tv);
14075 	ms_cts =  tcp_tv_to_mssectick(tv);
14076 	nsegs = m->m_pkthdr.lro_nsegs;
14077 	counter_u64_add(rack_proc_non_comp_ack, 1);
14078 	thflags = th->th_flags;
14079 #ifdef TCP_ACCOUNTING
14080 	sched_pin();
14081 	if (thflags & TH_ACK)
14082 		ts_val = get_cyclecount();
14083 #endif
14084 	if ((m->m_flags & M_TSTMP) ||
14085 	    (m->m_flags & M_TSTMP_LRO)) {
14086 		mbuf_tstmp2timespec(m, &ts);
14087 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
14088 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
14089 	} else
14090 		rack->r_ctl.act_rcv_time = *tv;
14091 	kern_prefetch(rack, &prev_state);
14092 	prev_state = 0;
14093 	/*
14094 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
14095 	 * the scale is zero.
14096 	 */
14097 	tiwin = th->th_win << tp->snd_scale;
14098 #ifdef TCP_ACCOUNTING
14099 	if (thflags & TH_ACK) {
14100 		/*
14101 		 * We have a tradeoff here. We can either do what we are
14102 		 * doing i.e. pinning to this CPU and then doing the accounting
14103 		 * <or> we could do a critical enter, setup the rdtsc and cpu
14104 		 * as in below, and then validate we are on the same CPU on
14105 		 * exit. I have choosen to not do the critical enter since
14106 		 * that often will gain you a context switch, and instead lock
14107 		 * us (line above this if) to the same CPU with sched_pin(). This
14108 		 * means we may be context switched out for a higher priority
14109 		 * interupt but we won't be moved to another CPU.
14110 		 *
14111 		 * If this occurs (which it won't very often since we most likely
14112 		 * are running this code in interupt context and only a higher
14113 		 * priority will bump us ... clock?) we will falsely add in
14114 		 * to the time the interupt processing time plus the ack processing
14115 		 * time. This is ok since its a rare event.
14116 		 */
14117 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
14118 						    ctf_fixed_maxseg(tp));
14119 	}
14120 #endif
14121 	/*
14122 	 * Parse options on any incoming segment.
14123 	 */
14124 	memset(&to, 0, sizeof(to));
14125 	tcp_dooptions(&to, (u_char *)(th + 1),
14126 	    (th->th_off << 2) - sizeof(struct tcphdr),
14127 	    (thflags & TH_SYN) ? TO_SYN : 0);
14128 	NET_EPOCH_ASSERT();
14129 	INP_WLOCK_ASSERT(tp->t_inpcb);
14130 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
14131 	    __func__));
14132 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
14133 	    __func__));
14134 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
14135 	    (tp->t_flags & TF_GPUTINPROG)) {
14136 		/*
14137 		 * We have a goodput in progress
14138 		 * and we have entered a late state.
14139 		 * Do we have enough data in the sb
14140 		 * to handle the GPUT request?
14141 		 */
14142 		uint32_t bytes;
14143 
14144 		bytes = tp->gput_ack - tp->gput_seq;
14145 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
14146 			bytes += tp->gput_seq - tp->snd_una;
14147 		if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) {
14148 			/*
14149 			 * There are not enough bytes in the socket
14150 			 * buffer that have been sent to cover this
14151 			 * measurement. Cancel it.
14152 			 */
14153 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
14154 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
14155 						   tp->gput_seq,
14156 						   0, 0, 18, __LINE__, NULL, 0);
14157 			tp->t_flags &= ~TF_GPUTINPROG;
14158 		}
14159 	}
14160 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14161 		union tcp_log_stackspecific log;
14162 		struct timeval ltv;
14163 #ifdef NETFLIX_HTTP_LOGGING
14164 		struct http_sendfile_track *http_req;
14165 
14166 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
14167 			http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
14168 		} else {
14169 			http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
14170 		}
14171 #endif
14172 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14173 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
14174 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
14175 		if (rack->rack_no_prr == 0)
14176 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
14177 		else
14178 			log.u_bbr.flex1 = 0;
14179 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
14180 		log.u_bbr.use_lt_bw <<= 1;
14181 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
14182 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
14183 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14184 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
14185 		log.u_bbr.flex3 = m->m_flags;
14186 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
14187 		log.u_bbr.lost = thflags;
14188 		log.u_bbr.pacing_gain = 0x1;
14189 #ifdef TCP_ACCOUNTING
14190 		log.u_bbr.cwnd_gain = ack_val_set;
14191 #endif
14192 		log.u_bbr.flex7 = 2;
14193 		if (m->m_flags & M_TSTMP) {
14194 			/* Record the hardware timestamp if present */
14195 			mbuf_tstmp2timespec(m, &ts);
14196 			ltv.tv_sec = ts.tv_sec;
14197 			ltv.tv_usec = ts.tv_nsec / 1000;
14198 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
14199 		} else if (m->m_flags & M_TSTMP_LRO) {
14200 			/* Record the LRO the arrival timestamp */
14201 			mbuf_tstmp2timespec(m, &ts);
14202 			ltv.tv_sec = ts.tv_sec;
14203 			ltv.tv_usec = ts.tv_nsec / 1000;
14204 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
14205 		}
14206 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
14207 		/* Log the rcv time */
14208 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
14209 #ifdef NETFLIX_HTTP_LOGGING
14210 		log.u_bbr.applimited = tp->t_http_closed;
14211 		log.u_bbr.applimited <<= 8;
14212 		log.u_bbr.applimited |= tp->t_http_open;
14213 		log.u_bbr.applimited <<= 8;
14214 		log.u_bbr.applimited |= tp->t_http_req;
14215 		if (http_req) {
14216 			/* Copy out any client req info */
14217 			/* seconds */
14218 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
14219 			/* useconds */
14220 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
14221 			log.u_bbr.rttProp = http_req->timestamp;
14222 			log.u_bbr.cur_del_rate = http_req->start;
14223 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
14224 				log.u_bbr.flex8 |= 1;
14225 			} else {
14226 				log.u_bbr.flex8 |= 2;
14227 				log.u_bbr.bw_inuse = http_req->end;
14228 			}
14229 			log.u_bbr.flex6 = http_req->start_seq;
14230 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
14231 				log.u_bbr.flex8 |= 4;
14232 				log.u_bbr.epoch = http_req->end_seq;
14233 			}
14234 		}
14235 #endif
14236 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
14237 		    tlen, &log, true, &ltv);
14238 	}
14239 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
14240 		way_out = 4;
14241 		retval = 0;
14242 		m_freem(m);
14243 		goto done_with_input;
14244 	}
14245 	/*
14246 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
14247 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
14248 	 */
14249 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
14250 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
14251 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14252 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14253 #ifdef TCP_ACCOUNTING
14254 		sched_unpin();
14255 #endif
14256 		return (1);
14257 	}
14258 	/*
14259 	 * If timestamps were negotiated during SYN/ACK and a
14260 	 * segment without a timestamp is received, silently drop
14261 	 * the segment, unless it is a RST segment or missing timestamps are
14262 	 * tolerated.
14263 	 * See section 3.2 of RFC 7323.
14264 	 */
14265 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
14266 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
14267 		way_out = 5;
14268 		retval = 0;
14269 		m_freem(m);
14270 		goto done_with_input;
14271 	}
14272 
14273 	/*
14274 	 * Segment received on connection. Reset idle time and keep-alive
14275 	 * timer. XXX: This should be done after segment validation to
14276 	 * ignore broken/spoofed segs.
14277 	 */
14278 	if  (tp->t_idle_reduce &&
14279 	     (tp->snd_max == tp->snd_una) &&
14280 	     ((ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
14281 		counter_u64_add(rack_input_idle_reduces, 1);
14282 		rack_cc_after_idle(rack, tp);
14283 	}
14284 	tp->t_rcvtime = ticks;
14285 #ifdef STATS
14286 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
14287 #endif
14288 	if (tiwin > rack->r_ctl.rc_high_rwnd)
14289 		rack->r_ctl.rc_high_rwnd = tiwin;
14290 	/*
14291 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
14292 	 * this to occur after we've validated the segment.
14293 	 */
14294 	if (tp->t_flags2 & TF2_ECN_PERMIT) {
14295 		if (thflags & TH_CWR) {
14296 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
14297 			tp->t_flags |= TF_ACKNOW;
14298 		}
14299 		switch (iptos & IPTOS_ECN_MASK) {
14300 		case IPTOS_ECN_CE:
14301 			tp->t_flags2 |= TF2_ECN_SND_ECE;
14302 			KMOD_TCPSTAT_INC(tcps_ecn_ce);
14303 			break;
14304 		case IPTOS_ECN_ECT0:
14305 			KMOD_TCPSTAT_INC(tcps_ecn_ect0);
14306 			break;
14307 		case IPTOS_ECN_ECT1:
14308 			KMOD_TCPSTAT_INC(tcps_ecn_ect1);
14309 			break;
14310 		}
14311 
14312 		/* Process a packet differently from RFC3168. */
14313 		cc_ecnpkt_handler(tp, th, iptos);
14314 
14315 		/* Congestion experienced. */
14316 		if (thflags & TH_ECE) {
14317 			rack_cong_signal(tp, CC_ECN, th->th_ack);
14318 		}
14319 	}
14320 
14321 	/*
14322 	 * If echoed timestamp is later than the current time, fall back to
14323 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
14324 	 * were used when this connection was established.
14325 	 */
14326 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
14327 		to.to_tsecr -= tp->ts_offset;
14328 		if (TSTMP_GT(to.to_tsecr, ms_cts))
14329 			to.to_tsecr = 0;
14330 	}
14331 
14332 	/*
14333 	 * If its the first time in we need to take care of options and
14334 	 * verify we can do SACK for rack!
14335 	 */
14336 	if (rack->r_state == 0) {
14337 		/* Should be init'd by rack_init() */
14338 		KASSERT(rack->rc_inp != NULL,
14339 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
14340 		if (rack->rc_inp == NULL) {
14341 			rack->rc_inp = tp->t_inpcb;
14342 		}
14343 
14344 		/*
14345 		 * Process options only when we get SYN/ACK back. The SYN
14346 		 * case for incoming connections is handled in tcp_syncache.
14347 		 * According to RFC1323 the window field in a SYN (i.e., a
14348 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
14349 		 * this is traditional behavior, may need to be cleaned up.
14350 		 */
14351 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
14352 			/* Handle parallel SYN for ECN */
14353 			if (!(thflags & TH_ACK) &&
14354 			    ((thflags & (TH_CWR | TH_ECE)) == (TH_CWR | TH_ECE)) &&
14355 			    ((V_tcp_do_ecn == 1) || (V_tcp_do_ecn == 2))) {
14356 				tp->t_flags2 |= TF2_ECN_PERMIT;
14357 				tp->t_flags2 |= TF2_ECN_SND_ECE;
14358 				TCPSTAT_INC(tcps_ecn_shs);
14359 			}
14360 			if ((to.to_flags & TOF_SCALE) &&
14361 			    (tp->t_flags & TF_REQ_SCALE)) {
14362 				tp->t_flags |= TF_RCVD_SCALE;
14363 				tp->snd_scale = to.to_wscale;
14364 			} else
14365 				tp->t_flags &= ~TF_REQ_SCALE;
14366 			/*
14367 			 * Initial send window.  It will be updated with the
14368 			 * next incoming segment to the scaled value.
14369 			 */
14370 			tp->snd_wnd = th->th_win;
14371 			rack_validate_fo_sendwin_up(tp, rack);
14372 			if ((to.to_flags & TOF_TS) &&
14373 			    (tp->t_flags & TF_REQ_TSTMP)) {
14374 				tp->t_flags |= TF_RCVD_TSTMP;
14375 				tp->ts_recent = to.to_tsval;
14376 				tp->ts_recent_age = cts;
14377 			} else
14378 				tp->t_flags &= ~TF_REQ_TSTMP;
14379 			if (to.to_flags & TOF_MSS) {
14380 				tcp_mss(tp, to.to_mss);
14381 			}
14382 			if ((tp->t_flags & TF_SACK_PERMIT) &&
14383 			    (to.to_flags & TOF_SACKPERM) == 0)
14384 				tp->t_flags &= ~TF_SACK_PERMIT;
14385 			if (IS_FASTOPEN(tp->t_flags)) {
14386 				if (to.to_flags & TOF_FASTOPEN) {
14387 					uint16_t mss;
14388 
14389 					if (to.to_flags & TOF_MSS)
14390 						mss = to.to_mss;
14391 					else
14392 						if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
14393 							mss = TCP6_MSS;
14394 						else
14395 							mss = TCP_MSS;
14396 					tcp_fastopen_update_cache(tp, mss,
14397 					    to.to_tfo_len, to.to_tfo_cookie);
14398 				} else
14399 					tcp_fastopen_disable_path(tp);
14400 			}
14401 		}
14402 		/*
14403 		 * At this point we are at the initial call. Here we decide
14404 		 * if we are doing RACK or not. We do this by seeing if
14405 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
14406 		 * The code now does do dup-ack counting so if you don't
14407 		 * switch back you won't get rack & TLP, but you will still
14408 		 * get this stack.
14409 		 */
14410 
14411 		if ((rack_sack_not_required == 0) &&
14412 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
14413 			tcp_switch_back_to_default(tp);
14414 			(*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
14415 			    tlen, iptos);
14416 #ifdef TCP_ACCOUNTING
14417 			sched_unpin();
14418 #endif
14419 			return (1);
14420 		}
14421 		tcp_set_hpts(tp->t_inpcb);
14422 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
14423 	}
14424 	if (thflags & TH_FIN)
14425 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
14426 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14427 	if ((rack->rc_gp_dyn_mul) &&
14428 	    (rack->use_fixed_rate == 0) &&
14429 	    (rack->rc_always_pace)) {
14430 		/* Check in on probertt */
14431 		rack_check_probe_rtt(rack, us_cts);
14432 	}
14433 	rack_clear_rate_sample(rack);
14434 	if (rack->forced_ack) {
14435 		uint32_t us_rtt;
14436 
14437 		/*
14438 		 * A persist or keep-alive was forced out, update our
14439 		 * min rtt time. Note we do not worry about lost
14440 		 * retransmissions since KEEP-ALIVES and persists
14441 		 * are usually way long on times of sending (though
14442 		 * if we were really paranoid or worried we could
14443 		 * at least use timestamps if available to validate).
14444 		 */
14445 		rack->forced_ack = 0;
14446 		if (tiwin == tp->snd_wnd) {
14447 			/*
14448 			 * Only apply the RTT update if this is
14449 			 * a response to our window probe. And that
14450 			 * means the rwnd sent must match the current
14451 			 * snd_wnd. If it does not, then we got a
14452 			 * window update ack instead.
14453 			 */
14454 			us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
14455 			if (us_rtt == 0)
14456 				us_rtt = 1;
14457 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
14458 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
14459 		}
14460 	}
14461 	/*
14462 	 * This is the one exception case where we set the rack state
14463 	 * always. All other times (timers etc) we must have a rack-state
14464 	 * set (so we assure we have done the checks above for SACK).
14465 	 */
14466 	rack->r_ctl.rc_rcvtime = cts;
14467 	if (rack->r_state != tp->t_state)
14468 		rack_set_state(tp, rack);
14469 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
14470 	    (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
14471 		kern_prefetch(rsm, &prev_state);
14472 	prev_state = rack->r_state;
14473 	retval = (*rack->r_substate) (m, th, so,
14474 	    tp, &to, drop_hdrlen,
14475 	    tlen, tiwin, thflags, nxt_pkt, iptos);
14476 #ifdef INVARIANTS
14477 	if ((retval == 0) &&
14478 	    (tp->t_inpcb == NULL)) {
14479 		panic("retval:%d tp:%p t_inpcb:NULL state:%d",
14480 		    retval, tp, prev_state);
14481 	}
14482 #endif
14483 	if (retval == 0) {
14484 		/*
14485 		 * If retval is 1 the tcb is unlocked and most likely the tp
14486 		 * is gone.
14487 		 */
14488 		INP_WLOCK_ASSERT(tp->t_inpcb);
14489 		if ((rack->rc_gp_dyn_mul) &&
14490 		    (rack->rc_always_pace) &&
14491 		    (rack->use_fixed_rate == 0) &&
14492 		    rack->in_probe_rtt &&
14493 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
14494 			/*
14495 			 * If we are going for target, lets recheck before
14496 			 * we output.
14497 			 */
14498 			rack_check_probe_rtt(rack, us_cts);
14499 		}
14500 		if (rack->set_pacing_done_a_iw == 0) {
14501 			/* How much has been acked? */
14502 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
14503 				/* We have enough to set in the pacing segment size */
14504 				rack->set_pacing_done_a_iw = 1;
14505 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
14506 			}
14507 		}
14508 		tcp_rack_xmit_timer_commit(rack, tp);
14509 #ifdef TCP_ACCOUNTING
14510 		/*
14511 		 * If we set the ack_val_se to what ack processing we are doing
14512 		 * we also want to track how many cycles we burned. Note
14513 		 * the bits after tcp_output we let be "free". This is because
14514 		 * we are also tracking the tcp_output times as well. Note the
14515 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
14516 		 * 0xf cannot be returned and is what we initialize it too to
14517 		 * indicate we are not doing the tabulations.
14518 		 */
14519 		if (ack_val_set != 0xf) {
14520 			uint64_t crtsc;
14521 
14522 			crtsc = get_cyclecount();
14523 			counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14524 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14525 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
14526 			}
14527 		}
14528 #endif
14529 		if (nxt_pkt == 0) {
14530 			if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14531 do_output_now:
14532 				did_out = 1;
14533 				(void)tp->t_fb->tfb_tcp_output(tp);
14534 			}
14535 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
14536 			rack_free_trim(rack);
14537 		}
14538 		/* Update any rounds needed */
14539 		if (SEQ_GEQ(tp->snd_una, rack->r_ctl.roundends)) {
14540 			rack->r_ctl.current_round++;
14541 			rack->r_ctl.roundends = tp->snd_max;
14542 			if (CC_ALGO(tp)->newround != NULL) {
14543 				CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round);
14544 			}
14545 		}
14546 		if ((nxt_pkt == 0) &&
14547 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
14548 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
14549 		     (tp->t_flags & TF_DELACK) ||
14550 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
14551 		      (tp->t_state <= TCPS_CLOSING)))) {
14552 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
14553 			if ((tp->snd_max == tp->snd_una) &&
14554 			    ((tp->t_flags & TF_DELACK) == 0) &&
14555 			    (rack->rc_inp->inp_in_hpts) &&
14556 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
14557 				/* keep alive not needed if we are hptsi output yet */
14558 				;
14559 			} else {
14560 				int late = 0;
14561 				if (rack->rc_inp->inp_in_hpts) {
14562 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14563 						us_cts = tcp_get_usecs(NULL);
14564 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
14565 							rack->r_early = 1;
14566 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
14567 						} else
14568 							late = 1;
14569 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
14570 					}
14571 					tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
14572 				}
14573 				if (late && (did_out == 0)) {
14574 					/*
14575 					 * We are late in the sending
14576 					 * and we did not call the output
14577 					 * (this probably should not happen).
14578 					 */
14579 					goto do_output_now;
14580 				}
14581 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14582 			}
14583 			way_out = 1;
14584 		} else if (nxt_pkt == 0) {
14585 			/* Do we have the correct timer running? */
14586 			rack_timer_audit(tp, rack, &so->so_snd);
14587 			way_out = 2;
14588 		}
14589 	done_with_input:
14590 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
14591 		if (did_out)
14592 			rack->r_wanted_output = 0;
14593 #ifdef INVARIANTS
14594 		if (tp->t_inpcb == NULL) {
14595 			panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
14596 			      did_out,
14597 			      retval, tp, prev_state);
14598 		}
14599 #endif
14600 #ifdef TCP_ACCOUNTING
14601 	} else {
14602 		/*
14603 		 * Track the time (see above).
14604 		 */
14605 		if (ack_val_set != 0xf) {
14606 			uint64_t crtsc;
14607 
14608 			crtsc = get_cyclecount();
14609 			counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14610 			/*
14611 			 * Note we *DO NOT* increment the per-tcb counters since
14612 			 * in the else the TP may be gone!!
14613 			 */
14614 		}
14615 #endif
14616 	}
14617 #ifdef TCP_ACCOUNTING
14618 	sched_unpin();
14619 #endif
14620 	return (retval);
14621 }
14622 
14623 void
14624 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
14625     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
14626 {
14627 	struct timeval tv;
14628 
14629 	/* First lets see if we have old packets */
14630 	if (tp->t_in_pkt) {
14631 		if (ctf_do_queued_segments(so, tp, 1)) {
14632 			m_freem(m);
14633 			return;
14634 		}
14635 	}
14636 	if (m->m_flags & M_TSTMP_LRO) {
14637 		tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000;
14638 		tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000;
14639 	} else {
14640 		/* Should not be should we kassert instead? */
14641 		tcp_get_usecs(&tv);
14642 	}
14643 	if (rack_do_segment_nounlock(m, th, so, tp,
14644 				     drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14645 		INP_WUNLOCK(tp->t_inpcb);
14646 	}
14647 }
14648 
14649 struct rack_sendmap *
14650 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14651 {
14652 	struct rack_sendmap *rsm = NULL;
14653 	int32_t idx;
14654 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
14655 
14656 	/* Return the next guy to be re-transmitted */
14657 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14658 		return (NULL);
14659 	}
14660 	if (tp->t_flags & TF_SENTFIN) {
14661 		/* retran the end FIN? */
14662 		return (NULL);
14663 	}
14664 	/* ok lets look at this one */
14665 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14666 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14667 		goto check_it;
14668 	}
14669 	rsm = rack_find_lowest_rsm(rack);
14670 	if (rsm == NULL) {
14671 		return (NULL);
14672 	}
14673 check_it:
14674 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14675 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14676 		/*
14677 		 * No sack so we automatically do the 3 strikes and
14678 		 * retransmit (no rack timer would be started).
14679 		 */
14680 
14681 		return (rsm);
14682 	}
14683 	if (rsm->r_flags & RACK_ACKED) {
14684 		return (NULL);
14685 	}
14686 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14687 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14688 		/* Its not yet ready */
14689 		return (NULL);
14690 	}
14691 	srtt = rack_grab_rtt(tp, rack);
14692 	idx = rsm->r_rtr_cnt - 1;
14693 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14694 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14695 	if ((tsused == ts_low) ||
14696 	    (TSTMP_LT(tsused, ts_low))) {
14697 		/* No time since sending */
14698 		return (NULL);
14699 	}
14700 	if ((tsused - ts_low) < thresh) {
14701 		/* It has not been long enough yet */
14702 		return (NULL);
14703 	}
14704 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14705 	    ((rsm->r_flags & RACK_SACK_PASSED) &&
14706 	     (rack->sack_attack_disable == 0))) {
14707 		/*
14708 		 * We have passed the dup-ack threshold <or>
14709 		 * a SACK has indicated this is missing.
14710 		 * Note that if you are a declared attacker
14711 		 * it is only the dup-ack threshold that
14712 		 * will cause retransmits.
14713 		 */
14714 		/* log retransmit reason */
14715 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14716 		rack->r_fast_output = 0;
14717 		return (rsm);
14718 	}
14719 	return (NULL);
14720 }
14721 
14722 static void
14723 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14724 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14725 			   int line, struct rack_sendmap *rsm, uint8_t quality)
14726 {
14727 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14728 		union tcp_log_stackspecific log;
14729 		struct timeval tv;
14730 
14731 		memset(&log, 0, sizeof(log));
14732 		log.u_bbr.flex1 = slot;
14733 		log.u_bbr.flex2 = len;
14734 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14735 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14736 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14737 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14738 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14739 		log.u_bbr.use_lt_bw <<= 1;
14740 		log.u_bbr.use_lt_bw |= rack->r_late;
14741 		log.u_bbr.use_lt_bw <<= 1;
14742 		log.u_bbr.use_lt_bw |= rack->r_early;
14743 		log.u_bbr.use_lt_bw <<= 1;
14744 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14745 		log.u_bbr.use_lt_bw <<= 1;
14746 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14747 		log.u_bbr.use_lt_bw <<= 1;
14748 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14749 		log.u_bbr.use_lt_bw <<= 1;
14750 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14751 		log.u_bbr.use_lt_bw <<= 1;
14752 		log.u_bbr.use_lt_bw |= rack->gp_ready;
14753 		log.u_bbr.pkt_epoch = line;
14754 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14755 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14756 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14757 		log.u_bbr.bw_inuse = bw_est;
14758 		log.u_bbr.delRate = bw;
14759 		if (rack->r_ctl.gp_bw == 0)
14760 			log.u_bbr.cur_del_rate = 0;
14761 		else
14762 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
14763 		log.u_bbr.rttProp = len_time;
14764 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14765 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14766 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14767 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14768 			/* We are in slow start */
14769 			log.u_bbr.flex7 = 1;
14770 		} else {
14771 			/* we are on congestion avoidance */
14772 			log.u_bbr.flex7 = 0;
14773 		}
14774 		log.u_bbr.flex8 = method;
14775 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14776 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14777 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14778 		log.u_bbr.cwnd_gain <<= 1;
14779 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14780 		log.u_bbr.cwnd_gain <<= 1;
14781 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14782 		log.u_bbr.bbr_substate = quality;
14783 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
14784 		    &rack->rc_inp->inp_socket->so_rcv,
14785 		    &rack->rc_inp->inp_socket->so_snd,
14786 		    BBR_LOG_HPTSI_CALC, 0,
14787 		    0, &log, false, &tv);
14788 	}
14789 }
14790 
14791 static uint32_t
14792 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14793 {
14794 	uint32_t new_tso, user_max;
14795 
14796 	user_max = rack->rc_user_set_max_segs * mss;
14797 	if (rack->rc_force_max_seg) {
14798 		return (user_max);
14799 	}
14800 	if (rack->use_fixed_rate &&
14801 	    ((rack->r_ctl.crte == NULL) ||
14802 	     (bw != rack->r_ctl.crte->rate))) {
14803 		/* Use the user mss since we are not exactly matched */
14804 		return (user_max);
14805 	}
14806 	new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14807 	if (new_tso > user_max)
14808 		new_tso = user_max;
14809 	return (new_tso);
14810 }
14811 
14812 static int32_t
14813 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)
14814 {
14815 	uint64_t lentim, fill_bw;
14816 
14817 	/* Lets first see if we are full, if so continue with normal rate */
14818 	rack->r_via_fill_cw = 0;
14819 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14820 		return (slot);
14821 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14822 		return (slot);
14823 	if (rack->r_ctl.rc_last_us_rtt == 0)
14824 		return (slot);
14825 	if (rack->rc_pace_fill_if_rttin_range &&
14826 	    (rack->r_ctl.rc_last_us_rtt >=
14827 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14828 		/* The rtt is huge, N * smallest, lets not fill */
14829 		return (slot);
14830 	}
14831 	/*
14832 	 * first lets calculate the b/w based on the last us-rtt
14833 	 * and the sndwnd.
14834 	 */
14835 	fill_bw = rack->r_ctl.cwnd_to_use;
14836 	/* Take the rwnd if its smaller */
14837 	if (fill_bw > rack->rc_tp->snd_wnd)
14838 		fill_bw = rack->rc_tp->snd_wnd;
14839 	if (rack->r_fill_less_agg) {
14840 		/*
14841 		 * Now take away the inflight (this will reduce our
14842 		 * aggressiveness and yeah, if we get that much out in 1RTT
14843 		 * we will have had acks come back and still be behind).
14844 		 */
14845 		fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14846 	}
14847 	/* Now lets make it into a b/w */
14848 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14849 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14850 	/* We are below the min b/w */
14851 	if (non_paced)
14852 		*rate_wanted = fill_bw;
14853 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14854 		return (slot);
14855 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14856 		fill_bw = rack->r_ctl.bw_rate_cap;
14857 	rack->r_via_fill_cw = 1;
14858 	if (rack->r_rack_hw_rate_caps &&
14859 	    (rack->r_ctl.crte != NULL)) {
14860 		uint64_t high_rate;
14861 
14862 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14863 		if (fill_bw > high_rate) {
14864 			/* We are capping bw at the highest rate table entry */
14865 			if (*rate_wanted > high_rate) {
14866 				/* The original rate was also capped */
14867 				rack->r_via_fill_cw = 0;
14868 			}
14869 			rack_log_hdwr_pacing(rack,
14870 					     fill_bw, high_rate, __LINE__,
14871 					     0, 3);
14872 			fill_bw = high_rate;
14873 			if (capped)
14874 				*capped = 1;
14875 		}
14876 	} else if ((rack->r_ctl.crte == NULL) &&
14877 		   (rack->rack_hdrw_pacing == 0) &&
14878 		   (rack->rack_hdw_pace_ena) &&
14879 		   rack->r_rack_hw_rate_caps &&
14880 		   (rack->rack_attempt_hdwr_pace == 0) &&
14881 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
14882 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14883 		/*
14884 		 * Ok we may have a first attempt that is greater than our top rate
14885 		 * lets check.
14886 		 */
14887 		uint64_t high_rate;
14888 
14889 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14890 		if (high_rate) {
14891 			if (fill_bw > high_rate) {
14892 				fill_bw = high_rate;
14893 				if (capped)
14894 					*capped = 1;
14895 			}
14896 		}
14897 	}
14898 	/*
14899 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
14900 	 * in a rtt, what does that time wise equate too?
14901 	 */
14902 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14903 	lentim /= fill_bw;
14904 	*rate_wanted = fill_bw;
14905 	if (non_paced || (lentim < slot)) {
14906 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14907 					   0, lentim, 12, __LINE__, NULL, 0);
14908 		return ((int32_t)lentim);
14909 	} else
14910 		return (slot);
14911 }
14912 
14913 static int32_t
14914 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14915 {
14916 	uint64_t srtt;
14917 	int32_t slot = 0;
14918 	int can_start_hw_pacing = 1;
14919 	int err;
14920 
14921 	if (rack->rc_always_pace == 0) {
14922 		/*
14923 		 * We use the most optimistic possible cwnd/srtt for
14924 		 * sending calculations. This will make our
14925 		 * calculation anticipate getting more through
14926 		 * quicker then possible. But thats ok we don't want
14927 		 * the peer to have a gap in data sending.
