xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision d5b0e70f7e04d971691517ce1304d86a1e367e2e) 
1 /*-
2  * Copyright (c) 2016-2020 Netflix, Inc.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. Redistributions in binary form must reproduce the above copyright
10  *    notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  */
26 
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
29 
30 #include "opt_inet.h"
31 #include "opt_inet6.h"
32 #include "opt_ipsec.h"
33 #include "opt_tcpdebug.h"
34 #include "opt_ratelimit.h"
35 #include "opt_kern_tls.h"
36 #include <sys/param.h>
37 #include <sys/arb.h>
38 #include <sys/module.h>
39 #include <sys/kernel.h>
40 #ifdef TCP_HHOOK
41 #include <sys/hhook.h>
42 #endif
43 #include <sys/lock.h>
44 #include <sys/malloc.h>
45 #include <sys/lock.h>
46 #include <sys/mutex.h>
47 #include <sys/mbuf.h>
48 #include <sys/proc.h>		/* for proc0 declaration */
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 #include <sys/systm.h>
53 #ifdef STATS
54 #include <sys/qmath.h>
55 #include <sys/tree.h>
56 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
57 #else
58 #include <sys/tree.h>
59 #endif
60 #include <sys/refcount.h>
61 #include <sys/queue.h>
62 #include <sys/tim_filter.h>
63 #include <sys/smp.h>
64 #include <sys/kthread.h>
65 #include <sys/kern_prefetch.h>
66 #include <sys/protosw.h>
67 #ifdef TCP_ACCOUNTING
68 #include <sys/sched.h>
69 #include <machine/cpu.h>
70 #endif
71 #include <vm/uma.h>
72 
73 #include <net/route.h>
74 #include <net/route/nhop.h>
75 #include <net/vnet.h>
76 
77 #define TCPSTATES		/* for logging */
78 
79 #include <netinet/in.h>
80 #include <netinet/in_kdtrace.h>
81 #include <netinet/in_pcb.h>
82 #include <netinet/ip.h>
83 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
84 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
85 #include <netinet/ip_var.h>
86 #include <netinet/ip6.h>
87 #include <netinet6/in6_pcb.h>
88 #include <netinet6/ip6_var.h>
89 #include <netinet/tcp.h>
90 #define	TCPOUTFLAGS
91 #include <netinet/tcp_fsm.h>
92 #include <netinet/tcp_log_buf.h>
93 #include <netinet/tcp_seq.h>
94 #include <netinet/tcp_timer.h>
95 #include <netinet/tcp_var.h>
96 #include <netinet/tcp_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 TCPDEBUG
109 #include <netinet/tcp_debug.h>
110 #endif				/* TCPDEBUG */
111 #ifdef TCP_OFFLOAD
112 #include <netinet/tcp_offload.h>
113 #endif
114 #ifdef INET6
115 #include <netinet6/tcp6_var.h>
116 #endif
117 #include <netinet/tcp_ecn.h>
118 
119 #include <netipsec/ipsec_support.h>
120 
121 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
122 #include <netipsec/ipsec.h>
123 #include <netipsec/ipsec6.h>
124 #endif				/* IPSEC */
125 
126 #include <netinet/udp.h>
127 #include <netinet/udp_var.h>
128 #include <machine/in_cksum.h>
129 
130 #ifdef MAC
131 #include <security/mac/mac_framework.h>
132 #endif
133 #include "sack_filter.h"
134 #include "tcp_rack.h"
135 #include "rack_bbr_common.h"
136 
137 uma_zone_t rack_zone;
138 uma_zone_t rack_pcb_zone;
139 
140 #ifndef TICKS2SBT
141 #define	TICKS2SBT(__t)	(tick_sbt * ((sbintime_t)(__t)))
142 #endif
143 
144 VNET_DECLARE(uint32_t, newreno_beta);
145 VNET_DECLARE(uint32_t, newreno_beta_ecn);
146 #define V_newreno_beta VNET(newreno_beta)
147 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
148 
149 
150 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
151 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
152 
153 struct sysctl_ctx_list rack_sysctl_ctx;
154 struct sysctl_oid *rack_sysctl_root;
155 
156 #define CUM_ACKED 1
157 #define SACKED 2
158 
159 /*
160  * The RACK module incorporates a number of
161  * TCP ideas that have been put out into the IETF
162  * over the last few years:
163  * - Matt Mathis's Rate Halving which slowly drops
164  *    the congestion window so that the ack clock can
165  *    be maintained during a recovery.
166  * - Yuchung Cheng's RACK TCP (for which its named) that
167  *    will stop us using the number of dup acks and instead
168  *    use time as the gage of when we retransmit.
169  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
170  *    of Dukkipati et.al.
171  * RACK depends on SACK, so if an endpoint arrives that
172  * cannot do SACK the state machine below will shuttle the
173  * connection back to using the "default" TCP stack that is
174  * in FreeBSD.
175  *
176  * To implement RACK the original TCP stack was first decomposed
177  * into a functional state machine with individual states
178  * for each of the possible TCP connection states. The do_segment
179  * functions role in life is to mandate the connection supports SACK
180  * initially and then assure that the RACK state matches the conenction
181  * state before calling the states do_segment function. Each
182  * state is simplified due to the fact that the original do_segment
183  * has been decomposed and we *know* what state we are in (no
184  * switches on the state) and all tests for SACK are gone. This
185  * greatly simplifies what each state does.
186  *
187  * TCP output is also over-written with a new version since it
188  * must maintain the new rack scoreboard.
189  *
190  */
191 static int32_t rack_tlp_thresh = 1;
192 static int32_t rack_tlp_limit = 2;	/* No more than 2 TLPs w-out new data */
193 static int32_t rack_tlp_use_greater = 1;
194 static int32_t rack_reorder_thresh = 2;
195 static int32_t rack_reorder_fade = 60000000;	/* 0 - never fade, def 60,000,000
196 						 * - 60 seconds */
197 static uint8_t rack_req_measurements = 1;
198 /* Attack threshold detections */
199 static uint32_t rack_highest_sack_thresh_seen = 0;
200 static uint32_t rack_highest_move_thresh_seen = 0;
201 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
202 static int32_t rack_hw_pace_extra_slots = 2;	/* 2 extra MSS time betweens */
203 static int32_t rack_hw_rate_caps = 1; /* 1; */
204 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
205 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
206 static int32_t rack_hw_up_only = 1;
207 static int32_t rack_stats_gets_ms_rtt = 1;
208 static int32_t rack_prr_addbackmax = 2;
209 static int32_t rack_do_hystart = 0;
210 static int32_t rack_apply_rtt_with_reduced_conf = 0;
211 
212 static int32_t rack_pkt_delay = 1000;
213 static int32_t rack_send_a_lot_in_prr = 1;
214 static int32_t rack_min_to = 1000;	/* Number of microsecond  min timeout */
215 static int32_t rack_verbose_logging = 0;
216 static int32_t rack_ignore_data_after_close = 1;
217 static int32_t rack_enable_shared_cwnd = 1;
218 static int32_t rack_use_cmp_acks = 1;
219 static int32_t rack_use_fsb = 1;
220 static int32_t rack_use_rfo = 1;
221 static int32_t rack_use_rsm_rfo = 1;
222 static int32_t rack_max_abc_post_recovery = 2;
223 static int32_t rack_client_low_buf = 0;
224 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 */
225 #ifdef TCP_ACCOUNTING
226 static int32_t rack_tcp_accounting = 0;
227 #endif
228 static int32_t rack_limits_scwnd = 1;
229 static int32_t rack_enable_mqueue_for_nonpaced = 0;
230 static int32_t rack_disable_prr = 0;
231 static int32_t use_rack_rr = 1;
232 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
233 static int32_t rack_persist_min = 250000;	/* 250usec */
234 static int32_t rack_persist_max = 2000000;	/* 2 Second in usec's */
235 static int32_t rack_sack_not_required = 1;	/* set to one to allow non-sack to use rack */
236 static int32_t rack_default_init_window = 0;	/* Use system default */
237 static int32_t rack_limit_time_with_srtt = 0;
238 static int32_t rack_autosndbuf_inc = 20;	/* In percentage form */
239 static int32_t rack_enobuf_hw_boost_mult = 2;	/* How many times the hw rate we boost slot using time_between */
240 static int32_t rack_enobuf_hw_max = 12000;	/* 12 ms in usecs */
241 static int32_t rack_enobuf_hw_min = 10000;	/* 10 ms in usecs */
242 static int32_t rack_hw_rwnd_factor = 2;		/* How many max_segs the rwnd must be before we hold off sending */
243 
244 /*
245  * Currently regular tcp has a rto_min of 30ms
246  * the backoff goes 12 times so that ends up
247  * being a total of 122.850 seconds before a
248  * connection is killed.
249  */
250 static uint32_t rack_def_data_window = 20;
251 static uint32_t rack_goal_bdp = 2;
252 static uint32_t rack_min_srtts = 1;
253 static uint32_t rack_min_measure_usec = 0;
254 static int32_t rack_tlp_min = 10000;	/* 10ms */
255 static int32_t rack_rto_min = 30000;	/* 30,000 usec same as main freebsd */
256 static int32_t rack_rto_max = 4000000;	/* 4 seconds in usec's */
257 static const int32_t rack_free_cache = 2;
258 static int32_t rack_hptsi_segments = 40;
259 static int32_t rack_rate_sample_method = USE_RTT_LOW;
260 static int32_t rack_pace_every_seg = 0;
261 static int32_t rack_delayed_ack_time = 40000;	/* 40ms in usecs */
262 static int32_t rack_slot_reduction = 4;
263 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
264 static int32_t rack_cwnd_block_ends_measure = 0;
265 static int32_t rack_rwnd_block_ends_measure = 0;
266 static int32_t rack_def_profile = 0;
267 
268 static int32_t rack_lower_cwnd_at_tlp = 0;
269 static int32_t rack_limited_retran = 0;
270 static int32_t rack_always_send_oldest = 0;
271 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
272 
273 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
274 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
275 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
276 
277 /* Probertt */
278 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
279 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
280 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
281 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
282 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
283 
284 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
285 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
286 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
287 static uint32_t rack_probertt_use_min_rtt_exit = 0;
288 static uint32_t rack_probe_rtt_sets_cwnd = 0;
289 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
290 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in usecs */
291 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
292 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction */
293 static uint32_t rack_min_probertt_hold = 40000;		/* Equal to delayed ack time */
294 static uint32_t rack_probertt_filter_life = 10000000;
295 static uint32_t rack_probertt_lower_within = 10;
296 static uint32_t rack_min_rtt_movement = 250000;	/* Must move at least 250ms (in microseconds)  to count as a lowering */
297 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
298 static int32_t rack_probertt_clear_is = 1;
299 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
300 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
301 
302 /* Part of pacing */
303 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
304 
305 /* Timely information */
306 /* Combine these two gives the range of 'no change' to bw */
307 /* ie the up/down provide the upper and lower bound */
308 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
309 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
310 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
311 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
312 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
313 static int32_t rack_gp_decrease_per = 20;	/* 20% decrease in multiplier */
314 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multiplier */
315 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
316 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
317 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
318 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
319 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
320 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
321 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
322 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
323 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
324 static int32_t rack_use_max_for_nobackoff = 0;
325 static int32_t rack_timely_int_timely_only = 0;	/* do interim timely's only use the timely algo (no b/w changes)? */
326 static int32_t rack_timely_no_stopping = 0;
327 static int32_t rack_down_raise_thresh = 100;
328 static int32_t rack_req_segs = 1;
329 static uint64_t rack_bw_rate_cap = 0;
330 static uint32_t rack_trace_point_config = 0;
331 static uint32_t rack_trace_point_bb_mode = 4;
332 static int32_t rack_trace_point_count = 0;
333 
334 
335 /* Weird delayed ack mode */
336 static int32_t rack_use_imac_dack = 0;
337 /* Rack specific counters */
338 counter_u64_t rack_saw_enobuf;
339 counter_u64_t rack_saw_enobuf_hw;
340 counter_u64_t rack_saw_enetunreach;
341 counter_u64_t rack_persists_sends;
342 counter_u64_t rack_persists_acks;
343 counter_u64_t rack_persists_loss;
344 counter_u64_t rack_persists_lost_ends;
345 #ifdef INVARIANTS
346 counter_u64_t rack_adjust_map_bw;
347 #endif
348 /* Tail loss probe counters */
349 counter_u64_t rack_tlp_tot;
350 counter_u64_t rack_tlp_newdata;
351 counter_u64_t rack_tlp_retran;
352 counter_u64_t rack_tlp_retran_bytes;
353 counter_u64_t rack_to_tot;
354 counter_u64_t rack_hot_alloc;
355 counter_u64_t rack_to_alloc;
356 counter_u64_t rack_to_alloc_hard;
357 counter_u64_t rack_to_alloc_emerg;
358 counter_u64_t rack_to_alloc_limited;
359 counter_u64_t rack_alloc_limited_conns;
360 counter_u64_t rack_split_limited;
361 
362 counter_u64_t rack_multi_single_eq;
363 counter_u64_t rack_proc_non_comp_ack;
364 
365 counter_u64_t rack_fto_send;
366 counter_u64_t rack_fto_rsm_send;
367 counter_u64_t rack_nfto_resend;
368 counter_u64_t rack_non_fto_send;
369 counter_u64_t rack_extended_rfo;
370 
371 counter_u64_t rack_sack_proc_all;
372 counter_u64_t rack_sack_proc_short;
373 counter_u64_t rack_sack_proc_restart;
374 counter_u64_t rack_sack_attacks_detected;
375 counter_u64_t rack_sack_attacks_reversed;
376 counter_u64_t rack_sack_used_next_merge;
377 counter_u64_t rack_sack_splits;
378 counter_u64_t rack_sack_used_prev_merge;
379 counter_u64_t rack_sack_skipped_acked;
380 counter_u64_t rack_ack_total;
381 counter_u64_t rack_express_sack;
382 counter_u64_t rack_sack_total;
383 counter_u64_t rack_move_none;
384 counter_u64_t rack_move_some;
385 
386 counter_u64_t rack_input_idle_reduces;
387 counter_u64_t rack_collapsed_win;
388 counter_u64_t rack_collapsed_win_seen;
389 counter_u64_t rack_collapsed_win_rxt;
390 counter_u64_t rack_collapsed_win_rxt_bytes;
391 counter_u64_t rack_try_scwnd;
392 counter_u64_t rack_hw_pace_init_fail;
393 counter_u64_t rack_hw_pace_lost;
394 
395 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
396 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
397 
398 
399 #define	RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
400 
401 #define	RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {	\
402 	(tv) = (value) + slop;	 \
403 	if ((u_long)(tv) < (u_long)(tvmin)) \
404 		(tv) = (tvmin); \
405 	if ((u_long)(tv) > (u_long)(tvmax)) \
406 		(tv) = (tvmax); \
407 } while (0)
408 
409 static void
410 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
411 
412 static int
413 rack_process_ack(struct mbuf *m, struct tcphdr *th,
414     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
415     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
416 static int
417 rack_process_data(struct mbuf *m, struct tcphdr *th,
418     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
419     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
420 static void
421 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
422    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
423 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
424 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
425     uint8_t limit_type);
426 static struct rack_sendmap *
427 rack_check_recovery_mode(struct tcpcb *tp,
428     uint32_t tsused);
429 static void
430 rack_cong_signal(struct tcpcb *tp,
431 		 uint32_t type, uint32_t ack, int );
432 static void rack_counter_destroy(void);
433 static int
434 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt);
435 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
436 static void
437 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
438 static void
439 rack_do_segment(struct mbuf *m, struct tcphdr *th,
440     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
441     uint8_t iptos);
442 static void rack_dtor(void *mem, int32_t size, void *arg);
443 static void
444 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
445     uint32_t flex1, uint32_t flex2,
446     uint32_t flex3, uint32_t flex4,
447     uint32_t flex5, uint32_t flex6,
448     uint16_t flex7, uint8_t mod);
449 
450 static void
451 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
452    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
453    struct rack_sendmap *rsm, uint8_t quality);
454 static struct rack_sendmap *
455 rack_find_high_nonack(struct tcp_rack *rack,
456     struct rack_sendmap *rsm);
457 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
458 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
459 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
460 static int rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt);
461 static void
462 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
463 			    tcp_seq th_ack, int line, uint8_t quality);
464 static uint32_t
465 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
466 static int32_t rack_handoff_ok(struct tcpcb *tp);
467 static int32_t rack_init(struct tcpcb *tp);
468 static void rack_init_sysctls(void);
469 static void
470 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
471     struct tcphdr *th, int entered_rec, int dup_ack_struck);
472 static void
473 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
474     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
475     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls);
476 
477 static void
478 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
479     struct rack_sendmap *rsm);
480 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
481 static int32_t rack_output(struct tcpcb *tp);
482 
483 static uint32_t
484 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
485     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
486     uint32_t cts, int *moved_two);
487 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
488 static void rack_remxt_tmr(struct tcpcb *tp);
489 static int rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt);
490 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
491 static int32_t rack_stopall(struct tcpcb *tp);
492 static void
493 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type,
494     uint32_t delta);
495 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type);
496 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
497 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type);
498 static uint32_t
499 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
500     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag);
501 static void
502 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
503     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag);
504 static int
505 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
506     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
507 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
508 static int
509 rack_do_close_wait(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_closing(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 thflags, int32_t nxt_pkt, uint8_t iptos);
516 static int
517 rack_do_established(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_fastnewdata(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 nxt_pkt, uint8_t iptos);
524 static int
525 rack_do_fin_wait_1(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_fin_wait_2(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_lastack(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 static int
537 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
538     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
539     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
540 static int
541 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
542     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
543     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
544 struct rack_sendmap *
545 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
546     uint32_t tsused);
547 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
548     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
549 static void
550      tcp_rack_partialack(struct tcpcb *tp);
551 static int
552 rack_set_profile(struct tcp_rack *rack, int prof);
553 static void
554 rack_apply_deferred_options(struct tcp_rack *rack);
555 
556 int32_t rack_clear_counter=0;
557 
558 static inline void
559 rack_trace_point(struct tcp_rack *rack, int num)
560 {
561 	if (((rack_trace_point_config == num)  ||
562 	     (rack_trace_point_config = 0xffffffff)) &&
563 	    (rack_trace_point_bb_mode != 0) &&
564 	    (rack_trace_point_count > 0) &&
565 	    (rack->rc_tp->t_logstate == 0)) {
566 		int res;
567 		res = atomic_fetchadd_int(&rack_trace_point_count, -1);
568 		if (res > 0) {
569 			rack->rc_tp->t_logstate = rack_trace_point_bb_mode;
570 		} else {
571 			/* Loss a race assure its zero now */
572 			rack_trace_point_count = 0;
573 		}
574 	}
575 }
576 
577 static void
578 rack_set_cc_pacing(struct tcp_rack *rack)
579 {
580 	struct sockopt sopt;
581 	struct cc_newreno_opts opt;
582 	struct newreno old, *ptr;
583 	struct tcpcb *tp;
584 	int error;
585 
586 	if (rack->rc_pacing_cc_set)
587 		return;
588 
589 	tp = rack->rc_tp;
590 	if (tp->cc_algo == NULL) {
591 		/* Tcb is leaving */
592 		return;
593 	}
594 	rack->rc_pacing_cc_set = 1;
595 	if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
596 		/* Not new-reno we can't play games with beta! */
597 		goto out;
598 	}
599 	ptr = ((struct newreno *)tp->ccv->cc_data);
600 	if (CC_ALGO(tp)->ctl_output == NULL)  {
601 		/* Huh, why does new_reno no longer have a set function? */
602 		goto out;
603 	}
604 	if (ptr == NULL) {
605 		/* Just the default values */
606 		old.beta = V_newreno_beta_ecn;
607 		old.beta_ecn = V_newreno_beta_ecn;
608 		old.newreno_flags = 0;
609 	} else {
610 		old.beta = ptr->beta;
611 		old.beta_ecn = ptr->beta_ecn;
612 		old.newreno_flags = ptr->newreno_flags;
613 	}
614 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
615 	sopt.sopt_dir = SOPT_SET;
616 	opt.name = CC_NEWRENO_BETA;
617 	opt.val = rack->r_ctl.rc_saved_beta.beta;
618 	error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
619 	if (error)  {
620 		goto out;
621 	}
622 	/*
623 	 * Hack alert we need to set in our newreno_flags
624 	 * so that Abe behavior is also applied.
625 	 */
626 	((struct newreno *)tp->ccv->cc_data)->newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
627 	opt.name = CC_NEWRENO_BETA_ECN;
628 	opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
629 	error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
630 	if (error) {
631 		goto out;
632 	}
633 	/* Save off the original values for restoral */
634 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
635 out:
636 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
637 		union tcp_log_stackspecific log;
638 		struct timeval tv;
639 
640 		ptr = ((struct newreno *)tp->ccv->cc_data);
641 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
642 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
643 		if (ptr) {
644 			log.u_bbr.flex1 = ptr->beta;
645 			log.u_bbr.flex2 = ptr->beta_ecn;
646 			log.u_bbr.flex3 = ptr->newreno_flags;
647 		}
648 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
649 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
650 		log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
651 		log.u_bbr.flex7 = rack->gp_ready;
652 		log.u_bbr.flex7 <<= 1;
653 		log.u_bbr.flex7 |= rack->use_fixed_rate;
654 		log.u_bbr.flex7 <<= 1;
655 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
656 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
657 		log.u_bbr.flex8 = 3;
658 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
659 			       0, &log, false, NULL, NULL, 0, &tv);
660 	}
661 }
662 
663 static void
664 rack_undo_cc_pacing(struct tcp_rack *rack)
665 {
666 	struct newreno old, *ptr;
667 	struct tcpcb *tp;
668 
669 	if (rack->rc_pacing_cc_set == 0)
670 		return;
671 	tp = rack->rc_tp;
672 	rack->rc_pacing_cc_set = 0;
673 	if (tp->cc_algo == NULL)
674 		/* Tcb is leaving */
675 		return;
676 	if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
677 		/* Not new-reno nothing to do! */
678 		return;
679 	}
680 	ptr = ((struct newreno *)tp->ccv->cc_data);
681 	if (ptr == NULL) {
682 		/*
683 		 * This happens at rack_fini() if the
684 		 * cc module gets freed on us. In that
685 		 * case we loose our "new" settings but
686 		 * thats ok, since the tcb is going away anyway.
687 		 */
688 		return;
689 	}
690 	/* Grab out our set values */
691 	memcpy(&old, ptr, sizeof(struct newreno));
692 	/* Copy back in the original values */
693 	memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno));
694 	/* Now save back the values we had set in (for when pacing is restored) */
695 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
696 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
697 		union tcp_log_stackspecific log;
698 		struct timeval tv;
699 
700 		ptr = ((struct newreno *)tp->ccv->cc_data);
701 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
702 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
703 		log.u_bbr.flex1 = ptr->beta;
704 		log.u_bbr.flex2 = ptr->beta_ecn;
705 		log.u_bbr.flex3 = ptr->newreno_flags;
706 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
707 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
708 		log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags;
709 		log.u_bbr.flex7 = rack->gp_ready;
710 		log.u_bbr.flex7 <<= 1;
711 		log.u_bbr.flex7 |= rack->use_fixed_rate;
712 		log.u_bbr.flex7 <<= 1;
713 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
714 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
715 		log.u_bbr.flex8 = 4;
716 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
717 			       0, &log, false, NULL, NULL, 0, &tv);
718 	}
719 }
720 
721 #ifdef NETFLIX_PEAKRATE
722 static inline void
723 rack_update_peakrate_thr(struct tcpcb *tp)
724 {
725 	/* Keep in mind that t_maxpeakrate is in B/s. */
726 	uint64_t peak;
727 	peak = uqmax((tp->t_maxseg * 2),
728 		     (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC));
729 	tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX);
730 }
731 #endif
732 
733 static int
734 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
735 {
736 	uint32_t stat;
737 	int32_t error;
738 
739 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
740 	if (error || req->newptr == NULL)
741 		return error;
742 
743 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
744 	if (error)
745 		return (error);
746 	if (stat == 1) {
747 #ifdef INVARIANTS
748 		printf("Clearing RACK counters\n");
749 #endif
750 		counter_u64_zero(rack_tlp_tot);
751 		counter_u64_zero(rack_tlp_newdata);
752 		counter_u64_zero(rack_tlp_retran);
753 		counter_u64_zero(rack_tlp_retran_bytes);
754 		counter_u64_zero(rack_to_tot);
755 		counter_u64_zero(rack_saw_enobuf);
756 		counter_u64_zero(rack_saw_enobuf_hw);
757 		counter_u64_zero(rack_saw_enetunreach);
758 		counter_u64_zero(rack_persists_sends);
759 		counter_u64_zero(rack_persists_acks);
760 		counter_u64_zero(rack_persists_loss);
761 		counter_u64_zero(rack_persists_lost_ends);
762 #ifdef INVARIANTS
763 		counter_u64_zero(rack_adjust_map_bw);
764 #endif
765 		counter_u64_zero(rack_to_alloc_hard);
766 		counter_u64_zero(rack_to_alloc_emerg);
767 		counter_u64_zero(rack_sack_proc_all);
768 		counter_u64_zero(rack_fto_send);
769 		counter_u64_zero(rack_fto_rsm_send);
770 		counter_u64_zero(rack_extended_rfo);
771 		counter_u64_zero(rack_hw_pace_init_fail);
772 		counter_u64_zero(rack_hw_pace_lost);
773 		counter_u64_zero(rack_non_fto_send);
774 		counter_u64_zero(rack_nfto_resend);
775 		counter_u64_zero(rack_sack_proc_short);
776 		counter_u64_zero(rack_sack_proc_restart);
777 		counter_u64_zero(rack_to_alloc);
778 		counter_u64_zero(rack_to_alloc_limited);
779 		counter_u64_zero(rack_alloc_limited_conns);
780 		counter_u64_zero(rack_split_limited);
781 		counter_u64_zero(rack_multi_single_eq);
782 		counter_u64_zero(rack_proc_non_comp_ack);
783 		counter_u64_zero(rack_sack_attacks_detected);
784 		counter_u64_zero(rack_sack_attacks_reversed);
785 		counter_u64_zero(rack_sack_used_next_merge);
786 		counter_u64_zero(rack_sack_used_prev_merge);
787 		counter_u64_zero(rack_sack_splits);
788 		counter_u64_zero(rack_sack_skipped_acked);
789 		counter_u64_zero(rack_ack_total);
790 		counter_u64_zero(rack_express_sack);
791 		counter_u64_zero(rack_sack_total);
792 		counter_u64_zero(rack_move_none);
793 		counter_u64_zero(rack_move_some);
794 		counter_u64_zero(rack_try_scwnd);
795 		counter_u64_zero(rack_collapsed_win);
796 		counter_u64_zero(rack_collapsed_win_rxt);
797 		counter_u64_zero(rack_collapsed_win_seen);
798 		counter_u64_zero(rack_collapsed_win_rxt_bytes);
799 	}
800 	rack_clear_counter = 0;
801 	return (0);
802 }
803 
804 static void
805 rack_init_sysctls(void)
806 {
807 	struct sysctl_oid *rack_counters;
808 	struct sysctl_oid *rack_attack;
809 	struct sysctl_oid *rack_pacing;
810 	struct sysctl_oid *rack_timely;
811 	struct sysctl_oid *rack_timers;
812 	struct sysctl_oid *rack_tlp;
813 	struct sysctl_oid *rack_misc;
814 	struct sysctl_oid *rack_features;
815 	struct sysctl_oid *rack_measure;
816 	struct sysctl_oid *rack_probertt;
817 	struct sysctl_oid *rack_hw_pacing;
818 	struct sysctl_oid *rack_tracepoint;
819 
820 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
821 	    SYSCTL_CHILDREN(rack_sysctl_root),
822 	    OID_AUTO,
823 	    "sack_attack",
824 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
825 	    "Rack Sack Attack Counters and Controls");
826 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
827 	    SYSCTL_CHILDREN(rack_sysctl_root),
828 	    OID_AUTO,
829 	    "stats",
830 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
831 	    "Rack Counters");
832 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
833 	    SYSCTL_CHILDREN(rack_sysctl_root),
834 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
835 	    &rack_rate_sample_method , USE_RTT_LOW,
836 	    "What method should we use for rate sampling 0=high, 1=low ");
837 	/* Probe rtt related controls */
838 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
839 	    SYSCTL_CHILDREN(rack_sysctl_root),
840 	    OID_AUTO,
841 	    "probertt",
842 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
843 	    "ProbeRTT related Controls");
844 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
845 	    SYSCTL_CHILDREN(rack_probertt),
846 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
847 	    &rack_atexit_prtt_hbp, 130,
848 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
849 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
850 	    SYSCTL_CHILDREN(rack_probertt),
851 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
852 	    &rack_atexit_prtt, 130,
853 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
854 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
855 	    SYSCTL_CHILDREN(rack_probertt),
856 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
857 	    &rack_per_of_gp_probertt, 60,
858 	    "What percentage of goodput do we pace at in probertt");
859 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
860 	    SYSCTL_CHILDREN(rack_probertt),
861 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
862 	    &rack_per_of_gp_probertt_reduce, 10,
863 	    "What percentage of goodput do we reduce every gp_srtt");
864 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
865 	    SYSCTL_CHILDREN(rack_probertt),
866 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
867 	    &rack_per_of_gp_lowthresh, 40,
868 	    "What percentage of goodput do we allow the multiplier to fall to");
869 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
870 	    SYSCTL_CHILDREN(rack_probertt),
871 	    OID_AUTO, "time_between", CTLFLAG_RW,
872 	    & rack_time_between_probertt, 96000000,
873 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
874 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
875 	    SYSCTL_CHILDREN(rack_probertt),
876 	    OID_AUTO, "safety", CTLFLAG_RW,
877 	    &rack_probe_rtt_safety_val, 2000000,
878 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
879 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
880 	    SYSCTL_CHILDREN(rack_probertt),
881 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
882 	    &rack_probe_rtt_sets_cwnd, 0,
883 	    "Do we set the cwnd too (if always_lower is on)");
884 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
885 	    SYSCTL_CHILDREN(rack_probertt),
886 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
887 	    &rack_max_drain_wait, 2,
888 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
889 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
890 	    SYSCTL_CHILDREN(rack_probertt),
891 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
892 	    &rack_must_drain, 1,
893 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
894 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
895 	    SYSCTL_CHILDREN(rack_probertt),
896 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
897 	    &rack_probertt_use_min_rtt_entry, 1,
898 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
899 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
900 	    SYSCTL_CHILDREN(rack_probertt),
901 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
902 	    &rack_probertt_use_min_rtt_exit, 0,
903 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
904 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
905 	    SYSCTL_CHILDREN(rack_probertt),
906 	    OID_AUTO, "length_div", CTLFLAG_RW,
907 	    &rack_probertt_gpsrtt_cnt_div, 0,
908 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
909 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
910 	    SYSCTL_CHILDREN(rack_probertt),
911 	    OID_AUTO, "length_mul", CTLFLAG_RW,
912 	    &rack_probertt_gpsrtt_cnt_mul, 0,
913 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
914 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
915 	    SYSCTL_CHILDREN(rack_probertt),
916 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
917 	    &rack_min_probertt_hold, 200000,
918 	    "What is the minimum time we hold probertt at target");
919 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
920 	    SYSCTL_CHILDREN(rack_probertt),
921 	    OID_AUTO, "filter_life", CTLFLAG_RW,
922 	    &rack_probertt_filter_life, 10000000,
923 	    "What is the time for the filters life in useconds");
924 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
925 	    SYSCTL_CHILDREN(rack_probertt),
926 	    OID_AUTO, "lower_within", CTLFLAG_RW,
927 	    &rack_probertt_lower_within, 10,
928 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
929 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
930 	    SYSCTL_CHILDREN(rack_probertt),
931 	    OID_AUTO, "must_move", CTLFLAG_RW,
932 	    &rack_min_rtt_movement, 250,
933 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
934 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
935 	    SYSCTL_CHILDREN(rack_probertt),
936 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
937 	    &rack_probertt_clear_is, 1,
938 	    "Do we clear I/S counts on exiting probe-rtt");
939 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
940 	    SYSCTL_CHILDREN(rack_probertt),
941 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
942 	    &rack_max_drain_hbp, 1,
943 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
944 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
945 	    SYSCTL_CHILDREN(rack_probertt),
946 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
947 	    &rack_hbp_thresh, 3,
948 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
949 
950 	rack_tracepoint = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
951 	    SYSCTL_CHILDREN(rack_sysctl_root),
952 	    OID_AUTO,
953 	    "tp",
954 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
955 	    "Rack tracepoint facility");
956 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
957 	    SYSCTL_CHILDREN(rack_tracepoint),
958 	    OID_AUTO, "number", CTLFLAG_RW,
959 	    &rack_trace_point_config, 0,
960 	    "What is the trace point number to activate (0=none, 0xffffffff = all)?");
961 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
962 	    SYSCTL_CHILDREN(rack_tracepoint),
963 	    OID_AUTO, "bbmode", CTLFLAG_RW,
964 	    &rack_trace_point_bb_mode, 4,
965 	    "What is BB logging mode that is activated?");
966 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
967 	    SYSCTL_CHILDREN(rack_tracepoint),
968 	    OID_AUTO, "count", CTLFLAG_RW,
969 	    &rack_trace_point_count, 0,
970 	    "How many connections will have BB logging turned on that hit the tracepoint?");
971 	/* Pacing related sysctls */
972 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
973 	    SYSCTL_CHILDREN(rack_sysctl_root),
974 	    OID_AUTO,
975 	    "pacing",
976 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
977 	    "Pacing related Controls");
978 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
979 	    SYSCTL_CHILDREN(rack_pacing),
980 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
981 	    &rack_max_per_above, 30,
982 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
983 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
984 	    SYSCTL_CHILDREN(rack_pacing),
985 	    OID_AUTO, "pace_to_one", CTLFLAG_RW,
986 	    &rack_pace_one_seg, 0,
987 	    "Do we allow low b/w pacing of 1MSS instead of two");
988 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
989 	    SYSCTL_CHILDREN(rack_pacing),
990 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
991 	    &rack_limit_time_with_srtt, 0,
992 	    "Do we limit pacing time based on srtt");
993 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
994 	    SYSCTL_CHILDREN(rack_pacing),
995 	    OID_AUTO, "init_win", CTLFLAG_RW,
996 	    &rack_default_init_window, 0,
997 	    "Do we have a rack initial window 0 = system default");
998 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
999 	    SYSCTL_CHILDREN(rack_pacing),
1000 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1001 	    &rack_per_of_gp_ss, 250,
1002 	    "If non zero, what percentage of goodput to pace at in slow start");
1003 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1004 	    SYSCTL_CHILDREN(rack_pacing),
1005 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1006 	    &rack_per_of_gp_ca, 150,
1007 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
1008 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1009 	    SYSCTL_CHILDREN(rack_pacing),
1010 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1011 	    &rack_per_of_gp_rec, 200,
1012 	    "If non zero, what percentage of goodput to pace at in recovery");
1013 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1014 	    SYSCTL_CHILDREN(rack_pacing),
1015 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1016 	    &rack_hptsi_segments, 40,
1017 	    "What size is the max for TSO segments in pacing and burst mitigation");
1018 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1019 	    SYSCTL_CHILDREN(rack_pacing),
1020 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
1021 	    &rack_slot_reduction, 4,
1022 	    "When doing only burst mitigation what is the reduce divisor");
1023 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1024 	    SYSCTL_CHILDREN(rack_sysctl_root),
1025 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
1026 	    &rack_pace_every_seg, 0,
1027 	    "If set we use pacing, if clear we use only the original burst mitigation");
1028 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1029 	    SYSCTL_CHILDREN(rack_pacing),
1030 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1031 	    &rack_bw_rate_cap, 0,
1032 	    "If set we apply this value to the absolute rate cap used by pacing");
1033 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1034 	    SYSCTL_CHILDREN(rack_sysctl_root),
1035 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1036 	    &rack_req_measurements, 1,
1037 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1038 	/* Hardware pacing */
1039 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1040 	    SYSCTL_CHILDREN(rack_sysctl_root),
1041 	    OID_AUTO,
1042 	    "hdwr_pacing",
1043 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1044 	    "Pacing related Controls");
1045 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1046 	    SYSCTL_CHILDREN(rack_hw_pacing),
1047 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1048 	    &rack_hw_rwnd_factor, 2,
1049 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1050 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1051 	    SYSCTL_CHILDREN(rack_hw_pacing),
1052 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1053 	    &rack_enobuf_hw_boost_mult, 2,
1054 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1055 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1056 	    SYSCTL_CHILDREN(rack_hw_pacing),
1057 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1058 	    &rack_enobuf_hw_max, 2,
1059 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1060 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1061 	    SYSCTL_CHILDREN(rack_hw_pacing),
1062 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1063 	    &rack_enobuf_hw_min, 2,
1064 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1065 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1066 	    SYSCTL_CHILDREN(rack_hw_pacing),
1067 	    OID_AUTO, "enable", CTLFLAG_RW,
1068 	    &rack_enable_hw_pacing, 0,
1069 	    "Should RACK attempt to use hw pacing?");
1070 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1071 	    SYSCTL_CHILDREN(rack_hw_pacing),
1072 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1073 	    &rack_hw_rate_caps, 1,
1074 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1075 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1076 	    SYSCTL_CHILDREN(rack_hw_pacing),
1077 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1078 	    &rack_hw_rate_min, 0,
1079 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1080 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1081 	    SYSCTL_CHILDREN(rack_hw_pacing),
1082 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1083 	    &rack_hw_rate_to_low, 0,
1084 	    "If we fall below this rate, dis-engage hw pacing?");
1085 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1086 	    SYSCTL_CHILDREN(rack_hw_pacing),
1087 	    OID_AUTO, "up_only", CTLFLAG_RW,
1088 	    &rack_hw_up_only, 1,
1089 	    "Do we allow hw pacing to lower the rate selected?");
1090 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1091 	    SYSCTL_CHILDREN(rack_hw_pacing),
1092 	    OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1093 	    &rack_hw_pace_extra_slots, 2,
1094 	    "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1095 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1096 	    SYSCTL_CHILDREN(rack_sysctl_root),
1097 	    OID_AUTO,
1098 	    "timely",
1099 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1100 	    "Rack Timely RTT Controls");
1101 	/* Timely based GP dynmics */
1102 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1103 	    SYSCTL_CHILDREN(rack_timely),
1104 	    OID_AUTO, "upper", CTLFLAG_RW,
1105 	    &rack_gp_per_bw_mul_up, 2,
1106 	    "Rack timely upper range for equal b/w (in percentage)");
1107 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1108 	    SYSCTL_CHILDREN(rack_timely),
1109 	    OID_AUTO, "lower", CTLFLAG_RW,
1110 	    &rack_gp_per_bw_mul_down, 4,
1111 	    "Rack timely lower range for equal b/w (in percentage)");
1112 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1113 	    SYSCTL_CHILDREN(rack_timely),
1114 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1115 	    &rack_gp_rtt_maxmul, 3,
1116 	    "Rack timely multiplier of lowest rtt for rtt_max");
1117 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1118 	    SYSCTL_CHILDREN(rack_timely),
1119 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1120 	    &rack_gp_rtt_mindiv, 4,
1121 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1122 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1123 	    SYSCTL_CHILDREN(rack_timely),
1124 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1125 	    &rack_gp_rtt_minmul, 1,
1126 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1127 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1128 	    SYSCTL_CHILDREN(rack_timely),
1129 	    OID_AUTO, "decrease", CTLFLAG_RW,
1130 	    &rack_gp_decrease_per, 20,
1131 	    "Rack timely decrease percentage of our GP multiplication factor");
1132 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1133 	    SYSCTL_CHILDREN(rack_timely),
1134 	    OID_AUTO, "increase", CTLFLAG_RW,
1135 	    &rack_gp_increase_per, 2,
1136 	    "Rack timely increase perentage of our GP multiplication factor");
1137 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1138 	    SYSCTL_CHILDREN(rack_timely),
1139 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1140 	    &rack_per_lower_bound, 50,
1141 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1142 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1143 	    SYSCTL_CHILDREN(rack_timely),
1144 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1145 	    &rack_per_upper_bound_ss, 0,
1146 	    "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1147 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1148 	    SYSCTL_CHILDREN(rack_timely),
1149 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1150 	    &rack_per_upper_bound_ca, 0,
1151 	    "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1152 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1153 	    SYSCTL_CHILDREN(rack_timely),
1154 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1155 	    &rack_do_dyn_mul, 0,
1156 	    "Rack timely do we enable dynmaic timely goodput by default");
1157 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1158 	    SYSCTL_CHILDREN(rack_timely),
1159 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1160 	    &rack_gp_no_rec_chg, 1,
1161 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1162 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1163 	    SYSCTL_CHILDREN(rack_timely),
1164 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1165 	    &rack_timely_dec_clear, 6,
1166 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1167 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1168 	    SYSCTL_CHILDREN(rack_timely),
1169 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1170 	    &rack_timely_max_push_rise, 3,
1171 	    "Rack timely how many times do we push up with b/w increase");
1172 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1173 	    SYSCTL_CHILDREN(rack_timely),
1174 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1175 	    &rack_timely_max_push_drop, 3,
1176 	    "Rack timely how many times do we push back on b/w decent");
1177 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1178 	    SYSCTL_CHILDREN(rack_timely),
1179 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1180 	    &rack_timely_min_segs, 4,
1181 	    "Rack timely when setting the cwnd what is the min num segments");
1182 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1183 	    SYSCTL_CHILDREN(rack_timely),
1184 	    OID_AUTO, "noback_max", CTLFLAG_RW,
1185 	    &rack_use_max_for_nobackoff, 0,
1186 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1187 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1188 	    SYSCTL_CHILDREN(rack_timely),
1189 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1190 	    &rack_timely_int_timely_only, 0,
1191 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1192 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1193 	    SYSCTL_CHILDREN(rack_timely),
1194 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1195 	    &rack_timely_no_stopping, 0,
1196 	    "Rack timely don't stop increase");
1197 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1198 	    SYSCTL_CHILDREN(rack_timely),
1199 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1200 	    &rack_down_raise_thresh, 100,
1201 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1202 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1203 	    SYSCTL_CHILDREN(rack_timely),
1204 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1205 	    &rack_req_segs, 1,
1206 	    "Bottom dragging if not these many segments outstanding and room");
1207 
1208 	/* TLP and Rack related parameters */
1209 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1210 	    SYSCTL_CHILDREN(rack_sysctl_root),
1211 	    OID_AUTO,
1212 	    "tlp",
1213 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1214 	    "TLP and Rack related Controls");
1215 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1216 	    SYSCTL_CHILDREN(rack_tlp),
1217 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1218 	    &use_rack_rr, 1,
1219 	    "Do we use Rack Rapid Recovery");
1220 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1221 	    SYSCTL_CHILDREN(rack_tlp),
1222 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1223 	    &rack_max_abc_post_recovery, 2,
1224 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1225 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1226 	    SYSCTL_CHILDREN(rack_tlp),
1227 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1228 	    &rack_non_rxt_use_cr, 0,
1229 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1230 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1231 	    SYSCTL_CHILDREN(rack_tlp),
1232 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1233 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1234 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1235 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1236 	    SYSCTL_CHILDREN(rack_tlp),
1237 	    OID_AUTO, "limit", CTLFLAG_RW,
1238 	    &rack_tlp_limit, 2,
1239 	    "How many TLP's can be sent without sending new data");
1240 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1241 	    SYSCTL_CHILDREN(rack_tlp),
1242 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1243 	    &rack_tlp_use_greater, 1,
1244 	    "Should we use the rack_rtt time if its greater than srtt");
1245 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1246 	    SYSCTL_CHILDREN(rack_tlp),
1247 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1248 	    &rack_tlp_min, 10000,
1249 	    "TLP minimum timeout per the specification (in microseconds)");
1250 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1251 	    SYSCTL_CHILDREN(rack_tlp),
1252 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1253 	    &rack_always_send_oldest, 0,
1254 	    "Should we always send the oldest TLP and RACK-TLP");
1255 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1256 	    SYSCTL_CHILDREN(rack_tlp),
1257 	    OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1258 	    &rack_limited_retran, 0,
1259 	    "How many times can a rack timeout drive out sends");
1260 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1261 	    SYSCTL_CHILDREN(rack_tlp),
1262 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1263 	    &rack_lower_cwnd_at_tlp, 0,
1264 	    "When a TLP completes a retran should we enter recovery");
1265 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1266 	    SYSCTL_CHILDREN(rack_tlp),
1267 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1268 	    &rack_reorder_thresh, 2,
1269 	    "What factor for rack will be added when seeing reordering (shift right)");
1270 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1271 	    SYSCTL_CHILDREN(rack_tlp),
1272 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1273 	    &rack_tlp_thresh, 1,
1274 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1275 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1276 	    SYSCTL_CHILDREN(rack_tlp),
1277 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1278 	    &rack_reorder_fade, 60000000,
1279 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1280 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1281 	    SYSCTL_CHILDREN(rack_tlp),
1282 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1283 	    &rack_pkt_delay, 1000,
1284 	    "Extra RACK time (in microseconds) besides reordering thresh");
1285 
1286 	/* Timer related controls */
1287 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1288 	    SYSCTL_CHILDREN(rack_sysctl_root),
1289 	    OID_AUTO,
1290 	    "timers",
1291 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1292 	    "Timer related controls");
1293 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1294 	    SYSCTL_CHILDREN(rack_timers),
1295 	    OID_AUTO, "persmin", CTLFLAG_RW,
1296 	    &rack_persist_min, 250000,
1297 	    "What is the minimum time in microseconds between persists");
1298 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1299 	    SYSCTL_CHILDREN(rack_timers),
1300 	    OID_AUTO, "persmax", CTLFLAG_RW,
1301 	    &rack_persist_max, 2000000,
1302 	    "What is the largest delay in microseconds between persists");
1303 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1304 	    SYSCTL_CHILDREN(rack_timers),
1305 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1306 	    &rack_delayed_ack_time, 40000,
1307 	    "Delayed ack time (40ms in microseconds)");
1308 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1309 	    SYSCTL_CHILDREN(rack_timers),
1310 	    OID_AUTO, "minrto", CTLFLAG_RW,
1311 	    &rack_rto_min, 30000,
1312 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1313 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1314 	    SYSCTL_CHILDREN(rack_timers),
1315 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1316 	    &rack_rto_max, 4000000,
1317 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1318 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1319 	    SYSCTL_CHILDREN(rack_timers),
1320 	    OID_AUTO, "minto", CTLFLAG_RW,
1321 	    &rack_min_to, 1000,
1322 	    "Minimum rack timeout in microseconds");
1323 	/* Measure controls */
1324 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1325 	    SYSCTL_CHILDREN(rack_sysctl_root),
1326 	    OID_AUTO,
1327 	    "measure",
1328 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1329 	    "Measure related controls");
1330 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1331 	    SYSCTL_CHILDREN(rack_measure),
1332 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1333 	    &rack_wma_divisor, 8,
1334 	    "When doing b/w calculation what is the  divisor for the WMA");
1335 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1336 	    SYSCTL_CHILDREN(rack_measure),
1337 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1338 	    &rack_cwnd_block_ends_measure, 0,
1339 	    "Does a cwnd just-return end the measurement window (app limited)");
1340 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1341 	    SYSCTL_CHILDREN(rack_measure),
1342 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1343 	    &rack_rwnd_block_ends_measure, 0,
1344 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1345 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1346 	    SYSCTL_CHILDREN(rack_measure),
1347 	    OID_AUTO, "min_target", CTLFLAG_RW,
1348 	    &rack_def_data_window, 20,
1349 	    "What is the minimum target window (in mss) for a GP measurements");
1350 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1351 	    SYSCTL_CHILDREN(rack_measure),
1352 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1353 	    &rack_goal_bdp, 2,
1354 	    "What is the goal BDP to measure");
1355 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1356 	    SYSCTL_CHILDREN(rack_measure),
1357 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1358 	    &rack_min_srtts, 1,
1359 	    "What is the goal BDP to measure");
1360 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1361 	    SYSCTL_CHILDREN(rack_measure),
1362 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1363 	    &rack_min_measure_usec, 0,
1364 	    "What is the Minimum time time for a measurement if 0, this is off");
1365 	/* Features */
1366 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1367 	    SYSCTL_CHILDREN(rack_sysctl_root),
1368 	    OID_AUTO,
1369 	    "features",
1370 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1371 	    "Feature controls");
1372 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1373 	    SYSCTL_CHILDREN(rack_features),
1374 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1375 	    &rack_use_cmp_acks, 1,
1376 	    "Should RACK have LRO send compressed acks");
1377 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1378 	    SYSCTL_CHILDREN(rack_features),
1379 	    OID_AUTO, "fsb", CTLFLAG_RW,
1380 	    &rack_use_fsb, 1,
1381 	    "Should RACK use the fast send block?");
1382 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1383 	    SYSCTL_CHILDREN(rack_features),
1384 	    OID_AUTO, "rfo", CTLFLAG_RW,
1385 	    &rack_use_rfo, 1,
1386 	    "Should RACK use rack_fast_output()?");
1387 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1388 	    SYSCTL_CHILDREN(rack_features),
1389 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1390 	    &rack_use_rsm_rfo, 1,
1391 	    "Should RACK use rack_fast_rsm_output()?");
1392 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1393 	    SYSCTL_CHILDREN(rack_features),
1394 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1395 	    &rack_enable_mqueue_for_nonpaced, 0,
1396 	    "Should RACK use mbuf queuing for non-paced connections");
1397 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1398 	    SYSCTL_CHILDREN(rack_features),
1399 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1400 	    &rack_do_hystart, 0,
1401 	    "Should RACK enable HyStart++ on connections?");
1402 	/* Misc rack controls */
1403 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1404 	    SYSCTL_CHILDREN(rack_sysctl_root),
1405 	    OID_AUTO,
1406 	    "misc",
1407 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1408 	    "Misc related controls");
1409 #ifdef TCP_ACCOUNTING
1410 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1411 	    SYSCTL_CHILDREN(rack_misc),
1412 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1413 	    &rack_tcp_accounting, 0,
1414 	    "Should we turn on TCP accounting for all rack sessions?");
1415 #endif
1416 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1417 	    SYSCTL_CHILDREN(rack_misc),
1418 	    OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1419 	    &rack_apply_rtt_with_reduced_conf, 0,
1420 	    "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1421 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1422 	    SYSCTL_CHILDREN(rack_misc),
1423 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1424 	    &rack_dsack_std_based, 3,
1425 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1426 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1427 	    SYSCTL_CHILDREN(rack_misc),
1428 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1429 	    &rack_prr_addbackmax, 2,
1430 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1431 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1432 	    SYSCTL_CHILDREN(rack_misc),
1433 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1434 	    &rack_stats_gets_ms_rtt, 1,
1435 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1436 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1437 	    SYSCTL_CHILDREN(rack_misc),
1438 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1439 	    &rack_client_low_buf, 0,
1440 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1441 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1442 	    SYSCTL_CHILDREN(rack_misc),
1443 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1444 	    &rack_def_profile, 0,
1445 	    "Should RACK use a default profile (0=no, num == profile num)?");
1446 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1447 	    SYSCTL_CHILDREN(rack_misc),
1448 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1449 	    &rack_enable_shared_cwnd, 1,
1450 	    "Should RACK try to use the shared cwnd on connections where allowed");
1451 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1452 	    SYSCTL_CHILDREN(rack_misc),
1453 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1454 	    &rack_limits_scwnd, 1,
1455 	    "Should RACK place low end time limits on the shared cwnd feature");
1456 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1457 	    SYSCTL_CHILDREN(rack_misc),
1458 	    OID_AUTO, "iMac_dack", CTLFLAG_RW,
1459 	    &rack_use_imac_dack, 0,
1460 	    "Should RACK try to emulate iMac delayed ack");
1461 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1462 	    SYSCTL_CHILDREN(rack_misc),
1463 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1464 	    &rack_disable_prr, 0,
1465 	    "Should RACK not use prr and only pace (must have pacing on)");
1466 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1467 	    SYSCTL_CHILDREN(rack_misc),
1468 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1469 	    &rack_verbose_logging, 0,
1470 	    "Should RACK black box logging be verbose");
1471 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1472 	    SYSCTL_CHILDREN(rack_misc),
1473 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1474 	    &rack_ignore_data_after_close, 1,
1475 	    "Do we hold off sending a RST until all pending data is ack'd");
1476 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1477 	    SYSCTL_CHILDREN(rack_misc),
1478 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1479 	    &rack_sack_not_required, 1,
1480 	    "Do we allow rack to run on connections not supporting SACK");
1481 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1482 	    SYSCTL_CHILDREN(rack_misc),
1483 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1484 	    &rack_send_a_lot_in_prr, 1,
1485 	    "Send a lot in prr");
1486 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1487 	    SYSCTL_CHILDREN(rack_misc),
1488 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1489 	    &rack_autosndbuf_inc, 20,
1490 	    "What percentage should rack scale up its snd buffer by?");
1491 	/* Sack Attacker detection stuff */
1492 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1493 	    SYSCTL_CHILDREN(rack_attack),
1494 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1495 	    &rack_highest_sack_thresh_seen, 0,
1496 	    "Highest sack to ack ratio seen");
1497 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1498 	    SYSCTL_CHILDREN(rack_attack),
1499 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1500 	    &rack_highest_move_thresh_seen, 0,
1501 	    "Highest move to non-move ratio seen");
1502 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1503 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1504 	    SYSCTL_CHILDREN(rack_attack),
1505 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1506 	    &rack_ack_total,
1507 	    "Total number of Ack's");
1508 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1509 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1510 	    SYSCTL_CHILDREN(rack_attack),
1511 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1512 	    &rack_express_sack,
1513 	    "Total expresss number of Sack's");
1514 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1515 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1516 	    SYSCTL_CHILDREN(rack_attack),
1517 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1518 	    &rack_sack_total,
1519 	    "Total number of SACKs");
1520 	rack_move_none = counter_u64_alloc(M_WAITOK);
1521 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1522 	    SYSCTL_CHILDREN(rack_attack),
1523 	    OID_AUTO, "move_none", CTLFLAG_RD,
1524 	    &rack_move_none,
1525 	    "Total number of SACK index reuse of positions under threshold");
1526 	rack_move_some = counter_u64_alloc(M_WAITOK);
1527 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1528 	    SYSCTL_CHILDREN(rack_attack),
1529 	    OID_AUTO, "move_some", CTLFLAG_RD,
1530 	    &rack_move_some,
1531 	    "Total number of SACK index reuse of positions over threshold");
1532 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1533 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1534 	    SYSCTL_CHILDREN(rack_attack),
1535 	    OID_AUTO, "attacks", CTLFLAG_RD,
1536 	    &rack_sack_attacks_detected,
1537 	    "Total number of SACK attackers that had sack disabled");
1538 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1539 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1540 	    SYSCTL_CHILDREN(rack_attack),
1541 	    OID_AUTO, "reversed", CTLFLAG_RD,
1542 	    &rack_sack_attacks_reversed,
1543 	    "Total number of SACK attackers that were later determined false positive");
1544 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1545 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1546 	    SYSCTL_CHILDREN(rack_attack),
1547 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1548 	    &rack_sack_used_next_merge,
1549 	    "Total number of times we used the next merge");
1550 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1551 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1552 	    SYSCTL_CHILDREN(rack_attack),
1553 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1554 	    &rack_sack_used_prev_merge,
1555 	    "Total number of times we used the prev merge");
1556 	/* Counters */
1557 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1558 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1559 	    SYSCTL_CHILDREN(rack_counters),
1560 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1561 	    &rack_fto_send, "Total number of rack_fast_output sends");
1562 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1563 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1564 	    SYSCTL_CHILDREN(rack_counters),
1565 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1566 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1567 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1568 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1569 	    SYSCTL_CHILDREN(rack_counters),
1570 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1571 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1572 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1573 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1574 	    SYSCTL_CHILDREN(rack_counters),
1575 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1576 	    &rack_non_fto_send, "Total number of rack_output first sends");
1577 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1578 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1579 	    SYSCTL_CHILDREN(rack_counters),
1580 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1581 	    &rack_extended_rfo, "Total number of times we extended rfo");
1582 
1583 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1584 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1585 	    SYSCTL_CHILDREN(rack_counters),
1586 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1587 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1588 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1589 
1590 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1591 	    SYSCTL_CHILDREN(rack_counters),
1592 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1593 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1594 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1595 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1596 	    SYSCTL_CHILDREN(rack_counters),
1597 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1598 	    &rack_tlp_tot,
1599 	    "Total number of tail loss probe expirations");
1600 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1601 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1602 	    SYSCTL_CHILDREN(rack_counters),
1603 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1604 	    &rack_tlp_newdata,
1605 	    "Total number of tail loss probe sending new data");
1606 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1607 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1608 	    SYSCTL_CHILDREN(rack_counters),
1609 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1610 	    &rack_tlp_retran,
1611 	    "Total number of tail loss probe sending retransmitted data");
1612 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1613 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1614 	    SYSCTL_CHILDREN(rack_counters),
1615 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1616 	    &rack_tlp_retran_bytes,
1617 	    "Total bytes of tail loss probe sending retransmitted data");
1618 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1619 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1620 	    SYSCTL_CHILDREN(rack_counters),
1621 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1622 	    &rack_to_tot,
1623 	    "Total number of times the rack to expired");
1624 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1625 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1626 	    SYSCTL_CHILDREN(rack_counters),
1627 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1628 	    &rack_saw_enobuf,
1629 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1630 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1631 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1632 	    SYSCTL_CHILDREN(rack_counters),
1633 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1634 	    &rack_saw_enobuf_hw,
1635 	    "Total number of times a send returned enobuf for hdwr paced connections");
1636 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1637 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1638 	    SYSCTL_CHILDREN(rack_counters),
1639 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1640 	    &rack_saw_enetunreach,
1641 	    "Total number of times a send received a enetunreachable");
1642 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1643 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1644 	    SYSCTL_CHILDREN(rack_counters),
1645 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1646 	    &rack_hot_alloc,
1647 	    "Total allocations from the top of our list");
1648 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1649 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1650 	    SYSCTL_CHILDREN(rack_counters),
1651 	    OID_AUTO, "allocs", CTLFLAG_RD,
1652 	    &rack_to_alloc,
1653 	    "Total allocations of tracking structures");
1654 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1655 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1656 	    SYSCTL_CHILDREN(rack_counters),
1657 	    OID_AUTO, "allochard", CTLFLAG_RD,
1658 	    &rack_to_alloc_hard,
1659 	    "Total allocations done with sleeping the hard way");
1660 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1661 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1662 	    SYSCTL_CHILDREN(rack_counters),
1663 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1664 	    &rack_to_alloc_emerg,
1665 	    "Total allocations done from emergency cache");
1666 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1667 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1668 	    SYSCTL_CHILDREN(rack_counters),
1669 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1670 	    &rack_to_alloc_limited,
1671 	    "Total allocations dropped due to limit");
1672 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1673 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1674 	    SYSCTL_CHILDREN(rack_counters),
1675 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1676 	    &rack_alloc_limited_conns,
1677 	    "Connections with allocations dropped due to limit");
1678 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1679 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1680 	    SYSCTL_CHILDREN(rack_counters),
1681 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1682 	    &rack_split_limited,
1683 	    "Split allocations dropped due to limit");
1684 	rack_persists_sends = counter_u64_alloc(M_WAITOK);
1685 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1686 	    SYSCTL_CHILDREN(rack_counters),
1687 	    OID_AUTO, "persist_sends", CTLFLAG_RD,
1688 	    &rack_persists_sends,
1689 	    "Number of times we sent a persist probe");
1690 	rack_persists_acks = counter_u64_alloc(M_WAITOK);
1691 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1692 	    SYSCTL_CHILDREN(rack_counters),
1693 	    OID_AUTO, "persist_acks", CTLFLAG_RD,
1694 	    &rack_persists_acks,
1695 	    "Number of times a persist probe was acked");
1696 	rack_persists_loss = counter_u64_alloc(M_WAITOK);
1697 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1698 	    SYSCTL_CHILDREN(rack_counters),
1699 	    OID_AUTO, "persist_loss", CTLFLAG_RD,
1700 	    &rack_persists_loss,
1701 	    "Number of times we detected a lost persist probe (no ack)");
1702 	rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1703 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1704 	    SYSCTL_CHILDREN(rack_counters),
1705 	    OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1706 	    &rack_persists_lost_ends,
1707 	    "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1708 #ifdef INVARIANTS
1709 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1710 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1711 	    SYSCTL_CHILDREN(rack_counters),
1712 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1713 	    &rack_adjust_map_bw,
1714 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1715 #endif
1716 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1717 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1718 	    SYSCTL_CHILDREN(rack_counters),
1719 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1720 	    &rack_multi_single_eq,
1721 	    "Number of compressed acks total represented");
1722 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1723 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1724 	    SYSCTL_CHILDREN(rack_counters),
1725 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1726 	    &rack_proc_non_comp_ack,
1727 	    "Number of non compresseds acks that we processed");
1728 
1729 
1730 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1731 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1732 	    SYSCTL_CHILDREN(rack_counters),
1733 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1734 	    &rack_sack_proc_all,
1735 	    "Total times we had to walk whole list for sack processing");
1736 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1737 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1738 	    SYSCTL_CHILDREN(rack_counters),
1739 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1740 	    &rack_sack_proc_restart,
1741 	    "Total times we had to walk whole list due to a restart");
1742 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1743 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1744 	    SYSCTL_CHILDREN(rack_counters),
1745 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1746 	    &rack_sack_proc_short,
1747 	    "Total times we took shortcut for sack processing");
1748 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1749 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1750 	    SYSCTL_CHILDREN(rack_attack),
1751 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1752 	    &rack_sack_skipped_acked,
1753 	    "Total number of times we skipped previously sacked");
1754 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1755 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1756 	    SYSCTL_CHILDREN(rack_attack),
1757 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1758 	    &rack_sack_splits,
1759 	    "Total number of times we did the old fashion tree split");
1760 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1761 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1762 	    SYSCTL_CHILDREN(rack_counters),
1763 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1764 	    &rack_input_idle_reduces,
1765 	    "Total number of idle reductions on input");
1766 	rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1767 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1768 	    SYSCTL_CHILDREN(rack_counters),
1769 	    OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1770 	    &rack_collapsed_win_seen,
1771 	    "Total number of collapsed window events seen (where our window shrinks)");
1772 
1773 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1774 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1775 	    SYSCTL_CHILDREN(rack_counters),
1776 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1777 	    &rack_collapsed_win,
1778 	    "Total number of collapsed window events where we mark packets");
1779 	rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1780 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1781 	    SYSCTL_CHILDREN(rack_counters),
1782 	    OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1783 	    &rack_collapsed_win_rxt,
1784 	    "Total number of packets that were retransmitted");
1785 	rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1786 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1787 	    SYSCTL_CHILDREN(rack_counters),
1788 	    OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1789 	    &rack_collapsed_win_rxt_bytes,
1790 	    "Total number of bytes that were retransmitted");
1791 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1792 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1793 	    SYSCTL_CHILDREN(rack_counters),
1794 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1795 	    &rack_try_scwnd,
1796 	    "Total number of scwnd attempts");
1797 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1798 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1799 	    OID_AUTO, "outsize", CTLFLAG_RD,
1800 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1801 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1802 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1803 	    OID_AUTO, "opts", CTLFLAG_RD,
1804 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1805 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1806 	    SYSCTL_CHILDREN(rack_sysctl_root),
1807 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1808 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1809 }
1810 
1811 static __inline int
1812 rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a)
1813 {
1814 	if (SEQ_GEQ(b->r_start, a->r_start) &&
1815 	    SEQ_LT(b->r_start, a->r_end)) {
1816 		/*
1817 		 * The entry b is within the
1818 		 * block a. i.e.:
1819 		 * a --   |-------------|
1820 		 * b --   |----|
1821 		 * <or>
1822 		 * b --       |------|
1823 		 * <or>
1824 		 * b --       |-----------|
1825 		 */
1826 		return (0);
1827 	} else if (SEQ_GEQ(b->r_start, a->r_end)) {
1828 		/*
1829 		 * b falls as either the next
1830 		 * sequence block after a so a
1831 		 * is said to be smaller than b.
1832 		 * i.e:
1833 		 * a --   |------|
1834 		 * b --          |--------|
1835 		 * or
1836 		 * b --              |-----|
1837 		 */
1838 		return (1);
1839 	}
1840 	/*
1841 	 * Whats left is where a is
1842 	 * larger than b. i.e:
1843 	 * a --         |-------|
1844 	 * b --  |---|
1845 	 * or even possibly
1846 	 * b --   |--------------|
1847 	 */
1848 	return (-1);
1849 }
1850 
1851 RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1852 RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp);
1853 
1854 static uint32_t
1855 rc_init_window(struct tcp_rack *rack)
1856 {
1857 	uint32_t win;
1858 
1859 	if (rack->rc_init_win == 0) {
1860 		/*
1861 		 * Nothing set by the user, use the system stack
1862 		 * default.
1863 		 */
1864 		return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1865 	}
1866 	win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1867 	return (win);
1868 }
1869 
1870 static uint64_t
1871 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1872 {
1873 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1874 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1875 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1876 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1877 	else
1878 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1879 }
1880 
1881 static uint64_t
1882 rack_get_bw(struct tcp_rack *rack)
1883 {
1884 	if (rack->use_fixed_rate) {
1885 		/* Return the fixed pacing rate */
1886 		return (rack_get_fixed_pacing_bw(rack));
1887 	}
1888 	if (rack->r_ctl.gp_bw == 0) {
1889 		/*
1890 		 * We have yet no b/w measurement,
1891 		 * if we have a user set initial bw
1892 		 * return it. If we don't have that and
1893 		 * we have an srtt, use the tcp IW (10) to
1894 		 * calculate a fictional b/w over the SRTT
1895 		 * which is more or less a guess. Note
1896 		 * we don't use our IW from rack on purpose
1897 		 * so if we have like IW=30, we are not
1898 		 * calculating a "huge" b/w.
