xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision 5740057b4388f116717b9614cd54cf2c755e6e7d) 
1 /*-
2  * Copyright (c) 2016-2020 Netflix, Inc.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include "opt_inet.h"
31 #include "opt_inet6.h"
32 #include "opt_ipsec.h"
33 #include "opt_ratelimit.h"
34 #include "opt_kern_tls.h"
35 #if defined(INET) || defined(INET6)
36 #include <sys/param.h>
37 #include <sys/arb.h>
38 #include <sys/module.h>
39 #include <sys/kernel.h>
40 #ifdef TCP_HHOOK
41 #include <sys/hhook.h>
42 #endif
43 #include <sys/lock.h>
44 #include <sys/malloc.h>
45 #include <sys/lock.h>
46 #include <sys/mutex.h>
47 #include <sys/mbuf.h>
48 #include <sys/proc.h>		/* for proc0 declaration */
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 #include <sys/systm.h>
53 #ifdef STATS
54 #include <sys/qmath.h>
55 #include <sys/tree.h>
56 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
57 #else
58 #include <sys/tree.h>
59 #endif
60 #include <sys/refcount.h>
61 #include <sys/queue.h>
62 #include <sys/tim_filter.h>
63 #include <sys/smp.h>
64 #include <sys/kthread.h>
65 #include <sys/kern_prefetch.h>
66 #include <sys/protosw.h>
67 #ifdef TCP_ACCOUNTING
68 #include <sys/sched.h>
69 #include <machine/cpu.h>
70 #endif
71 #include <vm/uma.h>
72 
73 #include <net/route.h>
74 #include <net/route/nhop.h>
75 #include <net/vnet.h>
76 
77 #define TCPSTATES		/* for logging */
78 
79 #include <netinet/in.h>
80 #include <netinet/in_kdtrace.h>
81 #include <netinet/in_pcb.h>
82 #include <netinet/ip.h>
83 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
84 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
85 #include <netinet/ip_var.h>
86 #include <netinet/ip6.h>
87 #include <netinet6/in6_pcb.h>
88 #include <netinet6/ip6_var.h>
89 #include <netinet/tcp.h>
90 #define	TCPOUTFLAGS
91 #include <netinet/tcp_fsm.h>
92 #include <netinet/tcp_log_buf.h>
93 #include <netinet/tcp_seq.h>
94 #include <netinet/tcp_timer.h>
95 #include <netinet/tcp_var.h>
96 #include <netinet/tcp_syncache.h>
97 #include <netinet/tcp_hpts.h>
98 #include <netinet/tcp_ratelimit.h>
99 #include <netinet/tcp_accounting.h>
100 #include <netinet/tcpip.h>
101 #include <netinet/cc/cc.h>
102 #include <netinet/cc/cc_newreno.h>
103 #include <netinet/tcp_fastopen.h>
104 #include <netinet/tcp_lro.h>
105 #ifdef NETFLIX_SHARED_CWND
106 #include <netinet/tcp_shared_cwnd.h>
107 #endif
108 #ifdef TCP_OFFLOAD
109 #include <netinet/tcp_offload.h>
110 #endif
111 #ifdef INET6
112 #include <netinet6/tcp6_var.h>
113 #endif
114 #include <netinet/tcp_ecn.h>
115 
116 #include <netipsec/ipsec_support.h>
117 
118 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
119 #include <netipsec/ipsec.h>
120 #include <netipsec/ipsec6.h>
121 #endif				/* IPSEC */
122 
123 #include <netinet/udp.h>
124 #include <netinet/udp_var.h>
125 #include <machine/in_cksum.h>
126 
127 #ifdef MAC
128 #include <security/mac/mac_framework.h>
129 #endif
130 #include "sack_filter.h"
131 #include "tcp_rack.h"
132 #include "rack_bbr_common.h"
133 
134 uma_zone_t rack_zone;
135 uma_zone_t rack_pcb_zone;
136 
137 #ifndef TICKS2SBT
138 #define	TICKS2SBT(__t)	(tick_sbt * ((sbintime_t)(__t)))
139 #endif
140 
141 VNET_DECLARE(uint32_t, newreno_beta);
142 VNET_DECLARE(uint32_t, newreno_beta_ecn);
143 #define V_newreno_beta VNET(newreno_beta)
144 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
145 
146 
147 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
148 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
149 
150 struct sysctl_ctx_list rack_sysctl_ctx;
151 struct sysctl_oid *rack_sysctl_root;
152 
153 #define CUM_ACKED 1
154 #define SACKED 2
155 
156 /*
157  * The RACK module incorporates a number of
158  * TCP ideas that have been put out into the IETF
159  * over the last few years:
160  * - Matt Mathis's Rate Halving which slowly drops
161  *    the congestion window so that the ack clock can
162  *    be maintained during a recovery.
163  * - Yuchung Cheng's RACK TCP (for which its named) that
164  *    will stop us using the number of dup acks and instead
165  *    use time as the gage of when we retransmit.
166  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
167  *    of Dukkipati et.al.
168  * RACK depends on SACK, so if an endpoint arrives that
169  * cannot do SACK the state machine below will shuttle the
170  * connection back to using the "default" TCP stack that is
171  * in FreeBSD.
172  *
173  * To implement RACK the original TCP stack was first decomposed
174  * into a functional state machine with individual states
175  * for each of the possible TCP connection states. The do_segment
176  * functions role in life is to mandate the connection supports SACK
177  * initially and then assure that the RACK state matches the conenction
178  * state before calling the states do_segment function. Each
179  * state is simplified due to the fact that the original do_segment
180  * has been decomposed and we *know* what state we are in (no
181  * switches on the state) and all tests for SACK are gone. This
182  * greatly simplifies what each state does.
183  *
184  * TCP output is also over-written with a new version since it
185  * must maintain the new rack scoreboard.
186  *
187  */
188 static int32_t rack_tlp_thresh = 1;
189 static int32_t rack_tlp_limit = 2;	/* No more than 2 TLPs w-out new data */
190 static int32_t rack_tlp_use_greater = 1;
191 static int32_t rack_reorder_thresh = 2;
192 static int32_t rack_reorder_fade = 60000000;	/* 0 - never fade, def 60,000,000
193 						 * - 60 seconds */
194 static uint8_t rack_req_measurements = 1;
195 /* Attack threshold detections */
196 static uint32_t rack_highest_sack_thresh_seen = 0;
197 static uint32_t rack_highest_move_thresh_seen = 0;
198 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
199 static int32_t rack_hw_pace_extra_slots = 2;	/* 2 extra MSS time betweens */
200 static int32_t rack_hw_rate_caps = 1; /* 1; */
201 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
202 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
203 static int32_t rack_hw_up_only = 1;
204 static int32_t rack_stats_gets_ms_rtt = 1;
205 static int32_t rack_prr_addbackmax = 2;
206 static int32_t rack_do_hystart = 0;
207 static int32_t rack_apply_rtt_with_reduced_conf = 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 /*
242  * Currently regular tcp has a rto_min of 30ms
243  * the backoff goes 12 times so that ends up
244  * being a total of 122.850 seconds before a
245  * connection is killed.
246  */
247 static uint32_t rack_def_data_window = 20;
248 static uint32_t rack_goal_bdp = 2;
249 static uint32_t rack_min_srtts = 1;
250 static uint32_t rack_min_measure_usec = 0;
251 static int32_t rack_tlp_min = 10000;	/* 10ms */
252 static int32_t rack_rto_min = 30000;	/* 30,000 usec same as main freebsd */
253 static int32_t rack_rto_max = 4000000;	/* 4 seconds in usec's */
254 static const int32_t rack_free_cache = 2;
255 static int32_t rack_hptsi_segments = 40;
256 static int32_t rack_rate_sample_method = USE_RTT_LOW;
257 static int32_t rack_pace_every_seg = 0;
258 static int32_t rack_delayed_ack_time = 40000;	/* 40ms in usecs */
259 static int32_t rack_slot_reduction = 4;
260 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
261 static int32_t rack_cwnd_block_ends_measure = 0;
262 static int32_t rack_rwnd_block_ends_measure = 0;
263 static int32_t rack_def_profile = 0;
264 
265 static int32_t rack_lower_cwnd_at_tlp = 0;
266 static int32_t rack_limited_retran = 0;
267 static int32_t rack_always_send_oldest = 0;
268 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
269 
270 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
271 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
272 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
273 
274 /* Probertt */
275 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
276 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
277 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
278 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
279 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
280 
281 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
282 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
283 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
284 static uint32_t rack_probertt_use_min_rtt_exit = 0;
285 static uint32_t rack_probe_rtt_sets_cwnd = 0;
286 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
287 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in usecs */
288 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
289 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction */
290 static uint32_t rack_min_probertt_hold = 40000;		/* Equal to delayed ack time */
291 static uint32_t rack_probertt_filter_life = 10000000;
292 static uint32_t rack_probertt_lower_within = 10;
293 static uint32_t rack_min_rtt_movement = 250000;	/* Must move at least 250ms (in microseconds)  to count as a lowering */
294 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
295 static int32_t rack_probertt_clear_is = 1;
296 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
297 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
298 
299 /* Part of pacing */
300 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
301 
302 /* Timely information */
303 /* Combine these two gives the range of 'no change' to bw */
304 /* ie the up/down provide the upper and lower bound */
305 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
306 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
307 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
308 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
309 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
310 static int32_t rack_gp_decrease_per = 20;	/* 20% decrease in multiplier */
311 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multiplier */
312 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
313 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
314 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
315 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
316 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
317 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
318 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
319 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
320 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
321 static int32_t rack_use_max_for_nobackoff = 0;
322 static int32_t rack_timely_int_timely_only = 0;	/* do interim timely's only use the timely algo (no b/w changes)? */
323 static int32_t rack_timely_no_stopping = 0;
324 static int32_t rack_down_raise_thresh = 100;
325 static int32_t rack_req_segs = 1;
326 static uint64_t rack_bw_rate_cap = 0;
327 static uint32_t rack_trace_point_config = 0;
328 static uint32_t rack_trace_point_bb_mode = 4;
329 static int32_t rack_trace_point_count = 0;
330 
331 
332 /* Weird delayed ack mode */
333 static int32_t rack_use_imac_dack = 0;
334 /* Rack specific counters */
335 counter_u64_t rack_saw_enobuf;
336 counter_u64_t rack_saw_enobuf_hw;
337 counter_u64_t rack_saw_enetunreach;
338 counter_u64_t rack_persists_sends;
339 counter_u64_t rack_persists_acks;
340 counter_u64_t rack_persists_loss;
341 counter_u64_t rack_persists_lost_ends;
342 #ifdef INVARIANTS
343 counter_u64_t rack_adjust_map_bw;
344 #endif
345 /* Tail loss probe counters */
346 counter_u64_t rack_tlp_tot;
347 counter_u64_t rack_tlp_newdata;
348 counter_u64_t rack_tlp_retran;
349 counter_u64_t rack_tlp_retran_bytes;
350 counter_u64_t rack_to_tot;
351 counter_u64_t rack_hot_alloc;
352 counter_u64_t rack_to_alloc;
353 counter_u64_t rack_to_alloc_hard;
354 counter_u64_t rack_to_alloc_emerg;
355 counter_u64_t rack_to_alloc_limited;
356 counter_u64_t rack_alloc_limited_conns;
357 counter_u64_t rack_split_limited;
358 
359 counter_u64_t rack_multi_single_eq;
360 counter_u64_t rack_proc_non_comp_ack;
361 
362 counter_u64_t rack_fto_send;
363 counter_u64_t rack_fto_rsm_send;
364 counter_u64_t rack_nfto_resend;
365 counter_u64_t rack_non_fto_send;
366 counter_u64_t rack_extended_rfo;
367 
368 counter_u64_t rack_sack_proc_all;
369 counter_u64_t rack_sack_proc_short;
370 counter_u64_t rack_sack_proc_restart;
371 counter_u64_t rack_sack_attacks_detected;
372 counter_u64_t rack_sack_attacks_reversed;
373 counter_u64_t rack_sack_used_next_merge;
374 counter_u64_t rack_sack_splits;
375 counter_u64_t rack_sack_used_prev_merge;
376 counter_u64_t rack_sack_skipped_acked;
377 counter_u64_t rack_ack_total;
378 counter_u64_t rack_express_sack;
379 counter_u64_t rack_sack_total;
380 counter_u64_t rack_move_none;
381 counter_u64_t rack_move_some;
382 
383 counter_u64_t rack_input_idle_reduces;
384 counter_u64_t rack_collapsed_win;
385 counter_u64_t rack_collapsed_win_seen;
386 counter_u64_t rack_collapsed_win_rxt;
387 counter_u64_t rack_collapsed_win_rxt_bytes;
388 counter_u64_t rack_try_scwnd;
389 counter_u64_t rack_hw_pace_init_fail;
390 counter_u64_t rack_hw_pace_lost;
391 
392 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
393 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
394 
395 
396 #define	RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
397 
398 #define	RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {	\
399 	(tv) = (value) + slop;	 \
400 	if ((u_long)(tv) < (u_long)(tvmin)) \
401 		(tv) = (tvmin); \
402 	if ((u_long)(tv) > (u_long)(tvmax)) \
403 		(tv) = (tvmax); \
404 } while (0)
405 
406 static void
407 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
408 
409 static int
410 rack_process_ack(struct mbuf *m, struct tcphdr *th,
411     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
412     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
413 static int
414 rack_process_data(struct mbuf *m, struct tcphdr *th,
415     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
416     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
417 static void
418 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
419    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
420 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
421 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
422     uint8_t limit_type);
423 static struct rack_sendmap *
424 rack_check_recovery_mode(struct tcpcb *tp,
425     uint32_t tsused);
426 static void
427 rack_cong_signal(struct tcpcb *tp,
428 		 uint32_t type, uint32_t ack, int );
429 static void rack_counter_destroy(void);
430 static int
431 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt);
432 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
433 static void
434 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
435 static void
436 rack_do_segment(struct mbuf *m, struct tcphdr *th,
437     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
438     uint8_t iptos);
439 static void rack_dtor(void *mem, int32_t size, void *arg);
440 static void
441 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
442     uint32_t flex1, uint32_t flex2,
443     uint32_t flex3, uint32_t flex4,
444     uint32_t flex5, uint32_t flex6,
445     uint16_t flex7, uint8_t mod);
446 
447 static void
448 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
449    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
450    struct rack_sendmap *rsm, uint8_t quality);
451 static struct rack_sendmap *
452 rack_find_high_nonack(struct tcp_rack *rack,
453     struct rack_sendmap *rsm);
454 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
455 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
456 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
457 static int rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt);
458 static void
459 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
460 			    tcp_seq th_ack, int line, uint8_t quality);
461 static uint32_t
462 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
463 static int32_t rack_handoff_ok(struct tcpcb *tp);
464 static int32_t rack_init(struct tcpcb *tp);
465 static void rack_init_sysctls(void);
466 static void
467 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
468     struct tcphdr *th, int entered_rec, int dup_ack_struck);
469 static void
470 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
471     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
472     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls);
473 
474 static void
475 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
476     struct rack_sendmap *rsm);
477 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
478 static int32_t rack_output(struct tcpcb *tp);
479 
480 static uint32_t
481 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
482     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
483     uint32_t cts, int *moved_two);
484 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
485 static void rack_remxt_tmr(struct tcpcb *tp);
486 static int rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt);
487 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
488 static int32_t rack_stopall(struct tcpcb *tp);
489 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
490 static uint32_t
491 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
492     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
493 static void
494 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
495     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
496 static int
497 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
498     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
499 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
500 static int
501 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
502     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
503     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
504 static int
505 rack_do_closing(struct mbuf *m, struct tcphdr *th,
506     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
507     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
508 static int
509 rack_do_established(struct mbuf *m, struct tcphdr *th,
510     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
511     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
512 static int
513 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
514     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
515     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
516 static int
517 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
518     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
519     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
520 static int
521 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
522     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
523     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
524 static int
525 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
526     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
527     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
528 static int
529 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
530     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
531     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
532 static int
533 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
534     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
535     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
536 struct rack_sendmap *
537 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
538     uint32_t tsused);
539 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
540     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
541 static void
542      tcp_rack_partialack(struct tcpcb *tp);
543 static int
544 rack_set_profile(struct tcp_rack *rack, int prof);
545 static void
546 rack_apply_deferred_options(struct tcp_rack *rack);
547 
548 int32_t rack_clear_counter=0;
549 
550 static inline void
551 rack_trace_point(struct tcp_rack *rack, int num)
552 {
553 	if (((rack_trace_point_config == num)  ||
554 	     (rack_trace_point_config = 0xffffffff)) &&
555 	    (rack_trace_point_bb_mode != 0) &&
556 	    (rack_trace_point_count > 0) &&
557 	    (rack->rc_tp->t_logstate == 0)) {
558 		int res;
559 		res = atomic_fetchadd_int(&rack_trace_point_count, -1);
560 		if (res > 0) {
561 			rack->rc_tp->t_logstate = rack_trace_point_bb_mode;
562 		} else {
563 			/* Loss a race assure its zero now */
564 			rack_trace_point_count = 0;
565 		}
566 	}
567 }
568 
569 static void
570 rack_swap_beta_values(struct tcp_rack *rack, uint8_t flex8)
571 {
572 	struct sockopt sopt;
573 	struct cc_newreno_opts opt;
574 	struct newreno old;
575 	struct tcpcb *tp;
576 	int error, failed = 0;
577 
578 	tp = rack->rc_tp;
579 	if (tp->t_cc == NULL) {
580 		/* Tcb is leaving */
581 		return;
582 	}
583 	rack->rc_pacing_cc_set = 1;
584 	if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
585 		/* Not new-reno we can't play games with beta! */
586 		failed = 1;
587 		goto out;
588 
589 	}
590 	if (CC_ALGO(tp)->ctl_output == NULL)  {
591 		/* Huh, not using new-reno so no swaps.? */
592 		failed = 2;
593 		goto out;
594 	}
595 	/* Get the current values out */
596 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
597 	sopt.sopt_dir = SOPT_GET;
598 	opt.name = CC_NEWRENO_BETA;
599 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
600 	if (error)  {
601 		failed = 3;
602 		goto out;
603 	}
604 	old.beta = opt.val;
605 	opt.name = CC_NEWRENO_BETA_ECN;
606 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
607 	if (error)  {
608 		failed = 4;
609 		goto out;
610 	}
611 	old.beta_ecn = opt.val;
612 
613 	/* Now lets set in the values we have stored */
614 	sopt.sopt_dir = SOPT_SET;
615 	opt.name = CC_NEWRENO_BETA;
616 	opt.val = rack->r_ctl.rc_saved_beta.beta;
617 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
618 	if (error)  {
619 		failed = 5;
620 		goto out;
621 	}
622 	opt.name = CC_NEWRENO_BETA_ECN;
623 	opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
624 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
625 	if (error) {
626 		failed = 6;
627 		goto out;
628 	}
629 	/* Save off the values for restoral */
630 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
631 out:
632 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
633 		union tcp_log_stackspecific log;
634 		struct timeval tv;
635 		struct newreno *ptr;
636 
637 		ptr = ((struct newreno *)tp->t_ccv.cc_data);
638 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
639 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
640 		log.u_bbr.flex1 = ptr->beta;
641 		log.u_bbr.flex2 = ptr->beta_ecn;
642 		log.u_bbr.flex3 = ptr->newreno_flags;
643 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
644 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
645 		log.u_bbr.flex6 = failed;
646 		log.u_bbr.flex7 = rack->gp_ready;
647 		log.u_bbr.flex7 <<= 1;
648 		log.u_bbr.flex7 |= rack->use_fixed_rate;
649 		log.u_bbr.flex7 <<= 1;
650 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
651 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
652 		log.u_bbr.flex8 = flex8;
653 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
654 			       0, &log, false, NULL, NULL, 0, &tv);
655 	}
656 }
657 
658 static void
659 rack_set_cc_pacing(struct tcp_rack *rack)
660 {
661 	if (rack->rc_pacing_cc_set)
662 		return;
663 	/*
664 	 * Use the swap utility placing in 3 for flex8 to id a
665 	 * set of a new set of values.
666 	 */
667 	rack->rc_pacing_cc_set = 1;
668 	rack_swap_beta_values(rack, 3);
669 }
670 
671 static void
672 rack_undo_cc_pacing(struct tcp_rack *rack)
673 {
674 	if (rack->rc_pacing_cc_set == 0)
675 		return;
676 	/*
677 	 * Use the swap utility placing in 4 for flex8 to id a
678 	 * restoral of the old values.
679 	 */
680 	rack->rc_pacing_cc_set = 0;
681 	rack_swap_beta_values(rack, 4);
682 }
683 
684 #ifdef NETFLIX_PEAKRATE
685 static inline void
686 rack_update_peakrate_thr(struct tcpcb *tp)
687 {
688 	/* Keep in mind that t_maxpeakrate is in B/s. */
689 	uint64_t peak;
690 	peak = uqmax((tp->t_maxseg * 2),
691 		     (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
692 	tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
693 }
694 #endif
695 
696 static int
697 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
698 {
699 	uint32_t stat;
700 	int32_t error;
701 
702 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
703 	if (error || req->newptr == NULL)
704 		return error;
705 
706 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
707 	if (error)
708 		return (error);
709 	if (stat == 1) {
710 #ifdef INVARIANTS
711 		printf("Clearing RACK counters\n");
712 #endif
713 		counter_u64_zero(rack_tlp_tot);
714 		counter_u64_zero(rack_tlp_newdata);
715 		counter_u64_zero(rack_tlp_retran);
716 		counter_u64_zero(rack_tlp_retran_bytes);
717 		counter_u64_zero(rack_to_tot);
718 		counter_u64_zero(rack_saw_enobuf);
719 		counter_u64_zero(rack_saw_enobuf_hw);
720 		counter_u64_zero(rack_saw_enetunreach);
721 		counter_u64_zero(rack_persists_sends);
722 		counter_u64_zero(rack_persists_acks);
723 		counter_u64_zero(rack_persists_loss);
724 		counter_u64_zero(rack_persists_lost_ends);
725 #ifdef INVARIANTS
726 		counter_u64_zero(rack_adjust_map_bw);
727 #endif
728 		counter_u64_zero(rack_to_alloc_hard);
729 		counter_u64_zero(rack_to_alloc_emerg);
730 		counter_u64_zero(rack_sack_proc_all);
731 		counter_u64_zero(rack_fto_send);
732 		counter_u64_zero(rack_fto_rsm_send);
733 		counter_u64_zero(rack_extended_rfo);
734 		counter_u64_zero(rack_hw_pace_init_fail);
735 		counter_u64_zero(rack_hw_pace_lost);
736 		counter_u64_zero(rack_non_fto_send);
737 		counter_u64_zero(rack_nfto_resend);
738 		counter_u64_zero(rack_sack_proc_short);
739 		counter_u64_zero(rack_sack_proc_restart);
740 		counter_u64_zero(rack_to_alloc);
741 		counter_u64_zero(rack_to_alloc_limited);
742 		counter_u64_zero(rack_alloc_limited_conns);
743 		counter_u64_zero(rack_split_limited);
744 		counter_u64_zero(rack_multi_single_eq);
745 		counter_u64_zero(rack_proc_non_comp_ack);
746 		counter_u64_zero(rack_sack_attacks_detected);
747 		counter_u64_zero(rack_sack_attacks_reversed);
748 		counter_u64_zero(rack_sack_used_next_merge);
749 		counter_u64_zero(rack_sack_used_prev_merge);
750 		counter_u64_zero(rack_sack_splits);
751 		counter_u64_zero(rack_sack_skipped_acked);
752 		counter_u64_zero(rack_ack_total);
753 		counter_u64_zero(rack_express_sack);
754 		counter_u64_zero(rack_sack_total);
755 		counter_u64_zero(rack_move_none);
756 		counter_u64_zero(rack_move_some);
757 		counter_u64_zero(rack_try_scwnd);
758 		counter_u64_zero(rack_collapsed_win);
759 		counter_u64_zero(rack_collapsed_win_rxt);
760 		counter_u64_zero(rack_collapsed_win_seen);
761 		counter_u64_zero(rack_collapsed_win_rxt_bytes);
762 	}
763 	rack_clear_counter = 0;
764 	return (0);
765 }
766 
767 static void
768 rack_init_sysctls(void)
769 {
770 	struct sysctl_oid *rack_counters;
771 	struct sysctl_oid *rack_attack;
772 	struct sysctl_oid *rack_pacing;
773 	struct sysctl_oid *rack_timely;
774 	struct sysctl_oid *rack_timers;
775 	struct sysctl_oid *rack_tlp;
776 	struct sysctl_oid *rack_misc;
777 	struct sysctl_oid *rack_features;
778 	struct sysctl_oid *rack_measure;
779 	struct sysctl_oid *rack_probertt;
780 	struct sysctl_oid *rack_hw_pacing;
781 	struct sysctl_oid *rack_tracepoint;
782 
783 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
784 	    SYSCTL_CHILDREN(rack_sysctl_root),
785 	    OID_AUTO,
786 	    "sack_attack",
787 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
788 	    "Rack Sack Attack Counters and Controls");
789 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
790 	    SYSCTL_CHILDREN(rack_sysctl_root),
791 	    OID_AUTO,
792 	    "stats",
793 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
794 	    "Rack Counters");
795 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
796 	    SYSCTL_CHILDREN(rack_sysctl_root),
797 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
798 	    &rack_rate_sample_method , USE_RTT_LOW,
799 	    "What method should we use for rate sampling 0=high, 1=low ");
800 	/* Probe rtt related controls */
801 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
802 	    SYSCTL_CHILDREN(rack_sysctl_root),
803 	    OID_AUTO,
804 	    "probertt",
805 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
806 	    "ProbeRTT related Controls");
807 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
808 	    SYSCTL_CHILDREN(rack_probertt),
809 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
810 	    &rack_atexit_prtt_hbp, 130,
811 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
812 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
813 	    SYSCTL_CHILDREN(rack_probertt),
814 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
815 	    &rack_atexit_prtt, 130,
816 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
817 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
818 	    SYSCTL_CHILDREN(rack_probertt),
819 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
820 	    &rack_per_of_gp_probertt, 60,
821 	    "What percentage of goodput do we pace at in probertt");
822 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
823 	    SYSCTL_CHILDREN(rack_probertt),
824 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
825 	    &rack_per_of_gp_probertt_reduce, 10,
826 	    "What percentage of goodput do we reduce every gp_srtt");
827 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
828 	    SYSCTL_CHILDREN(rack_probertt),
829 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
830 	    &rack_per_of_gp_lowthresh, 40,
831 	    "What percentage of goodput do we allow the multiplier to fall to");
832 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
833 	    SYSCTL_CHILDREN(rack_probertt),
834 	    OID_AUTO, "time_between", CTLFLAG_RW,
835 	    & rack_time_between_probertt, 96000000,
836 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
837 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
838 	    SYSCTL_CHILDREN(rack_probertt),
839 	    OID_AUTO, "safety", CTLFLAG_RW,
840 	    &rack_probe_rtt_safety_val, 2000000,
841 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
842 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
843 	    SYSCTL_CHILDREN(rack_probertt),
844 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
845 	    &rack_probe_rtt_sets_cwnd, 0,
846 	    "Do we set the cwnd too (if always_lower is on)");
847 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
848 	    SYSCTL_CHILDREN(rack_probertt),
849 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
850 	    &rack_max_drain_wait, 2,
851 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
852 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
853 	    SYSCTL_CHILDREN(rack_probertt),
854 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
855 	    &rack_must_drain, 1,
856 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
857 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
858 	    SYSCTL_CHILDREN(rack_probertt),
859 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
860 	    &rack_probertt_use_min_rtt_entry, 1,
861 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
862 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
863 	    SYSCTL_CHILDREN(rack_probertt),
864 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
865 	    &rack_probertt_use_min_rtt_exit, 0,
866 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
867 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
868 	    SYSCTL_CHILDREN(rack_probertt),
869 	    OID_AUTO, "length_div", CTLFLAG_RW,
870 	    &rack_probertt_gpsrtt_cnt_div, 0,
871 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
872 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
873 	    SYSCTL_CHILDREN(rack_probertt),
874 	    OID_AUTO, "length_mul", CTLFLAG_RW,
875 	    &rack_probertt_gpsrtt_cnt_mul, 0,
876 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
877 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
878 	    SYSCTL_CHILDREN(rack_probertt),
879 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
880 	    &rack_min_probertt_hold, 200000,
881 	    "What is the minimum time we hold probertt at target");
882 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
883 	    SYSCTL_CHILDREN(rack_probertt),
884 	    OID_AUTO, "filter_life", CTLFLAG_RW,
885 	    &rack_probertt_filter_life, 10000000,
886 	    "What is the time for the filters life in useconds");
887 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
888 	    SYSCTL_CHILDREN(rack_probertt),
889 	    OID_AUTO, "lower_within", CTLFLAG_RW,
890 	    &rack_probertt_lower_within, 10,
891 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
892 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
893 	    SYSCTL_CHILDREN(rack_probertt),
894 	    OID_AUTO, "must_move", CTLFLAG_RW,
895 	    &rack_min_rtt_movement, 250,
896 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
897 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
898 	    SYSCTL_CHILDREN(rack_probertt),
899 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
900 	    &rack_probertt_clear_is, 1,
901 	    "Do we clear I/S counts on exiting probe-rtt");
902 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
903 	    SYSCTL_CHILDREN(rack_probertt),
904 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
905 	    &rack_max_drain_hbp, 1,
906 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
907 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
908 	    SYSCTL_CHILDREN(rack_probertt),
909 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
910 	    &rack_hbp_thresh, 3,
911 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
912 
913 	rack_tracepoint = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
914 	    SYSCTL_CHILDREN(rack_sysctl_root),
915 	    OID_AUTO,
916 	    "tp",
917 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
918 	    "Rack tracepoint facility");
919 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
920 	    SYSCTL_CHILDREN(rack_tracepoint),
921 	    OID_AUTO, "number", CTLFLAG_RW,
922 	    &rack_trace_point_config, 0,
923 	    "What is the trace point number to activate (0=none, 0xffffffff = all)?");
924 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
925 	    SYSCTL_CHILDREN(rack_tracepoint),
926 	    OID_AUTO, "bbmode", CTLFLAG_RW,
927 	    &rack_trace_point_bb_mode, 4,
928 	    "What is BB logging mode that is activated?");
929 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
930 	    SYSCTL_CHILDREN(rack_tracepoint),
931 	    OID_AUTO, "count", CTLFLAG_RW,
932 	    &rack_trace_point_count, 0,
933 	    "How many connections will have BB logging turned on that hit the tracepoint?");
934 	/* Pacing related sysctls */
935 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
936 	    SYSCTL_CHILDREN(rack_sysctl_root),
937 	    OID_AUTO,
938 	    "pacing",
939 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
940 	    "Pacing related Controls");
941 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
942 	    SYSCTL_CHILDREN(rack_pacing),
943 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
944 	    &rack_max_per_above, 30,
945 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
946 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
947 	    SYSCTL_CHILDREN(rack_pacing),
948 	    OID_AUTO, "pace_to_one", CTLFLAG_RW,
949 	    &rack_pace_one_seg, 0,
950 	    "Do we allow low b/w pacing of 1MSS instead of two");
951 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
952 	    SYSCTL_CHILDREN(rack_pacing),
953 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
954 	    &rack_limit_time_with_srtt, 0,
955 	    "Do we limit pacing time based on srtt");
956 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
957 	    SYSCTL_CHILDREN(rack_pacing),
958 	    OID_AUTO, "init_win", CTLFLAG_RW,
959 	    &rack_default_init_window, 0,
960 	    "Do we have a rack initial window 0 = system default");
961 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
962 	    SYSCTL_CHILDREN(rack_pacing),
963 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
964 	    &rack_per_of_gp_ss, 250,
965 	    "If non zero, what percentage of goodput to pace at in slow start");
966 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
967 	    SYSCTL_CHILDREN(rack_pacing),
968 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
969 	    &rack_per_of_gp_ca, 150,
970 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
971 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
972 	    SYSCTL_CHILDREN(rack_pacing),
973 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
974 	    &rack_per_of_gp_rec, 200,
975 	    "If non zero, what percentage of goodput to pace at in recovery");
976 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
977 	    SYSCTL_CHILDREN(rack_pacing),
978 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
979 	    &rack_hptsi_segments, 40,
980 	    "What size is the max for TSO segments in pacing and burst mitigation");
981 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
982 	    SYSCTL_CHILDREN(rack_pacing),
983 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
984 	    &rack_slot_reduction, 4,
985 	    "When doing only burst mitigation what is the reduce divisor");
986 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
987 	    SYSCTL_CHILDREN(rack_sysctl_root),
988 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
989 	    &rack_pace_every_seg, 0,
990 	    "If set we use pacing, if clear we use only the original burst mitigation");
991 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
992 	    SYSCTL_CHILDREN(rack_pacing),
993 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
994 	    &rack_bw_rate_cap, 0,
995 	    "If set we apply this value to the absolute rate cap used by pacing");
996 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
997 	    SYSCTL_CHILDREN(rack_sysctl_root),
998 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
999 	    &rack_req_measurements, 1,
1000 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1001 	/* Hardware pacing */
1002 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1003 	    SYSCTL_CHILDREN(rack_sysctl_root),
1004 	    OID_AUTO,
1005 	    "hdwr_pacing",
1006 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1007 	    "Pacing related Controls");
1008 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1009 	    SYSCTL_CHILDREN(rack_hw_pacing),
1010 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1011 	    &rack_hw_rwnd_factor, 2,
1012 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1013 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1014 	    SYSCTL_CHILDREN(rack_hw_pacing),
1015 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1016 	    &rack_enobuf_hw_boost_mult, 2,
1017 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1018 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1019 	    SYSCTL_CHILDREN(rack_hw_pacing),
1020 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1021 	    &rack_enobuf_hw_max, 2,
1022 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1023 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1024 	    SYSCTL_CHILDREN(rack_hw_pacing),
1025 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1026 	    &rack_enobuf_hw_min, 2,
1027 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1028 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1029 	    SYSCTL_CHILDREN(rack_hw_pacing),
1030 	    OID_AUTO, "enable", CTLFLAG_RW,
1031 	    &rack_enable_hw_pacing, 0,
1032 	    "Should RACK attempt to use hw pacing?");
1033 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1034 	    SYSCTL_CHILDREN(rack_hw_pacing),
1035 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1036 	    &rack_hw_rate_caps, 1,
1037 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1038 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1039 	    SYSCTL_CHILDREN(rack_hw_pacing),
1040 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1041 	    &rack_hw_rate_min, 0,
1042 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1043 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1044 	    SYSCTL_CHILDREN(rack_hw_pacing),
1045 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1046 	    &rack_hw_rate_to_low, 0,
1047 	    "If we fall below this rate, dis-engage hw pacing?");
1048 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1049 	    SYSCTL_CHILDREN(rack_hw_pacing),
1050 	    OID_AUTO, "up_only", CTLFLAG_RW,
1051 	    &rack_hw_up_only, 1,
1052 	    "Do we allow hw pacing to lower the rate selected?");
1053 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1054 	    SYSCTL_CHILDREN(rack_hw_pacing),
1055 	    OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1056 	    &rack_hw_pace_extra_slots, 2,
1057 	    "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1058 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1059 	    SYSCTL_CHILDREN(rack_sysctl_root),
1060 	    OID_AUTO,
1061 	    "timely",
1062 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1063 	    "Rack Timely RTT Controls");
1064 	/* Timely based GP dynmics */
1065 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1066 	    SYSCTL_CHILDREN(rack_timely),
1067 	    OID_AUTO, "upper", CTLFLAG_RW,
1068 	    &rack_gp_per_bw_mul_up, 2,
1069 	    "Rack timely upper range for equal b/w (in percentage)");
1070 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1071 	    SYSCTL_CHILDREN(rack_timely),
1072 	    OID_AUTO, "lower", CTLFLAG_RW,
1073 	    &rack_gp_per_bw_mul_down, 4,
1074 	    "Rack timely lower range for equal b/w (in percentage)");
1075 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1076 	    SYSCTL_CHILDREN(rack_timely),
1077 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1078 	    &rack_gp_rtt_maxmul, 3,
1079 	    "Rack timely multiplier of lowest rtt for rtt_max");
1080 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1081 	    SYSCTL_CHILDREN(rack_timely),
1082 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1083 	    &rack_gp_rtt_mindiv, 4,
1084 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1085 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1086 	    SYSCTL_CHILDREN(rack_timely),
1087 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1088 	    &rack_gp_rtt_minmul, 1,
1089 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1090 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1091 	    SYSCTL_CHILDREN(rack_timely),
1092 	    OID_AUTO, "decrease", CTLFLAG_RW,
1093 	    &rack_gp_decrease_per, 20,
1094 	    "Rack timely decrease percentage of our GP multiplication factor");
1095 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1096 	    SYSCTL_CHILDREN(rack_timely),
1097 	    OID_AUTO, "increase", CTLFLAG_RW,
1098 	    &rack_gp_increase_per, 2,
1099 	    "Rack timely increase perentage of our GP multiplication factor");
1100 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1101 	    SYSCTL_CHILDREN(rack_timely),
1102 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1103 	    &rack_per_lower_bound, 50,
1104 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1105 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1106 	    SYSCTL_CHILDREN(rack_timely),
1107 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1108 	    &rack_per_upper_bound_ss, 0,
1109 	    "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1110 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1111 	    SYSCTL_CHILDREN(rack_timely),
1112 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1113 	    &rack_per_upper_bound_ca, 0,
1114 	    "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1115 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1116 	    SYSCTL_CHILDREN(rack_timely),
1117 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1118 	    &rack_do_dyn_mul, 0,
1119 	    "Rack timely do we enable dynmaic timely goodput by default");
1120 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1121 	    SYSCTL_CHILDREN(rack_timely),
1122 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1123 	    &rack_gp_no_rec_chg, 1,
1124 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1125 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1126 	    SYSCTL_CHILDREN(rack_timely),
1127 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1128 	    &rack_timely_dec_clear, 6,
1129 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1130 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1131 	    SYSCTL_CHILDREN(rack_timely),
1132 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1133 	    &rack_timely_max_push_rise, 3,
1134 	    "Rack timely how many times do we push up with b/w increase");
1135 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1136 	    SYSCTL_CHILDREN(rack_timely),
1137 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1138 	    &rack_timely_max_push_drop, 3,
1139 	    "Rack timely how many times do we push back on b/w decent");
1140 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1141 	    SYSCTL_CHILDREN(rack_timely),
1142 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1143 	    &rack_timely_min_segs, 4,
1144 	    "Rack timely when setting the cwnd what is the min num segments");
1145 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1146 	    SYSCTL_CHILDREN(rack_timely),
1147 	    OID_AUTO, "noback_max", CTLFLAG_RW,
1148 	    &rack_use_max_for_nobackoff, 0,
1149 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1150 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1151 	    SYSCTL_CHILDREN(rack_timely),
1152 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1153 	    &rack_timely_int_timely_only, 0,
1154 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1155 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1156 	    SYSCTL_CHILDREN(rack_timely),
1157 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1158 	    &rack_timely_no_stopping, 0,
1159 	    "Rack timely don't stop increase");
1160 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1161 	    SYSCTL_CHILDREN(rack_timely),
1162 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1163 	    &rack_down_raise_thresh, 100,
1164 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1165 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1166 	    SYSCTL_CHILDREN(rack_timely),
1167 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1168 	    &rack_req_segs, 1,
1169 	    "Bottom dragging if not these many segments outstanding and room");
1170 
1171 	/* TLP and Rack related parameters */
1172 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1173 	    SYSCTL_CHILDREN(rack_sysctl_root),
1174 	    OID_AUTO,
1175 	    "tlp",
1176 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1177 	    "TLP and Rack related Controls");
1178 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1179 	    SYSCTL_CHILDREN(rack_tlp),
1180 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1181 	    &use_rack_rr, 1,
1182 	    "Do we use Rack Rapid Recovery");
1183 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1184 	    SYSCTL_CHILDREN(rack_tlp),
1185 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1186 	    &rack_max_abc_post_recovery, 2,
1187 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1188 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1189 	    SYSCTL_CHILDREN(rack_tlp),
1190 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1191 	    &rack_non_rxt_use_cr, 0,
1192 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1193 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1194 	    SYSCTL_CHILDREN(rack_tlp),
1195 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1196 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1197 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1198 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1199 	    SYSCTL_CHILDREN(rack_tlp),
1200 	    OID_AUTO, "limit", CTLFLAG_RW,
1201 	    &rack_tlp_limit, 2,
1202 	    "How many TLP's can be sent without sending new data");
1203 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1204 	    SYSCTL_CHILDREN(rack_tlp),
1205 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1206 	    &rack_tlp_use_greater, 1,
1207 	    "Should we use the rack_rtt time if its greater than srtt");
1208 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1209 	    SYSCTL_CHILDREN(rack_tlp),
1210 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1211 	    &rack_tlp_min, 10000,
1212 	    "TLP minimum timeout per the specification (in microseconds)");
1213 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1214 	    SYSCTL_CHILDREN(rack_tlp),
1215 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1216 	    &rack_always_send_oldest, 0,
1217 	    "Should we always send the oldest TLP and RACK-TLP");
1218 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1219 	    SYSCTL_CHILDREN(rack_tlp),
1220 	    OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1221 	    &rack_limited_retran, 0,
1222 	    "How many times can a rack timeout drive out sends");
1223 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1224 	    SYSCTL_CHILDREN(rack_tlp),
1225 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1226 	    &rack_lower_cwnd_at_tlp, 0,
1227 	    "When a TLP completes a retran should we enter recovery");
1228 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1229 	    SYSCTL_CHILDREN(rack_tlp),
1230 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1231 	    &rack_reorder_thresh, 2,
1232 	    "What factor for rack will be added when seeing reordering (shift right)");
1233 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1234 	    SYSCTL_CHILDREN(rack_tlp),
1235 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1236 	    &rack_tlp_thresh, 1,
1237 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1238 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1239 	    SYSCTL_CHILDREN(rack_tlp),
1240 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1241 	    &rack_reorder_fade, 60000000,
1242 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1243 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1244 	    SYSCTL_CHILDREN(rack_tlp),
1245 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1246 	    &rack_pkt_delay, 1000,
1247 	    "Extra RACK time (in microseconds) besides reordering thresh");
1248 
1249 	/* Timer related controls */
1250 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1251 	    SYSCTL_CHILDREN(rack_sysctl_root),
1252 	    OID_AUTO,
1253 	    "timers",
1254 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1255 	    "Timer related controls");
1256 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1257 	    SYSCTL_CHILDREN(rack_timers),
1258 	    OID_AUTO, "persmin", CTLFLAG_RW,
1259 	    &rack_persist_min, 250000,
1260 	    "What is the minimum time in microseconds between persists");
1261 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1262 	    SYSCTL_CHILDREN(rack_timers),
1263 	    OID_AUTO, "persmax", CTLFLAG_RW,
1264 	    &rack_persist_max, 2000000,
1265 	    "What is the largest delay in microseconds between persists");
1266 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1267 	    SYSCTL_CHILDREN(rack_timers),
1268 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1269 	    &rack_delayed_ack_time, 40000,
1270 	    "Delayed ack time (40ms in microseconds)");
1271 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1272 	    SYSCTL_CHILDREN(rack_timers),
1273 	    OID_AUTO, "minrto", CTLFLAG_RW,
1274 	    &rack_rto_min, 30000,
1275 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1276 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1277 	    SYSCTL_CHILDREN(rack_timers),
1278 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1279 	    &rack_rto_max, 4000000,
1280 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1281 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1282 	    SYSCTL_CHILDREN(rack_timers),
1283 	    OID_AUTO, "minto", CTLFLAG_RW,
1284 	    &rack_min_to, 1000,
1285 	    "Minimum rack timeout in microseconds");
1286 	/* Measure controls */
1287 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1288 	    SYSCTL_CHILDREN(rack_sysctl_root),
1289 	    OID_AUTO,
1290 	    "measure",
1291 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1292 	    "Measure related controls");
1293 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1294 	    SYSCTL_CHILDREN(rack_measure),
1295 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1296 	    &rack_wma_divisor, 8,
1297 	    "When doing b/w calculation what is the  divisor for the WMA");
1298 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1299 	    SYSCTL_CHILDREN(rack_measure),
1300 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1301 	    &rack_cwnd_block_ends_measure, 0,
1302 	    "Does a cwnd just-return end the measurement window (app limited)");
1303 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1304 	    SYSCTL_CHILDREN(rack_measure),
1305 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1306 	    &rack_rwnd_block_ends_measure, 0,
1307 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1308 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1309 	    SYSCTL_CHILDREN(rack_measure),
1310 	    OID_AUTO, "min_target", CTLFLAG_RW,
1311 	    &rack_def_data_window, 20,
1312 	    "What is the minimum target window (in mss) for a GP measurements");
1313 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1314 	    SYSCTL_CHILDREN(rack_measure),
1315 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1316 	    &rack_goal_bdp, 2,
1317 	    "What is the goal BDP to measure");
1318 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1319 	    SYSCTL_CHILDREN(rack_measure),
1320 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1321 	    &rack_min_srtts, 1,
1322 	    "What is the goal BDP to measure");
1323 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1324 	    SYSCTL_CHILDREN(rack_measure),
1325 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1326 	    &rack_min_measure_usec, 0,
1327 	    "What is the Minimum time time for a measurement if 0, this is off");
1328 	/* Features */
1329 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1330 	    SYSCTL_CHILDREN(rack_sysctl_root),
1331 	    OID_AUTO,
1332 	    "features",
1333 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1334 	    "Feature controls");
1335 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1336 	    SYSCTL_CHILDREN(rack_features),
1337 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1338 	    &rack_use_cmp_acks, 1,
1339 	    "Should RACK have LRO send compressed acks");
1340 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1341 	    SYSCTL_CHILDREN(rack_features),
1342 	    OID_AUTO, "fsb", CTLFLAG_RW,
1343 	    &rack_use_fsb, 1,
1344 	    "Should RACK use the fast send block?");
1345 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1346 	    SYSCTL_CHILDREN(rack_features),
1347 	    OID_AUTO, "rfo", CTLFLAG_RW,
1348 	    &rack_use_rfo, 1,
1349 	    "Should RACK use rack_fast_output()?");
1350 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1351 	    SYSCTL_CHILDREN(rack_features),
1352 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1353 	    &rack_use_rsm_rfo, 1,
1354 	    "Should RACK use rack_fast_rsm_output()?");
1355 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1356 	    SYSCTL_CHILDREN(rack_features),
1357 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1358 	    &rack_enable_mqueue_for_nonpaced, 0,
1359 	    "Should RACK use mbuf queuing for non-paced connections");
1360 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1361 	    SYSCTL_CHILDREN(rack_features),
1362 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1363 	    &rack_do_hystart, 0,
1364 	    "Should RACK enable HyStart++ on connections?");
1365 	/* Misc rack controls */
1366 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1367 	    SYSCTL_CHILDREN(rack_sysctl_root),
1368 	    OID_AUTO,
1369 	    "misc",
1370 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1371 	    "Misc related controls");
1372 #ifdef TCP_ACCOUNTING
1373 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1374 	    SYSCTL_CHILDREN(rack_misc),
1375 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1376 	    &rack_tcp_accounting, 0,
1377 	    "Should we turn on TCP accounting for all rack sessions?");
1378 #endif
1379 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1380 	    SYSCTL_CHILDREN(rack_misc),
1381 	    OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1382 	    &rack_apply_rtt_with_reduced_conf, 0,
1383 	    "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1384 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1385 	    SYSCTL_CHILDREN(rack_misc),
1386 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1387 	    &rack_dsack_std_based, 3,
1388 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1389 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1390 	    SYSCTL_CHILDREN(rack_misc),
1391 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1392 	    &rack_prr_addbackmax, 2,
1393 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1394 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1395 	    SYSCTL_CHILDREN(rack_misc),
1396 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1397 	    &rack_stats_gets_ms_rtt, 1,
1398 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1399 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1400 	    SYSCTL_CHILDREN(rack_misc),
1401 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1402 	    &rack_client_low_buf, 0,
1403 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1404 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1405 	    SYSCTL_CHILDREN(rack_misc),
1406 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1407 	    &rack_def_profile, 0,
1408 	    "Should RACK use a default profile (0=no, num == profile num)?");
1409 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1410 	    SYSCTL_CHILDREN(rack_misc),
1411 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1412 	    &rack_enable_shared_cwnd, 1,
1413 	    "Should RACK try to use the shared cwnd on connections where allowed");
1414 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1415 	    SYSCTL_CHILDREN(rack_misc),
1416 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1417 	    &rack_limits_scwnd, 1,
1418 	    "Should RACK place low end time limits on the shared cwnd feature");
1419 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1420 	    SYSCTL_CHILDREN(rack_misc),
1421 	    OID_AUTO, "iMac_dack", CTLFLAG_RW,
1422 	    &rack_use_imac_dack, 0,
1423 	    "Should RACK try to emulate iMac delayed ack");
1424 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1425 	    SYSCTL_CHILDREN(rack_misc),
1426 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1427 	    &rack_disable_prr, 0,
1428 	    "Should RACK not use prr and only pace (must have pacing on)");
1429 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1430 	    SYSCTL_CHILDREN(rack_misc),
1431 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1432 	    &rack_verbose_logging, 0,
1433 	    "Should RACK black box logging be verbose");
1434 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1435 	    SYSCTL_CHILDREN(rack_misc),
1436 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1437 	    &rack_ignore_data_after_close, 1,
1438 	    "Do we hold off sending a RST until all pending data is ack'd");
1439 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1440 	    SYSCTL_CHILDREN(rack_misc),
1441 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1442 	    &rack_sack_not_required, 1,
1443 	    "Do we allow rack to run on connections not supporting SACK");
1444 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1445 	    SYSCTL_CHILDREN(rack_misc),
1446 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1447 	    &rack_send_a_lot_in_prr, 1,
1448 	    "Send a lot in prr");
1449 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1450 	    SYSCTL_CHILDREN(rack_misc),
1451 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1452 	    &rack_autosndbuf_inc, 20,
1453 	    "What percentage should rack scale up its snd buffer by?");
1454 	/* Sack Attacker detection stuff */
1455 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1456 	    SYSCTL_CHILDREN(rack_attack),
1457 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1458 	    &rack_highest_sack_thresh_seen, 0,
1459 	    "Highest sack to ack ratio seen");
1460 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1461 	    SYSCTL_CHILDREN(rack_attack),
1462 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1463 	    &rack_highest_move_thresh_seen, 0,
1464 	    "Highest move to non-move ratio seen");
1465 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1466 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1467 	    SYSCTL_CHILDREN(rack_attack),
1468 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1469 	    &rack_ack_total,
1470 	    "Total number of Ack's");
1471 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1472 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1473 	    SYSCTL_CHILDREN(rack_attack),
1474 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1475 	    &rack_express_sack,
1476 	    "Total expresss number of Sack's");
1477 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1478 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1479 	    SYSCTL_CHILDREN(rack_attack),
1480 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1481 	    &rack_sack_total,
1482 	    "Total number of SACKs");
1483 	rack_move_none = counter_u64_alloc(M_WAITOK);
1484 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1485 	    SYSCTL_CHILDREN(rack_attack),
1486 	    OID_AUTO, "move_none", CTLFLAG_RD,
1487 	    &rack_move_none,
1488 	    "Total number of SACK index reuse of positions under threshold");
1489 	rack_move_some = counter_u64_alloc(M_WAITOK);
1490 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1491 	    SYSCTL_CHILDREN(rack_attack),
1492 	    OID_AUTO, "move_some", CTLFLAG_RD,
1493 	    &rack_move_some,
1494 	    "Total number of SACK index reuse of positions over threshold");
1495 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1496 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1497 	    SYSCTL_CHILDREN(rack_attack),
1498 	    OID_AUTO, "attacks", CTLFLAG_RD,
1499 	    &rack_sack_attacks_detected,
1500 	    "Total number of SACK attackers that had sack disabled");
1501 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1502 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1503 	    SYSCTL_CHILDREN(rack_attack),
1504 	    OID_AUTO, "reversed", CTLFLAG_RD,
1505 	    &rack_sack_attacks_reversed,
1506 	    "Total number of SACK attackers that were later determined false positive");
1507 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1508 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1509 	    SYSCTL_CHILDREN(rack_attack),
1510 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1511 	    &rack_sack_used_next_merge,
1512 	    "Total number of times we used the next merge");
1513 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1514 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1515 	    SYSCTL_CHILDREN(rack_attack),
1516 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1517 	    &rack_sack_used_prev_merge,
1518 	    "Total number of times we used the prev merge");
1519 	/* Counters */
1520 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1521 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1522 	    SYSCTL_CHILDREN(rack_counters),
1523 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1524 	    &rack_fto_send, "Total number of rack_fast_output sends");
1525 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1526 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1527 	    SYSCTL_CHILDREN(rack_counters),
1528 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1529 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1530 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1531 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1532 	    SYSCTL_CHILDREN(rack_counters),
1533 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1534 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1535 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1536 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1537 	    SYSCTL_CHILDREN(rack_counters),
1538 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1539 	    &rack_non_fto_send, "Total number of rack_output first sends");
1540 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1541 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1542 	    SYSCTL_CHILDREN(rack_counters),
1543 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1544 	    &rack_extended_rfo, "Total number of times we extended rfo");
1545 
1546 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1547 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1548 	    SYSCTL_CHILDREN(rack_counters),
1549 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1550 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1551 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1552 
1553 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1554 	    SYSCTL_CHILDREN(rack_counters),
1555 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1556 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1557 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1558 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1559 	    SYSCTL_CHILDREN(rack_counters),
1560 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1561 	    &rack_tlp_tot,
1562 	    "Total number of tail loss probe expirations");
1563 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1564 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1565 	    SYSCTL_CHILDREN(rack_counters),
1566 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1567 	    &rack_tlp_newdata,
1568 	    "Total number of tail loss probe sending new data");
1569 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1570 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1571 	    SYSCTL_CHILDREN(rack_counters),
1572 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1573 	    &rack_tlp_retran,
1574 	    "Total number of tail loss probe sending retransmitted data");
1575 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1576 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1577 	    SYSCTL_CHILDREN(rack_counters),
1578 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1579 	    &rack_tlp_retran_bytes,
1580 	    "Total bytes of tail loss probe sending retransmitted data");
1581 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1582 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1583 	    SYSCTL_CHILDREN(rack_counters),
1584 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1585 	    &rack_to_tot,
1586 	    "Total number of times the rack to expired");
1587 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1588 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1589 	    SYSCTL_CHILDREN(rack_counters),
1590 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1591 	    &rack_saw_enobuf,
1592 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1593 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1594 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1595 	    SYSCTL_CHILDREN(rack_counters),
1596 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1597 	    &rack_saw_enobuf_hw,
1598 	    "Total number of times a send returned enobuf for hdwr paced connections");
1599 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1600 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1601 	    SYSCTL_CHILDREN(rack_counters),
1602 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1603 	    &rack_saw_enetunreach,
1604 	    "Total number of times a send received a enetunreachable");
1605 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1606 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1607 	    SYSCTL_CHILDREN(rack_counters),
1608 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1609 	    &rack_hot_alloc,
1610 	    "Total allocations from the top of our list");
1611 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1612 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1613 	    SYSCTL_CHILDREN(rack_counters),
1614 	    OID_AUTO, "allocs", CTLFLAG_RD,
1615 	    &rack_to_alloc,
1616 	    "Total allocations of tracking structures");
1617 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1618 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1619 	    SYSCTL_CHILDREN(rack_counters),
1620 	    OID_AUTO, "allochard", CTLFLAG_RD,
1621 	    &rack_to_alloc_hard,
1622 	    "Total allocations done with sleeping the hard way");
1623 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1624 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1625 	    SYSCTL_CHILDREN(rack_counters),
1626 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1627 	    &rack_to_alloc_emerg,
1628 	    "Total allocations done from emergency cache");
1629 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1630 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1631 	    SYSCTL_CHILDREN(rack_counters),
1632 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1633 	    &rack_to_alloc_limited,
1634 	    "Total allocations dropped due to limit");
1635 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1636 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1637 	    SYSCTL_CHILDREN(rack_counters),
1638 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1639 	    &rack_alloc_limited_conns,
1640 	    "Connections with allocations dropped due to limit");
1641 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1642 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1643 	    SYSCTL_CHILDREN(rack_counters),
1644 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1645 	    &rack_split_limited,
1646 	    "Split allocations dropped due to limit");
1647 	rack_persists_sends = counter_u64_alloc(M_WAITOK);
1648 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1649 	    SYSCTL_CHILDREN(rack_counters),
1650 	    OID_AUTO, "persist_sends", CTLFLAG_RD,
1651 	    &rack_persists_sends,
1652 	    "Number of times we sent a persist probe");
1653 	rack_persists_acks = counter_u64_alloc(M_WAITOK);
1654 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1655 	    SYSCTL_CHILDREN(rack_counters),
1656 	    OID_AUTO, "persist_acks", CTLFLAG_RD,
1657 	    &rack_persists_acks,
1658 	    "Number of times a persist probe was acked");
1659 	rack_persists_loss = counter_u64_alloc(M_WAITOK);
1660 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1661 	    SYSCTL_CHILDREN(rack_counters),
1662 	    OID_AUTO, "persist_loss", CTLFLAG_RD,
1663 	    &rack_persists_loss,
1664 	    "Number of times we detected a lost persist probe (no ack)");
1665 	rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1666 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1667 	    SYSCTL_CHILDREN(rack_counters),
1668 	    OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1669 	    &rack_persists_lost_ends,
1670 	    "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1671 #ifdef INVARIANTS
1672 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1673 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1674 	    SYSCTL_CHILDREN(rack_counters),
1675 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1676 	    &rack_adjust_map_bw,
1677 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1678 #endif
1679 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1680 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1681 	    SYSCTL_CHILDREN(rack_counters),
1682 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1683 	    &rack_multi_single_eq,
1684 	    "Number of compressed acks total represented");
1685 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1686 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1687 	    SYSCTL_CHILDREN(rack_counters),
1688 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1689 	    &rack_proc_non_comp_ack,
1690 	    "Number of non compresseds acks that we processed");
1691 
1692 
1693 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1694 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1695 	    SYSCTL_CHILDREN(rack_counters),
1696 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1697 	    &rack_sack_proc_all,
1698 	    "Total times we had to walk whole list for sack processing");
1699 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1700 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1701 	    SYSCTL_CHILDREN(rack_counters),
1702 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1703 	    &rack_sack_proc_restart,
1704 	    "Total times we had to walk whole list due to a restart");
1705 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1706 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1707 	    SYSCTL_CHILDREN(rack_counters),
1708 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1709 	    &rack_sack_proc_short,
1710 	    "Total times we took shortcut for sack processing");
1711 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1712 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1713 	    SYSCTL_CHILDREN(rack_attack),
1714 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1715 	    &rack_sack_skipped_acked,
1716 	    "Total number of times we skipped previously sacked");
1717 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1718 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1719 	    SYSCTL_CHILDREN(rack_attack),
1720 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1721 	    &rack_sack_splits,
1722 	    "Total number of times we did the old fashion tree split");
1723 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1724 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1725 	    SYSCTL_CHILDREN(rack_counters),
1726 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1727 	    &rack_input_idle_reduces,
1728 	    "Total number of idle reductions on input");
1729 	rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1730 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1731 	    SYSCTL_CHILDREN(rack_counters),
1732 	    OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1733 	    &rack_collapsed_win_seen,
1734 	    "Total number of collapsed window events seen (where our window shrinks)");
1735 
1736 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1737 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1738 	    SYSCTL_CHILDREN(rack_counters),
1739 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1740 	    &rack_collapsed_win,
1741 	    "Total number of collapsed window events where we mark packets");
1742 	rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1743 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1744 	    SYSCTL_CHILDREN(rack_counters),
1745 	    OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1746 	    &rack_collapsed_win_rxt,
1747 	    "Total number of packets that were retransmitted");
1748 	rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1749 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1750 	    SYSCTL_CHILDREN(rack_counters),
1751 	    OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1752 	    &rack_collapsed_win_rxt_bytes,
1753 	    "Total number of bytes that were retransmitted");
1754 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1755 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1756 	    SYSCTL_CHILDREN(rack_counters),
1757 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1758 	    &rack_try_scwnd,
1759 	    "Total number of scwnd attempts");
1760 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1761 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1762 	    OID_AUTO, "outsize", CTLFLAG_RD,
1763 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1764 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1765 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1766 	    OID_AUTO, "opts", CTLFLAG_RD,
1767 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1768 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1769 	    SYSCTL_CHILDREN(rack_sysctl_root),
1770 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1771 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1772 }
1773 
1774 static __inline int
1775 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1776 {
1777 	if (SEQ_GEQ(b->r_start, a->r_start) &&
1778 	    SEQ_LT(b->r_start, a->r_end)) {
1779 		/*
1780 		 * The entry b is within the
1781 		 * block a. i.e.:
1782 		 * a --   |-------------|
1783 		 * b --   |----|
1784 		 * <or>
1785 		 * b --       |------|
1786 		 * <or>
1787 		 * b --       |-----------|
1788 		 */
1789 		return (0);
1790 	} else if (SEQ_GEQ(b->r_start, a->r_end)) {
1791 		/*
1792 		 * b falls as either the next
1793 		 * sequence block after a so a
1794 		 * is said to be smaller than b.
1795 		 * i.e:
1796 		 * a --   |------|
1797 		 * b --          |--------|
1798 		 * or
1799 		 * b --              |-----|
1800 		 */
1801 		return (1);
1802 	}
1803 	/*
1804 	 * Whats left is where a is
1805 	 * larger than b. i.e:
1806 	 * a --         |-------|
1807 	 * b --  |---|
1808 	 * or even possibly
1809 	 * b --   |--------------|
1810 	 */
1811 	return (-1);
1812 }
1813 
1814 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1815 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1816 
1817 static uint32_t
1818 rc_init_window(struct tcp_rack *rack)
1819 {
1820 	uint32_t win;
1821 
1822 	if (rack->rc_init_win == 0) {
1823 		/*
1824 		 * Nothing set by the user, use the system stack
1825 		 * default.
1826 		 */
1827 		return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1828 	}
1829 	win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1830 	return (win);
1831 }
1832 
1833 static uint64_t
1834 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1835 {
1836 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1837 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1838 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1839 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1840 	else
1841 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1842 }
1843 
1844 static uint64_t
1845 rack_get_bw(struct tcp_rack *rack)
1846 {
1847 	if (rack->use_fixed_rate) {
1848 		/* Return the fixed pacing rate */
1849 		return (rack_get_fixed_pacing_bw(rack));
1850 	}
1851 	if (rack->r_ctl.gp_bw == 0) {
1852 		/*
1853 		 * We have yet no b/w measurement,
1854 		 * if we have a user set initial bw
1855 		 * return it. If we don't have that and
1856 		 * we have an srtt, use the tcp IW (10) to
1857 		 * calculate a fictional b/w over the SRTT
1858 		 * which is more or less a guess. Note
1859 		 * we don't use our IW from rack on purpose
1860 		 * so if we have like IW=30, we are not
1861 		 * calculating a "huge" b/w.
1862 		 */
1863 		uint64_t bw, srtt;
1864 		if (rack->r_ctl.init_rate)
1865 			return (rack->r_ctl.init_rate);
1866 
1867 		/* Has the user set a max peak rate? */
1868 #ifdef NETFLIX_PEAKRATE
1869 		if (rack->rc_tp->t_maxpeakrate)
1870 			return (rack->rc_tp->t_maxpeakrate);
1871 #endif
1872 		/* Ok lets come up with the IW guess, if we have a srtt */
1873 		if (rack->rc_tp->t_srtt == 0) {
1874 			/*
1875 			 * Go with old pacing method
1876 			 * i.e. burst mitigation only.
1877 			 */
1878 			return (0);
1879 		}
1880 		/* Ok lets get the initial TCP win (not racks) */
1881 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
1882 		srtt = (uint64_t)rack->rc_tp->t_srtt;
1883 		bw *= (uint64_t)USECS_IN_SECOND;
1884 		bw /= srtt;
1885 		if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1886 			bw = rack->r_ctl.bw_rate_cap;
1887 		return (bw);
1888 	} else {
1889 		uint64_t bw;
1890 
1891 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
1892 			/* Averaging is done, we can return the value */
1893 			bw = rack->r_ctl.gp_bw;
1894 		} else {
1895 			/* Still doing initial average must calculate */
1896 			bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
1897 		}
1898 #ifdef NETFLIX_PEAKRATE
1899 		if ((rack->rc_tp->t_maxpeakrate) &&
1900 		    (bw > rack->rc_tp->t_maxpeakrate)) {
1901 			/* The user has set a peak rate to pace at
1902 			 * don't allow us to pace faster than that.
1903 			 */
1904 			return (rack->rc_tp->t_maxpeakrate);
1905 		}
1906 #endif
1907 		if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1908 			bw = rack->r_ctl.bw_rate_cap;
1909 		return (bw);
1910 	}
1911 }
1912 
1913 static uint16_t
1914 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
1915 {
1916 	if (rack->use_fixed_rate) {
1917 		return (100);
1918 	} else if (rack->in_probe_rtt && (rsm == NULL))
1919 		return (rack->r_ctl.rack_per_of_gp_probertt);
1920 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
1921 		  rack->r_ctl.rack_per_of_gp_rec)) {
1922 		if (rsm) {
1923 			/* a retransmission always use the recovery rate */
1924 			return (rack->r_ctl.rack_per_of_gp_rec);
1925 		} else if (rack->rack_rec_nonrxt_use_cr) {
1926 			/* Directed to use the configured rate */
1927 			goto configured_rate;
1928 		} else if (rack->rack_no_prr &&
1929 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
1930 			/* No PRR, lets just use the b/w estimate only */
1931 			return (100);
1932 		} else {
1933 			/*
1934 			 * Here we may have a non-retransmit but we
1935 			 * have no overrides, so just use the recovery
1936 			 * rate (prr is in effect).
1937 			 */
1938 			return (rack->r_ctl.rack_per_of_gp_rec);
1939 		}
1940 	}
1941 configured_rate:
1942 	/* For the configured rate we look at our cwnd vs the ssthresh */
1943 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1944 		return (rack->r_ctl.rack_per_of_gp_ss);
1945 	else
1946 		return (rack->r_ctl.rack_per_of_gp_ca);
1947 }
1948 
1949 static void
1950 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
1951 {
1952 	/*
1953 	 * Types of logs (mod value)
1954 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
1955 	 * 2 = a dsack round begins, persist is reset to 16.
1956 	 * 3 = a dsack round ends
1957 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
1958 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
1959 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
1960 	 */
1961 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1962 		union tcp_log_stackspecific log;
1963 		struct timeval tv;
1964 
1965 		memset(&log, 0, sizeof(log));
1966 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
1967 		log.u_bbr.flex1 <<= 1;
1968 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
1969 		log.u_bbr.flex1 <<= 1;
1970 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
1971 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
1972 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
1973 		log.u_bbr.flex4 = flex4;
1974 		log.u_bbr.flex5 = flex5;
1975 		log.u_bbr.flex6 = flex6;
1976 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
1977 		log.u_bbr.flex8 = mod;
1978 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
1979 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
1980 		    &rack->rc_inp->inp_socket->so_rcv,
1981 		    &rack->rc_inp->inp_socket->so_snd,
1982 		    RACK_DSACK_HANDLING, 0,
1983 		    0, &log, false, &tv);
1984 	}
1985 }
1986 
1987 static void
1988 rack_log_hdwr_pacing(struct tcp_rack *rack,
1989 		     uint64_t rate, uint64_t hw_rate, int line,
1990 		     int error, uint16_t mod)
1991 {
1992 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1993 		union tcp_log_stackspecific log;
1994 		struct timeval tv;
1995 		const struct ifnet *ifp;
1996 
1997 		memset(&log, 0, sizeof(log));
1998 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
1999 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2000 		if (rack->r_ctl.crte) {
2001 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2002 		} else if (rack->rc_inp->inp_route.ro_nh &&
2003 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2004 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2005 		} else
2006 			ifp = NULL;
2007 		if (ifp) {
2008 			log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2009 			log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2010 		}
2011 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2012 		log.u_bbr.bw_inuse = rate;
2013 		log.u_bbr.flex5 = line;
2014 		log.u_bbr.flex6 = error;
2015 		log.u_bbr.flex7 = mod;
2016 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2017 		log.u_bbr.flex8 = rack->use_fixed_rate;
2018 		log.u_bbr.flex8 <<= 1;
2019 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2020 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2021 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2022 		if (rack->r_ctl.crte)
2023 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2024 		else
2025 			log.u_bbr.cur_del_rate = 0;
2026 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2027 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2028 		    &rack->rc_inp->inp_socket->so_rcv,
2029 		    &rack->rc_inp->inp_socket->so_snd,
2030 		    BBR_LOG_HDWR_PACE, 0,
2031 		    0, &log, false, &tv);
2032 	}
2033 }
2034 
2035 static uint64_t
2036 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2037 {
2038 	/*
2039 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2040 	 */
2041 	uint64_t bw_est, high_rate;
2042 	uint64_t gain;
2043 
2044 	gain = (uint64_t)rack_get_output_gain(rack, rsm);
2045 	bw_est = bw * gain;
2046 	bw_est /= (uint64_t)100;
2047 	/* Never fall below the minimum (def 64kbps) */
2048 	if (bw_est < RACK_MIN_BW)
2049 		bw_est = RACK_MIN_BW;
2050 	if (rack->r_rack_hw_rate_caps) {
2051 		/* Rate caps are in place */
2052 		if (rack->r_ctl.crte != NULL) {
2053 			/* We have a hdwr rate already */
2054 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2055 			if (bw_est >= high_rate) {
2056 				/* We are capping bw at the highest rate table entry */
2057 				rack_log_hdwr_pacing(rack,
2058 						     bw_est, high_rate, __LINE__,
2059 						     0, 3);
2060 				bw_est = high_rate;
2061 				if (capped)
2062 					*capped = 1;
2063 			}
2064 		} else if ((rack->rack_hdrw_pacing == 0) &&
2065 			   (rack->rack_hdw_pace_ena) &&
2066 			   (rack->rack_attempt_hdwr_pace == 0) &&
2067 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2068 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2069 			/*
2070 			 * Special case, we have not yet attempted hardware
2071 			 * pacing, and yet we may, when we do, find out if we are
2072 			 * above the highest rate. We need to know the maxbw for the interface
2073 			 * in question (if it supports ratelimiting). We get back
2074 			 * a 0, if the interface is not found in the RL lists.
2075 			 */
2076 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2077 			if (high_rate) {
2078 				/* Yep, we have a rate is it above this rate? */
2079 				if (bw_est > high_rate) {
2080 					bw_est = high_rate;
2081 					if (capped)
2082 						*capped = 1;
2083 				}
2084 			}
2085 		}
2086 	}
2087 	return (bw_est);
2088 }
2089 
2090 static void
2091 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2092 {
2093 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2094 		union tcp_log_stackspecific log;
2095 		struct timeval tv;
2096 
2097 		if ((mod != 1) && (rack_verbose_logging == 0)) {
2098 			/*
2099 			 * We get 3 values currently for mod
2100 			 * 1 - We are retransmitting and this tells the reason.
2101 			 * 2 - We are clearing a dup-ack count.
2102 			 * 3 - We are incrementing a dup-ack count.
2103 			 *
2104 			 * The clear/increment are only logged
2105 			 * if you have BBverbose on.
2106 			 */
2107 			return;
2108 		}
2109 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2110 		log.u_bbr.flex1 = tsused;
2111 		log.u_bbr.flex2 = thresh;
2112 		log.u_bbr.flex3 = rsm->r_flags;
2113 		log.u_bbr.flex4 = rsm->r_dupack;
2114 		log.u_bbr.flex5 = rsm->r_start;
2115 		log.u_bbr.flex6 = rsm->r_end;
2116 		log.u_bbr.flex8 = mod;
2117 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2118 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2119 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2120 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2121 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2122 		log.u_bbr.pacing_gain = rack->r_must_retran;
2123 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2124 		    &rack->rc_inp->inp_socket->so_rcv,
2125 		    &rack->rc_inp->inp_socket->so_snd,
2126 		    BBR_LOG_SETTINGS_CHG, 0,
2127 		    0, &log, false, &tv);
2128 	}
2129 }
2130 
2131 static void
2132 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2133 {
2134 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2135 		union tcp_log_stackspecific log;
2136 		struct timeval tv;
2137 
2138 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2139 		log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2140 		log.u_bbr.flex2 = to;
2141 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2142 		log.u_bbr.flex4 = slot;
2143 		log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2144 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2145 		log.u_bbr.flex7 = rack->rc_in_persist;
2146 		log.u_bbr.flex8 = which;
2147 		if (rack->rack_no_prr)
2148 			log.u_bbr.pkts_out = 0;
2149 		else
2150 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2151 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2152 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2153 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2154 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2155 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2156 		log.u_bbr.pacing_gain = rack->r_must_retran;
2157 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2158 		log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2159 		log.u_bbr.lost = rack_rto_min;
2160 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2161 		    &rack->rc_inp->inp_socket->so_rcv,
2162 		    &rack->rc_inp->inp_socket->so_snd,
2163 		    BBR_LOG_TIMERSTAR, 0,
2164 		    0, &log, false, &tv);
2165 	}
2166 }
2167 
2168 static void
2169 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2170 {
2171 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2172 		union tcp_log_stackspecific log;
2173 		struct timeval tv;
2174 
2175 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2176 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2177 		log.u_bbr.flex8 = to_num;
2178 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2179 		log.u_bbr.flex2 = rack->rc_rack_rtt;
2180 		if (rsm == NULL)
2181 			log.u_bbr.flex3 = 0;
2182 		else
2183 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2184 		if (rack->rack_no_prr)
2185 			log.u_bbr.flex5 = 0;
2186 		else
2187 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2188 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2189 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2190 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2191 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2192 		log.u_bbr.pacing_gain = rack->r_must_retran;
2193 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2194 		    &rack->rc_inp->inp_socket->so_rcv,
2195 		    &rack->rc_inp->inp_socket->so_snd,
2196 		    BBR_LOG_RTO, 0,
2197 		    0, &log, false, &tv);
2198 	}
2199 }
2200 
2201 static void
2202 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2203 		 struct rack_sendmap *prev,
2204 		 struct rack_sendmap *rsm,
2205 		 struct rack_sendmap *next,
2206 		 int flag, uint32_t th_ack, int line)
2207 {
2208 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2209 		union tcp_log_stackspecific log;
2210 		struct timeval tv;
2211 
2212 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2213 		log.u_bbr.flex8 = flag;
2214 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2215 		log.u_bbr.cur_del_rate = (uint64_t)prev;
2216 		log.u_bbr.delRate = (uint64_t)rsm;
2217 		log.u_bbr.rttProp = (uint64_t)next;
2218 		log.u_bbr.flex7 = 0;
2219 		if (prev) {
2220 			log.u_bbr.flex1 = prev->r_start;
2221 			log.u_bbr.flex2 = prev->r_end;
2222 			log.u_bbr.flex7 |= 0x4;
2223 		}
2224 		if (rsm) {
2225 			log.u_bbr.flex3 = rsm->r_start;
2226 			log.u_bbr.flex4 = rsm->r_end;
2227 			log.u_bbr.flex7 |= 0x2;
2228 		}
2229 		if (next) {
2230 			log.u_bbr.flex5 = next->r_start;
2231 			log.u_bbr.flex6 = next->r_end;
2232 			log.u_bbr.flex7 |= 0x1;
2233 		}
2234 		log.u_bbr.applimited = line;
2235 		log.u_bbr.pkts_out = th_ack;
2236 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2237 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2238 		if (rack->rack_no_prr)
2239 			log.u_bbr.lost = 0;
2240 		else
2241 			log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2242 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2243 		    &rack->rc_inp->inp_socket->so_rcv,
2244 		    &rack->rc_inp->inp_socket->so_snd,
2245 		    TCP_LOG_MAPCHG, 0,
2246 		    0, &log, false, &tv);
2247 	}
2248 }
2249 
2250 static void
2251 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2252 		 struct rack_sendmap *rsm, int conf)
2253 {
2254 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2255 		union tcp_log_stackspecific log;
2256 		struct timeval tv;
2257 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2258 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2259 		log.u_bbr.flex1 = t;
2260 		log.u_bbr.flex2 = len;
2261 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2262 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2263 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2264 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2265 		log.u_bbr.flex7 = conf;
2266 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2267 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2268 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2269 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2270 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2271 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2272 		if (rsm) {
2273 			log.u_bbr.pkt_epoch = rsm->r_start;
2274 			log.u_bbr.lost = rsm->r_end;
2275 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2276 			/* We loose any upper of the 24 bits */
2277 			log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2278 		} else {
2279 			/* Its a SYN */
2280 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2281 			log.u_bbr.lost = 0;
2282 			log.u_bbr.cwnd_gain = 0;
2283 			log.u_bbr.pacing_gain = 0;
2284 		}
2285 		/* Write out general bits of interest rrs here */
2286 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2287 		log.u_bbr.use_lt_bw <<= 1;
2288 		log.u_bbr.use_lt_bw |= rack->forced_ack;
2289 		log.u_bbr.use_lt_bw <<= 1;
2290 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2291 		log.u_bbr.use_lt_bw <<= 1;
2292 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2293 		log.u_bbr.use_lt_bw <<= 1;
2294 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2295 		log.u_bbr.use_lt_bw <<= 1;
2296 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2297 		log.u_bbr.use_lt_bw <<= 1;
2298 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2299 		log.u_bbr.use_lt_bw <<= 1;
2300 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2301 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2302 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2303 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2304 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2305 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2306 		log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2307 		log.u_bbr.bw_inuse <<= 32;
2308 		if (rsm)
2309 			log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2310 		TCP_LOG_EVENTP(tp, NULL,
2311 		    &rack->rc_inp->inp_socket->so_rcv,
2312 		    &rack->rc_inp->inp_socket->so_snd,
2313 		    BBR_LOG_BBRRTT, 0,
2314 		    0, &log, false, &tv);
2315 
2316 
2317 	}
2318 }
2319 
2320 static void
2321 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2322 {
2323 	/*
2324 	 * Log the rtt sample we are
2325 	 * applying to the srtt algorithm in
2326 	 * useconds.
2327 	 */
2328 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2329 		union tcp_log_stackspecific log;
2330 		struct timeval tv;
2331 
2332 		/* Convert our ms to a microsecond */
2333 		memset(&log, 0, sizeof(log));
2334 		log.u_bbr.flex1 = rtt;
2335 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2336 		log.u_bbr.flex3 = rack->r_ctl.sack_count;
2337 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2338 		log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2339 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2340 		log.u_bbr.flex7 = 1;
2341 		log.u_bbr.flex8 = rack->sack_attack_disable;
2342 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2343 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2344 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2345 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2346 		log.u_bbr.pacing_gain = rack->r_must_retran;
2347 		/*
2348 		 * We capture in delRate the upper 32 bits as
2349 		 * the confidence level we had declared, and the
2350 		 * lower 32 bits as the actual RTT using the arrival
2351 		 * timestamp.
2352 		 */
2353 		log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2354 		log.u_bbr.delRate <<= 32;
2355 		log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2356 		/* Lets capture all the things that make up t_rtxcur */
2357 		log.u_bbr.applimited = rack_rto_min;
2358 		log.u_bbr.epoch = rack_rto_max;
2359 		log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2360 		log.u_bbr.lost = rack_rto_min;
2361 		log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2362 		log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2363 		log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2364 		log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2365 		log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2366 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2367 		    &rack->rc_inp->inp_socket->so_rcv,
2368 		    &rack->rc_inp->inp_socket->so_snd,
2369 		    TCP_LOG_RTT, 0,
2370 		    0, &log, false, &tv);
2371 	}
2372 }
2373 
2374 static void
2375 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2376 {
2377 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2378 		union tcp_log_stackspecific log;
2379 		struct timeval tv;
2380 
2381 		/* Convert our ms to a microsecond */
2382 		memset(&log, 0, sizeof(log));
2383 		log.u_bbr.flex1 = rtt;
2384 		log.u_bbr.flex2 = send_time;
2385 		log.u_bbr.flex3 = ack_time;
2386 		log.u_bbr.flex4 = where;
2387 		log.u_bbr.flex7 = 2;
2388 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2389 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2390 		    &rack->rc_inp->inp_socket->so_rcv,
2391 		    &rack->rc_inp->inp_socket->so_snd,
2392 		    TCP_LOG_RTT, 0,
2393 		    0, &log, false, &tv);
2394 	}
2395 }
2396 
2397 
2398 
2399 static inline void
2400 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2401 {
2402 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2403 		union tcp_log_stackspecific log;
2404 		struct timeval tv;
2405 
2406 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2407 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2408 		log.u_bbr.flex1 = line;
2409 		log.u_bbr.flex2 = tick;
2410 		log.u_bbr.flex3 = tp->t_maxunacktime;
2411 		log.u_bbr.flex4 = tp->t_acktime;
2412 		log.u_bbr.flex8 = event;
2413 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2414 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2415 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2416 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2417 		log.u_bbr.pacing_gain = rack->r_must_retran;
2418 		TCP_LOG_EVENTP(tp, NULL,
2419 		    &rack->rc_inp->inp_socket->so_rcv,
2420 		    &rack->rc_inp->inp_socket->so_snd,
2421 		    BBR_LOG_PROGRESS, 0,
2422 		    0, &log, false, &tv);
2423 	}
2424 }
2425 
2426 static void
2427 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2428 {
2429 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2430 		union tcp_log_stackspecific log;
2431 
2432 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2433 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2434 		log.u_bbr.flex1 = slot;
2435 		if (rack->rack_no_prr)
2436 			log.u_bbr.flex2 = 0;
2437 		else
2438 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2439 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2440 		log.u_bbr.flex8 = rack->rc_in_persist;
2441 		log.u_bbr.timeStamp = cts;
2442 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2443 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2444 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2445 		log.u_bbr.pacing_gain = rack->r_must_retran;
2446 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2447 		    &rack->rc_inp->inp_socket->so_rcv,
2448 		    &rack->rc_inp->inp_socket->so_snd,
2449 		    BBR_LOG_BBRSND, 0,
2450 		    0, &log, false, tv);
2451 	}
2452 }
2453 
2454 static void
2455 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2456 {
2457 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2458 		union tcp_log_stackspecific log;
2459 		struct timeval tv;
2460 
2461 		memset(&log, 0, sizeof(log));
2462 		log.u_bbr.flex1 = did_out;
2463 		log.u_bbr.flex2 = nxt_pkt;
2464 		log.u_bbr.flex3 = way_out;
2465 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2466 		if (rack->rack_no_prr)
2467 			log.u_bbr.flex5 = 0;
2468 		else
2469 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2470 		log.u_bbr.flex6 = nsegs;
2471 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2472 		log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;	/* Do we have ack-can-send set */
2473 		log.u_bbr.flex7 <<= 1;
2474 		log.u_bbr.flex7 |= rack->r_fast_output;	/* is fast output primed */
2475 		log.u_bbr.flex7 <<= 1;
2476 		log.u_bbr.flex7 |= rack->r_wanted_output;	/* Do we want output */
2477 		log.u_bbr.flex8 = rack->rc_in_persist;
2478 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2479 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2480 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2481 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2482 		log.u_bbr.use_lt_bw <<= 1;
2483 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
2484 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2485 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2486 		log.u_bbr.pacing_gain = rack->r_must_retran;
2487 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2488 		    &rack->rc_inp->inp_socket->so_rcv,
2489 		    &rack->rc_inp->inp_socket->so_snd,
2490 		    BBR_LOG_DOSEG_DONE, 0,
2491 		    0, &log, false, &tv);
2492 	}
2493 }
2494 
2495 static void
2496 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2497 {
2498 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2499 		union tcp_log_stackspecific log;
2500 		struct timeval tv;
2501 
2502 		memset(&log, 0, sizeof(log));
2503 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2504 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2505 		log.u_bbr.flex4 = arg1;
2506 		log.u_bbr.flex5 = arg2;
2507 		log.u_bbr.flex6 = arg3;
2508 		log.u_bbr.flex8 = frm;
2509 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2510 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2511 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2512 		log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2513 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2514 		log.u_bbr.pacing_gain = rack->r_must_retran;
2515 		TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
2516 		    &tptosocket(tp)->so_snd,
2517 		    TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
2518 	}
2519 }
2520 
2521 static void
2522 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2523 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2524 {
2525 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2526 		union tcp_log_stackspecific log;
2527 		struct timeval tv;
2528 
2529 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2530 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2531 		log.u_bbr.flex1 = slot;
2532 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2533 		log.u_bbr.flex4 = reason;
2534 		if (rack->rack_no_prr)
2535 			log.u_bbr.flex5 = 0;
2536 		else
2537 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2538 		log.u_bbr.flex7 = hpts_calling;
2539 		log.u_bbr.flex8 = rack->rc_in_persist;
2540 		log.u_bbr.lt_epoch = cwnd_to_use;
2541 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2542 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2543 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2544 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2545 		log.u_bbr.pacing_gain = rack->r_must_retran;
2546 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2547 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2548 		    &rack->rc_inp->inp_socket->so_rcv,
2549 		    &rack->rc_inp->inp_socket->so_snd,
2550 		    BBR_LOG_JUSTRET, 0,
2551 		    tlen, &log, false, &tv);
2552 	}
2553 }
2554 
2555 static void
2556 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2557 		   struct timeval *tv, uint32_t flags_on_entry)
2558 {
2559 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2560 		union tcp_log_stackspecific log;
2561 
2562 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2563 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2564 		log.u_bbr.flex1 = line;
2565 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2566 		log.u_bbr.flex3 = flags_on_entry;
2567 		log.u_bbr.flex4 = us_cts;
2568 		if (rack->rack_no_prr)
2569 			log.u_bbr.flex5 = 0;
2570 		else
2571 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2572 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2573 		log.u_bbr.flex7 = hpts_removed;
2574 		log.u_bbr.flex8 = 1;
2575 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2576 		log.u_bbr.timeStamp = us_cts;
2577 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2578 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2579 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2580 		log.u_bbr.pacing_gain = rack->r_must_retran;
2581 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2582 		    &rack->rc_inp->inp_socket->so_rcv,
2583 		    &rack->rc_inp->inp_socket->so_snd,
2584 		    BBR_LOG_TIMERCANC, 0,
2585 		    0, &log, false, tv);
2586 	}
2587 }
2588 
2589 static void
2590 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2591 			  uint32_t flex1, uint32_t flex2,
2592 			  uint32_t flex3, uint32_t flex4,
2593 			  uint32_t flex5, uint32_t flex6,
2594 			  uint16_t flex7, uint8_t mod)
2595 {
2596 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2597 		union tcp_log_stackspecific log;
2598 		struct timeval tv;
2599 
2600 		if (mod == 1) {
2601 			/* No you can't use 1, its for the real to cancel */
2602 			return;
2603 		}
2604 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2605 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2606 		log.u_bbr.flex1 = flex1;
2607 		log.u_bbr.flex2 = flex2;
2608 		log.u_bbr.flex3 = flex3;
2609 		log.u_bbr.flex4 = flex4;
2610 		log.u_bbr.flex5 = flex5;
2611 		log.u_bbr.flex6 = flex6;
2612 		log.u_bbr.flex7 = flex7;
2613 		log.u_bbr.flex8 = mod;
2614 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2615 		    &rack->rc_inp->inp_socket->so_rcv,
2616 		    &rack->rc_inp->inp_socket->so_snd,
2617 		    BBR_LOG_TIMERCANC, 0,
2618 		    0, &log, false, &tv);
2619 	}
2620 }
2621 
2622 static void
2623 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2624 {
2625 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2626 		union tcp_log_stackspecific log;
2627 		struct timeval tv;
2628 
2629 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2630 		log.u_bbr.flex1 = timers;
2631 		log.u_bbr.flex2 = ret;
2632 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2633 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2634 		log.u_bbr.flex5 = cts;
2635 		if (rack->rack_no_prr)
2636 			log.u_bbr.flex6 = 0;
2637 		else
2638 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2639 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2640 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2641 		log.u_bbr.pacing_gain = rack->r_must_retran;
2642 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2643 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2644 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2645 		    &rack->rc_inp->inp_socket->so_rcv,
2646 		    &rack->rc_inp->inp_socket->so_snd,
2647 		    BBR_LOG_TO_PROCESS, 0,
2648 		    0, &log, false, &tv);
2649 	}
2650 }
2651 
2652 static void
2653 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
2654 {
2655 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2656 		union tcp_log_stackspecific log;
2657 		struct timeval tv;
2658 
2659 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2660 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2661 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2662 		if (rack->rack_no_prr)
2663 			log.u_bbr.flex3 = 0;
2664 		else
2665 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2666 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2667 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2668 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2669 		log.u_bbr.flex7 = line;
2670 		log.u_bbr.flex8 = frm;
2671 		log.u_bbr.pkts_out = orig_cwnd;
2672 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2673 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2674 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2675 		log.u_bbr.use_lt_bw <<= 1;
2676 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
2677 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2678 		    &rack->rc_inp->inp_socket->so_rcv,
2679 		    &rack->rc_inp->inp_socket->so_snd,
2680 		    BBR_LOG_BBRUPD, 0,
2681 		    0, &log, false, &tv);
2682 	}
2683 }
2684 
2685 #ifdef NETFLIX_EXP_DETECTION
2686 static void
2687 rack_log_sad(struct tcp_rack *rack, int event)
2688 {
2689 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2690 		union tcp_log_stackspecific log;
2691 		struct timeval tv;
2692 
2693 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2694 		log.u_bbr.flex1 = rack->r_ctl.sack_count;
2695 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2696 		log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2697 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2698 		log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2699 		log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2700 		log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2701 		log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2702 		log.u_bbr.lt_epoch |= rack->do_detection;
2703 		log.u_bbr.applimited = tcp_map_minimum;
2704 		log.u_bbr.flex7 = rack->sack_attack_disable;
2705 		log.u_bbr.flex8 = event;
2706 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2707 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2708 		log.u_bbr.delivered = tcp_sad_decay_val;
2709 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2710 		    &rack->rc_inp->inp_socket->so_rcv,
2711 		    &rack->rc_inp->inp_socket->so_snd,
2712 		    TCP_SAD_DETECTION, 0,
2713 		    0, &log, false, &tv);
2714 	}
2715 }
2716 #endif
2717 
2718 static void
2719 rack_counter_destroy(void)
2720 {
2721 	counter_u64_free(rack_fto_send);
2722 	counter_u64_free(rack_fto_rsm_send);
2723 	counter_u64_free(rack_nfto_resend);
2724 	counter_u64_free(rack_hw_pace_init_fail);
2725 	counter_u64_free(rack_hw_pace_lost);
2726 	counter_u64_free(rack_non_fto_send);
2727 	counter_u64_free(rack_extended_rfo);
2728 	counter_u64_free(rack_ack_total);
2729 	counter_u64_free(rack_express_sack);
2730 	counter_u64_free(rack_sack_total);
2731 	counter_u64_free(rack_move_none);
2732 	counter_u64_free(rack_move_some);
2733 	counter_u64_free(rack_sack_attacks_detected);
2734 	counter_u64_free(rack_sack_attacks_reversed);
2735 	counter_u64_free(rack_sack_used_next_merge);
2736 	counter_u64_free(rack_sack_used_prev_merge);
2737 	counter_u64_free(rack_tlp_tot);
2738 	counter_u64_free(rack_tlp_newdata);
2739 	counter_u64_free(rack_tlp_retran);
2740 	counter_u64_free(rack_tlp_retran_bytes);
2741 	counter_u64_free(rack_to_tot);
2742 	counter_u64_free(rack_saw_enobuf);
2743 	counter_u64_free(rack_saw_enobuf_hw);
2744 	counter_u64_free(rack_saw_enetunreach);
2745 	counter_u64_free(rack_hot_alloc);
2746 	counter_u64_free(rack_to_alloc);
2747 	counter_u64_free(rack_to_alloc_hard);
2748 	counter_u64_free(rack_to_alloc_emerg);
2749 	counter_u64_free(rack_to_alloc_limited);
2750 	counter_u64_free(rack_alloc_limited_conns);
2751 	counter_u64_free(rack_split_limited);
2752 	counter_u64_free(rack_multi_single_eq);
2753 	counter_u64_free(rack_proc_non_comp_ack);
2754 	counter_u64_free(rack_sack_proc_all);
2755 	counter_u64_free(rack_sack_proc_restart);
2756 	counter_u64_free(rack_sack_proc_short);
2757 	counter_u64_free(rack_sack_skipped_acked);
2758 	counter_u64_free(rack_sack_splits);
2759 	counter_u64_free(rack_input_idle_reduces);
2760 	counter_u64_free(rack_collapsed_win);
2761 	counter_u64_free(rack_collapsed_win_rxt);
2762 	counter_u64_free(rack_collapsed_win_rxt_bytes);
2763 	counter_u64_free(rack_collapsed_win_seen);
2764 	counter_u64_free(rack_try_scwnd);
2765 	counter_u64_free(rack_persists_sends);
2766 	counter_u64_free(rack_persists_acks);
2767 	counter_u64_free(rack_persists_loss);
2768 	counter_u64_free(rack_persists_lost_ends);
2769 #ifdef INVARIANTS
2770 	counter_u64_free(rack_adjust_map_bw);
2771 #endif
2772 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2773 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2774 }
2775 
2776 static struct rack_sendmap *
2777 rack_alloc(struct tcp_rack *rack)
2778 {
2779 	struct rack_sendmap *rsm;
2780 
2781 	/*
2782 	 * First get the top of the list it in
2783 	 * theory is the "hottest" rsm we have,
2784 	 * possibly just freed by ack processing.