14928 		 */
14929 		uint64_t cwnd, tr_perms = 0;
14930 		int32_t reduce = 0;
14931 
14932 	old_method:
14933 		/*
14934 		 * We keep no precise pacing with the old method
14935 		 * instead we use the pacer to mitigate bursts.
14936 		 */
14937 		if (rack->r_ctl.rc_rack_min_rtt)
14938 			srtt = rack->r_ctl.rc_rack_min_rtt;
14939 		else
14940 			srtt = max(tp->t_srtt, 1);
14941 		if (rack->r_ctl.rc_rack_largest_cwnd)
14942 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14943 		else
14944 			cwnd = rack->r_ctl.cwnd_to_use;
14945 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
14946 		tr_perms = (cwnd * 1000) / srtt;
14947 		if (tr_perms == 0) {
14948 			tr_perms = ctf_fixed_maxseg(tp);
14949 		}
14950 		/*
14951 		 * Calculate how long this will take to drain, if
14952 		 * the calculation comes out to zero, thats ok we
14953 		 * will use send_a_lot to possibly spin around for
14954 		 * more increasing tot_len_this_send to the point
14955 		 * that its going to require a pace, or we hit the
14956 		 * cwnd. Which in that case we are just waiting for
14957 		 * a ACK.
14958 		 */
14959 		slot = len / tr_perms;
14960 		/* Now do we reduce the time so we don't run dry? */
14961 		if (slot && rack_slot_reduction) {
14962 			reduce = (slot / rack_slot_reduction);
14963 			if (reduce < slot) {
14964 				slot -= reduce;
14965 			} else
14966 				slot = 0;
14967 		}
14968 		slot *= HPTS_USEC_IN_MSEC;
14969 		if (rack->rc_pace_to_cwnd) {
14970 			uint64_t rate_wanted = 0;
14971 
14972 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14973 			rack->rc_ack_can_sendout_data = 1;
14974 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
14975 		} else
14976 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
14977 	} else {
14978 		uint64_t bw_est, res, lentim, rate_wanted;
14979 		uint32_t orig_val, segs, oh;
14980 		int capped = 0;
14981 		int prev_fill;
14982 
14983 		if ((rack->r_rr_config == 1) && rsm) {
14984 			return (rack->r_ctl.rc_min_to);
14985 		}
14986 		if (rack->use_fixed_rate) {
14987 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14988 		} else if ((rack->r_ctl.init_rate == 0) &&
14989 #ifdef NETFLIX_PEAKRATE
14990 			   (rack->rc_tp->t_maxpeakrate == 0) &&
14991 #endif
14992 			   (rack->r_ctl.gp_bw == 0)) {
14993 			/* no way to yet do an estimate */
14994 			bw_est = rate_wanted = 0;
14995 		} else {
14996 			bw_est = rack_get_bw(rack);
14997 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14998 		}
14999 		if ((bw_est == 0) || (rate_wanted == 0) ||
15000 		    ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
15001 			/*
15002 			 * No way yet to make a b/w estimate or
15003 			 * our raise is set incorrectly.
15004 			 */
15005 			goto old_method;
15006 		}
15007 		/* We need to account for all the overheads */
15008 		segs = (len + segsiz - 1) / segsiz;
15009 		/*
15010 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
15011 		 * and how much data we put in each packet. Yes this
15012 		 * means we may be off if we are larger than 1500 bytes
15013 		 * or smaller. But this just makes us more conservative.
15014 		 */
15015 		if (rack_hw_rate_min &&
15016 		    (bw_est < rack_hw_rate_min))
15017 			can_start_hw_pacing = 0;
15018 		if (ETHERNET_SEGMENT_SIZE > segsiz)
15019 			oh = ETHERNET_SEGMENT_SIZE - segsiz;
15020 		else
15021 			oh = 0;
15022 		segs *= oh;
15023 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
15024 		res = lentim / rate_wanted;
15025 		slot = (uint32_t)res;
15026 		orig_val = rack->r_ctl.rc_pace_max_segs;
15027 		if (rack->r_ctl.crte == NULL) {
15028 			/*
15029 			 * Only do this if we are not hardware pacing
15030 			 * since if we are doing hw-pacing below we will
15031 			 * set make a call after setting up or changing
15032 			 * the rate.
15033 			 */
15034 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
15035 		} else if (rack->rc_inp->inp_snd_tag == NULL) {
15036 			/*
15037 			 * We lost our rate somehow, this can happen
15038 			 * if the interface changed underneath us.
15039 			 */
15040 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
15041 			rack->r_ctl.crte = NULL;
15042 			/* Lets re-allow attempting to setup pacing */
15043 			rack->rack_hdrw_pacing = 0;
15044 			rack->rack_attempt_hdwr_pace = 0;
15045 			rack_log_hdwr_pacing(rack,
15046 					     rate_wanted, bw_est, __LINE__,
15047 					     0, 6);
15048 		}
15049 		/* Did we change the TSO size, if so log it */
15050 		if (rack->r_ctl.rc_pace_max_segs != orig_val)
15051 			rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL, 0);
15052 		prev_fill = rack->r_via_fill_cw;
15053 		if ((rack->rc_pace_to_cwnd) &&
15054 		    (capped == 0) &&
15055 		    (rack->use_fixed_rate == 0) &&
15056 		    (rack->in_probe_rtt == 0) &&
15057 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
15058 			/*
15059 			 * We want to pace at our rate *or* faster to
15060 			 * fill the cwnd to the max if its not full.
15061 			 */
15062 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
15063 		}
15064 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
15065 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
15066 			if ((rack->rack_hdw_pace_ena) &&
15067 			    (can_start_hw_pacing > 0) &&
15068 			    (rack->rack_hdrw_pacing == 0) &&
15069 			    (rack->rack_attempt_hdwr_pace == 0)) {
15070 				/*
15071 				 * Lets attempt to turn on hardware pacing
15072 				 * if we can.
15073 				 */
15074 				rack->rack_attempt_hdwr_pace = 1;
15075 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
15076 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
15077 								       rate_wanted,
15078 								       RS_PACING_GEQ,
15079 								       &err, &rack->r_ctl.crte_prev_rate);
15080 				if (rack->r_ctl.crte) {
15081 					rack->rack_hdrw_pacing = 1;
15082 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
15083 												 0, rack->r_ctl.crte,
15084 												 NULL);
15085 					rack_log_hdwr_pacing(rack,
15086 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15087 							     err, 0);
15088 					rack->r_ctl.last_hw_bw_req = rate_wanted;
15089 				} else {
15090 					counter_u64_add(rack_hw_pace_init_fail, 1);
15091 				}
15092 			} else if (rack->rack_hdrw_pacing &&
15093 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
15094 				/* Do we need to adjust our rate? */
15095 				const struct tcp_hwrate_limit_table *nrte;
15096 
15097 				if (rack->r_up_only &&
15098 				    (rate_wanted < rack->r_ctl.crte->rate)) {
15099 					/**
15100 					 * We have four possible states here
15101 					 * having to do with the previous time
15102 					 * and this time.
15103 					 *   previous  |  this-time
15104 					 * A)     0      |     0   -- fill_cw not in the picture
15105 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
15106 					 * C)     1      |     1   -- all rates from fill_cw
15107 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
15108 					 *
15109 					 * For case A, C and D we don't allow a drop. But for
15110 					 * case B where we now our on our steady rate we do
15111 					 * allow a drop.
15112 					 *
15113 					 */
15114 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
15115 						goto done_w_hdwr;
15116 				}
15117 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
15118 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
15119 					if (rack_hw_rate_to_low &&
15120 					    (bw_est < rack_hw_rate_to_low)) {
15121 						/*
15122 						 * The pacing rate is too low for hardware, but
15123 						 * do allow hardware pacing to be restarted.
15124 						 */
15125 						rack_log_hdwr_pacing(rack,
15126 							     bw_est, rack->r_ctl.crte->rate, __LINE__,
15127 							     0, 5);
15128 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
15129 						rack->r_ctl.crte = NULL;
15130 						rack->rack_attempt_hdwr_pace = 0;
15131 						rack->rack_hdrw_pacing = 0;
15132 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15133 						goto done_w_hdwr;
15134 					}
15135 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
15136 								   rack->rc_tp,
15137 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
15138 								   rate_wanted,
15139 								   RS_PACING_GEQ,
15140 								   &err, &rack->r_ctl.crte_prev_rate);
15141 					if (nrte == NULL) {
15142 						/* Lost the rate */
15143 						rack->rack_hdrw_pacing = 0;
15144 						rack->r_ctl.crte = NULL;
15145 						rack_log_hdwr_pacing(rack,
15146 								     rate_wanted, 0, __LINE__,
15147 								     err, 1);
15148 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15149 						counter_u64_add(rack_hw_pace_lost, 1);
15150 					} else if (nrte != rack->r_ctl.crte) {
15151 						rack->r_ctl.crte = nrte;
15152 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
15153 													 segsiz, 0,
15154 													 rack->r_ctl.crte,
15155 													 NULL);
15156 						rack_log_hdwr_pacing(rack,
15157 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15158 								     err, 2);
15159 						rack->r_ctl.last_hw_bw_req = rate_wanted;
15160 					}
15161 				} else {
15162 					/* We just need to adjust the segment size */
15163 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15164 					rack_log_hdwr_pacing(rack,
15165 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15166 							     0, 4);
15167 					rack->r_ctl.last_hw_bw_req = rate_wanted;
15168 				}
15169 			}
15170 		}
15171 		if ((rack->r_ctl.crte != NULL) &&
15172 		    (rack->r_ctl.crte->rate == rate_wanted)) {
15173 			/*
15174 			 * We need to add a extra if the rates
15175 			 * are exactly matched. The idea is
15176 			 * we want the software to make sure the
15177 			 * queue is empty before adding more, this
15178 			 * gives us N MSS extra pace times where
15179 			 * N is our sysctl
15180 			 */
15181 			slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
15182 		}
15183 done_w_hdwr:
15184 		if (rack_limit_time_with_srtt &&
15185 		    (rack->use_fixed_rate == 0) &&
15186 #ifdef NETFLIX_PEAKRATE
15187 		    (rack->rc_tp->t_maxpeakrate == 0) &&
15188 #endif
15189 		    (rack->rack_hdrw_pacing == 0)) {
15190 			/*
15191 			 * Sanity check, we do not allow the pacing delay
15192 			 * to be longer than the SRTT of the path. If it is
15193 			 * a slow path, then adding a packet should increase
15194 			 * the RTT and compensate for this i.e. the srtt will
15195 			 * be greater so the allowed pacing time will be greater.
15196 			 *
15197 			 * Note this restriction is not for where a peak rate
15198 			 * is set, we are doing fixed pacing or hardware pacing.
15199 			 */
15200 			if (rack->rc_tp->t_srtt)
15201 				srtt = rack->rc_tp->t_srtt;
15202 			else
15203 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
15204 			if (srtt < (uint64_t)slot) {
15205 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
15206 				slot = srtt;
15207 			}
15208 		}
15209 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
15210 	}
15211 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
15212 		/*
15213 		 * If this rate is seeing enobufs when it
15214 		 * goes to send then either the nic is out
15215 		 * of gas or we are mis-estimating the time
15216 		 * somehow and not letting the queue empty
15217 		 * completely. Lets add to the pacing time.
15218 		 */
15219 		int hw_boost_delay;
15220 
15221 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
15222 		if (hw_boost_delay > rack_enobuf_hw_max)
15223 			hw_boost_delay = rack_enobuf_hw_max;
15224 		else if (hw_boost_delay < rack_enobuf_hw_min)
15225 			hw_boost_delay = rack_enobuf_hw_min;
15226 		slot += hw_boost_delay;
15227 	}
15228 	if (slot)
15229 		counter_u64_add(rack_calc_nonzero, 1);
15230 	else
15231 		counter_u64_add(rack_calc_zero, 1);
15232 	return (slot);
15233 }
15234 
15235 static void
15236 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
15237     tcp_seq startseq, uint32_t sb_offset)
15238 {
15239 	struct rack_sendmap *my_rsm = NULL;
15240 	struct rack_sendmap fe;
15241 
15242 	if (tp->t_state < TCPS_ESTABLISHED) {
15243 		/*
15244 		 * We don't start any measurements if we are
15245 		 * not at least established.
15246 		 */
15247 		return;
15248 	}
15249 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
15250 		/*
15251 		 * We will get no more data into the SB
15252 		 * this means we need to have the data available
15253 		 * before we start a measurement.
15254 		 */
15255 
15256 		if (sbavail(&tp->t_inpcb->inp_socket->so_snd) <
15257 		    max(rc_init_window(rack),
15258 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
15259 			/* Nope not enough data */
15260 			return;
15261 		}
15262 	}
15263 	tp->t_flags |= TF_GPUTINPROG;
15264 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
15265 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
15266 	tp->gput_seq = startseq;
15267 	rack->app_limited_needs_set = 0;
15268 	if (rack->in_probe_rtt)
15269 		rack->measure_saw_probe_rtt = 1;
15270 	else if ((rack->measure_saw_probe_rtt) &&
15271 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
15272 		rack->measure_saw_probe_rtt = 0;
15273 	if (rack->rc_gp_filled)
15274 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15275 	else {
15276 		/* Special case initial measurement */
15277 		struct timeval tv;
15278 
15279 		tp->gput_ts = tcp_get_usecs(&tv);
15280 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15281 	}
15282 	/*
15283 	 * We take a guess out into the future,
15284 	 * if we have no measurement and no
15285 	 * initial rate, we measure the first
15286 	 * initial-windows worth of data to
15287 	 * speed up getting some GP measurement and
15288 	 * thus start pacing.
15289 	 */
15290 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
15291 		rack->app_limited_needs_set = 1;
15292 		tp->gput_ack = startseq + max(rc_init_window(rack),
15293 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
15294 		rack_log_pacing_delay_calc(rack,
15295 					   tp->gput_seq,
15296 					   tp->gput_ack,
15297 					   0,
15298 					   tp->gput_ts,
15299 					   rack->r_ctl.rc_app_limited_cnt,
15300 					   9,
15301 					   __LINE__, NULL, 0);
15302 		return;
15303 	}
15304 	if (sb_offset) {
15305 		/*
15306 		 * We are out somewhere in the sb
15307 		 * can we use the already outstanding data?
15308 		 */
15309 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
15310 			/*
15311 			 * Yes first one is good and in this case
15312 			 * the tp->gput_ts is correctly set based on
15313 			 * the last ack that arrived (no need to
15314 			 * set things up when an ack comes in).
15315 			 */
15316 			my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
15317 			if ((my_rsm == NULL) ||
15318 			    (my_rsm->r_rtr_cnt != 1)) {
15319 				/* retransmission? */
15320 				goto use_latest;
15321 			}
15322 		} else {
15323 			if (rack->r_ctl.rc_first_appl == NULL) {
15324 				/*
15325 				 * If rc_first_appl is NULL
15326 				 * then the cnt should be 0.
15327 				 * This is probably an error, maybe
15328 				 * a KASSERT would be approprate.
15329 				 */
15330 				goto use_latest;
15331 			}
15332 			/*
15333 			 * If we have a marker pointer to the last one that is
15334 			 * app limited we can use that, but we need to set
15335 			 * things up so that when it gets ack'ed we record
15336 			 * the ack time (if its not already acked).
15337 			 */
15338 			rack->app_limited_needs_set = 1;
15339 			/*
15340 			 * We want to get to the rsm that is either
15341 			 * next with space i.e. over 1 MSS or the one
15342 			 * after that (after the app-limited).
15343 			 */
15344 			my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15345 					 rack->r_ctl.rc_first_appl);
15346 			if (my_rsm) {
15347 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
15348 					/* Have to use the next one */
15349 					my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15350 							 my_rsm);
15351 				else {
15352 					/* Use after the first MSS of it is acked */
15353 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
15354 					goto start_set;
15355 				}
15356 			}
15357 			if ((my_rsm == NULL) ||
15358 			    (my_rsm->r_rtr_cnt != 1)) {
15359 				/*
15360 				 * Either its a retransmit or
15361 				 * the last is the app-limited one.
15362 				 */
15363 				goto use_latest;
15364 			}
15365 		}
15366 		tp->gput_seq = my_rsm->r_start;
15367 start_set:
15368 		if (my_rsm->r_flags & RACK_ACKED) {
15369 			/*
15370 			 * This one has been acked use the arrival ack time
15371 			 */
15372 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15373 			rack->app_limited_needs_set = 0;
15374 		}
15375 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15376 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
15377 		rack_log_pacing_delay_calc(rack,
15378 					   tp->gput_seq,
15379 					   tp->gput_ack,
15380 					   (uint64_t)my_rsm,
15381 					   tp->gput_ts,
15382 					   rack->r_ctl.rc_app_limited_cnt,
15383 					   9,
15384 					   __LINE__, NULL, 0);
15385 		return;
15386 	}
15387 
15388 use_latest:
15389 	/*
15390 	 * We don't know how long we may have been
15391 	 * idle or if this is the first-send. Lets
15392 	 * setup the flag so we will trim off
15393 	 * the first ack'd data so we get a true
15394 	 * measurement.
15395 	 */
15396 	rack->app_limited_needs_set = 1;
15397 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
15398 	/* Find this guy so we can pull the send time */
15399 	fe.r_start = startseq;
15400 	my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
15401 	if (my_rsm) {
15402 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15403 		if (my_rsm->r_flags & RACK_ACKED) {
15404 			/*
15405 			 * Unlikely since its probably what was
15406 			 * just transmitted (but I am paranoid).
15407 			 */
15408 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15409 			rack->app_limited_needs_set = 0;
15410 		}
15411 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
15412 			/* This also is unlikely */
15413 			tp->gput_seq = my_rsm->r_start;
15414 		}
15415 	} else {
15416 		/*
15417 		 * TSNH unless we have some send-map limit,
15418 		 * and even at that it should not be hitting
15419 		 * that limit (we should have stopped sending).
15420 		 */
15421 		struct timeval tv;
15422 
15423 		microuptime(&tv);
15424 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15425 	}
15426 	rack_log_pacing_delay_calc(rack,
15427 				   tp->gput_seq,
15428 				   tp->gput_ack,
15429 				   (uint64_t)my_rsm,
15430 				   tp->gput_ts,
15431 				   rack->r_ctl.rc_app_limited_cnt,
15432 				   9, __LINE__, NULL, 0);
15433 }
15434 
15435 static inline uint32_t
15436 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
15437     uint32_t avail, int32_t sb_offset)
15438 {
15439 	uint32_t len;
15440 	uint32_t sendwin;
15441 
15442 	if (tp->snd_wnd > cwnd_to_use)
15443 		sendwin = cwnd_to_use;
15444 	else
15445 		sendwin = tp->snd_wnd;
15446 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
15447 		/* We never want to go over our peers rcv-window */
15448 		len = 0;
15449 	} else {
15450 		uint32_t flight;
15451 
15452 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15453 		if (flight >= sendwin) {
15454 			/*
15455 			 * We have in flight what we are allowed by cwnd (if
15456 			 * it was rwnd blocking it would have hit above out
15457 			 * >= tp->snd_wnd).
15458 			 */
15459 			return (0);
15460 		}
15461 		len = sendwin - flight;
15462 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
15463 			/* We would send too much (beyond the rwnd) */
15464 			len = tp->snd_wnd - ctf_outstanding(tp);
15465 		}
15466 		if ((len + sb_offset) > avail) {
15467 			/*
15468 			 * We don't have that much in the SB, how much is
15469 			 * there?
15470 			 */
15471 			len = avail - sb_offset;
15472 		}
15473 	}
15474 	return (len);
15475 }
15476 
15477 static void
15478 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
15479 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
15480 	     int rsm_is_null, int optlen, int line, uint16_t mode)
15481 {
15482 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15483 		union tcp_log_stackspecific log;
15484 		struct timeval tv;
15485 
15486 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15487 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
15488 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
15489 		log.u_bbr.flex1 = error;
15490 		log.u_bbr.flex2 = flags;
15491 		log.u_bbr.flex3 = rsm_is_null;
15492 		log.u_bbr.flex4 = ipoptlen;
15493 		log.u_bbr.flex5 = tp->rcv_numsacks;
15494 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15495 		log.u_bbr.flex7 = optlen;
15496 		log.u_bbr.flex8 = rack->r_fsb_inited;
15497 		log.u_bbr.applimited = rack->r_fast_output;
15498 		log.u_bbr.bw_inuse = rack_get_bw(rack);
15499 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15500 		log.u_bbr.cwnd_gain = mode;
15501 		log.u_bbr.pkts_out = orig_len;
15502 		log.u_bbr.lt_epoch = len;
15503 		log.u_bbr.delivered = line;
15504 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15505 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15506 		tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
15507 			       len, &log, false, NULL, NULL, 0, &tv);
15508 	}
15509 }
15510 
15511 
15512 static struct mbuf *
15513 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
15514 		   struct rack_fast_send_blk *fsb,
15515 		   int32_t seglimit, int32_t segsize, int hw_tls)
15516 {
15517 #ifdef KERN_TLS
15518 	struct ktls_session *tls, *ntls;
15519 	struct mbuf *start;
15520 #endif
15521 	struct mbuf *m, *n, **np, *smb;
15522 	struct mbuf *top;
15523 	int32_t off, soff;
15524 	int32_t len = *plen;
15525 	int32_t fragsize;
15526 	int32_t len_cp = 0;
15527 	uint32_t mlen, frags;
15528 
15529 	soff = off = the_off;
15530 	smb = m = the_m;
15531 	np = &top;
15532 	top = NULL;
15533 #ifdef KERN_TLS
15534 	if (hw_tls && (m->m_flags & M_EXTPG))
15535 		tls = m->m_epg_tls;
15536 	else
15537 		tls = NULL;
15538 	start = m;
15539 #endif
15540 	while (len > 0) {
15541 		if (m == NULL) {
15542 			*plen = len_cp;
15543 			break;
15544 		}
15545 #ifdef KERN_TLS
15546 		if (hw_tls) {
15547 			if (m->m_flags & M_EXTPG)
15548 				ntls = m->m_epg_tls;
15549 			else
15550 				ntls = NULL;
15551 
15552 			/*
15553 			 * Avoid mixing TLS records with handshake
15554 			 * data or TLS records from different
15555 			 * sessions.
15556 			 */
15557 			if (tls != ntls) {
15558 				MPASS(m != start);
15559 				*plen = len_cp;
15560 				break;
15561 			}
15562 		}
15563 #endif
15564 		mlen = min(len, m->m_len - off);
15565 		if (seglimit) {
15566 			/*
15567 			 * For M_EXTPG mbufs, add 3 segments
15568 			 * + 1 in case we are crossing page boundaries
15569 			 * + 2 in case the TLS hdr/trailer are used
15570 			 * It is cheaper to just add the segments
15571 			 * than it is to take the cache miss to look
15572 			 * at the mbuf ext_pgs state in detail.
15573 			 */
15574 			if (m->m_flags & M_EXTPG) {
15575 				fragsize = min(segsize, PAGE_SIZE);
15576 				frags = 3;
15577 			} else {
15578 				fragsize = segsize;
15579 				frags = 0;
15580 			}
15581 
15582 			/* Break if we really can't fit anymore. */
15583 			if ((frags + 1) >= seglimit) {
15584 				*plen =	len_cp;
15585 				break;
15586 			}
15587 
15588 			/*
15589 			 * Reduce size if you can't copy the whole
15590 			 * mbuf. If we can't copy the whole mbuf, also
15591 			 * adjust len so the loop will end after this
15592 			 * mbuf.
15593 			 */
15594 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
15595 				mlen = (seglimit - frags - 1) * fragsize;
15596 				len = mlen;
15597 				*plen = len_cp + len;
15598 			}
15599 			frags += howmany(mlen, fragsize);
15600 			if (frags == 0)
15601 				frags++;
15602 			seglimit -= frags;
15603 			KASSERT(seglimit > 0,
15604 			    ("%s: seglimit went too low", __func__));
15605 		}
15606 		n = m_get(M_NOWAIT, m->m_type);
15607 		*np = n;
15608 		if (n == NULL)
15609 			goto nospace;
15610 		n->m_len = mlen;
15611 		soff += mlen;
15612 		len_cp += n->m_len;
15613 		if (m->m_flags & (M_EXT|M_EXTPG)) {
15614 			n->m_data = m->m_data + off;
15615 			mb_dupcl(n, m);
15616 		} else {
15617 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
15618 			    (u_int)n->m_len);
15619 		}
15620 		len -= n->m_len;
15621 		off = 0;
15622 		m = m->m_next;
15623 		np = &n->m_next;
15624 		if (len || (soff == smb->m_len)) {
15625 			/*
15626 			 * We have more so we move forward  or
15627 			 * we have consumed the entire mbuf and
15628 			 * len has fell to 0.
15629 			 */
15630 			soff = 0;
15631 			smb = m;
15632 		}
15633 
15634 	}
15635 	if (fsb != NULL) {
15636 		fsb->m = smb;
15637 		fsb->off = soff;
15638 		if (smb) {
15639 			/*
15640 			 * Save off the size of the mbuf. We do
15641 			 * this so that we can recognize when it
15642 			 * has been trimmed by sbcut() as acks
15643 			 * come in.
15644 			 */
15645 			fsb->o_m_len = smb->m_len;
15646 		} else {
15647 			/*
15648 			 * This is the case where the next mbuf went to NULL. This
15649 			 * means with this copy we have sent everything in the sb.
15650 			 * In theory we could clear the fast_output flag, but lets
15651 			 * not since its possible that we could get more added
15652 			 * and acks that call the extend function which would let
15653 			 * us send more.
15654 			 */
15655 			fsb->o_m_len = 0;
15656 		}
15657 	}
15658 	return (top);
15659 nospace:
15660 	if (top)
15661 		m_freem(top);
15662 	return (NULL);
15663 
15664 }
15665 
15666 /*
15667  * This is a copy of m_copym(), taking the TSO segment size/limit
15668  * constraints into account, and advancing the sndptr as it goes.
15669  */
15670 static struct mbuf *
15671 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15672 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15673 {
15674 	struct mbuf *m, *n;
15675 	int32_t soff;
15676 
15677 	soff = rack->r_ctl.fsb.off;
15678 	m = rack->r_ctl.fsb.m;
15679 	if (rack->r_ctl.fsb.o_m_len > m->m_len) {
15680 		/*
15681 		 * The mbuf had the front of it chopped off by an ack
15682 		 * we need to adjust the soff/off by that difference.
15683 		 */
15684 		uint32_t delta;
15685 
15686 		delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15687 		soff -= delta;
15688 	} else if (rack->r_ctl.fsb.o_m_len < m->m_len) {
15689 		/*
15690 		 * The mbuf was expanded probably by
15691 		 * a m_compress. Just update o_m_len.
15692 		 */
15693 		rack->r_ctl.fsb.o_m_len = m->m_len;
15694 	}
15695 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15696 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15697 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15698 				 __FUNCTION__,
15699 				 rack, *plen, m, m->m_len));
15700 	/* Save off the right location before we copy and advance */
15701 	*s_soff = soff;
15702 	*s_mb = rack->r_ctl.fsb.m;
15703 	n = rack_fo_base_copym(m, soff, plen,
15704 			       &rack->r_ctl.fsb,
15705 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
15706 	return (n);
15707 }
15708 
15709 static int
15710 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15711 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
15712 {
15713 	/*
15714 	 * Enter the fast retransmit path. We are given that a sched_pin is
15715 	 * in place (if accounting is compliled in) and the cycle count taken
15716 	 * at the entry is in the ts_val. The concept her is that the rsm
15717 	 * now holds the mbuf offsets and such so we can directly transmit
15718 	 * without a lot of overhead, the len field is already set for
15719 	 * us to prohibit us from sending too much (usually its 1MSS).