1899 		 */
1900 		uint64_t bw, srtt;
1901 		if (rack->r_ctl.init_rate)
1902 			return (rack->r_ctl.init_rate);
1903 
1904 		/* Has the user set a max peak rate? */
1905 #ifdef NETFLIX_PEAKRATE
1906 		if (rack->rc_tp->t_maxpeakrate)
1907 			return (rack->rc_tp->t_maxpeakrate);
1908 #endif
1909 		/* Ok lets come up with the IW guess, if we have a srtt */
1910 		if (rack->rc_tp->t_srtt == 0) {
1911 			/*
1912 			 * Go with old pacing method
1913 			 * i.e. burst mitigation only.
1914 			 */
1915 			return (0);
1916 		}
1917 		/* Ok lets get the initial TCP win (not racks) */
1918 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
1919 		srtt = (uint64_t)rack->rc_tp->t_srtt;
1920 		bw *= (uint64_t)USECS_IN_SECOND;
1921 		bw /= srtt;
1922 		if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1923 			bw = rack->r_ctl.bw_rate_cap;
1924 		return (bw);
1925 	} else {
1926 		uint64_t bw;
1927 
1928 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
1929 			/* Averaging is done, we can return the value */
1930 			bw = rack->r_ctl.gp_bw;
1931 		} else {
1932 			/* Still doing initial average must calculate */
1933 			bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements;
1934 		}
1935 #ifdef NETFLIX_PEAKRATE
1936 		if ((rack->rc_tp->t_maxpeakrate) &&
1937 		    (bw > rack->rc_tp->t_maxpeakrate)) {
1938 			/* The user has set a peak rate to pace at
1939 			 * don't allow us to pace faster than that.
1940 			 */
1941 			return (rack->rc_tp->t_maxpeakrate);
1942 		}
1943 #endif
1944 		if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap))
1945 			bw = rack->r_ctl.bw_rate_cap;
1946 		return (bw);
1947 	}
1948 }
1949 
1950 static uint16_t
1951 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
1952 {
1953 	if (rack->use_fixed_rate) {
1954 		return (100);
1955 	} else if (rack->in_probe_rtt && (rsm == NULL))
1956 		return (rack->r_ctl.rack_per_of_gp_probertt);
1957 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
1958 		  rack->r_ctl.rack_per_of_gp_rec)) {
1959 		if (rsm) {
1960 			/* a retransmission always use the recovery rate */
1961 			return (rack->r_ctl.rack_per_of_gp_rec);
1962 		} else if (rack->rack_rec_nonrxt_use_cr) {
1963 			/* Directed to use the configured rate */
1964 			goto configured_rate;
1965 		} else if (rack->rack_no_prr &&
1966 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
1967 			/* No PRR, lets just use the b/w estimate only */
1968 			return (100);
1969 		} else {
1970 			/*
1971 			 * Here we may have a non-retransmit but we
1972 			 * have no overrides, so just use the recovery
1973 			 * rate (prr is in effect).
1974 			 */
1975 			return (rack->r_ctl.rack_per_of_gp_rec);
1976 		}
1977 	}
1978 configured_rate:
1979 	/* For the configured rate we look at our cwnd vs the ssthresh */
1980 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1981 		return (rack->r_ctl.rack_per_of_gp_ss);
1982 	else
1983 		return (rack->r_ctl.rack_per_of_gp_ca);
1984 }
1985 
1986 static void
1987 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
1988 {
1989 	/*
1990 	 * Types of logs (mod value)
1991 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
1992 	 * 2 = a dsack round begins, persist is reset to 16.
1993 	 * 3 = a dsack round ends
1994 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
1995 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
1996 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
1997 	 */
1998 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1999 		union tcp_log_stackspecific log;
2000 		struct timeval tv;
2001 
2002 		memset(&log, 0, sizeof(log));
2003 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2004 		log.u_bbr.flex1 <<= 1;
2005 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2006 		log.u_bbr.flex1 <<= 1;
2007 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2008 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2009 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2010 		log.u_bbr.flex4 = flex4;
2011 		log.u_bbr.flex5 = flex5;
2012 		log.u_bbr.flex6 = flex6;
2013 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2014 		log.u_bbr.flex8 = mod;
2015 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2016 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2017 		    &rack->rc_inp->inp_socket->so_rcv,
2018 		    &rack->rc_inp->inp_socket->so_snd,
2019 		    RACK_DSACK_HANDLING, 0,
2020 		    0, &log, false, &tv);
2021 	}
2022 }
2023 
2024 static void
2025 rack_log_hdwr_pacing(struct tcp_rack *rack,
2026 		     uint64_t rate, uint64_t hw_rate, int line,
2027 		     int error, uint16_t mod)
2028 {
2029 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2030 		union tcp_log_stackspecific log;
2031 		struct timeval tv;
2032 		const struct ifnet *ifp;
2033 
2034 		memset(&log, 0, sizeof(log));
2035 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2036 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2037 		if (rack->r_ctl.crte) {
2038 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2039 		} else if (rack->rc_inp->inp_route.ro_nh &&
2040 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2041 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2042 		} else
2043 			ifp = NULL;
2044 		if (ifp) {
2045 			log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2046 			log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2047 		}
2048 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2049 		log.u_bbr.bw_inuse = rate;
2050 		log.u_bbr.flex5 = line;
2051 		log.u_bbr.flex6 = error;
2052 		log.u_bbr.flex7 = mod;
2053 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2054 		log.u_bbr.flex8 = rack->use_fixed_rate;
2055 		log.u_bbr.flex8 <<= 1;
2056 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2057 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2058 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2059 		if (rack->r_ctl.crte)
2060 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2061 		else
2062 			log.u_bbr.cur_del_rate = 0;
2063 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2064 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2065 		    &rack->rc_inp->inp_socket->so_rcv,
2066 		    &rack->rc_inp->inp_socket->so_snd,
2067 		    BBR_LOG_HDWR_PACE, 0,
2068 		    0, &log, false, &tv);
2069 	}
2070 }
2071 
2072 static uint64_t
2073 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2074 {
2075 	/*
2076 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2077 	 */
2078 	uint64_t bw_est, high_rate;
2079 	uint64_t gain;
2080 
2081 	gain = (uint64_t)rack_get_output_gain(rack, rsm);
2082 	bw_est = bw * gain;
2083 	bw_est /= (uint64_t)100;
2084 	/* Never fall below the minimum (def 64kbps) */
2085 	if (bw_est < RACK_MIN_BW)
2086 		bw_est = RACK_MIN_BW;
2087 	if (rack->r_rack_hw_rate_caps) {
2088 		/* Rate caps are in place */
2089 		if (rack->r_ctl.crte != NULL) {
2090 			/* We have a hdwr rate already */
2091 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2092 			if (bw_est >= high_rate) {
2093 				/* We are capping bw at the highest rate table entry */
2094 				rack_log_hdwr_pacing(rack,
2095 						     bw_est, high_rate, __LINE__,
2096 						     0, 3);
2097 				bw_est = high_rate;
2098 				if (capped)
2099 					*capped = 1;
2100 			}
2101 		} else if ((rack->rack_hdrw_pacing == 0) &&
2102 			   (rack->rack_hdw_pace_ena) &&
2103 			   (rack->rack_attempt_hdwr_pace == 0) &&
2104 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2105 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2106 			/*
2107 			 * Special case, we have not yet attempted hardware
2108 			 * pacing, and yet we may, when we do, find out if we are
2109 			 * above the highest rate. We need to know the maxbw for the interface
2110 			 * in question (if it supports ratelimiting). We get back
2111 			 * a 0, if the interface is not found in the RL lists.
2112 			 */
2113 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2114 			if (high_rate) {
2115 				/* Yep, we have a rate is it above this rate? */
2116 				if (bw_est > high_rate) {
2117 					bw_est = high_rate;
2118 					if (capped)
2119 						*capped = 1;
2120 				}
2121 			}
2122 		}
2123 	}
2124 	return (bw_est);
2125 }
2126 
2127 static void
2128 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2129 {
2130 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2131 		union tcp_log_stackspecific log;
2132 		struct timeval tv;
2133 
2134 		if ((mod != 1) && (rack_verbose_logging == 0)) {
2135 			/*
2136 			 * We get 3 values currently for mod
2137 			 * 1 - We are retransmitting and this tells the reason.
2138 			 * 2 - We are clearing a dup-ack count.
2139 			 * 3 - We are incrementing a dup-ack count.
2140 			 *
2141 			 * The clear/increment are only logged
2142 			 * if you have BBverbose on.
2143 			 */
2144 			return;
2145 		}
2146 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2147 		log.u_bbr.flex1 = tsused;
2148 		log.u_bbr.flex2 = thresh;
2149 		log.u_bbr.flex3 = rsm->r_flags;
2150 		log.u_bbr.flex4 = rsm->r_dupack;
2151 		log.u_bbr.flex5 = rsm->r_start;
2152 		log.u_bbr.flex6 = rsm->r_end;
2153 		log.u_bbr.flex8 = mod;
2154 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2155 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2156 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2157 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2158 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2159 		log.u_bbr.pacing_gain = rack->r_must_retran;
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_SETTINGS_CHG, 0,
2164 		    0, &log, false, &tv);
2165 	}
2166 }
2167 
2168 static void
2169 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
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.flex1 = rack->rc_tp->t_srtt;
2177 		log.u_bbr.flex2 = to;
2178 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2179 		log.u_bbr.flex4 = slot;
2180 		log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
2181 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2182 		log.u_bbr.flex7 = rack->rc_in_persist;
2183 		log.u_bbr.flex8 = which;
2184 		if (rack->rack_no_prr)
2185 			log.u_bbr.pkts_out = 0;
2186 		else
2187 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2188 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2189 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2190 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2191 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2192 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2193 		log.u_bbr.pacing_gain = rack->r_must_retran;
2194 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2195 		log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2196 		log.u_bbr.lost = rack_rto_min;
2197 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2198 		    &rack->rc_inp->inp_socket->so_rcv,
2199 		    &rack->rc_inp->inp_socket->so_snd,
2200 		    BBR_LOG_TIMERSTAR, 0,
2201 		    0, &log, false, &tv);
2202 	}
2203 }
2204 
2205 static void
2206 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2207 {
2208 	if (rack->rc_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.inhpts = tcp_in_hpts(rack->rc_inp);
2214 		log.u_bbr.flex8 = to_num;
2215 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2216 		log.u_bbr.flex2 = rack->rc_rack_rtt;
2217 		if (rsm == NULL)
2218 			log.u_bbr.flex3 = 0;
2219 		else
2220 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2221 		if (rack->rack_no_prr)
2222 			log.u_bbr.flex5 = 0;
2223 		else
2224 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2225 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2226 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2227 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2228 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2229 		log.u_bbr.pacing_gain = rack->r_must_retran;
2230 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2231 		    &rack->rc_inp->inp_socket->so_rcv,
2232 		    &rack->rc_inp->inp_socket->so_snd,
2233 		    BBR_LOG_RTO, 0,
2234 		    0, &log, false, &tv);
2235 	}
2236 }
2237 
2238 static void
2239 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2240 		 struct rack_sendmap *prev,
2241 		 struct rack_sendmap *rsm,
2242 		 struct rack_sendmap *next,
2243 		 int flag, uint32_t th_ack, int line)
2244 {
2245 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2246 		union tcp_log_stackspecific log;
2247 		struct timeval tv;
2248 
2249 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2250 		log.u_bbr.flex8 = flag;
2251 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2252 		log.u_bbr.cur_del_rate = (uint64_t)prev;
2253 		log.u_bbr.delRate = (uint64_t)rsm;
2254 		log.u_bbr.rttProp = (uint64_t)next;
2255 		log.u_bbr.flex7 = 0;
2256 		if (prev) {
2257 			log.u_bbr.flex1 = prev->r_start;
2258 			log.u_bbr.flex2 = prev->r_end;
2259 			log.u_bbr.flex7 |= 0x4;
2260 		}
2261 		if (rsm) {
2262 			log.u_bbr.flex3 = rsm->r_start;
2263 			log.u_bbr.flex4 = rsm->r_end;
2264 			log.u_bbr.flex7 |= 0x2;
2265 		}
2266 		if (next) {
2267 			log.u_bbr.flex5 = next->r_start;
2268 			log.u_bbr.flex6 = next->r_end;
2269 			log.u_bbr.flex7 |= 0x1;
2270 		}
2271 		log.u_bbr.applimited = line;
2272 		log.u_bbr.pkts_out = th_ack;
2273 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2274 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2275 		if (rack->rack_no_prr)
2276 			log.u_bbr.lost = 0;
2277 		else
2278 			log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2279 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2280 		    &rack->rc_inp->inp_socket->so_rcv,
2281 		    &rack->rc_inp->inp_socket->so_snd,
2282 		    TCP_LOG_MAPCHG, 0,
2283 		    0, &log, false, &tv);
2284 	}
2285 }
2286 
2287 static void
2288 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2289 		 struct rack_sendmap *rsm, int conf)
2290 {
2291 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2292 		union tcp_log_stackspecific log;
2293 		struct timeval tv;
2294 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2295 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2296 		log.u_bbr.flex1 = t;
2297 		log.u_bbr.flex2 = len;
2298 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2299 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2300 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2301 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2302 		log.u_bbr.flex7 = conf;
2303 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2304 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2305 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2306 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2307 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2308 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2309 		if (rsm) {
2310 			log.u_bbr.pkt_epoch = rsm->r_start;
2311 			log.u_bbr.lost = rsm->r_end;
2312 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2313 			/* We loose any upper of the 24 bits */
2314 			log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2315 		} else {
2316 			/* Its a SYN */
2317 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2318 			log.u_bbr.lost = 0;
2319 			log.u_bbr.cwnd_gain = 0;
2320 			log.u_bbr.pacing_gain = 0;
2321 		}
2322 		/* Write out general bits of interest rrs here */
2323 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2324 		log.u_bbr.use_lt_bw <<= 1;
2325 		log.u_bbr.use_lt_bw |= rack->forced_ack;
2326 		log.u_bbr.use_lt_bw <<= 1;
2327 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2328 		log.u_bbr.use_lt_bw <<= 1;
2329 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2330 		log.u_bbr.use_lt_bw <<= 1;
2331 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2332 		log.u_bbr.use_lt_bw <<= 1;
2333 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2334 		log.u_bbr.use_lt_bw <<= 1;
2335 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2336 		log.u_bbr.use_lt_bw <<= 1;
2337 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2338 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2339 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2340 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2341 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2342 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2343 		log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2344 		log.u_bbr.bw_inuse <<= 32;
2345 		if (rsm)
2346 			log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2347 		TCP_LOG_EVENTP(tp, NULL,
2348 		    &rack->rc_inp->inp_socket->so_rcv,
2349 		    &rack->rc_inp->inp_socket->so_snd,
2350 		    BBR_LOG_BBRRTT, 0,
2351 		    0, &log, false, &tv);
2352 
2353 
2354 	}
2355 }
2356 
2357 static void
2358 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2359 {
2360 	/*
2361 	 * Log the rtt sample we are
2362 	 * applying to the srtt algorithm in
2363 	 * useconds.
2364 	 */
2365 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2366 		union tcp_log_stackspecific log;
2367 		struct timeval tv;
2368 
2369 		/* Convert our ms to a microsecond */
2370 		memset(&log, 0, sizeof(log));
2371 		log.u_bbr.flex1 = rtt;
2372 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2373 		log.u_bbr.flex3 = rack->r_ctl.sack_count;
2374 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2375 		log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2376 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2377 		log.u_bbr.flex7 = 1;
2378 		log.u_bbr.flex8 = rack->sack_attack_disable;
2379 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2380 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2381 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2382 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2383 		log.u_bbr.pacing_gain = rack->r_must_retran;
2384 		/*
2385 		 * We capture in delRate the upper 32 bits as
2386 		 * the confidence level we had declared, and the
2387 		 * lower 32 bits as the actual RTT using the arrival
2388 		 * timestamp.
2389 		 */
2390 		log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2391 		log.u_bbr.delRate <<= 32;
2392 		log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2393 		/* Lets capture all the things that make up t_rtxcur */
2394 		log.u_bbr.applimited = rack_rto_min;
2395 		log.u_bbr.epoch = rack_rto_max;
2396 		log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2397 		log.u_bbr.lost = rack_rto_min;
2398 		log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2399 		log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2400 		log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2401 		log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2402 		log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2403 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2404 		    &rack->rc_inp->inp_socket->so_rcv,
2405 		    &rack->rc_inp->inp_socket->so_snd,
2406 		    TCP_LOG_RTT, 0,
2407 		    0, &log, false, &tv);
2408 	}
2409 }
2410 
2411 static void
2412 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2413 {
2414 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
2415 		union tcp_log_stackspecific log;
2416 		struct timeval tv;
2417 
2418 		/* Convert our ms to a microsecond */
2419 		memset(&log, 0, sizeof(log));
2420 		log.u_bbr.flex1 = rtt;
2421 		log.u_bbr.flex2 = send_time;
2422 		log.u_bbr.flex3 = ack_time;
2423 		log.u_bbr.flex4 = where;
2424 		log.u_bbr.flex7 = 2;
2425 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2426 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2427 		    &rack->rc_inp->inp_socket->so_rcv,
2428 		    &rack->rc_inp->inp_socket->so_snd,
2429 		    TCP_LOG_RTT, 0,
2430 		    0, &log, false, &tv);
2431 	}
2432 }
2433 
2434 
2435 
2436 static inline void
2437 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2438 {
2439 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
2440 		union tcp_log_stackspecific log;
2441 		struct timeval tv;
2442 
2443 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2444 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2445 		log.u_bbr.flex1 = line;
2446 		log.u_bbr.flex2 = tick;
2447 		log.u_bbr.flex3 = tp->t_maxunacktime;
2448 		log.u_bbr.flex4 = tp->t_acktime;
2449 		log.u_bbr.flex8 = event;
2450 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2451 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2452 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2453 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2454 		log.u_bbr.pacing_gain = rack->r_must_retran;
2455 		TCP_LOG_EVENTP(tp, NULL,
2456 		    &rack->rc_inp->inp_socket->so_rcv,
2457 		    &rack->rc_inp->inp_socket->so_snd,
2458 		    BBR_LOG_PROGRESS, 0,
2459 		    0, &log, false, &tv);
2460 	}
2461 }
2462 
2463 static void
2464 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv)
2465 {
2466 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2467 		union tcp_log_stackspecific log;
2468 
2469 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2470 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2471 		log.u_bbr.flex1 = slot;
2472 		if (rack->rack_no_prr)
2473 			log.u_bbr.flex2 = 0;
2474 		else
2475 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
2476 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
2477 		log.u_bbr.flex8 = rack->rc_in_persist;
2478 		log.u_bbr.timeStamp = cts;
2479 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2480 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2481 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2482 		log.u_bbr.pacing_gain = rack->r_must_retran;
2483 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2484 		    &rack->rc_inp->inp_socket->so_rcv,
2485 		    &rack->rc_inp->inp_socket->so_snd,
2486 		    BBR_LOG_BBRSND, 0,
2487 		    0, &log, false, tv);
2488 	}
2489 }
2490 
2491 static void
2492 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
2493 {
2494 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2495 		union tcp_log_stackspecific log;
2496 		struct timeval tv;
2497 
2498 		memset(&log, 0, sizeof(log));
2499 		log.u_bbr.flex1 = did_out;
2500 		log.u_bbr.flex2 = nxt_pkt;
2501 		log.u_bbr.flex3 = way_out;
2502 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2503 		if (rack->rack_no_prr)
2504 			log.u_bbr.flex5 = 0;
2505 		else
2506 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2507 		log.u_bbr.flex6 = nsegs;
2508 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
2509 		log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;	/* Do we have ack-can-send set */
2510 		log.u_bbr.flex7 <<= 1;
2511 		log.u_bbr.flex7 |= rack->r_fast_output;	/* is fast output primed */
2512 		log.u_bbr.flex7 <<= 1;
2513 		log.u_bbr.flex7 |= rack->r_wanted_output;	/* Do we want output */
2514 		log.u_bbr.flex8 = rack->rc_in_persist;
2515 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2516 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2517 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2518 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2519 		log.u_bbr.use_lt_bw <<= 1;
2520 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
2521 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2522 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2523 		log.u_bbr.pacing_gain = rack->r_must_retran;
2524 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2525 		    &rack->rc_inp->inp_socket->so_rcv,
2526 		    &rack->rc_inp->inp_socket->so_snd,
2527 		    BBR_LOG_DOSEG_DONE, 0,
2528 		    0, &log, false, &tv);
2529 	}
2530 }
2531 
2532 static void
2533 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
2534 {
2535 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
2536 		union tcp_log_stackspecific log;
2537 		struct timeval tv;
2538 
2539 		memset(&log, 0, sizeof(log));
2540 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
2541 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
2542 		log.u_bbr.flex4 = arg1;
2543 		log.u_bbr.flex5 = arg2;
2544 		log.u_bbr.flex6 = arg3;
2545 		log.u_bbr.flex8 = frm;
2546 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2547 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2548 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2549 		log.u_bbr.applimited = rack->r_ctl.rc_sacked;
2550 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2551 		log.u_bbr.pacing_gain = rack->r_must_retran;
2552 		TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
2553 		    &tptosocket(tp)->so_snd,
2554 		    TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
2555 	}
2556 }
2557 
2558 static void
2559 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
2560 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
2561 {
2562 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2563 		union tcp_log_stackspecific log;
2564 		struct timeval tv;
2565 
2566 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2567 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2568 		log.u_bbr.flex1 = slot;
2569 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
2570 		log.u_bbr.flex4 = reason;
2571 		if (rack->rack_no_prr)
2572 			log.u_bbr.flex5 = 0;
2573 		else
2574 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2575 		log.u_bbr.flex7 = hpts_calling;
2576 		log.u_bbr.flex8 = rack->rc_in_persist;
2577 		log.u_bbr.lt_epoch = cwnd_to_use;
2578 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2579 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2580 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2581 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2582 		log.u_bbr.pacing_gain = rack->r_must_retran;
2583 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
2584 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2585 		    &rack->rc_inp->inp_socket->so_rcv,
2586 		    &rack->rc_inp->inp_socket->so_snd,
2587 		    BBR_LOG_JUSTRET, 0,
2588 		    tlen, &log, false, &tv);
2589 	}
2590 }
2591 
2592 static void
2593 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
2594 		   struct timeval *tv, uint32_t flags_on_entry)
2595 {
2596 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2597 		union tcp_log_stackspecific log;
2598 
2599 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2600 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
2601 		log.u_bbr.flex1 = line;
2602 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
2603 		log.u_bbr.flex3 = flags_on_entry;
2604 		log.u_bbr.flex4 = us_cts;
2605 		if (rack->rack_no_prr)
2606 			log.u_bbr.flex5 = 0;
2607 		else
2608 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2609 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2610 		log.u_bbr.flex7 = hpts_removed;
2611 		log.u_bbr.flex8 = 1;
2612 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
2613 		log.u_bbr.timeStamp = us_cts;
2614 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2615 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2616 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2617 		log.u_bbr.pacing_gain = rack->r_must_retran;
2618 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2619 		    &rack->rc_inp->inp_socket->so_rcv,
2620 		    &rack->rc_inp->inp_socket->so_snd,
2621 		    BBR_LOG_TIMERCANC, 0,
2622 		    0, &log, false, tv);
2623 	}
2624 }
2625 
2626 static void
2627 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
2628 			  uint32_t flex1, uint32_t flex2,
2629 			  uint32_t flex3, uint32_t flex4,
2630 			  uint32_t flex5, uint32_t flex6,
2631 			  uint16_t flex7, uint8_t mod)
2632 {
2633 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2634 		union tcp_log_stackspecific log;
2635 		struct timeval tv;
2636 
2637 		if (mod == 1) {
2638 			/* No you can't use 1, its for the real to cancel */
2639 			return;
2640 		}
2641 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2642 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2643 		log.u_bbr.flex1 = flex1;
2644 		log.u_bbr.flex2 = flex2;
2645 		log.u_bbr.flex3 = flex3;
2646 		log.u_bbr.flex4 = flex4;
2647 		log.u_bbr.flex5 = flex5;
2648 		log.u_bbr.flex6 = flex6;
2649 		log.u_bbr.flex7 = flex7;
2650 		log.u_bbr.flex8 = mod;
2651 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2652 		    &rack->rc_inp->inp_socket->so_rcv,
2653 		    &rack->rc_inp->inp_socket->so_snd,
2654 		    BBR_LOG_TIMERCANC, 0,
2655 		    0, &log, false, &tv);
2656 	}
2657 }
2658 
2659 static void
2660 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
2661 {
2662 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2663 		union tcp_log_stackspecific log;
2664 		struct timeval tv;
2665 
2666 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2667 		log.u_bbr.flex1 = timers;
2668 		log.u_bbr.flex2 = ret;
2669 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
2670 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
2671 		log.u_bbr.flex5 = cts;
2672 		if (rack->rack_no_prr)
2673 			log.u_bbr.flex6 = 0;
2674 		else
2675 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
2676 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2677 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2678 		log.u_bbr.pacing_gain = rack->r_must_retran;
2679 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2680 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2681 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2682 		    &rack->rc_inp->inp_socket->so_rcv,
2683 		    &rack->rc_inp->inp_socket->so_snd,
2684 		    BBR_LOG_TO_PROCESS, 0,
2685 		    0, &log, false, &tv);
2686 	}
2687 }
2688 
2689 static void
2690 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
2691 {
2692 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2693 		union tcp_log_stackspecific log;
2694 		struct timeval tv;
2695 
2696 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2697 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
2698 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
2699 		if (rack->rack_no_prr)
2700 			log.u_bbr.flex3 = 0;
2701 		else
2702 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
2703 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
2704 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
2705 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
2706 		log.u_bbr.flex7 = line;
2707 		log.u_bbr.flex8 = frm;
2708 		log.u_bbr.pkts_out = orig_cwnd;
2709 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2710 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2711 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
2712 		log.u_bbr.use_lt_bw <<= 1;
2713 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
2714 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2715 		    &rack->rc_inp->inp_socket->so_rcv,
2716 		    &rack->rc_inp->inp_socket->so_snd,
2717 		    BBR_LOG_BBRUPD, 0,
2718 		    0, &log, false, &tv);
2719 	}
2720 }
2721 
2722 #ifdef NETFLIX_EXP_DETECTION
2723 static void
2724 rack_log_sad(struct tcp_rack *rack, int event)
2725 {
2726 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
2727 		union tcp_log_stackspecific log;
2728 		struct timeval tv;
2729 
2730 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2731 		log.u_bbr.flex1 = rack->r_ctl.sack_count;
2732 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2733 		log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
2734 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2735 		log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
2736 		log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
2737 		log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
2738 		log.u_bbr.lt_epoch = (tcp_force_detection << 8);
2739 		log.u_bbr.lt_epoch |= rack->do_detection;
2740 		log.u_bbr.applimited = tcp_map_minimum;
2741 		log.u_bbr.flex7 = rack->sack_attack_disable;
2742 		log.u_bbr.flex8 = event;
2743 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2744 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2745 		log.u_bbr.delivered = tcp_sad_decay_val;
2746 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2747 		    &rack->rc_inp->inp_socket->so_rcv,
2748 		    &rack->rc_inp->inp_socket->so_snd,
2749 		    TCP_SAD_DETECTION, 0,
2750 		    0, &log, false, &tv);
2751 	}
2752 }
2753 #endif
2754 
2755 static void
2756 rack_counter_destroy(void)
2757 {
2758 	counter_u64_free(rack_fto_send);
2759 	counter_u64_free(rack_fto_rsm_send);
2760 	counter_u64_free(rack_nfto_resend);
2761 	counter_u64_free(rack_hw_pace_init_fail);
2762 	counter_u64_free(rack_hw_pace_lost);
2763 	counter_u64_free(rack_non_fto_send);
2764 	counter_u64_free(rack_extended_rfo);
2765 	counter_u64_free(rack_ack_total);
2766 	counter_u64_free(rack_express_sack);
2767 	counter_u64_free(rack_sack_total);
2768 	counter_u64_free(rack_move_none);
2769 	counter_u64_free(rack_move_some);
2770 	counter_u64_free(rack_sack_attacks_detected);
2771 	counter_u64_free(rack_sack_attacks_reversed);
2772 	counter_u64_free(rack_sack_used_next_merge);
2773 	counter_u64_free(rack_sack_used_prev_merge);
2774 	counter_u64_free(rack_tlp_tot);
2775 	counter_u64_free(rack_tlp_newdata);
2776 	counter_u64_free(rack_tlp_retran);
2777 	counter_u64_free(rack_tlp_retran_bytes);
2778 	counter_u64_free(rack_to_tot);
2779 	counter_u64_free(rack_saw_enobuf);
2780 	counter_u64_free(rack_saw_enobuf_hw);
2781 	counter_u64_free(rack_saw_enetunreach);
2782 	counter_u64_free(rack_hot_alloc);
2783 	counter_u64_free(rack_to_alloc);
2784 	counter_u64_free(rack_to_alloc_hard);
2785 	counter_u64_free(rack_to_alloc_emerg);
2786 	counter_u64_free(rack_to_alloc_limited);
2787 	counter_u64_free(rack_alloc_limited_conns);
2788 	counter_u64_free(rack_split_limited);
2789 	counter_u64_free(rack_multi_single_eq);
2790 	counter_u64_free(rack_proc_non_comp_ack);
2791 	counter_u64_free(rack_sack_proc_all);
2792 	counter_u64_free(rack_sack_proc_restart);
2793 	counter_u64_free(rack_sack_proc_short);
2794 	counter_u64_free(rack_sack_skipped_acked);
2795 	counter_u64_free(rack_sack_splits);
2796 	counter_u64_free(rack_input_idle_reduces);
2797 	counter_u64_free(rack_collapsed_win);
2798 	counter_u64_free(rack_collapsed_win_rxt);
2799 	counter_u64_free(rack_collapsed_win_rxt_bytes);
2800 	counter_u64_free(rack_collapsed_win_seen);
2801 	counter_u64_free(rack_try_scwnd);
2802 	counter_u64_free(rack_persists_sends);
2803 	counter_u64_free(rack_persists_acks);
2804 	counter_u64_free(rack_persists_loss);
2805 	counter_u64_free(rack_persists_lost_ends);
2806 #ifdef INVARIANTS
2807 	counter_u64_free(rack_adjust_map_bw);
2808 #endif
2809 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
2810 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
2811 }
2812 
2813 static struct rack_sendmap *
2814 rack_alloc(struct tcp_rack *rack)
2815 {
2816 	struct rack_sendmap *rsm;
2817 
2818 	/*
2819 	 * First get the top of the list it in
2820 	 * theory is the "hottest" rsm we have,
2821 	 * possibly just freed by ack processing.
2822 	 */
2823 	if (rack->rc_free_cnt > rack_free_cache) {
2824 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2825 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2826 		counter_u64_add(rack_hot_alloc, 1);
2827 		rack->rc_free_cnt--;
2828 		return (rsm);
2829 	}
2830 	/*
2831 	 * Once we get under our free cache we probably
2832 	 * no longer have a "hot" one available. Lets
2833 	 * get one from UMA.
2834 	 */
2835 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
2836 	if (rsm) {
2837 		rack->r_ctl.rc_num_maps_alloced++;
2838 		counter_u64_add(rack_to_alloc, 1);
2839 		return (rsm);
2840 	}
2841 	/*
2842 	 * Dig in to our aux rsm's (the last two) since
2843 	 * UMA failed to get us one.
2844 	 */
2845 	if (rack->rc_free_cnt) {
2846 		counter_u64_add(rack_to_alloc_emerg, 1);
2847 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
2848 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2849 		rack->rc_free_cnt--;
2850 		return (rsm);
2851 	}
2852 	return (NULL);
2853 }
2854 
2855 static struct rack_sendmap *
2856 rack_alloc_full_limit(struct tcp_rack *rack)
2857 {
2858 	if ((V_tcp_map_entries_limit > 0) &&
2859 	    (rack->do_detection == 0) &&
2860 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
2861 		counter_u64_add(rack_to_alloc_limited, 1);
2862 		if (!rack->alloc_limit_reported) {
2863 			rack->alloc_limit_reported = 1;
2864 			counter_u64_add(rack_alloc_limited_conns, 1);
2865 		}
2866 		return (NULL);
2867 	}
2868 	return (rack_alloc(rack));
2869 }
2870 
2871 /* wrapper to allocate a sendmap entry, subject to a specific limit */
2872 static struct rack_sendmap *
2873 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
2874 {
2875 	struct rack_sendmap *rsm;
2876 
2877 	if (limit_type) {
2878 		/* currently there is only one limit type */
2879 		if (V_tcp_map_split_limit > 0 &&
2880 		    (rack->do_detection == 0) &&
2881 		    rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) {
2882 			counter_u64_add(rack_split_limited, 1);
2883 			if (!rack->alloc_limit_reported) {
2884 				rack->alloc_limit_reported = 1;
2885 				counter_u64_add(rack_alloc_limited_conns, 1);
2886 			}
2887 			return (NULL);
2888 		}
2889 	}
2890 
2891 	/* allocate and mark in the limit type, if set */
2892 	rsm = rack_alloc(rack);
2893 	if (rsm != NULL && limit_type) {
2894 		rsm->r_limit_type = limit_type;
2895 		rack->r_ctl.rc_num_split_allocs++;
2896 	}
2897 	return (rsm);
2898 }
2899 
2900 static void
2901 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
2902 {
2903 	if (rsm->r_flags & RACK_APP_LIMITED) {
2904 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
2905 			rack->r_ctl.rc_app_limited_cnt--;
2906 		}
2907 	}
2908 	if (rsm->r_limit_type) {
2909 		/* currently there is only one limit type */
2910 		rack->r_ctl.rc_num_split_allocs--;
2911 	}
2912 	if (rsm == rack->r_ctl.rc_first_appl) {
2913 		if (rack->r_ctl.rc_app_limited_cnt == 0)
2914 			rack->r_ctl.rc_first_appl = NULL;
2915 		else {
2916 			/* Follow the next one out */
2917 			struct rack_sendmap fe;
2918 
2919 			fe.r_start = rsm->r_nseq_appl;
2920 			rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
2921 		}
2922 	}
2923 	if (rsm == rack->r_ctl.rc_resend)
2924 		rack->r_ctl.rc_resend = NULL;
2925 	if (rsm == rack->r_ctl.rc_end_appl)
2926 		rack->r_ctl.rc_end_appl = NULL;
2927 	if (rack->r_ctl.rc_tlpsend == rsm)
2928 		rack->r_ctl.rc_tlpsend = NULL;
2929 	if (rack->r_ctl.rc_sacklast == rsm)
2930 		rack->r_ctl.rc_sacklast = NULL;
2931 	memset(rsm, 0, sizeof(struct rack_sendmap));
2932 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
2933 	rack->rc_free_cnt++;
2934 }
2935 
2936 static void
2937 rack_free_trim(struct tcp_rack *rack)
2938 {
2939 	struct rack_sendmap *rsm;
2940 
2941 	/*
2942 	 * Free up all the tail entries until
2943 	 * we get our list down to the limit.
2944 	 */
2945 	while (rack->rc_free_cnt > rack_free_cache) {
2946 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
2947 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
2948 		rack->rc_free_cnt--;
2949 		uma_zfree(rack_zone, rsm);
2950 	}
2951 }
2952 
2953 
2954 static uint32_t
2955 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
2956 {
2957 	uint64_t srtt, bw, len, tim;
2958 	uint32_t segsiz, def_len, minl;
2959 
2960 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
2961 	def_len = rack_def_data_window * segsiz;
2962 	if (rack->rc_gp_filled == 0) {
2963 		/*
2964 		 * We have no measurement (IW is in flight?) so
2965 		 * we can only guess using our data_window sysctl
2966 		 * value (usually 20MSS).
2967 		 */
2968 		return (def_len);
2969 	}
2970 	/*
2971 	 * Now we have a number of factors to consider.
2972 	 *
2973 	 * 1) We have a desired BDP which is usually
2974 	 *    at least 2.
2975 	 * 2) We have a minimum number of rtt's usually 1 SRTT
2976 	 *    but we allow it too to be more.
2977 	 * 3) We want to make sure a measurement last N useconds (if
2978 	 *    we have set rack_min_measure_usec.
2979 	 *
2980 	 * We handle the first concern here by trying to create a data
2981 	 * window of max(rack_def_data_window, DesiredBDP). The
2982 	 * second concern we handle in not letting the measurement
2983 	 * window end normally until at least the required SRTT's
2984 	 * have gone by which is done further below in
2985 	 * rack_enough_for_measurement(). Finally the third concern
2986 	 * we also handle here by calculating how long that time
2987 	 * would take at the current BW and then return the
2988 	 * max of our first calculation and that length. Note
2989 	 * that if rack_min_measure_usec is 0, we don't deal
2990 	 * with concern 3. Also for both Concern 1 and 3 an
2991 	 * application limited period could end the measurement
2992 	 * earlier.
2993 	 *
2994 	 * So lets calculate the BDP with the "known" b/w using
2995 	 * the SRTT has our rtt and then multiply it by the
2996 	 * goal.
2997 	 */
2998 	bw = rack_get_bw(rack);
2999 	srtt = (uint64_t)tp->t_srtt;
3000 	len = bw * srtt;
3001 	len /= (uint64_t)HPTS_USEC_IN_SEC;
3002 	len *= max(1, rack_goal_bdp);
3003 	/* Now we need to round up to the nearest MSS */
3004 	len = roundup(len, segsiz);
3005 	if (rack_min_measure_usec) {
3006 		/* Now calculate our min length for this b/w */
3007 		tim = rack_min_measure_usec;
3008 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3009 		if (minl == 0)
3010 			minl = 1;
3011 		minl = roundup(minl, segsiz);
3012 		if (len < minl)
3013 			len = minl;
3014 	}
3015 	/*
3016 	 * Now if we have a very small window we want
3017 	 * to attempt to get the window that is
3018 	 * as small as possible. This happens on
3019 	 * low b/w connections and we don't want to
3020 	 * span huge numbers of rtt's between measurements.
3021 	 *
3022 	 * We basically include 2 over our "MIN window" so
3023 	 * that the measurement can be shortened (possibly) by
3024 	 * an ack'ed packet.
3025 	 */
3026 	if (len < def_len)
3027 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3028 	else
3029 		return (max((uint32_t)len, def_len));
3030 
3031 }
3032 
3033 static int
3034 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3035 {
3036 	uint32_t tim, srtts, segsiz;
3037 
3038 	/*
3039 	 * Has enough time passed for the GP measurement to be valid?
3040 	 */
3041 	if ((tp->snd_max == tp->snd_una) ||
3042 	    (th_ack == tp->snd_max)){
3043 		/* All is acked */
3044 		*quality = RACK_QUALITY_ALLACKED;
3045 		return (1);
3046 	}
3047 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3048 		/* Not enough bytes yet */
3049 		return (0);
3050 	}
3051 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3052 	if (SEQ_LT(th_ack, tp->gput_ack) &&
3053 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3054 		/* Not enough bytes yet */
3055 		return (0);
3056 	}
3057 	if (rack->r_ctl.rc_first_appl &&
3058 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3059 		/*
3060 		 * We are up to the app limited send point
3061 		 * we have to measure irrespective of the time..
3062 		 */
3063 		*quality = RACK_QUALITY_APPLIMITED;
3064 		return (1);
3065 	}
3066 	/* Now what about time? */
3067 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3068 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3069 	if (tim >= srtts) {
3070 		*quality = RACK_QUALITY_HIGH;
3071 		return (1);
3072 	}
3073 	/* Nope not even a full SRTT has passed */
3074 	return (0);
3075 }
3076 
3077 static void
3078 rack_log_timely(struct tcp_rack *rack,
3079 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3080 		uint64_t up_bnd, int line, uint8_t method)
3081 {
3082 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3083 		union tcp_log_stackspecific log;
3084 		struct timeval tv;
3085 
3086 		memset(&log, 0, sizeof(log));
3087 		log.u_bbr.flex1 = logged;
3088 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3089 		log.u_bbr.flex2 <<= 4;
3090 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3091 		log.u_bbr.flex2 <<= 4;
3092 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3093 		log.u_bbr.flex2 <<= 4;
3094 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3095 		log.u_bbr.flex3 = rack->rc_gp_incr;
3096 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3097 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3098 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3099 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3100 		log.u_bbr.flex8 = method;
3101 		log.u_bbr.cur_del_rate = cur_bw;
3102 		log.u_bbr.delRate = low_bnd;
3103 		log.u_bbr.bw_inuse = up_bnd;
3104 		log.u_bbr.rttProp = rack_get_bw(rack);
3105 		log.u_bbr.pkt_epoch = line;
3106 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3107 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3108 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3109 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3110 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3111 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3112 		log.u_bbr.cwnd_gain <<= 1;
3113 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3114 		log.u_bbr.cwnd_gain <<= 1;
3115 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3116 		log.u_bbr.cwnd_gain <<= 1;
3117 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3118 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3119 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3120 		    &rack->rc_inp->inp_socket->so_rcv,
3121 		    &rack->rc_inp->inp_socket->so_snd,
3122 		    TCP_TIMELY_WORK, 0,
3123 		    0, &log, false, &tv);
3124 	}
3125 }
3126 
3127 static int
3128 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3129 {
3130 	/*
3131 	 * Before we increase we need to know if
3132 	 * the estimate just made was less than
3133 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3134 	 *
3135 	 * If we already are pacing at a fast enough
3136 	 * rate to push us faster there is no sense of
3137 	 * increasing.
3138 	 *
3139 	 * We first caculate our actual pacing rate (ss or ca multiplier
3140 	 * times our cur_bw).
3141 	 *
3142 	 * Then we take the last measured rate and multipy by our
3143 	 * maximum pacing overage to give us a max allowable rate.
3144 	 *
3145 	 * If our act_rate is smaller than our max_allowable rate
3146 	 * then we should increase. Else we should hold steady.
3147 	 *
3148 	 */
3149 	uint64_t act_rate, max_allow_rate;
3150 
3151 	if (rack_timely_no_stopping)
3152 		return (1);
3153 
3154 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3155 		/*
3156 		 * Initial startup case or
3157 		 * everything is acked case.
3158 		 */
3159 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3160 				__LINE__, 9);
3161 		return (1);
3162 	}
3163 	if (mult <= 100) {
3164 		/*
3165 		 * We can always pace at or slightly above our rate.
3166 		 */
3167 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3168 				__LINE__, 9);
3169 		return (1);
3170 	}
3171 	act_rate = cur_bw * (uint64_t)mult;
3172 	act_rate /= 100;
3173 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3174 	max_allow_rate /= 100;
3175 	if (act_rate < max_allow_rate) {
3176 		/*
3177 		 * Here the rate we are actually pacing at
3178 		 * is smaller than 10% above our last measurement.
3179 		 * This means we are pacing below what we would
3180 		 * like to try to achieve (plus some wiggle room).
3181 		 */
3182 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3183 				__LINE__, 9);
3184 		return (1);
3185 	} else {
3186 		/*
3187 		 * Here we are already pacing at least rack_max_per_above(10%)
3188 		 * what we are getting back. This indicates most likely
3189 		 * that we are being limited (cwnd/rwnd/app) and can't
3190 		 * get any more b/w. There is no sense of trying to
3191 		 * raise up the pacing rate its not speeding us up
3192 		 * and we already are pacing faster than we are getting.
3193 		 */
3194 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3195 				__LINE__, 8);
3196 		return (0);
3197 	}
3198 }
3199 
3200 static void
3201 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3202 {
3203 	/*
3204 	 * When we drag bottom, we want to assure
3205 	 * that no multiplier is below 1.0, if so
3206 	 * we want to restore it to at least that.
3207 	 */
3208 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3209 		/* This is unlikely we usually do not touch recovery */
3210 		rack->r_ctl.rack_per_of_gp_rec = 100;
3211 	}
3212 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3213 		rack->r_ctl.rack_per_of_gp_ca = 100;
3214 	}
3215 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3216 		rack->r_ctl.rack_per_of_gp_ss = 100;
3217 	}
3218 }
3219 
3220 static void
3221 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3222 {
3223 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3224 		rack->r_ctl.rack_per_of_gp_ca = 100;
3225 	}
3226 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3227 		rack->r_ctl.rack_per_of_gp_ss = 100;
3228 	}
3229 }
3230 
3231 static void
3232 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3233 {
3234 	int32_t  calc, logged, plus;
3235 
3236 	logged = 0;
3237 
3238 	if (override) {
3239 		/*
3240 		 * override is passed when we are
3241 		 * loosing b/w and making one last
3242 		 * gasp at trying to not loose out
3243 		 * to a new-reno flow.
3244 		 */
3245 		goto extra_boost;
3246 	}
3247 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3248 	if (rack->rc_gp_incr &&
3249 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3250 		/*
3251 		 * Reset and get 5 strokes more before the boost. Note
3252 		 * that the count is 0 based so we have to add one.
3253 		 */
3254 extra_boost:
3255 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3256 		rack->rc_gp_timely_inc_cnt = 0;
3257 	} else
3258 		plus = (uint32_t)rack_gp_increase_per;
3259 	/* Must be at least 1% increase for true timely increases */
3260 	if ((plus < 1) &&
3261 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3262 		plus = 1;
3263 	if (rack->rc_gp_saw_rec &&
3264 	    (rack->rc_gp_no_rec_chg == 0) &&
3265 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3266 				  rack->r_ctl.rack_per_of_gp_rec)) {
3267 		/* We have been in recovery ding it too */
3268 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3269 		if (calc > 0xffff)
3270 			calc = 0xffff;
3271 		logged |= 1;
3272 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3273 		if (rack_per_upper_bound_ss &&
3274 		    (rack->rc_dragged_bottom == 0) &&
3275 		    (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss))
3276 			rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss;
3277 	}
3278 	if (rack->rc_gp_saw_ca &&
3279 	    (rack->rc_gp_saw_ss == 0) &&
3280 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3281 				  rack->r_ctl.rack_per_of_gp_ca)) {
3282 		/* In CA */
3283 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3284 		if (calc > 0xffff)
3285 			calc = 0xffff;
3286 		logged |= 2;
3287 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3288 		if (rack_per_upper_bound_ca &&
3289 		    (rack->rc_dragged_bottom == 0) &&
3290 		    (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca))
3291 			rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca;
3292 	}
3293 	if (rack->rc_gp_saw_ss &&
3294 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3295 				  rack->r_ctl.rack_per_of_gp_ss)) {
3296 		/* In SS */
3297 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3298 		if (calc > 0xffff)
3299 			calc = 0xffff;
3300 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3301 		if (rack_per_upper_bound_ss &&
3302 		    (rack->rc_dragged_bottom == 0) &&
3303 		    (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss))
3304 			rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss;
3305 		logged |= 4;
3306 	}
3307 	if (logged &&
3308 	    (rack->rc_gp_incr == 0)){
3309 		/* Go into increment mode */
3310 		rack->rc_gp_incr = 1;
3311 		rack->rc_gp_timely_inc_cnt = 0;
3312 	}
3313 	if (rack->rc_gp_incr &&
3314 	    logged &&
3315 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3316 		rack->rc_gp_timely_inc_cnt++;
3317 	}
3318 	rack_log_timely(rack,  logged, plus, 0, 0,
3319 			__LINE__, 1);
3320 }
3321 
3322 static uint32_t
3323 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3324 {
3325 	/*
3326 	 * norm_grad = rtt_diff / minrtt;
3327 	 * new_per = curper * (1 - B * norm_grad)
3328 	 *
3329 	 * B = rack_gp_decrease_per (default 10%)
3330 	 * rtt_dif = input var current rtt-diff
3331 	 * curper = input var current percentage
3332 	 * minrtt = from rack filter
3333 	 *
3334 	 */
3335 	uint64_t perf;
3336 
3337 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3338 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3339 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
3340 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3341 		     (uint64_t)1000000)) /
3342 		(uint64_t)1000000);
3343 	if (perf > curper) {
3344 		/* TSNH */
3345 		perf = curper - 1;
3346 	}
3347 	return ((uint32_t)perf);
3348 }
3349 
3350 static uint32_t
3351 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3352 {
3353 	/*
3354 	 *                                   highrttthresh
3355 	 * result = curper * (1 - (B * ( 1 -  ------          ))
3356 	 *                                     gp_srtt
3357 	 *
3358 	 * B = rack_gp_decrease_per (default 10%)
3359 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3360 	 */
3361 	uint64_t perf;
3362 	uint32_t highrttthresh;
3363 
3364 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3365 
3366 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3367 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3368 					((uint64_t)highrttthresh * (uint64_t)1000000) /
3369 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3370 	return (perf);
3371 }
3372 
3373 static void
3374 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3375 {
3376 	uint64_t logvar, logvar2, logvar3;
3377 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3378 
3379 	if (rack->rc_gp_incr) {
3380 		/* Turn off increment counting */
3381 		rack->rc_gp_incr = 0;
3382 		rack->rc_gp_timely_inc_cnt = 0;
3383 	}
3384 	ss_red = ca_red = rec_red = 0;
3385 	logged = 0;
3386 	/* Calculate the reduction value */
3387 	if (rtt_diff < 0) {
3388 		rtt_diff *= -1;
3389 	}
3390 	/* Must be at least 1% reduction */
3391 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3392 		/* We have been in recovery ding it too */
3393 		if (timely_says == 2) {
3394 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3395 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3396 			if (alt < new_per)
3397 				val = alt;
3398 			else
3399 				val = new_per;
3400 		} else
3401 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3402 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
3403 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
3404 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
3405 		} else {
3406 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3407 			rec_red = 0;
3408 		}
3409 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
3410 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
3411 		logged |= 1;
3412 	}
3413 	if (rack->rc_gp_saw_ss) {
3414 		/* Sent in SS */
3415 		if (timely_says == 2) {
3416 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
3417 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3418 			if (alt < new_per)
3419 				val = alt;
3420 			else
3421 				val = new_per;
3422 		} else
3423 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
3424 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
3425 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
3426 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
3427 		} else {
3428 			ss_red = new_per;
3429 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3430 			logvar = new_per;
3431 			logvar <<= 32;
3432 			logvar |= alt;
3433 			logvar2 = (uint32_t)rtt;
3434 			logvar2 <<= 32;
3435 			logvar2 |= (uint32_t)rtt_diff;
3436 			logvar3 = rack_gp_rtt_maxmul;
3437 			logvar3 <<= 32;
3438 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3439 			rack_log_timely(rack, timely_says,
3440 					logvar2, logvar3,
3441 					logvar, __LINE__, 10);
3442 		}
3443 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
3444 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
3445 		logged |= 4;
3446 	} else if (rack->rc_gp_saw_ca) {
3447 		/* Sent in CA */
3448 		if (timely_says == 2) {
3449 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
3450 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3451 			if (alt < new_per)
3452 				val = alt;
3453 			else
3454 				val = new_per;
3455 		} else
3456 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
3457 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
3458 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
3459 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
3460 		} else {
3461 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3462 			ca_red = 0;
3463 			logvar = new_per;
3464 			logvar <<= 32;
3465 			logvar |= alt;
3466 			logvar2 = (uint32_t)rtt;
3467 			logvar2 <<= 32;
3468 			logvar2 |= (uint32_t)rtt_diff;
3469 			logvar3 = rack_gp_rtt_maxmul;
3470 			logvar3 <<= 32;
3471 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3472 			rack_log_timely(rack, timely_says,
3473 					logvar2, logvar3,
3474 					logvar, __LINE__, 10);
3475 		}
3476 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
3477 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
3478 		logged |= 2;
3479 	}
3480 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
3481 		rack->rc_gp_timely_dec_cnt++;
3482 		if (rack_timely_dec_clear &&
3483 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
3484 			rack->rc_gp_timely_dec_cnt = 0;
3485 	}
3486 	logvar = ss_red;
3487 	logvar <<= 32;
3488 	logvar |= ca_red;
3489 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
3490 			__LINE__, 2);
3491 }
3492 
3493 static void
3494 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
3495 		     uint32_t rtt, uint32_t line, uint8_t reas)
3496 {
3497 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
3498 		union tcp_log_stackspecific log;
3499 		struct timeval tv;
3500 
3501 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3502 		log.u_bbr.flex1 = line;
3503 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
3504 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
3505 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3506 		log.u_bbr.flex5 = rtt;
3507 		log.u_bbr.flex6 = rack->rc_highly_buffered;
3508 		log.u_bbr.flex6 <<= 1;
3509 		log.u_bbr.flex6 |= rack->forced_ack;
3510 		log.u_bbr.flex6 <<= 1;
3511 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
3512 		log.u_bbr.flex6 <<= 1;
3513 		log.u_bbr.flex6 |= rack->in_probe_rtt;
3514 		log.u_bbr.flex6 <<= 1;
3515 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
3516 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
3517 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
3518 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
3519 		log.u_bbr.flex8 = reas;
3520 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3521 		log.u_bbr.delRate = rack_get_bw(rack);
3522 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
3523 		log.u_bbr.cur_del_rate <<= 32;
3524 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
3525 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
3526 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3527 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3528 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3529 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3530 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
3531 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
3532 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3533 		log.u_bbr.rttProp = us_cts;
3534 		log.u_bbr.rttProp <<= 32;
3535 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
3536 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3537 		    &rack->rc_inp->inp_socket->so_rcv,
3538 		    &rack->rc_inp->inp_socket->so_snd,
3539 		    BBR_LOG_RTT_SHRINKS, 0,
3540 		    0, &log, false, &rack->r_ctl.act_rcv_time);
3541 	}
3542 }
3543 
3544 static void
3545 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
3546 {
3547 	uint64_t bwdp;
3548 
3549 	bwdp = rack_get_bw(rack);
3550 	bwdp *= (uint64_t)rtt;
3551 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
3552 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
3553 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
3554 		/*
3555 		 * A window protocol must be able to have 4 packets
3556 		 * outstanding as the floor in order to function
3557 		 * (especially considering delayed ack :D).
3558 		 */
3559 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
3560 	}
3561 }
3562 
3563 static void
3564 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
3565 {
3566 	/**
3567 	 * ProbeRTT is a bit different in rack_pacing than in
3568 	 * BBR. It is like BBR in that it uses the lowering of
3569 	 * the RTT as a signal that we saw something new and
3570 	 * counts from there for how long between. But it is
3571 	 * different in that its quite simple. It does not
3572 	 * play with the cwnd and wait until we get down
3573 	 * to N segments outstanding and hold that for
3574 	 * 200ms. Instead it just sets the pacing reduction
3575 	 * rate to a set percentage (70 by default) and hold
3576 	 * that for a number of recent GP Srtt's.
3577 	 */
3578 	uint32_t segsiz;
3579 
3580 	if (rack->rc_gp_dyn_mul == 0)
3581 		return;
3582 
3583 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
3584 		/* We are idle */
3585 		return;
3586 	}
3587 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3588 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3589 		/*
3590 		 * Stop the goodput now, the idea here is
3591 		 * that future measurements with in_probe_rtt
3592 		 * won't register if they are not greater so
3593 		 * we want to get what info (if any) is available
3594 		 * now.
3595 		 */
3596 		rack_do_goodput_measurement(rack->rc_tp, rack,
3597 					    rack->rc_tp->snd_una, __LINE__,
3598 					    RACK_QUALITY_PROBERTT);
3599 	}
3600 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3601 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3602 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3603 		     rack->r_ctl.rc_pace_min_segs);
3604 	rack->in_probe_rtt = 1;
3605 	rack->measure_saw_probe_rtt = 1;
3606 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3607 	rack->r_ctl.rc_time_probertt_starts = 0;
3608 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
3609 	if (rack_probertt_use_min_rtt_entry)
3610 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3611 	else
3612 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
3613 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3614 			     __LINE__, RACK_RTTS_ENTERPROBE);
3615 }
3616 
3617 static void
3618 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
3619 {
3620 	struct rack_sendmap *rsm;
3621 	uint32_t segsiz;
3622 
3623 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
3624 		     rack->r_ctl.rc_pace_min_segs);
3625 	rack->in_probe_rtt = 0;
3626 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
3627 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
3628 		/*
3629 		 * Stop the goodput now, the idea here is
3630 		 * that future measurements with in_probe_rtt
3631 		 * won't register if they are not greater so
3632 		 * we want to get what info (if any) is available
3633 		 * now.
3634 		 */
3635 		rack_do_goodput_measurement(rack->rc_tp, rack,
3636 					    rack->rc_tp->snd_una, __LINE__,
3637 					    RACK_QUALITY_PROBERTT);
3638 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
3639 		/*
3640 		 * We don't have enough data to make a measurement.
3641 		 * So lets just stop and start here after exiting
3642 		 * probe-rtt. We probably are not interested in
3643 		 * the results anyway.
3644 		 */
3645 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
3646 	}
3647 	/*
3648 	 * Measurements through the current snd_max are going
3649 	 * to be limited by the slower pacing rate.
3650 	 *
3651 	 * We need to mark these as app-limited so we
3652 	 * don't collapse the b/w.
3653 	 */
3654 	rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
3655 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
3656 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3657 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
3658 		else {
3659 			/*
3660 			 * Go out to the end app limited and mark
3661 			 * this new one as next and move the end_appl up
3662 			 * to this guy.
3663 			 */
3664 			if (rack->r_ctl.rc_end_appl)
3665 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
3666 			rack->r_ctl.rc_end_appl = rsm;
3667 		}
3668 		rsm->r_flags |= RACK_APP_LIMITED;
3669 		rack->r_ctl.rc_app_limited_cnt++;
3670 	}
3671 	/*
3672 	 * Now, we need to examine our pacing rate multipliers.
3673 	 * If its under 100%, we need to kick it back up to
3674 	 * 100%. We also don't let it be over our "max" above
3675 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
3676 	 * Note setting clamp_atexit_prtt to 0 has the effect
3677 	 * of setting CA/SS to 100% always at exit (which is
3678 	 * the default behavior).
3679 	 */
3680 	if (rack_probertt_clear_is) {
3681 		rack->rc_gp_incr = 0;
3682 		rack->rc_gp_bwred = 0;
3683 		rack->rc_gp_timely_inc_cnt = 0;
3684 		rack->rc_gp_timely_dec_cnt = 0;
3685 	}
3686 	/* Do we do any clamping at exit? */
3687 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
3688 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
3689 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
3690 	}
3691 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
3692 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
3693 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
3694 	}
3695 	/*
3696 	 * Lets set rtt_diff to 0, so that we will get a "boost"
3697 	 * after exiting.
3698 	 */
3699 	rack->r_ctl.rc_rtt_diff = 0;
3700 
3701 	/* Clear all flags so we start fresh */
3702 	rack->rc_tp->t_bytes_acked = 0;
3703 	rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND;
3704 	/*
3705 	 * If configured to, set the cwnd and ssthresh to
3706 	 * our targets.
3707 	 */
3708 	if (rack_probe_rtt_sets_cwnd) {
3709 		uint64_t ebdp;
3710 		uint32_t setto;
3711 
3712 		/* Set ssthresh so we get into CA once we hit our target */
3713 		if (rack_probertt_use_min_rtt_exit == 1) {
3714 			/* Set to min rtt */
3715 			rack_set_prtt_target(rack, segsiz,
3716 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
3717 		} else if (rack_probertt_use_min_rtt_exit == 2) {
3718 			/* Set to current gp rtt */
3719 			rack_set_prtt_target(rack, segsiz,
3720 					     rack->r_ctl.rc_gp_srtt);
3721 		} else if (rack_probertt_use_min_rtt_exit == 3) {
3722 			/* Set to entry gp rtt */
3723 			rack_set_prtt_target(rack, segsiz,
3724 					     rack->r_ctl.rc_entry_gp_rtt);
3725 		} else {
3726 			uint64_t sum;
3727 			uint32_t setval;
3728 
3729 			sum = rack->r_ctl.rc_entry_gp_rtt;
3730 			sum *= 10;
3731 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
3732 			if (sum >= 20) {
3733 				/*
3734 				 * A highly buffered path needs
3735 				 * cwnd space for timely to work.
3736 				 * Lets set things up as if
3737 				 * we are heading back here again.
3738 				 */
3739 				setval = rack->r_ctl.rc_entry_gp_rtt;
3740 			} else if (sum >= 15) {
3741 				/*
3742 				 * Lets take the smaller of the
3743 				 * two since we are just somewhat
3744 				 * buffered.
3745 				 */
3746 				setval = rack->r_ctl.rc_gp_srtt;
3747 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
3748 					setval = rack->r_ctl.rc_entry_gp_rtt;
3749 			} else {
3750 				/*
3751 				 * Here we are not highly buffered
3752 				 * and should pick the min we can to
3753 				 * keep from causing loss.
3754 				 */
3755 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
3756 			}
3757 			rack_set_prtt_target(rack, segsiz,
3758 					     setval);
3759 		}
3760 		if (rack_probe_rtt_sets_cwnd > 1) {
3761 			/* There is a percentage here to boost */
3762 			ebdp = rack->r_ctl.rc_target_probertt_flight;
3763 			ebdp *= rack_probe_rtt_sets_cwnd;
3764 			ebdp /= 100;
3765 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
3766 		} else
3767 			setto = rack->r_ctl.rc_target_probertt_flight;
3768 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
3769 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
3770 			/* Enforce a min */
3771 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
3772 		}
3773 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
3774 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
3775 	}
3776 	rack_log_rtt_shrinks(rack,  us_cts,
3777 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3778 			     __LINE__, RACK_RTTS_EXITPROBE);
3779 	/* Clear times last so log has all the info */
3780 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
3781 	rack->r_ctl.rc_time_probertt_entered = us_cts;
3782 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
3783 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
3784 }
3785 
3786 static void
3787 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
3788 {
3789 	/* Check in on probe-rtt */
3790 	if (rack->rc_gp_filled == 0) {
3791 		/* We do not do p-rtt unless we have gp measurements */
3792 		return;
3793 	}
3794 	if (rack->in_probe_rtt) {
3795 		uint64_t no_overflow;
3796 		uint32_t endtime, must_stay;
3797 
3798 		if (rack->r_ctl.rc_went_idle_time &&
3799 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
3800 			/*
3801 			 * We went idle during prtt, just exit now.
3802 			 */
3803 			rack_exit_probertt(rack, us_cts);
3804 		} else if (rack_probe_rtt_safety_val &&
3805 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
3806 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
3807 			/*
3808 			 * Probe RTT safety value triggered!
3809 			 */
3810 			rack_log_rtt_shrinks(rack,  us_cts,
3811 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3812 					     __LINE__, RACK_RTTS_SAFETY);
3813 			rack_exit_probertt(rack, us_cts);
3814 		}
3815 		/* Calculate the max we will wait */
3816 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
3817 		if (rack->rc_highly_buffered)
3818 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
3819 		/* Calculate the min we must wait */
3820 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
3821 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
3822 		    TSTMP_LT(us_cts, endtime)) {
3823 			uint32_t calc;
3824 			/* Do we lower more? */
3825 no_exit:
3826 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
3827 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
3828 			else
3829 				calc = 0;
3830 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
3831 			if (calc) {
3832 				/* Maybe */
3833 				calc *= rack_per_of_gp_probertt_reduce;
3834 				rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
3835 				/* Limit it too */
3836 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
3837 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
3838 			}
3839 			/* We must reach target or the time set */
3840 			return;
3841 		}
3842 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
3843 			if ((TSTMP_LT(us_cts, must_stay) &&
3844 			     rack->rc_highly_buffered) ||
3845 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
3846 			      rack->r_ctl.rc_target_probertt_flight)) {
3847 				/* We are not past the must_stay time */
3848 				goto no_exit;
3849 			}
3850 			rack_log_rtt_shrinks(rack,  us_cts,
3851 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
3852 					     __LINE__, RACK_RTTS_REACHTARGET);
3853 			rack->r_ctl.rc_time_probertt_starts = us_cts;
3854 			if (rack->r_ctl.rc_time_probertt_starts == 0)
3855 				rack->r_ctl.rc_time_probertt_starts = 1;
3856 			/* Restore back to our rate we want to pace at in prtt */
3857 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
3858 		}
3859 		/*
3860 		 * Setup our end time, some number of gp_srtts plus 200ms.
3861 		 */
3862 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
3863 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
3864 		if (rack_probertt_gpsrtt_cnt_div)
3865 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
3866 		else
3867 			endtime = 0;
3868 		endtime += rack_min_probertt_hold;
3869 		endtime += rack->r_ctl.rc_time_probertt_starts;
3870 		if (TSTMP_GEQ(us_cts,  endtime)) {
3871 			/* yes, exit probertt */
3872 			rack_exit_probertt(rack, us_cts);
3873 		}
3874 
3875 	} else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
3876 		/* Go into probertt, its been too long since we went lower */
3877 		rack_enter_probertt(rack, us_cts);
3878 	}
3879 }
3880 
3881 static void
3882 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
3883 		       uint32_t rtt, int32_t rtt_diff)
3884 {
3885 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
3886 	uint32_t losses;
3887 
3888 	if ((rack->rc_gp_dyn_mul == 0) ||
3889 	    (rack->use_fixed_rate) ||
3890 	    (rack->in_probe_rtt) ||
3891 	    (rack->rc_always_pace == 0)) {
3892 		/* No dynamic GP multiplier in play */
3893 		return;
3894 	}
3895 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
3896 	cur_bw = rack_get_bw(rack);
3897 	/* Calculate our up and down range */
3898 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
3899 	up_bnd /= 100;
3900 	up_bnd += rack->r_ctl.last_gp_comp_bw;
3901 
3902 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
3903 	subfr /= 100;
3904 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
3905 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
3906 		/*
3907 		 * This is the case where our RTT is above
3908 		 * the max target and we have been configured
3909 		 * to just do timely no bonus up stuff in that case.
3910 		 *
3911 		 * There are two configurations, set to 1, and we
3912 		 * just do timely if we are over our max. If its
3913 		 * set above 1 then we slam the multipliers down
3914 		 * to 100 and then decrement per timely.
3915 		 */
3916 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3917 				__LINE__, 3);
3918 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
3919 			rack_validate_multipliers_at_or_below_100(rack);
3920 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
3921 	} else if ((last_bw_est < low_bnd) && !losses) {
3922 		/*
3923 		 * We are decreasing this is a bit complicated this
3924 		 * means we are loosing ground. This could be
3925 		 * because another flow entered and we are competing
3926 		 * for b/w with it. This will push the RTT up which
3927 		 * makes timely unusable unless we want to get shoved
3928 		 * into a corner and just be backed off (the age
3929 		 * old problem with delay based CC).