2785 	 */
2786 	if (rack->rc_free_cnt > rack_free_cache) {
2787 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2788 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2789 		counter_u64_add(rack_hot_alloc, 1);
2790 		rack->rc_free_cnt--;
2791 		return (rsm);
2792 	}
2793 	/*
2794 	 * Once we get under our free cache we probably
2795 	 * no longer have a "hot" one available. Lets
2796 	 * get one from UMA.
2797 	 */
2798 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
2799 	if (rsm) {
2800 		rack->r_ctl.rc_num_maps_alloced++;
2801 		counter_u64_add(rack_to_alloc, 1);
2802 		return (rsm);
2803 	}
2804 	/*
2805 	 * Dig in to our aux rsm's (the last two) since
2806 	 * UMA failed to get us one.
2807 	 */
2808 	if (rack->rc_free_cnt) {
2809 		counter_u64_add(rack_to_alloc_emerg, 1);
2810 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2811 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2812 		rack->rc_free_cnt--;
2813 		return (rsm);
2814 	}
2815 	return (NULL);
2816 }
2817 
2818 static struct rack_sendmap *
2819 rack_alloc_full_limit(struct tcp_rack *rack)
2820 {
2821 	if ((V_tcp_map_entries_limit > 0) &&
2822 	    (rack->do_detection == 0) &&
2823 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2824 		counter_u64_add(rack_to_alloc_limited, 1);
2825 		if (!rack->alloc_limit_reported) {
2826 			rack->alloc_limit_reported = 1;
2827 			counter_u64_add(rack_alloc_limited_conns, 1);
2828 		}
2829 		return (NULL);
2830 	}
2831 	return (rack_alloc(rack));
2832 }
2833 
2834 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2835 static struct rack_sendmap *
2836 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2837 {
2838 	struct rack_sendmap *rsm;
2839 
2840 	if (limit_type) {
2841 		/* currently there is only one limit type */
2842 		if (V_tcp_map_split_limit > 0 &&
2843 		    (rack->do_detection == 0) &&
2844 		    rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2845 			counter_u64_add(rack_split_limited, 1);
2846 			if (!rack->alloc_limit_reported) {
2847 				rack->alloc_limit_reported = 1;
2848 				counter_u64_add(rack_alloc_limited_conns, 1);
2849 			}
2850 			return (NULL);
2851 		}
2852 	}
2853 
2854 	/* allocate and mark in the limit type, if set */
2855 	rsm = rack_alloc(rack);
2856 	if (rsm != NULL && limit_type) {
2857 		rsm->r_limit_type = limit_type;
2858 		rack->r_ctl.rc_num_split_allocs++;
2859 	}
2860 	return (rsm);
2861 }
2862 
2863 static void
2864 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2865 {
2866 	if (rsm->r_flags & RACK_APP_LIMITED) {
2867 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
2868 			rack->r_ctl.rc_app_limited_cnt--;
2869 		}
2870 	}
2871 	if (rsm->r_limit_type) {
2872 		/* currently there is only one limit type */
2873 		rack->r_ctl.rc_num_split_allocs--;
2874 	}
2875 	if (rsm == rack->r_ctl.rc_first_appl) {
2876 		if (rack->r_ctl.rc_app_limited_cnt == 0)
2877 			rack->r_ctl.rc_first_appl = NULL;
2878 		else {
2879 			/* Follow the next one out */
2880 			struct rack_sendmap fe;
2881 
2882 			fe.r_start = rsm->r_nseq_appl;
2883 			rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
2884 		}
2885 	}
2886 	if (rsm == rack->r_ctl.rc_resend)
2887 		rack->r_ctl.rc_resend = NULL;
2888 	if (rsm == rack->r_ctl.rc_end_appl)
2889 		rack->r_ctl.rc_end_appl = NULL;
2890 	if (rack->r_ctl.rc_tlpsend == rsm)
2891 		rack->r_ctl.rc_tlpsend = NULL;
2892 	if (rack->r_ctl.rc_sacklast == rsm)
2893 		rack->r_ctl.rc_sacklast = NULL;
2894 	memset(rsm, 0, sizeof(struct rack_sendmap));
2895 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
2896 	rack->rc_free_cnt++;
2897 }
2898 
2899 static void
2900 rack_free_trim(struct tcp_rack *rack)
2901 {
2902 	struct rack_sendmap *rsm;
2903 
2904 	/*
2905 	 * Free up all the tail entries until
2906 	 * we get our list down to the limit.
2907 	 */
2908 	while (rack->rc_free_cnt > rack_free_cache) {
2909 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
2910 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2911 		rack->rc_free_cnt--;
2912 		uma_zfree(rack_zone, rsm);
2913 	}
2914 }
2915 
2916 
2917 static uint32_t
2918 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
2919 {
2920 	uint64_t srtt, bw, len, tim;
2921 	uint32_t segsiz, def_len, minl;
2922 
2923 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2924 	def_len = rack_def_data_window * segsiz;
2925 	if (rack->rc_gp_filled == 0) {
2926 		/*
2927 		 * We have no measurement (IW is in flight?) so
2928 		 * we can only guess using our data_window sysctl
2929 		 * value (usually 20MSS).
2930 		 */
2931 		return (def_len);
2932 	}
2933 	/*
2934 	 * Now we have a number of factors to consider.
2935 	 *
2936 	 * 1) We have a desired BDP which is usually
2937 	 *    at least 2.
2938 	 * 2) We have a minimum number of rtt's usually 1 SRTT
2939 	 *    but we allow it too to be more.
2940 	 * 3) We want to make sure a measurement last N useconds (if
2941 	 *    we have set rack_min_measure_usec.
2942 	 *
2943 	 * We handle the first concern here by trying to create a data
2944 	 * window of max(rack_def_data_window, DesiredBDP). The
2945 	 * second concern we handle in not letting the measurement
2946 	 * window end normally until at least the required SRTT's
2947 	 * have gone by which is done further below in
2948 	 * rack_enough_for_measurement(). Finally the third concern
2949 	 * we also handle here by calculating how long that time
2950 	 * would take at the current BW and then return the
2951 	 * max of our first calculation and that length. Note
2952 	 * that if rack_min_measure_usec is 0, we don't deal
2953 	 * with concern 3. Also for both Concern 1 and 3 an
2954 	 * application limited period could end the measurement
2955 	 * earlier.
2956 	 *
2957 	 * So lets calculate the BDP with the "known" b/w using
2958 	 * the SRTT has our rtt and then multiply it by the
2959 	 * goal.
2960 	 */
2961 	bw = rack_get_bw(rack);
2962 	srtt = (uint64_t)tp->t_srtt;
2963 	len = bw * srtt;
2964 	len /= (uint64_t)HPTS_USEC_IN_SEC;
2965 	len *= max(1, rack_goal_bdp);
2966 	/* Now we need to round up to the nearest MSS */
2967 	len = roundup(len, segsiz);
2968 	if (rack_min_measure_usec) {
2969 		/* Now calculate our min length for this b/w */
2970 		tim = rack_min_measure_usec;
2971 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
2972 		if (minl == 0)
2973 			minl = 1;
2974 		minl = roundup(minl, segsiz);
2975 		if (len < minl)
2976 			len = minl;
2977 	}
2978 	/*
2979 	 * Now if we have a very small window we want
2980 	 * to attempt to get the window that is
2981 	 * as small as possible. This happens on
2982 	 * low b/w connections and we don't want to
2983 	 * span huge numbers of rtt's between measurements.
2984 	 *
2985 	 * We basically include 2 over our "MIN window" so
2986 	 * that the measurement can be shortened (possibly) by
2987 	 * an ack'ed packet.
2988 	 */
2989 	if (len < def_len)
2990 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
2991 	else
2992 		return (max((uint32_t)len, def_len));
2993 
2994 }
2995 
2996 static int
2997 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
2998 {
2999 	uint32_t tim, srtts, segsiz;
3000 
3001 	/*
3002 	 * Has enough time passed for the GP measurement to be valid?
3003 	 */
3004 	if ((tp->snd_max == tp->snd_una) ||
3005 	    (th_ack == tp->snd_max)){
3006 		/* All is acked */
3007 		*quality = RACK_QUALITY_ALLACKED;
3008 		return (1);
3009 	}
3010 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3011 		/* Not enough bytes yet */
3012 		return (0);
3013 	}
3014 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3015 	if (SEQ_LT(th_ack, tp->gput_ack) &&
3016 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3017 		/* Not enough bytes yet */
3018 		return (0);
3019 	}
3020 	if (rack->r_ctl.rc_first_appl &&
3021 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3022 		/*
3023 		 * We are up to the app limited send point
3024 		 * we have to measure irrespective of the time..
3025 		 */
3026 		*quality = RACK_QUALITY_APPLIMITED;
3027 		return (1);
3028 	}
3029 	/* Now what about time? */
3030 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3031 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3032 	if (tim >= srtts) {
3033 		*quality = RACK_QUALITY_HIGH;
3034 		return (1);
3035 	}
3036 	/* Nope not even a full SRTT has passed */
3037 	return (0);
3038 }
3039 
3040 static void
3041 rack_log_timely(struct tcp_rack *rack,
3042 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3043 		uint64_t up_bnd, int line, uint8_t method)
3044 {
3045 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3046 		union tcp_log_stackspecific log;
3047 		struct timeval tv;
3048 
3049 		memset(&log, 0, sizeof(log));
3050 		log.u_bbr.flex1 = logged;
3051 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3052 		log.u_bbr.flex2 <<= 4;
3053 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3054 		log.u_bbr.flex2 <<= 4;
3055 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3056 		log.u_bbr.flex2 <<= 4;
3057 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3058 		log.u_bbr.flex3 = rack->rc_gp_incr;
3059 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3060 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3061 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3062 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3063 		log.u_bbr.flex8 = method;
3064 		log.u_bbr.cur_del_rate = cur_bw;
3065 		log.u_bbr.delRate = low_bnd;
3066 		log.u_bbr.bw_inuse = up_bnd;
3067 		log.u_bbr.rttProp = rack_get_bw(rack);
3068 		log.u_bbr.pkt_epoch = line;
3069 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3070 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3071 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3072 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3073 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3074 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3075 		log.u_bbr.cwnd_gain <<= 1;
3076 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3077 		log.u_bbr.cwnd_gain <<= 1;
3078 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3079 		log.u_bbr.cwnd_gain <<= 1;
3080 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3081 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3082 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3083 		    &rack->rc_inp->inp_socket->so_rcv,
3084 		    &rack->rc_inp->inp_socket->so_snd,
3085 		    TCP_TIMELY_WORK, 0,
3086 		    0, &log, false, &tv);
3087 	}
3088 }
3089 
3090 static int
3091 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3092 {
3093 	/*
3094 	 * Before we increase we need to know if
3095 	 * the estimate just made was less than
3096 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3097 	 *
3098 	 * If we already are pacing at a fast enough
3099 	 * rate to push us faster there is no sense of
3100 	 * increasing.
3101 	 *
3102 	 * We first caculate our actual pacing rate (ss or ca multiplier
3103 	 * times our cur_bw).
3104 	 *
3105 	 * Then we take the last measured rate and multipy by our
3106 	 * maximum pacing overage to give us a max allowable rate.
3107 	 *
3108 	 * If our act_rate is smaller than our max_allowable rate
3109 	 * then we should increase. Else we should hold steady.
3110 	 *
3111 	 */
3112 	uint64_t act_rate, max_allow_rate;
3113 
3114 	if (rack_timely_no_stopping)
3115 		return (1);
3116 
3117 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3118 		/*
3119 		 * Initial startup case or
3120 		 * everything is acked case.
3121 		 */
3122 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3123 				__LINE__, 9);
3124 		return (1);
3125 	}
3126 	if (mult <= 100) {
3127 		/*
3128 		 * We can always pace at or slightly above our rate.
3129 		 */
3130 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3131 				__LINE__, 9);
3132 		return (1);
3133 	}
3134 	act_rate = cur_bw * (uint64_t)mult;
3135 	act_rate /= 100;
3136 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3137 	max_allow_rate /= 100;
3138 	if (act_rate < max_allow_rate) {
3139 		/*
3140 		 * Here the rate we are actually pacing at
3141 		 * is smaller than 10% above our last measurement.
3142 		 * This means we are pacing below what we would
3143 		 * like to try to achieve (plus some wiggle room).
3144 		 */
3145 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3146 				__LINE__, 9);
3147 		return (1);
3148 	} else {
3149 		/*
3150 		 * Here we are already pacing at least rack_max_per_above(10%)
3151 		 * what we are getting back. This indicates most likely
3152 		 * that we are being limited (cwnd/rwnd/app) and can't
3153 		 * get any more b/w. There is no sense of trying to
3154 		 * raise up the pacing rate its not speeding us up
3155 		 * and we already are pacing faster than we are getting.
3156 		 */
3157 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3158 				__LINE__, 8);
3159 		return (0);
3160 	}
3161 }
3162 
3163 static void
3164 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3165 {
3166 	/*
3167 	 * When we drag bottom, we want to assure
3168 	 * that no multiplier is below 1.0, if so
3169 	 * we want to restore it to at least that.
3170 	 */
3171 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3172 		/* This is unlikely we usually do not touch recovery */
3173 		rack->r_ctl.rack_per_of_gp_rec = 100;
3174 	}
3175 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3176 		rack->r_ctl.rack_per_of_gp_ca = 100;
3177 	}
3178 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3179 		rack->r_ctl.rack_per_of_gp_ss = 100;
3180 	}
3181 }
3182 
3183 static void
3184 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3185 {
3186 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3187 		rack->r_ctl.rack_per_of_gp_ca = 100;
3188 	}
3189 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3190 		rack->r_ctl.rack_per_of_gp_ss = 100;
3191 	}
3192 }
3193 
3194 static void
3195 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3196 {
3197 	int32_t  calc, logged, plus;
3198 
3199 	logged = 0;
3200 
3201 	if (override) {
3202 		/*
3203 		 * override is passed when we are
3204 		 * loosing b/w and making one last
3205 		 * gasp at trying to not loose out
3206 		 * to a new-reno flow.
3207 		 */
3208 		goto extra_boost;
3209 	}
3210 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3211 	if (rack->rc_gp_incr &&
3212 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3213 		/*
3214 		 * Reset and get 5 strokes more before the boost. Note
3215 		 * that the count is 0 based so we have to add one.
3216 		 */
3217 extra_boost:
3218 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3219 		rack->rc_gp_timely_inc_cnt = 0;
3220 	} else
3221 		plus = (uint32_t)rack_gp_increase_per;
3222 	/* Must be at least 1% increase for true timely increases */
3223 	if ((plus < 1) &&
3224 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3225 		plus = 1;
3226 	if (rack->rc_gp_saw_rec &&
3227 	    (rack->rc_gp_no_rec_chg == 0) &&
3228 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3229 				  rack->r_ctl.rack_per_of_gp_rec)) {
3230 		/* We have been in recovery ding it too */
3231 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3232 		if (calc > 0xffff)
3233 			calc = 0xffff;
3234 		logged |= 1;
3235 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3236 		if (rack_per_upper_bound_ss &&
3237 		    (rack->rc_dragged_bottom == 0) &&
3238 		    (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3239 			rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3240 	}
3241 	if (rack->rc_gp_saw_ca &&
3242 	    (rack->rc_gp_saw_ss == 0) &&
3243 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3244 				  rack->r_ctl.rack_per_of_gp_ca)) {
3245 		/* In CA */
3246 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3247 		if (calc > 0xffff)
3248 			calc = 0xffff;
3249 		logged |= 2;
3250 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3251 		if (rack_per_upper_bound_ca &&
3252 		    (rack->rc_dragged_bottom == 0) &&
3253 		    (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3254 			rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3255 	}
3256 	if (rack->rc_gp_saw_ss &&
3257 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3258 				  rack->r_ctl.rack_per_of_gp_ss)) {
3259 		/* In SS */
3260 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3261 		if (calc > 0xffff)
3262 			calc = 0xffff;
3263 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3264 		if (rack_per_upper_bound_ss &&
3265 		    (rack->rc_dragged_bottom == 0) &&
3266 		    (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3267 			rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3268 		logged |= 4;
3269 	}
3270 	if (logged &&
3271 	    (rack->rc_gp_incr == 0)){
3272 		/* Go into increment mode */
3273 		rack->rc_gp_incr = 1;
3274 		rack->rc_gp_timely_inc_cnt = 0;
3275 	}
3276 	if (rack->rc_gp_incr &&
3277 	    logged &&
3278 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3279 		rack->rc_gp_timely_inc_cnt++;
3280 	}
3281 	rack_log_timely(rack,  logged, plus, 0, 0,
3282 			__LINE__, 1);
3283 }
3284 
3285 static uint32_t
3286 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3287 {
3288 	/*
3289 	 * norm_grad = rtt_diff / minrtt;
3290 	 * new_per = curper * (1 - B * norm_grad)
3291 	 *
3292 	 * B = rack_gp_decrease_per (default 10%)
3293 	 * rtt_dif = input var current rtt-diff
3294 	 * curper = input var current percentage
3295 	 * minrtt = from rack filter
3296 	 *
3297 	 */
3298 	uint64_t perf;
3299 
3300 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3301 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3302 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
3303 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3304 		     (uint64_t)1000000)) /
3305 		(uint64_t)1000000);
3306 	if (perf > curper) {
3307 		/* TSNH */
3308 		perf = curper - 1;
3309 	}
3310 	return ((uint32_t)perf);
3311 }
3312 
3313 static uint32_t
3314 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3315 {
3316 	/*
3317 	 *                                   highrttthresh
3318 	 * result = curper * (1 - (B * ( 1 -  ------          ))
3319 	 *                                     gp_srtt
3320 	 *
3321 	 * B = rack_gp_decrease_per (default 10%)
3322 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3323 	 */
3324 	uint64_t perf;
3325 	uint32_t highrttthresh;
3326 
3327 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3328 
3329 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3330 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3331 					((uint64_t)highrttthresh * (uint64_t)1000000) /
3332 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3333 	return (perf);
3334 }
3335 
3336 static void
3337 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3338 {
3339 	uint64_t logvar, logvar2, logvar3;
3340 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3341 
3342 	if (rack->rc_gp_incr) {
3343 		/* Turn off increment counting */
3344 		rack->rc_gp_incr = 0;
3345 		rack->rc_gp_timely_inc_cnt = 0;
3346 	}
3347 	ss_red = ca_red = rec_red = 0;
3348 	logged = 0;
3349 	/* Calculate the reduction value */
3350 	if (rtt_diff < 0) {
3351 		rtt_diff *= -1;
3352 	}
3353 	/* Must be at least 1% reduction */
3354 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3355 		/* We have been in recovery ding it too */
3356 		if (timely_says == 2) {
3357 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3358 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3359 			if (alt < new_per)
3360 				val = alt;
3361 			else
3362 				val = new_per;
3363 		} else
3364 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3365 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
3366 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3367 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3368 		} else {
3369 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3370 			rec_red = 0;
3371 		}
3372 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3373 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3374 		logged |= 1;
3375 	}
3376 	if (rack->rc_gp_saw_ss) {
3377 		/* Sent in SS */
3378 		if (timely_says == 2) {
3379 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3380 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3381 			if (alt < new_per)
3382 				val = alt;
3383 			else
3384 				val = new_per;
3385 		} else
3386 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3387 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3388 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3389 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3390 		} else {
3391 			ss_red = new_per;
3392 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3393 			logvar = new_per;
3394 			logvar <<= 32;
3395 			logvar |= alt;
3396 			logvar2 = (uint32_t)rtt;
3397 			logvar2 <<= 32;
3398 			logvar2 |= (uint32_t)rtt_diff;
3399 			logvar3 = rack_gp_rtt_maxmul;
3400 			logvar3 <<= 32;
3401 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3402 			rack_log_timely(rack, timely_says,
3403 					logvar2, logvar3,
3404 					logvar, __LINE__, 10);
3405 		}
3406 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3407 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3408 		logged |= 4;
3409 	} else if (rack->rc_gp_saw_ca) {
3410 		/* Sent in CA */
3411 		if (timely_says == 2) {
3412 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3413 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3414 			if (alt < new_per)
3415 				val = alt;
3416 			else
3417 				val = new_per;
3418 		} else
3419 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3420 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
3421 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3422 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3423 		} else {
3424 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3425 			ca_red = 0;
3426 			logvar = new_per;
3427 			logvar <<= 32;
3428 			logvar |= alt;
3429 			logvar2 = (uint32_t)rtt;
3430 			logvar2 <<= 32;
3431 			logvar2 |= (uint32_t)rtt_diff;
3432 			logvar3 = rack_gp_rtt_maxmul;
3433 			logvar3 <<= 32;
3434 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3435 			rack_log_timely(rack, timely_says,
3436 					logvar2, logvar3,
3437 					logvar, __LINE__, 10);
3438 		}
3439 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3440 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3441 		logged |= 2;
3442 	}
3443 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
3444 		rack->rc_gp_timely_dec_cnt++;
3445 		if (rack_timely_dec_clear &&
3446 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3447 			rack->rc_gp_timely_dec_cnt = 0;
3448 	}
3449 	logvar = ss_red;
3450 	logvar <<= 32;
3451 	logvar |= ca_red;
3452 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3453 			__LINE__, 2);
3454 }
3455 
3456 static void
3457 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3458 		     uint32_t rtt, uint32_t line, uint8_t reas)
3459 {
3460 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3461 		union tcp_log_stackspecific log;
3462 		struct timeval tv;
3463 
3464 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3465 		log.u_bbr.flex1 = line;
3466 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3467 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3468 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3469 		log.u_bbr.flex5 = rtt;
3470 		log.u_bbr.flex6 = rack->rc_highly_buffered;
3471 		log.u_bbr.flex6 <<= 1;
3472 		log.u_bbr.flex6 |= rack->forced_ack;
3473 		log.u_bbr.flex6 <<= 1;
3474 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3475 		log.u_bbr.flex6 <<= 1;
3476 		log.u_bbr.flex6 |= rack->in_probe_rtt;
3477 		log.u_bbr.flex6 <<= 1;
3478 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3479 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3480 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3481 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3482 		log.u_bbr.flex8 = reas;
3483 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3484 		log.u_bbr.delRate = rack_get_bw(rack);
3485 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3486 		log.u_bbr.cur_del_rate <<= 32;
3487 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3488 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3489 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3490 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3491 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3492 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3493 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3494 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3495 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3496 		log.u_bbr.rttProp = us_cts;
3497 		log.u_bbr.rttProp <<= 32;
3498 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3499 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3500 		    &rack->rc_inp->inp_socket->so_rcv,
3501 		    &rack->rc_inp->inp_socket->so_snd,
3502 		    BBR_LOG_RTT_SHRINKS, 0,
3503 		    0, &log, false, &rack->r_ctl.act_rcv_time);
3504 	}
3505 }
3506 
3507 static void
3508 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3509 {
3510 	uint64_t bwdp;
3511 
3512 	bwdp = rack_get_bw(rack);
3513 	bwdp *= (uint64_t)rtt;
3514 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3515 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3516 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3517 		/*
3518 		 * A window protocol must be able to have 4 packets
3519 		 * outstanding as the floor in order to function
3520 		 * (especially considering delayed ack :D).
3521 		 */
3522 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3523 	}
3524 }
3525 
3526 static void
3527 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3528 {
3529 	/**
3530 	 * ProbeRTT is a bit different in rack_pacing than in
3531 	 * BBR. It is like BBR in that it uses the lowering of
3532 	 * the RTT as a signal that we saw something new and
3533 	 * counts from there for how long between. But it is
3534 	 * different in that its quite simple. It does not
3535 	 * play with the cwnd and wait until we get down
3536 	 * to N segments outstanding and hold that for
3537 	 * 200ms. Instead it just sets the pacing reduction
3538 	 * rate to a set percentage (70 by default) and hold
3539 	 * that for a number of recent GP Srtt's.
3540 	 */
3541 	uint32_t segsiz;
3542 
3543 	if (rack->rc_gp_dyn_mul == 0)
3544 		return;
3545 
3546 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3547 		/* We are idle */
3548 		return;
3549 	}
3550 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3551 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3552 		/*
3553 		 * Stop the goodput now, the idea here is
3554 		 * that future measurements with in_probe_rtt
3555 		 * won't register if they are not greater so
3556 		 * we want to get what info (if any) is available
3557 		 * now.
3558 		 */
3559 		rack_do_goodput_measurement(rack->rc_tp, rack,
3560 					    rack->rc_tp->snd_una, __LINE__,
3561 					    RACK_QUALITY_PROBERTT);
3562 	}
3563 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3564 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3565 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3566 		     rack->r_ctl.rc_pace_min_segs);
3567 	rack->in_probe_rtt = 1;
3568 	rack->measure_saw_probe_rtt = 1;
3569 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3570 	rack->r_ctl.rc_time_probertt_starts = 0;
3571 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3572 	if (rack_probertt_use_min_rtt_entry)
3573 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3574 	else
3575 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3576 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3577 			     __LINE__, RACK_RTTS_ENTERPROBE);
3578 }
3579 
3580 static void
3581 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3582 {
3583 	struct rack_sendmap *rsm;
3584 	uint32_t segsiz;
3585 
3586 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3587 		     rack->r_ctl.rc_pace_min_segs);
3588 	rack->in_probe_rtt = 0;
3589 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3590 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3591 		/*
3592 		 * Stop the goodput now, the idea here is
3593 		 * that future measurements with in_probe_rtt
3594 		 * won't register if they are not greater so
3595 		 * we want to get what info (if any) is available
3596 		 * now.
3597 		 */
3598 		rack_do_goodput_measurement(rack->rc_tp, rack,
3599 					    rack->rc_tp->snd_una, __LINE__,
3600 					    RACK_QUALITY_PROBERTT);
3601 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3602 		/*
3603 		 * We don't have enough data to make a measurement.
3604 		 * So lets just stop and start here after exiting
3605 		 * probe-rtt. We probably are not interested in
3606 		 * the results anyway.
3607 		 */
3608 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3609 	}
3610 	/*
3611 	 * Measurements through the current snd_max are going
3612 	 * to be limited by the slower pacing rate.
3613 	 *
3614 	 * We need to mark these as app-limited so we
3615 	 * don't collapse the b/w.
3616 	 */
3617 	rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3618 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3619 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3620 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3621 		else {
3622 			/*
3623 			 * Go out to the end app limited and mark
3624 			 * this new one as next and move the end_appl up
3625 			 * to this guy.
3626 			 */
3627 			if (rack->r_ctl.rc_end_appl)
3628 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3629 			rack->r_ctl.rc_end_appl = rsm;
3630 		}
3631 		rsm->r_flags |= RACK_APP_LIMITED;
3632 		rack->r_ctl.rc_app_limited_cnt++;
3633 	}
3634 	/*
3635 	 * Now, we need to examine our pacing rate multipliers.
3636 	 * If its under 100%, we need to kick it back up to
3637 	 * 100%. We also don't let it be over our "max" above
3638 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3639 	 * Note setting clamp_atexit_prtt to 0 has the effect
3640 	 * of setting CA/SS to 100% always at exit (which is
3641 	 * the default behavior).
3642 	 */
3643 	if (rack_probertt_clear_is) {
3644 		rack->rc_gp_incr = 0;
3645 		rack->rc_gp_bwred = 0;
3646 		rack->rc_gp_timely_inc_cnt = 0;
3647 		rack->rc_gp_timely_dec_cnt = 0;
3648 	}
3649 	/* Do we do any clamping at exit? */
3650 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3651 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3652 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3653 	}
3654 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3655 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3656 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3657 	}
3658 	/*
3659 	 * Lets set rtt_diff to 0, so that we will get a "boost"
3660 	 * after exiting.
3661 	 */
3662 	rack->r_ctl.rc_rtt_diff = 0;
3663 
3664 	/* Clear all flags so we start fresh */
3665 	rack->rc_tp->t_bytes_acked = 0;
3666 	rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
3667 	/*
3668 	 * If configured to, set the cwnd and ssthresh to
3669 	 * our targets.
3670 	 */
3671 	if (rack_probe_rtt_sets_cwnd) {
3672 		uint64_t ebdp;
3673 		uint32_t setto;
3674 
3675 		/* Set ssthresh so we get into CA once we hit our target */
3676 		if (rack_probertt_use_min_rtt_exit == 1) {
3677 			/* Set to min rtt */
3678 			rack_set_prtt_target(rack, segsiz,
3679 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3680 		} else if (rack_probertt_use_min_rtt_exit == 2) {
3681 			/* Set to current gp rtt */
3682 			rack_set_prtt_target(rack, segsiz,
3683 					     rack->r_ctl.rc_gp_srtt);
3684 		} else if (rack_probertt_use_min_rtt_exit == 3) {
3685 			/* Set to entry gp rtt */
3686 			rack_set_prtt_target(rack, segsiz,
3687 					     rack->r_ctl.rc_entry_gp_rtt);
3688 		} else {
3689 			uint64_t sum;
3690 			uint32_t setval;
3691 
3692 			sum = rack->r_ctl.rc_entry_gp_rtt;
3693 			sum *= 10;
3694 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3695 			if (sum >= 20) {
3696 				/*
3697 				 * A highly buffered path needs
3698 				 * cwnd space for timely to work.
3699 				 * Lets set things up as if
3700 				 * we are heading back here again.
3701 				 */
3702 				setval = rack->r_ctl.rc_entry_gp_rtt;
3703 			} else if (sum >= 15) {
3704 				/*
3705 				 * Lets take the smaller of the
3706 				 * two since we are just somewhat
3707 				 * buffered.
3708 				 */
3709 				setval = rack->r_ctl.rc_gp_srtt;
3710 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
3711 					setval = rack->r_ctl.rc_entry_gp_rtt;
3712 			} else {
3713 				/*
3714 				 * Here we are not highly buffered
3715 				 * and should pick the min we can to
3716 				 * keep from causing loss.
3717 				 */
3718 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3719 			}
3720 			rack_set_prtt_target(rack, segsiz,
3721 					     setval);
3722 		}
3723 		if (rack_probe_rtt_sets_cwnd > 1) {
3724 			/* There is a percentage here to boost */
3725 			ebdp = rack->r_ctl.rc_target_probertt_flight;
3726 			ebdp *= rack_probe_rtt_sets_cwnd;
3727 			ebdp /= 100;
3728 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3729 		} else
3730 			setto = rack->r_ctl.rc_target_probertt_flight;
3731 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3732 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3733 			/* Enforce a min */
3734 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3735 		}
3736 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
3737 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3738 	}
3739 	rack_log_rtt_shrinks(rack,  us_cts,
3740 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3741 			     __LINE__, RACK_RTTS_EXITPROBE);
3742 	/* Clear times last so log has all the info */
3743 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3744 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3745 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3746 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
3747 }
3748 
3749 static void
3750 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3751 {
3752 	/* Check in on probe-rtt */
3753 	if (rack->rc_gp_filled == 0) {
3754 		/* We do not do p-rtt unless we have gp measurements */
3755 		return;
3756 	}
3757 	if (rack->in_probe_rtt) {
3758 		uint64_t no_overflow;
3759 		uint32_t endtime, must_stay;
3760 
3761 		if (rack->r_ctl.rc_went_idle_time &&
3762 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3763 			/*
3764 			 * We went idle during prtt, just exit now.
3765 			 */
3766 			rack_exit_probertt(rack, us_cts);
3767 		} else if (rack_probe_rtt_safety_val &&
3768 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3769 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3770 			/*
3771 			 * Probe RTT safety value triggered!
3772 			 */
3773 			rack_log_rtt_shrinks(rack,  us_cts,
3774 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3775 					     __LINE__, RACK_RTTS_SAFETY);
3776 			rack_exit_probertt(rack, us_cts);
3777 		}
3778 		/* Calculate the max we will wait */
3779 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3780 		if (rack->rc_highly_buffered)
3781 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3782 		/* Calculate the min we must wait */
3783 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3784 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3785 		    TSTMP_LT(us_cts, endtime)) {
3786 			uint32_t calc;
3787 			/* Do we lower more? */
3788 no_exit:
3789 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3790 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3791 			else
3792 				calc = 0;
3793 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3794 			if (calc) {
3795 				/* Maybe */
3796 				calc *= rack_per_of_gp_probertt_reduce;
3797 				rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3798 				/* Limit it too */
3799 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3800 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3801 			}
3802 			/* We must reach target or the time set */
3803 			return;
3804 		}
3805 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
3806 			if ((TSTMP_LT(us_cts, must_stay) &&
3807 			     rack->rc_highly_buffered) ||
3808 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3809 			      rack->r_ctl.rc_target_probertt_flight)) {
3810 				/* We are not past the must_stay time */
3811 				goto no_exit;
3812 			}
3813 			rack_log_rtt_shrinks(rack,  us_cts,
3814 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3815 					     __LINE__, RACK_RTTS_REACHTARGET);
3816 			rack->r_ctl.rc_time_probertt_starts = us_cts;
3817 			if (rack->r_ctl.rc_time_probertt_starts == 0)
3818 				rack->r_ctl.rc_time_probertt_starts = 1;
3819 			/* Restore back to our rate we want to pace at in prtt */
3820 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3821 		}
3822 		/*
3823 		 * Setup our end time, some number of gp_srtts plus 200ms.
3824 		 */
3825 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3826 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3827 		if (rack_probertt_gpsrtt_cnt_div)
3828 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3829 		else
3830 			endtime = 0;
3831 		endtime += rack_min_probertt_hold;
3832 		endtime += rack->r_ctl.rc_time_probertt_starts;
3833 		if (TSTMP_GEQ(us_cts,  endtime)) {
3834 			/* yes, exit probertt */
3835 			rack_exit_probertt(rack, us_cts);
3836 		}
3837 
3838 	} else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3839 		/* Go into probertt, its been too long since we went lower */
3840 		rack_enter_probertt(rack, us_cts);
3841 	}
3842 }
3843 
3844 static void
3845 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3846 		       uint32_t rtt, int32_t rtt_diff)
3847 {
3848 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
3849 	uint32_t losses;
3850 
3851 	if ((rack->rc_gp_dyn_mul == 0) ||
3852 	    (rack->use_fixed_rate) ||
3853 	    (rack->in_probe_rtt) ||
3854 	    (rack->rc_always_pace == 0)) {
3855 		/* No dynamic GP multiplier in play */
3856 		return;
3857 	}
3858 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3859 	cur_bw = rack_get_bw(rack);
3860 	/* Calculate our up and down range */
3861 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3862 	up_bnd /= 100;
3863 	up_bnd += rack->r_ctl.last_gp_comp_bw;
3864 
3865 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3866 	subfr /= 100;
3867 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3868 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3869 		/*
3870 		 * This is the case where our RTT is above
3871 		 * the max target and we have been configured
3872 		 * to just do timely no bonus up stuff in that case.
3873 		 *
3874 		 * There are two configurations, set to 1, and we
3875 		 * just do timely if we are over our max. If its
3876 		 * set above 1 then we slam the multipliers down
3877 		 * to 100 and then decrement per timely.
3878 		 */
3879 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3880 				__LINE__, 3);
3881 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3882 			rack_validate_multipliers_at_or_below_100(rack);
3883 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3884 	} else if ((last_bw_est < low_bnd) && !losses) {
3885 		/*
3886 		 * We are decreasing this is a bit complicated this
3887 		 * means we are loosing ground. This could be
3888 		 * because another flow entered and we are competing
3889 		 * for b/w with it. This will push the RTT up which
3890 		 * makes timely unusable unless we want to get shoved
3891 		 * into a corner and just be backed off (the age
3892 		 * old problem with delay based CC).
3893 		 *
3894 		 * On the other hand if it was a route change we
3895 		 * would like to stay somewhat contained and not
3896 		 * blow out the buffers.
3897 		 */
3898 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3899 				__LINE__, 3);
3900 		rack->r_ctl.last_gp_comp_bw = cur_bw;
3901 		if (rack->rc_gp_bwred == 0) {
3902 			/* Go into reduction counting */
3903 			rack->rc_gp_bwred = 1;
3904 			rack->rc_gp_timely_dec_cnt = 0;
3905 		}
3906 		if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
3907 		    (timely_says == 0)) {
3908 			/*
3909 			 * Push another time with a faster pacing
3910 			 * to try to gain back (we include override to
3911 			 * get a full raise factor).
3912 			 */
3913 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
3914 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
3915 			    (timely_says == 0) ||
3916 			    (rack_down_raise_thresh == 0)) {
3917 				/*
3918 				 * Do an override up in b/w if we were
3919 				 * below the threshold or if the threshold
3920 				 * is zero we always do the raise.
3921 				 */
3922 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
3923 			} else {
3924 				/* Log it stays the same */
3925 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
3926 						__LINE__, 11);
3927 			}
3928 			rack->rc_gp_timely_dec_cnt++;
3929 			/* We are not incrementing really no-count */
3930 			rack->rc_gp_incr = 0;
3931 			rack->rc_gp_timely_inc_cnt = 0;
3932 		} else {
3933 			/*
3934 			 * Lets just use the RTT
3935 			 * information and give up
3936 			 * pushing.
3937 			 */
3938 			goto use_timely;
3939 		}
3940 	} else if ((timely_says != 2) &&
3941 		    !losses &&
3942 		    (last_bw_est > up_bnd)) {
3943 		/*
3944 		 * We are increasing b/w lets keep going, updating
3945 		 * our b/w and ignoring any timely input, unless
3946 		 * of course we are at our max raise (if there is one).
3947 		 */
3948 
3949 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3950 				__LINE__, 3);
3951 		rack->r_ctl.last_gp_comp_bw = cur_bw;
3952 		if (rack->rc_gp_saw_ss &&
3953 		    rack_per_upper_bound_ss &&
3954 		     (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
3955 			    /*
3956 			     * In cases where we can't go higher
3957 			     * we should just use timely.
3958 			     */
3959 			    goto use_timely;
3960 		}
3961 		if (rack->rc_gp_saw_ca &&
3962 		    rack_per_upper_bound_ca &&
3963 		    (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
3964 			    /*
3965 			     * In cases where we can't go higher
3966 			     * we should just use timely.
3967 			     */
3968 			    goto use_timely;
3969 		}
3970 		rack->rc_gp_bwred = 0;
3971 		rack->rc_gp_timely_dec_cnt = 0;
3972 		/* You get a set number of pushes if timely is trying to reduce */
3973 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
3974 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3975 		} else {
3976 			/* Log it stays the same */
3977 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
3978 			    __LINE__, 12);
3979 		}
3980 		return;
3981 	} else {
3982 		/*
3983 		 * We are staying between the lower and upper range bounds
3984 		 * so use timely to decide.
3985 		 */
3986 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3987 				__LINE__, 3);
3988 use_timely:
3989 		if (timely_says) {
3990 			rack->rc_gp_incr = 0;
3991 			rack->rc_gp_timely_inc_cnt = 0;
3992 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
3993 			    !losses &&
3994 			    (last_bw_est < low_bnd)) {
3995 				/* We are loosing ground */
3996 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
3997 				rack->rc_gp_timely_dec_cnt++;
3998 				/* We are not incrementing really no-count */
3999 				rack->rc_gp_incr = 0;
4000 				rack->rc_gp_timely_inc_cnt = 0;
4001 			} else
4002 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4003 		} else {
4004 			rack->rc_gp_bwred = 0;
4005 			rack->rc_gp_timely_dec_cnt = 0;
4006 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4007 		}
4008 	}
4009 }
4010 
4011 static int32_t
4012 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4013 {
4014 	int32_t timely_says;
4015 	uint64_t log_mult, log_rtt_a_diff;
4016 
4017 	log_rtt_a_diff = rtt;
4018 	log_rtt_a_diff <<= 32;
4019 	log_rtt_a_diff |= (uint32_t)rtt_diff;
4020 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4021 		    rack_gp_rtt_maxmul)) {
4022 		/* Reduce the b/w multiplier */
4023 		timely_says = 2;
4024 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4025 		log_mult <<= 32;
4026 		log_mult |= prev_rtt;
4027 		rack_log_timely(rack,  timely_says, log_mult,
4028 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4029 				log_rtt_a_diff, __LINE__, 4);
4030 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4031 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4032 			    max(rack_gp_rtt_mindiv , 1)))) {
4033 		/* Increase the b/w multiplier */
4034 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4035 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4036 			 max(rack_gp_rtt_mindiv , 1));
4037 		log_mult <<= 32;
4038 		log_mult |= prev_rtt;
4039 		timely_says = 0;
4040 		rack_log_timely(rack,  timely_says, log_mult ,
4041 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4042 				log_rtt_a_diff, __LINE__, 5);
4043 	} else {
4044 		/*
4045 		 * Use a gradient to find it the timely gradient
4046 		 * is:
4047 		 * grad = rc_rtt_diff / min_rtt;
4048 		 *
4049 		 * anything below or equal to 0 will be
4050 		 * a increase indication. Anything above
4051 		 * zero is a decrease. Note we take care
4052 		 * of the actual gradient calculation
4053 		 * in the reduction (its not needed for
4054 		 * increase).
4055 		 */
4056 		log_mult = prev_rtt;
4057 		if (rtt_diff <= 0) {
4058 			/*
4059 			 * Rttdiff is less than zero, increase the
4060 			 * b/w multiplier (its 0 or negative)
4061 			 */
4062 			timely_says = 0;
4063 			rack_log_timely(rack,  timely_says, log_mult,
4064 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4065 		} else {
4066 			/* Reduce the b/w multiplier */
4067 			timely_says = 1;
4068 			rack_log_timely(rack,  timely_says, log_mult,
4069 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4070 		}
4071 	}
4072 	return (timely_says);
4073 }
4074 
4075 static void
4076 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4077 			    tcp_seq th_ack, int line, uint8_t quality)
4078 {
4079 	uint64_t tim, bytes_ps, ltim, stim, utim;
4080 	uint32_t segsiz, bytes, reqbytes, us_cts;
4081 	int32_t gput, new_rtt_diff, timely_says;
4082 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4083 	int did_add = 0;
4084 
4085 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4086 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4087 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
4088 		tim = us_cts - tp->gput_ts;
4089 	else
4090 		tim = 0;
4091 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4092 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4093 	else
4094 		stim = 0;
4095 	/*
4096 	 * Use the larger of the send time or ack time. This prevents us
4097 	 * from being influenced by ack artifacts to come up with too
4098 	 * high of measurement. Note that since we are spanning over many more
4099 	 * bytes in most of our measurements hopefully that is less likely to
4100 	 * occur.
4101 	 */
4102 	if (tim > stim)
4103 		utim = max(tim, 1);
4104 	else
4105 		utim = max(stim, 1);
4106 	/* Lets get a msec time ltim too for the old stuff */
4107 	ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4108 	gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4109 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4110 	if ((tim == 0) && (stim == 0)) {
4111 		/*
4112 		 * Invalid measurement time, maybe
4113 		 * all on one ack/one send?
4114 		 */
4115 		bytes = 0;
4116 		bytes_ps = 0;
4117 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4118 					   0, 0, 0, 10, __LINE__, NULL, quality);
4119 		goto skip_measurement;
4120 	}
4121 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4122 		/* We never made a us_rtt measurement? */
4123 		bytes = 0;
4124 		bytes_ps = 0;
4125 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4126 					   0, 0, 0, 10, __LINE__, NULL, quality);
4127 		goto skip_measurement;
4128 	}
4129 	/*
4130 	 * Calculate the maximum possible b/w this connection
4131 	 * could have. We base our calculation on the lowest
4132 	 * rtt we have seen during the measurement and the
4133 	 * largest rwnd the client has given us in that time. This
4134 	 * forms a BDP that is the maximum that we could ever
4135 	 * get to the client. Anything larger is not valid.
4136 	 *
4137 	 * I originally had code here that rejected measurements
4138 	 * where the time was less than 1/2 the latest us_rtt.
4139 	 * But after thinking on that I realized its wrong since
4140 	 * say you had a 150Mbps or even 1Gbps link, and you
4141 	 * were a long way away.. example I am in Europe (100ms rtt)
4142 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4143 	 * bytes my time would be 1.2ms, and yet my rtt would say
4144 	 * the measurement was invalid the time was < 50ms. The
4145 	 * same thing is true for 150Mb (8ms of time).
4146 	 *
4147 	 * A better way I realized is to look at what the maximum
4148 	 * the connection could possibly do. This is gated on
4149 	 * the lowest RTT we have seen and the highest rwnd.
4150 	 * We should in theory never exceed that, if we are
4151 	 * then something on the path is storing up packets
4152 	 * and then feeding them all at once to our endpoint
4153 	 * messing up our measurement.
4154 	 */
4155 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4156 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4157 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4158 	if (SEQ_LT(th_ack, tp->gput_seq)) {
4159 		/* No measurement can be made */
4160 		bytes = 0;
4161 		bytes_ps = 0;
4162 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4163 					   0, 0, 0, 10, __LINE__, NULL, quality);
4164 		goto skip_measurement;
4165 	} else
4166 		bytes = (th_ack - tp->gput_seq);
4167 	bytes_ps = (uint64_t)bytes;
4168 	/*
4169 	 * Don't measure a b/w for pacing unless we have gotten at least
4170 	 * an initial windows worth of data in this measurement interval.
4171 	 *
4172 	 * Small numbers of bytes get badly influenced by delayed ack and
4173 	 * other artifacts. Note we take the initial window or our
4174 	 * defined minimum GP (defaulting to 10 which hopefully is the
4175 	 * IW).
4176 	 */
4177 	if (rack->rc_gp_filled == 0) {
4178 		/*
4179 		 * The initial estimate is special. We
4180 		 * have blasted out an IW worth of packets
4181 		 * without a real valid ack ts results. We
4182 		 * then setup the app_limited_needs_set flag,
4183 		 * this should get the first ack in (probably 2
4184 		 * MSS worth) to be recorded as the timestamp.
4185 		 * We thus allow a smaller number of bytes i.e.
4186 		 * IW - 2MSS.
4187 		 */
4188 		reqbytes -= (2 * segsiz);
4189 		/* Also lets fill previous for our first measurement to be neutral */
4190 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4191 	}
4192 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4193 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4194 					   rack->r_ctl.rc_app_limited_cnt,
4195 					   0, 0, 10, __LINE__, NULL, quality);
4196 		goto skip_measurement;
4197 	}
4198 	/*
4199 	 * We now need to calculate the Timely like status so
4200 	 * we can update (possibly) the b/w multipliers.
4201 	 */
4202 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4203 	if (rack->rc_gp_filled == 0) {
4204 		/* No previous reading */
4205 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4206 	} else {
4207 		if (rack->measure_saw_probe_rtt == 0) {
4208 			/*
4209 			 * We don't want a probertt to be counted
4210 			 * since it will be negative incorrectly. We
4211 			 * expect to be reducing the RTT when we
4212 			 * pace at a slower rate.
4213 			 */
4214 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4215 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4216 		}
4217 	}
4218 	timely_says = rack_make_timely_judgement(rack,
4219 		rack->r_ctl.rc_gp_srtt,
4220 		rack->r_ctl.rc_rtt_diff,
4221 	        rack->r_ctl.rc_prev_gp_srtt
4222 		);
4223 	bytes_ps *= HPTS_USEC_IN_SEC;
4224 	bytes_ps /= utim;
4225 	if (bytes_ps > rack->r_ctl.last_max_bw) {
4226 		/*
4227 		 * Something is on path playing
4228 		 * since this b/w is not possible based
4229 		 * on our BDP (highest rwnd and lowest rtt
4230 		 * we saw in the measurement window).
4231 		 *
4232 		 * Another option here would be to
4233 		 * instead skip the measurement.
4234 		 */
4235 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4236 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
4237 					   11, __LINE__, NULL, quality);
4238 		bytes_ps = rack->r_ctl.last_max_bw;
4239 	}
4240 	/* We store gp for b/w in bytes per second */
4241 	if (rack->rc_gp_filled == 0) {
4242 		/* Initial measurement */
4243 		if (bytes_ps) {
4244 			rack->r_ctl.gp_bw = bytes_ps;
4245 			rack->rc_gp_filled = 1;
4246 			rack->r_ctl.num_measurements = 1;
4247 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4248 		} else {
4249 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4250 						   rack->r_ctl.rc_app_limited_cnt,
4251 						   0, 0, 10, __LINE__, NULL, quality);
4252 		}
4253 		if (tcp_in_hpts(rack->rc_inp) &&
4254 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4255 			/*
4256 			 * Ok we can't trust the pacer in this case
4257 			 * where we transition from un-paced to paced.
4258 			 * Or for that matter when the burst mitigation
4259 			 * was making a wild guess and got it wrong.
4260 			 * Stop the pacer and clear up all the aggregate
4261 			 * delays etc.
4262 			 */
4263 			tcp_hpts_remove(rack->rc_inp);
4264 			rack->r_ctl.rc_hpts_flags = 0;
4265 			rack->r_ctl.rc_last_output_to = 0;
4266 		}
4267 		did_add = 2;
4268 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4269 		/* Still a small number run an average */
4270 		rack->r_ctl.gp_bw += bytes_ps;
4271 		addpart = rack->r_ctl.num_measurements;
4272 		rack->r_ctl.num_measurements++;
4273 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4274 			/* We have collected enough to move forward */
4275 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4276 		}
4277 		did_add = 3;
4278 	} else {
4279 		/*
4280 		 * We want to take 1/wma of the goodput and add in to 7/8th
4281 		 * of the old value weighted by the srtt. So if your measurement
4282 		 * period is say 2 SRTT's long you would get 1/4 as the
4283 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
4284 		 *
4285 		 * But we must be careful not to take too much i.e. if the
4286 		 * srtt is say 20ms and the measurement is taken over
4287 		 * 400ms our weight would be 400/20 i.e. 20. On the
4288 		 * other hand if we get a measurement over 1ms with a
4289 		 * 10ms rtt we only want to take a much smaller portion.
4290 		 */
4291 		if (rack->r_ctl.num_measurements < 0xff) {
4292 			rack->r_ctl.num_measurements++;
4293 		}
4294 		srtt = (uint64_t)tp->t_srtt;
4295 		if (srtt == 0) {
4296 			/*
4297 			 * Strange why did t_srtt go back to zero?
4298 			 */
4299 			if (rack->r_ctl.rc_rack_min_rtt)
4300 				srtt = rack->r_ctl.rc_rack_min_rtt;
4301 			else
4302 				srtt = HPTS_USEC_IN_MSEC;
4303 		}
4304 		/*
4305 		 * XXXrrs: Note for reviewers, in playing with
4306 		 * dynamic pacing I discovered this GP calculation
4307 		 * as done originally leads to some undesired results.
4308 		 * Basically you can get longer measurements contributing
4309 		 * too much to the WMA. Thus I changed it if you are doing
4310 		 * dynamic adjustments to only do the aportioned adjustment
4311 		 * if we have a very small (time wise) measurement. Longer
4312 		 * measurements just get there weight (defaulting to 1/8)
4313 		 * add to the WMA. We may want to think about changing
4314 		 * this to always do that for both sides i.e. dynamic
4315 		 * and non-dynamic... but considering lots of folks
4316 		 * were playing with this I did not want to change the
4317 		 * calculation per.se. without your thoughts.. Lawerence?
4318 		 * Peter??
4319 		 */
4320 		if (rack->rc_gp_dyn_mul == 0) {
4321 			subpart = rack->r_ctl.gp_bw * utim;
4322 			subpart /= (srtt * 8);
4323 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
4324 				/*
4325 				 * The b/w update takes no more
4326 				 * away then 1/2 our running total
4327 				 * so factor it in.
4328 				 */
4329 				addpart = bytes_ps * utim;
4330 				addpart /= (srtt * 8);
4331 			} else {
4332 				/*
4333 				 * Don't allow a single measurement
4334 				 * to account for more than 1/2 of the
4335 				 * WMA. This could happen on a retransmission
4336 				 * where utim becomes huge compared to
4337 				 * srtt (multiple retransmissions when using
4338 				 * the sending rate which factors in all the
4339 				 * transmissions from the first one).
4340 				 */
4341 				subpart = rack->r_ctl.gp_bw / 2;
4342 				addpart = bytes_ps / 2;
4343 			}
4344 			resid_bw = rack->r_ctl.gp_bw - subpart;
4345 			rack->r_ctl.gp_bw = resid_bw + addpart;
4346 			did_add = 1;
4347 		} else {
4348 			if ((utim / srtt) <= 1) {
4349 				/*
4350 				 * The b/w update was over a small period
4351 				 * of time. The idea here is to prevent a small
4352 				 * measurement time period from counting
4353 				 * too much. So we scale it based on the
4354 				 * time so it attributes less than 1/rack_wma_divisor
4355 				 * of its measurement.
4356 				 */
4357 				subpart = rack->r_ctl.gp_bw * utim;
4358 				subpart /= (srtt * rack_wma_divisor);
4359 				addpart = bytes_ps * utim;
4360 				addpart /= (srtt * rack_wma_divisor);
4361 			} else {
4362 				/*
4363 				 * The scaled measurement was long
4364 				 * enough so lets just add in the
4365 				 * portion of the measurement i.e. 1/rack_wma_divisor
4366 				 */
4367 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4368 				addpart = bytes_ps / rack_wma_divisor;
4369 			}
4370 			if ((rack->measure_saw_probe_rtt == 0) ||
4371 		            (bytes_ps > rack->r_ctl.gp_bw)) {
4372 				/*
4373 				 * For probe-rtt we only add it in
4374 				 * if its larger, all others we just
4375 				 * add in.
4376 				 */
4377 				did_add = 1;
4378 				resid_bw = rack->r_ctl.gp_bw - subpart;
4379 				rack->r_ctl.gp_bw = resid_bw + addpart;
4380 			}
4381 		}
4382 	}
4383 	if ((rack->gp_ready == 0) &&
4384 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4385 		/* We have enough measurements now */
4386 		rack->gp_ready = 1;
4387 		rack_set_cc_pacing(rack);
4388 		if (rack->defer_options)
4389 			rack_apply_deferred_options(rack);
4390 	}
4391 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4392 				   rack_get_bw(rack), 22, did_add, NULL, quality);
4393 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
4394 	if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4395 		rack_update_multiplier(rack, timely_says, bytes_ps,
4396 				       rack->r_ctl.rc_gp_srtt,
4397 				       rack->r_ctl.rc_rtt_diff);
4398 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4399 				   rack_get_bw(rack), 3, line, NULL, quality);
4400 	/* reset the gp srtt and setup the new prev */
4401 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4402 	/* Record the lost count for the next measurement */
4403 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4404 	/*
4405 	 * We restart our diffs based on the gpsrtt in the
4406 	 * measurement window.
4407 	 */
4408 	rack->rc_gp_rtt_set = 0;
4409 	rack->rc_gp_saw_rec = 0;
4410 	rack->rc_gp_saw_ca = 0;
4411 	rack->rc_gp_saw_ss = 0;
4412 	rack->rc_dragged_bottom = 0;
4413 skip_measurement:
4414 
4415 #ifdef STATS
4416 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4417 				 gput);
4418 	/*
4419 	 * XXXLAS: This is a temporary hack, and should be
4420 	 * chained off VOI_TCP_GPUT when stats(9) grows an
4421 	 * API to deal with chained VOIs.
4422 	 */
4423 	if (tp->t_stats_gput_prev > 0)
4424 		stats_voi_update_abs_s32(tp->t_stats,
4425 					 VOI_TCP_GPUT_ND,
4426 					 ((gput - tp->t_stats_gput_prev) * 100) /
4427 					 tp->t_stats_gput_prev);
4428 #endif
4429 	tp->t_flags &= ~TF_GPUTINPROG;
4430 	tp->t_stats_gput_prev = gput;
4431 	/*
4432 	 * Now are we app limited now and there is space from where we
4433 	 * were to where we want to go?
4434 	 *
4435 	 * We don't do the other case i.e. non-applimited here since
4436 	 * the next send will trigger us picking up the missing data.
4437 	 */
4438 	if (rack->r_ctl.rc_first_appl &&
4439 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
4440 	    rack->r_ctl.rc_app_limited_cnt &&
4441 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4442 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
4443 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4444 		/*
4445 		 * Yep there is enough outstanding to make a measurement here.
4446 		 */
4447 		struct rack_sendmap *rsm, fe;
4448 
4449 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4450 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4451 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4452 		rack->app_limited_needs_set = 0;
4453 		tp->gput_seq = th_ack;
4454 		if (rack->in_probe_rtt)
4455 			rack->measure_saw_probe_rtt = 1;
4456 		else if ((rack->measure_saw_probe_rtt) &&
4457 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4458 			rack->measure_saw_probe_rtt = 0;
4459 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
4460 			/* There is a full window to gain info from */
4461 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4462 		} else {
4463 			/* We can only measure up to the applimited point */
4464 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
4465 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
4466 				/*
4467 				 * We don't have enough to make a measurement.
4468 				 */
4469 				tp->t_flags &= ~TF_GPUTINPROG;
4470 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
4471 							   0, 0, 0, 6, __LINE__, NULL, quality);
4472 				return;
4473 			}
4474 		}
4475 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
4476 			/*
4477 			 * We will get no more data into the SB
4478 			 * this means we need to have the data available
4479 			 * before we start a measurement.
4480 			 */
4481 			if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
4482 				/* Nope not enough data. */
4483 				return;
4484 			}
4485 		}
4486 		tp->t_flags |= TF_GPUTINPROG;
4487 		/*
4488 		 * Now we need to find the timestamp of the send at tp->gput_seq
4489 		 * for the send based measurement.
4490 		 */
4491 		fe.r_start = tp->gput_seq;
4492 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4493 		if (rsm) {
4494 			/* Ok send-based limit is set */
4495 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4496 				/*
4497 				 * Move back to include the earlier part
4498 				 * so our ack time lines up right (this may
4499 				 * make an overlapping measurement but thats
4500 				 * ok).
4501 				 */
4502 				tp->gput_seq = rsm->r_start;
4503 			}
4504 			if (rsm->r_flags & RACK_ACKED)
4505 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4506 			else
4507 				rack->app_limited_needs_set = 1;
4508 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4509 		} else {
4510 			/*
4511 			 * If we don't find the rsm due to some
4512 			 * send-limit set the current time, which
4513 			 * basically disables the send-limit.
4514 			 */
4515 			struct timeval tv;
4516 
4517 			microuptime(&tv);
4518 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4519 		}
4520 		rack_log_pacing_delay_calc(rack,
4521 					   tp->gput_seq,
4522 					   tp->gput_ack,
4523 					   (uint64_t)rsm,
4524 					   tp->gput_ts,
4525 					   rack->r_ctl.rc_app_limited_cnt,
4526 					   9,
4527 					   __LINE__, NULL, quality);
4528 	}
4529 }
4530 
4531 /*
4532  * CC wrapper hook functions
4533  */
4534 static void
4535 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4536     uint16_t type, int32_t recovery)
4537 {
4538 	uint32_t prior_cwnd, acked;
4539 	struct tcp_log_buffer *lgb = NULL;
4540 	uint8_t labc_to_use, quality;
4541 
4542 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4543 	tp->t_ccv.nsegs = nsegs;
4544 	acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
4545 	if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4546 		uint32_t max;
4547 
4548 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4549 		if (tp->t_ccv.bytes_this_ack > max) {
4550 			tp->t_ccv.bytes_this_ack = max;
4551 		}
4552 	}
4553 #ifdef STATS
4554 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4555 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4556 #endif
4557 	quality = RACK_QUALITY_NONE;
4558 	if ((tp->t_flags & TF_GPUTINPROG) &&
4559 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
4560 		/* Measure the Goodput */
4561 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
4562 #ifdef NETFLIX_PEAKRATE
4563 		if ((type == CC_ACK) &&
4564 		    (tp->t_maxpeakrate)) {
4565 			/*
4566 			 * We update t_peakrate_thr. This gives us roughly
4567 			 * one update per round trip time. Note
4568 			 * it will only be used if pace_always is off i.e
4569 			 * we don't do this for paced flows.
4570 			 */
4571 			rack_update_peakrate_thr(tp);
4572 		}
4573 #endif
4574 	}
4575 	/* Which way our we limited, if not cwnd limited no advance in CA */
4576 	if (tp->snd_cwnd <= tp->snd_wnd)
4577 		tp->t_ccv.flags |= CCF_CWND_LIMITED;
4578 	else
4579 		tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
4580 	if (tp->snd_cwnd > tp->snd_ssthresh) {
4581 		tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
4582 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4583 		/* For the setting of a window past use the actual scwnd we are using */
4584 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4585 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4586 			tp->t_ccv.flags |= CCF_ABC_SENTAWND;
4587 		}
4588 	} else {
4589 		tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4590 		tp->t_bytes_acked = 0;
4591 	}
4592 	prior_cwnd = tp->snd_cwnd;
4593 	if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4594 	    (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4595 		labc_to_use = rack->rc_labc;
4596 	else
4597 		labc_to_use = rack_max_abc_post_recovery;
4598 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4599 		union tcp_log_stackspecific log;
4600 		struct timeval tv;
4601 
4602 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4603 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4604 		log.u_bbr.flex1 = th_ack;
4605 		log.u_bbr.flex2 = tp->t_ccv.flags;
4606 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
4607 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
4608 		log.u_bbr.flex5 = labc_to_use;
4609 		log.u_bbr.flex6 = prior_cwnd;
4610 		log.u_bbr.flex7 = V_tcp_do_newsack;
4611 		log.u_bbr.flex8 = 1;
4612 		lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4613 				     0, &log, false, NULL, NULL, 0, &tv);
4614 	}
4615 	if (CC_ALGO(tp)->ack_received != NULL) {
4616 		/* XXXLAS: Find a way to live without this */
4617 		tp->t_ccv.curack = th_ack;
4618 		tp->t_ccv.labc = labc_to_use;
4619 		tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
4620 		CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
4621 	}
4622 	if (lgb) {
4623 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4624 	}
4625 	if (rack->r_must_retran) {
4626 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4627 			/*
4628 			 * We now are beyond the rxt point so lets disable
4629 			 * the flag.
4630 			 */
4631 			rack->r_ctl.rc_out_at_rto = 0;
4632 			rack->r_must_retran = 0;
4633 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4634 			/*
4635 			 * Only decrement the rc_out_at_rto if the cwnd advances
4636 			 * at least a whole segment. Otherwise next time the peer
4637 			 * acks, we won't be able to send this generaly happens
4638 			 * when we are in Congestion Avoidance.
4639 			 */
4640 			if (acked <= rack->r_ctl.rc_out_at_rto){
4641 				rack->r_ctl.rc_out_at_rto -= acked;
4642 			} else {
4643 				rack->r_ctl.rc_out_at_rto = 0;
4644 			}
4645 		}
4646 	}
4647 #ifdef STATS
4648 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4649 #endif
4650 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4651 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4652 	}
4653 #ifdef NETFLIX_PEAKRATE
4654 	/* we enforce max peak rate if it is set and we are not pacing */
4655 	if ((rack->rc_always_pace == 0) &&
4656 	    tp->t_peakrate_thr &&
4657 	    (tp->snd_cwnd > tp->t_peakrate_thr)) {
4658 		tp->snd_cwnd = tp->t_peakrate_thr;
4659 	}
4660 #endif
4661 }
4662 
4663 static void
4664 tcp_rack_partialack(struct tcpcb *tp)
4665 {
4666 	struct tcp_rack *rack;
4667 
4668 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4669 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4670 	/*
4671 	 * If we are doing PRR and have enough
4672 	 * room to send <or> we are pacing and prr
4673 	 * is disabled we will want to see if we
4674 	 * can send data (by setting r_wanted_output to
4675 	 * true).
4676 	 */
4677 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4678 	    rack->rack_no_prr)
4679 		rack->r_wanted_output = 1;
4680 }
4681 
4682 static void
4683 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4684 {
4685 	struct tcp_rack *rack;
4686 	uint32_t orig_cwnd;
4687 
4688 	orig_cwnd = tp->snd_cwnd;
4689 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4690 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4691 	/* only alert CC if we alerted when we entered */
4692 	if (CC_ALGO(tp)->post_recovery != NULL) {
4693 		tp->t_ccv.curack = th_ack;
4694 		CC_ALGO(tp)->post_recovery(&tp->t_ccv);
4695 		if (tp->snd_cwnd < tp->snd_ssthresh) {
4696 			/*
4697 			 * Rack has burst control and pacing
4698 			 * so lets not set this any lower than
4699 			 * snd_ssthresh per RFC-6582 (option 2).
4700 			 */
4701 			tp->snd_cwnd = tp->snd_ssthresh;
4702 		}
4703 	}
4704 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4705 		union tcp_log_stackspecific log;
4706 		struct timeval tv;
4707 
4708 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4709 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4710 		log.u_bbr.flex1 = th_ack;
4711 		log.u_bbr.flex2 = tp->t_ccv.flags;
4712 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
4713 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
4714 		log.u_bbr.flex5 = V_tcp_abc_l_var;
4715 		log.u_bbr.flex6 = orig_cwnd;
4716 		log.u_bbr.flex7 = V_tcp_do_newsack;
4717 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4718 		log.u_bbr.flex8 = 2;
4719 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4720 			       0, &log, false, NULL, NULL, 0, &tv);
4721 	}
4722 	if ((rack->rack_no_prr == 0) &&
4723 	    (rack->no_prr_addback == 0) &&
4724 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
4725 		/*
4726 		 * Suck the next prr cnt back into cwnd, but
4727 		 * only do that if we are not application limited.
4728 		 */
4729 		if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
4730 			/*
4731 			 * We are allowed to add back to the cwnd the amount we did
4732 			 * not get out if:
4733 			 * a) no_prr_addback is off.
4734 			 * b) we are not app limited
4735 			 * c) we are doing prr
4736 			 * <and>
4737 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4738 			 */
4739 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4740 					    rack->r_ctl.rc_prr_sndcnt);
4741 		}
4742 		rack->r_ctl.rc_prr_sndcnt = 0;
4743 		rack_log_to_prr(rack, 1, 0, __LINE__);
4744 	}
4745 	rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
4746 	tp->snd_recover = tp->snd_una;
4747 	if (rack->r_ctl.dsack_persist) {
4748 		rack->r_ctl.dsack_persist--;
4749 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
4750 			rack->r_ctl.num_dsack = 0;
4751 		}
4752 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
4753 	}
4754 	EXIT_RECOVERY(tp->t_flags);
4755 }
4756 
4757 static void
4758 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
4759 {
4760 	struct tcp_rack *rack;
4761 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4762 
4763 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4764 #ifdef STATS
4765 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4766 #endif
4767 	if (IN_RECOVERY(tp->t_flags) == 0) {
4768 		in_rec_at_entry = 0;
4769 		ssthresh_enter = tp->snd_ssthresh;
4770 		cwnd_enter = tp->snd_cwnd;
4771 	} else
4772 		in_rec_at_entry = 1;
4773 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4774 	switch (type) {
4775 	case CC_NDUPACK:
4776 		tp->t_flags &= ~TF_WASFRECOVERY;
4777 		tp->t_flags &= ~TF_WASCRECOVERY;
4778 		if (!IN_FASTRECOVERY(tp->t_flags)) {
4779 			rack->r_ctl.rc_prr_delivered = 0;
4780 			rack->r_ctl.rc_prr_out = 0;
4781 			if (rack->rack_no_prr == 0) {
4782 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4783 				rack_log_to_prr(rack, 2, in_rec_at_entry, line);
4784 			}
4785 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4786 			tp->snd_recover = tp->snd_max;
4787 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4788 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4789 		}
4790 		break;
4791 	case CC_ECN:
4792 		if (!IN_CONGRECOVERY(tp->t_flags) ||
4793 		    /*
4794 		     * Allow ECN reaction on ACK to CWR, if
4795 		     * that data segment was also CE marked.
4796 		     */
4797 		    SEQ_GEQ(ack, tp->snd_recover)) {
4798 			EXIT_CONGRECOVERY(tp->t_flags);
4799 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4800 			tp->snd_recover = tp->snd_max + 1;
4801 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4802 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4803 		}
4804 		break;
4805 	case CC_RTO:
4806 		tp->t_dupacks = 0;
4807 		tp->t_bytes_acked = 0;
4808 		EXIT_RECOVERY(tp->t_flags);
4809 		tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4810 		    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4811 		orig_cwnd = tp->snd_cwnd;
4812 		tp->snd_cwnd = ctf_fixed_maxseg(tp);
4813 		rack_log_to_prr(rack, 16, orig_cwnd, line);
4814 		if (tp->t_flags2 & TF2_ECN_PERMIT)
4815 			tp->t_flags2 |= TF2_ECN_SND_CWR;
4816 		break;
4817 	case CC_RTO_ERR:
4818 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4819 		/* RTO was unnecessary, so reset everything. */
4820 		tp->snd_cwnd = tp->snd_cwnd_prev;
4821 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
4822 		tp->snd_recover = tp->snd_recover_prev;
4823 		if (tp->t_flags & TF_WASFRECOVERY) {
4824 			ENTER_FASTRECOVERY(tp->t_flags);
4825 			tp->t_flags &= ~TF_WASFRECOVERY;
4826 		}
4827 		if (tp->t_flags & TF_WASCRECOVERY) {
4828 			ENTER_CONGRECOVERY(tp->t_flags);
4829 			tp->t_flags &= ~TF_WASCRECOVERY;
4830 		}
4831 		tp->snd_nxt = tp->snd_max;
4832 		tp->t_badrxtwin = 0;
4833 		break;
4834 	}
4835 	if ((CC_ALGO(tp)->cong_signal != NULL)  &&
4836 	    (type != CC_RTO)){
4837 		tp->t_ccv.curack = ack;
4838 		CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
4839 	}
4840 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4841 		rack_log_to_prr(rack, 15, cwnd_enter, line);
4842 		rack->r_ctl.dsack_byte_cnt = 0;
4843 		rack->r_ctl.retran_during_recovery = 0;
4844 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4845 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4846 		rack->r_ent_rec_ns = 1;
4847 	}
4848 }
4849 
4850 static inline void
4851 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4852 {
4853 	uint32_t i_cwnd;
4854 
4855 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4856 
4857 #ifdef NETFLIX_STATS
4858 	KMOD_TCPSTAT_INC(tcps_idle_restarts);
4859 	if (tp->t_state == TCPS_ESTABLISHED)
4860 		KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4861 #endif
4862 	if (CC_ALGO(tp)->after_idle != NULL)
4863 		CC_ALGO(tp)->after_idle(&tp->t_ccv);
4864 
4865 	if (tp->snd_cwnd == 1)
4866 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
4867 	else
4868 		i_cwnd = rc_init_window(rack);
4869 
4870 	/*
4871 	 * Being idle is no different than the initial window. If the cc
4872 	 * clamps it down below the initial window raise it to the initial
4873 	 * window.
4874 	 */
4875 	if (tp->snd_cwnd < i_cwnd) {
4876 		tp->snd_cwnd = i_cwnd;
4877 	}
4878 }
4879 
4880 /*
4881  * Indicate whether this ack should be delayed.  We can delay the ack if
4882  * following conditions are met:
4883  *	- There is no delayed ack timer in progress.
4884  *	- Our last ack wasn't a 0-sized window. We never want to delay
4885  *	  the ack that opens up a 0-sized window.
4886  *	- LRO wasn't used for this segment. We make sure by checking that the
4887  *	  segment size is not larger than the MSS.
4888  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
4889  *	  connection.
4890  */
4891 #define DELAY_ACK(tp, tlen)			 \
4892 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4893 	((tp->t_flags & TF_DELACK) == 0) &&	 \
4894 	(tlen <= tp->t_maxseg) &&		 \
4895 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4896 
4897 static struct rack_sendmap *
4898 rack_find_lowest_rsm(struct tcp_rack *rack)
4899 {
4900 	struct rack_sendmap *rsm;
4901 
4902 	/*
4903 	 * Walk the time-order transmitted list looking for an rsm that is
4904 	 * not acked. This will be the one that was sent the longest time
4905 	 * ago that is still outstanding.
4906 	 */
4907 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4908 		if (rsm->r_flags & RACK_ACKED) {
4909 			continue;
4910 		}
4911 		goto finish;
4912 	}
4913 finish:
4914 	return (rsm);
4915 }
4916 
4917 static struct rack_sendmap *
4918 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
4919 {
4920 	struct rack_sendmap *prsm;
4921 
4922 	/*
4923 	 * Walk the sequence order list backward until we hit and arrive at
4924 	 * the highest seq not acked. In theory when this is called it
4925 	 * should be the last segment (which it was not).
4926 	 */
4927 	prsm = rsm;
4928 	RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
4929 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
4930 			continue;
4931 		}
4932 		return (prsm);
4933 	}
4934 	return (NULL);
4935 }
4936 
4937 static uint32_t
4938 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
4939 {
4940 	int32_t lro;
4941 	uint32_t thresh;
4942 
4943 	/*
4944 	 * lro is the flag we use to determine if we have seen reordering.
4945 	 * If it gets set we have seen reordering. The reorder logic either
4946 	 * works in one of two ways:
4947 	 *
4948 	 * If reorder-fade is configured, then we track the last time we saw
4949 	 * re-ordering occur. If we reach the point where enough time as
4950 	 * passed we no longer consider reordering has occuring.
4951 	 *
4952 	 * Or if reorder-face is 0, then once we see reordering we consider
4953 	 * the connection to alway be subject to reordering and just set lro
4954 	 * to 1.
4955 	 *
4956 	 * In the end if lro is non-zero we add the extra time for
4957 	 * reordering in.
4958 	 */
4959 	if (srtt == 0)
4960 		srtt = 1;
4961 	if (rack->r_ctl.rc_reorder_ts) {
4962 		if (rack->r_ctl.rc_reorder_fade) {
4963 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
4964 				lro = cts - rack->r_ctl.rc_reorder_ts;
4965 				if (lro == 0) {
4966 					/*
4967 					 * No time as passed since the last
4968 					 * reorder, mark it as reordering.
4969 					 */
4970 					lro = 1;
4971 				}
4972 			} else {
4973 				/* Negative time? */
4974 				lro = 0;
4975 			}
4976 			if (lro > rack->r_ctl.rc_reorder_fade) {
4977 				/* Turn off reordering seen too */
4978 				rack->r_ctl.rc_reorder_ts = 0;
4979 				lro = 0;
4980 			}
4981 		} else {
4982 			/* Reodering does not fade */
4983 			lro = 1;
4984 		}
4985 	} else {
4986 		lro = 0;
4987 	}
4988 	if (rack->rc_rack_tmr_std_based == 0) {
4989 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
4990 	} else {
4991 		/* Standards based pkt-delay is 1/4 srtt */
4992 		thresh = srtt +  (srtt >> 2);
4993 	}
4994 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
4995 		/* It must be set, if not you get 1/4 rtt */
4996 		if (rack->r_ctl.rc_reorder_shift)
4997 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
4998 		else
4999 			thresh += (srtt >> 2);
5000 	}
5001 	if (rack->rc_rack_use_dsack &&
5002 	    lro &&
5003 	    (rack->r_ctl.num_dsack > 0)) {
5004 		/*
5005 		 * We only increase the reordering window if we
5006 		 * have seen reordering <and> we have a DSACK count.
5007 		 */
5008 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
5009 		rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
5010 	}
5011 	/* SRTT * 2 is the ceiling */
5012 	if (thresh > (srtt * 2)) {
5013 		thresh = srtt * 2;
5014 	}
5015 	/* And we don't want it above the RTO max either */
5016 	if (thresh > rack_rto_max) {
5017 		thresh = rack_rto_max;
5018 	}
5019 	rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
5020 	return (thresh);
5021 }
5022 
5023 static uint32_t
5024 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5025 		     struct rack_sendmap *rsm, uint32_t srtt)
5026 {
5027 	struct rack_sendmap *prsm;
5028 	uint32_t thresh, len;
5029 	int segsiz;
5030 
5031 	if (srtt == 0)
5032 		srtt = 1;
5033 	if (rack->r_ctl.rc_tlp_threshold)
5034 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5035 	else
5036 		thresh = (srtt * 2);
5037 
5038 	/* Get the previous sent packet, if any */
5039 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5040 	len = rsm->r_end - rsm->r_start;
5041 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5042 		/* Exactly like the ID */
5043 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5044 			uint32_t alt_thresh;
5045 			/*
5046 			 * Compensate for delayed-ack with the d-ack time.
5047 			 */
5048 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5049 			if (alt_thresh > thresh)
5050 				thresh = alt_thresh;
5051 		}
5052 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5053 		/* 2.1 behavior */
5054 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5055 		if (prsm && (len <= segsiz)) {
5056 			/*
5057 			 * Two packets outstanding, thresh should be (2*srtt) +
5058 			 * possible inter-packet delay (if any).
5059 			 */
5060 			uint32_t inter_gap = 0;
5061 			int idx, nidx;
5062 
5063 			idx = rsm->r_rtr_cnt - 1;
5064 			nidx = prsm->r_rtr_cnt - 1;
5065 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5066 				/* Yes it was sent later (or at the same time) */
5067 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5068 			}
5069 			thresh += inter_gap;
5070 		} else if (len <= segsiz) {
5071 			/*
5072 			 * Possibly compensate for delayed-ack.
5073 			 */
5074 			uint32_t alt_thresh;
5075 
5076 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5077 			if (alt_thresh > thresh)
5078 				thresh = alt_thresh;
5079 		}
5080 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5081 		/* 2.2 behavior */
5082 		if (len <= segsiz) {
5083 			uint32_t alt_thresh;
5084 			/*
5085 			 * Compensate for delayed-ack with the d-ack time.
5086 			 */
5087 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5088 			if (alt_thresh > thresh)
5089 				thresh = alt_thresh;
5090 		}
5091 	}
5092 	/* Not above an RTO */
5093 	if (thresh > tp->t_rxtcur) {
5094 		thresh = tp->t_rxtcur;
5095 	}
5096 	/* Not above a RTO max */
5097 	if (thresh > rack_rto_max) {
5098 		thresh = rack_rto_max;
5099 	}
5100 	/* Apply user supplied min TLP */
5101 	if (thresh < rack_tlp_min) {
5102 		thresh = rack_tlp_min;
5103 	}
5104 	return (thresh);
5105 }
5106 
5107 static uint32_t
5108 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5109 {
5110 	/*
5111 	 * We want the rack_rtt which is the
5112 	 * last rtt we measured. However if that
5113 	 * does not exist we fallback to the srtt (which
5114 	 * we probably will never do) and then as a last
5115 	 * resort we use RACK_INITIAL_RTO if no srtt is
5116 	 * yet set.
5117 	 */
5118 	if (rack->rc_rack_rtt)
5119 		return (rack->rc_rack_rtt);
5120 	else if (tp->t_srtt == 0)
5121 		return (RACK_INITIAL_RTO);
5122 	return (tp->t_srtt);
5123 }
5124 
5125 static struct rack_sendmap *
5126 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5127 {
5128 	/*
5129 	 * Check to see that we don't need to fall into recovery. We will
5130 	 * need to do so if our oldest transmit is past the time we should
5131 	 * have had an ack.
5132 	 */
5133 	struct tcp_rack *rack;
5134 	struct rack_sendmap *rsm;
5135 	int32_t idx;
5136 	uint32_t srtt, thresh;
5137 
5138 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5139 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5140 		return (NULL);
5141 	}
5142 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5143 	if (rsm == NULL)
5144 		return (NULL);
5145 
5146 
5147 	if (rsm->r_flags & RACK_ACKED) {
5148 		rsm = rack_find_lowest_rsm(rack);
5149 		if (rsm == NULL)
5150 			return (NULL);
5151 	}
5152 	idx = rsm->r_rtr_cnt - 1;
5153 	srtt = rack_grab_rtt(tp, rack);
5154 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5155 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5156 		return (NULL);
5157 	}
5158 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5159 		return (NULL);
5160 	}
5161 	/* Ok if we reach here we are over-due and this guy can be sent */
5162 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
5163 	return (rsm);
5164 }
5165 
5166 static uint32_t
5167 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5168 {
5169 	int32_t t;
5170 	int32_t tt;
5171 	uint32_t ret_val;
5172 
5173 	t = (tp->t_srtt + (tp->t_rttvar << 2));
5174 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5175  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5176 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5177 	ret_val = (uint32_t)tt;
5178 	return (ret_val);
5179 }
5180 
5181 static uint32_t
5182 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5183 {
5184 	/*
5185 	 * Start the FR timer, we do this based on getting the first one in
5186 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
5187 	 * events we need to stop the running timer (if its running) before
5188 	 * starting the new one.
5189 	 */
5190 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5191 	uint32_t srtt_cur;
5192 	int32_t idx;
5193 	int32_t is_tlp_timer = 0;
5194 	struct rack_sendmap *rsm;
5195 
5196 	if (rack->t_timers_stopped) {
5197 		/* All timers have been stopped none are to run */
5198 		return (0);
5199 	}
5200 	if (rack->rc_in_persist) {
5201 		/* We can't start any timer in persists */
5202 		return (rack_get_persists_timer_val(tp, rack));
5203 	}
5204 	rack->rc_on_min_to = 0;
5205 	if ((tp->t_state < TCPS_ESTABLISHED) ||
5206 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5207 		goto activate_rxt;
5208 	}
5209 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5210 	if ((rsm == NULL) || sup_rack) {
5211 		/* Nothing on the send map or no rack */
5212 activate_rxt:
5213 		time_since_sent = 0;
5214 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5215 		if (rsm) {
5216 			/*
5217 			 * Should we discount the RTX timer any?
5218 			 *
5219 			 * We want to discount it the smallest amount.
5220 			 * If a timer (Rack/TLP or RXT) has gone off more
5221 			 * recently thats the discount we want to use (now - timer time).
5222 			 * If the retransmit of the oldest packet was more recent then
5223 			 * we want to use that (now - oldest-packet-last_transmit_time).
5224 			 *
5225 			 */
5226 			idx = rsm->r_rtr_cnt - 1;
5227 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5228 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5229 			else
5230 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5231 			if (TSTMP_GT(cts, tstmp_touse))
5232 			    time_since_sent = cts - tstmp_touse;
5233 		}
5234 		if (SEQ_LT(tp->snd_una, tp->snd_max) ||
5235 		    sbavail(&tptosocket(tp)->so_snd)) {
5236 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5237 			to = tp->t_rxtcur;
5238 			if (to > time_since_sent)
5239 				to -= time_since_sent;
5240 			else
5241 				to = rack->r_ctl.rc_min_to;
5242 			if (to == 0)
5243 				to = 1;
5244 			/* Special case for KEEPINIT */
5245 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5246 			    (TP_KEEPINIT(tp) != 0) &&
5247 			    rsm) {
5248 				/*
5249 				 * We have to put a ceiling on the rxt timer
5250 				 * of the keep-init timeout.
5251 				 */
5252 				uint32_t max_time, red;
5253 
5254 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5255 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5256 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5257 					if (red < max_time)
5258 						max_time -= red;
5259 					else
5260 						max_time = 1;
5261 				}
5262 				/* Reduce timeout to the keep value if needed */
5263 				if (max_time < to)
5264 					to = max_time;
5265 			}
5266 			return (to);
5267 		}
5268 		return (0);
5269 	}
5270 	if (rsm->r_flags & RACK_ACKED) {
5271 		rsm = rack_find_lowest_rsm(rack);
5272 		if (rsm == NULL) {
5273 			/* No lowest? */
5274 			goto activate_rxt;
5275 		}
5276 	}
5277 	if (rack->sack_attack_disable) {
5278 		/*
5279 		 * We don't want to do
5280 		 * any TLP's if you are an attacker.
5281 		 * Though if you are doing what
5282 		 * is expected you may still have
5283 		 * SACK-PASSED marks.