15720 	 */
15721 	struct ip *ip = NULL;
15722 	struct udphdr *udp = NULL;
15723 	struct tcphdr *th = NULL;
15724 	struct mbuf *m = NULL;
15725 	struct inpcb *inp;
15726 	uint8_t *cpto;
15727 	struct tcp_log_buffer *lgb;
15728 #ifdef TCP_ACCOUNTING
15729 	uint64_t crtsc;
15730 	int cnt_thru = 1;
15731 #endif
15732 	struct tcpopt to;
15733 	u_char opt[TCP_MAXOLEN];
15734 	uint32_t hdrlen, optlen;
15735 	int32_t slot, segsiz, max_val, tso = 0, error, flags, ulen = 0;
15736 	uint32_t us_cts;
15737 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
15738 	uint32_t if_hw_tsomaxsegsize;
15739 
15740 #ifdef INET6
15741 	struct ip6_hdr *ip6 = NULL;
15742 
15743 	if (rack->r_is_v6) {
15744 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15745 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15746 	} else
15747 #endif				/* INET6 */
15748 	{
15749 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15750 		hdrlen = sizeof(struct tcpiphdr);
15751 	}
15752 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15753 		goto failed;
15754 	}
15755 	if (doing_tlp) {
15756 		/* Its a TLP add the flag, it may already be there but be sure */
15757 		rsm->r_flags |= RACK_TLP;
15758 	} else {
15759 		/* If it was a TLP it is not not on this retransmit */
15760 		rsm->r_flags &= ~RACK_TLP;
15761 	}
15762 	startseq = rsm->r_start;
15763 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15764 	inp = rack->rc_inp;
15765 	to.to_flags = 0;
15766 	flags = tcp_outflags[tp->t_state];
15767 	if (flags & (TH_SYN|TH_RST)) {
15768 		goto failed;
15769 	}
15770 	if (rsm->r_flags & RACK_HAS_FIN) {
15771 		/* We can't send a FIN here */
15772 		goto failed;
15773 	}
15774 	if (flags & TH_FIN) {
15775 		/* We never send a FIN */
15776 		flags &= ~TH_FIN;
15777 	}
15778 	if (tp->t_flags & TF_RCVD_TSTMP) {
15779 		to.to_tsval = ms_cts + tp->ts_offset;
15780 		to.to_tsecr = tp->ts_recent;
15781 		to.to_flags = TOF_TS;
15782 	}
15783 	optlen = tcp_addoptions(&to, opt);
15784 	hdrlen += optlen;
15785 	udp = rack->r_ctl.fsb.udp;
15786 	if (udp)
15787 		hdrlen += sizeof(struct udphdr);
15788 	if (rack->r_ctl.rc_pace_max_segs)
15789 		max_val = rack->r_ctl.rc_pace_max_segs;
15790 	else if (rack->rc_user_set_max_segs)
15791 		max_val = rack->rc_user_set_max_segs * segsiz;
15792 	else
15793 		max_val = len;
15794 	if ((tp->t_flags & TF_TSO) &&
15795 	    V_tcp_do_tso &&
15796 	    (len > segsiz) &&
15797 	    (tp->t_port == 0))
15798 		tso = 1;
15799 #ifdef INET6
15800 	if (MHLEN < hdrlen + max_linkhdr)
15801 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15802 	else
15803 #endif
15804 		m = m_gethdr(M_NOWAIT, MT_DATA);
15805 	if (m == NULL)
15806 		goto failed;
15807 	m->m_data += max_linkhdr;
15808 	m->m_len = hdrlen;
15809 	th = rack->r_ctl.fsb.th;
15810 	/* Establish the len to send */
15811 	if (len > max_val)
15812 		len = max_val;
15813 	if ((tso) && (len + optlen > tp->t_maxseg)) {
15814 		uint32_t if_hw_tsomax;
15815 		int32_t max_len;
15816 
15817 		/* extract TSO information */
15818 		if_hw_tsomax = tp->t_tsomax;
15819 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15820 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15821 		/*
15822 		 * Check if we should limit by maximum payload
15823 		 * length:
15824 		 */
15825 		if (if_hw_tsomax != 0) {
15826 			/* compute maximum TSO length */
15827 			max_len = (if_hw_tsomax - hdrlen -
15828 				   max_linkhdr);
15829 			if (max_len <= 0) {
15830 				goto failed;
15831 			} else if (len > max_len) {
15832 				len = max_len;
15833 			}
15834 		}
15835 		if (len <= segsiz) {
15836 			/*
15837 			 * In case there are too many small fragments don't
15838 			 * use TSO:
15839 			 */
15840 			tso = 0;
15841 		}
15842 	} else {
15843 		tso = 0;
15844 	}
15845 	if ((tso == 0) && (len > segsiz))
15846 		len = segsiz;
15847 	us_cts = tcp_get_usecs(tv);
15848 	if ((len == 0) ||
15849 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
15850 		goto failed;
15851 	}
15852 	th->th_seq = htonl(rsm->r_start);
15853 	th->th_ack = htonl(tp->rcv_nxt);
15854 	/*
15855 	 * The PUSH bit should only be applied
15856 	 * if the full retransmission is made. If
15857 	 * we are sending less than this is the
15858 	 * left hand edge and should not have
15859 	 * the PUSH bit.
15860 	 */
15861 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
15862 	    (len == (rsm->r_end - rsm->r_start)))
15863 		flags |= TH_PUSH;
15864 	th->th_flags = flags;
15865 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15866 	if (th->th_win == 0) {
15867 		tp->t_sndzerowin++;
15868 		tp->t_flags |= TF_RXWIN0SENT;
15869 	} else
15870 		tp->t_flags &= ~TF_RXWIN0SENT;
15871 	if (rsm->r_flags & RACK_TLP) {
15872 		/*
15873 		 * TLP should not count in retran count, but
15874 		 * in its own bin
15875 		 */
15876 		counter_u64_add(rack_tlp_retran, 1);
15877 		counter_u64_add(rack_tlp_retran_bytes, len);
15878 	} else {
15879 		tp->t_sndrexmitpack++;
15880 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15881 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15882 	}
15883 #ifdef STATS
15884 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15885 				 len);
15886 #endif
15887 	if (rsm->m == NULL)
15888 		goto failed;
15889 	if (rsm->orig_m_len != rsm->m->m_len) {
15890 		/* Fix up the orig_m_len and possibly the mbuf offset */
15891 		rack_adjust_orig_mlen(rsm);
15892 	}
15893 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
15894 	if (len <= segsiz) {
15895 		/*
15896 		 * Must have ran out of mbufs for the copy
15897 		 * shorten it to no longer need tso. Lets
15898 		 * not put on sendalot since we are low on
15899 		 * mbufs.
15900 		 */
15901 		tso = 0;
15902 	}
15903 	if ((m->m_next == NULL) || (len <= 0)){
15904 		goto failed;
15905 	}
15906 	if (udp) {
15907 		if (rack->r_is_v6)
15908 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
15909 		else
15910 			ulen = hdrlen + len - sizeof(struct ip);
15911 		udp->uh_ulen = htons(ulen);
15912 	}
15913 	m->m_pkthdr.rcvif = (struct ifnet *)0;
15914 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
15915 #ifdef INET6
15916 	if (rack->r_is_v6) {
15917 		if (tp->t_port) {
15918 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15919 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15920 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15921 			th->th_sum = htons(0);
15922 			UDPSTAT_INC(udps_opackets);
15923 		} else {
15924 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15925 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15926 			th->th_sum = in6_cksum_pseudo(ip6,
15927 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15928 						      0);
15929 		}
15930 	}
15931 #endif
15932 #if defined(INET6) && defined(INET)
15933 	else
15934 #endif
15935 #ifdef INET
15936 	{
15937 		if (tp->t_port) {
15938 			m->m_pkthdr.csum_flags = CSUM_UDP;
15939 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15940 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15941 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15942 			th->th_sum = htons(0);
15943 			UDPSTAT_INC(udps_opackets);
15944 		} else {
15945 			m->m_pkthdr.csum_flags = CSUM_TCP;
15946 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15947 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
15948 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15949 									IPPROTO_TCP + len + optlen));
15950 		}
15951 		/* IP version must be set here for ipv4/ipv6 checking later */
15952 		KASSERT(ip->ip_v == IPVERSION,
15953 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
15954 	}
15955 #endif
15956 	if (tso) {
15957 		KASSERT(len > tp->t_maxseg - optlen,
15958 			("%s: len <= tso_segsz tp:%p", __func__, tp));
15959 		m->m_pkthdr.csum_flags |= CSUM_TSO;
15960 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15961 	}
15962 #ifdef INET6
15963 	if (rack->r_is_v6) {
15964 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15965 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15966 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15967 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15968 		else
15969 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15970 	}
15971 #endif
15972 #if defined(INET) && defined(INET6)
15973 	else
15974 #endif
15975 #ifdef INET
15976 	{
15977 		ip->ip_len = htons(m->m_pkthdr.len);
15978 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15979 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15980 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15981 			if (tp->t_port == 0 || len < V_tcp_minmss) {
15982 				ip->ip_off |= htons(IP_DF);
15983 			}
15984 		} else {
15985 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15986 		}
15987 	}
15988 #endif
15989 	/* Time to copy in our header */
15990 	cpto = mtod(m, uint8_t *);
15991 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15992 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15993 	if (optlen) {
15994 		bcopy(opt, th + 1, optlen);
15995 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15996 	} else {
15997 		th->th_off = sizeof(struct tcphdr) >> 2;
15998 	}
15999 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
16000 		union tcp_log_stackspecific log;
16001 
16002 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16003 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
16004 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
16005 		if (rack->rack_no_prr)
16006 			log.u_bbr.flex1 = 0;
16007 		else
16008 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16009 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16010 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16011 		log.u_bbr.flex4 = max_val;
16012 		log.u_bbr.flex5 = 0;
16013 		/* Save off the early/late values */
16014 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16015 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16016 		log.u_bbr.bw_inuse = rack_get_bw(rack);
16017 		if (doing_tlp == 0)
16018 			log.u_bbr.flex8 = 1;
16019 		else
16020 			log.u_bbr.flex8 = 2;
16021 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16022 		log.u_bbr.flex7 = 55;
16023 		log.u_bbr.pkts_out = tp->t_maxseg;
16024 		log.u_bbr.timeStamp = cts;
16025 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16026 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16027 		log.u_bbr.delivered = 0;
16028 		lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16029 				     len, &log, false, NULL, NULL, 0, tv);
16030 	} else
16031 		lgb = NULL;
16032 #ifdef INET6
16033 	if (rack->r_is_v6) {
16034 		error = ip6_output(m, NULL,
16035 				   &inp->inp_route6,
16036 				   0, NULL, NULL, inp);
16037 	}
16038 #endif
16039 #if defined(INET) && defined(INET6)
16040 	else
16041 #endif
16042 #ifdef INET
16043 	{
16044 		error = ip_output(m, NULL,
16045 				  &inp->inp_route,
16046 				  0, 0, inp);
16047 	}
16048 #endif
16049 	m = NULL;
16050 	if (lgb) {
16051 		lgb->tlb_errno = error;
16052 		lgb = NULL;
16053 	}
16054 	if (error) {
16055 		goto failed;
16056 	}
16057 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
16058 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls);
16059 	if (doing_tlp && (rack->fast_rsm_hack == 0)) {
16060 		rack->rc_tlp_in_progress = 1;
16061 		rack->r_ctl.rc_tlp_cnt_out++;
16062 	}
16063 	if (error == 0) {
16064 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
16065 		if (doing_tlp) {
16066 			rack->rc_last_sent_tlp_past_cumack = 0;
16067 			rack->rc_last_sent_tlp_seq_valid = 1;
16068 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
16069 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
16070 		}
16071 	}
16072 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16073 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
16074 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
16075 		rack->r_ctl.retran_during_recovery += len;
16076 	{
16077 		int idx;
16078 
16079 		idx = (len / segsiz) + 3;
16080 		if (idx >= TCP_MSS_ACCT_ATIMER)
16081 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16082 		else
16083 			counter_u64_add(rack_out_size[idx], 1);
16084 	}
16085 	if (tp->t_rtttime == 0) {
16086 		tp->t_rtttime = ticks;
16087 		tp->t_rtseq = startseq;
16088 		KMOD_TCPSTAT_INC(tcps_segstimed);
16089 	}
16090 	counter_u64_add(rack_fto_rsm_send, 1);
16091 	if (error && (error == ENOBUFS)) {
16092 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
16093 		if (rack->rc_enobuf < 0x7f)
16094 			rack->rc_enobuf++;
16095 		if (slot < (10 * HPTS_USEC_IN_MSEC))
16096 			slot = 10 * HPTS_USEC_IN_MSEC;
16097 	} else
16098 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
16099 	if ((slot == 0) ||
16100 	    (rack->rc_always_pace == 0) ||
16101 	    (rack->r_rr_config == 1)) {
16102 		/*
16103 		 * We have no pacing set or we
16104 		 * are using old-style rack or
16105 		 * we are overriden to use the old 1ms pacing.
16106 		 */
16107 		slot = rack->r_ctl.rc_min_to;
16108 	}
16109 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
16110 	if (rack->r_must_retran) {
16111 		rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
16112 		if ((SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) ||
16113 		    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
16114 			/*
16115 			 * We have retransmitted all we need. If
16116 			 * RACK_MUST_RXT is not set then we need to
16117 			 * not retransmit this guy.
16118 			 */
16119 			rack->r_must_retran = 0;
16120 			rack->r_ctl.rc_out_at_rto = 0;
16121 			if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
16122 				/* Not one we should rxt */
16123 				goto failed;
16124 			} else {
16125 				/* Clear the flag */
16126 				rsm->r_flags &= ~RACK_MUST_RXT;
16127 			}
16128 		} else {
16129 			/* Remove  the flag */
16130 			rsm->r_flags &= ~RACK_MUST_RXT;
16131 		}
16132 	}
16133 #ifdef TCP_ACCOUNTING
16134 	crtsc = get_cyclecount();
16135 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16136 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16137 	}
16138 	counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16139 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16140 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16141 	}
16142 	counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16143 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16144 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
16145 	}
16146 	counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
16147 	sched_unpin();
16148 #endif
16149 	return (0);
16150 failed:
16151 	if (m)
16152 		m_free(m);
16153 	return (-1);
16154 }
16155 
16156 static void
16157 rack_sndbuf_autoscale(struct tcp_rack *rack)
16158 {
16159 	/*
16160 	 * Automatic sizing of send socket buffer.  Often the send buffer
16161 	 * size is not optimally adjusted to the actual network conditions
16162 	 * at hand (delay bandwidth product).  Setting the buffer size too
16163 	 * small limits throughput on links with high bandwidth and high
16164 	 * delay (eg. trans-continental/oceanic links).  Setting the
16165 	 * buffer size too big consumes too much real kernel memory,
16166 	 * especially with many connections on busy servers.
16167 	 *
16168 	 * The criteria to step up the send buffer one notch are:
16169 	 *  1. receive window of remote host is larger than send buffer
16170 	 *     (with a fudge factor of 5/4th);
16171 	 *  2. send buffer is filled to 7/8th with data (so we actually
16172 	 *     have data to make use of it);
16173 	 *  3. send buffer fill has not hit maximal automatic size;
16174 	 *  4. our send window (slow start and cogestion controlled) is
16175 	 *     larger than sent but unacknowledged data in send buffer.
16176 	 *
16177 	 * Note that the rack version moves things much faster since
16178 	 * we want to avoid hitting cache lines in the rack_fast_output()
16179 	 * path so this is called much less often and thus moves
16180 	 * the SB forward by a percentage.
16181 	 */
16182 	struct socket *so;
16183 	struct tcpcb *tp;
16184 	uint32_t sendwin, scaleup;
16185 
16186 	tp = rack->rc_tp;
16187 	so = rack->rc_inp->inp_socket;
16188 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
16189 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
16190 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
16191 		    sbused(&so->so_snd) >=
16192 		    (so->so_snd.sb_hiwat / 8 * 7) &&
16193 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
16194 		    sendwin >= (sbused(&so->so_snd) -
16195 		    (tp->snd_nxt - tp->snd_una))) {
16196 			if (rack_autosndbuf_inc)
16197 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
16198 			else
16199 				scaleup = V_tcp_autosndbuf_inc;
16200 			if (scaleup < V_tcp_autosndbuf_inc)
16201 				scaleup = V_tcp_autosndbuf_inc;
16202 			scaleup += so->so_snd.sb_hiwat;
16203 			if (scaleup > V_tcp_autosndbuf_max)
16204 				scaleup = V_tcp_autosndbuf_max;
16205 			if (!sbreserve_locked(&so->so_snd, scaleup, so, curthread))
16206 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
16207 		}
16208 	}
16209 }
16210 
16211 static int
16212 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
16213 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
16214 {
16215 	/*
16216 	 * Enter to do fast output. We are given that the sched_pin is
16217 	 * in place (if accounting is compiled in) and the cycle count taken
16218 	 * at entry is in place in ts_val. The idea here is that
16219 	 * we know how many more bytes needs to be sent (presumably either
16220 	 * during pacing or to fill the cwnd and that was greater than
16221 	 * the max-burst). We have how much to send and all the info we
16222 	 * need to just send.
16223 	 */
16224 	struct ip *ip = NULL;
16225 	struct udphdr *udp = NULL;
16226 	struct tcphdr *th = NULL;
16227 	struct mbuf *m, *s_mb;
16228 	struct inpcb *inp;
16229 	uint8_t *cpto;
16230 	struct tcp_log_buffer *lgb;
16231 #ifdef TCP_ACCOUNTING
16232 	uint64_t crtsc;
16233 #endif
16234 	struct tcpopt to;
16235 	u_char opt[TCP_MAXOLEN];
16236 	uint32_t hdrlen, optlen;
16237 	int cnt_thru = 1;
16238 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, flags, ulen = 0;
16239 	uint32_t us_cts, s_soff;
16240 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
16241 	uint32_t if_hw_tsomaxsegsize;
16242 	uint16_t add_flag = RACK_SENT_FP;
16243 #ifdef INET6
16244 	struct ip6_hdr *ip6 = NULL;
16245 
16246 	if (rack->r_is_v6) {
16247 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
16248 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
16249 	} else
16250 #endif				/* INET6 */
16251 	{
16252 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
16253 		hdrlen = sizeof(struct tcpiphdr);
16254 	}
16255 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
16256 		m = NULL;
16257 		goto failed;
16258 	}
16259 	startseq = tp->snd_max;
16260 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16261 	inp = rack->rc_inp;
16262 	len = rack->r_ctl.fsb.left_to_send;
16263 	to.to_flags = 0;
16264 	flags = rack->r_ctl.fsb.tcp_flags;
16265 	if (tp->t_flags & TF_RCVD_TSTMP) {
16266 		to.to_tsval = ms_cts + tp->ts_offset;
16267 		to.to_tsecr = tp->ts_recent;
16268 		to.to_flags = TOF_TS;
16269 	}
16270 	optlen = tcp_addoptions(&to, opt);
16271 	hdrlen += optlen;
16272 	udp = rack->r_ctl.fsb.udp;
16273 	if (udp)
16274 		hdrlen += sizeof(struct udphdr);
16275 	if (rack->r_ctl.rc_pace_max_segs)
16276 		max_val = rack->r_ctl.rc_pace_max_segs;
16277 	else if (rack->rc_user_set_max_segs)
16278 		max_val = rack->rc_user_set_max_segs * segsiz;
16279 	else
16280 		max_val = len;
16281 	if ((tp->t_flags & TF_TSO) &&
16282 	    V_tcp_do_tso &&
16283 	    (len > segsiz) &&
16284 	    (tp->t_port == 0))
16285 		tso = 1;
16286 again:
16287 #ifdef INET6
16288 	if (MHLEN < hdrlen + max_linkhdr)
16289 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
16290 	else
16291 #endif
16292 		m = m_gethdr(M_NOWAIT, MT_DATA);
16293 	if (m == NULL)
16294 		goto failed;
16295 	m->m_data += max_linkhdr;
16296 	m->m_len = hdrlen;
16297 	th = rack->r_ctl.fsb.th;
16298 	/* Establish the len to send */
16299 	if (len > max_val)
16300 		len = max_val;
16301 	if ((tso) && (len + optlen > tp->t_maxseg)) {
16302 		uint32_t if_hw_tsomax;
16303 		int32_t max_len;
16304 
16305 		/* extract TSO information */
16306 		if_hw_tsomax = tp->t_tsomax;
16307 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
16308 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
16309 		/*
16310 		 * Check if we should limit by maximum payload
16311 		 * length:
16312 		 */
16313 		if (if_hw_tsomax != 0) {
16314 			/* compute maximum TSO length */
16315 			max_len = (if_hw_tsomax - hdrlen -
16316 				   max_linkhdr);
16317 			if (max_len <= 0) {
16318 				goto failed;
16319 			} else if (len > max_len) {
16320 				len = max_len;
16321 			}
16322 		}
16323 		if (len <= segsiz) {
16324 			/*
16325 			 * In case there are too many small fragments don't
16326 			 * use TSO:
16327 			 */
16328 			tso = 0;
16329 		}
16330 	} else {
16331 		tso = 0;
16332 	}
16333 	if ((tso == 0) && (len > segsiz))
16334 		len = segsiz;
16335 	us_cts = tcp_get_usecs(tv);
16336 	if ((len == 0) ||
16337 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
16338 		goto failed;
16339 	}
16340 	sb_offset = tp->snd_max - tp->snd_una;
16341 	th->th_seq = htonl(tp->snd_max);
16342 	th->th_ack = htonl(tp->rcv_nxt);
16343 	th->th_flags = flags;
16344 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
16345 	if (th->th_win == 0) {
16346 		tp->t_sndzerowin++;
16347 		tp->t_flags |= TF_RXWIN0SENT;
16348 	} else
16349 		tp->t_flags &= ~TF_RXWIN0SENT;
16350 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
16351 	KMOD_TCPSTAT_INC(tcps_sndpack);
16352 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
16353 #ifdef STATS
16354 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
16355 				 len);
16356 #endif
16357 	if (rack->r_ctl.fsb.m == NULL)
16358 		goto failed;
16359 
16360 	/* s_mb and s_soff are saved for rack_log_output */
16361 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
16362 				    &s_mb, &s_soff);
16363 	if (len <= segsiz) {
16364 		/*
16365 		 * Must have ran out of mbufs for the copy
16366 		 * shorten it to no longer need tso. Lets
16367 		 * not put on sendalot since we are low on
16368 		 * mbufs.
16369 		 */
16370 		tso = 0;
16371 	}
16372 	if (rack->r_ctl.fsb.rfo_apply_push &&
16373 	    (len == rack->r_ctl.fsb.left_to_send)) {
16374 		th->th_flags |= TH_PUSH;
16375 		add_flag |= RACK_HAD_PUSH;
16376 	}
16377 	if ((m->m_next == NULL) || (len <= 0)){
16378 		goto failed;
16379 	}
16380 	if (udp) {
16381 		if (rack->r_is_v6)
16382 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
16383 		else
16384 			ulen = hdrlen + len - sizeof(struct ip);
16385 		udp->uh_ulen = htons(ulen);
16386 	}
16387 	m->m_pkthdr.rcvif = (struct ifnet *)0;
16388 	if (tp->t_state == TCPS_ESTABLISHED &&
16389 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
16390 		/*
16391 		 * If the peer has ECN, mark data packets with ECN capable
16392 		 * transmission (ECT). Ignore pure ack packets,
16393 		 * retransmissions.
16394 		 */
16395 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max)) {
16396 #ifdef INET6
16397 			if (rack->r_is_v6)
16398 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
16399 			else
16400 #endif
16401 				ip->ip_tos |= IPTOS_ECN_ECT0;
16402 			KMOD_TCPSTAT_INC(tcps_ecn_ect0);
16403 			/*
16404 			 * Reply with proper ECN notifications.
16405 			 * Only set CWR on new data segments.
16406 			 */
16407 			if (tp->t_flags2 & TF2_ECN_SND_CWR) {
16408 				flags |= TH_CWR;
16409 				tp->t_flags2 &= ~TF2_ECN_SND_CWR;
16410 			}
16411 		}
16412 		if (tp->t_flags2 & TF2_ECN_SND_ECE)
16413 			flags |= TH_ECE;
16414 	}
16415 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
16416 #ifdef INET6
16417 	if (rack->r_is_v6) {
16418 		if (tp->t_port) {
16419 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
16420 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16421 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
16422 			th->th_sum = htons(0);
16423 			UDPSTAT_INC(udps_opackets);
16424 		} else {
16425 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
16426 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16427 			th->th_sum = in6_cksum_pseudo(ip6,
16428 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
16429 						      0);
16430 		}
16431 	}
16432 #endif
16433 #if defined(INET6) && defined(INET)
16434 	else
16435 #endif
16436 #ifdef INET
16437 	{
16438 		if (tp->t_port) {
16439 			m->m_pkthdr.csum_flags = CSUM_UDP;
16440 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16441 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
16442 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
16443 			th->th_sum = htons(0);
16444 			UDPSTAT_INC(udps_opackets);
16445 		} else {
16446 			m->m_pkthdr.csum_flags = CSUM_TCP;
16447 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16448 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
16449 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
16450 									IPPROTO_TCP + len + optlen));
16451 		}
16452 		/* IP version must be set here for ipv4/ipv6 checking later */
16453 		KASSERT(ip->ip_v == IPVERSION,
16454 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
16455 	}
16456 #endif
16457 	if (tso) {
16458 		KASSERT(len > tp->t_maxseg - optlen,
16459 			("%s: len <= tso_segsz tp:%p", __func__, tp));
16460 		m->m_pkthdr.csum_flags |= CSUM_TSO;
16461 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
16462 	}
16463 #ifdef INET6
16464 	if (rack->r_is_v6) {
16465 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
16466 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
16467 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
16468 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16469 		else
16470 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16471 	}
16472 #endif
16473 #if defined(INET) && defined(INET6)
16474 	else
16475 #endif
16476 #ifdef INET
16477 	{
16478 		ip->ip_len = htons(m->m_pkthdr.len);
16479 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
16480 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
16481 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16482 			if (tp->t_port == 0 || len < V_tcp_minmss) {
16483 				ip->ip_off |= htons(IP_DF);
16484 			}
16485 		} else {
16486 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16487 		}
16488 	}
16489 #endif
16490 	/* Time to copy in our header */
16491 	cpto = mtod(m, uint8_t *);
16492 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
16493 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
16494 	if (optlen) {
16495 		bcopy(opt, th + 1, optlen);
16496 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
16497 	} else {
16498 		th->th_off = sizeof(struct tcphdr) >> 2;
16499 	}
16500 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
16501 		union tcp_log_stackspecific log;
16502 
16503 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16504 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
16505 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
16506 		if (rack->rack_no_prr)
16507 			log.u_bbr.flex1 = 0;
16508 		else
16509 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16510 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16511 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16512 		log.u_bbr.flex4 = max_val;
16513 		log.u_bbr.flex5 = 0;
16514 		/* Save off the early/late values */
16515 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16516 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16517 		log.u_bbr.bw_inuse = rack_get_bw(rack);
16518 		log.u_bbr.flex8 = 0;
16519 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16520 		log.u_bbr.flex7 = 44;
16521 		log.u_bbr.pkts_out = tp->t_maxseg;
16522 		log.u_bbr.timeStamp = cts;
16523 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16524 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16525 		log.u_bbr.delivered = 0;
16526 		lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16527 				     len, &log, false, NULL, NULL, 0, tv);
16528 	} else
16529 		lgb = NULL;
16530 #ifdef INET6
16531 	if (rack->r_is_v6) {
16532 		error = ip6_output(m, NULL,
16533 				   &inp->inp_route6,
16534 				   0, NULL, NULL, inp);
16535 	}
16536 #endif
16537 #if defined(INET) && defined(INET6)
16538 	else
16539 #endif
16540 #ifdef INET
16541 	{
16542 		error = ip_output(m, NULL,
16543 				  &inp->inp_route,
16544 				  0, 0, inp);
16545 	}
16546 #endif
16547 	if (lgb) {
16548 		lgb->tlb_errno = error;
16549 		lgb = NULL;
16550 	}
16551 	if (error) {
16552 		*send_err = error;
16553 		m = NULL;
16554 		goto failed;
16555 	}
16556 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
16557 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls);
16558 	m = NULL;
16559 	if (tp->snd_una == tp->snd_max) {
16560 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
16561 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
16562 		tp->t_acktime = ticks;
16563 	}
16564 	if (error == 0)
16565 		tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
16566 
16567 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
16568 	tot_len += len;
16569 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
16570 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
16571 	tp->snd_max += len;
16572 	tp->snd_nxt = tp->snd_max;
16573 	{
16574 		int idx;
16575 
16576 		idx = (len / segsiz) + 3;
16577 		if (idx >= TCP_MSS_ACCT_ATIMER)
16578 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16579 		else
16580 			counter_u64_add(rack_out_size[idx], 1);
16581 	}
16582 	if (len <= rack->r_ctl.fsb.left_to_send)
16583 		rack->r_ctl.fsb.left_to_send -= len;
16584 	else
16585 		rack->r_ctl.fsb.left_to_send = 0;
16586 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
16587 		rack->r_fast_output = 0;
16588 		rack->r_ctl.fsb.left_to_send = 0;
16589 		/* At the end of fast_output scale up the sb */
16590 		SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
16591 		rack_sndbuf_autoscale(rack);
16592 		SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
16593 	}
16594 	if (tp->t_rtttime == 0) {
16595 		tp->t_rtttime = ticks;
16596 		tp->t_rtseq = startseq;
16597 		KMOD_TCPSTAT_INC(tcps_segstimed);
16598 	}
16599 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
16600 	    (max_val > len) &&
16601 	    (tso == 0)) {
16602 		max_val -= len;
16603 		len = segsiz;
16604 		th = rack->r_ctl.fsb.th;
16605 		cnt_thru++;
16606 		goto again;
16607 	}
16608 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16609 	counter_u64_add(rack_fto_send, 1);
16610 	slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
16611 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
16612 #ifdef TCP_ACCOUNTING
16613 	crtsc = get_cyclecount();
16614 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16615 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16616 	}
16617 	counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16618 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16619 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16620 	}
16621 	counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16622 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16623 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
16624 	}
16625 	counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
16626 	sched_unpin();
16627 #endif
16628 	return (0);
16629 failed:
16630 	if (m)
16631 		m_free(m);
16632 	rack->r_fast_output = 0;
16633 	return (-1);
16634 }
16635 
16636 static int
16637 rack_output(struct tcpcb *tp)
16638 {
16639 	struct socket *so;
16640 	uint32_t recwin;
16641 	uint32_t sb_offset, s_moff = 0;
16642 	int32_t len, flags, error = 0;
16643 	struct mbuf *m, *s_mb = NULL;
16644 	struct mbuf *mb;
16645 	uint32_t if_hw_tsomaxsegcount = 0;
16646 	uint32_t if_hw_tsomaxsegsize;
16647 	int32_t segsiz, minseg;
16648 	long tot_len_this_send = 0;
16649 #ifdef INET
16650 	struct ip *ip = NULL;
16651 #endif
16652 #ifdef TCPDEBUG
16653 	struct ipovly *ipov = NULL;
16654 #endif
16655 	struct udphdr *udp = NULL;
16656 	struct tcp_rack *rack;
16657 	struct tcphdr *th;
16658 	uint8_t pass = 0;
16659 	uint8_t mark = 0;
16660 	uint8_t wanted_cookie = 0;
16661 	u_char opt[TCP_MAXOLEN];
16662 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
16663 	uint32_t rack_seq;
16664 
16665 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16666 	unsigned ipsec_optlen = 0;
16667 
16668 #endif
16669 	int32_t idle, sendalot;
16670 	int32_t sub_from_prr = 0;
16671 	volatile int32_t sack_rxmit;
16672 	struct rack_sendmap *rsm = NULL;
16673 	int32_t tso, mtu;
16674 	struct tcpopt to;
16675 	int32_t slot = 0;
16676 	int32_t sup_rack = 0;
16677 	uint32_t cts, ms_cts, delayed, early;
16678 	uint16_t add_flag = RACK_SENT_SP;
16679 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
16680 	uint8_t hpts_calling,  doing_tlp = 0;
16681 	uint32_t cwnd_to_use, pace_max_seg;
16682 	int32_t do_a_prefetch = 0;
16683 	int32_t prefetch_rsm = 0;
16684 	int32_t orig_len = 0;
16685 	struct timeval tv;
16686 	int32_t prefetch_so_done = 0;
16687 	struct tcp_log_buffer *lgb;
16688 	struct inpcb *inp;
16689 	struct sockbuf *sb;
16690 	uint64_t ts_val = 0;
16691 #ifdef TCP_ACCOUNTING
16692 	uint64_t crtsc;
16693 #endif
16694 #ifdef INET6
16695 	struct ip6_hdr *ip6 = NULL;
16696 	int32_t isipv6;
16697 #endif
16698 	uint8_t filled_all = 0;
16699 	bool hw_tls = false;
16700 
16701 	/* setup and take the cache hits here */
16702 	rack = (struct tcp_rack *)tp->t_fb_ptr;
16703 #ifdef TCP_ACCOUNTING
16704 	sched_pin();
16705 	ts_val = get_cyclecount();
16706 #endif
16707 	hpts_calling = rack->rc_inp->inp_hpts_calls;
16708 	NET_EPOCH_ASSERT();
16709 	INP_WLOCK_ASSERT(rack->rc_inp);
16710 #ifdef TCP_OFFLOAD
16711 	if (tp->t_flags & TF_TOE) {
16712 #ifdef TCP_ACCOUNTING
16713 		sched_unpin();
16714 #endif
16715 		return (tcp_offload_output(tp));
16716 	}
16717 #endif
16718 	/*
16719 	 * For TFO connections in SYN_RECEIVED, only allow the initial
16720 	 * SYN|ACK and those sent by the retransmit timer.