3930 		 *
3931 		 * On the other hand if it was a route change we
3932 		 * would like to stay somewhat contained and not
3933 		 * blow out the buffers.
3934 		 */
3935 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3936 				__LINE__, 3);
3937 		rack->r_ctl.last_gp_comp_bw = cur_bw;
3938 		if (rack->rc_gp_bwred == 0) {
3939 			/* Go into reduction counting */
3940 			rack->rc_gp_bwred = 1;
3941 			rack->rc_gp_timely_dec_cnt = 0;
3942 		}
3943 		if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) ||
3944 		    (timely_says == 0)) {
3945 			/*
3946 			 * Push another time with a faster pacing
3947 			 * to try to gain back (we include override to
3948 			 * get a full raise factor).
3949 			 */
3950 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
3951 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
3952 			    (timely_says == 0) ||
3953 			    (rack_down_raise_thresh == 0)) {
3954 				/*
3955 				 * Do an override up in b/w if we were
3956 				 * below the threshold or if the threshold
3957 				 * is zero we always do the raise.
3958 				 */
3959 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
3960 			} else {
3961 				/* Log it stays the same */
3962 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
3963 						__LINE__, 11);
3964 			}
3965 			rack->rc_gp_timely_dec_cnt++;
3966 			/* We are not incrementing really no-count */
3967 			rack->rc_gp_incr = 0;
3968 			rack->rc_gp_timely_inc_cnt = 0;
3969 		} else {
3970 			/*
3971 			 * Lets just use the RTT
3972 			 * information and give up
3973 			 * pushing.
3974 			 */
3975 			goto use_timely;
3976 		}
3977 	} else if ((timely_says != 2) &&
3978 		    !losses &&
3979 		    (last_bw_est > up_bnd)) {
3980 		/*
3981 		 * We are increasing b/w lets keep going, updating
3982 		 * our b/w and ignoring any timely input, unless
3983 		 * of course we are at our max raise (if there is one).
3984 		 */
3985 
3986 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
3987 				__LINE__, 3);
3988 		rack->r_ctl.last_gp_comp_bw = cur_bw;
3989 		if (rack->rc_gp_saw_ss &&
3990 		    rack_per_upper_bound_ss &&
3991 		     (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) {
3992 			    /*
3993 			     * In cases where we can't go higher
3994 			     * we should just use timely.
3995 			     */
3996 			    goto use_timely;
3997 		}
3998 		if (rack->rc_gp_saw_ca &&
3999 		    rack_per_upper_bound_ca &&
4000 		    (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) {
4001 			    /*
4002 			     * In cases where we can't go higher
4003 			     * we should just use timely.
4004 			     */
4005 			    goto use_timely;
4006 		}
4007 		rack->rc_gp_bwred = 0;
4008 		rack->rc_gp_timely_dec_cnt = 0;
4009 		/* You get a set number of pushes if timely is trying to reduce */
4010 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4011 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4012 		} else {
4013 			/* Log it stays the same */
4014 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4015 			    __LINE__, 12);
4016 		}
4017 		return;
4018 	} else {
4019 		/*
4020 		 * We are staying between the lower and upper range bounds
4021 		 * so use timely to decide.
4022 		 */
4023 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4024 				__LINE__, 3);
4025 use_timely:
4026 		if (timely_says) {
4027 			rack->rc_gp_incr = 0;
4028 			rack->rc_gp_timely_inc_cnt = 0;
4029 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4030 			    !losses &&
4031 			    (last_bw_est < low_bnd)) {
4032 				/* We are loosing ground */
4033 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4034 				rack->rc_gp_timely_dec_cnt++;
4035 				/* We are not incrementing really no-count */
4036 				rack->rc_gp_incr = 0;
4037 				rack->rc_gp_timely_inc_cnt = 0;
4038 			} else
4039 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4040 		} else {
4041 			rack->rc_gp_bwred = 0;
4042 			rack->rc_gp_timely_dec_cnt = 0;
4043 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4044 		}
4045 	}
4046 }
4047 
4048 static int32_t
4049 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4050 {
4051 	int32_t timely_says;
4052 	uint64_t log_mult, log_rtt_a_diff;
4053 
4054 	log_rtt_a_diff = rtt;
4055 	log_rtt_a_diff <<= 32;
4056 	log_rtt_a_diff |= (uint32_t)rtt_diff;
4057 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4058 		    rack_gp_rtt_maxmul)) {
4059 		/* Reduce the b/w multiplier */
4060 		timely_says = 2;
4061 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4062 		log_mult <<= 32;
4063 		log_mult |= prev_rtt;
4064 		rack_log_timely(rack,  timely_says, log_mult,
4065 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4066 				log_rtt_a_diff, __LINE__, 4);
4067 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4068 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4069 			    max(rack_gp_rtt_mindiv , 1)))) {
4070 		/* Increase the b/w multiplier */
4071 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4072 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4073 			 max(rack_gp_rtt_mindiv , 1));
4074 		log_mult <<= 32;
4075 		log_mult |= prev_rtt;
4076 		timely_says = 0;
4077 		rack_log_timely(rack,  timely_says, log_mult ,
4078 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4079 				log_rtt_a_diff, __LINE__, 5);
4080 	} else {
4081 		/*
4082 		 * Use a gradient to find it the timely gradient
4083 		 * is:
4084 		 * grad = rc_rtt_diff / min_rtt;
4085 		 *
4086 		 * anything below or equal to 0 will be
4087 		 * a increase indication. Anything above
4088 		 * zero is a decrease. Note we take care
4089 		 * of the actual gradient calculation
4090 		 * in the reduction (its not needed for
4091 		 * increase).
4092 		 */
4093 		log_mult = prev_rtt;
4094 		if (rtt_diff <= 0) {
4095 			/*
4096 			 * Rttdiff is less than zero, increase the
4097 			 * b/w multiplier (its 0 or negative)
4098 			 */
4099 			timely_says = 0;
4100 			rack_log_timely(rack,  timely_says, log_mult,
4101 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4102 		} else {
4103 			/* Reduce the b/w multiplier */
4104 			timely_says = 1;
4105 			rack_log_timely(rack,  timely_says, log_mult,
4106 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4107 		}
4108 	}
4109 	return (timely_says);
4110 }
4111 
4112 static void
4113 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4114 			    tcp_seq th_ack, int line, uint8_t quality)
4115 {
4116 	uint64_t tim, bytes_ps, ltim, stim, utim;
4117 	uint32_t segsiz, bytes, reqbytes, us_cts;
4118 	int32_t gput, new_rtt_diff, timely_says;
4119 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4120 	int did_add = 0;
4121 
4122 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4123 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4124 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
4125 		tim = us_cts - tp->gput_ts;
4126 	else
4127 		tim = 0;
4128 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4129 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4130 	else
4131 		stim = 0;
4132 	/*
4133 	 * Use the larger of the send time or ack time. This prevents us
4134 	 * from being influenced by ack artifacts to come up with too
4135 	 * high of measurement. Note that since we are spanning over many more
4136 	 * bytes in most of our measurements hopefully that is less likely to
4137 	 * occur.
4138 	 */
4139 	if (tim > stim)
4140 		utim = max(tim, 1);
4141 	else
4142 		utim = max(stim, 1);
4143 	/* Lets get a msec time ltim too for the old stuff */
4144 	ltim = max(1, (utim / HPTS_USEC_IN_MSEC));
4145 	gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim;
4146 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4147 	if ((tim == 0) && (stim == 0)) {
4148 		/*
4149 		 * Invalid measurement time, maybe
4150 		 * all on one ack/one send?
4151 		 */
4152 		bytes = 0;
4153 		bytes_ps = 0;
4154 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4155 					   0, 0, 0, 10, __LINE__, NULL, quality);
4156 		goto skip_measurement;
4157 	}
4158 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4159 		/* We never made a us_rtt measurement? */
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 	}
4166 	/*
4167 	 * Calculate the maximum possible b/w this connection
4168 	 * could have. We base our calculation on the lowest
4169 	 * rtt we have seen during the measurement and the
4170 	 * largest rwnd the client has given us in that time. This
4171 	 * forms a BDP that is the maximum that we could ever
4172 	 * get to the client. Anything larger is not valid.
4173 	 *
4174 	 * I originally had code here that rejected measurements
4175 	 * where the time was less than 1/2 the latest us_rtt.
4176 	 * But after thinking on that I realized its wrong since
4177 	 * say you had a 150Mbps or even 1Gbps link, and you
4178 	 * were a long way away.. example I am in Europe (100ms rtt)
4179 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4180 	 * bytes my time would be 1.2ms, and yet my rtt would say
4181 	 * the measurement was invalid the time was < 50ms. The
4182 	 * same thing is true for 150Mb (8ms of time).
4183 	 *
4184 	 * A better way I realized is to look at what the maximum
4185 	 * the connection could possibly do. This is gated on
4186 	 * the lowest RTT we have seen and the highest rwnd.
4187 	 * We should in theory never exceed that, if we are
4188 	 * then something on the path is storing up packets
4189 	 * and then feeding them all at once to our endpoint
4190 	 * messing up our measurement.
4191 	 */
4192 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4193 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4194 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4195 	if (SEQ_LT(th_ack, tp->gput_seq)) {
4196 		/* No measurement can be made */
4197 		bytes = 0;
4198 		bytes_ps = 0;
4199 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4200 					   0, 0, 0, 10, __LINE__, NULL, quality);
4201 		goto skip_measurement;
4202 	} else
4203 		bytes = (th_ack - tp->gput_seq);
4204 	bytes_ps = (uint64_t)bytes;
4205 	/*
4206 	 * Don't measure a b/w for pacing unless we have gotten at least
4207 	 * an initial windows worth of data in this measurement interval.
4208 	 *
4209 	 * Small numbers of bytes get badly influenced by delayed ack and
4210 	 * other artifacts. Note we take the initial window or our
4211 	 * defined minimum GP (defaulting to 10 which hopefully is the
4212 	 * IW).
4213 	 */
4214 	if (rack->rc_gp_filled == 0) {
4215 		/*
4216 		 * The initial estimate is special. We
4217 		 * have blasted out an IW worth of packets
4218 		 * without a real valid ack ts results. We
4219 		 * then setup the app_limited_needs_set flag,
4220 		 * this should get the first ack in (probably 2
4221 		 * MSS worth) to be recorded as the timestamp.
4222 		 * We thus allow a smaller number of bytes i.e.
4223 		 * IW - 2MSS.
4224 		 */
4225 		reqbytes -= (2 * segsiz);
4226 		/* Also lets fill previous for our first measurement to be neutral */
4227 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4228 	}
4229 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4230 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4231 					   rack->r_ctl.rc_app_limited_cnt,
4232 					   0, 0, 10, __LINE__, NULL, quality);
4233 		goto skip_measurement;
4234 	}
4235 	/*
4236 	 * We now need to calculate the Timely like status so
4237 	 * we can update (possibly) the b/w multipliers.
4238 	 */
4239 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4240 	if (rack->rc_gp_filled == 0) {
4241 		/* No previous reading */
4242 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4243 	} else {
4244 		if (rack->measure_saw_probe_rtt == 0) {
4245 			/*
4246 			 * We don't want a probertt to be counted
4247 			 * since it will be negative incorrectly. We
4248 			 * expect to be reducing the RTT when we
4249 			 * pace at a slower rate.
4250 			 */
4251 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4252 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4253 		}
4254 	}
4255 	timely_says = rack_make_timely_judgement(rack,
4256 		rack->r_ctl.rc_gp_srtt,
4257 		rack->r_ctl.rc_rtt_diff,
4258 	        rack->r_ctl.rc_prev_gp_srtt
4259 		);
4260 	bytes_ps *= HPTS_USEC_IN_SEC;
4261 	bytes_ps /= utim;
4262 	if (bytes_ps > rack->r_ctl.last_max_bw) {
4263 		/*
4264 		 * Something is on path playing
4265 		 * since this b/w is not possible based
4266 		 * on our BDP (highest rwnd and lowest rtt
4267 		 * we saw in the measurement window).
4268 		 *
4269 		 * Another option here would be to
4270 		 * instead skip the measurement.
4271 		 */
4272 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4273 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
4274 					   11, __LINE__, NULL, quality);
4275 		bytes_ps = rack->r_ctl.last_max_bw;
4276 	}
4277 	/* We store gp for b/w in bytes per second */
4278 	if (rack->rc_gp_filled == 0) {
4279 		/* Initial measurement */
4280 		if (bytes_ps) {
4281 			rack->r_ctl.gp_bw = bytes_ps;
4282 			rack->rc_gp_filled = 1;
4283 			rack->r_ctl.num_measurements = 1;
4284 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
4285 		} else {
4286 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4287 						   rack->r_ctl.rc_app_limited_cnt,
4288 						   0, 0, 10, __LINE__, NULL, quality);
4289 		}
4290 		if (tcp_in_hpts(rack->rc_inp) &&
4291 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
4292 			/*
4293 			 * Ok we can't trust the pacer in this case
4294 			 * where we transition from un-paced to paced.
4295 			 * Or for that matter when the burst mitigation
4296 			 * was making a wild guess and got it wrong.
4297 			 * Stop the pacer and clear up all the aggregate
4298 			 * delays etc.
4299 			 */
4300 			tcp_hpts_remove(rack->rc_inp);
4301 			rack->r_ctl.rc_hpts_flags = 0;
4302 			rack->r_ctl.rc_last_output_to = 0;
4303 		}
4304 		did_add = 2;
4305 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
4306 		/* Still a small number run an average */
4307 		rack->r_ctl.gp_bw += bytes_ps;
4308 		addpart = rack->r_ctl.num_measurements;
4309 		rack->r_ctl.num_measurements++;
4310 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
4311 			/* We have collected enough to move forward */
4312 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
4313 		}
4314 		did_add = 3;
4315 	} else {
4316 		/*
4317 		 * We want to take 1/wma of the goodput and add in to 7/8th
4318 		 * of the old value weighted by the srtt. So if your measurement
4319 		 * period is say 2 SRTT's long you would get 1/4 as the
4320 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
4321 		 *
4322 		 * But we must be careful not to take too much i.e. if the
4323 		 * srtt is say 20ms and the measurement is taken over
4324 		 * 400ms our weight would be 400/20 i.e. 20. On the
4325 		 * other hand if we get a measurement over 1ms with a
4326 		 * 10ms rtt we only want to take a much smaller portion.
4327 		 */
4328 		if (rack->r_ctl.num_measurements < 0xff) {
4329 			rack->r_ctl.num_measurements++;
4330 		}
4331 		srtt = (uint64_t)tp->t_srtt;
4332 		if (srtt == 0) {
4333 			/*
4334 			 * Strange why did t_srtt go back to zero?
4335 			 */
4336 			if (rack->r_ctl.rc_rack_min_rtt)
4337 				srtt = rack->r_ctl.rc_rack_min_rtt;
4338 			else
4339 				srtt = HPTS_USEC_IN_MSEC;
4340 		}
4341 		/*
4342 		 * XXXrrs: Note for reviewers, in playing with
4343 		 * dynamic pacing I discovered this GP calculation
4344 		 * as done originally leads to some undesired results.
4345 		 * Basically you can get longer measurements contributing
4346 		 * too much to the WMA. Thus I changed it if you are doing
4347 		 * dynamic adjustments to only do the aportioned adjustment
4348 		 * if we have a very small (time wise) measurement. Longer
4349 		 * measurements just get there weight (defaulting to 1/8)
4350 		 * add to the WMA. We may want to think about changing
4351 		 * this to always do that for both sides i.e. dynamic
4352 		 * and non-dynamic... but considering lots of folks
4353 		 * were playing with this I did not want to change the
4354 		 * calculation per.se. without your thoughts.. Lawerence?
4355 		 * Peter??
4356 		 */
4357 		if (rack->rc_gp_dyn_mul == 0) {
4358 			subpart = rack->r_ctl.gp_bw * utim;
4359 			subpart /= (srtt * 8);
4360 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
4361 				/*
4362 				 * The b/w update takes no more
4363 				 * away then 1/2 our running total
4364 				 * so factor it in.
4365 				 */
4366 				addpart = bytes_ps * utim;
4367 				addpart /= (srtt * 8);
4368 			} else {
4369 				/*
4370 				 * Don't allow a single measurement
4371 				 * to account for more than 1/2 of the
4372 				 * WMA. This could happen on a retransmission
4373 				 * where utim becomes huge compared to
4374 				 * srtt (multiple retransmissions when using
4375 				 * the sending rate which factors in all the
4376 				 * transmissions from the first one).
4377 				 */
4378 				subpart = rack->r_ctl.gp_bw / 2;
4379 				addpart = bytes_ps / 2;
4380 			}
4381 			resid_bw = rack->r_ctl.gp_bw - subpart;
4382 			rack->r_ctl.gp_bw = resid_bw + addpart;
4383 			did_add = 1;
4384 		} else {
4385 			if ((utim / srtt) <= 1) {
4386 				/*
4387 				 * The b/w update was over a small period
4388 				 * of time. The idea here is to prevent a small
4389 				 * measurement time period from counting
4390 				 * too much. So we scale it based on the
4391 				 * time so it attributes less than 1/rack_wma_divisor
4392 				 * of its measurement.
4393 				 */
4394 				subpart = rack->r_ctl.gp_bw * utim;
4395 				subpart /= (srtt * rack_wma_divisor);
4396 				addpart = bytes_ps * utim;
4397 				addpart /= (srtt * rack_wma_divisor);
4398 			} else {
4399 				/*
4400 				 * The scaled measurement was long
4401 				 * enough so lets just add in the
4402 				 * portion of the measurement i.e. 1/rack_wma_divisor
4403 				 */
4404 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
4405 				addpart = bytes_ps / rack_wma_divisor;
4406 			}
4407 			if ((rack->measure_saw_probe_rtt == 0) ||
4408 		            (bytes_ps > rack->r_ctl.gp_bw)) {
4409 				/*
4410 				 * For probe-rtt we only add it in
4411 				 * if its larger, all others we just
4412 				 * add in.
4413 				 */
4414 				did_add = 1;
4415 				resid_bw = rack->r_ctl.gp_bw - subpart;
4416 				rack->r_ctl.gp_bw = resid_bw + addpart;
4417 			}
4418 		}
4419 	}
4420 	if ((rack->gp_ready == 0) &&
4421 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
4422 		/* We have enough measurements now */
4423 		rack->gp_ready = 1;
4424 		rack_set_cc_pacing(rack);
4425 		if (rack->defer_options)
4426 			rack_apply_deferred_options(rack);
4427 	}
4428 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
4429 				   rack_get_bw(rack), 22, did_add, NULL, quality);
4430 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
4431 	if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
4432 		rack_update_multiplier(rack, timely_says, bytes_ps,
4433 				       rack->r_ctl.rc_gp_srtt,
4434 				       rack->r_ctl.rc_rtt_diff);
4435 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
4436 				   rack_get_bw(rack), 3, line, NULL, quality);
4437 	/* reset the gp srtt and setup the new prev */
4438 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4439 	/* Record the lost count for the next measurement */
4440 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
4441 	/*
4442 	 * We restart our diffs based on the gpsrtt in the
4443 	 * measurement window.
4444 	 */
4445 	rack->rc_gp_rtt_set = 0;
4446 	rack->rc_gp_saw_rec = 0;
4447 	rack->rc_gp_saw_ca = 0;
4448 	rack->rc_gp_saw_ss = 0;
4449 	rack->rc_dragged_bottom = 0;
4450 skip_measurement:
4451 
4452 #ifdef STATS
4453 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
4454 				 gput);
4455 	/*
4456 	 * XXXLAS: This is a temporary hack, and should be
4457 	 * chained off VOI_TCP_GPUT when stats(9) grows an
4458 	 * API to deal with chained VOIs.
4459 	 */
4460 	if (tp->t_stats_gput_prev > 0)
4461 		stats_voi_update_abs_s32(tp->t_stats,
4462 					 VOI_TCP_GPUT_ND,
4463 					 ((gput - tp->t_stats_gput_prev) * 100) /
4464 					 tp->t_stats_gput_prev);
4465 #endif
4466 	tp->t_flags &= ~TF_GPUTINPROG;
4467 	tp->t_stats_gput_prev = gput;
4468 	/*
4469 	 * Now are we app limited now and there is space from where we
4470 	 * were to where we want to go?
4471 	 *
4472 	 * We don't do the other case i.e. non-applimited here since
4473 	 * the next send will trigger us picking up the missing data.
4474 	 */
4475 	if (rack->r_ctl.rc_first_appl &&
4476 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
4477 	    rack->r_ctl.rc_app_limited_cnt &&
4478 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
4479 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
4480 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
4481 		/*
4482 		 * Yep there is enough outstanding to make a measurement here.
4483 		 */
4484 		struct rack_sendmap *rsm, fe;
4485 
4486 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
4487 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
4488 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4489 		rack->app_limited_needs_set = 0;
4490 		tp->gput_seq = th_ack;
4491 		if (rack->in_probe_rtt)
4492 			rack->measure_saw_probe_rtt = 1;
4493 		else if ((rack->measure_saw_probe_rtt) &&
4494 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
4495 			rack->measure_saw_probe_rtt = 0;
4496 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
4497 			/* There is a full window to gain info from */
4498 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
4499 		} else {
4500 			/* We can only measure up to the applimited point */
4501 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
4502 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
4503 				/*
4504 				 * We don't have enough to make a measurement.
4505 				 */
4506 				tp->t_flags &= ~TF_GPUTINPROG;
4507 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
4508 							   0, 0, 0, 6, __LINE__, NULL, quality);
4509 				return;
4510 			}
4511 		}
4512 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
4513 			/*
4514 			 * We will get no more data into the SB
4515 			 * this means we need to have the data available
4516 			 * before we start a measurement.
4517 			 */
4518 			if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
4519 				/* Nope not enough data. */
4520 				return;
4521 			}
4522 		}
4523 		tp->t_flags |= TF_GPUTINPROG;
4524 		/*
4525 		 * Now we need to find the timestamp of the send at tp->gput_seq
4526 		 * for the send based measurement.
4527 		 */
4528 		fe.r_start = tp->gput_seq;
4529 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
4530 		if (rsm) {
4531 			/* Ok send-based limit is set */
4532 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
4533 				/*
4534 				 * Move back to include the earlier part
4535 				 * so our ack time lines up right (this may
4536 				 * make an overlapping measurement but thats
4537 				 * ok).
4538 				 */
4539 				tp->gput_seq = rsm->r_start;
4540 			}
4541 			if (rsm->r_flags & RACK_ACKED)
4542 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
4543 			else
4544 				rack->app_limited_needs_set = 1;
4545 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
4546 		} else {
4547 			/*
4548 			 * If we don't find the rsm due to some
4549 			 * send-limit set the current time, which
4550 			 * basically disables the send-limit.
4551 			 */
4552 			struct timeval tv;
4553 
4554 			microuptime(&tv);
4555 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
4556 		}
4557 		rack_log_pacing_delay_calc(rack,
4558 					   tp->gput_seq,
4559 					   tp->gput_ack,
4560 					   (uint64_t)rsm,
4561 					   tp->gput_ts,
4562 					   rack->r_ctl.rc_app_limited_cnt,
4563 					   9,
4564 					   __LINE__, NULL, quality);
4565 	}
4566 }
4567 
4568 /*
4569  * CC wrapper hook functions
4570  */
4571 static void
4572 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
4573     uint16_t type, int32_t recovery)
4574 {
4575 	uint32_t prior_cwnd, acked;
4576 	struct tcp_log_buffer *lgb = NULL;
4577 	uint8_t labc_to_use, quality;
4578 
4579 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4580 	tp->ccv->nsegs = nsegs;
4581 	acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una);
4582 	if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
4583 		uint32_t max;
4584 
4585 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
4586 		if (tp->ccv->bytes_this_ack > max) {
4587 			tp->ccv->bytes_this_ack = max;
4588 		}
4589 	}
4590 #ifdef STATS
4591 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
4592 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
4593 #endif
4594 	quality = RACK_QUALITY_NONE;
4595 	if ((tp->t_flags & TF_GPUTINPROG) &&
4596 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
4597 		/* Measure the Goodput */
4598 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
4599 #ifdef NETFLIX_PEAKRATE
4600 		if ((type == CC_ACK) &&
4601 		    (tp->t_maxpeakrate)) {
4602 			/*
4603 			 * We update t_peakrate_thr. This gives us roughly
4604 			 * one update per round trip time. Note
4605 			 * it will only be used if pace_always is off i.e
4606 			 * we don't do this for paced flows.
4607 			 */
4608 			rack_update_peakrate_thr(tp);
4609 		}
4610 #endif
4611 	}
4612 	/* Which way our we limited, if not cwnd limited no advance in CA */
4613 	if (tp->snd_cwnd <= tp->snd_wnd)
4614 		tp->ccv->flags |= CCF_CWND_LIMITED;
4615 	else
4616 		tp->ccv->flags &= ~CCF_CWND_LIMITED;
4617 	if (tp->snd_cwnd > tp->snd_ssthresh) {
4618 		tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
4619 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
4620 		/* For the setting of a window past use the actual scwnd we are using */
4621 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
4622 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
4623 			tp->ccv->flags |= CCF_ABC_SENTAWND;
4624 		}
4625 	} else {
4626 		tp->ccv->flags &= ~CCF_ABC_SENTAWND;
4627 		tp->t_bytes_acked = 0;
4628 	}
4629 	prior_cwnd = tp->snd_cwnd;
4630 	if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
4631 	    (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf)))
4632 		labc_to_use = rack->rc_labc;
4633 	else
4634 		labc_to_use = rack_max_abc_post_recovery;
4635 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4636 		union tcp_log_stackspecific log;
4637 		struct timeval tv;
4638 
4639 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4640 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4641 		log.u_bbr.flex1 = th_ack;
4642 		log.u_bbr.flex2 = tp->ccv->flags;
4643 		log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4644 		log.u_bbr.flex4 = tp->ccv->nsegs;
4645 		log.u_bbr.flex5 = labc_to_use;
4646 		log.u_bbr.flex6 = prior_cwnd;
4647 		log.u_bbr.flex7 = V_tcp_do_newsack;
4648 		log.u_bbr.flex8 = 1;
4649 		lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4650 				     0, &log, false, NULL, NULL, 0, &tv);
4651 	}
4652 	if (CC_ALGO(tp)->ack_received != NULL) {
4653 		/* XXXLAS: Find a way to live without this */
4654 		tp->ccv->curack = th_ack;
4655 		tp->ccv->labc = labc_to_use;
4656 		tp->ccv->flags |= CCF_USE_LOCAL_ABC;
4657 		CC_ALGO(tp)->ack_received(tp->ccv, type);
4658 	}
4659 	if (lgb) {
4660 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
4661 	}
4662 	if (rack->r_must_retran) {
4663 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
4664 			/*
4665 			 * We now are beyond the rxt point so lets disable
4666 			 * the flag.
4667 			 */
4668 			rack->r_ctl.rc_out_at_rto = 0;
4669 			rack->r_must_retran = 0;
4670 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
4671 			/*
4672 			 * Only decrement the rc_out_at_rto if the cwnd advances
4673 			 * at least a whole segment. Otherwise next time the peer
4674 			 * acks, we won't be able to send this generaly happens
4675 			 * when we are in Congestion Avoidance.
4676 			 */
4677 			if (acked <= rack->r_ctl.rc_out_at_rto){
4678 				rack->r_ctl.rc_out_at_rto -= acked;
4679 			} else {
4680 				rack->r_ctl.rc_out_at_rto = 0;
4681 			}
4682 		}
4683 	}
4684 #ifdef STATS
4685 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
4686 #endif
4687 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
4688 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
4689 	}
4690 #ifdef NETFLIX_PEAKRATE
4691 	/* we enforce max peak rate if it is set and we are not pacing */
4692 	if ((rack->rc_always_pace == 0) &&
4693 	    tp->t_peakrate_thr &&
4694 	    (tp->snd_cwnd > tp->t_peakrate_thr)) {
4695 		tp->snd_cwnd = tp->t_peakrate_thr;
4696 	}
4697 #endif
4698 }
4699 
4700 static void
4701 tcp_rack_partialack(struct tcpcb *tp)
4702 {
4703 	struct tcp_rack *rack;
4704 
4705 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4706 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4707 	/*
4708 	 * If we are doing PRR and have enough
4709 	 * room to send <or> we are pacing and prr
4710 	 * is disabled we will want to see if we
4711 	 * can send data (by setting r_wanted_output to
4712 	 * true).
4713 	 */
4714 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
4715 	    rack->rack_no_prr)
4716 		rack->r_wanted_output = 1;
4717 }
4718 
4719 static void
4720 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
4721 {
4722 	struct tcp_rack *rack;
4723 	uint32_t orig_cwnd;
4724 
4725 	orig_cwnd = tp->snd_cwnd;
4726 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4727 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4728 	/* only alert CC if we alerted when we entered */
4729 	if (CC_ALGO(tp)->post_recovery != NULL) {
4730 		tp->ccv->curack = th_ack;
4731 		CC_ALGO(tp)->post_recovery(tp->ccv);
4732 		if (tp->snd_cwnd < tp->snd_ssthresh) {
4733 			/*
4734 			 * Rack has burst control and pacing
4735 			 * so lets not set this any lower than
4736 			 * snd_ssthresh per RFC-6582 (option 2).
4737 			 */
4738 			tp->snd_cwnd = tp->snd_ssthresh;
4739 		}
4740 	}
4741 	if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
4742 		union tcp_log_stackspecific log;
4743 		struct timeval tv;
4744 
4745 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4746 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4747 		log.u_bbr.flex1 = th_ack;
4748 		log.u_bbr.flex2 = tp->ccv->flags;
4749 		log.u_bbr.flex3 = tp->ccv->bytes_this_ack;
4750 		log.u_bbr.flex4 = tp->ccv->nsegs;
4751 		log.u_bbr.flex5 = V_tcp_abc_l_var;
4752 		log.u_bbr.flex6 = orig_cwnd;
4753 		log.u_bbr.flex7 = V_tcp_do_newsack;
4754 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
4755 		log.u_bbr.flex8 = 2;
4756 		tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
4757 			       0, &log, false, NULL, NULL, 0, &tv);
4758 	}
4759 	if ((rack->rack_no_prr == 0) &&
4760 	    (rack->no_prr_addback == 0) &&
4761 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
4762 		/*
4763 		 * Suck the next prr cnt back into cwnd, but
4764 		 * only do that if we are not application limited.
4765 		 */
4766 		if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
4767 			/*
4768 			 * We are allowed to add back to the cwnd the amount we did
4769 			 * not get out if:
4770 			 * a) no_prr_addback is off.
4771 			 * b) we are not app limited
4772 			 * c) we are doing prr
4773 			 * <and>
4774 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
4775 			 */
4776 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
4777 					    rack->r_ctl.rc_prr_sndcnt);
4778 		}
4779 		rack->r_ctl.rc_prr_sndcnt = 0;
4780 		rack_log_to_prr(rack, 1, 0, __LINE__);
4781 	}
4782 	rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
4783 	tp->snd_recover = tp->snd_una;
4784 	if (rack->r_ctl.dsack_persist) {
4785 		rack->r_ctl.dsack_persist--;
4786 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
4787 			rack->r_ctl.num_dsack = 0;
4788 		}
4789 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
4790 	}
4791 	EXIT_RECOVERY(tp->t_flags);
4792 }
4793 
4794 static void
4795 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
4796 {
4797 	struct tcp_rack *rack;
4798 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
4799 
4800 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4801 #ifdef STATS
4802 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
4803 #endif
4804 	if (IN_RECOVERY(tp->t_flags) == 0) {
4805 		in_rec_at_entry = 0;
4806 		ssthresh_enter = tp->snd_ssthresh;
4807 		cwnd_enter = tp->snd_cwnd;
4808 	} else
4809 		in_rec_at_entry = 1;
4810 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4811 	switch (type) {
4812 	case CC_NDUPACK:
4813 		tp->t_flags &= ~TF_WASFRECOVERY;
4814 		tp->t_flags &= ~TF_WASCRECOVERY;
4815 		if (!IN_FASTRECOVERY(tp->t_flags)) {
4816 			rack->r_ctl.rc_prr_delivered = 0;
4817 			rack->r_ctl.rc_prr_out = 0;
4818 			if (rack->rack_no_prr == 0) {
4819 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
4820 				rack_log_to_prr(rack, 2, in_rec_at_entry, line);
4821 			}
4822 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
4823 			tp->snd_recover = tp->snd_max;
4824 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4825 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4826 		}
4827 		break;
4828 	case CC_ECN:
4829 		if (!IN_CONGRECOVERY(tp->t_flags) ||
4830 		    /*
4831 		     * Allow ECN reaction on ACK to CWR, if
4832 		     * that data segment was also CE marked.
4833 		     */
4834 		    SEQ_GEQ(ack, tp->snd_recover)) {
4835 			EXIT_CONGRECOVERY(tp->t_flags);
4836 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
4837 			tp->snd_recover = tp->snd_max + 1;
4838 			if (tp->t_flags2 & TF2_ECN_PERMIT)
4839 				tp->t_flags2 |= TF2_ECN_SND_CWR;
4840 		}
4841 		break;
4842 	case CC_RTO:
4843 		tp->t_dupacks = 0;
4844 		tp->t_bytes_acked = 0;
4845 		EXIT_RECOVERY(tp->t_flags);
4846 		tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
4847 		    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
4848 		orig_cwnd = tp->snd_cwnd;
4849 		tp->snd_cwnd = ctf_fixed_maxseg(tp);
4850 		rack_log_to_prr(rack, 16, orig_cwnd, line);
4851 		if (tp->t_flags2 & TF2_ECN_PERMIT)
4852 			tp->t_flags2 |= TF2_ECN_SND_CWR;
4853 		break;
4854 	case CC_RTO_ERR:
4855 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
4856 		/* RTO was unnecessary, so reset everything. */
4857 		tp->snd_cwnd = tp->snd_cwnd_prev;
4858 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
4859 		tp->snd_recover = tp->snd_recover_prev;
4860 		if (tp->t_flags & TF_WASFRECOVERY) {
4861 			ENTER_FASTRECOVERY(tp->t_flags);
4862 			tp->t_flags &= ~TF_WASFRECOVERY;
4863 		}
4864 		if (tp->t_flags & TF_WASCRECOVERY) {
4865 			ENTER_CONGRECOVERY(tp->t_flags);
4866 			tp->t_flags &= ~TF_WASCRECOVERY;
4867 		}
4868 		tp->snd_nxt = tp->snd_max;
4869 		tp->t_badrxtwin = 0;
4870 		break;
4871 	}
4872 	if ((CC_ALGO(tp)->cong_signal != NULL)  &&
4873 	    (type != CC_RTO)){
4874 		tp->ccv->curack = ack;
4875 		CC_ALGO(tp)->cong_signal(tp->ccv, type);
4876 	}
4877 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
4878 		rack_log_to_prr(rack, 15, cwnd_enter, line);
4879 		rack->r_ctl.dsack_byte_cnt = 0;
4880 		rack->r_ctl.retran_during_recovery = 0;
4881 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
4882 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
4883 		rack->r_ent_rec_ns = 1;
4884 	}
4885 }
4886 
4887 static inline void
4888 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
4889 {
4890 	uint32_t i_cwnd;
4891 
4892 	INP_WLOCK_ASSERT(tptoinpcb(tp));
4893 
4894 #ifdef NETFLIX_STATS
4895 	KMOD_TCPSTAT_INC(tcps_idle_restarts);
4896 	if (tp->t_state == TCPS_ESTABLISHED)
4897 		KMOD_TCPSTAT_INC(tcps_idle_estrestarts);
4898 #endif
4899 	if (CC_ALGO(tp)->after_idle != NULL)
4900 		CC_ALGO(tp)->after_idle(tp->ccv);
4901 
4902 	if (tp->snd_cwnd == 1)
4903 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
4904 	else
4905 		i_cwnd = rc_init_window(rack);
4906 
4907 	/*
4908 	 * Being idle is no different than the initial window. If the cc
4909 	 * clamps it down below the initial window raise it to the initial
4910 	 * window.
4911 	 */
4912 	if (tp->snd_cwnd < i_cwnd) {
4913 		tp->snd_cwnd = i_cwnd;
4914 	}
4915 }
4916 
4917 /*
4918  * Indicate whether this ack should be delayed.  We can delay the ack if
4919  * following conditions are met:
4920  *	- There is no delayed ack timer in progress.
4921  *	- Our last ack wasn't a 0-sized window. We never want to delay
4922  *	  the ack that opens up a 0-sized window.
4923  *	- LRO wasn't used for this segment. We make sure by checking that the
4924  *	  segment size is not larger than the MSS.
4925  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
4926  *	  connection.
4927  */
4928 #define DELAY_ACK(tp, tlen)			 \
4929 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
4930 	((tp->t_flags & TF_DELACK) == 0) &&	 \
4931 	(tlen <= tp->t_maxseg) &&		 \
4932 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
4933 
4934 static struct rack_sendmap *
4935 rack_find_lowest_rsm(struct tcp_rack *rack)
4936 {
4937 	struct rack_sendmap *rsm;
4938 
4939 	/*
4940 	 * Walk the time-order transmitted list looking for an rsm that is
4941 	 * not acked. This will be the one that was sent the longest time
4942 	 * ago that is still outstanding.
4943 	 */
4944 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
4945 		if (rsm->r_flags & RACK_ACKED) {
4946 			continue;
4947 		}
4948 		goto finish;
4949 	}
4950 finish:
4951 	return (rsm);
4952 }
4953 
4954 static struct rack_sendmap *
4955 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
4956 {
4957 	struct rack_sendmap *prsm;
4958 
4959 	/*
4960 	 * Walk the sequence order list backward until we hit and arrive at
4961 	 * the highest seq not acked. In theory when this is called it
4962 	 * should be the last segment (which it was not).
4963 	 */
4964 	prsm = rsm;
4965 	RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) {
4966 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
4967 			continue;
4968 		}
4969 		return (prsm);
4970 	}
4971 	return (NULL);
4972 }
4973 
4974 static uint32_t
4975 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
4976 {
4977 	int32_t lro;
4978 	uint32_t thresh;
4979 
4980 	/*
4981 	 * lro is the flag we use to determine if we have seen reordering.
4982 	 * If it gets set we have seen reordering. The reorder logic either
4983 	 * works in one of two ways:
4984 	 *
4985 	 * If reorder-fade is configured, then we track the last time we saw
4986 	 * re-ordering occur. If we reach the point where enough time as
4987 	 * passed we no longer consider reordering has occuring.
4988 	 *
4989 	 * Or if reorder-face is 0, then once we see reordering we consider
4990 	 * the connection to alway be subject to reordering and just set lro
4991 	 * to 1.
4992 	 *
4993 	 * In the end if lro is non-zero we add the extra time for
4994 	 * reordering in.
4995 	 */
4996 	if (srtt == 0)
4997 		srtt = 1;
4998 	if (rack->r_ctl.rc_reorder_ts) {
4999 		if (rack->r_ctl.rc_reorder_fade) {
5000 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
5001 				lro = cts - rack->r_ctl.rc_reorder_ts;
5002 				if (lro == 0) {
5003 					/*
5004 					 * No time as passed since the last
5005 					 * reorder, mark it as reordering.
5006 					 */
5007 					lro = 1;
5008 				}
5009 			} else {
5010 				/* Negative time? */
5011 				lro = 0;
5012 			}
5013 			if (lro > rack->r_ctl.rc_reorder_fade) {
5014 				/* Turn off reordering seen too */
5015 				rack->r_ctl.rc_reorder_ts = 0;
5016 				lro = 0;
5017 			}
5018 		} else {
5019 			/* Reodering does not fade */
5020 			lro = 1;
5021 		}
5022 	} else {
5023 		lro = 0;
5024 	}
5025 	if (rack->rc_rack_tmr_std_based == 0) {
5026 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
5027 	} else {
5028 		/* Standards based pkt-delay is 1/4 srtt */
5029 		thresh = srtt +  (srtt >> 2);
5030 	}
5031 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
5032 		/* It must be set, if not you get 1/4 rtt */
5033 		if (rack->r_ctl.rc_reorder_shift)
5034 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
5035 		else
5036 			thresh += (srtt >> 2);
5037 	}
5038 	if (rack->rc_rack_use_dsack &&
5039 	    lro &&
5040 	    (rack->r_ctl.num_dsack > 0)) {
5041 		/*
5042 		 * We only increase the reordering window if we
5043 		 * have seen reordering <and> we have a DSACK count.
5044 		 */
5045 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
5046 		rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
5047 	}
5048 	/* SRTT * 2 is the ceiling */
5049 	if (thresh > (srtt * 2)) {
5050 		thresh = srtt * 2;
5051 	}
5052 	/* And we don't want it above the RTO max either */
5053 	if (thresh > rack_rto_max) {
5054 		thresh = rack_rto_max;
5055 	}
5056 	rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
5057 	return (thresh);
5058 }
5059 
5060 static uint32_t
5061 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
5062 		     struct rack_sendmap *rsm, uint32_t srtt)
5063 {
5064 	struct rack_sendmap *prsm;
5065 	uint32_t thresh, len;
5066 	int segsiz;
5067 
5068 	if (srtt == 0)
5069 		srtt = 1;
5070 	if (rack->r_ctl.rc_tlp_threshold)
5071 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
5072 	else
5073 		thresh = (srtt * 2);
5074 
5075 	/* Get the previous sent packet, if any */
5076 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5077 	len = rsm->r_end - rsm->r_start;
5078 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
5079 		/* Exactly like the ID */
5080 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
5081 			uint32_t alt_thresh;
5082 			/*
5083 			 * Compensate for delayed-ack with the d-ack time.
5084 			 */
5085 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5086 			if (alt_thresh > thresh)
5087 				thresh = alt_thresh;
5088 		}
5089 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
5090 		/* 2.1 behavior */
5091 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
5092 		if (prsm && (len <= segsiz)) {
5093 			/*
5094 			 * Two packets outstanding, thresh should be (2*srtt) +
5095 			 * possible inter-packet delay (if any).
5096 			 */
5097 			uint32_t inter_gap = 0;
5098 			int idx, nidx;
5099 
5100 			idx = rsm->r_rtr_cnt - 1;
5101 			nidx = prsm->r_rtr_cnt - 1;
5102 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
5103 				/* Yes it was sent later (or at the same time) */
5104 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
5105 			}
5106 			thresh += inter_gap;
5107 		} else if (len <= segsiz) {
5108 			/*
5109 			 * Possibly compensate for delayed-ack.
5110 			 */
5111 			uint32_t alt_thresh;
5112 
5113 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5114 			if (alt_thresh > thresh)
5115 				thresh = alt_thresh;
5116 		}
5117 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
5118 		/* 2.2 behavior */
5119 		if (len <= segsiz) {
5120 			uint32_t alt_thresh;
5121 			/*
5122 			 * Compensate for delayed-ack with the d-ack time.
5123 			 */
5124 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
5125 			if (alt_thresh > thresh)
5126 				thresh = alt_thresh;
5127 		}
5128 	}
5129 	/* Not above an RTO */
5130 	if (thresh > tp->t_rxtcur) {
5131 		thresh = tp->t_rxtcur;
5132 	}
5133 	/* Not above a RTO max */
5134 	if (thresh > rack_rto_max) {
5135 		thresh = rack_rto_max;
5136 	}
5137 	/* Apply user supplied min TLP */
5138 	if (thresh < rack_tlp_min) {
5139 		thresh = rack_tlp_min;
5140 	}
5141 	return (thresh);
5142 }
5143 
5144 static uint32_t
5145 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
5146 {
5147 	/*
5148 	 * We want the rack_rtt which is the
5149 	 * last rtt we measured. However if that
5150 	 * does not exist we fallback to the srtt (which
5151 	 * we probably will never do) and then as a last
5152 	 * resort we use RACK_INITIAL_RTO if no srtt is
5153 	 * yet set.
5154 	 */
5155 	if (rack->rc_rack_rtt)
5156 		return (rack->rc_rack_rtt);
5157 	else if (tp->t_srtt == 0)
5158 		return (RACK_INITIAL_RTO);
5159 	return (tp->t_srtt);
5160 }
5161 
5162 static struct rack_sendmap *
5163 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
5164 {
5165 	/*
5166 	 * Check to see that we don't need to fall into recovery. We will
5167 	 * need to do so if our oldest transmit is past the time we should
5168 	 * have had an ack.
5169 	 */
5170 	struct tcp_rack *rack;
5171 	struct rack_sendmap *rsm;
5172 	int32_t idx;
5173 	uint32_t srtt, thresh;
5174 
5175 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5176 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
5177 		return (NULL);
5178 	}
5179 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5180 	if (rsm == NULL)
5181 		return (NULL);
5182 
5183 
5184 	if (rsm->r_flags & RACK_ACKED) {
5185 		rsm = rack_find_lowest_rsm(rack);
5186 		if (rsm == NULL)
5187 			return (NULL);
5188 	}
5189 	idx = rsm->r_rtr_cnt - 1;
5190 	srtt = rack_grab_rtt(tp, rack);
5191 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
5192 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
5193 		return (NULL);
5194 	}
5195 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
5196 		return (NULL);
5197 	}
5198 	/* Ok if we reach here we are over-due and this guy can be sent */
5199 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
5200 	return (rsm);
5201 }
5202 
5203 static uint32_t
5204 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
5205 {
5206 	int32_t t;
5207 	int32_t tt;
5208 	uint32_t ret_val;
5209 
5210 	t = (tp->t_srtt + (tp->t_rttvar << 2));
5211 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
5212  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
5213 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
5214 	ret_val = (uint32_t)tt;
5215 	return (ret_val);
5216 }
5217 
5218 static uint32_t
5219 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
5220 {
5221 	/*
5222 	 * Start the FR timer, we do this based on getting the first one in
5223 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
5224 	 * events we need to stop the running timer (if its running) before
5225 	 * starting the new one.
5226 	 */
5227 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
5228 	uint32_t srtt_cur;
5229 	int32_t idx;
5230 	int32_t is_tlp_timer = 0;
5231 	struct rack_sendmap *rsm;
5232 
5233 	if (rack->t_timers_stopped) {
5234 		/* All timers have been stopped none are to run */
5235 		return (0);
5236 	}
5237 	if (rack->rc_in_persist) {
5238 		/* We can't start any timer in persists */
5239 		return (rack_get_persists_timer_val(tp, rack));
5240 	}
5241 	rack->rc_on_min_to = 0;
5242 	if ((tp->t_state < TCPS_ESTABLISHED) ||
5243 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
5244 		goto activate_rxt;
5245 	}
5246 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5247 	if ((rsm == NULL) || sup_rack) {
5248 		/* Nothing on the send map or no rack */
5249 activate_rxt:
5250 		time_since_sent = 0;
5251 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
5252 		if (rsm) {
5253 			/*
5254 			 * Should we discount the RTX timer any?
5255 			 *
5256 			 * We want to discount it the smallest amount.
5257 			 * If a timer (Rack/TLP or RXT) has gone off more
5258 			 * recently thats the discount we want to use (now - timer time).
5259 			 * If the retransmit of the oldest packet was more recent then
5260 			 * we want to use that (now - oldest-packet-last_transmit_time).
5261 			 *
5262 			 */
5263 			idx = rsm->r_rtr_cnt - 1;
5264 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
5265 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5266 			else
5267 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5268 			if (TSTMP_GT(cts, tstmp_touse))
5269 			    time_since_sent = cts - tstmp_touse;
5270 		}
5271 		if (SEQ_LT(tp->snd_una, tp->snd_max) ||
5272 		    sbavail(&tptosocket(tp)->so_snd)) {
5273 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
5274 			to = tp->t_rxtcur;
5275 			if (to > time_since_sent)
5276 				to -= time_since_sent;
5277 			else
5278 				to = rack->r_ctl.rc_min_to;
5279 			if (to == 0)
5280 				to = 1;
5281 			/* Special case for KEEPINIT */
5282 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
5283 			    (TP_KEEPINIT(tp) != 0) &&
5284 			    rsm) {
5285 				/*
5286 				 * We have to put a ceiling on the rxt timer
5287 				 * of the keep-init timeout.
5288 				 */
5289 				uint32_t max_time, red;
5290 
5291 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
5292 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
5293 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
5294 					if (red < max_time)
5295 						max_time -= red;
5296 					else
5297 						max_time = 1;
5298 				}
5299 				/* Reduce timeout to the keep value if needed */
5300 				if (max_time < to)
5301 					to = max_time;
5302 			}
5303 			return (to);
5304 		}
5305 		return (0);
5306 	}
5307 	if (rsm->r_flags & RACK_ACKED) {
5308 		rsm = rack_find_lowest_rsm(rack);
5309 		if (rsm == NULL) {
5310 			/* No lowest? */
5311 			goto activate_rxt;
5312 		}
5313 	}
5314 	if (rack->sack_attack_disable) {
5315 		/*
5316 		 * We don't want to do
5317 		 * any TLP's if you are an attacker.
5318 		 * Though if you are doing what
5319 		 * is expected you may still have
5320 		 * SACK-PASSED marks.
5321 		 */
5322 		goto activate_rxt;
5323 	}
5324 	/* Convert from ms to usecs */
5325 	if ((rsm->r_flags & RACK_SACK_PASSED) ||
5326 	    (rsm->r_flags & RACK_RWND_COLLAPSED) ||
5327 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
5328 		if ((tp->t_flags & TF_SENTFIN) &&
5329 		    ((tp->snd_max - tp->snd_una) == 1) &&
5330 		    (rsm->r_flags & RACK_HAS_FIN)) {
5331 			/*
5332 			 * We don't start a rack timer if all we have is a
5333 			 * FIN outstanding.
5334 			 */
5335 			goto activate_rxt;
5336 		}
5337 		if ((rack->use_rack_rr == 0) &&
5338 		    (IN_FASTRECOVERY(tp->t_flags)) &&
5339 		    (rack->rack_no_prr == 0) &&
5340 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
5341 			/*
5342 			 * We are not cheating, in recovery  and
5343 			 * not enough ack's to yet get our next
5344 			 * retransmission out.
5345 			 *
5346 			 * Note that classified attackers do not
5347 			 * get to use the rack-cheat.
5348 			 */
5349 			goto activate_tlp;
5350 		}
5351 		srtt = rack_grab_rtt(tp, rack);
5352 		thresh = rack_calc_thresh_rack(rack, srtt, cts);
5353 		idx = rsm->r_rtr_cnt - 1;
5354 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
5355 		if (SEQ_GEQ(exp, cts)) {
5356 			to = exp - cts;
5357 			if (to < rack->r_ctl.rc_min_to) {
5358 				to = rack->r_ctl.rc_min_to;
5359 				if (rack->r_rr_config == 3)
5360 					rack->rc_on_min_to = 1;
5361 			}
5362 		} else {
5363 			to = rack->r_ctl.rc_min_to;
5364 			if (rack->r_rr_config == 3)
5365 				rack->rc_on_min_to = 1;
5366 		}
5367 	} else {
5368 		/* Ok we need to do a TLP not RACK */
5369 activate_tlp:
5370 		if ((rack->rc_tlp_in_progress != 0) &&
5371 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
5372 			/*
5373 			 * The previous send was a TLP and we have sent
5374 			 * N TLP's without sending new data.
5375 			 */
5376 			goto activate_rxt;
5377 		}
5378 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
5379 		if (rsm == NULL) {
5380 			/* We found no rsm to TLP with. */
5381 			goto activate_rxt;
5382 		}
5383 		if (rsm->r_flags & RACK_HAS_FIN) {
5384 			/* If its a FIN we dont do TLP */
5385 			rsm = NULL;
5386 			goto activate_rxt;
5387 		}
5388 		idx = rsm->r_rtr_cnt - 1;
5389 		time_since_sent = 0;
5390 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
5391 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
5392 		else
5393 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
5394 		if (TSTMP_GT(cts, tstmp_touse))
5395 		    time_since_sent = cts - tstmp_touse;
5396 		is_tlp_timer = 1;
5397 		if (tp->t_srtt) {
5398 			if ((rack->rc_srtt_measure_made == 0) &&
5399 			    (tp->t_srtt == 1)) {
5400 				/*
5401 				 * If another stack as run and set srtt to 1,
5402 				 * then the srtt was 0, so lets use the initial.
5403 				 */
5404 				srtt = RACK_INITIAL_RTO;
5405 			} else {
5406 				srtt_cur = tp->t_srtt;
5407 				srtt = srtt_cur;
5408 			}
5409 		} else
5410 			srtt = RACK_INITIAL_RTO;
5411 		/*
5412 		 * If the SRTT is not keeping up and the
5413 		 * rack RTT has spiked we want to use
5414 		 * the last RTT not the smoothed one.
5415 		 */
5416 		if (rack_tlp_use_greater &&
5417 		    tp->t_srtt &&
5418 		    (srtt < rack_grab_rtt(tp, rack))) {
5419 			srtt = rack_grab_rtt(tp, rack);
5420 		}
5421 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
5422 		if (thresh > time_since_sent) {
5423 			to = thresh - time_since_sent;
5424 		} else {
5425 			to = rack->r_ctl.rc_min_to;
5426 			rack_log_alt_to_to_cancel(rack,
5427 						  thresh,		/* flex1 */
5428 						  time_since_sent,	/* flex2 */
5429 						  tstmp_touse,		/* flex3 */
5430 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
5431 						  (uint32_t)rsm->r_tim_lastsent[idx],
5432 						  srtt,
5433 						  idx, 99);
5434 		}
5435 		if (to < rack_tlp_min) {
5436 			to = rack_tlp_min;
5437 		}
5438 		if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
5439 			/*
5440 			 * If the TLP time works out to larger than the max
5441 			 * RTO lets not do TLP.. just RTO.
5442 			 */
5443 			goto activate_rxt;
5444 		}
5445 	}
5446 	if (is_tlp_timer == 0) {
5447 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
5448 	} else {
5449 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
5450 	}
5451 	if (to == 0)
5452 		to = 1;
5453 	return (to);
5454 }
5455 
5456 static void
5457 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5458 {
5459 	if (rack->rc_in_persist == 0) {
5460 		if (tp->t_flags & TF_GPUTINPROG) {
5461 			/*
5462 			 * Stop the goodput now, the calling of the
5463 			 * measurement function clears the flag.
5464 			 */
5465 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
5466 						    RACK_QUALITY_PERSIST);
5467 		}
5468 #ifdef NETFLIX_SHARED_CWND
5469 		if (rack->r_ctl.rc_scw) {
5470 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5471 			rack->rack_scwnd_is_idle = 1;
5472 		}
5473 #endif
5474 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
5475 		if (rack->r_ctl.rc_went_idle_time == 0)
5476 			rack->r_ctl.rc_went_idle_time = 1;
5477 		rack_timer_cancel(tp, rack, cts, __LINE__);
5478 		rack->r_ctl.persist_lost_ends = 0;
5479 		rack->probe_not_answered = 0;
5480 		rack->forced_ack = 0;
5481 		tp->t_rxtshift = 0;
5482 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5483 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5484 		rack->rc_in_persist = 1;
5485 	}
5486 }
5487 
5488 static void
5489 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5490 {
5491 	if (tcp_in_hpts(rack->rc_inp)) {
5492 		tcp_hpts_remove(rack->rc_inp);
5493 		rack->r_ctl.rc_hpts_flags = 0;
5494 	}
5495 #ifdef NETFLIX_SHARED_CWND
5496 	if (rack->r_ctl.rc_scw) {
5497 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
5498 		rack->rack_scwnd_is_idle = 0;
5499 	}
5500 #endif
5501 	if (rack->rc_gp_dyn_mul &&
5502 	    (rack->use_fixed_rate == 0) &&
5503 	    (rack->rc_always_pace)) {
5504 		/*
5505 		 * Do we count this as if a probe-rtt just
5506 		 * finished?
5507 		 */
5508 		uint32_t time_idle, idle_min;
5509 
5510 		time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time;
5511 		idle_min = rack_min_probertt_hold;
5512 		if (rack_probertt_gpsrtt_cnt_div) {
5513 			uint64_t extra;
5514 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
5515 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
5516 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
5517 			idle_min += (uint32_t)extra;
5518 		}
5519 		if (time_idle >= idle_min) {
5520 			/* Yes, we count it as a probe-rtt. */
5521 			uint32_t us_cts;
5522 
5523 			us_cts = tcp_get_usecs(NULL);
5524 			if (rack->in_probe_rtt == 0) {
5525 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
5526 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
5527 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
5528 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
5529 			} else {
5530 				rack_exit_probertt(rack, us_cts);
5531 			}
5532 		}
5533 	}
5534 	rack->rc_in_persist = 0;
5535 	rack->r_ctl.rc_went_idle_time = 0;
5536 	tp->t_rxtshift = 0;
5537 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
5538 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
5539 	rack->r_ctl.rc_agg_delayed = 0;
5540 	rack->r_early = 0;
5541 	rack->r_late = 0;
5542 	rack->r_ctl.rc_agg_early = 0;
5543 }
5544 
5545 static void
5546 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
5547 		   struct hpts_diag *diag, struct timeval *tv)
5548 {
5549 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5550 		union tcp_log_stackspecific log;
5551 
5552 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5553 		log.u_bbr.flex1 = diag->p_nxt_slot;
5554 		log.u_bbr.flex2 = diag->p_cur_slot;
5555 		log.u_bbr.flex3 = diag->slot_req;
5556 		log.u_bbr.flex4 = diag->inp_hptsslot;
5557 		log.u_bbr.flex5 = diag->slot_remaining;
5558 		log.u_bbr.flex6 = diag->need_new_to;
5559 		log.u_bbr.flex7 = diag->p_hpts_active;
5560 		log.u_bbr.flex8 = diag->p_on_min_sleep;
5561 		/* Hijack other fields as needed */
5562 		log.u_bbr.epoch = diag->have_slept;
5563 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
5564 		log.u_bbr.pkts_out = diag->co_ret;
5565 		log.u_bbr.applimited = diag->hpts_sleep_time;
5566 		log.u_bbr.delivered = diag->p_prev_slot;
5567 		log.u_bbr.inflight = diag->p_runningslot;
5568 		log.u_bbr.bw_inuse = diag->wheel_slot;
5569 		log.u_bbr.rttProp = diag->wheel_cts;
5570 		log.u_bbr.timeStamp = cts;
5571 		log.u_bbr.delRate = diag->maxslots;
5572 		log.u_bbr.cur_del_rate = diag->p_curtick;
5573 		log.u_bbr.cur_del_rate <<= 32;
5574 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
5575 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5576 		    &rack->rc_inp->inp_socket->so_rcv,
5577 		    &rack->rc_inp->inp_socket->so_snd,
5578 		    BBR_LOG_HPTSDIAG, 0,
5579 		    0, &log, false, tv);
5580 	}
5581 
5582 }
5583 
5584 static void
5585 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
5586 {
5587 	if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
5588 		union tcp_log_stackspecific log;
5589 		struct timeval tv;
5590 
5591 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5592 		log.u_bbr.flex1 = sb->sb_flags;
5593 		log.u_bbr.flex2 = len;
5594 		log.u_bbr.flex3 = sb->sb_state;
5595 		log.u_bbr.flex8 = type;
5596 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5597 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
5598 		    &rack->rc_inp->inp_socket->so_rcv,
5599 		    &rack->rc_inp->inp_socket->so_snd,
5600 		    TCP_LOG_SB_WAKE, 0,
5601 		    len, &log, false, &tv);
5602 	}
5603 }
5604 
5605 static void
5606 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
5607       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
5608 {
5609 	struct hpts_diag diag;
5610 	struct inpcb *inp = tptoinpcb(tp);
5611 	struct timeval tv;
5612 	uint32_t delayed_ack = 0;
5613 	uint32_t hpts_timeout;
5614 	uint32_t entry_slot = slot;
5615 	uint8_t stopped;
5616 	uint32_t left = 0;
5617 	uint32_t us_cts;
5618 
5619 	if ((tp->t_state == TCPS_CLOSED) ||
5620 	    (tp->t_state == TCPS_LISTEN)) {
5621 		return;
5622 	}
5623 	if (tcp_in_hpts(inp)) {
5624 		/* Already on the pacer */
5625 		return;
5626 	}
5627 	stopped = rack->rc_tmr_stopped;
5628 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
5629 		left = rack->r_ctl.rc_timer_exp - cts;
5630 	}
5631 	rack->r_ctl.rc_timer_exp = 0;
5632 	rack->r_ctl.rc_hpts_flags = 0;
5633 	us_cts = tcp_get_usecs(&tv);
5634 	/* Now early/late accounting */
5635 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
5636 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
5637 		/*
5638 		 * We have a early carry over set,
5639 		 * we can always add more time so we
5640 		 * can always make this compensation.
5641 		 *
5642 		 * Note if ack's are allowed to wake us do not
5643 		 * penalize the next timer for being awoke
5644 		 * by an ack aka the rc_agg_early (non-paced mode).
5645 		 */
5646 		slot += rack->r_ctl.rc_agg_early;
5647 		rack->r_early = 0;
5648 		rack->r_ctl.rc_agg_early = 0;
5649 	}
5650 	if (rack->r_late) {
5651 		/*
5652 		 * This is harder, we can
5653 		 * compensate some but it
5654 		 * really depends on what
5655 		 * the current pacing time is.
5656 		 */
5657 		if (rack->r_ctl.rc_agg_delayed >= slot) {
5658 			/*
5659 			 * We can't compensate for it all.
5660 			 * And we have to have some time
5661 			 * on the clock. We always have a min
5662 			 * 10 slots (10 x 10 i.e. 100 usecs).
5663 			 */
5664 			if (slot <= HPTS_TICKS_PER_SLOT) {
5665 				/* We gain delay */
5666 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
5667 				slot = HPTS_TICKS_PER_SLOT;
5668 			} else {
5669 				/* We take off some */
5670 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
5671 				slot = HPTS_TICKS_PER_SLOT;
5672 			}
5673 		} else {
5674 			slot -= rack->r_ctl.rc_agg_delayed;
5675 			rack->r_ctl.rc_agg_delayed = 0;
5676 			/* Make sure we have 100 useconds at minimum */
5677 			if (slot < HPTS_TICKS_PER_SLOT) {
5678 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
5679 				slot = HPTS_TICKS_PER_SLOT;
5680 			}
5681 			if (rack->r_ctl.rc_agg_delayed == 0)
5682 				rack->r_late = 0;
5683 		}
5684 	}
5685 	if (slot) {
5686 		/* We are pacing too */
5687 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
5688 	}
5689 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
5690 #ifdef NETFLIX_EXP_DETECTION
5691 	if (rack->sack_attack_disable &&
5692 	    (slot < tcp_sad_pacing_interval)) {
5693 		/*
5694 		 * We have a potential attacker on
5695 		 * the line. We have possibly some
5696 		 * (or now) pacing time set. We want to
5697 		 * slow down the processing of sacks by some
5698 		 * amount (if it is an attacker). Set the default
5699 		 * slot for attackers in place (unless the orginal
5700 		 * interval is longer). Its stored in
5701 		 * micro-seconds, so lets convert to msecs.
5702 		 */
5703 		slot = tcp_sad_pacing_interval;
5704 	}
5705 #endif
5706 	if (tp->t_flags & TF_DELACK) {
5707 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
5708 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
5709 	}
5710 	if (delayed_ack && ((hpts_timeout == 0) ||
5711 			    (delayed_ack < hpts_timeout)))
5712 		hpts_timeout = delayed_ack;
5713 	else
5714 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
5715 	/*
5716 	 * If no timers are going to run and we will fall off the hptsi
5717 	 * wheel, we resort to a keep-alive timer if its configured.
5718 	 */
5719 	if ((hpts_timeout == 0) &&
5720 	    (slot == 0)) {
5721 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
5722 		    (tp->t_state <= TCPS_CLOSING)) {
5723 			/*
5724 			 * Ok we have no timer (persists, rack, tlp, rxt  or
5725 			 * del-ack), we don't have segments being paced. So
5726 			 * all that is left is the keepalive timer.
5727 			 */
5728 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
5729 				/* Get the established keep-alive time */
5730 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
5731 			} else {
5732 				/*
5733 				 * Get the initial setup keep-alive time,
5734 				 * note that this is probably not going to
5735 				 * happen, since rack will be running a rxt timer
5736 				 * if a SYN of some sort is outstanding. It is
5737 				 * actually handled in rack_timeout_rxt().
5738 				 */
5739 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
5740 			}
5741 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
5742 			if (rack->in_probe_rtt) {
5743 				/*
5744 				 * We want to instead not wake up a long time from
5745 				 * now but to wake up about the time we would
5746 				 * exit probe-rtt and initiate a keep-alive ack.
5747 				 * This will get us out of probe-rtt and update
5748 				 * our min-rtt.
5749 				 */
5750 				hpts_timeout = rack_min_probertt_hold;
5751 			}
5752 		}
5753 	}
5754 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
5755 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
5756 		/*
5757 		 * RACK, TLP, persists and RXT timers all are restartable
5758 		 * based on actions input .. i.e we received a packet (ack
5759 		 * or sack) and that changes things (rw, or snd_una etc).
5760 		 * Thus we can restart them with a new value. For
5761 		 * keep-alive, delayed_ack we keep track of what was left
5762 		 * and restart the timer with a smaller value.
5763 		 */
5764 		if (left < hpts_timeout)
5765 			hpts_timeout = left;
5766 	}
5767 	if (hpts_timeout) {
5768 		/*
5769 		 * Hack alert for now we can't time-out over 2,147,483
5770 		 * seconds (a bit more than 596 hours), which is probably ok
5771 		 * :).
5772 		 */
5773 		if (hpts_timeout > 0x7ffffffe)
5774 			hpts_timeout = 0x7ffffffe;
5775 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
5776 	}
5777 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
5778 	if ((rack->gp_ready == 0) &&
5779 	    (rack->use_fixed_rate == 0) &&
5780 	    (hpts_timeout < slot) &&
5781 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
5782 		/*
5783 		 * We have no good estimate yet for the
5784 		 * old clunky burst mitigation or the
5785 		 * real pacing. And the tlp or rxt is smaller
5786 		 * than the pacing calculation. Lets not
5787 		 * pace that long since we know the calculation
5788 		 * so far is not accurate.