5284 		 */
5285 		goto activate_rxt;
5286 	}
5287 	/* Convert from ms to usecs */
5288 	if ((rsm->r_flags & RACK_SACK_PASSED) ||
5289 	    (rsm->r_flags & RACK_RWND_COLLAPSED) ||
5290 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5291 		if ((tp->t_flags & TF_SENTFIN) &&
5292 		    ((tp->snd_max - tp->snd_una) == 1) &&
5293 		    (rsm->r_flags & RACK_HAS_FIN)) {
5294 			/*
5295 			 * We don't start a rack timer if all we have is a
5296 			 * FIN outstanding.
5297 			 */
5298 			goto activate_rxt;
5299 		}
5300 		if ((rack->use_rack_rr == 0) &&
5301 		    (IN_FASTRECOVERY(tp->t_flags)) &&
5302 		    (rack->rack_no_prr == 0) &&
5303 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5304 			/*
5305 			 * We are not cheating, in recovery  and
5306 			 * not enough ack's to yet get our next
5307 			 * retransmission out.
5308 			 *
5309 			 * Note that classified attackers do not
5310 			 * get to use the rack-cheat.
5311 			 */
5312 			goto activate_tlp;
5313 		}
5314 		srtt = rack_grab_rtt(tp, rack);
5315 		thresh = rack_calc_thresh_rack(rack, srtt, cts);
5316 		idx = rsm->r_rtr_cnt - 1;
5317 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5318 		if (SEQ_GEQ(exp, cts)) {
5319 			to = exp - cts;
5320 			if (to < rack->r_ctl.rc_min_to) {
5321 				to = rack->r_ctl.rc_min_to;
5322 				if (rack->r_rr_config == 3)
5323 					rack->rc_on_min_to = 1;
5324 			}
5325 		} else {
5326 			to = rack->r_ctl.rc_min_to;
5327 			if (rack->r_rr_config == 3)
5328 				rack->rc_on_min_to = 1;
5329 		}
5330 	} else {
5331 		/* Ok we need to do a TLP not RACK */
5332 activate_tlp:
5333 		if ((rack->rc_tlp_in_progress != 0) &&
5334 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5335 			/*
5336 			 * The previous send was a TLP and we have sent
5337 			 * N TLP's without sending new data.
5338 			 */
5339 			goto activate_rxt;
5340 		}
5341 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5342 		if (rsm == NULL) {
5343 			/* We found no rsm to TLP with. */
5344 			goto activate_rxt;
5345 		}
5346 		if (rsm->r_flags & RACK_HAS_FIN) {
5347 			/* If its a FIN we dont do TLP */
5348 			rsm = NULL;
5349 			goto activate_rxt;
5350 		}
5351 		idx = rsm->r_rtr_cnt - 1;
5352 		time_since_sent = 0;
5353 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5354 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5355 		else
5356 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5357 		if (TSTMP_GT(cts, tstmp_touse))
5358 		    time_since_sent = cts - tstmp_touse;
5359 		is_tlp_timer = 1;
5360 		if (tp->t_srtt) {
5361 			if ((rack->rc_srtt_measure_made == 0) &&
5362 			    (tp->t_srtt == 1)) {
5363 				/*
5364 				 * If another stack as run and set srtt to 1,
5365 				 * then the srtt was 0, so lets use the initial.
5366 				 */
5367 				srtt = RACK_INITIAL_RTO;
5368 			} else {
5369 				srtt_cur = tp->t_srtt;
5370 				srtt = srtt_cur;
5371 			}
5372 		} else
5373 			srtt = RACK_INITIAL_RTO;
5374 		/*
5375 		 * If the SRTT is not keeping up and the
5376 		 * rack RTT has spiked we want to use
5377 		 * the last RTT not the smoothed one.
5378 		 */
5379 		if (rack_tlp_use_greater &&
5380 		    tp->t_srtt &&
5381 		    (srtt < rack_grab_rtt(tp, rack))) {
5382 			srtt = rack_grab_rtt(tp, rack);
5383 		}
5384 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5385 		if (thresh > time_since_sent) {
5386 			to = thresh - time_since_sent;
5387 		} else {
5388 			to = rack->r_ctl.rc_min_to;
5389 			rack_log_alt_to_to_cancel(rack,
5390 						  thresh,		/* flex1 */
5391 						  time_since_sent,	/* flex2 */
5392 						  tstmp_touse,		/* flex3 */
5393 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5394 						  (uint32_t)rsm->r_tim_lastsent[idx],
5395 						  srtt,
5396 						  idx, 99);
5397 		}
5398 		if (to < rack_tlp_min) {
5399 			to = rack_tlp_min;
5400 		}
5401 		if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5402 			/*
5403 			 * If the TLP time works out to larger than the max
5404 			 * RTO lets not do TLP.. just RTO.
5405 			 */
5406 			goto activate_rxt;
5407 		}
5408 	}
5409 	if (is_tlp_timer == 0) {
5410 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5411 	} else {
5412 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5413 	}
5414 	if (to == 0)
5415 		to = 1;
5416 	return (to);
5417 }
5418 
5419 static void
5420 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5421 {
5422 	if (rack->rc_in_persist == 0) {
5423 		if (tp->t_flags & TF_GPUTINPROG) {
5424 			/*
5425 			 * Stop the goodput now, the calling of the
5426 			 * measurement function clears the flag.
5427 			 */
5428 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
5429 						    RACK_QUALITY_PERSIST);
5430 		}
5431 #ifdef NETFLIX_SHARED_CWND
5432 		if (rack->r_ctl.rc_scw) {
5433 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5434 			rack->rack_scwnd_is_idle = 1;
5435 		}
5436 #endif
5437 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5438 		if (rack->r_ctl.rc_went_idle_time == 0)
5439 			rack->r_ctl.rc_went_idle_time = 1;
5440 		rack_timer_cancel(tp, rack, cts, __LINE__);
5441 		rack->r_ctl.persist_lost_ends = 0;
5442 		rack->probe_not_answered = 0;
5443 		rack->forced_ack = 0;
5444 		tp->t_rxtshift = 0;
5445 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5446 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5447 		rack->rc_in_persist = 1;
5448 	}
5449 }
5450 
5451 static void
5452 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5453 {
5454 	if (tcp_in_hpts(rack->rc_inp)) {
5455 		tcp_hpts_remove(rack->rc_inp);
5456 		rack->r_ctl.rc_hpts_flags = 0;
5457 	}
5458 #ifdef NETFLIX_SHARED_CWND
5459 	if (rack->r_ctl.rc_scw) {
5460 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5461 		rack->rack_scwnd_is_idle = 0;
5462 	}
5463 #endif
5464 	if (rack->rc_gp_dyn_mul &&
5465 	    (rack->use_fixed_rate == 0) &&
5466 	    (rack->rc_always_pace)) {
5467 		/*
5468 		 * Do we count this as if a probe-rtt just
5469 		 * finished?
5470 		 */
5471 		uint32_t time_idle, idle_min;
5472 
5473 		time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5474 		idle_min = rack_min_probertt_hold;
5475 		if (rack_probertt_gpsrtt_cnt_div) {
5476 			uint64_t extra;
5477 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5478 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
5479 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5480 			idle_min += (uint32_t)extra;
5481 		}
5482 		if (time_idle >= idle_min) {
5483 			/* Yes, we count it as a probe-rtt. */
5484 			uint32_t us_cts;
5485 
5486 			us_cts = tcp_get_usecs(NULL);
5487 			if (rack->in_probe_rtt == 0) {
5488 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5489 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5490 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5491 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5492 			} else {
5493 				rack_exit_probertt(rack, us_cts);
5494 			}
5495 		}
5496 	}
5497 	rack->rc_in_persist = 0;
5498 	rack->r_ctl.rc_went_idle_time = 0;
5499 	tp->t_rxtshift = 0;
5500 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5501 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5502 	rack->r_ctl.rc_agg_delayed = 0;
5503 	rack->r_early = 0;
5504 	rack->r_late = 0;
5505 	rack->r_ctl.rc_agg_early = 0;
5506 }
5507 
5508 static void
5509 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5510 		   struct hpts_diag *diag, struct timeval *tv)
5511 {
5512 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5513 		union tcp_log_stackspecific log;
5514 
5515 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5516 		log.u_bbr.flex1 = diag->p_nxt_slot;
5517 		log.u_bbr.flex2 = diag->p_cur_slot;
5518 		log.u_bbr.flex3 = diag->slot_req;
5519 		log.u_bbr.flex4 = diag->inp_hptsslot;
5520 		log.u_bbr.flex5 = diag->slot_remaining;
5521 		log.u_bbr.flex6 = diag->need_new_to;
5522 		log.u_bbr.flex7 = diag->p_hpts_active;
5523 		log.u_bbr.flex8 = diag->p_on_min_sleep;
5524 		/* Hijack other fields as needed */
5525 		log.u_bbr.epoch = diag->have_slept;
5526 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
5527 		log.u_bbr.pkts_out = diag->co_ret;
5528 		log.u_bbr.applimited = diag->hpts_sleep_time;
5529 		log.u_bbr.delivered = diag->p_prev_slot;
5530 		log.u_bbr.inflight = diag->p_runningslot;
5531 		log.u_bbr.bw_inuse = diag->wheel_slot;
5532 		log.u_bbr.rttProp = diag->wheel_cts;
5533 		log.u_bbr.timeStamp = cts;
5534 		log.u_bbr.delRate = diag->maxslots;
5535 		log.u_bbr.cur_del_rate = diag->p_curtick;
5536 		log.u_bbr.cur_del_rate <<= 32;
5537 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
5538 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5539 		    &rack->rc_inp->inp_socket->so_rcv,
5540 		    &rack->rc_inp->inp_socket->so_snd,
5541 		    BBR_LOG_HPTSDIAG, 0,
5542 		    0, &log, false, tv);
5543 	}
5544 
5545 }
5546 
5547 static void
5548 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5549 {
5550 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5551 		union tcp_log_stackspecific log;
5552 		struct timeval tv;
5553 
5554 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5555 		log.u_bbr.flex1 = sb->sb_flags;
5556 		log.u_bbr.flex2 = len;
5557 		log.u_bbr.flex3 = sb->sb_state;
5558 		log.u_bbr.flex8 = type;
5559 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5560 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5561 		    &rack->rc_inp->inp_socket->so_rcv,
5562 		    &rack->rc_inp->inp_socket->so_snd,
5563 		    TCP_LOG_SB_WAKE, 0,
5564 		    len, &log, false, &tv);
5565 	}
5566 }
5567 
5568 static void
5569 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5570       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5571 {
5572 	struct hpts_diag diag;
5573 	struct inpcb *inp = tptoinpcb(tp);
5574 	struct timeval tv;
5575 	uint32_t delayed_ack = 0;
5576 	uint32_t hpts_timeout;
5577 	uint32_t entry_slot = slot;
5578 	uint8_t stopped;
5579 	uint32_t left = 0;
5580 	uint32_t us_cts;
5581 
5582 	if ((tp->t_state == TCPS_CLOSED) ||
5583 	    (tp->t_state == TCPS_LISTEN)) {
5584 		return;
5585 	}
5586 	if (tcp_in_hpts(inp)) {
5587 		/* Already on the pacer */
5588 		return;
5589 	}
5590 	stopped = rack->rc_tmr_stopped;
5591 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5592 		left = rack->r_ctl.rc_timer_exp - cts;
5593 	}
5594 	rack->r_ctl.rc_timer_exp = 0;
5595 	rack->r_ctl.rc_hpts_flags = 0;
5596 	us_cts = tcp_get_usecs(&tv);
5597 	/* Now early/late accounting */
5598 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
5599 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5600 		/*
5601 		 * We have a early carry over set,
5602 		 * we can always add more time so we
5603 		 * can always make this compensation.
5604 		 *
5605 		 * Note if ack's are allowed to wake us do not
5606 		 * penalize the next timer for being awoke
5607 		 * by an ack aka the rc_agg_early (non-paced mode).
5608 		 */
5609 		slot += rack->r_ctl.rc_agg_early;
5610 		rack->r_early = 0;
5611 		rack->r_ctl.rc_agg_early = 0;
5612 	}
5613 	if (rack->r_late) {
5614 		/*
5615 		 * This is harder, we can
5616 		 * compensate some but it
5617 		 * really depends on what
5618 		 * the current pacing time is.
5619 		 */
5620 		if (rack->r_ctl.rc_agg_delayed >= slot) {
5621 			/*
5622 			 * We can't compensate for it all.
5623 			 * And we have to have some time
5624 			 * on the clock. We always have a min
5625 			 * 10 slots (10 x 10 i.e. 100 usecs).
5626 			 */
5627 			if (slot <= HPTS_TICKS_PER_SLOT) {
5628 				/* We gain delay */
5629 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
5630 				slot = HPTS_TICKS_PER_SLOT;
5631 			} else {
5632 				/* We take off some */
5633 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
5634 				slot = HPTS_TICKS_PER_SLOT;
5635 			}
5636 		} else {
5637 			slot -= rack->r_ctl.rc_agg_delayed;
5638 			rack->r_ctl.rc_agg_delayed = 0;
5639 			/* Make sure we have 100 useconds at minimum */
5640 			if (slot < HPTS_TICKS_PER_SLOT) {
5641 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
5642 				slot = HPTS_TICKS_PER_SLOT;
5643 			}
5644 			if (rack->r_ctl.rc_agg_delayed == 0)
5645 				rack->r_late = 0;
5646 		}
5647 	}
5648 	if (slot) {
5649 		/* We are pacing too */
5650 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5651 	}
5652 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5653 #ifdef NETFLIX_EXP_DETECTION
5654 	if (rack->sack_attack_disable &&
5655 	    (slot < tcp_sad_pacing_interval)) {
5656 		/*
5657 		 * We have a potential attacker on
5658 		 * the line. We have possibly some
5659 		 * (or now) pacing time set. We want to
5660 		 * slow down the processing of sacks by some
5661 		 * amount (if it is an attacker). Set the default
5662 		 * slot for attackers in place (unless the orginal
5663 		 * interval is longer). Its stored in
5664 		 * micro-seconds, so lets convert to msecs.
5665 		 */
5666 		slot = tcp_sad_pacing_interval;
5667 	}
5668 #endif
5669 	if (tp->t_flags & TF_DELACK) {
5670 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
5671 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5672 	}
5673 	if (delayed_ack && ((hpts_timeout == 0) ||
5674 			    (delayed_ack < hpts_timeout)))
5675 		hpts_timeout = delayed_ack;
5676 	else
5677 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5678 	/*
5679 	 * If no timers are going to run and we will fall off the hptsi
5680 	 * wheel, we resort to a keep-alive timer if its configured.
5681 	 */
5682 	if ((hpts_timeout == 0) &&
5683 	    (slot == 0)) {
5684 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5685 		    (tp->t_state <= TCPS_CLOSING)) {
5686 			/*
5687 			 * Ok we have no timer (persists, rack, tlp, rxt  or
5688 			 * del-ack), we don't have segments being paced. So
5689 			 * all that is left is the keepalive timer.
5690 			 */
5691 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5692 				/* Get the established keep-alive time */
5693 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5694 			} else {
5695 				/*
5696 				 * Get the initial setup keep-alive time,
5697 				 * note that this is probably not going to
5698 				 * happen, since rack will be running a rxt timer
5699 				 * if a SYN of some sort is outstanding. It is
5700 				 * actually handled in rack_timeout_rxt().
5701 				 */
5702 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5703 			}
5704 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5705 			if (rack->in_probe_rtt) {
5706 				/*
5707 				 * We want to instead not wake up a long time from
5708 				 * now but to wake up about the time we would
5709 				 * exit probe-rtt and initiate a keep-alive ack.
5710 				 * This will get us out of probe-rtt and update
5711 				 * our min-rtt.
5712 				 */
5713 				hpts_timeout = rack_min_probertt_hold;
5714 			}
5715 		}
5716 	}
5717 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5718 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5719 		/*
5720 		 * RACK, TLP, persists and RXT timers all are restartable
5721 		 * based on actions input .. i.e we received a packet (ack
5722 		 * or sack) and that changes things (rw, or snd_una etc).
5723 		 * Thus we can restart them with a new value. For
5724 		 * keep-alive, delayed_ack we keep track of what was left
5725 		 * and restart the timer with a smaller value.
5726 		 */
5727 		if (left < hpts_timeout)
5728 			hpts_timeout = left;
5729 	}
5730 	if (hpts_timeout) {
5731 		/*
5732 		 * Hack alert for now we can't time-out over 2,147,483
5733 		 * seconds (a bit more than 596 hours), which is probably ok
5734 		 * :).
5735 		 */
5736 		if (hpts_timeout > 0x7ffffffe)
5737 			hpts_timeout = 0x7ffffffe;
5738 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5739 	}
5740 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
5741 	if ((rack->gp_ready == 0) &&
5742 	    (rack->use_fixed_rate == 0) &&
5743 	    (hpts_timeout < slot) &&
5744 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5745 		/*
5746 		 * We have no good estimate yet for the
5747 		 * old clunky burst mitigation or the
5748 		 * real pacing. And the tlp or rxt is smaller
5749 		 * than the pacing calculation. Lets not
5750 		 * pace that long since we know the calculation
5751 		 * so far is not accurate.
5752 		 */
5753 		slot = hpts_timeout;
5754 	}
5755 	/**
5756 	 * Turn off all the flags for queuing by default. The
5757 	 * flags have important meanings to what happens when
5758 	 * LRO interacts with the transport. Most likely (by default now)
5759 	 * mbuf_queueing and ack compression are on. So the transport
5760 	 * has a couple of flags that control what happens (if those
5761 	 * are not on then these flags won't have any effect since it
5762 	 * won't go through the queuing LRO path).
5763 	 *
5764 	 * INP_MBUF_QUEUE_READY - This flags says that I am busy
5765 	 *                        pacing output, so don't disturb. But
5766 	 *                        it also means LRO can wake me if there
5767 	 *                        is a SACK arrival.
5768 	 *
5769 	 * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5770 	 *                       with the above flag (QUEUE_READY) and
5771 	 *                       when present it says don't even wake me
5772 	 *                       if a SACK arrives.
5773 	 *
5774 	 * The idea behind these flags is that if we are pacing we
5775 	 * set the MBUF_QUEUE_READY and only get woken up if
5776 	 * a SACK arrives (which could change things) or if
5777 	 * our pacing timer expires. If, however, we have a rack
5778 	 * timer running, then we don't even want a sack to wake
5779 	 * us since the rack timer has to expire before we can send.
5780 	 *
5781 	 * Other cases should usually have none of the flags set
5782 	 * so LRO can call into us.
5783 	 */
5784 	inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5785 	if (slot) {
5786 		rack->r_ctl.rc_last_output_to = us_cts + slot;
5787 		/*
5788 		 * A pacing timer (slot) is being set, in
5789 		 * such a case we cannot send (we are blocked by
5790 		 * the timer). So lets tell LRO that it should not
5791 		 * wake us unless there is a SACK. Note this only
5792 		 * will be effective if mbuf queueing is on or
5793 		 * compressed acks are being processed.
5794 		 */
5795 		inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5796 		/*
5797 		 * But wait if we have a Rack timer running
5798 		 * even a SACK should not disturb us (with
5799 		 * the exception of r_rr_config 3).
5800 		 */
5801 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5802 		    (rack->r_rr_config != 3))
5803 			inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5804 		if (rack->rc_ack_can_sendout_data) {
5805 			/*
5806 			 * Ahh but wait, this is that special case
5807 			 * where the pacing timer can be disturbed
5808 			 * backout the changes (used for non-paced
5809 			 * burst limiting).
5810 			 */
5811 			inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5812 		}
5813 		if ((rack->use_rack_rr) &&
5814 		    (rack->r_rr_config < 2) &&
5815 		    ((hpts_timeout) && (hpts_timeout < slot))) {
5816 			/*
5817 			 * Arrange for the hpts to kick back in after the
5818 			 * t-o if the t-o does not cause a send.
5819 			 */
5820 			(void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
5821 						   __LINE__, &diag);
5822 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5823 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5824 		} else {
5825 			(void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(slot),
5826 						   __LINE__, &diag);
5827 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5828 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5829 		}
5830 	} else if (hpts_timeout) {
5831 		/*
5832 		 * With respect to inp_flags2 here, lets let any new acks wake
5833 		 * us up here. Since we are not pacing (no pacing timer), output
5834 		 * can happen so we should let it. If its a Rack timer, then any inbound
5835 		 * packet probably won't change the sending (we will be blocked)
5836 		 * but it may change the prr stats so letting it in (the set defaults
5837 		 * at the start of this block) are good enough.
5838 		 */
5839 		(void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
5840 					   __LINE__, &diag);
5841 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5842 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5843 	} else {
5844 		/* No timer starting */
5845 #ifdef INVARIANTS
5846 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5847 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5848 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5849 		}
5850 #endif
5851 	}
5852 	rack->rc_tmr_stopped = 0;
5853 	if (slot)
5854 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5855 }
5856 
5857 /*
5858  * RACK Timer, here we simply do logging and house keeping.
5859  * the normal rack_output() function will call the
5860  * appropriate thing to check if we need to do a RACK retransmit.
5861  * We return 1, saying don't proceed with rack_output only
5862  * when all timers have been stopped (destroyed PCB?).
5863  */
5864 static int
5865 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5866 {
5867 	/*
5868 	 * This timer simply provides an internal trigger to send out data.
5869 	 * The check_recovery_mode call will see if there are needed
5870 	 * retransmissions, if so we will enter fast-recovery. The output
5871 	 * call may or may not do the same thing depending on sysctl
5872 	 * settings.
5873 	 */
5874 	struct rack_sendmap *rsm;
5875 
5876 	counter_u64_add(rack_to_tot, 1);
5877 	if (rack->r_state && (rack->r_state != tp->t_state))
5878 		rack_set_state(tp, rack);
5879 	rack->rc_on_min_to = 0;
5880 	rsm = rack_check_recovery_mode(tp, cts);
5881 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5882 	if (rsm) {
5883 		rack->r_ctl.rc_resend = rsm;
5884 		rack->r_timer_override = 1;
5885 		if (rack->use_rack_rr) {
5886 			/*
5887 			 * Don't accumulate extra pacing delay
5888 			 * we are allowing the rack timer to
5889 			 * over-ride pacing i.e. rrr takes precedence
5890 			 * if the pacing interval is longer than the rrr
5891 			 * time (in other words we get the min pacing
5892 			 * time versus rrr pacing time).
5893 			 */
5894 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5895 		}
5896 	}
5897 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5898 	if (rsm == NULL) {
5899 		/* restart a timer and return 1 */
5900 		rack_start_hpts_timer(rack, tp, cts,
5901 				      0, 0, 0);
5902 		return (1);
5903 	}
5904 	return (0);
5905 }
5906 
5907 static void
5908 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
5909 {
5910 	if (rsm->m->m_len > rsm->orig_m_len) {
5911 		/*
5912 		 * Mbuf grew, caused by sbcompress, our offset does
5913 		 * not change.
5914 		 */
5915 		rsm->orig_m_len = rsm->m->m_len;
5916 	} else if (rsm->m->m_len < rsm->orig_m_len) {
5917 		/*
5918 		 * Mbuf shrank, trimmed off the top by an ack, our
5919 		 * offset changes.
5920 		 */
5921 		rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
5922 		rsm->orig_m_len = rsm->m->m_len;
5923 	}
5924 }
5925 
5926 static void
5927 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
5928 {
5929 	struct mbuf *m;
5930 	uint32_t soff;
5931 
5932 	if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
5933 		/* Fix up the orig_m_len and possibly the mbuf offset */
5934 		rack_adjust_orig_mlen(src_rsm);
5935 	}
5936 	m = src_rsm->m;
5937 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
5938 	while (soff >= m->m_len) {
5939 		/* Move out past this mbuf */
5940 		soff -= m->m_len;
5941 		m = m->m_next;
5942 		KASSERT((m != NULL),
5943 			("rsm:%p nrsm:%p hit at soff:%u null m",
5944 			 src_rsm, rsm, soff));
5945 	}
5946 	rsm->m = m;
5947 	rsm->soff = soff;
5948 	rsm->orig_m_len = m->m_len;
5949 }
5950 
5951 static __inline void
5952 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
5953 	       struct rack_sendmap *rsm, uint32_t start)
5954 {
5955 	int idx;
5956 
5957 	nrsm->r_start = start;
5958 	nrsm->r_end = rsm->r_end;
5959 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
5960 	nrsm->r_flags = rsm->r_flags;
5961 	nrsm->r_dupack = rsm->r_dupack;
5962 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
5963 	nrsm->r_rtr_bytes = 0;
5964 	nrsm->r_fas = rsm->r_fas;
5965 	rsm->r_end = nrsm->r_start;
5966 	nrsm->r_just_ret = rsm->r_just_ret;
5967 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
5968 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
5969 	}
5970 	/* Now if we have SYN flag we keep it on the left edge */
5971 	if (nrsm->r_flags & RACK_HAS_SYN)
5972 		nrsm->r_flags &= ~RACK_HAS_SYN;
5973 	/* Now if we have a FIN flag we keep it on the right edge */
5974 	if (rsm->r_flags & RACK_HAS_FIN)
5975 		rsm->r_flags &= ~RACK_HAS_FIN;
5976 	/* Push bit must go to the right edge as well */
5977 	if (rsm->r_flags & RACK_HAD_PUSH)
5978 		rsm->r_flags &= ~RACK_HAD_PUSH;
5979 	/* Clone over the state of the hw_tls flag */
5980 	nrsm->r_hw_tls = rsm->r_hw_tls;
5981 	/*
5982 	 * Now we need to find nrsm's new location in the mbuf chain
5983 	 * we basically calculate a new offset, which is soff +
5984 	 * how much is left in original rsm. Then we walk out the mbuf
5985 	 * chain to find the righ position, it may be the same mbuf
5986 	 * or maybe not.
5987 	 */
5988 	KASSERT(((rsm->m != NULL) ||
5989 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
5990 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
5991 	if (rsm->m)
5992 		rack_setup_offset_for_rsm(rsm, nrsm);
5993 }
5994 
5995 static struct rack_sendmap *
5996 rack_merge_rsm(struct tcp_rack *rack,
5997 	       struct rack_sendmap *l_rsm,
5998 	       struct rack_sendmap *r_rsm)
5999 {
6000 	/*
6001 	 * We are merging two ack'd RSM's,
6002 	 * the l_rsm is on the left (lower seq
6003 	 * values) and the r_rsm is on the right
6004 	 * (higher seq value). The simplest way
6005 	 * to merge these is to move the right
6006 	 * one into the left. I don't think there
6007 	 * is any reason we need to try to find
6008 	 * the oldest (or last oldest retransmitted).
6009 	 */
6010 #ifdef INVARIANTS
6011 	struct rack_sendmap *rm;
6012 #endif
6013 	rack_log_map_chg(rack->rc_tp, rack, NULL,
6014 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
6015 	l_rsm->r_end = r_rsm->r_end;
6016 	if (l_rsm->r_dupack < r_rsm->r_dupack)
6017 		l_rsm->r_dupack = r_rsm->r_dupack;
6018 	if (r_rsm->r_rtr_bytes)
6019 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
6020 	if (r_rsm->r_in_tmap) {
6021 		/* This really should not happen */
6022 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
6023 		r_rsm->r_in_tmap = 0;
6024 	}
6025 
6026 	/* Now the flags */
6027 	if (r_rsm->r_flags & RACK_HAS_FIN)
6028 		l_rsm->r_flags |= RACK_HAS_FIN;
6029 	if (r_rsm->r_flags & RACK_TLP)
6030 		l_rsm->r_flags |= RACK_TLP;
6031 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
6032 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
6033 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
6034 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6035 		/*
6036 		 * If both are app-limited then let the
6037 		 * free lower the count. If right is app
6038 		 * limited and left is not, transfer.
6039 		 */
6040 		l_rsm->r_flags |= RACK_APP_LIMITED;
6041 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
6042 		if (r_rsm == rack->r_ctl.rc_first_appl)
6043 			rack->r_ctl.rc_first_appl = l_rsm;
6044 	}
6045 #ifndef INVARIANTS
6046 	(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6047 #else
6048 	rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6049 	if (rm != r_rsm) {
6050 		panic("removing head in rack:%p rsm:%p rm:%p",
6051 		      rack, r_rsm, rm);
6052 	}
6053 #endif
6054 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6055 		/* Transfer the split limit to the map we free */
6056 		r_rsm->r_limit_type = l_rsm->r_limit_type;
6057 		l_rsm->r_limit_type = 0;
6058 	}
6059 	rack_free(rack, r_rsm);
6060 	return (l_rsm);
6061 }
6062 
6063 /*
6064  * TLP Timer, here we simply setup what segment we want to
6065  * have the TLP expire on, the normal rack_output() will then
6066  * send it out.
6067  *
6068  * We return 1, saying don't proceed with rack_output only
6069  * when all timers have been stopped (destroyed PCB?).
6070  */
6071 static int
6072 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
6073 {
6074 	/*
6075 	 * Tail Loss Probe.
6076 	 */
6077 	struct rack_sendmap *rsm = NULL;
6078 #ifdef INVARIANTS
6079 	struct rack_sendmap *insret;
6080 #endif
6081 	struct socket *so = tptosocket(tp);
6082 	uint32_t amm;
6083 	uint32_t out, avail;
6084 	int collapsed_win = 0;
6085 
6086 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6087 		/* Its not time yet */
6088 		return (0);
6089 	}
6090 	if (ctf_progress_timeout_check(tp, true)) {
6091 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6092 		return (-ETIMEDOUT);	/* tcp_drop() */
6093 	}
6094 	/*
6095 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
6096 	 * need to figure out how to force a full MSS segment out.
6097 	 */
6098 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6099 	rack->r_ctl.retran_during_recovery = 0;
6100 	rack->r_ctl.dsack_byte_cnt = 0;
6101 	counter_u64_add(rack_tlp_tot, 1);
6102 	if (rack->r_state && (rack->r_state != tp->t_state))
6103 		rack_set_state(tp, rack);
6104 	avail = sbavail(&so->so_snd);
6105 	out = tp->snd_max - tp->snd_una;
6106 	if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
6107 		/* special case, we need a retransmission */
6108 		collapsed_win = 1;
6109 		goto need_retran;
6110 	}
6111 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
6112 		rack->r_ctl.dsack_persist--;
6113 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6114 			rack->r_ctl.num_dsack = 0;
6115 		}
6116 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6117 	}
6118 	if ((tp->t_flags & TF_GPUTINPROG) &&
6119 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
6120 		/*
6121 		 * If this is the second in a row
6122 		 * TLP and we are doing a measurement
6123 		 * its time to abandon the measurement.
6124 		 * Something is likely broken on
6125 		 * the clients network and measuring a
6126 		 * broken network does us no good.
6127 		 */
6128 		tp->t_flags &= ~TF_GPUTINPROG;
6129 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6130 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
6131 					   tp->gput_seq,
6132 					   0, 0, 18, __LINE__, NULL, 0);
6133 	}
6134 	/*
6135 	 * Check our send oldest always settings, and if
6136 	 * there is an oldest to send jump to the need_retran.
6137 	 */
6138 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6139 		goto need_retran;
6140 
6141 	if (avail > out) {
6142 		/* New data is available */
6143 		amm = avail - out;
6144 		if (amm > ctf_fixed_maxseg(tp)) {
6145 			amm = ctf_fixed_maxseg(tp);
6146 			if ((amm + out) > tp->snd_wnd) {
6147 				/* We are rwnd limited */
6148 				goto need_retran;
6149 			}
6150 		} else if (amm < ctf_fixed_maxseg(tp)) {
6151 			/* not enough to fill a MTU */
6152 			goto need_retran;
6153 		}
6154 		if (IN_FASTRECOVERY(tp->t_flags)) {
6155 			/* Unlikely */
6156 			if (rack->rack_no_prr == 0) {
6157 				if (out + amm <= tp->snd_wnd) {
6158 					rack->r_ctl.rc_prr_sndcnt = amm;
6159 					rack->r_ctl.rc_tlp_new_data = amm;
6160 					rack_log_to_prr(rack, 4, 0, __LINE__);
6161 				}
6162 			} else
6163 				goto need_retran;
6164 		} else {
6165 			/* Set the send-new override */
6166 			if (out + amm <= tp->snd_wnd)
6167 				rack->r_ctl.rc_tlp_new_data = amm;
6168 			else
6169 				goto need_retran;
6170 		}
6171 		rack->r_ctl.rc_tlpsend = NULL;
6172 		counter_u64_add(rack_tlp_newdata, 1);
6173 		goto send;
6174 	}
6175 need_retran:
6176 	/*
6177 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
6178 	 * optionally the first un-acked segment.
6179 	 */
6180 	if (collapsed_win == 0) {
6181 		if (rack_always_send_oldest)
6182 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6183 		else {
6184 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6185 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6186 				rsm = rack_find_high_nonack(rack, rsm);
6187 			}
6188 		}
6189 		if (rsm == NULL) {
6190 #ifdef TCP_BLACKBOX
6191 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6192 #endif
6193 			goto out;
6194 		}
6195 	} else {
6196 		/*
6197 		 * We must find the last segment
6198 		 * that was acceptable by the client.
6199 		 */
6200 		RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6201 			if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6202 				/* Found one */
6203 				break;
6204 			}
6205 		}
6206 		if (rsm == NULL) {
6207 			/* None? if so send the first */
6208 			rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6209 			if (rsm == NULL) {
6210 #ifdef TCP_BLACKBOX
6211 				tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6212 #endif
6213 				goto out;
6214 			}
6215 		}
6216 	}
6217 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6218 		/*
6219 		 * We need to split this the last segment in two.
6220 		 */
6221 		struct rack_sendmap *nrsm;
6222 
6223 		nrsm = rack_alloc_full_limit(rack);
6224 		if (nrsm == NULL) {
6225 			/*
6226 			 * No memory to split, we will just exit and punt
6227 			 * off to the RXT timer.
6228 			 */
6229 			goto out;
6230 		}
6231 		rack_clone_rsm(rack, nrsm, rsm,
6232 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
6233 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6234 #ifndef INVARIANTS
6235 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6236 #else
6237 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6238 		if (insret != NULL) {
6239 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6240 			      nrsm, insret, rack, rsm);
6241 		}
6242 #endif
6243 		if (rsm->r_in_tmap) {
6244 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6245 			nrsm->r_in_tmap = 1;
6246 		}
6247 		rsm = nrsm;
6248 	}
6249 	rack->r_ctl.rc_tlpsend = rsm;
6250 send:
6251 	/* Make sure output path knows we are doing a TLP */
6252 	*doing_tlp = 1;
6253 	rack->r_timer_override = 1;
6254 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6255 	return (0);
6256 out:
6257 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6258 	return (0);
6259 }
6260 
6261 /*
6262  * Delayed ack Timer, here we simply need to setup the
6263  * ACK_NOW flag and remove the DELACK flag. From there
6264  * the output routine will send the ack out.
6265  *
6266  * We only return 1, saying don't proceed, if all timers
6267  * are stopped (destroyed PCB?).
6268  */
6269 static int
6270 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6271 {
6272 
6273 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6274 	tp->t_flags &= ~TF_DELACK;
6275 	tp->t_flags |= TF_ACKNOW;
6276 	KMOD_TCPSTAT_INC(tcps_delack);
6277 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6278 	return (0);
6279 }
6280 
6281 /*
6282  * Persists timer, here we simply send the
6283  * same thing as a keepalive will.
6284  * the one byte send.
6285  *
6286  * We only return 1, saying don't proceed, if all timers
6287  * are stopped (destroyed PCB?).
6288  */
6289 static int
6290 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6291 {
6292 	struct tcptemp *t_template;
6293 	int32_t retval = 1;
6294 
6295 	if (rack->rc_in_persist == 0)
6296 		return (0);
6297 	if (ctf_progress_timeout_check(tp, false)) {
6298 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6299 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6300 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6301 		return (-ETIMEDOUT);	/* tcp_drop() */
6302 	}
6303 	/*
6304 	 * Persistence timer into zero window. Force a byte to be output, if
6305 	 * possible.
6306 	 */
6307 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
6308 	/*
6309 	 * Hack: if the peer is dead/unreachable, we do not time out if the
6310 	 * window is closed.  After a full backoff, drop the connection if
6311 	 * the idle time (no responses to probes) reaches the maximum
6312 	 * backoff that we would use if retransmitting.
6313 	 */
6314 	if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6315 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6316 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6317 		KMOD_TCPSTAT_INC(tcps_persistdrop);
6318 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6319 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6320 		retval = -ETIMEDOUT;	/* tcp_drop() */
6321 		goto out;
6322 	}
6323 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6324 	    tp->snd_una == tp->snd_max)
6325 		rack_exit_persist(tp, rack, cts);
6326 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6327 	/*
6328 	 * If the user has closed the socket then drop a persisting
6329 	 * connection after a much reduced timeout.
6330 	 */
6331 	if (tp->t_state > TCPS_CLOSE_WAIT &&
6332 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6333 		KMOD_TCPSTAT_INC(tcps_persistdrop);
6334 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6335 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6336 		retval = -ETIMEDOUT;	/* tcp_drop() */
6337 		goto out;
6338 	}
6339 	t_template = tcpip_maketemplate(rack->rc_inp);
6340 	if (t_template) {
6341 		/* only set it if we were answered */
6342 		if (rack->forced_ack == 0) {
6343 			rack->forced_ack = 1;
6344 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6345 		} else {
6346 			rack->probe_not_answered = 1;
6347 			counter_u64_add(rack_persists_loss, 1);
6348 			rack->r_ctl.persist_lost_ends++;
6349 		}
6350 		counter_u64_add(rack_persists_sends, 1);
6351 		tcp_respond(tp, t_template->tt_ipgen,
6352 			    &t_template->tt_t, (struct mbuf *)NULL,
6353 			    tp->rcv_nxt, tp->snd_una - 1, 0);
6354 		/* This sends an ack */
6355 		if (tp->t_flags & TF_DELACK)
6356 			tp->t_flags &= ~TF_DELACK;
6357 		free(t_template, M_TEMP);
6358 	}
6359 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6360 		tp->t_rxtshift++;
6361 out:
6362 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6363 	rack_start_hpts_timer(rack, tp, cts,
6364 			      0, 0, 0);
6365 	return (retval);
6366 }
6367 
6368 /*
6369  * If a keepalive goes off, we had no other timers
6370  * happening. We always return 1 here since this
6371  * routine either drops the connection or sends
6372  * out a segment with respond.
6373  */
6374 static int
6375 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6376 {
6377 	struct tcptemp *t_template;
6378 	struct inpcb *inp = tptoinpcb(tp);
6379 
6380 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6381 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6382 	/*
6383 	 * Keep-alive timer went off; send something or drop connection if
6384 	 * idle for too long.
6385 	 */
6386 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
6387 	if (tp->t_state < TCPS_ESTABLISHED)
6388 		goto dropit;
6389 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6390 	    tp->t_state <= TCPS_CLOSING) {
6391 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6392 			goto dropit;
6393 		/*
6394 		 * Send a packet designed to force a response if the peer is
6395 		 * up and reachable: either an ACK if the connection is
6396 		 * still alive, or an RST if the peer has closed the
6397 		 * connection due to timeout or reboot. Using sequence
6398 		 * number tp->snd_una-1 causes the transmitted zero-length
6399 		 * segment to lie outside the receive window; by the
6400 		 * protocol spec, this requires the correspondent TCP to
6401 		 * respond.
6402 		 */
6403 		KMOD_TCPSTAT_INC(tcps_keepprobe);
6404 		t_template = tcpip_maketemplate(inp);
6405 		if (t_template) {
6406 			if (rack->forced_ack == 0) {
6407 				rack->forced_ack = 1;
6408 				rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6409 			} else {
6410 				rack->probe_not_answered = 1;
6411 			}
6412 			tcp_respond(tp, t_template->tt_ipgen,
6413 			    &t_template->tt_t, (struct mbuf *)NULL,
6414 			    tp->rcv_nxt, tp->snd_una - 1, 0);
6415 			free(t_template, M_TEMP);
6416 		}
6417 	}
6418 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6419 	return (1);
6420 dropit:
6421 	KMOD_TCPSTAT_INC(tcps_keepdrops);
6422 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6423 	return (-ETIMEDOUT);	/* tcp_drop() */
6424 }
6425 
6426 /*
6427  * Retransmit helper function, clear up all the ack
6428  * flags and take care of important book keeping.
6429  */
6430 static void
6431 rack_remxt_tmr(struct tcpcb *tp)
6432 {
6433 	/*
6434 	 * The retransmit timer went off, all sack'd blocks must be
6435 	 * un-acked.
6436 	 */
6437 	struct rack_sendmap *rsm, *trsm = NULL;
6438 	struct tcp_rack *rack;
6439 
6440 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6441 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6442 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6443 	if (rack->r_state && (rack->r_state != tp->t_state))
6444 		rack_set_state(tp, rack);
6445 	/*
6446 	 * Ideally we would like to be able to
6447 	 * mark SACK-PASS on anything not acked here.
6448 	 *
6449 	 * However, if we do that we would burst out
6450 	 * all that data 1ms apart. This would be unwise,
6451 	 * so for now we will just let the normal rxt timer
6452 	 * and tlp timer take care of it.
6453 	 *
6454 	 * Also we really need to stick them back in sequence
6455 	 * order. This way we send in the proper order and any
6456 	 * sacks that come floating in will "re-ack" the data.
6457 	 * To do this we zap the tmap with an INIT and then
6458 	 * walk through and place every rsm in the RB tree
6459 	 * back in its seq ordered place.
6460 	 */
6461 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
6462 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6463 		rsm->r_dupack = 0;
6464 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6465 		/* We must re-add it back to the tlist */
6466 		if (trsm == NULL) {
6467 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6468 		} else {
6469 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6470 		}
6471 		rsm->r_in_tmap = 1;
6472 		trsm = rsm;
6473 		if (rsm->r_flags & RACK_ACKED)
6474 			rsm->r_flags |= RACK_WAS_ACKED;
6475 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
6476 		rsm->r_flags |= RACK_MUST_RXT;
6477 	}
6478 	/* Clear the count (we just un-acked them) */
6479 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6480 	rack->r_ctl.rc_sacked = 0;
6481 	rack->r_ctl.rc_sacklast = NULL;
6482 	rack->r_ctl.rc_agg_delayed = 0;
6483 	rack->r_early = 0;
6484 	rack->r_ctl.rc_agg_early = 0;
6485 	rack->r_late = 0;
6486 	/* Clear the tlp rtx mark */
6487 	rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6488 	if (rack->r_ctl.rc_resend != NULL)
6489 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6490 	rack->r_ctl.rc_prr_sndcnt = 0;
6491 	rack_log_to_prr(rack, 6, 0, __LINE__);
6492 	rack->r_timer_override = 1;
6493 	if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6494 #ifdef NETFLIX_EXP_DETECTION
6495 	    || (rack->sack_attack_disable != 0)
6496 #endif
6497 		    ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6498 		/*
6499 		 * For non-sack customers new data
6500 		 * needs to go out as retransmits until
6501 		 * we retransmit up to snd_max.
6502 		 */
6503 		rack->r_must_retran = 1;
6504 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6505 						rack->r_ctl.rc_sacked);
6506 	}
6507 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6508 }
6509 
6510 static void
6511 rack_convert_rtts(struct tcpcb *tp)
6512 {
6513 	if (tp->t_srtt > 1) {
6514 		uint32_t val, frac;
6515 
6516 		val = tp->t_srtt >> TCP_RTT_SHIFT;
6517 		frac = tp->t_srtt & 0x1f;
6518 		tp->t_srtt = TICKS_2_USEC(val);
6519 		/*
6520 		 * frac is the fractional part of the srtt (if any)
6521 		 * but its in ticks and every bit represents
6522 		 * 1/32nd of a hz.
6523 		 */
6524 		if (frac) {
6525 			if (hz == 1000) {
6526 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6527 			} else {
6528 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6529 			}
6530 			tp->t_srtt += frac;
6531 		}
6532 	}
6533 	if (tp->t_rttvar) {
6534 		uint32_t val, frac;
6535 
6536 		val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6537 		frac = tp->t_rttvar & 0x1f;
6538 		tp->t_rttvar = TICKS_2_USEC(val);
6539 		/*
6540 		 * frac is the fractional part of the srtt (if any)
6541 		 * but its in ticks and every bit represents
6542 		 * 1/32nd of a hz.
6543 		 */
6544 		if (frac) {
6545 			if (hz == 1000) {
6546 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6547 			} else {
6548 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6549 			}
6550 			tp->t_rttvar += frac;
6551 		}
6552 	}
6553 	tp->t_rxtcur = RACK_REXMTVAL(tp);
6554 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6555 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6556 	}
6557 	if (tp->t_rxtcur > rack_rto_max) {
6558 		tp->t_rxtcur = rack_rto_max;
6559 	}
6560 }
6561 
6562 static void
6563 rack_cc_conn_init(struct tcpcb *tp)
6564 {
6565 	struct tcp_rack *rack;
6566 	uint32_t srtt;
6567 
6568 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6569 	srtt = tp->t_srtt;
6570 	cc_conn_init(tp);
6571 	/*
6572 	 * Now convert to rack's internal format,
6573 	 * if required.
6574 	 */
6575 	if ((srtt == 0) && (tp->t_srtt != 0))
6576 		rack_convert_rtts(tp);
6577 	/*
6578 	 * We want a chance to stay in slowstart as
6579 	 * we create a connection. TCP spec says that
6580 	 * initially ssthresh is infinite. For our
6581 	 * purposes that is the snd_wnd.
6582 	 */
6583 	if (tp->snd_ssthresh < tp->snd_wnd) {
6584 		tp->snd_ssthresh = tp->snd_wnd;
6585 	}
6586 	/*
6587 	 * We also want to assure a IW worth of
6588 	 * data can get inflight.
6589 	 */
6590 	if (rc_init_window(rack) < tp->snd_cwnd)
6591 		tp->snd_cwnd = rc_init_window(rack);
6592 }
6593 
6594 /*
6595  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6596  * we will setup to retransmit the lowest seq number outstanding.
6597  */
6598 static int
6599 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6600 {
6601 	struct inpcb *inp = tptoinpcb(tp);
6602 	int32_t rexmt;
6603 	int32_t retval = 0;
6604 	bool isipv6;
6605 
6606 	if ((tp->t_flags & TF_GPUTINPROG) &&
6607 	    (tp->t_rxtshift)) {
6608 		/*
6609 		 * We have had a second timeout
6610 		 * measurements on successive rxt's are not profitable.
6611 		 * It is unlikely to be of any use (the network is
6612 		 * broken or the client went away).
6613 		 */
6614 		tp->t_flags &= ~TF_GPUTINPROG;
6615 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6616 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
6617 					   tp->gput_seq,
6618 					   0, 0, 18, __LINE__, NULL, 0);
6619 	}
6620 	if (ctf_progress_timeout_check(tp, false)) {
6621 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6622 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6623 		return (-ETIMEDOUT);	/* tcp_drop() */
6624 	}
6625 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6626 	rack->r_ctl.retran_during_recovery = 0;
6627 	rack->rc_ack_required = 1;
6628 	rack->r_ctl.dsack_byte_cnt = 0;
6629 	if (IN_FASTRECOVERY(tp->t_flags))
6630 		tp->t_flags |= TF_WASFRECOVERY;
6631 	else
6632 		tp->t_flags &= ~TF_WASFRECOVERY;
6633 	if (IN_CONGRECOVERY(tp->t_flags))
6634 		tp->t_flags |= TF_WASCRECOVERY;
6635 	else
6636 		tp->t_flags &= ~TF_WASCRECOVERY;
6637 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6638 	    (tp->snd_una == tp->snd_max)) {
6639 		/* Nothing outstanding .. nothing to do */
6640 		return (0);
6641 	}
6642 	if (rack->r_ctl.dsack_persist) {
6643 		rack->r_ctl.dsack_persist--;
6644 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6645 			rack->r_ctl.num_dsack = 0;
6646 		}
6647 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6648 	}
6649 	/*
6650 	 * Rack can only run one timer  at a time, so we cannot
6651 	 * run a KEEPINIT (gating SYN sending) and a retransmit
6652 	 * timer for the SYN. So if we are in a front state and
6653 	 * have a KEEPINIT timer we need to check the first transmit
6654 	 * against now to see if we have exceeded the KEEPINIT time
6655 	 * (if one is set).
6656 	 */
6657 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6658 	    (TP_KEEPINIT(tp) != 0)) {
6659 		struct rack_sendmap *rsm;
6660 
6661 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6662 		if (rsm) {
6663 			/* Ok we have something outstanding to test keepinit with */
6664 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6665 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6666 				/* We have exceeded the KEEPINIT time */
6667 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6668 				goto drop_it;
6669 			}
6670 		}
6671 	}
6672 	/*
6673 	 * Retransmission timer went off.  Message has not been acked within
6674 	 * retransmit interval.  Back off to a longer retransmit interval
6675 	 * and retransmit one segment.
6676 	 */
6677 	rack_remxt_tmr(tp);
6678 	if ((rack->r_ctl.rc_resend == NULL) ||
6679 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6680 		/*
6681 		 * If the rwnd collapsed on
6682 		 * the one we are retransmitting
6683 		 * it does not count against the
6684 		 * rxt count.
6685 		 */
6686 		tp->t_rxtshift++;
6687 	}
6688 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6689 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6690 drop_it:
6691 		tp->t_rxtshift = TCP_MAXRXTSHIFT;
6692 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6693 		/* XXXGL: previously t_softerror was casted to uint16_t */
6694 		MPASS(tp->t_softerror >= 0);
6695 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
6696 		goto out;	/* tcp_drop() */
6697 	}
6698 	if (tp->t_state == TCPS_SYN_SENT) {
6699 		/*
6700 		 * If the SYN was retransmitted, indicate CWND to be limited
6701 		 * to 1 segment in cc_conn_init().
6702 		 */
6703 		tp->snd_cwnd = 1;
6704 	} else if (tp->t_rxtshift == 1) {
6705 		/*
6706 		 * first retransmit; record ssthresh and cwnd so they can be
6707 		 * recovered if this turns out to be a "bad" retransmit. A
6708 		 * retransmit is considered "bad" if an ACK for this segment
6709 		 * is received within RTT/2 interval; the assumption here is
6710 		 * that the ACK was already in flight.  See "On Estimating
6711 		 * End-to-End Network Path Properties" by Allman and Paxson
6712 		 * for more details.
6713 		 */
6714 		tp->snd_cwnd_prev = tp->snd_cwnd;
6715 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
6716 		tp->snd_recover_prev = tp->snd_recover;
6717 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6718 		tp->t_flags |= TF_PREVVALID;
6719 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6720 		tp->t_flags &= ~TF_PREVVALID;
6721 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6722 	if ((tp->t_state == TCPS_SYN_SENT) ||
6723 	    (tp->t_state == TCPS_SYN_RECEIVED))
6724 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6725 	else
6726 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6727 
6728 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6729 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6730 	/*
6731 	 * We enter the path for PLMTUD if connection is established or, if
6732 	 * connection is FIN_WAIT_1 status, reason for the last is that if
6733 	 * amount of data we send is very small, we could send it in couple
6734 	 * of packets and process straight to FIN. In that case we won't
6735 	 * catch ESTABLISHED state.
6736 	 */
6737 #ifdef INET6
6738 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
6739 #else
6740 	isipv6 = false;
6741 #endif
6742 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
6743 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6744 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6745 	    ((tp->t_state == TCPS_ESTABLISHED) ||
6746 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
6747 		/*
6748 		 * Idea here is that at each stage of mtu probe (usually,
6749 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
6750 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6751 		 * should take care of that.
6752 		 */
6753 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6754 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6755 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6756 		    tp->t_rxtshift % 2 == 0)) {
6757 			/*
6758 			 * Enter Path MTU Black-hole Detection mechanism: -
6759 			 * Disable Path MTU Discovery (IP "DF" bit). -
6760 			 * Reduce MTU to lower value than what we negotiated
6761 			 * with peer.
6762 			 */
6763 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6764 				/* Record that we may have found a black hole. */
6765 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6766 				/* Keep track of previous MSS. */
6767 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6768 			}
6769 
6770 			/*
6771 			 * Reduce the MSS to blackhole value or to the
6772 			 * default in an attempt to retransmit.
6773 			 */
6774 #ifdef INET6
6775 			if (isipv6 &&
6776 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6777 				/* Use the sysctl tuneable blackhole MSS. */
6778 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6779 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6780 			} else if (isipv6) {
6781 				/* Use the default MSS. */
6782 				tp->t_maxseg = V_tcp_v6mssdflt;
6783 				/*
6784 				 * Disable Path MTU Discovery when we switch
6785 				 * to minmss.
6786 				 */
6787 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6788 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6789 			}
6790 #endif
6791 #if defined(INET6) && defined(INET)
6792 			else
6793 #endif
6794 #ifdef INET
6795 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6796 				/* Use the sysctl tuneable blackhole MSS. */
6797 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6798 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6799 			} else {
6800 				/* Use the default MSS. */
6801 				tp->t_maxseg = V_tcp_mssdflt;
6802 				/*
6803 				 * Disable Path MTU Discovery when we switch
6804 				 * to minmss.
6805 				 */
6806 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6807 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6808 			}
6809 #endif
6810 		} else {
6811 			/*
6812 			 * If further retransmissions are still unsuccessful
6813 			 * with a lowered MTU, maybe this isn't a blackhole
6814 			 * and we restore the previous MSS and blackhole
6815 			 * detection flags. The limit '6' is determined by
6816 			 * giving each probe stage (1448, 1188, 524) 2
6817 			 * chances to recover.
6818 			 */
6819 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6820 			    (tp->t_rxtshift >= 6)) {
6821 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6822 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6823 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6824 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6825 			}
6826 		}
6827 	}
6828 	/*
6829 	 * Disable RFC1323 and SACK if we haven't got any response to
6830 	 * our third SYN to work-around some broken terminal servers
6831 	 * (most of which have hopefully been retired) that have bad VJ
6832 	 * header compression code which trashes TCP segments containing
6833 	 * unknown-to-them TCP options.
6834 	 */
6835 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6836 	    (tp->t_rxtshift == 3))
6837 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6838 	/*
6839 	 * If we backed off this far, our srtt estimate is probably bogus.
6840 	 * Clobber it so we'll take the next rtt measurement as our srtt;
6841 	 * move the current srtt into rttvar to keep the current retransmit
6842 	 * times until then.
6843 	 */
6844 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6845 #ifdef INET6
6846 		if ((inp->inp_vflag & INP_IPV6) != 0)
6847 			in6_losing(inp);
6848 		else
6849 #endif
6850 			in_losing(inp);
6851 		tp->t_rttvar += tp->t_srtt;
6852 		tp->t_srtt = 0;
6853 	}
6854 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6855 	tp->snd_recover = tp->snd_max;
6856 	tp->t_flags |= TF_ACKNOW;
6857 	tp->t_rtttime = 0;
6858 	rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
6859 out:
6860 	return (retval);
6861 }
6862 
6863 static int
6864 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
6865 {
6866 	int32_t ret = 0;
6867 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6868 
6869 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
6870 	    (tp->t_flags & TF_GPUTINPROG)) {
6871 		/*
6872 		 * We have a goodput in progress
6873 		 * and we have entered a late state.
6874 		 * Do we have enough data in the sb
6875 		 * to handle the GPUT request?
6876 		 */
6877 		uint32_t bytes;
6878 
6879 		bytes = tp->gput_ack - tp->gput_seq;
6880 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
6881 			bytes += tp->gput_seq - tp->snd_una;
6882 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
6883 			/*
6884 			 * There are not enough bytes in the socket
6885 			 * buffer that have been sent to cover this
6886 			 * measurement. Cancel it.
6887 			 */
6888 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6889 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
6890 						   tp->gput_seq,
6891 						   0, 0, 18, __LINE__, NULL, 0);
6892 			tp->t_flags &= ~TF_GPUTINPROG;
6893 		}
6894 	}
6895 	if (timers == 0) {
6896 		return (0);
6897 	}
6898 	if (tp->t_state == TCPS_LISTEN) {
6899 		/* no timers on listen sockets */
6900 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6901 			return (0);
6902 		return (1);
6903 	}
6904 	if ((timers & PACE_TMR_RACK) &&
6905 	    rack->rc_on_min_to) {
6906 		/*
6907 		 * For the rack timer when we
6908 		 * are on a min-timeout (which means rrr_conf = 3)
6909 		 * we don't want to check the timer. It may
6910 		 * be going off for a pace and thats ok we
6911 		 * want to send the retransmit (if its ready).
6912 		 *
6913 		 * If its on a normal rack timer (non-min) then
6914 		 * we will check if its expired.
6915 		 */
6916 		goto skip_time_check;
6917 	}
6918 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6919 		uint32_t left;
6920 
6921 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6922 			ret = -1;
6923 			rack_log_to_processing(rack, cts, ret, 0);
6924 			return (0);
6925 		}
6926 		if (hpts_calling == 0) {
6927 			/*
6928 			 * A user send or queued mbuf (sack) has called us? We
6929 			 * return 0 and let the pacing guards
6930 			 * deal with it if they should or
6931 			 * should not cause a send.
6932 			 */
6933 			ret = -2;
6934 			rack_log_to_processing(rack, cts, ret, 0);
6935 			return (0);
6936 		}
6937 		/*
6938 		 * Ok our timer went off early and we are not paced false
6939 		 * alarm, go back to sleep.
6940 		 */
6941 		ret = -3;
6942 		left = rack->r_ctl.rc_timer_exp - cts;
6943 		tcp_hpts_insert(tptoinpcb(tp), HPTS_MS_TO_SLOTS(left));
6944 		rack_log_to_processing(rack, cts, ret, left);
6945 		return (1);
6946 	}
6947 skip_time_check:
6948 	rack->rc_tmr_stopped = 0;
6949 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
6950 	if (timers & PACE_TMR_DELACK) {
6951 		ret = rack_timeout_delack(tp, rack, cts);
6952 	} else if (timers & PACE_TMR_RACK) {
6953 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
6954 		rack->r_fast_output = 0;
6955 		ret = rack_timeout_rack(tp, rack, cts);
6956 	} else if (timers & PACE_TMR_TLP) {
6957 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
6958 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
6959 	} else if (timers & PACE_TMR_RXT) {
6960 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
6961 		rack->r_fast_output = 0;
6962 		ret = rack_timeout_rxt(tp, rack, cts);
6963 	} else if (timers & PACE_TMR_PERSIT) {
6964 		ret = rack_timeout_persist(tp, rack, cts);
6965 	} else if (timers & PACE_TMR_KEEP) {
6966 		ret = rack_timeout_keepalive(tp, rack, cts);
6967 	}
6968 	rack_log_to_processing(rack, cts, ret, timers);
6969 	return (ret);
6970 }
6971 
6972 static void
6973 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
6974 {
6975 	struct timeval tv;
6976 	uint32_t us_cts, flags_on_entry;
6977 	uint8_t hpts_removed = 0;
6978 
6979 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
6980 	us_cts = tcp_get_usecs(&tv);
6981 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
6982 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
6983 	     ((tp->snd_max - tp->snd_una) == 0))) {
6984 		tcp_hpts_remove(rack->rc_inp);
6985 		hpts_removed = 1;
6986 		/* If we were not delayed cancel out the flag. */
6987 		if ((tp->snd_max - tp->snd_una) == 0)
6988 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
6989 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
6990 	}
6991 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
6992 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
6993 		if (tcp_in_hpts(rack->rc_inp) &&
6994 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
6995 			/*
6996 			 * Canceling timer's when we have no output being
6997 			 * paced. We also must remove ourselves from the
6998 			 * hpts.
6999 			 */
7000 			tcp_hpts_remove(rack->rc_inp);
7001 			hpts_removed = 1;
7002 		}
7003 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7004 	}
7005 	if (hpts_removed == 0)
7006 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7007 }
7008 
7009 static int
7010 rack_stopall(struct tcpcb *tp)
7011 {
7012 	struct tcp_rack *rack;
7013 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7014 	rack->t_timers_stopped = 1;
7015 	return (0);
7016 }
7017 
7018 static void
7019 rack_stop_all_timers(struct tcpcb *tp)
7020 {
7021 	struct tcp_rack *rack;
7022 
7023 	/*
7024 	 * Assure no timers are running.
7025 	 */
7026 	if (tcp_timer_active(tp, TT_PERSIST)) {
7027 		/* We enter in persists, set the flag appropriately */
7028 		rack = (struct tcp_rack *)tp->t_fb_ptr;
7029 		rack->rc_in_persist = 1;
7030 	}
7031 }
7032 
7033 static void
7034 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7035     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7036 {
7037 	int32_t idx;
7038 
7039 	rsm->r_rtr_cnt++;
7040 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7041 	rsm->r_dupack = 0;
7042 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7043 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7044 		rsm->r_flags |= RACK_OVERMAX;
7045 	}
7046 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7047 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7048 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7049 	}
7050 	idx = rsm->r_rtr_cnt - 1;
7051 	rsm->r_tim_lastsent[idx] = ts;
7052 	/*
7053 	 * Here we don't add in the len of send, since its already
7054 	 * in snduna <->snd_max.
7055 	 */
7056 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
7057 				     rack->r_ctl.rc_sacked);
7058 	if (rsm->r_flags & RACK_ACKED) {
7059 		/* Problably MTU discovery messing with us */
7060 		rsm->r_flags &= ~RACK_ACKED;
7061 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7062 	}
7063 	if (rsm->r_in_tmap) {
7064 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7065 		rsm->r_in_tmap = 0;
7066 	}
7067 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7068 	rsm->r_in_tmap = 1;
7069 	/* Take off the must retransmit flag, if its on */
7070 	if (rsm->r_flags & RACK_MUST_RXT) {
7071 		if (rack->r_must_retran)
7072 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
7073 		if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
7074 			/*
7075 			 * We have retransmitted all we need. Clear
7076 			 * any must retransmit flags.
7077 			 */
7078 			rack->r_must_retran = 0;
7079 			rack->r_ctl.rc_out_at_rto = 0;
7080 		}
7081 		rsm->r_flags &= ~RACK_MUST_RXT;
7082 	}
7083 	if (rsm->r_flags & RACK_SACK_PASSED) {
7084 		/* We have retransmitted due to the SACK pass */
7085 		rsm->r_flags &= ~RACK_SACK_PASSED;
7086 		rsm->r_flags |= RACK_WAS_SACKPASS;
7087 	}
7088 }
7089 
7090 static uint32_t
7091 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7092     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7093 {
7094 	/*
7095 	 * We (re-)transmitted starting at rsm->r_start for some length
7096 	 * (possibly less than r_end.
7097 	 */
7098 	struct rack_sendmap *nrsm;
7099 #ifdef INVARIANTS
7100 	struct rack_sendmap *insret;
7101 #endif
7102 	uint32_t c_end;
7103 	int32_t len;
7104 
7105 	len = *lenp;
7106 	c_end = rsm->r_start + len;
7107 	if (SEQ_GEQ(c_end, rsm->r_end)) {
7108 		/*
7109 		 * We retransmitted the whole piece or more than the whole
7110 		 * slopping into the next rsm.
7111 		 */
7112 		rack_update_rsm(tp, rack, rsm, ts, add_flag);
7113 		if (c_end == rsm->r_end) {
7114 			*lenp = 0;
7115 			return (0);
7116 		} else {
7117 			int32_t act_len;
7118 
7119 			/* Hangs over the end return whats left */
7120 			act_len = rsm->r_end - rsm->r_start;
7121 			*lenp = (len - act_len);
7122 			return (rsm->r_end);
7123 		}
7124 		/* We don't get out of this block. */
7125 	}
7126 	/*
7127 	 * Here we retransmitted less than the whole thing which means we
7128 	 * have to split this into what was transmitted and what was not.
7129 	 */
7130 	nrsm = rack_alloc_full_limit(rack);
7131 	if (nrsm == NULL) {
7132 		/*
7133 		 * We can't get memory, so lets not proceed.
7134 		 */
7135 		*lenp = 0;
7136 		return (0);
7137 	}
7138 	/*
7139 	 * So here we are going to take the original rsm and make it what we
7140 	 * retransmitted. nrsm will be the tail portion we did not
7141 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7142 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7143 	 * 1, 6 and the new piece will be 6, 11.
7144 	 */
7145 	rack_clone_rsm(rack, nrsm, rsm, c_end);
7146 	nrsm->r_dupack = 0;
7147 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7148 #ifndef INVARIANTS
7149 	(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7150 #else
7151 	insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7152 	if (insret != NULL) {
7153 		panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7154 		      nrsm, insret, rack, rsm);
7155 	}
7156 #endif
7157 	if (rsm->r_in_tmap) {
7158 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7159 		nrsm->r_in_tmap = 1;
7160 	}
7161 	rsm->r_flags &= (~RACK_HAS_FIN);
7162 	rack_update_rsm(tp, rack, rsm, ts, add_flag);
7163 	/* Log a split of rsm into rsm and nrsm */
7164 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7165 	*lenp = 0;
7166 	return (0);
7167 }
7168 
7169 static void
7170 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7171 		uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
7172 		struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls)
7173 {
7174 	struct tcp_rack *rack;
7175 	struct rack_sendmap *rsm, *nrsm, fe;
7176 #ifdef INVARIANTS
7177 	struct rack_sendmap *insret;
7178 #endif
7179 	register uint32_t snd_max, snd_una;
7180 
7181 	/*
7182 	 * Add to the RACK log of packets in flight or retransmitted. If
7183 	 * there is a TS option we will use the TS echoed, if not we will
7184 	 * grab a TS.
7185 	 *
7186 	 * Retransmissions will increment the count and move the ts to its
7187 	 * proper place. Note that if options do not include TS's then we
7188 	 * won't be able to effectively use the ACK for an RTT on a retran.
7189 	 *
7190 	 * Notes about r_start and r_end. Lets consider a send starting at
7191 	 * sequence 1 for 10 bytes. In such an example the r_start would be
7192 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7193 	 * This means that r_end is actually the first sequence for the next
7194 	 * slot (11).
7195 	 *
7196 	 */
7197 	/*
7198 	 * If err is set what do we do XXXrrs? should we not add the thing?
7199 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
7200 	 * i.e. proceed with add ** do this for now.
7201 	 */
7202 	INP_WLOCK_ASSERT(tptoinpcb(tp));
7203 	if (err)
7204 		/*
7205 		 * We don't log errors -- we could but snd_max does not
7206 		 * advance in this case either.
7207 		 */
7208 		return;
7209 
7210 	if (th_flags & TH_RST) {
7211 		/*
7212 		 * We don't log resets and we return immediately from
7213 		 * sending
7214 		 */
7215 		return;
7216 	}
7217 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7218 	snd_una = tp->snd_una;
7219 	snd_max = tp->snd_max;
7220 	if (th_flags & (TH_SYN | TH_FIN)) {
7221 		/*
7222 		 * The call to rack_log_output is made before bumping
7223 		 * snd_max. This means we can record one extra byte on a SYN
7224 		 * or FIN if seq_out is adding more on and a FIN is present
7225 		 * (and we are not resending).
7226 		 */
7227 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7228 			len++;
7229 		if (th_flags & TH_FIN)
7230 			len++;
7231 		if (SEQ_LT(snd_max, tp->snd_nxt)) {
7232 			/*
7233 			 * The add/update as not been done for the FIN/SYN
7234 			 * yet.
7235 			 */
7236 			snd_max = tp->snd_nxt;
7237 		}
7238 	}
7239 	if (SEQ_LEQ((seq_out + len), snd_una)) {
7240 		/* Are sending an old segment to induce an ack (keep-alive)? */
7241 		return;
7242 	}
7243 	if (SEQ_LT(seq_out, snd_una)) {
7244 		/* huh? should we panic? */
7245 		uint32_t end;
7246 
7247 		end = seq_out + len;
7248 		seq_out = snd_una;
7249 		if (SEQ_GEQ(end, seq_out))
7250 			len = end - seq_out;
7251 		else
7252 			len = 0;
7253 	}
7254 	if (len == 0) {
7255 		/* We don't log zero window probes */
7256 		return;
7257 	}
7258 	if (IN_FASTRECOVERY(tp->t_flags)) {
7259 		rack->r_ctl.rc_prr_out += len;
7260 	}
7261 	/* First question is it a retransmission or new? */
7262 	if (seq_out == snd_max) {
7263 		/* Its new */
7264 again:
7265 		rsm = rack_alloc(rack);
7266 		if (rsm == NULL) {
7267 			/*
7268 			 * Hmm out of memory and the tcb got destroyed while
7269 			 * we tried to wait.
7270 			 */
7271 			return;
7272 		}
7273 		if (th_flags & TH_FIN) {
7274 			rsm->r_flags = RACK_HAS_FIN|add_flag;
7275 		} else {
7276 			rsm->r_flags = add_flag;
7277 		}
7278 		if (hw_tls)
7279 			rsm->r_hw_tls = 1;
7280 		rsm->r_tim_lastsent[0] = cts;
7281 		rsm->r_rtr_cnt = 1;
7282 		rsm->r_rtr_bytes = 0;
7283 		if (th_flags & TH_SYN) {
7284 			/* The data space is one beyond snd_una */
7285 			rsm->r_flags |= RACK_HAS_SYN;
7286 		}
7287 		rsm->r_start = seq_out;
7288 		rsm->r_end = rsm->r_start + len;
7289 		rsm->r_dupack = 0;
7290 		/*
7291 		 * save off the mbuf location that
7292 		 * sndmbuf_noadv returned (which is
7293 		 * where we started copying from)..
7294 		 */
7295 		rsm->m = s_mb;
7296 		rsm->soff = s_moff;
7297 		/*
7298 		 * Here we do add in the len of send, since its not yet
7299 		 * reflected in in snduna <->snd_max
7300 		 */
7301 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
7302 					      rack->r_ctl.rc_sacked) +
7303 			      (rsm->r_end - rsm->r_start));
7304 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7305 		if (rsm->m) {
7306 			if (rsm->m->m_len <= rsm->soff) {
7307 				/*
7308 				 * XXXrrs Question, will this happen?
7309 				 *
7310 				 * If sbsndptr is set at the correct place
7311 				 * then s_moff should always be somewhere
7312 				 * within rsm->m. But if the sbsndptr was
7313 				 * off then that won't be true. If it occurs
7314 				 * we need to walkout to the correct location.
7315 				 */
7316 				struct mbuf *lm;
7317 
7318 				lm = rsm->m;
7319 				while (lm->m_len <= rsm->soff) {
7320 					rsm->soff -= lm->m_len;
7321 					lm = lm->m_next;
7322 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7323 							     __func__, rack, s_moff, s_mb, rsm->soff));
7324 				}
7325 				rsm->m = lm;
7326 			}
7327 			rsm->orig_m_len = rsm->m->m_len;
7328 		} else
7329 			rsm->orig_m_len = 0;
7330 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7331 		/* Log a new rsm */
7332 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7333 #ifndef INVARIANTS
7334 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7335 #else
7336 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7337 		if (insret != NULL) {
7338 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7339 			      nrsm, insret, rack, rsm);
7340 		}
7341 #endif
7342 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7343 		rsm->r_in_tmap = 1;
7344 		/*
7345 		 * Special case detection, is there just a single
7346 		 * packet outstanding when we are not in recovery?
7347 		 *
7348 		 * If this is true mark it so.
7349 		 */
7350 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7351 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7352 			struct rack_sendmap *prsm;
7353 
7354 			prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7355 			if (prsm)
7356 				prsm->r_one_out_nr = 1;
7357 		}
7358 		return;
7359 	}
7360 	/*
7361 	 * If we reach here its a retransmission and we need to find it.
7362 	 */
7363 	memset(&fe, 0, sizeof(fe));
7364 more:
7365 	if (hintrsm && (hintrsm->r_start == seq_out)) {
7366 		rsm = hintrsm;
7367 		hintrsm = NULL;
7368 	} else {
7369 		/* No hints sorry */
7370 		rsm = NULL;
7371 	}
7372 	if ((rsm) && (rsm->r_start == seq_out)) {
7373 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7374 		if (len == 0) {
7375 			return;
7376 		} else {
7377 			goto more;
7378 		}
7379 	}
7380 	/* Ok it was not the last pointer go through it the hard way. */
7381 refind:
7382 	fe.r_start = seq_out;
7383 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7384 	if (rsm) {
7385 		if (rsm->r_start == seq_out) {
7386 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7387 			if (len == 0) {
7388 				return;
7389 			} else {
7390 				goto refind;
7391 			}
7392 		}
7393 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7394 			/* Transmitted within this piece */
7395 			/*
7396 			 * Ok we must split off the front and then let the
7397 			 * update do the rest
7398 			 */
7399 			nrsm = rack_alloc_full_limit(rack);
7400 			if (nrsm == NULL) {
7401 				rack_update_rsm(tp, rack, rsm, cts, add_flag);
7402 				return;
7403 			}
7404 			/*
7405 			 * copy rsm to nrsm and then trim the front of rsm
7406 			 * to not include this part.
7407 			 */
7408 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
7409 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7410 #ifndef INVARIANTS
7411 			(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7412 #else
7413 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7414 			if (insret != NULL) {
7415 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7416 				      nrsm, insret, rack, rsm);
7417 			}
7418 #endif
7419 			if (rsm->r_in_tmap) {
7420 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7421 				nrsm->r_in_tmap = 1;
7422 			}
7423 			rsm->r_flags &= (~RACK_HAS_FIN);
7424 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7425 			if (len == 0) {
7426 				return;
7427 			} else if (len > 0)
7428 				goto refind;
7429 		}
7430 	}
7431 	/*
7432 	 * Hmm not found in map did they retransmit both old and on into the
7433 	 * new?
7434 	 */
7435 	if (seq_out == tp->snd_max) {
7436 		goto again;
7437 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
7438 #ifdef INVARIANTS
7439 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7440 		       seq_out, len, tp->snd_una, tp->snd_max);
7441 		printf("Starting Dump of all rack entries\n");
7442 		RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7443 			printf("rsm:%p start:%u end:%u\n",
7444 			       rsm, rsm->r_start, rsm->r_end);
7445 		}
7446 		printf("Dump complete\n");
7447 		panic("seq_out not found rack:%p tp:%p",
7448 		      rack, tp);
7449 #endif
7450 	} else {
7451 #ifdef INVARIANTS
7452 		/*
7453 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
7454 		 * flag)
7455 		 */
7456 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7457 		      seq_out, len, tp->snd_max, tp);
7458 #endif
7459 	}
7460 }
7461 
7462 /*
7463  * Record one of the RTT updates from an ack into
7464  * our sample structure.
7465  */
7466 
7467 static void
7468 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7469 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7470 {
7471 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7472 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7473 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7474 	}
7475 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7476 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7477 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7478 	}
7479 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7480 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7481 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
7482 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7483 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7484 	}
7485 	if ((confidence == 1) &&
7486 	    ((rsm == NULL) ||
7487 	     (rsm->r_just_ret) ||
7488 	     (rsm->r_one_out_nr &&
7489 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7490 		/*
7491 		 * If the rsm had a just return
7492 		 * hit it then we can't trust the
7493 		 * rtt measurement for buffer deterimination
7494 		 * Note that a confidence of 2, indicates
7495 		 * SACK'd which overrides the r_just_ret or
7496 		 * the r_one_out_nr. If it was a CUM-ACK and
7497 		 * we had only two outstanding, but get an
7498 		 * ack for only 1. Then that also lowers our
7499 		 * confidence.
7500 		 */
7501 		confidence = 0;
7502 	}
7503 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7504 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7505 		if (rack->r_ctl.rack_rs.confidence == 0) {
7506 			/*
7507 			 * We take anything with no current confidence
7508 			 * saved.
7509 			 */
7510 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7511 			rack->r_ctl.rack_rs.confidence = confidence;
7512 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7513 		} else if (confidence || rack->r_ctl.rack_rs.confidence) {
7514 			/*
7515 			 * Once we have a confident number,
7516 			 * we can update it with a smaller
7517 			 * value since this confident number
7518 			 * may include the DSACK time until
7519 			 * the next segment (the second one) arrived.
7520 			 */
7521 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7522 			rack->r_ctl.rack_rs.confidence = confidence;
7523 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7524 		}
7525 	}
7526 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7527 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7528 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7529 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
7530 }
7531 
7532 /*
7533  * Collect new round-trip time estimate
7534  * and update averages and current timeout.
7535  */
7536 static void
7537 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7538 {
7539 	int32_t delta;
7540 	int32_t rtt;
7541 
7542 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7543 		/* No valid sample */
7544 		return;
7545 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7546 		/* We are to use the lowest RTT seen in a single ack */
7547 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7548 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7549 		/* We are to use the highest RTT seen in a single ack */
7550 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7551 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7552 		/* We are to use the average RTT seen in a single ack */
7553 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7554 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7555 	} else {
7556 #ifdef INVARIANTS
7557 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7558 #endif
7559 		return;
7560 	}
7561 	if (rtt == 0)
7562 		rtt = 1;
7563 	if (rack->rc_gp_rtt_set == 0) {
7564 		/*
7565 		 * With no RTT we have to accept
7566 		 * even one we are not confident of.
7567 		 */
7568 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7569 		rack->rc_gp_rtt_set = 1;
7570 	} else if (rack->r_ctl.rack_rs.confidence) {
7571 		/* update the running gp srtt */
7572 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7573 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7574 	}
7575 	if (rack->r_ctl.rack_rs.confidence) {
7576 		/*
7577 		 * record the low and high for highly buffered path computation,
7578 		 * we only do this if we are confident (not a retransmission).
7579 		 */
7580 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7581 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7582 		}
7583 		if (rack->rc_highly_buffered == 0) {
7584 			/*
7585 			 * Currently once we declare a path has
7586 			 * highly buffered there is no going
7587 			 * back, which may be a problem...
7588 			 */
7589 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7590 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7591 						     rack->r_ctl.rc_highest_us_rtt,
7592 						     rack->r_ctl.rc_lowest_us_rtt,
7593 						     RACK_RTTS_SEEHBP);
7594 				rack->rc_highly_buffered = 1;
7595 			}
7596 		}
7597 	}
7598 	if ((rack->r_ctl.rack_rs.confidence) ||
7599 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7600 		/*
7601 		 * If we are highly confident of it <or> it was
7602 		 * never retransmitted we accept it as the last us_rtt.
7603 		 */
7604 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7605 		/* The lowest rtt can be set if its was not retransmited */
7606 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7607 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7608 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
7609 				rack->r_ctl.rc_lowest_us_rtt = 1;
7610 		}
7611 	}
7612 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7613 	if (tp->t_srtt != 0) {
7614 		/*
7615 		 * We keep a simple srtt in microseconds, like our rtt
7616 		 * measurement. We don't need to do any tricks with shifting
7617 		 * etc. Instead we just add in 1/8th of the new measurement
7618 		 * and subtract out 1/8 of the old srtt. We do the same with
7619 		 * the variance after finding the absolute value of the
7620 		 * difference between this sample and the current srtt.
7621 		 */
7622 		delta = tp->t_srtt - rtt;
7623 		/* Take off 1/8th of the current sRTT */
7624 		tp->t_srtt -= (tp->t_srtt >> 3);
7625 		/* Add in 1/8th of the new RTT just measured */
7626 		tp->t_srtt += (rtt >> 3);
7627 		if (tp->t_srtt <= 0)
7628 			tp->t_srtt = 1;
7629 		/* Now lets make the absolute value of the variance */
7630 		if (delta < 0)
7631 			delta = -delta;
7632 		/* Subtract out 1/8th */
7633 		tp->t_rttvar -= (tp->t_rttvar >> 3);
7634 		/* Add in 1/8th of the new variance we just saw */
7635 		tp->t_rttvar += (delta >> 3);
7636 		if (tp->t_rttvar <= 0)
7637 			tp->t_rttvar = 1;
7638 	} else {
7639 		/*
7640 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
7641 		 * variance to half the rtt (so our first retransmit happens
7642 		 * at 3*rtt).
7643 		 */
7644 		tp->t_srtt = rtt;
7645 		tp->t_rttvar = rtt >> 1;
7646 	}
7647 	rack->rc_srtt_measure_made = 1;
7648 	KMOD_TCPSTAT_INC(tcps_rttupdated);
7649 	tp->t_rttupdated++;
7650 #ifdef STATS
7651 	if (rack_stats_gets_ms_rtt == 0) {
7652 		/* Send in the microsecond rtt used for rxt timeout purposes */
7653 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7654 	} else if (rack_stats_gets_ms_rtt == 1) {
7655 		/* Send in the millisecond rtt used for rxt timeout purposes */
7656 		int32_t ms_rtt;
7657 
7658 		/* Round up */
7659 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7660 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7661 	} else if (rack_stats_gets_ms_rtt == 2) {
7662 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
7663 		int32_t ms_rtt;
7664 
7665 		/* Round up */
7666 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7667 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7668 	}  else {
7669 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
7670 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7671 	}
7672 
7673 #endif
7674 	/*
7675 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
7676 	 * way we do the smoothing, srtt and rttvar will each average +1/2
7677 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
7678 	 * tick of rounding and 1 extra tick because of +-1/2 tick
7679 	 * uncertainty in the firing of the timer.  The bias will give us
7680 	 * exactly the 1.5 tick we need.  But, because the bias is
7681 	 * statistical, we have to test that we don't drop below the minimum
7682 	 * feasible timer (which is 2 ticks).
7683 	 */
7684 	tp->t_rxtshift = 0;
7685 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7686 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7687 	rack_log_rtt_sample(rack, rtt);
7688 	tp->t_softerror = 0;
7689 }
7690 
7691 
7692 static void
7693 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7694 {
7695 	/*
7696 	 * Apply to filter the inbound us-rtt at us_cts.
7697 	 */
7698 	uint32_t old_rtt;
7699 
7700 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7701 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7702 			       us_rtt, us_cts);
7703 	if (old_rtt > us_rtt) {
7704 		/* We just hit a new lower rtt time */
7705 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7706 				     __LINE__, RACK_RTTS_NEWRTT);
7707 		/*
7708 		 * Only count it if its lower than what we saw within our
7709 		 * calculated range.
7710 		 */
7711 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7712 			if (rack_probertt_lower_within &&
7713 			    rack->rc_gp_dyn_mul &&
7714 			    (rack->use_fixed_rate == 0) &&
7715 			    (rack->rc_always_pace)) {
7716 				/*
7717 				 * We are seeing a new lower rtt very close
7718 				 * to the time that we would have entered probe-rtt.
7719 				 * This is probably due to the fact that a peer flow
7720 				 * has entered probe-rtt. Lets go in now too.
7721 				 */
7722 				uint32_t val;
7723 
7724 				val = rack_probertt_lower_within * rack_time_between_probertt;
7725 				val /= 100;
7726 				if ((rack->in_probe_rtt == 0)  &&
7727 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
7728 					rack_enter_probertt(rack, us_cts);
7729 				}
7730 			}
7731 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7732 		}
7733 	}
7734 }
7735 
7736 static int
7737 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7738     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7739 {
7740 	uint32_t us_rtt;
7741 	int32_t i, all;
7742 	uint32_t t, len_acked;
7743 
7744 	if ((rsm->r_flags & RACK_ACKED) ||
7745 	    (rsm->r_flags & RACK_WAS_ACKED))
7746 		/* Already done */
7747 		return (0);
7748 	if (rsm->r_no_rtt_allowed) {
7749 		/* Not allowed */
7750 		return (0);
7751 	}
7752 	if (ack_type == CUM_ACKED) {
7753 		if (SEQ_GT(th_ack, rsm->r_end)) {
7754 			len_acked = rsm->r_end - rsm->r_start;
7755 			all = 1;
7756 		} else {
7757 			len_acked = th_ack - rsm->r_start;
7758 			all = 0;
7759 		}
7760 	} else {
7761 		len_acked = rsm->r_end - rsm->r_start;
7762 		all = 0;
7763 	}
7764 	if (rsm->r_rtr_cnt == 1) {
7765 
7766 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7767 		if ((int)t <= 0)
7768 			t = 1;
7769 		if (!tp->t_rttlow || tp->t_rttlow > t)
7770 			tp->t_rttlow = t;
7771 		if (!rack->r_ctl.rc_rack_min_rtt ||
7772 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7773 			rack->r_ctl.rc_rack_min_rtt = t;
7774 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
7775 				rack->r_ctl.rc_rack_min_rtt = 1;
7776 			}
7777 		}
7778 		if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7779 			us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7780 		else
7781 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7782 		if (us_rtt == 0)
7783 			us_rtt = 1;
7784 		if (CC_ALGO(tp)->rttsample != NULL) {
7785 			/* Kick the RTT to the CC */
7786 			CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
7787 		}
7788 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7789 		if (ack_type == SACKED) {
7790 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7791 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7792 		} else {
7793 			/*
7794 			 * We need to setup what our confidence
7795 			 * is in this ack.
7796 			 *
7797 			 * If the rsm was app limited and it is
7798 			 * less than a mss in length (the end
7799 			 * of the send) then we have a gap. If we
7800 			 * were app limited but say we were sending
7801 			 * multiple MSS's then we are more confident
7802 			 * int it.
7803 			 *
7804 			 * When we are not app-limited then we see if
7805 			 * the rsm is being included in the current
7806 			 * measurement, we tell this by the app_limited_needs_set
7807 			 * flag.
7808 			 *
7809 			 * Note that being cwnd blocked is not applimited
7810 			 * as well as the pacing delay between packets which
7811 			 * are sending only 1 or 2 MSS's also will show up
7812 			 * in the RTT. We probably need to examine this algorithm
7813 			 * a bit more and enhance it to account for the delay
7814 			 * between rsm's. We could do that by saving off the
7815 			 * pacing delay of each rsm (in an rsm) and then
7816 			 * factoring that in somehow though for now I am
7817 			 * not sure how :)
7818 			 */
7819 			int calc_conf = 0;
7820 
7821 			if (rsm->r_flags & RACK_APP_LIMITED) {
7822 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7823 					calc_conf = 0;
7824 				else
7825 					calc_conf = 1;
7826 			} else if (rack->app_limited_needs_set == 0) {
7827 				calc_conf = 1;
7828 			} else {
7829 				calc_conf = 0;
7830 			}
7831 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7832 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7833 					    calc_conf, rsm, rsm->r_rtr_cnt);
7834 		}
7835 		if ((rsm->r_flags & RACK_TLP) &&
7836 		    (!IN_FASTRECOVERY(tp->t_flags))) {
7837 			/* Segment was a TLP and our retrans matched */
7838 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7839 				rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
7840 			}
7841 		}
7842 		if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7843 			/* New more recent rack_tmit_time */
7844 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7845 			rack->rc_rack_rtt = t;
7846 		}
7847 		return (1);
7848 	}
7849 	/*
7850 	 * We clear the soft/rxtshift since we got an ack.
7851 	 * There is no assurance we will call the commit() function
7852 	 * so we need to clear these to avoid incorrect handling.
7853 	 */
7854 	tp->t_rxtshift = 0;
7855 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7856 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7857 	tp->t_softerror = 0;
7858 	if (to && (to->to_flags & TOF_TS) &&
7859 	    (ack_type == CUM_ACKED) &&
7860 	    (to->to_tsecr) &&
7861 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7862 		/*
7863 		 * Now which timestamp does it match? In this block the ACK
7864 		 * must be coming from a previous transmission.
7865 		 */
7866 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
7867 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7868 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7869 				if ((int)t <= 0)
7870 					t = 1;
7871 				if (CC_ALGO(tp)->rttsample != NULL) {
7872 					/*
7873 					 * Kick the RTT to the CC, here
7874 					 * we lie a bit in that we know the
7875 					 * retransmission is correct even though
7876 					 * we retransmitted. This is because
7877 					 * we match the timestamps.
7878 					 */
7879 					if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
7880 						us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
7881 					else
7882 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
7883 					CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
7884 				}
7885 				if ((i + 1) < rsm->r_rtr_cnt) {
7886 					/*
7887 					 * The peer ack'd from our previous
7888 					 * transmission. We have a spurious
7889 					 * retransmission and thus we dont
7890 					 * want to update our rack_rtt.
7891 					 *
7892 					 * Hmm should there be a CC revert here?
7893 					 *
7894 					 */
7895 					return (0);
7896 				}
7897 				if (!tp->t_rttlow || tp->t_rttlow > t)
7898 					tp->t_rttlow = t;
7899 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7900 					rack->r_ctl.rc_rack_min_rtt = t;
7901 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
7902 						rack->r_ctl.rc_rack_min_rtt = 1;
7903 					}
7904 				}
7905 				if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7906 					   (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7907 					/* New more recent rack_tmit_time */
7908 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7909 					rack->rc_rack_rtt = t;
7910 				}
7911 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7912 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7913 						    rsm->r_rtr_cnt);
7914 				return (1);
7915 			}
7916 		}
7917 		goto ts_not_found;
7918 	} else {
7919 		/*
7920 		 * Ok its a SACK block that we retransmitted. or a windows
7921 		 * machine without timestamps. We can tell nothing from the
7922 		 * time-stamp since its not there or the time the peer last
7923 		 * recieved a segment that moved forward its cum-ack point.