16721 	 */
16722 	if (IS_FASTOPEN(tp->t_flags) &&
16723 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
16724 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16725 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16726 #ifdef TCP_ACCOUNTING
16727 		sched_unpin();
16728 #endif
16729 		return (0);
16730 	}
16731 #ifdef INET6
16732 	if (rack->r_state) {
16733 		/* Use the cache line loaded if possible */
16734 		isipv6 = rack->r_is_v6;
16735 	} else {
16736 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16737 	}
16738 #endif
16739 	early = 0;
16740 	cts = tcp_get_usecs(&tv);
16741 	ms_cts = tcp_tv_to_mssectick(&tv);
16742 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16743 	    rack->rc_inp->inp_in_hpts) {
16744 		/*
16745 		 * We are on the hpts for some timer but not hptsi output.
16746 		 * Remove from the hpts unconditionally.
16747 		 */
16748 		rack_timer_cancel(tp, rack, cts, __LINE__);
16749 	}
16750 	/* Are we pacing and late? */
16751 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16752 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16753 		/* We are delayed */
16754 		delayed = cts - rack->r_ctl.rc_last_output_to;
16755 	} else {
16756 		delayed = 0;
16757 	}
16758 	/* Do the timers, which may override the pacer */
16759 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16760 		if (rack_process_timers(tp, rack, cts, hpts_calling, &doing_tlp)) {
16761 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16762 #ifdef TCP_ACCOUNTING
16763 			sched_unpin();
16764 #endif
16765 			return (0);
16766 		}
16767 	}
16768 	if (rack->rc_in_persist) {
16769 		if (rack->rc_inp->inp_in_hpts == 0) {
16770 			/* Timer is not running */
16771 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16772 		}
16773 #ifdef TCP_ACCOUNTING
16774 		sched_unpin();
16775 #endif
16776 		return (0);
16777 	}
16778 	if ((rack->r_timer_override) ||
16779 	    (rack->rc_ack_can_sendout_data) ||
16780 	    (delayed) ||
16781 	    (tp->t_state < TCPS_ESTABLISHED)) {
16782 		rack->rc_ack_can_sendout_data = 0;
16783 		if (rack->rc_inp->inp_in_hpts)
16784 			tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
16785 	} else if (rack->rc_inp->inp_in_hpts) {
16786 		/*
16787 		 * On the hpts you can't pass even if ACKNOW is on, we will
16788 		 * when the hpts fires.
16789 		 */
16790 #ifdef TCP_ACCOUNTING
16791 		crtsc = get_cyclecount();
16792 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16793 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16794 		}
16795 		counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16796 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16797 			tp->tcp_cnt_counters[SND_BLOCKED]++;
16798 		}
16799 		counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16800 		sched_unpin();
16801 #endif
16802 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16803 		return (0);
16804 	}
16805 	rack->rc_inp->inp_hpts_calls = 0;
16806 	/* Finish out both pacing early and late accounting */
16807 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16808 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16809 		early = rack->r_ctl.rc_last_output_to - cts;
16810 	} else
16811 		early = 0;
16812 	if (delayed) {
16813 		rack->r_ctl.rc_agg_delayed += delayed;
16814 		rack->r_late = 1;
16815 	} else if (early) {
16816 		rack->r_ctl.rc_agg_early += early;
16817 		rack->r_early = 1;
16818 	}
16819 	/* Now that early/late accounting is done turn off the flag */
16820 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16821 	rack->r_wanted_output = 0;
16822 	rack->r_timer_override = 0;
16823 	if ((tp->t_state != rack->r_state) &&
16824 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
16825 		rack_set_state(tp, rack);
16826 	}
16827 	if ((rack->r_fast_output) &&
16828 	    (doing_tlp == 0) &&
16829 	    (tp->rcv_numsacks == 0)) {
16830 		int ret;
16831 
16832 		error = 0;
16833 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16834 		if (ret >= 0)
16835 			return(ret);
16836 		else if (error) {
16837 			inp = rack->rc_inp;
16838 			so = inp->inp_socket;
16839 			sb = &so->so_snd;
16840 			goto nomore;
16841 		}
16842 	}
16843 	inp = rack->rc_inp;
16844 	/*
16845 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
16846 	 * only allow the initial SYN or SYN|ACK and those sent
16847 	 * by the retransmit timer.
16848 	 */
16849 	if (IS_FASTOPEN(tp->t_flags) &&
16850 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
16851 	     (tp->t_state == TCPS_SYN_SENT)) &&
16852 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16853 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
16854 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16855 		so = inp->inp_socket;
16856 		sb = &so->so_snd;
16857 		goto just_return_nolock;
16858 	}
16859 	/*
16860 	 * Determine length of data that should be transmitted, and flags
16861 	 * that will be used. If there is some data or critical controls
16862 	 * (SYN, RST) to send, then transmit; otherwise, investigate
16863 	 * further.
16864 	 */
16865 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16866 	if (tp->t_idle_reduce) {
16867 		if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16868 			rack_cc_after_idle(rack, tp);
16869 	}
16870 	tp->t_flags &= ~TF_LASTIDLE;
16871 	if (idle) {
16872 		if (tp->t_flags & TF_MORETOCOME) {
16873 			tp->t_flags |= TF_LASTIDLE;
16874 			idle = 0;
16875 		}
16876 	}
16877 	if ((tp->snd_una == tp->snd_max) &&
16878 	    rack->r_ctl.rc_went_idle_time &&
16879 	    TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16880 		idle = cts - rack->r_ctl.rc_went_idle_time;
16881 		if (idle > rack_min_probertt_hold) {
16882 			/* Count as a probe rtt */
16883 			if (rack->in_probe_rtt == 0) {
16884 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
16885 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16886 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16887 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16888 			} else {
16889 				rack_exit_probertt(rack, cts);
16890 			}
16891 		}
16892 		idle = 0;
16893 	}
16894 	if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED))
16895 		rack_init_fsb_block(tp, rack);
16896 again:
16897 	/*
16898 	 * If we've recently taken a timeout, snd_max will be greater than
16899 	 * snd_nxt.  There may be SACK information that allows us to avoid
16900 	 * resending already delivered data.  Adjust snd_nxt accordingly.
16901 	 */
16902 	sendalot = 0;
16903 	cts = tcp_get_usecs(&tv);
16904 	ms_cts = tcp_tv_to_mssectick(&tv);
16905 	tso = 0;
16906 	mtu = 0;
16907 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16908 	minseg = segsiz;
16909 	if (rack->r_ctl.rc_pace_max_segs == 0)
16910 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16911 	else
16912 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16913 	sb_offset = tp->snd_max - tp->snd_una;
16914 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16915 	flags = tcp_outflags[tp->t_state];
16916 	while (rack->rc_free_cnt < rack_free_cache) {
16917 		rsm = rack_alloc(rack);
16918 		if (rsm == NULL) {
16919 			if (inp->inp_hpts_calls)
16920 				/* Retry in a ms */
16921 				slot = (1 * HPTS_USEC_IN_MSEC);
16922 			so = inp->inp_socket;
16923 			sb = &so->so_snd;
16924 			goto just_return_nolock;
16925 		}
16926 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16927 		rack->rc_free_cnt++;
16928 		rsm = NULL;
16929 	}
16930 	if (inp->inp_hpts_calls)
16931 		inp->inp_hpts_calls = 0;
16932 	sack_rxmit = 0;
16933 	len = 0;
16934 	rsm = NULL;
16935 	if (flags & TH_RST) {
16936 		SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16937 		so = inp->inp_socket;
16938 		sb = &so->so_snd;
16939 		goto send;
16940 	}
16941 	if (rack->r_ctl.rc_resend) {
16942 		/* Retransmit timer */
16943 		rsm = rack->r_ctl.rc_resend;
16944 		rack->r_ctl.rc_resend = NULL;
16945 		len = rsm->r_end - rsm->r_start;
16946 		sack_rxmit = 1;
16947 		sendalot = 0;
16948 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16949 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16950 			 __func__, __LINE__,
16951 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
16952 		sb_offset = rsm->r_start - tp->snd_una;
16953 		if (len >= segsiz)
16954 			len = segsiz;
16955 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16956 		/* We have a retransmit that takes precedence */
16957 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16958 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16959 			/* Enter recovery if not induced by a time-out */
16960 			rack->r_ctl.rc_rsm_start = rsm->r_start;
16961 			rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
16962 			rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
16963 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una);
16964 		}
16965 #ifdef INVARIANTS
16966 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16967 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16968 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
16969 		}
16970 #endif
16971 		len = rsm->r_end - rsm->r_start;
16972 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16973 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16974 			 __func__, __LINE__,
16975 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
16976 		sb_offset = rsm->r_start - tp->snd_una;
16977 		sendalot = 0;
16978 		if (len >= segsiz)
16979 			len = segsiz;
16980 		if (len > 0) {
16981 			sack_rxmit = 1;
16982 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
16983 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
16984 			    min(len, segsiz));
16985 			counter_u64_add(rack_rtm_prr_retran, 1);
16986 		}
16987 	} else if (rack->r_ctl.rc_tlpsend) {
16988 		/* Tail loss probe */
16989 		long cwin;
16990 		long tlen;
16991 
16992 		/*
16993 		 * Check if we can do a TLP with a RACK'd packet
16994 		 * this can happen if we are not doing the rack
16995 		 * cheat and we skipped to a TLP and it
16996 		 * went off.
16997 		 */
16998 		rsm = rack->r_ctl.rc_tlpsend;
16999 		/* We are doing a TLP make sure the flag is preent */
17000 		rsm->r_flags |= RACK_TLP;
17001 		rack->r_ctl.rc_tlpsend = NULL;
17002 		sack_rxmit = 1;
17003 		tlen = rsm->r_end - rsm->r_start;
17004 		if (tlen > segsiz)
17005 			tlen = segsiz;
17006 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
17007 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
17008 			 __func__, __LINE__,
17009 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
17010 		sb_offset = rsm->r_start - tp->snd_una;
17011 		cwin = min(tp->snd_wnd, tlen);
17012 		len = cwin;
17013 	}
17014 	if (rack->r_must_retran &&
17015 	    (rsm == NULL)) {
17016 		/*
17017 		 * Non-Sack and we had a RTO or Sack/non-Sack and a
17018 		 * MTU change, we need to retransmit until we reach
17019 		 * the former snd_max (rack->r_ctl.rc_snd_max_at_rto).
17020 		 */
17021 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
17022 			int sendwin, flight;
17023 
17024 			sendwin = min(tp->snd_wnd, tp->snd_cwnd);
17025 			flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
17026 			if (flight >= sendwin) {
17027 				so = inp->inp_socket;
17028 				sb = &so->so_snd;
17029 				goto just_return_nolock;
17030 			}
17031 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17032 			if (rsm == NULL) {
17033 				/* TSNH */
17034 				rack->r_must_retran = 0;
17035 				rack->r_ctl.rc_out_at_rto = 0;
17036 				rack->r_must_retran = 0;
17037 				so = inp->inp_socket;
17038 				sb = &so->so_snd;
17039 				goto just_return_nolock;
17040 			}
17041 			if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
17042 				/* It does not have the flag, we are done */
17043 				rack->r_must_retran = 0;
17044 				rack->r_ctl.rc_out_at_rto = 0;
17045 			} else {
17046 				sack_rxmit = 1;
17047 				len = rsm->r_end - rsm->r_start;
17048 				sendalot = 0;
17049 				sb_offset = rsm->r_start - tp->snd_una;
17050 				if (len >= segsiz)
17051 					len = segsiz;
17052 				/*
17053 				 * Delay removing the flag RACK_MUST_RXT so
17054 				 * that the fastpath for retransmit will
17055 				 * work with this rsm.
17056 				 */
17057 
17058 			}
17059 		} else {
17060 			/* We must be done if there is nothing outstanding */
17061 			rack->r_must_retran = 0;
17062 			rack->r_ctl.rc_out_at_rto = 0;
17063 		}
17064 	}
17065 	/*
17066 	 * Enforce a connection sendmap count limit if set
17067 	 * as long as we are not retransmiting.
17068 	 */
17069 	if ((rsm == NULL) &&
17070 	    (rack->do_detection == 0) &&
17071 	    (V_tcp_map_entries_limit > 0) &&
17072 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
17073 		counter_u64_add(rack_to_alloc_limited, 1);
17074 		if (!rack->alloc_limit_reported) {
17075 			rack->alloc_limit_reported = 1;
17076 			counter_u64_add(rack_alloc_limited_conns, 1);
17077 		}
17078 		so = inp->inp_socket;
17079 		sb = &so->so_snd;
17080 		goto just_return_nolock;
17081 	}
17082 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
17083 		/* we are retransmitting the fin */
17084 		len--;
17085 		if (len) {
17086 			/*
17087 			 * When retransmitting data do *not* include the
17088 			 * FIN. This could happen from a TLP probe.
17089 			 */
17090 			flags &= ~TH_FIN;
17091 		}
17092 	}
17093 #ifdef INVARIANTS
17094 	/* For debugging */
17095 	rack->r_ctl.rc_rsm_at_retran = rsm;
17096 #endif
17097 	if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
17098 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
17099 		int ret;
17100 
17101 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
17102 		if (ret == 0)
17103 			return (0);
17104 	}
17105 	if (rsm && (rsm->r_flags & RACK_MUST_RXT)) {
17106 		/*
17107 		 * Clear the flag in prep for the send
17108 		 * note that if we can't get an mbuf
17109 		 * and fail, we won't retransmit this
17110 		 * rsm but that should be ok (its rare).
17111 		 */
17112 		rsm->r_flags &= ~RACK_MUST_RXT;
17113 	}
17114 	so = inp->inp_socket;
17115 	sb = &so->so_snd;
17116 	if (do_a_prefetch == 0) {
17117 		kern_prefetch(sb, &do_a_prefetch);
17118 		do_a_prefetch = 1;
17119 	}
17120 #ifdef NETFLIX_SHARED_CWND
17121 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
17122 	    rack->rack_enable_scwnd) {
17123 		/* We are doing cwnd sharing */
17124 		if (rack->gp_ready &&
17125 		    (rack->rack_attempted_scwnd == 0) &&
17126 		    (rack->r_ctl.rc_scw == NULL) &&
17127 		    tp->t_lib) {
17128 			/* The pcbid is in, lets make an attempt */
17129 			counter_u64_add(rack_try_scwnd, 1);
17130 			rack->rack_attempted_scwnd = 1;
17131 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
17132 								   &rack->r_ctl.rc_scw_index,
17133 								   segsiz);
17134 		}
17135 		if (rack->r_ctl.rc_scw &&
17136 		    (rack->rack_scwnd_is_idle == 1) &&
17137 		    sbavail(&so->so_snd)) {
17138 			/* we are no longer out of data */
17139 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17140 			rack->rack_scwnd_is_idle = 0;
17141 		}
17142 		if (rack->r_ctl.rc_scw) {
17143 			/* First lets update and get the cwnd */
17144 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
17145 								    rack->r_ctl.rc_scw_index,
17146 								    tp->snd_cwnd, tp->snd_wnd, segsiz);
17147 		}
17148 	}
17149 #endif
17150 	/*
17151 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
17152 	 * state flags.
17153 	 */
17154 	if (tp->t_flags & TF_NEEDFIN)
17155 		flags |= TH_FIN;
17156 	if (tp->t_flags & TF_NEEDSYN)
17157 		flags |= TH_SYN;
17158 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
17159 		void *end_rsm;
17160 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
17161 		if (end_rsm)
17162 			kern_prefetch(end_rsm, &prefetch_rsm);
17163 		prefetch_rsm = 1;
17164 	}
17165 	SOCKBUF_LOCK(sb);
17166 	/*
17167 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
17168 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
17169 	 * negative length.  This can also occur when TCP opens up its
17170 	 * congestion window while receiving additional duplicate acks after
17171 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
17172 	 * the fast-retransmit.
17173 	 *
17174 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
17175 	 * set to snd_una, the sb_offset will be 0, and the length may wind
17176 	 * up 0.
17177 	 *
17178 	 * If sack_rxmit is true we are retransmitting from the scoreboard
17179 	 * in which case len is already set.
17180 	 */
17181 	if ((sack_rxmit == 0) &&
17182 	    (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
17183 		uint32_t avail;
17184 
17185 		avail = sbavail(sb);
17186 		if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
17187 			sb_offset = tp->snd_nxt - tp->snd_una;
17188 		else
17189 			sb_offset = 0;
17190 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
17191 			if (rack->r_ctl.rc_tlp_new_data) {
17192 				/* TLP is forcing out new data */
17193 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
17194 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
17195 				}
17196 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
17197 					if (tp->snd_wnd > sb_offset)
17198 						len = tp->snd_wnd - sb_offset;
17199 					else
17200 						len = 0;
17201 				} else {
17202 					len = rack->r_ctl.rc_tlp_new_data;
17203 				}
17204 			}  else {
17205 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
17206 			}
17207 			if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
17208 				/*
17209 				 * For prr=off, we need to send only 1 MSS
17210 				 * at a time. We do this because another sack could
17211 				 * be arriving that causes us to send retransmits and
17212 				 * we don't want to be on a long pace due to a larger send
17213 				 * that keeps us from sending out the retransmit.
17214 				 */
17215 				len = segsiz;
17216 			}
17217 		} else {
17218 			uint32_t outstanding;
17219 			/*
17220 			 * We are inside of a Fast recovery episode, this
17221 			 * is caused by a SACK or 3 dup acks. At this point
17222 			 * we have sent all the retransmissions and we rely
17223 			 * on PRR to dictate what we will send in the form of
17224 			 * new data.
17225 			 */
17226 
17227 			outstanding = tp->snd_max - tp->snd_una;
17228 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
17229 				if (tp->snd_wnd > outstanding) {
17230 					len = tp->snd_wnd - outstanding;
17231 					/* Check to see if we have the data */
17232 					if ((sb_offset + len) > avail) {
17233 						/* It does not all fit */
17234 						if (avail > sb_offset)
17235 							len = avail - sb_offset;
17236 						else
17237 							len = 0;
17238 					}
17239 				} else {
17240 					len = 0;
17241 				}
17242 			} else if (avail > sb_offset) {
17243 				len = avail - sb_offset;
17244 			} else {
17245 				len = 0;
17246 			}
17247 			if (len > 0) {
17248 				if (len > rack->r_ctl.rc_prr_sndcnt) {
17249 					len = rack->r_ctl.rc_prr_sndcnt;
17250 				}
17251 				if (len > 0) {
17252 					sub_from_prr = 1;
17253 					counter_u64_add(rack_rtm_prr_newdata, 1);
17254 				}
17255 			}
17256 			if (len > segsiz) {
17257 				/*
17258 				 * We should never send more than a MSS when
17259 				 * retransmitting or sending new data in prr
17260 				 * mode unless the override flag is on. Most
17261 				 * likely the PRR algorithm is not going to
17262 				 * let us send a lot as well :-)
17263 				 */
17264 				if (rack->r_ctl.rc_prr_sendalot == 0) {
17265 					len = segsiz;
17266 				}
17267 			} else if (len < segsiz) {
17268 				/*
17269 				 * Do we send any? The idea here is if the
17270 				 * send empty's the socket buffer we want to
17271 				 * do it. However if not then lets just wait
17272 				 * for our prr_sndcnt to get bigger.
17273 				 */
17274 				long leftinsb;
17275 
17276 				leftinsb = sbavail(sb) - sb_offset;
17277 				if (leftinsb > len) {
17278 					/* This send does not empty the sb */
17279 					len = 0;
17280 				}
17281 			}
17282 		}
17283 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
17284 		/*
17285 		 * If you have not established
17286 		 * and are not doing FAST OPEN
17287 		 * no data please.
17288 		 */
17289 		if ((sack_rxmit == 0) &&
17290 		    (!IS_FASTOPEN(tp->t_flags))){
17291 			len = 0;
17292 			sb_offset = 0;
17293 		}
17294 	}
17295 	if (prefetch_so_done == 0) {
17296 		kern_prefetch(so, &prefetch_so_done);
17297 		prefetch_so_done = 1;
17298 	}
17299 	/*
17300 	 * Lop off SYN bit if it has already been sent.  However, if this is
17301 	 * SYN-SENT state and if segment contains data and if we don't know
17302 	 * that foreign host supports TAO, suppress sending segment.
17303 	 */
17304 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
17305 	    ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
17306 		/*
17307 		 * When sending additional segments following a TFO SYN|ACK,
17308 		 * do not include the SYN bit.
17309 		 */
17310 		if (IS_FASTOPEN(tp->t_flags) &&
17311 		    (tp->t_state == TCPS_SYN_RECEIVED))
17312 			flags &= ~TH_SYN;
17313 	}
17314 	/*
17315 	 * Be careful not to send data and/or FIN on SYN segments. This
17316 	 * measure is needed to prevent interoperability problems with not
17317 	 * fully conformant TCP implementations.
17318 	 */
17319 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
17320 		len = 0;
17321 		flags &= ~TH_FIN;
17322 	}
17323 	/*
17324 	 * On TFO sockets, ensure no data is sent in the following cases:
17325 	 *
17326 	 *  - When retransmitting SYN|ACK on a passively-created socket
17327 	 *
17328 	 *  - When retransmitting SYN on an actively created socket
17329 	 *
17330 	 *  - When sending a zero-length cookie (cookie request) on an
17331 	 *    actively created socket
17332 	 *
17333 	 *  - When the socket is in the CLOSED state (RST is being sent)
17334 	 */
17335 	if (IS_FASTOPEN(tp->t_flags) &&
17336 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
17337 	     ((tp->t_state == TCPS_SYN_SENT) &&
17338 	      (tp->t_tfo_client_cookie_len == 0)) ||
17339 	     (flags & TH_RST))) {
17340 		sack_rxmit = 0;
17341 		len = 0;
17342 	}
17343 	/* Without fast-open there should never be data sent on a SYN */
17344 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
17345 		tp->snd_nxt = tp->iss;
17346 		len = 0;
17347 	}
17348 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
17349 		/* We only send 1 MSS if we have a DSACK block */
17350 		add_flag |= RACK_SENT_W_DSACK;
17351 		len = segsiz;
17352 	}
17353 	orig_len = len;
17354 	if (len <= 0) {
17355 		/*
17356 		 * If FIN has been sent but not acked, but we haven't been
17357 		 * called to retransmit, len will be < 0.  Otherwise, window
17358 		 * shrank after we sent into it.  If window shrank to 0,
17359 		 * cancel pending retransmit, pull snd_nxt back to (closed)
17360 		 * window, and set the persist timer if it isn't already
17361 		 * going.  If the window didn't close completely, just wait
17362 		 * for an ACK.
17363 		 *
17364 		 * We also do a general check here to ensure that we will
17365 		 * set the persist timer when we have data to send, but a
17366 		 * 0-byte window. This makes sure the persist timer is set
17367 		 * even if the packet hits one of the "goto send" lines
17368 		 * below.
17369 		 */
17370 		len = 0;
17371 		if ((tp->snd_wnd == 0) &&
17372 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17373 		    (tp->snd_una == tp->snd_max) &&
17374 		    (sb_offset < (int)sbavail(sb))) {
17375 			rack_enter_persist(tp, rack, cts);
17376 		}
17377 	} else if ((rsm == NULL) &&
17378 		   (doing_tlp == 0) &&
17379 		   (len < pace_max_seg)) {
17380 		/*
17381 		 * We are not sending a maximum sized segment for
17382 		 * some reason. Should we not send anything (think
17383 		 * sws or persists)?
17384 		 */
17385 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17386 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17387 		    (len < minseg) &&
17388 		    (len < (int)(sbavail(sb) - sb_offset))) {
17389 			/*
17390 			 * Here the rwnd is less than
17391 			 * the minimum pacing size, this is not a retransmit,
17392 			 * we are established and
17393 			 * the send is not the last in the socket buffer
17394 			 * we send nothing, and we may enter persists
17395 			 * if nothing is outstanding.
17396 			 */
17397 			len = 0;
17398 			if (tp->snd_max == tp->snd_una) {
17399 				/*
17400 				 * Nothing out we can
17401 				 * go into persists.
17402 				 */
17403 				rack_enter_persist(tp, rack, cts);
17404 			}
17405 		     } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
17406 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17407 			   (len < (int)(sbavail(sb) - sb_offset)) &&
17408 			   (len < minseg)) {
17409 			/*
17410 			 * Here we are not retransmitting, and
17411 			 * the cwnd is not so small that we could
17412 			 * not send at least a min size (rxt timer
17413 			 * not having gone off), We have 2 segments or
17414 			 * more already in flight, its not the tail end
17415 			 * of the socket buffer  and the cwnd is blocking
17416 			 * us from sending out a minimum pacing segment size.
17417 			 * Lets not send anything.
17418 			 */
17419 			len = 0;
17420 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
17421 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17422 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17423 			   (len < (int)(sbavail(sb) - sb_offset)) &&
17424 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
17425 			/*
17426 			 * Here we have a send window but we have
17427 			 * filled it up and we can't send another pacing segment.
17428 			 * We also have in flight more than 2 segments
17429 			 * and we are not completing the sb i.e. we allow
17430 			 * the last bytes of the sb to go out even if
17431 			 * its not a full pacing segment.
17432 			 */
17433 			len = 0;
17434 		} else if ((rack->r_ctl.crte != NULL) &&
17435 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
17436 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
17437 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
17438 			   (len < (int)(sbavail(sb) - sb_offset))) {
17439 			/*
17440 			 * Here we are doing hardware pacing, this is not a TLP,
17441 			 * we are not sending a pace max segment size, there is rwnd
17442 			 * room to send at least N pace_max_seg, the cwnd is greater
17443 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
17444 			 * more segments in flight and its not the tail of the socket buffer.
17445 			 *
17446 			 * We don't want to send instead we need to get more ack's in to
17447 			 * allow us to send a full pacing segment. Normally, if we are pacing
17448 			 * about the right speed, we should have finished our pacing
17449 			 * send as most of the acks have come back if we are at the
17450 			 * right rate. This is a bit fuzzy since return path delay
17451 			 * can delay the acks, which is why we want to make sure we
17452 			 * have cwnd space to have a bit more than a max pace segments in flight.
17453 			 *
17454 			 * If we have not gotten our acks back we are pacing at too high a
17455 			 * rate delaying will not hurt and will bring our GP estimate down by
17456 			 * injecting the delay. If we don't do this we will send
17457 			 * 2 MSS out in response to the acks being clocked in which
17458 			 * defeats the point of hw-pacing (i.e. to help us get
17459 			 * larger TSO's out).
17460 			 */
17461 			len = 0;
17462 
17463 		}
17464 
17465 	}
17466 	/* len will be >= 0 after this point. */
17467 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17468 	rack_sndbuf_autoscale(rack);
17469 	/*
17470 	 * Decide if we can use TCP Segmentation Offloading (if supported by
17471 	 * hardware).
17472 	 *
17473 	 * TSO may only be used if we are in a pure bulk sending state.  The
17474 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
17475 	 * options prevent using TSO.  With TSO the TCP header is the same
17476 	 * (except for the sequence number) for all generated packets.  This
17477 	 * makes it impossible to transmit any options which vary per
17478 	 * generated segment or packet.
17479 	 *
17480 	 * IPv4 handling has a clear separation of ip options and ip header
17481 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
17482 	 * the right thing below to provide length of just ip options and thus
17483 	 * checking for ipoptlen is enough to decide if ip options are present.
17484 	 */
17485 	ipoptlen = 0;
17486 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17487 	/*
17488 	 * Pre-calculate here as we save another lookup into the darknesses
17489 	 * of IPsec that way and can actually decide if TSO is ok.