5789 		 */
5790 		slot = hpts_timeout;
5791 	}
5792 	/**
5793 	 * Turn off all the flags for queuing by default. The
5794 	 * flags have important meanings to what happens when
5795 	 * LRO interacts with the transport. Most likely (by default now)
5796 	 * mbuf_queueing and ack compression are on. So the transport
5797 	 * has a couple of flags that control what happens (if those
5798 	 * are not on then these flags won't have any effect since it
5799 	 * won't go through the queuing LRO path).
5800 	 *
5801 	 * INP_MBUF_QUEUE_READY - This flags says that I am busy
5802 	 *                        pacing output, so don't disturb. But
5803 	 *                        it also means LRO can wake me if there
5804 	 *                        is a SACK arrival.
5805 	 *
5806 	 * INP_DONT_SACK_QUEUE - This flag is used in conjunction
5807 	 *                       with the above flag (QUEUE_READY) and
5808 	 *                       when present it says don't even wake me
5809 	 *                       if a SACK arrives.
5810 	 *
5811 	 * The idea behind these flags is that if we are pacing we
5812 	 * set the MBUF_QUEUE_READY and only get woken up if
5813 	 * a SACK arrives (which could change things) or if
5814 	 * our pacing timer expires. If, however, we have a rack
5815 	 * timer running, then we don't even want a sack to wake
5816 	 * us since the rack timer has to expire before we can send.
5817 	 *
5818 	 * Other cases should usually have none of the flags set
5819 	 * so LRO can call into us.
5820 	 */
5821 	inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5822 	if (slot) {
5823 		rack->r_ctl.rc_last_output_to = us_cts + slot;
5824 		/*
5825 		 * A pacing timer (slot) is being set, in
5826 		 * such a case we cannot send (we are blocked by
5827 		 * the timer). So lets tell LRO that it should not
5828 		 * wake us unless there is a SACK. Note this only
5829 		 * will be effective if mbuf queueing is on or
5830 		 * compressed acks are being processed.
5831 		 */
5832 		inp->inp_flags2 |= INP_MBUF_QUEUE_READY;
5833 		/*
5834 		 * But wait if we have a Rack timer running
5835 		 * even a SACK should not disturb us (with
5836 		 * the exception of r_rr_config 3).
5837 		 */
5838 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) &&
5839 		    (rack->r_rr_config != 3))
5840 			inp->inp_flags2 |= INP_DONT_SACK_QUEUE;
5841 		if (rack->rc_ack_can_sendout_data) {
5842 			/*
5843 			 * Ahh but wait, this is that special case
5844 			 * where the pacing timer can be disturbed
5845 			 * backout the changes (used for non-paced
5846 			 * burst limiting).
5847 			 */
5848 			inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY);
5849 		}
5850 		if ((rack->use_rack_rr) &&
5851 		    (rack->r_rr_config < 2) &&
5852 		    ((hpts_timeout) && (hpts_timeout < slot))) {
5853 			/*
5854 			 * Arrange for the hpts to kick back in after the
5855 			 * t-o if the t-o does not cause a send.
5856 			 */
5857 			(void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
5858 						   __LINE__, &diag);
5859 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5860 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5861 		} else {
5862 			(void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(slot),
5863 						   __LINE__, &diag);
5864 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5865 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
5866 		}
5867 	} else if (hpts_timeout) {
5868 		/*
5869 		 * With respect to inp_flags2 here, lets let any new acks wake
5870 		 * us up here. Since we are not pacing (no pacing timer), output
5871 		 * can happen so we should let it. If its a Rack timer, then any inbound
5872 		 * packet probably won't change the sending (we will be blocked)
5873 		 * but it may change the prr stats so letting it in (the set defaults
5874 		 * at the start of this block) are good enough.
5875 		 */
5876 		(void)tcp_hpts_insert_diag(inp, HPTS_USEC_TO_SLOTS(hpts_timeout),
5877 					   __LINE__, &diag);
5878 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
5879 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
5880 	} else {
5881 		/* No timer starting */
5882 #ifdef INVARIANTS
5883 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
5884 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
5885 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
5886 		}
5887 #endif
5888 	}
5889 	rack->rc_tmr_stopped = 0;
5890 	if (slot)
5891 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv);
5892 }
5893 
5894 /*
5895  * RACK Timer, here we simply do logging and house keeping.
5896  * the normal rack_output() function will call the
5897  * appropriate thing to check if we need to do a RACK retransmit.
5898  * We return 1, saying don't proceed with rack_output only
5899  * when all timers have been stopped (destroyed PCB?).
5900  */
5901 static int
5902 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
5903 {
5904 	/*
5905 	 * This timer simply provides an internal trigger to send out data.
5906 	 * The check_recovery_mode call will see if there are needed
5907 	 * retransmissions, if so we will enter fast-recovery. The output
5908 	 * call may or may not do the same thing depending on sysctl
5909 	 * settings.
5910 	 */
5911 	struct rack_sendmap *rsm;
5912 
5913 	if (tp->t_timers->tt_flags & TT_STOPPED) {
5914 		return (1);
5915 	}
5916 	counter_u64_add(rack_to_tot, 1);
5917 	if (rack->r_state && (rack->r_state != tp->t_state))
5918 		rack_set_state(tp, rack);
5919 	rack->rc_on_min_to = 0;
5920 	rsm = rack_check_recovery_mode(tp, cts);
5921 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
5922 	if (rsm) {
5923 		rack->r_ctl.rc_resend = rsm;
5924 		rack->r_timer_override = 1;
5925 		if (rack->use_rack_rr) {
5926 			/*
5927 			 * Don't accumulate extra pacing delay
5928 			 * we are allowing the rack timer to
5929 			 * over-ride pacing i.e. rrr takes precedence
5930 			 * if the pacing interval is longer than the rrr
5931 			 * time (in other words we get the min pacing
5932 			 * time versus rrr pacing time).
5933 			 */
5934 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
5935 		}
5936 	}
5937 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
5938 	if (rsm == NULL) {
5939 		/* restart a timer and return 1 */
5940 		rack_start_hpts_timer(rack, tp, cts,
5941 				      0, 0, 0);
5942 		return (1);
5943 	}
5944 	return (0);
5945 }
5946 
5947 static void
5948 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
5949 {
5950 	if (rsm->m->m_len > rsm->orig_m_len) {
5951 		/*
5952 		 * Mbuf grew, caused by sbcompress, our offset does
5953 		 * not change.
5954 		 */
5955 		rsm->orig_m_len = rsm->m->m_len;
5956 	} else if (rsm->m->m_len < rsm->orig_m_len) {
5957 		/*
5958 		 * Mbuf shrank, trimmed off the top by an ack, our
5959 		 * offset changes.
5960 		 */
5961 		rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
5962 		rsm->orig_m_len = rsm->m->m_len;
5963 	}
5964 }
5965 
5966 static void
5967 rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
5968 {
5969 	struct mbuf *m;
5970 	uint32_t soff;
5971 
5972 	if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) {
5973 		/* Fix up the orig_m_len and possibly the mbuf offset */
5974 		rack_adjust_orig_mlen(src_rsm);
5975 	}
5976 	m = src_rsm->m;
5977 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
5978 	while (soff >= m->m_len) {
5979 		/* Move out past this mbuf */
5980 		soff -= m->m_len;
5981 		m = m->m_next;
5982 		KASSERT((m != NULL),
5983 			("rsm:%p nrsm:%p hit at soff:%u null m",
5984 			 src_rsm, rsm, soff));
5985 	}
5986 	rsm->m = m;
5987 	rsm->soff = soff;
5988 	rsm->orig_m_len = m->m_len;
5989 }
5990 
5991 static __inline void
5992 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
5993 	       struct rack_sendmap *rsm, uint32_t start)
5994 {
5995 	int idx;
5996 
5997 	nrsm->r_start = start;
5998 	nrsm->r_end = rsm->r_end;
5999 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
6000 	nrsm->r_flags = rsm->r_flags;
6001 	nrsm->r_dupack = rsm->r_dupack;
6002 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
6003 	nrsm->r_rtr_bytes = 0;
6004 	nrsm->r_fas = rsm->r_fas;
6005 	rsm->r_end = nrsm->r_start;
6006 	nrsm->r_just_ret = rsm->r_just_ret;
6007 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
6008 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
6009 	}
6010 	/* Now if we have SYN flag we keep it on the left edge */
6011 	if (nrsm->r_flags & RACK_HAS_SYN)
6012 		nrsm->r_flags &= ~RACK_HAS_SYN;
6013 	/* Now if we have a FIN flag we keep it on the right edge */
6014 	if (rsm->r_flags & RACK_HAS_FIN)
6015 		rsm->r_flags &= ~RACK_HAS_FIN;
6016 	/* Push bit must go to the right edge as well */
6017 	if (rsm->r_flags & RACK_HAD_PUSH)
6018 		rsm->r_flags &= ~RACK_HAD_PUSH;
6019 	/* Clone over the state of the hw_tls flag */
6020 	nrsm->r_hw_tls = rsm->r_hw_tls;
6021 	/*
6022 	 * Now we need to find nrsm's new location in the mbuf chain
6023 	 * we basically calculate a new offset, which is soff +
6024 	 * how much is left in original rsm. Then we walk out the mbuf
6025 	 * chain to find the righ position, it may be the same mbuf
6026 	 * or maybe not.
6027 	 */
6028 	KASSERT(((rsm->m != NULL) ||
6029 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
6030 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
6031 	if (rsm->m)
6032 		rack_setup_offset_for_rsm(rsm, nrsm);
6033 }
6034 
6035 static struct rack_sendmap *
6036 rack_merge_rsm(struct tcp_rack *rack,
6037 	       struct rack_sendmap *l_rsm,
6038 	       struct rack_sendmap *r_rsm)
6039 {
6040 	/*
6041 	 * We are merging two ack'd RSM's,
6042 	 * the l_rsm is on the left (lower seq
6043 	 * values) and the r_rsm is on the right
6044 	 * (higher seq value). The simplest way
6045 	 * to merge these is to move the right
6046 	 * one into the left. I don't think there
6047 	 * is any reason we need to try to find
6048 	 * the oldest (or last oldest retransmitted).
6049 	 */
6050 #ifdef INVARIANTS
6051 	struct rack_sendmap *rm;
6052 #endif
6053 	rack_log_map_chg(rack->rc_tp, rack, NULL,
6054 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
6055 	l_rsm->r_end = r_rsm->r_end;
6056 	if (l_rsm->r_dupack < r_rsm->r_dupack)
6057 		l_rsm->r_dupack = r_rsm->r_dupack;
6058 	if (r_rsm->r_rtr_bytes)
6059 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
6060 	if (r_rsm->r_in_tmap) {
6061 		/* This really should not happen */
6062 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
6063 		r_rsm->r_in_tmap = 0;
6064 	}
6065 
6066 	/* Now the flags */
6067 	if (r_rsm->r_flags & RACK_HAS_FIN)
6068 		l_rsm->r_flags |= RACK_HAS_FIN;
6069 	if (r_rsm->r_flags & RACK_TLP)
6070 		l_rsm->r_flags |= RACK_TLP;
6071 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
6072 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
6073 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
6074 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
6075 		/*
6076 		 * If both are app-limited then let the
6077 		 * free lower the count. If right is app
6078 		 * limited and left is not, transfer.
6079 		 */
6080 		l_rsm->r_flags |= RACK_APP_LIMITED;
6081 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
6082 		if (r_rsm == rack->r_ctl.rc_first_appl)
6083 			rack->r_ctl.rc_first_appl = l_rsm;
6084 	}
6085 #ifndef INVARIANTS
6086 	(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6087 #else
6088 	rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm);
6089 	if (rm != r_rsm) {
6090 		panic("removing head in rack:%p rsm:%p rm:%p",
6091 		      rack, r_rsm, rm);
6092 	}
6093 #endif
6094 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
6095 		/* Transfer the split limit to the map we free */
6096 		r_rsm->r_limit_type = l_rsm->r_limit_type;
6097 		l_rsm->r_limit_type = 0;
6098 	}
6099 	rack_free(rack, r_rsm);
6100 	return (l_rsm);
6101 }
6102 
6103 /*
6104  * TLP Timer, here we simply setup what segment we want to
6105  * have the TLP expire on, the normal rack_output() will then
6106  * send it out.
6107  *
6108  * We return 1, saying don't proceed with rack_output only
6109  * when all timers have been stopped (destroyed PCB?).
6110  */
6111 static int
6112 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
6113 {
6114 	/*
6115 	 * Tail Loss Probe.
6116 	 */
6117 	struct rack_sendmap *rsm = NULL;
6118 #ifdef INVARIANTS
6119 	struct rack_sendmap *insret;
6120 #endif
6121 	struct socket *so = tptosocket(tp);
6122 	uint32_t amm;
6123 	uint32_t out, avail;
6124 	int collapsed_win = 0;
6125 
6126 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6127 		return (1);
6128 	}
6129 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6130 		/* Its not time yet */
6131 		return (0);
6132 	}
6133 	if (ctf_progress_timeout_check(tp, true)) {
6134 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6135 		return (-ETIMEDOUT);	/* tcp_drop() */
6136 	}
6137 	/*
6138 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
6139 	 * need to figure out how to force a full MSS segment out.
6140 	 */
6141 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
6142 	rack->r_ctl.retran_during_recovery = 0;
6143 	rack->r_ctl.dsack_byte_cnt = 0;
6144 	counter_u64_add(rack_tlp_tot, 1);
6145 	if (rack->r_state && (rack->r_state != tp->t_state))
6146 		rack_set_state(tp, rack);
6147 	avail = sbavail(&so->so_snd);
6148 	out = tp->snd_max - tp->snd_una;
6149 	if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
6150 		/* special case, we need a retransmission */
6151 		collapsed_win = 1;
6152 		goto need_retran;
6153 	}
6154 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
6155 		rack->r_ctl.dsack_persist--;
6156 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6157 			rack->r_ctl.num_dsack = 0;
6158 		}
6159 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6160 	}
6161 	if ((tp->t_flags & TF_GPUTINPROG) &&
6162 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
6163 		/*
6164 		 * If this is the second in a row
6165 		 * TLP and we are doing a measurement
6166 		 * its time to abandon the measurement.
6167 		 * Something is likely broken on
6168 		 * the clients network and measuring a
6169 		 * broken network does us no good.
6170 		 */
6171 		tp->t_flags &= ~TF_GPUTINPROG;
6172 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6173 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
6174 					   tp->gput_seq,
6175 					   0, 0, 18, __LINE__, NULL, 0);
6176 	}
6177 	/*
6178 	 * Check our send oldest always settings, and if
6179 	 * there is an oldest to send jump to the need_retran.
6180 	 */
6181 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
6182 		goto need_retran;
6183 
6184 	if (avail > out) {
6185 		/* New data is available */
6186 		amm = avail - out;
6187 		if (amm > ctf_fixed_maxseg(tp)) {
6188 			amm = ctf_fixed_maxseg(tp);
6189 			if ((amm + out) > tp->snd_wnd) {
6190 				/* We are rwnd limited */
6191 				goto need_retran;
6192 			}
6193 		} else if (amm < ctf_fixed_maxseg(tp)) {
6194 			/* not enough to fill a MTU */
6195 			goto need_retran;
6196 		}
6197 		if (IN_FASTRECOVERY(tp->t_flags)) {
6198 			/* Unlikely */
6199 			if (rack->rack_no_prr == 0) {
6200 				if (out + amm <= tp->snd_wnd) {
6201 					rack->r_ctl.rc_prr_sndcnt = amm;
6202 					rack->r_ctl.rc_tlp_new_data = amm;
6203 					rack_log_to_prr(rack, 4, 0, __LINE__);
6204 				}
6205 			} else
6206 				goto need_retran;
6207 		} else {
6208 			/* Set the send-new override */
6209 			if (out + amm <= tp->snd_wnd)
6210 				rack->r_ctl.rc_tlp_new_data = amm;
6211 			else
6212 				goto need_retran;
6213 		}
6214 		rack->r_ctl.rc_tlpsend = NULL;
6215 		counter_u64_add(rack_tlp_newdata, 1);
6216 		goto send;
6217 	}
6218 need_retran:
6219 	/*
6220 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
6221 	 * optionally the first un-acked segment.
6222 	 */
6223 	if (collapsed_win == 0) {
6224 		if (rack_always_send_oldest)
6225 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6226 		else {
6227 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6228 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
6229 				rsm = rack_find_high_nonack(rack, rsm);
6230 			}
6231 		}
6232 		if (rsm == NULL) {
6233 #ifdef TCP_BLACKBOX
6234 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6235 #endif
6236 			goto out;
6237 		}
6238 	} else {
6239 		/*
6240 		 * We must find the last segment
6241 		 * that was acceptable by the client.
6242 		 */
6243 		RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6244 			if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) {
6245 				/* Found one */
6246 				break;
6247 			}
6248 		}
6249 		if (rsm == NULL) {
6250 			/* None? if so send the first */
6251 			rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6252 			if (rsm == NULL) {
6253 #ifdef TCP_BLACKBOX
6254 				tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
6255 #endif
6256 				goto out;
6257 			}
6258 		}
6259 	}
6260 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
6261 		/*
6262 		 * We need to split this the last segment in two.
6263 		 */
6264 		struct rack_sendmap *nrsm;
6265 
6266 		nrsm = rack_alloc_full_limit(rack);
6267 		if (nrsm == NULL) {
6268 			/*
6269 			 * No memory to split, we will just exit and punt
6270 			 * off to the RXT timer.
6271 			 */
6272 			goto out;
6273 		}
6274 		rack_clone_rsm(rack, nrsm, rsm,
6275 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
6276 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
6277 #ifndef INVARIANTS
6278 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6279 #else
6280 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
6281 		if (insret != NULL) {
6282 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
6283 			      nrsm, insret, rack, rsm);
6284 		}
6285 #endif
6286 		if (rsm->r_in_tmap) {
6287 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
6288 			nrsm->r_in_tmap = 1;
6289 		}
6290 		rsm = nrsm;
6291 	}
6292 	rack->r_ctl.rc_tlpsend = rsm;
6293 send:
6294 	/* Make sure output path knows we are doing a TLP */
6295 	*doing_tlp = 1;
6296 	rack->r_timer_override = 1;
6297 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6298 	return (0);
6299 out:
6300 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
6301 	return (0);
6302 }
6303 
6304 /*
6305  * Delayed ack Timer, here we simply need to setup the
6306  * ACK_NOW flag and remove the DELACK flag. From there
6307  * the output routine will send the ack out.
6308  *
6309  * We only return 1, saying don't proceed, if all timers
6310  * are stopped (destroyed PCB?).
6311  */
6312 static int
6313 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6314 {
6315 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6316 		return (1);
6317 	}
6318 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
6319 	tp->t_flags &= ~TF_DELACK;
6320 	tp->t_flags |= TF_ACKNOW;
6321 	KMOD_TCPSTAT_INC(tcps_delack);
6322 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6323 	return (0);
6324 }
6325 
6326 /*
6327  * Persists timer, here we simply send the
6328  * same thing as a keepalive will.
6329  * the one byte send.
6330  *
6331  * We only return 1, saying don't proceed, if all timers
6332  * are stopped (destroyed PCB?).
6333  */
6334 static int
6335 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6336 {
6337 	struct tcptemp *t_template;
6338 	int32_t retval = 1;
6339 
6340 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6341 		return (1);
6342 	}
6343 	if (rack->rc_in_persist == 0)
6344 		return (0);
6345 	if (ctf_progress_timeout_check(tp, false)) {
6346 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6347 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6348 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6349 		return (-ETIMEDOUT);	/* tcp_drop() */
6350 	}
6351 	/*
6352 	 * Persistence timer into zero window. Force a byte to be output, if
6353 	 * possible.
6354 	 */
6355 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
6356 	/*
6357 	 * Hack: if the peer is dead/unreachable, we do not time out if the
6358 	 * window is closed.  After a full backoff, drop the connection if
6359 	 * the idle time (no responses to probes) reaches the maximum
6360 	 * backoff that we would use if retransmitting.
6361 	 */
6362 	if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
6363 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
6364 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
6365 		KMOD_TCPSTAT_INC(tcps_persistdrop);
6366 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6367 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6368 		retval = -ETIMEDOUT;	/* tcp_drop() */
6369 		goto out;
6370 	}
6371 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
6372 	    tp->snd_una == tp->snd_max)
6373 		rack_exit_persist(tp, rack, cts);
6374 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
6375 	/*
6376 	 * If the user has closed the socket then drop a persisting
6377 	 * connection after a much reduced timeout.
6378 	 */
6379 	if (tp->t_state > TCPS_CLOSE_WAIT &&
6380 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
6381 		KMOD_TCPSTAT_INC(tcps_persistdrop);
6382 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
6383 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
6384 		retval = -ETIMEDOUT;	/* tcp_drop() */
6385 		goto out;
6386 	}
6387 	t_template = tcpip_maketemplate(rack->rc_inp);
6388 	if (t_template) {
6389 		/* only set it if we were answered */
6390 		if (rack->forced_ack == 0) {
6391 			rack->forced_ack = 1;
6392 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6393 		} else {
6394 			rack->probe_not_answered = 1;
6395 			counter_u64_add(rack_persists_loss, 1);
6396 			rack->r_ctl.persist_lost_ends++;
6397 		}
6398 		counter_u64_add(rack_persists_sends, 1);
6399 		tcp_respond(tp, t_template->tt_ipgen,
6400 			    &t_template->tt_t, (struct mbuf *)NULL,
6401 			    tp->rcv_nxt, tp->snd_una - 1, 0);
6402 		/* This sends an ack */
6403 		if (tp->t_flags & TF_DELACK)
6404 			tp->t_flags &= ~TF_DELACK;
6405 		free(t_template, M_TEMP);
6406 	}
6407 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
6408 		tp->t_rxtshift++;
6409 out:
6410 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
6411 	rack_start_hpts_timer(rack, tp, cts,
6412 			      0, 0, 0);
6413 	return (retval);
6414 }
6415 
6416 /*
6417  * If a keepalive goes off, we had no other timers
6418  * happening. We always return 1 here since this
6419  * routine either drops the connection or sends
6420  * out a segment with respond.
6421  */
6422 static int
6423 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6424 {
6425 	struct tcptemp *t_template;
6426 	struct inpcb *inp = tptoinpcb(tp);
6427 
6428 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6429 		return (1);
6430 	}
6431 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
6432 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
6433 	/*
6434 	 * Keep-alive timer went off; send something or drop connection if
6435 	 * idle for too long.
6436 	 */
6437 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
6438 	if (tp->t_state < TCPS_ESTABLISHED)
6439 		goto dropit;
6440 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6441 	    tp->t_state <= TCPS_CLOSING) {
6442 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
6443 			goto dropit;
6444 		/*
6445 		 * Send a packet designed to force a response if the peer is
6446 		 * up and reachable: either an ACK if the connection is
6447 		 * still alive, or an RST if the peer has closed the
6448 		 * connection due to timeout or reboot. Using sequence
6449 		 * number tp->snd_una-1 causes the transmitted zero-length
6450 		 * segment to lie outside the receive window; by the
6451 		 * protocol spec, this requires the correspondent TCP to
6452 		 * respond.
6453 		 */
6454 		KMOD_TCPSTAT_INC(tcps_keepprobe);
6455 		t_template = tcpip_maketemplate(inp);
6456 		if (t_template) {
6457 			if (rack->forced_ack == 0) {
6458 				rack->forced_ack = 1;
6459 				rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
6460 			} else {
6461 				rack->probe_not_answered = 1;
6462 			}
6463 			tcp_respond(tp, t_template->tt_ipgen,
6464 			    &t_template->tt_t, (struct mbuf *)NULL,
6465 			    tp->rcv_nxt, tp->snd_una - 1, 0);
6466 			free(t_template, M_TEMP);
6467 		}
6468 	}
6469 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
6470 	return (1);
6471 dropit:
6472 	KMOD_TCPSTAT_INC(tcps_keepdrops);
6473 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6474 	return (-ETIMEDOUT);	/* tcp_drop() */
6475 }
6476 
6477 /*
6478  * Retransmit helper function, clear up all the ack
6479  * flags and take care of important book keeping.
6480  */
6481 static void
6482 rack_remxt_tmr(struct tcpcb *tp)
6483 {
6484 	/*
6485 	 * The retransmit timer went off, all sack'd blocks must be
6486 	 * un-acked.
6487 	 */
6488 	struct rack_sendmap *rsm, *trsm = NULL;
6489 	struct tcp_rack *rack;
6490 
6491 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6492 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
6493 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
6494 	if (rack->r_state && (rack->r_state != tp->t_state))
6495 		rack_set_state(tp, rack);
6496 	/*
6497 	 * Ideally we would like to be able to
6498 	 * mark SACK-PASS on anything not acked here.
6499 	 *
6500 	 * However, if we do that we would burst out
6501 	 * all that data 1ms apart. This would be unwise,
6502 	 * so for now we will just let the normal rxt timer
6503 	 * and tlp timer take care of it.
6504 	 *
6505 	 * Also we really need to stick them back in sequence
6506 	 * order. This way we send in the proper order and any
6507 	 * sacks that come floating in will "re-ack" the data.
6508 	 * To do this we zap the tmap with an INIT and then
6509 	 * walk through and place every rsm in the RB tree
6510 	 * back in its seq ordered place.
6511 	 */
6512 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
6513 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
6514 		rsm->r_dupack = 0;
6515 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
6516 		/* We must re-add it back to the tlist */
6517 		if (trsm == NULL) {
6518 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6519 		} else {
6520 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
6521 		}
6522 		rsm->r_in_tmap = 1;
6523 		trsm = rsm;
6524 		if (rsm->r_flags & RACK_ACKED)
6525 			rsm->r_flags |= RACK_WAS_ACKED;
6526 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
6527 		rsm->r_flags |= RACK_MUST_RXT;
6528 	}
6529 	/* Clear the count (we just un-acked them) */
6530 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
6531 	rack->r_ctl.rc_sacked = 0;
6532 	rack->r_ctl.rc_sacklast = NULL;
6533 	rack->r_ctl.rc_agg_delayed = 0;
6534 	rack->r_early = 0;
6535 	rack->r_ctl.rc_agg_early = 0;
6536 	rack->r_late = 0;
6537 	/* Clear the tlp rtx mark */
6538 	rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6539 	if (rack->r_ctl.rc_resend != NULL)
6540 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
6541 	rack->r_ctl.rc_prr_sndcnt = 0;
6542 	rack_log_to_prr(rack, 6, 0, __LINE__);
6543 	rack->r_timer_override = 1;
6544 	if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
6545 #ifdef NETFLIX_EXP_DETECTION
6546 	    || (rack->sack_attack_disable != 0)
6547 #endif
6548 		    ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
6549 		/*
6550 		 * For non-sack customers new data
6551 		 * needs to go out as retransmits until
6552 		 * we retransmit up to snd_max.
6553 		 */
6554 		rack->r_must_retran = 1;
6555 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
6556 						rack->r_ctl.rc_sacked);
6557 	}
6558 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
6559 }
6560 
6561 static void
6562 rack_convert_rtts(struct tcpcb *tp)
6563 {
6564 	if (tp->t_srtt > 1) {
6565 		uint32_t val, frac;
6566 
6567 		val = tp->t_srtt >> TCP_RTT_SHIFT;
6568 		frac = tp->t_srtt & 0x1f;
6569 		tp->t_srtt = TICKS_2_USEC(val);
6570 		/*
6571 		 * frac is the fractional part of the srtt (if any)
6572 		 * but its in ticks and every bit represents
6573 		 * 1/32nd of a hz.
6574 		 */
6575 		if (frac) {
6576 			if (hz == 1000) {
6577 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6578 			} else {
6579 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6580 			}
6581 			tp->t_srtt += frac;
6582 		}
6583 	}
6584 	if (tp->t_rttvar) {
6585 		uint32_t val, frac;
6586 
6587 		val = tp->t_rttvar >> TCP_RTTVAR_SHIFT;
6588 		frac = tp->t_rttvar & 0x1f;
6589 		tp->t_rttvar = TICKS_2_USEC(val);
6590 		/*
6591 		 * frac is the fractional part of the srtt (if any)
6592 		 * but its in ticks and every bit represents
6593 		 * 1/32nd of a hz.
6594 		 */
6595 		if (frac) {
6596 			if (hz == 1000) {
6597 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE);
6598 			} else {
6599 				frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE));
6600 			}
6601 			tp->t_rttvar += frac;
6602 		}
6603 	}
6604 	tp->t_rxtcur = RACK_REXMTVAL(tp);
6605 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6606 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
6607 	}
6608 	if (tp->t_rxtcur > rack_rto_max) {
6609 		tp->t_rxtcur = rack_rto_max;
6610 	}
6611 }
6612 
6613 static void
6614 rack_cc_conn_init(struct tcpcb *tp)
6615 {
6616 	struct tcp_rack *rack;
6617 	uint32_t srtt;
6618 
6619 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6620 	srtt = tp->t_srtt;
6621 	cc_conn_init(tp);
6622 	/*
6623 	 * Now convert to rack's internal format,
6624 	 * if required.
6625 	 */
6626 	if ((srtt == 0) && (tp->t_srtt != 0))
6627 		rack_convert_rtts(tp);
6628 	/*
6629 	 * We want a chance to stay in slowstart as
6630 	 * we create a connection. TCP spec says that
6631 	 * initially ssthresh is infinite. For our
6632 	 * purposes that is the snd_wnd.
6633 	 */
6634 	if (tp->snd_ssthresh < tp->snd_wnd) {
6635 		tp->snd_ssthresh = tp->snd_wnd;
6636 	}
6637 	/*
6638 	 * We also want to assure a IW worth of
6639 	 * data can get inflight.
6640 	 */
6641 	if (rc_init_window(rack) < tp->snd_cwnd)
6642 		tp->snd_cwnd = rc_init_window(rack);
6643 }
6644 
6645 /*
6646  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
6647  * we will setup to retransmit the lowest seq number outstanding.
6648  */
6649 static int
6650 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6651 {
6652 	struct inpcb *inp = tptoinpcb(tp);
6653 	int32_t rexmt;
6654 	int32_t retval = 0;
6655 	bool isipv6;
6656 
6657 	if (tp->t_timers->tt_flags & TT_STOPPED) {
6658 		return (1);
6659 	}
6660 	if ((tp->t_flags & TF_GPUTINPROG) &&
6661 	    (tp->t_rxtshift)) {
6662 		/*
6663 		 * We have had a second timeout
6664 		 * measurements on successive rxt's are not profitable.
6665 		 * It is unlikely to be of any use (the network is
6666 		 * broken or the client went away).
6667 		 */
6668 		tp->t_flags &= ~TF_GPUTINPROG;
6669 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6670 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
6671 					   tp->gput_seq,
6672 					   0, 0, 18, __LINE__, NULL, 0);
6673 	}
6674 	if (ctf_progress_timeout_check(tp, false)) {
6675 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6676 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
6677 		return (-ETIMEDOUT);	/* tcp_drop() */
6678 	}
6679 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
6680 	rack->r_ctl.retran_during_recovery = 0;
6681 	rack->rc_ack_required = 1;
6682 	rack->r_ctl.dsack_byte_cnt = 0;
6683 	if (IN_FASTRECOVERY(tp->t_flags))
6684 		tp->t_flags |= TF_WASFRECOVERY;
6685 	else
6686 		tp->t_flags &= ~TF_WASFRECOVERY;
6687 	if (IN_CONGRECOVERY(tp->t_flags))
6688 		tp->t_flags |= TF_WASCRECOVERY;
6689 	else
6690 		tp->t_flags &= ~TF_WASCRECOVERY;
6691 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
6692 	    (tp->snd_una == tp->snd_max)) {
6693 		/* Nothing outstanding .. nothing to do */
6694 		return (0);
6695 	}
6696 	if (rack->r_ctl.dsack_persist) {
6697 		rack->r_ctl.dsack_persist--;
6698 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
6699 			rack->r_ctl.num_dsack = 0;
6700 		}
6701 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
6702 	}
6703 	/*
6704 	 * Rack can only run one timer  at a time, so we cannot
6705 	 * run a KEEPINIT (gating SYN sending) and a retransmit
6706 	 * timer for the SYN. So if we are in a front state and
6707 	 * have a KEEPINIT timer we need to check the first transmit
6708 	 * against now to see if we have exceeded the KEEPINIT time
6709 	 * (if one is set).
6710 	 */
6711 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6712 	    (TP_KEEPINIT(tp) != 0)) {
6713 		struct rack_sendmap *rsm;
6714 
6715 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
6716 		if (rsm) {
6717 			/* Ok we have something outstanding to test keepinit with */
6718 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
6719 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
6720 				/* We have exceeded the KEEPINIT time */
6721 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
6722 				goto drop_it;
6723 			}
6724 		}
6725 	}
6726 	/*
6727 	 * Retransmission timer went off.  Message has not been acked within
6728 	 * retransmit interval.  Back off to a longer retransmit interval
6729 	 * and retransmit one segment.
6730 	 */
6731 	rack_remxt_tmr(tp);
6732 	if ((rack->r_ctl.rc_resend == NULL) ||
6733 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
6734 		/*
6735 		 * If the rwnd collapsed on
6736 		 * the one we are retransmitting
6737 		 * it does not count against the
6738 		 * rxt count.
6739 		 */
6740 		tp->t_rxtshift++;
6741 	}
6742 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT) {
6743 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
6744 drop_it:
6745 		tp->t_rxtshift = TCP_MAXRXTSHIFT;
6746 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
6747 		/* XXXGL: previously t_softerror was casted to uint16_t */
6748 		MPASS(tp->t_softerror >= 0);
6749 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
6750 		goto out;	/* tcp_drop() */
6751 	}
6752 	if (tp->t_state == TCPS_SYN_SENT) {
6753 		/*
6754 		 * If the SYN was retransmitted, indicate CWND to be limited
6755 		 * to 1 segment in cc_conn_init().
6756 		 */
6757 		tp->snd_cwnd = 1;
6758 	} else if (tp->t_rxtshift == 1) {
6759 		/*
6760 		 * first retransmit; record ssthresh and cwnd so they can be
6761 		 * recovered if this turns out to be a "bad" retransmit. A
6762 		 * retransmit is considered "bad" if an ACK for this segment
6763 		 * is received within RTT/2 interval; the assumption here is
6764 		 * that the ACK was already in flight.  See "On Estimating
6765 		 * End-to-End Network Path Properties" by Allman and Paxson
6766 		 * for more details.
6767 		 */
6768 		tp->snd_cwnd_prev = tp->snd_cwnd;
6769 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
6770 		tp->snd_recover_prev = tp->snd_recover;
6771 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
6772 		tp->t_flags |= TF_PREVVALID;
6773 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
6774 		tp->t_flags &= ~TF_PREVVALID;
6775 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
6776 	if ((tp->t_state == TCPS_SYN_SENT) ||
6777 	    (tp->t_state == TCPS_SYN_RECEIVED))
6778 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
6779 	else
6780 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
6781 
6782 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
6783 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
6784 	/*
6785 	 * We enter the path for PLMTUD if connection is established or, if
6786 	 * connection is FIN_WAIT_1 status, reason for the last is that if
6787 	 * amount of data we send is very small, we could send it in couple
6788 	 * of packets and process straight to FIN. In that case we won't
6789 	 * catch ESTABLISHED state.
6790 	 */
6791 #ifdef INET6
6792 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
6793 #else
6794 	isipv6 = false;
6795 #endif
6796 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
6797 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
6798 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
6799 	    ((tp->t_state == TCPS_ESTABLISHED) ||
6800 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
6801 		/*
6802 		 * Idea here is that at each stage of mtu probe (usually,
6803 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
6804 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
6805 		 * should take care of that.
6806 		 */
6807 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
6808 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
6809 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
6810 		    tp->t_rxtshift % 2 == 0)) {
6811 			/*
6812 			 * Enter Path MTU Black-hole Detection mechanism: -
6813 			 * Disable Path MTU Discovery (IP "DF" bit). -
6814 			 * Reduce MTU to lower value than what we negotiated
6815 			 * with peer.
6816 			 */
6817 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
6818 				/* Record that we may have found a black hole. */
6819 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
6820 				/* Keep track of previous MSS. */
6821 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
6822 			}
6823 
6824 			/*
6825 			 * Reduce the MSS to blackhole value or to the
6826 			 * default in an attempt to retransmit.
6827 			 */
6828 #ifdef INET6
6829 			if (isipv6 &&
6830 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
6831 				/* Use the sysctl tuneable blackhole MSS. */
6832 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
6833 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6834 			} else if (isipv6) {
6835 				/* Use the default MSS. */
6836 				tp->t_maxseg = V_tcp_v6mssdflt;
6837 				/*
6838 				 * Disable Path MTU Discovery when we switch
6839 				 * to minmss.
6840 				 */
6841 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6842 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6843 			}
6844 #endif
6845 #if defined(INET6) && defined(INET)
6846 			else
6847 #endif
6848 #ifdef INET
6849 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
6850 				/* Use the sysctl tuneable blackhole MSS. */
6851 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
6852 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
6853 			} else {
6854 				/* Use the default MSS. */
6855 				tp->t_maxseg = V_tcp_mssdflt;
6856 				/*
6857 				 * Disable Path MTU Discovery when we switch
6858 				 * to minmss.
6859 				 */
6860 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
6861 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
6862 			}
6863 #endif
6864 		} else {
6865 			/*
6866 			 * If further retransmissions are still unsuccessful
6867 			 * with a lowered MTU, maybe this isn't a blackhole
6868 			 * and we restore the previous MSS and blackhole
6869 			 * detection flags. The limit '6' is determined by
6870 			 * giving each probe stage (1448, 1188, 524) 2
6871 			 * chances to recover.
6872 			 */
6873 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
6874 			    (tp->t_rxtshift >= 6)) {
6875 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
6876 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
6877 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
6878 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
6879 			}
6880 		}
6881 	}
6882 	/*
6883 	 * Disable RFC1323 and SACK if we haven't got any response to
6884 	 * our third SYN to work-around some broken terminal servers
6885 	 * (most of which have hopefully been retired) that have bad VJ
6886 	 * header compression code which trashes TCP segments containing
6887 	 * unknown-to-them TCP options.
6888 	 */
6889 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
6890 	    (tp->t_rxtshift == 3))
6891 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
6892 	/*
6893 	 * If we backed off this far, our srtt estimate is probably bogus.
6894 	 * Clobber it so we'll take the next rtt measurement as our srtt;
6895 	 * move the current srtt into rttvar to keep the current retransmit
6896 	 * times until then.
6897 	 */
6898 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
6899 #ifdef INET6
6900 		if ((inp->inp_vflag & INP_IPV6) != 0)
6901 			in6_losing(inp);
6902 		else
6903 #endif
6904 			in_losing(inp);
6905 		tp->t_rttvar += tp->t_srtt;
6906 		tp->t_srtt = 0;
6907 	}
6908 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
6909 	tp->snd_recover = tp->snd_max;
6910 	tp->t_flags |= TF_ACKNOW;
6911 	tp->t_rtttime = 0;
6912 	rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
6913 out:
6914 	return (retval);
6915 }
6916 
6917 static int
6918 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
6919 {
6920 	int32_t ret = 0;
6921 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
6922 
6923 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
6924 	    (tp->t_flags & TF_GPUTINPROG)) {
6925 		/*
6926 		 * We have a goodput in progress
6927 		 * and we have entered a late state.
6928 		 * Do we have enough data in the sb
6929 		 * to handle the GPUT request?
6930 		 */
6931 		uint32_t bytes;
6932 
6933 		bytes = tp->gput_ack - tp->gput_seq;
6934 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
6935 			bytes += tp->gput_seq - tp->snd_una;
6936 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
6937 			/*
6938 			 * There are not enough bytes in the socket
6939 			 * buffer that have been sent to cover this
6940 			 * measurement. Cancel it.
6941 			 */
6942 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
6943 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
6944 						   tp->gput_seq,
6945 						   0, 0, 18, __LINE__, NULL, 0);
6946 			tp->t_flags &= ~TF_GPUTINPROG;
6947 		}
6948 	}
6949 	if (timers == 0) {
6950 		return (0);
6951 	}
6952 	if (tp->t_state == TCPS_LISTEN) {
6953 		/* no timers on listen sockets */
6954 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
6955 			return (0);
6956 		return (1);
6957 	}
6958 	if ((timers & PACE_TMR_RACK) &&
6959 	    rack->rc_on_min_to) {
6960 		/*
6961 		 * For the rack timer when we
6962 		 * are on a min-timeout (which means rrr_conf = 3)
6963 		 * we don't want to check the timer. It may
6964 		 * be going off for a pace and thats ok we
6965 		 * want to send the retransmit (if its ready).
6966 		 *
6967 		 * If its on a normal rack timer (non-min) then
6968 		 * we will check if its expired.
6969 		 */
6970 		goto skip_time_check;
6971 	}
6972 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
6973 		uint32_t left;
6974 
6975 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
6976 			ret = -1;
6977 			rack_log_to_processing(rack, cts, ret, 0);
6978 			return (0);
6979 		}
6980 		if (hpts_calling == 0) {
6981 			/*
6982 			 * A user send or queued mbuf (sack) has called us? We
6983 			 * return 0 and let the pacing guards
6984 			 * deal with it if they should or
6985 			 * should not cause a send.
6986 			 */
6987 			ret = -2;
6988 			rack_log_to_processing(rack, cts, ret, 0);
6989 			return (0);
6990 		}
6991 		/*
6992 		 * Ok our timer went off early and we are not paced false
6993 		 * alarm, go back to sleep.
6994 		 */
6995 		ret = -3;
6996 		left = rack->r_ctl.rc_timer_exp - cts;
6997 		tcp_hpts_insert(tptoinpcb(tp), HPTS_MS_TO_SLOTS(left));
6998 		rack_log_to_processing(rack, cts, ret, left);
6999 		return (1);
7000 	}
7001 skip_time_check:
7002 	rack->rc_tmr_stopped = 0;
7003 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
7004 	if (timers & PACE_TMR_DELACK) {
7005 		ret = rack_timeout_delack(tp, rack, cts);
7006 	} else if (timers & PACE_TMR_RACK) {
7007 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
7008 		rack->r_fast_output = 0;
7009 		ret = rack_timeout_rack(tp, rack, cts);
7010 	} else if (timers & PACE_TMR_TLP) {
7011 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
7012 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
7013 	} else if (timers & PACE_TMR_RXT) {
7014 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
7015 		rack->r_fast_output = 0;
7016 		ret = rack_timeout_rxt(tp, rack, cts);
7017 	} else if (timers & PACE_TMR_PERSIT) {
7018 		ret = rack_timeout_persist(tp, rack, cts);
7019 	} else if (timers & PACE_TMR_KEEP) {
7020 		ret = rack_timeout_keepalive(tp, rack, cts);
7021 	}
7022 	rack_log_to_processing(rack, cts, ret, timers);
7023 	return (ret);
7024 }
7025 
7026 static void
7027 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
7028 {
7029 	struct timeval tv;
7030 	uint32_t us_cts, flags_on_entry;
7031 	uint8_t hpts_removed = 0;
7032 
7033 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
7034 	us_cts = tcp_get_usecs(&tv);
7035 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
7036 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
7037 	     ((tp->snd_max - tp->snd_una) == 0))) {
7038 		tcp_hpts_remove(rack->rc_inp);
7039 		hpts_removed = 1;
7040 		/* If we were not delayed cancel out the flag. */
7041 		if ((tp->snd_max - tp->snd_una) == 0)
7042 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7043 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7044 	}
7045 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
7046 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
7047 		if (tcp_in_hpts(rack->rc_inp) &&
7048 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
7049 			/*
7050 			 * Canceling timer's when we have no output being
7051 			 * paced. We also must remove ourselves from the
7052 			 * hpts.
7053 			 */
7054 			tcp_hpts_remove(rack->rc_inp);
7055 			hpts_removed = 1;
7056 		}
7057 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
7058 	}
7059 	if (hpts_removed == 0)
7060 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
7061 }
7062 
7063 static void
7064 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
7065 {
7066 	return;
7067 }
7068 
7069 static int
7070 rack_stopall(struct tcpcb *tp)
7071 {
7072 	struct tcp_rack *rack;
7073 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7074 	rack->t_timers_stopped = 1;
7075 	return (0);
7076 }
7077 
7078 static void
7079 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
7080 {
7081 	return;
7082 }
7083 
7084 static int
7085 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
7086 {
7087 	return (0);
7088 }
7089 
7090 static void
7091 rack_stop_all_timers(struct tcpcb *tp)
7092 {
7093 	struct tcp_rack *rack;
7094 
7095 	/*
7096 	 * Assure no timers are running.
7097 	 */
7098 	if (tcp_timer_active(tp, TT_PERSIST)) {
7099 		/* We enter in persists, set the flag appropriately */
7100 		rack = (struct tcp_rack *)tp->t_fb_ptr;
7101 		rack->rc_in_persist = 1;
7102 	}
7103 	tcp_timer_suspend(tp, TT_PERSIST);
7104 	tcp_timer_suspend(tp, TT_REXMT);
7105 	tcp_timer_suspend(tp, TT_KEEP);
7106 	tcp_timer_suspend(tp, TT_DELACK);
7107 }
7108 
7109 static void
7110 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
7111     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag)
7112 {
7113 	int32_t idx;
7114 
7115 	rsm->r_rtr_cnt++;
7116 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7117 	rsm->r_dupack = 0;
7118 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
7119 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
7120 		rsm->r_flags |= RACK_OVERMAX;
7121 	}
7122 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
7123 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
7124 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
7125 	}
7126 	idx = rsm->r_rtr_cnt - 1;
7127 	rsm->r_tim_lastsent[idx] = ts;
7128 	/*
7129 	 * Here we don't add in the len of send, since its already
7130 	 * in snduna <->snd_max.
7131 	 */
7132 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
7133 				     rack->r_ctl.rc_sacked);
7134 	if (rsm->r_flags & RACK_ACKED) {
7135 		/* Problably MTU discovery messing with us */
7136 		rsm->r_flags &= ~RACK_ACKED;
7137 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
7138 	}
7139 	if (rsm->r_in_tmap) {
7140 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7141 		rsm->r_in_tmap = 0;
7142 	}
7143 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7144 	rsm->r_in_tmap = 1;
7145 	/* Take off the must retransmit flag, if its on */
7146 	if (rsm->r_flags & RACK_MUST_RXT) {
7147 		if (rack->r_must_retran)
7148 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
7149 		if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
7150 			/*
7151 			 * We have retransmitted all we need. Clear
7152 			 * any must retransmit flags.
7153 			 */
7154 			rack->r_must_retran = 0;
7155 			rack->r_ctl.rc_out_at_rto = 0;
7156 		}
7157 		rsm->r_flags &= ~RACK_MUST_RXT;
7158 	}
7159 	if (rsm->r_flags & RACK_SACK_PASSED) {
7160 		/* We have retransmitted due to the SACK pass */
7161 		rsm->r_flags &= ~RACK_SACK_PASSED;
7162 		rsm->r_flags |= RACK_WAS_SACKPASS;
7163 	}
7164 }
7165 
7166 static uint32_t
7167 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
7168     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag)
7169 {
7170 	/*
7171 	 * We (re-)transmitted starting at rsm->r_start for some length
7172 	 * (possibly less than r_end.
7173 	 */
7174 	struct rack_sendmap *nrsm;
7175 #ifdef INVARIANTS
7176 	struct rack_sendmap *insret;
7177 #endif
7178 	uint32_t c_end;
7179 	int32_t len;
7180 
7181 	len = *lenp;
7182 	c_end = rsm->r_start + len;
7183 	if (SEQ_GEQ(c_end, rsm->r_end)) {
7184 		/*
7185 		 * We retransmitted the whole piece or more than the whole
7186 		 * slopping into the next rsm.
7187 		 */
7188 		rack_update_rsm(tp, rack, rsm, ts, add_flag);
7189 		if (c_end == rsm->r_end) {
7190 			*lenp = 0;
7191 			return (0);
7192 		} else {
7193 			int32_t act_len;
7194 
7195 			/* Hangs over the end return whats left */
7196 			act_len = rsm->r_end - rsm->r_start;
7197 			*lenp = (len - act_len);
7198 			return (rsm->r_end);
7199 		}
7200 		/* We don't get out of this block. */
7201 	}
7202 	/*
7203 	 * Here we retransmitted less than the whole thing which means we
7204 	 * have to split this into what was transmitted and what was not.
7205 	 */
7206 	nrsm = rack_alloc_full_limit(rack);
7207 	if (nrsm == NULL) {
7208 		/*
7209 		 * We can't get memory, so lets not proceed.
7210 		 */
7211 		*lenp = 0;
7212 		return (0);
7213 	}
7214 	/*
7215 	 * So here we are going to take the original rsm and make it what we
7216 	 * retransmitted. nrsm will be the tail portion we did not
7217 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
7218 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
7219 	 * 1, 6 and the new piece will be 6, 11.
7220 	 */
7221 	rack_clone_rsm(rack, nrsm, rsm, c_end);
7222 	nrsm->r_dupack = 0;
7223 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
7224 #ifndef INVARIANTS
7225 	(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7226 #else
7227 	insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7228 	if (insret != NULL) {
7229 		panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7230 		      nrsm, insret, rack, rsm);
7231 	}
7232 #endif
7233 	if (rsm->r_in_tmap) {
7234 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7235 		nrsm->r_in_tmap = 1;
7236 	}
7237 	rsm->r_flags &= (~RACK_HAS_FIN);
7238 	rack_update_rsm(tp, rack, rsm, ts, add_flag);
7239 	/* Log a split of rsm into rsm and nrsm */
7240 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7241 	*lenp = 0;
7242 	return (0);
7243 }
7244 
7245 static void
7246 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
7247 		uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
7248 		struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls)
7249 {
7250 	struct tcp_rack *rack;
7251 	struct rack_sendmap *rsm, *nrsm, fe;
7252 #ifdef INVARIANTS
7253 	struct rack_sendmap *insret;
7254 #endif
7255 	register uint32_t snd_max, snd_una;
7256 
7257 	/*
7258 	 * Add to the RACK log of packets in flight or retransmitted. If
7259 	 * there is a TS option we will use the TS echoed, if not we will
7260 	 * grab a TS.
7261 	 *
7262 	 * Retransmissions will increment the count and move the ts to its
7263 	 * proper place. Note that if options do not include TS's then we
7264 	 * won't be able to effectively use the ACK for an RTT on a retran.
7265 	 *
7266 	 * Notes about r_start and r_end. Lets consider a send starting at
7267 	 * sequence 1 for 10 bytes. In such an example the r_start would be
7268 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
7269 	 * This means that r_end is actually the first sequence for the next
7270 	 * slot (11).
7271 	 *
7272 	 */
7273 	/*
7274 	 * If err is set what do we do XXXrrs? should we not add the thing?
7275 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
7276 	 * i.e. proceed with add ** do this for now.
7277 	 */
7278 	INP_WLOCK_ASSERT(tptoinpcb(tp));
7279 	if (err)
7280 		/*
7281 		 * We don't log errors -- we could but snd_max does not
7282 		 * advance in this case either.
7283 		 */
7284 		return;
7285 
7286 	if (th_flags & TH_RST) {
7287 		/*
7288 		 * We don't log resets and we return immediately from
7289 		 * sending
7290 		 */
7291 		return;
7292 	}
7293 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7294 	snd_una = tp->snd_una;
7295 	snd_max = tp->snd_max;
7296 	if (th_flags & (TH_SYN | TH_FIN)) {
7297 		/*
7298 		 * The call to rack_log_output is made before bumping
7299 		 * snd_max. This means we can record one extra byte on a SYN
7300 		 * or FIN if seq_out is adding more on and a FIN is present
7301 		 * (and we are not resending).
7302 		 */
7303 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
7304 			len++;
7305 		if (th_flags & TH_FIN)
7306 			len++;
7307 		if (SEQ_LT(snd_max, tp->snd_nxt)) {
7308 			/*
7309 			 * The add/update as not been done for the FIN/SYN
7310 			 * yet.
7311 			 */
7312 			snd_max = tp->snd_nxt;
7313 		}
7314 	}
7315 	if (SEQ_LEQ((seq_out + len), snd_una)) {
7316 		/* Are sending an old segment to induce an ack (keep-alive)? */
7317 		return;
7318 	}
7319 	if (SEQ_LT(seq_out, snd_una)) {
7320 		/* huh? should we panic? */
7321 		uint32_t end;
7322 
7323 		end = seq_out + len;
7324 		seq_out = snd_una;
7325 		if (SEQ_GEQ(end, seq_out))
7326 			len = end - seq_out;
7327 		else
7328 			len = 0;
7329 	}
7330 	if (len == 0) {
7331 		/* We don't log zero window probes */
7332 		return;
7333 	}
7334 	if (IN_FASTRECOVERY(tp->t_flags)) {
7335 		rack->r_ctl.rc_prr_out += len;
7336 	}
7337 	/* First question is it a retransmission or new? */
7338 	if (seq_out == snd_max) {
7339 		/* Its new */
7340 again:
7341 		rsm = rack_alloc(rack);
7342 		if (rsm == NULL) {
7343 			/*
7344 			 * Hmm out of memory and the tcb got destroyed while
7345 			 * we tried to wait.
7346 			 */
7347 			return;
7348 		}
7349 		if (th_flags & TH_FIN) {
7350 			rsm->r_flags = RACK_HAS_FIN|add_flag;
7351 		} else {
7352 			rsm->r_flags = add_flag;
7353 		}
7354 		if (hw_tls)
7355 			rsm->r_hw_tls = 1;
7356 		rsm->r_tim_lastsent[0] = cts;
7357 		rsm->r_rtr_cnt = 1;
7358 		rsm->r_rtr_bytes = 0;
7359 		if (th_flags & TH_SYN) {
7360 			/* The data space is one beyond snd_una */
7361 			rsm->r_flags |= RACK_HAS_SYN;
7362 		}
7363 		rsm->r_start = seq_out;
7364 		rsm->r_end = rsm->r_start + len;
7365 		rsm->r_dupack = 0;
7366 		/*
7367 		 * save off the mbuf location that
7368 		 * sndmbuf_noadv returned (which is
7369 		 * where we started copying from)..
7370 		 */
7371 		rsm->m = s_mb;
7372 		rsm->soff = s_moff;
7373 		/*
7374 		 * Here we do add in the len of send, since its not yet
7375 		 * reflected in in snduna <->snd_max
7376 		 */
7377 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
7378 					      rack->r_ctl.rc_sacked) +
7379 			      (rsm->r_end - rsm->r_start));
7380 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
7381 		if (rsm->m) {
7382 			if (rsm->m->m_len <= rsm->soff) {
7383 				/*
7384 				 * XXXrrs Question, will this happen?
7385 				 *
7386 				 * If sbsndptr is set at the correct place
7387 				 * then s_moff should always be somewhere
7388 				 * within rsm->m. But if the sbsndptr was
7389 				 * off then that won't be true. If it occurs
7390 				 * we need to walkout to the correct location.
7391 				 */
7392 				struct mbuf *lm;
7393 
7394 				lm = rsm->m;
7395 				while (lm->m_len <= rsm->soff) {
7396 					rsm->soff -= lm->m_len;
7397 					lm = lm->m_next;
7398 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
7399 							     __func__, rack, s_moff, s_mb, rsm->soff));
7400 				}
7401 				rsm->m = lm;
7402 			}
7403 			rsm->orig_m_len = rsm->m->m_len;
7404 		} else
7405 			rsm->orig_m_len = 0;
7406 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7407 		/* Log a new rsm */
7408 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
7409 #ifndef INVARIANTS
7410 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7411 #else
7412 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7413 		if (insret != NULL) {
7414 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7415 			      nrsm, insret, rack, rsm);
7416 		}
7417 #endif
7418 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7419 		rsm->r_in_tmap = 1;
7420 		/*
7421 		 * Special case detection, is there just a single
7422 		 * packet outstanding when we are not in recovery?
7423 		 *
7424 		 * If this is true mark it so.
7425 		 */
7426 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
7427 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
7428 			struct rack_sendmap *prsm;
7429 
7430 			prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
7431 			if (prsm)
7432 				prsm->r_one_out_nr = 1;
7433 		}
7434 		return;
7435 	}
7436 	/*
7437 	 * If we reach here its a retransmission and we need to find it.
7438 	 */
7439 	memset(&fe, 0, sizeof(fe));
7440 more:
7441 	if (hintrsm && (hintrsm->r_start == seq_out)) {
7442 		rsm = hintrsm;
7443 		hintrsm = NULL;
7444 	} else {
7445 		/* No hints sorry */
7446 		rsm = NULL;
7447 	}
7448 	if ((rsm) && (rsm->r_start == seq_out)) {
7449 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7450 		if (len == 0) {
7451 			return;
7452 		} else {
7453 			goto more;
7454 		}
7455 	}
7456 	/* Ok it was not the last pointer go through it the hard way. */
7457 refind:
7458 	fe.r_start = seq_out;
7459 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
7460 	if (rsm) {
7461 		if (rsm->r_start == seq_out) {
7462 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag);
7463 			if (len == 0) {
7464 				return;
7465 			} else {
7466 				goto refind;
7467 			}
7468 		}
7469 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
7470 			/* Transmitted within this piece */
7471 			/*
7472 			 * Ok we must split off the front and then let the
7473 			 * update do the rest
7474 			 */
7475 			nrsm = rack_alloc_full_limit(rack);
7476 			if (nrsm == NULL) {
7477 				rack_update_rsm(tp, rack, rsm, cts, add_flag);
7478 				return;
7479 			}
7480 			/*
7481 			 * copy rsm to nrsm and then trim the front of rsm
7482 			 * to not include this part.
7483 			 */
7484 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
7485 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7486 #ifndef INVARIANTS
7487 			(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7488 #else
7489 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
7490 			if (insret != NULL) {
7491 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
7492 				      nrsm, insret, rack, rsm);
7493 			}
7494 #endif
7495 			if (rsm->r_in_tmap) {
7496 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7497 				nrsm->r_in_tmap = 1;
7498 			}
7499 			rsm->r_flags &= (~RACK_HAS_FIN);
7500 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag);
7501 			if (len == 0) {
7502 				return;
7503 			} else if (len > 0)
7504 				goto refind;
7505 		}
7506 	}
7507 	/*
7508 	 * Hmm not found in map did they retransmit both old and on into the
7509 	 * new?
7510 	 */
7511 	if (seq_out == tp->snd_max) {
7512 		goto again;
7513 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
7514 #ifdef INVARIANTS
7515 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
7516 		       seq_out, len, tp->snd_una, tp->snd_max);
7517 		printf("Starting Dump of all rack entries\n");
7518 		RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
7519 			printf("rsm:%p start:%u end:%u\n",
7520 			       rsm, rsm->r_start, rsm->r_end);
7521 		}
7522 		printf("Dump complete\n");
7523 		panic("seq_out not found rack:%p tp:%p",
7524 		      rack, tp);
7525 #endif
7526 	} else {
7527 #ifdef INVARIANTS
7528 		/*
7529 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
7530 		 * flag)
7531 		 */
7532 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
7533 		      seq_out, len, tp->snd_max, tp);
7534 #endif
7535 	}
7536 }
7537 
7538 /*
7539  * Record one of the RTT updates from an ack into
7540  * our sample structure.
7541  */
7542 
7543 static void
7544 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
7545 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
7546 {
7547 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7548 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
7549 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
7550 	}
7551 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7552 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
7553 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
7554 	}
7555 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
7556 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
7557 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
7558 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
7559 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
7560 	}
7561 	if ((confidence == 1) &&
7562 	    ((rsm == NULL) ||
7563 	     (rsm->r_just_ret) ||
7564 	     (rsm->r_one_out_nr &&
7565 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
7566 		/*
7567 		 * If the rsm had a just return
7568 		 * hit it then we can't trust the
7569 		 * rtt measurement for buffer deterimination
7570 		 * Note that a confidence of 2, indicates
7571 		 * SACK'd which overrides the r_just_ret or
7572 		 * the r_one_out_nr. If it was a CUM-ACK and
7573 		 * we had only two outstanding, but get an
7574 		 * ack for only 1. Then that also lowers our
7575 		 * confidence.
7576 		 */
7577 		confidence = 0;
7578 	}
7579 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
7580 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
7581 		if (rack->r_ctl.rack_rs.confidence == 0) {
7582 			/*
7583 			 * We take anything with no current confidence
7584 			 * saved.
7585 			 */
7586 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7587 			rack->r_ctl.rack_rs.confidence = confidence;
7588 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7589 		} else if (confidence || rack->r_ctl.rack_rs.confidence) {
7590 			/*
7591 			 * Once we have a confident number,
7592 			 * we can update it with a smaller
7593 			 * value since this confident number
7594 			 * may include the DSACK time until
7595 			 * the next segment (the second one) arrived.
7596 			 */
7597 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
7598 			rack->r_ctl.rack_rs.confidence = confidence;
7599 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
7600 		}
7601 	}
7602 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
7603 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
7604 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
7605 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
7606 }
7607 
7608 /*
7609  * Collect new round-trip time estimate
7610  * and update averages and current timeout.
7611  */
7612 static void
7613 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
7614 {
7615 	int32_t delta;
7616 	int32_t rtt;
7617 
7618 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
7619 		/* No valid sample */
7620 		return;
7621 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
7622 		/* We are to use the lowest RTT seen in a single ack */
7623 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
7624 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
7625 		/* We are to use the highest RTT seen in a single ack */
7626 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
7627 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
7628 		/* We are to use the average RTT seen in a single ack */
7629 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
7630 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
7631 	} else {
7632 #ifdef INVARIANTS
7633 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
7634 #endif
7635 		return;
7636 	}
7637 	if (rtt == 0)
7638 		rtt = 1;
7639 	if (rack->rc_gp_rtt_set == 0) {
7640 		/*
7641 		 * With no RTT we have to accept
7642 		 * even one we are not confident of.
7643 		 */
7644 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
7645 		rack->rc_gp_rtt_set = 1;
7646 	} else if (rack->r_ctl.rack_rs.confidence) {
7647 		/* update the running gp srtt */
7648 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
7649 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
7650 	}
7651 	if (rack->r_ctl.rack_rs.confidence) {
7652 		/*
7653 		 * record the low and high for highly buffered path computation,
7654 		 * we only do this if we are confident (not a retransmission).
7655 		 */
7656 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
7657 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7658 		}
7659 		if (rack->rc_highly_buffered == 0) {
7660 			/*
7661 			 * Currently once we declare a path has
7662 			 * highly buffered there is no going
7663 			 * back, which may be a problem...
7664 			 */
7665 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
7666 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
7667 						     rack->r_ctl.rc_highest_us_rtt,
7668 						     rack->r_ctl.rc_lowest_us_rtt,
7669 						     RACK_RTTS_SEEHBP);
7670 				rack->rc_highly_buffered = 1;
7671 			}
7672 		}
7673 	}
7674 	if ((rack->r_ctl.rack_rs.confidence) ||
7675 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
7676 		/*
7677 		 * If we are highly confident of it <or> it was
7678 		 * never retransmitted we accept it as the last us_rtt.
7679 		 */
7680 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7681 		/* The lowest rtt can be set if its was not retransmited */
7682 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
7683 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
7684 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
7685 				rack->r_ctl.rc_lowest_us_rtt = 1;
7686 		}
7687 	}
7688 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7689 	if (tp->t_srtt != 0) {
7690 		/*
7691 		 * We keep a simple srtt in microseconds, like our rtt
7692 		 * measurement. We don't need to do any tricks with shifting
7693 		 * etc. Instead we just add in 1/8th of the new measurement
7694 		 * and subtract out 1/8 of the old srtt. We do the same with
7695 		 * the variance after finding the absolute value of the
7696 		 * difference between this sample and the current srtt.
7697 		 */
7698 		delta = tp->t_srtt - rtt;
7699 		/* Take off 1/8th of the current sRTT */
7700 		tp->t_srtt -= (tp->t_srtt >> 3);
7701 		/* Add in 1/8th of the new RTT just measured */
7702 		tp->t_srtt += (rtt >> 3);
7703 		if (tp->t_srtt <= 0)
7704 			tp->t_srtt = 1;
7705 		/* Now lets make the absolute value of the variance */
7706 		if (delta < 0)
7707 			delta = -delta;
7708 		/* Subtract out 1/8th */
7709 		tp->t_rttvar -= (tp->t_rttvar >> 3);
7710 		/* Add in 1/8th of the new variance we just saw */
7711 		tp->t_rttvar += (delta >> 3);
7712 		if (tp->t_rttvar <= 0)
7713 			tp->t_rttvar = 1;
7714 		if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
7715 			tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7716 	} else {
7717 		/*
7718 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
7719 		 * variance to half the rtt (so our first retransmit happens
7720 		 * at 3*rtt).
7721 		 */
7722 		tp->t_srtt = rtt;
7723 		tp->t_rttvar = rtt >> 1;
7724 		tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
7725 	}
7726 	rack->rc_srtt_measure_made = 1;
7727 	KMOD_TCPSTAT_INC(tcps_rttupdated);
7728 	tp->t_rttupdated++;
7729 #ifdef STATS
7730 	if (rack_stats_gets_ms_rtt == 0) {
7731 		/* Send in the microsecond rtt used for rxt timeout purposes */
7732 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
7733 	} else if (rack_stats_gets_ms_rtt == 1) {
7734 		/* Send in the millisecond rtt used for rxt timeout purposes */
7735 		int32_t ms_rtt;
7736 
7737 		/* Round up */
7738 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7739 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7740 	} else if (rack_stats_gets_ms_rtt == 2) {
7741 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
7742 		int32_t ms_rtt;
7743 
7744 		/* Round up */
7745 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
7746 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
7747 	}  else {
7748 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
7749 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
7750 	}
7751 
7752 #endif
7753 	/*
7754 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
7755 	 * way we do the smoothing, srtt and rttvar will each average +1/2
7756 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
7757 	 * tick of rounding and 1 extra tick because of +-1/2 tick
7758 	 * uncertainty in the firing of the timer.  The bias will give us
7759 	 * exactly the 1.5 tick we need.  But, because the bias is
7760 	 * statistical, we have to test that we don't drop below the minimum
7761 	 * feasible timer (which is 2 ticks).