7924 		 */
7925 ts_not_found:
7926 		i = rsm->r_rtr_cnt - 1;
7927 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7928 		if ((int)t <= 0)
7929 			t = 1;
7930 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7931 			/*
7932 			 * We retransmitted and the ack came back in less
7933 			 * than the smallest rtt we have observed. We most
7934 			 * likely did an improper retransmit as outlined in
7935 			 * 6.2 Step 2 point 2 in the rack-draft so we
7936 			 * don't want to update our rack_rtt. We in
7937 			 * theory (in future) might want to think about reverting our
7938 			 * cwnd state but we won't for now.
7939 			 */
7940 			return (0);
7941 		} else if (rack->r_ctl.rc_rack_min_rtt) {
7942 			/*
7943 			 * We retransmitted it and the retransmit did the
7944 			 * job.
7945 			 */
7946 			if (!rack->r_ctl.rc_rack_min_rtt ||
7947 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7948 				rack->r_ctl.rc_rack_min_rtt = t;
7949 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
7950 					rack->r_ctl.rc_rack_min_rtt = 1;
7951 				}
7952 			}
7953 			if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
7954 				/* New more recent rack_tmit_time */
7955 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
7956 				rack->rc_rack_rtt = t;
7957 			}
7958 			return (1);
7959 		}
7960 	}
7961 	return (0);
7962 }
7963 
7964 /*
7965  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
7966  */
7967 static void
7968 rack_log_sack_passed(struct tcpcb *tp,
7969     struct tcp_rack *rack, struct rack_sendmap *rsm)
7970 {
7971 	struct rack_sendmap *nrsm;
7972 
7973 	nrsm = rsm;
7974 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
7975 	    rack_head, r_tnext) {
7976 		if (nrsm == rsm) {
7977 			/* Skip orginal segment he is acked */
7978 			continue;
7979 		}
7980 		if (nrsm->r_flags & RACK_ACKED) {
7981 			/*
7982 			 * Skip ack'd segments, though we
7983 			 * should not see these, since tmap
7984 			 * should not have ack'd segments.
7985 			 */
7986 			continue;
7987 		}
7988 		if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
7989 			/*
7990 			 * If the peer dropped the rwnd on
7991 			 * these then we don't worry about them.
7992 			 */
7993 			continue;
7994 		}
7995 		if (nrsm->r_flags & RACK_SACK_PASSED) {
7996 			/*
7997 			 * We found one that is already marked
7998 			 * passed, we have been here before and
7999 			 * so all others below this are marked.
8000 			 */
8001 			break;
8002 		}
8003 		nrsm->r_flags |= RACK_SACK_PASSED;
8004 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
8005 	}
8006 }
8007 
8008 static void
8009 rack_need_set_test(struct tcpcb *tp,
8010 		   struct tcp_rack *rack,
8011 		   struct rack_sendmap *rsm,
8012 		   tcp_seq th_ack,
8013 		   int line,
8014 		   int use_which)
8015 {
8016 
8017 	if ((tp->t_flags & TF_GPUTINPROG) &&
8018 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8019 		/*
8020 		 * We were app limited, and this ack
8021 		 * butts up or goes beyond the point where we want
8022 		 * to start our next measurement. We need
8023 		 * to record the new gput_ts as here and
8024 		 * possibly update the start sequence.
8025 		 */
8026 		uint32_t seq, ts;
8027 
8028 		if (rsm->r_rtr_cnt > 1) {
8029 			/*
8030 			 * This is a retransmit, can we
8031 			 * really make any assessment at this
8032 			 * point?  We are not really sure of
8033 			 * the timestamp, is it this or the
8034 			 * previous transmission?
8035 			 *
8036 			 * Lets wait for something better that
8037 			 * is not retransmitted.
8038 			 */
8039 			return;
8040 		}
8041 		seq = tp->gput_seq;
8042 		ts = tp->gput_ts;
8043 		rack->app_limited_needs_set = 0;
8044 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8045 		/* Do we start at a new end? */
8046 		if ((use_which == RACK_USE_BEG) &&
8047 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8048 			/*
8049 			 * When we get an ACK that just eats
8050 			 * up some of the rsm, we set RACK_USE_BEG
8051 			 * since whats at r_start (i.e. th_ack)
8052 			 * is left unacked and thats where the
8053 			 * measurement not starts.
8054 			 */
8055 			tp->gput_seq = rsm->r_start;
8056 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8057 		}
8058 		if ((use_which == RACK_USE_END) &&
8059 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8060 			    /*
8061 			     * We use the end when the cumack
8062 			     * is moving forward and completely
8063 			     * deleting the rsm passed so basically
8064 			     * r_end holds th_ack.
8065 			     *
8066 			     * For SACK's we also want to use the end
8067 			     * since this piece just got sacked and
8068 			     * we want to target anything after that
8069 			     * in our measurement.
8070 			     */
8071 			    tp->gput_seq = rsm->r_end;
8072 			    rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8073 		}
8074 		if (use_which == RACK_USE_END_OR_THACK) {
8075 			/*
8076 			 * special case for ack moving forward,
8077 			 * not a sack, we need to move all the
8078 			 * way up to where this ack cum-ack moves
8079 			 * to.
8080 			 */
8081 			if (SEQ_GT(th_ack, rsm->r_end))
8082 				tp->gput_seq = th_ack;
8083 			else
8084 				tp->gput_seq = rsm->r_end;
8085 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8086 		}
8087 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8088 			/*
8089 			 * We moved beyond this guy's range, re-calculate
8090 			 * the new end point.
8091 			 */
8092 			if (rack->rc_gp_filled == 0) {
8093 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8094 			} else {
8095 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8096 			}
8097 		}
8098 		/*
8099 		 * We are moving the goal post, we may be able to clear the
8100 		 * measure_saw_probe_rtt flag.
8101 		 */
8102 		if ((rack->in_probe_rtt == 0) &&
8103 		    (rack->measure_saw_probe_rtt) &&
8104 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8105 			rack->measure_saw_probe_rtt = 0;
8106 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8107 					   seq, tp->gput_seq, 0, 5, line, NULL, 0);
8108 		if (rack->rc_gp_filled &&
8109 		    ((tp->gput_ack - tp->gput_seq) <
8110 		     max(rc_init_window(rack), (MIN_GP_WIN *
8111 						ctf_fixed_maxseg(tp))))) {
8112 			uint32_t ideal_amount;
8113 
8114 			ideal_amount = rack_get_measure_window(tp, rack);
8115 			if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
8116 				/*
8117 				 * There is no sense of continuing this measurement
8118 				 * because its too small to gain us anything we
8119 				 * trust. Skip it and that way we can start a new
8120 				 * measurement quicker.
8121 				 */
8122 				tp->t_flags &= ~TF_GPUTINPROG;
8123 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8124 							   0, 0, 0, 6, __LINE__, NULL, 0);
8125 			} else {
8126 				/*
8127 				 * Reset the window further out.
8128 				 */
8129 				tp->gput_ack = tp->gput_seq + ideal_amount;
8130 			}
8131 		}
8132 	}
8133 }
8134 
8135 static inline int
8136 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
8137 {
8138 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
8139 		/* Behind our TLP definition or right at */
8140 		return (0);
8141 	}
8142 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
8143 		/* The start is beyond or right at our end of TLP definition */
8144 		return (0);
8145 	}
8146 	/* It has to be a sub-part of the original TLP recorded */
8147 	return (1);
8148 }
8149 
8150 
8151 static uint32_t
8152 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8153 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8154 {
8155 	uint32_t start, end, changed = 0;
8156 	struct rack_sendmap stack_map;
8157 	struct rack_sendmap *rsm, *nrsm, fe, *prev, *next;
8158 #ifdef INVARIANTS
8159 	struct rack_sendmap *insret;
8160 #endif
8161 	int32_t used_ref = 1;
8162 	int moved = 0;
8163 
8164 	start = sack->start;
8165 	end = sack->end;
8166 	rsm = *prsm;
8167 	memset(&fe, 0, sizeof(fe));
8168 do_rest_ofb:
8169 	if ((rsm == NULL) ||
8170 	    (SEQ_LT(end, rsm->r_start)) ||
8171 	    (SEQ_GEQ(start, rsm->r_end)) ||
8172 	    (SEQ_LT(start, rsm->r_start))) {
8173 		/*
8174 		 * We are not in the right spot,
8175 		 * find the correct spot in the tree.
8176 		 */
8177 		used_ref = 0;
8178 		fe.r_start = start;
8179 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8180 		moved++;
8181 	}
8182 	if (rsm == NULL) {
8183 		/* TSNH */
8184 		goto out;
8185 	}
8186 	/* Ok we have an ACK for some piece of this rsm */
8187 	if (rsm->r_start != start) {
8188 		if ((rsm->r_flags & RACK_ACKED) == 0) {
8189 			/*
8190 			 * Before any splitting or hookery is
8191 			 * done is it a TLP of interest i.e. rxt?
8192 			 */
8193 			if ((rsm->r_flags & RACK_TLP) &&
8194 			    (rsm->r_rtr_cnt > 1)) {
8195 				/*
8196 				 * We are splitting a rxt TLP, check
8197 				 * if we need to save off the start/end
8198 				 */
8199 				if (rack->rc_last_tlp_acked_set &&
8200 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8201 					/*
8202 					 * We already turned this on since we are inside
8203 					 * the previous one was a partially sack now we
8204 					 * are getting another one (maybe all of it).
8205 					 *
8206 					 */
8207 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8208 					/*
8209 					 * Lets make sure we have all of it though.
8210 					 */
8211 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8212 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8213 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8214 								     rack->r_ctl.last_tlp_acked_end);
8215 					}
8216 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8217 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8218 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8219 								     rack->r_ctl.last_tlp_acked_end);
8220 					}
8221 				} else {
8222 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8223 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8224 					rack->rc_last_tlp_past_cumack = 0;
8225 					rack->rc_last_tlp_acked_set = 1;
8226 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8227 				}
8228 			}
8229 			/**
8230 			 * Need to split this in two pieces the before and after,
8231 			 * the before remains in the map, the after must be
8232 			 * added. In other words we have:
8233 			 * rsm        |--------------|
8234 			 * sackblk        |------->
8235 			 * rsm will become
8236 			 *     rsm    |---|
8237 			 * and nrsm will be  the sacked piece
8238 			 *     nrsm       |----------|
8239 			 *
8240 			 * But before we start down that path lets
8241 			 * see if the sack spans over on top of
8242 			 * the next guy and it is already sacked.
8243 			 *
8244 			 */
8245 			next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8246 			if (next && (next->r_flags & RACK_ACKED) &&
8247 			    SEQ_GEQ(end, next->r_start)) {
8248 				/**
8249 				 * So the next one is already acked, and
8250 				 * we can thus by hookery use our stack_map
8251 				 * to reflect the piece being sacked and
8252 				 * then adjust the two tree entries moving
8253 				 * the start and ends around. So we start like:
8254 				 *  rsm     |------------|             (not-acked)
8255 				 *  next                 |-----------| (acked)
8256 				 *  sackblk        |-------->
8257 				 *  We want to end like so:
8258 				 *  rsm     |------|                   (not-acked)
8259 				 *  next           |-----------------| (acked)
8260 				 *  nrsm           |-----|
8261 				 * Where nrsm is a temporary stack piece we
8262 				 * use to update all the gizmos.
8263 				 */
8264 				/* Copy up our fudge block */
8265 				nrsm = &stack_map;
8266 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8267 				/* Now adjust our tree blocks */
8268 				rsm->r_end = start;
8269 				next->r_start = start;
8270 				/* Now we must adjust back where next->m is */
8271 				rack_setup_offset_for_rsm(rsm, next);
8272 
8273 				/* We don't need to adjust rsm, it did not change */
8274 				/* Clear out the dup ack count of the remainder */
8275 				rsm->r_dupack = 0;
8276 				rsm->r_just_ret = 0;
8277 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8278 				/* Now lets make sure our fudge block is right */
8279 				nrsm->r_start = start;
8280 				/* Now lets update all the stats and such */
8281 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8282 				if (rack->app_limited_needs_set)
8283 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8284 				changed += (nrsm->r_end - nrsm->r_start);
8285 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8286 				if (nrsm->r_flags & RACK_SACK_PASSED) {
8287 					rack->r_ctl.rc_reorder_ts = cts;
8288 				}
8289 				/*
8290 				 * Now we want to go up from rsm (the
8291 				 * one left un-acked) to the next one
8292 				 * in the tmap. We do this so when
8293 				 * we walk backwards we include marking
8294 				 * sack-passed on rsm (The one passed in
8295 				 * is skipped since it is generally called
8296 				 * on something sacked before removing it
8297 				 * from the tmap).
8298 				 */
8299 				if (rsm->r_in_tmap) {
8300 					nrsm = TAILQ_NEXT(rsm, r_tnext);
8301 					/*
8302 					 * Now that we have the next
8303 					 * one walk backwards from there.
8304 					 */
8305 					if (nrsm && nrsm->r_in_tmap)
8306 						rack_log_sack_passed(tp, rack, nrsm);
8307 				}
8308 				/* Now are we done? */
8309 				if (SEQ_LT(end, next->r_end) ||
8310 				    (end == next->r_end)) {
8311 					/* Done with block */
8312 					goto out;
8313 				}
8314 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8315 				counter_u64_add(rack_sack_used_next_merge, 1);
8316 				/* Postion for the next block */
8317 				start = next->r_end;
8318 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8319 				if (rsm == NULL)
8320 					goto out;
8321 			} else {
8322 				/**
8323 				 * We can't use any hookery here, so we
8324 				 * need to split the map. We enter like
8325 				 * so:
8326 				 *  rsm      |--------|
8327 				 *  sackblk       |----->
8328 				 * We will add the new block nrsm and
8329 				 * that will be the new portion, and then
8330 				 * fall through after reseting rsm. So we
8331 				 * split and look like this:
8332 				 *  rsm      |----|
8333 				 *  sackblk       |----->
8334 				 *  nrsm          |---|
8335 				 * We then fall through reseting
8336 				 * rsm to nrsm, so the next block
8337 				 * picks it up.
8338 				 */
8339 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8340 				if (nrsm == NULL) {
8341 					/*
8342 					 * failed XXXrrs what can we do but loose the sack
8343 					 * info?
8344 					 */
8345 					goto out;
8346 				}
8347 				counter_u64_add(rack_sack_splits, 1);
8348 				rack_clone_rsm(rack, nrsm, rsm, start);
8349 				rsm->r_just_ret = 0;
8350 #ifndef INVARIANTS
8351 				(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8352 #else
8353 				insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8354 				if (insret != NULL) {
8355 					panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8356 					      nrsm, insret, rack, rsm);
8357 				}
8358 #endif
8359 				if (rsm->r_in_tmap) {
8360 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8361 					nrsm->r_in_tmap = 1;
8362 				}
8363 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8364 				rsm->r_flags &= (~RACK_HAS_FIN);
8365 				/* Position us to point to the new nrsm that starts the sack blk */
8366 				rsm = nrsm;
8367 			}
8368 		} else {
8369 			/* Already sacked this piece */
8370 			counter_u64_add(rack_sack_skipped_acked, 1);
8371 			moved++;
8372 			if (end == rsm->r_end) {
8373 				/* Done with block */
8374 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8375 				goto out;
8376 			} else if (SEQ_LT(end, rsm->r_end)) {
8377 				/* A partial sack to a already sacked block */
8378 				moved++;
8379 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8380 				goto out;
8381 			} else {
8382 				/*
8383 				 * The end goes beyond this guy
8384 				 * reposition the start to the
8385 				 * next block.
8386 				 */
8387 				start = rsm->r_end;
8388 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8389 				if (rsm == NULL)
8390 					goto out;
8391 			}
8392 		}
8393 	}
8394 	if (SEQ_GEQ(end, rsm->r_end)) {
8395 		/**
8396 		 * The end of this block is either beyond this guy or right
8397 		 * at this guy. I.e.:
8398 		 *  rsm ---                 |-----|
8399 		 *  end                     |-----|
8400 		 *  <or>
8401 		 *  end                     |---------|
8402 		 */
8403 		if ((rsm->r_flags & RACK_ACKED) == 0) {
8404 			/*
8405 			 * Is it a TLP of interest?
8406 			 */
8407 			if ((rsm->r_flags & RACK_TLP) &&
8408 			    (rsm->r_rtr_cnt > 1)) {
8409 				/*
8410 				 * We are splitting a rxt TLP, check
8411 				 * if we need to save off the start/end
8412 				 */
8413 				if (rack->rc_last_tlp_acked_set &&
8414 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8415 					/*
8416 					 * We already turned this on since we are inside
8417 					 * the previous one was a partially sack now we
8418 					 * are getting another one (maybe all of it).
8419 					 */
8420 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8421 					/*
8422 					 * Lets make sure we have all of it though.
8423 					 */
8424 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8425 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8426 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8427 								     rack->r_ctl.last_tlp_acked_end);
8428 					}
8429 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8430 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8431 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8432 								     rack->r_ctl.last_tlp_acked_end);
8433 					}
8434 				} else {
8435 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8436 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8437 					rack->rc_last_tlp_past_cumack = 0;
8438 					rack->rc_last_tlp_acked_set = 1;
8439 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8440 				}
8441 			}
8442 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8443 			changed += (rsm->r_end - rsm->r_start);
8444 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8445 			if (rsm->r_in_tmap) /* should be true */
8446 				rack_log_sack_passed(tp, rack, rsm);
8447 			/* Is Reordering occuring? */
8448 			if (rsm->r_flags & RACK_SACK_PASSED) {
8449 				rsm->r_flags &= ~RACK_SACK_PASSED;
8450 				rack->r_ctl.rc_reorder_ts = cts;
8451 			}
8452 			if (rack->app_limited_needs_set)
8453 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8454 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8455 			rsm->r_flags |= RACK_ACKED;
8456 			if (rsm->r_in_tmap) {
8457 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8458 				rsm->r_in_tmap = 0;
8459 			}
8460 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8461 		} else {
8462 			counter_u64_add(rack_sack_skipped_acked, 1);
8463 			moved++;
8464 		}
8465 		if (end == rsm->r_end) {
8466 			/* This block only - done, setup for next */
8467 			goto out;
8468 		}
8469 		/*
8470 		 * There is more not coverend by this rsm move on
8471 		 * to the next block in the RB tree.
8472 		 */
8473 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8474 		start = rsm->r_end;
8475 		rsm = nrsm;
8476 		if (rsm == NULL)
8477 			goto out;
8478 		goto do_rest_ofb;
8479 	}
8480 	/**
8481 	 * The end of this sack block is smaller than
8482 	 * our rsm i.e.:
8483 	 *  rsm ---                 |-----|
8484 	 *  end                     |--|
8485 	 */
8486 	if ((rsm->r_flags & RACK_ACKED) == 0) {
8487 		/*
8488 		 * Is it a TLP of interest?
8489 		 */
8490 		if ((rsm->r_flags & RACK_TLP) &&
8491 		    (rsm->r_rtr_cnt > 1)) {
8492 			/*
8493 			 * We are splitting a rxt TLP, check
8494 			 * if we need to save off the start/end
8495 			 */
8496 			if (rack->rc_last_tlp_acked_set &&
8497 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8498 				/*
8499 				 * We already turned this on since we are inside
8500 				 * the previous one was a partially sack now we
8501 				 * are getting another one (maybe all of it).
8502 				 */
8503 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8504 				/*
8505 				 * Lets make sure we have all of it though.
8506 				 */
8507 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8508 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8509 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8510 							     rack->r_ctl.last_tlp_acked_end);
8511 				}
8512 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8513 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8514 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8515 							     rack->r_ctl.last_tlp_acked_end);
8516 				}
8517 			} else {
8518 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8519 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8520 				rack->rc_last_tlp_past_cumack = 0;
8521 				rack->rc_last_tlp_acked_set = 1;
8522 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8523 			}
8524 		}
8525 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8526 		if (prev &&
8527 		    (prev->r_flags & RACK_ACKED)) {
8528 			/**
8529 			 * Goal, we want the right remainder of rsm to shrink
8530 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
8531 			 * We want to expand prev to go all the way
8532 			 * to prev->r_end <- end.
8533 			 * so in the tree we have before:
8534 			 *   prev     |--------|         (acked)
8535 			 *   rsm               |-------| (non-acked)
8536 			 *   sackblk           |-|
8537 			 * We churn it so we end up with
8538 			 *   prev     |----------|       (acked)
8539 			 *   rsm                 |-----| (non-acked)
8540 			 *   nrsm              |-| (temporary)
8541 			 *
8542 			 * Note if either prev/rsm is a TLP we don't
8543 			 * do this.
8544 			 */
8545 			nrsm = &stack_map;
8546 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8547 			prev->r_end = end;
8548 			rsm->r_start = end;
8549 			/* Now adjust nrsm (stack copy) to be
8550 			 * the one that is the small
8551 			 * piece that was "sacked".
8552 			 */
8553 			nrsm->r_end = end;
8554 			rsm->r_dupack = 0;
8555 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8556 			/*
8557 			 * Now that the rsm has had its start moved forward
8558 			 * lets go ahead and get its new place in the world.
8559 			 */
8560 			rack_setup_offset_for_rsm(prev, rsm);
8561 			/*
8562 			 * Now nrsm is our new little piece
8563 			 * that is acked (which was merged
8564 			 * to prev). Update the rtt and changed
8565 			 * based on that. Also check for reordering.
8566 			 */
8567 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8568 			if (rack->app_limited_needs_set)
8569 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8570 			changed += (nrsm->r_end - nrsm->r_start);
8571 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8572 			if (nrsm->r_flags & RACK_SACK_PASSED) {
8573 				rack->r_ctl.rc_reorder_ts = cts;
8574 			}
8575 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8576 			rsm = prev;
8577 			counter_u64_add(rack_sack_used_prev_merge, 1);
8578 		} else {
8579 			/**
8580 			 * This is the case where our previous
8581 			 * block is not acked either, so we must
8582 			 * split the block in two.
8583 			 */
8584 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8585 			if (nrsm == NULL) {
8586 				/* failed rrs what can we do but loose the sack info? */
8587 				goto out;
8588 			}
8589 			if ((rsm->r_flags & RACK_TLP) &&
8590 			    (rsm->r_rtr_cnt > 1)) {
8591 				/*
8592 				 * We are splitting a rxt TLP, check
8593 				 * if we need to save off the start/end
8594 				 */
8595 				if (rack->rc_last_tlp_acked_set &&
8596 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8597 					    /*
8598 					     * We already turned this on since this block is inside
8599 					     * the previous one was a partially sack now we
8600 					     * are getting another one (maybe all of it).
8601 					     */
8602 					    rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8603 					    /*
8604 					     * Lets make sure we have all of it though.
8605 					     */
8606 					    if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8607 						    rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8608 						    rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8609 									 rack->r_ctl.last_tlp_acked_end);
8610 					    }
8611 					    if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8612 						    rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8613 						    rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8614 									 rack->r_ctl.last_tlp_acked_end);
8615 					    }
8616 				    } else {
8617 					    rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8618 					    rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8619 					    rack->rc_last_tlp_acked_set = 1;
8620 					    rack->rc_last_tlp_past_cumack = 0;
8621 					    rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8622 				    }
8623 			}
8624 			/**
8625 			 * In this case nrsm becomes
8626 			 * nrsm->r_start = end;
8627 			 * nrsm->r_end = rsm->r_end;
8628 			 * which is un-acked.
8629 			 * <and>
8630 			 * rsm->r_end = nrsm->r_start;
8631 			 * i.e. the remaining un-acked
8632 			 * piece is left on the left
8633 			 * hand side.
8634 			 *
8635 			 * So we start like this
8636 			 * rsm      |----------| (not acked)
8637 			 * sackblk  |---|
8638 			 * build it so we have
8639 			 * rsm      |---|         (acked)
8640 			 * nrsm         |------|  (not acked)
8641 			 */
8642 			counter_u64_add(rack_sack_splits, 1);
8643 			rack_clone_rsm(rack, nrsm, rsm, end);
8644 			rsm->r_flags &= (~RACK_HAS_FIN);
8645 			rsm->r_just_ret = 0;
8646 #ifndef INVARIANTS
8647 			(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8648 #else
8649 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8650 			if (insret != NULL) {
8651 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8652 				      nrsm, insret, rack, rsm);
8653 			}
8654 #endif
8655 			if (rsm->r_in_tmap) {
8656 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8657 				nrsm->r_in_tmap = 1;
8658 			}
8659 			nrsm->r_dupack = 0;
8660 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8661 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8662 			changed += (rsm->r_end - rsm->r_start);
8663 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8664 			if (rsm->r_in_tmap) /* should be true */
8665 				rack_log_sack_passed(tp, rack, rsm);
8666 			/* Is Reordering occuring? */
8667 			if (rsm->r_flags & RACK_SACK_PASSED) {
8668 				rsm->r_flags &= ~RACK_SACK_PASSED;
8669 				rack->r_ctl.rc_reorder_ts = cts;
8670 			}
8671 			if (rack->app_limited_needs_set)
8672 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8673 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8674 			rsm->r_flags |= RACK_ACKED;
8675 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8676 			if (rsm->r_in_tmap) {
8677 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8678 				rsm->r_in_tmap = 0;
8679 			}
8680 		}
8681 	} else if (start != end){
8682 		/*
8683 		 * The block was already acked.
8684 		 */
8685 		counter_u64_add(rack_sack_skipped_acked, 1);
8686 		moved++;
8687 	}
8688 out:
8689 	if (rsm &&
8690 	    ((rsm->r_flags & RACK_TLP) == 0) &&
8691 	    (rsm->r_flags & RACK_ACKED)) {
8692 		/*
8693 		 * Now can we merge where we worked
8694 		 * with either the previous or
8695 		 * next block?
8696 		 */
8697 		next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8698 		while (next) {
8699 			if (next->r_flags & RACK_TLP)
8700 				break;
8701 			if (next->r_flags & RACK_ACKED) {
8702 			/* yep this and next can be merged */
8703 				rsm = rack_merge_rsm(rack, rsm, next);
8704 				next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8705 			} else
8706 				break;
8707 		}
8708 		/* Now what about the previous? */
8709 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8710 		while (prev) {
8711 			if (prev->r_flags & RACK_TLP)
8712 				break;
8713 			if (prev->r_flags & RACK_ACKED) {
8714 				/* yep the previous and this can be merged */
8715 				rsm = rack_merge_rsm(rack, prev, rsm);
8716 				prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8717 			} else
8718 				break;
8719 		}
8720 	}
8721 	if (used_ref == 0) {
8722 		counter_u64_add(rack_sack_proc_all, 1);
8723 	} else {
8724 		counter_u64_add(rack_sack_proc_short, 1);
8725 	}
8726 	/* Save off the next one for quick reference. */
8727 	if (rsm)
8728 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8729 	else
8730 		nrsm = NULL;
8731 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
8732 	/* Pass back the moved. */
8733 	*moved_two = moved;
8734 	return (changed);
8735 }
8736 
8737 static void inline
8738 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8739 {
8740 	struct rack_sendmap *tmap;
8741 
8742 	tmap = NULL;
8743 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
8744 		/* Its no longer sacked, mark it so */
8745 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8746 #ifdef INVARIANTS
8747 		if (rsm->r_in_tmap) {
8748 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
8749 			      rack, rsm, rsm->r_flags);
8750 		}
8751 #endif
8752 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8753 		/* Rebuild it into our tmap */
8754 		if (tmap == NULL) {
8755 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8756 			tmap = rsm;
8757 		} else {
8758 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8759 			tmap = rsm;
8760 		}
8761 		tmap->r_in_tmap = 1;
8762 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8763 	}
8764 	/*
8765 	 * Now lets possibly clear the sack filter so we start
8766 	 * recognizing sacks that cover this area.
8767 	 */
8768 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8769 
8770 }
8771 
8772 static void
8773 rack_do_decay(struct tcp_rack *rack)
8774 {
8775 	struct timeval res;
8776 
8777 #define	timersub(tvp, uvp, vvp)						\
8778 	do {								\
8779 		(vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;		\
8780 		(vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;	\
8781 		if ((vvp)->tv_usec < 0) {				\
8782 			(vvp)->tv_sec--;				\
8783 			(vvp)->tv_usec += 1000000;			\
8784 		}							\
8785 	} while (0)
8786 
8787 	timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8788 #undef timersub
8789 
8790 	rack->r_ctl.input_pkt++;
8791 	if ((rack->rc_in_persist) ||
8792 	    (res.tv_sec >= 1) ||
8793 	    (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8794 		/*
8795 		 * Check for decay of non-SAD,
8796 		 * we want all SAD detection metrics to
8797 		 * decay 1/4 per second (or more) passed.
8798 		 */
8799 #ifdef NETFLIX_EXP_DETECTION
8800 		uint32_t pkt_delta;
8801 
8802 		pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8803 #endif
8804 		/* Update our saved tracking values */
8805 		rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8806 		rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8807 		/* Now do we escape without decay? */
8808 #ifdef NETFLIX_EXP_DETECTION
8809 		if (rack->rc_in_persist ||
8810 		    (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8811 		    (pkt_delta < tcp_sad_low_pps)){
8812 			/*
8813 			 * We don't decay idle connections
8814 			 * or ones that have a low input pps.
8815 			 */
8816 			return;
8817 		}
8818 		/* Decay the counters */
8819 		rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8820 							tcp_sad_decay_val);
8821 		rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8822 							 tcp_sad_decay_val);
8823 		rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8824 							       tcp_sad_decay_val);
8825 		rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8826 								tcp_sad_decay_val);
8827 #endif
8828 	}
8829 }
8830 
8831 static void
8832 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8833 {
8834 	struct rack_sendmap *rsm;
8835 #ifdef INVARIANTS
8836 	struct rack_sendmap *rm;
8837 #endif
8838 
8839 	/*
8840 	 * The ACK point is advancing to th_ack, we must drop off
8841 	 * the packets in the rack log and calculate any eligble
8842 	 * RTT's.
8843 	 */
8844 	rack->r_wanted_output = 1;
8845 
8846 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
8847 	if ((rack->rc_last_tlp_acked_set == 1)&&
8848 	    (rack->rc_last_tlp_past_cumack == 1) &&
8849 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
8850 		/*
8851 		 * We have reached the point where our last rack
8852 		 * tlp retransmit sequence is ahead of the cum-ack.
8853 		 * This can only happen when the cum-ack moves all
8854 		 * the way around (its been a full 2^^31+1 bytes
8855 		 * or more since we sent a retransmitted TLP). Lets
8856 		 * turn off the valid flag since its not really valid.
8857 		 *
8858 		 * Note since sack's also turn on this event we have
8859 		 * a complication, we have to wait to age it out until
8860 		 * the cum-ack is by the TLP before checking which is
8861 		 * what the next else clause does.
8862 		 */
8863 		rack_log_dsack_event(rack, 9, __LINE__,
8864 				     rack->r_ctl.last_tlp_acked_start,
8865 				     rack->r_ctl.last_tlp_acked_end);
8866 		rack->rc_last_tlp_acked_set = 0;
8867 		rack->rc_last_tlp_past_cumack = 0;
8868 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
8869 		   (rack->rc_last_tlp_past_cumack == 0) &&
8870 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
8871 		/*
8872 		 * It is safe to start aging TLP's out.
8873 		 */
8874 		rack->rc_last_tlp_past_cumack = 1;
8875 	}
8876 	/* We do the same for the tlp send seq as well */
8877 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8878 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
8879 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
8880 		rack_log_dsack_event(rack, 9, __LINE__,
8881 				     rack->r_ctl.last_sent_tlp_seq,
8882 				     (rack->r_ctl.last_sent_tlp_seq +
8883 				      rack->r_ctl.last_sent_tlp_len));
8884 		rack->rc_last_sent_tlp_seq_valid = 0;
8885 		rack->rc_last_sent_tlp_past_cumack = 0;
8886 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8887 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
8888 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
8889 		/*
8890 		 * It is safe to start aging TLP's send.
8891 		 */
8892 		rack->rc_last_sent_tlp_past_cumack = 1;
8893 	}
8894 more:
8895 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8896 	if (rsm == NULL) {
8897 		if ((th_ack - 1) == tp->iss) {
8898 			/*
8899 			 * For the SYN incoming case we will not
8900 			 * have called tcp_output for the sending of
8901 			 * the SYN, so there will be no map. All
8902 			 * other cases should probably be a panic.
8903 			 */
8904 			return;
8905 		}
8906 		if (tp->t_flags & TF_SENTFIN) {
8907 			/* if we sent a FIN we often will not have map */
8908 			return;
8909 		}
8910 #ifdef INVARIANTS
8911 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8912 		      tp,
8913 		      tp->t_state, th_ack, rack,
8914 		      tp->snd_una, tp->snd_max, tp->snd_nxt);
8915 #endif
8916 		return;
8917 	}
8918 	if (SEQ_LT(th_ack, rsm->r_start)) {
8919 		/* Huh map is missing this */
8920 #ifdef INVARIANTS
8921 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
8922 		       rsm->r_start,
8923 		       th_ack, tp->t_state, rack->r_state);
8924 #endif
8925 		return;
8926 	}
8927 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
8928 
8929 	/* Now was it a retransmitted TLP? */
8930 	if ((rsm->r_flags & RACK_TLP) &&
8931 	    (rsm->r_rtr_cnt > 1)) {
8932 		/*
8933 		 * Yes, this rsm was a TLP and retransmitted, remember that
8934 		 * since if a DSACK comes back on this we don't want
8935 		 * to think of it as a reordered segment. This may
8936 		 * get updated again with possibly even other TLPs
8937 		 * in flight, but thats ok. Only when we don't send
8938 		 * a retransmitted TLP for 1/2 the sequences space
8939 		 * will it get turned off (above).
8940 		 */
8941 		if (rack->rc_last_tlp_acked_set &&
8942 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8943 			/*
8944 			 * We already turned this on since the end matches,
8945 			 * the previous one was a partially ack now we
8946 			 * are getting another one (maybe all of it).
8947 			 */
8948 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8949 			/*
8950 			 * Lets make sure we have all of it though.
8951 			 */
8952 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8953 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8954 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8955 						     rack->r_ctl.last_tlp_acked_end);
8956 			}
8957 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8958 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8959 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8960 						     rack->r_ctl.last_tlp_acked_end);
8961 			}
8962 		} else {
8963 			rack->rc_last_tlp_past_cumack = 1;
8964 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8965 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8966 			rack->rc_last_tlp_acked_set = 1;
8967 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8968 		}
8969 	}
8970 	/* Now do we consume the whole thing? */
8971 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
8972 		/* Its all consumed. */
8973 		uint32_t left;
8974 		uint8_t newly_acked;
8975 
8976 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
8977 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
8978 		rsm->r_rtr_bytes = 0;
8979 		/* Record the time of highest cumack sent */
8980 		rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8981 #ifndef INVARIANTS
8982 		(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8983 #else
8984 		rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8985 		if (rm != rsm) {
8986 			panic("removing head in rack:%p rsm:%p rm:%p",
8987 			      rack, rsm, rm);
8988 		}
8989 #endif
8990 		if (rsm->r_in_tmap) {
8991 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8992 			rsm->r_in_tmap = 0;
8993 		}
8994 		newly_acked = 1;
8995 		if (rsm->r_flags & RACK_ACKED) {
8996 			/*
8997 			 * It was acked on the scoreboard -- remove
8998 			 * it from total
8999 			 */
9000 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
9001 			newly_acked = 0;
9002 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
9003 			/*
9004 			 * There are segments ACKED on the
9005 			 * scoreboard further up. We are seeing
9006 			 * reordering.
9007 			 */
9008 			rsm->r_flags &= ~RACK_SACK_PASSED;
9009 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9010 			rsm->r_flags |= RACK_ACKED;
9011 			rack->r_ctl.rc_reorder_ts = cts;
9012 			if (rack->r_ent_rec_ns) {
9013 				/*
9014 				 * We have sent no more, and we saw an sack
9015 				 * then ack arrive.
9016 				 */
9017 				rack->r_might_revert = 1;
9018 			}
9019 		}
9020 		if ((rsm->r_flags & RACK_TO_REXT) &&
9021 		    (tp->t_flags & TF_RCVD_TSTMP) &&
9022 		    (to->to_flags & TOF_TS) &&
9023 		    (to->to_tsecr != 0) &&
9024 		    (tp->t_flags & TF_PREVVALID)) {
9025 			/*
9026 			 * We can use the timestamp to see
9027 			 * if this retransmission was from the
9028 			 * first transmit. If so we made a mistake.
9029 			 */
9030 			tp->t_flags &= ~TF_PREVVALID;
9031 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
9032 				/* The first transmit is what this ack is for */
9033 				rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
9034 			}
9035 		}
9036 		left = th_ack - rsm->r_end;
9037 		if (rack->app_limited_needs_set && newly_acked)
9038 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
9039 		/* Free back to zone */
9040 		rack_free(rack, rsm);
9041 		if (left) {
9042 			goto more;
9043 		}
9044 		/* Check for reneging */
9045 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9046 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
9047 			/*
9048 			 * The peer has moved snd_una up to
9049 			 * the edge of this send, i.e. one
9050 			 * that it had previously acked. The only
9051 			 * way that can be true if the peer threw
9052 			 * away data (space issues) that it had
9053 			 * previously sacked (else it would have
9054 			 * given us snd_una up to (rsm->r_end).
9055 			 * We need to undo the acked markings here.
9056 			 *
9057 			 * Note we have to look to make sure th_ack is
9058 			 * our rsm->r_start in case we get an old ack
9059 			 * where th_ack is behind snd_una.
9060 			 */
9061 			rack_peer_reneges(rack, rsm, th_ack);
9062 		}
9063 		return;
9064 	}
9065 	if (rsm->r_flags & RACK_ACKED) {
9066 		/*
9067 		 * It was acked on the scoreboard -- remove it from
9068 		 * total for the part being cum-acked.
9069 		 */
9070 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
9071 	}
9072 	/*
9073 	 * Clear the dup ack count for
9074 	 * the piece that remains.
9075 	 */
9076 	rsm->r_dupack = 0;
9077 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9078 	if (rsm->r_rtr_bytes) {
9079 		/*
9080 		 * It was retransmitted adjust the
9081 		 * sack holes for what was acked.
9082 		 */
9083 		int ack_am;
9084 
9085 		ack_am = (th_ack - rsm->r_start);
9086 		if (ack_am >= rsm->r_rtr_bytes) {
9087 			rack->r_ctl.rc_holes_rxt -= ack_am;
9088 			rsm->r_rtr_bytes -= ack_am;
9089 		}
9090 	}
9091 	/*
9092 	 * Update where the piece starts and record
9093 	 * the time of send of highest cumack sent.
9094 	 */
9095 	rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9096 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
9097 	/* Now we need to move our offset forward too */
9098 	if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
9099 		/* Fix up the orig_m_len and possibly the mbuf offset */
9100 		rack_adjust_orig_mlen(rsm);
9101 	}
9102 	rsm->soff += (th_ack - rsm->r_start);
9103 	rsm->r_start = th_ack;
9104 	/* Now do we need to move the mbuf fwd too? */
9105 	if (rsm->m) {
9106 		while (rsm->soff >= rsm->m->m_len) {
9107 			rsm->soff -= rsm->m->m_len;
9108 			rsm->m = rsm->m->m_next;
9109 			KASSERT((rsm->m != NULL),
9110 				(" nrsm:%p hit at soff:%u null m",
9111 				 rsm, rsm->soff));
9112 		}
9113 		rsm->orig_m_len = rsm->m->m_len;
9114 	}
9115 	if (rack->app_limited_needs_set)
9116 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
9117 }
9118 
9119 static void
9120 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
9121 {
9122 	struct rack_sendmap *rsm;
9123 	int sack_pass_fnd = 0;
9124 
9125 	if (rack->r_might_revert) {
9126 		/*
9127 		 * Ok we have reordering, have not sent anything, we
9128 		 * might want to revert the congestion state if nothing
9129 		 * further has SACK_PASSED on it. Lets check.
9130 		 *
9131 		 * We also get here when we have DSACKs come in for
9132 		 * all the data that we FR'd. Note that a rxt or tlp
9133 		 * timer clears this from happening.
9134 		 */
9135 
9136 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
9137 			if (rsm->r_flags & RACK_SACK_PASSED) {
9138 				sack_pass_fnd = 1;
9139 				break;
9140 			}
9141 		}
9142 		if (sack_pass_fnd == 0) {
9143 			/*
9144 			 * We went into recovery
9145 			 * incorrectly due to reordering!
9146 			 */
9147 			int orig_cwnd;
9148 
9149 			rack->r_ent_rec_ns = 0;
9150 			orig_cwnd = tp->snd_cwnd;
9151 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
9152 			tp->snd_recover = tp->snd_una;
9153 			rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
9154 			EXIT_RECOVERY(tp->t_flags);
9155 		}
9156 		rack->r_might_revert = 0;
9157 	}
9158 }
9159 
9160 #ifdef NETFLIX_EXP_DETECTION
9161 static void
9162 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
9163 {
9164 	if ((rack->do_detection || tcp_force_detection) &&
9165 	    tcp_sack_to_ack_thresh &&
9166 	    tcp_sack_to_move_thresh &&
9167 	    ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
9168 		/*
9169 		 * We have thresholds set to find
9170 		 * possible attackers and disable sack.
9171 		 * Check them.
9172 		 */
9173 		uint64_t ackratio, moveratio, movetotal;
9174 
9175 		/* Log detecting */
9176 		rack_log_sad(rack, 1);
9177 		ackratio = (uint64_t)(rack->r_ctl.sack_count);
9178 		ackratio *= (uint64_t)(1000);
9179 		if (rack->r_ctl.ack_count)
9180 			ackratio /= (uint64_t)(rack->r_ctl.ack_count);
9181 		else {
9182 			/* We really should not hit here */
9183 			ackratio = 1000;
9184 		}
9185 		if ((rack->sack_attack_disable == 0) &&
9186 		    (ackratio > rack_highest_sack_thresh_seen))
9187 			rack_highest_sack_thresh_seen = (uint32_t)ackratio;
9188 		movetotal = rack->r_ctl.sack_moved_extra;
9189 		movetotal += rack->r_ctl.sack_noextra_move;
9190 		moveratio = rack->r_ctl.sack_moved_extra;
9191 		moveratio *= (uint64_t)1000;
9192 		if (movetotal)
9193 			moveratio /= movetotal;
9194 		else {
9195 			/* No moves, thats pretty good */
9196 			moveratio = 0;
9197 		}
9198 		if ((rack->sack_attack_disable == 0) &&
9199 		    (moveratio > rack_highest_move_thresh_seen))
9200 			rack_highest_move_thresh_seen = (uint32_t)moveratio;
9201 		if (rack->sack_attack_disable == 0) {
9202 			if ((ackratio > tcp_sack_to_ack_thresh) &&
9203 			    (moveratio > tcp_sack_to_move_thresh)) {
9204 				/* Disable sack processing */
9205 				rack->sack_attack_disable = 1;
9206 				if (rack->r_rep_attack == 0) {
9207 					rack->r_rep_attack = 1;
9208 					counter_u64_add(rack_sack_attacks_detected, 1);
9209 				}
9210 				if (tcp_attack_on_turns_on_logging) {
9211 					/*
9212 					 * Turn on logging, used for debugging
9213 					 * false positives.
9214 					 */
9215 					rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
9216 				}
9217 				/* Clamp the cwnd at flight size */
9218 				rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
9219 				rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9220 				rack_log_sad(rack, 2);
9221 			}
9222 		} else {
9223 			/* We are sack-disabled check for false positives */
9224 			if ((ackratio <= tcp_restoral_thresh) ||
9225 			    (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
9226 				rack->sack_attack_disable = 0;
9227 				rack_log_sad(rack, 3);
9228 				/* Restart counting */
9229 				rack->r_ctl.sack_count = 0;
9230 				rack->r_ctl.sack_moved_extra = 0;
9231 				rack->r_ctl.sack_noextra_move = 1;
9232 				rack->r_ctl.ack_count = max(1,
9233 				      (bytes_this_ack / segsiz));
9234 
9235 				if (rack->r_rep_reverse == 0) {
9236 					rack->r_rep_reverse = 1;
9237 					counter_u64_add(rack_sack_attacks_reversed, 1);
9238 				}
9239 				/* Restore the cwnd */
9240 				if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
9241 					rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
9242 			}
9243 		}
9244 	}
9245 }
9246 #endif
9247 
9248 static int
9249 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
9250 {
9251 
9252 	uint32_t am, l_end;
9253 	int was_tlp = 0;
9254 
9255 	if (SEQ_GT(end, start))
9256 		am = end - start;
9257 	else
9258 		am = 0;
9259 	if ((rack->rc_last_tlp_acked_set ) &&
9260 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
9261 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
9262 		/*
9263 		 * The DSACK is because of a TLP which we don't
9264 		 * do anything with the reordering window over since
9265 		 * it was not reordering that caused the DSACK but
9266 		 * our previous retransmit TLP.
9267 		 */
9268 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
9269 		was_tlp = 1;
9270 		goto skip_dsack_round;
9271 	}
9272 	if (rack->rc_last_sent_tlp_seq_valid) {
9273 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
9274 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
9275 		    (SEQ_LEQ(end, l_end))) {
9276 			/*
9277 			 * This dsack is from the last sent TLP, ignore it
9278 			 * for reordering purposes.
9279 			 */
9280 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
9281 			was_tlp = 1;
9282 			goto skip_dsack_round;
9283 		}
9284 	}
9285 	if (rack->rc_dsack_round_seen == 0) {
9286 		rack->rc_dsack_round_seen = 1;
9287 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
9288 		rack->r_ctl.num_dsack++;
9289 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
9290 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
9291 	}
9292 skip_dsack_round:
9293 	/*
9294 	 * We keep track of how many DSACK blocks we get
9295 	 * after a recovery incident.
9296 	 */
9297 	rack->r_ctl.dsack_byte_cnt += am;
9298 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
9299 	    rack->r_ctl.retran_during_recovery &&
9300 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
9301 		/*
9302 		 * False recovery most likely culprit is reordering. If
9303 		 * nothing else is missing we need to revert.
9304 		 */
9305 		rack->r_might_revert = 1;
9306 		rack_handle_might_revert(rack->rc_tp, rack);
9307 		rack->r_might_revert = 0;
9308 		rack->r_ctl.retran_during_recovery = 0;
9309 		rack->r_ctl.dsack_byte_cnt = 0;
9310 	}
9311 	return (was_tlp);
9312 }
9313 
9314 static uint32_t
9315 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
9316 {
9317 	return (((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt);
9318 }
9319 
9320 static int32_t
9321 rack_compute_pipe(struct tcpcb *tp)
9322 {
9323 	return ((int32_t)do_rack_compute_pipe(tp,
9324 					      (struct tcp_rack *)tp->t_fb_ptr,
9325 					      tp->snd_una));
9326 }
9327 
9328 static void
9329 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9330 {
9331 	/* Deal with changed and PRR here (in recovery only) */
9332 	uint32_t pipe, snd_una;
9333 
9334 	rack->r_ctl.rc_prr_delivered += changed;
9335 
9336 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9337 		/*
9338 		 * It is all outstanding, we are application limited
9339 		 * and thus we don't need more room to send anything.
9340 		 * Note we use tp->snd_una here and not th_ack because
9341 		 * the data as yet not been cut from the sb.
9342 		 */
9343 		rack->r_ctl.rc_prr_sndcnt = 0;
9344 		return;
9345 	}
9346 	/* Compute prr_sndcnt */
9347 	if (SEQ_GT(tp->snd_una, th_ack)) {
9348 		snd_una = tp->snd_una;
9349 	} else {
9350 		snd_una = th_ack;
9351 	}
9352 	pipe = do_rack_compute_pipe(tp, rack, snd_una);
9353 	if (pipe > tp->snd_ssthresh) {
9354 		long sndcnt;
9355 
9356 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9357 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
9358 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9359 		else {
9360 			rack->r_ctl.rc_prr_sndcnt = 0;
9361 			rack_log_to_prr(rack, 9, 0, __LINE__);
9362 			sndcnt = 0;
9363 		}
9364 		sndcnt++;
9365 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9366 			sndcnt -= rack->r_ctl.rc_prr_out;
9367 		else
9368 			sndcnt = 0;
9369 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
9370 		rack_log_to_prr(rack, 10, 0, __LINE__);
9371 	} else {
9372 		uint32_t limit;
9373 
9374 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9375 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9376 		else
9377 			limit = 0;
9378 		if (changed > limit)
9379 			limit = changed;
9380 		limit += ctf_fixed_maxseg(tp);
9381 		if (tp->snd_ssthresh > pipe) {
9382 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9383 			rack_log_to_prr(rack, 11, 0, __LINE__);
9384 		} else {
9385 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9386 			rack_log_to_prr(rack, 12, 0, __LINE__);
9387 		}
9388 	}
9389 }
9390 
9391 static void
9392 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9393 {
9394 	uint32_t changed;
9395 	struct tcp_rack *rack;
9396 	struct rack_sendmap *rsm;
9397 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9398 	register uint32_t th_ack;
9399 	int32_t i, j, k, num_sack_blks = 0;
9400 	uint32_t cts, acked, ack_point;
9401 	int loop_start = 0, moved_two = 0;
9402 	uint32_t tsused;
9403 
9404 
9405 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9406 	if (tcp_get_flags(th) & TH_RST) {
9407 		/* We don't log resets */
9408 		return;
9409 	}
9410 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9411 	cts = tcp_get_usecs(NULL);
9412 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9413 	changed = 0;
9414 	th_ack = th->th_ack;
9415 	if (rack->sack_attack_disable == 0)
9416 		rack_do_decay(rack);
9417 	if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9418 		/*
9419 		 * You only get credit for
9420 		 * MSS and greater (and you get extra
9421 		 * credit for larger cum-ack moves).
9422 		 */
9423 		int ac;
9424 
9425 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9426 		rack->r_ctl.ack_count += ac;
9427 		counter_u64_add(rack_ack_total, ac);
9428 	}
9429 	if (rack->r_ctl.ack_count > 0xfff00000) {
9430 		/*
9431 		 * reduce the number to keep us under
9432 		 * a uint32_t.
9433 		 */
9434 		rack->r_ctl.ack_count /= 2;
9435 		rack->r_ctl.sack_count /= 2;
9436 	}
9437 	if (SEQ_GT(th_ack, tp->snd_una)) {
9438 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9439 		tp->t_acktime = ticks;
9440 	}
9441 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
9442 		changed = th_ack - rsm->r_start;
9443 	if (changed) {
9444 		rack_process_to_cumack(tp, rack, th_ack, cts, to);
9445 	}
9446 	if ((to->to_flags & TOF_SACK) == 0) {
9447 		/* We are done nothing left and no sack. */
9448 		rack_handle_might_revert(tp, rack);
9449 		/*
9450 		 * For cases where we struck a dup-ack
9451 		 * with no SACK, add to the changes so
9452 		 * PRR will work right.
9453 		 */
9454 		if (dup_ack_struck && (changed == 0)) {
9455 			changed += ctf_fixed_maxseg(rack->rc_tp);
9456 		}
9457 		goto out;
9458 	}
9459 	/* Sack block processing */
9460 	if (SEQ_GT(th_ack, tp->snd_una))
9461 		ack_point = th_ack;
9462 	else
9463 		ack_point = tp->snd_una;
9464 	for (i = 0; i < to->to_nsacks; i++) {
9465 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
9466 		      &sack, sizeof(sack));
9467 		sack.start = ntohl(sack.start);
9468 		sack.end = ntohl(sack.end);
9469 		if (SEQ_GT(sack.end, sack.start) &&
9470 		    SEQ_GT(sack.start, ack_point) &&
9471 		    SEQ_LT(sack.start, tp->snd_max) &&
9472 		    SEQ_GT(sack.end, ack_point) &&
9473 		    SEQ_LEQ(sack.end, tp->snd_max)) {
9474 			sack_blocks[num_sack_blks] = sack;
9475 			num_sack_blks++;
9476 		} else if (SEQ_LEQ(sack.start, th_ack) &&
9477 			   SEQ_LEQ(sack.end, th_ack)) {
9478 			int was_tlp;
9479 
9480 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
9481 			/*
9482 			 * Its a D-SACK block.
9483 			 */
9484 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
9485 		}
9486 	}
9487 	if (rack->rc_dsack_round_seen) {
9488 		/* Is the dsack roound over? */
9489 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
9490 			/* Yes it is */
9491 			rack->rc_dsack_round_seen = 0;
9492 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
9493 		}
9494 	}
9495 	/*
9496 	 * Sort the SACK blocks so we can update the rack scoreboard with
9497 	 * just one pass.
9498 	 */
9499 	num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9500 					 num_sack_blks, th->th_ack);
9501 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9502 	if (num_sack_blks == 0) {
9503 		/* Nothing to sack (DSACKs?) */
9504 		goto out_with_totals;
9505 	}
9506 	if (num_sack_blks < 2) {
9507 		/* Only one, we don't need to sort */
9508 		goto do_sack_work;
9509 	}
9510 	/* Sort the sacks */
9511 	for (i = 0; i < num_sack_blks; i++) {
9512 		for (j = i + 1; j < num_sack_blks; j++) {
9513 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9514 				sack = sack_blocks[i];
9515 				sack_blocks[i] = sack_blocks[j];
9516 				sack_blocks[j] = sack;
9517 			}
9518 		}
9519 	}
9520 	/*
9521 	 * Now are any of the sack block ends the same (yes some
9522 	 * implementations send these)?
9523 	 */
9524 again:
9525 	if (num_sack_blks == 0)
9526 		goto out_with_totals;
9527 	if (num_sack_blks > 1) {
9528 		for (i = 0; i < num_sack_blks; i++) {
9529 			for (j = i + 1; j < num_sack_blks; j++) {
9530 				if (sack_blocks[i].end == sack_blocks[j].end) {
9531 					/*
9532 					 * Ok these two have the same end we
9533 					 * want the smallest end and then
9534 					 * throw away the larger and start
9535 					 * again.
9536 					 */
9537 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9538 						/*
9539 						 * The second block covers
9540 						 * more area use that
9541 						 */
9542 						sack_blocks[i].start = sack_blocks[j].start;
9543 					}
9544 					/*
9545 					 * Now collapse out the dup-sack and
9546 					 * lower the count
9547 					 */
9548 					for (k = (j + 1); k < num_sack_blks; k++) {
9549 						sack_blocks[j].start = sack_blocks[k].start;
9550 						sack_blocks[j].end = sack_blocks[k].end;
9551 						j++;
9552 					}
9553 					num_sack_blks--;
9554 					goto again;
9555 				}
9556 			}
9557 		}
9558 	}
9559 do_sack_work:
9560 	/*
9561 	 * First lets look to see if
9562 	 * we have retransmitted and
9563 	 * can use the transmit next?
9564 	 */
9565 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9566 	if (rsm &&
9567 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9568 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9569 		/*
9570 		 * We probably did the FR and the next
9571 		 * SACK in continues as we would expect.
9572 		 */
9573 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9574 		if (acked) {
9575 			rack->r_wanted_output = 1;
9576 			changed += acked;
9577 		}
9578 		if (num_sack_blks == 1) {
9579 			/*
9580 			 * This is what we would expect from
9581 			 * a normal implementation to happen
9582 			 * after we have retransmitted the FR,
9583 			 * i.e the sack-filter pushes down
9584 			 * to 1 block and the next to be retransmitted
9585 			 * is the sequence in the sack block (has more
9586 			 * are acked). Count this as ACK'd data to boost
9587 			 * up the chances of recovering any false positives.
9588 			 */
9589 			rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9590 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9591 			counter_u64_add(rack_express_sack, 1);
9592 			if (rack->r_ctl.ack_count > 0xfff00000) {
9593 				/*
9594 				 * reduce the number to keep us under
9595 				 * a uint32_t.
9596 				 */
9597 				rack->r_ctl.ack_count /= 2;
9598 				rack->r_ctl.sack_count /= 2;
9599 			}
9600 			goto out_with_totals;
9601 		} else {
9602 			/*
9603 			 * Start the loop through the
9604 			 * rest of blocks, past the first block.
9605 			 */
9606 			moved_two = 0;
9607 			loop_start = 1;
9608 		}
9609 	}
9610 	/* Its a sack of some sort */
9611 	rack->r_ctl.sack_count++;
9612 	if (rack->r_ctl.sack_count > 0xfff00000) {
9613 		/*
9614 		 * reduce the number to keep us under
9615 		 * a uint32_t.
9616 		 */
9617 		rack->r_ctl.ack_count /= 2;
9618 		rack->r_ctl.sack_count /= 2;
9619 	}
9620 	counter_u64_add(rack_sack_total, 1);
9621 	if (rack->sack_attack_disable) {
9622 		/* An attacker disablement is in place */
9623 		if (num_sack_blks > 1) {
9624 			rack->r_ctl.sack_count += (num_sack_blks - 1);
9625 			rack->r_ctl.sack_moved_extra++;
9626 			counter_u64_add(rack_move_some, 1);
9627 			if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9628 				rack->r_ctl.sack_moved_extra /= 2;
9629 				rack->r_ctl.sack_noextra_move /= 2;
9630 			}
9631 		}
9632 		goto out;
9633 	}
9634 	rsm = rack->r_ctl.rc_sacklast;
9635 	for (i = loop_start; i < num_sack_blks; i++) {
9636 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9637 		if (acked) {
9638 			rack->r_wanted_output = 1;
9639 			changed += acked;
9640 		}
9641 		if (moved_two) {
9642 			/*
9643 			 * If we did not get a SACK for at least a MSS and
9644 			 * had to move at all, or if we moved more than our
9645 			 * threshold, it counts against the "extra" move.
9646 			 */
9647 			rack->r_ctl.sack_moved_extra += moved_two;
9648 			counter_u64_add(rack_move_some, 1);
9649 		} else {
9650 			/*
9651 			 * else we did not have to move
9652 			 * any more than we would expect.
9653 			 */
9654 			rack->r_ctl.sack_noextra_move++;
9655 			counter_u64_add(rack_move_none, 1);
9656 		}
9657 		if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9658 			/*
9659 			 * If the SACK was not a full MSS then
9660 			 * we add to sack_count the number of
9661 			 * MSS's (or possibly more than
9662 			 * a MSS if its a TSO send) we had to skip by.
9663 			 */
9664 			rack->r_ctl.sack_count += moved_two;
9665 			counter_u64_add(rack_sack_total, moved_two);
9666 		}
9667 		/*
9668 		 * Now we need to setup for the next
9669 		 * round. First we make sure we won't
9670 		 * exceed the size of our uint32_t on
9671 		 * the various counts, and then clear out
9672 		 * moved_two.
9673 		 */
9674 		if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9675 		    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9676 			rack->r_ctl.sack_moved_extra /= 2;
9677 			rack->r_ctl.sack_noextra_move /= 2;
9678 		}
9679 		if (rack->r_ctl.sack_count > 0xfff00000) {
9680 			rack->r_ctl.ack_count /= 2;
9681 			rack->r_ctl.sack_count /= 2;
9682 		}
9683 		moved_two = 0;
9684 	}
9685 out_with_totals:
9686 	if (num_sack_blks > 1) {
9687 		/*
9688 		 * You get an extra stroke if
9689 		 * you have more than one sack-blk, this
9690 		 * could be where we are skipping forward
9691 		 * and the sack-filter is still working, or
9692 		 * it could be an attacker constantly
9693 		 * moving us.
9694 		 */
9695 		rack->r_ctl.sack_moved_extra++;
9696 		counter_u64_add(rack_move_some, 1);
9697 	}
9698 out:
9699 #ifdef NETFLIX_EXP_DETECTION
9700 	rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9701 #endif
9702 	if (changed) {
9703 		/* Something changed cancel the rack timer */
9704 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9705 	}
9706 	tsused = tcp_get_usecs(NULL);
9707 	rsm = tcp_rack_output(tp, rack, tsused);
9708 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9709 	    rsm &&
9710 	    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
9711 		/* Enter recovery */
9712 		entered_recovery = 1;
9713 		rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
9714 		/*
9715 		 * When we enter recovery we need to assure we send
9716 		 * one packet.
9717 		 */
9718 		if (rack->rack_no_prr == 0) {
9719 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9720 			rack_log_to_prr(rack, 8, 0, __LINE__);
9721 		}
9722 		rack->r_timer_override = 1;
9723 		rack->r_early = 0;
9724 		rack->r_ctl.rc_agg_early = 0;
9725 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
9726 		   rsm &&
9727 		   (rack->r_rr_config == 3)) {
9728 		/*
9729 		 * Assure we can output and we get no
9730 		 * remembered pace time except the retransmit.
9731 		 */
9732 		rack->r_timer_override = 1;
9733 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9734 		rack->r_ctl.rc_resend = rsm;
9735 	}
9736 	if (IN_FASTRECOVERY(tp->t_flags) &&
9737 	    (rack->rack_no_prr == 0) &&
9738 	    (entered_recovery == 0)) {
9739 		rack_update_prr(tp, rack, changed, th_ack);
9740 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9741 		     ((tcp_in_hpts(rack->rc_inp) == 0) &&
9742 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9743 			/*
9744 			 * If you are pacing output you don't want
9745 			 * to override.
9746 			 */
9747 			rack->r_early = 0;
9748 			rack->r_ctl.rc_agg_early = 0;
9749 			rack->r_timer_override = 1;
9750 		}
9751 	}
9752 }
9753 
9754 static void
9755 rack_strike_dupack(struct tcp_rack *rack)
9756 {
9757 	struct rack_sendmap *rsm;
9758 
9759 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9760 	while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9761 		rsm = TAILQ_NEXT(rsm, r_tnext);
9762 		if (rsm->r_flags & RACK_MUST_RXT) {
9763 			/* Sendmap entries that are marked to
9764 			 * be retransmitted do not need dupack's
9765 			 * struck. We get these marks for a number
9766 			 * of reasons (rxt timeout with no sack,
9767 			 * mtu change, or rwnd collapses). When
9768 			 * these events occur, we know we must retransmit
9769 			 * them and mark the sendmap entries. Dupack counting
9770 			 * is not needed since we are already set to retransmit
9771 			 * it as soon as we can.
9772 			 */
9773 			continue;
9774 		}
9775 	}
9776 	if (rsm && (rsm->r_dupack < 0xff)) {
9777 		rsm->r_dupack++;
9778 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9779 			struct timeval tv;
9780 			uint32_t cts;
9781 			/*
9782 			 * Here we see if we need to retransmit. For
9783 			 * a SACK type connection if enough time has passed
9784 			 * we will get a return of the rsm. For a non-sack
9785 			 * connection we will get the rsm returned if the
9786 			 * dupack value is 3 or more.
9787 			 */
9788 			cts = tcp_get_usecs(&tv);
9789 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9790 			if (rack->r_ctl.rc_resend != NULL) {
9791 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9792 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9793 							 rack->rc_tp->snd_una, __LINE__);
9794 				}
9795 				rack->r_wanted_output = 1;
9796 				rack->r_timer_override = 1;
9797 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9798 			}
9799 		} else {
9800 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9801 		}
9802 	}
9803 }
9804 
9805 static void
9806 rack_check_bottom_drag(struct tcpcb *tp,
9807 		       struct tcp_rack *rack,
9808 		       struct socket *so, int32_t acked)
9809 {
9810 	uint32_t segsiz, minseg;
9811 
9812 	segsiz = ctf_fixed_maxseg(tp);
9813 	minseg = segsiz;
9814 
9815 	if (tp->snd_max == tp->snd_una) {
9816 		/*
9817 		 * We are doing dynamic pacing and we are way
9818 		 * under. Basically everything got acked while
9819 		 * we were still waiting on the pacer to expire.
9820 		 *
9821 		 * This means we need to boost the b/w in
9822 		 * addition to any earlier boosting of
9823 		 * the multiplier.
9824 		 */
9825 		rack->rc_dragged_bottom = 1;
9826 		rack_validate_multipliers_at_or_above100(rack);
9827 		/*
9828 		 * Lets use the segment bytes acked plus
9829 		 * the lowest RTT seen as the basis to
9830 		 * form a b/w estimate. This will be off
9831 		 * due to the fact that the true estimate
9832 		 * should be around 1/2 the time of the RTT
9833 		 * but we can settle for that.
9834 		 */
9835 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9836 		    acked) {
9837 			uint64_t bw, calc_bw, rtt;
9838 
9839 			rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9840 			if (rtt == 0) {
9841 				/* no us sample is there a ms one? */
9842 				if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9843 					rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9844 				} else {
9845 					goto no_measurement;
9846 				}
9847 			}
9848 			bw = acked;
9849 			calc_bw = bw * 1000000;
9850 			calc_bw /= rtt;
9851 			if (rack->r_ctl.last_max_bw &&
9852 			    (rack->r_ctl.last_max_bw < calc_bw)) {
9853 				/*
9854 				 * If we have a last calculated max bw
9855 				 * enforce it.
9856 				 */
9857 				calc_bw = rack->r_ctl.last_max_bw;
9858 			}
9859 			/* now plop it in */
9860 			if (rack->rc_gp_filled == 0) {
9861 				if (calc_bw > ONE_POINT_TWO_MEG) {
9862 					/*
9863 					 * If we have no measurement
9864 					 * don't let us set in more than
9865 					 * 1.2Mbps. If we are still too
9866 					 * low after pacing with this we
9867 					 * will hopefully have a max b/w
9868 					 * available to sanity check things.
9869 					 */
9870 					calc_bw = ONE_POINT_TWO_MEG;
9871 				}
9872 				rack->r_ctl.rc_rtt_diff = 0;
9873 				rack->r_ctl.gp_bw = calc_bw;
9874 				rack->rc_gp_filled = 1;
9875 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9876 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
9877 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9878 			} else if (calc_bw > rack->r_ctl.gp_bw) {
9879 				rack->r_ctl.rc_rtt_diff = 0;
9880 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9881 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
9882 				rack->r_ctl.gp_bw = calc_bw;
9883 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9884 			} else
9885 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
9886 			if ((rack->gp_ready == 0) &&
9887 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9888 				/* We have enough measurements now */
9889 				rack->gp_ready = 1;
9890 				rack_set_cc_pacing(rack);
9891 				if (rack->defer_options)
9892 					rack_apply_deferred_options(rack);
9893 			}
9894 			/*
9895 			 * For acks over 1mss we do a extra boost to simulate
9896 			 * where we would get 2 acks (we want 110 for the mul).
9897 			 */
9898 			if (acked > segsiz)
9899 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
9900 		} else {
9901 			/*
9902 			 * zero rtt possibly?, settle for just an old increase.
9903 			 */
9904 no_measurement:
9905 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
9906 		}
9907 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9908 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9909 					       minseg)) &&
9910 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9911 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9912 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9913 		    (segsiz * rack_req_segs))) {
9914 		/*
9915 		 * We are doing dynamic GP pacing and
9916 		 * we have everything except 1MSS or less
9917 		 * bytes left out. We are still pacing away.
9918 		 * And there is data that could be sent, This
9919 		 * means we are inserting delayed ack time in
9920 		 * our measurements because we are pacing too slow.
9921 		 */
9922 		rack_validate_multipliers_at_or_above100(rack);
9923 		rack->rc_dragged_bottom = 1;
9924 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
9925 	}
9926 }
9927 
9928 
9929 
9930 static void
9931 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
9932 {
9933 	/*
9934 	 * The fast output path is enabled and we
9935 	 * have moved the cumack forward. Lets see if
9936 	 * we can expand forward the fast path length by
9937 	 * that amount. What we would ideally like to
9938 	 * do is increase the number of bytes in the
9939 	 * fast path block (left_to_send) by the
9940 	 * acked amount. However we have to gate that
9941 	 * by two factors:
9942 	 * 1) The amount outstanding and the rwnd of the peer
9943 	 *    (i.e. we don't want to exceed the rwnd of the peer).
9944 	 *    <and>
9945 	 * 2) The amount of data left in the socket buffer (i.e.
9946 	 *    we can't send beyond what is in the buffer).
9947 	 *
9948 	 * Note that this does not take into account any increase
9949 	 * in the cwnd. We will only extend the fast path by
9950 	 * what was acked.
9951 	 */
9952 	uint32_t new_total, gating_val;
9953 
9954 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
9955 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
9956 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
9957 	if (new_total <= gating_val) {
9958 		/* We can increase left_to_send by the acked amount */
9959 		counter_u64_add(rack_extended_rfo, 1);
9960 		rack->r_ctl.fsb.left_to_send = new_total;
9961 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
9962 			("rack:%p left_to_send:%u sbavail:%u out:%u",
9963 			 rack, rack->r_ctl.fsb.left_to_send,
9964 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
9965 			 (tp->snd_max - tp->snd_una)));
9966 
9967 	}
9968 }
9969 
9970 static void
9971 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
9972 {
9973 	/*
9974 	 * Here any sendmap entry that points to the
9975 	 * beginning mbuf must be adjusted to the correct
9976 	 * offset. This must be called with:
9977 	 * 1) The socket buffer locked
9978 	 * 2) snd_una adjusted to its new position.
9979 	 *
9980 	 * Note that (2) implies rack_ack_received has also
9981 	 * been called.
9982 	 *
9983 	 * We grab the first mbuf in the socket buffer and
9984 	 * then go through the front of the sendmap, recalculating
9985 	 * the stored offset for any sendmap entry that has
9986 	 * that mbuf. We must use the sb functions to do this
9987 	 * since its possible an add was done has well as
9988 	 * the subtraction we may have just completed. This should
9989 	 * not be a penalty though, since we just referenced the sb
9990 	 * to go in and trim off the mbufs that we freed (of course
9991 	 * there will be a penalty for the sendmap references though).
9992 	 */
9993 	struct mbuf *m;
9994 	struct rack_sendmap *rsm;
9995 
9996 	SOCKBUF_LOCK_ASSERT(sb);
9997 	m = sb->sb_mb;
9998 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9999 	if ((rsm == NULL) || (m == NULL)) {
10000 		/* Nothing outstanding */
10001 		return;
10002 	}
10003 	while (rsm->m && (rsm->m == m)) {
10004 		/* one to adjust */
10005 #ifdef INVARIANTS
10006 		struct mbuf *tm;
10007 		uint32_t soff;
10008 
10009 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
10010 		if (rsm->orig_m_len != m->m_len) {
10011 			rack_adjust_orig_mlen(rsm);
10012 		}
10013 		if (rsm->soff != soff) {
10014 			/*
10015 			 * This is not a fatal error, we anticipate it
10016 			 * might happen (the else code), so we count it here
10017 			 * so that under invariant we can see that it really
10018 			 * does happen.
10019 			 */
10020 			counter_u64_add(rack_adjust_map_bw, 1);
10021 		}
10022 		rsm->m = tm;
10023 		rsm->soff = soff;
10024 		if (tm)
10025 			rsm->orig_m_len = rsm->m->m_len;
10026 		else
10027 			rsm->orig_m_len = 0;
10028 #else
10029 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
10030 		if (rsm->m)
10031 			rsm->orig_m_len = rsm->m->m_len;
10032 		else
10033 			rsm->orig_m_len = 0;
10034 #endif
10035 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
10036 			      rsm);
10037 		if (rsm == NULL)
10038 			break;
10039 	}
10040 }
10041 
10042 /*
10043  * Return value of 1, we do not need to call rack_process_data().
10044  * return value of 0, rack_process_data can be called.
10045  * For ret_val if its 0 the TCP is locked, if its non-zero
10046  * its unlocked and probably unsafe to touch the TCB.
10047  */
10048 static int
10049 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10050     struct tcpcb *tp, struct tcpopt *to,
10051     uint32_t tiwin, int32_t tlen,
10052     int32_t * ofia, int32_t thflags, int32_t *ret_val)
10053 {
10054 	int32_t ourfinisacked = 0;
10055 	int32_t nsegs, acked_amount;
10056 	int32_t acked;
10057 	struct mbuf *mfree;
10058 	struct tcp_rack *rack;
10059 	int32_t under_pacing = 0;
10060 	int32_t recovery = 0;
10061 
10062 	INP_WLOCK_ASSERT(tptoinpcb(tp));
10063 
10064 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10065 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
10066 		__ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
10067 				      &rack->r_ctl.challenge_ack_ts,
10068 				      &rack->r_ctl.challenge_ack_cnt);
10069 		rack->r_wanted_output = 1;
10070 		return (1);
10071 	}
10072 	if (rack->gp_ready &&
10073 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10074 		under_pacing = 1;
10075 	}
10076 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
10077 		int in_rec, dup_ack_struck = 0;
10078 
10079 		in_rec = IN_FASTRECOVERY(tp->t_flags);
10080 		if (rack->rc_in_persist) {
10081 			tp->t_rxtshift = 0;
10082 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10083 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10084 		}
10085 		if ((th->th_ack == tp->snd_una) &&
10086 		    (tiwin == tp->snd_wnd) &&
10087 		    ((to->to_flags & TOF_SACK) == 0)) {
10088 			rack_strike_dupack(rack);
10089 			dup_ack_struck = 1;
10090 		}
10091 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
10092 	}
10093 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10094 		/*
10095 		 * Old ack, behind (or duplicate to) the last one rcv'd
10096 		 * Note: We mark reordering is occuring if its
10097 		 * less than and we have not closed our window.
10098 		 */
10099 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
10100 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
10101 		}
10102 		return (0);
10103 	}
10104 	/*
10105 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
10106 	 * something we sent.
10107 	 */
10108 	if (tp->t_flags & TF_NEEDSYN) {
10109 		/*
10110 		 * T/TCP: Connection was half-synchronized, and our SYN has
10111 		 * been ACK'd (so connection is now fully synchronized).  Go
10112 		 * to non-starred state, increment snd_una for ACK of SYN,
10113 		 * and check if we can do window scaling.
10114 		 */
10115 		tp->t_flags &= ~TF_NEEDSYN;
10116 		tp->snd_una++;
10117 		/* Do window scaling? */
10118 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10119 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10120 			tp->rcv_scale = tp->request_r_scale;
10121 			/* Send window already scaled. */
10122 		}
10123 	}
10124 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10125 
10126 	acked = BYTES_THIS_ACK(tp, th);
10127 	if (acked) {
10128 		/*
10129 		 * Any time we move the cum-ack forward clear
10130 		 * keep-alive tied probe-not-answered. The
10131 		 * persists clears its own on entry.
10132 		 */
10133 		rack->probe_not_answered = 0;
10134 	}
10135 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10136 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10137 	/*
10138 	 * If we just performed our first retransmit, and the ACK arrives
10139 	 * within our recovery window, then it was a mistake to do the
10140 	 * retransmit in the first place.  Recover our original cwnd and
10141 	 * ssthresh, and proceed to transmit where we left off.
10142 	 */
10143 	if ((tp->t_flags & TF_PREVVALID) &&
10144 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10145 		tp->t_flags &= ~TF_PREVVALID;
10146 		if (tp->t_rxtshift == 1 &&
10147 		    (int)(ticks - tp->t_badrxtwin) < 0)
10148 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
10149 	}
10150 	if (acked) {
10151 		/* assure we are not backed off */
10152 		tp->t_rxtshift = 0;
10153 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10154 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10155 		rack->rc_tlp_in_progress = 0;
10156 		rack->r_ctl.rc_tlp_cnt_out = 0;
10157 		/*
10158 		 * If it is the RXT timer we want to
10159 		 * stop it, so we can restart a TLP.
10160 		 */
10161 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10162 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10163 #ifdef NETFLIX_HTTP_LOGGING
10164 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10165 #endif
10166 	}
10167 	/*
10168 	 * If we have a timestamp reply, update smoothed round trip time. If
10169 	 * no timestamp is present but transmit timer is running and timed
10170 	 * sequence number was acked, update smoothed round trip time. Since
10171 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
10172 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
10173 	 * timer.
10174 	 *
10175 	 * Some boxes send broken timestamp replies during the SYN+ACK
10176 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
10177 	 * and blow up the retransmit timer.
10178 	 */
10179 	/*
10180 	 * If all outstanding data is acked, stop retransmit timer and
10181 	 * remember to restart (more output or persist). If there is more
10182 	 * data to be acked, restart retransmit timer, using current
10183 	 * (possibly backed-off) value.
10184 	 */
10185 	if (acked == 0) {
10186 		if (ofia)
10187 			*ofia = ourfinisacked;
10188 		return (0);
10189 	}
10190 	if (IN_RECOVERY(tp->t_flags)) {
10191 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
10192 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
10193 			tcp_rack_partialack(tp);
10194 		} else {
10195 			rack_post_recovery(tp, th->th_ack);
10196 			recovery = 1;
10197 		}
10198 	}
10199 	/*
10200 	 * Let the congestion control algorithm update congestion control
10201 	 * related information. This typically means increasing the
10202 	 * congestion window.
10203 	 */
10204 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
10205 	SOCKBUF_LOCK(&so->so_snd);
10206 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
10207 	tp->snd_wnd -= acked_amount;
10208 	mfree = sbcut_locked(&so->so_snd, acked_amount);
10209 	if ((sbused(&so->so_snd) == 0) &&
10210 	    (acked > acked_amount) &&
10211 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
10212 	    (tp->t_flags & TF_SENTFIN)) {
10213 		/*
10214 		 * We must be sure our fin
10215 		 * was sent and acked (we can be
10216 		 * in FIN_WAIT_1 without having
10217 		 * sent the fin).
10218 		 */
10219 		ourfinisacked = 1;
10220 	}
10221 	tp->snd_una = th->th_ack;
10222 	if (acked_amount && sbavail(&so->so_snd))
10223 		rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10224 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10225 	/* NB: sowwakeup_locked() does an implicit unlock. */
10226 	sowwakeup_locked(so);
10227 	m_freem(mfree);
10228 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
10229 		tp->snd_recover = tp->snd_una;
10230 
10231 	if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
10232 		tp->snd_nxt = tp->snd_una;
10233 	}
10234 	if (under_pacing &&
10235 	    (rack->use_fixed_rate == 0) &&
10236 	    (rack->in_probe_rtt == 0) &&
10237 	    rack->rc_gp_dyn_mul &&
10238 	    rack->rc_always_pace) {
10239 		/* Check if we are dragging bottom */
10240 		rack_check_bottom_drag(tp, rack, so, acked);
10241 	}
10242 	if (tp->snd_una == tp->snd_max) {
10243 		/* Nothing left outstanding */
10244 		tp->t_flags &= ~TF_PREVVALID;
10245 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10246 		rack->r_ctl.retran_during_recovery = 0;
10247 		rack->r_ctl.dsack_byte_cnt = 0;
10248 		if (rack->r_ctl.rc_went_idle_time == 0)
10249 			rack->r_ctl.rc_went_idle_time = 1;
10250 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10251 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
10252 			tp->t_acktime = 0;
10253 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10254 		/* Set need output so persist might get set */
10255 		rack->r_wanted_output = 1;
10256 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10257 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
10258 		    (sbavail(&so->so_snd) == 0) &&
10259 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
10260 			/*
10261 			 * The socket was gone and the
10262 			 * peer sent data (now or in the past), time to
10263 			 * reset him.
10264 			 */
10265 			*ret_val = 1;
10266 			/* tcp_close will kill the inp pre-log the Reset */
10267 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
10268 			tp = tcp_close(tp);
10269 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
10270 			return (1);
10271 		}
10272 	}
10273 	if (ofia)
10274 		*ofia = ourfinisacked;
10275 	return (0);
10276 }
10277 
10278 
10279 static void
10280 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
10281 		  int dir, uint32_t flags, struct rack_sendmap *rsm)
10282 {
10283 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
10284 		union tcp_log_stackspecific log;
10285 		struct timeval tv;
10286 
10287 		memset(&log, 0, sizeof(log));
10288 		log.u_bbr.flex1 = cnt;
10289 		log.u_bbr.flex2 = split;
10290 		log.u_bbr.flex3 = out;
10291 		log.u_bbr.flex4 = line;
10292 		log.u_bbr.flex5 = rack->r_must_retran;
10293 		log.u_bbr.flex6 = flags;
10294 		log.u_bbr.flex7 = rack->rc_has_collapsed;
10295 		log.u_bbr.flex8 = dir;	/*
10296 					 * 1 is collapsed, 0 is uncollapsed,
10297 					 * 2 is log of a rsm being marked, 3 is a split.
10298 					 */
10299 		if (rsm == NULL)
10300 			log.u_bbr.rttProp = 0;
10301 		else
10302 			log.u_bbr.rttProp = (uint64_t)rsm;
10303 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
10304 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10305 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
10306 		    &rack->rc_inp->inp_socket->so_rcv,
10307 		    &rack->rc_inp->inp_socket->so_snd,
10308 		    TCP_RACK_LOG_COLLAPSE, 0,
10309 		    0, &log, false, &tv);
10310 	}
10311 }
10312 
10313 static void
10314 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, int line)
10315 {
10316 	/*
10317 	 * Here all we do is mark the collapsed point and set the flag.
10318 	 * This may happen again and again, but there is no
10319 	 * sense splitting our map until we know where the
10320 	 * peer finally lands in the collapse.
10321 	 */
10322 	rack_trace_point(rack, RACK_TP_COLLAPSED_WND);
10323 	if ((rack->rc_has_collapsed == 0) ||
10324 	    (rack->r_ctl.last_collapse_point != (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)))
10325 		counter_u64_add(rack_collapsed_win_seen, 1);
10326 	rack->r_ctl.last_collapse_point = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
10327 	rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
10328 	rack->rc_has_collapsed = 1;
10329 	rack->r_collapse_point_valid = 1;
10330 	rack_log_collapse(rack, 0, 0, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
10331 }
10332 
10333 static void
10334 rack_un_collapse_window(struct tcp_rack *rack, int line)
10335 {
10336 	struct rack_sendmap *nrsm, *rsm, fe;
10337 	int cnt = 0, split = 0;
10338 #ifdef INVARIANTS
10339 	struct rack_sendmap *insret;
10340 #endif
10341 
10342 	memset(&fe, 0, sizeof(fe));
10343 	rack->rc_has_collapsed = 0;
10344 	fe.r_start = rack->r_ctl.last_collapse_point;
10345 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
10346 	if (rsm == NULL) {
10347 		/* Nothing to do maybe the peer ack'ed it all */
10348 		rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
10349 		return;
10350 	}
10351 	/* Now do we need to split this one? */
10352 	if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
10353 		rack_log_collapse(rack, rsm->r_start, rsm->r_end,
10354 				  rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
10355 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10356 		if (nrsm == NULL) {
10357 			/* We can't get a rsm, mark all? */
10358 			nrsm = rsm;
10359 			goto no_split;
10360 		}
10361 		/* Clone it */
10362 		split = 1;
10363 		rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
10364 #ifndef INVARIANTS
10365 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10366 #else
10367 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10368 		if (insret != NULL) {
10369 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
10370 			      nrsm, insret, rack, rsm);
10371 		}
10372 #endif
10373 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
10374 				 rack->r_ctl.last_collapse_point, __LINE__);
10375 		if (rsm->r_in_tmap) {
10376 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10377 			nrsm->r_in_tmap = 1;
10378 		}
10379 		/*
10380 		 * Set in the new RSM as the
10381 		 * collapsed starting point
10382 		 */
10383 		rsm = nrsm;
10384 	}
10385 no_split:
10386 	RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
10387 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
10388 		rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
10389 		cnt++;
10390 	}
10391 	if (cnt) {
10392 		counter_u64_add(rack_collapsed_win, 1);
10393 	}
10394 	rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
10395 }
10396 
10397 static void
10398 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10399 			int32_t tlen, int32_t tfo_syn)
10400 {
10401 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
10402 		if (rack->rc_dack_mode &&
10403 		    (tlen > 500) &&
10404 		    (rack->rc_dack_toggle == 1)) {
10405 			goto no_delayed_ack;
10406 		}
10407 		rack_timer_cancel(tp, rack,
10408 				  rack->r_ctl.rc_rcvtime, __LINE__);
10409 		tp->t_flags |= TF_DELACK;
10410 	} else {
10411 no_delayed_ack:
10412 		rack->r_wanted_output = 1;
10413 		tp->t_flags |= TF_ACKNOW;
10414 		if (rack->rc_dack_mode) {
10415 			if (tp->t_flags & TF_DELACK)
10416 				rack->rc_dack_toggle = 1;
10417 			else
10418 				rack->rc_dack_toggle = 0;
10419 		}
10420 	}
10421 }
10422 
10423 static void
10424 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10425 {
10426 	/*
10427 	 * If fast output is in progress, lets validate that
10428 	 * the new window did not shrink on us and make it
10429 	 * so fast output should end.
10430 	 */
10431 	if (rack->r_fast_output) {
10432 		uint32_t out;
10433 
10434 		/*
10435 		 * Calculate what we will send if left as is
10436 		 * and compare that to our send window.
10437 		 */
10438 		out = ctf_outstanding(tp);
10439 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10440 			/* ok we have an issue */
10441 			if (out >= tp->snd_wnd) {
10442 				/* Turn off fast output the window is met or collapsed */
10443 				rack->r_fast_output = 0;
10444 			} else {
10445 				/* we have some room left */
10446 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10447 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10448 					/* If not at least 1 full segment never mind */
10449 					rack->r_fast_output = 0;
10450 				}
10451 			}
10452 		}
10453 	}
10454 }
10455 
10456 
10457 /*
10458  * Return value of 1, the TCB is unlocked and most
10459  * likely gone, return value of 0, the TCP is still
10460  * locked.
10461  */
10462 static int
10463 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10464     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10465     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10466 {
10467 	/*
10468 	 * Update window information. Don't look at window if no ACK: TAC's
10469 	 * send garbage on first SYN.
10470 	 */
10471 	int32_t nsegs;
10472 	int32_t tfo_syn;
10473 	struct tcp_rack *rack;
10474 
10475 	INP_WLOCK_ASSERT(tptoinpcb(tp));
10476 
10477 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10478 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10479 	if ((thflags & TH_ACK) &&
10480 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10481 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10482 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10483 		/* keep track of pure window updates */
10484 		if (tlen == 0 &&
10485 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10486 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10487 		tp->snd_wnd = tiwin;
10488 		rack_validate_fo_sendwin_up(tp, rack);
10489 		tp->snd_wl1 = th->th_seq;
10490 		tp->snd_wl2 = th->th_ack;
10491 		if (tp->snd_wnd > tp->max_sndwnd)
10492 			tp->max_sndwnd = tp->snd_wnd;
10493 		rack->r_wanted_output = 1;
10494 	} else if (thflags & TH_ACK) {
10495 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10496 			tp->snd_wnd = tiwin;
10497 			rack_validate_fo_sendwin_up(tp, rack);
10498 			tp->snd_wl1 = th->th_seq;
10499 			tp->snd_wl2 = th->th_ack;
10500 		}
10501 	}
10502 	if (tp->snd_wnd < ctf_outstanding(tp))
10503 		/* The peer collapsed the window */
10504 		rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__);
10505 	else if (rack->rc_has_collapsed)
10506 		rack_un_collapse_window(rack, __LINE__);
10507 	if ((rack->r_collapse_point_valid) &&
10508 	    (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
10509 		rack->r_collapse_point_valid = 0;
10510 	/* Was persist timer active and now we have window space? */
10511 	if ((rack->rc_in_persist != 0) &&
10512 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10513 				rack->r_ctl.rc_pace_min_segs))) {
10514 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10515 		tp->snd_nxt = tp->snd_max;
10516 		/* Make sure we output to start the timer */
10517 		rack->r_wanted_output = 1;
10518 	}
10519 	/* Do we enter persists? */
10520 	if ((rack->rc_in_persist == 0) &&
10521 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10522 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
10523 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
10524 	    sbavail(&tptosocket(tp)->so_snd) &&
10525 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
10526 		/*
10527 		 * Here the rwnd is less than
10528 		 * the pacing size, we are established,
10529 		 * nothing is outstanding, and there is
10530 		 * data to send. Enter persists.
10531 		 */
10532 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10533 	}
10534 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10535 		m_freem(m);
10536 		return (0);
10537 	}
10538 	/*
10539 	 * don't process the URG bit, ignore them drag
10540 	 * along the up.
10541 	 */
10542 	tp->rcv_up = tp->rcv_nxt;
10543 
10544 	/*
10545 	 * Process the segment text, merging it into the TCP sequencing
10546 	 * queue, and arranging for acknowledgment of receipt if necessary.
10547 	 * This process logically involves adjusting tp->rcv_wnd as data is
10548 	 * presented to the user (this happens in tcp_usrreq.c, case
10549 	 * PRU_RCVD).  If a FIN has already been received on this connection
10550 	 * then we just ignore the text.