17490 	 */
17491 #ifdef INET6
17492 	if (isipv6 && IPSEC_ENABLED(ipv6))
17493 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
17494 #ifdef INET
17495 	else
17496 #endif
17497 #endif				/* INET6 */
17498 #ifdef INET
17499 		if (IPSEC_ENABLED(ipv4))
17500 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
17501 #endif				/* INET */
17502 #endif
17503 
17504 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17505 	ipoptlen += ipsec_optlen;
17506 #endif
17507 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
17508 	    (tp->t_port == 0) &&
17509 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
17510 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
17511 	    ipoptlen == 0)
17512 		tso = 1;
17513 	{
17514 		uint32_t outstanding;
17515 
17516 		outstanding = tp->snd_max - tp->snd_una;
17517 		if (tp->t_flags & TF_SENTFIN) {
17518 			/*
17519 			 * If we sent a fin, snd_max is 1 higher than
17520 			 * snd_una
17521 			 */
17522 			outstanding--;
17523 		}
17524 		if (sack_rxmit) {
17525 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
17526 				flags &= ~TH_FIN;
17527 		} else {
17528 			if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
17529 				   sbused(sb)))
17530 				flags &= ~TH_FIN;
17531 		}
17532 	}
17533 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
17534 	    (long)TCP_MAXWIN << tp->rcv_scale);
17535 
17536 	/*
17537 	 * Sender silly window avoidance.   We transmit under the following
17538 	 * conditions when len is non-zero:
17539 	 *
17540 	 * - We have a full segment (or more with TSO) - This is the last
17541 	 * buffer in a write()/send() and we are either idle or running
17542 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
17543 	 * then 1/2 the maximum send window's worth of data (receiver may be
17544 	 * limited the window size) - we need to retransmit
17545 	 */
17546 	if (len) {
17547 		if (len >= segsiz) {
17548 			goto send;
17549 		}
17550 		/*
17551 		 * NOTE! on localhost connections an 'ack' from the remote
17552 		 * end may occur synchronously with the output and cause us
17553 		 * to flush a buffer queued with moretocome.  XXX
17554 		 *
17555 		 */
17556 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
17557 		    (idle || (tp->t_flags & TF_NODELAY)) &&
17558 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17559 		    (tp->t_flags & TF_NOPUSH) == 0) {
17560 			pass = 2;
17561 			goto send;
17562 		}
17563 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
17564 			pass = 22;
17565 			goto send;
17566 		}
17567 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
17568 			pass = 4;
17569 			goto send;
17570 		}
17571 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {	/* retransmit case */
17572 			pass = 5;
17573 			goto send;
17574 		}
17575 		if (sack_rxmit) {
17576 			pass = 6;
17577 			goto send;
17578 		}
17579 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
17580 		    (ctf_outstanding(tp) < (segsiz * 2))) {
17581 			/*
17582 			 * We have less than two MSS outstanding (delayed ack)
17583 			 * and our rwnd will not let us send a full sized
17584 			 * MSS. Lets go ahead and let this small segment
17585 			 * out because we want to try to have at least two
17586 			 * packets inflight to not be caught by delayed ack.
17587 			 */
17588 			pass = 12;
17589 			goto send;
17590 		}
17591 	}
17592 	/*
17593 	 * Sending of standalone window updates.
17594 	 *
17595 	 * Window updates are important when we close our window due to a
17596 	 * full socket buffer and are opening it again after the application
17597 	 * reads data from it.  Once the window has opened again and the
17598 	 * remote end starts to send again the ACK clock takes over and
17599 	 * provides the most current window information.
17600 	 *
17601 	 * We must avoid the silly window syndrome whereas every read from
17602 	 * the receive buffer, no matter how small, causes a window update
17603 	 * to be sent.  We also should avoid sending a flurry of window
17604 	 * updates when the socket buffer had queued a lot of data and the
17605 	 * application is doing small reads.
17606 	 *
17607 	 * Prevent a flurry of pointless window updates by only sending an
17608 	 * update when we can increase the advertized window by more than
17609 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
17610 	 * full or is very small be more aggressive and send an update
17611 	 * whenever we can increase by two mss sized segments. In all other
17612 	 * situations the ACK's to new incoming data will carry further
17613 	 * window increases.
17614 	 *
17615 	 * Don't send an independent window update if a delayed ACK is
17616 	 * pending (it will get piggy-backed on it) or the remote side
17617 	 * already has done a half-close and won't send more data.  Skip
17618 	 * this if the connection is in T/TCP half-open state.
17619 	 */
17620 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
17621 	    !(tp->t_flags & TF_DELACK) &&
17622 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
17623 		/*
17624 		 * "adv" is the amount we could increase the window, taking
17625 		 * into account that we are limited by TCP_MAXWIN <<
17626 		 * tp->rcv_scale.
17627 		 */
17628 		int32_t adv;
17629 		int oldwin;
17630 
17631 		adv = recwin;
17632 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
17633 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
17634 			if (adv > oldwin)
17635 			    adv -= oldwin;
17636 			else {
17637 				/* We can't increase the window */
17638 				adv = 0;
17639 			}
17640 		} else
17641 			oldwin = 0;
17642 
17643 		/*
17644 		 * If the new window size ends up being the same as or less
17645 		 * than the old size when it is scaled, then don't force
17646 		 * a window update.
17647 		 */
17648 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
17649 			goto dontupdate;
17650 
17651 		if (adv >= (int32_t)(2 * segsiz) &&
17652 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
17653 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
17654 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
17655 			pass = 7;
17656 			goto send;
17657 		}
17658 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
17659 			pass = 23;
17660 			goto send;
17661 		}
17662 	}
17663 dontupdate:
17664 
17665 	/*
17666 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
17667 	 * is also a catch-all for the retransmit timer timeout case.
17668 	 */
17669 	if (tp->t_flags & TF_ACKNOW) {
17670 		pass = 8;
17671 		goto send;
17672 	}
17673 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
17674 		pass = 9;
17675 		goto send;
17676 	}
17677 	/*
17678 	 * If our state indicates that FIN should be sent and we have not
17679 	 * yet done so, then we need to send.
17680 	 */
17681 	if ((flags & TH_FIN) &&
17682 	    (tp->snd_nxt == tp->snd_una)) {
17683 		pass = 11;
17684 		goto send;
17685 	}
17686 	/*
17687 	 * No reason to send a segment, just return.
17688 	 */
17689 just_return:
17690 	SOCKBUF_UNLOCK(sb);
17691 just_return_nolock:
17692 	{
17693 		int app_limited = CTF_JR_SENT_DATA;
17694 
17695 		if (tot_len_this_send > 0) {
17696 			/* Make sure snd_nxt is up to max */
17697 			rack->r_ctl.fsb.recwin = recwin;
17698 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17699 			if ((error == 0) &&
17700 			    rack_use_rfo &&
17701 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
17702 			    (ipoptlen == 0) &&
17703 			    (tp->snd_nxt == tp->snd_max) &&
17704 			    (tp->rcv_numsacks == 0) &&
17705 			    rack->r_fsb_inited &&
17706 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
17707 			    (rack->r_must_retran == 0) &&
17708 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
17709 			    (len > 0) && (orig_len > 0) &&
17710 			    (orig_len > len) &&
17711 			    ((orig_len - len) >= segsiz) &&
17712 			    ((optlen == 0) ||
17713 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17714 				/* We can send at least one more MSS using our fsb */
17715 
17716 				rack->r_fast_output = 1;
17717 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17718 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17719 				rack->r_ctl.fsb.tcp_flags = flags;
17720 				rack->r_ctl.fsb.left_to_send = orig_len - len;
17721 				if (hw_tls)
17722 					rack->r_ctl.fsb.hw_tls = 1;
17723 				else
17724 					rack->r_ctl.fsb.hw_tls = 0;
17725 				KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17726 					("rack:%p left_to_send:%u sbavail:%u out:%u",
17727 					rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17728 					 (tp->snd_max - tp->snd_una)));
17729 				if (rack->r_ctl.fsb.left_to_send < segsiz)
17730 					rack->r_fast_output = 0;
17731 				else {
17732 					if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17733 						rack->r_ctl.fsb.rfo_apply_push = 1;
17734 					else
17735 						rack->r_ctl.fsb.rfo_apply_push = 0;
17736 				}
17737 			} else
17738 				rack->r_fast_output = 0;
17739 
17740 
17741 			rack_log_fsb(rack, tp, so, flags,
17742 				     ipoptlen, orig_len, len, 0,
17743 				     1, optlen, __LINE__, 1);
17744 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17745 				tp->snd_nxt = tp->snd_max;
17746 		} else {
17747 			int end_window = 0;
17748 			uint32_t seq = tp->gput_ack;
17749 
17750 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17751 			if (rsm) {
17752 				/*
17753 				 * Mark the last sent that we just-returned (hinting
17754 				 * that delayed ack may play a role in any rtt measurement).
17755 				 */
17756 				rsm->r_just_ret = 1;
17757 			}
17758 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17759 			rack->r_ctl.rc_agg_delayed = 0;
17760 			rack->r_early = 0;
17761 			rack->r_late = 0;
17762 			rack->r_ctl.rc_agg_early = 0;
17763 			if ((ctf_outstanding(tp) +
17764 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17765 				 minseg)) >= tp->snd_wnd) {
17766 				/* We are limited by the rwnd */
17767 				app_limited = CTF_JR_RWND_LIMITED;
17768 				if (IN_FASTRECOVERY(tp->t_flags))
17769 				    rack->r_ctl.rc_prr_sndcnt = 0;
17770 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
17771 				/* We are limited by whats available -- app limited */
17772 				app_limited = CTF_JR_APP_LIMITED;
17773 				if (IN_FASTRECOVERY(tp->t_flags))
17774 				    rack->r_ctl.rc_prr_sndcnt = 0;
17775 			} else if ((idle == 0) &&
17776 				   ((tp->t_flags & TF_NODELAY) == 0) &&
17777 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17778 				   (len < segsiz)) {
17779 				/*
17780 				 * No delay is not on and the
17781 				 * user is sending less than 1MSS. This
17782 				 * brings out SWS avoidance so we
17783 				 * don't send. Another app-limited case.
17784 				 */
17785 				app_limited = CTF_JR_APP_LIMITED;
17786 			} else if (tp->t_flags & TF_NOPUSH) {
17787 				/*
17788 				 * The user has requested no push of
17789 				 * the last segment and we are
17790 				 * at the last segment. Another app
17791 				 * limited case.
17792 				 */
17793 				app_limited = CTF_JR_APP_LIMITED;
17794 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17795 				/* Its the cwnd */
17796 				app_limited = CTF_JR_CWND_LIMITED;
17797 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
17798 				   (rack->rack_no_prr == 0) &&
17799 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17800 				app_limited = CTF_JR_PRR;
17801 			} else {
17802 				/* Now why here are we not sending? */
17803 #ifdef NOW
17804 #ifdef INVARIANTS
17805 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17806 #endif
17807 #endif
17808 				app_limited = CTF_JR_ASSESSING;
17809 			}
17810 			/*
17811 			 * App limited in some fashion, for our pacing GP
17812 			 * measurements we don't want any gap (even cwnd).
17813 			 * Close  down the measurement window.
17814 			 */
17815 			if (rack_cwnd_block_ends_measure &&
17816 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
17817 			     (app_limited == CTF_JR_PRR))) {
17818 				/*
17819 				 * The reason we are not sending is
17820 				 * the cwnd (or prr). We have been configured
17821 				 * to end the measurement window in
17822 				 * this case.
17823 				 */
17824 				end_window = 1;
17825 			} else if (rack_rwnd_block_ends_measure &&
17826 				   (app_limited == CTF_JR_RWND_LIMITED)) {
17827 				/*
17828 				 * We are rwnd limited and have been
17829 				 * configured to end the measurement
17830 				 * window in this case.
17831 				 */
17832 				end_window = 1;
17833 			} else if (app_limited == CTF_JR_APP_LIMITED) {
17834 				/*
17835 				 * A true application limited period, we have
17836 				 * ran out of data.
17837 				 */
17838 				end_window = 1;
17839 			} else if (app_limited == CTF_JR_ASSESSING) {
17840 				/*
17841 				 * In the assessing case we hit the end of
17842 				 * the if/else and had no known reason
17843 				 * This will panic us under invariants..
17844 				 *
17845 				 * If we get this out in logs we need to
17846 				 * investagate which reason we missed.
17847 				 */
17848 				end_window = 1;
17849 			}
17850 			if (end_window) {
17851 				uint8_t log = 0;
17852 
17853 				/* Adjust the Gput measurement */
17854 				if ((tp->t_flags & TF_GPUTINPROG) &&
17855 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
17856 					tp->gput_ack = tp->snd_max;
17857 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
17858 						/*
17859 						 * There is not enough to measure.
17860 						 */
17861 						tp->t_flags &= ~TF_GPUTINPROG;
17862 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17863 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
17864 									   tp->gput_seq,
17865 									   0, 0, 18, __LINE__, NULL, 0);
17866 					} else
17867 						log = 1;
17868 				}
17869 				/* Mark the last packet has app limited */
17870 				rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17871 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17872 					if (rack->r_ctl.rc_app_limited_cnt == 0)
17873 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17874 					else {
17875 						/*
17876 						 * Go out to the end app limited and mark
17877 						 * this new one as next and move the end_appl up
17878 						 * to this guy.
17879 						 */
17880 						if (rack->r_ctl.rc_end_appl)
17881 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17882 						rack->r_ctl.rc_end_appl = rsm;
17883 					}
17884 					rsm->r_flags |= RACK_APP_LIMITED;
17885 					rack->r_ctl.rc_app_limited_cnt++;
17886 				}
17887 				if (log)
17888 					rack_log_pacing_delay_calc(rack,
17889 								   rack->r_ctl.rc_app_limited_cnt, seq,
17890 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
17891 			}
17892 		}
17893 		if (slot) {
17894 			/* set the rack tcb into the slot N */
17895 			counter_u64_add(rack_paced_segments, 1);
17896 		} else if (tot_len_this_send) {
17897 			counter_u64_add(rack_unpaced_segments, 1);
17898 		}
17899 		/* Check if we need to go into persists or not */
17900 		if ((tp->snd_max == tp->snd_una) &&
17901 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
17902 		    sbavail(sb) &&
17903 		    (sbavail(sb) > tp->snd_wnd) &&
17904 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17905 			/* Yes lets make sure to move to persist before timer-start */
17906 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17907 		}
17908 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17909 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17910 	}
17911 #ifdef NETFLIX_SHARED_CWND
17912 	if ((sbavail(sb) == 0) &&
17913 	    rack->r_ctl.rc_scw) {
17914 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17915 		rack->rack_scwnd_is_idle = 1;
17916 	}
17917 #endif
17918 #ifdef TCP_ACCOUNTING
17919 	if (tot_len_this_send > 0) {
17920 		crtsc = get_cyclecount();
17921 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17922 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
17923 		}
17924 		counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17925 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17926 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17927 		}
17928 		counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
17929 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17930 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17931 		}
17932 		counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
17933 	} else {
17934 		crtsc = get_cyclecount();
17935 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17936 			tp->tcp_cnt_counters[SND_LIMITED]++;
17937 		}
17938 		counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
17939 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17940 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17941 		}
17942 		counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
17943 	}
17944 	sched_unpin();
17945 #endif
17946 	return (0);
17947 
17948 send:
17949 	if (rsm || sack_rxmit)
17950 		counter_u64_add(rack_nfto_resend, 1);
17951 	else
17952 		counter_u64_add(rack_non_fto_send, 1);
17953 	if ((flags & TH_FIN) &&
17954 	    sbavail(sb)) {
17955 		/*
17956 		 * We do not transmit a FIN
17957 		 * with data outstanding. We
17958 		 * need to make it so all data
17959 		 * is acked first.
17960 		 */
17961 		flags &= ~TH_FIN;
17962 	}
17963 	/* Enforce stack imposed max seg size if we have one */
17964 	if (rack->r_ctl.rc_pace_max_segs &&
17965 	    (len > rack->r_ctl.rc_pace_max_segs)) {
17966 		mark = 1;
17967 		len = rack->r_ctl.rc_pace_max_segs;
17968 	}
17969 	SOCKBUF_LOCK_ASSERT(sb);
17970 	if (len > 0) {
17971 		if (len >= segsiz)
17972 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
17973 		else
17974 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
17975 	}
17976 	/*
17977 	 * Before ESTABLISHED, force sending of initial options unless TCP
17978 	 * set not to do any options. NOTE: we assume that the IP/TCP header
17979 	 * plus TCP options always fit in a single mbuf, leaving room for a
17980 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
17981 	 * + optlen <= MCLBYTES
17982 	 */
17983 	optlen = 0;
17984 #ifdef INET6
17985 	if (isipv6)
17986 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
17987 	else
17988 #endif
17989 		hdrlen = sizeof(struct tcpiphdr);
17990 
17991 	/*
17992 	 * Compute options for segment. We only have to care about SYN and
17993 	 * established connection segments.  Options for SYN-ACK segments
17994 	 * are handled in TCP syncache.
17995 	 */
17996 	to.to_flags = 0;
17997 	if ((tp->t_flags & TF_NOOPT) == 0) {
17998 		/* Maximum segment size. */
17999 		if (flags & TH_SYN) {
18000 			tp->snd_nxt = tp->iss;
18001 			to.to_mss = tcp_mssopt(&inp->inp_inc);
18002 			if (tp->t_port)
18003 				to.to_mss -= V_tcp_udp_tunneling_overhead;
18004 			to.to_flags |= TOF_MSS;
18005 
18006 			/*
18007 			 * On SYN or SYN|ACK transmits on TFO connections,
18008 			 * only include the TFO option if it is not a
18009 			 * retransmit, as the presence of the TFO option may
18010 			 * have caused the original SYN or SYN|ACK to have
18011 			 * been dropped by a middlebox.
18012 			 */
18013 			if (IS_FASTOPEN(tp->t_flags) &&
18014 			    (tp->t_rxtshift == 0)) {
18015 				if (tp->t_state == TCPS_SYN_RECEIVED) {
18016 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
18017 					to.to_tfo_cookie =
18018 						(u_int8_t *)&tp->t_tfo_cookie.server;
18019 					to.to_flags |= TOF_FASTOPEN;
18020 					wanted_cookie = 1;
18021 				} else if (tp->t_state == TCPS_SYN_SENT) {
18022 					to.to_tfo_len =
18023 						tp->t_tfo_client_cookie_len;
18024 					to.to_tfo_cookie =
18025 						tp->t_tfo_cookie.client;
18026 					to.to_flags |= TOF_FASTOPEN;
18027 					wanted_cookie = 1;
18028 					/*
18029 					 * If we wind up having more data to
18030 					 * send with the SYN than can fit in
18031 					 * one segment, don't send any more
18032 					 * until the SYN|ACK comes back from
18033 					 * the other end.
18034 					 */
18035 					sendalot = 0;
18036 				}
18037 			}
18038 		}
18039 		/* Window scaling. */
18040 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
18041 			to.to_wscale = tp->request_r_scale;
18042 			to.to_flags |= TOF_SCALE;
18043 		}
18044 		/* Timestamps. */
18045 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
18046 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
18047 			to.to_tsval = ms_cts + tp->ts_offset;
18048 			to.to_tsecr = tp->ts_recent;
18049 			to.to_flags |= TOF_TS;
18050 		}
18051 		/* Set receive buffer autosizing timestamp. */
18052 		if (tp->rfbuf_ts == 0 &&
18053 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
18054 			tp->rfbuf_ts = tcp_ts_getticks();
18055 		/* Selective ACK's. */
18056 		if (tp->t_flags & TF_SACK_PERMIT) {
18057 			if (flags & TH_SYN)
18058 				to.to_flags |= TOF_SACKPERM;
18059 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18060 				 tp->rcv_numsacks > 0) {
18061 				to.to_flags |= TOF_SACK;
18062 				to.to_nsacks = tp->rcv_numsacks;
18063 				to.to_sacks = (u_char *)tp->sackblks;
18064 			}
18065 		}
18066 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18067 		/* TCP-MD5 (RFC2385). */
18068 		if (tp->t_flags & TF_SIGNATURE)
18069 			to.to_flags |= TOF_SIGNATURE;
18070 #endif				/* TCP_SIGNATURE */
18071 
18072 		/* Processing the options. */
18073 		hdrlen += optlen = tcp_addoptions(&to, opt);
18074 		/*
18075 		 * If we wanted a TFO option to be added, but it was unable
18076 		 * to fit, ensure no data is sent.
18077 		 */
18078 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
18079 		    !(to.to_flags & TOF_FASTOPEN))
18080 			len = 0;
18081 	}
18082 	if (tp->t_port) {
18083 		if (V_tcp_udp_tunneling_port == 0) {
18084 			/* The port was removed?? */
18085 			SOCKBUF_UNLOCK(&so->so_snd);
18086 #ifdef TCP_ACCOUNTING
18087 			crtsc = get_cyclecount();
18088 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18089 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18090 			}
18091 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18092 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18093 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18094 			}
18095 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18096 			sched_unpin();
18097 #endif
18098 			return (EHOSTUNREACH);
18099 		}
18100 		hdrlen += sizeof(struct udphdr);
18101 	}
18102 #ifdef INET6
18103 	if (isipv6)
18104 		ipoptlen = ip6_optlen(tp->t_inpcb);
18105 	else
18106 #endif
18107 		if (tp->t_inpcb->inp_options)
18108 			ipoptlen = tp->t_inpcb->inp_options->m_len -
18109 				offsetof(struct ipoption, ipopt_list);
18110 		else
18111 			ipoptlen = 0;
18112 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18113 	ipoptlen += ipsec_optlen;
18114 #endif
18115 
18116 	/*
18117 	 * Adjust data length if insertion of options will bump the packet
18118 	 * length beyond the t_maxseg length. Clear the FIN bit because we
18119 	 * cut off the tail of the segment.
18120 	 */
18121 	if (len + optlen + ipoptlen > tp->t_maxseg) {
18122 		if (tso) {
18123 			uint32_t if_hw_tsomax;
18124 			uint32_t moff;
18125 			int32_t max_len;
18126 
18127 			/* extract TSO information */
18128 			if_hw_tsomax = tp->t_tsomax;
18129 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18130 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18131 			KASSERT(ipoptlen == 0,
18132 				("%s: TSO can't do IP options", __func__));
18133 
18134 			/*
18135 			 * Check if we should limit by maximum payload
18136 			 * length:
18137 			 */
18138 			if (if_hw_tsomax != 0) {
18139 				/* compute maximum TSO length */
18140 				max_len = (if_hw_tsomax - hdrlen -
18141 					   max_linkhdr);
18142 				if (max_len <= 0) {
18143 					len = 0;
18144 				} else if (len > max_len) {
18145 					sendalot = 1;
18146 					len = max_len;
18147 					mark = 2;
18148 				}
18149 			}
18150 			/*
18151 			 * Prevent the last segment from being fractional
18152 			 * unless the send sockbuf can be emptied:
18153 			 */
18154 			max_len = (tp->t_maxseg - optlen);
18155 			if ((sb_offset + len) < sbavail(sb)) {
18156 				moff = len % (u_int)max_len;
18157 				if (moff != 0) {
18158 					mark = 3;
18159 					len -= moff;
18160 				}
18161 			}
18162 			/*
18163 			 * In case there are too many small fragments don't
18164 			 * use TSO:
18165 			 */
18166 			if (len <= segsiz) {
18167 				mark = 4;
18168 				tso = 0;
18169 			}
18170 			/*
18171 			 * Send the FIN in a separate segment after the bulk
18172 			 * sending is done. We don't trust the TSO
18173 			 * implementations to clear the FIN flag on all but
18174 			 * the last segment.
18175 			 */
18176 			if (tp->t_flags & TF_NEEDFIN) {
18177 				sendalot = 4;
18178 			}
18179 		} else {
18180 			mark = 5;
18181 			if (optlen + ipoptlen >= tp->t_maxseg) {
18182 				/*
18183 				 * Since we don't have enough space to put
18184 				 * the IP header chain and the TCP header in
18185 				 * one packet as required by RFC 7112, don't
18186 				 * send it. Also ensure that at least one
18187 				 * byte of the payload can be put into the
18188 				 * TCP segment.
18189 				 */
18190 				SOCKBUF_UNLOCK(&so->so_snd);
18191 				error = EMSGSIZE;
18192 				sack_rxmit = 0;
18193 				goto out;
18194 			}
18195 			len = tp->t_maxseg - optlen - ipoptlen;
18196 			sendalot = 5;
18197 		}
18198 	} else {
18199 		tso = 0;
18200 		mark = 6;
18201 	}
18202 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
18203 		("%s: len > IP_MAXPACKET", __func__));
18204 #ifdef DIAGNOSTIC
18205 #ifdef INET6
18206 	if (max_linkhdr + hdrlen > MCLBYTES)
18207 #else
18208 		if (max_linkhdr + hdrlen > MHLEN)
18209 #endif
18210 			panic("tcphdr too big");
18211 #endif
18212 
18213 	/*
18214 	 * This KASSERT is here to catch edge cases at a well defined place.
18215 	 * Before, those had triggered (random) panic conditions further
18216 	 * down.
18217 	 */
18218 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
18219 	if ((len == 0) &&
18220 	    (flags & TH_FIN) &&
18221 	    (sbused(sb))) {
18222 		/*
18223 		 * We have outstanding data, don't send a fin by itself!.
18224 		 */
18225 		goto just_return;
18226 	}
18227 	/*
18228 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
18229 	 * and initialize the header from the template for sends on this
18230 	 * connection.
18231 	 */
18232 	hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
18233 	if (len) {
18234 		uint32_t max_val;
18235 		uint32_t moff;
18236 
18237 		if (rack->r_ctl.rc_pace_max_segs)
18238 			max_val = rack->r_ctl.rc_pace_max_segs;
18239 		else if (rack->rc_user_set_max_segs)
18240 			max_val = rack->rc_user_set_max_segs * segsiz;
18241 		else
18242 			max_val = len;
18243 		/*
18244 		 * We allow a limit on sending with hptsi.
18245 		 */
18246 		if (len > max_val) {
18247 			mark = 7;
18248 			len = max_val;
18249 		}
18250 #ifdef INET6
18251 		if (MHLEN < hdrlen + max_linkhdr)
18252 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18253 		else
18254 #endif
18255 			m = m_gethdr(M_NOWAIT, MT_DATA);
18256 
18257 		if (m == NULL) {
18258 			SOCKBUF_UNLOCK(sb);
18259 			error = ENOBUFS;
18260 			sack_rxmit = 0;
18261 			goto out;
18262 		}
18263 		m->m_data += max_linkhdr;
18264 		m->m_len = hdrlen;
18265 
18266 		/*
18267 		 * Start the m_copy functions from the closest mbuf to the
18268 		 * sb_offset in the socket buffer chain.
18269 		 */
18270 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
18271 		s_mb = mb;
18272 		s_moff = moff;
18273 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
18274 			m_copydata(mb, moff, (int)len,
18275 				   mtod(m, caddr_t)+hdrlen);
18276 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18277 				sbsndptr_adv(sb, mb, len);
18278 			m->m_len += len;
18279 		} else {
18280 			struct sockbuf *msb;
18281 
18282 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18283 				msb = NULL;
18284 			else
18285 				msb = sb;
18286 			m->m_next = tcp_m_copym(
18287 				mb, moff, &len,
18288 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
18289 				((rsm == NULL) ? hw_tls : 0)
18290 #ifdef NETFLIX_COPY_ARGS
18291 				, &filled_all
18292 #endif
18293 				);
18294 			if (len <= (tp->t_maxseg - optlen)) {
18295 				/*
18296 				 * Must have ran out of mbufs for the copy
18297 				 * shorten it to no longer need tso. Lets
18298 				 * not put on sendalot since we are low on
18299 				 * mbufs.
18300 				 */
18301 				tso = 0;
18302 			}
18303 			if (m->m_next == NULL) {
18304 				SOCKBUF_UNLOCK(sb);
18305 				(void)m_free(m);
18306 				error = ENOBUFS;
18307 				sack_rxmit = 0;
18308 				goto out;
18309 			}
18310 		}
18311 		if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
18312 			if (rsm && (rsm->r_flags & RACK_TLP)) {
18313 				/*
18314 				 * TLP should not count in retran count, but
18315 				 * in its own bin
18316 				 */
18317 				counter_u64_add(rack_tlp_retran, 1);
18318 				counter_u64_add(rack_tlp_retran_bytes, len);
18319 			} else {
18320 				tp->t_sndrexmitpack++;
18321 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18322 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18323 			}
18324 #ifdef STATS
18325 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18326 						 len);
18327 #endif
18328 		} else {
18329 			KMOD_TCPSTAT_INC(tcps_sndpack);
18330 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
18331 #ifdef STATS
18332 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
18333 						 len);
18334 #endif
18335 		}
18336 		/*
18337 		 * If we're sending everything we've got, set PUSH. (This
18338 		 * will keep happy those implementations which only give
18339 		 * data to the user when a buffer fills or a PUSH comes in.)