7762 	 */
7763 	tp->t_rxtshift = 0;
7764 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7765 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
7766 	rack_log_rtt_sample(rack, rtt);
7767 	tp->t_softerror = 0;
7768 }
7769 
7770 
7771 static void
7772 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
7773 {
7774 	/*
7775 	 * Apply to filter the inbound us-rtt at us_cts.
7776 	 */
7777 	uint32_t old_rtt;
7778 
7779 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
7780 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
7781 			       us_rtt, us_cts);
7782 	if (old_rtt > us_rtt) {
7783 		/* We just hit a new lower rtt time */
7784 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
7785 				     __LINE__, RACK_RTTS_NEWRTT);
7786 		/*
7787 		 * Only count it if its lower than what we saw within our
7788 		 * calculated range.
7789 		 */
7790 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
7791 			if (rack_probertt_lower_within &&
7792 			    rack->rc_gp_dyn_mul &&
7793 			    (rack->use_fixed_rate == 0) &&
7794 			    (rack->rc_always_pace)) {
7795 				/*
7796 				 * We are seeing a new lower rtt very close
7797 				 * to the time that we would have entered probe-rtt.
7798 				 * This is probably due to the fact that a peer flow
7799 				 * has entered probe-rtt. Lets go in now too.
7800 				 */
7801 				uint32_t val;
7802 
7803 				val = rack_probertt_lower_within * rack_time_between_probertt;
7804 				val /= 100;
7805 				if ((rack->in_probe_rtt == 0)  &&
7806 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
7807 					rack_enter_probertt(rack, us_cts);
7808 				}
7809 			}
7810 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
7811 		}
7812 	}
7813 }
7814 
7815 static int
7816 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
7817     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
7818 {
7819 	uint32_t us_rtt;
7820 	int32_t i, all;
7821 	uint32_t t, len_acked;
7822 
7823 	if ((rsm->r_flags & RACK_ACKED) ||
7824 	    (rsm->r_flags & RACK_WAS_ACKED))
7825 		/* Already done */
7826 		return (0);
7827 	if (rsm->r_no_rtt_allowed) {
7828 		/* Not allowed */
7829 		return (0);
7830 	}
7831 	if (ack_type == CUM_ACKED) {
7832 		if (SEQ_GT(th_ack, rsm->r_end)) {
7833 			len_acked = rsm->r_end - rsm->r_start;
7834 			all = 1;
7835 		} else {
7836 			len_acked = th_ack - rsm->r_start;
7837 			all = 0;
7838 		}
7839 	} else {
7840 		len_acked = rsm->r_end - rsm->r_start;
7841 		all = 0;
7842 	}
7843 	if (rsm->r_rtr_cnt == 1) {
7844 
7845 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7846 		if ((int)t <= 0)
7847 			t = 1;
7848 		if (!tp->t_rttlow || tp->t_rttlow > t)
7849 			tp->t_rttlow = t;
7850 		if (!rack->r_ctl.rc_rack_min_rtt ||
7851 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7852 			rack->r_ctl.rc_rack_min_rtt = t;
7853 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
7854 				rack->r_ctl.rc_rack_min_rtt = 1;
7855 			}
7856 		}
7857 		if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
7858 			us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7859 		else
7860 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
7861 		if (us_rtt == 0)
7862 			us_rtt = 1;
7863 		if (CC_ALGO(tp)->rttsample != NULL) {
7864 			/* Kick the RTT to the CC */
7865 			CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas);
7866 		}
7867 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
7868 		if (ack_type == SACKED) {
7869 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
7870 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
7871 		} else {
7872 			/*
7873 			 * We need to setup what our confidence
7874 			 * is in this ack.
7875 			 *
7876 			 * If the rsm was app limited and it is
7877 			 * less than a mss in length (the end
7878 			 * of the send) then we have a gap. If we
7879 			 * were app limited but say we were sending
7880 			 * multiple MSS's then we are more confident
7881 			 * int it.
7882 			 *
7883 			 * When we are not app-limited then we see if
7884 			 * the rsm is being included in the current
7885 			 * measurement, we tell this by the app_limited_needs_set
7886 			 * flag.
7887 			 *
7888 			 * Note that being cwnd blocked is not applimited
7889 			 * as well as the pacing delay between packets which
7890 			 * are sending only 1 or 2 MSS's also will show up
7891 			 * in the RTT. We probably need to examine this algorithm
7892 			 * a bit more and enhance it to account for the delay
7893 			 * between rsm's. We could do that by saving off the
7894 			 * pacing delay of each rsm (in an rsm) and then
7895 			 * factoring that in somehow though for now I am
7896 			 * not sure how :)
7897 			 */
7898 			int calc_conf = 0;
7899 
7900 			if (rsm->r_flags & RACK_APP_LIMITED) {
7901 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
7902 					calc_conf = 0;
7903 				else
7904 					calc_conf = 1;
7905 			} else if (rack->app_limited_needs_set == 0) {
7906 				calc_conf = 1;
7907 			} else {
7908 				calc_conf = 0;
7909 			}
7910 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
7911 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
7912 					    calc_conf, rsm, rsm->r_rtr_cnt);
7913 		}
7914 		if ((rsm->r_flags & RACK_TLP) &&
7915 		    (!IN_FASTRECOVERY(tp->t_flags))) {
7916 			/* Segment was a TLP and our retrans matched */
7917 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
7918 				rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
7919 			}
7920 		}
7921 		if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7922 			/* New more recent rack_tmit_time */
7923 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7924 			rack->rc_rack_rtt = t;
7925 		}
7926 		return (1);
7927 	}
7928 	/*
7929 	 * We clear the soft/rxtshift since we got an ack.
7930 	 * There is no assurance we will call the commit() function
7931 	 * so we need to clear these to avoid incorrect handling.
7932 	 */
7933 	tp->t_rxtshift = 0;
7934 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
7935 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
7936 	tp->t_softerror = 0;
7937 	if (to && (to->to_flags & TOF_TS) &&
7938 	    (ack_type == CUM_ACKED) &&
7939 	    (to->to_tsecr) &&
7940 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
7941 		/*
7942 		 * Now which timestamp does it match? In this block the ACK
7943 		 * must be coming from a previous transmission.
7944 		 */
7945 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
7946 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
7947 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
7948 				if ((int)t <= 0)
7949 					t = 1;
7950 				if (CC_ALGO(tp)->rttsample != NULL) {
7951 					/*
7952 					 * Kick the RTT to the CC, here
7953 					 * we lie a bit in that we know the
7954 					 * retransmission is correct even though
7955 					 * we retransmitted. This is because
7956 					 * we match the timestamps.
7957 					 */
7958 					if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
7959 						us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
7960 					else
7961 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
7962 					CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas);
7963 				}
7964 				if ((i + 1) < rsm->r_rtr_cnt) {
7965 					/*
7966 					 * The peer ack'd from our previous
7967 					 * transmission. We have a spurious
7968 					 * retransmission and thus we dont
7969 					 * want to update our rack_rtt.
7970 					 *
7971 					 * Hmm should there be a CC revert here?
7972 					 *
7973 					 */
7974 					return (0);
7975 				}
7976 				if (!tp->t_rttlow || tp->t_rttlow > t)
7977 					tp->t_rttlow = t;
7978 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
7979 					rack->r_ctl.rc_rack_min_rtt = t;
7980 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
7981 						rack->r_ctl.rc_rack_min_rtt = 1;
7982 					}
7983 				}
7984 				if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
7985 					   (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
7986 					/* New more recent rack_tmit_time */
7987 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
7988 					rack->rc_rack_rtt = t;
7989 				}
7990 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
7991 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
7992 						    rsm->r_rtr_cnt);
7993 				return (1);
7994 			}
7995 		}
7996 		goto ts_not_found;
7997 	} else {
7998 		/*
7999 		 * Ok its a SACK block that we retransmitted. or a windows
8000 		 * machine without timestamps. We can tell nothing from the
8001 		 * time-stamp since its not there or the time the peer last
8002 		 * recieved a segment that moved forward its cum-ack point.
8003 		 */
8004 ts_not_found:
8005 		i = rsm->r_rtr_cnt - 1;
8006 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
8007 		if ((int)t <= 0)
8008 			t = 1;
8009 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8010 			/*
8011 			 * We retransmitted and the ack came back in less
8012 			 * than the smallest rtt we have observed. We most
8013 			 * likely did an improper retransmit as outlined in
8014 			 * 6.2 Step 2 point 2 in the rack-draft so we
8015 			 * don't want to update our rack_rtt. We in
8016 			 * theory (in future) might want to think about reverting our
8017 			 * cwnd state but we won't for now.
8018 			 */
8019 			return (0);
8020 		} else if (rack->r_ctl.rc_rack_min_rtt) {
8021 			/*
8022 			 * We retransmitted it and the retransmit did the
8023 			 * job.
8024 			 */
8025 			if (!rack->r_ctl.rc_rack_min_rtt ||
8026 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8027 				rack->r_ctl.rc_rack_min_rtt = t;
8028 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
8029 					rack->r_ctl.rc_rack_min_rtt = 1;
8030 				}
8031 			}
8032 			if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) {
8033 				/* New more recent rack_tmit_time */
8034 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
8035 				rack->rc_rack_rtt = t;
8036 			}
8037 			return (1);
8038 		}
8039 	}
8040 	return (0);
8041 }
8042 
8043 /*
8044  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
8045  */
8046 static void
8047 rack_log_sack_passed(struct tcpcb *tp,
8048     struct tcp_rack *rack, struct rack_sendmap *rsm)
8049 {
8050 	struct rack_sendmap *nrsm;
8051 
8052 	nrsm = rsm;
8053 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
8054 	    rack_head, r_tnext) {
8055 		if (nrsm == rsm) {
8056 			/* Skip orginal segment he is acked */
8057 			continue;
8058 		}
8059 		if (nrsm->r_flags & RACK_ACKED) {
8060 			/*
8061 			 * Skip ack'd segments, though we
8062 			 * should not see these, since tmap
8063 			 * should not have ack'd segments.
8064 			 */
8065 			continue;
8066 		}
8067 		if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
8068 			/*
8069 			 * If the peer dropped the rwnd on
8070 			 * these then we don't worry about them.
8071 			 */
8072 			continue;
8073 		}
8074 		if (nrsm->r_flags & RACK_SACK_PASSED) {
8075 			/*
8076 			 * We found one that is already marked
8077 			 * passed, we have been here before and
8078 			 * so all others below this are marked.
8079 			 */
8080 			break;
8081 		}
8082 		nrsm->r_flags |= RACK_SACK_PASSED;
8083 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
8084 	}
8085 }
8086 
8087 static void
8088 rack_need_set_test(struct tcpcb *tp,
8089 		   struct tcp_rack *rack,
8090 		   struct rack_sendmap *rsm,
8091 		   tcp_seq th_ack,
8092 		   int line,
8093 		   int use_which)
8094 {
8095 
8096 	if ((tp->t_flags & TF_GPUTINPROG) &&
8097 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8098 		/*
8099 		 * We were app limited, and this ack
8100 		 * butts up or goes beyond the point where we want
8101 		 * to start our next measurement. We need
8102 		 * to record the new gput_ts as here and
8103 		 * possibly update the start sequence.
8104 		 */
8105 		uint32_t seq, ts;
8106 
8107 		if (rsm->r_rtr_cnt > 1) {
8108 			/*
8109 			 * This is a retransmit, can we
8110 			 * really make any assessment at this
8111 			 * point?  We are not really sure of
8112 			 * the timestamp, is it this or the
8113 			 * previous transmission?
8114 			 *
8115 			 * Lets wait for something better that
8116 			 * is not retransmitted.
8117 			 */
8118 			return;
8119 		}
8120 		seq = tp->gput_seq;
8121 		ts = tp->gput_ts;
8122 		rack->app_limited_needs_set = 0;
8123 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
8124 		/* Do we start at a new end? */
8125 		if ((use_which == RACK_USE_BEG) &&
8126 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
8127 			/*
8128 			 * When we get an ACK that just eats
8129 			 * up some of the rsm, we set RACK_USE_BEG
8130 			 * since whats at r_start (i.e. th_ack)
8131 			 * is left unacked and thats where the
8132 			 * measurement not starts.
8133 			 */
8134 			tp->gput_seq = rsm->r_start;
8135 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8136 		}
8137 		if ((use_which == RACK_USE_END) &&
8138 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
8139 			    /*
8140 			     * We use the end when the cumack
8141 			     * is moving forward and completely
8142 			     * deleting the rsm passed so basically
8143 			     * r_end holds th_ack.
8144 			     *
8145 			     * For SACK's we also want to use the end
8146 			     * since this piece just got sacked and
8147 			     * we want to target anything after that
8148 			     * in our measurement.
8149 			     */
8150 			    tp->gput_seq = rsm->r_end;
8151 			    rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8152 		}
8153 		if (use_which == RACK_USE_END_OR_THACK) {
8154 			/*
8155 			 * special case for ack moving forward,
8156 			 * not a sack, we need to move all the
8157 			 * way up to where this ack cum-ack moves
8158 			 * to.
8159 			 */
8160 			if (SEQ_GT(th_ack, rsm->r_end))
8161 				tp->gput_seq = th_ack;
8162 			else
8163 				tp->gput_seq = rsm->r_end;
8164 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8165 		}
8166 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
8167 			/*
8168 			 * We moved beyond this guy's range, re-calculate
8169 			 * the new end point.
8170 			 */
8171 			if (rack->rc_gp_filled == 0) {
8172 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
8173 			} else {
8174 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
8175 			}
8176 		}
8177 		/*
8178 		 * We are moving the goal post, we may be able to clear the
8179 		 * measure_saw_probe_rtt flag.
8180 		 */
8181 		if ((rack->in_probe_rtt == 0) &&
8182 		    (rack->measure_saw_probe_rtt) &&
8183 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
8184 			rack->measure_saw_probe_rtt = 0;
8185 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
8186 					   seq, tp->gput_seq, 0, 5, line, NULL, 0);
8187 		if (rack->rc_gp_filled &&
8188 		    ((tp->gput_ack - tp->gput_seq) <
8189 		     max(rc_init_window(rack), (MIN_GP_WIN *
8190 						ctf_fixed_maxseg(tp))))) {
8191 			uint32_t ideal_amount;
8192 
8193 			ideal_amount = rack_get_measure_window(tp, rack);
8194 			if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
8195 				/*
8196 				 * There is no sense of continuing this measurement
8197 				 * because its too small to gain us anything we
8198 				 * trust. Skip it and that way we can start a new
8199 				 * measurement quicker.
8200 				 */
8201 				tp->t_flags &= ~TF_GPUTINPROG;
8202 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
8203 							   0, 0, 0, 6, __LINE__, NULL, 0);
8204 			} else {
8205 				/*
8206 				 * Reset the window further out.
8207 				 */
8208 				tp->gput_ack = tp->gput_seq + ideal_amount;
8209 			}
8210 		}
8211 	}
8212 }
8213 
8214 static inline int
8215 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
8216 {
8217 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
8218 		/* Behind our TLP definition or right at */
8219 		return (0);
8220 	}
8221 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
8222 		/* The start is beyond or right at our end of TLP definition */
8223 		return (0);
8224 	}
8225 	/* It has to be a sub-part of the original TLP recorded */
8226 	return (1);
8227 }
8228 
8229 
8230 static uint32_t
8231 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
8232 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two)
8233 {
8234 	uint32_t start, end, changed = 0;
8235 	struct rack_sendmap stack_map;
8236 	struct rack_sendmap *rsm, *nrsm, fe, *prev, *next;
8237 #ifdef INVARIANTS
8238 	struct rack_sendmap *insret;
8239 #endif
8240 	int32_t used_ref = 1;
8241 	int moved = 0;
8242 
8243 	start = sack->start;
8244 	end = sack->end;
8245 	rsm = *prsm;
8246 	memset(&fe, 0, sizeof(fe));
8247 do_rest_ofb:
8248 	if ((rsm == NULL) ||
8249 	    (SEQ_LT(end, rsm->r_start)) ||
8250 	    (SEQ_GEQ(start, rsm->r_end)) ||
8251 	    (SEQ_LT(start, rsm->r_start))) {
8252 		/*
8253 		 * We are not in the right spot,
8254 		 * find the correct spot in the tree.
8255 		 */
8256 		used_ref = 0;
8257 		fe.r_start = start;
8258 		rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
8259 		moved++;
8260 	}
8261 	if (rsm == NULL) {
8262 		/* TSNH */
8263 		goto out;
8264 	}
8265 	/* Ok we have an ACK for some piece of this rsm */
8266 	if (rsm->r_start != start) {
8267 		if ((rsm->r_flags & RACK_ACKED) == 0) {
8268 			/*
8269 			 * Before any splitting or hookery is
8270 			 * done is it a TLP of interest i.e. rxt?
8271 			 */
8272 			if ((rsm->r_flags & RACK_TLP) &&
8273 			    (rsm->r_rtr_cnt > 1)) {
8274 				/*
8275 				 * We are splitting a rxt TLP, check
8276 				 * if we need to save off the start/end
8277 				 */
8278 				if (rack->rc_last_tlp_acked_set &&
8279 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8280 					/*
8281 					 * We already turned this on since we are inside
8282 					 * the previous one was a partially sack now we
8283 					 * are getting another one (maybe all of it).
8284 					 *
8285 					 */
8286 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8287 					/*
8288 					 * Lets make sure we have all of it though.
8289 					 */
8290 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8291 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8292 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8293 								     rack->r_ctl.last_tlp_acked_end);
8294 					}
8295 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8296 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8297 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8298 								     rack->r_ctl.last_tlp_acked_end);
8299 					}
8300 				} else {
8301 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8302 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8303 					rack->rc_last_tlp_past_cumack = 0;
8304 					rack->rc_last_tlp_acked_set = 1;
8305 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8306 				}
8307 			}
8308 			/**
8309 			 * Need to split this in two pieces the before and after,
8310 			 * the before remains in the map, the after must be
8311 			 * added. In other words we have:
8312 			 * rsm        |--------------|
8313 			 * sackblk        |------->
8314 			 * rsm will become
8315 			 *     rsm    |---|
8316 			 * and nrsm will be  the sacked piece
8317 			 *     nrsm       |----------|
8318 			 *
8319 			 * But before we start down that path lets
8320 			 * see if the sack spans over on top of
8321 			 * the next guy and it is already sacked.
8322 			 *
8323 			 */
8324 			next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8325 			if (next && (next->r_flags & RACK_ACKED) &&
8326 			    SEQ_GEQ(end, next->r_start)) {
8327 				/**
8328 				 * So the next one is already acked, and
8329 				 * we can thus by hookery use our stack_map
8330 				 * to reflect the piece being sacked and
8331 				 * then adjust the two tree entries moving
8332 				 * the start and ends around. So we start like:
8333 				 *  rsm     |------------|             (not-acked)
8334 				 *  next                 |-----------| (acked)
8335 				 *  sackblk        |-------->
8336 				 *  We want to end like so:
8337 				 *  rsm     |------|                   (not-acked)
8338 				 *  next           |-----------------| (acked)
8339 				 *  nrsm           |-----|
8340 				 * Where nrsm is a temporary stack piece we
8341 				 * use to update all the gizmos.
8342 				 */
8343 				/* Copy up our fudge block */
8344 				nrsm = &stack_map;
8345 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8346 				/* Now adjust our tree blocks */
8347 				rsm->r_end = start;
8348 				next->r_start = start;
8349 				/* Now we must adjust back where next->m is */
8350 				rack_setup_offset_for_rsm(rsm, next);
8351 
8352 				/* We don't need to adjust rsm, it did not change */
8353 				/* Clear out the dup ack count of the remainder */
8354 				rsm->r_dupack = 0;
8355 				rsm->r_just_ret = 0;
8356 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8357 				/* Now lets make sure our fudge block is right */
8358 				nrsm->r_start = start;
8359 				/* Now lets update all the stats and such */
8360 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8361 				if (rack->app_limited_needs_set)
8362 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8363 				changed += (nrsm->r_end - nrsm->r_start);
8364 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8365 				if (nrsm->r_flags & RACK_SACK_PASSED) {
8366 					rack->r_ctl.rc_reorder_ts = cts;
8367 				}
8368 				/*
8369 				 * Now we want to go up from rsm (the
8370 				 * one left un-acked) to the next one
8371 				 * in the tmap. We do this so when
8372 				 * we walk backwards we include marking
8373 				 * sack-passed on rsm (The one passed in
8374 				 * is skipped since it is generally called
8375 				 * on something sacked before removing it
8376 				 * from the tmap).
8377 				 */
8378 				if (rsm->r_in_tmap) {
8379 					nrsm = TAILQ_NEXT(rsm, r_tnext);
8380 					/*
8381 					 * Now that we have the next
8382 					 * one walk backwards from there.
8383 					 */
8384 					if (nrsm && nrsm->r_in_tmap)
8385 						rack_log_sack_passed(tp, rack, nrsm);
8386 				}
8387 				/* Now are we done? */
8388 				if (SEQ_LT(end, next->r_end) ||
8389 				    (end == next->r_end)) {
8390 					/* Done with block */
8391 					goto out;
8392 				}
8393 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
8394 				counter_u64_add(rack_sack_used_next_merge, 1);
8395 				/* Postion for the next block */
8396 				start = next->r_end;
8397 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next);
8398 				if (rsm == NULL)
8399 					goto out;
8400 			} else {
8401 				/**
8402 				 * We can't use any hookery here, so we
8403 				 * need to split the map. We enter like
8404 				 * so:
8405 				 *  rsm      |--------|
8406 				 *  sackblk       |----->
8407 				 * We will add the new block nrsm and
8408 				 * that will be the new portion, and then
8409 				 * fall through after reseting rsm. So we
8410 				 * split and look like this:
8411 				 *  rsm      |----|
8412 				 *  sackblk       |----->
8413 				 *  nrsm          |---|
8414 				 * We then fall through reseting
8415 				 * rsm to nrsm, so the next block
8416 				 * picks it up.
8417 				 */
8418 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8419 				if (nrsm == NULL) {
8420 					/*
8421 					 * failed XXXrrs what can we do but loose the sack
8422 					 * info?
8423 					 */
8424 					goto out;
8425 				}
8426 				counter_u64_add(rack_sack_splits, 1);
8427 				rack_clone_rsm(rack, nrsm, rsm, start);
8428 				rsm->r_just_ret = 0;
8429 #ifndef INVARIANTS
8430 				(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8431 #else
8432 				insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8433 				if (insret != NULL) {
8434 					panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8435 					      nrsm, insret, rack, rsm);
8436 				}
8437 #endif
8438 				if (rsm->r_in_tmap) {
8439 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8440 					nrsm->r_in_tmap = 1;
8441 				}
8442 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
8443 				rsm->r_flags &= (~RACK_HAS_FIN);
8444 				/* Position us to point to the new nrsm that starts the sack blk */
8445 				rsm = nrsm;
8446 			}
8447 		} else {
8448 			/* Already sacked this piece */
8449 			counter_u64_add(rack_sack_skipped_acked, 1);
8450 			moved++;
8451 			if (end == rsm->r_end) {
8452 				/* Done with block */
8453 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8454 				goto out;
8455 			} else if (SEQ_LT(end, rsm->r_end)) {
8456 				/* A partial sack to a already sacked block */
8457 				moved++;
8458 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8459 				goto out;
8460 			} else {
8461 				/*
8462 				 * The end goes beyond this guy
8463 				 * reposition the start to the
8464 				 * next block.
8465 				 */
8466 				start = rsm->r_end;
8467 				rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8468 				if (rsm == NULL)
8469 					goto out;
8470 			}
8471 		}
8472 	}
8473 	if (SEQ_GEQ(end, rsm->r_end)) {
8474 		/**
8475 		 * The end of this block is either beyond this guy or right
8476 		 * at this guy. I.e.:
8477 		 *  rsm ---                 |-----|
8478 		 *  end                     |-----|
8479 		 *  <or>
8480 		 *  end                     |---------|
8481 		 */
8482 		if ((rsm->r_flags & RACK_ACKED) == 0) {
8483 			/*
8484 			 * Is it a TLP of interest?
8485 			 */
8486 			if ((rsm->r_flags & RACK_TLP) &&
8487 			    (rsm->r_rtr_cnt > 1)) {
8488 				/*
8489 				 * We are splitting a rxt TLP, check
8490 				 * if we need to save off the start/end
8491 				 */
8492 				if (rack->rc_last_tlp_acked_set &&
8493 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8494 					/*
8495 					 * We already turned this on since we are inside
8496 					 * the previous one was a partially sack now we
8497 					 * are getting another one (maybe all of it).
8498 					 */
8499 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8500 					/*
8501 					 * Lets make sure we have all of it though.
8502 					 */
8503 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8504 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8505 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8506 								     rack->r_ctl.last_tlp_acked_end);
8507 					}
8508 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8509 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8510 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8511 								     rack->r_ctl.last_tlp_acked_end);
8512 					}
8513 				} else {
8514 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8515 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8516 					rack->rc_last_tlp_past_cumack = 0;
8517 					rack->rc_last_tlp_acked_set = 1;
8518 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8519 				}
8520 			}
8521 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8522 			changed += (rsm->r_end - rsm->r_start);
8523 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8524 			if (rsm->r_in_tmap) /* should be true */
8525 				rack_log_sack_passed(tp, rack, rsm);
8526 			/* Is Reordering occuring? */
8527 			if (rsm->r_flags & RACK_SACK_PASSED) {
8528 				rsm->r_flags &= ~RACK_SACK_PASSED;
8529 				rack->r_ctl.rc_reorder_ts = cts;
8530 			}
8531 			if (rack->app_limited_needs_set)
8532 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8533 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8534 			rsm->r_flags |= RACK_ACKED;
8535 			if (rsm->r_in_tmap) {
8536 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8537 				rsm->r_in_tmap = 0;
8538 			}
8539 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
8540 		} else {
8541 			counter_u64_add(rack_sack_skipped_acked, 1);
8542 			moved++;
8543 		}
8544 		if (end == rsm->r_end) {
8545 			/* This block only - done, setup for next */
8546 			goto out;
8547 		}
8548 		/*
8549 		 * There is more not coverend by this rsm move on
8550 		 * to the next block in the RB tree.
8551 		 */
8552 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8553 		start = rsm->r_end;
8554 		rsm = nrsm;
8555 		if (rsm == NULL)
8556 			goto out;
8557 		goto do_rest_ofb;
8558 	}
8559 	/**
8560 	 * The end of this sack block is smaller than
8561 	 * our rsm i.e.:
8562 	 *  rsm ---                 |-----|
8563 	 *  end                     |--|
8564 	 */
8565 	if ((rsm->r_flags & RACK_ACKED) == 0) {
8566 		/*
8567 		 * Is it a TLP of interest?
8568 		 */
8569 		if ((rsm->r_flags & RACK_TLP) &&
8570 		    (rsm->r_rtr_cnt > 1)) {
8571 			/*
8572 			 * We are splitting a rxt TLP, check
8573 			 * if we need to save off the start/end
8574 			 */
8575 			if (rack->rc_last_tlp_acked_set &&
8576 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8577 				/*
8578 				 * We already turned this on since we are inside
8579 				 * the previous one was a partially sack now we
8580 				 * are getting another one (maybe all of it).
8581 				 */
8582 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8583 				/*
8584 				 * Lets make sure we have all of it though.
8585 				 */
8586 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8587 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8588 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8589 							     rack->r_ctl.last_tlp_acked_end);
8590 				}
8591 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8592 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8593 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8594 							     rack->r_ctl.last_tlp_acked_end);
8595 				}
8596 			} else {
8597 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8598 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8599 				rack->rc_last_tlp_past_cumack = 0;
8600 				rack->rc_last_tlp_acked_set = 1;
8601 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8602 			}
8603 		}
8604 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8605 		if (prev &&
8606 		    (prev->r_flags & RACK_ACKED)) {
8607 			/**
8608 			 * Goal, we want the right remainder of rsm to shrink
8609 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
8610 			 * We want to expand prev to go all the way
8611 			 * to prev->r_end <- end.
8612 			 * so in the tree we have before:
8613 			 *   prev     |--------|         (acked)
8614 			 *   rsm               |-------| (non-acked)
8615 			 *   sackblk           |-|
8616 			 * We churn it so we end up with
8617 			 *   prev     |----------|       (acked)
8618 			 *   rsm                 |-----| (non-acked)
8619 			 *   nrsm              |-| (temporary)
8620 			 *
8621 			 * Note if either prev/rsm is a TLP we don't
8622 			 * do this.
8623 			 */
8624 			nrsm = &stack_map;
8625 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
8626 			prev->r_end = end;
8627 			rsm->r_start = end;
8628 			/* Now adjust nrsm (stack copy) to be
8629 			 * the one that is the small
8630 			 * piece that was "sacked".
8631 			 */
8632 			nrsm->r_end = end;
8633 			rsm->r_dupack = 0;
8634 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8635 			/*
8636 			 * Now that the rsm has had its start moved forward
8637 			 * lets go ahead and get its new place in the world.
8638 			 */
8639 			rack_setup_offset_for_rsm(prev, rsm);
8640 			/*
8641 			 * Now nrsm is our new little piece
8642 			 * that is acked (which was merged
8643 			 * to prev). Update the rtt and changed
8644 			 * based on that. Also check for reordering.
8645 			 */
8646 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
8647 			if (rack->app_limited_needs_set)
8648 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
8649 			changed += (nrsm->r_end - nrsm->r_start);
8650 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
8651 			if (nrsm->r_flags & RACK_SACK_PASSED) {
8652 				rack->r_ctl.rc_reorder_ts = cts;
8653 			}
8654 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
8655 			rsm = prev;
8656 			counter_u64_add(rack_sack_used_prev_merge, 1);
8657 		} else {
8658 			/**
8659 			 * This is the case where our previous
8660 			 * block is not acked either, so we must
8661 			 * split the block in two.
8662 			 */
8663 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
8664 			if (nrsm == NULL) {
8665 				/* failed rrs what can we do but loose the sack info? */
8666 				goto out;
8667 			}
8668 			if ((rsm->r_flags & RACK_TLP) &&
8669 			    (rsm->r_rtr_cnt > 1)) {
8670 				/*
8671 				 * We are splitting a rxt TLP, check
8672 				 * if we need to save off the start/end
8673 				 */
8674 				if (rack->rc_last_tlp_acked_set &&
8675 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
8676 					    /*
8677 					     * We already turned this on since this block is inside
8678 					     * the previous one was a partially sack now we
8679 					     * are getting another one (maybe all of it).
8680 					     */
8681 					    rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
8682 					    /*
8683 					     * Lets make sure we have all of it though.
8684 					     */
8685 					    if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
8686 						    rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8687 						    rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8688 									 rack->r_ctl.last_tlp_acked_end);
8689 					    }
8690 					    if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
8691 						    rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8692 						    rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
8693 									 rack->r_ctl.last_tlp_acked_end);
8694 					    }
8695 				    } else {
8696 					    rack->r_ctl.last_tlp_acked_start = rsm->r_start;
8697 					    rack->r_ctl.last_tlp_acked_end = rsm->r_end;
8698 					    rack->rc_last_tlp_acked_set = 1;
8699 					    rack->rc_last_tlp_past_cumack = 0;
8700 					    rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
8701 				    }
8702 			}
8703 			/**
8704 			 * In this case nrsm becomes
8705 			 * nrsm->r_start = end;
8706 			 * nrsm->r_end = rsm->r_end;
8707 			 * which is un-acked.
8708 			 * <and>
8709 			 * rsm->r_end = nrsm->r_start;
8710 			 * i.e. the remaining un-acked
8711 			 * piece is left on the left
8712 			 * hand side.
8713 			 *
8714 			 * So we start like this
8715 			 * rsm      |----------| (not acked)
8716 			 * sackblk  |---|
8717 			 * build it so we have
8718 			 * rsm      |---|         (acked)
8719 			 * nrsm         |------|  (not acked)
8720 			 */
8721 			counter_u64_add(rack_sack_splits, 1);
8722 			rack_clone_rsm(rack, nrsm, rsm, end);
8723 			rsm->r_flags &= (~RACK_HAS_FIN);
8724 			rsm->r_just_ret = 0;
8725 #ifndef INVARIANTS
8726 			(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8727 #else
8728 			insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
8729 			if (insret != NULL) {
8730 				panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
8731 				      nrsm, insret, rack, rsm);
8732 			}
8733 #endif
8734 			if (rsm->r_in_tmap) {
8735 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8736 				nrsm->r_in_tmap = 1;
8737 			}
8738 			nrsm->r_dupack = 0;
8739 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8740 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
8741 			changed += (rsm->r_end - rsm->r_start);
8742 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
8743 			if (rsm->r_in_tmap) /* should be true */
8744 				rack_log_sack_passed(tp, rack, rsm);
8745 			/* Is Reordering occuring? */
8746 			if (rsm->r_flags & RACK_SACK_PASSED) {
8747 				rsm->r_flags &= ~RACK_SACK_PASSED;
8748 				rack->r_ctl.rc_reorder_ts = cts;
8749 			}
8750 			if (rack->app_limited_needs_set)
8751 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
8752 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
8753 			rsm->r_flags |= RACK_ACKED;
8754 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
8755 			if (rsm->r_in_tmap) {
8756 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8757 				rsm->r_in_tmap = 0;
8758 			}
8759 		}
8760 	} else if (start != end){
8761 		/*
8762 		 * The block was already acked.
8763 		 */
8764 		counter_u64_add(rack_sack_skipped_acked, 1);
8765 		moved++;
8766 	}
8767 out:
8768 	if (rsm &&
8769 	    ((rsm->r_flags & RACK_TLP) == 0) &&
8770 	    (rsm->r_flags & RACK_ACKED)) {
8771 		/*
8772 		 * Now can we merge where we worked
8773 		 * with either the previous or
8774 		 * next block?
8775 		 */
8776 		next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8777 		while (next) {
8778 			if (next->r_flags & RACK_TLP)
8779 				break;
8780 			if (next->r_flags & RACK_ACKED) {
8781 			/* yep this and next can be merged */
8782 				rsm = rack_merge_rsm(rack, rsm, next);
8783 				next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8784 			} else
8785 				break;
8786 		}
8787 		/* Now what about the previous? */
8788 		prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8789 		while (prev) {
8790 			if (prev->r_flags & RACK_TLP)
8791 				break;
8792 			if (prev->r_flags & RACK_ACKED) {
8793 				/* yep the previous and this can be merged */
8794 				rsm = rack_merge_rsm(rack, prev, rsm);
8795 				prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8796 			} else
8797 				break;
8798 		}
8799 	}
8800 	if (used_ref == 0) {
8801 		counter_u64_add(rack_sack_proc_all, 1);
8802 	} else {
8803 		counter_u64_add(rack_sack_proc_short, 1);
8804 	}
8805 	/* Save off the next one for quick reference. */
8806 	if (rsm)
8807 		nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8808 	else
8809 		nrsm = NULL;
8810 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
8811 	/* Pass back the moved. */
8812 	*moved_two = moved;
8813 	return (changed);
8814 }
8815 
8816 static void inline
8817 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
8818 {
8819 	struct rack_sendmap *tmap;
8820 
8821 	tmap = NULL;
8822 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
8823 		/* Its no longer sacked, mark it so */
8824 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8825 #ifdef INVARIANTS
8826 		if (rsm->r_in_tmap) {
8827 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
8828 			      rack, rsm, rsm->r_flags);
8829 		}
8830 #endif
8831 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
8832 		/* Rebuild it into our tmap */
8833 		if (tmap == NULL) {
8834 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8835 			tmap = rsm;
8836 		} else {
8837 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
8838 			tmap = rsm;
8839 		}
8840 		tmap->r_in_tmap = 1;
8841 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
8842 	}
8843 	/*
8844 	 * Now lets possibly clear the sack filter so we start
8845 	 * recognizing sacks that cover this area.
8846 	 */
8847 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
8848 
8849 }
8850 
8851 static void
8852 rack_do_decay(struct tcp_rack *rack)
8853 {
8854 	struct timeval res;
8855 
8856 #define	timersub(tvp, uvp, vvp)						\
8857 	do {								\
8858 		(vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;		\
8859 		(vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;	\
8860 		if ((vvp)->tv_usec < 0) {				\
8861 			(vvp)->tv_sec--;				\
8862 			(vvp)->tv_usec += 1000000;			\
8863 		}							\
8864 	} while (0)
8865 
8866 	timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
8867 #undef timersub
8868 
8869 	rack->r_ctl.input_pkt++;
8870 	if ((rack->rc_in_persist) ||
8871 	    (res.tv_sec >= 1) ||
8872 	    (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
8873 		/*
8874 		 * Check for decay of non-SAD,
8875 		 * we want all SAD detection metrics to
8876 		 * decay 1/4 per second (or more) passed.
8877 		 */
8878 #ifdef NETFLIX_EXP_DETECTION
8879 		uint32_t pkt_delta;
8880 
8881 		pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
8882 #endif
8883 		/* Update our saved tracking values */
8884 		rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
8885 		rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
8886 		/* Now do we escape without decay? */
8887 #ifdef NETFLIX_EXP_DETECTION
8888 		if (rack->rc_in_persist ||
8889 		    (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
8890 		    (pkt_delta < tcp_sad_low_pps)){
8891 			/*
8892 			 * We don't decay idle connections
8893 			 * or ones that have a low input pps.
8894 			 */
8895 			return;
8896 		}
8897 		/* Decay the counters */
8898 		rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
8899 							tcp_sad_decay_val);
8900 		rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
8901 							 tcp_sad_decay_val);
8902 		rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
8903 							       tcp_sad_decay_val);
8904 		rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
8905 								tcp_sad_decay_val);
8906 #endif
8907 	}
8908 }
8909 
8910 static void
8911 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to)
8912 {
8913 	struct rack_sendmap *rsm;
8914 #ifdef INVARIANTS
8915 	struct rack_sendmap *rm;
8916 #endif
8917 
8918 	/*
8919 	 * The ACK point is advancing to th_ack, we must drop off
8920 	 * the packets in the rack log and calculate any eligble
8921 	 * RTT's.
8922 	 */
8923 	rack->r_wanted_output = 1;
8924 
8925 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
8926 	if ((rack->rc_last_tlp_acked_set == 1)&&
8927 	    (rack->rc_last_tlp_past_cumack == 1) &&
8928 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
8929 		/*
8930 		 * We have reached the point where our last rack
8931 		 * tlp retransmit sequence is ahead of the cum-ack.
8932 		 * This can only happen when the cum-ack moves all
8933 		 * the way around (its been a full 2^^31+1 bytes
8934 		 * or more since we sent a retransmitted TLP). Lets
8935 		 * turn off the valid flag since its not really valid.
8936 		 *
8937 		 * Note since sack's also turn on this event we have
8938 		 * a complication, we have to wait to age it out until
8939 		 * the cum-ack is by the TLP before checking which is
8940 		 * what the next else clause does.
8941 		 */
8942 		rack_log_dsack_event(rack, 9, __LINE__,
8943 				     rack->r_ctl.last_tlp_acked_start,
8944 				     rack->r_ctl.last_tlp_acked_end);
8945 		rack->rc_last_tlp_acked_set = 0;
8946 		rack->rc_last_tlp_past_cumack = 0;
8947 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
8948 		   (rack->rc_last_tlp_past_cumack == 0) &&
8949 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
8950 		/*
8951 		 * It is safe to start aging TLP's out.
8952 		 */
8953 		rack->rc_last_tlp_past_cumack = 1;
8954 	}
8955 	/* We do the same for the tlp send seq as well */
8956 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8957 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
8958 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
8959 		rack_log_dsack_event(rack, 9, __LINE__,
8960 				     rack->r_ctl.last_sent_tlp_seq,
8961 				     (rack->r_ctl.last_sent_tlp_seq +
8962 				      rack->r_ctl.last_sent_tlp_len));
8963 		rack->rc_last_sent_tlp_seq_valid = 0;
8964 		rack->rc_last_sent_tlp_past_cumack = 0;
8965 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
8966 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
8967 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
8968 		/*
8969 		 * It is safe to start aging TLP's send.
8970 		 */
8971 		rack->rc_last_sent_tlp_past_cumack = 1;
8972 	}
8973 more:
8974 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
8975 	if (rsm == NULL) {
8976 		if ((th_ack - 1) == tp->iss) {
8977 			/*
8978 			 * For the SYN incoming case we will not
8979 			 * have called tcp_output for the sending of
8980 			 * the SYN, so there will be no map. All
8981 			 * other cases should probably be a panic.
8982 			 */
8983 			return;
8984 		}
8985 		if (tp->t_flags & TF_SENTFIN) {
8986 			/* if we sent a FIN we often will not have map */
8987 			return;
8988 		}
8989 #ifdef INVARIANTS
8990 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
8991 		      tp,
8992 		      tp->t_state, th_ack, rack,
8993 		      tp->snd_una, tp->snd_max, tp->snd_nxt);
8994 #endif
8995 		return;
8996 	}
8997 	if (SEQ_LT(th_ack, rsm->r_start)) {
8998 		/* Huh map is missing this */
8999 #ifdef INVARIANTS
9000 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
9001 		       rsm->r_start,
9002 		       th_ack, tp->t_state, rack->r_state);
9003 #endif
9004 		return;
9005 	}
9006 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
9007 
9008 	/* Now was it a retransmitted TLP? */
9009 	if ((rsm->r_flags & RACK_TLP) &&
9010 	    (rsm->r_rtr_cnt > 1)) {
9011 		/*
9012 		 * Yes, this rsm was a TLP and retransmitted, remember that
9013 		 * since if a DSACK comes back on this we don't want
9014 		 * to think of it as a reordered segment. This may
9015 		 * get updated again with possibly even other TLPs
9016 		 * in flight, but thats ok. Only when we don't send
9017 		 * a retransmitted TLP for 1/2 the sequences space
9018 		 * will it get turned off (above).
9019 		 */
9020 		if (rack->rc_last_tlp_acked_set &&
9021 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9022 			/*
9023 			 * We already turned this on since the end matches,
9024 			 * the previous one was a partially ack now we
9025 			 * are getting another one (maybe all of it).
9026 			 */
9027 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9028 			/*
9029 			 * Lets make sure we have all of it though.
9030 			 */
9031 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9032 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9033 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9034 						     rack->r_ctl.last_tlp_acked_end);
9035 			}
9036 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9037 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9038 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9039 						     rack->r_ctl.last_tlp_acked_end);
9040 			}
9041 		} else {
9042 			rack->rc_last_tlp_past_cumack = 1;
9043 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9044 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9045 			rack->rc_last_tlp_acked_set = 1;
9046 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9047 		}
9048 	}
9049 	/* Now do we consume the whole thing? */
9050 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
9051 		/* Its all consumed. */
9052 		uint32_t left;
9053 		uint8_t newly_acked;
9054 
9055 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
9056 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
9057 		rsm->r_rtr_bytes = 0;
9058 		/* Record the time of highest cumack sent */
9059 		rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9060 #ifndef INVARIANTS
9061 		(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
9062 #else
9063 		rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
9064 		if (rm != rsm) {
9065 			panic("removing head in rack:%p rsm:%p rm:%p",
9066 			      rack, rsm, rm);
9067 		}
9068 #endif
9069 		if (rsm->r_in_tmap) {
9070 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9071 			rsm->r_in_tmap = 0;
9072 		}
9073 		newly_acked = 1;
9074 		if (rsm->r_flags & RACK_ACKED) {
9075 			/*
9076 			 * It was acked on the scoreboard -- remove
9077 			 * it from total
9078 			 */
9079 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
9080 			newly_acked = 0;
9081 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
9082 			/*
9083 			 * There are segments ACKED on the
9084 			 * scoreboard further up. We are seeing
9085 			 * reordering.
9086 			 */
9087 			rsm->r_flags &= ~RACK_SACK_PASSED;
9088 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9089 			rsm->r_flags |= RACK_ACKED;
9090 			rack->r_ctl.rc_reorder_ts = cts;
9091 			if (rack->r_ent_rec_ns) {
9092 				/*
9093 				 * We have sent no more, and we saw an sack
9094 				 * then ack arrive.
9095 				 */
9096 				rack->r_might_revert = 1;
9097 			}
9098 		}
9099 		if ((rsm->r_flags & RACK_TO_REXT) &&
9100 		    (tp->t_flags & TF_RCVD_TSTMP) &&
9101 		    (to->to_flags & TOF_TS) &&
9102 		    (to->to_tsecr != 0) &&
9103 		    (tp->t_flags & TF_PREVVALID)) {
9104 			/*
9105 			 * We can use the timestamp to see
9106 			 * if this retransmission was from the
9107 			 * first transmit. If so we made a mistake.
9108 			 */
9109 			tp->t_flags &= ~TF_PREVVALID;
9110 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
9111 				/* The first transmit is what this ack is for */
9112 				rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
9113 			}
9114 		}
9115 		left = th_ack - rsm->r_end;
9116 		if (rack->app_limited_needs_set && newly_acked)
9117 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
9118 		/* Free back to zone */
9119 		rack_free(rack, rsm);
9120 		if (left) {
9121 			goto more;
9122 		}
9123 		/* Check for reneging */
9124 		rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9125 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
9126 			/*
9127 			 * The peer has moved snd_una up to
9128 			 * the edge of this send, i.e. one
9129 			 * that it had previously acked. The only
9130 			 * way that can be true if the peer threw
9131 			 * away data (space issues) that it had
9132 			 * previously sacked (else it would have
9133 			 * given us snd_una up to (rsm->r_end).
9134 			 * We need to undo the acked markings here.
9135 			 *
9136 			 * Note we have to look to make sure th_ack is
9137 			 * our rsm->r_start in case we get an old ack
9138 			 * where th_ack is behind snd_una.
9139 			 */
9140 			rack_peer_reneges(rack, rsm, th_ack);
9141 		}
9142 		return;
9143 	}
9144 	if (rsm->r_flags & RACK_ACKED) {
9145 		/*
9146 		 * It was acked on the scoreboard -- remove it from
9147 		 * total for the part being cum-acked.
9148 		 */
9149 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
9150 	}
9151 	/*
9152 	 * Clear the dup ack count for
9153 	 * the piece that remains.
9154 	 */
9155 	rsm->r_dupack = 0;
9156 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9157 	if (rsm->r_rtr_bytes) {
9158 		/*
9159 		 * It was retransmitted adjust the
9160 		 * sack holes for what was acked.
9161 		 */
9162 		int ack_am;
9163 
9164 		ack_am = (th_ack - rsm->r_start);
9165 		if (ack_am >= rsm->r_rtr_bytes) {
9166 			rack->r_ctl.rc_holes_rxt -= ack_am;
9167 			rsm->r_rtr_bytes -= ack_am;
9168 		}
9169 	}
9170 	/*
9171 	 * Update where the piece starts and record
9172 	 * the time of send of highest cumack sent.
9173 	 */
9174 	rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
9175 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
9176 	/* Now we need to move our offset forward too */
9177 	if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) {
9178 		/* Fix up the orig_m_len and possibly the mbuf offset */
9179 		rack_adjust_orig_mlen(rsm);
9180 	}
9181 	rsm->soff += (th_ack - rsm->r_start);
9182 	rsm->r_start = th_ack;
9183 	/* Now do we need to move the mbuf fwd too? */
9184 	if (rsm->m) {
9185 		while (rsm->soff >= rsm->m->m_len) {
9186 			rsm->soff -= rsm->m->m_len;
9187 			rsm->m = rsm->m->m_next;
9188 			KASSERT((rsm->m != NULL),
9189 				(" nrsm:%p hit at soff:%u null m",
9190 				 rsm, rsm->soff));
9191 		}
9192 		rsm->orig_m_len = rsm->m->m_len;
9193 	}
9194 	if (rack->app_limited_needs_set)
9195 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
9196 }
9197 
9198 static void
9199 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
9200 {
9201 	struct rack_sendmap *rsm;
9202 	int sack_pass_fnd = 0;
9203 
9204 	if (rack->r_might_revert) {
9205 		/*
9206 		 * Ok we have reordering, have not sent anything, we
9207 		 * might want to revert the congestion state if nothing
9208 		 * further has SACK_PASSED on it. Lets check.
9209 		 *
9210 		 * We also get here when we have DSACKs come in for
9211 		 * all the data that we FR'd. Note that a rxt or tlp
9212 		 * timer clears this from happening.
9213 		 */
9214 
9215 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
9216 			if (rsm->r_flags & RACK_SACK_PASSED) {
9217 				sack_pass_fnd = 1;
9218 				break;
9219 			}
9220 		}
9221 		if (sack_pass_fnd == 0) {
9222 			/*
9223 			 * We went into recovery
9224 			 * incorrectly due to reordering!
9225 			 */
9226 			int orig_cwnd;
9227 
9228 			rack->r_ent_rec_ns = 0;
9229 			orig_cwnd = tp->snd_cwnd;
9230 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
9231 			tp->snd_recover = tp->snd_una;
9232 			rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
9233 			EXIT_RECOVERY(tp->t_flags);
9234 		}
9235 		rack->r_might_revert = 0;
9236 	}
9237 }
9238 
9239 #ifdef NETFLIX_EXP_DETECTION
9240 static void
9241 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
9242 {
9243 	if ((rack->do_detection || tcp_force_detection) &&
9244 	    tcp_sack_to_ack_thresh &&
9245 	    tcp_sack_to_move_thresh &&
9246 	    ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) {
9247 		/*
9248 		 * We have thresholds set to find
9249 		 * possible attackers and disable sack.
9250 		 * Check them.
9251 		 */
9252 		uint64_t ackratio, moveratio, movetotal;
9253 
9254 		/* Log detecting */
9255 		rack_log_sad(rack, 1);
9256 		ackratio = (uint64_t)(rack->r_ctl.sack_count);
9257 		ackratio *= (uint64_t)(1000);
9258 		if (rack->r_ctl.ack_count)
9259 			ackratio /= (uint64_t)(rack->r_ctl.ack_count);
9260 		else {
9261 			/* We really should not hit here */
9262 			ackratio = 1000;
9263 		}
9264 		if ((rack->sack_attack_disable == 0) &&
9265 		    (ackratio > rack_highest_sack_thresh_seen))
9266 			rack_highest_sack_thresh_seen = (uint32_t)ackratio;
9267 		movetotal = rack->r_ctl.sack_moved_extra;
9268 		movetotal += rack->r_ctl.sack_noextra_move;
9269 		moveratio = rack->r_ctl.sack_moved_extra;
9270 		moveratio *= (uint64_t)1000;
9271 		if (movetotal)
9272 			moveratio /= movetotal;
9273 		else {
9274 			/* No moves, thats pretty good */
9275 			moveratio = 0;
9276 		}
9277 		if ((rack->sack_attack_disable == 0) &&
9278 		    (moveratio > rack_highest_move_thresh_seen))
9279 			rack_highest_move_thresh_seen = (uint32_t)moveratio;
9280 		if (rack->sack_attack_disable == 0) {
9281 			if ((ackratio > tcp_sack_to_ack_thresh) &&
9282 			    (moveratio > tcp_sack_to_move_thresh)) {
9283 				/* Disable sack processing */
9284 				rack->sack_attack_disable = 1;
9285 				if (rack->r_rep_attack == 0) {
9286 					rack->r_rep_attack = 1;
9287 					counter_u64_add(rack_sack_attacks_detected, 1);
9288 				}
9289 				if (tcp_attack_on_turns_on_logging) {
9290 					/*
9291 					 * Turn on logging, used for debugging
9292 					 * false positives.
9293 					 */
9294 					rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging;
9295 				}
9296 				/* Clamp the cwnd at flight size */
9297 				rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
9298 				rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9299 				rack_log_sad(rack, 2);
9300 			}
9301 		} else {
9302 			/* We are sack-disabled check for false positives */
9303 			if ((ackratio <= tcp_restoral_thresh) ||
9304 			    (rack->r_ctl.rc_num_maps_alloced  < tcp_map_minimum)) {
9305 				rack->sack_attack_disable = 0;
9306 				rack_log_sad(rack, 3);
9307 				/* Restart counting */
9308 				rack->r_ctl.sack_count = 0;
9309 				rack->r_ctl.sack_moved_extra = 0;
9310 				rack->r_ctl.sack_noextra_move = 1;
9311 				rack->r_ctl.ack_count = max(1,
9312 				      (bytes_this_ack / segsiz));
9313 
9314 				if (rack->r_rep_reverse == 0) {
9315 					rack->r_rep_reverse = 1;
9316 					counter_u64_add(rack_sack_attacks_reversed, 1);
9317 				}
9318 				/* Restore the cwnd */
9319 				if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
9320 					rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
9321 			}
9322 		}
9323 	}
9324 }
9325 #endif
9326 
9327 static int
9328 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
9329 {
9330 
9331 	uint32_t am, l_end;
9332 	int was_tlp = 0;
9333 
9334 	if (SEQ_GT(end, start))
9335 		am = end - start;
9336 	else
9337 		am = 0;
9338 	if ((rack->rc_last_tlp_acked_set ) &&
9339 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
9340 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
9341 		/*
9342 		 * The DSACK is because of a TLP which we don't
9343 		 * do anything with the reordering window over since
9344 		 * it was not reordering that caused the DSACK but
9345 		 * our previous retransmit TLP.
9346 		 */
9347 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
9348 		was_tlp = 1;
9349 		goto skip_dsack_round;
9350 	}
9351 	if (rack->rc_last_sent_tlp_seq_valid) {
9352 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
9353 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
9354 		    (SEQ_LEQ(end, l_end))) {
9355 			/*
9356 			 * This dsack is from the last sent TLP, ignore it
9357 			 * for reordering purposes.
9358 			 */
9359 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
9360 			was_tlp = 1;
9361 			goto skip_dsack_round;
9362 		}
9363 	}
9364 	if (rack->rc_dsack_round_seen == 0) {
9365 		rack->rc_dsack_round_seen = 1;
9366 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
9367 		rack->r_ctl.num_dsack++;
9368 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
9369 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
9370 	}
9371 skip_dsack_round:
9372 	/*
9373 	 * We keep track of how many DSACK blocks we get
9374 	 * after a recovery incident.
9375 	 */
9376 	rack->r_ctl.dsack_byte_cnt += am;
9377 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
9378 	    rack->r_ctl.retran_during_recovery &&
9379 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
9380 		/*
9381 		 * False recovery most likely culprit is reordering. If
9382 		 * nothing else is missing we need to revert.
9383 		 */
9384 		rack->r_might_revert = 1;
9385 		rack_handle_might_revert(rack->rc_tp, rack);
9386 		rack->r_might_revert = 0;
9387 		rack->r_ctl.retran_during_recovery = 0;
9388 		rack->r_ctl.dsack_byte_cnt = 0;
9389 	}
9390 	return (was_tlp);
9391 }
9392 
9393 static uint32_t
9394 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
9395 {
9396 	return (((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt);
9397 }
9398 
9399 static int32_t
9400 rack_compute_pipe(struct tcpcb *tp)
9401 {
9402 	return ((int32_t)do_rack_compute_pipe(tp,
9403 					      (struct tcp_rack *)tp->t_fb_ptr,
9404 					      tp->snd_una));
9405 }
9406 
9407 static void
9408 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
9409 {
9410 	/* Deal with changed and PRR here (in recovery only) */
9411 	uint32_t pipe, snd_una;
9412 
9413 	rack->r_ctl.rc_prr_delivered += changed;
9414 
9415 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
9416 		/*
9417 		 * It is all outstanding, we are application limited
9418 		 * and thus we don't need more room to send anything.
9419 		 * Note we use tp->snd_una here and not th_ack because
9420 		 * the data as yet not been cut from the sb.
9421 		 */
9422 		rack->r_ctl.rc_prr_sndcnt = 0;
9423 		return;
9424 	}
9425 	/* Compute prr_sndcnt */
9426 	if (SEQ_GT(tp->snd_una, th_ack)) {
9427 		snd_una = tp->snd_una;
9428 	} else {
9429 		snd_una = th_ack;
9430 	}
9431 	pipe = do_rack_compute_pipe(tp, rack, snd_una);
9432 	if (pipe > tp->snd_ssthresh) {
9433 		long sndcnt;
9434 
9435 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
9436 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
9437 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
9438 		else {
9439 			rack->r_ctl.rc_prr_sndcnt = 0;
9440 			rack_log_to_prr(rack, 9, 0, __LINE__);
9441 			sndcnt = 0;
9442 		}
9443 		sndcnt++;
9444 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
9445 			sndcnt -= rack->r_ctl.rc_prr_out;
9446 		else
9447 			sndcnt = 0;
9448 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
9449 		rack_log_to_prr(rack, 10, 0, __LINE__);
9450 	} else {
9451 		uint32_t limit;
9452 
9453 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
9454 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
9455 		else
9456 			limit = 0;
9457 		if (changed > limit)
9458 			limit = changed;
9459 		limit += ctf_fixed_maxseg(tp);
9460 		if (tp->snd_ssthresh > pipe) {
9461 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
9462 			rack_log_to_prr(rack, 11, 0, __LINE__);
9463 		} else {
9464 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
9465 			rack_log_to_prr(rack, 12, 0, __LINE__);
9466 		}
9467 	}
9468 }
9469 
9470 static void
9471 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck)
9472 {
9473 	uint32_t changed;
9474 	struct tcp_rack *rack;
9475 	struct rack_sendmap *rsm;
9476 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
9477 	register uint32_t th_ack;
9478 	int32_t i, j, k, num_sack_blks = 0;
9479 	uint32_t cts, acked, ack_point;
9480 	int loop_start = 0, moved_two = 0;
9481 	uint32_t tsused;
9482 
9483 
9484 	INP_WLOCK_ASSERT(tptoinpcb(tp));
9485 	if (tcp_get_flags(th) & TH_RST) {
9486 		/* We don't log resets */
9487 		return;
9488 	}
9489 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9490 	cts = tcp_get_usecs(NULL);
9491 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
9492 	changed = 0;
9493 	th_ack = th->th_ack;
9494 	if (rack->sack_attack_disable == 0)
9495 		rack_do_decay(rack);
9496 	if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) {
9497 		/*
9498 		 * You only get credit for
9499 		 * MSS and greater (and you get extra
9500 		 * credit for larger cum-ack moves).
9501 		 */
9502 		int ac;
9503 
9504 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
9505 		rack->r_ctl.ack_count += ac;
9506 		counter_u64_add(rack_ack_total, ac);
9507 	}
9508 	if (rack->r_ctl.ack_count > 0xfff00000) {
9509 		/*
9510 		 * reduce the number to keep us under
9511 		 * a uint32_t.
9512 		 */
9513 		rack->r_ctl.ack_count /= 2;
9514 		rack->r_ctl.sack_count /= 2;
9515 	}
9516 	if (SEQ_GT(th_ack, tp->snd_una)) {
9517 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
9518 		tp->t_acktime = ticks;
9519 	}
9520 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
9521 		changed = th_ack - rsm->r_start;
9522 	if (changed) {
9523 		rack_process_to_cumack(tp, rack, th_ack, cts, to);
9524 	}
9525 	if ((to->to_flags & TOF_SACK) == 0) {
9526 		/* We are done nothing left and no sack. */
9527 		rack_handle_might_revert(tp, rack);
9528 		/*
9529 		 * For cases where we struck a dup-ack
9530 		 * with no SACK, add to the changes so
9531 		 * PRR will work right.
9532 		 */
9533 		if (dup_ack_struck && (changed == 0)) {
9534 			changed += ctf_fixed_maxseg(rack->rc_tp);
9535 		}
9536 		goto out;
9537 	}
9538 	/* Sack block processing */
9539 	if (SEQ_GT(th_ack, tp->snd_una))
9540 		ack_point = th_ack;
9541 	else
9542 		ack_point = tp->snd_una;
9543 	for (i = 0; i < to->to_nsacks; i++) {
9544 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
9545 		      &sack, sizeof(sack));
9546 		sack.start = ntohl(sack.start);
9547 		sack.end = ntohl(sack.end);
9548 		if (SEQ_GT(sack.end, sack.start) &&
9549 		    SEQ_GT(sack.start, ack_point) &&
9550 		    SEQ_LT(sack.start, tp->snd_max) &&
9551 		    SEQ_GT(sack.end, ack_point) &&
9552 		    SEQ_LEQ(sack.end, tp->snd_max)) {
9553 			sack_blocks[num_sack_blks] = sack;
9554 			num_sack_blks++;
9555 		} else if (SEQ_LEQ(sack.start, th_ack) &&
9556 			   SEQ_LEQ(sack.end, th_ack)) {
9557 			int was_tlp;
9558 
9559 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
9560 			/*
9561 			 * Its a D-SACK block.
9562 			 */
9563 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
9564 		}
9565 	}
9566 	if (rack->rc_dsack_round_seen) {
9567 		/* Is the dsack roound over? */
9568 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
9569 			/* Yes it is */
9570 			rack->rc_dsack_round_seen = 0;
9571 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
9572 		}
9573 	}
9574 	/*
9575 	 * Sort the SACK blocks so we can update the rack scoreboard with
9576 	 * just one pass.
9577 	 */
9578 	num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
9579 					 num_sack_blks, th->th_ack);
9580 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
9581 	if (num_sack_blks == 0) {
9582 		/* Nothing to sack (DSACKs?) */
9583 		goto out_with_totals;
9584 	}
9585 	if (num_sack_blks < 2) {
9586 		/* Only one, we don't need to sort */
9587 		goto do_sack_work;
9588 	}
9589 	/* Sort the sacks */
9590 	for (i = 0; i < num_sack_blks; i++) {
9591 		for (j = i + 1; j < num_sack_blks; j++) {
9592 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
9593 				sack = sack_blocks[i];
9594 				sack_blocks[i] = sack_blocks[j];
9595 				sack_blocks[j] = sack;
9596 			}
9597 		}
9598 	}
9599 	/*
9600 	 * Now are any of the sack block ends the same (yes some
9601 	 * implementations send these)?
9602 	 */
9603 again:
9604 	if (num_sack_blks == 0)
9605 		goto out_with_totals;
9606 	if (num_sack_blks > 1) {
9607 		for (i = 0; i < num_sack_blks; i++) {
9608 			for (j = i + 1; j < num_sack_blks; j++) {
9609 				if (sack_blocks[i].end == sack_blocks[j].end) {
9610 					/*
9611 					 * Ok these two have the same end we
9612 					 * want the smallest end and then
9613 					 * throw away the larger and start
9614 					 * again.
9615 					 */
9616 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
9617 						/*
9618 						 * The second block covers
9619 						 * more area use that
9620 						 */
9621 						sack_blocks[i].start = sack_blocks[j].start;
9622 					}
9623 					/*
9624 					 * Now collapse out the dup-sack and
9625 					 * lower the count
9626 					 */
9627 					for (k = (j + 1); k < num_sack_blks; k++) {
9628 						sack_blocks[j].start = sack_blocks[k].start;
9629 						sack_blocks[j].end = sack_blocks[k].end;
9630 						j++;
9631 					}
9632 					num_sack_blks--;
9633 					goto again;
9634 				}
9635 			}
9636 		}
9637 	}
9638 do_sack_work:
9639 	/*
9640 	 * First lets look to see if
9641 	 * we have retransmitted and
9642 	 * can use the transmit next?
9643 	 */
9644 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9645 	if (rsm &&
9646 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
9647 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
9648 		/*
9649 		 * We probably did the FR and the next
9650 		 * SACK in continues as we would expect.
9651 		 */
9652 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two);
9653 		if (acked) {
9654 			rack->r_wanted_output = 1;
9655 			changed += acked;
9656 		}
9657 		if (num_sack_blks == 1) {
9658 			/*
9659 			 * This is what we would expect from
9660 			 * a normal implementation to happen
9661 			 * after we have retransmitted the FR,
9662 			 * i.e the sack-filter pushes down
9663 			 * to 1 block and the next to be retransmitted
9664 			 * is the sequence in the sack block (has more
9665 			 * are acked). Count this as ACK'd data to boost
9666 			 * up the chances of recovering any false positives.
9667 			 */
9668 			rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
9669 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
9670 			counter_u64_add(rack_express_sack, 1);
9671 			if (rack->r_ctl.ack_count > 0xfff00000) {
9672 				/*
9673 				 * reduce the number to keep us under
9674 				 * a uint32_t.
9675 				 */
9676 				rack->r_ctl.ack_count /= 2;
9677 				rack->r_ctl.sack_count /= 2;
9678 			}
9679 			goto out_with_totals;
9680 		} else {
9681 			/*
9682 			 * Start the loop through the
9683 			 * rest of blocks, past the first block.
9684 			 */
9685 			moved_two = 0;
9686 			loop_start = 1;
9687 		}
9688 	}
9689 	/* Its a sack of some sort */
9690 	rack->r_ctl.sack_count++;
9691 	if (rack->r_ctl.sack_count > 0xfff00000) {
9692 		/*
9693 		 * reduce the number to keep us under
9694 		 * a uint32_t.
9695 		 */
9696 		rack->r_ctl.ack_count /= 2;
9697 		rack->r_ctl.sack_count /= 2;
9698 	}
9699 	counter_u64_add(rack_sack_total, 1);
9700 	if (rack->sack_attack_disable) {
9701 		/* An attacker disablement is in place */
9702 		if (num_sack_blks > 1) {
9703 			rack->r_ctl.sack_count += (num_sack_blks - 1);
9704 			rack->r_ctl.sack_moved_extra++;
9705 			counter_u64_add(rack_move_some, 1);
9706 			if (rack->r_ctl.sack_moved_extra > 0xfff00000) {
9707 				rack->r_ctl.sack_moved_extra /= 2;
9708 				rack->r_ctl.sack_noextra_move /= 2;
9709 			}
9710 		}
9711 		goto out;
9712 	}
9713 	rsm = rack->r_ctl.rc_sacklast;
9714 	for (i = loop_start; i < num_sack_blks; i++) {
9715 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two);
9716 		if (acked) {
9717 			rack->r_wanted_output = 1;
9718 			changed += acked;
9719 		}
9720 		if (moved_two) {
9721 			/*
9722 			 * If we did not get a SACK for at least a MSS and
9723 			 * had to move at all, or if we moved more than our
9724 			 * threshold, it counts against the "extra" move.
9725 			 */
9726 			rack->r_ctl.sack_moved_extra += moved_two;
9727 			counter_u64_add(rack_move_some, 1);
9728 		} else {
9729 			/*
9730 			 * else we did not have to move
9731 			 * any more than we would expect.