10551 	 */
10552 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10553 		   IS_FASTOPEN(tp->t_flags));
10554 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10555 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10556 		tcp_seq save_start = th->th_seq;
10557 		tcp_seq save_rnxt  = tp->rcv_nxt;
10558 		int     save_tlen  = tlen;
10559 
10560 		m_adj(m, drop_hdrlen);	/* delayed header drop */
10561 		/*
10562 		 * Insert segment which includes th into TCP reassembly
10563 		 * queue with control block tp.  Set thflags to whether
10564 		 * reassembly now includes a segment with FIN.  This handles
10565 		 * the common case inline (segment is the next to be
10566 		 * received on an established connection, and the queue is
10567 		 * empty), avoiding linkage into and removal from the queue
10568 		 * and repetition of various conversions. Set DELACK for
10569 		 * segments received in order, but ack immediately when
10570 		 * segments are out of order (so fast retransmit can work).
10571 		 */
10572 		if (th->th_seq == tp->rcv_nxt &&
10573 		    SEGQ_EMPTY(tp) &&
10574 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
10575 		    tfo_syn)) {
10576 #ifdef NETFLIX_SB_LIMITS
10577 			u_int mcnt, appended;
10578 
10579 			if (so->so_rcv.sb_shlim) {
10580 				mcnt = m_memcnt(m);
10581 				appended = 0;
10582 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10583 				    CFO_NOSLEEP, NULL) == false) {
10584 					counter_u64_add(tcp_sb_shlim_fails, 1);
10585 					m_freem(m);
10586 					return (0);
10587 				}
10588 			}
10589 #endif
10590 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10591 			tp->rcv_nxt += tlen;
10592 			if (tlen &&
10593 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10594 			    (tp->t_fbyte_in == 0)) {
10595 				tp->t_fbyte_in = ticks;
10596 				if (tp->t_fbyte_in == 0)
10597 					tp->t_fbyte_in = 1;
10598 				if (tp->t_fbyte_out && tp->t_fbyte_in)
10599 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10600 			}
10601 			thflags = tcp_get_flags(th) & TH_FIN;
10602 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10603 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10604 			SOCKBUF_LOCK(&so->so_rcv);
10605 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10606 				m_freem(m);
10607 			} else
10608 #ifdef NETFLIX_SB_LIMITS
10609 				appended =
10610 #endif
10611 					sbappendstream_locked(&so->so_rcv, m, 0);
10612 
10613 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10614 			/* NB: sorwakeup_locked() does an implicit unlock. */
10615 			sorwakeup_locked(so);
10616 #ifdef NETFLIX_SB_LIMITS
10617 			if (so->so_rcv.sb_shlim && appended != mcnt)
10618 				counter_fo_release(so->so_rcv.sb_shlim,
10619 				    mcnt - appended);
10620 #endif
10621 		} else {
10622 			/*
10623 			 * XXX: Due to the header drop above "th" is
10624 			 * theoretically invalid by now.  Fortunately
10625 			 * m_adj() doesn't actually frees any mbufs when
10626 			 * trimming from the head.
10627 			 */
10628 			tcp_seq temp = save_start;
10629 
10630 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
10631 			tp->t_flags |= TF_ACKNOW;
10632 			if (tp->t_flags & TF_WAKESOR) {
10633 				tp->t_flags &= ~TF_WAKESOR;
10634 				/* NB: sorwakeup_locked() does an implicit unlock. */
10635 				sorwakeup_locked(so);
10636 			}
10637 		}
10638 		if ((tp->t_flags & TF_SACK_PERMIT) &&
10639 		    (save_tlen > 0) &&
10640 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
10641 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10642 				/*
10643 				 * DSACK actually handled in the fastpath
10644 				 * above.
10645 				 */
10646 				RACK_OPTS_INC(tcp_sack_path_1);
10647 				tcp_update_sack_list(tp, save_start,
10648 				    save_start + save_tlen);
10649 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10650 				if ((tp->rcv_numsacks >= 1) &&
10651 				    (tp->sackblks[0].end == save_start)) {
10652 					/*
10653 					 * Partial overlap, recorded at todrop
10654 					 * above.
10655 					 */
10656 					RACK_OPTS_INC(tcp_sack_path_2a);
10657 					tcp_update_sack_list(tp,
10658 					    tp->sackblks[0].start,
10659 					    tp->sackblks[0].end);
10660 				} else {
10661 					RACK_OPTS_INC(tcp_sack_path_2b);
10662 					tcp_update_dsack_list(tp, save_start,
10663 					    save_start + save_tlen);
10664 				}
10665 			} else if (tlen >= save_tlen) {
10666 				/* Update of sackblks. */
10667 				RACK_OPTS_INC(tcp_sack_path_3);
10668 				tcp_update_dsack_list(tp, save_start,
10669 				    save_start + save_tlen);
10670 			} else if (tlen > 0) {
10671 				RACK_OPTS_INC(tcp_sack_path_4);
10672 				tcp_update_dsack_list(tp, save_start,
10673 				    save_start + tlen);
10674 			}
10675 		}
10676 	} else {
10677 		m_freem(m);
10678 		thflags &= ~TH_FIN;
10679 	}
10680 
10681 	/*
10682 	 * If FIN is received ACK the FIN and let the user know that the
10683 	 * connection is closing.
10684 	 */
10685 	if (thflags & TH_FIN) {
10686 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10687 			/* The socket upcall is handled by socantrcvmore. */
10688 			socantrcvmore(so);
10689 			/*
10690 			 * If connection is half-synchronized (ie NEEDSYN
10691 			 * flag on) then delay ACK, so it may be piggybacked
10692 			 * when SYN is sent. Otherwise, since we received a
10693 			 * FIN then no more input can be expected, send ACK
10694 			 * now.
10695 			 */
10696 			if (tp->t_flags & TF_NEEDSYN) {
10697 				rack_timer_cancel(tp, rack,
10698 				    rack->r_ctl.rc_rcvtime, __LINE__);
10699 				tp->t_flags |= TF_DELACK;
10700 			} else {
10701 				tp->t_flags |= TF_ACKNOW;
10702 			}
10703 			tp->rcv_nxt++;
10704 		}
10705 		switch (tp->t_state) {
10706 			/*
10707 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
10708 			 * CLOSE_WAIT state.
10709 			 */
10710 		case TCPS_SYN_RECEIVED:
10711 			tp->t_starttime = ticks;
10712 			/* FALLTHROUGH */
10713 		case TCPS_ESTABLISHED:
10714 			rack_timer_cancel(tp, rack,
10715 			    rack->r_ctl.rc_rcvtime, __LINE__);
10716 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
10717 			break;
10718 
10719 			/*
10720 			 * If still in FIN_WAIT_1 STATE FIN has not been
10721 			 * acked so enter the CLOSING state.
10722 			 */
10723 		case TCPS_FIN_WAIT_1:
10724 			rack_timer_cancel(tp, rack,
10725 			    rack->r_ctl.rc_rcvtime, __LINE__);
10726 			tcp_state_change(tp, TCPS_CLOSING);
10727 			break;
10728 
10729 			/*
10730 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
10731 			 * starting the time-wait timer, turning off the
10732 			 * other standard timers.
10733 			 */
10734 		case TCPS_FIN_WAIT_2:
10735 			rack_timer_cancel(tp, rack,
10736 			    rack->r_ctl.rc_rcvtime, __LINE__);
10737 			tcp_twstart(tp);
10738 			return (1);
10739 		}
10740 	}
10741 	/*
10742 	 * Return any desired output.
10743 	 */
10744 	if ((tp->t_flags & TF_ACKNOW) ||
10745 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10746 		rack->r_wanted_output = 1;
10747 	}
10748 	return (0);
10749 }
10750 
10751 /*
10752  * Here nothing is really faster, its just that we
10753  * have broken out the fast-data path also just like
10754  * the fast-ack.
10755  */
10756 static int
10757 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10758     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10759     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10760 {
10761 	int32_t nsegs;
10762 	int32_t newsize = 0;	/* automatic sockbuf scaling */
10763 	struct tcp_rack *rack;
10764 #ifdef NETFLIX_SB_LIMITS
10765 	u_int mcnt, appended;
10766 #endif
10767 
10768 	/*
10769 	 * If last ACK falls within this segment's sequence numbers, record
10770 	 * the timestamp. NOTE that the test is modified according to the
10771 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10772 	 */
10773 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10774 		return (0);
10775 	}
10776 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10777 		return (0);
10778 	}
10779 	if (tiwin && tiwin != tp->snd_wnd) {
10780 		return (0);
10781 	}
10782 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10783 		return (0);
10784 	}
10785 	if (__predict_false((to->to_flags & TOF_TS) &&
10786 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10787 		return (0);
10788 	}
10789 	if (__predict_false((th->th_ack != tp->snd_una))) {
10790 		return (0);
10791 	}
10792 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10793 		return (0);
10794 	}
10795 	if ((to->to_flags & TOF_TS) != 0 &&
10796 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10797 		tp->ts_recent_age = tcp_ts_getticks();
10798 		tp->ts_recent = to->to_tsval;
10799 	}
10800 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10801 	/*
10802 	 * This is a pure, in-sequence data packet with nothing on the
10803 	 * reassembly queue and we have enough buffer space to take it.
10804 	 */
10805 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10806 
10807 #ifdef NETFLIX_SB_LIMITS
10808 	if (so->so_rcv.sb_shlim) {
10809 		mcnt = m_memcnt(m);
10810 		appended = 0;
10811 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10812 		    CFO_NOSLEEP, NULL) == false) {
10813 			counter_u64_add(tcp_sb_shlim_fails, 1);
10814 			m_freem(m);
10815 			return (1);
10816 		}
10817 	}
10818 #endif
10819 	/* Clean receiver SACK report if present */
10820 	if (tp->rcv_numsacks)
10821 		tcp_clean_sackreport(tp);
10822 	KMOD_TCPSTAT_INC(tcps_preddat);
10823 	tp->rcv_nxt += tlen;
10824 	if (tlen &&
10825 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10826 	    (tp->t_fbyte_in == 0)) {
10827 		tp->t_fbyte_in = ticks;
10828 		if (tp->t_fbyte_in == 0)
10829 			tp->t_fbyte_in = 1;
10830 		if (tp->t_fbyte_out && tp->t_fbyte_in)
10831 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10832 	}
10833 	/*
10834 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10835 	 */
10836 	tp->snd_wl1 = th->th_seq;
10837 	/*
10838 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10839 	 */
10840 	tp->rcv_up = tp->rcv_nxt;
10841 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10842 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10843 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10844 
10845 	/* Add data to socket buffer. */
10846 	SOCKBUF_LOCK(&so->so_rcv);
10847 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10848 		m_freem(m);
10849 	} else {
10850 		/*
10851 		 * Set new socket buffer size. Give up when limit is
10852 		 * reached.
10853 		 */
10854 		if (newsize)
10855 			if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
10856 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10857 		m_adj(m, drop_hdrlen);	/* delayed header drop */
10858 #ifdef NETFLIX_SB_LIMITS
10859 		appended =
10860 #endif
10861 			sbappendstream_locked(&so->so_rcv, m, 0);
10862 		ctf_calc_rwin(so, tp);
10863 	}
10864 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10865 	/* NB: sorwakeup_locked() does an implicit unlock. */
10866 	sorwakeup_locked(so);
10867 #ifdef NETFLIX_SB_LIMITS
10868 	if (so->so_rcv.sb_shlim && mcnt != appended)
10869 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10870 #endif
10871 	rack_handle_delayed_ack(tp, rack, tlen, 0);
10872 	if (tp->snd_una == tp->snd_max)
10873 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10874 	return (1);
10875 }
10876 
10877 /*
10878  * This subfunction is used to try to highly optimize the
10879  * fast path. We again allow window updates that are
10880  * in sequence to remain in the fast-path. We also add
10881  * in the __predict's to attempt to help the compiler.
10882  * Note that if we return a 0, then we can *not* process
10883  * it and the caller should push the packet into the
10884  * slow-path.
10885  */
10886 static int
10887 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10888     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10889     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10890 {
10891 	int32_t acked;
10892 	int32_t nsegs;
10893 	int32_t under_pacing = 0;
10894 	struct tcp_rack *rack;
10895 
10896 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10897 		/* Old ack, behind (or duplicate to) the last one rcv'd */
10898 		return (0);
10899 	}
10900 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
10901 		/* Above what we have sent? */
10902 		return (0);
10903 	}
10904 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10905 		/* We are retransmitting */
10906 		return (0);
10907 	}
10908 	if (__predict_false(tiwin == 0)) {
10909 		/* zero window */
10910 		return (0);
10911 	}
10912 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
10913 		/* We need a SYN or a FIN, unlikely.. */
10914 		return (0);
10915 	}
10916 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
10917 		/* Timestamp is behind .. old ack with seq wrap? */
10918 		return (0);
10919 	}
10920 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
10921 		/* Still recovering */
10922 		return (0);
10923 	}
10924 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10925 	if (rack->r_ctl.rc_sacked) {
10926 		/* We have sack holes on our scoreboard */
10927 		return (0);
10928 	}
10929 	/* Ok if we reach here, we can process a fast-ack */
10930 	if (rack->gp_ready &&
10931 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10932 		under_pacing = 1;
10933 	}
10934 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10935 	rack_log_ack(tp, to, th, 0, 0);
10936 	/* Did the window get updated? */
10937 	if (tiwin != tp->snd_wnd) {
10938 		tp->snd_wnd = tiwin;
10939 		rack_validate_fo_sendwin_up(tp, rack);
10940 		tp->snd_wl1 = th->th_seq;
10941 		if (tp->snd_wnd > tp->max_sndwnd)
10942 			tp->max_sndwnd = tp->snd_wnd;
10943 	}
10944 	/* Do we exit persists? */
10945 	if ((rack->rc_in_persist != 0) &&
10946 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10947 			       rack->r_ctl.rc_pace_min_segs))) {
10948 		rack_exit_persist(tp, rack, cts);
10949 	}
10950 	/* Do we enter persists? */
10951 	if ((rack->rc_in_persist == 0) &&
10952 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10953 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
10954 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
10955 	    sbavail(&tptosocket(tp)->so_snd) &&
10956 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
10957 		/*
10958 		 * Here the rwnd is less than
10959 		 * the pacing size, we are established,
10960 		 * nothing is outstanding, and there is
10961 		 * data to send. Enter persists.
10962 		 */
10963 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10964 	}
10965 	/*
10966 	 * If last ACK falls within this segment's sequence numbers, record
10967 	 * the timestamp. NOTE that the test is modified according to the
10968 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10969 	 */
10970 	if ((to->to_flags & TOF_TS) != 0 &&
10971 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10972 		tp->ts_recent_age = tcp_ts_getticks();
10973 		tp->ts_recent = to->to_tsval;
10974 	}
10975 	/*
10976 	 * This is a pure ack for outstanding data.
10977 	 */
10978 	KMOD_TCPSTAT_INC(tcps_predack);
10979 
10980 	/*
10981 	 * "bad retransmit" recovery.
10982 	 */
10983 	if ((tp->t_flags & TF_PREVVALID) &&
10984 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10985 		tp->t_flags &= ~TF_PREVVALID;
10986 		if (tp->t_rxtshift == 1 &&
10987 		    (int)(ticks - tp->t_badrxtwin) < 0)
10988 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
10989 	}
10990 	/*
10991 	 * Recalculate the transmit timer / rtt.
10992 	 *
10993 	 * Some boxes send broken timestamp replies during the SYN+ACK
10994 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
10995 	 * and blow up the retransmit timer.
10996 	 */
10997 	acked = BYTES_THIS_ACK(tp, th);
10998 
10999 #ifdef TCP_HHOOK
11000 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
11001 	hhook_run_tcp_est_in(tp, th, to);
11002 #endif
11003 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
11004 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
11005 	if (acked) {
11006 		struct mbuf *mfree;
11007 
11008 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
11009 		SOCKBUF_LOCK(&so->so_snd);
11010 		mfree = sbcut_locked(&so->so_snd, acked);
11011 		tp->snd_una = th->th_ack;
11012 		/* Note we want to hold the sb lock through the sendmap adjust */
11013 		rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
11014 		/* Wake up the socket if we have room to write more */
11015 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
11016 		sowwakeup_locked(so);
11017 		m_freem(mfree);
11018 		tp->t_rxtshift = 0;
11019 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11020 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11021 		rack->rc_tlp_in_progress = 0;
11022 		rack->r_ctl.rc_tlp_cnt_out = 0;
11023 		/*
11024 		 * If it is the RXT timer we want to
11025 		 * stop it, so we can restart a TLP.
11026 		 */
11027 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
11028 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11029 #ifdef NETFLIX_HTTP_LOGGING
11030 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
11031 #endif
11032 	}
11033 	/*
11034 	 * Let the congestion control algorithm update congestion control
11035 	 * related information. This typically means increasing the
11036 	 * congestion window.
11037 	 */
11038 	if (tp->snd_wnd < ctf_outstanding(tp)) {
11039 		/* The peer collapsed the window */
11040 		rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__);
11041 	} else if (rack->rc_has_collapsed)
11042 		rack_un_collapse_window(rack, __LINE__);
11043 	if ((rack->r_collapse_point_valid) &&
11044 	    (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
11045 		rack->r_collapse_point_valid = 0;
11046 	/*
11047 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
11048 	 */
11049 	tp->snd_wl2 = th->th_ack;
11050 	tp->t_dupacks = 0;
11051 	m_freem(m);
11052 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
11053 
11054 	/*
11055 	 * If all outstanding data are acked, stop retransmit timer,
11056 	 * otherwise restart timer using current (possibly backed-off)
11057 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
11058 	 * If data are ready to send, let tcp_output decide between more
11059 	 * output or persist.
11060 	 */
11061 	if (under_pacing &&
11062 	    (rack->use_fixed_rate == 0) &&
11063 	    (rack->in_probe_rtt == 0) &&
11064 	    rack->rc_gp_dyn_mul &&
11065 	    rack->rc_always_pace) {
11066 		/* Check if we are dragging bottom */
11067 		rack_check_bottom_drag(tp, rack, so, acked);
11068 	}
11069 	if (tp->snd_una == tp->snd_max) {
11070 		tp->t_flags &= ~TF_PREVVALID;
11071 		rack->r_ctl.retran_during_recovery = 0;
11072 		rack->r_ctl.dsack_byte_cnt = 0;
11073 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
11074 		if (rack->r_ctl.rc_went_idle_time == 0)
11075 			rack->r_ctl.rc_went_idle_time = 1;
11076 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
11077 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
11078 			tp->t_acktime = 0;
11079 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11080 	}
11081 	if (acked && rack->r_fast_output)
11082 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
11083 	if (sbavail(&so->so_snd)) {
11084 		rack->r_wanted_output = 1;
11085 	}
11086 	return (1);
11087 }
11088 
11089 /*
11090  * Return value of 1, the TCB is unlocked and most
11091  * likely gone, return value of 0, the TCP is still
11092  * locked.
11093  */
11094 static int
11095 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
11096     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11097     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11098 {
11099 	int32_t ret_val = 0;
11100 	int32_t todrop;
11101 	int32_t ourfinisacked = 0;
11102 	struct tcp_rack *rack;
11103 
11104 	INP_WLOCK_ASSERT(tptoinpcb(tp));
11105 
11106 	ctf_calc_rwin(so, tp);
11107 	/*
11108 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
11109 	 * SYN, drop the input. if seg contains a RST, then drop the
11110 	 * connection. if seg does not contain SYN, then drop it. Otherwise
11111 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
11112 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
11113 	 * contains an ECE and ECN support is enabled, the stream is ECN
11114 	 * capable. if SYN has been acked change to ESTABLISHED else
11115 	 * SYN_RCVD state arrange for segment to be acked (eventually)
11116 	 * continue processing rest of data/controls.
11117 	 */
11118 	if ((thflags & TH_ACK) &&
11119 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
11120 	    SEQ_GT(th->th_ack, tp->snd_max))) {
11121 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11122 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11123 		return (1);
11124 	}
11125 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
11126 		TCP_PROBE5(connect__refused, NULL, tp,
11127 		    mtod(m, const char *), tp, th);
11128 		tp = tcp_drop(tp, ECONNREFUSED);
11129 		ctf_do_drop(m, tp);
11130 		return (1);
11131 	}
11132 	if (thflags & TH_RST) {
11133 		ctf_do_drop(m, tp);
11134 		return (1);
11135 	}
11136 	if (!(thflags & TH_SYN)) {
11137 		ctf_do_drop(m, tp);
11138 		return (1);
11139 	}
11140 	tp->irs = th->th_seq;
11141 	tcp_rcvseqinit(tp);
11142 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11143 	if (thflags & TH_ACK) {
11144 		int tfo_partial = 0;
11145 
11146 		KMOD_TCPSTAT_INC(tcps_connects);
11147 		soisconnected(so);
11148 #ifdef MAC
11149 		mac_socketpeer_set_from_mbuf(m, so);
11150 #endif
11151 		/* Do window scaling on this connection? */
11152 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11153 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11154 			tp->rcv_scale = tp->request_r_scale;
11155 		}
11156 		tp->rcv_adv += min(tp->rcv_wnd,
11157 		    TCP_MAXWIN << tp->rcv_scale);
11158 		/*
11159 		 * If not all the data that was sent in the TFO SYN
11160 		 * has been acked, resend the remainder right away.
11161 		 */
11162 		if (IS_FASTOPEN(tp->t_flags) &&
11163 		    (tp->snd_una != tp->snd_max)) {
11164 			tp->snd_nxt = th->th_ack;
11165 			tfo_partial = 1;
11166 		}
11167 		/*
11168 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
11169 		 * will be turned on later.
11170 		 */
11171 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
11172 			rack_timer_cancel(tp, rack,
11173 					  rack->r_ctl.rc_rcvtime, __LINE__);
11174 			tp->t_flags |= TF_DELACK;
11175 		} else {
11176 			rack->r_wanted_output = 1;
11177 			tp->t_flags |= TF_ACKNOW;
11178 			rack->rc_dack_toggle = 0;
11179 		}
11180 
11181 		tcp_ecn_input_syn_sent(tp, thflags, iptos);
11182 
11183 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
11184 			/*
11185 			 * We advance snd_una for the
11186 			 * fast open case. If th_ack is
11187 			 * acknowledging data beyond
11188 			 * snd_una we can't just call
11189 			 * ack-processing since the
11190 			 * data stream in our send-map
11191 			 * will start at snd_una + 1 (one
11192 			 * beyond the SYN). If its just
11193 			 * equal we don't need to do that
11194 			 * and there is no send_map.
11195 			 */
11196 			tp->snd_una++;
11197 		}
11198 		/*
11199 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
11200 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
11201 		 */
11202 		tp->t_starttime = ticks;
11203 		if (tp->t_flags & TF_NEEDFIN) {
11204 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
11205 			tp->t_flags &= ~TF_NEEDFIN;
11206 			thflags &= ~TH_SYN;
11207 		} else {
11208 			tcp_state_change(tp, TCPS_ESTABLISHED);
11209 			TCP_PROBE5(connect__established, NULL, tp,
11210 			    mtod(m, const char *), tp, th);
11211 			rack_cc_conn_init(tp);
11212 		}
11213 	} else {
11214 		/*
11215 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
11216 		 * open.  If segment contains CC option and there is a
11217 		 * cached CC, apply TAO test. If it succeeds, connection is *
11218 		 * half-synchronized. Otherwise, do 3-way handshake:
11219 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
11220 		 * there was no CC option, clear cached CC value.
11221 		 */
11222 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
11223 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
11224 	}
11225 	/*
11226 	 * Advance th->th_seq to correspond to first data byte. If data,
11227 	 * trim to stay within window, dropping FIN if necessary.
11228 	 */
11229 	th->th_seq++;
11230 	if (tlen > tp->rcv_wnd) {
11231 		todrop = tlen - tp->rcv_wnd;
11232 		m_adj(m, -todrop);
11233 		tlen = tp->rcv_wnd;
11234 		thflags &= ~TH_FIN;
11235 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
11236 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
11237 	}
11238 	tp->snd_wl1 = th->th_seq - 1;
11239 	tp->rcv_up = th->th_seq;
11240 	/*
11241 	 * Client side of transaction: already sent SYN and data. If the
11242 	 * remote host used T/TCP to validate the SYN, our data will be
11243 	 * ACK'd; if so, enter normal data segment processing in the middle
11244 	 * of step 5, ack processing. Otherwise, goto step 6.
11245 	 */
11246 	if (thflags & TH_ACK) {
11247 		/* For syn-sent we need to possibly update the rtt */
11248 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11249 			uint32_t t, mcts;
11250 
11251 			mcts = tcp_ts_getticks();
11252 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11253 			if (!tp->t_rttlow || tp->t_rttlow > t)
11254 				tp->t_rttlow = t;
11255 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
11256 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11257 			tcp_rack_xmit_timer_commit(rack, tp);
11258 		}
11259 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
11260 			return (ret_val);
11261 		/* We may have changed to FIN_WAIT_1 above */
11262 		if (tp->t_state == TCPS_FIN_WAIT_1) {
11263 			/*
11264 			 * In FIN_WAIT_1 STATE in addition to the processing
11265 			 * for the ESTABLISHED state if our FIN is now
11266 			 * acknowledged then enter FIN_WAIT_2.
11267 			 */
11268 			if (ourfinisacked) {
11269 				/*
11270 				 * If we can't receive any more data, then
11271 				 * closing user can proceed. Starting the
11272 				 * timer is contrary to the specification,
11273 				 * but if we don't get a FIN we'll hang
11274 				 * forever.
11275 				 *
11276 				 * XXXjl: we should release the tp also, and
11277 				 * use a compressed state.
11278 				 */
11279 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11280 					soisdisconnected(so);
11281 					tcp_timer_activate(tp, TT_2MSL,
11282 					    (tcp_fast_finwait2_recycle ?
11283 					    tcp_finwait2_timeout :
11284 					    TP_MAXIDLE(tp)));
11285 				}
11286 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
11287 			}
11288 		}
11289 	}
11290 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11291 	   tiwin, thflags, nxt_pkt));
11292 }
11293 
11294 /*
11295  * Return value of 1, the TCB is unlocked and most
11296  * likely gone, return value of 0, the TCP is still
11297  * locked.
11298  */
11299 static int
11300 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
11301     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11302     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11303 {
11304 	struct tcp_rack *rack;
11305 	int32_t ret_val = 0;
11306 	int32_t ourfinisacked = 0;
11307 
11308 	ctf_calc_rwin(so, tp);
11309 	if ((thflags & TH_ACK) &&
11310 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
11311 	    SEQ_GT(th->th_ack, tp->snd_max))) {
11312 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11313 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11314 		return (1);
11315 	}
11316 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11317 	if (IS_FASTOPEN(tp->t_flags)) {
11318 		/*
11319 		 * When a TFO connection is in SYN_RECEIVED, the
11320 		 * only valid packets are the initial SYN, a
11321 		 * retransmit/copy of the initial SYN (possibly with
11322 		 * a subset of the original data), a valid ACK, a
11323 		 * FIN, or a RST.
11324 		 */
11325 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
11326 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11327 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11328 			return (1);
11329 		} else if (thflags & TH_SYN) {
11330 			/* non-initial SYN is ignored */
11331 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
11332 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
11333 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
11334 				ctf_do_drop(m, NULL);
11335 				return (0);
11336 			}
11337 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
11338 			ctf_do_drop(m, NULL);
11339 			return (0);
11340 		}
11341 	}
11342 
11343 	if ((thflags & TH_RST) ||
11344 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11345 		return (__ctf_process_rst(m, th, so, tp,
11346 					  &rack->r_ctl.challenge_ack_ts,
11347 					  &rack->r_ctl.challenge_ack_cnt));
11348 	/*
11349 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11350 	 * it's less than ts_recent, drop it.
11351 	 */
11352 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11353 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11354 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11355 			return (ret_val);
11356 	}
11357 	/*
11358 	 * In the SYN-RECEIVED state, validate that the packet belongs to
11359 	 * this connection before trimming the data to fit the receive
11360 	 * window.  Check the sequence number versus IRS since we know the
11361 	 * sequence numbers haven't wrapped.  This is a partial fix for the
11362 	 * "LAND" DoS attack.
11363 	 */
11364 	if (SEQ_LT(th->th_seq, tp->irs)) {
11365 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11366 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11367 		return (1);
11368 	}
11369 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11370 			      &rack->r_ctl.challenge_ack_ts,
11371 			      &rack->r_ctl.challenge_ack_cnt)) {
11372 		return (ret_val);
11373 	}
11374 	/*
11375 	 * If last ACK falls within this segment's sequence numbers, record
11376 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11377 	 * from the latest proposal of the tcplw@cray.com list (Braden
11378 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11379 	 * with our earlier PAWS tests, so this check should be solely
11380 	 * predicated on the sequence space of this segment. 3) That we
11381 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11382 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11383 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11384 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11385 	 * p.869. In such cases, we can still calculate the RTT correctly
11386 	 * when RCV.NXT == Last.ACK.Sent.
11387 	 */
11388 	if ((to->to_flags & TOF_TS) != 0 &&
11389 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11390 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11391 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11392 		tp->ts_recent_age = tcp_ts_getticks();
11393 		tp->ts_recent = to->to_tsval;
11394 	}
11395 	tp->snd_wnd = tiwin;
11396 	rack_validate_fo_sendwin_up(tp, rack);
11397 	/*
11398 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11399 	 * is on (half-synchronized state), then queue data for later
11400 	 * processing; else drop segment and return.
11401 	 */
11402 	if ((thflags & TH_ACK) == 0) {
11403 		if (IS_FASTOPEN(tp->t_flags)) {
11404 			rack_cc_conn_init(tp);
11405 		}
11406 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11407 		    tiwin, thflags, nxt_pkt));
11408 	}
11409 	KMOD_TCPSTAT_INC(tcps_connects);
11410 	if (tp->t_flags & TF_SONOTCONN) {
11411 		tp->t_flags &= ~TF_SONOTCONN;
11412 		soisconnected(so);
11413 	}
11414 	/* Do window scaling? */
11415 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11416 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11417 		tp->rcv_scale = tp->request_r_scale;
11418 	}
11419 	/*
11420 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11421 	 * FIN-WAIT-1
11422 	 */
11423 	tp->t_starttime = ticks;
11424 	if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11425 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11426 		tp->t_tfo_pending = NULL;
11427 	}
11428 	if (tp->t_flags & TF_NEEDFIN) {
11429 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
11430 		tp->t_flags &= ~TF_NEEDFIN;
11431 	} else {
11432 		tcp_state_change(tp, TCPS_ESTABLISHED);
11433 		TCP_PROBE5(accept__established, NULL, tp,
11434 		    mtod(m, const char *), tp, th);
11435 		/*
11436 		 * TFO connections call cc_conn_init() during SYN
11437 		 * processing.  Calling it again here for such connections
11438 		 * is not harmless as it would undo the snd_cwnd reduction
11439 		 * that occurs when a TFO SYN|ACK is retransmitted.
11440 		 */
11441 		if (!IS_FASTOPEN(tp->t_flags))
11442 			rack_cc_conn_init(tp);
11443 	}
11444 	/*
11445 	 * Account for the ACK of our SYN prior to
11446 	 * regular ACK processing below, except for
11447 	 * simultaneous SYN, which is handled later.
11448 	 */
11449 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11450 		tp->snd_una++;
11451 	/*
11452 	 * If segment contains data or ACK, will call tcp_reass() later; if
11453 	 * not, do so now to pass queued data to user.
11454 	 */
11455 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
11456 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11457 		    (struct mbuf *)0);
11458 		if (tp->t_flags & TF_WAKESOR) {
11459 			tp->t_flags &= ~TF_WAKESOR;
11460 			/* NB: sorwakeup_locked() does an implicit unlock. */
11461 			sorwakeup_locked(so);
11462 		}
11463 	}
11464 	tp->snd_wl1 = th->th_seq - 1;
11465 	/* For syn-recv we need to possibly update the rtt */
11466 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11467 		uint32_t t, mcts;
11468 
11469 		mcts = tcp_ts_getticks();
11470 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11471 		if (!tp->t_rttlow || tp->t_rttlow > t)
11472 			tp->t_rttlow = t;
11473 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11474 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11475 		tcp_rack_xmit_timer_commit(rack, tp);
11476 	}
11477 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11478 		return (ret_val);
11479 	}
11480 	if (tp->t_state == TCPS_FIN_WAIT_1) {
11481 		/* We could have went to FIN_WAIT_1 (or EST) above */
11482 		/*
11483 		 * In FIN_WAIT_1 STATE in addition to the processing for the
11484 		 * ESTABLISHED state if our FIN is now acknowledged then
11485 		 * enter FIN_WAIT_2.
11486 		 */
11487 		if (ourfinisacked) {
11488 			/*
11489 			 * If we can't receive any more data, then closing
11490 			 * user can proceed. Starting the timer is contrary
11491 			 * to the specification, but if we don't get a FIN
11492 			 * we'll hang forever.
11493 			 *
11494 			 * XXXjl: we should release the tp also, and use a
11495 			 * compressed state.
11496 			 */
11497 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11498 				soisdisconnected(so);
11499 				tcp_timer_activate(tp, TT_2MSL,
11500 				    (tcp_fast_finwait2_recycle ?
11501 				    tcp_finwait2_timeout :
11502 				    TP_MAXIDLE(tp)));
11503 			}
11504 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
11505 		}
11506 	}
11507 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11508 	    tiwin, thflags, nxt_pkt));
11509 }
11510 
11511 /*
11512  * Return value of 1, the TCB is unlocked and most
11513  * likely gone, return value of 0, the TCP is still
11514  * locked.
11515  */
11516 static int
11517 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11518     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11519     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11520 {
11521 	int32_t ret_val = 0;
11522 	struct tcp_rack *rack;
11523 
11524 	/*
11525 	 * Header prediction: check for the two common cases of a
11526 	 * uni-directional data xfer.  If the packet has no control flags,
11527 	 * is in-sequence, the window didn't change and we're not
11528 	 * retransmitting, it's a candidate.  If the length is zero and the
11529 	 * ack moved forward, we're the sender side of the xfer.  Just free
11530 	 * the data acked & wake any higher level process that was blocked
11531 	 * waiting for space.  If the length is non-zero and the ack didn't
11532 	 * move, we're the receiver side.  If we're getting packets in-order
11533 	 * (the reassembly queue is empty), add the data toc The socket
11534 	 * buffer and note that we need a delayed ack. Make sure that the
11535 	 * hidden state-flags are also off. Since we check for
11536 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11537 	 */
11538 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11539 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11540 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11541 	    __predict_true(SEGQ_EMPTY(tp)) &&
11542 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
11543 		if (tlen == 0) {
11544 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11545 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11546 				return (0);
11547 			}
11548 		} else {
11549 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11550 			    tiwin, nxt_pkt, iptos)) {
11551 				return (0);
11552 			}
11553 		}
11554 	}
11555 	ctf_calc_rwin(so, tp);
11556 
11557 	if ((thflags & TH_RST) ||
11558 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11559 		return (__ctf_process_rst(m, th, so, tp,
11560 					  &rack->r_ctl.challenge_ack_ts,
11561 					  &rack->r_ctl.challenge_ack_cnt));
11562 
11563 	/*
11564 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11565 	 * synchronized state.
11566 	 */
11567 	if (thflags & TH_SYN) {
11568 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11569 		return (ret_val);
11570 	}
11571 	/*
11572 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11573 	 * it's less than ts_recent, drop it.
11574 	 */
11575 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11576 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11577 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11578 			return (ret_val);
11579 	}
11580 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11581 			      &rack->r_ctl.challenge_ack_ts,
11582 			      &rack->r_ctl.challenge_ack_cnt)) {
11583 		return (ret_val);
11584 	}
11585 	/*
11586 	 * If last ACK falls within this segment's sequence numbers, record
11587 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11588 	 * from the latest proposal of the tcplw@cray.com list (Braden
11589 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11590 	 * with our earlier PAWS tests, so this check should be solely
11591 	 * predicated on the sequence space of this segment. 3) That we
11592 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11593 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11594 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11595 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11596 	 * p.869. In such cases, we can still calculate the RTT correctly
11597 	 * when RCV.NXT == Last.ACK.Sent.
11598 	 */
11599 	if ((to->to_flags & TOF_TS) != 0 &&
11600 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11601 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11602 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11603 		tp->ts_recent_age = tcp_ts_getticks();
11604 		tp->ts_recent = to->to_tsval;
11605 	}
11606 	/*
11607 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11608 	 * is on (half-synchronized state), then queue data for later
11609 	 * processing; else drop segment and return.
11610 	 */
11611 	if ((thflags & TH_ACK) == 0) {
11612 		if (tp->t_flags & TF_NEEDSYN) {
11613 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11614 			    tiwin, thflags, nxt_pkt));
11615 
11616 		} else if (tp->t_flags & TF_ACKNOW) {
11617 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11618 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11619 			return (ret_val);
11620 		} else {
11621 			ctf_do_drop(m, NULL);
11622 			return (0);
11623 		}
11624 	}
11625 	/*
11626 	 * Ack processing.
11627 	 */
11628 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11629 		return (ret_val);
11630 	}
11631 	if (sbavail(&so->so_snd)) {
11632 		if (ctf_progress_timeout_check(tp, true)) {
11633 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11634 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11635 			return (1);
11636 		}
11637 	}
11638 	/* State changes only happen in rack_process_data() */
11639 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11640 	    tiwin, thflags, nxt_pkt));
11641 }
11642 
11643 /*
11644  * Return value of 1, the TCB is unlocked and most
11645  * likely gone, return value of 0, the TCP is still
11646  * locked.
11647  */
11648 static int
11649 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11650     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11651     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11652 {
11653 	int32_t ret_val = 0;
11654 	struct tcp_rack *rack;
11655 
11656 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11657 	ctf_calc_rwin(so, tp);
11658 	if ((thflags & TH_RST) ||
11659 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11660 		return (__ctf_process_rst(m, th, so, tp,
11661 					  &rack->r_ctl.challenge_ack_ts,
11662 					  &rack->r_ctl.challenge_ack_cnt));
11663 	/*
11664 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11665 	 * synchronized state.
11666 	 */
11667 	if (thflags & TH_SYN) {
11668 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11669 		return (ret_val);
11670 	}
11671 	/*
11672 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11673 	 * it's less than ts_recent, drop it.
11674 	 */
11675 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11676 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11677 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11678 			return (ret_val);
11679 	}
11680 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11681 			      &rack->r_ctl.challenge_ack_ts,
11682 			      &rack->r_ctl.challenge_ack_cnt)) {
11683 		return (ret_val);
11684 	}
11685 	/*
11686 	 * If last ACK falls within this segment's sequence numbers, record
11687 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11688 	 * from the latest proposal of the tcplw@cray.com list (Braden
11689 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11690 	 * with our earlier PAWS tests, so this check should be solely
11691 	 * predicated on the sequence space of this segment. 3) That we
11692 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11693 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11694 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11695 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11696 	 * p.869. In such cases, we can still calculate the RTT correctly
11697 	 * when RCV.NXT == Last.ACK.Sent.
11698 	 */
11699 	if ((to->to_flags & TOF_TS) != 0 &&
11700 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11701 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11702 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11703 		tp->ts_recent_age = tcp_ts_getticks();
11704 		tp->ts_recent = to->to_tsval;
11705 	}
11706 	/*
11707 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11708 	 * is on (half-synchronized state), then queue data for later
11709 	 * processing; else drop segment and return.
11710 	 */
11711 	if ((thflags & TH_ACK) == 0) {
11712 		if (tp->t_flags & TF_NEEDSYN) {
11713 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11714 			    tiwin, thflags, nxt_pkt));
11715 
11716 		} else if (tp->t_flags & TF_ACKNOW) {
11717 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11718 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11719 			return (ret_val);
11720 		} else {
11721 			ctf_do_drop(m, NULL);
11722 			return (0);
11723 		}
11724 	}
11725 	/*
11726 	 * Ack processing.
11727 	 */
11728 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11729 		return (ret_val);
11730 	}
11731 	if (sbavail(&so->so_snd)) {
11732 		if (ctf_progress_timeout_check(tp, true)) {
11733 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11734 						tp, tick, PROGRESS_DROP, __LINE__);
11735 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11736 			return (1);
11737 		}
11738 	}
11739 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11740 	    tiwin, thflags, nxt_pkt));
11741 }
11742 
11743 static int
11744 rack_check_data_after_close(struct mbuf *m,
11745     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11746 {
11747 	struct tcp_rack *rack;
11748 
11749 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11750 	if (rack->rc_allow_data_af_clo == 0) {
11751 	close_now:
11752 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11753 		/* tcp_close will kill the inp pre-log the Reset */
11754 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11755 		tp = tcp_close(tp);
11756 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11757 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11758 		return (1);
11759 	}
11760 	if (sbavail(&so->so_snd) == 0)
11761 		goto close_now;
11762 	/* Ok we allow data that is ignored and a followup reset */
11763 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11764 	tp->rcv_nxt = th->th_seq + *tlen;
11765 	tp->t_flags2 |= TF2_DROP_AF_DATA;
11766 	rack->r_wanted_output = 1;
11767 	*tlen = 0;
11768 	return (0);
11769 }
11770 
11771 /*
11772  * Return value of 1, the TCB is unlocked and most
11773  * likely gone, return value of 0, the TCP is still
11774  * locked.
11775  */
11776 static int
11777 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11778     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11779     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11780 {
11781 	int32_t ret_val = 0;
11782 	int32_t ourfinisacked = 0;
11783 	struct tcp_rack *rack;
11784 
11785 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11786 	ctf_calc_rwin(so, tp);
11787 
11788 	if ((thflags & TH_RST) ||
11789 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11790 		return (__ctf_process_rst(m, th, so, tp,
11791 					  &rack->r_ctl.challenge_ack_ts,
11792 					  &rack->r_ctl.challenge_ack_cnt));
11793 	/*
11794 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11795 	 * synchronized state.
11796 	 */
11797 	if (thflags & TH_SYN) {
11798 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11799 		return (ret_val);
11800 	}
11801 	/*
11802 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11803 	 * it's less than ts_recent, drop it.
11804 	 */
11805 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11806 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11807 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11808 			return (ret_val);
11809 	}
11810 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11811 			      &rack->r_ctl.challenge_ack_ts,
11812 			      &rack->r_ctl.challenge_ack_cnt)) {
11813 		return (ret_val);
11814 	}
11815 	/*
11816 	 * If new data are received on a connection after the user processes
11817 	 * are gone, then RST the other end.
11818 	 */
11819 	if ((tp->t_flags & TF_CLOSED) && tlen &&
11820 	    rack_check_data_after_close(m, tp, &tlen, th, so))
11821 		return (1);
11822 	/*
11823 	 * If last ACK falls within this segment's sequence numbers, record
11824 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11825 	 * from the latest proposal of the tcplw@cray.com list (Braden
11826 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11827 	 * with our earlier PAWS tests, so this check should be solely
11828 	 * predicated on the sequence space of this segment. 3) That we
11829 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11830 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11831 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11832 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11833 	 * p.869. In such cases, we can still calculate the RTT correctly
11834 	 * when RCV.NXT == Last.ACK.Sent.
11835 	 */
11836 	if ((to->to_flags & TOF_TS) != 0 &&
11837 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11838 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11839 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11840 		tp->ts_recent_age = tcp_ts_getticks();
11841 		tp->ts_recent = to->to_tsval;
11842 	}
11843 	/*
11844 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11845 	 * is on (half-synchronized state), then queue data for later
11846 	 * processing; else drop segment and return.
11847 	 */
11848 	if ((thflags & TH_ACK) == 0) {
11849 		if (tp->t_flags & TF_NEEDSYN) {
11850 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11851 			    tiwin, thflags, nxt_pkt));
11852 		} else if (tp->t_flags & TF_ACKNOW) {
11853 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11854 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11855 			return (ret_val);
11856 		} else {
11857 			ctf_do_drop(m, NULL);
11858 			return (0);
11859 		}
11860 	}
11861 	/*
11862 	 * Ack processing.
11863 	 */
11864 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11865 		return (ret_val);
11866 	}
11867 	if (ourfinisacked) {
11868 		/*
11869 		 * If we can't receive any more data, then closing user can
11870 		 * proceed. Starting the timer is contrary to the
11871 		 * specification, but if we don't get a FIN we'll hang
11872 		 * forever.
11873 		 *
11874 		 * XXXjl: we should release the tp also, and use a
11875 		 * compressed state.
11876 		 */
11877 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11878 			soisdisconnected(so);
11879 			tcp_timer_activate(tp, TT_2MSL,
11880 			    (tcp_fast_finwait2_recycle ?
11881 			    tcp_finwait2_timeout :
11882 			    TP_MAXIDLE(tp)));
11883 		}
11884 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
11885 	}
11886 	if (sbavail(&so->so_snd)) {
11887 		if (ctf_progress_timeout_check(tp, true)) {
11888 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11889 						tp, tick, PROGRESS_DROP, __LINE__);
11890 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11891 			return (1);
11892 		}
11893 	}
11894 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11895 	    tiwin, thflags, nxt_pkt));
11896 }
11897 
11898 /*
11899  * Return value of 1, the TCB is unlocked and most
11900  * likely gone, return value of 0, the TCP is still
11901  * locked.
11902  */
11903 static int
11904 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
11905     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11906     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11907 {
11908 	int32_t ret_val = 0;
11909 	int32_t ourfinisacked = 0;
11910 	struct tcp_rack *rack;
11911 
11912 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11913 	ctf_calc_rwin(so, tp);
11914 
11915 	if ((thflags & TH_RST) ||
11916 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11917 		return (__ctf_process_rst(m, th, so, tp,
11918 					  &rack->r_ctl.challenge_ack_ts,
11919 					  &rack->r_ctl.challenge_ack_cnt));
11920 	/*
11921 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11922 	 * synchronized state.
11923 	 */
11924 	if (thflags & TH_SYN) {
11925 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11926 		return (ret_val);
11927 	}
11928 	/*
11929 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11930 	 * it's less than ts_recent, drop it.
11931 	 */
11932 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11933 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11934 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11935 			return (ret_val);
11936 	}
11937 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11938 			      &rack->r_ctl.challenge_ack_ts,
11939 			      &rack->r_ctl.challenge_ack_cnt)) {
11940 		return (ret_val);
11941 	}
11942 	/*
11943 	 * If new data are received on a connection after the user processes
11944 	 * are gone, then RST the other end.
11945 	 */
11946 	if ((tp->t_flags & TF_CLOSED) && tlen &&
11947 	    rack_check_data_after_close(m, tp, &tlen, th, so))
11948 		return (1);
11949 	/*
11950 	 * If last ACK falls within this segment's sequence numbers, record
11951 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11952 	 * from the latest proposal of the tcplw@cray.com list (Braden
11953 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11954 	 * with our earlier PAWS tests, so this check should be solely
11955 	 * predicated on the sequence space of this segment. 3) That we
11956 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11957 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11958 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11959 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11960 	 * p.869. In such cases, we can still calculate the RTT correctly
11961 	 * when RCV.NXT == Last.ACK.Sent.
11962 	 */
11963 	if ((to->to_flags & TOF_TS) != 0 &&
11964 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11965 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11966 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11967 		tp->ts_recent_age = tcp_ts_getticks();
11968 		tp->ts_recent = to->to_tsval;
11969 	}
11970 	/*
11971 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11972 	 * is on (half-synchronized state), then queue data for later
11973 	 * processing; else drop segment and return.
11974 	 */
11975 	if ((thflags & TH_ACK) == 0) {
11976 		if (tp->t_flags & TF_NEEDSYN) {
11977 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11978 			    tiwin, thflags, nxt_pkt));
11979 		} else if (tp->t_flags & TF_ACKNOW) {
11980 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11981 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11982 			return (ret_val);
11983 		} else {
11984 			ctf_do_drop(m, NULL);
11985 			return (0);
11986 		}
11987 	}
11988 	/*
11989 	 * Ack processing.
11990 	 */
11991 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11992 		return (ret_val);
11993 	}
11994 	if (ourfinisacked) {
11995 		tcp_twstart(tp);
11996 		m_freem(m);
11997 		return (1);
11998 	}
11999 	if (sbavail(&so->so_snd)) {
12000 		if (ctf_progress_timeout_check(tp, true)) {
12001 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12002 						tp, tick, PROGRESS_DROP, __LINE__);
12003 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12004 			return (1);
12005 		}
12006 	}
12007 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12008 	    tiwin, thflags, nxt_pkt));
12009 }
12010 
12011 /*
12012  * Return value of 1, the TCB is unlocked and most
12013  * likely gone, return value of 0, the TCP is still
12014  * locked.
12015  */
12016 static int
12017 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12018     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12019     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12020 {
12021 	int32_t ret_val = 0;
12022 	int32_t ourfinisacked = 0;
12023 	struct tcp_rack *rack;
12024 
12025 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12026 	ctf_calc_rwin(so, tp);
12027 
12028 	if ((thflags & TH_RST) ||
12029 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12030 		return (__ctf_process_rst(m, th, so, tp,
12031 					  &rack->r_ctl.challenge_ack_ts,
12032 					  &rack->r_ctl.challenge_ack_cnt));
12033 	/*
12034 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12035 	 * synchronized state.
12036 	 */
12037 	if (thflags & TH_SYN) {
12038 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
12039 		return (ret_val);
12040 	}
12041 	/*
12042 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12043 	 * it's less than ts_recent, drop it.
12044 	 */
12045 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12046 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12047 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12048 			return (ret_val);
12049 	}
12050 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12051 			      &rack->r_ctl.challenge_ack_ts,
12052 			      &rack->r_ctl.challenge_ack_cnt)) {
12053 		return (ret_val);
12054 	}
12055 	/*
12056 	 * If new data are received on a connection after the user processes
12057 	 * are gone, then RST the other end.
12058 	 */
12059 	if ((tp->t_flags & TF_CLOSED) && tlen &&
12060 	    rack_check_data_after_close(m, tp, &tlen, th, so))
12061 		return (1);
12062 	/*
12063 	 * If last ACK falls within this segment's sequence numbers, record
12064 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12065 	 * from the latest proposal of the tcplw@cray.com list (Braden
12066 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12067 	 * with our earlier PAWS tests, so this check should be solely
12068 	 * predicated on the sequence space of this segment. 3) That we
12069 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12070 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12071 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12072 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12073 	 * p.869. In such cases, we can still calculate the RTT correctly
12074 	 * when RCV.NXT == Last.ACK.Sent.
12075 	 */
12076 	if ((to->to_flags & TOF_TS) != 0 &&
12077 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12078 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12079 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12080 		tp->ts_recent_age = tcp_ts_getticks();
12081 		tp->ts_recent = to->to_tsval;
12082 	}
12083 	/*
12084 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12085 	 * is on (half-synchronized state), then queue data for later
12086 	 * processing; else drop segment and return.
12087 	 */
12088 	if ((thflags & TH_ACK) == 0) {
12089 		if (tp->t_flags & TF_NEEDSYN) {
12090 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12091 			    tiwin, thflags, nxt_pkt));
12092 		} else if (tp->t_flags & TF_ACKNOW) {
12093 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12094 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12095 			return (ret_val);
12096 		} else {
12097 			ctf_do_drop(m, NULL);
12098 			return (0);
12099 		}
12100 	}
12101 	/*
12102 	 * case TCPS_LAST_ACK: Ack processing.
12103 	 */
12104 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12105 		return (ret_val);
12106 	}
12107 	if (ourfinisacked) {
12108 		tp = tcp_close(tp);
12109 		ctf_do_drop(m, tp);
12110 		return (1);
12111 	}
12112 	if (sbavail(&so->so_snd)) {
12113 		if (ctf_progress_timeout_check(tp, true)) {
12114 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12115 						tp, tick, PROGRESS_DROP, __LINE__);
12116 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12117 			return (1);
12118 		}
12119 	}
12120 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12121 	    tiwin, thflags, nxt_pkt));
12122 }
12123 
12124 /*
12125  * Return value of 1, the TCB is unlocked and most
12126  * likely gone, return value of 0, the TCP is still
12127  * locked.
12128  */
12129 static int
12130 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
12131     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12132     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12133 {
12134 	int32_t ret_val = 0;
12135 	int32_t ourfinisacked = 0;
12136 	struct tcp_rack *rack;
12137 
12138 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12139 	ctf_calc_rwin(so, tp);
12140 
12141 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
12142 	if ((thflags & TH_RST) ||
12143 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12144 		return (__ctf_process_rst(m, th, so, tp,
12145 					  &rack->r_ctl.challenge_ack_ts,
12146 					  &rack->r_ctl.challenge_ack_cnt));
12147 	/*
12148 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12149 	 * synchronized state.
12150 	 */
12151 	if (thflags & TH_SYN) {
12152 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
12153 		return (ret_val);
12154 	}
12155 	/*
12156 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12157 	 * it's less than ts_recent, drop it.
12158 	 */
12159 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12160 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12161 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12162 			return (ret_val);
12163 	}
12164 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12165 			      &rack->r_ctl.challenge_ack_ts,
12166 			      &rack->r_ctl.challenge_ack_cnt)) {
12167 		return (ret_val);
12168 	}
12169 	/*
12170 	 * If new data are received on a connection after the user processes
12171 	 * are gone, then RST the other end.
12172 	 */
12173 	if ((tp->t_flags & TF_CLOSED) && tlen &&
12174 	    rack_check_data_after_close(m, tp, &tlen, th, so))
12175 		return (1);
12176 	/*
12177 	 * If last ACK falls within this segment's sequence numbers, record
12178 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12179 	 * from the latest proposal of the tcplw@cray.com list (Braden
12180 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12181 	 * with our earlier PAWS tests, so this check should be solely
12182 	 * predicated on the sequence space of this segment. 3) That we
12183 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12184 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12185 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12186 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12187 	 * p.869. In such cases, we can still calculate the RTT correctly
12188 	 * when RCV.NXT == Last.ACK.Sent.
12189 	 */
12190 	if ((to->to_flags & TOF_TS) != 0 &&
12191 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12192 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12193 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12194 		tp->ts_recent_age = tcp_ts_getticks();
12195 		tp->ts_recent = to->to_tsval;
12196 	}
12197 	/*
12198 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12199 	 * is on (half-synchronized state), then queue data for later
12200 	 * processing; else drop segment and return.
12201 	 */
12202 	if ((thflags & TH_ACK) == 0) {
12203 		if (tp->t_flags & TF_NEEDSYN) {
12204 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12205 			    tiwin, thflags, nxt_pkt));
12206 		} else if (tp->t_flags & TF_ACKNOW) {
12207 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12208 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12209 			return (ret_val);
12210 		} else {
12211 			ctf_do_drop(m, NULL);
12212 			return (0);
12213 		}
12214 	}
12215 	/*
12216 	 * Ack processing.
12217 	 */
12218 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12219 		return (ret_val);
12220 	}
12221 	if (sbavail(&so->so_snd)) {
12222 		if (ctf_progress_timeout_check(tp, true)) {
12223 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12224 						tp, tick, PROGRESS_DROP, __LINE__);
12225 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12226 			return (1);
12227 		}
12228 	}
12229 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12230 	    tiwin, thflags, nxt_pkt));
12231 }
12232 
12233 static void inline
12234 rack_clear_rate_sample(struct tcp_rack *rack)
12235 {
12236 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
12237 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
12238 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
12239 }
12240 
12241 static void
12242 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
12243 {
12244 	uint64_t bw_est, rate_wanted;
12245 	int chged = 0;
12246 	uint32_t user_max, orig_min, orig_max;
12247 
12248 	orig_min = rack->r_ctl.rc_pace_min_segs;
12249 	orig_max = rack->r_ctl.rc_pace_max_segs;
12250 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
12251 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
12252 		chged = 1;
12253 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
12254 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
12255 		if (user_max != rack->r_ctl.rc_pace_max_segs)
12256 			chged = 1;
12257 	}
12258 	if (rack->rc_force_max_seg) {
12259 		rack->r_ctl.rc_pace_max_segs = user_max;
12260 	} else if (rack->use_fixed_rate) {
12261 		bw_est = rack_get_bw(rack);
12262 		if ((rack->r_ctl.crte == NULL) ||
12263 		    (bw_est != rack->r_ctl.crte->rate)) {
12264 			rack->r_ctl.rc_pace_max_segs = user_max;
12265 		} else {
12266 			/* We are pacing right at the hardware rate */
12267 			uint32_t segsiz;
12268 
12269 			segsiz = min(ctf_fixed_maxseg(tp),
12270 				     rack->r_ctl.rc_pace_min_segs);
12271 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
12272 				                           tp, bw_est, segsiz, 0,
12273 							   rack->r_ctl.crte, NULL);
12274 		}
12275 	} else if (rack->rc_always_pace) {
12276 		if (rack->r_ctl.gp_bw ||
12277 #ifdef NETFLIX_PEAKRATE
12278 		    rack->rc_tp->t_maxpeakrate ||
12279 #endif
12280 		    rack->r_ctl.init_rate) {
12281 			/* We have a rate of some sort set */
12282 			uint32_t  orig;
12283 
12284 			bw_est = rack_get_bw(rack);
12285 			orig = rack->r_ctl.rc_pace_max_segs;
12286 			if (fill_override)
12287 				rate_wanted = *fill_override;
12288 			else
12289 				rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
12290 			if (rate_wanted) {
12291 				/* We have something */
12292 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
12293 										   rate_wanted,
12294 										   ctf_fixed_maxseg(rack->rc_tp));
12295 			} else
12296 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
12297 			if (orig != rack->r_ctl.rc_pace_max_segs)
12298 				chged = 1;
12299 		} else if ((rack->r_ctl.gp_bw == 0) &&
12300 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
12301 			/*
12302 			 * If we have nothing limit us to bursting
12303 			 * out IW sized pieces.
12304 			 */
12305 			chged = 1;
12306 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
12307 		}
12308 	}
12309 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
12310 		chged = 1;
12311 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
12312 	}
12313 	if (chged)
12314 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
12315 }
12316 
12317 
12318 static void
12319 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
12320 {
12321 #ifdef INET6
12322 	struct ip6_hdr *ip6 = NULL;
12323 #endif
12324 #ifdef INET
12325 	struct ip *ip = NULL;
12326 #endif
12327 	struct udphdr *udp = NULL;
12328 
12329 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
12330 #ifdef INET6
12331 	if (rack->r_is_v6) {
12332 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12333 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
12334 		if (tp->t_port) {
12335 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12336 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
12337 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12338 			udp->uh_dport = tp->t_port;
12339 			rack->r_ctl.fsb.udp = udp;
12340 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12341 		} else
12342 		{
12343 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
12344 			rack->r_ctl.fsb.udp = NULL;
12345 		}
12346 		tcpip_fillheaders(rack->rc_inp,
12347 				  tp->t_port,
12348 				  ip6, rack->r_ctl.fsb.th);
12349 	} else
12350 #endif				/* INET6 */
12351 #ifdef INET
12352 	{
12353 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
12354 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
12355 		if (tp->t_port) {
12356 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12357 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
12358 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12359 			udp->uh_dport = tp->t_port;
12360 			rack->r_ctl.fsb.udp = udp;
12361 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12362 		} else
12363 		{
12364 			rack->r_ctl.fsb.udp = NULL;
12365 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
12366 		}
12367 		tcpip_fillheaders(rack->rc_inp,
12368 				  tp->t_port,
12369 				  ip, rack->r_ctl.fsb.th);
12370 	}
12371 #endif
12372 	rack->r_fsb_inited = 1;
12373 }
12374 
12375 static int
12376 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12377 {
12378 	/*
12379 	 * Allocate the larger of spaces V6 if available else just
12380 	 * V4 and include udphdr (overbook)
12381 	 */
12382 #ifdef INET6
12383 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12384 #else
12385 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12386 #endif
12387 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12388 					    M_TCPFSB, M_NOWAIT|M_ZERO);
12389 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12390 		return (ENOMEM);
12391 	}
12392 	rack->r_fsb_inited = 0;
12393 	return (0);
12394 }
12395 
12396 static int
12397 rack_init(struct tcpcb *tp)
12398 {
12399 	struct inpcb *inp = tptoinpcb(tp);
12400 	struct tcp_rack *rack = NULL;
12401 #ifdef INVARIANTS
12402 	struct rack_sendmap *insret;
12403 #endif
12404 	uint32_t iwin, snt, us_cts;
12405 	int err;
12406 
12407 	tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12408 	if (tp->t_fb_ptr == NULL) {
12409 		/*
12410 		 * We need to allocate memory but cant. The INP and INP_INFO
12411 		 * locks and they are recursive (happens during setup. So a
12412 		 * scheme to drop the locks fails :(
12413 		 *
12414 		 */
12415 		return (ENOMEM);
12416 	}
12417 	memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12418 
12419 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12420 	RB_INIT(&rack->r_ctl.rc_mtree);
12421 	TAILQ_INIT(&rack->r_ctl.rc_free);
12422 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
12423 	rack->rc_tp = tp;
12424 	rack->rc_inp = inp;
12425 	/* Set the flag */
12426 	rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
12427 	/* Probably not needed but lets be sure */
12428 	rack_clear_rate_sample(rack);
12429 	/*
12430 	 * Save off the default values, socket options will poke
12431 	 * at these if pacing is not on or we have not yet
12432 	 * reached where pacing is on (gp_ready/fixed enabled).
12433 	 * When they get set into the CC module (when gp_ready
12434 	 * is enabled or we enable fixed) then we will set these
12435 	 * values into the CC and place in here the old values
12436 	 * so we have a restoral. Then we will set the flag
12437 	 * rc_pacing_cc_set. That way whenever we turn off pacing
12438 	 * or switch off this stack, we will know to go restore
12439 	 * the saved values.
12440 	 */
12441 	rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12442 	rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12443 	/* We want abe like behavior as well */
12444 	rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
12445 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12446 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12447 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12448 	rack->r_ctl.roundends = tp->snd_max;
12449 	if (use_rack_rr)
12450 		rack->use_rack_rr = 1;
12451 	if (V_tcp_delack_enabled)
12452 		tp->t_delayed_ack = 1;
12453 	else
12454 		tp->t_delayed_ack = 0;
12455 #ifdef TCP_ACCOUNTING
12456 	if (rack_tcp_accounting) {
12457 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12458 	}
12459 #endif
12460 	if (rack_enable_shared_cwnd)
12461 		rack->rack_enable_scwnd = 1;
12462 	rack->rc_user_set_max_segs = rack_hptsi_segments;
12463 	rack->rc_force_max_seg = 0;
12464 	if (rack_use_imac_dack)
12465 		rack->rc_dack_mode = 1;
12466 	TAILQ_INIT(&rack->r_ctl.opt_list);
12467 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12468 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12469 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12470 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12471 	rack->r_ctl.rc_highest_us_rtt = 0;
12472 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12473 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12474 	if (rack_use_cmp_acks)
12475 		rack->r_use_cmp_ack = 1;
12476 	if (rack_disable_prr)
12477 		rack->rack_no_prr = 1;
12478 	if (rack_gp_no_rec_chg)
12479 		rack->rc_gp_no_rec_chg = 1;
12480 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12481 		rack->rc_always_pace = 1;
12482 		if (rack->use_fixed_rate || rack->gp_ready)
12483 			rack_set_cc_pacing(rack);
12484 	} else
12485 		rack->rc_always_pace = 0;
12486 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12487 		rack->r_mbuf_queue = 1;
12488 	else
12489 		rack->r_mbuf_queue = 0;
12490 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12491 		inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12492 	else
12493 		inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12494 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
12495 	if (rack_limits_scwnd)
12496 		rack->r_limit_scw = 1;
12497 	else
12498 		rack->r_limit_scw = 0;
12499 	rack->rc_labc = V_tcp_abc_l_var;
12500 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12501 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12502 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12503 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12504 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12505 	rack->r_ctl.rc_min_to = rack_min_to;
12506 	microuptime(&rack->r_ctl.act_rcv_time);
12507 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12508 	rack->rc_init_win = rack_default_init_window;
12509 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12510 	if (rack_hw_up_only)
12511 		rack->r_up_only = 1;
12512 	if (rack_do_dyn_mul) {
12513 		/* When dynamic adjustment is on CA needs to start at 100% */
12514 		rack->rc_gp_dyn_mul = 1;
12515 		if (rack_do_dyn_mul >= 100)
12516 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12517 	} else
12518 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12519 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12520 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12521 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12522 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12523 				rack_probertt_filter_life);
12524 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12525 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12526 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
12527 	rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12528 	rack->r_ctl.rc_time_probertt_starts = 0;
12529 	if (rack_dsack_std_based & 0x1) {
12530 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
12531 		rack->rc_rack_tmr_std_based = 1;
12532 	}
12533 	if (rack_dsack_std_based & 0x2) {
12534 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
12535 		rack->rc_rack_use_dsack = 1;
12536 	}
12537 	/* We require at least one measurement, even if the sysctl is 0 */
12538 	if (rack_req_measurements)
12539 		rack->r_ctl.req_measurements = rack_req_measurements;
12540 	else
12541 		rack->r_ctl.req_measurements = 1;
12542 	if (rack_enable_hw_pacing)
12543 		rack->rack_hdw_pace_ena = 1;
12544 	if (rack_hw_rate_caps)
12545 		rack->r_rack_hw_rate_caps = 1;
12546 	/* Do we force on detection? */
12547 #ifdef NETFLIX_EXP_DETECTION
12548 	if (tcp_force_detection)
12549 		rack->do_detection = 1;
12550 	else
12551 #endif
12552 		rack->do_detection = 0;
12553 	if (rack_non_rxt_use_cr)
12554 		rack->rack_rec_nonrxt_use_cr = 1;
12555 	err = rack_init_fsb(tp, rack);
12556 	if (err) {
12557 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12558 		tp->t_fb_ptr = NULL;
12559 		return (err);
12560 	}
12561 	if (tp->snd_una != tp->snd_max) {
12562 		/* Create a send map for the current outstanding data */
12563 		struct rack_sendmap *rsm;
12564 
12565 		rsm = rack_alloc(rack);
12566 		if (rsm == NULL) {
12567 			uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12568 			tp->t_fb_ptr = NULL;
12569 			return (ENOMEM);
12570 		}
12571 		rsm->r_no_rtt_allowed = 1;
12572 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12573 		rsm->r_rtr_cnt = 1;
12574 		rsm->r_rtr_bytes = 0;
12575 		if (tp->t_flags & TF_SENTFIN)
12576 			rsm->r_flags |= RACK_HAS_FIN;
12577 		if ((tp->snd_una == tp->iss) &&
12578 		    !TCPS_HAVEESTABLISHED(tp->t_state))
12579 			rsm->r_flags |= RACK_HAS_SYN;
12580 		rsm->r_start = tp->snd_una;
12581 		rsm->r_end = tp->snd_max;
12582 		rsm->r_dupack = 0;
12583 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12584 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12585 			if (rsm->m)
12586 				rsm->orig_m_len = rsm->m->m_len;
12587 			else
12588 				rsm->orig_m_len = 0;
12589 		} else {
12590 			/*
12591 			 * This can happen if we have a stand-alone FIN or
12592 			 *  SYN.
12593 			 */
12594 			rsm->m = NULL;
12595 			rsm->orig_m_len = 0;
12596 			rsm->soff = 0;
12597 		}
12598 #ifndef INVARIANTS
12599 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12600 #else
12601 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12602 		if (insret != NULL) {
12603 			panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12604 			      insret, rack, rsm);
12605 		}
12606 #endif
12607 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12608 		rsm->r_in_tmap = 1;
12609 	}
12610 	/*
12611 	 * Timers in Rack are kept in microseconds so lets
12612 	 * convert any initial incoming variables
12613 	 * from ticks into usecs. Note that we
12614 	 * also change the values of t_srtt and t_rttvar, if
12615 	 * they are non-zero. They are kept with a 5
12616 	 * bit decimal so we have to carefully convert
12617 	 * these to get the full precision.
12618 	 */
12619 	rack_convert_rtts(tp);
12620 	tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12621 	if (rack_do_hystart) {
12622 		tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
12623 		if (rack_do_hystart > 1)
12624 			tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
12625 		if (rack_do_hystart > 2)
12626 			tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
12627 	}
12628 	if (rack_def_profile)
12629 		rack_set_profile(rack, rack_def_profile);
12630 	/* Cancel the GP measurement in progress */
12631 	tp->t_flags &= ~TF_GPUTINPROG;
12632 	if (SEQ_GT(tp->snd_max, tp->iss))
12633 		snt = tp->snd_max - tp->iss;
12634 	else
12635 		snt = 0;
12636 	iwin = rc_init_window(rack);
12637 	if (snt < iwin) {
12638 		/* We are not past the initial window
12639 		 * so we need to make sure cwnd is
12640 		 * correct.
12641 		 */
12642 		if (tp->snd_cwnd < iwin)
12643 			tp->snd_cwnd = iwin;
12644 		/*
12645 		 * If we are within the initial window
12646 		 * we want ssthresh to be unlimited. Setting
12647 		 * it to the rwnd (which the default stack does
12648 		 * and older racks) is not really a good idea
12649 		 * since we want to be in SS and grow both the
12650 		 * cwnd and the rwnd (via dynamic rwnd growth). If
12651 		 * we set it to the rwnd then as the peer grows its
12652 		 * rwnd we will be stuck in CA and never hit SS.
12653 		 *
12654 		 * Its far better to raise it up high (this takes the
12655 		 * risk that there as been a loss already, probably
12656 		 * we should have an indicator in all stacks of loss
12657 		 * but we don't), but considering the normal use this
12658 		 * is a risk worth taking. The consequences of not
12659 		 * hitting SS are far worse than going one more time
12660 		 * into it early on (before we have sent even a IW).
12661 		 * It is highly unlikely that we will have had a loss
12662 		 * before getting the IW out.
12663 		 */
12664 		tp->snd_ssthresh = 0xffffffff;
12665 	}
12666 	rack_stop_all_timers(tp);
12667 	/* Lets setup the fsb block */
12668 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12669 	rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12670 			     __LINE__, RACK_RTTS_INIT);
12671 	return (0);
12672 }
12673 
12674 static int
12675 rack_handoff_ok(struct tcpcb *tp)
12676 {
12677 	if ((tp->t_state == TCPS_CLOSED) ||
12678 	    (tp->t_state == TCPS_LISTEN)) {
12679 		/* Sure no problem though it may not stick */
12680 		return (0);
12681 	}
12682 	if ((tp->t_state == TCPS_SYN_SENT) ||
12683 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
12684 		/*
12685 		 * We really don't know if you support sack,
12686 		 * you have to get to ESTAB or beyond to tell.
12687 		 */
12688 		return (EAGAIN);
12689 	}
12690 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12691 		/*
12692 		 * Rack will only send a FIN after all data is acknowledged.
12693 		 * So in this case we have more data outstanding. We can't
12694 		 * switch stacks until either all data and only the FIN
12695 		 * is left (in which case rack_init() now knows how
12696 		 * to deal with that) <or> all is acknowledged and we
12697 		 * are only left with incoming data, though why you
12698 		 * would want to switch to rack after all data is acknowledged
12699 		 * I have no idea (rrs)!
12700 		 */
12701 		return (EAGAIN);
12702 	}
12703 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12704 		return (0);
12705 	}
12706 	/*
12707 	 * If we reach here we don't do SACK on this connection so we can
12708 	 * never do rack.
12709 	 */
12710 	return (EINVAL);
12711 }
12712 
12713 
12714 static void
12715 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12716 {
12717 	struct inpcb *inp = tptoinpcb(tp);
12718 
12719 	if (tp->t_fb_ptr) {
12720 		struct tcp_rack *rack;
12721 		struct rack_sendmap *rsm, *nrsm;
12722 #ifdef INVARIANTS
12723 		struct rack_sendmap *rm;
12724 #endif
12725 
12726 		rack = (struct tcp_rack *)tp->t_fb_ptr;
12727 		if (tp->t_in_pkt) {
12728 			/*
12729 			 * It is unsafe to process the packets since a
12730 			 * reset may be lurking in them (its rare but it
12731 			 * can occur). If we were to find a RST, then we
12732 			 * would end up dropping the connection and the
12733 			 * INP lock, so when we return the caller (tcp_usrreq)
12734 			 * will blow up when it trys to unlock the inp.
12735 			 */
12736 			struct mbuf *save, *m;
12737 
12738 			m = tp->t_in_pkt;
12739 			tp->t_in_pkt = NULL;
12740 			tp->t_tail_pkt = NULL;
12741 			while (m) {
12742 				save = m->m_nextpkt;
12743 				m->m_nextpkt = NULL;
12744 				m_freem(m);
12745 				m = save;
12746 			}
12747 		}
12748 		tp->t_flags &= ~TF_FORCEDATA;
12749 #ifdef NETFLIX_SHARED_CWND
12750 		if (rack->r_ctl.rc_scw) {
12751 			uint32_t limit;
12752 
12753 			if (rack->r_limit_scw)
12754 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12755 			else
12756 				limit = 0;
12757 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12758 						  rack->r_ctl.rc_scw_index,
12759 						  limit);
12760 			rack->r_ctl.rc_scw = NULL;
12761 		}
12762 #endif
12763 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
12764 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12765 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12766 			rack->r_ctl.fsb.th = NULL;
12767 		}
12768 		/* Convert back to ticks, with  */
12769 		if (tp->t_srtt > 1) {
12770 			uint32_t val, frac;
12771 
12772 			val = USEC_2_TICKS(tp->t_srtt);
12773 			frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12774 			tp->t_srtt = val << TCP_RTT_SHIFT;
12775 			/*
12776 			 * frac is the fractional part here is left
12777 			 * over from converting to hz and shifting.
12778 			 * We need to convert this to the 5 bit
12779 			 * remainder.
12780 			 */
12781 			if (frac) {
12782 				if (hz == 1000) {
12783 					frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12784 				} else {
12785 					frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12786 				}
12787 				tp->t_srtt += frac;
12788 			}
12789 		}
12790 		if (tp->t_rttvar) {
12791 			uint32_t val, frac;
12792 
12793 			val = USEC_2_TICKS(tp->t_rttvar);
12794 			frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12795 			tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12796 			/*
12797 			 * frac is the fractional part here is left
12798 			 * over from converting to hz and shifting.
12799 			 * We need to convert this to the 5 bit
12800 			 * remainder.
12801 			 */
12802 			if (frac) {
12803 				if (hz == 1000) {
12804 					frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12805 				} else {
12806 					frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12807 				}
12808 				tp->t_rttvar += frac;
12809 			}
12810 		}
12811 		tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12812 		tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12813 		if (rack->rc_always_pace) {
12814 			tcp_decrement_paced_conn();
12815 			rack_undo_cc_pacing(rack);
12816 			rack->rc_always_pace = 0;
12817 		}
12818 		/* Clean up any options if they were not applied */
12819 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12820 			struct deferred_opt_list *dol;
12821 
12822 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12823 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12824 			free(dol, M_TCPDO);
12825 		}
12826 		/* rack does not use force data but other stacks may clear it */
12827 		if (rack->r_ctl.crte != NULL) {
12828 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12829 			rack->rack_hdrw_pacing = 0;
12830 			rack->r_ctl.crte = NULL;
12831 		}
12832 #ifdef TCP_BLACKBOX
12833 		tcp_log_flowend(tp);
12834 #endif
12835 		RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12836 #ifndef INVARIANTS
12837 			(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12838 #else
12839 			rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12840 			if (rm != rsm) {
12841 				panic("At fini, rack:%p rsm:%p rm:%p",
12842 				      rack, rsm, rm);
12843 			}
12844 #endif
12845 			uma_zfree(rack_zone, rsm);
12846 		}
12847 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12848 		while (rsm) {
12849 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12850 			uma_zfree(rack_zone, rsm);
12851 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12852 		}
12853 		rack->rc_free_cnt = 0;
12854 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12855 		tp->t_fb_ptr = NULL;
12856 	}
12857 	inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12858 	inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12859 	inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12860 	inp->inp_flags2 &= ~INP_MBUF_ACKCMP;
12861 	/* Cancel the GP measurement in progress */
12862 	tp->t_flags &= ~TF_GPUTINPROG;
12863 	inp->inp_flags2 &= ~INP_MBUF_L_ACKS;
12864 	/* Make sure snd_nxt is correctly set */
12865 	tp->snd_nxt = tp->snd_max;
12866 }
12867 
12868 static void
12869 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12870 {
12871 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
12872 		rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
12873 	}
12874 	switch (tp->t_state) {
12875 	case TCPS_SYN_SENT:
12876 		rack->r_state = TCPS_SYN_SENT;
12877 		rack->r_substate = rack_do_syn_sent;
12878 		break;
12879 	case TCPS_SYN_RECEIVED:
12880 		rack->r_state = TCPS_SYN_RECEIVED;
12881 		rack->r_substate = rack_do_syn_recv;
12882 		break;
12883 	case TCPS_ESTABLISHED:
12884 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12885 		rack->r_state = TCPS_ESTABLISHED;
12886 		rack->r_substate = rack_do_established;
12887 		break;
12888 	case TCPS_CLOSE_WAIT:
12889 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12890 		rack->r_state = TCPS_CLOSE_WAIT;
12891 		rack->r_substate = rack_do_close_wait;
12892 		break;
12893 	case TCPS_FIN_WAIT_1:
12894 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12895 		rack->r_state = TCPS_FIN_WAIT_1;
12896 		rack->r_substate = rack_do_fin_wait_1;
12897 		break;
12898 	case TCPS_CLOSING:
12899 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12900 		rack->r_state = TCPS_CLOSING;
12901 		rack->r_substate = rack_do_closing;
12902 		break;
12903 	case TCPS_LAST_ACK:
12904 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12905 		rack->r_state = TCPS_LAST_ACK;
12906 		rack->r_substate = rack_do_lastack;
12907 		break;
12908 	case TCPS_FIN_WAIT_2:
12909 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12910 		rack->r_state = TCPS_FIN_WAIT_2;
12911 		rack->r_substate = rack_do_fin_wait_2;
12912 		break;
12913 	case TCPS_LISTEN:
12914 	case TCPS_CLOSED:
12915 	case TCPS_TIME_WAIT:
12916 	default:
12917 		break;
12918 	};
12919 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
12920 		rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
12921 
12922 }
12923 
12924 static void
12925 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
12926 {
12927 	/*
12928 	 * We received an ack, and then did not
12929 	 * call send or were bounced out due to the
12930 	 * hpts was running. Now a timer is up as well, is
12931 	 * it the right timer?
12932 	 */
12933 	struct rack_sendmap *rsm;
12934 	int tmr_up;
12935 
12936 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
12937 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
12938 		return;
12939 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
12940 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
12941 	    (tmr_up == PACE_TMR_RXT)) {
12942 		/* Should be an RXT */
12943 		return;
12944 	}
12945 	if (rsm == NULL) {
12946 		/* Nothing outstanding? */
12947 		if (tp->t_flags & TF_DELACK) {
12948 			if (tmr_up == PACE_TMR_DELACK)
12949 				/* We are supposed to have delayed ack up and we do */
12950 				return;
12951 		} else if (sbavail(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) {
12952 			/*
12953 			 * if we hit enobufs then we would expect the possibility
12954 			 * of nothing outstanding and the RXT up (and the hptsi timer).
12955 			 */
12956 			return;
12957 		} else if (((V_tcp_always_keepalive ||
12958 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
12959 			    (tp->t_state <= TCPS_CLOSING)) &&
12960 			   (tmr_up == PACE_TMR_KEEP) &&
12961 			   (tp->snd_max == tp->snd_una)) {
12962 			/* We should have keep alive up and we do */
12963 			return;
12964 		}
12965 	}
12966 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
12967 		   ((tmr_up == PACE_TMR_TLP) ||
12968 		    (tmr_up == PACE_TMR_RACK) ||
12969 		    (tmr_up == PACE_TMR_RXT))) {
12970 		/*
12971 		 * Either a Rack, TLP or RXT is fine if  we
12972 		 * have outstanding data.
12973 		 */
12974 		return;
12975 	} else if (tmr_up == PACE_TMR_DELACK) {
12976 		/*
12977 		 * If the delayed ack was going to go off
12978 		 * before the rtx/tlp/rack timer were going to
12979 		 * expire, then that would be the timer in control.
12980 		 * Note we don't check the time here trusting the
12981 		 * code is correct.
12982 		 */
12983 		return;
12984 	}
12985 	/*
12986 	 * Ok the timer originally started is not what we want now.
12987 	 * We will force the hpts to be stopped if any, and restart
12988 	 * with the slot set to what was in the saved slot.
12989 	 */
12990 	if (tcp_in_hpts(rack->rc_inp)) {
12991 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
12992 			uint32_t us_cts;
12993 
12994 			us_cts = tcp_get_usecs(NULL);
12995 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
12996 				rack->r_early = 1;
12997 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
12998 			}
12999 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
13000 		}
13001 		tcp_hpts_remove(rack->rc_inp);
13002 	}
13003 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13004 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13005 }
13006 
13007 
13008 static void
13009 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts, uint32_t high_seq)
13010 {
13011 	if ((SEQ_LT(tp->snd_wl1, seq) ||
13012 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
13013 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
13014 		/* keep track of pure window updates */
13015 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
13016 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
13017 		tp->snd_wnd = tiwin;
13018 		rack_validate_fo_sendwin_up(tp, rack);
13019 		tp->snd_wl1 = seq;
13020 		tp->snd_wl2 = ack;
13021 		if (tp->snd_wnd > tp->max_sndwnd)
13022 			tp->max_sndwnd = tp->snd_wnd;
13023 	    rack->r_wanted_output = 1;
13024 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
13025 		tp->snd_wnd = tiwin;
13026 		rack_validate_fo_sendwin_up(tp, rack);
13027 		tp->snd_wl1 = seq;
13028 		tp->snd_wl2 = ack;
13029 	} else {
13030 		/* Not a valid win update */
13031 		return;
13032 	}
13033 	/* Do we exit persists? */
13034 	if ((rack->rc_in_persist != 0) &&
13035 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13036 				rack->r_ctl.rc_pace_min_segs))) {
13037 		rack_exit_persist(tp, rack, cts);
13038 	}
13039 	/* Do we enter persists? */
13040 	if ((rack->rc_in_persist == 0) &&
13041 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13042 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
13043 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
13044 	    sbavail(&tptosocket(tp)->so_snd) &&
13045 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
13046 		/*
13047 		 * Here the rwnd is less than
13048 		 * the pacing size, we are established,
13049 		 * nothing is outstanding, and there is
13050 		 * data to send. Enter persists.
13051 		 */
13052 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
13053 	}
13054 }
13055 
13056 static void
13057 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
13058 {
13059 
13060 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13061 		struct inpcb *inp = tptoinpcb(tp);
13062 		union tcp_log_stackspecific log;
13063 		struct timeval ltv;
13064 		char tcp_hdr_buf[60];
13065 		struct tcphdr *th;
13066 		struct timespec ts;
13067 		uint32_t orig_snd_una;
13068 		uint8_t xx = 0;
13069 
13070 #ifdef NETFLIX_HTTP_LOGGING
13071 		struct http_sendfile_track *http_req;
13072 
13073 		if (SEQ_GT(ae->ack, tp->snd_una)) {
13074 			http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
13075 		} else {
13076 			http_req = tcp_http_find_req_for_seq(tp, ae->ack);
13077 		}
13078 #endif
13079 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13080 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
13081 		if (rack->rack_no_prr == 0)
13082 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13083 		else
13084 			log.u_bbr.flex1 = 0;
13085 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13086 		log.u_bbr.use_lt_bw <<= 1;
13087 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
13088 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13089 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
13090 		log.u_bbr.pkts_out = tp->t_maxseg;
13091 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13092 		log.u_bbr.flex7 = 1;
13093 		log.u_bbr.lost = ae->flags;
13094 		log.u_bbr.cwnd_gain = ackval;
13095 		log.u_bbr.pacing_gain = 0x2;
13096 		if (ae->flags & TSTMP_HDWR) {
13097 			/* Record the hardware timestamp if present */
13098 			log.u_bbr.flex3 = M_TSTMP;
13099 			ts.tv_sec = ae->timestamp / 1000000000;
13100 			ts.tv_nsec = ae->timestamp % 1000000000;
13101 			ltv.tv_sec = ts.tv_sec;
13102 			ltv.tv_usec = ts.tv_nsec / 1000;
13103 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13104 		} else if (ae->flags & TSTMP_LRO) {
13105 			/* Record the LRO the arrival timestamp */
13106 			log.u_bbr.flex3 = M_TSTMP_LRO;
13107 			ts.tv_sec = ae->timestamp / 1000000000;
13108 			ts.tv_nsec = ae->timestamp % 1000000000;
13109 			ltv.tv_sec = ts.tv_sec;
13110 			ltv.tv_usec = ts.tv_nsec / 1000;
13111 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13112 		}
13113 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13114 		/* Log the rcv time */
13115 		log.u_bbr.delRate = ae->timestamp;
13116 #ifdef NETFLIX_HTTP_LOGGING
13117 		log.u_bbr.applimited = tp->t_http_closed;
13118 		log.u_bbr.applimited <<= 8;
13119 		log.u_bbr.applimited |= tp->t_http_open;
13120 		log.u_bbr.applimited <<= 8;
13121 		log.u_bbr.applimited |= tp->t_http_req;
13122 		if (http_req) {
13123 			/* Copy out any client req info */
13124 			/* seconds */
13125 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13126 			/* useconds */
13127 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13128 			log.u_bbr.rttProp = http_req->timestamp;
13129 			log.u_bbr.cur_del_rate = http_req->start;
13130 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13131 				log.u_bbr.flex8 |= 1;
13132 			} else {
13133 				log.u_bbr.flex8 |= 2;
13134 				log.u_bbr.bw_inuse = http_req->end;
13135 			}
13136 			log.u_bbr.flex6 = http_req->start_seq;
13137 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13138 				log.u_bbr.flex8 |= 4;
13139 				log.u_bbr.epoch = http_req->end_seq;
13140 			}
13141 		}
13142 #endif
13143 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
13144 		th = (struct tcphdr *)tcp_hdr_buf;
13145 		th->th_seq = ae->seq;
13146 		th->th_ack = ae->ack;
13147 		th->th_win = ae->win;
13148 		/* Now fill in the ports */
13149 		th->th_sport = inp->inp_fport;
13150 		th->th_dport = inp->inp_lport;
13151 		tcp_set_flags(th, ae->flags);
13152 		/* Now do we have a timestamp option? */
13153 		if (ae->flags & HAS_TSTMP) {
13154 			u_char *cp;
13155 			uint32_t val;
13156 
13157 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
13158 			cp = (u_char *)(th + 1);
13159 			*cp = TCPOPT_NOP;
13160 			cp++;
13161 			*cp = TCPOPT_NOP;
13162 			cp++;
13163 			*cp = TCPOPT_TIMESTAMP;
13164 			cp++;
13165 			*cp = TCPOLEN_TIMESTAMP;
13166 			cp++;
13167 			val = htonl(ae->ts_value);
13168 			bcopy((char *)&val,
13169 			      (char *)cp, sizeof(uint32_t));
13170 			val = htonl(ae->ts_echo);
13171 			bcopy((char *)&val,
13172 			      (char *)(cp + 4), sizeof(uint32_t));
13173 		} else
13174 			th->th_off = (sizeof(struct tcphdr) >> 2);
13175 
13176 		/*
13177 		 * For sane logging we need to play a little trick.
13178 		 * If the ack were fully processed we would have moved
13179 		 * snd_una to high_seq, but since compressed acks are
13180 		 * processed in two phases, at this point (logging) snd_una
13181 		 * won't be advanced. So we would see multiple acks showing
13182 		 * the advancement. We can prevent that by "pretending" that
13183 		 * snd_una was advanced and then un-advancing it so that the
13184 		 * logging code has the right value for tlb_snd_una.
13185 		 */
13186 		if (tp->snd_una != high_seq) {
13187 			orig_snd_una = tp->snd_una;
13188 			tp->snd_una = high_seq;
13189 			xx = 1;
13190 		} else
13191 			xx = 0;
13192 		TCP_LOG_EVENTP(tp, th,
13193 			       &tptosocket(tp)->so_rcv,
13194 			       &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
13195 			       0, &log, true, &ltv);
13196 		if (xx) {
13197 			tp->snd_una = orig_snd_una;
13198 		}
13199 	}
13200 
13201 }
13202 
13203 static void
13204 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
13205 {
13206 	uint32_t us_rtt;
13207 	/*
13208 	 * A persist or keep-alive was forced out, update our
13209 	 * min rtt time. Note now worry about lost responses.