18340 		 */
18341 		if (sb_offset + len == sbused(sb) &&
18342 		    sbused(sb) &&
18343 		    !(flags & TH_SYN)) {
18344 			flags |= TH_PUSH;
18345 			add_flag |= RACK_HAD_PUSH;
18346 		}
18347 
18348 		SOCKBUF_UNLOCK(sb);
18349 	} else {
18350 		SOCKBUF_UNLOCK(sb);
18351 		if (tp->t_flags & TF_ACKNOW)
18352 			KMOD_TCPSTAT_INC(tcps_sndacks);
18353 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
18354 			KMOD_TCPSTAT_INC(tcps_sndctrl);
18355 		else
18356 			KMOD_TCPSTAT_INC(tcps_sndwinup);
18357 
18358 		m = m_gethdr(M_NOWAIT, MT_DATA);
18359 		if (m == NULL) {
18360 			error = ENOBUFS;
18361 			sack_rxmit = 0;
18362 			goto out;
18363 		}
18364 #ifdef INET6
18365 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
18366 		    MHLEN >= hdrlen) {
18367 			M_ALIGN(m, hdrlen);
18368 		} else
18369 #endif
18370 			m->m_data += max_linkhdr;
18371 		m->m_len = hdrlen;
18372 	}
18373 	SOCKBUF_UNLOCK_ASSERT(sb);
18374 	m->m_pkthdr.rcvif = (struct ifnet *)0;
18375 #ifdef MAC
18376 	mac_inpcb_create_mbuf(inp, m);
18377 #endif
18378 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
18379 #ifdef INET6
18380 		if (isipv6)
18381 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18382 		else
18383 #endif				/* INET6 */
18384 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18385 		th = rack->r_ctl.fsb.th;
18386 		udp = rack->r_ctl.fsb.udp;
18387 		if (udp) {
18388 #ifdef INET6
18389 			if (isipv6)
18390 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
18391 			else
18392 #endif				/* INET6 */
18393 				ulen = hdrlen + len - sizeof(struct ip);
18394 			udp->uh_ulen = htons(ulen);
18395 		}
18396 	} else {
18397 #ifdef INET6
18398 		if (isipv6) {
18399 			ip6 = mtod(m, struct ip6_hdr *);
18400 			if (tp->t_port) {
18401 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
18402 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18403 				udp->uh_dport = tp->t_port;
18404 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
18405 				udp->uh_ulen = htons(ulen);
18406 				th = (struct tcphdr *)(udp + 1);
18407 			} else
18408 				th = (struct tcphdr *)(ip6 + 1);
18409 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
18410 		} else
18411 #endif				/* INET6 */
18412 		{
18413 			ip = mtod(m, struct ip *);
18414 #ifdef TCPDEBUG
18415 			ipov = (struct ipovly *)ip;
18416 #endif
18417 			if (tp->t_port) {
18418 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
18419 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18420 				udp->uh_dport = tp->t_port;
18421 				ulen = hdrlen + len - sizeof(struct ip);
18422 				udp->uh_ulen = htons(ulen);
18423 				th = (struct tcphdr *)(udp + 1);
18424 			} else
18425 				th = (struct tcphdr *)(ip + 1);
18426 			tcpip_fillheaders(inp, tp->t_port, ip, th);
18427 		}
18428 	}
18429 	/*
18430 	 * Fill in fields, remembering maximum advertised window for use in
18431 	 * delaying messages about window sizes. If resending a FIN, be sure
18432 	 * not to use a new sequence number.
18433 	 */
18434 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
18435 	    tp->snd_nxt == tp->snd_max)
18436 		tp->snd_nxt--;
18437 	/*
18438 	 * If we are starting a connection, send ECN setup SYN packet. If we
18439 	 * are on a retransmit, we may resend those bits a number of times
18440 	 * as per RFC 3168.
18441 	 */
18442 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
18443 		if (tp->t_rxtshift >= 1) {
18444 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
18445 				flags |= TH_ECE | TH_CWR;
18446 		} else
18447 			flags |= TH_ECE | TH_CWR;
18448 	}
18449 	/* Handle parallel SYN for ECN */
18450 	if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18451 	    (tp->t_flags2 & TF2_ECN_SND_ECE)) {
18452 		flags |= TH_ECE;
18453 		tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18454 	}
18455 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18456 	    (tp->t_flags2 & TF2_ECN_PERMIT)) {
18457 		/*
18458 		 * If the peer has ECN, mark data packets with ECN capable
18459 		 * transmission (ECT). Ignore pure ack packets,
18460 		 * retransmissions.
18461 		 */
18462 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
18463 		    (sack_rxmit == 0)) {
18464 #ifdef INET6
18465 			if (isipv6)
18466 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
18467 			else
18468 #endif
18469 				ip->ip_tos |= IPTOS_ECN_ECT0;
18470 			KMOD_TCPSTAT_INC(tcps_ecn_ect0);
18471 			/*
18472 			 * Reply with proper ECN notifications.
18473 			 * Only set CWR on new data segments.
18474 			 */
18475 			if (tp->t_flags2 & TF2_ECN_SND_CWR) {
18476 				flags |= TH_CWR;
18477 				tp->t_flags2 &= ~TF2_ECN_SND_CWR;
18478 			}
18479 		}
18480 		if (tp->t_flags2 & TF2_ECN_SND_ECE)
18481 			flags |= TH_ECE;
18482 	}
18483 	/*
18484 	 * If we are doing retransmissions, then snd_nxt will not reflect
18485 	 * the first unsent octet.  For ACK only packets, we do not want the
18486 	 * sequence number of the retransmitted packet, we want the sequence
18487 	 * number of the next unsent octet.  So, if there is no data (and no
18488 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
18489 	 * ti_seq.  But if we are in persist state, snd_max might reflect
18490 	 * one byte beyond the right edge of the window, so use snd_nxt in
18491 	 * that case, since we know we aren't doing a retransmission.
18492 	 * (retransmit and persist are mutually exclusive...)
18493 	 */
18494 	if (sack_rxmit == 0) {
18495 		if (len || (flags & (TH_SYN | TH_FIN))) {
18496 			th->th_seq = htonl(tp->snd_nxt);
18497 			rack_seq = tp->snd_nxt;
18498 		} else {
18499 			th->th_seq = htonl(tp->snd_max);
18500 			rack_seq = tp->snd_max;
18501 		}
18502 	} else {
18503 		th->th_seq = htonl(rsm->r_start);
18504 		rack_seq = rsm->r_start;
18505 	}
18506 	th->th_ack = htonl(tp->rcv_nxt);
18507 	th->th_flags = flags;
18508 	/*
18509 	 * Calculate receive window.  Don't shrink window, but avoid silly
18510 	 * window syndrome.
18511 	 * If a RST segment is sent, advertise a window of zero.
18512 	 */
18513 	if (flags & TH_RST) {
18514 		recwin = 0;
18515 	} else {
18516 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
18517 		    recwin < (long)segsiz) {
18518 			recwin = 0;
18519 		}
18520 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
18521 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
18522 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
18523 	}
18524 
18525 	/*
18526 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
18527 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
18528 	 * handled in syncache.
18529 	 */
18530 	if (flags & TH_SYN)
18531 		th->th_win = htons((u_short)
18532 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
18533 	else {
18534 		/* Avoid shrinking window with window scaling. */
18535 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
18536 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
18537 	}
18538 	/*
18539 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
18540 	 * window.  This may cause the remote transmitter to stall.  This
18541 	 * flag tells soreceive() to disable delayed acknowledgements when
18542 	 * draining the buffer.  This can occur if the receiver is
18543 	 * attempting to read more data than can be buffered prior to
18544 	 * transmitting on the connection.
18545 	 */
18546 	if (th->th_win == 0) {
18547 		tp->t_sndzerowin++;
18548 		tp->t_flags |= TF_RXWIN0SENT;
18549 	} else
18550 		tp->t_flags &= ~TF_RXWIN0SENT;
18551 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
18552 	/* Now are we using fsb?, if so copy the template data to the mbuf */
18553 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
18554 		uint8_t *cpto;
18555 
18556 		cpto = mtod(m, uint8_t *);
18557 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18558 		/*
18559 		 * We have just copied in:
18560 		 * IP/IP6
18561 		 * <optional udphdr>
18562 		 * tcphdr (no options)
18563 		 *
18564 		 * We need to grab the correct pointers into the mbuf
18565 		 * for both the tcp header, and possibly the udp header (if tunneling).
18566 		 * We do this by using the offset in the copy buffer and adding it
18567 		 * to the mbuf base pointer (cpto).
18568 		 */
18569 #ifdef INET6
18570 		if (isipv6)
18571 			ip6 = mtod(m, struct ip6_hdr *);
18572 		else
18573 #endif				/* INET6 */
18574 			ip = mtod(m, struct ip *);
18575 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18576 		/* If we have a udp header lets set it into the mbuf as well */
18577 		if (udp)
18578 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
18579 	}
18580 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18581 	if (to.to_flags & TOF_SIGNATURE) {
18582 		/*
18583 		 * Calculate MD5 signature and put it into the place
18584 		 * determined before.
18585 		 * NOTE: since TCP options buffer doesn't point into
18586 		 * mbuf's data, calculate offset and use it.
18587 		 */
18588 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18589 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18590 			/*
18591 			 * Do not send segment if the calculation of MD5
18592 			 * digest has failed.
18593 			 */
18594 			goto out;
18595 		}
18596 	}
18597 #endif
18598 	if (optlen) {
18599 		bcopy(opt, th + 1, optlen);
18600 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18601 	}
18602 	/*
18603 	 * Put TCP length in extended header, and then checksum extended
18604 	 * header and data.
18605 	 */
18606 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
18607 #ifdef INET6
18608 	if (isipv6) {
18609 		/*
18610 		 * ip6_plen is not need to be filled now, and will be filled
18611 		 * in ip6_output.
18612 		 */
18613 		if (tp->t_port) {
18614 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18615 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18616 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18617 			th->th_sum = htons(0);
18618 			UDPSTAT_INC(udps_opackets);
18619 		} else {
18620 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18621 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18622 			th->th_sum = in6_cksum_pseudo(ip6,
18623 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18624 						      0);
18625 		}
18626 	}
18627 #endif
18628 #if defined(INET6) && defined(INET)
18629 	else
18630 #endif
18631 #ifdef INET
18632 	{
18633 		if (tp->t_port) {
18634 			m->m_pkthdr.csum_flags = CSUM_UDP;
18635 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18636 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18637 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18638 			th->th_sum = htons(0);
18639 			UDPSTAT_INC(udps_opackets);
18640 		} else {
18641 			m->m_pkthdr.csum_flags = CSUM_TCP;
18642 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18643 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
18644 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18645 									IPPROTO_TCP + len + optlen));
18646 		}
18647 		/* IP version must be set here for ipv4/ipv6 checking later */
18648 		KASSERT(ip->ip_v == IPVERSION,
18649 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
18650 	}
18651 #endif
18652 	/*
18653 	 * Enable TSO and specify the size of the segments. The TCP pseudo
18654 	 * header checksum is always provided. XXX: Fixme: This is currently
18655 	 * not the case for IPv6.
18656 	 */
18657 	if (tso) {
18658 		KASSERT(len > tp->t_maxseg - optlen,
18659 			("%s: len <= tso_segsz", __func__));
18660 		m->m_pkthdr.csum_flags |= CSUM_TSO;
18661 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
18662 	}
18663 	KASSERT(len + hdrlen == m_length(m, NULL),
18664 		("%s: mbuf chain different than expected: %d + %u != %u",
18665 		 __func__, len, hdrlen, m_length(m, NULL)));
18666 
18667 #ifdef TCP_HHOOK
18668 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
18669 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
18670 #endif
18671 	/* We're getting ready to send; log now. */
18672 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
18673 		union tcp_log_stackspecific log;
18674 
18675 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18676 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
18677 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
18678 		if (rack->rack_no_prr)
18679 			log.u_bbr.flex1 = 0;
18680 		else
18681 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18682 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18683 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18684 		log.u_bbr.flex4 = orig_len;
18685 		if (filled_all)
18686 			log.u_bbr.flex5 = 0x80000000;
18687 		else
18688 			log.u_bbr.flex5 = 0;
18689 		/* Save off the early/late values */
18690 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18691 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18692 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18693 		if (rsm || sack_rxmit) {
18694 			if (doing_tlp)
18695 				log.u_bbr.flex8 = 2;
18696 			else
18697 				log.u_bbr.flex8 = 1;
18698 		} else {
18699 			if (doing_tlp)
18700 				log.u_bbr.flex8 = 3;
18701 			else
18702 				log.u_bbr.flex8 = 0;
18703 		}
18704 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
18705 		log.u_bbr.flex7 = mark;
18706 		log.u_bbr.flex7 <<= 8;
18707 		log.u_bbr.flex7 |= pass;
18708 		log.u_bbr.pkts_out = tp->t_maxseg;
18709 		log.u_bbr.timeStamp = cts;
18710 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18711 		log.u_bbr.lt_epoch = cwnd_to_use;
18712 		log.u_bbr.delivered = sendalot;
18713 		lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
18714 				     len, &log, false, NULL, NULL, 0, &tv);
18715 	} else
18716 		lgb = NULL;
18717 
18718 	/*
18719 	 * Fill in IP length and desired time to live and send to IP level.
18720 	 * There should be a better way to handle ttl and tos; we could keep
18721 	 * them in the template, but need a way to checksum without them.
18722 	 */
18723 	/*
18724 	 * m->m_pkthdr.len should have been set before cksum calcuration,
18725 	 * because in6_cksum() need it.
18726 	 */
18727 #ifdef INET6
18728 	if (isipv6) {
18729 		/*
18730 		 * we separately set hoplimit for every segment, since the
18731 		 * user might want to change the value via setsockopt. Also,
18732 		 * desired default hop limit might be changed via Neighbor
18733 		 * Discovery.
18734 		 */
18735 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18736 
18737 		/*
18738 		 * Set the packet size here for the benefit of DTrace
18739 		 * probes. ip6_output() will set it properly; it's supposed
18740 		 * to include the option header lengths as well.
18741 		 */
18742 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18743 
18744 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18745 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18746 		else
18747 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18748 
18749 		if (tp->t_state == TCPS_SYN_SENT)
18750 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18751 
18752 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18753 		/* TODO: IPv6 IP6TOS_ECT bit on */
18754 		error = ip6_output(m,
18755 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18756 				   inp->in6p_outputopts,
18757 #else
18758 				   NULL,
18759 #endif
18760 				   &inp->inp_route6,
18761 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18762 				   NULL, NULL, inp);
18763 
18764 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18765 			mtu = inp->inp_route6.ro_nh->nh_mtu;
18766 	}
18767 #endif				/* INET6 */
18768 #if defined(INET) && defined(INET6)
18769 	else
18770 #endif
18771 #ifdef INET
18772 	{
18773 		ip->ip_len = htons(m->m_pkthdr.len);
18774 #ifdef INET6
18775 		if (inp->inp_vflag & INP_IPV6PROTO)
18776 			ip->ip_ttl = in6_selecthlim(inp, NULL);
18777 #endif				/* INET6 */
18778 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18779 		/*
18780 		 * If we do path MTU discovery, then we set DF on every
18781 		 * packet. This might not be the best thing to do according
18782 		 * to RFC3390 Section 2. However the tcp hostcache migitates
18783 		 * the problem so it affects only the first tcp connection
18784 		 * with a host.
18785 		 *
18786 		 * NB: Don't set DF on small MTU/MSS to have a safe
18787 		 * fallback.
18788 		 */
18789 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18790 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18791 			if (tp->t_port == 0 || len < V_tcp_minmss) {
18792 				ip->ip_off |= htons(IP_DF);
18793 			}
18794 		} else {
18795 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18796 		}
18797 
18798 		if (tp->t_state == TCPS_SYN_SENT)
18799 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18800 
18801 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
18802 
18803 		error = ip_output(m,
18804 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18805 				  inp->inp_options,
18806 #else
18807 				  NULL,
18808 #endif
18809 				  &inp->inp_route,
18810 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18811 				  inp);
18812 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18813 			mtu = inp->inp_route.ro_nh->nh_mtu;
18814 	}
18815 #endif				/* INET */
18816 
18817 out:
18818 	if (lgb) {
18819 		lgb->tlb_errno = error;
18820 		lgb = NULL;
18821 	}
18822 	/*
18823 	 * In transmit state, time the transmission and arrange for the
18824 	 * retransmit.  In persist state, just set snd_max.
18825 	 */
18826 	if (error == 0) {
18827 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
18828 		if (rsm && doing_tlp) {
18829 			rack->rc_last_sent_tlp_past_cumack = 0;
18830 			rack->rc_last_sent_tlp_seq_valid = 1;
18831 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18832 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18833 		}
18834 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
18835 		if (rsm && (doing_tlp == 0)) {
18836 			/* Set we retransmitted */
18837 			rack->rc_gp_saw_rec = 1;
18838 		} else {
18839 			if (cwnd_to_use > tp->snd_ssthresh) {
18840 				/* Set we sent in CA */
18841 				rack->rc_gp_saw_ca = 1;
18842 			} else {
18843 				/* Set we sent in SS */
18844 				rack->rc_gp_saw_ss = 1;
18845 			}
18846 		}
18847 		if (doing_tlp && (rsm == NULL)) {
18848 			/* Make sure new data TLP cnt is clear */
18849 			rack->r_ctl.rc_tlp_new_data = 0;
18850 		}
18851 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18852 		    (tp->t_flags & TF_SACK_PERMIT) &&
18853 		    tp->rcv_numsacks > 0)
18854 			tcp_clean_dsack_blocks(tp);
18855 		tot_len_this_send += len;
18856 		if (len == 0)
18857 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18858 		else if (len == 1) {
18859 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18860 		} else if (len > 1) {
18861 			int idx;
18862 
18863 			idx = (len / segsiz) + 3;
18864 			if (idx >= TCP_MSS_ACCT_ATIMER)
18865 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18866 			else
18867 				counter_u64_add(rack_out_size[idx], 1);
18868 		}
18869 	}
18870 	if ((rack->rack_no_prr == 0) &&
18871 	    sub_from_prr &&
18872 	    (error == 0)) {
18873 		if (rack->r_ctl.rc_prr_sndcnt >= len)
18874 			rack->r_ctl.rc_prr_sndcnt -= len;
18875 		else
18876 			rack->r_ctl.rc_prr_sndcnt = 0;
18877 	}
18878 	sub_from_prr = 0;
18879 	if (doing_tlp) {
18880 		/* Make sure the TLP is added */
18881 		add_flag |= RACK_TLP;
18882 	} else if (rsm) {
18883 		/* If its a resend without TLP then it must not have the flag */
18884 		rsm->r_flags &= ~RACK_TLP;
18885 	}
18886 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18887 			rack_to_usec_ts(&tv),
18888 			rsm, add_flag, s_mb, s_moff, hw_tls);
18889 
18890 
18891 	if ((error == 0) &&
18892 	    (len > 0) &&
18893 	    (tp->snd_una == tp->snd_max))
18894 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
18895 	{
18896 		tcp_seq startseq = tp->snd_nxt;
18897 
18898 		/* Track our lost count */
18899 		if (rsm && (doing_tlp == 0))
18900 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18901 		/*
18902 		 * Advance snd_nxt over sequence space of this segment.
18903 		 */
18904 		if (error)
18905 			/* We don't log or do anything with errors */
18906 			goto nomore;
18907 		if (doing_tlp == 0) {
18908 			if (rsm == NULL) {
18909 				/*
18910 				 * Not a retransmission of some
18911 				 * sort, new data is going out so
18912 				 * clear our TLP count and flag.
18913 				 */
18914 				rack->rc_tlp_in_progress = 0;
18915 				rack->r_ctl.rc_tlp_cnt_out = 0;
18916 			}
18917 		} else {
18918 			/*
18919 			 * We have just sent a TLP, mark that it is true
18920 			 * and make sure our in progress is set so we
18921 			 * continue to check the count.
18922 			 */
18923 			rack->rc_tlp_in_progress = 1;
18924 			rack->r_ctl.rc_tlp_cnt_out++;
18925 		}
18926 		if (flags & (TH_SYN | TH_FIN)) {
18927 			if (flags & TH_SYN)
18928 				tp->snd_nxt++;
18929 			if (flags & TH_FIN) {
18930 				tp->snd_nxt++;
18931 				tp->t_flags |= TF_SENTFIN;
18932 			}
18933 		}
18934 		/* In the ENOBUFS case we do *not* update snd_max */
18935 		if (sack_rxmit)
18936 			goto nomore;
18937 
18938 		tp->snd_nxt += len;
18939 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18940 			if (tp->snd_una == tp->snd_max) {
18941 				/*
18942 				 * Update the time we just added data since
18943 				 * none was outstanding.
18944 				 */
18945 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18946 				tp->t_acktime = ticks;
18947 			}
18948 			tp->snd_max = tp->snd_nxt;
18949 			/*
18950 			 * Time this transmission if not a retransmission and
18951 			 * not currently timing anything.
18952 			 * This is only relevant in case of switching back to
18953 			 * the base stack.
18954 			 */
18955 			if (tp->t_rtttime == 0) {
18956 				tp->t_rtttime = ticks;
18957 				tp->t_rtseq = startseq;
18958 				KMOD_TCPSTAT_INC(tcps_segstimed);
18959 			}
18960 			if (len &&
18961 			    ((tp->t_flags & TF_GPUTINPROG) == 0))
18962 				rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18963 		}
18964 		/*
18965 		 * If we are doing FO we need to update the mbuf position and subtract
18966 		 * this happens when the peer sends us duplicate information and
18967 		 * we thus want to send a DSACK.
18968 		 *
18969 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18970 		 * turned off? If not then we are going to echo multiple DSACK blocks
18971 		 * out (with the TSO), which we should not be doing.
18972 		 */
18973 		if (rack->r_fast_output && len) {
18974 			if (rack->r_ctl.fsb.left_to_send > len)
18975 				rack->r_ctl.fsb.left_to_send -= len;
18976 			else
18977 				rack->r_ctl.fsb.left_to_send = 0;
18978 			if (rack->r_ctl.fsb.left_to_send < segsiz)
18979 				rack->r_fast_output = 0;
18980 			if (rack->r_fast_output) {
18981 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18982 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18983 			}
18984 		}
18985 	}
18986 nomore:
18987 	if (error) {
18988 		rack->r_ctl.rc_agg_delayed = 0;
18989 		rack->r_early = 0;
18990 		rack->r_late = 0;
18991 		rack->r_ctl.rc_agg_early = 0;
18992 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
18993 		/*
18994 		 * Failures do not advance the seq counter above. For the
18995 		 * case of ENOBUFS we will fall out and retry in 1ms with
18996 		 * the hpts. Everything else will just have to retransmit
18997 		 * with the timer.
18998 		 *
18999 		 * In any case, we do not want to loop around for another
19000 		 * send without a good reason.
19001 		 */
19002 		sendalot = 0;
19003 		switch (error) {
19004 		case EPERM:
19005 			tp->t_softerror = error;
19006 #ifdef TCP_ACCOUNTING
19007 			crtsc = get_cyclecount();
19008 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19009 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19010 			}
19011 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19012 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19013 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19014 			}
19015 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19016 			sched_unpin();
19017 #endif
19018 			return (error);
19019 		case ENOBUFS:
19020 			/*
19021 			 * Pace us right away to retry in a some
19022 			 * time
19023 			 */
19024 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
19025 			if (rack->rc_enobuf < 0x7f)
19026 				rack->rc_enobuf++;
19027 			if (slot < (10 * HPTS_USEC_IN_MSEC))
19028 				slot = 10 * HPTS_USEC_IN_MSEC;
19029 			if (rack->r_ctl.crte != NULL) {
19030 				counter_u64_add(rack_saw_enobuf_hw, 1);
19031 				tcp_rl_log_enobuf(rack->r_ctl.crte);
19032 			}
19033 			counter_u64_add(rack_saw_enobuf, 1);
19034 			goto enobufs;
19035 		case EMSGSIZE:
19036 			/*
19037 			 * For some reason the interface we used initially
19038 			 * to send segments changed to another or lowered
19039 			 * its MTU. If TSO was active we either got an
19040 			 * interface without TSO capabilits or TSO was
19041 			 * turned off. If we obtained mtu from ip_output()
19042 			 * then update it and try again.
19043 			 */
19044 			if (tso)
19045 				tp->t_flags &= ~TF_TSO;
19046 			if (mtu != 0) {
19047 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
19048 				goto again;
19049 			}
19050 			slot = 10 * HPTS_USEC_IN_MSEC;
19051 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19052 #ifdef TCP_ACCOUNTING
19053 			crtsc = get_cyclecount();
19054 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19055 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19056 			}
19057 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19058 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19059 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19060 			}
19061 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19062 			sched_unpin();
19063 #endif
19064 			return (error);
19065 		case ENETUNREACH:
19066 			counter_u64_add(rack_saw_enetunreach, 1);
19067 		case EHOSTDOWN:
19068 		case EHOSTUNREACH:
19069 		case ENETDOWN:
19070 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
19071 				tp->t_softerror = error;
19072 			}
19073 			/* FALLTHROUGH */
19074 		default:
19075 			slot = 10 * HPTS_USEC_IN_MSEC;
19076 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19077 #ifdef TCP_ACCOUNTING
19078 			crtsc = get_cyclecount();
19079 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19080 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19081 			}
19082 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19083 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19084 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19085 			}
19086 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19087 			sched_unpin();
19088 #endif
19089 			return (error);
19090 		}
19091 	} else {
19092 		rack->rc_enobuf = 0;
19093 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
19094 			rack->r_ctl.retran_during_recovery += len;
19095 	}
19096 	KMOD_TCPSTAT_INC(tcps_sndtotal);
19097 
19098 	/*
19099 	 * Data sent (as far as we can tell). If this advertises a larger
19100 	 * window than any other segment, then remember the size of the
19101 	 * advertised window. Any pending ACK has now been sent.
19102 	 */
19103 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
19104 		tp->rcv_adv = tp->rcv_nxt + recwin;
19105 
19106 	tp->last_ack_sent = tp->rcv_nxt;
19107 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19108 enobufs:
19109 	if (sendalot) {
19110 		/* Do we need to turn off sendalot? */
19111 		if (rack->r_ctl.rc_pace_max_segs &&
19112 		    (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
19113 			/* We hit our max. */
19114 			sendalot = 0;
19115 		} else if ((rack->rc_user_set_max_segs) &&
19116 			   (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
19117 			/* We hit the user defined max */
19118 			sendalot = 0;
19119 		}
19120 	}
19121 	if ((error == 0) && (flags & TH_FIN))
19122 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
19123 	if (flags & TH_RST) {
19124 		/*
19125 		 * We don't send again after sending a RST.
19126 		 */
19127 		slot = 0;
19128 		sendalot = 0;
19129 		if (error == 0)
19130 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
19131 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
19132 		/*
19133 		 * Get our pacing rate, if an error
19134 		 * occurred in sending (ENOBUF) we would
19135 		 * hit the else if with slot preset. Other
19136 		 * errors return.
19137 		 */
19138 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
19139 	}
19140 	if (rsm &&
19141 	    (rsm->r_flags & RACK_HAS_SYN) == 0 &&
19142 	    rack->use_rack_rr) {
19143 		/* Its a retransmit and we use the rack cheat? */
19144 		if ((slot == 0) ||
19145 		    (rack->rc_always_pace == 0) ||
19146 		    (rack->r_rr_config == 1)) {
19147 			/*
19148 			 * We have no pacing set or we
19149 			 * are using old-style rack or
19150 			 * we are overriden to use the old 1ms pacing.
19151 			 */
19152 			slot = rack->r_ctl.rc_min_to;
19153 		}
19154 	}
19155 	/* We have sent clear the flag */
19156 	rack->r_ent_rec_ns = 0;
19157 	if (rack->r_must_retran) {
19158 		if (rsm) {
19159 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
19160 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
19161 				/*
19162 				 * We have retransmitted all.
19163 				 */
19164 				rack->r_must_retran = 0;
19165 				rack->r_ctl.rc_out_at_rto = 0;
19166 			}
19167 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19168 			/*
19169 			 * Sending new data will also kill
19170 			 * the loop.
19171 			 */
19172 			rack->r_must_retran = 0;
19173 			rack->r_ctl.rc_out_at_rto = 0;
19174 		}
19175 	}
19176 	rack->r_ctl.fsb.recwin = recwin;
19177 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
19178 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19179 		/*
19180 		 * We hit an RTO and now have past snd_max at the RTO
19181 		 * clear all the WAS flags.