9732 			 */
9733 			rack->r_ctl.sack_noextra_move++;
9734 			counter_u64_add(rack_move_none, 1);
9735 		}
9736 		if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
9737 			/*
9738 			 * If the SACK was not a full MSS then
9739 			 * we add to sack_count the number of
9740 			 * MSS's (or possibly more than
9741 			 * a MSS if its a TSO send) we had to skip by.
9742 			 */
9743 			rack->r_ctl.sack_count += moved_two;
9744 			counter_u64_add(rack_sack_total, moved_two);
9745 		}
9746 		/*
9747 		 * Now we need to setup for the next
9748 		 * round. First we make sure we won't
9749 		 * exceed the size of our uint32_t on
9750 		 * the various counts, and then clear out
9751 		 * moved_two.
9752 		 */
9753 		if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
9754 		    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
9755 			rack->r_ctl.sack_moved_extra /= 2;
9756 			rack->r_ctl.sack_noextra_move /= 2;
9757 		}
9758 		if (rack->r_ctl.sack_count > 0xfff00000) {
9759 			rack->r_ctl.ack_count /= 2;
9760 			rack->r_ctl.sack_count /= 2;
9761 		}
9762 		moved_two = 0;
9763 	}
9764 out_with_totals:
9765 	if (num_sack_blks > 1) {
9766 		/*
9767 		 * You get an extra stroke if
9768 		 * you have more than one sack-blk, this
9769 		 * could be where we are skipping forward
9770 		 * and the sack-filter is still working, or
9771 		 * it could be an attacker constantly
9772 		 * moving us.
9773 		 */
9774 		rack->r_ctl.sack_moved_extra++;
9775 		counter_u64_add(rack_move_some, 1);
9776 	}
9777 out:
9778 #ifdef NETFLIX_EXP_DETECTION
9779 	rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
9780 #endif
9781 	if (changed) {
9782 		/* Something changed cancel the rack timer */
9783 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
9784 	}
9785 	tsused = tcp_get_usecs(NULL);
9786 	rsm = tcp_rack_output(tp, rack, tsused);
9787 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
9788 	    rsm &&
9789 	    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
9790 		/* Enter recovery */
9791 		entered_recovery = 1;
9792 		rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
9793 		/*
9794 		 * When we enter recovery we need to assure we send
9795 		 * one packet.
9796 		 */
9797 		if (rack->rack_no_prr == 0) {
9798 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
9799 			rack_log_to_prr(rack, 8, 0, __LINE__);
9800 		}
9801 		rack->r_timer_override = 1;
9802 		rack->r_early = 0;
9803 		rack->r_ctl.rc_agg_early = 0;
9804 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
9805 		   rsm &&
9806 		   (rack->r_rr_config == 3)) {
9807 		/*
9808 		 * Assure we can output and we get no
9809 		 * remembered pace time except the retransmit.
9810 		 */
9811 		rack->r_timer_override = 1;
9812 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
9813 		rack->r_ctl.rc_resend = rsm;
9814 	}
9815 	if (IN_FASTRECOVERY(tp->t_flags) &&
9816 	    (rack->rack_no_prr == 0) &&
9817 	    (entered_recovery == 0)) {
9818 		rack_update_prr(tp, rack, changed, th_ack);
9819 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
9820 		     ((tcp_in_hpts(rack->rc_inp) == 0) &&
9821 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
9822 			/*
9823 			 * If you are pacing output you don't want
9824 			 * to override.
9825 			 */
9826 			rack->r_early = 0;
9827 			rack->r_ctl.rc_agg_early = 0;
9828 			rack->r_timer_override = 1;
9829 		}
9830 	}
9831 }
9832 
9833 static void
9834 rack_strike_dupack(struct tcp_rack *rack)
9835 {
9836 	struct rack_sendmap *rsm;
9837 
9838 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
9839 	while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
9840 		rsm = TAILQ_NEXT(rsm, r_tnext);
9841 		if (rsm->r_flags & RACK_MUST_RXT) {
9842 			/* Sendmap entries that are marked to
9843 			 * be retransmitted do not need dupack's
9844 			 * struck. We get these marks for a number
9845 			 * of reasons (rxt timeout with no sack,
9846 			 * mtu change, or rwnd collapses). When
9847 			 * these events occur, we know we must retransmit
9848 			 * them and mark the sendmap entries. Dupack counting
9849 			 * is not needed since we are already set to retransmit
9850 			 * it as soon as we can.
9851 			 */
9852 			continue;
9853 		}
9854 	}
9855 	if (rsm && (rsm->r_dupack < 0xff)) {
9856 		rsm->r_dupack++;
9857 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
9858 			struct timeval tv;
9859 			uint32_t cts;
9860 			/*
9861 			 * Here we see if we need to retransmit. For
9862 			 * a SACK type connection if enough time has passed
9863 			 * we will get a return of the rsm. For a non-sack
9864 			 * connection we will get the rsm returned if the
9865 			 * dupack value is 3 or more.
9866 			 */
9867 			cts = tcp_get_usecs(&tv);
9868 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
9869 			if (rack->r_ctl.rc_resend != NULL) {
9870 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
9871 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
9872 							 rack->rc_tp->snd_una, __LINE__);
9873 				}
9874 				rack->r_wanted_output = 1;
9875 				rack->r_timer_override = 1;
9876 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
9877 			}
9878 		} else {
9879 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
9880 		}
9881 	}
9882 }
9883 
9884 static void
9885 rack_check_bottom_drag(struct tcpcb *tp,
9886 		       struct tcp_rack *rack,
9887 		       struct socket *so, int32_t acked)
9888 {
9889 	uint32_t segsiz, minseg;
9890 
9891 	segsiz = ctf_fixed_maxseg(tp);
9892 	minseg = segsiz;
9893 
9894 	if (tp->snd_max == tp->snd_una) {
9895 		/*
9896 		 * We are doing dynamic pacing and we are way
9897 		 * under. Basically everything got acked while
9898 		 * we were still waiting on the pacer to expire.
9899 		 *
9900 		 * This means we need to boost the b/w in
9901 		 * addition to any earlier boosting of
9902 		 * the multiplier.
9903 		 */
9904 		rack->rc_dragged_bottom = 1;
9905 		rack_validate_multipliers_at_or_above100(rack);
9906 		/*
9907 		 * Lets use the segment bytes acked plus
9908 		 * the lowest RTT seen as the basis to
9909 		 * form a b/w estimate. This will be off
9910 		 * due to the fact that the true estimate
9911 		 * should be around 1/2 the time of the RTT
9912 		 * but we can settle for that.
9913 		 */
9914 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
9915 		    acked) {
9916 			uint64_t bw, calc_bw, rtt;
9917 
9918 			rtt = rack->r_ctl.rack_rs.rs_us_rtt;
9919 			if (rtt == 0) {
9920 				/* no us sample is there a ms one? */
9921 				if (rack->r_ctl.rack_rs.rs_rtt_lowest) {
9922 					rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
9923 				} else {
9924 					goto no_measurement;
9925 				}
9926 			}
9927 			bw = acked;
9928 			calc_bw = bw * 1000000;
9929 			calc_bw /= rtt;
9930 			if (rack->r_ctl.last_max_bw &&
9931 			    (rack->r_ctl.last_max_bw < calc_bw)) {
9932 				/*
9933 				 * If we have a last calculated max bw
9934 				 * enforce it.
9935 				 */
9936 				calc_bw = rack->r_ctl.last_max_bw;
9937 			}
9938 			/* now plop it in */
9939 			if (rack->rc_gp_filled == 0) {
9940 				if (calc_bw > ONE_POINT_TWO_MEG) {
9941 					/*
9942 					 * If we have no measurement
9943 					 * don't let us set in more than
9944 					 * 1.2Mbps. If we are still too
9945 					 * low after pacing with this we
9946 					 * will hopefully have a max b/w
9947 					 * available to sanity check things.
9948 					 */
9949 					calc_bw = ONE_POINT_TWO_MEG;
9950 				}
9951 				rack->r_ctl.rc_rtt_diff = 0;
9952 				rack->r_ctl.gp_bw = calc_bw;
9953 				rack->rc_gp_filled = 1;
9954 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9955 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
9956 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9957 			} else if (calc_bw > rack->r_ctl.gp_bw) {
9958 				rack->r_ctl.rc_rtt_diff = 0;
9959 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
9960 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
9961 				rack->r_ctl.gp_bw = calc_bw;
9962 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
9963 			} else
9964 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
9965 			if ((rack->gp_ready == 0) &&
9966 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
9967 				/* We have enough measurements now */
9968 				rack->gp_ready = 1;
9969 				rack_set_cc_pacing(rack);
9970 				if (rack->defer_options)
9971 					rack_apply_deferred_options(rack);
9972 			}
9973 			/*
9974 			 * For acks over 1mss we do a extra boost to simulate
9975 			 * where we would get 2 acks (we want 110 for the mul).
9976 			 */
9977 			if (acked > segsiz)
9978 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
9979 		} else {
9980 			/*
9981 			 * zero rtt possibly?, settle for just an old increase.
9982 			 */
9983 no_measurement:
9984 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
9985 		}
9986 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
9987 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
9988 					       minseg)) &&
9989 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9990 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
9991 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
9992 		    (segsiz * rack_req_segs))) {
9993 		/*
9994 		 * We are doing dynamic GP pacing and
9995 		 * we have everything except 1MSS or less
9996 		 * bytes left out. We are still pacing away.
9997 		 * And there is data that could be sent, This
9998 		 * means we are inserting delayed ack time in
9999 		 * our measurements because we are pacing too slow.
10000 		 */
10001 		rack_validate_multipliers_at_or_above100(rack);
10002 		rack->rc_dragged_bottom = 1;
10003 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
10004 	}
10005 }
10006 
10007 
10008 
10009 static void
10010 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
10011 {
10012 	/*
10013 	 * The fast output path is enabled and we
10014 	 * have moved the cumack forward. Lets see if
10015 	 * we can expand forward the fast path length by
10016 	 * that amount. What we would ideally like to
10017 	 * do is increase the number of bytes in the
10018 	 * fast path block (left_to_send) by the
10019 	 * acked amount. However we have to gate that
10020 	 * by two factors:
10021 	 * 1) The amount outstanding and the rwnd of the peer
10022 	 *    (i.e. we don't want to exceed the rwnd of the peer).
10023 	 *    <and>
10024 	 * 2) The amount of data left in the socket buffer (i.e.
10025 	 *    we can't send beyond what is in the buffer).
10026 	 *
10027 	 * Note that this does not take into account any increase
10028 	 * in the cwnd. We will only extend the fast path by
10029 	 * what was acked.
10030 	 */
10031 	uint32_t new_total, gating_val;
10032 
10033 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
10034 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
10035 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
10036 	if (new_total <= gating_val) {
10037 		/* We can increase left_to_send by the acked amount */
10038 		counter_u64_add(rack_extended_rfo, 1);
10039 		rack->r_ctl.fsb.left_to_send = new_total;
10040 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
10041 			("rack:%p left_to_send:%u sbavail:%u out:%u",
10042 			 rack, rack->r_ctl.fsb.left_to_send,
10043 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
10044 			 (tp->snd_max - tp->snd_una)));
10045 
10046 	}
10047 }
10048 
10049 static void
10050 rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una)
10051 {
10052 	/*
10053 	 * Here any sendmap entry that points to the
10054 	 * beginning mbuf must be adjusted to the correct
10055 	 * offset. This must be called with:
10056 	 * 1) The socket buffer locked
10057 	 * 2) snd_una adjusted to its new postion.
10058 	 *
10059 	 * Note that (2) implies rack_ack_received has also
10060 	 * been called.
10061 	 *
10062 	 * We grab the first mbuf in the socket buffer and
10063 	 * then go through the front of the sendmap, recalculating
10064 	 * the stored offset for any sendmap entry that has
10065 	 * that mbuf. We must use the sb functions to do this
10066 	 * since its possible an add was done has well as
10067 	 * the subtraction we may have just completed. This should
10068 	 * not be a penalty though, since we just referenced the sb
10069 	 * to go in and trim off the mbufs that we freed (of course
10070 	 * there will be a penalty for the sendmap references though).
10071 	 */
10072 	struct mbuf *m;
10073 	struct rack_sendmap *rsm;
10074 
10075 	SOCKBUF_LOCK_ASSERT(sb);
10076 	m = sb->sb_mb;
10077 	rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
10078 	if ((rsm == NULL) || (m == NULL)) {
10079 		/* Nothing outstanding */
10080 		return;
10081 	}
10082 	while (rsm->m && (rsm->m == m)) {
10083 		/* one to adjust */
10084 #ifdef INVARIANTS
10085 		struct mbuf *tm;
10086 		uint32_t soff;
10087 
10088 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
10089 		if (rsm->orig_m_len != m->m_len) {
10090 			rack_adjust_orig_mlen(rsm);
10091 		}
10092 		if (rsm->soff != soff) {
10093 			/*
10094 			 * This is not a fatal error, we anticipate it
10095 			 * might happen (the else code), so we count it here
10096 			 * so that under invariant we can see that it really
10097 			 * does happen.
10098 			 */
10099 			counter_u64_add(rack_adjust_map_bw, 1);
10100 		}
10101 		rsm->m = tm;
10102 		rsm->soff = soff;
10103 		if (tm)
10104 			rsm->orig_m_len = rsm->m->m_len;
10105 		else
10106 			rsm->orig_m_len = 0;
10107 #else
10108 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
10109 		if (rsm->m)
10110 			rsm->orig_m_len = rsm->m->m_len;
10111 		else
10112 			rsm->orig_m_len = 0;
10113 #endif
10114 		rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
10115 			      rsm);
10116 		if (rsm == NULL)
10117 			break;
10118 	}
10119 }
10120 
10121 /*
10122  * Return value of 1, we do not need to call rack_process_data().
10123  * return value of 0, rack_process_data can be called.
10124  * For ret_val if its 0 the TCP is locked, if its non-zero
10125  * its unlocked and probably unsafe to touch the TCB.
10126  */
10127 static int
10128 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10129     struct tcpcb *tp, struct tcpopt *to,
10130     uint32_t tiwin, int32_t tlen,
10131     int32_t * ofia, int32_t thflags, int32_t *ret_val)
10132 {
10133 	int32_t ourfinisacked = 0;
10134 	int32_t nsegs, acked_amount;
10135 	int32_t acked;
10136 	struct mbuf *mfree;
10137 	struct tcp_rack *rack;
10138 	int32_t under_pacing = 0;
10139 	int32_t recovery = 0;
10140 
10141 	INP_WLOCK_ASSERT(tptoinpcb(tp));
10142 
10143 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10144 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
10145 		__ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
10146 				      &rack->r_ctl.challenge_ack_ts,
10147 				      &rack->r_ctl.challenge_ack_cnt);
10148 		rack->r_wanted_output = 1;
10149 		return (1);
10150 	}
10151 	if (rack->gp_ready &&
10152 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
10153 		under_pacing = 1;
10154 	}
10155 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
10156 		int in_rec, dup_ack_struck = 0;
10157 
10158 		in_rec = IN_FASTRECOVERY(tp->t_flags);
10159 		if (rack->rc_in_persist) {
10160 			tp->t_rxtshift = 0;
10161 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10162 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10163 		}
10164 		if ((th->th_ack == tp->snd_una) &&
10165 		    (tiwin == tp->snd_wnd) &&
10166 		    ((to->to_flags & TOF_SACK) == 0)) {
10167 			rack_strike_dupack(rack);
10168 			dup_ack_struck = 1;
10169 		}
10170 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck);
10171 	}
10172 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10173 		/*
10174 		 * Old ack, behind (or duplicate to) the last one rcv'd
10175 		 * Note: We mark reordering is occuring if its
10176 		 * less than and we have not closed our window.
10177 		 */
10178 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
10179 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
10180 		}
10181 		return (0);
10182 	}
10183 	/*
10184 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
10185 	 * something we sent.
10186 	 */
10187 	if (tp->t_flags & TF_NEEDSYN) {
10188 		/*
10189 		 * T/TCP: Connection was half-synchronized, and our SYN has
10190 		 * been ACK'd (so connection is now fully synchronized).  Go
10191 		 * to non-starred state, increment snd_una for ACK of SYN,
10192 		 * and check if we can do window scaling.
10193 		 */
10194 		tp->t_flags &= ~TF_NEEDSYN;
10195 		tp->snd_una++;
10196 		/* Do window scaling? */
10197 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
10198 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
10199 			tp->rcv_scale = tp->request_r_scale;
10200 			/* Send window already scaled. */
10201 		}
10202 	}
10203 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10204 
10205 	acked = BYTES_THIS_ACK(tp, th);
10206 	if (acked) {
10207 		/*
10208 		 * Any time we move the cum-ack forward clear
10209 		 * keep-alive tied probe-not-answered. The
10210 		 * persists clears its own on entry.
10211 		 */
10212 		rack->probe_not_answered = 0;
10213 	}
10214 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
10215 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
10216 	/*
10217 	 * If we just performed our first retransmit, and the ACK arrives
10218 	 * within our recovery window, then it was a mistake to do the
10219 	 * retransmit in the first place.  Recover our original cwnd and
10220 	 * ssthresh, and proceed to transmit where we left off.
10221 	 */
10222 	if ((tp->t_flags & TF_PREVVALID) &&
10223 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
10224 		tp->t_flags &= ~TF_PREVVALID;
10225 		if (tp->t_rxtshift == 1 &&
10226 		    (int)(ticks - tp->t_badrxtwin) < 0)
10227 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
10228 	}
10229 	if (acked) {
10230 		/* assure we are not backed off */
10231 		tp->t_rxtshift = 0;
10232 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
10233 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
10234 		rack->rc_tlp_in_progress = 0;
10235 		rack->r_ctl.rc_tlp_cnt_out = 0;
10236 		/*
10237 		 * If it is the RXT timer we want to
10238 		 * stop it, so we can restart a TLP.
10239 		 */
10240 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
10241 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10242 #ifdef NETFLIX_HTTP_LOGGING
10243 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
10244 #endif
10245 	}
10246 	/*
10247 	 * If we have a timestamp reply, update smoothed round trip time. If
10248 	 * no timestamp is present but transmit timer is running and timed
10249 	 * sequence number was acked, update smoothed round trip time. Since
10250 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
10251 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
10252 	 * timer.
10253 	 *
10254 	 * Some boxes send broken timestamp replies during the SYN+ACK
10255 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
10256 	 * and blow up the retransmit timer.
10257 	 */
10258 	/*
10259 	 * If all outstanding data is acked, stop retransmit timer and
10260 	 * remember to restart (more output or persist). If there is more
10261 	 * data to be acked, restart retransmit timer, using current
10262 	 * (possibly backed-off) value.
10263 	 */
10264 	if (acked == 0) {
10265 		if (ofia)
10266 			*ofia = ourfinisacked;
10267 		return (0);
10268 	}
10269 	if (IN_RECOVERY(tp->t_flags)) {
10270 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
10271 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
10272 			tcp_rack_partialack(tp);
10273 		} else {
10274 			rack_post_recovery(tp, th->th_ack);
10275 			recovery = 1;
10276 		}
10277 	}
10278 	/*
10279 	 * Let the congestion control algorithm update congestion control
10280 	 * related information. This typically means increasing the
10281 	 * congestion window.
10282 	 */
10283 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
10284 	SOCKBUF_LOCK(&so->so_snd);
10285 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
10286 	tp->snd_wnd -= acked_amount;
10287 	mfree = sbcut_locked(&so->so_snd, acked_amount);
10288 	if ((sbused(&so->so_snd) == 0) &&
10289 	    (acked > acked_amount) &&
10290 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
10291 	    (tp->t_flags & TF_SENTFIN)) {
10292 		/*
10293 		 * We must be sure our fin
10294 		 * was sent and acked (we can be
10295 		 * in FIN_WAIT_1 without having
10296 		 * sent the fin).
10297 		 */
10298 		ourfinisacked = 1;
10299 	}
10300 	tp->snd_una = th->th_ack;
10301 	if (acked_amount && sbavail(&so->so_snd))
10302 		rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
10303 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
10304 	/* NB: sowwakeup_locked() does an implicit unlock. */
10305 	sowwakeup_locked(so);
10306 	m_freem(mfree);
10307 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
10308 		tp->snd_recover = tp->snd_una;
10309 
10310 	if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
10311 		tp->snd_nxt = tp->snd_una;
10312 	}
10313 	if (under_pacing &&
10314 	    (rack->use_fixed_rate == 0) &&
10315 	    (rack->in_probe_rtt == 0) &&
10316 	    rack->rc_gp_dyn_mul &&
10317 	    rack->rc_always_pace) {
10318 		/* Check if we are dragging bottom */
10319 		rack_check_bottom_drag(tp, rack, so, acked);
10320 	}
10321 	if (tp->snd_una == tp->snd_max) {
10322 		/* Nothing left outstanding */
10323 		tp->t_flags &= ~TF_PREVVALID;
10324 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
10325 		rack->r_ctl.retran_during_recovery = 0;
10326 		rack->r_ctl.dsack_byte_cnt = 0;
10327 		if (rack->r_ctl.rc_went_idle_time == 0)
10328 			rack->r_ctl.rc_went_idle_time = 1;
10329 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
10330 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
10331 			tp->t_acktime = 0;
10332 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
10333 		/* Set need output so persist might get set */
10334 		rack->r_wanted_output = 1;
10335 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10336 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
10337 		    (sbavail(&so->so_snd) == 0) &&
10338 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
10339 			/*
10340 			 * The socket was gone and the
10341 			 * peer sent data (now or in the past), time to
10342 			 * reset him.
10343 			 */
10344 			*ret_val = 1;
10345 			/* tcp_close will kill the inp pre-log the Reset */
10346 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
10347 			tp = tcp_close(tp);
10348 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
10349 			return (1);
10350 		}
10351 	}
10352 	if (ofia)
10353 		*ofia = ourfinisacked;
10354 	return (0);
10355 }
10356 
10357 
10358 static void
10359 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
10360 		  int dir, uint32_t flags, struct rack_sendmap *rsm)
10361 {
10362 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
10363 		union tcp_log_stackspecific log;
10364 		struct timeval tv;
10365 
10366 		memset(&log, 0, sizeof(log));
10367 		log.u_bbr.flex1 = cnt;
10368 		log.u_bbr.flex2 = split;
10369 		log.u_bbr.flex3 = out;
10370 		log.u_bbr.flex4 = line;
10371 		log.u_bbr.flex5 = rack->r_must_retran;
10372 		log.u_bbr.flex6 = flags;
10373 		log.u_bbr.flex7 = rack->rc_has_collapsed;
10374 		log.u_bbr.flex8 = dir;	/*
10375 					 * 1 is collapsed, 0 is uncollapsed,
10376 					 * 2 is log of a rsm being marked, 3 is a split.
10377 					 */
10378 		if (rsm == NULL)
10379 			log.u_bbr.rttProp = 0;
10380 		else
10381 			log.u_bbr.rttProp = (uint64_t)rsm;
10382 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
10383 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10384 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
10385 		    &rack->rc_inp->inp_socket->so_rcv,
10386 		    &rack->rc_inp->inp_socket->so_snd,
10387 		    TCP_RACK_LOG_COLLAPSE, 0,
10388 		    0, &log, false, &tv);
10389 	}
10390 }
10391 
10392 static void
10393 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, int line)
10394 {
10395 	/*
10396 	 * Here all we do is mark the collapsed point and set the flag.
10397 	 * This may happen again and again, but there is no
10398 	 * sense splitting our map until we know where the
10399 	 * peer finally lands in the collapse.
10400 	 */
10401 	rack_trace_point(rack, RACK_TP_COLLAPSED_WND);
10402 	if ((rack->rc_has_collapsed == 0) ||
10403 	    (rack->r_ctl.last_collapse_point != (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)))
10404 		counter_u64_add(rack_collapsed_win_seen, 1);
10405 	rack->r_ctl.last_collapse_point = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd;
10406 	rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
10407 	rack->rc_has_collapsed = 1;
10408 	rack->r_collapse_point_valid = 1;
10409 	rack_log_collapse(rack, 0, 0, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
10410 }
10411 
10412 static void
10413 rack_un_collapse_window(struct tcp_rack *rack, int line)
10414 {
10415 	struct rack_sendmap *nrsm, *rsm, fe;
10416 	int cnt = 0, split = 0;
10417 #ifdef INVARIANTS
10418 	struct rack_sendmap *insret;
10419 #endif
10420 
10421 	memset(&fe, 0, sizeof(fe));
10422 	rack->rc_has_collapsed = 0;
10423 	fe.r_start = rack->r_ctl.last_collapse_point;
10424 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
10425 	if (rsm == NULL) {
10426 		/* Nothing to do maybe the peer ack'ed it all */
10427 		rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
10428 		return;
10429 	}
10430 	/* Now do we need to split this one? */
10431 	if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
10432 		rack_log_collapse(rack, rsm->r_start, rsm->r_end,
10433 				  rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
10434 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10435 		if (nrsm == NULL) {
10436 			/* We can't get a rsm, mark all? */
10437 			nrsm = rsm;
10438 			goto no_split;
10439 		}
10440 		/* Clone it */
10441 		split = 1;
10442 		rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
10443 #ifndef INVARIANTS
10444 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10445 #else
10446 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm);
10447 		if (insret != NULL) {
10448 			panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p",
10449 			      nrsm, insret, rack, rsm);
10450 		}
10451 #endif
10452 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
10453 				 rack->r_ctl.last_collapse_point, __LINE__);
10454 		if (rsm->r_in_tmap) {
10455 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10456 			nrsm->r_in_tmap = 1;
10457 		}
10458 		/*
10459 		 * Set in the new RSM as the
10460 		 * collapsed starting point
10461 		 */
10462 		rsm = nrsm;
10463 	}
10464 no_split:
10465 	RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) {
10466 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
10467 		rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
10468 		cnt++;
10469 	}
10470 	if (cnt) {
10471 		counter_u64_add(rack_collapsed_win, 1);
10472 	}
10473 	rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
10474 }
10475 
10476 static void
10477 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
10478 			int32_t tlen, int32_t tfo_syn)
10479 {
10480 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
10481 		if (rack->rc_dack_mode &&
10482 		    (tlen > 500) &&
10483 		    (rack->rc_dack_toggle == 1)) {
10484 			goto no_delayed_ack;
10485 		}
10486 		rack_timer_cancel(tp, rack,
10487 				  rack->r_ctl.rc_rcvtime, __LINE__);
10488 		tp->t_flags |= TF_DELACK;
10489 	} else {
10490 no_delayed_ack:
10491 		rack->r_wanted_output = 1;
10492 		tp->t_flags |= TF_ACKNOW;
10493 		if (rack->rc_dack_mode) {
10494 			if (tp->t_flags & TF_DELACK)
10495 				rack->rc_dack_toggle = 1;
10496 			else
10497 				rack->rc_dack_toggle = 0;
10498 		}
10499 	}
10500 }
10501 
10502 static void
10503 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
10504 {
10505 	/*
10506 	 * If fast output is in progress, lets validate that
10507 	 * the new window did not shrink on us and make it
10508 	 * so fast output should end.
10509 	 */
10510 	if (rack->r_fast_output) {
10511 		uint32_t out;
10512 
10513 		/*
10514 		 * Calculate what we will send if left as is
10515 		 * and compare that to our send window.
10516 		 */
10517 		out = ctf_outstanding(tp);
10518 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
10519 			/* ok we have an issue */
10520 			if (out >= tp->snd_wnd) {
10521 				/* Turn off fast output the window is met or collapsed */
10522 				rack->r_fast_output = 0;
10523 			} else {
10524 				/* we have some room left */
10525 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
10526 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
10527 					/* If not at least 1 full segment never mind */
10528 					rack->r_fast_output = 0;
10529 				}
10530 			}
10531 		}
10532 	}
10533 }
10534 
10535 
10536 /*
10537  * Return value of 1, the TCB is unlocked and most
10538  * likely gone, return value of 0, the TCP is still
10539  * locked.
10540  */
10541 static int
10542 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
10543     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
10544     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
10545 {
10546 	/*
10547 	 * Update window information. Don't look at window if no ACK: TAC's
10548 	 * send garbage on first SYN.
10549 	 */
10550 	int32_t nsegs;
10551 	int32_t tfo_syn;
10552 	struct tcp_rack *rack;
10553 
10554 	INP_WLOCK_ASSERT(tptoinpcb(tp));
10555 
10556 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10557 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10558 	if ((thflags & TH_ACK) &&
10559 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
10560 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
10561 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
10562 		/* keep track of pure window updates */
10563 		if (tlen == 0 &&
10564 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
10565 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
10566 		tp->snd_wnd = tiwin;
10567 		rack_validate_fo_sendwin_up(tp, rack);
10568 		tp->snd_wl1 = th->th_seq;
10569 		tp->snd_wl2 = th->th_ack;
10570 		if (tp->snd_wnd > tp->max_sndwnd)
10571 			tp->max_sndwnd = tp->snd_wnd;
10572 		rack->r_wanted_output = 1;
10573 	} else if (thflags & TH_ACK) {
10574 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
10575 			tp->snd_wnd = tiwin;
10576 			rack_validate_fo_sendwin_up(tp, rack);
10577 			tp->snd_wl1 = th->th_seq;
10578 			tp->snd_wl2 = th->th_ack;
10579 		}
10580 	}
10581 	if (tp->snd_wnd < ctf_outstanding(tp))
10582 		/* The peer collapsed the window */
10583 		rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__);
10584 	else if (rack->rc_has_collapsed)
10585 		rack_un_collapse_window(rack, __LINE__);
10586 	if ((rack->r_collapse_point_valid) &&
10587 	    (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
10588 		rack->r_collapse_point_valid = 0;
10589 	/* Was persist timer active and now we have window space? */
10590 	if ((rack->rc_in_persist != 0) &&
10591 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
10592 				rack->r_ctl.rc_pace_min_segs))) {
10593 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10594 		tp->snd_nxt = tp->snd_max;
10595 		/* Make sure we output to start the timer */
10596 		rack->r_wanted_output = 1;
10597 	}
10598 	/* Do we enter persists? */
10599 	if ((rack->rc_in_persist == 0) &&
10600 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
10601 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
10602 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
10603 	    sbavail(&tptosocket(tp)->so_snd) &&
10604 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
10605 		/*
10606 		 * Here the rwnd is less than
10607 		 * the pacing size, we are established,
10608 		 * nothing is outstanding, and there is
10609 		 * data to send. Enter persists.
10610 		 */
10611 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
10612 	}
10613 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
10614 		m_freem(m);
10615 		return (0);
10616 	}
10617 	/*
10618 	 * don't process the URG bit, ignore them drag
10619 	 * along the up.
10620 	 */
10621 	tp->rcv_up = tp->rcv_nxt;
10622 
10623 	/*
10624 	 * Process the segment text, merging it into the TCP sequencing
10625 	 * queue, and arranging for acknowledgment of receipt if necessary.
10626 	 * This process logically involves adjusting tp->rcv_wnd as data is
10627 	 * presented to the user (this happens in tcp_usrreq.c, case
10628 	 * PRU_RCVD).  If a FIN has already been received on this connection
10629 	 * then we just ignore the text.
10630 	 */
10631 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
10632 		   IS_FASTOPEN(tp->t_flags));
10633 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
10634 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10635 		tcp_seq save_start = th->th_seq;
10636 		tcp_seq save_rnxt  = tp->rcv_nxt;
10637 		int     save_tlen  = tlen;
10638 
10639 		m_adj(m, drop_hdrlen);	/* delayed header drop */
10640 		/*
10641 		 * Insert segment which includes th into TCP reassembly
10642 		 * queue with control block tp.  Set thflags to whether
10643 		 * reassembly now includes a segment with FIN.  This handles
10644 		 * the common case inline (segment is the next to be
10645 		 * received on an established connection, and the queue is
10646 		 * empty), avoiding linkage into and removal from the queue
10647 		 * and repetition of various conversions. Set DELACK for
10648 		 * segments received in order, but ack immediately when
10649 		 * segments are out of order (so fast retransmit can work).
10650 		 */
10651 		if (th->th_seq == tp->rcv_nxt &&
10652 		    SEGQ_EMPTY(tp) &&
10653 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
10654 		    tfo_syn)) {
10655 #ifdef NETFLIX_SB_LIMITS
10656 			u_int mcnt, appended;
10657 
10658 			if (so->so_rcv.sb_shlim) {
10659 				mcnt = m_memcnt(m);
10660 				appended = 0;
10661 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10662 				    CFO_NOSLEEP, NULL) == false) {
10663 					counter_u64_add(tcp_sb_shlim_fails, 1);
10664 					m_freem(m);
10665 					return (0);
10666 				}
10667 			}
10668 #endif
10669 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
10670 			tp->rcv_nxt += tlen;
10671 			if (tlen &&
10672 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10673 			    (tp->t_fbyte_in == 0)) {
10674 				tp->t_fbyte_in = ticks;
10675 				if (tp->t_fbyte_in == 0)
10676 					tp->t_fbyte_in = 1;
10677 				if (tp->t_fbyte_out && tp->t_fbyte_in)
10678 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10679 			}
10680 			thflags = tcp_get_flags(th) & TH_FIN;
10681 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10682 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10683 			SOCKBUF_LOCK(&so->so_rcv);
10684 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10685 				m_freem(m);
10686 			} else
10687 #ifdef NETFLIX_SB_LIMITS
10688 				appended =
10689 #endif
10690 					sbappendstream_locked(&so->so_rcv, m, 0);
10691 
10692 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10693 			/* NB: sorwakeup_locked() does an implicit unlock. */
10694 			sorwakeup_locked(so);
10695 #ifdef NETFLIX_SB_LIMITS
10696 			if (so->so_rcv.sb_shlim && appended != mcnt)
10697 				counter_fo_release(so->so_rcv.sb_shlim,
10698 				    mcnt - appended);
10699 #endif
10700 		} else {
10701 			/*
10702 			 * XXX: Due to the header drop above "th" is
10703 			 * theoretically invalid by now.  Fortunately
10704 			 * m_adj() doesn't actually frees any mbufs when
10705 			 * trimming from the head.
10706 			 */
10707 			tcp_seq temp = save_start;
10708 
10709 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
10710 			tp->t_flags |= TF_ACKNOW;
10711 			if (tp->t_flags & TF_WAKESOR) {
10712 				tp->t_flags &= ~TF_WAKESOR;
10713 				/* NB: sorwakeup_locked() does an implicit unlock. */
10714 				sorwakeup_locked(so);
10715 			}
10716 		}
10717 		if ((tp->t_flags & TF_SACK_PERMIT) &&
10718 		    (save_tlen > 0) &&
10719 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
10720 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
10721 				/*
10722 				 * DSACK actually handled in the fastpath
10723 				 * above.
10724 				 */
10725 				RACK_OPTS_INC(tcp_sack_path_1);
10726 				tcp_update_sack_list(tp, save_start,
10727 				    save_start + save_tlen);
10728 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
10729 				if ((tp->rcv_numsacks >= 1) &&
10730 				    (tp->sackblks[0].end == save_start)) {
10731 					/*
10732 					 * Partial overlap, recorded at todrop
10733 					 * above.
10734 					 */
10735 					RACK_OPTS_INC(tcp_sack_path_2a);
10736 					tcp_update_sack_list(tp,
10737 					    tp->sackblks[0].start,
10738 					    tp->sackblks[0].end);
10739 				} else {
10740 					RACK_OPTS_INC(tcp_sack_path_2b);
10741 					tcp_update_dsack_list(tp, save_start,
10742 					    save_start + save_tlen);
10743 				}
10744 			} else if (tlen >= save_tlen) {
10745 				/* Update of sackblks. */
10746 				RACK_OPTS_INC(tcp_sack_path_3);
10747 				tcp_update_dsack_list(tp, save_start,
10748 				    save_start + save_tlen);
10749 			} else if (tlen > 0) {
10750 				RACK_OPTS_INC(tcp_sack_path_4);
10751 				tcp_update_dsack_list(tp, save_start,
10752 				    save_start + tlen);
10753 			}
10754 		}
10755 	} else {
10756 		m_freem(m);
10757 		thflags &= ~TH_FIN;
10758 	}
10759 
10760 	/*
10761 	 * If FIN is received ACK the FIN and let the user know that the
10762 	 * connection is closing.
10763 	 */
10764 	if (thflags & TH_FIN) {
10765 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
10766 			/* The socket upcall is handled by socantrcvmore. */
10767 			socantrcvmore(so);
10768 			/*
10769 			 * If connection is half-synchronized (ie NEEDSYN
10770 			 * flag on) then delay ACK, so it may be piggybacked
10771 			 * when SYN is sent. Otherwise, since we received a
10772 			 * FIN then no more input can be expected, send ACK
10773 			 * now.
10774 			 */
10775 			if (tp->t_flags & TF_NEEDSYN) {
10776 				rack_timer_cancel(tp, rack,
10777 				    rack->r_ctl.rc_rcvtime, __LINE__);
10778 				tp->t_flags |= TF_DELACK;
10779 			} else {
10780 				tp->t_flags |= TF_ACKNOW;
10781 			}
10782 			tp->rcv_nxt++;
10783 		}
10784 		switch (tp->t_state) {
10785 			/*
10786 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
10787 			 * CLOSE_WAIT state.
10788 			 */
10789 		case TCPS_SYN_RECEIVED:
10790 			tp->t_starttime = ticks;
10791 			/* FALLTHROUGH */
10792 		case TCPS_ESTABLISHED:
10793 			rack_timer_cancel(tp, rack,
10794 			    rack->r_ctl.rc_rcvtime, __LINE__);
10795 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
10796 			break;
10797 
10798 			/*
10799 			 * If still in FIN_WAIT_1 STATE FIN has not been
10800 			 * acked so enter the CLOSING state.
10801 			 */
10802 		case TCPS_FIN_WAIT_1:
10803 			rack_timer_cancel(tp, rack,
10804 			    rack->r_ctl.rc_rcvtime, __LINE__);
10805 			tcp_state_change(tp, TCPS_CLOSING);
10806 			break;
10807 
10808 			/*
10809 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
10810 			 * starting the time-wait timer, turning off the
10811 			 * other standard timers.
10812 			 */
10813 		case TCPS_FIN_WAIT_2:
10814 			rack_timer_cancel(tp, rack,
10815 			    rack->r_ctl.rc_rcvtime, __LINE__);
10816 			tcp_twstart(tp);
10817 			return (1);
10818 		}
10819 	}
10820 	/*
10821 	 * Return any desired output.
10822 	 */
10823 	if ((tp->t_flags & TF_ACKNOW) ||
10824 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
10825 		rack->r_wanted_output = 1;
10826 	}
10827 	return (0);
10828 }
10829 
10830 /*
10831  * Here nothing is really faster, its just that we
10832  * have broken out the fast-data path also just like
10833  * the fast-ack.
10834  */
10835 static int
10836 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
10837     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10838     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
10839 {
10840 	int32_t nsegs;
10841 	int32_t newsize = 0;	/* automatic sockbuf scaling */
10842 	struct tcp_rack *rack;
10843 #ifdef NETFLIX_SB_LIMITS
10844 	u_int mcnt, appended;
10845 #endif
10846 #ifdef TCPDEBUG
10847 	/*
10848 	 * The size of tcp_saveipgen must be the size of the max ip header,
10849 	 * now IPv6.
10850 	 */
10851 	u_char tcp_saveipgen[IP6_HDR_LEN];
10852 	struct tcphdr tcp_savetcp;
10853 	short ostate = 0;
10854 
10855 #endif
10856 	/*
10857 	 * If last ACK falls within this segment's sequence numbers, record
10858 	 * the timestamp. NOTE that the test is modified according to the
10859 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
10860 	 */
10861 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
10862 		return (0);
10863 	}
10864 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
10865 		return (0);
10866 	}
10867 	if (tiwin && tiwin != tp->snd_wnd) {
10868 		return (0);
10869 	}
10870 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
10871 		return (0);
10872 	}
10873 	if (__predict_false((to->to_flags & TOF_TS) &&
10874 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
10875 		return (0);
10876 	}
10877 	if (__predict_false((th->th_ack != tp->snd_una))) {
10878 		return (0);
10879 	}
10880 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
10881 		return (0);
10882 	}
10883 	if ((to->to_flags & TOF_TS) != 0 &&
10884 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
10885 		tp->ts_recent_age = tcp_ts_getticks();
10886 		tp->ts_recent = to->to_tsval;
10887 	}
10888 	rack = (struct tcp_rack *)tp->t_fb_ptr;
10889 	/*
10890 	 * This is a pure, in-sequence data packet with nothing on the
10891 	 * reassembly queue and we have enough buffer space to take it.
10892 	 */
10893 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
10894 
10895 #ifdef NETFLIX_SB_LIMITS
10896 	if (so->so_rcv.sb_shlim) {
10897 		mcnt = m_memcnt(m);
10898 		appended = 0;
10899 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
10900 		    CFO_NOSLEEP, NULL) == false) {
10901 			counter_u64_add(tcp_sb_shlim_fails, 1);
10902 			m_freem(m);
10903 			return (1);
10904 		}
10905 	}
10906 #endif
10907 	/* Clean receiver SACK report if present */
10908 	if (tp->rcv_numsacks)
10909 		tcp_clean_sackreport(tp);
10910 	KMOD_TCPSTAT_INC(tcps_preddat);
10911 	tp->rcv_nxt += tlen;
10912 	if (tlen &&
10913 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
10914 	    (tp->t_fbyte_in == 0)) {
10915 		tp->t_fbyte_in = ticks;
10916 		if (tp->t_fbyte_in == 0)
10917 			tp->t_fbyte_in = 1;
10918 		if (tp->t_fbyte_out && tp->t_fbyte_in)
10919 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
10920 	}
10921 	/*
10922 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
10923 	 */
10924 	tp->snd_wl1 = th->th_seq;
10925 	/*
10926 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
10927 	 */
10928 	tp->rcv_up = tp->rcv_nxt;
10929 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
10930 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
10931 #ifdef TCPDEBUG
10932 	if (so->so_options & SO_DEBUG)
10933 		tcp_trace(TA_INPUT, ostate, tp,
10934 		    (void *)tcp_saveipgen, &tcp_savetcp, 0);
10935 #endif
10936 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
10937 
10938 	/* Add data to socket buffer. */
10939 	SOCKBUF_LOCK(&so->so_rcv);
10940 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
10941 		m_freem(m);
10942 	} else {
10943 		/*
10944 		 * Set new socket buffer size. Give up when limit is
10945 		 * reached.
10946 		 */
10947 		if (newsize)
10948 			if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
10949 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
10950 		m_adj(m, drop_hdrlen);	/* delayed header drop */
10951 #ifdef NETFLIX_SB_LIMITS
10952 		appended =
10953 #endif
10954 			sbappendstream_locked(&so->so_rcv, m, 0);
10955 		ctf_calc_rwin(so, tp);
10956 	}
10957 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
10958 	/* NB: sorwakeup_locked() does an implicit unlock. */
10959 	sorwakeup_locked(so);
10960 #ifdef NETFLIX_SB_LIMITS
10961 	if (so->so_rcv.sb_shlim && mcnt != appended)
10962 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
10963 #endif
10964 	rack_handle_delayed_ack(tp, rack, tlen, 0);
10965 	if (tp->snd_una == tp->snd_max)
10966 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
10967 	return (1);
10968 }
10969 
10970 /*
10971  * This subfunction is used to try to highly optimize the
10972  * fast path. We again allow window updates that are
10973  * in sequence to remain in the fast-path. We also add
10974  * in the __predict's to attempt to help the compiler.
10975  * Note that if we return a 0, then we can *not* process
10976  * it and the caller should push the packet into the
10977  * slow-path.
10978  */
10979 static int
10980 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
10981     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
10982     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
10983 {
10984 	int32_t acked;
10985 	int32_t nsegs;
10986 #ifdef TCPDEBUG
10987 	/*
10988 	 * The size of tcp_saveipgen must be the size of the max ip header,
10989 	 * now IPv6.
10990 	 */
10991 	u_char tcp_saveipgen[IP6_HDR_LEN];
10992 	struct tcphdr tcp_savetcp;
10993 	short ostate = 0;
10994 #endif
10995 	int32_t under_pacing = 0;
10996 	struct tcp_rack *rack;
10997 
10998 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
10999 		/* Old ack, behind (or duplicate to) the last one rcv'd */
11000 		return (0);
11001 	}
11002 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
11003 		/* Above what we have sent? */
11004 		return (0);
11005 	}
11006 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
11007 		/* We are retransmitting */
11008 		return (0);
11009 	}
11010 	if (__predict_false(tiwin == 0)) {
11011 		/* zero window */
11012 		return (0);
11013 	}
11014 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
11015 		/* We need a SYN or a FIN, unlikely.. */
11016 		return (0);
11017 	}
11018 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
11019 		/* Timestamp is behind .. old ack with seq wrap? */
11020 		return (0);
11021 	}
11022 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
11023 		/* Still recovering */
11024 		return (0);
11025 	}
11026 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11027 	if (rack->r_ctl.rc_sacked) {
11028 		/* We have sack holes on our scoreboard */
11029 		return (0);
11030 	}
11031 	/* Ok if we reach here, we can process a fast-ack */
11032 	if (rack->gp_ready &&
11033 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
11034 		under_pacing = 1;
11035 	}
11036 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
11037 	rack_log_ack(tp, to, th, 0, 0);
11038 	/* Did the window get updated? */
11039 	if (tiwin != tp->snd_wnd) {
11040 		tp->snd_wnd = tiwin;
11041 		rack_validate_fo_sendwin_up(tp, rack);
11042 		tp->snd_wl1 = th->th_seq;
11043 		if (tp->snd_wnd > tp->max_sndwnd)
11044 			tp->max_sndwnd = tp->snd_wnd;
11045 	}
11046 	/* Do we exit persists? */
11047 	if ((rack->rc_in_persist != 0) &&
11048 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
11049 			       rack->r_ctl.rc_pace_min_segs))) {
11050 		rack_exit_persist(tp, rack, cts);
11051 	}
11052 	/* Do we enter persists? */
11053 	if ((rack->rc_in_persist == 0) &&
11054 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
11055 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
11056 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
11057 	    sbavail(&tptosocket(tp)->so_snd) &&
11058 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
11059 		/*
11060 		 * Here the rwnd is less than
11061 		 * the pacing size, we are established,
11062 		 * nothing is outstanding, and there is
11063 		 * data to send. Enter persists.
11064 		 */
11065 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
11066 	}
11067 	/*
11068 	 * If last ACK falls within this segment's sequence numbers, record
11069 	 * the timestamp. NOTE that the test is modified according to the
11070 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
11071 	 */
11072 	if ((to->to_flags & TOF_TS) != 0 &&
11073 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
11074 		tp->ts_recent_age = tcp_ts_getticks();
11075 		tp->ts_recent = to->to_tsval;
11076 	}
11077 	/*
11078 	 * This is a pure ack for outstanding data.
11079 	 */
11080 	KMOD_TCPSTAT_INC(tcps_predack);
11081 
11082 	/*
11083 	 * "bad retransmit" recovery.
11084 	 */
11085 	if ((tp->t_flags & TF_PREVVALID) &&
11086 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
11087 		tp->t_flags &= ~TF_PREVVALID;
11088 		if (tp->t_rxtshift == 1 &&
11089 		    (int)(ticks - tp->t_badrxtwin) < 0)
11090 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
11091 	}
11092 	/*
11093 	 * Recalculate the transmit timer / rtt.
11094 	 *
11095 	 * Some boxes send broken timestamp replies during the SYN+ACK
11096 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
11097 	 * and blow up the retransmit timer.
11098 	 */
11099 	acked = BYTES_THIS_ACK(tp, th);
11100 
11101 #ifdef TCP_HHOOK
11102 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
11103 	hhook_run_tcp_est_in(tp, th, to);
11104 #endif
11105 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
11106 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
11107 	if (acked) {
11108 		struct mbuf *mfree;
11109 
11110 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
11111 		SOCKBUF_LOCK(&so->so_snd);
11112 		mfree = sbcut_locked(&so->so_snd, acked);
11113 		tp->snd_una = th->th_ack;
11114 		/* Note we want to hold the sb lock through the sendmap adjust */
11115 		rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
11116 		/* Wake up the socket if we have room to write more */
11117 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
11118 		sowwakeup_locked(so);
11119 		m_freem(mfree);
11120 		tp->t_rxtshift = 0;
11121 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
11122 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
11123 		rack->rc_tlp_in_progress = 0;
11124 		rack->r_ctl.rc_tlp_cnt_out = 0;
11125 		/*
11126 		 * If it is the RXT timer we want to
11127 		 * stop it, so we can restart a TLP.
11128 		 */
11129 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
11130 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11131 #ifdef NETFLIX_HTTP_LOGGING
11132 		tcp_http_check_for_comp(rack->rc_tp, th->th_ack);
11133 #endif
11134 	}
11135 	/*
11136 	 * Let the congestion control algorithm update congestion control
11137 	 * related information. This typically means increasing the
11138 	 * congestion window.
11139 	 */
11140 	if (tp->snd_wnd < ctf_outstanding(tp)) {
11141 		/* The peer collapsed the window */
11142 		rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__);
11143 	} else if (rack->rc_has_collapsed)
11144 		rack_un_collapse_window(rack, __LINE__);
11145 	if ((rack->r_collapse_point_valid) &&
11146 	    (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
11147 		rack->r_collapse_point_valid = 0;
11148 	/*
11149 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
11150 	 */
11151 	tp->snd_wl2 = th->th_ack;
11152 	tp->t_dupacks = 0;
11153 	m_freem(m);
11154 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
11155 
11156 	/*
11157 	 * If all outstanding data are acked, stop retransmit timer,
11158 	 * otherwise restart timer using current (possibly backed-off)
11159 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
11160 	 * If data are ready to send, let tcp_output decide between more
11161 	 * output or persist.
11162 	 */
11163 #ifdef TCPDEBUG
11164 	if (so->so_options & SO_DEBUG)
11165 		tcp_trace(TA_INPUT, ostate, tp,
11166 		    (void *)tcp_saveipgen,
11167 		    &tcp_savetcp, 0);
11168 #endif
11169 	if (under_pacing &&
11170 	    (rack->use_fixed_rate == 0) &&
11171 	    (rack->in_probe_rtt == 0) &&
11172 	    rack->rc_gp_dyn_mul &&
11173 	    rack->rc_always_pace) {
11174 		/* Check if we are dragging bottom */
11175 		rack_check_bottom_drag(tp, rack, so, acked);
11176 	}
11177 	if (tp->snd_una == tp->snd_max) {
11178 		tp->t_flags &= ~TF_PREVVALID;
11179 		rack->r_ctl.retran_during_recovery = 0;
11180 		rack->r_ctl.dsack_byte_cnt = 0;
11181 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
11182 		if (rack->r_ctl.rc_went_idle_time == 0)
11183 			rack->r_ctl.rc_went_idle_time = 1;
11184 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
11185 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
11186 			tp->t_acktime = 0;
11187 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11188 	}
11189 	if (acked && rack->r_fast_output)
11190 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
11191 	if (sbavail(&so->so_snd)) {
11192 		rack->r_wanted_output = 1;
11193 	}
11194 	return (1);
11195 }
11196 
11197 /*
11198  * Return value of 1, the TCB is unlocked and most
11199  * likely gone, return value of 0, the TCP is still
11200  * locked.
11201  */
11202 static int
11203 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
11204     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11205     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11206 {
11207 	int32_t ret_val = 0;
11208 	int32_t todrop;
11209 	int32_t ourfinisacked = 0;
11210 	struct tcp_rack *rack;
11211 
11212 	INP_WLOCK_ASSERT(tptoinpcb(tp));
11213 
11214 	ctf_calc_rwin(so, tp);
11215 	/*
11216 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
11217 	 * SYN, drop the input. if seg contains a RST, then drop the
11218 	 * connection. if seg does not contain SYN, then drop it. Otherwise
11219 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
11220 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
11221 	 * contains an ECE and ECN support is enabled, the stream is ECN
11222 	 * capable. if SYN has been acked change to ESTABLISHED else
11223 	 * SYN_RCVD state arrange for segment to be acked (eventually)
11224 	 * continue processing rest of data/controls.
11225 	 */
11226 	if ((thflags & TH_ACK) &&
11227 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
11228 	    SEQ_GT(th->th_ack, tp->snd_max))) {
11229 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11230 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11231 		return (1);
11232 	}
11233 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
11234 		TCP_PROBE5(connect__refused, NULL, tp,
11235 		    mtod(m, const char *), tp, th);
11236 		tp = tcp_drop(tp, ECONNREFUSED);
11237 		ctf_do_drop(m, tp);
11238 		return (1);
11239 	}
11240 	if (thflags & TH_RST) {
11241 		ctf_do_drop(m, tp);
11242 		return (1);
11243 	}
11244 	if (!(thflags & TH_SYN)) {
11245 		ctf_do_drop(m, tp);
11246 		return (1);
11247 	}
11248 	tp->irs = th->th_seq;
11249 	tcp_rcvseqinit(tp);
11250 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11251 	if (thflags & TH_ACK) {
11252 		int tfo_partial = 0;
11253 
11254 		KMOD_TCPSTAT_INC(tcps_connects);
11255 		soisconnected(so);
11256 #ifdef MAC
11257 		mac_socketpeer_set_from_mbuf(m, so);
11258 #endif
11259 		/* Do window scaling on this connection? */
11260 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11261 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11262 			tp->rcv_scale = tp->request_r_scale;
11263 		}
11264 		tp->rcv_adv += min(tp->rcv_wnd,
11265 		    TCP_MAXWIN << tp->rcv_scale);
11266 		/*
11267 		 * If not all the data that was sent in the TFO SYN
11268 		 * has been acked, resend the remainder right away.
11269 		 */
11270 		if (IS_FASTOPEN(tp->t_flags) &&
11271 		    (tp->snd_una != tp->snd_max)) {
11272 			tp->snd_nxt = th->th_ack;
11273 			tfo_partial = 1;
11274 		}
11275 		/*
11276 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
11277 		 * will be turned on later.
11278 		 */
11279 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
11280 			rack_timer_cancel(tp, rack,
11281 					  rack->r_ctl.rc_rcvtime, __LINE__);
11282 			tp->t_flags |= TF_DELACK;
11283 		} else {
11284 			rack->r_wanted_output = 1;
11285 			tp->t_flags |= TF_ACKNOW;
11286 			rack->rc_dack_toggle = 0;
11287 		}
11288 
11289 		tcp_ecn_input_syn_sent(tp, thflags, iptos);
11290 
11291 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
11292 			/*
11293 			 * We advance snd_una for the
11294 			 * fast open case. If th_ack is
11295 			 * acknowledging data beyond
11296 			 * snd_una we can't just call
11297 			 * ack-processing since the
11298 			 * data stream in our send-map
11299 			 * will start at snd_una + 1 (one
11300 			 * beyond the SYN). If its just
11301 			 * equal we don't need to do that
11302 			 * and there is no send_map.
11303 			 */
11304 			tp->snd_una++;
11305 		}
11306 		/*
11307 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
11308 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
11309 		 */
11310 		tp->t_starttime = ticks;
11311 		if (tp->t_flags & TF_NEEDFIN) {
11312 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
11313 			tp->t_flags &= ~TF_NEEDFIN;
11314 			thflags &= ~TH_SYN;
11315 		} else {
11316 			tcp_state_change(tp, TCPS_ESTABLISHED);
11317 			TCP_PROBE5(connect__established, NULL, tp,
11318 			    mtod(m, const char *), tp, th);
11319 			rack_cc_conn_init(tp);
11320 		}
11321 	} else {
11322 		/*
11323 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
11324 		 * open.  If segment contains CC option and there is a
11325 		 * cached CC, apply TAO test. If it succeeds, connection is *
11326 		 * half-synchronized. Otherwise, do 3-way handshake:
11327 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
11328 		 * there was no CC option, clear cached CC value.
11329 		 */
11330 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
11331 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
11332 	}
11333 	/*
11334 	 * Advance th->th_seq to correspond to first data byte. If data,
11335 	 * trim to stay within window, dropping FIN if necessary.
11336 	 */
11337 	th->th_seq++;
11338 	if (tlen > tp->rcv_wnd) {
11339 		todrop = tlen - tp->rcv_wnd;
11340 		m_adj(m, -todrop);
11341 		tlen = tp->rcv_wnd;
11342 		thflags &= ~TH_FIN;
11343 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
11344 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
11345 	}
11346 	tp->snd_wl1 = th->th_seq - 1;
11347 	tp->rcv_up = th->th_seq;
11348 	/*
11349 	 * Client side of transaction: already sent SYN and data. If the
11350 	 * remote host used T/TCP to validate the SYN, our data will be
11351 	 * ACK'd; if so, enter normal data segment processing in the middle
11352 	 * of step 5, ack processing. Otherwise, goto step 6.
11353 	 */
11354 	if (thflags & TH_ACK) {
11355 		/* For syn-sent we need to possibly update the rtt */
11356 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11357 			uint32_t t, mcts;
11358 
11359 			mcts = tcp_ts_getticks();
11360 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11361 			if (!tp->t_rttlow || tp->t_rttlow > t)
11362 				tp->t_rttlow = t;
11363 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
11364 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11365 			tcp_rack_xmit_timer_commit(rack, tp);
11366 		}
11367 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
11368 			return (ret_val);
11369 		/* We may have changed to FIN_WAIT_1 above */
11370 		if (tp->t_state == TCPS_FIN_WAIT_1) {
11371 			/*
11372 			 * In FIN_WAIT_1 STATE in addition to the processing
11373 			 * for the ESTABLISHED state if our FIN is now
11374 			 * acknowledged then enter FIN_WAIT_2.
11375 			 */
11376 			if (ourfinisacked) {
11377 				/*
11378 				 * If we can't receive any more data, then
11379 				 * closing user can proceed. Starting the
11380 				 * timer is contrary to the specification,
11381 				 * but if we don't get a FIN we'll hang
11382 				 * forever.
11383 				 *
11384 				 * XXXjl: we should release the tp also, and
11385 				 * use a compressed state.
11386 				 */
11387 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11388 					soisdisconnected(so);
11389 					tcp_timer_activate(tp, TT_2MSL,
11390 					    (tcp_fast_finwait2_recycle ?
11391 					    tcp_finwait2_timeout :
11392 					    TP_MAXIDLE(tp)));
11393 				}
11394 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
11395 			}
11396 		}
11397 	}
11398 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11399 	   tiwin, thflags, nxt_pkt));
11400 }
11401 
11402 /*
11403  * Return value of 1, the TCB is unlocked and most
11404  * likely gone, return value of 0, the TCP is still
11405  * locked.
11406  */
11407 static int
11408 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
11409     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11410     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11411 {
11412 	struct tcp_rack *rack;
11413 	int32_t ret_val = 0;
11414 	int32_t ourfinisacked = 0;
11415 
11416 	ctf_calc_rwin(so, tp);
11417 	if ((thflags & TH_ACK) &&
11418 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
11419 	    SEQ_GT(th->th_ack, tp->snd_max))) {
11420 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11421 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11422 		return (1);
11423 	}
11424 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11425 	if (IS_FASTOPEN(tp->t_flags)) {
11426 		/*
11427 		 * When a TFO connection is in SYN_RECEIVED, the
11428 		 * only valid packets are the initial SYN, a
11429 		 * retransmit/copy of the initial SYN (possibly with
11430 		 * a subset of the original data), a valid ACK, a
11431 		 * FIN, or a RST.
11432 		 */
11433 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
11434 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11435 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11436 			return (1);
11437 		} else if (thflags & TH_SYN) {
11438 			/* non-initial SYN is ignored */
11439 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
11440 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
11441 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
11442 				ctf_do_drop(m, NULL);
11443 				return (0);
11444 			}
11445 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
11446 			ctf_do_drop(m, NULL);
11447 			return (0);
11448 		}
11449 	}
11450 
11451 	if ((thflags & TH_RST) ||
11452 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11453 		return (__ctf_process_rst(m, th, so, tp,
11454 					  &rack->r_ctl.challenge_ack_ts,
11455 					  &rack->r_ctl.challenge_ack_cnt));
11456 	/*
11457 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11458 	 * it's less than ts_recent, drop it.
11459 	 */
11460 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11461 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11462 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11463 			return (ret_val);
11464 	}
11465 	/*
11466 	 * In the SYN-RECEIVED state, validate that the packet belongs to
11467 	 * this connection before trimming the data to fit the receive
11468 	 * window.  Check the sequence number versus IRS since we know the
11469 	 * sequence numbers haven't wrapped.  This is a partial fix for the
11470 	 * "LAND" DoS attack.
11471 	 */
11472 	if (SEQ_LT(th->th_seq, tp->irs)) {
11473 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
11474 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11475 		return (1);
11476 	}
11477 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11478 			      &rack->r_ctl.challenge_ack_ts,
11479 			      &rack->r_ctl.challenge_ack_cnt)) {
11480 		return (ret_val);
11481 	}
11482 	/*
11483 	 * If last ACK falls within this segment's sequence numbers, record
11484 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11485 	 * from the latest proposal of the tcplw@cray.com list (Braden
11486 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11487 	 * with our earlier PAWS tests, so this check should be solely
11488 	 * predicated on the sequence space of this segment. 3) That we
11489 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11490 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11491 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11492 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11493 	 * p.869. In such cases, we can still calculate the RTT correctly
11494 	 * when RCV.NXT == Last.ACK.Sent.
11495 	 */
11496 	if ((to->to_flags & TOF_TS) != 0 &&
11497 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11498 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11499 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11500 		tp->ts_recent_age = tcp_ts_getticks();
11501 		tp->ts_recent = to->to_tsval;
11502 	}
11503 	tp->snd_wnd = tiwin;
11504 	rack_validate_fo_sendwin_up(tp, rack);
11505 	/*
11506 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11507 	 * is on (half-synchronized state), then queue data for later
11508 	 * processing; else drop segment and return.
11509 	 */
11510 	if ((thflags & TH_ACK) == 0) {
11511 		if (IS_FASTOPEN(tp->t_flags)) {
11512 			rack_cc_conn_init(tp);
11513 		}
11514 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11515 		    tiwin, thflags, nxt_pkt));
11516 	}
11517 	KMOD_TCPSTAT_INC(tcps_connects);
11518 	if (tp->t_flags & TF_SONOTCONN) {
11519 		tp->t_flags &= ~TF_SONOTCONN;
11520 		soisconnected(so);
11521 	}
11522 	/* Do window scaling? */
11523 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
11524 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
11525 		tp->rcv_scale = tp->request_r_scale;
11526 	}
11527 	/*
11528 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
11529 	 * FIN-WAIT-1
11530 	 */
11531 	tp->t_starttime = ticks;
11532 	if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
11533 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
11534 		tp->t_tfo_pending = NULL;
11535 	}
11536 	if (tp->t_flags & TF_NEEDFIN) {
11537 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
11538 		tp->t_flags &= ~TF_NEEDFIN;
11539 	} else {
11540 		tcp_state_change(tp, TCPS_ESTABLISHED);
11541 		TCP_PROBE5(accept__established, NULL, tp,
11542 		    mtod(m, const char *), tp, th);
11543 		/*
11544 		 * TFO connections call cc_conn_init() during SYN
11545 		 * processing.  Calling it again here for such connections
11546 		 * is not harmless as it would undo the snd_cwnd reduction
11547 		 * that occurs when a TFO SYN|ACK is retransmitted.
11548 		 */
11549 		if (!IS_FASTOPEN(tp->t_flags))
11550 			rack_cc_conn_init(tp);
11551 	}
11552 	/*
11553 	 * Account for the ACK of our SYN prior to
11554 	 * regular ACK processing below, except for
11555 	 * simultaneous SYN, which is handled later.
11556 	 */
11557 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
11558 		tp->snd_una++;
11559 	/*
11560 	 * If segment contains data or ACK, will call tcp_reass() later; if
11561 	 * not, do so now to pass queued data to user.
11562 	 */
11563 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
11564 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
11565 		    (struct mbuf *)0);
11566 		if (tp->t_flags & TF_WAKESOR) {
11567 			tp->t_flags &= ~TF_WAKESOR;
11568 			/* NB: sorwakeup_locked() does an implicit unlock. */
11569 			sorwakeup_locked(so);
11570 		}
11571 	}
11572 	tp->snd_wl1 = th->th_seq - 1;
11573 	/* For syn-recv we need to possibly update the rtt */
11574 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
11575 		uint32_t t, mcts;
11576 
11577 		mcts = tcp_ts_getticks();
11578 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
11579 		if (!tp->t_rttlow || tp->t_rttlow > t)
11580 			tp->t_rttlow = t;
11581 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
11582 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
11583 		tcp_rack_xmit_timer_commit(rack, tp);
11584 	}
11585 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11586 		return (ret_val);
11587 	}
11588 	if (tp->t_state == TCPS_FIN_WAIT_1) {
11589 		/* We could have went to FIN_WAIT_1 (or EST) above */
11590 		/*
11591 		 * In FIN_WAIT_1 STATE in addition to the processing for the
11592 		 * ESTABLISHED state if our FIN is now acknowledged then
11593 		 * enter FIN_WAIT_2.
11594 		 */
11595 		if (ourfinisacked) {
11596 			/*
11597 			 * If we can't receive any more data, then closing
11598 			 * user can proceed. Starting the timer is contrary
11599 			 * to the specification, but if we don't get a FIN
11600 			 * we'll hang forever.
11601 			 *
11602 			 * XXXjl: we should release the tp also, and use a
11603 			 * compressed state.
11604 			 */
11605 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11606 				soisdisconnected(so);
11607 				tcp_timer_activate(tp, TT_2MSL,
11608 				    (tcp_fast_finwait2_recycle ?
11609 				    tcp_finwait2_timeout :
11610 				    TP_MAXIDLE(tp)));
11611 			}
11612 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
11613 		}
11614 	}
11615 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11616 	    tiwin, thflags, nxt_pkt));
11617 }
11618 
11619 /*
11620  * Return value of 1, the TCB is unlocked and most
11621  * likely gone, return value of 0, the TCP is still
11622  * locked.