13210 	 * When a subsequent keep-alive or persist times out
13211 	 * and forced_ack is still on, then the last probe
13212 	 * was not responded to. In such cases we have a
13213 	 * sysctl that controls the behavior. Either we apply
13214 	 * the rtt but with reduced confidence (0). Or we just
13215 	 * plain don't apply the rtt estimate. Having data flow
13216 	 * will clear the probe_not_answered flag i.e. cum-ack
13217 	 * move forward <or> exiting and reentering persists.
13218 	 */
13219 
13220 	rack->forced_ack = 0;
13221 	rack->rc_tp->t_rxtshift = 0;
13222 	if ((rack->rc_in_persist &&
13223 	     (tiwin == rack->rc_tp->snd_wnd)) ||
13224 	    (rack->rc_in_persist == 0)) {
13225 		/*
13226 		 * In persists only apply the RTT update if this is
13227 		 * a response to our window probe. And that
13228 		 * means the rwnd sent must match the current
13229 		 * snd_wnd. If it does not, then we got a
13230 		 * window update ack instead. For keepalive
13231 		 * we allow the answer no matter what the window.
13232 		 *
13233 		 * Note that if the probe_not_answered is set then
13234 		 * the forced_ack_ts is the oldest one i.e. the first
13235 		 * probe sent that might have been lost. This assures
13236 		 * us that if we do calculate an RTT it is longer not
13237 		 * some short thing.
13238 		 */
13239 		if (rack->rc_in_persist)
13240 			counter_u64_add(rack_persists_acks, 1);
13241 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13242 		if (us_rtt == 0)
13243 			us_rtt = 1;
13244 		if (rack->probe_not_answered == 0) {
13245 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13246 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
13247 		} else {
13248 			/* We have a retransmitted probe here too */
13249 			if (rack_apply_rtt_with_reduced_conf) {
13250 				rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13251 				tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
13252 			}
13253 		}
13254 	}
13255 }
13256 
13257 static int
13258 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
13259 {
13260 	/*
13261 	 * Handle a "special" compressed ack mbuf. Each incoming
13262 	 * ack has only four possible dispositions:
13263 	 *
13264 	 * A) It moves the cum-ack forward
13265 	 * B) It is behind the cum-ack.
13266 	 * C) It is a window-update ack.
13267 	 * D) It is a dup-ack.
13268 	 *
13269 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
13270 	 * in the incoming mbuf. We also need to still pay attention
13271 	 * to nxt_pkt since there may be another packet after this
13272 	 * one.
13273 	 */
13274 #ifdef TCP_ACCOUNTING
13275 	uint64_t ts_val;
13276 	uint64_t rdstc;
13277 #endif
13278 	int segsiz;
13279 	struct timespec ts;
13280 	struct tcp_rack *rack;
13281 	struct tcp_ackent *ae;
13282 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
13283 	int cnt, i, did_out, ourfinisacked = 0;
13284 	struct tcpopt to_holder, *to = NULL;
13285 #ifdef TCP_ACCOUNTING
13286 	int win_up_req = 0;
13287 #endif
13288 	int nsegs = 0;
13289 	int under_pacing = 1;
13290 	int recovery = 0;
13291 #ifdef TCP_ACCOUNTING
13292 	sched_pin();
13293 #endif
13294 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13295 	if (rack->gp_ready &&
13296 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
13297 		under_pacing = 0;
13298 	else
13299 		under_pacing = 1;
13300 
13301 	if (rack->r_state != tp->t_state)
13302 		rack_set_state(tp, rack);
13303 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13304 	    (tp->t_flags & TF_GPUTINPROG)) {
13305 		/*
13306 		 * We have a goodput in progress
13307 		 * and we have entered a late state.
13308 		 * Do we have enough data in the sb
13309 		 * to handle the GPUT request?
13310 		 */
13311 		uint32_t bytes;
13312 
13313 		bytes = tp->gput_ack - tp->gput_seq;
13314 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
13315 			bytes += tp->gput_seq - tp->snd_una;
13316 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
13317 			/*
13318 			 * There are not enough bytes in the socket
13319 			 * buffer that have been sent to cover this
13320 			 * measurement. Cancel it.
13321 			 */
13322 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
13323 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
13324 						   tp->gput_seq,
13325 						   0, 0, 18, __LINE__, NULL, 0);
13326 			tp->t_flags &= ~TF_GPUTINPROG;
13327 		}
13328 	}
13329 	to = &to_holder;
13330 	to->to_flags = 0;
13331 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
13332 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
13333 	cnt = m->m_len / sizeof(struct tcp_ackent);
13334 	counter_u64_add(rack_multi_single_eq, cnt);
13335 	high_seq = tp->snd_una;
13336 	the_win = tp->snd_wnd;
13337 	win_seq = tp->snd_wl1;
13338 	win_upd_ack = tp->snd_wl2;
13339 	cts = tcp_tv_to_usectick(tv);
13340 	ms_cts = tcp_tv_to_mssectick(tv);
13341 	rack->r_ctl.rc_rcvtime = cts;
13342 	segsiz = ctf_fixed_maxseg(tp);
13343 	if ((rack->rc_gp_dyn_mul) &&
13344 	    (rack->use_fixed_rate == 0) &&
13345 	    (rack->rc_always_pace)) {
13346 		/* Check in on probertt */
13347 		rack_check_probe_rtt(rack, cts);
13348 	}
13349 	for (i = 0; i < cnt; i++) {
13350 #ifdef TCP_ACCOUNTING
13351 		ts_val = get_cyclecount();
13352 #endif
13353 		rack_clear_rate_sample(rack);
13354 		ae = ((mtod(m, struct tcp_ackent *)) + i);
13355 		/* Setup the window */
13356 		tiwin = ae->win << tp->snd_scale;
13357 		if (tiwin > rack->r_ctl.rc_high_rwnd)
13358 			rack->r_ctl.rc_high_rwnd = tiwin;
13359 		/* figure out the type of ack */
13360 		if (SEQ_LT(ae->ack, high_seq)) {
13361 			/* Case B*/
13362 			ae->ack_val_set = ACK_BEHIND;
13363 		} else if (SEQ_GT(ae->ack, high_seq)) {
13364 			/* Case A */
13365 			ae->ack_val_set = ACK_CUMACK;
13366 		} else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
13367 			/* Case D */
13368 			ae->ack_val_set = ACK_DUPACK;
13369 		} else {
13370 			/* Case C */
13371 			ae->ack_val_set = ACK_RWND;
13372 		}
13373 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
13374 		/* Validate timestamp */
13375 		if (ae->flags & HAS_TSTMP) {
13376 			/* Setup for a timestamp */
13377 			to->to_flags = TOF_TS;
13378 			ae->ts_echo -= tp->ts_offset;
13379 			to->to_tsecr = ae->ts_echo;
13380 			to->to_tsval = ae->ts_value;
13381 			/*
13382 			 * If echoed timestamp is later than the current time, fall back to
13383 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13384 			 * were used when this connection was established.
13385 			 */
13386 			if (TSTMP_GT(ae->ts_echo, ms_cts))
13387 				to->to_tsecr = 0;
13388 			if (tp->ts_recent &&
13389 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
13390 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
13391 #ifdef TCP_ACCOUNTING
13392 					rdstc = get_cyclecount();
13393 					if (rdstc > ts_val) {
13394 						counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
13395 								(rdstc - ts_val));
13396 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13397 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13398 						}
13399 					}
13400 #endif
13401 					continue;
13402 				}
13403 			}
13404 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
13405 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
13406 				tp->ts_recent_age = tcp_ts_getticks();
13407 				tp->ts_recent = ae->ts_value;
13408 			}
13409 		} else {
13410 			/* Setup for a no options */
13411 			to->to_flags = 0;
13412 		}
13413 		/* Update the rcv time and perform idle reduction possibly */
13414 		if  (tp->t_idle_reduce &&
13415 		     (tp->snd_max == tp->snd_una) &&
13416 		     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13417 			counter_u64_add(rack_input_idle_reduces, 1);
13418 			rack_cc_after_idle(rack, tp);
13419 		}
13420 		tp->t_rcvtime = ticks;
13421 		/* Now what about ECN of a chain of pure ACKs? */
13422 		if (tcp_ecn_input_segment(tp, ae->flags, 0,
13423 			tcp_packets_this_ack(tp, ae->ack),
13424 			ae->codepoint))
13425 			rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
13426 #ifdef TCP_ACCOUNTING
13427 		/* Count for the specific type of ack in */
13428 		counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
13429 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13430 			tp->tcp_cnt_counters[ae->ack_val_set]++;
13431 		}
13432 #endif
13433 		/*
13434 		 * Note how we could move up these in the determination
13435 		 * above, but we don't so that way the timestamp checks (and ECN)
13436 		 * is done first before we do any processing on the ACK.
13437 		 * The non-compressed path through the code has this
13438 		 * weakness (noted by @jtl) that it actually does some
13439 		 * processing before verifying the timestamp information.
13440 		 * We don't take that path here which is why we set
13441 		 * the ack_val_set first, do the timestamp and ecn
13442 		 * processing, and then look at what we have setup.
13443 		 */
13444 		if (ae->ack_val_set == ACK_BEHIND) {
13445 			/*
13446 			 * Case B flag reordering, if window is not closed
13447 			 * or it could be a keep-alive or persists
13448 			 */
13449 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
13450 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13451 			}
13452 		} else if (ae->ack_val_set == ACK_DUPACK) {
13453 			/* Case D */
13454 			rack_strike_dupack(rack);
13455 		} else if (ae->ack_val_set == ACK_RWND) {
13456 			/* Case C */
13457 			if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13458 				ts.tv_sec = ae->timestamp / 1000000000;
13459 				ts.tv_nsec = ae->timestamp % 1000000000;
13460 				rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13461 				rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13462 			} else {
13463 				rack->r_ctl.act_rcv_time = *tv;
13464 			}
13465 			if (rack->forced_ack) {
13466 				rack_handle_probe_response(rack, tiwin,
13467 							   tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
13468 			}
13469 #ifdef TCP_ACCOUNTING
13470 			win_up_req = 1;
13471 #endif
13472 			win_upd_ack = ae->ack;
13473 			win_seq = ae->seq;
13474 			the_win = tiwin;
13475 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13476 		} else {
13477 			/* Case A */
13478 			if (SEQ_GT(ae->ack, tp->snd_max)) {
13479 				/*
13480 				 * We just send an ack since the incoming
13481 				 * ack is beyond the largest seq we sent.
13482 				 */
13483 				if ((tp->t_flags & TF_ACKNOW) == 0) {
13484 					ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13485 					if (tp->t_flags && TF_ACKNOW)
13486 						rack->r_wanted_output = 1;
13487 				}
13488 			} else {
13489 				nsegs++;
13490 				/* If the window changed setup to update */
13491 				if (tiwin != tp->snd_wnd) {
13492 					win_upd_ack = ae->ack;
13493 					win_seq = ae->seq;
13494 					the_win = tiwin;
13495 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13496 				}
13497 #ifdef TCP_ACCOUNTING
13498 				/* Account for the acks */
13499 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13500 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13501 				}
13502 				counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13503 						(((ae->ack - high_seq) + segsiz - 1) / segsiz));
13504 #endif
13505 				high_seq = ae->ack;
13506 				if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
13507 					union tcp_log_stackspecific log;
13508 					struct timeval tv;
13509 
13510 					memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13511 					log.u_bbr.timeStamp = tcp_get_usecs(&tv);
13512 					log.u_bbr.flex1 = high_seq;
13513 					log.u_bbr.flex2 = rack->r_ctl.roundends;
13514 					log.u_bbr.flex3 = rack->r_ctl.current_round;
13515 					log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
13516 					log.u_bbr.flex8 = 8;
13517 					tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
13518 						       0, &log, false, NULL, NULL, 0, &tv);
13519 				}
13520 				/*
13521 				 * The draft (v3) calls for us to use SEQ_GEQ, but that
13522 				 * causes issues when we are just going app limited. Lets
13523 				 * instead use SEQ_GT <or> where its equal but more data
13524 				 * is outstanding.
13525 				 */
13526 				if ((SEQ_GT(high_seq, rack->r_ctl.roundends)) ||
13527 				    ((high_seq == rack->r_ctl.roundends) &&
13528 				     SEQ_GT(tp->snd_max, tp->snd_una))) {
13529 					rack->r_ctl.current_round++;
13530 					rack->r_ctl.roundends = tp->snd_max;
13531 					if (CC_ALGO(tp)->newround != NULL) {
13532 						CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
13533 					}
13534 				}
13535 				/* Setup our act_rcv_time */
13536 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13537 					ts.tv_sec = ae->timestamp / 1000000000;
13538 					ts.tv_nsec = ae->timestamp % 1000000000;
13539 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13540 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13541 				} else {
13542 					rack->r_ctl.act_rcv_time = *tv;
13543 				}
13544 				rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13545 				if (rack->rc_dsack_round_seen) {
13546 					/* Is the dsack round over? */
13547 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
13548 						/* Yes it is */
13549 						rack->rc_dsack_round_seen = 0;
13550 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
13551 					}
13552 				}
13553 			}
13554 		}
13555 		/* And lets be sure to commit the rtt measurements for this ack */
13556 		tcp_rack_xmit_timer_commit(rack, tp);
13557 #ifdef TCP_ACCOUNTING
13558 		rdstc = get_cyclecount();
13559 		if (rdstc > ts_val) {
13560 			counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13561 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13562 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13563 				if (ae->ack_val_set == ACK_CUMACK)
13564 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13565 			}
13566 		}
13567 #endif
13568 	}
13569 #ifdef TCP_ACCOUNTING
13570 	ts_val = get_cyclecount();
13571 #endif
13572 	/* Tend to any collapsed window */
13573 	if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
13574 		/* The peer collapsed the window */
13575 		rack_collapsed_window(rack, (tp->snd_max - high_seq), __LINE__);
13576 	} else if (rack->rc_has_collapsed)
13577 		rack_un_collapse_window(rack, __LINE__);
13578 	if ((rack->r_collapse_point_valid) &&
13579 	    (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
13580 		rack->r_collapse_point_valid = 0;
13581 	acked_amount = acked = (high_seq - tp->snd_una);
13582 	if (acked) {
13583 		/*
13584 		 * Clear the probe not answered flag
13585 		 * since cum-ack moved forward.
13586 		 */
13587 		rack->probe_not_answered = 0;
13588 		if (rack->sack_attack_disable == 0)
13589 			rack_do_decay(rack);
13590 		if (acked >= segsiz) {
13591 			/*
13592 			 * You only get credit for
13593 			 * MSS and greater (and you get extra
13594 			 * credit for larger cum-ack moves).
13595 			 */
13596 			int ac;
13597 
13598 			ac = acked / segsiz;
13599 			rack->r_ctl.ack_count += ac;
13600 			counter_u64_add(rack_ack_total, ac);
13601 		}
13602 		if (rack->r_ctl.ack_count > 0xfff00000) {
13603 			/*
13604 			 * reduce the number to keep us under
13605 			 * a uint32_t.
13606 			 */
13607 			rack->r_ctl.ack_count /= 2;
13608 			rack->r_ctl.sack_count /= 2;
13609 		}
13610 		if (tp->t_flags & TF_NEEDSYN) {
13611 			/*
13612 			 * T/TCP: Connection was half-synchronized, and our SYN has
13613 			 * been ACK'd (so connection is now fully synchronized).  Go
13614 			 * to non-starred state, increment snd_una for ACK of SYN,
13615 			 * and check if we can do window scaling.
13616 			 */
13617 			tp->t_flags &= ~TF_NEEDSYN;
13618 			tp->snd_una++;
13619 			acked_amount = acked = (high_seq - tp->snd_una);
13620 		}
13621 		if (acked > sbavail(&so->so_snd))
13622 			acked_amount = sbavail(&so->so_snd);
13623 #ifdef NETFLIX_EXP_DETECTION
13624 		/*
13625 		 * We only care on a cum-ack move if we are in a sack-disabled
13626 		 * state. We have already added in to the ack_count, and we never
13627 		 * would disable on a cum-ack move, so we only care to do the
13628 		 * detection if it may "undo" it, i.e. we were in disabled already.
13629 		 */
13630 		if (rack->sack_attack_disable)
13631 			rack_do_detection(tp, rack, acked_amount, segsiz);
13632 #endif
13633 		if (IN_FASTRECOVERY(tp->t_flags) &&
13634 		    (rack->rack_no_prr == 0))
13635 			rack_update_prr(tp, rack, acked_amount, high_seq);
13636 		if (IN_RECOVERY(tp->t_flags)) {
13637 			if (SEQ_LT(high_seq, tp->snd_recover) &&
13638 			    (SEQ_LT(high_seq, tp->snd_max))) {
13639 				tcp_rack_partialack(tp);
13640 			} else {
13641 				rack_post_recovery(tp, high_seq);
13642 				recovery = 1;
13643 			}
13644 		}
13645 		/* Handle the rack-log-ack part (sendmap) */
13646 		if ((sbused(&so->so_snd) == 0) &&
13647 		    (acked > acked_amount) &&
13648 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
13649 		    (tp->t_flags & TF_SENTFIN)) {
13650 			/*
13651 			 * We must be sure our fin
13652 			 * was sent and acked (we can be
13653 			 * in FIN_WAIT_1 without having
13654 			 * sent the fin).
13655 			 */
13656 			ourfinisacked = 1;
13657 			/*
13658 			 * Lets make sure snd_una is updated
13659 			 * since most likely acked_amount = 0 (it
13660 			 * should be).
13661 			 */
13662 			tp->snd_una = high_seq;
13663 		}
13664 		/* Did we make a RTO error? */
13665 		if ((tp->t_flags & TF_PREVVALID) &&
13666 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13667 			tp->t_flags &= ~TF_PREVVALID;
13668 			if (tp->t_rxtshift == 1 &&
13669 			    (int)(ticks - tp->t_badrxtwin) < 0)
13670 				rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
13671 		}
13672 		/* Handle the data in the socket buffer */
13673 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13674 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13675 		if (acked_amount > 0) {
13676 			struct mbuf *mfree;
13677 
13678 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13679 			SOCKBUF_LOCK(&so->so_snd);
13680 			mfree = sbcut_locked(&so->so_snd, acked_amount);
13681 			tp->snd_una = high_seq;
13682 			/* Note we want to hold the sb lock through the sendmap adjust */
13683 			rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13684 			/* Wake up the socket if we have room to write more */
13685 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13686 			sowwakeup_locked(so);
13687 			m_freem(mfree);
13688 		}
13689 		/* update progress */
13690 		tp->t_acktime = ticks;
13691 		rack_log_progress_event(rack, tp, tp->t_acktime,
13692 					PROGRESS_UPDATE, __LINE__);
13693 		/* Clear out shifts and such */
13694 		tp->t_rxtshift = 0;
13695 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13696 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13697 		rack->rc_tlp_in_progress = 0;
13698 		rack->r_ctl.rc_tlp_cnt_out = 0;
13699 		/* Send recover and snd_nxt must be dragged along */
13700 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
13701 			tp->snd_recover = tp->snd_una;
13702 		if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13703 			tp->snd_nxt = tp->snd_una;
13704 		/*
13705 		 * If the RXT timer is running we want to
13706 		 * stop it, so we can restart a TLP (or new RXT).
13707 		 */
13708 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13709 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13710 #ifdef NETFLIX_HTTP_LOGGING
13711 		tcp_http_check_for_comp(rack->rc_tp, high_seq);
13712 #endif
13713 		tp->snd_wl2 = high_seq;
13714 		tp->t_dupacks = 0;
13715 		if (under_pacing &&
13716 		    (rack->use_fixed_rate == 0) &&
13717 		    (rack->in_probe_rtt == 0) &&
13718 		    rack->rc_gp_dyn_mul &&
13719 		    rack->rc_always_pace) {
13720 			/* Check if we are dragging bottom */
13721 			rack_check_bottom_drag(tp, rack, so, acked);
13722 		}
13723 		if (tp->snd_una == tp->snd_max) {
13724 			tp->t_flags &= ~TF_PREVVALID;
13725 			rack->r_ctl.retran_during_recovery = 0;
13726 			rack->r_ctl.dsack_byte_cnt = 0;
13727 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13728 			if (rack->r_ctl.rc_went_idle_time == 0)
13729 				rack->r_ctl.rc_went_idle_time = 1;
13730 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13731 			if (sbavail(&tptosocket(tp)->so_snd) == 0)
13732 				tp->t_acktime = 0;
13733 			/* Set so we might enter persists... */
13734 			rack->r_wanted_output = 1;
13735 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13736 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13737 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13738 			    (sbavail(&so->so_snd) == 0) &&
13739 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13740 				/*
13741 				 * The socket was gone and the
13742 				 * peer sent data (not now in the past), time to
13743 				 * reset him.
13744 				 */
13745 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13746 				/* tcp_close will kill the inp pre-log the Reset */
13747 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13748 #ifdef TCP_ACCOUNTING
13749 				rdstc = get_cyclecount();
13750 				if (rdstc > ts_val) {
13751 					counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13752 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13753 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13754 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13755 					}
13756 				}
13757 #endif
13758 				m_freem(m);
13759 				tp = tcp_close(tp);
13760 				if (tp == NULL) {
13761 #ifdef TCP_ACCOUNTING
13762 					sched_unpin();
13763 #endif
13764 					return (1);
13765 				}
13766 				/*
13767 				 * We would normally do drop-with-reset which would
13768 				 * send back a reset. We can't since we don't have
13769 				 * all the needed bits. Instead lets arrange for
13770 				 * a call to tcp_output(). That way since we
13771 				 * are in the closed state we will generate a reset.
13772 				 *
13773 				 * Note if tcp_accounting is on we don't unpin since
13774 				 * we do that after the goto label.
13775 				 */
13776 				goto send_out_a_rst;
13777 			}
13778 			if ((sbused(&so->so_snd) == 0) &&
13779 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
13780 			    (tp->t_flags & TF_SENTFIN)) {
13781 				/*
13782 				 * If we can't receive any more data, then closing user can
13783 				 * proceed. Starting the timer is contrary to the
13784 				 * specification, but if we don't get a FIN we'll hang
13785 				 * forever.
13786 				 *
13787 				 */
13788 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13789 					soisdisconnected(so);
13790 					tcp_timer_activate(tp, TT_2MSL,
13791 							   (tcp_fast_finwait2_recycle ?
13792 							    tcp_finwait2_timeout :
13793 							    TP_MAXIDLE(tp)));
13794 				}
13795 				if (ourfinisacked == 0) {
13796 					/*
13797 					 * We don't change to fin-wait-2 if we have our fin acked
13798 					 * which means we are probably in TCPS_CLOSING.
13799 					 */
13800 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
13801 				}
13802 			}
13803 		}
13804 		/* Wake up the socket if we have room to write more */
13805 		if (sbavail(&so->so_snd)) {
13806 			rack->r_wanted_output = 1;
13807 			if (ctf_progress_timeout_check(tp, true)) {
13808 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13809 							tp, tick, PROGRESS_DROP, __LINE__);
13810 				/*
13811 				 * We cheat here and don't send a RST, we should send one
13812 				 * when the pacer drops the connection.
13813 				 */
13814 #ifdef TCP_ACCOUNTING
13815 				rdstc = get_cyclecount();
13816 				if (rdstc > ts_val) {
13817 					counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13818 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13819 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13820 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13821 					}
13822 				}
13823 				sched_unpin();
13824 #endif
13825 				(void)tcp_drop(tp, ETIMEDOUT);
13826 				m_freem(m);
13827 				return (1);
13828 			}
13829 		}
13830 		if (ourfinisacked) {
13831 			switch(tp->t_state) {
13832 			case TCPS_CLOSING:
13833 #ifdef TCP_ACCOUNTING
13834 				rdstc = get_cyclecount();
13835 				if (rdstc > ts_val) {
13836 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13837 							(rdstc - ts_val));
13838 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13839 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13840 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13841 					}
13842 				}
13843 				sched_unpin();
13844 #endif
13845 				tcp_twstart(tp);
13846 				m_freem(m);
13847 				return (1);
13848 				break;
13849 			case TCPS_LAST_ACK:
13850 #ifdef TCP_ACCOUNTING
13851 				rdstc = get_cyclecount();
13852 				if (rdstc > ts_val) {
13853 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13854 							(rdstc - ts_val));
13855 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13856 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13857 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13858 					}
13859 				}
13860 				sched_unpin();
13861 #endif
13862 				tp = tcp_close(tp);
13863 				ctf_do_drop(m, tp);
13864 				return (1);
13865 				break;
13866 			case TCPS_FIN_WAIT_1:
13867 #ifdef TCP_ACCOUNTING
13868 				rdstc = get_cyclecount();
13869 				if (rdstc > ts_val) {
13870 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13871 							(rdstc - ts_val));
13872 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13873 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13874 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13875 					}
13876 				}
13877 #endif
13878 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13879 					soisdisconnected(so);
13880 					tcp_timer_activate(tp, TT_2MSL,
13881 							   (tcp_fast_finwait2_recycle ?
13882 							    tcp_finwait2_timeout :
13883 							    TP_MAXIDLE(tp)));
13884 				}
13885 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
13886 				break;
13887 			default:
13888 				break;
13889 			}
13890 		}
13891 		if (rack->r_fast_output) {
13892 			/*
13893 			 * We re doing fast output.. can we expand that?
13894 			 */
13895 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
13896 		}
13897 #ifdef TCP_ACCOUNTING
13898 		rdstc = get_cyclecount();
13899 		if (rdstc > ts_val) {
13900 			counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13901 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13902 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13903 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13904 			}
13905 		}
13906 
13907 	} else if (win_up_req) {
13908 		rdstc = get_cyclecount();
13909 		if (rdstc > ts_val) {
13910 			counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
13911 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13912 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
13913 			}
13914 		}
13915 #endif
13916 	}
13917 	/* Now is there a next packet, if so we are done */
13918 	m_freem(m);
13919 	did_out = 0;
13920 	if (nxt_pkt) {
13921 #ifdef TCP_ACCOUNTING
13922 		sched_unpin();
13923 #endif
13924 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
13925 		return (0);
13926 	}
13927 	rack_handle_might_revert(tp, rack);
13928 	ctf_calc_rwin(so, tp);
13929 	if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
13930 	send_out_a_rst:
13931 		if (tcp_output(tp) < 0) {
13932 #ifdef TCP_ACCOUNTING
13933 			sched_unpin();
13934 #endif
13935 			return (1);
13936 		}
13937 		did_out = 1;
13938 	}
13939 	rack_free_trim(rack);
13940 #ifdef TCP_ACCOUNTING
13941 	sched_unpin();
13942 #endif
13943 	rack_timer_audit(tp, rack, &so->so_snd);
13944 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
13945 	return (0);
13946 }
13947 
13948 
13949 static int
13950 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
13951     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
13952     int32_t nxt_pkt, struct timeval *tv)
13953 {
13954 	struct inpcb *inp = tptoinpcb(tp);
13955 #ifdef TCP_ACCOUNTING
13956 	uint64_t ts_val;
13957 #endif
13958 	int32_t thflags, retval, did_out = 0;
13959 	int32_t way_out = 0;
13960 	/*
13961 	 * cts - is the current time from tv (caller gets ts) in microseconds.
13962 	 * ms_cts - is the current time from tv in milliseconds.
13963 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
13964 	 */
13965 	uint32_t cts, us_cts, ms_cts;
13966 	uint32_t tiwin, high_seq;
13967 	struct timespec ts;
13968 	struct tcpopt to;
13969 	struct tcp_rack *rack;
13970 	struct rack_sendmap *rsm;
13971 	int32_t prev_state = 0;
13972 #ifdef TCP_ACCOUNTING
13973 	int ack_val_set = 0xf;
13974 #endif
13975 	int nsegs;
13976 
13977 	NET_EPOCH_ASSERT();
13978 	INP_WLOCK_ASSERT(inp);
13979 
13980 	/*
13981 	 * tv passed from common code is from either M_TSTMP_LRO or
13982 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
13983 	 */
13984 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13985 	if (m->m_flags & M_ACKCMP) {
13986 		/*
13987 		 * All compressed ack's are ack's by definition so
13988 		 * remove any ack required flag and then do the processing.
13989 		 */
13990 		rack->rc_ack_required = 0;
13991 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
13992 	}
13993 	if (m->m_flags & M_ACKCMP) {
13994 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
13995 	}
13996 	cts = tcp_tv_to_usectick(tv);
13997 	ms_cts =  tcp_tv_to_mssectick(tv);
13998 	nsegs = m->m_pkthdr.lro_nsegs;
13999 	counter_u64_add(rack_proc_non_comp_ack, 1);
14000 	thflags = tcp_get_flags(th);
14001 #ifdef TCP_ACCOUNTING
14002 	sched_pin();
14003 	if (thflags & TH_ACK)
14004 		ts_val = get_cyclecount();
14005 #endif
14006 	if ((m->m_flags & M_TSTMP) ||
14007 	    (m->m_flags & M_TSTMP_LRO)) {
14008 		mbuf_tstmp2timespec(m, &ts);
14009 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
14010 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
14011 	} else
14012 		rack->r_ctl.act_rcv_time = *tv;
14013 	kern_prefetch(rack, &prev_state);
14014 	prev_state = 0;
14015 	/*
14016 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
14017 	 * the scale is zero.
14018 	 */
14019 	tiwin = th->th_win << tp->snd_scale;
14020 #ifdef TCP_ACCOUNTING
14021 	if (thflags & TH_ACK) {
14022 		/*
14023 		 * We have a tradeoff here. We can either do what we are
14024 		 * doing i.e. pinning to this CPU and then doing the accounting
14025 		 * <or> we could do a critical enter, setup the rdtsc and cpu
14026 		 * as in below, and then validate we are on the same CPU on
14027 		 * exit. I have choosen to not do the critical enter since
14028 		 * that often will gain you a context switch, and instead lock
14029 		 * us (line above this if) to the same CPU with sched_pin(). This
14030 		 * means we may be context switched out for a higher priority
14031 		 * interupt but we won't be moved to another CPU.
14032 		 *
14033 		 * If this occurs (which it won't very often since we most likely
14034 		 * are running this code in interupt context and only a higher
14035 		 * priority will bump us ... clock?) we will falsely add in
14036 		 * to the time the interupt processing time plus the ack processing
14037 		 * time. This is ok since its a rare event.
14038 		 */
14039 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
14040 						    ctf_fixed_maxseg(tp));
14041 	}
14042 #endif
14043 	/*
14044 	 * Parse options on any incoming segment.
14045 	 */
14046 	memset(&to, 0, sizeof(to));
14047 	tcp_dooptions(&to, (u_char *)(th + 1),
14048 	    (th->th_off << 2) - sizeof(struct tcphdr),
14049 	    (thflags & TH_SYN) ? TO_SYN : 0);
14050 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
14051 	    __func__));
14052 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
14053 	    __func__));
14054 
14055 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
14056 	    (tp->t_flags & TF_GPUTINPROG)) {
14057 		/*
14058 		 * We have a goodput in progress
14059 		 * and we have entered a late state.
14060 		 * Do we have enough data in the sb
14061 		 * to handle the GPUT request?
14062 		 */
14063 		uint32_t bytes;
14064 
14065 		bytes = tp->gput_ack - tp->gput_seq;
14066 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
14067 			bytes += tp->gput_seq - tp->snd_una;
14068 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
14069 			/*
14070 			 * There are not enough bytes in the socket
14071 			 * buffer that have been sent to cover this
14072 			 * measurement. Cancel it.
14073 			 */
14074 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
14075 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
14076 						   tp->gput_seq,
14077 						   0, 0, 18, __LINE__, NULL, 0);
14078 			tp->t_flags &= ~TF_GPUTINPROG;
14079 		}
14080 	}
14081 	high_seq = th->th_ack;
14082 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14083 		union tcp_log_stackspecific log;
14084 		struct timeval ltv;
14085 #ifdef NETFLIX_HTTP_LOGGING
14086 		struct http_sendfile_track *http_req;
14087 
14088 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
14089 			http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
14090 		} else {
14091 			http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
14092 		}
14093 #endif
14094 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14095 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
14096 		if (rack->rack_no_prr == 0)
14097 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
14098 		else
14099 			log.u_bbr.flex1 = 0;
14100 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
14101 		log.u_bbr.use_lt_bw <<= 1;
14102 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
14103 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
14104 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14105 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
14106 		log.u_bbr.flex3 = m->m_flags;
14107 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
14108 		log.u_bbr.lost = thflags;
14109 		log.u_bbr.pacing_gain = 0x1;
14110 #ifdef TCP_ACCOUNTING
14111 		log.u_bbr.cwnd_gain = ack_val_set;
14112 #endif
14113 		log.u_bbr.flex7 = 2;
14114 		if (m->m_flags & M_TSTMP) {
14115 			/* Record the hardware timestamp if present */
14116 			mbuf_tstmp2timespec(m, &ts);
14117 			ltv.tv_sec = ts.tv_sec;
14118 			ltv.tv_usec = ts.tv_nsec / 1000;
14119 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
14120 		} else if (m->m_flags & M_TSTMP_LRO) {
14121 			/* Record the LRO the arrival timestamp */
14122 			mbuf_tstmp2timespec(m, &ts);
14123 			ltv.tv_sec = ts.tv_sec;
14124 			ltv.tv_usec = ts.tv_nsec / 1000;
14125 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
14126 		}
14127 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
14128 		/* Log the rcv time */
14129 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
14130 #ifdef NETFLIX_HTTP_LOGGING
14131 		log.u_bbr.applimited = tp->t_http_closed;
14132 		log.u_bbr.applimited <<= 8;
14133 		log.u_bbr.applimited |= tp->t_http_open;
14134 		log.u_bbr.applimited <<= 8;
14135 		log.u_bbr.applimited |= tp->t_http_req;
14136 		if (http_req) {
14137 			/* Copy out any client req info */
14138 			/* seconds */
14139 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
14140 			/* useconds */
14141 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
14142 			log.u_bbr.rttProp = http_req->timestamp;
14143 			log.u_bbr.cur_del_rate = http_req->start;
14144 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
14145 				log.u_bbr.flex8 |= 1;
14146 			} else {
14147 				log.u_bbr.flex8 |= 2;
14148 				log.u_bbr.bw_inuse = http_req->end;
14149 			}
14150 			log.u_bbr.flex6 = http_req->start_seq;
14151 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
14152 				log.u_bbr.flex8 |= 4;
14153 				log.u_bbr.epoch = http_req->end_seq;
14154 			}
14155 		}
14156 #endif
14157 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
14158 		    tlen, &log, true, &ltv);
14159 	}
14160 	/* Remove ack required flag if set, we have one  */
14161 	if (thflags & TH_ACK)
14162 		rack->rc_ack_required = 0;
14163 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
14164 		way_out = 4;
14165 		retval = 0;
14166 		m_freem(m);
14167 		goto done_with_input;
14168 	}
14169 	/*
14170 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
14171 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
14172 	 */
14173 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
14174 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
14175 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14176 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14177 #ifdef TCP_ACCOUNTING
14178 		sched_unpin();
14179 #endif
14180 		return (1);
14181 	}
14182 	/*
14183 	 * If timestamps were negotiated during SYN/ACK and a
14184 	 * segment without a timestamp is received, silently drop
14185 	 * the segment, unless it is a RST segment or missing timestamps are
14186 	 * tolerated.
14187 	 * See section 3.2 of RFC 7323.
14188 	 */
14189 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
14190 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
14191 		way_out = 5;
14192 		retval = 0;
14193 		m_freem(m);
14194 		goto done_with_input;
14195 	}
14196 
14197 	/*
14198 	 * Segment received on connection. Reset idle time and keep-alive
14199 	 * timer. XXX: This should be done after segment validation to
14200 	 * ignore broken/spoofed segs.
14201 	 */
14202 	if  (tp->t_idle_reduce &&
14203 	     (tp->snd_max == tp->snd_una) &&
14204 	     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
14205 		counter_u64_add(rack_input_idle_reduces, 1);
14206 		rack_cc_after_idle(rack, tp);
14207 	}
14208 	tp->t_rcvtime = ticks;
14209 #ifdef STATS
14210 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
14211 #endif
14212 	if (tiwin > rack->r_ctl.rc_high_rwnd)
14213 		rack->r_ctl.rc_high_rwnd = tiwin;
14214 	/*
14215 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
14216 	 * this to occur after we've validated the segment.
14217 	 */
14218 	if (tcp_ecn_input_segment(tp, thflags, tlen,
14219 	    tcp_packets_this_ack(tp, th->th_ack),
14220 	    iptos))
14221 		rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
14222 
14223 	/*
14224 	 * If echoed timestamp is later than the current time, fall back to
14225 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
14226 	 * were used when this connection was established.
14227 	 */
14228 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
14229 		to.to_tsecr -= tp->ts_offset;
14230 		if (TSTMP_GT(to.to_tsecr, ms_cts))
14231 			to.to_tsecr = 0;
14232 	}
14233 
14234 	/*
14235 	 * If its the first time in we need to take care of options and
14236 	 * verify we can do SACK for rack!
14237 	 */
14238 	if (rack->r_state == 0) {
14239 		/* Should be init'd by rack_init() */
14240 		KASSERT(rack->rc_inp != NULL,
14241 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
14242 		if (rack->rc_inp == NULL) {
14243 			rack->rc_inp = inp;
14244 		}
14245 
14246 		/*
14247 		 * Process options only when we get SYN/ACK back. The SYN
14248 		 * case for incoming connections is handled in tcp_syncache.
14249 		 * According to RFC1323 the window field in a SYN (i.e., a
14250 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
14251 		 * this is traditional behavior, may need to be cleaned up.
14252 		 */
14253 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
14254 			/* Handle parallel SYN for ECN */
14255 			tcp_ecn_input_parallel_syn(tp, thflags, iptos);
14256 			if ((to.to_flags & TOF_SCALE) &&
14257 			    (tp->t_flags & TF_REQ_SCALE)) {
14258 				tp->t_flags |= TF_RCVD_SCALE;
14259 				tp->snd_scale = to.to_wscale;
14260 			} else
14261 				tp->t_flags &= ~TF_REQ_SCALE;
14262 			/*
14263 			 * Initial send window.  It will be updated with the
14264 			 * next incoming segment to the scaled value.
14265 			 */
14266 			tp->snd_wnd = th->th_win;
14267 			rack_validate_fo_sendwin_up(tp, rack);
14268 			if ((to.to_flags & TOF_TS) &&
14269 			    (tp->t_flags & TF_REQ_TSTMP)) {
14270 				tp->t_flags |= TF_RCVD_TSTMP;
14271 				tp->ts_recent = to.to_tsval;
14272 				tp->ts_recent_age = cts;
14273 			} else
14274 				tp->t_flags &= ~TF_REQ_TSTMP;
14275 			if (to.to_flags & TOF_MSS) {
14276 				tcp_mss(tp, to.to_mss);
14277 			}
14278 			if ((tp->t_flags & TF_SACK_PERMIT) &&
14279 			    (to.to_flags & TOF_SACKPERM) == 0)
14280 				tp->t_flags &= ~TF_SACK_PERMIT;
14281 			if (IS_FASTOPEN(tp->t_flags)) {
14282 				if (to.to_flags & TOF_FASTOPEN) {
14283 					uint16_t mss;
14284 
14285 					if (to.to_flags & TOF_MSS)
14286 						mss = to.to_mss;
14287 					else
14288 						if ((inp->inp_vflag & INP_IPV6) != 0)
14289 							mss = TCP6_MSS;
14290 						else
14291 							mss = TCP_MSS;
14292 					tcp_fastopen_update_cache(tp, mss,
14293 					    to.to_tfo_len, to.to_tfo_cookie);
14294 				} else
14295 					tcp_fastopen_disable_path(tp);
14296 			}
14297 		}
14298 		/*
14299 		 * At this point we are at the initial call. Here we decide
14300 		 * if we are doing RACK or not. We do this by seeing if
14301 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
14302 		 * The code now does do dup-ack counting so if you don't
14303 		 * switch back you won't get rack & TLP, but you will still
14304 		 * get this stack.
14305 		 */
14306 
14307 		if ((rack_sack_not_required == 0) &&
14308 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
14309 			tcp_switch_back_to_default(tp);
14310 			(*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
14311 			    tlen, iptos);
14312 #ifdef TCP_ACCOUNTING
14313 			sched_unpin();
14314 #endif
14315 			return (1);
14316 		}
14317 		tcp_set_hpts(inp);
14318 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
14319 	}
14320 	if (thflags & TH_FIN)
14321 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
14322 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14323 	if ((rack->rc_gp_dyn_mul) &&
14324 	    (rack->use_fixed_rate == 0) &&
14325 	    (rack->rc_always_pace)) {
14326 		/* Check in on probertt */
14327 		rack_check_probe_rtt(rack, us_cts);
14328 	}
14329 	rack_clear_rate_sample(rack);
14330 	if ((rack->forced_ack) &&
14331 	    ((tcp_get_flags(th) & TH_RST) == 0)) {
14332 		rack_handle_probe_response(rack, tiwin, us_cts);
14333 	}
14334 	/*
14335 	 * This is the one exception case where we set the rack state
14336 	 * always. All other times (timers etc) we must have a rack-state
14337 	 * set (so we assure we have done the checks above for SACK).
14338 	 */
14339 	rack->r_ctl.rc_rcvtime = cts;
14340 	if (rack->r_state != tp->t_state)
14341 		rack_set_state(tp, rack);
14342 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
14343 	    (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
14344 		kern_prefetch(rsm, &prev_state);
14345 	prev_state = rack->r_state;
14346 	retval = (*rack->r_substate) (m, th, so,
14347 	    tp, &to, drop_hdrlen,
14348 	    tlen, tiwin, thflags, nxt_pkt, iptos);
14349 	if (retval == 0) {
14350 		/*
14351 		 * If retval is 1 the tcb is unlocked and most likely the tp
14352 		 * is gone.
14353 		 */
14354 		INP_WLOCK_ASSERT(inp);
14355 		if ((rack->rc_gp_dyn_mul) &&
14356 		    (rack->rc_always_pace) &&
14357 		    (rack->use_fixed_rate == 0) &&
14358 		    rack->in_probe_rtt &&
14359 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
14360 			/*
14361 			 * If we are going for target, lets recheck before
14362 			 * we output.
14363 			 */
14364 			rack_check_probe_rtt(rack, us_cts);
14365 		}
14366 		if (rack->set_pacing_done_a_iw == 0) {
14367 			/* How much has been acked? */
14368 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
14369 				/* We have enough to set in the pacing segment size */
14370 				rack->set_pacing_done_a_iw = 1;
14371 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
14372 			}
14373 		}
14374 		tcp_rack_xmit_timer_commit(rack, tp);
14375 #ifdef TCP_ACCOUNTING
14376 		/*
14377 		 * If we set the ack_val_se to what ack processing we are doing
14378 		 * we also want to track how many cycles we burned. Note
14379 		 * the bits after tcp_output we let be "free". This is because
14380 		 * we are also tracking the tcp_output times as well. Note the
14381 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
14382 		 * 0xf cannot be returned and is what we initialize it too to
14383 		 * indicate we are not doing the tabulations.
14384 		 */
14385 		if (ack_val_set != 0xf) {
14386 			uint64_t crtsc;
14387 
14388 			crtsc = get_cyclecount();
14389 			counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14390 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14391 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
14392 			}
14393 		}
14394 #endif
14395 		if (nxt_pkt == 0) {
14396 			if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14397 do_output_now:
14398 				if (tcp_output(tp) < 0)
14399 					return (1);
14400 				did_out = 1;
14401 			}
14402 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
14403 			rack_free_trim(rack);
14404 		}
14405 		/* Update any rounds needed */
14406 		if (rack_verbose_logging &&  (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
14407 			union tcp_log_stackspecific log;
14408 			struct timeval tv;
14409 
14410 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14411 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14412 			log.u_bbr.flex1 = high_seq;
14413 			log.u_bbr.flex2 = rack->r_ctl.roundends;
14414 			log.u_bbr.flex3 = rack->r_ctl.current_round;
14415 			log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
14416 			log.u_bbr.flex8 = 9;
14417 			tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
14418 				       0, &log, false, NULL, NULL, 0, &tv);
14419 		}
14420 		/*
14421 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
14422 		 * causes issues when we are just going app limited. Lets
14423 		 * instead use SEQ_GT <or> where its equal but more data
14424 		 * is outstanding.
14425 		 */
14426 		if ((SEQ_GT(tp->snd_una, rack->r_ctl.roundends)) ||
14427 		    ((tp->snd_una == rack->r_ctl.roundends) && SEQ_GT(tp->snd_max, tp->snd_una))) {
14428 			rack->r_ctl.current_round++;
14429 			rack->r_ctl.roundends = tp->snd_max;
14430 			if (CC_ALGO(tp)->newround != NULL) {
14431 				CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
14432 			}
14433 		}
14434 		if ((nxt_pkt == 0) &&
14435 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
14436 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
14437 		     (tp->t_flags & TF_DELACK) ||
14438 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
14439 		      (tp->t_state <= TCPS_CLOSING)))) {
14440 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
14441 			if ((tp->snd_max == tp->snd_una) &&
14442 			    ((tp->t_flags & TF_DELACK) == 0) &&
14443 			    (tcp_in_hpts(rack->rc_inp)) &&
14444 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
14445 				/* keep alive not needed if we are hptsi output yet */
14446 				;
14447 			} else {
14448 				int late = 0;
14449 				if (tcp_in_hpts(inp)) {
14450 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14451 						us_cts = tcp_get_usecs(NULL);
14452 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
14453 							rack->r_early = 1;
14454 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
14455 						} else
14456 							late = 1;
14457 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
14458 					}
14459 					tcp_hpts_remove(inp);
14460 				}
14461 				if (late && (did_out == 0)) {
14462 					/*
14463 					 * We are late in the sending
14464 					 * and we did not call the output
14465 					 * (this probably should not happen).
14466 					 */
14467 					goto do_output_now;
14468 				}
14469 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14470 			}
14471 			way_out = 1;
14472 		} else if (nxt_pkt == 0) {
14473 			/* Do we have the correct timer running? */
14474 			rack_timer_audit(tp, rack, &so->so_snd);
14475 			way_out = 2;
14476 		}
14477 	done_with_input:
14478 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
14479 		if (did_out)
14480 			rack->r_wanted_output = 0;
14481 #ifdef TCP_ACCOUNTING
14482 	} else {
14483 		/*
14484 		 * Track the time (see above).
14485 		 */
14486 		if (ack_val_set != 0xf) {
14487 			uint64_t crtsc;
14488 
14489 			crtsc = get_cyclecount();
14490 			counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14491 			/*
14492 			 * Note we *DO NOT* increment the per-tcb counters since
14493 			 * in the else the TP may be gone!!
14494 			 */
14495 		}
14496 #endif
14497 	}
14498 #ifdef TCP_ACCOUNTING
14499 	sched_unpin();
14500 #endif
14501 	return (retval);
14502 }
14503 
14504 void
14505 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
14506     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
14507 {
14508 	struct timeval tv;
14509 
14510 	/* First lets see if we have old packets */
14511 	if (tp->t_in_pkt) {
14512 		if (ctf_do_queued_segments(so, tp, 1)) {
14513 			m_freem(m);
14514 			return;
14515 		}
14516 	}
14517 	if (m->m_flags & M_TSTMP_LRO) {
14518 		mbuf_tstmp2timeval(m, &tv);
14519 	} else {
14520 		/* Should not be should we kassert instead? */
14521 		tcp_get_usecs(&tv);
14522 	}
14523 	if (rack_do_segment_nounlock(m, th, so, tp,
14524 				     drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14525 		INP_WUNLOCK(tptoinpcb(tp));
14526 	}
14527 }
14528 
14529 struct rack_sendmap *
14530 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14531 {
14532 	struct rack_sendmap *rsm = NULL;
14533 	int32_t idx;
14534 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
14535 
14536 	/* Return the next guy to be re-transmitted */
14537 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14538 		return (NULL);
14539 	}
14540 	if (tp->t_flags & TF_SENTFIN) {
14541 		/* retran the end FIN? */
14542 		return (NULL);
14543 	}
14544 	/* ok lets look at this one */
14545 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14546 	if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
14547 		return (rsm);
14548 	}
14549 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14550 		goto check_it;
14551 	}
14552 	rsm = rack_find_lowest_rsm(rack);
14553 	if (rsm == NULL) {
14554 		return (NULL);
14555 	}
14556 check_it:
14557 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14558 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14559 		/*
14560 		 * No sack so we automatically do the 3 strikes and
14561 		 * retransmit (no rack timer would be started).
14562 		 */
14563 
14564 		return (rsm);
14565 	}
14566 	if (rsm->r_flags & RACK_ACKED) {
14567 		return (NULL);
14568 	}
14569 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14570 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14571 		/* Its not yet ready */
14572 		return (NULL);
14573 	}
14574 	srtt = rack_grab_rtt(tp, rack);
14575 	idx = rsm->r_rtr_cnt - 1;
14576 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14577 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14578 	if ((tsused == ts_low) ||
14579 	    (TSTMP_LT(tsused, ts_low))) {
14580 		/* No time since sending */
14581 		return (NULL);
14582 	}
14583 	if ((tsused - ts_low) < thresh) {
14584 		/* It has not been long enough yet */
14585 		return (NULL);
14586 	}
14587 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14588 	    ((rsm->r_flags & RACK_SACK_PASSED) &&
14589 	     (rack->sack_attack_disable == 0))) {
14590 		/*
14591 		 * We have passed the dup-ack threshold <or>
14592 		 * a SACK has indicated this is missing.
14593 		 * Note that if you are a declared attacker
14594 		 * it is only the dup-ack threshold that
14595 		 * will cause retransmits.
14596 		 */
14597 		/* log retransmit reason */
14598 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14599 		rack->r_fast_output = 0;
14600 		return (rsm);
14601 	}
14602 	return (NULL);
14603 }
14604 
14605 static void
14606 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14607 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14608 			   int line, struct rack_sendmap *rsm, uint8_t quality)
14609 {
14610 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14611 		union tcp_log_stackspecific log;
14612 		struct timeval tv;
14613 
14614 		memset(&log, 0, sizeof(log));
14615 		log.u_bbr.flex1 = slot;
14616 		log.u_bbr.flex2 = len;
14617 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14618 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14619 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14620 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14621 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14622 		log.u_bbr.use_lt_bw <<= 1;
14623 		log.u_bbr.use_lt_bw |= rack->r_late;
14624 		log.u_bbr.use_lt_bw <<= 1;
14625 		log.u_bbr.use_lt_bw |= rack->r_early;
14626 		log.u_bbr.use_lt_bw <<= 1;
14627 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14628 		log.u_bbr.use_lt_bw <<= 1;
14629 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14630 		log.u_bbr.use_lt_bw <<= 1;
14631 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14632 		log.u_bbr.use_lt_bw <<= 1;
14633 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14634 		log.u_bbr.use_lt_bw <<= 1;
14635 		log.u_bbr.use_lt_bw |= rack->gp_ready;
14636 		log.u_bbr.pkt_epoch = line;
14637 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14638 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14639 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14640 		log.u_bbr.bw_inuse = bw_est;
14641 		log.u_bbr.delRate = bw;
14642 		if (rack->r_ctl.gp_bw == 0)
14643 			log.u_bbr.cur_del_rate = 0;
14644 		else
14645 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
14646 		log.u_bbr.rttProp = len_time;
14647 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14648 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14649 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14650 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14651 			/* We are in slow start */
14652 			log.u_bbr.flex7 = 1;
14653 		} else {
14654 			/* we are on congestion avoidance */
14655 			log.u_bbr.flex7 = 0;
14656 		}
14657 		log.u_bbr.flex8 = method;
14658 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14659 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14660 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14661 		log.u_bbr.cwnd_gain <<= 1;
14662 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14663 		log.u_bbr.cwnd_gain <<= 1;
14664 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14665 		log.u_bbr.bbr_substate = quality;
14666 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
14667 		    &rack->rc_inp->inp_socket->so_rcv,
14668 		    &rack->rc_inp->inp_socket->so_snd,
14669 		    BBR_LOG_HPTSI_CALC, 0,
14670 		    0, &log, false, &tv);
14671 	}
14672 }
14673 
14674 static uint32_t
14675 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14676 {
14677 	uint32_t new_tso, user_max;
14678 
14679 	user_max = rack->rc_user_set_max_segs * mss;
14680 	if (rack->rc_force_max_seg) {
14681 		return (user_max);
14682 	}
14683 	if (rack->use_fixed_rate &&
14684 	    ((rack->r_ctl.crte == NULL) ||
14685 	     (bw != rack->r_ctl.crte->rate))) {
14686 		/* Use the user mss since we are not exactly matched */
14687 		return (user_max);
14688 	}
14689 	new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14690 	if (new_tso > user_max)
14691 		new_tso = user_max;
14692 	return (new_tso);
14693 }
14694 
14695 static int32_t
14696 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)
14697 {
14698 	uint64_t lentim, fill_bw;
14699 
14700 	/* Lets first see if we are full, if so continue with normal rate */
14701 	rack->r_via_fill_cw = 0;
14702 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14703 		return (slot);
14704 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14705 		return (slot);
14706 	if (rack->r_ctl.rc_last_us_rtt == 0)
14707 		return (slot);
14708 	if (rack->rc_pace_fill_if_rttin_range &&
14709 	    (rack->r_ctl.rc_last_us_rtt >=
14710 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14711 		/* The rtt is huge, N * smallest, lets not fill */
14712 		return (slot);
14713 	}
14714 	/*
14715 	 * first lets calculate the b/w based on the last us-rtt
14716 	 * and the sndwnd.
14717 	 */
14718 	fill_bw = rack->r_ctl.cwnd_to_use;
14719 	/* Take the rwnd if its smaller */
14720 	if (fill_bw > rack->rc_tp->snd_wnd)
14721 		fill_bw = rack->rc_tp->snd_wnd;
14722 	if (rack->r_fill_less_agg) {
14723 		/*
14724 		 * Now take away the inflight (this will reduce our
14725 		 * aggressiveness and yeah, if we get that much out in 1RTT
14726 		 * we will have had acks come back and still be behind).
14727 		 */
14728 		fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14729 	}
14730 	/* Now lets make it into a b/w */
14731 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14732 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14733 	/* We are below the min b/w */
14734 	if (non_paced)
14735 		*rate_wanted = fill_bw;
14736 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14737 		return (slot);
14738 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14739 		fill_bw = rack->r_ctl.bw_rate_cap;
14740 	rack->r_via_fill_cw = 1;
14741 	if (rack->r_rack_hw_rate_caps &&
14742 	    (rack->r_ctl.crte != NULL)) {
14743 		uint64_t high_rate;
14744 
14745 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14746 		if (fill_bw > high_rate) {
14747 			/* We are capping bw at the highest rate table entry */
14748 			if (*rate_wanted > high_rate) {
14749 				/* The original rate was also capped */
14750 				rack->r_via_fill_cw = 0;
14751 			}
14752 			rack_log_hdwr_pacing(rack,
14753 					     fill_bw, high_rate, __LINE__,
14754 					     0, 3);
14755 			fill_bw = high_rate;
14756 			if (capped)
14757 				*capped = 1;
14758 		}
14759 	} else if ((rack->r_ctl.crte == NULL) &&
14760 		   (rack->rack_hdrw_pacing == 0) &&
14761 		   (rack->rack_hdw_pace_ena) &&
14762 		   rack->r_rack_hw_rate_caps &&
14763 		   (rack->rack_attempt_hdwr_pace == 0) &&
14764 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
14765 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14766 		/*
14767 		 * Ok we may have a first attempt that is greater than our top rate
14768 		 * lets check.
14769 		 */
14770 		uint64_t high_rate;
14771 
14772 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14773 		if (high_rate) {
14774 			if (fill_bw > high_rate) {
14775 				fill_bw = high_rate;
14776 				if (capped)
14777 					*capped = 1;
14778 			}
14779 		}
14780 	}
14781 	/*
14782 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
14783 	 * in a rtt, what does that time wise equate too?
14784 	 */
14785 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14786 	lentim /= fill_bw;
14787 	*rate_wanted = fill_bw;
14788 	if (non_paced || (lentim < slot)) {
14789 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14790 					   0, lentim, 12, __LINE__, NULL, 0);
14791 		return ((int32_t)lentim);
14792 	} else
14793 		return (slot);
14794 }
14795 
14796 static int32_t
14797 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14798 {
14799 	uint64_t srtt;
14800 	int32_t slot = 0;
14801 	int can_start_hw_pacing = 1;
14802 	int err;
14803 
14804 	if (rack->rc_always_pace == 0) {
14805 		/*
14806 		 * We use the most optimistic possible cwnd/srtt for
14807 		 * sending calculations. This will make our
14808 		 * calculation anticipate getting more through
14809 		 * quicker then possible. But thats ok we don't want
14810 		 * the peer to have a gap in data sending.
14811 		 */
14812 		uint64_t cwnd, tr_perms = 0;
14813 		int32_t reduce = 0;
14814 
14815 	old_method:
14816 		/*
14817 		 * We keep no precise pacing with the old method
14818 		 * instead we use the pacer to mitigate bursts.
14819 		 */
14820 		if (rack->r_ctl.rc_rack_min_rtt)
14821 			srtt = rack->r_ctl.rc_rack_min_rtt;
14822 		else
14823 			srtt = max(tp->t_srtt, 1);
14824 		if (rack->r_ctl.rc_rack_largest_cwnd)
14825 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14826 		else
14827 			cwnd = rack->r_ctl.cwnd_to_use;
14828 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
14829 		tr_perms = (cwnd * 1000) / srtt;
14830 		if (tr_perms == 0) {
14831 			tr_perms = ctf_fixed_maxseg(tp);
14832 		}
14833 		/*
14834 		 * Calculate how long this will take to drain, if
14835 		 * the calculation comes out to zero, thats ok we
14836 		 * will use send_a_lot to possibly spin around for
14837 		 * more increasing tot_len_this_send to the point
14838 		 * that its going to require a pace, or we hit the
14839 		 * cwnd. Which in that case we are just waiting for
14840 		 * a ACK.
14841 		 */
14842 		slot = len / tr_perms;
14843 		/* Now do we reduce the time so we don't run dry? */
14844 		if (slot && rack_slot_reduction) {
14845 			reduce = (slot / rack_slot_reduction);
14846 			if (reduce < slot) {
14847 				slot -= reduce;
14848 			} else
14849 				slot = 0;
14850 		}
14851 		slot *= HPTS_USEC_IN_MSEC;
14852 		if (rack->rc_pace_to_cwnd) {
14853 			uint64_t rate_wanted = 0;
14854 
14855 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14856 			rack->rc_ack_can_sendout_data = 1;
14857 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
14858 		} else
14859 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
14860 	} else {
14861 		uint64_t bw_est, res, lentim, rate_wanted;
14862 		uint32_t orig_val, segs, oh;
14863 		int capped = 0;
14864 		int prev_fill;
14865 
14866 		if ((rack->r_rr_config == 1) && rsm) {
14867 			return (rack->r_ctl.rc_min_to);
14868 		}
14869 		if (rack->use_fixed_rate) {
14870 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14871 		} else if ((rack->r_ctl.init_rate == 0) &&
14872 #ifdef NETFLIX_PEAKRATE
14873 			   (rack->rc_tp->t_maxpeakrate == 0) &&
14874 #endif
14875 			   (rack->r_ctl.gp_bw == 0)) {
14876 			/* no way to yet do an estimate */
14877 			bw_est = rate_wanted = 0;
14878 		} else {
14879 			bw_est = rack_get_bw(rack);
14880 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14881 		}
14882 		if ((bw_est == 0) || (rate_wanted == 0) ||
14883 		    ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14884 			/*
14885 			 * No way yet to make a b/w estimate or
14886 			 * our raise is set incorrectly.
14887 			 */
14888 			goto old_method;
14889 		}
14890 		/* We need to account for all the overheads */
14891 		segs = (len + segsiz - 1) / segsiz;
14892 		/*
14893 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
14894 		 * and how much data we put in each packet. Yes this
14895 		 * means we may be off if we are larger than 1500 bytes
14896 		 * or smaller. But this just makes us more conservative.
14897 		 */
14898 		if (rack_hw_rate_min &&
14899 		    (bw_est < rack_hw_rate_min))
14900 			can_start_hw_pacing = 0;
14901 		if (ETHERNET_SEGMENT_SIZE > segsiz)
14902 			oh = ETHERNET_SEGMENT_SIZE - segsiz;
14903 		else
14904 			oh = 0;
14905 		segs *= oh;
14906 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
14907 		res = lentim / rate_wanted;
14908 		slot = (uint32_t)res;
14909 		orig_val = rack->r_ctl.rc_pace_max_segs;
14910 		if (rack->r_ctl.crte == NULL) {
14911 			/*
14912 			 * Only do this if we are not hardware pacing
14913 			 * since if we are doing hw-pacing below we will
14914 			 * set make a call after setting up or changing
14915 			 * the rate.
14916 			 */
14917 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
14918 		} else if (rack->rc_inp->inp_snd_tag == NULL) {
14919 			/*
14920 			 * We lost our rate somehow, this can happen
14921 			 * if the interface changed underneath us.
14922 			 */
14923 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
14924 			rack->r_ctl.crte = NULL;
14925 			/* Lets re-allow attempting to setup pacing */
14926 			rack->rack_hdrw_pacing = 0;
14927 			rack->rack_attempt_hdwr_pace = 0;
14928 			rack_log_hdwr_pacing(rack,
14929 					     rate_wanted, bw_est, __LINE__,
14930 					     0, 6);
14931 		}
14932 		/* Did we change the TSO size, if so log it */
14933 		if (rack->r_ctl.rc_pace_max_segs != orig_val)
14934 			rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL, 0);
14935 		prev_fill = rack->r_via_fill_cw;
14936 		if ((rack->rc_pace_to_cwnd) &&
14937 		    (capped == 0) &&
14938 		    (rack->use_fixed_rate == 0) &&
14939 		    (rack->in_probe_rtt == 0) &&
14940 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
14941 			/*
14942 			 * We want to pace at our rate *or* faster to
14943 			 * fill the cwnd to the max if its not full.
14944 			 */
14945 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
14946 		}
14947 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
14948 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14949 			if ((rack->rack_hdw_pace_ena) &&
14950 			    (can_start_hw_pacing > 0) &&
14951 			    (rack->rack_hdrw_pacing == 0) &&
14952 			    (rack->rack_attempt_hdwr_pace == 0)) {
14953 				/*
14954 				 * Lets attempt to turn on hardware pacing
14955 				 * if we can.
14956 				 */
14957 				rack->rack_attempt_hdwr_pace = 1;
14958 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
14959 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
14960 								       rate_wanted,
14961 								       RS_PACING_GEQ,
14962 								       &err, &rack->r_ctl.crte_prev_rate);
14963 				if (rack->r_ctl.crte) {
14964 					rack->rack_hdrw_pacing = 1;
14965 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
14966 												 0, rack->r_ctl.crte,
14967 												 NULL);
14968 					rack_log_hdwr_pacing(rack,
14969 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
14970 							     err, 0);
14971 					rack->r_ctl.last_hw_bw_req = rate_wanted;
14972 				} else {
14973 					counter_u64_add(rack_hw_pace_init_fail, 1);
14974 				}
14975 			} else if (rack->rack_hdrw_pacing &&
14976 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
14977 				/* Do we need to adjust our rate? */
14978 				const struct tcp_hwrate_limit_table *nrte;
14979 
14980 				if (rack->r_up_only &&
14981 				    (rate_wanted < rack->r_ctl.crte->rate)) {
14982 					/**
14983 					 * We have four possible states here
14984 					 * having to do with the previous time
14985 					 * and this time.
14986 					 *   previous  |  this-time
14987 					 * A)     0      |     0   -- fill_cw not in the picture
14988 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
14989 					 * C)     1      |     1   -- all rates from fill_cw
14990 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
14991 					 *
14992 					 * For case A, C and D we don't allow a drop. But for
14993 					 * case B where we now our on our steady rate we do
14994 					 * allow a drop.
14995 					 *
14996 					 */
14997 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
14998 						goto done_w_hdwr;
14999 				}
15000 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
15001 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
15002 					if (rack_hw_rate_to_low &&
15003 					    (bw_est < rack_hw_rate_to_low)) {
15004 						/*
15005 						 * The pacing rate is too low for hardware, but
15006 						 * do allow hardware pacing to be restarted.
15007 						 */
15008 						rack_log_hdwr_pacing(rack,
15009 							     bw_est, rack->r_ctl.crte->rate, __LINE__,
15010 							     0, 5);
15011 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
15012 						rack->r_ctl.crte = NULL;
15013 						rack->rack_attempt_hdwr_pace = 0;
15014 						rack->rack_hdrw_pacing = 0;
15015 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15016 						goto done_w_hdwr;
15017 					}
15018 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
15019 								   rack->rc_tp,
15020 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
15021 								   rate_wanted,
15022 								   RS_PACING_GEQ,
15023 								   &err, &rack->r_ctl.crte_prev_rate);
15024 					if (nrte == NULL) {
15025 						/* Lost the rate */
15026 						rack->rack_hdrw_pacing = 0;
15027 						rack->r_ctl.crte = NULL;
15028 						rack_log_hdwr_pacing(rack,
15029 								     rate_wanted, 0, __LINE__,
15030 								     err, 1);
15031 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15032 						counter_u64_add(rack_hw_pace_lost, 1);
15033 					} else if (nrte != rack->r_ctl.crte) {
15034 						rack->r_ctl.crte = nrte;
15035 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
15036 													 segsiz, 0,
15037 													 rack->r_ctl.crte,
15038 													 NULL);
15039 						rack_log_hdwr_pacing(rack,
15040 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15041 								     err, 2);
15042 						rack->r_ctl.last_hw_bw_req = rate_wanted;
15043 					}
15044 				} else {
15045 					/* We just need to adjust the segment size */
15046 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15047 					rack_log_hdwr_pacing(rack,
15048 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15049 							     0, 4);
15050 					rack->r_ctl.last_hw_bw_req = rate_wanted;
15051 				}
15052 			}
15053 		}
15054 		if ((rack->r_ctl.crte != NULL) &&
15055 		    (rack->r_ctl.crte->rate == rate_wanted)) {
15056 			/*
15057 			 * We need to add a extra if the rates
15058 			 * are exactly matched. The idea is
15059 			 * we want the software to make sure the
15060 			 * queue is empty before adding more, this
15061 			 * gives us N MSS extra pace times where
15062 			 * N is our sysctl
15063 			 */
15064 			slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
15065 		}
15066 done_w_hdwr:
15067 		if (rack_limit_time_with_srtt &&
15068 		    (rack->use_fixed_rate == 0) &&
15069 #ifdef NETFLIX_PEAKRATE
15070 		    (rack->rc_tp->t_maxpeakrate == 0) &&
15071 #endif
15072 		    (rack->rack_hdrw_pacing == 0)) {
15073 			/*
15074 			 * Sanity check, we do not allow the pacing delay
15075 			 * to be longer than the SRTT of the path. If it is
15076 			 * a slow path, then adding a packet should increase
15077 			 * the RTT and compensate for this i.e. the srtt will
15078 			 * be greater so the allowed pacing time will be greater.
15079 			 *
15080 			 * Note this restriction is not for where a peak rate
15081 			 * is set, we are doing fixed pacing or hardware pacing.
15082 			 */
15083 			if (rack->rc_tp->t_srtt)
15084 				srtt = rack->rc_tp->t_srtt;
15085 			else
15086 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
15087 			if (srtt < (uint64_t)slot) {
15088 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
15089 				slot = srtt;
15090 			}
15091 		}
15092 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
15093 	}
15094 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
15095 		/*
15096 		 * If this rate is seeing enobufs when it
15097 		 * goes to send then either the nic is out
15098 		 * of gas or we are mis-estimating the time
15099 		 * somehow and not letting the queue empty
15100 		 * completely. Lets add to the pacing time.
15101 		 */
15102 		int hw_boost_delay;
15103 
15104 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
15105 		if (hw_boost_delay > rack_enobuf_hw_max)
15106 			hw_boost_delay = rack_enobuf_hw_max;
15107 		else if (hw_boost_delay < rack_enobuf_hw_min)
15108 			hw_boost_delay = rack_enobuf_hw_min;
15109 		slot += hw_boost_delay;
15110 	}
15111 	return (slot);
15112 }
15113 
15114 static void
15115 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
15116     tcp_seq startseq, uint32_t sb_offset)
15117 {
15118 	struct rack_sendmap *my_rsm = NULL;
15119 	struct rack_sendmap fe;
15120 
15121 	if (tp->t_state < TCPS_ESTABLISHED) {
15122 		/*
15123 		 * We don't start any measurements if we are
15124 		 * not at least established.
15125 		 */
15126 		return;
15127 	}
15128 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
15129 		/*
15130 		 * We will get no more data into the SB
15131 		 * this means we need to have the data available
15132 		 * before we start a measurement.
15133 		 */
15134 
15135 		if (sbavail(&tptosocket(tp)->so_snd) <
15136 		    max(rc_init_window(rack),
15137 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
15138 			/* Nope not enough data */
15139 			return;
15140 		}
15141 	}
15142 	tp->t_flags |= TF_GPUTINPROG;
15143 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
15144 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
15145 	tp->gput_seq = startseq;
15146 	rack->app_limited_needs_set = 0;
15147 	if (rack->in_probe_rtt)
15148 		rack->measure_saw_probe_rtt = 1;
15149 	else if ((rack->measure_saw_probe_rtt) &&
15150 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
15151 		rack->measure_saw_probe_rtt = 0;
15152 	if (rack->rc_gp_filled)
15153 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15154 	else {
15155 		/* Special case initial measurement */
15156 		struct timeval tv;
15157 
15158 		tp->gput_ts = tcp_get_usecs(&tv);
15159 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15160 	}
15161 	/*
15162 	 * We take a guess out into the future,
15163 	 * if we have no measurement and no
15164 	 * initial rate, we measure the first
15165 	 * initial-windows worth of data to
15166 	 * speed up getting some GP measurement and
15167 	 * thus start pacing.
15168 	 */
15169 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
15170 		rack->app_limited_needs_set = 1;
15171 		tp->gput_ack = startseq + max(rc_init_window(rack),
15172 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
15173 		rack_log_pacing_delay_calc(rack,
15174 					   tp->gput_seq,
15175 					   tp->gput_ack,
15176 					   0,
15177 					   tp->gput_ts,
15178 					   rack->r_ctl.rc_app_limited_cnt,
15179 					   9,
15180 					   __LINE__, NULL, 0);
15181 		return;
15182 	}
15183 	if (sb_offset) {
15184 		/*
15185 		 * We are out somewhere in the sb
15186 		 * can we use the already outstanding data?
15187 		 */
15188 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
15189 			/*
15190 			 * Yes first one is good and in this case
15191 			 * the tp->gput_ts is correctly set based on
15192 			 * the last ack that arrived (no need to
15193 			 * set things up when an ack comes in).
15194 			 */
15195 			my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
15196 			if ((my_rsm == NULL) ||
15197 			    (my_rsm->r_rtr_cnt != 1)) {
15198 				/* retransmission? */
15199 				goto use_latest;
15200 			}
15201 		} else {
15202 			if (rack->r_ctl.rc_first_appl == NULL) {
15203 				/*
15204 				 * If rc_first_appl is NULL
15205 				 * then the cnt should be 0.
15206 				 * This is probably an error, maybe
15207 				 * a KASSERT would be approprate.
15208 				 */
15209 				goto use_latest;
15210 			}
15211 			/*
15212 			 * If we have a marker pointer to the last one that is
15213 			 * app limited we can use that, but we need to set
15214 			 * things up so that when it gets ack'ed we record
15215 			 * the ack time (if its not already acked).
15216 			 */
15217 			rack->app_limited_needs_set = 1;
15218 			/*
15219 			 * We want to get to the rsm that is either
15220 			 * next with space i.e. over 1 MSS or the one
15221 			 * after that (after the app-limited).
15222 			 */
15223 			my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15224 					 rack->r_ctl.rc_first_appl);
15225 			if (my_rsm) {
15226 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
15227 					/* Have to use the next one */
15228 					my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15229 							 my_rsm);
15230 				else {
15231 					/* Use after the first MSS of it is acked */
15232 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
15233 					goto start_set;
15234 				}
15235 			}
15236 			if ((my_rsm == NULL) ||
15237 			    (my_rsm->r_rtr_cnt != 1)) {
15238 				/*
15239 				 * Either its a retransmit or
15240 				 * the last is the app-limited one.
15241 				 */
15242 				goto use_latest;
15243 			}
15244 		}
15245 		tp->gput_seq = my_rsm->r_start;
15246 start_set:
15247 		if (my_rsm->r_flags & RACK_ACKED) {
15248 			/*
15249 			 * This one has been acked use the arrival ack time
15250 			 */
15251 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15252 			rack->app_limited_needs_set = 0;
15253 		}
15254 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15255 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
15256 		rack_log_pacing_delay_calc(rack,
15257 					   tp->gput_seq,
15258 					   tp->gput_ack,
15259 					   (uint64_t)my_rsm,
15260 					   tp->gput_ts,
15261 					   rack->r_ctl.rc_app_limited_cnt,
15262 					   9,
15263 					   __LINE__, NULL, 0);
15264 		return;
15265 	}
15266 
15267 use_latest:
15268 	/*
15269 	 * We don't know how long we may have been
15270 	 * idle or if this is the first-send. Lets
15271 	 * setup the flag so we will trim off
15272 	 * the first ack'd data so we get a true
15273 	 * measurement.
15274 	 */
15275 	rack->app_limited_needs_set = 1;
15276 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
15277 	/* Find this guy so we can pull the send time */
15278 	fe.r_start = startseq;
15279 	my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
15280 	if (my_rsm) {
15281 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15282 		if (my_rsm->r_flags & RACK_ACKED) {
15283 			/*
15284 			 * Unlikely since its probably what was
15285 			 * just transmitted (but I am paranoid).
15286 			 */
15287 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15288 			rack->app_limited_needs_set = 0;
15289 		}
15290 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
15291 			/* This also is unlikely */
15292 			tp->gput_seq = my_rsm->r_start;
15293 		}
15294 	} else {
15295 		/*
15296 		 * TSNH unless we have some send-map limit,
15297 		 * and even at that it should not be hitting
15298 		 * that limit (we should have stopped sending).
15299 		 */
15300 		struct timeval tv;
15301 
15302 		microuptime(&tv);
15303 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15304 	}
15305 	rack_log_pacing_delay_calc(rack,
15306 				   tp->gput_seq,
15307 				   tp->gput_ack,
15308 				   (uint64_t)my_rsm,
15309 				   tp->gput_ts,
15310 				   rack->r_ctl.rc_app_limited_cnt,
15311 				   9, __LINE__, NULL, 0);
15312 }
15313 
15314 static inline uint32_t
15315 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
15316     uint32_t avail, int32_t sb_offset)
15317 {
15318 	uint32_t len;
15319 	uint32_t sendwin;
15320 
15321 	if (tp->snd_wnd > cwnd_to_use)
15322 		sendwin = cwnd_to_use;
15323 	else
15324 		sendwin = tp->snd_wnd;
15325 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
15326 		/* We never want to go over our peers rcv-window */
15327 		len = 0;
15328 	} else {
15329 		uint32_t flight;
15330 
15331 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15332 		if (flight >= sendwin) {
15333 			/*
15334 			 * We have in flight what we are allowed by cwnd (if
15335 			 * it was rwnd blocking it would have hit above out
15336 			 * >= tp->snd_wnd).
15337 			 */
15338 			return (0);
15339 		}
15340 		len = sendwin - flight;
15341 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
15342 			/* We would send too much (beyond the rwnd) */
15343 			len = tp->snd_wnd - ctf_outstanding(tp);
15344 		}
15345 		if ((len + sb_offset) > avail) {
15346 			/*
15347 			 * We don't have that much in the SB, how much is
15348 			 * there?
15349 			 */
15350 			len = avail - sb_offset;
15351 		}
15352 	}
15353 	return (len);
15354 }
15355 
15356 static void
15357 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
15358 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
15359 	     int rsm_is_null, int optlen, int line, uint16_t mode)
15360 {
15361 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15362 		union tcp_log_stackspecific log;
15363 		struct timeval tv;
15364 
15365 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15366 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15367 		log.u_bbr.flex1 = error;
15368 		log.u_bbr.flex2 = flags;
15369 		log.u_bbr.flex3 = rsm_is_null;
15370 		log.u_bbr.flex4 = ipoptlen;
15371 		log.u_bbr.flex5 = tp->rcv_numsacks;
15372 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15373 		log.u_bbr.flex7 = optlen;
15374 		log.u_bbr.flex8 = rack->r_fsb_inited;
15375 		log.u_bbr.applimited = rack->r_fast_output;
15376 		log.u_bbr.bw_inuse = rack_get_bw(rack);
15377 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15378 		log.u_bbr.cwnd_gain = mode;
15379 		log.u_bbr.pkts_out = orig_len;
15380 		log.u_bbr.lt_epoch = len;
15381 		log.u_bbr.delivered = line;
15382 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15383 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15384 		tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
15385 			       len, &log, false, NULL, NULL, 0, &tv);
15386 	}
15387 }
15388 
15389 
15390 static struct mbuf *
15391 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
15392 		   struct rack_fast_send_blk *fsb,
15393 		   int32_t seglimit, int32_t segsize, int hw_tls)
15394 {
15395 #ifdef KERN_TLS
15396 	struct ktls_session *tls, *ntls;
15397 #ifdef INVARIANTS
15398 	struct mbuf *start;
15399 #endif
15400 #endif
15401 	struct mbuf *m, *n, **np, *smb;
15402 	struct mbuf *top;
15403 	int32_t off, soff;
15404 	int32_t len = *plen;
15405 	int32_t fragsize;
15406 	int32_t len_cp = 0;
15407 	uint32_t mlen, frags;
15408 
15409 	soff = off = the_off;
15410 	smb = m = the_m;
15411 	np = &top;
15412 	top = NULL;
15413 #ifdef KERN_TLS
15414 	if (hw_tls && (m->m_flags & M_EXTPG))
15415 		tls = m->m_epg_tls;
15416 	else
15417 		tls = NULL;
15418 #ifdef INVARIANTS
15419 	start = m;
15420 #endif
15421 #endif
15422 	while (len > 0) {
15423 		if (m == NULL) {
15424 			*plen = len_cp;
15425 			break;
15426 		}
15427 #ifdef KERN_TLS
15428 		if (hw_tls) {
15429 			if (m->m_flags & M_EXTPG)
15430 				ntls = m->m_epg_tls;
15431 			else
15432 				ntls = NULL;
15433 
15434 			/*
15435 			 * Avoid mixing TLS records with handshake
15436 			 * data or TLS records from different
15437 			 * sessions.
15438 			 */
15439 			if (tls != ntls) {
15440 				MPASS(m != start);
15441 				*plen = len_cp;
15442 				break;
15443 			}
15444 		}
15445 #endif
15446 		mlen = min(len, m->m_len - off);
15447 		if (seglimit) {
15448 			/*
15449 			 * For M_EXTPG mbufs, add 3 segments
15450 			 * + 1 in case we are crossing page boundaries
15451 			 * + 2 in case the TLS hdr/trailer are used
15452 			 * It is cheaper to just add the segments
15453 			 * than it is to take the cache miss to look
15454 			 * at the mbuf ext_pgs state in detail.
15455 			 */
15456 			if (m->m_flags & M_EXTPG) {
15457 				fragsize = min(segsize, PAGE_SIZE);
15458 				frags = 3;
15459 			} else {
15460 				fragsize = segsize;
15461 				frags = 0;
15462 			}
15463 
15464 			/* Break if we really can't fit anymore. */
15465 			if ((frags + 1) >= seglimit) {
15466 				*plen =	len_cp;
15467 				break;
15468 			}
15469 
15470 			/*
15471 			 * Reduce size if you can't copy the whole
15472 			 * mbuf. If we can't copy the whole mbuf, also
15473 			 * adjust len so the loop will end after this
15474 			 * mbuf.
15475 			 */
15476 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
15477 				mlen = (seglimit - frags - 1) * fragsize;
15478 				len = mlen;
15479 				*plen = len_cp + len;
15480 			}
15481 			frags += howmany(mlen, fragsize);
15482 			if (frags == 0)
15483 				frags++;
15484 			seglimit -= frags;
15485 			KASSERT(seglimit > 0,
15486 			    ("%s: seglimit went too low", __func__));
15487 		}
15488 		n = m_get(M_NOWAIT, m->m_type);
15489 		*np = n;
15490 		if (n == NULL)
15491 			goto nospace;
15492 		n->m_len = mlen;
15493 		soff += mlen;
15494 		len_cp += n->m_len;
15495 		if (m->m_flags & (M_EXT|M_EXTPG)) {
15496 			n->m_data = m->m_data + off;
15497 			mb_dupcl(n, m);
15498 		} else {
15499 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
15500 			    (u_int)n->m_len);
15501 		}
15502 		len -= n->m_len;
15503 		off = 0;
15504 		m = m->m_next;
15505 		np = &n->m_next;
15506 		if (len || (soff == smb->m_len)) {
15507 			/*
15508 			 * We have more so we move forward  or
15509 			 * we have consumed the entire mbuf and
15510 			 * len has fell to 0.
15511 			 */
15512 			soff = 0;
15513 			smb = m;
15514 		}
15515 
15516 	}
15517 	if (fsb != NULL) {
15518 		fsb->m = smb;
15519 		fsb->off = soff;
15520 		if (smb) {
15521 			/*
15522 			 * Save off the size of the mbuf. We do
15523 			 * this so that we can recognize when it
15524 			 * has been trimmed by sbcut() as acks
15525 			 * come in.
15526 			 */
15527 			fsb->o_m_len = smb->m_len;
15528 		} else {
15529 			/*
15530 			 * This is the case where the next mbuf went to NULL. This
15531 			 * means with this copy we have sent everything in the sb.
15532 			 * In theory we could clear the fast_output flag, but lets
15533 			 * not since its possible that we could get more added
15534 			 * and acks that call the extend function which would let
15535 			 * us send more.
15536 			 */
15537 			fsb->o_m_len = 0;
15538 		}
15539 	}
15540 	return (top);
15541 nospace:
15542 	if (top)
15543 		m_freem(top);
15544 	return (NULL);
15545 
15546 }
15547 
15548 /*
15549  * This is a copy of m_copym(), taking the TSO segment size/limit
15550  * constraints into account, and advancing the sndptr as it goes.
15551  */
15552 static struct mbuf *
15553 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15554 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15555 {
15556 	struct mbuf *m, *n;
15557 	int32_t soff;
15558 
15559 	soff = rack->r_ctl.fsb.off;
15560 	m = rack->r_ctl.fsb.m;
15561 	if (rack->r_ctl.fsb.o_m_len > m->m_len) {
15562 		/*
15563 		 * The mbuf had the front of it chopped off by an ack
15564 		 * we need to adjust the soff/off by that difference.
15565 		 */
15566 		uint32_t delta;
15567 
15568 		delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15569 		soff -= delta;
15570 	} else if (rack->r_ctl.fsb.o_m_len < m->m_len) {
15571 		/*
15572 		 * The mbuf was expanded probably by
15573 		 * a m_compress. Just update o_m_len.
15574 		 */
15575 		rack->r_ctl.fsb.o_m_len = m->m_len;
15576 	}
15577 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15578 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15579 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15580 				 __FUNCTION__,
15581 				 rack, *plen, m, m->m_len));
15582 	/* Save off the right location before we copy and advance */
15583 	*s_soff = soff;
15584 	*s_mb = rack->r_ctl.fsb.m;
15585 	n = rack_fo_base_copym(m, soff, plen,
15586 			       &rack->r_ctl.fsb,
15587 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
15588 	return (n);
15589 }
15590 
15591 static int
15592 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15593 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
15594 {
15595 	/*
15596 	 * Enter the fast retransmit path. We are given that a sched_pin is
15597 	 * in place (if accounting is compliled in) and the cycle count taken
15598 	 * at the entry is in the ts_val. The concept her is that the rsm
15599 	 * now holds the mbuf offsets and such so we can directly transmit
15600 	 * without a lot of overhead, the len field is already set for
15601 	 * us to prohibit us from sending too much (usually its 1MSS).
15602 	 */
15603 	struct ip *ip = NULL;
15604 	struct udphdr *udp = NULL;
15605 	struct tcphdr *th = NULL;
15606 	struct mbuf *m = NULL;
15607 	struct inpcb *inp;
15608 	uint8_t *cpto;
15609 	struct tcp_log_buffer *lgb;
15610 #ifdef TCP_ACCOUNTING
15611 	uint64_t crtsc;
15612 	int cnt_thru = 1;
15613 #endif
15614 	struct tcpopt to;
15615 	u_char opt[TCP_MAXOLEN];
15616 	uint32_t hdrlen, optlen;
15617 	int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
15618 	uint16_t flags;
15619 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
15620 	uint32_t if_hw_tsomaxsegsize;
15621 
15622 #ifdef INET6
15623 	struct ip6_hdr *ip6 = NULL;
15624 
15625 	if (rack->r_is_v6) {
15626 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15627 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15628 	} else
15629 #endif				/* INET6 */
15630 	{
15631 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15632 		hdrlen = sizeof(struct tcpiphdr);
15633 	}
15634 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15635 		goto failed;
15636 	}
15637 	if (doing_tlp) {
15638 		/* Its a TLP add the flag, it may already be there but be sure */
15639 		rsm->r_flags |= RACK_TLP;
15640 	} else {
15641 		/* If it was a TLP it is not not on this retransmit */
15642 		rsm->r_flags &= ~RACK_TLP;
15643 	}
15644 	startseq = rsm->r_start;
15645 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15646 	inp = rack->rc_inp;
15647 	to.to_flags = 0;
15648 	flags = tcp_outflags[tp->t_state];
15649 	if (flags & (TH_SYN|TH_RST)) {
15650 		goto failed;
15651 	}
15652 	if (rsm->r_flags & RACK_HAS_FIN) {
15653 		/* We can't send a FIN here */
15654 		goto failed;
15655 	}
15656 	if (flags & TH_FIN) {
15657 		/* We never send a FIN */
15658 		flags &= ~TH_FIN;
15659 	}
15660 	if (tp->t_flags & TF_RCVD_TSTMP) {
15661 		to.to_tsval = ms_cts + tp->ts_offset;
15662 		to.to_tsecr = tp->ts_recent;
15663 		to.to_flags = TOF_TS;
15664 	}
15665 	optlen = tcp_addoptions(&to, opt);
15666 	hdrlen += optlen;
15667 	udp = rack->r_ctl.fsb.udp;
15668 	if (udp)
15669 		hdrlen += sizeof(struct udphdr);
15670 	if (rack->r_ctl.rc_pace_max_segs)
15671 		max_val = rack->r_ctl.rc_pace_max_segs;
15672 	else if (rack->rc_user_set_max_segs)
15673 		max_val = rack->rc_user_set_max_segs * segsiz;
15674 	else
15675 		max_val = len;
15676 	if ((tp->t_flags & TF_TSO) &&
15677 	    V_tcp_do_tso &&
15678 	    (len > segsiz) &&
15679 	    (tp->t_port == 0))
15680 		tso = 1;
15681 #ifdef INET6
15682 	if (MHLEN < hdrlen + max_linkhdr)
15683 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15684 	else
15685 #endif
15686 		m = m_gethdr(M_NOWAIT, MT_DATA);
15687 	if (m == NULL)
15688 		goto failed;
15689 	m->m_data += max_linkhdr;
15690 	m->m_len = hdrlen;
15691 	th = rack->r_ctl.fsb.th;
15692 	/* Establish the len to send */
15693 	if (len > max_val)
15694 		len = max_val;
15695 	if ((tso) && (len + optlen > tp->t_maxseg)) {
15696 		uint32_t if_hw_tsomax;
15697 		int32_t max_len;
15698 
15699 		/* extract TSO information */
15700 		if_hw_tsomax = tp->t_tsomax;
15701 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15702 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15703 		/*
15704 		 * Check if we should limit by maximum payload
15705 		 * length:
15706 		 */
15707 		if (if_hw_tsomax != 0) {
15708 			/* compute maximum TSO length */
15709 			max_len = (if_hw_tsomax - hdrlen -
15710 				   max_linkhdr);
15711 			if (max_len <= 0) {
15712 				goto failed;
15713 			} else if (len > max_len) {
15714 				len = max_len;
15715 			}
15716 		}
15717 		if (len <= segsiz) {
15718 			/*
15719 			 * In case there are too many small fragments don't
15720 			 * use TSO:
15721 			 */
15722 			tso = 0;
15723 		}
15724 	} else {
15725 		tso = 0;
15726 	}
15727 	if ((tso == 0) && (len > segsiz))
15728 		len = segsiz;
15729 	if ((len == 0) ||
15730 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
15731 		goto failed;
15732 	}
15733 	th->th_seq = htonl(rsm->r_start);
15734 	th->th_ack = htonl(tp->rcv_nxt);
15735 	/*
15736 	 * The PUSH bit should only be applied
15737 	 * if the full retransmission is made. If
15738 	 * we are sending less than this is the
15739 	 * left hand edge and should not have
15740 	 * the PUSH bit.