19182 		 */
19183 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
19184 	}
19185 	if (slot) {
19186 		/* set the rack tcb into the slot N */
19187 		counter_u64_add(rack_paced_segments, 1);
19188 		if ((error == 0) &&
19189 		    rack_use_rfo &&
19190 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
19191 		    (rsm == NULL) &&
19192 		    (tp->snd_nxt == tp->snd_max) &&
19193 		    (ipoptlen == 0) &&
19194 		    (tp->rcv_numsacks == 0) &&
19195 		    rack->r_fsb_inited &&
19196 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
19197 		    (rack->r_must_retran == 0) &&
19198 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
19199 		    (len > 0) && (orig_len > 0) &&
19200 		    (orig_len > len) &&
19201 		    ((orig_len - len) >= segsiz) &&
19202 		    ((optlen == 0) ||
19203 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19204 			/* We can send at least one more MSS using our fsb */
19205 
19206 			rack->r_fast_output = 1;
19207 			rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19208 			rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19209 			rack->r_ctl.fsb.tcp_flags = flags;
19210 			rack->r_ctl.fsb.left_to_send = orig_len - len;
19211 			if (hw_tls)
19212 				rack->r_ctl.fsb.hw_tls = 1;
19213 			else
19214 				rack->r_ctl.fsb.hw_tls = 0;
19215 			KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19216 				("rack:%p left_to_send:%u sbavail:%u out:%u",
19217 				 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19218 				 (tp->snd_max - tp->snd_una)));
19219 			if (rack->r_ctl.fsb.left_to_send < segsiz)
19220 				rack->r_fast_output = 0;
19221 			else {
19222 				if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19223 					rack->r_ctl.fsb.rfo_apply_push = 1;
19224 				else
19225 					rack->r_ctl.fsb.rfo_apply_push = 0;
19226 			}
19227 		} else
19228 			rack->r_fast_output = 0;
19229 		rack_log_fsb(rack, tp, so, flags,
19230 			     ipoptlen, orig_len, len, error,
19231 			     (rsm == NULL), optlen, __LINE__, 2);
19232 	} else if (sendalot) {
19233 		int ret;
19234 
19235 		if (len)
19236 			counter_u64_add(rack_unpaced_segments, 1);
19237 		sack_rxmit = 0;
19238 		if ((error == 0) &&
19239 		    rack_use_rfo &&
19240 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
19241 		    (rsm == NULL) &&
19242 		    (ipoptlen == 0) &&
19243 		    (tp->rcv_numsacks == 0) &&
19244 		    (tp->snd_nxt == tp->snd_max) &&
19245 		    (rack->r_must_retran == 0) &&
19246 		    rack->r_fsb_inited &&
19247 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
19248 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
19249 		    (len > 0) && (orig_len > 0) &&
19250 		    (orig_len > len) &&
19251 		    ((orig_len - len) >= segsiz) &&
19252 		    ((optlen == 0) ||
19253 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19254 			/* we can use fast_output for more */
19255 
19256 			rack->r_fast_output = 1;
19257 			rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19258 			rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19259 			rack->r_ctl.fsb.tcp_flags = flags;
19260 			rack->r_ctl.fsb.left_to_send = orig_len - len;
19261 			if (hw_tls)
19262 				rack->r_ctl.fsb.hw_tls = 1;
19263 			else
19264 				rack->r_ctl.fsb.hw_tls = 0;
19265 			KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19266 				("rack:%p left_to_send:%u sbavail:%u out:%u",
19267 				 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19268 				 (tp->snd_max - tp->snd_una)));
19269 			if (rack->r_ctl.fsb.left_to_send < segsiz) {
19270 				rack->r_fast_output = 0;
19271 			}
19272 			if (rack->r_fast_output) {
19273 				if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19274 					rack->r_ctl.fsb.rfo_apply_push = 1;
19275 				else
19276 					rack->r_ctl.fsb.rfo_apply_push = 0;
19277 				rack_log_fsb(rack, tp, so, flags,
19278 					     ipoptlen, orig_len, len, error,
19279 					     (rsm == NULL), optlen, __LINE__, 3);
19280 				error = 0;
19281 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19282 				if (ret >= 0)
19283 					return (ret);
19284 			        else if (error)
19285 					goto nomore;
19286 
19287 			}
19288 		}
19289 		goto again;
19290 	} else if (len) {
19291 		counter_u64_add(rack_unpaced_segments, 1);
19292 	}
19293 	/* Assure when we leave that snd_nxt will point to top */
19294 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
19295 		tp->snd_nxt = tp->snd_max;
19296 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
19297 #ifdef TCP_ACCOUNTING
19298 	crtsc = get_cyclecount() - ts_val;
19299 	if (tot_len_this_send) {
19300 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19301 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
19302 		}
19303 		counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
19304 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19305 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
19306 		}
19307 		counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
19308 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19309 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
19310 		}
19311 		counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
19312 	} else {
19313 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19314 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
19315 		}
19316 		counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
19317 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19318 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
19319 		}
19320 		counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
19321 	}
19322 	sched_unpin();
19323 #endif
19324 	if (error == ENOBUFS)
19325 		error = 0;
19326 	return (error);
19327 }
19328 
19329 static void
19330 rack_update_seg(struct tcp_rack *rack)
19331 {
19332 	uint32_t orig_val;
19333 
19334 	orig_val = rack->r_ctl.rc_pace_max_segs;
19335 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
19336 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
19337 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
19338 }
19339 
19340 static void
19341 rack_mtu_change(struct tcpcb *tp)
19342 {
19343 	/*
19344 	 * The MSS may have changed
19345 	 */
19346 	struct tcp_rack *rack;
19347 	struct rack_sendmap *rsm;
19348 
19349 	rack = (struct tcp_rack *)tp->t_fb_ptr;
19350 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
19351 		/*
19352 		 * The MTU has changed we need to resend everything
19353 		 * since all we have sent is lost. We first fix
19354 		 * up the mtu though.
19355 		 */
19356 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
19357 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
19358 		rack_remxt_tmr(tp);
19359 		rack->r_fast_output = 0;
19360 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
19361 						rack->r_ctl.rc_sacked);
19362 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
19363 		rack->r_must_retran = 1;
19364 		/* Mark all inflight to needing to be rxt'd */
19365 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
19366 			rsm->r_flags |= RACK_MUST_RXT;
19367 		}
19368 	}
19369 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
19370 	/* We don't use snd_nxt to retransmit */
19371 	tp->snd_nxt = tp->snd_max;
19372 }
19373 
19374 static int
19375 rack_set_profile(struct tcp_rack *rack, int prof)
19376 {
19377 	int err = EINVAL;
19378 	if (prof == 1) {
19379 		/* pace_always=1 */
19380 		if (rack->rc_always_pace == 0) {
19381 			if (tcp_can_enable_pacing() == 0)
19382 				return (EBUSY);
19383 		}
19384 		rack->rc_always_pace = 1;
19385 		if (rack->use_fixed_rate || rack->gp_ready)
19386 			rack_set_cc_pacing(rack);
19387 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19388 		rack->rack_attempt_hdwr_pace = 0;
19389 		/* cmpack=1 */
19390 		if (rack_use_cmp_acks)
19391 			rack->r_use_cmp_ack = 1;
19392 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19393 		    rack->r_use_cmp_ack)
19394 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19395 		/* scwnd=1 */
19396 		rack->rack_enable_scwnd = 1;
19397 		/* dynamic=100 */
19398 		rack->rc_gp_dyn_mul = 1;
19399 		/* gp_inc_ca */
19400 		rack->r_ctl.rack_per_of_gp_ca = 100;
19401 		/* rrr_conf=3 */
19402 		rack->r_rr_config = 3;
19403 		/* npush=2 */
19404 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19405 		/* fillcw=1 */
19406 		rack->rc_pace_to_cwnd = 1;
19407 		rack->rc_pace_fill_if_rttin_range = 0;
19408 		rack->rtt_limit_mul = 0;
19409 		/* noprr=1 */
19410 		rack->rack_no_prr = 1;
19411 		/* lscwnd=1 */
19412 		rack->r_limit_scw = 1;
19413 		/* gp_inc_rec */
19414 		rack->r_ctl.rack_per_of_gp_rec = 90;
19415 		err = 0;
19416 
19417 	} else if (prof == 3) {
19418 		/* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
19419 		/* pace_always=1 */
19420 		if (rack->rc_always_pace == 0) {
19421 			if (tcp_can_enable_pacing() == 0)
19422 				return (EBUSY);
19423 		}
19424 		rack->rc_always_pace = 1;
19425 		if (rack->use_fixed_rate || rack->gp_ready)
19426 			rack_set_cc_pacing(rack);
19427 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19428 		rack->rack_attempt_hdwr_pace = 0;
19429 		/* cmpack=1 */
19430 		if (rack_use_cmp_acks)
19431 			rack->r_use_cmp_ack = 1;
19432 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19433 		    rack->r_use_cmp_ack)
19434 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19435 		/* scwnd=1 */
19436 		rack->rack_enable_scwnd = 1;
19437 		/* dynamic=100 */
19438 		rack->rc_gp_dyn_mul = 1;
19439 		/* gp_inc_ca */
19440 		rack->r_ctl.rack_per_of_gp_ca = 100;
19441 		/* rrr_conf=3 */
19442 		rack->r_rr_config = 3;
19443 		/* npush=2 */
19444 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19445 		/* fillcw=2 */
19446 		rack->rc_pace_to_cwnd = 1;
19447 		rack->r_fill_less_agg = 1;
19448 		rack->rc_pace_fill_if_rttin_range = 0;
19449 		rack->rtt_limit_mul = 0;
19450 		/* noprr=1 */
19451 		rack->rack_no_prr = 1;
19452 		/* lscwnd=1 */
19453 		rack->r_limit_scw = 1;
19454 		/* gp_inc_rec */
19455 		rack->r_ctl.rack_per_of_gp_rec = 90;
19456 		err = 0;
19457 
19458 
19459 	} else if (prof == 2) {
19460 		/* cmpack=1 */
19461 		if (rack->rc_always_pace == 0) {
19462 			if (tcp_can_enable_pacing() == 0)
19463 				return (EBUSY);
19464 		}
19465 		rack->rc_always_pace = 1;
19466 		if (rack->use_fixed_rate || rack->gp_ready)
19467 			rack_set_cc_pacing(rack);
19468 		rack->r_use_cmp_ack = 1;
19469 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19470 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19471 		/* pace_always=1 */
19472 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19473 		/* scwnd=1 */
19474 		rack->rack_enable_scwnd = 1;
19475 		/* dynamic=100 */
19476 		rack->rc_gp_dyn_mul = 1;
19477 		rack->r_ctl.rack_per_of_gp_ca = 100;
19478 		/* rrr_conf=3 */
19479 		rack->r_rr_config = 3;
19480 		/* npush=2 */
19481 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19482 		/* fillcw=1 */
19483 		rack->rc_pace_to_cwnd = 1;
19484 		rack->rc_pace_fill_if_rttin_range = 0;
19485 		rack->rtt_limit_mul = 0;
19486 		/* noprr=1 */
19487 		rack->rack_no_prr = 1;
19488 		/* lscwnd=0 */
19489 		rack->r_limit_scw = 0;
19490 		err = 0;
19491 	} else if (prof == 0) {
19492 		/* This changes things back to the default settings */
19493 		err = 0;
19494 		if (rack->rc_always_pace) {
19495 			tcp_decrement_paced_conn();
19496 			rack_undo_cc_pacing(rack);
19497 			rack->rc_always_pace = 0;
19498 		}
19499 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
19500 			rack->rc_always_pace = 1;
19501 			if (rack->use_fixed_rate || rack->gp_ready)
19502 				rack_set_cc_pacing(rack);
19503 		} else
19504 			rack->rc_always_pace = 0;
19505 		if (rack_dsack_std_based & 0x1) {
19506 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
19507 			rack->rc_rack_tmr_std_based = 1;
19508 		}
19509 		if (rack_dsack_std_based & 0x2) {
19510 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
19511 			rack->rc_rack_use_dsack = 1;
19512 		}
19513 		if (rack_use_cmp_acks)
19514 			rack->r_use_cmp_ack = 1;
19515 		else
19516 			rack->r_use_cmp_ack = 0;
19517 		if (rack_disable_prr)
19518 			rack->rack_no_prr = 1;
19519 		else
19520 			rack->rack_no_prr = 0;
19521 		if (rack_gp_no_rec_chg)
19522 			rack->rc_gp_no_rec_chg = 1;
19523 		else
19524 			rack->rc_gp_no_rec_chg = 0;
19525 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
19526 			rack->r_mbuf_queue = 1;
19527 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19528 				rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19529 			rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19530 		} else {
19531 			rack->r_mbuf_queue = 0;
19532 			rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19533 		}
19534 		if (rack_enable_shared_cwnd)
19535 			rack->rack_enable_scwnd = 1;
19536 		else
19537 			rack->rack_enable_scwnd = 0;
19538 		if (rack_do_dyn_mul) {
19539 			/* When dynamic adjustment is on CA needs to start at 100% */
19540 			rack->rc_gp_dyn_mul = 1;
19541 			if (rack_do_dyn_mul >= 100)
19542 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
19543 		} else {
19544 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
19545 			rack->rc_gp_dyn_mul = 0;
19546 		}
19547 		rack->r_rr_config = 0;
19548 		rack->r_ctl.rc_no_push_at_mrtt = 0;
19549 		rack->rc_pace_to_cwnd = 0;
19550 		rack->rc_pace_fill_if_rttin_range = 0;
19551 		rack->rtt_limit_mul = 0;
19552 
19553 		if (rack_enable_hw_pacing)
19554 			rack->rack_hdw_pace_ena = 1;
19555 		else
19556 			rack->rack_hdw_pace_ena = 0;
19557 		if (rack_disable_prr)
19558 			rack->rack_no_prr = 1;
19559 		else
19560 			rack->rack_no_prr = 0;
19561 		if (rack_limits_scwnd)
19562 			rack->r_limit_scw  = 1;
19563 		else
19564 			rack->r_limit_scw  = 0;
19565 		err = 0;
19566 	}
19567 	return (err);
19568 }
19569 
19570 static int
19571 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
19572 {
19573 	struct deferred_opt_list *dol;
19574 
19575 	dol = malloc(sizeof(struct deferred_opt_list),
19576 		     M_TCPFSB, M_NOWAIT|M_ZERO);
19577 	if (dol == NULL) {
19578 		/*
19579 		 * No space yikes -- fail out..
19580 		 */
19581 		return (0);
19582 	}
19583 	dol->optname = sopt_name;
19584 	dol->optval = loptval;
19585 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
19586 	return (1);
19587 }
19588 
19589 static int
19590 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
19591 		    uint32_t optval, uint64_t loptval)
19592 {
19593 	struct epoch_tracker et;
19594 	struct sockopt sopt;
19595 	struct cc_newreno_opts opt;
19596 	uint64_t val;
19597 	int error = 0;
19598 	uint16_t ca, ss;
19599 
19600 	switch (sopt_name) {
19601 
19602 	case TCP_RACK_DSACK_OPT:
19603 		RACK_OPTS_INC(tcp_rack_dsack_opt);
19604 		if (optval & 0x1) {
19605 			rack->rc_rack_tmr_std_based = 1;
19606 		} else {
19607 			rack->rc_rack_tmr_std_based = 0;
19608 		}
19609 		if (optval & 0x2) {
19610 			rack->rc_rack_use_dsack = 1;
19611 		} else {
19612 			rack->rc_rack_use_dsack = 0;
19613 		}
19614 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
19615 		break;
19616 	case TCP_RACK_PACING_BETA:
19617 		RACK_OPTS_INC(tcp_rack_beta);
19618 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
19619 			/* This only works for newreno. */
19620 			error = EINVAL;
19621 			break;
19622 		}
19623 		if (rack->rc_pacing_cc_set) {
19624 			/*
19625 			 * Set them into the real CC module
19626 			 * whats in the rack pcb is the old values
19627 			 * to be used on restoral/
19628 			 */
19629 			sopt.sopt_dir = SOPT_SET;
19630 			opt.name = CC_NEWRENO_BETA;
19631 			opt.val = optval;
19632 			if (CC_ALGO(tp)->ctl_output != NULL)
19633 				error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
19634 			else {
19635 				error = ENOENT;
19636 				break;
19637 			}
19638 		} else {
19639 			/*
19640 			 * Not pacing yet so set it into our local
19641 			 * rack pcb storage.
19642 			 */
19643 			rack->r_ctl.rc_saved_beta.beta = optval;
19644 		}
19645 		break;
19646 	case TCP_RACK_TIMER_SLOP:
19647 		RACK_OPTS_INC(tcp_rack_timer_slop);
19648 		rack->r_ctl.timer_slop = optval;
19649 		if (rack->rc_tp->t_srtt) {
19650 			/*
19651 			 * If we have an SRTT lets update t_rxtcur
19652 			 * to have the new slop.
19653 			 */
19654 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
19655 					   rack_rto_min, rack_rto_max,
19656 					   rack->r_ctl.timer_slop);
19657 		}
19658 		break;
19659 	case TCP_RACK_PACING_BETA_ECN:
19660 		RACK_OPTS_INC(tcp_rack_beta_ecn);
19661 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
19662 			/* This only works for newreno. */
19663 			error = EINVAL;
19664 			break;
19665 		}
19666 		if (rack->rc_pacing_cc_set) {
19667 			/*
19668 			 * Set them into the real CC module
19669 			 * whats in the rack pcb is the old values
19670 			 * to be used on restoral/
19671 			 */
19672 			sopt.sopt_dir = SOPT_SET;
19673 			opt.name = CC_NEWRENO_BETA_ECN;
19674 			opt.val = optval;
19675 			if (CC_ALGO(tp)->ctl_output != NULL)
19676 				error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
19677 			else
19678 				error = ENOENT;
19679 		} else {
19680 			/*
19681 			 * Not pacing yet so set it into our local
19682 			 * rack pcb storage.
19683 			 */
19684 			rack->r_ctl.rc_saved_beta.beta_ecn = optval;
19685 			rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
19686 		}
19687 		break;
19688 	case TCP_DEFER_OPTIONS:
19689 		RACK_OPTS_INC(tcp_defer_opt);
19690 		if (optval) {
19691 			if (rack->gp_ready) {
19692 				/* Too late */
19693 				error = EINVAL;
19694 				break;
19695 			}
19696 			rack->defer_options = 1;
19697 		} else
19698 			rack->defer_options = 0;
19699 		break;
19700 	case TCP_RACK_MEASURE_CNT:
19701 		RACK_OPTS_INC(tcp_rack_measure_cnt);
19702 		if (optval && (optval <= 0xff)) {
19703 			rack->r_ctl.req_measurements = optval;
19704 		} else
19705 			error = EINVAL;
19706 		break;
19707 	case TCP_REC_ABC_VAL:
19708 		RACK_OPTS_INC(tcp_rec_abc_val);
19709 		if (optval > 0)
19710 			rack->r_use_labc_for_rec = 1;
19711 		else
19712 			rack->r_use_labc_for_rec = 0;
19713 		break;
19714 	case TCP_RACK_ABC_VAL:
19715 		RACK_OPTS_INC(tcp_rack_abc_val);
19716 		if ((optval > 0) && (optval < 255))
19717 			rack->rc_labc = optval;
19718 		else
19719 			error = EINVAL;
19720 		break;
19721 	case TCP_HDWR_UP_ONLY:
19722 		RACK_OPTS_INC(tcp_pacing_up_only);
19723 		if (optval)
19724 			rack->r_up_only = 1;
19725 		else
19726 			rack->r_up_only = 0;
19727 		break;
19728 	case TCP_PACING_RATE_CAP:
19729 		RACK_OPTS_INC(tcp_pacing_rate_cap);
19730 		rack->r_ctl.bw_rate_cap = loptval;
19731 		break;
19732 	case TCP_RACK_PROFILE:
19733 		RACK_OPTS_INC(tcp_profile);
19734 		error = rack_set_profile(rack, optval);
19735 		break;
19736 	case TCP_USE_CMP_ACKS:
19737 		RACK_OPTS_INC(tcp_use_cmp_acks);
19738 		if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
19739 			/* You can't turn it off once its on! */
19740 			error = EINVAL;
19741 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
19742 			rack->r_use_cmp_ack = 1;
19743 			rack->r_mbuf_queue = 1;
19744 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19745 		}
19746 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
19747 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19748 		break;
19749 	case TCP_SHARED_CWND_TIME_LIMIT:
19750 		RACK_OPTS_INC(tcp_lscwnd);
19751 		if (optval)
19752 			rack->r_limit_scw = 1;
19753 		else
19754 			rack->r_limit_scw = 0;
19755 		break;
19756  	case TCP_RACK_PACE_TO_FILL:
19757 		RACK_OPTS_INC(tcp_fillcw);
19758 		if (optval == 0)
19759 			rack->rc_pace_to_cwnd = 0;
19760 		else {
19761 			rack->rc_pace_to_cwnd = 1;
19762 			if (optval > 1)
19763 				rack->r_fill_less_agg = 1;
19764 		}
19765 		if ((optval >= rack_gp_rtt_maxmul) &&
19766 		    rack_gp_rtt_maxmul &&
19767 		    (optval < 0xf)) {
19768 			rack->rc_pace_fill_if_rttin_range = 1;
19769 			rack->rtt_limit_mul = optval;
19770 		} else {
19771 			rack->rc_pace_fill_if_rttin_range = 0;
19772 			rack->rtt_limit_mul = 0;
19773 		}
19774 		break;
19775 	case TCP_RACK_NO_PUSH_AT_MAX:
19776 		RACK_OPTS_INC(tcp_npush);
19777 		if (optval == 0)
19778 			rack->r_ctl.rc_no_push_at_mrtt = 0;
19779 		else if (optval < 0xff)
19780 			rack->r_ctl.rc_no_push_at_mrtt = optval;
19781 		else
19782 			error = EINVAL;
19783 		break;
19784 	case TCP_SHARED_CWND_ENABLE:
19785 		RACK_OPTS_INC(tcp_rack_scwnd);
19786 		if (optval == 0)
19787 			rack->rack_enable_scwnd = 0;
19788 		else
19789 			rack->rack_enable_scwnd = 1;
19790 		break;
19791 	case TCP_RACK_MBUF_QUEUE:
19792 		/* Now do we use the LRO mbuf-queue feature */
19793 		RACK_OPTS_INC(tcp_rack_mbufq);
19794 		if (optval || rack->r_use_cmp_ack)
19795 			rack->r_mbuf_queue = 1;
19796 		else
19797 			rack->r_mbuf_queue = 0;
19798 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19799 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19800 		else
19801 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19802 		break;
19803 	case TCP_RACK_NONRXT_CFG_RATE:
19804 		RACK_OPTS_INC(tcp_rack_cfg_rate);
19805 		if (optval == 0)
19806 			rack->rack_rec_nonrxt_use_cr = 0;
19807 		else
19808 			rack->rack_rec_nonrxt_use_cr = 1;
19809 		break;
19810 	case TCP_NO_PRR:
19811 		RACK_OPTS_INC(tcp_rack_noprr);
19812 		if (optval == 0)
19813 			rack->rack_no_prr = 0;
19814 		else if (optval == 1)
19815 			rack->rack_no_prr = 1;
19816 		else if (optval == 2)
19817 			rack->no_prr_addback = 1;
19818 		else
19819 			error = EINVAL;
19820 		break;
19821 	case TCP_TIMELY_DYN_ADJ:
19822 		RACK_OPTS_INC(tcp_timely_dyn);
19823 		if (optval == 0)
19824 			rack->rc_gp_dyn_mul = 0;
19825 		else {
19826 			rack->rc_gp_dyn_mul = 1;
19827 			if (optval >= 100) {
19828 				/*
19829 				 * If the user sets something 100 or more
19830 				 * its the gp_ca value.
19831 				 */
19832 				rack->r_ctl.rack_per_of_gp_ca  = optval;
19833 			}
19834 		}
19835 		break;
19836 	case TCP_RACK_DO_DETECTION:
19837 		RACK_OPTS_INC(tcp_rack_do_detection);
19838 		if (optval == 0)
19839 			rack->do_detection = 0;
19840 		else
19841 			rack->do_detection = 1;
19842 		break;
19843 	case TCP_RACK_TLP_USE:
19844 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19845 			error = EINVAL;
19846 			break;
19847 		}
19848 		RACK_OPTS_INC(tcp_tlp_use);
19849 		rack->rack_tlp_threshold_use = optval;
19850 		break;
19851 	case TCP_RACK_TLP_REDUCE:
19852 		/* RACK TLP cwnd reduction (bool) */
19853 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
19854 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19855 		break;
19856 	/*  Pacing related ones */
19857 	case TCP_RACK_PACE_ALWAYS:
19858 		/*
19859 		 * zero is old rack method, 1 is new
19860 		 * method using a pacing rate.
19861 		 */
19862 		RACK_OPTS_INC(tcp_rack_pace_always);
19863 		if (optval > 0) {
19864 			if (rack->rc_always_pace) {
19865 				error = EALREADY;
19866 				break;
19867 			} else if (tcp_can_enable_pacing()) {
19868 				rack->rc_always_pace = 1;
19869 				if (rack->use_fixed_rate || rack->gp_ready)
19870 					rack_set_cc_pacing(rack);
19871 			}
19872 			else {
19873 				error = ENOSPC;
19874 				break;
19875 			}
19876 		} else {
19877 			if (rack->rc_always_pace) {
19878 				tcp_decrement_paced_conn();
19879 				rack->rc_always_pace = 0;
19880 				rack_undo_cc_pacing(rack);
19881 			}
19882 		}
19883 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19884 			tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19885 		else
19886 			tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19887 		/* A rate may be set irate or other, if so set seg size */
19888 		rack_update_seg(rack);
19889 		break;
19890 	case TCP_BBR_RACK_INIT_RATE:
19891 		RACK_OPTS_INC(tcp_initial_rate);
19892 		val = optval;
19893 		/* Change from kbits per second to bytes per second */
19894 		val *= 1000;
19895 		val /= 8;
19896 		rack->r_ctl.init_rate = val;
19897 		if (rack->rc_init_win != rack_default_init_window) {
19898 			uint32_t win, snt;
19899 
19900 			/*
19901 			 * Options don't always get applied
19902 			 * in the order you think. So in order
19903 			 * to assure we update a cwnd we need
19904 			 * to check and see if we are still
19905 			 * where we should raise the cwnd.
19906 			 */
19907 			win = rc_init_window(rack);
19908 			if (SEQ_GT(tp->snd_max, tp->iss))
19909 				snt = tp->snd_max - tp->iss;
19910 			else
19911 				snt = 0;
19912 			if ((snt < win) &&
19913 			    (tp->snd_cwnd < win))
19914 				tp->snd_cwnd = win;
19915 		}
19916 		if (rack->rc_always_pace)
19917 			rack_update_seg(rack);
19918 		break;
19919 	case TCP_BBR_IWINTSO:
19920 		RACK_OPTS_INC(tcp_initial_win);
19921 		if (optval && (optval <= 0xff)) {
19922 			uint32_t win, snt;
19923 
19924 			rack->rc_init_win = optval;
19925 			win = rc_init_window(rack);
19926 			if (SEQ_GT(tp->snd_max, tp->iss))
19927 				snt = tp->snd_max - tp->iss;
19928 			else
19929 				snt = 0;
19930 			if ((snt < win) &&
19931 			    (tp->t_srtt |
19932 #ifdef NETFLIX_PEAKRATE
19933 			     tp->t_maxpeakrate |
19934 #endif
19935 			     rack->r_ctl.init_rate)) {
19936 				/*
19937 				 * We are not past the initial window
19938 				 * and we have some bases for pacing,
19939 				 * so we need to possibly adjust up
19940 				 * the cwnd. Note even if we don't set
19941 				 * the cwnd, its still ok to raise the rc_init_win
19942 				 * which can be used coming out of idle when we
19943 				 * would have a rate.
19944 				 */
19945 				if (tp->snd_cwnd < win)
19946 					tp->snd_cwnd = win;
19947 			}
19948 			if (rack->rc_always_pace)
19949 				rack_update_seg(rack);
19950 		} else
19951 			error = EINVAL;
19952 		break;
19953 	case TCP_RACK_FORCE_MSEG:
19954 		RACK_OPTS_INC(tcp_rack_force_max_seg);
19955 		if (optval)
19956 			rack->rc_force_max_seg = 1;
19957 		else
19958 			rack->rc_force_max_seg = 0;
19959 		break;
19960 	case TCP_RACK_PACE_MAX_SEG:
19961 		/* Max segments size in a pace in bytes */
19962 		RACK_OPTS_INC(tcp_rack_max_seg);
19963 		rack->rc_user_set_max_segs = optval;
19964 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
19965 		break;
19966 	case TCP_RACK_PACE_RATE_REC:
19967 		/* Set the fixed pacing rate in Bytes per second ca */
19968 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19969 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19970 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19971 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19972 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19973 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19974 		rack->use_fixed_rate = 1;
19975 		if (rack->rc_always_pace)
19976 			rack_set_cc_pacing(rack);
19977 		rack_log_pacing_delay_calc(rack,
19978 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
19979 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
19980 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19981 					   __LINE__, NULL,0);
19982 		break;
19983 
19984 	case TCP_RACK_PACE_RATE_SS:
19985 		/* Set the fixed pacing rate in Bytes per second ca */
19986 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19987 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19988 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19989 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19990 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19991 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19992 		rack->use_fixed_rate = 1;
19993 		if (rack->rc_always_pace)
19994 			rack_set_cc_pacing(rack);
19995 		rack_log_pacing_delay_calc(rack,
19996 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
19997 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
19998 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19999 					   __LINE__, NULL, 0);
20000 		break;
20001 
20002 	case TCP_RACK_PACE_RATE_CA:
20003 		/* Set the fixed pacing rate in Bytes per second ca */
20004 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
20005 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
20006 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
20007 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
20008 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
20009 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
20010 		rack->use_fixed_rate = 1;
20011 		if (rack->rc_always_pace)
20012 			rack_set_cc_pacing(rack);
20013 		rack_log_pacing_delay_calc(rack,
20014 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
20015 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
20016 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
20017 					   __LINE__, NULL, 0);
20018 		break;
20019 	case TCP_RACK_GP_INCREASE_REC:
20020 		RACK_OPTS_INC(tcp_gp_inc_rec);
20021 		rack->r_ctl.rack_per_of_gp_rec = optval;
20022 		rack_log_pacing_delay_calc(rack,
20023 					   rack->r_ctl.rack_per_of_gp_ss,
20024 					   rack->r_ctl.rack_per_of_gp_ca,
20025 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20026 					   __LINE__, NULL, 0);
20027 		break;
20028 	case TCP_RACK_GP_INCREASE_CA:
20029 		RACK_OPTS_INC(tcp_gp_inc_ca);
20030 		ca = optval;
20031 		if (ca < 100) {
20032 			/*
20033 			 * We don't allow any reduction
20034 			 * over the GP b/w.
20035 			 */
20036 			error = EINVAL;
20037 			break;
20038 		}
20039 		rack->r_ctl.rack_per_of_gp_ca = ca;
20040 		rack_log_pacing_delay_calc(rack,
20041 					   rack->r_ctl.rack_per_of_gp_ss,
20042 					   rack->r_ctl.rack_per_of_gp_ca,
20043 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20044 					   __LINE__, NULL, 0);
20045 		break;
20046 	case TCP_RACK_GP_INCREASE_SS:
20047 		RACK_OPTS_INC(tcp_gp_inc_ss);
20048 		ss = optval;
20049 		if (ss < 100) {
20050 			/*
20051 			 * We don't allow any reduction
20052 			 * over the GP b/w.