11623  */
11624 static int
11625 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
11626     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11627     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11628 {
11629 	int32_t ret_val = 0;
11630 	struct tcp_rack *rack;
11631 
11632 	/*
11633 	 * Header prediction: check for the two common cases of a
11634 	 * uni-directional data xfer.  If the packet has no control flags,
11635 	 * is in-sequence, the window didn't change and we're not
11636 	 * retransmitting, it's a candidate.  If the length is zero and the
11637 	 * ack moved forward, we're the sender side of the xfer.  Just free
11638 	 * the data acked & wake any higher level process that was blocked
11639 	 * waiting for space.  If the length is non-zero and the ack didn't
11640 	 * move, we're the receiver side.  If we're getting packets in-order
11641 	 * (the reassembly queue is empty), add the data toc The socket
11642 	 * buffer and note that we need a delayed ack. Make sure that the
11643 	 * hidden state-flags are also off. Since we check for
11644 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
11645 	 */
11646 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11647 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
11648 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
11649 	    __predict_true(SEGQ_EMPTY(tp)) &&
11650 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
11651 		if (tlen == 0) {
11652 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
11653 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
11654 				return (0);
11655 			}
11656 		} else {
11657 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
11658 			    tiwin, nxt_pkt, iptos)) {
11659 				return (0);
11660 			}
11661 		}
11662 	}
11663 	ctf_calc_rwin(so, tp);
11664 
11665 	if ((thflags & TH_RST) ||
11666 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11667 		return (__ctf_process_rst(m, th, so, tp,
11668 					  &rack->r_ctl.challenge_ack_ts,
11669 					  &rack->r_ctl.challenge_ack_cnt));
11670 
11671 	/*
11672 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11673 	 * synchronized state.
11674 	 */
11675 	if (thflags & TH_SYN) {
11676 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11677 		return (ret_val);
11678 	}
11679 	/*
11680 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11681 	 * it's less than ts_recent, drop it.
11682 	 */
11683 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11684 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11685 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11686 			return (ret_val);
11687 	}
11688 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11689 			      &rack->r_ctl.challenge_ack_ts,
11690 			      &rack->r_ctl.challenge_ack_cnt)) {
11691 		return (ret_val);
11692 	}
11693 	/*
11694 	 * If last ACK falls within this segment's sequence numbers, record
11695 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11696 	 * from the latest proposal of the tcplw@cray.com list (Braden
11697 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11698 	 * with our earlier PAWS tests, so this check should be solely
11699 	 * predicated on the sequence space of this segment. 3) That we
11700 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11701 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11702 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11703 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11704 	 * p.869. In such cases, we can still calculate the RTT correctly
11705 	 * when RCV.NXT == Last.ACK.Sent.
11706 	 */
11707 	if ((to->to_flags & TOF_TS) != 0 &&
11708 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11709 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11710 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11711 		tp->ts_recent_age = tcp_ts_getticks();
11712 		tp->ts_recent = to->to_tsval;
11713 	}
11714 	/*
11715 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11716 	 * is on (half-synchronized state), then queue data for later
11717 	 * processing; else drop segment and return.
11718 	 */
11719 	if ((thflags & TH_ACK) == 0) {
11720 		if (tp->t_flags & TF_NEEDSYN) {
11721 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11722 			    tiwin, thflags, nxt_pkt));
11723 
11724 		} else if (tp->t_flags & TF_ACKNOW) {
11725 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11726 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11727 			return (ret_val);
11728 		} else {
11729 			ctf_do_drop(m, NULL);
11730 			return (0);
11731 		}
11732 	}
11733 	/*
11734 	 * Ack processing.
11735 	 */
11736 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11737 		return (ret_val);
11738 	}
11739 	if (sbavail(&so->so_snd)) {
11740 		if (ctf_progress_timeout_check(tp, true)) {
11741 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
11742 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11743 			return (1);
11744 		}
11745 	}
11746 	/* State changes only happen in rack_process_data() */
11747 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11748 	    tiwin, thflags, nxt_pkt));
11749 }
11750 
11751 /*
11752  * Return value of 1, the TCB is unlocked and most
11753  * likely gone, return value of 0, the TCP is still
11754  * locked.
11755  */
11756 static int
11757 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
11758     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11759     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11760 {
11761 	int32_t ret_val = 0;
11762 	struct tcp_rack *rack;
11763 
11764 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11765 	ctf_calc_rwin(so, tp);
11766 	if ((thflags & TH_RST) ||
11767 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11768 		return (__ctf_process_rst(m, th, so, tp,
11769 					  &rack->r_ctl.challenge_ack_ts,
11770 					  &rack->r_ctl.challenge_ack_cnt));
11771 	/*
11772 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11773 	 * synchronized state.
11774 	 */
11775 	if (thflags & TH_SYN) {
11776 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11777 		return (ret_val);
11778 	}
11779 	/*
11780 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11781 	 * it's less than ts_recent, drop it.
11782 	 */
11783 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11784 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11785 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11786 			return (ret_val);
11787 	}
11788 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11789 			      &rack->r_ctl.challenge_ack_ts,
11790 			      &rack->r_ctl.challenge_ack_cnt)) {
11791 		return (ret_val);
11792 	}
11793 	/*
11794 	 * If last ACK falls within this segment's sequence numbers, record
11795 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11796 	 * from the latest proposal of the tcplw@cray.com list (Braden
11797 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11798 	 * with our earlier PAWS tests, so this check should be solely
11799 	 * predicated on the sequence space of this segment. 3) That we
11800 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11801 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11802 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11803 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11804 	 * p.869. In such cases, we can still calculate the RTT correctly
11805 	 * when RCV.NXT == Last.ACK.Sent.
11806 	 */
11807 	if ((to->to_flags & TOF_TS) != 0 &&
11808 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11809 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11810 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11811 		tp->ts_recent_age = tcp_ts_getticks();
11812 		tp->ts_recent = to->to_tsval;
11813 	}
11814 	/*
11815 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11816 	 * is on (half-synchronized state), then queue data for later
11817 	 * processing; else drop segment and return.
11818 	 */
11819 	if ((thflags & TH_ACK) == 0) {
11820 		if (tp->t_flags & TF_NEEDSYN) {
11821 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11822 			    tiwin, thflags, nxt_pkt));
11823 
11824 		} else if (tp->t_flags & TF_ACKNOW) {
11825 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11826 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11827 			return (ret_val);
11828 		} else {
11829 			ctf_do_drop(m, NULL);
11830 			return (0);
11831 		}
11832 	}
11833 	/*
11834 	 * Ack processing.
11835 	 */
11836 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
11837 		return (ret_val);
11838 	}
11839 	if (sbavail(&so->so_snd)) {
11840 		if (ctf_progress_timeout_check(tp, true)) {
11841 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11842 						tp, tick, PROGRESS_DROP, __LINE__);
11843 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11844 			return (1);
11845 		}
11846 	}
11847 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11848 	    tiwin, thflags, nxt_pkt));
11849 }
11850 
11851 static int
11852 rack_check_data_after_close(struct mbuf *m,
11853     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
11854 {
11855 	struct tcp_rack *rack;
11856 
11857 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11858 	if (rack->rc_allow_data_af_clo == 0) {
11859 	close_now:
11860 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11861 		/* tcp_close will kill the inp pre-log the Reset */
11862 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
11863 		tp = tcp_close(tp);
11864 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
11865 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
11866 		return (1);
11867 	}
11868 	if (sbavail(&so->so_snd) == 0)
11869 		goto close_now;
11870 	/* Ok we allow data that is ignored and a followup reset */
11871 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
11872 	tp->rcv_nxt = th->th_seq + *tlen;
11873 	tp->t_flags2 |= TF2_DROP_AF_DATA;
11874 	rack->r_wanted_output = 1;
11875 	*tlen = 0;
11876 	return (0);
11877 }
11878 
11879 /*
11880  * Return value of 1, the TCB is unlocked and most
11881  * likely gone, return value of 0, the TCP is still
11882  * locked.
11883  */
11884 static int
11885 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
11886     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
11887     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
11888 {
11889 	int32_t ret_val = 0;
11890 	int32_t ourfinisacked = 0;
11891 	struct tcp_rack *rack;
11892 
11893 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11894 	ctf_calc_rwin(so, tp);
11895 
11896 	if ((thflags & TH_RST) ||
11897 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
11898 		return (__ctf_process_rst(m, th, so, tp,
11899 					  &rack->r_ctl.challenge_ack_ts,
11900 					  &rack->r_ctl.challenge_ack_cnt));
11901 	/*
11902 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
11903 	 * synchronized state.
11904 	 */
11905 	if (thflags & TH_SYN) {
11906 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
11907 		return (ret_val);
11908 	}
11909 	/*
11910 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
11911 	 * it's less than ts_recent, drop it.
11912 	 */
11913 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
11914 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
11915 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
11916 			return (ret_val);
11917 	}
11918 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
11919 			      &rack->r_ctl.challenge_ack_ts,
11920 			      &rack->r_ctl.challenge_ack_cnt)) {
11921 		return (ret_val);
11922 	}
11923 	/*
11924 	 * If new data are received on a connection after the user processes
11925 	 * are gone, then RST the other end.
11926 	 */
11927 	if ((tp->t_flags & TF_CLOSED) && tlen &&
11928 	    rack_check_data_after_close(m, tp, &tlen, th, so))
11929 		return (1);
11930 	/*
11931 	 * If last ACK falls within this segment's sequence numbers, record
11932 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
11933 	 * from the latest proposal of the tcplw@cray.com list (Braden
11934 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
11935 	 * with our earlier PAWS tests, so this check should be solely
11936 	 * predicated on the sequence space of this segment. 3) That we
11937 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
11938 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
11939 	 * SEG.Len, This modified check allows us to overcome RFC1323's
11940 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
11941 	 * p.869. In such cases, we can still calculate the RTT correctly
11942 	 * when RCV.NXT == Last.ACK.Sent.
11943 	 */
11944 	if ((to->to_flags & TOF_TS) != 0 &&
11945 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
11946 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
11947 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
11948 		tp->ts_recent_age = tcp_ts_getticks();
11949 		tp->ts_recent = to->to_tsval;
11950 	}
11951 	/*
11952 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
11953 	 * is on (half-synchronized state), then queue data for later
11954 	 * processing; else drop segment and return.
11955 	 */
11956 	if ((thflags & TH_ACK) == 0) {
11957 		if (tp->t_flags & TF_NEEDSYN) {
11958 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
11959 			    tiwin, thflags, nxt_pkt));
11960 		} else if (tp->t_flags & TF_ACKNOW) {
11961 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
11962 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
11963 			return (ret_val);
11964 		} else {
11965 			ctf_do_drop(m, NULL);
11966 			return (0);
11967 		}
11968 	}
11969 	/*
11970 	 * Ack processing.
11971 	 */
11972 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
11973 		return (ret_val);
11974 	}
11975 	if (ourfinisacked) {
11976 		/*
11977 		 * If we can't receive any more data, then closing user can
11978 		 * proceed. Starting the timer is contrary to the
11979 		 * specification, but if we don't get a FIN we'll hang
11980 		 * forever.
11981 		 *
11982 		 * XXXjl: we should release the tp also, and use a
11983 		 * compressed state.
11984 		 */
11985 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
11986 			soisdisconnected(so);
11987 			tcp_timer_activate(tp, TT_2MSL,
11988 			    (tcp_fast_finwait2_recycle ?
11989 			    tcp_finwait2_timeout :
11990 			    TP_MAXIDLE(tp)));
11991 		}
11992 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
11993 	}
11994 	if (sbavail(&so->so_snd)) {
11995 		if (ctf_progress_timeout_check(tp, true)) {
11996 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
11997 						tp, tick, PROGRESS_DROP, __LINE__);
11998 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
11999 			return (1);
12000 		}
12001 	}
12002 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12003 	    tiwin, thflags, nxt_pkt));
12004 }
12005 
12006 /*
12007  * Return value of 1, the TCB is unlocked and most
12008  * likely gone, return value of 0, the TCP is still
12009  * locked.
12010  */
12011 static int
12012 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
12013     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12014     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12015 {
12016 	int32_t ret_val = 0;
12017 	int32_t ourfinisacked = 0;
12018 	struct tcp_rack *rack;
12019 
12020 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12021 	ctf_calc_rwin(so, tp);
12022 
12023 	if ((thflags & TH_RST) ||
12024 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12025 		return (__ctf_process_rst(m, th, so, tp,
12026 					  &rack->r_ctl.challenge_ack_ts,
12027 					  &rack->r_ctl.challenge_ack_cnt));
12028 	/*
12029 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12030 	 * synchronized state.
12031 	 */
12032 	if (thflags & TH_SYN) {
12033 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
12034 		return (ret_val);
12035 	}
12036 	/*
12037 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12038 	 * it's less than ts_recent, drop it.
12039 	 */
12040 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12041 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12042 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12043 			return (ret_val);
12044 	}
12045 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12046 			      &rack->r_ctl.challenge_ack_ts,
12047 			      &rack->r_ctl.challenge_ack_cnt)) {
12048 		return (ret_val);
12049 	}
12050 	/*
12051 	 * If new data are received on a connection after the user processes
12052 	 * are gone, then RST the other end.
12053 	 */
12054 	if ((tp->t_flags & TF_CLOSED) && tlen &&
12055 	    rack_check_data_after_close(m, tp, &tlen, th, so))
12056 		return (1);
12057 	/*
12058 	 * If last ACK falls within this segment's sequence numbers, record
12059 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12060 	 * from the latest proposal of the tcplw@cray.com list (Braden
12061 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12062 	 * with our earlier PAWS tests, so this check should be solely
12063 	 * predicated on the sequence space of this segment. 3) That we
12064 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12065 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12066 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12067 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12068 	 * p.869. In such cases, we can still calculate the RTT correctly
12069 	 * when RCV.NXT == Last.ACK.Sent.
12070 	 */
12071 	if ((to->to_flags & TOF_TS) != 0 &&
12072 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12073 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12074 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12075 		tp->ts_recent_age = tcp_ts_getticks();
12076 		tp->ts_recent = to->to_tsval;
12077 	}
12078 	/*
12079 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12080 	 * is on (half-synchronized state), then queue data for later
12081 	 * processing; else drop segment and return.
12082 	 */
12083 	if ((thflags & TH_ACK) == 0) {
12084 		if (tp->t_flags & TF_NEEDSYN) {
12085 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12086 			    tiwin, thflags, nxt_pkt));
12087 		} else if (tp->t_flags & TF_ACKNOW) {
12088 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12089 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12090 			return (ret_val);
12091 		} else {
12092 			ctf_do_drop(m, NULL);
12093 			return (0);
12094 		}
12095 	}
12096 	/*
12097 	 * Ack processing.
12098 	 */
12099 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12100 		return (ret_val);
12101 	}
12102 	if (ourfinisacked) {
12103 		tcp_twstart(tp);
12104 		m_freem(m);
12105 		return (1);
12106 	}
12107 	if (sbavail(&so->so_snd)) {
12108 		if (ctf_progress_timeout_check(tp, true)) {
12109 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12110 						tp, tick, PROGRESS_DROP, __LINE__);
12111 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12112 			return (1);
12113 		}
12114 	}
12115 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12116 	    tiwin, thflags, nxt_pkt));
12117 }
12118 
12119 /*
12120  * Return value of 1, the TCB is unlocked and most
12121  * likely gone, return value of 0, the TCP is still
12122  * locked.
12123  */
12124 static int
12125 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12126     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12127     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12128 {
12129 	int32_t ret_val = 0;
12130 	int32_t ourfinisacked = 0;
12131 	struct tcp_rack *rack;
12132 
12133 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12134 	ctf_calc_rwin(so, tp);
12135 
12136 	if ((thflags & TH_RST) ||
12137 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12138 		return (__ctf_process_rst(m, th, so, tp,
12139 					  &rack->r_ctl.challenge_ack_ts,
12140 					  &rack->r_ctl.challenge_ack_cnt));
12141 	/*
12142 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12143 	 * synchronized state.
12144 	 */
12145 	if (thflags & TH_SYN) {
12146 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
12147 		return (ret_val);
12148 	}
12149 	/*
12150 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12151 	 * it's less than ts_recent, drop it.
12152 	 */
12153 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12154 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12155 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12156 			return (ret_val);
12157 	}
12158 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12159 			      &rack->r_ctl.challenge_ack_ts,
12160 			      &rack->r_ctl.challenge_ack_cnt)) {
12161 		return (ret_val);
12162 	}
12163 	/*
12164 	 * If new data are received on a connection after the user processes
12165 	 * are gone, then RST the other end.
12166 	 */
12167 	if ((tp->t_flags & TF_CLOSED) && tlen &&
12168 	    rack_check_data_after_close(m, tp, &tlen, th, so))
12169 		return (1);
12170 	/*
12171 	 * If last ACK falls within this segment's sequence numbers, record
12172 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12173 	 * from the latest proposal of the tcplw@cray.com list (Braden
12174 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12175 	 * with our earlier PAWS tests, so this check should be solely
12176 	 * predicated on the sequence space of this segment. 3) That we
12177 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12178 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12179 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12180 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12181 	 * p.869. In such cases, we can still calculate the RTT correctly
12182 	 * when RCV.NXT == Last.ACK.Sent.
12183 	 */
12184 	if ((to->to_flags & TOF_TS) != 0 &&
12185 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12186 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12187 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12188 		tp->ts_recent_age = tcp_ts_getticks();
12189 		tp->ts_recent = to->to_tsval;
12190 	}
12191 	/*
12192 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12193 	 * is on (half-synchronized state), then queue data for later
12194 	 * processing; else drop segment and return.
12195 	 */
12196 	if ((thflags & TH_ACK) == 0) {
12197 		if (tp->t_flags & TF_NEEDSYN) {
12198 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12199 			    tiwin, thflags, nxt_pkt));
12200 		} else if (tp->t_flags & TF_ACKNOW) {
12201 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12202 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12203 			return (ret_val);
12204 		} else {
12205 			ctf_do_drop(m, NULL);
12206 			return (0);
12207 		}
12208 	}
12209 	/*
12210 	 * case TCPS_LAST_ACK: Ack processing.
12211 	 */
12212 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12213 		return (ret_val);
12214 	}
12215 	if (ourfinisacked) {
12216 		tp = tcp_close(tp);
12217 		ctf_do_drop(m, tp);
12218 		return (1);
12219 	}
12220 	if (sbavail(&so->so_snd)) {
12221 		if (ctf_progress_timeout_check(tp, true)) {
12222 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12223 						tp, tick, PROGRESS_DROP, __LINE__);
12224 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12225 			return (1);
12226 		}
12227 	}
12228 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12229 	    tiwin, thflags, nxt_pkt));
12230 }
12231 
12232 /*
12233  * Return value of 1, the TCB is unlocked and most
12234  * likely gone, return value of 0, the TCP is still
12235  * locked.
12236  */
12237 static int
12238 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
12239     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12240     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
12241 {
12242 	int32_t ret_val = 0;
12243 	int32_t ourfinisacked = 0;
12244 	struct tcp_rack *rack;
12245 
12246 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12247 	ctf_calc_rwin(so, tp);
12248 
12249 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
12250 	if ((thflags & TH_RST) ||
12251 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
12252 		return (__ctf_process_rst(m, th, so, tp,
12253 					  &rack->r_ctl.challenge_ack_ts,
12254 					  &rack->r_ctl.challenge_ack_cnt));
12255 	/*
12256 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
12257 	 * synchronized state.
12258 	 */
12259 	if (thflags & TH_SYN) {
12260 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
12261 		return (ret_val);
12262 	}
12263 	/*
12264 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
12265 	 * it's less than ts_recent, drop it.
12266 	 */
12267 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
12268 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
12269 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
12270 			return (ret_val);
12271 	}
12272 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
12273 			      &rack->r_ctl.challenge_ack_ts,
12274 			      &rack->r_ctl.challenge_ack_cnt)) {
12275 		return (ret_val);
12276 	}
12277 	/*
12278 	 * If new data are received on a connection after the user processes
12279 	 * are gone, then RST the other end.
12280 	 */
12281 	if ((tp->t_flags & TF_CLOSED) && tlen &&
12282 	    rack_check_data_after_close(m, tp, &tlen, th, so))
12283 		return (1);
12284 	/*
12285 	 * If last ACK falls within this segment's sequence numbers, record
12286 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
12287 	 * from the latest proposal of the tcplw@cray.com list (Braden
12288 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
12289 	 * with our earlier PAWS tests, so this check should be solely
12290 	 * predicated on the sequence space of this segment. 3) That we
12291 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
12292 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
12293 	 * SEG.Len, This modified check allows us to overcome RFC1323's
12294 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
12295 	 * p.869. In such cases, we can still calculate the RTT correctly
12296 	 * when RCV.NXT == Last.ACK.Sent.
12297 	 */
12298 	if ((to->to_flags & TOF_TS) != 0 &&
12299 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
12300 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
12301 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
12302 		tp->ts_recent_age = tcp_ts_getticks();
12303 		tp->ts_recent = to->to_tsval;
12304 	}
12305 	/*
12306 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
12307 	 * is on (half-synchronized state), then queue data for later
12308 	 * processing; else drop segment and return.
12309 	 */
12310 	if ((thflags & TH_ACK) == 0) {
12311 		if (tp->t_flags & TF_NEEDSYN) {
12312 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12313 			    tiwin, thflags, nxt_pkt));
12314 		} else if (tp->t_flags & TF_ACKNOW) {
12315 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
12316 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
12317 			return (ret_val);
12318 		} else {
12319 			ctf_do_drop(m, NULL);
12320 			return (0);
12321 		}
12322 	}
12323 	/*
12324 	 * Ack processing.
12325 	 */
12326 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
12327 		return (ret_val);
12328 	}
12329 	if (sbavail(&so->so_snd)) {
12330 		if (ctf_progress_timeout_check(tp, true)) {
12331 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
12332 						tp, tick, PROGRESS_DROP, __LINE__);
12333 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
12334 			return (1);
12335 		}
12336 	}
12337 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
12338 	    tiwin, thflags, nxt_pkt));
12339 }
12340 
12341 static void inline
12342 rack_clear_rate_sample(struct tcp_rack *rack)
12343 {
12344 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
12345 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
12346 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
12347 }
12348 
12349 static void
12350 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
12351 {
12352 	uint64_t bw_est, rate_wanted;
12353 	int chged = 0;
12354 	uint32_t user_max, orig_min, orig_max;
12355 
12356 	orig_min = rack->r_ctl.rc_pace_min_segs;
12357 	orig_max = rack->r_ctl.rc_pace_max_segs;
12358 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
12359 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
12360 		chged = 1;
12361 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
12362 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
12363 		if (user_max != rack->r_ctl.rc_pace_max_segs)
12364 			chged = 1;
12365 	}
12366 	if (rack->rc_force_max_seg) {
12367 		rack->r_ctl.rc_pace_max_segs = user_max;
12368 	} else if (rack->use_fixed_rate) {
12369 		bw_est = rack_get_bw(rack);
12370 		if ((rack->r_ctl.crte == NULL) ||
12371 		    (bw_est != rack->r_ctl.crte->rate)) {
12372 			rack->r_ctl.rc_pace_max_segs = user_max;
12373 		} else {
12374 			/* We are pacing right at the hardware rate */
12375 			uint32_t segsiz;
12376 
12377 			segsiz = min(ctf_fixed_maxseg(tp),
12378 				     rack->r_ctl.rc_pace_min_segs);
12379 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(
12380 				                           tp, bw_est, segsiz, 0,
12381 							   rack->r_ctl.crte, NULL);
12382 		}
12383 	} else if (rack->rc_always_pace) {
12384 		if (rack->r_ctl.gp_bw ||
12385 #ifdef NETFLIX_PEAKRATE
12386 		    rack->rc_tp->t_maxpeakrate ||
12387 #endif
12388 		    rack->r_ctl.init_rate) {
12389 			/* We have a rate of some sort set */
12390 			uint32_t  orig;
12391 
12392 			bw_est = rack_get_bw(rack);
12393 			orig = rack->r_ctl.rc_pace_max_segs;
12394 			if (fill_override)
12395 				rate_wanted = *fill_override;
12396 			else
12397 				rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL);
12398 			if (rate_wanted) {
12399 				/* We have something */
12400 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
12401 										   rate_wanted,
12402 										   ctf_fixed_maxseg(rack->rc_tp));
12403 			} else
12404 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
12405 			if (orig != rack->r_ctl.rc_pace_max_segs)
12406 				chged = 1;
12407 		} else if ((rack->r_ctl.gp_bw == 0) &&
12408 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
12409 			/*
12410 			 * If we have nothing limit us to bursting
12411 			 * out IW sized pieces.
12412 			 */
12413 			chged = 1;
12414 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
12415 		}
12416 	}
12417 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
12418 		chged = 1;
12419 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
12420 	}
12421 	if (chged)
12422 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
12423 }
12424 
12425 
12426 static void
12427 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack)
12428 {
12429 #ifdef INET6
12430 	struct ip6_hdr *ip6 = NULL;
12431 #endif
12432 #ifdef INET
12433 	struct ip *ip = NULL;
12434 #endif
12435 	struct udphdr *udp = NULL;
12436 
12437 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
12438 #ifdef INET6
12439 	if (rack->r_is_v6) {
12440 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
12441 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
12442 		if (tp->t_port) {
12443 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12444 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
12445 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12446 			udp->uh_dport = tp->t_port;
12447 			rack->r_ctl.fsb.udp = udp;
12448 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12449 		} else
12450 		{
12451 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
12452 			rack->r_ctl.fsb.udp = NULL;
12453 		}
12454 		tcpip_fillheaders(rack->rc_inp,
12455 				  tp->t_port,
12456 				  ip6, rack->r_ctl.fsb.th);
12457 	} else
12458 #endif				/* INET6 */
12459 	{
12460 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
12461 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
12462 		if (tp->t_port) {
12463 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
12464 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
12465 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
12466 			udp->uh_dport = tp->t_port;
12467 			rack->r_ctl.fsb.udp = udp;
12468 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
12469 		} else
12470 		{
12471 			rack->r_ctl.fsb.udp = NULL;
12472 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
12473 		}
12474 		tcpip_fillheaders(rack->rc_inp,
12475 				  tp->t_port,
12476 				  ip, rack->r_ctl.fsb.th);
12477 	}
12478 	rack->r_fsb_inited = 1;
12479 }
12480 
12481 static int
12482 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
12483 {
12484 	/*
12485 	 * Allocate the larger of spaces V6 if available else just
12486 	 * V4 and include udphdr (overbook)
12487 	 */
12488 #ifdef INET6
12489 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
12490 #else
12491 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
12492 #endif
12493 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
12494 					    M_TCPFSB, M_NOWAIT|M_ZERO);
12495 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
12496 		return (ENOMEM);
12497 	}
12498 	rack->r_fsb_inited = 0;
12499 	return (0);
12500 }
12501 
12502 static int
12503 rack_init(struct tcpcb *tp)
12504 {
12505 	struct inpcb *inp = tptoinpcb(tp);
12506 	struct tcp_rack *rack = NULL;
12507 #ifdef INVARIANTS
12508 	struct rack_sendmap *insret;
12509 #endif
12510 	uint32_t iwin, snt, us_cts;
12511 	int err;
12512 
12513 	tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
12514 	if (tp->t_fb_ptr == NULL) {
12515 		/*
12516 		 * We need to allocate memory but cant. The INP and INP_INFO
12517 		 * locks and they are recursive (happens during setup. So a
12518 		 * scheme to drop the locks fails :(
12519 		 *
12520 		 */
12521 		return (ENOMEM);
12522 	}
12523 	memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
12524 
12525 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12526 	RB_INIT(&rack->r_ctl.rc_mtree);
12527 	TAILQ_INIT(&rack->r_ctl.rc_free);
12528 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
12529 	rack->rc_tp = tp;
12530 	rack->rc_inp = inp;
12531 	/* Set the flag */
12532 	rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
12533 	/* Probably not needed but lets be sure */
12534 	rack_clear_rate_sample(rack);
12535 	/*
12536 	 * Save off the default values, socket options will poke
12537 	 * at these if pacing is not on or we have not yet
12538 	 * reached where pacing is on (gp_ready/fixed enabled).
12539 	 * When they get set into the CC module (when gp_ready
12540 	 * is enabled or we enable fixed) then we will set these
12541 	 * values into the CC and place in here the old values
12542 	 * so we have a restoral. Then we will set the flag
12543 	 * rc_pacing_cc_set. That way whenever we turn off pacing
12544 	 * or switch off this stack, we will know to go restore
12545 	 * the saved values.
12546 	 */
12547 	rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
12548 	rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
12549 	/* We want abe like behavior as well */
12550 	rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
12551 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
12552 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
12553 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
12554 	rack->r_ctl.roundends = tp->snd_max;
12555 	if (use_rack_rr)
12556 		rack->use_rack_rr = 1;
12557 	if (V_tcp_delack_enabled)
12558 		tp->t_delayed_ack = 1;
12559 	else
12560 		tp->t_delayed_ack = 0;
12561 #ifdef TCP_ACCOUNTING
12562 	if (rack_tcp_accounting) {
12563 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
12564 	}
12565 #endif
12566 	if (rack_enable_shared_cwnd)
12567 		rack->rack_enable_scwnd = 1;
12568 	rack->rc_user_set_max_segs = rack_hptsi_segments;
12569 	rack->rc_force_max_seg = 0;
12570 	if (rack_use_imac_dack)
12571 		rack->rc_dack_mode = 1;
12572 	TAILQ_INIT(&rack->r_ctl.opt_list);
12573 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
12574 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
12575 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
12576 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
12577 	rack->r_ctl.rc_highest_us_rtt = 0;
12578 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
12579 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
12580 	if (rack_use_cmp_acks)
12581 		rack->r_use_cmp_ack = 1;
12582 	if (rack_disable_prr)
12583 		rack->rack_no_prr = 1;
12584 	if (rack_gp_no_rec_chg)
12585 		rack->rc_gp_no_rec_chg = 1;
12586 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
12587 		rack->rc_always_pace = 1;
12588 		if (rack->use_fixed_rate || rack->gp_ready)
12589 			rack_set_cc_pacing(rack);
12590 	} else
12591 		rack->rc_always_pace = 0;
12592 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
12593 		rack->r_mbuf_queue = 1;
12594 	else
12595 		rack->r_mbuf_queue = 0;
12596 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
12597 		inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
12598 	else
12599 		inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12600 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
12601 	if (rack_limits_scwnd)
12602 		rack->r_limit_scw = 1;
12603 	else
12604 		rack->r_limit_scw = 0;
12605 	rack->rc_labc = V_tcp_abc_l_var;
12606 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
12607 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
12608 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
12609 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
12610 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
12611 	rack->r_ctl.rc_min_to = rack_min_to;
12612 	microuptime(&rack->r_ctl.act_rcv_time);
12613 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
12614 	rack->rc_init_win = rack_default_init_window;
12615 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
12616 	if (rack_hw_up_only)
12617 		rack->r_up_only = 1;
12618 	if (rack_do_dyn_mul) {
12619 		/* When dynamic adjustment is on CA needs to start at 100% */
12620 		rack->rc_gp_dyn_mul = 1;
12621 		if (rack_do_dyn_mul >= 100)
12622 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
12623 	} else
12624 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
12625 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
12626 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
12627 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
12628 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
12629 				rack_probertt_filter_life);
12630 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12631 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
12632 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
12633 	rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
12634 	rack->r_ctl.rc_time_probertt_starts = 0;
12635 	if (rack_dsack_std_based & 0x1) {
12636 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
12637 		rack->rc_rack_tmr_std_based = 1;
12638 	}
12639 	if (rack_dsack_std_based & 0x2) {
12640 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
12641 		rack->rc_rack_use_dsack = 1;
12642 	}
12643 	/* We require at least one measurement, even if the sysctl is 0 */
12644 	if (rack_req_measurements)
12645 		rack->r_ctl.req_measurements = rack_req_measurements;
12646 	else
12647 		rack->r_ctl.req_measurements = 1;
12648 	if (rack_enable_hw_pacing)
12649 		rack->rack_hdw_pace_ena = 1;
12650 	if (rack_hw_rate_caps)
12651 		rack->r_rack_hw_rate_caps = 1;
12652 	/* Do we force on detection? */
12653 #ifdef NETFLIX_EXP_DETECTION
12654 	if (tcp_force_detection)
12655 		rack->do_detection = 1;
12656 	else
12657 #endif
12658 		rack->do_detection = 0;
12659 	if (rack_non_rxt_use_cr)
12660 		rack->rack_rec_nonrxt_use_cr = 1;
12661 	err = rack_init_fsb(tp, rack);
12662 	if (err) {
12663 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12664 		tp->t_fb_ptr = NULL;
12665 		return (err);
12666 	}
12667 	if (tp->snd_una != tp->snd_max) {
12668 		/* Create a send map for the current outstanding data */
12669 		struct rack_sendmap *rsm;
12670 
12671 		rsm = rack_alloc(rack);
12672 		if (rsm == NULL) {
12673 			uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12674 			tp->t_fb_ptr = NULL;
12675 			return (ENOMEM);
12676 		}
12677 		rsm->r_no_rtt_allowed = 1;
12678 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
12679 		rsm->r_rtr_cnt = 1;
12680 		rsm->r_rtr_bytes = 0;
12681 		if (tp->t_flags & TF_SENTFIN)
12682 			rsm->r_flags |= RACK_HAS_FIN;
12683 		if ((tp->snd_una == tp->iss) &&
12684 		    !TCPS_HAVEESTABLISHED(tp->t_state))
12685 			rsm->r_flags |= RACK_HAS_SYN;
12686 		rsm->r_start = tp->snd_una;
12687 		rsm->r_end = tp->snd_max;
12688 		rsm->r_dupack = 0;
12689 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
12690 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
12691 			if (rsm->m)
12692 				rsm->orig_m_len = rsm->m->m_len;
12693 			else
12694 				rsm->orig_m_len = 0;
12695 		} else {
12696 			/*
12697 			 * This can happen if we have a stand-alone FIN or
12698 			 *  SYN.
12699 			 */
12700 			rsm->m = NULL;
12701 			rsm->orig_m_len = 0;
12702 			rsm->soff = 0;
12703 		}
12704 #ifndef INVARIANTS
12705 		(void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12706 #else
12707 		insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12708 		if (insret != NULL) {
12709 			panic("Insert in rb tree fails ret:%p rack:%p rsm:%p",
12710 			      insret, rack, rsm);
12711 		}
12712 #endif
12713 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
12714 		rsm->r_in_tmap = 1;
12715 	}
12716 	/*
12717 	 * Timers in Rack are kept in microseconds so lets
12718 	 * convert any initial incoming variables
12719 	 * from ticks into usecs. Note that we
12720 	 * also change the values of t_srtt and t_rttvar, if
12721 	 * they are non-zero. They are kept with a 5
12722 	 * bit decimal so we have to carefully convert
12723 	 * these to get the full precision.
12724 	 */
12725 	rack_convert_rtts(tp);
12726 	tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow);
12727 	if (rack_do_hystart) {
12728 		tp->ccv->flags |= CCF_HYSTART_ALLOWED;
12729 		if (rack_do_hystart > 1)
12730 			tp->ccv->flags |= CCF_HYSTART_CAN_SH_CWND;
12731 		if (rack_do_hystart > 2)
12732 			tp->ccv->flags |= CCF_HYSTART_CONS_SSTH;
12733 	}
12734 	if (rack_def_profile)
12735 		rack_set_profile(rack, rack_def_profile);
12736 	/* Cancel the GP measurement in progress */
12737 	tp->t_flags &= ~TF_GPUTINPROG;
12738 	if (SEQ_GT(tp->snd_max, tp->iss))
12739 		snt = tp->snd_max - tp->iss;
12740 	else
12741 		snt = 0;
12742 	iwin = rc_init_window(rack);
12743 	if (snt < iwin) {
12744 		/* We are not past the initial window
12745 		 * so we need to make sure cwnd is
12746 		 * correct.
12747 		 */
12748 		if (tp->snd_cwnd < iwin)
12749 			tp->snd_cwnd = iwin;
12750 		/*
12751 		 * If we are within the initial window
12752 		 * we want ssthresh to be unlimited. Setting
12753 		 * it to the rwnd (which the default stack does
12754 		 * and older racks) is not really a good idea
12755 		 * since we want to be in SS and grow both the
12756 		 * cwnd and the rwnd (via dynamic rwnd growth). If
12757 		 * we set it to the rwnd then as the peer grows its
12758 		 * rwnd we will be stuck in CA and never hit SS.
12759 		 *
12760 		 * Its far better to raise it up high (this takes the
12761 		 * risk that there as been a loss already, probably
12762 		 * we should have an indicator in all stacks of loss
12763 		 * but we don't), but considering the normal use this
12764 		 * is a risk worth taking. The consequences of not
12765 		 * hitting SS are far worse than going one more time
12766 		 * into it early on (before we have sent even a IW).
12767 		 * It is highly unlikely that we will have had a loss
12768 		 * before getting the IW out.
12769 		 */
12770 		tp->snd_ssthresh = 0xffffffff;
12771 	}
12772 	rack_stop_all_timers(tp);
12773 	/* Lets setup the fsb block */
12774 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
12775 	rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
12776 			     __LINE__, RACK_RTTS_INIT);
12777 	return (0);
12778 }
12779 
12780 static int
12781 rack_handoff_ok(struct tcpcb *tp)
12782 {
12783 	if ((tp->t_state == TCPS_CLOSED) ||
12784 	    (tp->t_state == TCPS_LISTEN)) {
12785 		/* Sure no problem though it may not stick */
12786 		return (0);
12787 	}
12788 	if ((tp->t_state == TCPS_SYN_SENT) ||
12789 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
12790 		/*
12791 		 * We really don't know if you support sack,
12792 		 * you have to get to ESTAB or beyond to tell.
12793 		 */
12794 		return (EAGAIN);
12795 	}
12796 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
12797 		/*
12798 		 * Rack will only send a FIN after all data is acknowledged.
12799 		 * So in this case we have more data outstanding. We can't
12800 		 * switch stacks until either all data and only the FIN
12801 		 * is left (in which case rack_init() now knows how
12802 		 * to deal with that) <or> all is acknowledged and we
12803 		 * are only left with incoming data, though why you
12804 		 * would want to switch to rack after all data is acknowledged
12805 		 * I have no idea (rrs)!
12806 		 */
12807 		return (EAGAIN);
12808 	}
12809 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
12810 		return (0);
12811 	}
12812 	/*
12813 	 * If we reach here we don't do SACK on this connection so we can
12814 	 * never do rack.
12815 	 */
12816 	return (EINVAL);
12817 }
12818 
12819 
12820 static void
12821 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
12822 {
12823 	struct inpcb *inp = tptoinpcb(tp);
12824 
12825 	if (tp->t_fb_ptr) {
12826 		struct tcp_rack *rack;
12827 		struct rack_sendmap *rsm, *nrsm;
12828 #ifdef INVARIANTS
12829 		struct rack_sendmap *rm;
12830 #endif
12831 
12832 		rack = (struct tcp_rack *)tp->t_fb_ptr;
12833 		if (tp->t_in_pkt) {
12834 			/*
12835 			 * It is unsafe to process the packets since a
12836 			 * reset may be lurking in them (its rare but it
12837 			 * can occur). If we were to find a RST, then we
12838 			 * would end up dropping the connection and the
12839 			 * INP lock, so when we return the caller (tcp_usrreq)
12840 			 * will blow up when it trys to unlock the inp.
12841 			 */
12842 			struct mbuf *save, *m;
12843 
12844 			m = tp->t_in_pkt;
12845 			tp->t_in_pkt = NULL;
12846 			tp->t_tail_pkt = NULL;
12847 			while (m) {
12848 				save = m->m_nextpkt;
12849 				m->m_nextpkt = NULL;
12850 				m_freem(m);
12851 				m = save;
12852 			}
12853 		}
12854 		tp->t_flags &= ~TF_FORCEDATA;
12855 #ifdef NETFLIX_SHARED_CWND
12856 		if (rack->r_ctl.rc_scw) {
12857 			uint32_t limit;
12858 
12859 			if (rack->r_limit_scw)
12860 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
12861 			else
12862 				limit = 0;
12863 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
12864 						  rack->r_ctl.rc_scw_index,
12865 						  limit);
12866 			rack->r_ctl.rc_scw = NULL;
12867 		}
12868 #endif
12869 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
12870 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
12871 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
12872 			rack->r_ctl.fsb.th = NULL;
12873 		}
12874 		/* Convert back to ticks, with  */
12875 		if (tp->t_srtt > 1) {
12876 			uint32_t val, frac;
12877 
12878 			val = USEC_2_TICKS(tp->t_srtt);
12879 			frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12880 			tp->t_srtt = val << TCP_RTT_SHIFT;
12881 			/*
12882 			 * frac is the fractional part here is left
12883 			 * over from converting to hz and shifting.
12884 			 * We need to convert this to the 5 bit
12885 			 * remainder.
12886 			 */
12887 			if (frac) {
12888 				if (hz == 1000) {
12889 					frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12890 				} else {
12891 					frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12892 				}
12893 				tp->t_srtt += frac;
12894 			}
12895 		}
12896 		if (tp->t_rttvar) {
12897 			uint32_t val, frac;
12898 
12899 			val = USEC_2_TICKS(tp->t_rttvar);
12900 			frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz);
12901 			tp->t_rttvar = val <<  TCP_RTTVAR_SHIFT;
12902 			/*
12903 			 * frac is the fractional part here is left
12904 			 * over from converting to hz and shifting.
12905 			 * We need to convert this to the 5 bit
12906 			 * remainder.
12907 			 */
12908 			if (frac) {
12909 				if (hz == 1000) {
12910 					frac = (((uint64_t)frac *  (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC);
12911 				} else {
12912 					frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC);
12913 				}
12914 				tp->t_rttvar += frac;
12915 			}
12916 		}
12917 		tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur);
12918 		tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow);
12919 		if (rack->rc_always_pace) {
12920 			tcp_decrement_paced_conn();
12921 			rack_undo_cc_pacing(rack);
12922 			rack->rc_always_pace = 0;
12923 		}
12924 		/* Clean up any options if they were not applied */
12925 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
12926 			struct deferred_opt_list *dol;
12927 
12928 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
12929 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
12930 			free(dol, M_TCPDO);
12931 		}
12932 		/* rack does not use force data but other stacks may clear it */
12933 		if (rack->r_ctl.crte != NULL) {
12934 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
12935 			rack->rack_hdrw_pacing = 0;
12936 			rack->r_ctl.crte = NULL;
12937 		}
12938 #ifdef TCP_BLACKBOX
12939 		tcp_log_flowend(tp);
12940 #endif
12941 		RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) {
12942 #ifndef INVARIANTS
12943 			(void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12944 #else
12945 			rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm);
12946 			if (rm != rsm) {
12947 				panic("At fini, rack:%p rsm:%p rm:%p",
12948 				      rack, rsm, rm);
12949 			}
12950 #endif
12951 			uma_zfree(rack_zone, rsm);
12952 		}
12953 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12954 		while (rsm) {
12955 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
12956 			uma_zfree(rack_zone, rsm);
12957 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
12958 		}
12959 		rack->rc_free_cnt = 0;
12960 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
12961 		tp->t_fb_ptr = NULL;
12962 	}
12963 	inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
12964 	inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY;
12965 	inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE;
12966 	inp->inp_flags2 &= ~INP_MBUF_ACKCMP;
12967 	/* Cancel the GP measurement in progress */
12968 	tp->t_flags &= ~TF_GPUTINPROG;
12969 	inp->inp_flags2 &= ~INP_MBUF_L_ACKS;
12970 	/* Make sure snd_nxt is correctly set */
12971 	tp->snd_nxt = tp->snd_max;
12972 }
12973 
12974 static void
12975 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
12976 {
12977 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
12978 		rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
12979 	}
12980 	switch (tp->t_state) {
12981 	case TCPS_SYN_SENT:
12982 		rack->r_state = TCPS_SYN_SENT;
12983 		rack->r_substate = rack_do_syn_sent;
12984 		break;
12985 	case TCPS_SYN_RECEIVED:
12986 		rack->r_state = TCPS_SYN_RECEIVED;
12987 		rack->r_substate = rack_do_syn_recv;
12988 		break;
12989 	case TCPS_ESTABLISHED:
12990 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12991 		rack->r_state = TCPS_ESTABLISHED;
12992 		rack->r_substate = rack_do_established;
12993 		break;
12994 	case TCPS_CLOSE_WAIT:
12995 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
12996 		rack->r_state = TCPS_CLOSE_WAIT;
12997 		rack->r_substate = rack_do_close_wait;
12998 		break;
12999 	case TCPS_FIN_WAIT_1:
13000 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13001 		rack->r_state = TCPS_FIN_WAIT_1;
13002 		rack->r_substate = rack_do_fin_wait_1;
13003 		break;
13004 	case TCPS_CLOSING:
13005 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13006 		rack->r_state = TCPS_CLOSING;
13007 		rack->r_substate = rack_do_closing;
13008 		break;
13009 	case TCPS_LAST_ACK:
13010 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13011 		rack->r_state = TCPS_LAST_ACK;
13012 		rack->r_substate = rack_do_lastack;
13013 		break;
13014 	case TCPS_FIN_WAIT_2:
13015 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
13016 		rack->r_state = TCPS_FIN_WAIT_2;
13017 		rack->r_substate = rack_do_fin_wait_2;
13018 		break;
13019 	case TCPS_LISTEN:
13020 	case TCPS_CLOSED:
13021 	case TCPS_TIME_WAIT:
13022 	default:
13023 		break;
13024 	};
13025 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
13026 		rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
13027 
13028 }
13029 
13030 static void
13031 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
13032 {
13033 	/*
13034 	 * We received an ack, and then did not
13035 	 * call send or were bounced out due to the
13036 	 * hpts was running. Now a timer is up as well, is
13037 	 * it the right timer?
13038 	 */
13039 	struct rack_sendmap *rsm;
13040 	int tmr_up;
13041 
13042 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
13043 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
13044 		return;
13045 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
13046 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
13047 	    (tmr_up == PACE_TMR_RXT)) {
13048 		/* Should be an RXT */
13049 		return;
13050 	}
13051 	if (rsm == NULL) {
13052 		/* Nothing outstanding? */
13053 		if (tp->t_flags & TF_DELACK) {
13054 			if (tmr_up == PACE_TMR_DELACK)
13055 				/* We are supposed to have delayed ack up and we do */
13056 				return;
13057 		} else if (sbavail(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) {
13058 			/*
13059 			 * if we hit enobufs then we would expect the possibility
13060 			 * of nothing outstanding and the RXT up (and the hptsi timer).
13061 			 */
13062 			return;
13063 		} else if (((V_tcp_always_keepalive ||
13064 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
13065 			    (tp->t_state <= TCPS_CLOSING)) &&
13066 			   (tmr_up == PACE_TMR_KEEP) &&
13067 			   (tp->snd_max == tp->snd_una)) {
13068 			/* We should have keep alive up and we do */
13069 			return;
13070 		}
13071 	}
13072 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
13073 		   ((tmr_up == PACE_TMR_TLP) ||
13074 		    (tmr_up == PACE_TMR_RACK) ||
13075 		    (tmr_up == PACE_TMR_RXT))) {
13076 		/*
13077 		 * Either a Rack, TLP or RXT is fine if  we
13078 		 * have outstanding data.
13079 		 */
13080 		return;
13081 	} else if (tmr_up == PACE_TMR_DELACK) {
13082 		/*
13083 		 * If the delayed ack was going to go off
13084 		 * before the rtx/tlp/rack timer were going to
13085 		 * expire, then that would be the timer in control.
13086 		 * Note we don't check the time here trusting the
13087 		 * code is correct.
13088 		 */
13089 		return;
13090 	}
13091 	/*
13092 	 * Ok the timer originally started is not what we want now.
13093 	 * We will force the hpts to be stopped if any, and restart
13094 	 * with the slot set to what was in the saved slot.
13095 	 */
13096 	if (tcp_in_hpts(rack->rc_inp)) {
13097 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
13098 			uint32_t us_cts;
13099 
13100 			us_cts = tcp_get_usecs(NULL);
13101 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
13102 				rack->r_early = 1;
13103 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
13104 			}
13105 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
13106 		}
13107 		tcp_hpts_remove(rack->rc_inp);
13108 	}
13109 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13110 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
13111 }
13112 
13113 
13114 static void
13115 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)
13116 {
13117 	if ((SEQ_LT(tp->snd_wl1, seq) ||
13118 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
13119 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
13120 		/* keep track of pure window updates */
13121 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
13122 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
13123 		tp->snd_wnd = tiwin;
13124 		rack_validate_fo_sendwin_up(tp, rack);
13125 		tp->snd_wl1 = seq;
13126 		tp->snd_wl2 = ack;
13127 		if (tp->snd_wnd > tp->max_sndwnd)
13128 			tp->max_sndwnd = tp->snd_wnd;
13129 	    rack->r_wanted_output = 1;
13130 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
13131 		tp->snd_wnd = tiwin;
13132 		rack_validate_fo_sendwin_up(tp, rack);
13133 		tp->snd_wl1 = seq;
13134 		tp->snd_wl2 = ack;
13135 	} else {
13136 		/* Not a valid win update */
13137 		return;
13138 	}
13139 	/* Do we exit persists? */
13140 	if ((rack->rc_in_persist != 0) &&
13141 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13142 				rack->r_ctl.rc_pace_min_segs))) {
13143 		rack_exit_persist(tp, rack, cts);
13144 	}
13145 	/* Do we enter persists? */
13146 	if ((rack->rc_in_persist == 0) &&
13147 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13148 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
13149 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
13150 	    sbavail(&tptosocket(tp)->so_snd) &&
13151 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
13152 		/*
13153 		 * Here the rwnd is less than
13154 		 * the pacing size, we are established,
13155 		 * nothing is outstanding, and there is
13156 		 * data to send. Enter persists.
13157 		 */
13158 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
13159 	}
13160 }
13161 
13162 static void
13163 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
13164 {
13165 
13166 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
13167 		struct inpcb *inp = tptoinpcb(tp);
13168 		union tcp_log_stackspecific log;
13169 		struct timeval ltv;
13170 		char tcp_hdr_buf[60];
13171 		struct tcphdr *th;
13172 		struct timespec ts;
13173 		uint32_t orig_snd_una;
13174 		uint8_t xx = 0;
13175 
13176 #ifdef NETFLIX_HTTP_LOGGING
13177 		struct http_sendfile_track *http_req;
13178 
13179 		if (SEQ_GT(ae->ack, tp->snd_una)) {
13180 			http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1));
13181 		} else {
13182 			http_req = tcp_http_find_req_for_seq(tp, ae->ack);
13183 		}
13184 #endif
13185 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13186 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
13187 		if (rack->rack_no_prr == 0)
13188 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
13189 		else
13190 			log.u_bbr.flex1 = 0;
13191 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
13192 		log.u_bbr.use_lt_bw <<= 1;
13193 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
13194 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
13195 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
13196 		log.u_bbr.pkts_out = tp->t_maxseg;
13197 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
13198 		log.u_bbr.flex7 = 1;
13199 		log.u_bbr.lost = ae->flags;
13200 		log.u_bbr.cwnd_gain = ackval;
13201 		log.u_bbr.pacing_gain = 0x2;
13202 		if (ae->flags & TSTMP_HDWR) {
13203 			/* Record the hardware timestamp if present */
13204 			log.u_bbr.flex3 = M_TSTMP;
13205 			ts.tv_sec = ae->timestamp / 1000000000;
13206 			ts.tv_nsec = ae->timestamp % 1000000000;
13207 			ltv.tv_sec = ts.tv_sec;
13208 			ltv.tv_usec = ts.tv_nsec / 1000;
13209 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
13210 		} else if (ae->flags & TSTMP_LRO) {
13211 			/* Record the LRO the arrival timestamp */
13212 			log.u_bbr.flex3 = M_TSTMP_LRO;
13213 			ts.tv_sec = ae->timestamp / 1000000000;
13214 			ts.tv_nsec = ae->timestamp % 1000000000;
13215 			ltv.tv_sec = ts.tv_sec;
13216 			ltv.tv_usec = ts.tv_nsec / 1000;
13217 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
13218 		}
13219 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
13220 		/* Log the rcv time */
13221 		log.u_bbr.delRate = ae->timestamp;
13222 #ifdef NETFLIX_HTTP_LOGGING
13223 		log.u_bbr.applimited = tp->t_http_closed;
13224 		log.u_bbr.applimited <<= 8;
13225 		log.u_bbr.applimited |= tp->t_http_open;
13226 		log.u_bbr.applimited <<= 8;
13227 		log.u_bbr.applimited |= tp->t_http_req;
13228 		if (http_req) {
13229 			/* Copy out any client req info */
13230 			/* seconds */
13231 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
13232 			/* useconds */
13233 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
13234 			log.u_bbr.rttProp = http_req->timestamp;
13235 			log.u_bbr.cur_del_rate = http_req->start;
13236 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
13237 				log.u_bbr.flex8 |= 1;
13238 			} else {
13239 				log.u_bbr.flex8 |= 2;
13240 				log.u_bbr.bw_inuse = http_req->end;
13241 			}
13242 			log.u_bbr.flex6 = http_req->start_seq;
13243 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
13244 				log.u_bbr.flex8 |= 4;
13245 				log.u_bbr.epoch = http_req->end_seq;
13246 			}
13247 		}
13248 #endif
13249 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
13250 		th = (struct tcphdr *)tcp_hdr_buf;
13251 		th->th_seq = ae->seq;
13252 		th->th_ack = ae->ack;
13253 		th->th_win = ae->win;
13254 		/* Now fill in the ports */
13255 		th->th_sport = inp->inp_fport;
13256 		th->th_dport = inp->inp_lport;
13257 		tcp_set_flags(th, ae->flags);
13258 		/* Now do we have a timestamp option? */
13259 		if (ae->flags & HAS_TSTMP) {
13260 			u_char *cp;
13261 			uint32_t val;
13262 
13263 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
13264 			cp = (u_char *)(th + 1);
13265 			*cp = TCPOPT_NOP;
13266 			cp++;
13267 			*cp = TCPOPT_NOP;
13268 			cp++;
13269 			*cp = TCPOPT_TIMESTAMP;
13270 			cp++;
13271 			*cp = TCPOLEN_TIMESTAMP;
13272 			cp++;
13273 			val = htonl(ae->ts_value);
13274 			bcopy((char *)&val,
13275 			      (char *)cp, sizeof(uint32_t));
13276 			val = htonl(ae->ts_echo);
13277 			bcopy((char *)&val,
13278 			      (char *)(cp + 4), sizeof(uint32_t));
13279 		} else
13280 			th->th_off = (sizeof(struct tcphdr) >> 2);
13281 
13282 		/*
13283 		 * For sane logging we need to play a little trick.
13284 		 * If the ack were fully processed we would have moved
13285 		 * snd_una to high_seq, but since compressed acks are
13286 		 * processed in two phases, at this point (logging) snd_una
13287 		 * won't be advanced. So we would see multiple acks showing
13288 		 * the advancement. We can prevent that by "pretending" that
13289 		 * snd_una was advanced and then un-advancing it so that the
13290 		 * logging code has the right value for tlb_snd_una.
13291 		 */
13292 		if (tp->snd_una != high_seq) {
13293 			orig_snd_una = tp->snd_una;
13294 			tp->snd_una = high_seq;
13295 			xx = 1;
13296 		} else
13297 			xx = 0;
13298 		TCP_LOG_EVENTP(tp, th,
13299 			       &tptosocket(tp)->so_rcv,
13300 			       &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
13301 			       0, &log, true, &ltv);
13302 		if (xx) {
13303 			tp->snd_una = orig_snd_una;
13304 		}
13305 	}
13306 
13307 }
13308 
13309 static void
13310 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
13311 {
13312 	uint32_t us_rtt;
13313 	/*
13314 	 * A persist or keep-alive was forced out, update our
13315 	 * min rtt time. Note now worry about lost responses.
13316 	 * When a subsequent keep-alive or persist times out
13317 	 * and forced_ack is still on, then the last probe
13318 	 * was not responded to. In such cases we have a
13319 	 * sysctl that controls the behavior. Either we apply
13320 	 * the rtt but with reduced confidence (0). Or we just
13321 	 * plain don't apply the rtt estimate. Having data flow
13322 	 * will clear the probe_not_answered flag i.e. cum-ack
13323 	 * move forward <or> exiting and reentering persists.
13324 	 */
13325 
13326 	rack->forced_ack = 0;
13327 	rack->rc_tp->t_rxtshift = 0;
13328 	if ((rack->rc_in_persist &&
13329 	     (tiwin == rack->rc_tp->snd_wnd)) ||
13330 	    (rack->rc_in_persist == 0)) {
13331 		/*
13332 		 * In persists only apply the RTT update if this is
13333 		 * a response to our window probe. And that
13334 		 * means the rwnd sent must match the current
13335 		 * snd_wnd. If it does not, then we got a
13336 		 * window update ack instead. For keepalive
13337 		 * we allow the answer no matter what the window.
13338 		 *
13339 		 * Note that if the probe_not_answered is set then
13340 		 * the forced_ack_ts is the oldest one i.e. the first
13341 		 * probe sent that might have been lost. This assures
13342 		 * us that if we do calculate an RTT it is longer not
13343 		 * some short thing.
13344 		 */
13345 		if (rack->rc_in_persist)
13346 			counter_u64_add(rack_persists_acks, 1);
13347 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
13348 		if (us_rtt == 0)
13349 			us_rtt = 1;
13350 		if (rack->probe_not_answered == 0) {
13351 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13352 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
13353 		} else {
13354 			/* We have a retransmitted probe here too */
13355 			if (rack_apply_rtt_with_reduced_conf) {
13356 				rack_apply_updated_usrtt(rack, us_rtt, us_cts);
13357 				tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
13358 			}
13359 		}
13360 	}
13361 }
13362 
13363 static int
13364 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
13365 {
13366 	/*
13367 	 * Handle a "special" compressed ack mbuf. Each incoming
13368 	 * ack has only four possible dispositions:
13369 	 *
13370 	 * A) It moves the cum-ack forward
13371 	 * B) It is behind the cum-ack.
13372 	 * C) It is a window-update ack.
13373 	 * D) It is a dup-ack.
13374 	 *
13375 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
13376 	 * in the incoming mbuf. We also need to still pay attention
13377 	 * to nxt_pkt since there may be another packet after this
13378 	 * one.
13379 	 */
13380 #ifdef TCP_ACCOUNTING
13381 	uint64_t ts_val;
13382 	uint64_t rdstc;
13383 #endif
13384 	int segsiz;
13385 	struct timespec ts;
13386 	struct tcp_rack *rack;
13387 	struct tcp_ackent *ae;
13388 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
13389 	int cnt, i, did_out, ourfinisacked = 0;
13390 	struct tcpopt to_holder, *to = NULL;
13391 #ifdef TCP_ACCOUNTING
13392 	int win_up_req = 0;
13393 #endif
13394 	int nsegs = 0;
13395 	int under_pacing = 1;
13396 	int recovery = 0;
13397 #ifdef TCP_ACCOUNTING
13398 	sched_pin();
13399 #endif
13400 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13401 	if (rack->gp_ready &&
13402 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
13403 		under_pacing = 0;
13404 	else
13405 		under_pacing = 1;
13406 
13407 	if (rack->r_state != tp->t_state)
13408 		rack_set_state(tp, rack);
13409 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13410 	    (tp->t_flags & TF_GPUTINPROG)) {
13411 		/*
13412 		 * We have a goodput in progress
13413 		 * and we have entered a late state.
13414 		 * Do we have enough data in the sb
13415 		 * to handle the GPUT request?
13416 		 */
13417 		uint32_t bytes;
13418 
13419 		bytes = tp->gput_ack - tp->gput_seq;
13420 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
13421 			bytes += tp->gput_seq - tp->snd_una;
13422 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
13423 			/*
13424 			 * There are not enough bytes in the socket
13425 			 * buffer that have been sent to cover this
13426 			 * measurement. Cancel it.
13427 			 */
13428 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
13429 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
13430 						   tp->gput_seq,
13431 						   0, 0, 18, __LINE__, NULL, 0);
13432 			tp->t_flags &= ~TF_GPUTINPROG;
13433 		}
13434 	}
13435 	to = &to_holder;
13436 	to->to_flags = 0;
13437 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
13438 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
13439 	cnt = m->m_len / sizeof(struct tcp_ackent);
13440 	counter_u64_add(rack_multi_single_eq, cnt);
13441 	high_seq = tp->snd_una;
13442 	the_win = tp->snd_wnd;
13443 	win_seq = tp->snd_wl1;
13444 	win_upd_ack = tp->snd_wl2;
13445 	cts = tcp_tv_to_usectick(tv);
13446 	ms_cts = tcp_tv_to_mssectick(tv);
13447 	rack->r_ctl.rc_rcvtime = cts;
13448 	segsiz = ctf_fixed_maxseg(tp);
13449 	if ((rack->rc_gp_dyn_mul) &&
13450 	    (rack->use_fixed_rate == 0) &&
13451 	    (rack->rc_always_pace)) {
13452 		/* Check in on probertt */
13453 		rack_check_probe_rtt(rack, cts);
13454 	}
13455 	for (i = 0; i < cnt; i++) {
13456 #ifdef TCP_ACCOUNTING
13457 		ts_val = get_cyclecount();
13458 #endif
13459 		rack_clear_rate_sample(rack);
13460 		ae = ((mtod(m, struct tcp_ackent *)) + i);
13461 		/* Setup the window */
13462 		tiwin = ae->win << tp->snd_scale;
13463 		if (tiwin > rack->r_ctl.rc_high_rwnd)
13464 			rack->r_ctl.rc_high_rwnd = tiwin;
13465 		/* figure out the type of ack */
13466 		if (SEQ_LT(ae->ack, high_seq)) {
13467 			/* Case B*/
13468 			ae->ack_val_set = ACK_BEHIND;
13469 		} else if (SEQ_GT(ae->ack, high_seq)) {
13470 			/* Case A */
13471 			ae->ack_val_set = ACK_CUMACK;
13472 		} else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
13473 			/* Case D */
13474 			ae->ack_val_set = ACK_DUPACK;
13475 		} else {
13476 			/* Case C */
13477 			ae->ack_val_set = ACK_RWND;
13478 		}
13479 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
13480 		/* Validate timestamp */
13481 		if (ae->flags & HAS_TSTMP) {
13482 			/* Setup for a timestamp */
13483 			to->to_flags = TOF_TS;
13484 			ae->ts_echo -= tp->ts_offset;
13485 			to->to_tsecr = ae->ts_echo;
13486 			to->to_tsval = ae->ts_value;
13487 			/*
13488 			 * If echoed timestamp is later than the current time, fall back to
13489 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
13490 			 * were used when this connection was established.
13491 			 */
13492 			if (TSTMP_GT(ae->ts_echo, ms_cts))
13493 				to->to_tsecr = 0;
13494 			if (tp->ts_recent &&
13495 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
13496 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
13497 #ifdef TCP_ACCOUNTING
13498 					rdstc = get_cyclecount();
13499 					if (rdstc > ts_val) {
13500 						counter_u64_add(tcp_proc_time[ae->ack_val_set] ,
13501 								(rdstc - ts_val));
13502 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13503 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13504 						}
13505 					}
13506 #endif
13507 					continue;
13508 				}
13509 			}
13510 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
13511 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
13512 				tp->ts_recent_age = tcp_ts_getticks();
13513 				tp->ts_recent = ae->ts_value;
13514 			}
13515 		} else {
13516 			/* Setup for a no options */
13517 			to->to_flags = 0;
13518 		}
13519 		/* Update the rcv time and perform idle reduction possibly */
13520 		if  (tp->t_idle_reduce &&
13521 		     (tp->snd_max == tp->snd_una) &&
13522 		     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
13523 			counter_u64_add(rack_input_idle_reduces, 1);
13524 			rack_cc_after_idle(rack, tp);
13525 		}
13526 		tp->t_rcvtime = ticks;
13527 		/* Now what about ECN of a chain of pure ACKs? */
13528 		if (tcp_ecn_input_segment(tp, ae->flags, 0,
13529 			tcp_packets_this_ack(tp, ae->ack),
13530 			ae->codepoint))
13531 			rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
13532 #ifdef TCP_ACCOUNTING
13533 		/* Count for the specific type of ack in */
13534 		counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1);
13535 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13536 			tp->tcp_cnt_counters[ae->ack_val_set]++;
13537 		}
13538 #endif
13539 		/*
13540 		 * Note how we could move up these in the determination
13541 		 * above, but we don't so that way the timestamp checks (and ECN)
13542 		 * is done first before we do any processing on the ACK.
13543 		 * The non-compressed path through the code has this
13544 		 * weakness (noted by @jtl) that it actually does some
13545 		 * processing before verifying the timestamp information.
13546 		 * We don't take that path here which is why we set
13547 		 * the ack_val_set first, do the timestamp and ecn
13548 		 * processing, and then look at what we have setup.
13549 		 */
13550 		if (ae->ack_val_set == ACK_BEHIND) {
13551 			/*
13552 			 * Case B flag reordering, if window is not closed
13553 			 * or it could be a keep-alive or persists
13554 			 */
13555 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
13556 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
13557 			}
13558 		} else if (ae->ack_val_set == ACK_DUPACK) {
13559 			/* Case D */
13560 			rack_strike_dupack(rack);
13561 		} else if (ae->ack_val_set == ACK_RWND) {
13562 			/* Case C */
13563 			if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13564 				ts.tv_sec = ae->timestamp / 1000000000;
13565 				ts.tv_nsec = ae->timestamp % 1000000000;
13566 				rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13567 				rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13568 			} else {
13569 				rack->r_ctl.act_rcv_time = *tv;
13570 			}
13571 			if (rack->forced_ack) {
13572 				rack_handle_probe_response(rack, tiwin,
13573 							   tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
13574 			}
13575 #ifdef TCP_ACCOUNTING
13576 			win_up_req = 1;
13577 #endif
13578 			win_upd_ack = ae->ack;
13579 			win_seq = ae->seq;
13580 			the_win = tiwin;
13581 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13582 		} else {
13583 			/* Case A */
13584 			if (SEQ_GT(ae->ack, tp->snd_max)) {
13585 				/*
13586 				 * We just send an ack since the incoming
13587 				 * ack is beyond the largest seq we sent.