15741 	 */
15742 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
15743 	    (len == (rsm->r_end - rsm->r_start)))
15744 		flags |= TH_PUSH;
15745 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15746 	if (th->th_win == 0) {
15747 		tp->t_sndzerowin++;
15748 		tp->t_flags |= TF_RXWIN0SENT;
15749 	} else
15750 		tp->t_flags &= ~TF_RXWIN0SENT;
15751 	if (rsm->r_flags & RACK_TLP) {
15752 		/*
15753 		 * TLP should not count in retran count, but
15754 		 * in its own bin
15755 		 */
15756 		counter_u64_add(rack_tlp_retran, 1);
15757 		counter_u64_add(rack_tlp_retran_bytes, len);
15758 	} else {
15759 		tp->t_sndrexmitpack++;
15760 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15761 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15762 	}
15763 #ifdef STATS
15764 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15765 				 len);
15766 #endif
15767 	if (rsm->m == NULL)
15768 		goto failed;
15769 	if (rsm->orig_m_len != rsm->m->m_len) {
15770 		/* Fix up the orig_m_len and possibly the mbuf offset */
15771 		rack_adjust_orig_mlen(rsm);
15772 	}
15773 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
15774 	if (len <= segsiz) {
15775 		/*
15776 		 * Must have ran out of mbufs for the copy
15777 		 * shorten it to no longer need tso. Lets
15778 		 * not put on sendalot since we are low on
15779 		 * mbufs.
15780 		 */
15781 		tso = 0;
15782 	}
15783 	if ((m->m_next == NULL) || (len <= 0)){
15784 		goto failed;
15785 	}
15786 	if (udp) {
15787 		if (rack->r_is_v6)
15788 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
15789 		else
15790 			ulen = hdrlen + len - sizeof(struct ip);
15791 		udp->uh_ulen = htons(ulen);
15792 	}
15793 	m->m_pkthdr.rcvif = (struct ifnet *)0;
15794 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
15795 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
15796 		int ect = tcp_ecn_output_established(tp, &flags, len, true);
15797 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
15798 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
15799 		    tp->t_flags2 &= ~TF2_ECN_SND_ECE;
15800 #ifdef INET6
15801 		if (rack->r_is_v6) {
15802 		    ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
15803 		    ip6->ip6_flow |= htonl(ect << 20);
15804 		}
15805 		else
15806 #endif
15807 		{
15808 		    ip->ip_tos &= ~IPTOS_ECN_MASK;
15809 		    ip->ip_tos |= ect;
15810 		}
15811 	}
15812 	tcp_set_flags(th, flags);
15813 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
15814 #ifdef INET6
15815 	if (rack->r_is_v6) {
15816 		if (tp->t_port) {
15817 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15818 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15819 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15820 			th->th_sum = htons(0);
15821 			UDPSTAT_INC(udps_opackets);
15822 		} else {
15823 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15824 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15825 			th->th_sum = in6_cksum_pseudo(ip6,
15826 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15827 						      0);
15828 		}
15829 	}
15830 #endif
15831 #if defined(INET6) && defined(INET)
15832 	else
15833 #endif
15834 #ifdef INET
15835 	{
15836 		if (tp->t_port) {
15837 			m->m_pkthdr.csum_flags = CSUM_UDP;
15838 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15839 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15840 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15841 			th->th_sum = htons(0);
15842 			UDPSTAT_INC(udps_opackets);
15843 		} else {
15844 			m->m_pkthdr.csum_flags = CSUM_TCP;
15845 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15846 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
15847 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15848 									IPPROTO_TCP + len + optlen));
15849 		}
15850 		/* IP version must be set here for ipv4/ipv6 checking later */
15851 		KASSERT(ip->ip_v == IPVERSION,
15852 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
15853 	}
15854 #endif
15855 	if (tso) {
15856 		KASSERT(len > tp->t_maxseg - optlen,
15857 			("%s: len <= tso_segsz tp:%p", __func__, tp));
15858 		m->m_pkthdr.csum_flags |= CSUM_TSO;
15859 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15860 	}
15861 #ifdef INET6
15862 	if (rack->r_is_v6) {
15863 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15864 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15865 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15866 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15867 		else
15868 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15869 	}
15870 #endif
15871 #if defined(INET) && defined(INET6)
15872 	else
15873 #endif
15874 #ifdef INET
15875 	{
15876 		ip->ip_len = htons(m->m_pkthdr.len);
15877 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15878 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15879 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15880 			if (tp->t_port == 0 || len < V_tcp_minmss) {
15881 				ip->ip_off |= htons(IP_DF);
15882 			}
15883 		} else {
15884 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15885 		}
15886 	}
15887 #endif
15888 	/* Time to copy in our header */
15889 	cpto = mtod(m, uint8_t *);
15890 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15891 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15892 	if (optlen) {
15893 		bcopy(opt, th + 1, optlen);
15894 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
15895 	} else {
15896 		th->th_off = sizeof(struct tcphdr) >> 2;
15897 	}
15898 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15899 		union tcp_log_stackspecific log;
15900 
15901 		if (rsm->r_flags & RACK_RWND_COLLAPSED) {
15902 			rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
15903 			counter_u64_add(rack_collapsed_win_rxt, 1);
15904 			counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
15905 		}
15906 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15907 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15908 		if (rack->rack_no_prr)
15909 			log.u_bbr.flex1 = 0;
15910 		else
15911 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15912 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
15913 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
15914 		log.u_bbr.flex4 = max_val;
15915 		log.u_bbr.flex5 = 0;
15916 		/* Save off the early/late values */
15917 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15918 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
15919 		log.u_bbr.bw_inuse = rack_get_bw(rack);
15920 		if (doing_tlp == 0)
15921 			log.u_bbr.flex8 = 1;
15922 		else
15923 			log.u_bbr.flex8 = 2;
15924 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15925 		log.u_bbr.flex7 = 55;
15926 		log.u_bbr.pkts_out = tp->t_maxseg;
15927 		log.u_bbr.timeStamp = cts;
15928 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15929 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
15930 		log.u_bbr.delivered = 0;
15931 		lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15932 				     len, &log, false, NULL, NULL, 0, tv);
15933 	} else
15934 		lgb = NULL;
15935 #ifdef INET6
15936 	if (rack->r_is_v6) {
15937 		error = ip6_output(m, NULL,
15938 				   &inp->inp_route6,
15939 				   0, NULL, NULL, inp);
15940 	}
15941 	else
15942 #endif
15943 #ifdef INET
15944 	{
15945 		error = ip_output(m, NULL,
15946 				  &inp->inp_route,
15947 				  0, 0, inp);
15948 	}
15949 #endif
15950 	m = NULL;
15951 	if (lgb) {
15952 		lgb->tlb_errno = error;
15953 		lgb = NULL;
15954 	}
15955 	if (error) {
15956 		goto failed;
15957 	}
15958 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
15959 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls);
15960 	if (doing_tlp && (rack->fast_rsm_hack == 0)) {
15961 		rack->rc_tlp_in_progress = 1;
15962 		rack->r_ctl.rc_tlp_cnt_out++;
15963 	}
15964 	if (error == 0) {
15965 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
15966 		if (doing_tlp) {
15967 			rack->rc_last_sent_tlp_past_cumack = 0;
15968 			rack->rc_last_sent_tlp_seq_valid = 1;
15969 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
15970 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
15971 		}
15972 	}
15973 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
15974 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
15975 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
15976 		rack->r_ctl.retran_during_recovery += len;
15977 	{
15978 		int idx;
15979 
15980 		idx = (len / segsiz) + 3;
15981 		if (idx >= TCP_MSS_ACCT_ATIMER)
15982 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
15983 		else
15984 			counter_u64_add(rack_out_size[idx], 1);
15985 	}
15986 	if (tp->t_rtttime == 0) {
15987 		tp->t_rtttime = ticks;
15988 		tp->t_rtseq = startseq;
15989 		KMOD_TCPSTAT_INC(tcps_segstimed);
15990 	}
15991 	counter_u64_add(rack_fto_rsm_send, 1);
15992 	if (error && (error == ENOBUFS)) {
15993 		if (rack->r_ctl.crte != NULL) {
15994 			rack_trace_point(rack, RACK_TP_HWENOBUF);
15995 		} else
15996 			rack_trace_point(rack, RACK_TP_ENOBUF);
15997 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
15998 		if (rack->rc_enobuf < 0x7f)
15999 			rack->rc_enobuf++;
16000 		if (slot < (10 * HPTS_USEC_IN_MSEC))
16001 			slot = 10 * HPTS_USEC_IN_MSEC;
16002 	} else
16003 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
16004 	if ((slot == 0) ||
16005 	    (rack->rc_always_pace == 0) ||
16006 	    (rack->r_rr_config == 1)) {
16007 		/*
16008 		 * We have no pacing set or we
16009 		 * are using old-style rack or
16010 		 * we are overridden to use the old 1ms pacing.
16011 		 */
16012 		slot = rack->r_ctl.rc_min_to;
16013 	}
16014 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
16015 #ifdef TCP_ACCOUNTING
16016 	crtsc = get_cyclecount();
16017 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16018 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16019 	}
16020 	counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16021 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16022 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16023 	}
16024 	counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16025 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16026 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
16027 	}
16028 	counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
16029 	sched_unpin();
16030 #endif
16031 	return (0);
16032 failed:
16033 	if (m)
16034 		m_free(m);
16035 	return (-1);
16036 }
16037 
16038 static void
16039 rack_sndbuf_autoscale(struct tcp_rack *rack)
16040 {
16041 	/*
16042 	 * Automatic sizing of send socket buffer.  Often the send buffer
16043 	 * size is not optimally adjusted to the actual network conditions
16044 	 * at hand (delay bandwidth product).  Setting the buffer size too
16045 	 * small limits throughput on links with high bandwidth and high
16046 	 * delay (eg. trans-continental/oceanic links).  Setting the
16047 	 * buffer size too big consumes too much real kernel memory,
16048 	 * especially with many connections on busy servers.
16049 	 *
16050 	 * The criteria to step up the send buffer one notch are:
16051 	 *  1. receive window of remote host is larger than send buffer
16052 	 *     (with a fudge factor of 5/4th);
16053 	 *  2. send buffer is filled to 7/8th with data (so we actually
16054 	 *     have data to make use of it);
16055 	 *  3. send buffer fill has not hit maximal automatic size;
16056 	 *  4. our send window (slow start and cogestion controlled) is
16057 	 *     larger than sent but unacknowledged data in send buffer.
16058 	 *
16059 	 * Note that the rack version moves things much faster since
16060 	 * we want to avoid hitting cache lines in the rack_fast_output()
16061 	 * path so this is called much less often and thus moves
16062 	 * the SB forward by a percentage.
16063 	 */
16064 	struct socket *so;
16065 	struct tcpcb *tp;
16066 	uint32_t sendwin, scaleup;
16067 
16068 	tp = rack->rc_tp;
16069 	so = rack->rc_inp->inp_socket;
16070 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
16071 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
16072 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
16073 		    sbused(&so->so_snd) >=
16074 		    (so->so_snd.sb_hiwat / 8 * 7) &&
16075 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
16076 		    sendwin >= (sbused(&so->so_snd) -
16077 		    (tp->snd_nxt - tp->snd_una))) {
16078 			if (rack_autosndbuf_inc)
16079 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
16080 			else
16081 				scaleup = V_tcp_autosndbuf_inc;
16082 			if (scaleup < V_tcp_autosndbuf_inc)
16083 				scaleup = V_tcp_autosndbuf_inc;
16084 			scaleup += so->so_snd.sb_hiwat;
16085 			if (scaleup > V_tcp_autosndbuf_max)
16086 				scaleup = V_tcp_autosndbuf_max;
16087 			if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
16088 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
16089 		}
16090 	}
16091 }
16092 
16093 static int
16094 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
16095 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
16096 {
16097 	/*
16098 	 * Enter to do fast output. We are given that the sched_pin is
16099 	 * in place (if accounting is compiled in) and the cycle count taken
16100 	 * at entry is in place in ts_val. The idea here is that
16101 	 * we know how many more bytes needs to be sent (presumably either
16102 	 * during pacing or to fill the cwnd and that was greater than
16103 	 * the max-burst). We have how much to send and all the info we
16104 	 * need to just send.
16105 	 */
16106 #ifdef INET
16107 	struct ip *ip = NULL;
16108 #endif
16109 	struct udphdr *udp = NULL;
16110 	struct tcphdr *th = NULL;
16111 	struct mbuf *m, *s_mb;
16112 	struct inpcb *inp;
16113 	uint8_t *cpto;
16114 	struct tcp_log_buffer *lgb;
16115 #ifdef TCP_ACCOUNTING
16116 	uint64_t crtsc;
16117 #endif
16118 	struct tcpopt to;
16119 	u_char opt[TCP_MAXOLEN];
16120 	uint32_t hdrlen, optlen;
16121 #ifdef TCP_ACCOUNTING
16122 	int cnt_thru = 1;
16123 #endif
16124 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
16125 	uint16_t flags;
16126 	uint32_t s_soff;
16127 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
16128 	uint32_t if_hw_tsomaxsegsize;
16129 	uint16_t add_flag = RACK_SENT_FP;
16130 #ifdef INET6
16131 	struct ip6_hdr *ip6 = NULL;
16132 
16133 	if (rack->r_is_v6) {
16134 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
16135 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
16136 	} else
16137 #endif				/* INET6 */
16138 	{
16139 #ifdef INET
16140 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
16141 		hdrlen = sizeof(struct tcpiphdr);
16142 #endif
16143 	}
16144 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
16145 		m = NULL;
16146 		goto failed;
16147 	}
16148 	startseq = tp->snd_max;
16149 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16150 	inp = rack->rc_inp;
16151 	len = rack->r_ctl.fsb.left_to_send;
16152 	to.to_flags = 0;
16153 	flags = rack->r_ctl.fsb.tcp_flags;
16154 	if (tp->t_flags & TF_RCVD_TSTMP) {
16155 		to.to_tsval = ms_cts + tp->ts_offset;
16156 		to.to_tsecr = tp->ts_recent;
16157 		to.to_flags = TOF_TS;
16158 	}
16159 	optlen = tcp_addoptions(&to, opt);
16160 	hdrlen += optlen;
16161 	udp = rack->r_ctl.fsb.udp;
16162 	if (udp)
16163 		hdrlen += sizeof(struct udphdr);
16164 	if (rack->r_ctl.rc_pace_max_segs)
16165 		max_val = rack->r_ctl.rc_pace_max_segs;
16166 	else if (rack->rc_user_set_max_segs)
16167 		max_val = rack->rc_user_set_max_segs * segsiz;
16168 	else
16169 		max_val = len;
16170 	if ((tp->t_flags & TF_TSO) &&
16171 	    V_tcp_do_tso &&
16172 	    (len > segsiz) &&
16173 	    (tp->t_port == 0))
16174 		tso = 1;
16175 again:
16176 #ifdef INET6
16177 	if (MHLEN < hdrlen + max_linkhdr)
16178 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
16179 	else
16180 #endif
16181 		m = m_gethdr(M_NOWAIT, MT_DATA);
16182 	if (m == NULL)
16183 		goto failed;
16184 	m->m_data += max_linkhdr;
16185 	m->m_len = hdrlen;
16186 	th = rack->r_ctl.fsb.th;
16187 	/* Establish the len to send */
16188 	if (len > max_val)
16189 		len = max_val;
16190 	if ((tso) && (len + optlen > tp->t_maxseg)) {
16191 		uint32_t if_hw_tsomax;
16192 		int32_t max_len;
16193 
16194 		/* extract TSO information */
16195 		if_hw_tsomax = tp->t_tsomax;
16196 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
16197 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
16198 		/*
16199 		 * Check if we should limit by maximum payload
16200 		 * length:
16201 		 */
16202 		if (if_hw_tsomax != 0) {
16203 			/* compute maximum TSO length */
16204 			max_len = (if_hw_tsomax - hdrlen -
16205 				   max_linkhdr);
16206 			if (max_len <= 0) {
16207 				goto failed;
16208 			} else if (len > max_len) {
16209 				len = max_len;
16210 			}
16211 		}
16212 		if (len <= segsiz) {
16213 			/*
16214 			 * In case there are too many small fragments don't
16215 			 * use TSO:
16216 			 */
16217 			tso = 0;
16218 		}
16219 	} else {
16220 		tso = 0;
16221 	}
16222 	if ((tso == 0) && (len > segsiz))
16223 		len = segsiz;
16224 	if ((len == 0) ||
16225 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
16226 		goto failed;
16227 	}
16228 	sb_offset = tp->snd_max - tp->snd_una;
16229 	th->th_seq = htonl(tp->snd_max);
16230 	th->th_ack = htonl(tp->rcv_nxt);
16231 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
16232 	if (th->th_win == 0) {
16233 		tp->t_sndzerowin++;
16234 		tp->t_flags |= TF_RXWIN0SENT;
16235 	} else
16236 		tp->t_flags &= ~TF_RXWIN0SENT;
16237 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
16238 	KMOD_TCPSTAT_INC(tcps_sndpack);
16239 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
16240 #ifdef STATS
16241 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
16242 				 len);
16243 #endif
16244 	if (rack->r_ctl.fsb.m == NULL)
16245 		goto failed;
16246 
16247 	/* s_mb and s_soff are saved for rack_log_output */
16248 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
16249 				    &s_mb, &s_soff);
16250 	if (len <= segsiz) {
16251 		/*
16252 		 * Must have ran out of mbufs for the copy
16253 		 * shorten it to no longer need tso. Lets
16254 		 * not put on sendalot since we are low on
16255 		 * mbufs.
16256 		 */
16257 		tso = 0;
16258 	}
16259 	if (rack->r_ctl.fsb.rfo_apply_push &&
16260 	    (len == rack->r_ctl.fsb.left_to_send)) {
16261 		flags |= TH_PUSH;
16262 		add_flag |= RACK_HAD_PUSH;
16263 	}
16264 	if ((m->m_next == NULL) || (len <= 0)){
16265 		goto failed;
16266 	}
16267 	if (udp) {
16268 		if (rack->r_is_v6)
16269 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
16270 		else
16271 			ulen = hdrlen + len - sizeof(struct ip);
16272 		udp->uh_ulen = htons(ulen);
16273 	}
16274 	m->m_pkthdr.rcvif = (struct ifnet *)0;
16275 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
16276 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
16277 		int ect = tcp_ecn_output_established(tp, &flags, len, false);
16278 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
16279 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
16280 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
16281 #ifdef INET6
16282 		if (rack->r_is_v6) {
16283 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
16284 			ip6->ip6_flow |= htonl(ect << 20);
16285 		}
16286 		else
16287 #endif
16288 		{
16289 #ifdef INET
16290 			ip->ip_tos &= ~IPTOS_ECN_MASK;
16291 			ip->ip_tos |= ect;
16292 #endif
16293 		}
16294 	}
16295 	tcp_set_flags(th, flags);
16296 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
16297 #ifdef INET6
16298 	if (rack->r_is_v6) {
16299 		if (tp->t_port) {
16300 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
16301 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16302 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
16303 			th->th_sum = htons(0);
16304 			UDPSTAT_INC(udps_opackets);
16305 		} else {
16306 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
16307 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16308 			th->th_sum = in6_cksum_pseudo(ip6,
16309 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
16310 						      0);
16311 		}
16312 	}
16313 #endif
16314 #if defined(INET6) && defined(INET)
16315 	else
16316 #endif
16317 #ifdef INET
16318 	{
16319 		if (tp->t_port) {
16320 			m->m_pkthdr.csum_flags = CSUM_UDP;
16321 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16322 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
16323 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
16324 			th->th_sum = htons(0);
16325 			UDPSTAT_INC(udps_opackets);
16326 		} else {
16327 			m->m_pkthdr.csum_flags = CSUM_TCP;
16328 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16329 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
16330 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
16331 									IPPROTO_TCP + len + optlen));
16332 		}
16333 		/* IP version must be set here for ipv4/ipv6 checking later */
16334 		KASSERT(ip->ip_v == IPVERSION,
16335 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
16336 	}
16337 #endif
16338 	if (tso) {
16339 		KASSERT(len > tp->t_maxseg - optlen,
16340 			("%s: len <= tso_segsz tp:%p", __func__, tp));
16341 		m->m_pkthdr.csum_flags |= CSUM_TSO;
16342 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
16343 	}
16344 #ifdef INET6
16345 	if (rack->r_is_v6) {
16346 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
16347 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
16348 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
16349 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16350 		else
16351 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16352 	}
16353 #endif
16354 #if defined(INET) && defined(INET6)
16355 	else
16356 #endif
16357 #ifdef INET
16358 	{
16359 		ip->ip_len = htons(m->m_pkthdr.len);
16360 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
16361 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
16362 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16363 			if (tp->t_port == 0 || len < V_tcp_minmss) {
16364 				ip->ip_off |= htons(IP_DF);
16365 			}
16366 		} else {
16367 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16368 		}
16369 	}
16370 #endif
16371 	/* Time to copy in our header */
16372 	cpto = mtod(m, uint8_t *);
16373 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
16374 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
16375 	if (optlen) {
16376 		bcopy(opt, th + 1, optlen);
16377 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
16378 	} else {
16379 		th->th_off = sizeof(struct tcphdr) >> 2;
16380 	}
16381 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
16382 		union tcp_log_stackspecific log;
16383 
16384 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16385 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
16386 		if (rack->rack_no_prr)
16387 			log.u_bbr.flex1 = 0;
16388 		else
16389 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16390 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16391 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16392 		log.u_bbr.flex4 = max_val;
16393 		log.u_bbr.flex5 = 0;
16394 		/* Save off the early/late values */
16395 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16396 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16397 		log.u_bbr.bw_inuse = rack_get_bw(rack);
16398 		log.u_bbr.flex8 = 0;
16399 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16400 		log.u_bbr.flex7 = 44;
16401 		log.u_bbr.pkts_out = tp->t_maxseg;
16402 		log.u_bbr.timeStamp = cts;
16403 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16404 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16405 		log.u_bbr.delivered = 0;
16406 		lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16407 				     len, &log, false, NULL, NULL, 0, tv);
16408 	} else
16409 		lgb = NULL;
16410 #ifdef INET6
16411 	if (rack->r_is_v6) {
16412 		error = ip6_output(m, NULL,
16413 				   &inp->inp_route6,
16414 				   0, NULL, NULL, inp);
16415 	}
16416 #endif
16417 #if defined(INET) && defined(INET6)
16418 	else
16419 #endif
16420 #ifdef INET
16421 	{
16422 		error = ip_output(m, NULL,
16423 				  &inp->inp_route,
16424 				  0, 0, inp);
16425 	}
16426 #endif
16427 	if (lgb) {
16428 		lgb->tlb_errno = error;
16429 		lgb = NULL;
16430 	}
16431 	if (error) {
16432 		*send_err = error;
16433 		m = NULL;
16434 		goto failed;
16435 	}
16436 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
16437 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls);
16438 	m = NULL;
16439 	if (tp->snd_una == tp->snd_max) {
16440 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
16441 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
16442 		tp->t_acktime = ticks;
16443 	}
16444 	if (error == 0)
16445 		tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
16446 
16447 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
16448 	tot_len += len;
16449 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
16450 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
16451 	tp->snd_max += len;
16452 	tp->snd_nxt = tp->snd_max;
16453 	{
16454 		int idx;
16455 
16456 		idx = (len / segsiz) + 3;
16457 		if (idx >= TCP_MSS_ACCT_ATIMER)
16458 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16459 		else
16460 			counter_u64_add(rack_out_size[idx], 1);
16461 	}
16462 	if (len <= rack->r_ctl.fsb.left_to_send)
16463 		rack->r_ctl.fsb.left_to_send -= len;
16464 	else
16465 		rack->r_ctl.fsb.left_to_send = 0;
16466 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
16467 		rack->r_fast_output = 0;
16468 		rack->r_ctl.fsb.left_to_send = 0;
16469 		/* At the end of fast_output scale up the sb */
16470 		SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
16471 		rack_sndbuf_autoscale(rack);
16472 		SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
16473 	}
16474 	if (tp->t_rtttime == 0) {
16475 		tp->t_rtttime = ticks;
16476 		tp->t_rtseq = startseq;
16477 		KMOD_TCPSTAT_INC(tcps_segstimed);
16478 	}
16479 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
16480 	    (max_val > len) &&
16481 	    (tso == 0)) {
16482 		max_val -= len;
16483 		len = segsiz;
16484 		th = rack->r_ctl.fsb.th;
16485 #ifdef TCP_ACCOUNTING
16486 		cnt_thru++;
16487 #endif
16488 		goto again;
16489 	}
16490 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16491 	counter_u64_add(rack_fto_send, 1);
16492 	slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
16493 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
16494 #ifdef TCP_ACCOUNTING
16495 	crtsc = get_cyclecount();
16496 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16497 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16498 	}
16499 	counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16500 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16501 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16502 	}
16503 	counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16504 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16505 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
16506 	}
16507 	counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
16508 	sched_unpin();
16509 #endif
16510 	return (0);
16511 failed:
16512 	if (m)
16513 		m_free(m);
16514 	rack->r_fast_output = 0;
16515 	return (-1);
16516 }
16517 
16518 static struct rack_sendmap *
16519 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
16520 {
16521 	struct rack_sendmap *rsm = NULL;
16522 	struct rack_sendmap fe;
16523 	int thresh;
16524 
16525 restart:
16526 	fe.r_start = rack->r_ctl.last_collapse_point;
16527 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
16528 	if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
16529 		/* Nothing, strange turn off validity  */
16530 		rack->r_collapse_point_valid = 0;
16531 		return (NULL);
16532 	}
16533 	/* Can we send it yet? */
16534 	if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
16535 		/*
16536 		 * Receiver window has not grown enough for
16537 		 * the segment to be put on the wire.
16538 		 */
16539 		return (NULL);
16540 	}
16541 	if (rsm->r_flags & RACK_ACKED) {
16542 		/*
16543 		 * It has been sacked, lets move to the
16544 		 * next one if possible.
16545 		 */
16546 		rack->r_ctl.last_collapse_point = rsm->r_end;
16547 		/* Are we done? */
16548 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
16549 			    rack->r_ctl.high_collapse_point)) {
16550 			rack->r_collapse_point_valid = 0;
16551 			return (NULL);
16552 		}
16553 		goto restart;
16554 	}
16555 	/* Now has it been long enough ? */
16556 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts);
16557 	if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
16558 		rack_log_collapse(rack, rsm->r_start,
16559 				  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
16560 				  thresh, __LINE__, 6, rsm->r_flags, rsm);
16561 		return (rsm);
16562 	}
16563 	/* Not enough time */
16564 	rack_log_collapse(rack, rsm->r_start,
16565 			  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
16566 			  thresh, __LINE__, 7, rsm->r_flags, rsm);
16567 	return (NULL);
16568 }
16569 
16570 static int
16571 rack_output(struct tcpcb *tp)
16572 {
16573 	struct socket *so;
16574 	uint32_t recwin;
16575 	uint32_t sb_offset, s_moff = 0;
16576 	int32_t len, error = 0;
16577 	uint16_t flags;
16578 	struct mbuf *m, *s_mb = NULL;
16579 	struct mbuf *mb;
16580 	uint32_t if_hw_tsomaxsegcount = 0;
16581 	uint32_t if_hw_tsomaxsegsize;
16582 	int32_t segsiz, minseg;
16583 	long tot_len_this_send = 0;
16584 #ifdef INET
16585 	struct ip *ip = NULL;
16586 #endif
16587 	struct udphdr *udp = NULL;
16588 	struct tcp_rack *rack;
16589 	struct tcphdr *th;
16590 	uint8_t pass = 0;
16591 	uint8_t mark = 0;
16592 	uint8_t wanted_cookie = 0;
16593 	u_char opt[TCP_MAXOLEN];
16594 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
16595 	uint32_t rack_seq;
16596 
16597 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16598 	unsigned ipsec_optlen = 0;
16599 
16600 #endif
16601 	int32_t idle, sendalot;
16602 	int32_t sub_from_prr = 0;
16603 	volatile int32_t sack_rxmit;
16604 	struct rack_sendmap *rsm = NULL;
16605 	int32_t tso, mtu;
16606 	struct tcpopt to;
16607 	int32_t slot = 0;
16608 	int32_t sup_rack = 0;
16609 	uint32_t cts, ms_cts, delayed, early;
16610 	uint16_t add_flag = RACK_SENT_SP;
16611 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
16612 	uint8_t hpts_calling,  doing_tlp = 0;
16613 	uint32_t cwnd_to_use, pace_max_seg;
16614 	int32_t do_a_prefetch = 0;
16615 	int32_t prefetch_rsm = 0;
16616 	int32_t orig_len = 0;
16617 	struct timeval tv;
16618 	int32_t prefetch_so_done = 0;
16619 	struct tcp_log_buffer *lgb;
16620 	struct inpcb *inp = tptoinpcb(tp);
16621 	struct sockbuf *sb;
16622 	uint64_t ts_val = 0;
16623 #ifdef TCP_ACCOUNTING
16624 	uint64_t crtsc;
16625 #endif
16626 #ifdef INET6
16627 	struct ip6_hdr *ip6 = NULL;
16628 	int32_t isipv6;
16629 #endif
16630 	bool hw_tls = false;
16631 
16632 	NET_EPOCH_ASSERT();
16633 	INP_WLOCK_ASSERT(inp);
16634 
16635 	/* setup and take the cache hits here */
16636 	rack = (struct tcp_rack *)tp->t_fb_ptr;
16637 #ifdef TCP_ACCOUNTING
16638 	sched_pin();
16639 	ts_val = get_cyclecount();
16640 #endif
16641 	hpts_calling = inp->inp_hpts_calls;
16642 #ifdef TCP_OFFLOAD
16643 	if (tp->t_flags & TF_TOE) {
16644 #ifdef TCP_ACCOUNTING
16645 		sched_unpin();
16646 #endif
16647 		return (tcp_offload_output(tp));
16648 	}
16649 #endif
16650 	/*
16651 	 * For TFO connections in SYN_RECEIVED, only allow the initial
16652 	 * SYN|ACK and those sent by the retransmit timer.
16653 	 */
16654 	if (IS_FASTOPEN(tp->t_flags) &&
16655 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
16656 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16657 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16658 #ifdef TCP_ACCOUNTING
16659 		sched_unpin();
16660 #endif
16661 		return (0);
16662 	}
16663 #ifdef INET6
16664 	if (rack->r_state) {
16665 		/* Use the cache line loaded if possible */
16666 		isipv6 = rack->r_is_v6;
16667 	} else {
16668 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16669 	}
16670 #endif
16671 	early = 0;
16672 	cts = tcp_get_usecs(&tv);
16673 	ms_cts = tcp_tv_to_mssectick(&tv);
16674 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16675 	    tcp_in_hpts(rack->rc_inp)) {
16676 		/*
16677 		 * We are on the hpts for some timer but not hptsi output.
16678 		 * Remove from the hpts unconditionally.
16679 		 */
16680 		rack_timer_cancel(tp, rack, cts, __LINE__);
16681 	}
16682 	/* Are we pacing and late? */
16683 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16684 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16685 		/* We are delayed */
16686 		delayed = cts - rack->r_ctl.rc_last_output_to;
16687 	} else {
16688 		delayed = 0;
16689 	}
16690 	/* Do the timers, which may override the pacer */
16691 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16692 		int retval;
16693 
16694 		retval = rack_process_timers(tp, rack, cts, hpts_calling,
16695 		    &doing_tlp);
16696 		if (retval != 0) {
16697 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16698 #ifdef TCP_ACCOUNTING
16699 			sched_unpin();
16700 #endif
16701 			/*
16702 			 * If timers want tcp_drop(), then pass error out,
16703 			 * otherwise suppress it.
16704 			 */
16705 			return (retval < 0 ? retval : 0);
16706 		}
16707 	}
16708 	if (rack->rc_in_persist) {
16709 		if (tcp_in_hpts(rack->rc_inp) == 0) {
16710 			/* Timer is not running */
16711 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16712 		}
16713 #ifdef TCP_ACCOUNTING
16714 		sched_unpin();
16715 #endif
16716 		return (0);
16717 	}
16718 	if ((rack->rc_ack_required == 1) &&
16719 	    (rack->r_timer_override == 0)){
16720 		/* A timeout occurred and no ack has arrived */
16721 		if (tcp_in_hpts(rack->rc_inp) == 0) {
16722 			/* Timer is not running */
16723 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16724 		}
16725 #ifdef TCP_ACCOUNTING
16726 		sched_unpin();
16727 #endif
16728 		return (0);
16729 	}
16730 	if ((rack->r_timer_override) ||
16731 	    (rack->rc_ack_can_sendout_data) ||
16732 	    (delayed) ||
16733 	    (tp->t_state < TCPS_ESTABLISHED)) {
16734 		rack->rc_ack_can_sendout_data = 0;
16735 		if (tcp_in_hpts(rack->rc_inp))
16736 			tcp_hpts_remove(rack->rc_inp);
16737 	} else if (tcp_in_hpts(rack->rc_inp)) {
16738 		/*
16739 		 * On the hpts you can't pass even if ACKNOW is on, we will
16740 		 * when the hpts fires.
16741 		 */
16742 #ifdef TCP_ACCOUNTING
16743 		crtsc = get_cyclecount();
16744 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16745 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16746 		}
16747 		counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16748 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16749 			tp->tcp_cnt_counters[SND_BLOCKED]++;
16750 		}
16751 		counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16752 		sched_unpin();
16753 #endif
16754 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16755 		return (0);
16756 	}
16757 	rack->rc_inp->inp_hpts_calls = 0;
16758 	/* Finish out both pacing early and late accounting */
16759 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16760 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16761 		early = rack->r_ctl.rc_last_output_to - cts;
16762 	} else
16763 		early = 0;
16764 	if (delayed) {
16765 		rack->r_ctl.rc_agg_delayed += delayed;
16766 		rack->r_late = 1;
16767 	} else if (early) {
16768 		rack->r_ctl.rc_agg_early += early;
16769 		rack->r_early = 1;
16770 	}
16771 	/* Now that early/late accounting is done turn off the flag */
16772 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16773 	rack->r_wanted_output = 0;
16774 	rack->r_timer_override = 0;
16775 	if ((tp->t_state != rack->r_state) &&
16776 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
16777 		rack_set_state(tp, rack);
16778 	}
16779 	if ((rack->r_fast_output) &&
16780 	    (doing_tlp == 0) &&
16781 	    (tp->rcv_numsacks == 0)) {
16782 		int ret;
16783 
16784 		error = 0;
16785 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16786 		if (ret >= 0)
16787 			return(ret);
16788 		else if (error) {
16789 			inp = rack->rc_inp;
16790 			so = inp->inp_socket;
16791 			sb = &so->so_snd;
16792 			goto nomore;
16793 		}
16794 	}
16795 	inp = rack->rc_inp;
16796 	/*
16797 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
16798 	 * only allow the initial SYN or SYN|ACK and those sent
16799 	 * by the retransmit timer.
16800 	 */
16801 	if (IS_FASTOPEN(tp->t_flags) &&
16802 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
16803 	     (tp->t_state == TCPS_SYN_SENT)) &&
16804 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16805 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
16806 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16807 		so = inp->inp_socket;
16808 		sb = &so->so_snd;
16809 		goto just_return_nolock;
16810 	}
16811 	/*
16812 	 * Determine length of data that should be transmitted, and flags
16813 	 * that will be used. If there is some data or critical controls
16814 	 * (SYN, RST) to send, then transmit; otherwise, investigate
16815 	 * further.
16816 	 */
16817 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16818 	if (tp->t_idle_reduce) {
16819 		if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16820 			rack_cc_after_idle(rack, tp);
16821 	}
16822 	tp->t_flags &= ~TF_LASTIDLE;
16823 	if (idle) {
16824 		if (tp->t_flags & TF_MORETOCOME) {
16825 			tp->t_flags |= TF_LASTIDLE;
16826 			idle = 0;
16827 		}
16828 	}
16829 	if ((tp->snd_una == tp->snd_max) &&
16830 	    rack->r_ctl.rc_went_idle_time &&
16831 	    TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16832 		idle = cts - rack->r_ctl.rc_went_idle_time;
16833 		if (idle > rack_min_probertt_hold) {
16834 			/* Count as a probe rtt */
16835 			if (rack->in_probe_rtt == 0) {
16836 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
16837 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16838 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16839 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16840 			} else {
16841 				rack_exit_probertt(rack, cts);
16842 			}
16843 		}
16844 		idle = 0;
16845 	}
16846 	if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED))
16847 		rack_init_fsb_block(tp, rack);
16848 again:
16849 	/*
16850 	 * If we've recently taken a timeout, snd_max will be greater than
16851 	 * snd_nxt.  There may be SACK information that allows us to avoid
16852 	 * resending already delivered data.  Adjust snd_nxt accordingly.
16853 	 */
16854 	sendalot = 0;
16855 	cts = tcp_get_usecs(&tv);
16856 	ms_cts = tcp_tv_to_mssectick(&tv);
16857 	tso = 0;
16858 	mtu = 0;
16859 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16860 	minseg = segsiz;
16861 	if (rack->r_ctl.rc_pace_max_segs == 0)
16862 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16863 	else
16864 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16865 	sb_offset = tp->snd_max - tp->snd_una;
16866 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16867 	flags = tcp_outflags[tp->t_state];
16868 	while (rack->rc_free_cnt < rack_free_cache) {
16869 		rsm = rack_alloc(rack);
16870 		if (rsm == NULL) {
16871 			if (inp->inp_hpts_calls)
16872 				/* Retry in a ms */
16873 				slot = (1 * HPTS_USEC_IN_MSEC);
16874 			so = inp->inp_socket;
16875 			sb = &so->so_snd;
16876 			goto just_return_nolock;
16877 		}
16878 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16879 		rack->rc_free_cnt++;
16880 		rsm = NULL;
16881 	}
16882 	if (inp->inp_hpts_calls)
16883 		inp->inp_hpts_calls = 0;
16884 	sack_rxmit = 0;
16885 	len = 0;
16886 	rsm = NULL;
16887 	if (flags & TH_RST) {
16888 		SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16889 		so = inp->inp_socket;
16890 		sb = &so->so_snd;
16891 		goto send;
16892 	}
16893 	if (rack->r_ctl.rc_resend) {
16894 		/* Retransmit timer */
16895 		rsm = rack->r_ctl.rc_resend;
16896 		rack->r_ctl.rc_resend = NULL;
16897 		len = rsm->r_end - rsm->r_start;
16898 		sack_rxmit = 1;
16899 		sendalot = 0;
16900 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16901 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16902 			 __func__, __LINE__,
16903 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
16904 		sb_offset = rsm->r_start - tp->snd_una;
16905 		if (len >= segsiz)
16906 			len = segsiz;
16907 	} else if (rack->r_collapse_point_valid &&
16908 		   ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
16909 		/*
16910 		 * If an RSM is returned then enough time has passed
16911 		 * for us to retransmit it. Move up the collapse point,
16912 		 * since this rsm has its chance to retransmit now.
16913 		 */
16914 		rack_trace_point(rack, RACK_TP_COLLAPSED_RXT);
16915 		rack->r_ctl.last_collapse_point = rsm->r_end;
16916 		/* Are we done? */
16917 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
16918 			    rack->r_ctl.high_collapse_point))
16919 			rack->r_collapse_point_valid = 0;
16920 		sack_rxmit = 1;
16921 		/* We are not doing a TLP */
16922 		doing_tlp = 0;
16923 		len = rsm->r_end - rsm->r_start;
16924 		sb_offset = rsm->r_start - tp->snd_una;
16925 		sendalot = 0;
16926 		if ((rack->full_size_rxt == 0) &&
16927 		    (rack->shape_rxt_to_pacing_min == 0) &&
16928 		    (len >= segsiz))
16929 			len = segsiz;
16930 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
16931 		/* We have a retransmit that takes precedence */
16932 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
16933 		    ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
16934 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
16935 			/* Enter recovery if not induced by a time-out */
16936 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
16937 		}
16938 #ifdef INVARIANTS
16939 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
16940 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
16941 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
16942 		}
16943 #endif
16944 		len = rsm->r_end - rsm->r_start;
16945 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16946 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16947 			 __func__, __LINE__,
16948 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
16949 		sb_offset = rsm->r_start - tp->snd_una;
16950 		sendalot = 0;
16951 		if (len >= segsiz)
16952 			len = segsiz;
16953 		if (len > 0) {
16954 			sack_rxmit = 1;
16955 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
16956 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
16957 			    min(len, segsiz));
16958 		}
16959 	} else if (rack->r_ctl.rc_tlpsend) {
16960 		/* Tail loss probe */
16961 		long cwin;
16962 		long tlen;
16963 
16964 		/*
16965 		 * Check if we can do a TLP with a RACK'd packet
16966 		 * this can happen if we are not doing the rack
16967 		 * cheat and we skipped to a TLP and it
16968 		 * went off.
16969 		 */
16970 		rsm = rack->r_ctl.rc_tlpsend;
16971 		/* We are doing a TLP make sure the flag is preent */
16972 		rsm->r_flags |= RACK_TLP;
16973 		rack->r_ctl.rc_tlpsend = NULL;
16974 		sack_rxmit = 1;
16975 		tlen = rsm->r_end - rsm->r_start;
16976 		if (tlen > segsiz)
16977 			tlen = segsiz;
16978 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
16979 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
16980 			 __func__, __LINE__,
16981 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
16982 		sb_offset = rsm->r_start - tp->snd_una;
16983 		cwin = min(tp->snd_wnd, tlen);
16984 		len = cwin;
16985 	}
16986 	if (rack->r_must_retran &&
16987 	    (doing_tlp == 0) &&
16988 	    (SEQ_GT(tp->snd_max, tp->snd_una)) &&
16989 	    (rsm == NULL)) {
16990 		/*
16991 		 * There are two different ways that we
16992 		 * can get into this block:
16993 		 * a) This is a non-sack connection, we had a time-out
16994 		 *    and thus r_must_retran was set and everything
16995 		 *    left outstanding as been marked for retransmit.
16996 		 * b) The MTU of the path shrank, so that everything
16997 		 *    was marked to be retransmitted with the smaller
16998 		 *    mtu and r_must_retran was set.
16999 		 *
17000 		 * This means that we expect the sendmap (outstanding)
17001 		 * to all be marked must. We can use the tmap to
17002 		 * look at them.
17003 		 *
17004 		 */
17005 		int sendwin, flight;
17006 
17007 		sendwin = min(tp->snd_wnd, tp->snd_cwnd);
17008 		flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
17009 		if (flight >= sendwin) {
17010 			/*
17011 			 * We can't send yet.
17012 			 */
17013 			so = inp->inp_socket;
17014 			sb = &so->so_snd;
17015 			goto just_return_nolock;
17016 		}
17017 		/*
17018 		 * This is the case a/b mentioned above. All
17019 		 * outstanding/not-acked should be marked.
17020 		 * We can use the tmap to find them.
17021 		 */
17022 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17023 		if (rsm == NULL) {
17024 			/* TSNH */
17025 			rack->r_must_retran = 0;
17026 			rack->r_ctl.rc_out_at_rto = 0;
17027 			so = inp->inp_socket;
17028 			sb = &so->so_snd;
17029 			goto just_return_nolock;
17030 		}
17031 		if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
17032 			/*
17033 			 * The first one does not have the flag, did we collapse
17034 			 * further up in our list?
17035 			 */
17036 			rack->r_must_retran = 0;
17037 			rack->r_ctl.rc_out_at_rto = 0;
17038 			rsm = NULL;
17039 			sack_rxmit = 0;
17040 		} else {
17041 			sack_rxmit = 1;
17042 			len = rsm->r_end - rsm->r_start;
17043 			sb_offset = rsm->r_start - tp->snd_una;
17044 			sendalot = 0;
17045 			if ((rack->full_size_rxt == 0) &&
17046 			    (rack->shape_rxt_to_pacing_min == 0) &&
17047 			    (len >= segsiz))
17048 				len = segsiz;
17049 			/*
17050 			 * Delay removing the flag RACK_MUST_RXT so
17051 			 * that the fastpath for retransmit will
17052 			 * work with this rsm.
17053 			 */
17054 		}
17055 	}
17056 	/*
17057 	 * Enforce a connection sendmap count limit if set
17058 	 * as long as we are not retransmiting.
17059 	 */
17060 	if ((rsm == NULL) &&
17061 	    (rack->do_detection == 0) &&
17062 	    (V_tcp_map_entries_limit > 0) &&
17063 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
17064 		counter_u64_add(rack_to_alloc_limited, 1);
17065 		if (!rack->alloc_limit_reported) {
17066 			rack->alloc_limit_reported = 1;
17067 			counter_u64_add(rack_alloc_limited_conns, 1);
17068 		}
17069 		so = inp->inp_socket;
17070 		sb = &so->so_snd;
17071 		goto just_return_nolock;
17072 	}
17073 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
17074 		/* we are retransmitting the fin */
17075 		len--;
17076 		if (len) {
17077 			/*
17078 			 * When retransmitting data do *not* include the
17079 			 * FIN. This could happen from a TLP probe.
17080 			 */
17081 			flags &= ~TH_FIN;
17082 		}
17083 	}
17084 	if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
17085 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
17086 		int ret;
17087 
17088 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
17089 		if (ret == 0)
17090 			return (0);
17091 	}
17092 	so = inp->inp_socket;
17093 	sb = &so->so_snd;
17094 	if (do_a_prefetch == 0) {
17095 		kern_prefetch(sb, &do_a_prefetch);
17096 		do_a_prefetch = 1;
17097 	}
17098 #ifdef NETFLIX_SHARED_CWND
17099 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
17100 	    rack->rack_enable_scwnd) {
17101 		/* We are doing cwnd sharing */
17102 		if (rack->gp_ready &&
17103 		    (rack->rack_attempted_scwnd == 0) &&
17104 		    (rack->r_ctl.rc_scw == NULL) &&
17105 		    tp->t_lib) {
17106 			/* The pcbid is in, lets make an attempt */
17107 			counter_u64_add(rack_try_scwnd, 1);
17108 			rack->rack_attempted_scwnd = 1;
17109 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
17110 								   &rack->r_ctl.rc_scw_index,
17111 								   segsiz);
17112 		}
17113 		if (rack->r_ctl.rc_scw &&
17114 		    (rack->rack_scwnd_is_idle == 1) &&
17115 		    sbavail(&so->so_snd)) {
17116 			/* we are no longer out of data */
17117 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17118 			rack->rack_scwnd_is_idle = 0;
17119 		}
17120 		if (rack->r_ctl.rc_scw) {
17121 			/* First lets update and get the cwnd */
17122 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
17123 								    rack->r_ctl.rc_scw_index,
17124 								    tp->snd_cwnd, tp->snd_wnd, segsiz);
17125 		}
17126 	}
17127 #endif
17128 	/*
17129 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
17130 	 * state flags.
17131 	 */
17132 	if (tp->t_flags & TF_NEEDFIN)
17133 		flags |= TH_FIN;
17134 	if (tp->t_flags & TF_NEEDSYN)
17135 		flags |= TH_SYN;
17136 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
17137 		void *end_rsm;
17138 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
17139 		if (end_rsm)
17140 			kern_prefetch(end_rsm, &prefetch_rsm);
17141 		prefetch_rsm = 1;
17142 	}
17143 	SOCKBUF_LOCK(sb);
17144 	/*
17145 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
17146 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
17147 	 * negative length.  This can also occur when TCP opens up its
17148 	 * congestion window while receiving additional duplicate acks after
17149 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
17150 	 * the fast-retransmit.
17151 	 *
17152 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
17153 	 * set to snd_una, the sb_offset will be 0, and the length may wind
17154 	 * up 0.
17155 	 *
17156 	 * If sack_rxmit is true we are retransmitting from the scoreboard
17157 	 * in which case len is already set.
17158 	 */
17159 	if ((sack_rxmit == 0) &&
17160 	    (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
17161 		uint32_t avail;
17162 
17163 		avail = sbavail(sb);
17164 		if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
17165 			sb_offset = tp->snd_nxt - tp->snd_una;
17166 		else
17167 			sb_offset = 0;
17168 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
17169 			if (rack->r_ctl.rc_tlp_new_data) {
17170 				/* TLP is forcing out new data */
17171 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
17172 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
17173 				}
17174 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
17175 					if (tp->snd_wnd > sb_offset)
17176 						len = tp->snd_wnd - sb_offset;
17177 					else
17178 						len = 0;
17179 				} else {
17180 					len = rack->r_ctl.rc_tlp_new_data;
17181 				}
17182 				rack->r_ctl.rc_tlp_new_data = 0;
17183 			}  else {
17184 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
17185 			}
17186 			if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
17187 				/*
17188 				 * For prr=off, we need to send only 1 MSS
17189 				 * at a time. We do this because another sack could
17190 				 * be arriving that causes us to send retransmits and
17191 				 * we don't want to be on a long pace due to a larger send
17192 				 * that keeps us from sending out the retransmit.
17193 				 */
17194 				len = segsiz;
17195 			}
17196 		} else {
17197 			uint32_t outstanding;
17198 			/*
17199 			 * We are inside of a Fast recovery episode, this
17200 			 * is caused by a SACK or 3 dup acks. At this point
17201 			 * we have sent all the retransmissions and we rely
17202 			 * on PRR to dictate what we will send in the form of
17203 			 * new data.
17204 			 */
17205 
17206 			outstanding = tp->snd_max - tp->snd_una;
17207 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
17208 				if (tp->snd_wnd > outstanding) {
17209 					len = tp->snd_wnd - outstanding;
17210 					/* Check to see if we have the data */
17211 					if ((sb_offset + len) > avail) {
17212 						/* It does not all fit */
17213 						if (avail > sb_offset)
17214 							len = avail - sb_offset;
17215 						else
17216 							len = 0;
17217 					}
17218 				} else {
17219 					len = 0;
17220 				}
17221 			} else if (avail > sb_offset) {
17222 				len = avail - sb_offset;
17223 			} else {
17224 				len = 0;
17225 			}
17226 			if (len > 0) {
17227 				if (len > rack->r_ctl.rc_prr_sndcnt) {
17228 					len = rack->r_ctl.rc_prr_sndcnt;
17229 				}
17230 				if (len > 0) {
17231 					sub_from_prr = 1;
17232 				}
17233 			}
17234 			if (len > segsiz) {
17235 				/*
17236 				 * We should never send more than a MSS when
17237 				 * retransmitting or sending new data in prr
17238 				 * mode unless the override flag is on. Most
17239 				 * likely the PRR algorithm is not going to
17240 				 * let us send a lot as well :-)
17241 				 */
17242 				if (rack->r_ctl.rc_prr_sendalot == 0) {
17243 					len = segsiz;
17244 				}
17245 			} else if (len < segsiz) {
17246 				/*
17247 				 * Do we send any? The idea here is if the
17248 				 * send empty's the socket buffer we want to
17249 				 * do it. However if not then lets just wait
17250 				 * for our prr_sndcnt to get bigger.
17251 				 */
17252 				long leftinsb;
17253 
17254 				leftinsb = sbavail(sb) - sb_offset;
17255 				if (leftinsb > len) {
17256 					/* This send does not empty the sb */
17257 					len = 0;
17258 				}
17259 			}
17260 		}
17261 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
17262 		/*
17263 		 * If you have not established
17264 		 * and are not doing FAST OPEN
17265 		 * no data please.
17266 		 */
17267 		if ((sack_rxmit == 0) &&
17268 		    (!IS_FASTOPEN(tp->t_flags))){
17269 			len = 0;
17270 			sb_offset = 0;
17271 		}
17272 	}
17273 	if (prefetch_so_done == 0) {
17274 		kern_prefetch(so, &prefetch_so_done);
17275 		prefetch_so_done = 1;
17276 	}
17277 	/*
17278 	 * Lop off SYN bit if it has already been sent.  However, if this is
17279 	 * SYN-SENT state and if segment contains data and if we don't know
17280 	 * that foreign host supports TAO, suppress sending segment.
17281 	 */
17282 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
17283 	    ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
17284 		/*
17285 		 * When sending additional segments following a TFO SYN|ACK,
17286 		 * do not include the SYN bit.
17287 		 */
17288 		if (IS_FASTOPEN(tp->t_flags) &&
17289 		    (tp->t_state == TCPS_SYN_RECEIVED))
17290 			flags &= ~TH_SYN;
17291 	}
17292 	/*
17293 	 * Be careful not to send data and/or FIN on SYN segments. This
17294 	 * measure is needed to prevent interoperability problems with not
17295 	 * fully conformant TCP implementations.
17296 	 */
17297 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
17298 		len = 0;
17299 		flags &= ~TH_FIN;
17300 	}
17301 	/*
17302 	 * On TFO sockets, ensure no data is sent in the following cases:
17303 	 *
17304 	 *  - When retransmitting SYN|ACK on a passively-created socket
17305 	 *
17306 	 *  - When retransmitting SYN on an actively created socket
17307 	 *
17308 	 *  - When sending a zero-length cookie (cookie request) on an
17309 	 *    actively created socket
17310 	 *
17311 	 *  - When the socket is in the CLOSED state (RST is being sent)
17312 	 */
17313 	if (IS_FASTOPEN(tp->t_flags) &&
17314 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
17315 	     ((tp->t_state == TCPS_SYN_SENT) &&
17316 	      (tp->t_tfo_client_cookie_len == 0)) ||
17317 	     (flags & TH_RST))) {
17318 		sack_rxmit = 0;
17319 		len = 0;
17320 	}
17321 	/* Without fast-open there should never be data sent on a SYN */
17322 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
17323 		tp->snd_nxt = tp->iss;
17324 		len = 0;
17325 	}
17326 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
17327 		/* We only send 1 MSS if we have a DSACK block */
17328 		add_flag |= RACK_SENT_W_DSACK;
17329 		len = segsiz;
17330 	}
17331 	orig_len = len;
17332 	if (len <= 0) {
17333 		/*
17334 		 * If FIN has been sent but not acked, but we haven't been
17335 		 * called to retransmit, len will be < 0.  Otherwise, window
17336 		 * shrank after we sent into it.  If window shrank to 0,
17337 		 * cancel pending retransmit, pull snd_nxt back to (closed)
17338 		 * window, and set the persist timer if it isn't already
17339 		 * going.  If the window didn't close completely, just wait
17340 		 * for an ACK.
17341 		 *
17342 		 * We also do a general check here to ensure that we will
17343 		 * set the persist timer when we have data to send, but a
17344 		 * 0-byte window. This makes sure the persist timer is set
17345 		 * even if the packet hits one of the "goto send" lines
17346 		 * below.
17347 		 */
17348 		len = 0;
17349 		if ((tp->snd_wnd == 0) &&
17350 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17351 		    (tp->snd_una == tp->snd_max) &&
17352 		    (sb_offset < (int)sbavail(sb))) {
17353 			rack_enter_persist(tp, rack, cts);
17354 		}
17355 	} else if ((rsm == NULL) &&
17356 		   (doing_tlp == 0) &&
17357 		   (len < pace_max_seg)) {
17358 		/*
17359 		 * We are not sending a maximum sized segment for
17360 		 * some reason. Should we not send anything (think
17361 		 * sws or persists)?
17362 		 */
17363 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17364 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17365 		    (len < minseg) &&
17366 		    (len < (int)(sbavail(sb) - sb_offset))) {
17367 			/*
17368 			 * Here the rwnd is less than
17369 			 * the minimum pacing size, this is not a retransmit,
17370 			 * we are established and
17371 			 * the send is not the last in the socket buffer
17372 			 * we send nothing, and we may enter persists
17373 			 * if nothing is outstanding.
17374 			 */
17375 			len = 0;
17376 			if (tp->snd_max == tp->snd_una) {
17377 				/*
17378 				 * Nothing out we can
17379 				 * go into persists.
17380 				 */
17381 				rack_enter_persist(tp, rack, cts);
17382 			}
17383 		     } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
17384 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17385 			   (len < (int)(sbavail(sb) - sb_offset)) &&
17386 			   (len < minseg)) {
17387 			/*
17388 			 * Here we are not retransmitting, and
17389 			 * the cwnd is not so small that we could
17390 			 * not send at least a min size (rxt timer
17391 			 * not having gone off), We have 2 segments or
17392 			 * more already in flight, its not the tail end
17393 			 * of the socket buffer  and the cwnd is blocking
17394 			 * us from sending out a minimum pacing segment size.
17395 			 * Lets not send anything.
17396 			 */
17397 			len = 0;
17398 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
17399 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17400 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17401 			   (len < (int)(sbavail(sb) - sb_offset)) &&
17402 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
17403 			/*
17404 			 * Here we have a send window but we have
17405 			 * filled it up and we can't send another pacing segment.
17406 			 * We also have in flight more than 2 segments
17407 			 * and we are not completing the sb i.e. we allow
17408 			 * the last bytes of the sb to go out even if
17409 			 * its not a full pacing segment.
17410 			 */
17411 			len = 0;
17412 		} else if ((rack->r_ctl.crte != NULL) &&
17413 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
17414 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
17415 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
17416 			   (len < (int)(sbavail(sb) - sb_offset))) {
17417 			/*
17418 			 * Here we are doing hardware pacing, this is not a TLP,
17419 			 * we are not sending a pace max segment size, there is rwnd
17420 			 * room to send at least N pace_max_seg, the cwnd is greater
17421 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
17422 			 * more segments in flight and its not the tail of the socket buffer.
17423 			 *
17424 			 * We don't want to send instead we need to get more ack's in to
17425 			 * allow us to send a full pacing segment. Normally, if we are pacing
17426 			 * about the right speed, we should have finished our pacing
17427 			 * send as most of the acks have come back if we are at the
17428 			 * right rate. This is a bit fuzzy since return path delay
17429 			 * can delay the acks, which is why we want to make sure we
17430 			 * have cwnd space to have a bit more than a max pace segments in flight.
17431 			 *
17432 			 * If we have not gotten our acks back we are pacing at too high a
17433 			 * rate delaying will not hurt and will bring our GP estimate down by
17434 			 * injecting the delay. If we don't do this we will send
17435 			 * 2 MSS out in response to the acks being clocked in which
17436 			 * defeats the point of hw-pacing (i.e. to help us get
17437 			 * larger TSO's out).
17438 			 */
17439 			len = 0;
17440 
17441 		}
17442 
17443 	}
17444 	/* len will be >= 0 after this point. */
17445 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17446 	rack_sndbuf_autoscale(rack);
17447 	/*
17448 	 * Decide if we can use TCP Segmentation Offloading (if supported by
17449 	 * hardware).
17450 	 *
17451 	 * TSO may only be used if we are in a pure bulk sending state.  The
17452 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
17453 	 * options prevent using TSO.  With TSO the TCP header is the same
17454 	 * (except for the sequence number) for all generated packets.  This
17455 	 * makes it impossible to transmit any options which vary per
17456 	 * generated segment or packet.
17457 	 *
17458 	 * IPv4 handling has a clear separation of ip options and ip header
17459 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
17460 	 * the right thing below to provide length of just ip options and thus
17461 	 * checking for ipoptlen is enough to decide if ip options are present.
17462 	 */
17463 	ipoptlen = 0;
17464 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17465 	/*
17466 	 * Pre-calculate here as we save another lookup into the darknesses
17467 	 * of IPsec that way and can actually decide if TSO is ok.
17468 	 */
17469 #ifdef INET6
17470 	if (isipv6 && IPSEC_ENABLED(ipv6))
17471 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
17472 #ifdef INET
17473 	else
17474 #endif
17475 #endif				/* INET6 */
17476 #ifdef INET
17477 		if (IPSEC_ENABLED(ipv4))
17478 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
17479 #endif				/* INET */
17480 #endif
17481 
17482 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17483 	ipoptlen += ipsec_optlen;
17484 #endif
17485 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
17486 	    (tp->t_port == 0) &&
17487 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
17488 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
17489 	    ipoptlen == 0)
17490 		tso = 1;
17491 	{
17492 		uint32_t outstanding __unused;
17493 
17494 		outstanding = tp->snd_max - tp->snd_una;
17495 		if (tp->t_flags & TF_SENTFIN) {
17496 			/*
17497 			 * If we sent a fin, snd_max is 1 higher than
17498 			 * snd_una
17499 			 */
17500 			outstanding--;
17501 		}
17502 		if (sack_rxmit) {
17503 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
17504 				flags &= ~TH_FIN;
17505 		} else {
17506 			if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
17507 				   sbused(sb)))
17508 				flags &= ~TH_FIN;
17509 		}
17510 	}
17511 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
17512 	    (long)TCP_MAXWIN << tp->rcv_scale);
17513 
17514 	/*
17515 	 * Sender silly window avoidance.   We transmit under the following
17516 	 * conditions when len is non-zero:
17517 	 *
17518 	 * - We have a full segment (or more with TSO) - This is the last
17519 	 * buffer in a write()/send() and we are either idle or running
17520 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
17521 	 * then 1/2 the maximum send window's worth of data (receiver may be
17522 	 * limited the window size) - we need to retransmit
17523 	 */
17524 	if (len) {
17525 		if (len >= segsiz) {
17526 			goto send;
17527 		}
17528 		/*
17529 		 * NOTE! on localhost connections an 'ack' from the remote
17530 		 * end may occur synchronously with the output and cause us
17531 		 * to flush a buffer queued with moretocome.  XXX
17532 		 *
17533 		 */
17534 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
17535 		    (idle || (tp->t_flags & TF_NODELAY)) &&
17536 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17537 		    (tp->t_flags & TF_NOPUSH) == 0) {
17538 			pass = 2;
17539 			goto send;
17540 		}
17541 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
17542 			pass = 22;
17543 			goto send;
17544 		}
17545 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
17546 			pass = 4;
17547 			goto send;
17548 		}
17549 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {	/* retransmit case */
17550 			pass = 5;
17551 			goto send;
17552 		}
17553 		if (sack_rxmit) {
17554 			pass = 6;
17555 			goto send;
17556 		}
17557 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
17558 		    (ctf_outstanding(tp) < (segsiz * 2))) {
17559 			/*
17560 			 * We have less than two MSS outstanding (delayed ack)
17561 			 * and our rwnd will not let us send a full sized
17562 			 * MSS. Lets go ahead and let this small segment
17563 			 * out because we want to try to have at least two
17564 			 * packets inflight to not be caught by delayed ack.
17565 			 */
17566 			pass = 12;
17567 			goto send;
17568 		}
17569 	}
17570 	/*
17571 	 * Sending of standalone window updates.
17572 	 *
17573 	 * Window updates are important when we close our window due to a
17574 	 * full socket buffer and are opening it again after the application
17575 	 * reads data from it.  Once the window has opened again and the
17576 	 * remote end starts to send again the ACK clock takes over and
17577 	 * provides the most current window information.
17578 	 *
17579 	 * We must avoid the silly window syndrome whereas every read from
17580 	 * the receive buffer, no matter how small, causes a window update
17581 	 * to be sent.  We also should avoid sending a flurry of window
17582 	 * updates when the socket buffer had queued a lot of data and the
17583 	 * application is doing small reads.
17584 	 *
17585 	 * Prevent a flurry of pointless window updates by only sending an
17586 	 * update when we can increase the advertized window by more than
17587 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
17588 	 * full or is very small be more aggressive and send an update
17589 	 * whenever we can increase by two mss sized segments. In all other
17590 	 * situations the ACK's to new incoming data will carry further
17591 	 * window increases.
17592 	 *
17593 	 * Don't send an independent window update if a delayed ACK is
17594 	 * pending (it will get piggy-backed on it) or the remote side
17595 	 * already has done a half-close and won't send more data.  Skip
17596 	 * this if the connection is in T/TCP half-open state.
17597 	 */
17598 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
17599 	    !(tp->t_flags & TF_DELACK) &&
17600 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
17601 		/*
17602 		 * "adv" is the amount we could increase the window, taking
17603 		 * into account that we are limited by TCP_MAXWIN <<
17604 		 * tp->rcv_scale.
17605 		 */
17606 		int32_t adv;
17607 		int oldwin;
17608 
17609 		adv = recwin;
17610 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
17611 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
17612 			if (adv > oldwin)
17613 			    adv -= oldwin;
17614 			else {
17615 				/* We can't increase the window */
17616 				adv = 0;
17617 			}
17618 		} else
17619 			oldwin = 0;
17620 
17621 		/*
17622 		 * If the new window size ends up being the same as or less
17623 		 * than the old size when it is scaled, then don't force
17624 		 * a window update.
17625 		 */
17626 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
17627 			goto dontupdate;
17628 
17629 		if (adv >= (int32_t)(2 * segsiz) &&
17630 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
17631 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
17632 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
17633 			pass = 7;
17634 			goto send;
17635 		}
17636 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
17637 			pass = 23;
17638 			goto send;
17639 		}
17640 	}
17641 dontupdate:
17642 
17643 	/*
17644 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
17645 	 * is also a catch-all for the retransmit timer timeout case.
17646 	 */
17647 	if (tp->t_flags & TF_ACKNOW) {
17648 		pass = 8;
17649 		goto send;
17650 	}
17651 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
17652 		pass = 9;
17653 		goto send;
17654 	}
17655 	/*
17656 	 * If our state indicates that FIN should be sent and we have not
17657 	 * yet done so, then we need to send.
17658 	 */
17659 	if ((flags & TH_FIN) &&
17660 	    (tp->snd_nxt == tp->snd_una)) {
17661 		pass = 11;
17662 		goto send;
17663 	}
17664 	/*
17665 	 * No reason to send a segment, just return.
17666 	 */
17667 just_return:
17668 	SOCKBUF_UNLOCK(sb);
17669 just_return_nolock:
17670 	{
17671 		int app_limited = CTF_JR_SENT_DATA;
17672 
17673 		if (tot_len_this_send > 0) {
17674 			/* Make sure snd_nxt is up to max */
17675 			rack->r_ctl.fsb.recwin = recwin;
17676 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17677 			if ((error == 0) &&
17678 			    rack_use_rfo &&
17679 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
17680 			    (ipoptlen == 0) &&
17681 			    (tp->snd_nxt == tp->snd_max) &&
17682 			    (tp->rcv_numsacks == 0) &&
17683 			    rack->r_fsb_inited &&
17684 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
17685 			    (rack->r_must_retran == 0) &&
17686 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
17687 			    (len > 0) && (orig_len > 0) &&
17688 			    (orig_len > len) &&
17689 			    ((orig_len - len) >= segsiz) &&
17690 			    ((optlen == 0) ||
17691 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17692 				/* We can send at least one more MSS using our fsb */
17693 
17694 				rack->r_fast_output = 1;
17695 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17696 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17697 				rack->r_ctl.fsb.tcp_flags = flags;
17698 				rack->r_ctl.fsb.left_to_send = orig_len - len;
17699 				if (hw_tls)
17700 					rack->r_ctl.fsb.hw_tls = 1;
17701 				else
17702 					rack->r_ctl.fsb.hw_tls = 0;
17703 				KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17704 					("rack:%p left_to_send:%u sbavail:%u out:%u",
17705 					rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17706 					 (tp->snd_max - tp->snd_una)));
17707 				if (rack->r_ctl.fsb.left_to_send < segsiz)
17708 					rack->r_fast_output = 0;
17709 				else {
17710 					if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17711 						rack->r_ctl.fsb.rfo_apply_push = 1;
17712 					else
17713 						rack->r_ctl.fsb.rfo_apply_push = 0;
17714 				}
17715 			} else
17716 				rack->r_fast_output = 0;
17717 
17718 
17719 			rack_log_fsb(rack, tp, so, flags,
17720 				     ipoptlen, orig_len, len, 0,
17721 				     1, optlen, __LINE__, 1);
17722 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17723 				tp->snd_nxt = tp->snd_max;
17724 		} else {
17725 			int end_window = 0;
17726 			uint32_t seq = tp->gput_ack;
17727 
17728 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17729 			if (rsm) {
17730 				/*
17731 				 * Mark the last sent that we just-returned (hinting
17732 				 * that delayed ack may play a role in any rtt measurement).
17733 				 */
17734 				rsm->r_just_ret = 1;
17735 			}
17736 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17737 			rack->r_ctl.rc_agg_delayed = 0;
17738 			rack->r_early = 0;
17739 			rack->r_late = 0;
17740 			rack->r_ctl.rc_agg_early = 0;
17741 			if ((ctf_outstanding(tp) +
17742 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17743 				 minseg)) >= tp->snd_wnd) {
17744 				/* We are limited by the rwnd */
17745 				app_limited = CTF_JR_RWND_LIMITED;
17746 				if (IN_FASTRECOVERY(tp->t_flags))
17747 				    rack->r_ctl.rc_prr_sndcnt = 0;
17748 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
17749 				/* We are limited by whats available -- app limited */
17750 				app_limited = CTF_JR_APP_LIMITED;
17751 				if (IN_FASTRECOVERY(tp->t_flags))
17752 				    rack->r_ctl.rc_prr_sndcnt = 0;
17753 			} else if ((idle == 0) &&
17754 				   ((tp->t_flags & TF_NODELAY) == 0) &&
17755 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17756 				   (len < segsiz)) {
17757 				/*
17758 				 * No delay is not on and the
17759 				 * user is sending less than 1MSS. This
17760 				 * brings out SWS avoidance so we
17761 				 * don't send. Another app-limited case.
17762 				 */
17763 				app_limited = CTF_JR_APP_LIMITED;
17764 			} else if (tp->t_flags & TF_NOPUSH) {
17765 				/*
17766 				 * The user has requested no push of
17767 				 * the last segment and we are
17768 				 * at the last segment. Another app
17769 				 * limited case.
17770 				 */
17771 				app_limited = CTF_JR_APP_LIMITED;
17772 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17773 				/* Its the cwnd */
17774 				app_limited = CTF_JR_CWND_LIMITED;
17775 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
17776 				   (rack->rack_no_prr == 0) &&
17777 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17778 				app_limited = CTF_JR_PRR;
17779 			} else {
17780 				/* Now why here are we not sending? */
17781 #ifdef NOW
17782 #ifdef INVARIANTS
17783 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17784 #endif
17785 #endif
17786 				app_limited = CTF_JR_ASSESSING;
17787 			}
17788 			/*
17789 			 * App limited in some fashion, for our pacing GP
17790 			 * measurements we don't want any gap (even cwnd).
17791 			 * Close  down the measurement window.
17792 			 */
17793 			if (rack_cwnd_block_ends_measure &&
17794 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
17795 			     (app_limited == CTF_JR_PRR))) {
17796 				/*
17797 				 * The reason we are not sending is
17798 				 * the cwnd (or prr). We have been configured
17799 				 * to end the measurement window in
17800 				 * this case.
17801 				 */
17802 				end_window = 1;
17803 			} else if (rack_rwnd_block_ends_measure &&
17804 				   (app_limited == CTF_JR_RWND_LIMITED)) {
17805 				/*
17806 				 * We are rwnd limited and have been
17807 				 * configured to end the measurement
17808 				 * window in this case.
17809 				 */
17810 				end_window = 1;
17811 			} else if (app_limited == CTF_JR_APP_LIMITED) {
17812 				/*
17813 				 * A true application limited period, we have
17814 				 * ran out of data.
17815 				 */
17816 				end_window = 1;
17817 			} else if (app_limited == CTF_JR_ASSESSING) {
17818 				/*
17819 				 * In the assessing case we hit the end of
17820 				 * the if/else and had no known reason
17821 				 * This will panic us under invariants..
17822 				 *
17823 				 * If we get this out in logs we need to
17824 				 * investagate which reason we missed.
17825 				 */
17826 				end_window = 1;
17827 			}
17828 			if (end_window) {
17829 				uint8_t log = 0;
17830 
17831 				/* Adjust the Gput measurement */
17832 				if ((tp->t_flags & TF_GPUTINPROG) &&
17833 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
17834 					tp->gput_ack = tp->snd_max;
17835 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
17836 						/*
17837 						 * There is not enough to measure.
17838 						 */
17839 						tp->t_flags &= ~TF_GPUTINPROG;
17840 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17841 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
17842 									   tp->gput_seq,
17843 									   0, 0, 18, __LINE__, NULL, 0);
17844 					} else
17845 						log = 1;
17846 				}
17847 				/* Mark the last packet has app limited */
17848 				rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17849 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17850 					if (rack->r_ctl.rc_app_limited_cnt == 0)
17851 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17852 					else {
17853 						/*
17854 						 * Go out to the end app limited and mark
17855 						 * this new one as next and move the end_appl up
17856 						 * to this guy.
17857 						 */
17858 						if (rack->r_ctl.rc_end_appl)
17859 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17860 						rack->r_ctl.rc_end_appl = rsm;
17861 					}
17862 					rsm->r_flags |= RACK_APP_LIMITED;
17863 					rack->r_ctl.rc_app_limited_cnt++;
17864 				}
17865 				if (log)
17866 					rack_log_pacing_delay_calc(rack,
17867 								   rack->r_ctl.rc_app_limited_cnt, seq,
17868 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
17869 			}
17870 		}
17871 		/* Check if we need to go into persists or not */
17872 		if ((tp->snd_max == tp->snd_una) &&
17873 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
17874 		    sbavail(sb) &&
17875 		    (sbavail(sb) > tp->snd_wnd) &&
17876 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17877 			/* Yes lets make sure to move to persist before timer-start */
17878 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17879 		}
17880 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17881 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17882 	}
17883 #ifdef NETFLIX_SHARED_CWND
17884 	if ((sbavail(sb) == 0) &&
17885 	    rack->r_ctl.rc_scw) {
17886 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17887 		rack->rack_scwnd_is_idle = 1;
17888 	}
17889 #endif
17890 #ifdef TCP_ACCOUNTING
17891 	if (tot_len_this_send > 0) {
17892 		crtsc = get_cyclecount();
17893 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17894 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
17895 		}
17896 		counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17897 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17898 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
17899 		}
17900 		counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
17901 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17902 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
17903 		}
17904 		counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
17905 	} else {
17906 		crtsc = get_cyclecount();
17907 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17908 			tp->tcp_cnt_counters[SND_LIMITED]++;
17909 		}
17910 		counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
17911 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17912 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
17913 		}
17914 		counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
17915 	}
17916 	sched_unpin();
17917 #endif
17918 	return (0);
17919 
17920 send:
17921 	if (rsm || sack_rxmit)
17922 		counter_u64_add(rack_nfto_resend, 1);
17923 	else
17924 		counter_u64_add(rack_non_fto_send, 1);
17925 	if ((flags & TH_FIN) &&
17926 	    sbavail(sb)) {
17927 		/*
17928 		 * We do not transmit a FIN
17929 		 * with data outstanding. We
17930 		 * need to make it so all data
17931 		 * is acked first.
17932 		 */
17933 		flags &= ~TH_FIN;
17934 	}
17935 	/* Enforce stack imposed max seg size if we have one */
17936 	if (rack->r_ctl.rc_pace_max_segs &&
17937 	    (len > rack->r_ctl.rc_pace_max_segs)) {
17938 		mark = 1;
17939 		len = rack->r_ctl.rc_pace_max_segs;
17940 	}
17941 	SOCKBUF_LOCK_ASSERT(sb);
17942 	if (len > 0) {
17943 		if (len >= segsiz)
17944 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
17945 		else
17946 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
17947 	}
17948 	/*
17949 	 * Before ESTABLISHED, force sending of initial options unless TCP
17950 	 * set not to do any options. NOTE: we assume that the IP/TCP header
17951 	 * plus TCP options always fit in a single mbuf, leaving room for a
17952 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
17953 	 * + optlen <= MCLBYTES
17954 	 */
17955 	optlen = 0;
17956 #ifdef INET6
17957 	if (isipv6)
17958 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
17959 	else
17960 #endif
17961 		hdrlen = sizeof(struct tcpiphdr);
17962 
17963 	/*
17964 	 * Compute options for segment. We only have to care about SYN and
17965 	 * established connection segments.  Options for SYN-ACK segments
17966 	 * are handled in TCP syncache.
17967 	 */
17968 	to.to_flags = 0;
17969 	if ((tp->t_flags & TF_NOOPT) == 0) {
17970 		/* Maximum segment size. */
17971 		if (flags & TH_SYN) {
17972 			tp->snd_nxt = tp->iss;
17973 			to.to_mss = tcp_mssopt(&inp->inp_inc);
17974 			if (tp->t_port)
17975 				to.to_mss -= V_tcp_udp_tunneling_overhead;
17976 			to.to_flags |= TOF_MSS;
17977 
17978 			/*
17979 			 * On SYN or SYN|ACK transmits on TFO connections,
17980 			 * only include the TFO option if it is not a
17981 			 * retransmit, as the presence of the TFO option may
17982 			 * have caused the original SYN or SYN|ACK to have
17983 			 * been dropped by a middlebox.
17984 			 */
17985 			if (IS_FASTOPEN(tp->t_flags) &&
17986 			    (tp->t_rxtshift == 0)) {
17987 				if (tp->t_state == TCPS_SYN_RECEIVED) {
17988 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
17989 					to.to_tfo_cookie =
17990 						(u_int8_t *)&tp->t_tfo_cookie.server;
17991 					to.to_flags |= TOF_FASTOPEN;
17992 					wanted_cookie = 1;
17993 				} else if (tp->t_state == TCPS_SYN_SENT) {
17994 					to.to_tfo_len =
17995 						tp->t_tfo_client_cookie_len;
17996 					to.to_tfo_cookie =
17997 						tp->t_tfo_cookie.client;
17998 					to.to_flags |= TOF_FASTOPEN;
17999 					wanted_cookie = 1;
18000 					/*
18001 					 * If we wind up having more data to
18002 					 * send with the SYN than can fit in
18003 					 * one segment, don't send any more
18004 					 * until the SYN|ACK comes back from
18005 					 * the other end.
18006 					 */
18007 					sendalot = 0;
18008 				}
18009 			}
18010 		}
18011 		/* Window scaling. */
18012 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
18013 			to.to_wscale = tp->request_r_scale;
18014 			to.to_flags |= TOF_SCALE;
18015 		}
18016 		/* Timestamps. */
18017 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
18018 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
18019 			to.to_tsval = ms_cts + tp->ts_offset;
18020 			to.to_tsecr = tp->ts_recent;
18021 			to.to_flags |= TOF_TS;
18022 		}
18023 		/* Set receive buffer autosizing timestamp. */
18024 		if (tp->rfbuf_ts == 0 &&
18025 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
18026 			tp->rfbuf_ts = tcp_ts_getticks();
18027 		/* Selective ACK's. */
18028 		if (tp->t_flags & TF_SACK_PERMIT) {
18029 			if (flags & TH_SYN)
18030 				to.to_flags |= TOF_SACKPERM;
18031 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18032 				 tp->rcv_numsacks > 0) {
18033 				to.to_flags |= TOF_SACK;
18034 				to.to_nsacks = tp->rcv_numsacks;
18035 				to.to_sacks = (u_char *)tp->sackblks;
18036 			}
18037 		}
18038 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18039 		/* TCP-MD5 (RFC2385). */
18040 		if (tp->t_flags & TF_SIGNATURE)
18041 			to.to_flags |= TOF_SIGNATURE;
18042 #endif				/* TCP_SIGNATURE */
18043 
18044 		/* Processing the options. */
18045 		hdrlen += optlen = tcp_addoptions(&to, opt);
18046 		/*
18047 		 * If we wanted a TFO option to be added, but it was unable
18048 		 * to fit, ensure no data is sent.
18049 		 */
18050 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
18051 		    !(to.to_flags & TOF_FASTOPEN))
18052 			len = 0;
18053 	}
18054 	if (tp->t_port) {
18055 		if (V_tcp_udp_tunneling_port == 0) {
18056 			/* The port was removed?? */
18057 			SOCKBUF_UNLOCK(&so->so_snd);
18058 #ifdef TCP_ACCOUNTING
18059 			crtsc = get_cyclecount();
18060 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18061 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18062 			}
18063 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18064 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18065 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18066 			}
18067 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18068 			sched_unpin();
18069 #endif
18070 			return (EHOSTUNREACH);
18071 		}
18072 		hdrlen += sizeof(struct udphdr);
18073 	}
18074 #ifdef INET6
18075 	if (isipv6)
18076 		ipoptlen = ip6_optlen(inp);
18077 	else
18078 #endif
18079 		if (inp->inp_options)
18080 			ipoptlen = inp->inp_options->m_len -
18081 				offsetof(struct ipoption, ipopt_list);
18082 		else
18083 			ipoptlen = 0;
18084 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18085 	ipoptlen += ipsec_optlen;
18086 #endif
18087 
18088 	/*
18089 	 * Adjust data length if insertion of options will bump the packet
18090 	 * length beyond the t_maxseg length. Clear the FIN bit because we
18091 	 * cut off the tail of the segment.
18092 	 */
18093 	if (len + optlen + ipoptlen > tp->t_maxseg) {
18094 		if (tso) {
18095 			uint32_t if_hw_tsomax;
18096 			uint32_t moff;
18097 			int32_t max_len;
18098 
18099 			/* extract TSO information */
18100 			if_hw_tsomax = tp->t_tsomax;
18101 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18102 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18103 			KASSERT(ipoptlen == 0,
18104 				("%s: TSO can't do IP options", __func__));
18105 
18106 			/*
18107 			 * Check if we should limit by maximum payload
18108 			 * length:
18109 			 */
18110 			if (if_hw_tsomax != 0) {
18111 				/* compute maximum TSO length */
18112 				max_len = (if_hw_tsomax - hdrlen -
18113 					   max_linkhdr);
18114 				if (max_len <= 0) {
18115 					len = 0;
18116 				} else if (len > max_len) {
18117 					sendalot = 1;
18118 					len = max_len;
18119 					mark = 2;
18120 				}
18121 			}
18122 			/*
18123 			 * Prevent the last segment from being fractional
18124 			 * unless the send sockbuf can be emptied:
18125 			 */
18126 			max_len = (tp->t_maxseg - optlen);
18127 			if ((sb_offset + len) < sbavail(sb)) {
18128 				moff = len % (u_int)max_len;
18129 				if (moff != 0) {
18130 					mark = 3;
18131 					len -= moff;
18132 				}
18133 			}
18134 			/*
18135 			 * In case there are too many small fragments don't
18136 			 * use TSO:
18137 			 */
18138 			if (len <= segsiz) {
18139 				mark = 4;
18140 				tso = 0;
18141 			}
18142 			/*
18143 			 * Send the FIN in a separate segment after the bulk
18144 			 * sending is done. We don't trust the TSO
18145 			 * implementations to clear the FIN flag on all but
18146 			 * the last segment.
18147 			 */
18148 			if (tp->t_flags & TF_NEEDFIN) {
18149 				sendalot = 4;
18150 			}
18151 		} else {
18152 			mark = 5;
18153 			if (optlen + ipoptlen >= tp->t_maxseg) {
18154 				/*
18155 				 * Since we don't have enough space to put
18156 				 * the IP header chain and the TCP header in
18157 				 * one packet as required by RFC 7112, don't
18158 				 * send it. Also ensure that at least one
18159 				 * byte of the payload can be put into the
18160 				 * TCP segment.
18161 				 */
18162 				SOCKBUF_UNLOCK(&so->so_snd);
18163 				error = EMSGSIZE;
18164 				sack_rxmit = 0;
18165 				goto out;
18166 			}
18167 			len = tp->t_maxseg - optlen - ipoptlen;
18168 			sendalot = 5;
18169 		}
18170 	} else {
18171 		tso = 0;
18172 		mark = 6;
18173 	}
18174 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
18175 		("%s: len > IP_MAXPACKET", __func__));
18176 #ifdef DIAGNOSTIC
18177 #ifdef INET6
18178 	if (max_linkhdr + hdrlen > MCLBYTES)
18179 #else
18180 		if (max_linkhdr + hdrlen > MHLEN)
18181 #endif
18182 			panic("tcphdr too big");
18183 #endif
18184 
18185 	/*
18186 	 * This KASSERT is here to catch edge cases at a well defined place.
18187 	 * Before, those had triggered (random) panic conditions further
18188 	 * down.
18189 	 */
18190 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
18191 	if ((len == 0) &&
18192 	    (flags & TH_FIN) &&
18193 	    (sbused(sb))) {
18194 		/*
18195 		 * We have outstanding data, don't send a fin by itself!.
18196 		 */
18197 		goto just_return;
18198 	}
18199 	/*
18200 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
18201 	 * and initialize the header from the template for sends on this
18202 	 * connection.
18203 	 */
18204 	hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
18205 	if (len) {
18206 		uint32_t max_val;
18207 		uint32_t moff;
18208 
18209 		if (rack->r_ctl.rc_pace_max_segs)
18210 			max_val = rack->r_ctl.rc_pace_max_segs;
18211 		else if (rack->rc_user_set_max_segs)
18212 			max_val = rack->rc_user_set_max_segs * segsiz;
18213 		else
18214 			max_val = len;
18215 		/*
18216 		 * We allow a limit on sending with hptsi.
18217 		 */
18218 		if (len > max_val) {
18219 			mark = 7;
18220 			len = max_val;
18221 		}
18222 #ifdef INET6
18223 		if (MHLEN < hdrlen + max_linkhdr)
18224 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18225 		else
18226 #endif
18227 			m = m_gethdr(M_NOWAIT, MT_DATA);
18228 
18229 		if (m == NULL) {
18230 			SOCKBUF_UNLOCK(sb);
18231 			error = ENOBUFS;
18232 			sack_rxmit = 0;
18233 			goto out;
18234 		}
18235 		m->m_data += max_linkhdr;
18236 		m->m_len = hdrlen;
18237 
18238 		/*
18239 		 * Start the m_copy functions from the closest mbuf to the
18240 		 * sb_offset in the socket buffer chain.
18241 		 */
18242 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
18243 		s_mb = mb;
18244 		s_moff = moff;
18245 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
18246 			m_copydata(mb, moff, (int)len,
18247 				   mtod(m, caddr_t)+hdrlen);
18248 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18249 				sbsndptr_adv(sb, mb, len);
18250 			m->m_len += len;
18251 		} else {
18252 			struct sockbuf *msb;
18253 
18254 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18255 				msb = NULL;
18256 			else
18257 				msb = sb;
18258 			m->m_next = tcp_m_copym(
18259 				mb, moff, &len,
18260 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
18261 				((rsm == NULL) ? hw_tls : 0)
18262 #ifdef NETFLIX_COPY_ARGS
18263 				, &s_mb, &s_moff
18264 #endif
18265 				);
18266 			if (len <= (tp->t_maxseg - optlen)) {
18267 				/*
18268 				 * Must have ran out of mbufs for the copy
18269 				 * shorten it to no longer need tso. Lets
18270 				 * not put on sendalot since we are low on
18271 				 * mbufs.
18272 				 */
18273 				tso = 0;
18274 			}
18275 			if (m->m_next == NULL) {
18276 				SOCKBUF_UNLOCK(sb);
18277 				(void)m_free(m);
18278 				error = ENOBUFS;
18279 				sack_rxmit = 0;
18280 				goto out;
18281 			}
18282 		}
18283 		if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
18284 			if (rsm && (rsm->r_flags & RACK_TLP)) {
18285 				/*
18286 				 * TLP should not count in retran count, but
18287 				 * in its own bin
18288 				 */
18289 				counter_u64_add(rack_tlp_retran, 1);
18290 				counter_u64_add(rack_tlp_retran_bytes, len);
18291 			} else {
18292 				tp->t_sndrexmitpack++;
18293 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18294 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18295 			}
18296 #ifdef STATS
18297 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18298 						 len);
18299 #endif
18300 		} else {
18301 			KMOD_TCPSTAT_INC(tcps_sndpack);
18302 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
18303 #ifdef STATS
18304 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
18305 						 len);
18306 #endif
18307 		}
18308 		/*
18309 		 * If we're sending everything we've got, set PUSH. (This
18310 		 * will keep happy those implementations which only give
18311 		 * data to the user when a buffer fills or a PUSH comes in.)