20053 			 */
20054 			error = EINVAL;
20055 			break;
20056 		}
20057 		rack->r_ctl.rack_per_of_gp_ss = ss;
20058 		rack_log_pacing_delay_calc(rack,
20059 					   rack->r_ctl.rack_per_of_gp_ss,
20060 					   rack->r_ctl.rack_per_of_gp_ca,
20061 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20062 					   __LINE__, NULL, 0);
20063 		break;
20064 	case TCP_RACK_RR_CONF:
20065 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
20066 		if (optval && optval <= 3)
20067 			rack->r_rr_config = optval;
20068 		else
20069 			rack->r_rr_config = 0;
20070 		break;
20071 	case TCP_HDWR_RATE_CAP:
20072 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
20073 		if (optval) {
20074 			if (rack->r_rack_hw_rate_caps == 0)
20075 				rack->r_rack_hw_rate_caps = 1;
20076 			else
20077 				error = EALREADY;
20078 		} else {
20079 			rack->r_rack_hw_rate_caps = 0;
20080 		}
20081 		break;
20082 	case TCP_BBR_HDWR_PACE:
20083 		RACK_OPTS_INC(tcp_hdwr_pacing);
20084 		if (optval){
20085 			if (rack->rack_hdrw_pacing == 0) {
20086 				rack->rack_hdw_pace_ena = 1;
20087 				rack->rack_attempt_hdwr_pace = 0;
20088 			} else
20089 				error = EALREADY;
20090 		} else {
20091 			rack->rack_hdw_pace_ena = 0;
20092 #ifdef RATELIMIT
20093 			if (rack->r_ctl.crte != NULL) {
20094 				rack->rack_hdrw_pacing = 0;
20095 				rack->rack_attempt_hdwr_pace = 0;
20096 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
20097 				rack->r_ctl.crte = NULL;
20098 			}
20099 #endif
20100 		}
20101 		break;
20102 	/*  End Pacing related ones */
20103 	case TCP_RACK_PRR_SENDALOT:
20104 		/* Allow PRR to send more than one seg */
20105 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
20106 		rack->r_ctl.rc_prr_sendalot = optval;
20107 		break;
20108 	case TCP_RACK_MIN_TO:
20109 		/* Minimum time between rack t-o's in ms */
20110 		RACK_OPTS_INC(tcp_rack_min_to);
20111 		rack->r_ctl.rc_min_to = optval;
20112 		break;
20113 	case TCP_RACK_EARLY_SEG:
20114 		/* If early recovery max segments */
20115 		RACK_OPTS_INC(tcp_rack_early_seg);
20116 		rack->r_ctl.rc_early_recovery_segs = optval;
20117 		break;
20118 	case TCP_RACK_ENABLE_HYSTART:
20119 	{
20120 		struct sockopt sopt;
20121 		struct cc_newreno_opts opt;
20122 
20123 		sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
20124 		sopt.sopt_dir = SOPT_SET;
20125 		opt.name = CC_NEWRENO_ENABLE_HYSTART;
20126 		opt.val = optval;
20127 		if (CC_ALGO(tp)->ctl_output != NULL)
20128 			error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
20129 		else
20130 			error = EINVAL;
20131 	}
20132 	break;
20133 	case TCP_RACK_REORD_THRESH:
20134 		/* RACK reorder threshold (shift amount) */
20135 		RACK_OPTS_INC(tcp_rack_reord_thresh);
20136 		if ((optval > 0) && (optval < 31))
20137 			rack->r_ctl.rc_reorder_shift = optval;
20138 		else
20139 			error = EINVAL;
20140 		break;
20141 	case TCP_RACK_REORD_FADE:
20142 		/* Does reordering fade after ms time */
20143 		RACK_OPTS_INC(tcp_rack_reord_fade);
20144 		rack->r_ctl.rc_reorder_fade = optval;
20145 		break;
20146 	case TCP_RACK_TLP_THRESH:
20147 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
20148 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
20149 		if (optval)
20150 			rack->r_ctl.rc_tlp_threshold = optval;
20151 		else
20152 			error = EINVAL;
20153 		break;
20154 	case TCP_BBR_USE_RACK_RR:
20155 		RACK_OPTS_INC(tcp_rack_rr);
20156 		if (optval)
20157 			rack->use_rack_rr = 1;
20158 		else
20159 			rack->use_rack_rr = 0;
20160 		break;
20161 	case TCP_FAST_RSM_HACK:
20162 		RACK_OPTS_INC(tcp_rack_fastrsm_hack);
20163 		if (optval)
20164 			rack->fast_rsm_hack = 1;
20165 		else
20166 			rack->fast_rsm_hack = 0;
20167 		break;
20168 	case TCP_RACK_PKT_DELAY:
20169 		/* RACK added ms i.e. rack-rtt + reord + N */
20170 		RACK_OPTS_INC(tcp_rack_pkt_delay);
20171 		rack->r_ctl.rc_pkt_delay = optval;
20172 		break;
20173 	case TCP_DELACK:
20174 		RACK_OPTS_INC(tcp_rack_delayed_ack);
20175 		if (optval == 0)
20176 			tp->t_delayed_ack = 0;
20177 		else
20178 			tp->t_delayed_ack = 1;
20179 		if (tp->t_flags & TF_DELACK) {
20180 			tp->t_flags &= ~TF_DELACK;
20181 			tp->t_flags |= TF_ACKNOW;
20182 			NET_EPOCH_ENTER(et);
20183 			rack_output(tp);
20184 			NET_EPOCH_EXIT(et);
20185 		}
20186 		break;
20187 
20188 	case TCP_BBR_RACK_RTT_USE:
20189 		RACK_OPTS_INC(tcp_rack_rtt_use);
20190 		if ((optval != USE_RTT_HIGH) &&
20191 		    (optval != USE_RTT_LOW) &&
20192 		    (optval != USE_RTT_AVG))
20193 			error = EINVAL;
20194 		else
20195 			rack->r_ctl.rc_rate_sample_method = optval;
20196 		break;
20197 	case TCP_DATA_AFTER_CLOSE:
20198 		RACK_OPTS_INC(tcp_data_after_close);
20199 		if (optval)
20200 			rack->rc_allow_data_af_clo = 1;
20201 		else
20202 			rack->rc_allow_data_af_clo = 0;
20203 		break;
20204 	default:
20205 		break;
20206 	}
20207 #ifdef NETFLIX_STATS
20208 	tcp_log_socket_option(tp, sopt_name, optval, error);
20209 #endif
20210 	return (error);
20211 }
20212 
20213 
20214 static void
20215 rack_apply_deferred_options(struct tcp_rack *rack)
20216 {
20217 	struct deferred_opt_list *dol, *sdol;
20218 	uint32_t s_optval;
20219 
20220 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
20221 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
20222 		/* Disadvantage of deferal is you loose the error return */
20223 		s_optval = (uint32_t)dol->optval;
20224 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
20225 		free(dol, M_TCPDO);
20226 	}
20227 }
20228 
20229 static void
20230 rack_hw_tls_change(struct tcpcb *tp, int chg)
20231 {
20232 	/*
20233 	 * HW tls state has changed.. fix all
20234 	 * rsm's in flight.
20235 	 */
20236 	struct tcp_rack *rack;
20237 	struct rack_sendmap *rsm;
20238 
20239 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20240 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
20241 		if (chg)
20242 			rsm->r_hw_tls = 1;
20243 		else
20244 			rsm->r_hw_tls = 0;
20245 	}
20246 	if (chg)
20247 		rack->r_ctl.fsb.hw_tls = 1;
20248 	else
20249 		rack->r_ctl.fsb.hw_tls = 0;
20250 }
20251 
20252 static int
20253 rack_pru_options(struct tcpcb *tp, int flags)
20254 {
20255 	if (flags & PRUS_OOB)
20256 		return (EOPNOTSUPP);
20257 	return (0);
20258 }
20259 
20260 static struct tcp_function_block __tcp_rack = {
20261 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
20262 	.tfb_tcp_output = rack_output,
20263 	.tfb_do_queued_segments = ctf_do_queued_segments,
20264 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
20265 	.tfb_tcp_do_segment = rack_do_segment,
20266 	.tfb_tcp_ctloutput = rack_ctloutput,
20267 	.tfb_tcp_fb_init = rack_init,
20268 	.tfb_tcp_fb_fini = rack_fini,
20269 	.tfb_tcp_timer_stop_all = rack_stopall,
20270 	.tfb_tcp_timer_activate = rack_timer_activate,
20271 	.tfb_tcp_timer_active = rack_timer_active,
20272 	.tfb_tcp_timer_stop = rack_timer_stop,
20273 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
20274 	.tfb_tcp_handoff_ok = rack_handoff_ok,
20275 	.tfb_tcp_mtu_chg = rack_mtu_change,
20276 	.tfb_pru_options = rack_pru_options,
20277 	.tfb_hwtls_change = rack_hw_tls_change,
20278 };
20279 
20280 /*
20281  * rack_ctloutput() must drop the inpcb lock before performing copyin on
20282  * socket option arguments.  When it re-acquires the lock after the copy, it
20283  * has to revalidate that the connection is still valid for the socket
20284  * option.
20285  */
20286 static int
20287 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
20288     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
20289 {
20290 #ifdef INET6
20291 	struct ip6_hdr *ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
20292 #endif
20293 #ifdef INET
20294 	struct ip *ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
20295 #endif
20296 	uint64_t loptval;
20297 	int32_t error = 0, optval;
20298 
20299 	switch (sopt->sopt_level) {
20300 #ifdef INET6
20301 	case IPPROTO_IPV6:
20302 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
20303 		switch (sopt->sopt_name) {
20304 		case IPV6_USE_MIN_MTU:
20305 			tcp6_use_min_mtu(tp);
20306 			break;
20307 		case IPV6_TCLASS:
20308 			/*
20309 			 * The DSCP codepoint has changed, update the fsb.
20310 			 */
20311 			ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
20312 			    (rack->rc_inp->inp_flow & IPV6_FLOWINFO_MASK);
20313 			break;
20314 		}
20315 		INP_WUNLOCK(inp);
20316 		return (0);
20317 #endif
20318 #ifdef INET
20319 	case IPPROTO_IP:
20320 		switch (sopt->sopt_name) {
20321 		case IP_TOS:
20322 			/*
20323 			 * The DSCP codepoint has changed, update the fsb.
20324 			 */
20325 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
20326 			break;
20327 		case IP_TTL:
20328 			/*
20329 			 * The TTL has changed, update the fsb.
20330 			 */
20331 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
20332 			break;
20333 		}
20334 		INP_WUNLOCK(inp);
20335 		return (0);
20336 #endif
20337 	}
20338 
20339 	switch (sopt->sopt_name) {
20340 	case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
20341 	/*  Pacing related ones */
20342 	case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
20343 	case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
20344 	case TCP_BBR_IWINTSO:			/*  URL:tso_iwin */
20345 	case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
20346 	case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
20347 	case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
20348 	case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
20349 	case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
20350 	case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
20351 	case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
20352 	case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
20353 	case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
20354 	case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
20355 	case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
20356 	case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
20357 	case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
20358        /* End pacing related */
20359 	case TCP_FAST_RSM_HACK:			/*  URL:frsm_hack */
20360 	case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
20361 	case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
20362 	case TCP_RACK_MIN_TO:			/*  URL:min_to */
20363 	case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
20364 	case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
20365 	case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
20366 	case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
20367 	case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
20368 	case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
20369 	case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
20370 	case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
20371 	case TCP_RACK_DO_DETECTION:		/*  URL:detect */
20372 	case TCP_NO_PRR:			/*  URL:noprr */
20373 	case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
20374 	case TCP_DATA_AFTER_CLOSE:		/*  no URL */
20375 	case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
20376 	case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
20377 	case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
20378 	case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
20379 	case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
20380 	case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
20381 	case TCP_RACK_PROFILE:			/*  URL:profile */
20382 	case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
20383 	case TCP_RACK_ABC_VAL:			/*  URL:labc */
20384 	case TCP_REC_ABC_VAL:			/*  URL:reclabc */
20385 	case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
20386 	case TCP_DEFER_OPTIONS:			/*  URL:defer */
20387 	case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
20388 	case TCP_RACK_PACING_BETA:		/*  URL:pacing_beta */
20389 	case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
20390 	case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
20391 	case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
20392 		break;
20393 	default:
20394 		/* Filter off all unknown options to the base stack */
20395 		return (tcp_default_ctloutput(so, sopt, inp, tp));
20396 		break;
20397 	}
20398 	INP_WUNLOCK(inp);
20399 	if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
20400 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
20401 		/*
20402 		 * We truncate it down to 32 bits for the socket-option trace this
20403 		 * means rates > 34Gbps won't show right, but thats probably ok.
20404 		 */
20405 		optval = (uint32_t)loptval;
20406 	} else {
20407 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
20408 		/* Save it in 64 bit form too */
20409 		loptval = optval;
20410 	}
20411 	if (error)
20412 		return (error);
20413 	INP_WLOCK(inp);
20414 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
20415 		INP_WUNLOCK(inp);
20416 		return (ECONNRESET);
20417 	}
20418 	if (tp->t_fb != &__tcp_rack) {
20419 		INP_WUNLOCK(inp);
20420 		return (ENOPROTOOPT);
20421 	}
20422 	if (rack->defer_options && (rack->gp_ready == 0) &&
20423 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
20424 	    (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
20425 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
20426 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
20427 		/* Options are beind deferred */
20428 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
20429 			INP_WUNLOCK(inp);
20430 			return (0);
20431 		} else {
20432 			/* No memory to defer, fail */
20433 			INP_WUNLOCK(inp);
20434 			return (ENOMEM);
20435 		}
20436 	}
20437 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
20438 	INP_WUNLOCK(inp);
20439 	return (error);
20440 }
20441 
20442 static void
20443 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
20444 {
20445 
20446 	INP_WLOCK_ASSERT(tp->t_inpcb);
20447 	bzero(ti, sizeof(*ti));
20448 
20449 	ti->tcpi_state = tp->t_state;
20450 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
20451 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
20452 	if (tp->t_flags & TF_SACK_PERMIT)
20453 		ti->tcpi_options |= TCPI_OPT_SACK;
20454 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
20455 		ti->tcpi_options |= TCPI_OPT_WSCALE;
20456 		ti->tcpi_snd_wscale = tp->snd_scale;
20457 		ti->tcpi_rcv_wscale = tp->rcv_scale;
20458 	}
20459 	if (tp->t_flags2 & TF2_ECN_PERMIT)
20460 		ti->tcpi_options |= TCPI_OPT_ECN;
20461 	if (tp->t_flags & TF_FASTOPEN)
20462 		ti->tcpi_options |= TCPI_OPT_TFO;
20463 	/* still kept in ticks is t_rcvtime */
20464 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
20465 	/* Since we hold everything in precise useconds this is easy */
20466 	ti->tcpi_rtt = tp->t_srtt;
20467 	ti->tcpi_rttvar = tp->t_rttvar;
20468 	ti->tcpi_rto = tp->t_rxtcur;
20469 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
20470 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
20471 	/*
20472 	 * FreeBSD-specific extension fields for tcp_info.
20473 	 */
20474 	ti->tcpi_rcv_space = tp->rcv_wnd;
20475 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
20476 	ti->tcpi_snd_wnd = tp->snd_wnd;
20477 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
20478 	ti->tcpi_snd_nxt = tp->snd_nxt;
20479 	ti->tcpi_snd_mss = tp->t_maxseg;
20480 	ti->tcpi_rcv_mss = tp->t_maxseg;
20481 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
20482 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
20483 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
20484 #ifdef NETFLIX_STATS
20485 	ti->tcpi_total_tlp = tp->t_sndtlppack;
20486 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
20487 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
20488 #endif
20489 #ifdef TCP_OFFLOAD
20490 	if (tp->t_flags & TF_TOE) {
20491 		ti->tcpi_options |= TCPI_OPT_TOE;
20492 		tcp_offload_tcp_info(tp, ti);
20493 	}
20494 #endif
20495 }
20496 
20497 static int
20498 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
20499     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
20500 {
20501 	int32_t error, optval;
20502 	uint64_t val, loptval;
20503 	struct	tcp_info ti;
20504 	/*
20505 	 * Because all our options are either boolean or an int, we can just
20506 	 * pull everything into optval and then unlock and copy. If we ever
20507 	 * add a option that is not a int, then this will have quite an
20508 	 * impact to this routine.
20509 	 */
20510 	error = 0;
20511 	switch (sopt->sopt_name) {
20512 	case TCP_INFO:
20513 		/* First get the info filled */
20514 		rack_fill_info(tp, &ti);
20515 		/* Fix up the rtt related fields if needed */
20516 		INP_WUNLOCK(inp);
20517 		error = sooptcopyout(sopt, &ti, sizeof ti);
20518 		return (error);
20519 	/*
20520 	 * Beta is the congestion control value for NewReno that influences how
20521 	 * much of a backoff happens when loss is detected. It is normally set
20522 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
20523 	 * when you exit recovery.
20524 	 */
20525 	case TCP_RACK_PACING_BETA:
20526 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
20527 			error = EINVAL;
20528 		else if (rack->rc_pacing_cc_set == 0)
20529 			optval = rack->r_ctl.rc_saved_beta.beta;
20530 		else {
20531 			/*
20532 			 * Reach out into the CC data and report back what
20533 			 * I have previously set. Yeah it looks hackish but
20534 			 * we don't want to report the saved values.
20535 			 */
20536 			if (tp->ccv->cc_data)
20537 				optval = ((struct newreno *)tp->ccv->cc_data)->beta;
20538 			else
20539 				error = EINVAL;
20540 		}
20541 		break;
20542 		/*
20543 		 * Beta_ecn is the congestion control value for NewReno that influences how
20544 		 * much of a backoff happens when a ECN mark is detected. It is normally set
20545 		 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
20546 		 * you exit recovery. Note that classic ECN has a beta of 50, it is only
20547 		 * ABE Ecn that uses this "less" value, but we do too with pacing :)
20548 		 */
20549 
20550 	case TCP_RACK_PACING_BETA_ECN:
20551 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
20552 			error = EINVAL;
20553 		else if (rack->rc_pacing_cc_set == 0)
20554 			optval = rack->r_ctl.rc_saved_beta.beta_ecn;
20555 		else {
20556 			/*
20557 			 * Reach out into the CC data and report back what
20558 			 * I have previously set. Yeah it looks hackish but
20559 			 * we don't want to report the saved values.
20560 			 */
20561 			if (tp->ccv->cc_data)
20562 				optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn;
20563 			else
20564 				error = EINVAL;
20565 		}
20566 		break;
20567 	case TCP_RACK_DSACK_OPT:
20568 		optval = 0;
20569 		if (rack->rc_rack_tmr_std_based) {
20570 			optval |= 1;
20571 		}
20572 		if (rack->rc_rack_use_dsack) {
20573 			optval |= 2;
20574 		}
20575 		break;
20576  	case TCP_RACK_ENABLE_HYSTART:
20577 	{
20578 		struct sockopt sopt;
20579 		struct cc_newreno_opts opt;
20580 
20581 		sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
20582 		sopt.sopt_dir = SOPT_GET;
20583 		opt.name = CC_NEWRENO_ENABLE_HYSTART;
20584 		if (CC_ALGO(tp)->ctl_output != NULL)
20585 			error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
20586 		else
20587 			error = EINVAL;
20588 		optval = opt.val;
20589 	}
20590 	break;
20591 	case TCP_FAST_RSM_HACK:
20592 		optval = rack->fast_rsm_hack;
20593 		break;
20594 	case TCP_DEFER_OPTIONS:
20595 		optval = rack->defer_options;
20596 		break;
20597 	case TCP_RACK_MEASURE_CNT:
20598 		optval = rack->r_ctl.req_measurements;
20599 		break;
20600 	case TCP_REC_ABC_VAL:
20601 		optval = rack->r_use_labc_for_rec;
20602 		break;
20603 	case TCP_RACK_ABC_VAL:
20604 		optval = rack->rc_labc;
20605 		break;
20606 	case TCP_HDWR_UP_ONLY:
20607 		optval= rack->r_up_only;
20608 		break;
20609 	case TCP_PACING_RATE_CAP:
20610 		loptval = rack->r_ctl.bw_rate_cap;
20611 		break;
20612 	case TCP_RACK_PROFILE:
20613 		/* You cannot retrieve a profile, its write only */
20614 		error = EINVAL;
20615 		break;
20616 	case TCP_USE_CMP_ACKS:
20617 		optval = rack->r_use_cmp_ack;
20618 		break;
20619 	case TCP_RACK_PACE_TO_FILL:
20620 		optval = rack->rc_pace_to_cwnd;
20621 		if (optval && rack->r_fill_less_agg)
20622 			optval++;
20623 		break;
20624 	case TCP_RACK_NO_PUSH_AT_MAX:
20625 		optval = rack->r_ctl.rc_no_push_at_mrtt;
20626 		break;
20627 	case TCP_SHARED_CWND_ENABLE:
20628 		optval = rack->rack_enable_scwnd;
20629 		break;
20630 	case TCP_RACK_NONRXT_CFG_RATE:
20631 		optval = rack->rack_rec_nonrxt_use_cr;
20632 		break;
20633 	case TCP_NO_PRR:
20634 		if (rack->rack_no_prr  == 1)
20635 			optval = 1;
20636 		else if (rack->no_prr_addback == 1)
20637 			optval = 2;
20638 		else
20639 			optval = 0;
20640 		break;
20641 	case TCP_RACK_DO_DETECTION:
20642 		optval = rack->do_detection;
20643 		break;
20644 	case TCP_RACK_MBUF_QUEUE:
20645 		/* Now do we use the LRO mbuf-queue feature */
20646 		optval = rack->r_mbuf_queue;
20647 		break;
20648 	case TCP_TIMELY_DYN_ADJ:
20649 		optval = rack->rc_gp_dyn_mul;
20650 		break;
20651 	case TCP_BBR_IWINTSO:
20652 		optval = rack->rc_init_win;
20653 		break;
20654 	case TCP_RACK_TLP_REDUCE:
20655 		/* RACK TLP cwnd reduction (bool) */
20656 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
20657 		break;
20658 	case TCP_BBR_RACK_INIT_RATE:
20659 		val = rack->r_ctl.init_rate;
20660 		/* convert to kbits per sec */
20661 		val *= 8;
20662 		val /= 1000;
20663 		optval = (uint32_t)val;
20664 		break;
20665 	case TCP_RACK_FORCE_MSEG:
20666 		optval = rack->rc_force_max_seg;
20667 		break;
20668 	case TCP_RACK_PACE_MAX_SEG:
20669 		/* Max segments in a pace */
20670 		optval = rack->rc_user_set_max_segs;
20671 		break;
20672 	case TCP_RACK_PACE_ALWAYS:
20673 		/* Use the always pace method */
20674 		optval = rack->rc_always_pace;
20675 		break;
20676 	case TCP_RACK_PRR_SENDALOT:
20677 		/* Allow PRR to send more than one seg */
20678 		optval = rack->r_ctl.rc_prr_sendalot;
20679 		break;
20680 	case TCP_RACK_MIN_TO:
20681 		/* Minimum time between rack t-o's in ms */
20682 		optval = rack->r_ctl.rc_min_to;
20683 		break;
20684 	case TCP_RACK_EARLY_SEG:
20685 		/* If early recovery max segments */
20686 		optval = rack->r_ctl.rc_early_recovery_segs;
20687 		break;
20688 	case TCP_RACK_REORD_THRESH:
20689 		/* RACK reorder threshold (shift amount) */
20690 		optval = rack->r_ctl.rc_reorder_shift;
20691 		break;
20692 	case TCP_RACK_REORD_FADE:
20693 		/* Does reordering fade after ms time */
20694 		optval = rack->r_ctl.rc_reorder_fade;
20695 		break;
20696 	case TCP_BBR_USE_RACK_RR:
20697 		/* Do we use the rack cheat for rxt */
20698 		optval = rack->use_rack_rr;
20699 		break;
20700 	case TCP_RACK_RR_CONF:
20701 		optval = rack->r_rr_config;
20702 		break;
20703 	case TCP_HDWR_RATE_CAP:
20704 		optval = rack->r_rack_hw_rate_caps;
20705 		break;
20706 	case TCP_BBR_HDWR_PACE:
20707 		optval = rack->rack_hdw_pace_ena;
20708 		break;
20709 	case TCP_RACK_TLP_THRESH:
20710 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
20711 		optval = rack->r_ctl.rc_tlp_threshold;
20712 		break;
20713 	case TCP_RACK_PKT_DELAY:
20714 		/* RACK added ms i.e. rack-rtt + reord + N */
20715 		optval = rack->r_ctl.rc_pkt_delay;
20716 		break;
20717 	case TCP_RACK_TLP_USE:
20718 		optval = rack->rack_tlp_threshold_use;
20719 		break;
20720 	case TCP_RACK_PACE_RATE_CA:
20721 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
20722 		break;
20723 	case TCP_RACK_PACE_RATE_SS:
20724 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
20725 		break;
20726 	case TCP_RACK_PACE_RATE_REC:
20727 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
20728 		break;
20729 	case TCP_RACK_GP_INCREASE_SS:
20730 		optval = rack->r_ctl.rack_per_of_gp_ca;
20731 		break;
20732 	case TCP_RACK_GP_INCREASE_CA:
20733 		optval = rack->r_ctl.rack_per_of_gp_ss;
20734 		break;
20735 	case TCP_BBR_RACK_RTT_USE:
20736 		optval = rack->r_ctl.rc_rate_sample_method;
20737 		break;
20738 	case TCP_DELACK:
20739 		optval = tp->t_delayed_ack;
20740 		break;
20741 	case TCP_DATA_AFTER_CLOSE:
20742 		optval = rack->rc_allow_data_af_clo;
20743 		break;
20744 	case TCP_SHARED_CWND_TIME_LIMIT:
20745 		optval = rack->r_limit_scw;
20746 		break;
20747 	case TCP_RACK_TIMER_SLOP:
20748 		optval = rack->r_ctl.timer_slop;
20749 		break;
20750 	default:
20751 		return (tcp_default_ctloutput(so, sopt, inp, tp));
20752 		break;
20753 	}
20754 	INP_WUNLOCK(inp);
20755 	if (error == 0) {
20756 		if (TCP_PACING_RATE_CAP)
20757 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
20758 		else
20759 			error = sooptcopyout(sopt, &optval, sizeof optval);
20760 	}
20761 	return (error);
20762 }
20763 
20764 static int
20765 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
20766 {
20767 	int32_t error = EINVAL;
20768 	struct tcp_rack *rack;
20769 
20770 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20771 	if (rack == NULL) {
20772 		/* Huh? */
20773 		goto out;
20774 	}
20775 	if (sopt->sopt_dir == SOPT_SET) {
20776 		return (rack_set_sockopt(so, sopt, inp, tp, rack));
20777 	} else if (sopt->sopt_dir == SOPT_GET) {
20778 		return (rack_get_sockopt(so, sopt, inp, tp, rack));
20779 	}
20780 out:
20781 	INP_WUNLOCK(inp);
20782 	return (error);
20783 }
20784 
20785 static const char *rack_stack_names[] = {
20786 	__XSTRING(STACKNAME),
20787 #ifdef STACKALIAS
20788 	__XSTRING(STACKALIAS),
20789 #endif
20790 };
20791 
20792 static int
20793 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
20794 {
20795 	memset(mem, 0, size);
20796 	return (0);
20797 }
20798 
20799 static void
20800 rack_dtor(void *mem, int32_t size, void *arg)
20801 {
20802 
20803 }
20804 
20805 static bool rack_mod_inited = false;
20806 
20807 static int
20808 tcp_addrack(module_t mod, int32_t type, void *data)
20809 {
20810 	int32_t err = 0;
20811 	int num_stacks;
20812 
20813 	switch (type) {
20814 	case MOD_LOAD:
20815 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
20816 		    sizeof(struct rack_sendmap),
20817 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
20818 
20819 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
20820 		    sizeof(struct tcp_rack),
20821 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
20822 
20823 		sysctl_ctx_init(&rack_sysctl_ctx);
20824 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
20825 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
20826 		    OID_AUTO,
20827 #ifdef STACKALIAS
20828 		    __XSTRING(STACKALIAS),
20829 #else
20830 		    __XSTRING(STACKNAME),
20831 #endif
20832 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
20833 		    "");
20834 		if (rack_sysctl_root == NULL) {
20835 			printf("Failed to add sysctl node\n");
20836 			err = EFAULT;
20837 			goto free_uma;
20838 		}
20839 		rack_init_sysctls();
20840 		num_stacks = nitems(rack_stack_names);
20841 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
20842 		    rack_stack_names, &num_stacks);
20843 		if (err) {
20844 			printf("Failed to register %s stack name for "
20845 			    "%s module\n", rack_stack_names[num_stacks],
20846 			    __XSTRING(MODNAME));
20847 			sysctl_ctx_free(&rack_sysctl_ctx);
20848 free_uma:
20849 			uma_zdestroy(rack_zone);
20850 			uma_zdestroy(rack_pcb_zone);
20851 			rack_counter_destroy();
20852 			printf("Failed to register rack module -- err:%d\n", err);
20853 			return (err);
20854 		}
20855 		tcp_lro_reg_mbufq();
20856 		rack_mod_inited = true;
20857 		break;
20858 	case MOD_QUIESCE:
20859 		err = deregister_tcp_functions(&__tcp_rack, true, false);
20860 		break;
20861 	case MOD_UNLOAD:
20862 		err = deregister_tcp_functions(&__tcp_rack, false, true);
20863 		if (err == EBUSY)
20864 			break;
20865 		if (rack_mod_inited) {
20866 			uma_zdestroy(rack_zone);
20867 			uma_zdestroy(rack_pcb_zone);
20868 			sysctl_ctx_free(&rack_sysctl_ctx);
20869 			rack_counter_destroy();
20870 			rack_mod_inited = false;
20871 		}
20872 		tcp_lro_dereg_mbufq();
20873 		err = 0;
20874 		break;
20875 	default:
20876 		return (EOPNOTSUPP);
20877 	}
20878 	return (err);
20879 }
20880 
20881 static moduledata_t tcp_rack = {
20882 	.name = __XSTRING(MODNAME),
20883 	.evhand = tcp_addrack,
20884 	.priv = 0
20885 };
20886 
20887 MODULE_VERSION(MODNAME, 1);
20888 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
20889 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
20890