18312 		 */
18313 		if (sb_offset + len == sbused(sb) &&
18314 		    sbused(sb) &&
18315 		    !(flags & TH_SYN)) {
18316 			flags |= TH_PUSH;
18317 			add_flag |= RACK_HAD_PUSH;
18318 		}
18319 
18320 		SOCKBUF_UNLOCK(sb);
18321 	} else {
18322 		SOCKBUF_UNLOCK(sb);
18323 		if (tp->t_flags & TF_ACKNOW)
18324 			KMOD_TCPSTAT_INC(tcps_sndacks);
18325 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
18326 			KMOD_TCPSTAT_INC(tcps_sndctrl);
18327 		else
18328 			KMOD_TCPSTAT_INC(tcps_sndwinup);
18329 
18330 		m = m_gethdr(M_NOWAIT, MT_DATA);
18331 		if (m == NULL) {
18332 			error = ENOBUFS;
18333 			sack_rxmit = 0;
18334 			goto out;
18335 		}
18336 #ifdef INET6
18337 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
18338 		    MHLEN >= hdrlen) {
18339 			M_ALIGN(m, hdrlen);
18340 		} else
18341 #endif
18342 			m->m_data += max_linkhdr;
18343 		m->m_len = hdrlen;
18344 	}
18345 	SOCKBUF_UNLOCK_ASSERT(sb);
18346 	m->m_pkthdr.rcvif = (struct ifnet *)0;
18347 #ifdef MAC
18348 	mac_inpcb_create_mbuf(inp, m);
18349 #endif
18350 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
18351 #ifdef INET6
18352 		if (isipv6)
18353 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18354 		else
18355 #endif				/* INET6 */
18356 #ifdef INET
18357 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18358 #endif
18359 		th = rack->r_ctl.fsb.th;
18360 		udp = rack->r_ctl.fsb.udp;
18361 		if (udp) {
18362 #ifdef INET6
18363 			if (isipv6)
18364 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
18365 			else
18366 #endif				/* INET6 */
18367 				ulen = hdrlen + len - sizeof(struct ip);
18368 			udp->uh_ulen = htons(ulen);
18369 		}
18370 	} else {
18371 #ifdef INET6
18372 		if (isipv6) {
18373 			ip6 = mtod(m, struct ip6_hdr *);
18374 			if (tp->t_port) {
18375 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
18376 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18377 				udp->uh_dport = tp->t_port;
18378 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
18379 				udp->uh_ulen = htons(ulen);
18380 				th = (struct tcphdr *)(udp + 1);
18381 			} else
18382 				th = (struct tcphdr *)(ip6 + 1);
18383 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
18384 		} else
18385 #endif				/* INET6 */
18386 		{
18387 #ifdef INET
18388 			ip = mtod(m, struct ip *);
18389 			if (tp->t_port) {
18390 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
18391 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18392 				udp->uh_dport = tp->t_port;
18393 				ulen = hdrlen + len - sizeof(struct ip);
18394 				udp->uh_ulen = htons(ulen);
18395 				th = (struct tcphdr *)(udp + 1);
18396 			} else
18397 				th = (struct tcphdr *)(ip + 1);
18398 			tcpip_fillheaders(inp, tp->t_port, ip, th);
18399 #endif
18400 		}
18401 	}
18402 	/*
18403 	 * Fill in fields, remembering maximum advertised window for use in
18404 	 * delaying messages about window sizes. If resending a FIN, be sure
18405 	 * not to use a new sequence number.
18406 	 */
18407 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
18408 	    tp->snd_nxt == tp->snd_max)
18409 		tp->snd_nxt--;
18410 	/*
18411 	 * If we are starting a connection, send ECN setup SYN packet. If we
18412 	 * are on a retransmit, we may resend those bits a number of times
18413 	 * as per RFC 3168.
18414 	 */
18415 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
18416 		flags |= tcp_ecn_output_syn_sent(tp);
18417 	}
18418 	/* Also handle parallel SYN for ECN */
18419 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
18420 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18421 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
18422 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18423 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
18424 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18425 #ifdef INET6
18426 		if (isipv6) {
18427 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18428 			ip6->ip6_flow |= htonl(ect << 20);
18429 		}
18430 		else
18431 #endif
18432 		{
18433 #ifdef INET
18434 			ip->ip_tos &= ~IPTOS_ECN_MASK;
18435 			ip->ip_tos |= ect;
18436 #endif
18437 		}
18438 	}
18439 	/*
18440 	 * If we are doing retransmissions, then snd_nxt will not reflect
18441 	 * the first unsent octet.  For ACK only packets, we do not want the
18442 	 * sequence number of the retransmitted packet, we want the sequence
18443 	 * number of the next unsent octet.  So, if there is no data (and no
18444 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
18445 	 * ti_seq.  But if we are in persist state, snd_max might reflect
18446 	 * one byte beyond the right edge of the window, so use snd_nxt in
18447 	 * that case, since we know we aren't doing a retransmission.
18448 	 * (retransmit and persist are mutually exclusive...)
18449 	 */
18450 	if (sack_rxmit == 0) {
18451 		if (len || (flags & (TH_SYN | TH_FIN))) {
18452 			th->th_seq = htonl(tp->snd_nxt);
18453 			rack_seq = tp->snd_nxt;
18454 		} else {
18455 			th->th_seq = htonl(tp->snd_max);
18456 			rack_seq = tp->snd_max;
18457 		}
18458 	} else {
18459 		th->th_seq = htonl(rsm->r_start);
18460 		rack_seq = rsm->r_start;
18461 	}
18462 	th->th_ack = htonl(tp->rcv_nxt);
18463 	tcp_set_flags(th, flags);
18464 	/*
18465 	 * Calculate receive window.  Don't shrink window, but avoid silly
18466 	 * window syndrome.
18467 	 * If a RST segment is sent, advertise a window of zero.
18468 	 */
18469 	if (flags & TH_RST) {
18470 		recwin = 0;
18471 	} else {
18472 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
18473 		    recwin < (long)segsiz) {
18474 			recwin = 0;
18475 		}
18476 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
18477 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
18478 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
18479 	}
18480 
18481 	/*
18482 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
18483 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
18484 	 * handled in syncache.
18485 	 */
18486 	if (flags & TH_SYN)
18487 		th->th_win = htons((u_short)
18488 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
18489 	else {
18490 		/* Avoid shrinking window with window scaling. */
18491 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
18492 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
18493 	}
18494 	/*
18495 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
18496 	 * window.  This may cause the remote transmitter to stall.  This
18497 	 * flag tells soreceive() to disable delayed acknowledgements when
18498 	 * draining the buffer.  This can occur if the receiver is
18499 	 * attempting to read more data than can be buffered prior to
18500 	 * transmitting on the connection.
18501 	 */
18502 	if (th->th_win == 0) {
18503 		tp->t_sndzerowin++;
18504 		tp->t_flags |= TF_RXWIN0SENT;
18505 	} else
18506 		tp->t_flags &= ~TF_RXWIN0SENT;
18507 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
18508 	/* Now are we using fsb?, if so copy the template data to the mbuf */
18509 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
18510 		uint8_t *cpto;
18511 
18512 		cpto = mtod(m, uint8_t *);
18513 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18514 		/*
18515 		 * We have just copied in:
18516 		 * IP/IP6
18517 		 * <optional udphdr>
18518 		 * tcphdr (no options)
18519 		 *
18520 		 * We need to grab the correct pointers into the mbuf
18521 		 * for both the tcp header, and possibly the udp header (if tunneling).
18522 		 * We do this by using the offset in the copy buffer and adding it
18523 		 * to the mbuf base pointer (cpto).
18524 		 */
18525 #ifdef INET6
18526 		if (isipv6)
18527 			ip6 = mtod(m, struct ip6_hdr *);
18528 		else
18529 #endif				/* INET6 */
18530 #ifdef INET
18531 			ip = mtod(m, struct ip *);
18532 #endif
18533 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18534 		/* If we have a udp header lets set it into the mbuf as well */
18535 		if (udp)
18536 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
18537 	}
18538 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18539 	if (to.to_flags & TOF_SIGNATURE) {
18540 		/*
18541 		 * Calculate MD5 signature and put it into the place
18542 		 * determined before.
18543 		 * NOTE: since TCP options buffer doesn't point into
18544 		 * mbuf's data, calculate offset and use it.
18545 		 */
18546 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18547 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18548 			/*
18549 			 * Do not send segment if the calculation of MD5
18550 			 * digest has failed.
18551 			 */
18552 			goto out;
18553 		}
18554 	}
18555 #endif
18556 	if (optlen) {
18557 		bcopy(opt, th + 1, optlen);
18558 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18559 	}
18560 	/*
18561 	 * Put TCP length in extended header, and then checksum extended
18562 	 * header and data.
18563 	 */
18564 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
18565 #ifdef INET6
18566 	if (isipv6) {
18567 		/*
18568 		 * ip6_plen is not need to be filled now, and will be filled
18569 		 * in ip6_output.
18570 		 */
18571 		if (tp->t_port) {
18572 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18573 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18574 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18575 			th->th_sum = htons(0);
18576 			UDPSTAT_INC(udps_opackets);
18577 		} else {
18578 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18579 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18580 			th->th_sum = in6_cksum_pseudo(ip6,
18581 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18582 						      0);
18583 		}
18584 	}
18585 #endif
18586 #if defined(INET6) && defined(INET)
18587 	else
18588 #endif
18589 #ifdef INET
18590 	{
18591 		if (tp->t_port) {
18592 			m->m_pkthdr.csum_flags = CSUM_UDP;
18593 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18594 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18595 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18596 			th->th_sum = htons(0);
18597 			UDPSTAT_INC(udps_opackets);
18598 		} else {
18599 			m->m_pkthdr.csum_flags = CSUM_TCP;
18600 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18601 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
18602 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18603 									IPPROTO_TCP + len + optlen));
18604 		}
18605 		/* IP version must be set here for ipv4/ipv6 checking later */
18606 		KASSERT(ip->ip_v == IPVERSION,
18607 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
18608 	}
18609 #endif
18610 	/*
18611 	 * Enable TSO and specify the size of the segments. The TCP pseudo
18612 	 * header checksum is always provided. XXX: Fixme: This is currently
18613 	 * not the case for IPv6.
18614 	 */
18615 	if (tso) {
18616 		KASSERT(len > tp->t_maxseg - optlen,
18617 			("%s: len <= tso_segsz", __func__));
18618 		m->m_pkthdr.csum_flags |= CSUM_TSO;
18619 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
18620 	}
18621 	KASSERT(len + hdrlen == m_length(m, NULL),
18622 		("%s: mbuf chain different than expected: %d + %u != %u",
18623 		 __func__, len, hdrlen, m_length(m, NULL)));
18624 
18625 #ifdef TCP_HHOOK
18626 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
18627 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
18628 #endif
18629 	/* We're getting ready to send; log now. */
18630 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
18631 		union tcp_log_stackspecific log;
18632 
18633 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18634 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
18635 		if (rack->rack_no_prr)
18636 			log.u_bbr.flex1 = 0;
18637 		else
18638 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18639 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18640 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18641 		log.u_bbr.flex4 = orig_len;
18642 		/* Save off the early/late values */
18643 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18644 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18645 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18646 		log.u_bbr.flex8 = 0;
18647 		if (rsm) {
18648 			if (rsm->r_flags & RACK_RWND_COLLAPSED) {
18649 				rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
18650 				counter_u64_add(rack_collapsed_win_rxt, 1);
18651 				counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
18652 			}
18653 			if (doing_tlp)
18654 				log.u_bbr.flex8 = 2;
18655 			else
18656 				log.u_bbr.flex8 = 1;
18657 		} else {
18658 			if (doing_tlp)
18659 				log.u_bbr.flex8 = 3;
18660 			else
18661 				log.u_bbr.flex8 = 0;
18662 		}
18663 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
18664 		log.u_bbr.flex7 = mark;
18665 		log.u_bbr.flex7 <<= 8;
18666 		log.u_bbr.flex7 |= pass;
18667 		log.u_bbr.pkts_out = tp->t_maxseg;
18668 		log.u_bbr.timeStamp = cts;
18669 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18670 		log.u_bbr.lt_epoch = cwnd_to_use;
18671 		log.u_bbr.delivered = sendalot;
18672 		lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
18673 				     len, &log, false, NULL, NULL, 0, &tv);
18674 	} else
18675 		lgb = NULL;
18676 
18677 	/*
18678 	 * Fill in IP length and desired time to live and send to IP level.
18679 	 * There should be a better way to handle ttl and tos; we could keep
18680 	 * them in the template, but need a way to checksum without them.
18681 	 */
18682 	/*
18683 	 * m->m_pkthdr.len should have been set before cksum calcuration,
18684 	 * because in6_cksum() need it.
18685 	 */
18686 #ifdef INET6
18687 	if (isipv6) {
18688 		/*
18689 		 * we separately set hoplimit for every segment, since the
18690 		 * user might want to change the value via setsockopt. Also,
18691 		 * desired default hop limit might be changed via Neighbor
18692 		 * Discovery.
18693 		 */
18694 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18695 
18696 		/*
18697 		 * Set the packet size here for the benefit of DTrace
18698 		 * probes. ip6_output() will set it properly; it's supposed
18699 		 * to include the option header lengths as well.
18700 		 */
18701 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18702 
18703 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18704 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18705 		else
18706 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18707 
18708 		if (tp->t_state == TCPS_SYN_SENT)
18709 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18710 
18711 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18712 		/* TODO: IPv6 IP6TOS_ECT bit on */
18713 		error = ip6_output(m,
18714 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18715 				   inp->in6p_outputopts,
18716 #else
18717 				   NULL,
18718 #endif
18719 				   &inp->inp_route6,
18720 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18721 				   NULL, NULL, inp);
18722 
18723 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18724 			mtu = inp->inp_route6.ro_nh->nh_mtu;
18725 	}
18726 #endif				/* INET6 */
18727 #if defined(INET) && defined(INET6)
18728 	else
18729 #endif
18730 #ifdef INET
18731 	{
18732 		ip->ip_len = htons(m->m_pkthdr.len);
18733 #ifdef INET6
18734 		if (inp->inp_vflag & INP_IPV6PROTO)
18735 			ip->ip_ttl = in6_selecthlim(inp, NULL);
18736 #endif				/* INET6 */
18737 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18738 		/*
18739 		 * If we do path MTU discovery, then we set DF on every
18740 		 * packet. This might not be the best thing to do according
18741 		 * to RFC3390 Section 2. However the tcp hostcache migitates
18742 		 * the problem so it affects only the first tcp connection
18743 		 * with a host.
18744 		 *
18745 		 * NB: Don't set DF on small MTU/MSS to have a safe
18746 		 * fallback.
18747 		 */
18748 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18749 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18750 			if (tp->t_port == 0 || len < V_tcp_minmss) {
18751 				ip->ip_off |= htons(IP_DF);
18752 			}
18753 		} else {
18754 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18755 		}
18756 
18757 		if (tp->t_state == TCPS_SYN_SENT)
18758 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18759 
18760 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
18761 
18762 		error = ip_output(m,
18763 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18764 				  inp->inp_options,
18765 #else
18766 				  NULL,
18767 #endif
18768 				  &inp->inp_route,
18769 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18770 				  inp);
18771 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18772 			mtu = inp->inp_route.ro_nh->nh_mtu;
18773 	}
18774 #endif				/* INET */
18775 
18776 out:
18777 	if (lgb) {
18778 		lgb->tlb_errno = error;
18779 		lgb = NULL;
18780 	}
18781 	/*
18782 	 * In transmit state, time the transmission and arrange for the
18783 	 * retransmit.  In persist state, just set snd_max.
18784 	 */
18785 	if (error == 0) {
18786 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
18787 		if (rsm && doing_tlp) {
18788 			rack->rc_last_sent_tlp_past_cumack = 0;
18789 			rack->rc_last_sent_tlp_seq_valid = 1;
18790 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18791 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18792 		}
18793 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
18794 		if (rsm && (doing_tlp == 0)) {
18795 			/* Set we retransmitted */
18796 			rack->rc_gp_saw_rec = 1;
18797 		} else {
18798 			if (cwnd_to_use > tp->snd_ssthresh) {
18799 				/* Set we sent in CA */
18800 				rack->rc_gp_saw_ca = 1;
18801 			} else {
18802 				/* Set we sent in SS */
18803 				rack->rc_gp_saw_ss = 1;
18804 			}
18805 		}
18806 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18807 		    (tp->t_flags & TF_SACK_PERMIT) &&
18808 		    tp->rcv_numsacks > 0)
18809 			tcp_clean_dsack_blocks(tp);
18810 		tot_len_this_send += len;
18811 		if (len == 0)
18812 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18813 		else if (len == 1) {
18814 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18815 		} else if (len > 1) {
18816 			int idx;
18817 
18818 			idx = (len / segsiz) + 3;
18819 			if (idx >= TCP_MSS_ACCT_ATIMER)
18820 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18821 			else
18822 				counter_u64_add(rack_out_size[idx], 1);
18823 		}
18824 	}
18825 	if ((rack->rack_no_prr == 0) &&
18826 	    sub_from_prr &&
18827 	    (error == 0)) {
18828 		if (rack->r_ctl.rc_prr_sndcnt >= len)
18829 			rack->r_ctl.rc_prr_sndcnt -= len;
18830 		else
18831 			rack->r_ctl.rc_prr_sndcnt = 0;
18832 	}
18833 	sub_from_prr = 0;
18834 	if (doing_tlp) {
18835 		/* Make sure the TLP is added */
18836 		add_flag |= RACK_TLP;
18837 	} else if (rsm) {
18838 		/* If its a resend without TLP then it must not have the flag */
18839 		rsm->r_flags &= ~RACK_TLP;
18840 	}
18841 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18842 			rack_to_usec_ts(&tv),
18843 			rsm, add_flag, s_mb, s_moff, hw_tls);
18844 
18845 
18846 	if ((error == 0) &&
18847 	    (len > 0) &&
18848 	    (tp->snd_una == tp->snd_max))
18849 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
18850 	{
18851 		tcp_seq startseq = tp->snd_nxt;
18852 
18853 		/* Track our lost count */
18854 		if (rsm && (doing_tlp == 0))
18855 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18856 		/*
18857 		 * Advance snd_nxt over sequence space of this segment.
18858 		 */
18859 		if (error)
18860 			/* We don't log or do anything with errors */
18861 			goto nomore;
18862 		if (doing_tlp == 0) {
18863 			if (rsm == NULL) {
18864 				/*
18865 				 * Not a retransmission of some
18866 				 * sort, new data is going out so
18867 				 * clear our TLP count and flag.
18868 				 */
18869 				rack->rc_tlp_in_progress = 0;
18870 				rack->r_ctl.rc_tlp_cnt_out = 0;
18871 			}
18872 		} else {
18873 			/*
18874 			 * We have just sent a TLP, mark that it is true
18875 			 * and make sure our in progress is set so we
18876 			 * continue to check the count.
18877 			 */
18878 			rack->rc_tlp_in_progress = 1;
18879 			rack->r_ctl.rc_tlp_cnt_out++;
18880 		}
18881 		if (flags & (TH_SYN | TH_FIN)) {
18882 			if (flags & TH_SYN)
18883 				tp->snd_nxt++;
18884 			if (flags & TH_FIN) {
18885 				tp->snd_nxt++;
18886 				tp->t_flags |= TF_SENTFIN;
18887 			}
18888 		}
18889 		/* In the ENOBUFS case we do *not* update snd_max */
18890 		if (sack_rxmit)
18891 			goto nomore;
18892 
18893 		tp->snd_nxt += len;
18894 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18895 			if (tp->snd_una == tp->snd_max) {
18896 				/*
18897 				 * Update the time we just added data since
18898 				 * none was outstanding.
18899 				 */
18900 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18901 				tp->t_acktime = ticks;
18902 			}
18903 			tp->snd_max = tp->snd_nxt;
18904 			/*
18905 			 * Time this transmission if not a retransmission and
18906 			 * not currently timing anything.
18907 			 * This is only relevant in case of switching back to
18908 			 * the base stack.
18909 			 */
18910 			if (tp->t_rtttime == 0) {
18911 				tp->t_rtttime = ticks;
18912 				tp->t_rtseq = startseq;
18913 				KMOD_TCPSTAT_INC(tcps_segstimed);
18914 			}
18915 			if (len &&
18916 			    ((tp->t_flags & TF_GPUTINPROG) == 0))
18917 				rack_start_gp_measurement(tp, rack, startseq, sb_offset);
18918 		}
18919 		/*
18920 		 * If we are doing FO we need to update the mbuf position and subtract
18921 		 * this happens when the peer sends us duplicate information and
18922 		 * we thus want to send a DSACK.
18923 		 *
18924 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
18925 		 * turned off? If not then we are going to echo multiple DSACK blocks
18926 		 * out (with the TSO), which we should not be doing.
18927 		 */
18928 		if (rack->r_fast_output && len) {
18929 			if (rack->r_ctl.fsb.left_to_send > len)
18930 				rack->r_ctl.fsb.left_to_send -= len;
18931 			else
18932 				rack->r_ctl.fsb.left_to_send = 0;
18933 			if (rack->r_ctl.fsb.left_to_send < segsiz)
18934 				rack->r_fast_output = 0;
18935 			if (rack->r_fast_output) {
18936 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
18937 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
18938 			}
18939 		}
18940 	}
18941 nomore:
18942 	if (error) {
18943 		rack->r_ctl.rc_agg_delayed = 0;
18944 		rack->r_early = 0;
18945 		rack->r_late = 0;
18946 		rack->r_ctl.rc_agg_early = 0;
18947 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
18948 		/*
18949 		 * Failures do not advance the seq counter above. For the
18950 		 * case of ENOBUFS we will fall out and retry in 1ms with
18951 		 * the hpts. Everything else will just have to retransmit
18952 		 * with the timer.
18953 		 *
18954 		 * In any case, we do not want to loop around for another
18955 		 * send without a good reason.
18956 		 */
18957 		sendalot = 0;
18958 		switch (error) {
18959 		case EPERM:
18960 			tp->t_softerror = error;
18961 #ifdef TCP_ACCOUNTING
18962 			crtsc = get_cyclecount();
18963 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18964 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18965 			}
18966 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18967 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18968 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18969 			}
18970 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18971 			sched_unpin();
18972 #endif
18973 			return (error);
18974 		case ENOBUFS:
18975 			/*
18976 			 * Pace us right away to retry in a some
18977 			 * time
18978 			 */
18979 			if (rack->r_ctl.crte != NULL) {
18980 				rack_trace_point(rack, RACK_TP_HWENOBUF);
18981 			} else
18982 				rack_trace_point(rack, RACK_TP_ENOBUF);
18983 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
18984 			if (rack->rc_enobuf < 0x7f)
18985 				rack->rc_enobuf++;
18986 			if (slot < (10 * HPTS_USEC_IN_MSEC))
18987 				slot = 10 * HPTS_USEC_IN_MSEC;
18988 			if (rack->r_ctl.crte != NULL) {
18989 				counter_u64_add(rack_saw_enobuf_hw, 1);
18990 				tcp_rl_log_enobuf(rack->r_ctl.crte);
18991 			}
18992 			counter_u64_add(rack_saw_enobuf, 1);
18993 			goto enobufs;
18994 		case EMSGSIZE:
18995 			/*
18996 			 * For some reason the interface we used initially
18997 			 * to send segments changed to another or lowered
18998 			 * its MTU. If TSO was active we either got an
18999 			 * interface without TSO capabilits or TSO was
19000 			 * turned off. If we obtained mtu from ip_output()
19001 			 * then update it and try again.
19002 			 */
19003 			if (tso)
19004 				tp->t_flags &= ~TF_TSO;
19005 			if (mtu != 0) {
19006 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
19007 				goto again;
19008 			}
19009 			slot = 10 * HPTS_USEC_IN_MSEC;
19010 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19011 #ifdef TCP_ACCOUNTING
19012 			crtsc = get_cyclecount();
19013 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19014 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19015 			}
19016 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19017 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19018 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19019 			}
19020 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19021 			sched_unpin();
19022 #endif
19023 			return (error);
19024 		case ENETUNREACH:
19025 			counter_u64_add(rack_saw_enetunreach, 1);
19026 		case EHOSTDOWN:
19027 		case EHOSTUNREACH:
19028 		case ENETDOWN:
19029 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
19030 				tp->t_softerror = error;
19031 			}
19032 			/* FALLTHROUGH */
19033 		default:
19034 			slot = 10 * HPTS_USEC_IN_MSEC;
19035 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19036 #ifdef TCP_ACCOUNTING
19037 			crtsc = get_cyclecount();
19038 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19039 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19040 			}
19041 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19042 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19043 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19044 			}
19045 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19046 			sched_unpin();
19047 #endif
19048 			return (error);
19049 		}
19050 	} else {
19051 		rack->rc_enobuf = 0;
19052 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
19053 			rack->r_ctl.retran_during_recovery += len;
19054 	}
19055 	KMOD_TCPSTAT_INC(tcps_sndtotal);
19056 
19057 	/*
19058 	 * Data sent (as far as we can tell). If this advertises a larger
19059 	 * window than any other segment, then remember the size of the
19060 	 * advertised window. Any pending ACK has now been sent.
19061 	 */
19062 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
19063 		tp->rcv_adv = tp->rcv_nxt + recwin;
19064 
19065 	tp->last_ack_sent = tp->rcv_nxt;
19066 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19067 enobufs:
19068 	if (sendalot) {
19069 		/* Do we need to turn off sendalot? */
19070 		if (rack->r_ctl.rc_pace_max_segs &&
19071 		    (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
19072 			/* We hit our max. */
19073 			sendalot = 0;
19074 		} else if ((rack->rc_user_set_max_segs) &&
19075 			   (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
19076 			/* We hit the user defined max */
19077 			sendalot = 0;
19078 		}
19079 	}
19080 	if ((error == 0) && (flags & TH_FIN))
19081 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
19082 	if (flags & TH_RST) {
19083 		/*
19084 		 * We don't send again after sending a RST.
19085 		 */
19086 		slot = 0;
19087 		sendalot = 0;
19088 		if (error == 0)
19089 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
19090 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
19091 		/*
19092 		 * Get our pacing rate, if an error
19093 		 * occurred in sending (ENOBUF) we would
19094 		 * hit the else if with slot preset. Other
19095 		 * errors return.
19096 		 */
19097 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
19098 	}
19099 	if (rsm &&
19100 	    (rsm->r_flags & RACK_HAS_SYN) == 0 &&
19101 	    rack->use_rack_rr) {
19102 		/* Its a retransmit and we use the rack cheat? */
19103 		if ((slot == 0) ||
19104 		    (rack->rc_always_pace == 0) ||
19105 		    (rack->r_rr_config == 1)) {
19106 			/*
19107 			 * We have no pacing set or we
19108 			 * are using old-style rack or
19109 			 * we are overridden to use the old 1ms pacing.
19110 			 */
19111 			slot = rack->r_ctl.rc_min_to;
19112 		}
19113 	}
19114 	/* We have sent clear the flag */
19115 	rack->r_ent_rec_ns = 0;
19116 	if (rack->r_must_retran) {
19117 		if (rsm) {
19118 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
19119 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
19120 				/*
19121 				 * We have retransmitted all.
19122 				 */
19123 				rack->r_must_retran = 0;
19124 				rack->r_ctl.rc_out_at_rto = 0;
19125 			}
19126 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19127 			/*
19128 			 * Sending new data will also kill
19129 			 * the loop.
19130 			 */
19131 			rack->r_must_retran = 0;
19132 			rack->r_ctl.rc_out_at_rto = 0;
19133 		}
19134 	}
19135 	rack->r_ctl.fsb.recwin = recwin;
19136 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
19137 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19138 		/*
19139 		 * We hit an RTO and now have past snd_max at the RTO
19140 		 * clear all the WAS flags.
19141 		 */
19142 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
19143 	}
19144 	if (slot) {
19145 		/* set the rack tcb into the slot N */
19146 		if ((error == 0) &&
19147 		    rack_use_rfo &&
19148 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
19149 		    (rsm == NULL) &&
19150 		    (tp->snd_nxt == tp->snd_max) &&
19151 		    (ipoptlen == 0) &&
19152 		    (tp->rcv_numsacks == 0) &&
19153 		    rack->r_fsb_inited &&
19154 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
19155 		    (rack->r_must_retran == 0) &&
19156 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
19157 		    (len > 0) && (orig_len > 0) &&
19158 		    (orig_len > len) &&
19159 		    ((orig_len - len) >= segsiz) &&
19160 		    ((optlen == 0) ||
19161 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19162 			/* We can send at least one more MSS using our fsb */
19163 
19164 			rack->r_fast_output = 1;
19165 			rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19166 			rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19167 			rack->r_ctl.fsb.tcp_flags = flags;
19168 			rack->r_ctl.fsb.left_to_send = orig_len - len;
19169 			if (hw_tls)
19170 				rack->r_ctl.fsb.hw_tls = 1;
19171 			else
19172 				rack->r_ctl.fsb.hw_tls = 0;
19173 			KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19174 				("rack:%p left_to_send:%u sbavail:%u out:%u",
19175 				 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19176 				 (tp->snd_max - tp->snd_una)));
19177 			if (rack->r_ctl.fsb.left_to_send < segsiz)
19178 				rack->r_fast_output = 0;
19179 			else {
19180 				if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19181 					rack->r_ctl.fsb.rfo_apply_push = 1;
19182 				else
19183 					rack->r_ctl.fsb.rfo_apply_push = 0;
19184 			}
19185 		} else
19186 			rack->r_fast_output = 0;
19187 		rack_log_fsb(rack, tp, so, flags,
19188 			     ipoptlen, orig_len, len, error,
19189 			     (rsm == NULL), optlen, __LINE__, 2);
19190 	} else if (sendalot) {
19191 		int ret;
19192 
19193 		sack_rxmit = 0;
19194 		if ((error == 0) &&
19195 		    rack_use_rfo &&
19196 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
19197 		    (rsm == NULL) &&
19198 		    (ipoptlen == 0) &&
19199 		    (tp->rcv_numsacks == 0) &&
19200 		    (tp->snd_nxt == tp->snd_max) &&
19201 		    (rack->r_must_retran == 0) &&
19202 		    rack->r_fsb_inited &&
19203 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
19204 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
19205 		    (len > 0) && (orig_len > 0) &&
19206 		    (orig_len > len) &&
19207 		    ((orig_len - len) >= segsiz) &&
19208 		    ((optlen == 0) ||
19209 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19210 			/* we can use fast_output for more */
19211 
19212 			rack->r_fast_output = 1;
19213 			rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19214 			rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19215 			rack->r_ctl.fsb.tcp_flags = flags;
19216 			rack->r_ctl.fsb.left_to_send = orig_len - len;
19217 			if (hw_tls)
19218 				rack->r_ctl.fsb.hw_tls = 1;
19219 			else
19220 				rack->r_ctl.fsb.hw_tls = 0;
19221 			KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19222 				("rack:%p left_to_send:%u sbavail:%u out:%u",
19223 				 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19224 				 (tp->snd_max - tp->snd_una)));
19225 			if (rack->r_ctl.fsb.left_to_send < segsiz) {
19226 				rack->r_fast_output = 0;
19227 			}
19228 			if (rack->r_fast_output) {
19229 				if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19230 					rack->r_ctl.fsb.rfo_apply_push = 1;
19231 				else
19232 					rack->r_ctl.fsb.rfo_apply_push = 0;
19233 				rack_log_fsb(rack, tp, so, flags,
19234 					     ipoptlen, orig_len, len, error,
19235 					     (rsm == NULL), optlen, __LINE__, 3);
19236 				error = 0;
19237 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19238 				if (ret >= 0)
19239 					return (ret);
19240 			        else if (error)
19241 					goto nomore;
19242 
19243 			}
19244 		}
19245 		goto again;
19246 	}
19247 	/* Assure when we leave that snd_nxt will point to top */
19248 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
19249 		tp->snd_nxt = tp->snd_max;
19250 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
19251 #ifdef TCP_ACCOUNTING
19252 	crtsc = get_cyclecount() - ts_val;
19253 	if (tot_len_this_send) {
19254 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19255 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
19256 		}
19257 		counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
19258 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19259 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
19260 		}
19261 		counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
19262 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19263 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
19264 		}
19265 		counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
19266 	} else {
19267 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19268 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
19269 		}
19270 		counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
19271 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19272 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
19273 		}
19274 		counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
19275 	}
19276 	sched_unpin();
19277 #endif
19278 	if (error == ENOBUFS)
19279 		error = 0;
19280 	return (error);
19281 }
19282 
19283 static void
19284 rack_update_seg(struct tcp_rack *rack)
19285 {
19286 	uint32_t orig_val;
19287 
19288 	orig_val = rack->r_ctl.rc_pace_max_segs;
19289 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
19290 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
19291 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
19292 }
19293 
19294 static void
19295 rack_mtu_change(struct tcpcb *tp)
19296 {
19297 	/*
19298 	 * The MSS may have changed
19299 	 */
19300 	struct tcp_rack *rack;
19301 	struct rack_sendmap *rsm;
19302 
19303 	rack = (struct tcp_rack *)tp->t_fb_ptr;
19304 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
19305 		/*
19306 		 * The MTU has changed we need to resend everything
19307 		 * since all we have sent is lost. We first fix
19308 		 * up the mtu though.
19309 		 */
19310 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
19311 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
19312 		rack_remxt_tmr(tp);
19313 		rack->r_fast_output = 0;
19314 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
19315 						rack->r_ctl.rc_sacked);
19316 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
19317 		rack->r_must_retran = 1;
19318 		/* Mark all inflight to needing to be rxt'd */
19319 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
19320 			rsm->r_flags |= RACK_MUST_RXT;
19321 		}
19322 	}
19323 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
19324 	/* We don't use snd_nxt to retransmit */
19325 	tp->snd_nxt = tp->snd_max;
19326 }
19327 
19328 static int
19329 rack_set_profile(struct tcp_rack *rack, int prof)
19330 {
19331 	int err = EINVAL;
19332 	if (prof == 1) {
19333 		/* pace_always=1 */
19334 		if (rack->rc_always_pace == 0) {
19335 			if (tcp_can_enable_pacing() == 0)
19336 				return (EBUSY);
19337 		}
19338 		rack->rc_always_pace = 1;
19339 		if (rack->use_fixed_rate || rack->gp_ready)
19340 			rack_set_cc_pacing(rack);
19341 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19342 		rack->rack_attempt_hdwr_pace = 0;
19343 		/* cmpack=1 */
19344 		if (rack_use_cmp_acks)
19345 			rack->r_use_cmp_ack = 1;
19346 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19347 		    rack->r_use_cmp_ack)
19348 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19349 		/* scwnd=1 */
19350 		rack->rack_enable_scwnd = 1;
19351 		/* dynamic=100 */
19352 		rack->rc_gp_dyn_mul = 1;
19353 		/* gp_inc_ca */
19354 		rack->r_ctl.rack_per_of_gp_ca = 100;
19355 		/* rrr_conf=3 */
19356 		rack->r_rr_config = 3;
19357 		/* npush=2 */
19358 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19359 		/* fillcw=1 */
19360 		rack->rc_pace_to_cwnd = 1;
19361 		rack->rc_pace_fill_if_rttin_range = 0;
19362 		rack->rtt_limit_mul = 0;
19363 		/* noprr=1 */
19364 		rack->rack_no_prr = 1;
19365 		/* lscwnd=1 */
19366 		rack->r_limit_scw = 1;
19367 		/* gp_inc_rec */
19368 		rack->r_ctl.rack_per_of_gp_rec = 90;
19369 		err = 0;
19370 
19371 	} else if (prof == 3) {
19372 		/* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
19373 		/* pace_always=1 */
19374 		if (rack->rc_always_pace == 0) {
19375 			if (tcp_can_enable_pacing() == 0)
19376 				return (EBUSY);
19377 		}
19378 		rack->rc_always_pace = 1;
19379 		if (rack->use_fixed_rate || rack->gp_ready)
19380 			rack_set_cc_pacing(rack);
19381 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19382 		rack->rack_attempt_hdwr_pace = 0;
19383 		/* cmpack=1 */
19384 		if (rack_use_cmp_acks)
19385 			rack->r_use_cmp_ack = 1;
19386 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19387 		    rack->r_use_cmp_ack)
19388 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19389 		/* scwnd=1 */
19390 		rack->rack_enable_scwnd = 1;
19391 		/* dynamic=100 */
19392 		rack->rc_gp_dyn_mul = 1;
19393 		/* gp_inc_ca */
19394 		rack->r_ctl.rack_per_of_gp_ca = 100;
19395 		/* rrr_conf=3 */
19396 		rack->r_rr_config = 3;
19397 		/* npush=2 */
19398 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19399 		/* fillcw=2 */
19400 		rack->rc_pace_to_cwnd = 1;
19401 		rack->r_fill_less_agg = 1;
19402 		rack->rc_pace_fill_if_rttin_range = 0;
19403 		rack->rtt_limit_mul = 0;
19404 		/* noprr=1 */
19405 		rack->rack_no_prr = 1;
19406 		/* lscwnd=1 */
19407 		rack->r_limit_scw = 1;
19408 		/* gp_inc_rec */
19409 		rack->r_ctl.rack_per_of_gp_rec = 90;
19410 		err = 0;
19411 
19412 
19413 	} else if (prof == 2) {
19414 		/* cmpack=1 */
19415 		if (rack->rc_always_pace == 0) {
19416 			if (tcp_can_enable_pacing() == 0)
19417 				return (EBUSY);
19418 		}
19419 		rack->rc_always_pace = 1;
19420 		if (rack->use_fixed_rate || rack->gp_ready)
19421 			rack_set_cc_pacing(rack);
19422 		rack->r_use_cmp_ack = 1;
19423 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19424 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19425 		/* pace_always=1 */
19426 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19427 		/* scwnd=1 */
19428 		rack->rack_enable_scwnd = 1;
19429 		/* dynamic=100 */
19430 		rack->rc_gp_dyn_mul = 1;
19431 		rack->r_ctl.rack_per_of_gp_ca = 100;
19432 		/* rrr_conf=3 */
19433 		rack->r_rr_config = 3;
19434 		/* npush=2 */
19435 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19436 		/* fillcw=1 */
19437 		rack->rc_pace_to_cwnd = 1;
19438 		rack->rc_pace_fill_if_rttin_range = 0;
19439 		rack->rtt_limit_mul = 0;
19440 		/* noprr=1 */
19441 		rack->rack_no_prr = 1;
19442 		/* lscwnd=0 */
19443 		rack->r_limit_scw = 0;
19444 		err = 0;
19445 	} else if (prof == 0) {
19446 		/* This changes things back to the default settings */
19447 		err = 0;
19448 		if (rack->rc_always_pace) {
19449 			tcp_decrement_paced_conn();
19450 			rack_undo_cc_pacing(rack);
19451 			rack->rc_always_pace = 0;
19452 		}
19453 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
19454 			rack->rc_always_pace = 1;
19455 			if (rack->use_fixed_rate || rack->gp_ready)
19456 				rack_set_cc_pacing(rack);
19457 		} else
19458 			rack->rc_always_pace = 0;
19459 		if (rack_dsack_std_based & 0x1) {
19460 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
19461 			rack->rc_rack_tmr_std_based = 1;
19462 		}
19463 		if (rack_dsack_std_based & 0x2) {
19464 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
19465 			rack->rc_rack_use_dsack = 1;
19466 		}
19467 		if (rack_use_cmp_acks)
19468 			rack->r_use_cmp_ack = 1;
19469 		else
19470 			rack->r_use_cmp_ack = 0;
19471 		if (rack_disable_prr)
19472 			rack->rack_no_prr = 1;
19473 		else
19474 			rack->rack_no_prr = 0;
19475 		if (rack_gp_no_rec_chg)
19476 			rack->rc_gp_no_rec_chg = 1;
19477 		else
19478 			rack->rc_gp_no_rec_chg = 0;
19479 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
19480 			rack->r_mbuf_queue = 1;
19481 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19482 				rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19483 			rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19484 		} else {
19485 			rack->r_mbuf_queue = 0;
19486 			rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19487 		}
19488 		if (rack_enable_shared_cwnd)
19489 			rack->rack_enable_scwnd = 1;
19490 		else
19491 			rack->rack_enable_scwnd = 0;
19492 		if (rack_do_dyn_mul) {
19493 			/* When dynamic adjustment is on CA needs to start at 100% */
19494 			rack->rc_gp_dyn_mul = 1;
19495 			if (rack_do_dyn_mul >= 100)
19496 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
19497 		} else {
19498 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
19499 			rack->rc_gp_dyn_mul = 0;
19500 		}
19501 		rack->r_rr_config = 0;
19502 		rack->r_ctl.rc_no_push_at_mrtt = 0;
19503 		rack->rc_pace_to_cwnd = 0;
19504 		rack->rc_pace_fill_if_rttin_range = 0;
19505 		rack->rtt_limit_mul = 0;
19506 
19507 		if (rack_enable_hw_pacing)
19508 			rack->rack_hdw_pace_ena = 1;
19509 		else
19510 			rack->rack_hdw_pace_ena = 0;
19511 		if (rack_disable_prr)
19512 			rack->rack_no_prr = 1;
19513 		else
19514 			rack->rack_no_prr = 0;
19515 		if (rack_limits_scwnd)
19516 			rack->r_limit_scw  = 1;
19517 		else
19518 			rack->r_limit_scw  = 0;
19519 		err = 0;
19520 	}
19521 	return (err);
19522 }
19523 
19524 static int
19525 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
19526 {
19527 	struct deferred_opt_list *dol;
19528 
19529 	dol = malloc(sizeof(struct deferred_opt_list),
19530 		     M_TCPFSB, M_NOWAIT|M_ZERO);
19531 	if (dol == NULL) {
19532 		/*
19533 		 * No space yikes -- fail out..
19534 		 */
19535 		return (0);
19536 	}
19537 	dol->optname = sopt_name;
19538 	dol->optval = loptval;
19539 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
19540 	return (1);
19541 }
19542 
19543 static int
19544 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
19545 		    uint32_t optval, uint64_t loptval)
19546 {
19547 	struct epoch_tracker et;
19548 	struct sockopt sopt;
19549 	struct cc_newreno_opts opt;
19550 	struct inpcb *inp = tptoinpcb(tp);
19551 	uint64_t val;
19552 	int error = 0;
19553 	uint16_t ca, ss;
19554 
19555 	switch (sopt_name) {
19556 
19557 	case TCP_RACK_DSACK_OPT:
19558 		RACK_OPTS_INC(tcp_rack_dsack_opt);
19559 		if (optval & 0x1) {
19560 			rack->rc_rack_tmr_std_based = 1;
19561 		} else {
19562 			rack->rc_rack_tmr_std_based = 0;
19563 		}
19564 		if (optval & 0x2) {
19565 			rack->rc_rack_use_dsack = 1;
19566 		} else {
19567 			rack->rc_rack_use_dsack = 0;
19568 		}
19569 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
19570 		break;
19571 	case TCP_RACK_PACING_BETA:
19572 		RACK_OPTS_INC(tcp_rack_beta);
19573 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
19574 			/* This only works for newreno. */
19575 			error = EINVAL;
19576 			break;
19577 		}
19578 		if (rack->rc_pacing_cc_set) {
19579 			/*
19580 			 * Set them into the real CC module
19581 			 * whats in the rack pcb is the old values
19582 			 * to be used on restoral/
19583 			 */
19584 			sopt.sopt_dir = SOPT_SET;
19585 			opt.name = CC_NEWRENO_BETA;
19586 			opt.val = optval;
19587 			if (CC_ALGO(tp)->ctl_output != NULL)
19588 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
19589 			else {
19590 				error = ENOENT;
19591 				break;
19592 			}
19593 		} else {
19594 			/*
19595 			 * Not pacing yet so set it into our local
19596 			 * rack pcb storage.
19597 			 */
19598 			rack->r_ctl.rc_saved_beta.beta = optval;
19599 		}
19600 		break;
19601 	case TCP_RACK_TIMER_SLOP:
19602 		RACK_OPTS_INC(tcp_rack_timer_slop);
19603 		rack->r_ctl.timer_slop = optval;
19604 		if (rack->rc_tp->t_srtt) {
19605 			/*
19606 			 * If we have an SRTT lets update t_rxtcur
19607 			 * to have the new slop.
19608 			 */
19609 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
19610 					   rack_rto_min, rack_rto_max,
19611 					   rack->r_ctl.timer_slop);
19612 		}
19613 		break;
19614 	case TCP_RACK_PACING_BETA_ECN:
19615 		RACK_OPTS_INC(tcp_rack_beta_ecn);
19616 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
19617 			/* This only works for newreno. */
19618 			error = EINVAL;
19619 			break;
19620 		}
19621 		if (rack->rc_pacing_cc_set) {
19622 			/*
19623 			 * Set them into the real CC module
19624 			 * whats in the rack pcb is the old values
19625 			 * to be used on restoral/
19626 			 */
19627 			sopt.sopt_dir = SOPT_SET;
19628 			opt.name = CC_NEWRENO_BETA_ECN;
19629 			opt.val = optval;
19630 			if (CC_ALGO(tp)->ctl_output != NULL)
19631 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
19632 			else
19633 				error = ENOENT;
19634 		} else {
19635 			/*
19636 			 * Not pacing yet so set it into our local
19637 			 * rack pcb storage.
19638 			 */
19639 			rack->r_ctl.rc_saved_beta.beta_ecn = optval;
19640 			rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
19641 		}
19642 		break;
19643 	case TCP_DEFER_OPTIONS:
19644 		RACK_OPTS_INC(tcp_defer_opt);
19645 		if (optval) {
19646 			if (rack->gp_ready) {
19647 				/* Too late */
19648 				error = EINVAL;
19649 				break;
19650 			}
19651 			rack->defer_options = 1;
19652 		} else
19653 			rack->defer_options = 0;
19654 		break;
19655 	case TCP_RACK_MEASURE_CNT:
19656 		RACK_OPTS_INC(tcp_rack_measure_cnt);
19657 		if (optval && (optval <= 0xff)) {
19658 			rack->r_ctl.req_measurements = optval;
19659 		} else
19660 			error = EINVAL;
19661 		break;
19662 	case TCP_REC_ABC_VAL:
19663 		RACK_OPTS_INC(tcp_rec_abc_val);
19664 		if (optval > 0)
19665 			rack->r_use_labc_for_rec = 1;
19666 		else
19667 			rack->r_use_labc_for_rec = 0;
19668 		break;
19669 	case TCP_RACK_ABC_VAL:
19670 		RACK_OPTS_INC(tcp_rack_abc_val);
19671 		if ((optval > 0) && (optval < 255))
19672 			rack->rc_labc = optval;
19673 		else
19674 			error = EINVAL;
19675 		break;
19676 	case TCP_HDWR_UP_ONLY:
19677 		RACK_OPTS_INC(tcp_pacing_up_only);
19678 		if (optval)
19679 			rack->r_up_only = 1;
19680 		else
19681 			rack->r_up_only = 0;
19682 		break;
19683 	case TCP_PACING_RATE_CAP:
19684 		RACK_OPTS_INC(tcp_pacing_rate_cap);
19685 		rack->r_ctl.bw_rate_cap = loptval;
19686 		break;
19687 	case TCP_RACK_PROFILE:
19688 		RACK_OPTS_INC(tcp_profile);
19689 		error = rack_set_profile(rack, optval);
19690 		break;
19691 	case TCP_USE_CMP_ACKS:
19692 		RACK_OPTS_INC(tcp_use_cmp_acks);
19693 		if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
19694 			/* You can't turn it off once its on! */
19695 			error = EINVAL;
19696 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
19697 			rack->r_use_cmp_ack = 1;
19698 			rack->r_mbuf_queue = 1;
19699 			inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19700 		}
19701 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
19702 			inp->inp_flags2 |= INP_MBUF_ACKCMP;
19703 		break;
19704 	case TCP_SHARED_CWND_TIME_LIMIT:
19705 		RACK_OPTS_INC(tcp_lscwnd);
19706 		if (optval)
19707 			rack->r_limit_scw = 1;
19708 		else
19709 			rack->r_limit_scw = 0;
19710 		break;
19711  	case TCP_RACK_PACE_TO_FILL:
19712 		RACK_OPTS_INC(tcp_fillcw);
19713 		if (optval == 0)
19714 			rack->rc_pace_to_cwnd = 0;
19715 		else {
19716 			rack->rc_pace_to_cwnd = 1;
19717 			if (optval > 1)
19718 				rack->r_fill_less_agg = 1;
19719 		}
19720 		if ((optval >= rack_gp_rtt_maxmul) &&
19721 		    rack_gp_rtt_maxmul &&
19722 		    (optval < 0xf)) {
19723 			rack->rc_pace_fill_if_rttin_range = 1;
19724 			rack->rtt_limit_mul = optval;
19725 		} else {
19726 			rack->rc_pace_fill_if_rttin_range = 0;
19727 			rack->rtt_limit_mul = 0;
19728 		}
19729 		break;
19730 	case TCP_RACK_NO_PUSH_AT_MAX:
19731 		RACK_OPTS_INC(tcp_npush);
19732 		if (optval == 0)
19733 			rack->r_ctl.rc_no_push_at_mrtt = 0;
19734 		else if (optval < 0xff)
19735 			rack->r_ctl.rc_no_push_at_mrtt = optval;
19736 		else
19737 			error = EINVAL;
19738 		break;
19739 	case TCP_SHARED_CWND_ENABLE:
19740 		RACK_OPTS_INC(tcp_rack_scwnd);
19741 		if (optval == 0)
19742 			rack->rack_enable_scwnd = 0;
19743 		else
19744 			rack->rack_enable_scwnd = 1;
19745 		break;
19746 	case TCP_RACK_MBUF_QUEUE:
19747 		/* Now do we use the LRO mbuf-queue feature */
19748 		RACK_OPTS_INC(tcp_rack_mbufq);
19749 		if (optval || rack->r_use_cmp_ack)
19750 			rack->r_mbuf_queue = 1;
19751 		else
19752 			rack->r_mbuf_queue = 0;
19753 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19754 			inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19755 		else
19756 			inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19757 		break;
19758 	case TCP_RACK_NONRXT_CFG_RATE:
19759 		RACK_OPTS_INC(tcp_rack_cfg_rate);
19760 		if (optval == 0)
19761 			rack->rack_rec_nonrxt_use_cr = 0;
19762 		else
19763 			rack->rack_rec_nonrxt_use_cr = 1;
19764 		break;
19765 	case TCP_NO_PRR:
19766 		RACK_OPTS_INC(tcp_rack_noprr);
19767 		if (optval == 0)
19768 			rack->rack_no_prr = 0;
19769 		else if (optval == 1)
19770 			rack->rack_no_prr = 1;
19771 		else if (optval == 2)
19772 			rack->no_prr_addback = 1;
19773 		else
19774 			error = EINVAL;
19775 		break;
19776 	case TCP_TIMELY_DYN_ADJ:
19777 		RACK_OPTS_INC(tcp_timely_dyn);
19778 		if (optval == 0)
19779 			rack->rc_gp_dyn_mul = 0;
19780 		else {
19781 			rack->rc_gp_dyn_mul = 1;
19782 			if (optval >= 100) {
19783 				/*
19784 				 * If the user sets something 100 or more
19785 				 * its the gp_ca value.
19786 				 */
19787 				rack->r_ctl.rack_per_of_gp_ca  = optval;
19788 			}
19789 		}
19790 		break;
19791 	case TCP_RACK_DO_DETECTION:
19792 		RACK_OPTS_INC(tcp_rack_do_detection);
19793 		if (optval == 0)
19794 			rack->do_detection = 0;
19795 		else
19796 			rack->do_detection = 1;
19797 		break;
19798 	case TCP_RACK_TLP_USE:
19799 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19800 			error = EINVAL;
19801 			break;
19802 		}
19803 		RACK_OPTS_INC(tcp_tlp_use);
19804 		rack->rack_tlp_threshold_use = optval;
19805 		break;
19806 	case TCP_RACK_TLP_REDUCE:
19807 		/* RACK TLP cwnd reduction (bool) */
19808 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
19809 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19810 		break;
19811 	/*  Pacing related ones */
19812 	case TCP_RACK_PACE_ALWAYS:
19813 		/*
19814 		 * zero is old rack method, 1 is new
19815 		 * method using a pacing rate.
19816 		 */
19817 		RACK_OPTS_INC(tcp_rack_pace_always);
19818 		if (optval > 0) {
19819 			if (rack->rc_always_pace) {
19820 				error = EALREADY;
19821 				break;
19822 			} else if (tcp_can_enable_pacing()) {
19823 				rack->rc_always_pace = 1;
19824 				if (rack->use_fixed_rate || rack->gp_ready)
19825 					rack_set_cc_pacing(rack);
19826 			}
19827 			else {
19828 				error = ENOSPC;
19829 				break;
19830 			}
19831 		} else {
19832 			if (rack->rc_always_pace) {
19833 				tcp_decrement_paced_conn();
19834 				rack->rc_always_pace = 0;
19835 				rack_undo_cc_pacing(rack);
19836 			}
19837 		}
19838 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19839 			inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19840 		else
19841 			inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19842 		/* A rate may be set irate or other, if so set seg size */
19843 		rack_update_seg(rack);
19844 		break;
19845 	case TCP_BBR_RACK_INIT_RATE:
19846 		RACK_OPTS_INC(tcp_initial_rate);
19847 		val = optval;
19848 		/* Change from kbits per second to bytes per second */
19849 		val *= 1000;
19850 		val /= 8;
19851 		rack->r_ctl.init_rate = val;
19852 		if (rack->rc_init_win != rack_default_init_window) {
19853 			uint32_t win, snt;
19854 
19855 			/*
19856 			 * Options don't always get applied
19857 			 * in the order you think. So in order
19858 			 * to assure we update a cwnd we need
19859 			 * to check and see if we are still
19860 			 * where we should raise the cwnd.
19861 			 */
19862 			win = rc_init_window(rack);
19863 			if (SEQ_GT(tp->snd_max, tp->iss))
19864 				snt = tp->snd_max - tp->iss;
19865 			else
19866 				snt = 0;
19867 			if ((snt < win) &&
19868 			    (tp->snd_cwnd < win))
19869 				tp->snd_cwnd = win;
19870 		}
19871 		if (rack->rc_always_pace)
19872 			rack_update_seg(rack);
19873 		break;
19874 	case TCP_BBR_IWINTSO:
19875 		RACK_OPTS_INC(tcp_initial_win);
19876 		if (optval && (optval <= 0xff)) {
19877 			uint32_t win, snt;
19878 
19879 			rack->rc_init_win = optval;
19880 			win = rc_init_window(rack);
19881 			if (SEQ_GT(tp->snd_max, tp->iss))
19882 				snt = tp->snd_max - tp->iss;
19883 			else
19884 				snt = 0;
19885 			if ((snt < win) &&
19886 			    (tp->t_srtt |
19887 #ifdef NETFLIX_PEAKRATE
19888 			     tp->t_maxpeakrate |
19889 #endif
19890 			     rack->r_ctl.init_rate)) {
19891 				/*
19892 				 * We are not past the initial window
19893 				 * and we have some bases for pacing,
19894 				 * so we need to possibly adjust up
19895 				 * the cwnd. Note even if we don't set
19896 				 * the cwnd, its still ok to raise the rc_init_win
19897 				 * which can be used coming out of idle when we
19898 				 * would have a rate.
19899 				 */
19900 				if (tp->snd_cwnd < win)
19901 					tp->snd_cwnd = win;
19902 			}
19903 			if (rack->rc_always_pace)
19904 				rack_update_seg(rack);
19905 		} else
19906 			error = EINVAL;
19907 		break;
19908 	case TCP_RACK_FORCE_MSEG:
19909 		RACK_OPTS_INC(tcp_rack_force_max_seg);
19910 		if (optval)
19911 			rack->rc_force_max_seg = 1;
19912 		else
19913 			rack->rc_force_max_seg = 0;
19914 		break;
19915 	case TCP_RACK_PACE_MAX_SEG:
19916 		/* Max segments size in a pace in bytes */
19917 		RACK_OPTS_INC(tcp_rack_max_seg);
19918 		rack->rc_user_set_max_segs = optval;
19919 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
19920 		break;
19921 	case TCP_RACK_PACE_RATE_REC:
19922 		/* Set the fixed pacing rate in Bytes per second ca */
19923 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
19924 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19925 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19926 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19927 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19928 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19929 		rack->use_fixed_rate = 1;
19930 		if (rack->rc_always_pace)
19931 			rack_set_cc_pacing(rack);
19932 		rack_log_pacing_delay_calc(rack,
19933 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
19934 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
19935 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19936 					   __LINE__, NULL,0);
19937 		break;
19938 
19939 	case TCP_RACK_PACE_RATE_SS:
19940 		/* Set the fixed pacing rate in Bytes per second ca */
19941 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
19942 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19943 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
19944 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19945 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19946 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19947 		rack->use_fixed_rate = 1;
19948 		if (rack->rc_always_pace)
19949 			rack_set_cc_pacing(rack);
19950 		rack_log_pacing_delay_calc(rack,
19951 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
19952 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
19953 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19954 					   __LINE__, NULL, 0);
19955 		break;
19956 
19957 	case TCP_RACK_PACE_RATE_CA:
19958 		/* Set the fixed pacing rate in Bytes per second ca */
19959 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
19960 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
19961 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
19962 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
19963 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
19964 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
19965 		rack->use_fixed_rate = 1;
19966 		if (rack->rc_always_pace)
19967 			rack_set_cc_pacing(rack);
19968 		rack_log_pacing_delay_calc(rack,
19969 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
19970 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
19971 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
19972 					   __LINE__, NULL, 0);
19973 		break;
19974 	case TCP_RACK_GP_INCREASE_REC:
19975 		RACK_OPTS_INC(tcp_gp_inc_rec);
19976 		rack->r_ctl.rack_per_of_gp_rec = optval;
19977 		rack_log_pacing_delay_calc(rack,
19978 					   rack->r_ctl.rack_per_of_gp_ss,
19979 					   rack->r_ctl.rack_per_of_gp_ca,
19980 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19981 					   __LINE__, NULL, 0);
19982 		break;
19983 	case TCP_RACK_GP_INCREASE_CA:
19984 		RACK_OPTS_INC(tcp_gp_inc_ca);
19985 		ca = optval;
19986 		if (ca < 100) {
19987 			/*
19988 			 * We don't allow any reduction
19989 			 * over the GP b/w.
19990 			 */
19991 			error = EINVAL;
19992 			break;
19993 		}
19994 		rack->r_ctl.rack_per_of_gp_ca = ca;
19995 		rack_log_pacing_delay_calc(rack,
19996 					   rack->r_ctl.rack_per_of_gp_ss,
19997 					   rack->r_ctl.rack_per_of_gp_ca,
19998 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
19999 					   __LINE__, NULL, 0);
20000 		break;
20001 	case TCP_RACK_GP_INCREASE_SS:
20002 		RACK_OPTS_INC(tcp_gp_inc_ss);
20003 		ss = optval;
20004 		if (ss < 100) {
20005 			/*
20006 			 * We don't allow any reduction
20007 			 * over the GP b/w.
20008 			 */
20009 			error = EINVAL;
20010 			break;
20011 		}
20012 		rack->r_ctl.rack_per_of_gp_ss = ss;
20013 		rack_log_pacing_delay_calc(rack,
20014 					   rack->r_ctl.rack_per_of_gp_ss,
20015 					   rack->r_ctl.rack_per_of_gp_ca,
20016 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20017 					   __LINE__, NULL, 0);
20018 		break;
20019 	case TCP_RACK_RR_CONF:
20020 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
20021 		if (optval && optval <= 3)
20022 			rack->r_rr_config = optval;
20023 		else
20024 			rack->r_rr_config = 0;
20025 		break;
20026 	case TCP_HDWR_RATE_CAP:
20027 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
20028 		if (optval) {
20029 			if (rack->r_rack_hw_rate_caps == 0)
20030 				rack->r_rack_hw_rate_caps = 1;
20031 			else
20032 				error = EALREADY;
20033 		} else {
20034 			rack->r_rack_hw_rate_caps = 0;
20035 		}
20036 		break;
20037 	case TCP_BBR_HDWR_PACE:
20038 		RACK_OPTS_INC(tcp_hdwr_pacing);
20039 		if (optval){
20040 			if (rack->rack_hdrw_pacing == 0) {
20041 				rack->rack_hdw_pace_ena = 1;
20042 				rack->rack_attempt_hdwr_pace = 0;
20043 			} else
20044 				error = EALREADY;
20045 		} else {
20046 			rack->rack_hdw_pace_ena = 0;
20047 #ifdef RATELIMIT
20048 			if (rack->r_ctl.crte != NULL) {
20049 				rack->rack_hdrw_pacing = 0;
20050 				rack->rack_attempt_hdwr_pace = 0;
20051 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
20052 				rack->r_ctl.crte = NULL;
20053 			}
20054 #endif
20055 		}
20056 		break;
20057 	/*  End Pacing related ones */
20058 	case TCP_RACK_PRR_SENDALOT:
20059 		/* Allow PRR to send more than one seg */
20060 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
20061 		rack->r_ctl.rc_prr_sendalot = optval;
20062 		break;
20063 	case TCP_RACK_MIN_TO:
20064 		/* Minimum time between rack t-o's in ms */
20065 		RACK_OPTS_INC(tcp_rack_min_to);
20066 		rack->r_ctl.rc_min_to = optval;
20067 		break;
20068 	case TCP_RACK_EARLY_SEG:
20069 		/* If early recovery max segments */
20070 		RACK_OPTS_INC(tcp_rack_early_seg);
20071 		rack->r_ctl.rc_early_recovery_segs = optval;
20072 		break;
20073 	case TCP_RACK_ENABLE_HYSTART:
20074 	{
20075 		if (optval) {
20076 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
20077 			if (rack_do_hystart > RACK_HYSTART_ON)
20078 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
20079 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
20080 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
20081 		} else {
20082 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
20083 		}
20084 	}
20085 	break;
20086 	case TCP_RACK_REORD_THRESH:
20087 		/* RACK reorder threshold (shift amount) */
20088 		RACK_OPTS_INC(tcp_rack_reord_thresh);
20089 		if ((optval > 0) && (optval < 31))
20090 			rack->r_ctl.rc_reorder_shift = optval;
20091 		else
20092 			error = EINVAL;
20093 		break;
20094 	case TCP_RACK_REORD_FADE:
20095 		/* Does reordering fade after ms time */
20096 		RACK_OPTS_INC(tcp_rack_reord_fade);
20097 		rack->r_ctl.rc_reorder_fade = optval;
20098 		break;
20099 	case TCP_RACK_TLP_THRESH:
20100 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
20101 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
20102 		if (optval)
20103 			rack->r_ctl.rc_tlp_threshold = optval;
20104 		else
20105 			error = EINVAL;
20106 		break;
20107 	case TCP_BBR_USE_RACK_RR:
20108 		RACK_OPTS_INC(tcp_rack_rr);
20109 		if (optval)
20110 			rack->use_rack_rr = 1;
20111 		else
20112 			rack->use_rack_rr = 0;
20113 		break;
20114 	case TCP_FAST_RSM_HACK:
20115 		RACK_OPTS_INC(tcp_rack_fastrsm_hack);
20116 		if (optval)
20117 			rack->fast_rsm_hack = 1;
20118 		else
20119 			rack->fast_rsm_hack = 0;
20120 		break;
20121 	case TCP_RACK_PKT_DELAY:
20122 		/* RACK added ms i.e. rack-rtt + reord + N */
20123 		RACK_OPTS_INC(tcp_rack_pkt_delay);
20124 		rack->r_ctl.rc_pkt_delay = optval;
20125 		break;
20126 	case TCP_DELACK:
20127 		RACK_OPTS_INC(tcp_rack_delayed_ack);
20128 		if (optval == 0)
20129 			tp->t_delayed_ack = 0;
20130 		else
20131 			tp->t_delayed_ack = 1;
20132 		if (tp->t_flags & TF_DELACK) {
20133 			tp->t_flags &= ~TF_DELACK;
20134 			tp->t_flags |= TF_ACKNOW;
20135 			NET_EPOCH_ENTER(et);
20136 			rack_output(tp);
20137 			NET_EPOCH_EXIT(et);
20138 		}
20139 		break;
20140 
20141 	case TCP_BBR_RACK_RTT_USE:
20142 		RACK_OPTS_INC(tcp_rack_rtt_use);
20143 		if ((optval != USE_RTT_HIGH) &&
20144 		    (optval != USE_RTT_LOW) &&
20145 		    (optval != USE_RTT_AVG))
20146 			error = EINVAL;
20147 		else
20148 			rack->r_ctl.rc_rate_sample_method = optval;
20149 		break;
20150 	case TCP_DATA_AFTER_CLOSE:
20151 		RACK_OPTS_INC(tcp_data_after_close);
20152 		if (optval)
20153 			rack->rc_allow_data_af_clo = 1;
20154 		else
20155 			rack->rc_allow_data_af_clo = 0;
20156 		break;
20157 	default:
20158 		break;
20159 	}
20160 #ifdef NETFLIX_STATS
20161 	tcp_log_socket_option(tp, sopt_name, optval, error);
20162 #endif
20163 	return (error);
20164 }
20165 
20166 
20167 static void
20168 rack_apply_deferred_options(struct tcp_rack *rack)
20169 {
20170 	struct deferred_opt_list *dol, *sdol;
20171 	uint32_t s_optval;
20172 
20173 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
20174 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
20175 		/* Disadvantage of deferal is you loose the error return */
20176 		s_optval = (uint32_t)dol->optval;
20177 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
20178 		free(dol, M_TCPDO);
20179 	}
20180 }
20181 
20182 static void
20183 rack_hw_tls_change(struct tcpcb *tp, int chg)
20184 {
20185 	/*
20186 	 * HW tls state has changed.. fix all
20187 	 * rsm's in flight.
20188 	 */
20189 	struct tcp_rack *rack;
20190 	struct rack_sendmap *rsm;
20191 
20192 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20193 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
20194 		if (chg)
20195 			rsm->r_hw_tls = 1;
20196 		else
20197 			rsm->r_hw_tls = 0;
20198 	}
20199 	if (chg)
20200 		rack->r_ctl.fsb.hw_tls = 1;
20201 	else
20202 		rack->r_ctl.fsb.hw_tls = 0;
20203 }
20204 
20205 static int
20206 rack_pru_options(struct tcpcb *tp, int flags)
20207 {
20208 	if (flags & PRUS_OOB)
20209 		return (EOPNOTSUPP);
20210 	return (0);
20211 }
20212 
20213 static struct tcp_function_block __tcp_rack = {
20214 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
20215 	.tfb_tcp_output = rack_output,
20216 	.tfb_do_queued_segments = ctf_do_queued_segments,
20217 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
20218 	.tfb_tcp_do_segment = rack_do_segment,
20219 	.tfb_tcp_ctloutput = rack_ctloutput,
20220 	.tfb_tcp_fb_init = rack_init,
20221 	.tfb_tcp_fb_fini = rack_fini,
20222 	.tfb_tcp_timer_stop_all = rack_stopall,
20223 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
20224 	.tfb_tcp_handoff_ok = rack_handoff_ok,
20225 	.tfb_tcp_mtu_chg = rack_mtu_change,
20226 	.tfb_pru_options = rack_pru_options,
20227 	.tfb_hwtls_change = rack_hw_tls_change,
20228 	.tfb_compute_pipe = rack_compute_pipe,
20229 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
20230 };
20231 
20232 /*
20233  * rack_ctloutput() must drop the inpcb lock before performing copyin on
20234  * socket option arguments.  When it re-acquires the lock after the copy, it
20235  * has to revalidate that the connection is still valid for the socket
20236  * option.
20237  */
20238 static int
20239 rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt)
20240 {
20241 #ifdef INET6
20242 	struct ip6_hdr *ip6;
20243 #endif
20244 #ifdef INET
20245 	struct ip *ip;
20246 #endif
20247 	struct tcpcb *tp;
20248 	struct tcp_rack *rack;
20249 	uint64_t loptval;
20250 	int32_t error = 0, optval;
20251 
20252 	tp = intotcpcb(inp);
20253 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20254 	if (rack == NULL) {
20255 		INP_WUNLOCK(inp);
20256 		return (EINVAL);
20257 	}
20258 #ifdef INET6
20259 	ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
20260 #endif
20261 #ifdef INET
20262 	ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
20263 #endif
20264 
20265 	switch (sopt->sopt_level) {
20266 #ifdef INET6
20267 	case IPPROTO_IPV6:
20268 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
20269 		switch (sopt->sopt_name) {
20270 		case IPV6_USE_MIN_MTU:
20271 			tcp6_use_min_mtu(tp);
20272 			break;
20273 		case IPV6_TCLASS:
20274 			/*
20275 			 * The DSCP codepoint has changed, update the fsb.
20276 			 */
20277 			ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
20278 			    (rack->rc_inp->inp_flow & IPV6_FLOWINFO_MASK);
20279 			break;
20280 		}
20281 		INP_WUNLOCK(inp);
20282 		return (0);
20283 #endif
20284 #ifdef INET
20285 	case IPPROTO_IP:
20286 		switch (sopt->sopt_name) {
20287 		case IP_TOS:
20288 			/*
20289 			 * The DSCP codepoint has changed, update the fsb.
20290 			 */
20291 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
20292 			break;
20293 		case IP_TTL:
20294 			/*
20295 			 * The TTL has changed, update the fsb.
20296 			 */
20297 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
20298 			break;
20299 		}
20300 		INP_WUNLOCK(inp);
20301 		return (0);
20302 #endif
20303 	}
20304 
20305 	switch (sopt->sopt_name) {
20306 	case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
20307 	/*  Pacing related ones */
20308 	case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
20309 	case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
20310 	case TCP_BBR_IWINTSO:			/*  URL:tso_iwin */
20311 	case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
20312 	case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
20313 	case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
20314 	case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
20315 	case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
20316 	case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
20317 	case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
20318 	case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
20319 	case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
20320 	case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
20321 	case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
20322 	case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
20323 	case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
20324        /* End pacing related */
20325 	case TCP_FAST_RSM_HACK:			/*  URL:frsm_hack */
20326 	case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
20327 	case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
20328 	case TCP_RACK_MIN_TO:			/*  URL:min_to */
20329 	case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
20330 	case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
20331 	case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
20332 	case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
20333 	case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
20334 	case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
20335 	case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
20336 	case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
20337 	case TCP_RACK_DO_DETECTION:		/*  URL:detect */
20338 	case TCP_NO_PRR:			/*  URL:noprr */
20339 	case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
20340 	case TCP_DATA_AFTER_CLOSE:		/*  no URL */
20341 	case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
20342 	case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
20343 	case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
20344 	case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
20345 	case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
20346 	case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
20347 	case TCP_RACK_PROFILE:			/*  URL:profile */
20348 	case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
20349 	case TCP_RACK_ABC_VAL:			/*  URL:labc */
20350 	case TCP_REC_ABC_VAL:			/*  URL:reclabc */
20351 	case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
20352 	case TCP_DEFER_OPTIONS:			/*  URL:defer */
20353 	case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
20354 	case TCP_RACK_PACING_BETA:		/*  URL:pacing_beta */
20355 	case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
20356 	case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
20357 	case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
20358 		break;
20359 	default:
20360 		/* Filter off all unknown options to the base stack */
20361 		return (tcp_default_ctloutput(inp, sopt));
20362 		break;
20363 	}
20364 	INP_WUNLOCK(inp);
20365 	if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
20366 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
20367 		/*
20368 		 * We truncate it down to 32 bits for the socket-option trace this
20369 		 * means rates > 34Gbps won't show right, but thats probably ok.
20370 		 */
20371 		optval = (uint32_t)loptval;
20372 	} else {
20373 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
20374 		/* Save it in 64 bit form too */
20375 		loptval = optval;
20376 	}
20377 	if (error)
20378 		return (error);
20379 	INP_WLOCK(inp);
20380 	if (inp->inp_flags & INP_DROPPED) {
20381 		INP_WUNLOCK(inp);
20382 		return (ECONNRESET);
20383 	}
20384 	if (tp->t_fb != &__tcp_rack) {
20385 		INP_WUNLOCK(inp);
20386 		return (ENOPROTOOPT);
20387 	}
20388 	if (rack->defer_options && (rack->gp_ready == 0) &&
20389 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
20390 	    (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
20391 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
20392 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
20393 		/* Options are beind deferred */
20394 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
20395 			INP_WUNLOCK(inp);
20396 			return (0);
20397 		} else {
20398 			/* No memory to defer, fail */
20399 			INP_WUNLOCK(inp);
20400 			return (ENOMEM);
20401 		}
20402 	}
20403 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
20404 	INP_WUNLOCK(inp);
20405 	return (error);
20406 }
20407 
20408 static void
20409 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
20410 {
20411 
20412 	INP_WLOCK_ASSERT(tptoinpcb(tp));
20413 	bzero(ti, sizeof(*ti));
20414 
20415 	ti->tcpi_state = tp->t_state;
20416 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
20417 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
20418 	if (tp->t_flags & TF_SACK_PERMIT)
20419 		ti->tcpi_options |= TCPI_OPT_SACK;
20420 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
20421 		ti->tcpi_options |= TCPI_OPT_WSCALE;
20422 		ti->tcpi_snd_wscale = tp->snd_scale;
20423 		ti->tcpi_rcv_wscale = tp->rcv_scale;
20424 	}
20425 	if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
20426 		ti->tcpi_options |= TCPI_OPT_ECN;
20427 	if (tp->t_flags & TF_FASTOPEN)
20428 		ti->tcpi_options |= TCPI_OPT_TFO;
20429 	/* still kept in ticks is t_rcvtime */
20430 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
20431 	/* Since we hold everything in precise useconds this is easy */
20432 	ti->tcpi_rtt = tp->t_srtt;
20433 	ti->tcpi_rttvar = tp->t_rttvar;
20434 	ti->tcpi_rto = tp->t_rxtcur;
20435 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
20436 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
20437 	/*
20438 	 * FreeBSD-specific extension fields for tcp_info.
20439 	 */
20440 	ti->tcpi_rcv_space = tp->rcv_wnd;
20441 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
20442 	ti->tcpi_snd_wnd = tp->snd_wnd;
20443 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
20444 	ti->tcpi_snd_nxt = tp->snd_nxt;
20445 	ti->tcpi_snd_mss = tp->t_maxseg;
20446 	ti->tcpi_rcv_mss = tp->t_maxseg;
20447 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
20448 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
20449 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
20450 #ifdef NETFLIX_STATS
20451 	ti->tcpi_total_tlp = tp->t_sndtlppack;
20452 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
20453 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
20454 #endif
20455 #ifdef TCP_OFFLOAD
20456 	if (tp->t_flags & TF_TOE) {
20457 		ti->tcpi_options |= TCPI_OPT_TOE;
20458 		tcp_offload_tcp_info(tp, ti);
20459 	}
20460 #endif
20461 }
20462 
20463 static int
20464 rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt)
20465 {
20466 	struct tcpcb *tp;
20467 	struct tcp_rack *rack;
20468 	int32_t error, optval;
20469 	uint64_t val, loptval;
20470 	struct	tcp_info ti;
20471 	/*
20472 	 * Because all our options are either boolean or an int, we can just
20473 	 * pull everything into optval and then unlock and copy. If we ever
20474 	 * add a option that is not a int, then this will have quite an
20475 	 * impact to this routine.
20476 	 */
20477 	error = 0;
20478 	tp = intotcpcb(inp);
20479 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20480 	if (rack == NULL) {
20481 		INP_WUNLOCK(inp);
20482 		return (EINVAL);
20483 	}
20484 	switch (sopt->sopt_name) {
20485 	case TCP_INFO:
20486 		/* First get the info filled */
20487 		rack_fill_info(tp, &ti);
20488 		/* Fix up the rtt related fields if needed */
20489 		INP_WUNLOCK(inp);
20490 		error = sooptcopyout(sopt, &ti, sizeof ti);
20491 		return (error);
20492 	/*
20493 	 * Beta is the congestion control value for NewReno that influences how
20494 	 * much of a backoff happens when loss is detected. It is normally set
20495 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
20496 	 * when you exit recovery.
20497 	 */
20498 	case TCP_RACK_PACING_BETA:
20499 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
20500 			error = EINVAL;
20501 		else if (rack->rc_pacing_cc_set == 0)
20502 			optval = rack->r_ctl.rc_saved_beta.beta;
20503 		else {
20504 			/*
20505 			 * Reach out into the CC data and report back what
20506 			 * I have previously set. Yeah it looks hackish but
20507 			 * we don't want to report the saved values.
20508 			 */
20509 			if (tp->t_ccv.cc_data)
20510 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta;
20511 			else
20512 				error = EINVAL;
20513 		}
20514 		break;
20515 		/*
20516 		 * Beta_ecn is the congestion control value for NewReno that influences how
20517 		 * much of a backoff happens when a ECN mark is detected. It is normally set
20518 		 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
20519 		 * you exit recovery. Note that classic ECN has a beta of 50, it is only
20520 		 * ABE Ecn that uses this "less" value, but we do too with pacing :)
20521 		 */
20522 
20523 	case TCP_RACK_PACING_BETA_ECN:
20524 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
20525 			error = EINVAL;
20526 		else if (rack->rc_pacing_cc_set == 0)
20527 			optval = rack->r_ctl.rc_saved_beta.beta_ecn;
20528 		else {
20529 			/*
20530 			 * Reach out into the CC data and report back what
20531 			 * I have previously set. Yeah it looks hackish but
20532 			 * we don't want to report the saved values.
20533 			 */
20534 			if (tp->t_ccv.cc_data)
20535 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
20536 			else
20537 				error = EINVAL;
20538 		}
20539 		break;
20540 	case TCP_RACK_DSACK_OPT:
20541 		optval = 0;
20542 		if (rack->rc_rack_tmr_std_based) {
20543 			optval |= 1;
20544 		}
20545 		if (rack->rc_rack_use_dsack) {
20546 			optval |= 2;
20547 		}
20548 		break;
20549  	case TCP_RACK_ENABLE_HYSTART:
20550 	{
20551 		if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
20552 			optval = RACK_HYSTART_ON;
20553 			if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
20554 				optval = RACK_HYSTART_ON_W_SC;
20555 			if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
20556 				optval = RACK_HYSTART_ON_W_SC_C;
20557 		} else {
20558 			optval = RACK_HYSTART_OFF;
20559 		}
20560 	}
20561 	break;
20562 	case TCP_FAST_RSM_HACK:
20563 		optval = rack->fast_rsm_hack;
20564 		break;
20565 	case TCP_DEFER_OPTIONS:
20566 		optval = rack->defer_options;
20567 		break;
20568 	case TCP_RACK_MEASURE_CNT:
20569 		optval = rack->r_ctl.req_measurements;
20570 		break;
20571 	case TCP_REC_ABC_VAL:
20572 		optval = rack->r_use_labc_for_rec;
20573 		break;
20574 	case TCP_RACK_ABC_VAL:
20575 		optval = rack->rc_labc;
20576 		break;
20577 	case TCP_HDWR_UP_ONLY:
20578 		optval= rack->r_up_only;
20579 		break;
20580 	case TCP_PACING_RATE_CAP:
20581 		loptval = rack->r_ctl.bw_rate_cap;
20582 		break;
20583 	case TCP_RACK_PROFILE:
20584 		/* You cannot retrieve a profile, its write only */
20585 		error = EINVAL;
20586 		break;
20587 	case TCP_USE_CMP_ACKS:
20588 		optval = rack->r_use_cmp_ack;
20589 		break;
20590 	case TCP_RACK_PACE_TO_FILL:
20591 		optval = rack->rc_pace_to_cwnd;
20592 		if (optval && rack->r_fill_less_agg)
20593 			optval++;
20594 		break;
20595 	case TCP_RACK_NO_PUSH_AT_MAX:
20596 		optval = rack->r_ctl.rc_no_push_at_mrtt;
20597 		break;
20598 	case TCP_SHARED_CWND_ENABLE:
20599 		optval = rack->rack_enable_scwnd;
20600 		break;
20601 	case TCP_RACK_NONRXT_CFG_RATE:
20602 		optval = rack->rack_rec_nonrxt_use_cr;
20603 		break;
20604 	case TCP_NO_PRR:
20605 		if (rack->rack_no_prr  == 1)
20606 			optval = 1;
20607 		else if (rack->no_prr_addback == 1)
20608 			optval = 2;
20609 		else
20610 			optval = 0;
20611 		break;
20612 	case TCP_RACK_DO_DETECTION:
20613 		optval = rack->do_detection;
20614 		break;
20615 	case TCP_RACK_MBUF_QUEUE:
20616 		/* Now do we use the LRO mbuf-queue feature */
20617 		optval = rack->r_mbuf_queue;
20618 		break;
20619 	case TCP_TIMELY_DYN_ADJ:
20620 		optval = rack->rc_gp_dyn_mul;
20621 		break;
20622 	case TCP_BBR_IWINTSO:
20623 		optval = rack->rc_init_win;
20624 		break;
20625 	case TCP_RACK_TLP_REDUCE:
20626 		/* RACK TLP cwnd reduction (bool) */
20627 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
20628 		break;
20629 	case TCP_BBR_RACK_INIT_RATE:
20630 		val = rack->r_ctl.init_rate;
20631 		/* convert to kbits per sec */
20632 		val *= 8;
20633 		val /= 1000;
20634 		optval = (uint32_t)val;
20635 		break;
20636 	case TCP_RACK_FORCE_MSEG:
20637 		optval = rack->rc_force_max_seg;
20638 		break;
20639 	case TCP_RACK_PACE_MAX_SEG:
20640 		/* Max segments in a pace */
20641 		optval = rack->rc_user_set_max_segs;
20642 		break;
20643 	case TCP_RACK_PACE_ALWAYS:
20644 		/* Use the always pace method */
20645 		optval = rack->rc_always_pace;
20646 		break;
20647 	case TCP_RACK_PRR_SENDALOT:
20648 		/* Allow PRR to send more than one seg */
20649 		optval = rack->r_ctl.rc_prr_sendalot;
20650 		break;
20651 	case TCP_RACK_MIN_TO:
20652 		/* Minimum time between rack t-o's in ms */
20653 		optval = rack->r_ctl.rc_min_to;
20654 		break;
20655 	case TCP_RACK_EARLY_SEG:
20656 		/* If early recovery max segments */
20657 		optval = rack->r_ctl.rc_early_recovery_segs;
20658 		break;
20659 	case TCP_RACK_REORD_THRESH:
20660 		/* RACK reorder threshold (shift amount) */
20661 		optval = rack->r_ctl.rc_reorder_shift;
20662 		break;
20663 	case TCP_RACK_REORD_FADE:
20664 		/* Does reordering fade after ms time */
20665 		optval = rack->r_ctl.rc_reorder_fade;
20666 		break;
20667 	case TCP_BBR_USE_RACK_RR:
20668 		/* Do we use the rack cheat for rxt */
20669 		optval = rack->use_rack_rr;
20670 		break;
20671 	case TCP_RACK_RR_CONF:
20672 		optval = rack->r_rr_config;
20673 		break;
20674 	case TCP_HDWR_RATE_CAP:
20675 		optval = rack->r_rack_hw_rate_caps;
20676 		break;
20677 	case TCP_BBR_HDWR_PACE:
20678 		optval = rack->rack_hdw_pace_ena;
20679 		break;
20680 	case TCP_RACK_TLP_THRESH:
20681 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
20682 		optval = rack->r_ctl.rc_tlp_threshold;
20683 		break;
20684 	case TCP_RACK_PKT_DELAY:
20685 		/* RACK added ms i.e. rack-rtt + reord + N */
20686 		optval = rack->r_ctl.rc_pkt_delay;
20687 		break;
20688 	case TCP_RACK_TLP_USE:
20689 		optval = rack->rack_tlp_threshold_use;
20690 		break;
20691 	case TCP_RACK_PACE_RATE_CA:
20692 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
20693 		break;
20694 	case TCP_RACK_PACE_RATE_SS:
20695 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
20696 		break;
20697 	case TCP_RACK_PACE_RATE_REC:
20698 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
20699 		break;
20700 	case TCP_RACK_GP_INCREASE_SS:
20701 		optval = rack->r_ctl.rack_per_of_gp_ca;
20702 		break;
20703 	case TCP_RACK_GP_INCREASE_CA:
20704 		optval = rack->r_ctl.rack_per_of_gp_ss;
20705 		break;
20706 	case TCP_BBR_RACK_RTT_USE:
20707 		optval = rack->r_ctl.rc_rate_sample_method;
20708 		break;
20709 	case TCP_DELACK:
20710 		optval = tp->t_delayed_ack;
20711 		break;
20712 	case TCP_DATA_AFTER_CLOSE:
20713 		optval = rack->rc_allow_data_af_clo;
20714 		break;
20715 	case TCP_SHARED_CWND_TIME_LIMIT:
20716 		optval = rack->r_limit_scw;
20717 		break;
20718 	case TCP_RACK_TIMER_SLOP:
20719 		optval = rack->r_ctl.timer_slop;
20720 		break;
20721 	default:
20722 		return (tcp_default_ctloutput(inp, sopt));
20723 		break;
20724 	}
20725 	INP_WUNLOCK(inp);
20726 	if (error == 0) {
20727 		if (TCP_PACING_RATE_CAP)
20728 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
20729 		else
20730 			error = sooptcopyout(sopt, &optval, sizeof optval);
20731 	}
20732 	return (error);
20733 }
20734 
20735 static int
20736 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt)
20737 {
20738 	if (sopt->sopt_dir == SOPT_SET) {
20739 		return (rack_set_sockopt(inp, sopt));
20740 	} else if (sopt->sopt_dir == SOPT_GET) {
20741 		return (rack_get_sockopt(inp, sopt));
20742 	} else {
20743 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
20744 	}
20745 }
20746 
20747 static const char *rack_stack_names[] = {
20748 	__XSTRING(STACKNAME),
20749 #ifdef STACKALIAS
20750 	__XSTRING(STACKALIAS),
20751 #endif
20752 };
20753 
20754 static int
20755 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
20756 {
20757 	memset(mem, 0, size);
20758 	return (0);
20759 }
20760 
20761 static void
20762 rack_dtor(void *mem, int32_t size, void *arg)
20763 {
20764 
20765 }
20766 
20767 static bool rack_mod_inited = false;
20768 
20769 static int
20770 tcp_addrack(module_t mod, int32_t type, void *data)
20771 {
20772 	int32_t err = 0;
20773 	int num_stacks;
20774 
20775 	switch (type) {
20776 	case MOD_LOAD:
20777 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
20778 		    sizeof(struct rack_sendmap),
20779 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
20780 
20781 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
20782 		    sizeof(struct tcp_rack),
20783 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
20784 
20785 		sysctl_ctx_init(&rack_sysctl_ctx);
20786 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
20787 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
20788 		    OID_AUTO,
20789 #ifdef STACKALIAS
20790 		    __XSTRING(STACKALIAS),
20791 #else
20792 		    __XSTRING(STACKNAME),
20793 #endif
20794 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
20795 		    "");
20796 		if (rack_sysctl_root == NULL) {
20797 			printf("Failed to add sysctl node\n");
20798 			err = EFAULT;
20799 			goto free_uma;
20800 		}
20801 		rack_init_sysctls();
20802 		num_stacks = nitems(rack_stack_names);
20803 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
20804 		    rack_stack_names, &num_stacks);
20805 		if (err) {
20806 			printf("Failed to register %s stack name for "
20807 			    "%s module\n", rack_stack_names[num_stacks],
20808 			    __XSTRING(MODNAME));
20809 			sysctl_ctx_free(&rack_sysctl_ctx);
20810 free_uma:
20811 			uma_zdestroy(rack_zone);
20812 			uma_zdestroy(rack_pcb_zone);
20813 			rack_counter_destroy();
20814 			printf("Failed to register rack module -- err:%d\n", err);
20815 			return (err);
20816 		}
20817 		tcp_lro_reg_mbufq();
20818 		rack_mod_inited = true;
20819 		break;
20820 	case MOD_QUIESCE:
20821 		err = deregister_tcp_functions(&__tcp_rack, true, false);
20822 		break;
20823 	case MOD_UNLOAD:
20824 		err = deregister_tcp_functions(&__tcp_rack, false, true);
20825 		if (err == EBUSY)
20826 			break;
20827 		if (rack_mod_inited) {
20828 			uma_zdestroy(rack_zone);
20829 			uma_zdestroy(rack_pcb_zone);
20830 			sysctl_ctx_free(&rack_sysctl_ctx);
20831 			rack_counter_destroy();
20832 			rack_mod_inited = false;
20833 		}
20834 		tcp_lro_dereg_mbufq();
20835 		err = 0;
20836 		break;
20837 	default:
20838 		return (EOPNOTSUPP);
20839 	}
20840 	return (err);
20841 }
20842 
20843 static moduledata_t tcp_rack = {
20844 	.name = __XSTRING(MODNAME),
20845 	.evhand = tcp_addrack,
20846 	.priv = 0
20847 };
20848 
20849 MODULE_VERSION(MODNAME, 1);
20850 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
20851 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
20852 
20853 #endif /* #if !defined(INET) && !defined(INET6) */
20854