13588 				 */
13589 				if ((tp->t_flags & TF_ACKNOW) == 0) {
13590 					ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
13591 					if (tp->t_flags && TF_ACKNOW)
13592 						rack->r_wanted_output = 1;
13593 				}
13594 			} else {
13595 				nsegs++;
13596 				/* If the window changed setup to update */
13597 				if (tiwin != tp->snd_wnd) {
13598 					win_upd_ack = ae->ack;
13599 					win_seq = ae->seq;
13600 					the_win = tiwin;
13601 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq);
13602 				}
13603 #ifdef TCP_ACCOUNTING
13604 				/* Account for the acks */
13605 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13606 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
13607 				}
13608 				counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN],
13609 						(((ae->ack - high_seq) + segsiz - 1) / segsiz));
13610 #endif
13611 				high_seq = ae->ack;
13612 				if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
13613 					union tcp_log_stackspecific log;
13614 					struct timeval tv;
13615 
13616 					memset(&log.u_bbr, 0, sizeof(log.u_bbr));
13617 					log.u_bbr.timeStamp = tcp_get_usecs(&tv);
13618 					log.u_bbr.flex1 = high_seq;
13619 					log.u_bbr.flex2 = rack->r_ctl.roundends;
13620 					log.u_bbr.flex3 = rack->r_ctl.current_round;
13621 					log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
13622 					log.u_bbr.flex8 = 8;
13623 					tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
13624 						       0, &log, false, NULL, NULL, 0, &tv);
13625 				}
13626 				/*
13627 				 * The draft (v3) calls for us to use SEQ_GEQ, but that
13628 				 * causes issues when we are just going app limited. Lets
13629 				 * instead use SEQ_GT <or> where its equal but more data
13630 				 * is outstanding.
13631 				 */
13632 				if ((SEQ_GT(high_seq, rack->r_ctl.roundends)) ||
13633 				    ((high_seq == rack->r_ctl.roundends) &&
13634 				     SEQ_GT(tp->snd_max, tp->snd_una))) {
13635 					rack->r_ctl.current_round++;
13636 					rack->r_ctl.roundends = tp->snd_max;
13637 					if (CC_ALGO(tp)->newround != NULL) {
13638 						CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round);
13639 					}
13640 				}
13641 				/* Setup our act_rcv_time */
13642 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
13643 					ts.tv_sec = ae->timestamp / 1000000000;
13644 					ts.tv_nsec = ae->timestamp % 1000000000;
13645 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
13646 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
13647 				} else {
13648 					rack->r_ctl.act_rcv_time = *tv;
13649 				}
13650 				rack_process_to_cumack(tp, rack, ae->ack, cts, to);
13651 				if (rack->rc_dsack_round_seen) {
13652 					/* Is the dsack round over? */
13653 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
13654 						/* Yes it is */
13655 						rack->rc_dsack_round_seen = 0;
13656 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
13657 					}
13658 				}
13659 			}
13660 		}
13661 		/* And lets be sure to commit the rtt measurements for this ack */
13662 		tcp_rack_xmit_timer_commit(rack, tp);
13663 #ifdef TCP_ACCOUNTING
13664 		rdstc = get_cyclecount();
13665 		if (rdstc > ts_val) {
13666 			counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val));
13667 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13668 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
13669 				if (ae->ack_val_set == ACK_CUMACK)
13670 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
13671 			}
13672 		}
13673 #endif
13674 	}
13675 #ifdef TCP_ACCOUNTING
13676 	ts_val = get_cyclecount();
13677 #endif
13678 	/* Tend to any collapsed window */
13679 	if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
13680 		/* The peer collapsed the window */
13681 		rack_collapsed_window(rack, (tp->snd_max - high_seq), __LINE__);
13682 	} else if (rack->rc_has_collapsed)
13683 		rack_un_collapse_window(rack, __LINE__);
13684 	if ((rack->r_collapse_point_valid) &&
13685 	    (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
13686 		rack->r_collapse_point_valid = 0;
13687 	acked_amount = acked = (high_seq - tp->snd_una);
13688 	if (acked) {
13689 		/*
13690 		 * Clear the probe not answered flag
13691 		 * since cum-ack moved forward.
13692 		 */
13693 		rack->probe_not_answered = 0;
13694 		if (rack->sack_attack_disable == 0)
13695 			rack_do_decay(rack);
13696 		if (acked >= segsiz) {
13697 			/*
13698 			 * You only get credit for
13699 			 * MSS and greater (and you get extra
13700 			 * credit for larger cum-ack moves).
13701 			 */
13702 			int ac;
13703 
13704 			ac = acked / segsiz;
13705 			rack->r_ctl.ack_count += ac;
13706 			counter_u64_add(rack_ack_total, ac);
13707 		}
13708 		if (rack->r_ctl.ack_count > 0xfff00000) {
13709 			/*
13710 			 * reduce the number to keep us under
13711 			 * a uint32_t.
13712 			 */
13713 			rack->r_ctl.ack_count /= 2;
13714 			rack->r_ctl.sack_count /= 2;
13715 		}
13716 		if (tp->t_flags & TF_NEEDSYN) {
13717 			/*
13718 			 * T/TCP: Connection was half-synchronized, and our SYN has
13719 			 * been ACK'd (so connection is now fully synchronized).  Go
13720 			 * to non-starred state, increment snd_una for ACK of SYN,
13721 			 * and check if we can do window scaling.
13722 			 */
13723 			tp->t_flags &= ~TF_NEEDSYN;
13724 			tp->snd_una++;
13725 			acked_amount = acked = (high_seq - tp->snd_una);
13726 		}
13727 		if (acked > sbavail(&so->so_snd))
13728 			acked_amount = sbavail(&so->so_snd);
13729 #ifdef NETFLIX_EXP_DETECTION
13730 		/*
13731 		 * We only care on a cum-ack move if we are in a sack-disabled
13732 		 * state. We have already added in to the ack_count, and we never
13733 		 * would disable on a cum-ack move, so we only care to do the
13734 		 * detection if it may "undo" it, i.e. we were in disabled already.
13735 		 */
13736 		if (rack->sack_attack_disable)
13737 			rack_do_detection(tp, rack, acked_amount, segsiz);
13738 #endif
13739 		if (IN_FASTRECOVERY(tp->t_flags) &&
13740 		    (rack->rack_no_prr == 0))
13741 			rack_update_prr(tp, rack, acked_amount, high_seq);
13742 		if (IN_RECOVERY(tp->t_flags)) {
13743 			if (SEQ_LT(high_seq, tp->snd_recover) &&
13744 			    (SEQ_LT(high_seq, tp->snd_max))) {
13745 				tcp_rack_partialack(tp);
13746 			} else {
13747 				rack_post_recovery(tp, high_seq);
13748 				recovery = 1;
13749 			}
13750 		}
13751 		/* Handle the rack-log-ack part (sendmap) */
13752 		if ((sbused(&so->so_snd) == 0) &&
13753 		    (acked > acked_amount) &&
13754 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
13755 		    (tp->t_flags & TF_SENTFIN)) {
13756 			/*
13757 			 * We must be sure our fin
13758 			 * was sent and acked (we can be
13759 			 * in FIN_WAIT_1 without having
13760 			 * sent the fin).
13761 			 */
13762 			ourfinisacked = 1;
13763 			/*
13764 			 * Lets make sure snd_una is updated
13765 			 * since most likely acked_amount = 0 (it
13766 			 * should be).
13767 			 */
13768 			tp->snd_una = high_seq;
13769 		}
13770 		/* Did we make a RTO error? */
13771 		if ((tp->t_flags & TF_PREVVALID) &&
13772 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13773 			tp->t_flags &= ~TF_PREVVALID;
13774 			if (tp->t_rxtshift == 1 &&
13775 			    (int)(ticks - tp->t_badrxtwin) < 0)
13776 				rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
13777 		}
13778 		/* Handle the data in the socket buffer */
13779 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
13780 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13781 		if (acked_amount > 0) {
13782 			struct mbuf *mfree;
13783 
13784 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
13785 			SOCKBUF_LOCK(&so->so_snd);
13786 			mfree = sbcut_locked(&so->so_snd, acked_amount);
13787 			tp->snd_una = high_seq;
13788 			/* Note we want to hold the sb lock through the sendmap adjust */
13789 			rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una);
13790 			/* Wake up the socket if we have room to write more */
13791 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13792 			sowwakeup_locked(so);
13793 			m_freem(mfree);
13794 		}
13795 		/* update progress */
13796 		tp->t_acktime = ticks;
13797 		rack_log_progress_event(rack, tp, tp->t_acktime,
13798 					PROGRESS_UPDATE, __LINE__);
13799 		/* Clear out shifts and such */
13800 		tp->t_rxtshift = 0;
13801 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13802 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13803 		rack->rc_tlp_in_progress = 0;
13804 		rack->r_ctl.rc_tlp_cnt_out = 0;
13805 		/* Send recover and snd_nxt must be dragged along */
13806 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
13807 			tp->snd_recover = tp->snd_una;
13808 		if (SEQ_LT(tp->snd_nxt, tp->snd_una))
13809 			tp->snd_nxt = tp->snd_una;
13810 		/*
13811 		 * If the RXT timer is running we want to
13812 		 * stop it, so we can restart a TLP (or new RXT).
13813 		 */
13814 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13815 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13816 #ifdef NETFLIX_HTTP_LOGGING
13817 		tcp_http_check_for_comp(rack->rc_tp, high_seq);
13818 #endif
13819 		tp->snd_wl2 = high_seq;
13820 		tp->t_dupacks = 0;
13821 		if (under_pacing &&
13822 		    (rack->use_fixed_rate == 0) &&
13823 		    (rack->in_probe_rtt == 0) &&
13824 		    rack->rc_gp_dyn_mul &&
13825 		    rack->rc_always_pace) {
13826 			/* Check if we are dragging bottom */
13827 			rack_check_bottom_drag(tp, rack, so, acked);
13828 		}
13829 		if (tp->snd_una == tp->snd_max) {
13830 			tp->t_flags &= ~TF_PREVVALID;
13831 			rack->r_ctl.retran_during_recovery = 0;
13832 			rack->r_ctl.dsack_byte_cnt = 0;
13833 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13834 			if (rack->r_ctl.rc_went_idle_time == 0)
13835 				rack->r_ctl.rc_went_idle_time = 1;
13836 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13837 			if (sbavail(&tptosocket(tp)->so_snd) == 0)
13838 				tp->t_acktime = 0;
13839 			/* Set so we might enter persists... */
13840 			rack->r_wanted_output = 1;
13841 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13842 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
13843 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
13844 			    (sbavail(&so->so_snd) == 0) &&
13845 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
13846 				/*
13847 				 * The socket was gone and the
13848 				 * peer sent data (not now in the past), time to
13849 				 * reset him.
13850 				 */
13851 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13852 				/* tcp_close will kill the inp pre-log the Reset */
13853 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13854 #ifdef TCP_ACCOUNTING
13855 				rdstc = get_cyclecount();
13856 				if (rdstc > ts_val) {
13857 					counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13858 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13859 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13860 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13861 					}
13862 				}
13863 #endif
13864 				m_freem(m);
13865 				tp = tcp_close(tp);
13866 				if (tp == NULL) {
13867 #ifdef TCP_ACCOUNTING
13868 					sched_unpin();
13869 #endif
13870 					return (1);
13871 				}
13872 				/*
13873 				 * We would normally do drop-with-reset which would
13874 				 * send back a reset. We can't since we don't have
13875 				 * all the needed bits. Instead lets arrange for
13876 				 * a call to tcp_output(). That way since we
13877 				 * are in the closed state we will generate a reset.
13878 				 *
13879 				 * Note if tcp_accounting is on we don't unpin since
13880 				 * we do that after the goto label.
13881 				 */
13882 				goto send_out_a_rst;
13883 			}
13884 			if ((sbused(&so->so_snd) == 0) &&
13885 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
13886 			    (tp->t_flags & TF_SENTFIN)) {
13887 				/*
13888 				 * If we can't receive any more data, then closing user can
13889 				 * proceed. Starting the timer is contrary to the
13890 				 * specification, but if we don't get a FIN we'll hang
13891 				 * forever.
13892 				 *
13893 				 */
13894 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13895 					soisdisconnected(so);
13896 					tcp_timer_activate(tp, TT_2MSL,
13897 							   (tcp_fast_finwait2_recycle ?
13898 							    tcp_finwait2_timeout :
13899 							    TP_MAXIDLE(tp)));
13900 				}
13901 				if (ourfinisacked == 0) {
13902 					/*
13903 					 * We don't change to fin-wait-2 if we have our fin acked
13904 					 * which means we are probably in TCPS_CLOSING.
13905 					 */
13906 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
13907 				}
13908 			}
13909 		}
13910 		/* Wake up the socket if we have room to write more */
13911 		if (sbavail(&so->so_snd)) {
13912 			rack->r_wanted_output = 1;
13913 			if (ctf_progress_timeout_check(tp, true)) {
13914 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13915 							tp, tick, PROGRESS_DROP, __LINE__);
13916 				/*
13917 				 * We cheat here and don't send a RST, we should send one
13918 				 * when the pacer drops the connection.
13919 				 */
13920 #ifdef TCP_ACCOUNTING
13921 				rdstc = get_cyclecount();
13922 				if (rdstc > ts_val) {
13923 					counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
13924 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13925 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13926 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13927 					}
13928 				}
13929 				sched_unpin();
13930 #endif
13931 				(void)tcp_drop(tp, ETIMEDOUT);
13932 				m_freem(m);
13933 				return (1);
13934 			}
13935 		}
13936 		if (ourfinisacked) {
13937 			switch(tp->t_state) {
13938 			case TCPS_CLOSING:
13939 #ifdef TCP_ACCOUNTING
13940 				rdstc = get_cyclecount();
13941 				if (rdstc > ts_val) {
13942 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13943 							(rdstc - ts_val));
13944 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13945 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13946 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13947 					}
13948 				}
13949 				sched_unpin();
13950 #endif
13951 				tcp_twstart(tp);
13952 				m_freem(m);
13953 				return (1);
13954 				break;
13955 			case TCPS_LAST_ACK:
13956 #ifdef TCP_ACCOUNTING
13957 				rdstc = get_cyclecount();
13958 				if (rdstc > ts_val) {
13959 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13960 							(rdstc - ts_val));
13961 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13962 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13963 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13964 					}
13965 				}
13966 				sched_unpin();
13967 #endif
13968 				tp = tcp_close(tp);
13969 				ctf_do_drop(m, tp);
13970 				return (1);
13971 				break;
13972 			case TCPS_FIN_WAIT_1:
13973 #ifdef TCP_ACCOUNTING
13974 				rdstc = get_cyclecount();
13975 				if (rdstc > ts_val) {
13976 					counter_u64_add(tcp_proc_time[ACK_CUMACK] ,
13977 							(rdstc - ts_val));
13978 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
13979 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
13980 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
13981 					}
13982 				}
13983 #endif
13984 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13985 					soisdisconnected(so);
13986 					tcp_timer_activate(tp, TT_2MSL,
13987 							   (tcp_fast_finwait2_recycle ?
13988 							    tcp_finwait2_timeout :
13989 							    TP_MAXIDLE(tp)));
13990 				}
13991 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
13992 				break;
13993 			default:
13994 				break;
13995 			}
13996 		}
13997 		if (rack->r_fast_output) {
13998 			/*
13999 			 * We re doing fast output.. can we expand that?
14000 			 */
14001 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
14002 		}
14003 #ifdef TCP_ACCOUNTING
14004 		rdstc = get_cyclecount();
14005 		if (rdstc > ts_val) {
14006 			counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val));
14007 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14008 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
14009 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
14010 			}
14011 		}
14012 
14013 	} else if (win_up_req) {
14014 		rdstc = get_cyclecount();
14015 		if (rdstc > ts_val) {
14016 			counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val));
14017 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14018 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
14019 			}
14020 		}
14021 #endif
14022 	}
14023 	/* Now is there a next packet, if so we are done */
14024 	m_freem(m);
14025 	did_out = 0;
14026 	if (nxt_pkt) {
14027 #ifdef TCP_ACCOUNTING
14028 		sched_unpin();
14029 #endif
14030 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
14031 		return (0);
14032 	}
14033 	rack_handle_might_revert(tp, rack);
14034 	ctf_calc_rwin(so, tp);
14035 	if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14036 	send_out_a_rst:
14037 		if (tcp_output(tp) < 0) {
14038 #ifdef TCP_ACCOUNTING
14039 			sched_unpin();
14040 #endif
14041 			return (1);
14042 		}
14043 		did_out = 1;
14044 	}
14045 	rack_free_trim(rack);
14046 #ifdef TCP_ACCOUNTING
14047 	sched_unpin();
14048 #endif
14049 	rack_timer_audit(tp, rack, &so->so_snd);
14050 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
14051 	return (0);
14052 }
14053 
14054 
14055 static int
14056 rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so,
14057     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
14058     int32_t nxt_pkt, struct timeval *tv)
14059 {
14060 	struct inpcb *inp = tptoinpcb(tp);
14061 #ifdef TCP_ACCOUNTING
14062 	uint64_t ts_val;
14063 #endif
14064 	int32_t thflags, retval, did_out = 0;
14065 	int32_t way_out = 0;
14066 	/*
14067 	 * cts - is the current time from tv (caller gets ts) in microseconds.
14068 	 * ms_cts - is the current time from tv in milliseconds.
14069 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
14070 	 */
14071 	uint32_t cts, us_cts, ms_cts;
14072 	uint32_t tiwin, high_seq;
14073 	struct timespec ts;
14074 	struct tcpopt to;
14075 	struct tcp_rack *rack;
14076 	struct rack_sendmap *rsm;
14077 	int32_t prev_state = 0;
14078 #ifdef TCP_ACCOUNTING
14079 	int ack_val_set = 0xf;
14080 #endif
14081 	int nsegs;
14082 
14083 	NET_EPOCH_ASSERT();
14084 	INP_WLOCK_ASSERT(inp);
14085 
14086 	/*
14087 	 * tv passed from common code is from either M_TSTMP_LRO or
14088 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
14089 	 */
14090 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14091 	if (m->m_flags & M_ACKCMP) {
14092 		/*
14093 		 * All compressed ack's are ack's by definition so
14094 		 * remove any ack required flag and then do the processing.
14095 		 */
14096 		rack->rc_ack_required = 0;
14097 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
14098 	}
14099 	if (m->m_flags & M_ACKCMP) {
14100 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
14101 	}
14102 	cts = tcp_tv_to_usectick(tv);
14103 	ms_cts =  tcp_tv_to_mssectick(tv);
14104 	nsegs = m->m_pkthdr.lro_nsegs;
14105 	counter_u64_add(rack_proc_non_comp_ack, 1);
14106 	thflags = tcp_get_flags(th);
14107 #ifdef TCP_ACCOUNTING
14108 	sched_pin();
14109 	if (thflags & TH_ACK)
14110 		ts_val = get_cyclecount();
14111 #endif
14112 	if ((m->m_flags & M_TSTMP) ||
14113 	    (m->m_flags & M_TSTMP_LRO)) {
14114 		mbuf_tstmp2timespec(m, &ts);
14115 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
14116 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
14117 	} else
14118 		rack->r_ctl.act_rcv_time = *tv;
14119 	kern_prefetch(rack, &prev_state);
14120 	prev_state = 0;
14121 	/*
14122 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
14123 	 * the scale is zero.
14124 	 */
14125 	tiwin = th->th_win << tp->snd_scale;
14126 #ifdef TCP_ACCOUNTING
14127 	if (thflags & TH_ACK) {
14128 		/*
14129 		 * We have a tradeoff here. We can either do what we are
14130 		 * doing i.e. pinning to this CPU and then doing the accounting
14131 		 * <or> we could do a critical enter, setup the rdtsc and cpu
14132 		 * as in below, and then validate we are on the same CPU on
14133 		 * exit. I have choosen to not do the critical enter since
14134 		 * that often will gain you a context switch, and instead lock
14135 		 * us (line above this if) to the same CPU with sched_pin(). This
14136 		 * means we may be context switched out for a higher priority
14137 		 * interupt but we won't be moved to another CPU.
14138 		 *
14139 		 * If this occurs (which it won't very often since we most likely
14140 		 * are running this code in interupt context and only a higher
14141 		 * priority will bump us ... clock?) we will falsely add in
14142 		 * to the time the interupt processing time plus the ack processing
14143 		 * time. This is ok since its a rare event.
14144 		 */
14145 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
14146 						    ctf_fixed_maxseg(tp));
14147 	}
14148 #endif
14149 	/*
14150 	 * Parse options on any incoming segment.
14151 	 */
14152 	memset(&to, 0, sizeof(to));
14153 	tcp_dooptions(&to, (u_char *)(th + 1),
14154 	    (th->th_off << 2) - sizeof(struct tcphdr),
14155 	    (thflags & TH_SYN) ? TO_SYN : 0);
14156 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
14157 	    __func__));
14158 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
14159 	    __func__));
14160 
14161 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
14162 	    (tp->t_flags & TF_GPUTINPROG)) {
14163 		/*
14164 		 * We have a goodput in progress
14165 		 * and we have entered a late state.
14166 		 * Do we have enough data in the sb
14167 		 * to handle the GPUT request?
14168 		 */
14169 		uint32_t bytes;
14170 
14171 		bytes = tp->gput_ack - tp->gput_seq;
14172 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
14173 			bytes += tp->gput_seq - tp->snd_una;
14174 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
14175 			/*
14176 			 * There are not enough bytes in the socket
14177 			 * buffer that have been sent to cover this
14178 			 * measurement. Cancel it.
14179 			 */
14180 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
14181 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
14182 						   tp->gput_seq,
14183 						   0, 0, 18, __LINE__, NULL, 0);
14184 			tp->t_flags &= ~TF_GPUTINPROG;
14185 		}
14186 	}
14187 	high_seq = th->th_ack;
14188 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
14189 		union tcp_log_stackspecific log;
14190 		struct timeval ltv;
14191 #ifdef NETFLIX_HTTP_LOGGING
14192 		struct http_sendfile_track *http_req;
14193 
14194 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
14195 			http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1));
14196 		} else {
14197 			http_req = tcp_http_find_req_for_seq(tp, th->th_ack);
14198 		}
14199 #endif
14200 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14201 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
14202 		if (rack->rack_no_prr == 0)
14203 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
14204 		else
14205 			log.u_bbr.flex1 = 0;
14206 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
14207 		log.u_bbr.use_lt_bw <<= 1;
14208 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
14209 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
14210 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14211 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
14212 		log.u_bbr.flex3 = m->m_flags;
14213 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
14214 		log.u_bbr.lost = thflags;
14215 		log.u_bbr.pacing_gain = 0x1;
14216 #ifdef TCP_ACCOUNTING
14217 		log.u_bbr.cwnd_gain = ack_val_set;
14218 #endif
14219 		log.u_bbr.flex7 = 2;
14220 		if (m->m_flags & M_TSTMP) {
14221 			/* Record the hardware timestamp if present */
14222 			mbuf_tstmp2timespec(m, &ts);
14223 			ltv.tv_sec = ts.tv_sec;
14224 			ltv.tv_usec = ts.tv_nsec / 1000;
14225 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
14226 		} else if (m->m_flags & M_TSTMP_LRO) {
14227 			/* Record the LRO the arrival timestamp */
14228 			mbuf_tstmp2timespec(m, &ts);
14229 			ltv.tv_sec = ts.tv_sec;
14230 			ltv.tv_usec = ts.tv_nsec / 1000;
14231 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
14232 		}
14233 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
14234 		/* Log the rcv time */
14235 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
14236 #ifdef NETFLIX_HTTP_LOGGING
14237 		log.u_bbr.applimited = tp->t_http_closed;
14238 		log.u_bbr.applimited <<= 8;
14239 		log.u_bbr.applimited |= tp->t_http_open;
14240 		log.u_bbr.applimited <<= 8;
14241 		log.u_bbr.applimited |= tp->t_http_req;
14242 		if (http_req) {
14243 			/* Copy out any client req info */
14244 			/* seconds */
14245 			log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC);
14246 			/* useconds */
14247 			log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC);
14248 			log.u_bbr.rttProp = http_req->timestamp;
14249 			log.u_bbr.cur_del_rate = http_req->start;
14250 			if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) {
14251 				log.u_bbr.flex8 |= 1;
14252 			} else {
14253 				log.u_bbr.flex8 |= 2;
14254 				log.u_bbr.bw_inuse = http_req->end;
14255 			}
14256 			log.u_bbr.flex6 = http_req->start_seq;
14257 			if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) {
14258 				log.u_bbr.flex8 |= 4;
14259 				log.u_bbr.epoch = http_req->end_seq;
14260 			}
14261 		}
14262 #endif
14263 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
14264 		    tlen, &log, true, &ltv);
14265 	}
14266 	/* Remove ack required flag if set, we have one  */
14267 	if (thflags & TH_ACK)
14268 		rack->rc_ack_required = 0;
14269 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
14270 		way_out = 4;
14271 		retval = 0;
14272 		m_freem(m);
14273 		goto done_with_input;
14274 	}
14275 	/*
14276 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
14277 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
14278 	 */
14279 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
14280 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
14281 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
14282 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14283 #ifdef TCP_ACCOUNTING
14284 		sched_unpin();
14285 #endif
14286 		return (1);
14287 	}
14288 	/*
14289 	 * If timestamps were negotiated during SYN/ACK and a
14290 	 * segment without a timestamp is received, silently drop
14291 	 * the segment, unless it is a RST segment or missing timestamps are
14292 	 * tolerated.
14293 	 * See section 3.2 of RFC 7323.
14294 	 */
14295 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
14296 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
14297 		way_out = 5;
14298 		retval = 0;
14299 		m_freem(m);
14300 		goto done_with_input;
14301 	}
14302 
14303 	/*
14304 	 * Segment received on connection. Reset idle time and keep-alive
14305 	 * timer. XXX: This should be done after segment validation to
14306 	 * ignore broken/spoofed segs.
14307 	 */
14308 	if  (tp->t_idle_reduce &&
14309 	     (tp->snd_max == tp->snd_una) &&
14310 	     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
14311 		counter_u64_add(rack_input_idle_reduces, 1);
14312 		rack_cc_after_idle(rack, tp);
14313 	}
14314 	tp->t_rcvtime = ticks;
14315 #ifdef STATS
14316 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
14317 #endif
14318 	if (tiwin > rack->r_ctl.rc_high_rwnd)
14319 		rack->r_ctl.rc_high_rwnd = tiwin;
14320 	/*
14321 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
14322 	 * this to occur after we've validated the segment.
14323 	 */
14324 	if (tcp_ecn_input_segment(tp, thflags, tlen,
14325 	    tcp_packets_this_ack(tp, th->th_ack),
14326 	    iptos))
14327 		rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
14328 
14329 	/*
14330 	 * If echoed timestamp is later than the current time, fall back to
14331 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
14332 	 * were used when this connection was established.
14333 	 */
14334 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
14335 		to.to_tsecr -= tp->ts_offset;
14336 		if (TSTMP_GT(to.to_tsecr, ms_cts))
14337 			to.to_tsecr = 0;
14338 	}
14339 
14340 	/*
14341 	 * If its the first time in we need to take care of options and
14342 	 * verify we can do SACK for rack!
14343 	 */
14344 	if (rack->r_state == 0) {
14345 		/* Should be init'd by rack_init() */
14346 		KASSERT(rack->rc_inp != NULL,
14347 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
14348 		if (rack->rc_inp == NULL) {
14349 			rack->rc_inp = inp;
14350 		}
14351 
14352 		/*
14353 		 * Process options only when we get SYN/ACK back. The SYN
14354 		 * case for incoming connections is handled in tcp_syncache.
14355 		 * According to RFC1323 the window field in a SYN (i.e., a
14356 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
14357 		 * this is traditional behavior, may need to be cleaned up.
14358 		 */
14359 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
14360 			/* Handle parallel SYN for ECN */
14361 			tcp_ecn_input_parallel_syn(tp, thflags, iptos);
14362 			if ((to.to_flags & TOF_SCALE) &&
14363 			    (tp->t_flags & TF_REQ_SCALE)) {
14364 				tp->t_flags |= TF_RCVD_SCALE;
14365 				tp->snd_scale = to.to_wscale;
14366 			} else
14367 				tp->t_flags &= ~TF_REQ_SCALE;
14368 			/*
14369 			 * Initial send window.  It will be updated with the
14370 			 * next incoming segment to the scaled value.
14371 			 */
14372 			tp->snd_wnd = th->th_win;
14373 			rack_validate_fo_sendwin_up(tp, rack);
14374 			if ((to.to_flags & TOF_TS) &&
14375 			    (tp->t_flags & TF_REQ_TSTMP)) {
14376 				tp->t_flags |= TF_RCVD_TSTMP;
14377 				tp->ts_recent = to.to_tsval;
14378 				tp->ts_recent_age = cts;
14379 			} else
14380 				tp->t_flags &= ~TF_REQ_TSTMP;
14381 			if (to.to_flags & TOF_MSS) {
14382 				tcp_mss(tp, to.to_mss);
14383 			}
14384 			if ((tp->t_flags & TF_SACK_PERMIT) &&
14385 			    (to.to_flags & TOF_SACKPERM) == 0)
14386 				tp->t_flags &= ~TF_SACK_PERMIT;
14387 			if (IS_FASTOPEN(tp->t_flags)) {
14388 				if (to.to_flags & TOF_FASTOPEN) {
14389 					uint16_t mss;
14390 
14391 					if (to.to_flags & TOF_MSS)
14392 						mss = to.to_mss;
14393 					else
14394 						if ((inp->inp_vflag & INP_IPV6) != 0)
14395 							mss = TCP6_MSS;
14396 						else
14397 							mss = TCP_MSS;
14398 					tcp_fastopen_update_cache(tp, mss,
14399 					    to.to_tfo_len, to.to_tfo_cookie);
14400 				} else
14401 					tcp_fastopen_disable_path(tp);
14402 			}
14403 		}
14404 		/*
14405 		 * At this point we are at the initial call. Here we decide
14406 		 * if we are doing RACK or not. We do this by seeing if
14407 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
14408 		 * The code now does do dup-ack counting so if you don't
14409 		 * switch back you won't get rack & TLP, but you will still
14410 		 * get this stack.
14411 		 */
14412 
14413 		if ((rack_sack_not_required == 0) &&
14414 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
14415 			tcp_switch_back_to_default(tp);
14416 			(*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
14417 			    tlen, iptos);
14418 #ifdef TCP_ACCOUNTING
14419 			sched_unpin();
14420 #endif
14421 			return (1);
14422 		}
14423 		tcp_set_hpts(inp);
14424 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
14425 	}
14426 	if (thflags & TH_FIN)
14427 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
14428 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
14429 	if ((rack->rc_gp_dyn_mul) &&
14430 	    (rack->use_fixed_rate == 0) &&
14431 	    (rack->rc_always_pace)) {
14432 		/* Check in on probertt */
14433 		rack_check_probe_rtt(rack, us_cts);
14434 	}
14435 	rack_clear_rate_sample(rack);
14436 	if ((rack->forced_ack) &&
14437 	    ((tcp_get_flags(th) & TH_RST) == 0)) {
14438 		rack_handle_probe_response(rack, tiwin, us_cts);
14439 	}
14440 	/*
14441 	 * This is the one exception case where we set the rack state
14442 	 * always. All other times (timers etc) we must have a rack-state
14443 	 * set (so we assure we have done the checks above for SACK).
14444 	 */
14445 	rack->r_ctl.rc_rcvtime = cts;
14446 	if (rack->r_state != tp->t_state)
14447 		rack_set_state(tp, rack);
14448 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
14449 	    (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL)
14450 		kern_prefetch(rsm, &prev_state);
14451 	prev_state = rack->r_state;
14452 	retval = (*rack->r_substate) (m, th, so,
14453 	    tp, &to, drop_hdrlen,
14454 	    tlen, tiwin, thflags, nxt_pkt, iptos);
14455 	if (retval == 0) {
14456 		/*
14457 		 * If retval is 1 the tcb is unlocked and most likely the tp
14458 		 * is gone.
14459 		 */
14460 		INP_WLOCK_ASSERT(inp);
14461 		if ((rack->rc_gp_dyn_mul) &&
14462 		    (rack->rc_always_pace) &&
14463 		    (rack->use_fixed_rate == 0) &&
14464 		    rack->in_probe_rtt &&
14465 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
14466 			/*
14467 			 * If we are going for target, lets recheck before
14468 			 * we output.
14469 			 */
14470 			rack_check_probe_rtt(rack, us_cts);
14471 		}
14472 		if (rack->set_pacing_done_a_iw == 0) {
14473 			/* How much has been acked? */
14474 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
14475 				/* We have enough to set in the pacing segment size */
14476 				rack->set_pacing_done_a_iw = 1;
14477 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
14478 			}
14479 		}
14480 		tcp_rack_xmit_timer_commit(rack, tp);
14481 #ifdef TCP_ACCOUNTING
14482 		/*
14483 		 * If we set the ack_val_se to what ack processing we are doing
14484 		 * we also want to track how many cycles we burned. Note
14485 		 * the bits after tcp_output we let be "free". This is because
14486 		 * we are also tracking the tcp_output times as well. Note the
14487 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
14488 		 * 0xf cannot be returned and is what we initialize it too to
14489 		 * indicate we are not doing the tabulations.
14490 		 */
14491 		if (ack_val_set != 0xf) {
14492 			uint64_t crtsc;
14493 
14494 			crtsc = get_cyclecount();
14495 			counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14496 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
14497 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
14498 			}
14499 		}
14500 #endif
14501 		if (nxt_pkt == 0) {
14502 			if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
14503 do_output_now:
14504 				if (tcp_output(tp) < 0)
14505 					return (1);
14506 				did_out = 1;
14507 			}
14508 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
14509 			rack_free_trim(rack);
14510 		}
14511 		/* Update any rounds needed */
14512 		if (rack_verbose_logging &&  (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) {
14513 			union tcp_log_stackspecific log;
14514 			struct timeval tv;
14515 
14516 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14517 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14518 			log.u_bbr.flex1 = high_seq;
14519 			log.u_bbr.flex2 = rack->r_ctl.roundends;
14520 			log.u_bbr.flex3 = rack->r_ctl.current_round;
14521 			log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround;
14522 			log.u_bbr.flex8 = 9;
14523 			tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
14524 				       0, &log, false, NULL, NULL, 0, &tv);
14525 		}
14526 		/*
14527 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
14528 		 * causes issues when we are just going app limited. Lets
14529 		 * instead use SEQ_GT <or> where its equal but more data
14530 		 * is outstanding.
14531 		 */
14532 		if ((SEQ_GT(tp->snd_una, rack->r_ctl.roundends)) ||
14533 		    ((tp->snd_una == rack->r_ctl.roundends) && SEQ_GT(tp->snd_max, tp->snd_una))) {
14534 			rack->r_ctl.current_round++;
14535 			rack->r_ctl.roundends = tp->snd_max;
14536 			if (CC_ALGO(tp)->newround != NULL) {
14537 				CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round);
14538 			}
14539 		}
14540 		if ((nxt_pkt == 0) &&
14541 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
14542 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
14543 		     (tp->t_flags & TF_DELACK) ||
14544 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
14545 		      (tp->t_state <= TCPS_CLOSING)))) {
14546 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
14547 			if ((tp->snd_max == tp->snd_una) &&
14548 			    ((tp->t_flags & TF_DELACK) == 0) &&
14549 			    (tcp_in_hpts(rack->rc_inp)) &&
14550 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
14551 				/* keep alive not needed if we are hptsi output yet */
14552 				;
14553 			} else {
14554 				int late = 0;
14555 				if (tcp_in_hpts(inp)) {
14556 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14557 						us_cts = tcp_get_usecs(NULL);
14558 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
14559 							rack->r_early = 1;
14560 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
14561 						} else
14562 							late = 1;
14563 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
14564 					}
14565 					tcp_hpts_remove(inp);
14566 				}
14567 				if (late && (did_out == 0)) {
14568 					/*
14569 					 * We are late in the sending
14570 					 * and we did not call the output
14571 					 * (this probably should not happen).
14572 					 */
14573 					goto do_output_now;
14574 				}
14575 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
14576 			}
14577 			way_out = 1;
14578 		} else if (nxt_pkt == 0) {
14579 			/* Do we have the correct timer running? */
14580 			rack_timer_audit(tp, rack, &so->so_snd);
14581 			way_out = 2;
14582 		}
14583 	done_with_input:
14584 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
14585 		if (did_out)
14586 			rack->r_wanted_output = 0;
14587 #ifdef TCP_ACCOUNTING
14588 	} else {
14589 		/*
14590 		 * Track the time (see above).
14591 		 */
14592 		if (ack_val_set != 0xf) {
14593 			uint64_t crtsc;
14594 
14595 			crtsc = get_cyclecount();
14596 			counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val));
14597 			/*
14598 			 * Note we *DO NOT* increment the per-tcb counters since
14599 			 * in the else the TP may be gone!!
14600 			 */
14601 		}
14602 #endif
14603 	}
14604 #ifdef TCP_ACCOUNTING
14605 	sched_unpin();
14606 #endif
14607 	return (retval);
14608 }
14609 
14610 void
14611 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
14612     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
14613 {
14614 	struct timeval tv;
14615 
14616 	/* First lets see if we have old packets */
14617 	if (tp->t_in_pkt) {
14618 		if (ctf_do_queued_segments(so, tp, 1)) {
14619 			m_freem(m);
14620 			return;
14621 		}
14622 	}
14623 	if (m->m_flags & M_TSTMP_LRO) {
14624 		mbuf_tstmp2timeval(m, &tv);
14625 	} else {
14626 		/* Should not be should we kassert instead? */
14627 		tcp_get_usecs(&tv);
14628 	}
14629 	if (rack_do_segment_nounlock(m, th, so, tp,
14630 				     drop_hdrlen, tlen, iptos, 0, &tv) == 0) {
14631 		INP_WUNLOCK(tptoinpcb(tp));
14632 	}
14633 }
14634 
14635 struct rack_sendmap *
14636 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
14637 {
14638 	struct rack_sendmap *rsm = NULL;
14639 	int32_t idx;
14640 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
14641 
14642 	/* Return the next guy to be re-transmitted */
14643 	if (RB_EMPTY(&rack->r_ctl.rc_mtree)) {
14644 		return (NULL);
14645 	}
14646 	if (tp->t_flags & TF_SENTFIN) {
14647 		/* retran the end FIN? */
14648 		return (NULL);
14649 	}
14650 	/* ok lets look at this one */
14651 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
14652 	if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
14653 		return (rsm);
14654 	}
14655 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
14656 		goto check_it;
14657 	}
14658 	rsm = rack_find_lowest_rsm(rack);
14659 	if (rsm == NULL) {
14660 		return (NULL);
14661 	}
14662 check_it:
14663 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) &&
14664 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
14665 		/*
14666 		 * No sack so we automatically do the 3 strikes and
14667 		 * retransmit (no rack timer would be started).
14668 		 */
14669 
14670 		return (rsm);
14671 	}
14672 	if (rsm->r_flags & RACK_ACKED) {
14673 		return (NULL);
14674 	}
14675 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
14676 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
14677 		/* Its not yet ready */
14678 		return (NULL);
14679 	}
14680 	srtt = rack_grab_rtt(tp, rack);
14681 	idx = rsm->r_rtr_cnt - 1;
14682 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
14683 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
14684 	if ((tsused == ts_low) ||
14685 	    (TSTMP_LT(tsused, ts_low))) {
14686 		/* No time since sending */
14687 		return (NULL);
14688 	}
14689 	if ((tsused - ts_low) < thresh) {
14690 		/* It has not been long enough yet */
14691 		return (NULL);
14692 	}
14693 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
14694 	    ((rsm->r_flags & RACK_SACK_PASSED) &&
14695 	     (rack->sack_attack_disable == 0))) {
14696 		/*
14697 		 * We have passed the dup-ack threshold <or>
14698 		 * a SACK has indicated this is missing.
14699 		 * Note that if you are a declared attacker
14700 		 * it is only the dup-ack threshold that
14701 		 * will cause retransmits.
14702 		 */
14703 		/* log retransmit reason */
14704 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
14705 		rack->r_fast_output = 0;
14706 		return (rsm);
14707 	}
14708 	return (NULL);
14709 }
14710 
14711 static void
14712 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
14713 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
14714 			   int line, struct rack_sendmap *rsm, uint8_t quality)
14715 {
14716 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
14717 		union tcp_log_stackspecific log;
14718 		struct timeval tv;
14719 
14720 		memset(&log, 0, sizeof(log));
14721 		log.u_bbr.flex1 = slot;
14722 		log.u_bbr.flex2 = len;
14723 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
14724 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
14725 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
14726 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
14727 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
14728 		log.u_bbr.use_lt_bw <<= 1;
14729 		log.u_bbr.use_lt_bw |= rack->r_late;
14730 		log.u_bbr.use_lt_bw <<= 1;
14731 		log.u_bbr.use_lt_bw |= rack->r_early;
14732 		log.u_bbr.use_lt_bw <<= 1;
14733 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
14734 		log.u_bbr.use_lt_bw <<= 1;
14735 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
14736 		log.u_bbr.use_lt_bw <<= 1;
14737 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
14738 		log.u_bbr.use_lt_bw <<= 1;
14739 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
14740 		log.u_bbr.use_lt_bw <<= 1;
14741 		log.u_bbr.use_lt_bw |= rack->gp_ready;
14742 		log.u_bbr.pkt_epoch = line;
14743 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
14744 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
14745 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
14746 		log.u_bbr.bw_inuse = bw_est;
14747 		log.u_bbr.delRate = bw;
14748 		if (rack->r_ctl.gp_bw == 0)
14749 			log.u_bbr.cur_del_rate = 0;
14750 		else
14751 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
14752 		log.u_bbr.rttProp = len_time;
14753 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
14754 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
14755 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
14756 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
14757 			/* We are in slow start */
14758 			log.u_bbr.flex7 = 1;
14759 		} else {
14760 			/* we are on congestion avoidance */
14761 			log.u_bbr.flex7 = 0;
14762 		}
14763 		log.u_bbr.flex8 = method;
14764 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14765 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14766 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
14767 		log.u_bbr.cwnd_gain <<= 1;
14768 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
14769 		log.u_bbr.cwnd_gain <<= 1;
14770 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
14771 		log.u_bbr.bbr_substate = quality;
14772 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
14773 		    &rack->rc_inp->inp_socket->so_rcv,
14774 		    &rack->rc_inp->inp_socket->so_snd,
14775 		    BBR_LOG_HPTSI_CALC, 0,
14776 		    0, &log, false, &tv);
14777 	}
14778 }
14779 
14780 static uint32_t
14781 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
14782 {
14783 	uint32_t new_tso, user_max;
14784 
14785 	user_max = rack->rc_user_set_max_segs * mss;
14786 	if (rack->rc_force_max_seg) {
14787 		return (user_max);
14788 	}
14789 	if (rack->use_fixed_rate &&
14790 	    ((rack->r_ctl.crte == NULL) ||
14791 	     (bw != rack->r_ctl.crte->rate))) {
14792 		/* Use the user mss since we are not exactly matched */
14793 		return (user_max);
14794 	}
14795 	new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL);
14796 	if (new_tso > user_max)
14797 		new_tso = user_max;
14798 	return (new_tso);
14799 }
14800 
14801 static int32_t
14802 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)
14803 {
14804 	uint64_t lentim, fill_bw;
14805 
14806 	/* Lets first see if we are full, if so continue with normal rate */
14807 	rack->r_via_fill_cw = 0;
14808 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
14809 		return (slot);
14810 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
14811 		return (slot);
14812 	if (rack->r_ctl.rc_last_us_rtt == 0)
14813 		return (slot);
14814 	if (rack->rc_pace_fill_if_rttin_range &&
14815 	    (rack->r_ctl.rc_last_us_rtt >=
14816 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
14817 		/* The rtt is huge, N * smallest, lets not fill */
14818 		return (slot);
14819 	}
14820 	/*
14821 	 * first lets calculate the b/w based on the last us-rtt
14822 	 * and the sndwnd.
14823 	 */
14824 	fill_bw = rack->r_ctl.cwnd_to_use;
14825 	/* Take the rwnd if its smaller */
14826 	if (fill_bw > rack->rc_tp->snd_wnd)
14827 		fill_bw = rack->rc_tp->snd_wnd;
14828 	if (rack->r_fill_less_agg) {
14829 		/*
14830 		 * Now take away the inflight (this will reduce our
14831 		 * aggressiveness and yeah, if we get that much out in 1RTT
14832 		 * we will have had acks come back and still be behind).
14833 		 */
14834 		fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
14835 	}
14836 	/* Now lets make it into a b/w */
14837 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
14838 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
14839 	/* We are below the min b/w */
14840 	if (non_paced)
14841 		*rate_wanted = fill_bw;
14842 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
14843 		return (slot);
14844 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap))
14845 		fill_bw = rack->r_ctl.bw_rate_cap;
14846 	rack->r_via_fill_cw = 1;
14847 	if (rack->r_rack_hw_rate_caps &&
14848 	    (rack->r_ctl.crte != NULL)) {
14849 		uint64_t high_rate;
14850 
14851 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
14852 		if (fill_bw > high_rate) {
14853 			/* We are capping bw at the highest rate table entry */
14854 			if (*rate_wanted > high_rate) {
14855 				/* The original rate was also capped */
14856 				rack->r_via_fill_cw = 0;
14857 			}
14858 			rack_log_hdwr_pacing(rack,
14859 					     fill_bw, high_rate, __LINE__,
14860 					     0, 3);
14861 			fill_bw = high_rate;
14862 			if (capped)
14863 				*capped = 1;
14864 		}
14865 	} else if ((rack->r_ctl.crte == NULL) &&
14866 		   (rack->rack_hdrw_pacing == 0) &&
14867 		   (rack->rack_hdw_pace_ena) &&
14868 		   rack->r_rack_hw_rate_caps &&
14869 		   (rack->rack_attempt_hdwr_pace == 0) &&
14870 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
14871 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
14872 		/*
14873 		 * Ok we may have a first attempt that is greater than our top rate
14874 		 * lets check.
14875 		 */
14876 		uint64_t high_rate;
14877 
14878 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
14879 		if (high_rate) {
14880 			if (fill_bw > high_rate) {
14881 				fill_bw = high_rate;
14882 				if (capped)
14883 					*capped = 1;
14884 			}
14885 		}
14886 	}
14887 	/*
14888 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
14889 	 * in a rtt, what does that time wise equate too?
14890 	 */
14891 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
14892 	lentim /= fill_bw;
14893 	*rate_wanted = fill_bw;
14894 	if (non_paced || (lentim < slot)) {
14895 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
14896 					   0, lentim, 12, __LINE__, NULL, 0);
14897 		return ((int32_t)lentim);
14898 	} else
14899 		return (slot);
14900 }
14901 
14902 static int32_t
14903 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
14904 {
14905 	uint64_t srtt;
14906 	int32_t slot = 0;
14907 	int can_start_hw_pacing = 1;
14908 	int err;
14909 
14910 	if (rack->rc_always_pace == 0) {
14911 		/*
14912 		 * We use the most optimistic possible cwnd/srtt for
14913 		 * sending calculations. This will make our
14914 		 * calculation anticipate getting more through
14915 		 * quicker then possible. But thats ok we don't want
14916 		 * the peer to have a gap in data sending.
14917 		 */
14918 		uint64_t cwnd, tr_perms = 0;
14919 		int32_t reduce = 0;
14920 
14921 	old_method:
14922 		/*
14923 		 * We keep no precise pacing with the old method
14924 		 * instead we use the pacer to mitigate bursts.
14925 		 */
14926 		if (rack->r_ctl.rc_rack_min_rtt)
14927 			srtt = rack->r_ctl.rc_rack_min_rtt;
14928 		else
14929 			srtt = max(tp->t_srtt, 1);
14930 		if (rack->r_ctl.rc_rack_largest_cwnd)
14931 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
14932 		else
14933 			cwnd = rack->r_ctl.cwnd_to_use;
14934 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
14935 		tr_perms = (cwnd * 1000) / srtt;
14936 		if (tr_perms == 0) {
14937 			tr_perms = ctf_fixed_maxseg(tp);
14938 		}
14939 		/*
14940 		 * Calculate how long this will take to drain, if
14941 		 * the calculation comes out to zero, thats ok we
14942 		 * will use send_a_lot to possibly spin around for
14943 		 * more increasing tot_len_this_send to the point
14944 		 * that its going to require a pace, or we hit the
14945 		 * cwnd. Which in that case we are just waiting for
14946 		 * a ACK.
14947 		 */
14948 		slot = len / tr_perms;
14949 		/* Now do we reduce the time so we don't run dry? */
14950 		if (slot && rack_slot_reduction) {
14951 			reduce = (slot / rack_slot_reduction);
14952 			if (reduce < slot) {
14953 				slot -= reduce;
14954 			} else
14955 				slot = 0;
14956 		}
14957 		slot *= HPTS_USEC_IN_MSEC;
14958 		if (rack->rc_pace_to_cwnd) {
14959 			uint64_t rate_wanted = 0;
14960 
14961 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
14962 			rack->rc_ack_can_sendout_data = 1;
14963 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
14964 		} else
14965 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
14966 	} else {
14967 		uint64_t bw_est, res, lentim, rate_wanted;
14968 		uint32_t orig_val, segs, oh;
14969 		int capped = 0;
14970 		int prev_fill;
14971 
14972 		if ((rack->r_rr_config == 1) && rsm) {
14973 			return (rack->r_ctl.rc_min_to);
14974 		}
14975 		if (rack->use_fixed_rate) {
14976 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
14977 		} else if ((rack->r_ctl.init_rate == 0) &&
14978 #ifdef NETFLIX_PEAKRATE
14979 			   (rack->rc_tp->t_maxpeakrate == 0) &&
14980 #endif
14981 			   (rack->r_ctl.gp_bw == 0)) {
14982 			/* no way to yet do an estimate */
14983 			bw_est = rate_wanted = 0;
14984 		} else {
14985 			bw_est = rack_get_bw(rack);
14986 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
14987 		}
14988 		if ((bw_est == 0) || (rate_wanted == 0) ||
14989 		    ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
14990 			/*
14991 			 * No way yet to make a b/w estimate or
14992 			 * our raise is set incorrectly.
14993 			 */
14994 			goto old_method;
14995 		}
14996 		/* We need to account for all the overheads */
14997 		segs = (len + segsiz - 1) / segsiz;
14998 		/*
14999 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
15000 		 * and how much data we put in each packet. Yes this
15001 		 * means we may be off if we are larger than 1500 bytes
15002 		 * or smaller. But this just makes us more conservative.
15003 		 */
15004 		if (rack_hw_rate_min &&
15005 		    (bw_est < rack_hw_rate_min))
15006 			can_start_hw_pacing = 0;
15007 		if (ETHERNET_SEGMENT_SIZE > segsiz)
15008 			oh = ETHERNET_SEGMENT_SIZE - segsiz;
15009 		else
15010 			oh = 0;
15011 		segs *= oh;
15012 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
15013 		res = lentim / rate_wanted;
15014 		slot = (uint32_t)res;
15015 		orig_val = rack->r_ctl.rc_pace_max_segs;
15016 		if (rack->r_ctl.crte == NULL) {
15017 			/*
15018 			 * Only do this if we are not hardware pacing
15019 			 * since if we are doing hw-pacing below we will
15020 			 * set make a call after setting up or changing
15021 			 * the rate.
15022 			 */
15023 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
15024 		} else if (rack->rc_inp->inp_snd_tag == NULL) {
15025 			/*
15026 			 * We lost our rate somehow, this can happen
15027 			 * if the interface changed underneath us.
15028 			 */
15029 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
15030 			rack->r_ctl.crte = NULL;
15031 			/* Lets re-allow attempting to setup pacing */
15032 			rack->rack_hdrw_pacing = 0;
15033 			rack->rack_attempt_hdwr_pace = 0;
15034 			rack_log_hdwr_pacing(rack,
15035 					     rate_wanted, bw_est, __LINE__,
15036 					     0, 6);
15037 		}
15038 		/* Did we change the TSO size, if so log it */
15039 		if (rack->r_ctl.rc_pace_max_segs != orig_val)
15040 			rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL, 0);
15041 		prev_fill = rack->r_via_fill_cw;
15042 		if ((rack->rc_pace_to_cwnd) &&
15043 		    (capped == 0) &&
15044 		    (rack->use_fixed_rate == 0) &&
15045 		    (rack->in_probe_rtt == 0) &&
15046 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
15047 			/*
15048 			 * We want to pace at our rate *or* faster to
15049 			 * fill the cwnd to the max if its not full.
15050 			 */
15051 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
15052 		}
15053 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
15054 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
15055 			if ((rack->rack_hdw_pace_ena) &&
15056 			    (can_start_hw_pacing > 0) &&
15057 			    (rack->rack_hdrw_pacing == 0) &&
15058 			    (rack->rack_attempt_hdwr_pace == 0)) {
15059 				/*
15060 				 * Lets attempt to turn on hardware pacing
15061 				 * if we can.
15062 				 */
15063 				rack->rack_attempt_hdwr_pace = 1;
15064 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
15065 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
15066 								       rate_wanted,
15067 								       RS_PACING_GEQ,
15068 								       &err, &rack->r_ctl.crte_prev_rate);
15069 				if (rack->r_ctl.crte) {
15070 					rack->rack_hdrw_pacing = 1;
15071 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz,
15072 												 0, rack->r_ctl.crte,
15073 												 NULL);
15074 					rack_log_hdwr_pacing(rack,
15075 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15076 							     err, 0);
15077 					rack->r_ctl.last_hw_bw_req = rate_wanted;
15078 				} else {
15079 					counter_u64_add(rack_hw_pace_init_fail, 1);
15080 				}
15081 			} else if (rack->rack_hdrw_pacing &&
15082 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
15083 				/* Do we need to adjust our rate? */
15084 				const struct tcp_hwrate_limit_table *nrte;
15085 
15086 				if (rack->r_up_only &&
15087 				    (rate_wanted < rack->r_ctl.crte->rate)) {
15088 					/**
15089 					 * We have four possible states here
15090 					 * having to do with the previous time
15091 					 * and this time.
15092 					 *   previous  |  this-time
15093 					 * A)     0      |     0   -- fill_cw not in the picture
15094 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
15095 					 * C)     1      |     1   -- all rates from fill_cw
15096 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
15097 					 *
15098 					 * For case A, C and D we don't allow a drop. But for
15099 					 * case B where we now our on our steady rate we do
15100 					 * allow a drop.
15101 					 *
15102 					 */
15103 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
15104 						goto done_w_hdwr;
15105 				}
15106 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
15107 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
15108 					if (rack_hw_rate_to_low &&
15109 					    (bw_est < rack_hw_rate_to_low)) {
15110 						/*
15111 						 * The pacing rate is too low for hardware, but
15112 						 * do allow hardware pacing to be restarted.
15113 						 */
15114 						rack_log_hdwr_pacing(rack,
15115 							     bw_est, rack->r_ctl.crte->rate, __LINE__,
15116 							     0, 5);
15117 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
15118 						rack->r_ctl.crte = NULL;
15119 						rack->rack_attempt_hdwr_pace = 0;
15120 						rack->rack_hdrw_pacing = 0;
15121 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15122 						goto done_w_hdwr;
15123 					}
15124 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
15125 								   rack->rc_tp,
15126 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
15127 								   rate_wanted,
15128 								   RS_PACING_GEQ,
15129 								   &err, &rack->r_ctl.crte_prev_rate);
15130 					if (nrte == NULL) {
15131 						/* Lost the rate */
15132 						rack->rack_hdrw_pacing = 0;
15133 						rack->r_ctl.crte = NULL;
15134 						rack_log_hdwr_pacing(rack,
15135 								     rate_wanted, 0, __LINE__,
15136 								     err, 1);
15137 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15138 						counter_u64_add(rack_hw_pace_lost, 1);
15139 					} else if (nrte != rack->r_ctl.crte) {
15140 						rack->r_ctl.crte = nrte;
15141 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted,
15142 													 segsiz, 0,
15143 													 rack->r_ctl.crte,
15144 													 NULL);
15145 						rack_log_hdwr_pacing(rack,
15146 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15147 								     err, 2);
15148 						rack->r_ctl.last_hw_bw_req = rate_wanted;
15149 					}
15150 				} else {
15151 					/* We just need to adjust the segment size */
15152 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
15153 					rack_log_hdwr_pacing(rack,
15154 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
15155 							     0, 4);
15156 					rack->r_ctl.last_hw_bw_req = rate_wanted;
15157 				}
15158 			}
15159 		}
15160 		if ((rack->r_ctl.crte != NULL) &&
15161 		    (rack->r_ctl.crte->rate == rate_wanted)) {
15162 			/*
15163 			 * We need to add a extra if the rates
15164 			 * are exactly matched. The idea is
15165 			 * we want the software to make sure the
15166 			 * queue is empty before adding more, this
15167 			 * gives us N MSS extra pace times where
15168 			 * N is our sysctl
15169 			 */
15170 			slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots);
15171 		}
15172 done_w_hdwr:
15173 		if (rack_limit_time_with_srtt &&
15174 		    (rack->use_fixed_rate == 0) &&
15175 #ifdef NETFLIX_PEAKRATE
15176 		    (rack->rc_tp->t_maxpeakrate == 0) &&
15177 #endif
15178 		    (rack->rack_hdrw_pacing == 0)) {
15179 			/*
15180 			 * Sanity check, we do not allow the pacing delay
15181 			 * to be longer than the SRTT of the path. If it is
15182 			 * a slow path, then adding a packet should increase
15183 			 * the RTT and compensate for this i.e. the srtt will
15184 			 * be greater so the allowed pacing time will be greater.
15185 			 *
15186 			 * Note this restriction is not for where a peak rate
15187 			 * is set, we are doing fixed pacing or hardware pacing.
15188 			 */
15189 			if (rack->rc_tp->t_srtt)
15190 				srtt = rack->rc_tp->t_srtt;
15191 			else
15192 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
15193 			if (srtt < (uint64_t)slot) {
15194 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
15195 				slot = srtt;
15196 			}
15197 		}
15198 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
15199 	}
15200 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
15201 		/*
15202 		 * If this rate is seeing enobufs when it
15203 		 * goes to send then either the nic is out
15204 		 * of gas or we are mis-estimating the time
15205 		 * somehow and not letting the queue empty
15206 		 * completely. Lets add to the pacing time.
15207 		 */
15208 		int hw_boost_delay;
15209 
15210 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
15211 		if (hw_boost_delay > rack_enobuf_hw_max)
15212 			hw_boost_delay = rack_enobuf_hw_max;
15213 		else if (hw_boost_delay < rack_enobuf_hw_min)
15214 			hw_boost_delay = rack_enobuf_hw_min;
15215 		slot += hw_boost_delay;
15216 	}
15217 	return (slot);
15218 }
15219 
15220 static void
15221 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
15222     tcp_seq startseq, uint32_t sb_offset)
15223 {
15224 	struct rack_sendmap *my_rsm = NULL;
15225 	struct rack_sendmap fe;
15226 
15227 	if (tp->t_state < TCPS_ESTABLISHED) {
15228 		/*
15229 		 * We don't start any measurements if we are
15230 		 * not at least established.
15231 		 */
15232 		return;
15233 	}
15234 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
15235 		/*
15236 		 * We will get no more data into the SB
15237 		 * this means we need to have the data available
15238 		 * before we start a measurement.
15239 		 */
15240 
15241 		if (sbavail(&tptosocket(tp)->so_snd) <
15242 		    max(rc_init_window(rack),
15243 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
15244 			/* Nope not enough data */
15245 			return;
15246 		}
15247 	}
15248 	tp->t_flags |= TF_GPUTINPROG;
15249 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
15250 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
15251 	tp->gput_seq = startseq;
15252 	rack->app_limited_needs_set = 0;
15253 	if (rack->in_probe_rtt)
15254 		rack->measure_saw_probe_rtt = 1;
15255 	else if ((rack->measure_saw_probe_rtt) &&
15256 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
15257 		rack->measure_saw_probe_rtt = 0;
15258 	if (rack->rc_gp_filled)
15259 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15260 	else {
15261 		/* Special case initial measurement */
15262 		struct timeval tv;
15263 
15264 		tp->gput_ts = tcp_get_usecs(&tv);
15265 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15266 	}
15267 	/*
15268 	 * We take a guess out into the future,
15269 	 * if we have no measurement and no
15270 	 * initial rate, we measure the first
15271 	 * initial-windows worth of data to
15272 	 * speed up getting some GP measurement and
15273 	 * thus start pacing.
15274 	 */
15275 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
15276 		rack->app_limited_needs_set = 1;
15277 		tp->gput_ack = startseq + max(rc_init_window(rack),
15278 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
15279 		rack_log_pacing_delay_calc(rack,
15280 					   tp->gput_seq,
15281 					   tp->gput_ack,
15282 					   0,
15283 					   tp->gput_ts,
15284 					   rack->r_ctl.rc_app_limited_cnt,
15285 					   9,
15286 					   __LINE__, NULL, 0);
15287 		return;
15288 	}
15289 	if (sb_offset) {
15290 		/*
15291 		 * We are out somewhere in the sb
15292 		 * can we use the already outstanding data?
15293 		 */
15294 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
15295 			/*
15296 			 * Yes first one is good and in this case
15297 			 * the tp->gput_ts is correctly set based on
15298 			 * the last ack that arrived (no need to
15299 			 * set things up when an ack comes in).
15300 			 */
15301 			my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
15302 			if ((my_rsm == NULL) ||
15303 			    (my_rsm->r_rtr_cnt != 1)) {
15304 				/* retransmission? */
15305 				goto use_latest;
15306 			}
15307 		} else {
15308 			if (rack->r_ctl.rc_first_appl == NULL) {
15309 				/*
15310 				 * If rc_first_appl is NULL
15311 				 * then the cnt should be 0.
15312 				 * This is probably an error, maybe
15313 				 * a KASSERT would be approprate.
15314 				 */
15315 				goto use_latest;
15316 			}
15317 			/*
15318 			 * If we have a marker pointer to the last one that is
15319 			 * app limited we can use that, but we need to set
15320 			 * things up so that when it gets ack'ed we record
15321 			 * the ack time (if its not already acked).
15322 			 */
15323 			rack->app_limited_needs_set = 1;
15324 			/*
15325 			 * We want to get to the rsm that is either
15326 			 * next with space i.e. over 1 MSS or the one
15327 			 * after that (after the app-limited).
15328 			 */
15329 			my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15330 					 rack->r_ctl.rc_first_appl);
15331 			if (my_rsm) {
15332 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
15333 					/* Have to use the next one */
15334 					my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree,
15335 							 my_rsm);
15336 				else {
15337 					/* Use after the first MSS of it is acked */
15338 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
15339 					goto start_set;
15340 				}
15341 			}
15342 			if ((my_rsm == NULL) ||
15343 			    (my_rsm->r_rtr_cnt != 1)) {
15344 				/*
15345 				 * Either its a retransmit or
15346 				 * the last is the app-limited one.
15347 				 */
15348 				goto use_latest;
15349 			}
15350 		}
15351 		tp->gput_seq = my_rsm->r_start;
15352 start_set:
15353 		if (my_rsm->r_flags & RACK_ACKED) {
15354 			/*
15355 			 * This one has been acked use the arrival ack time
15356 			 */
15357 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15358 			rack->app_limited_needs_set = 0;
15359 		}
15360 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15361 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
15362 		rack_log_pacing_delay_calc(rack,
15363 					   tp->gput_seq,
15364 					   tp->gput_ack,
15365 					   (uint64_t)my_rsm,
15366 					   tp->gput_ts,
15367 					   rack->r_ctl.rc_app_limited_cnt,
15368 					   9,
15369 					   __LINE__, NULL, 0);
15370 		return;
15371 	}
15372 
15373 use_latest:
15374 	/*
15375 	 * We don't know how long we may have been
15376 	 * idle or if this is the first-send. Lets
15377 	 * setup the flag so we will trim off
15378 	 * the first ack'd data so we get a true
15379 	 * measurement.
15380 	 */
15381 	rack->app_limited_needs_set = 1;
15382 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
15383 	/* Find this guy so we can pull the send time */
15384 	fe.r_start = startseq;
15385 	my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
15386 	if (my_rsm) {
15387 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)];
15388 		if (my_rsm->r_flags & RACK_ACKED) {
15389 			/*
15390 			 * Unlikely since its probably what was
15391 			 * just transmitted (but I am paranoid).
15392 			 */
15393 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
15394 			rack->app_limited_needs_set = 0;
15395 		}
15396 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
15397 			/* This also is unlikely */
15398 			tp->gput_seq = my_rsm->r_start;
15399 		}
15400 	} else {
15401 		/*
15402 		 * TSNH unless we have some send-map limit,
15403 		 * and even at that it should not be hitting
15404 		 * that limit (we should have stopped sending).
15405 		 */
15406 		struct timeval tv;
15407 
15408 		microuptime(&tv);
15409 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
15410 	}
15411 	rack_log_pacing_delay_calc(rack,
15412 				   tp->gput_seq,
15413 				   tp->gput_ack,
15414 				   (uint64_t)my_rsm,
15415 				   tp->gput_ts,
15416 				   rack->r_ctl.rc_app_limited_cnt,
15417 				   9, __LINE__, NULL, 0);
15418 }
15419 
15420 static inline uint32_t
15421 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
15422     uint32_t avail, int32_t sb_offset)
15423 {
15424 	uint32_t len;
15425 	uint32_t sendwin;
15426 
15427 	if (tp->snd_wnd > cwnd_to_use)
15428 		sendwin = cwnd_to_use;
15429 	else
15430 		sendwin = tp->snd_wnd;
15431 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
15432 		/* We never want to go over our peers rcv-window */
15433 		len = 0;
15434 	} else {
15435 		uint32_t flight;
15436 
15437 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15438 		if (flight >= sendwin) {
15439 			/*
15440 			 * We have in flight what we are allowed by cwnd (if
15441 			 * it was rwnd blocking it would have hit above out
15442 			 * >= tp->snd_wnd).
15443 			 */
15444 			return (0);
15445 		}
15446 		len = sendwin - flight;
15447 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
15448 			/* We would send too much (beyond the rwnd) */
15449 			len = tp->snd_wnd - ctf_outstanding(tp);
15450 		}
15451 		if ((len + sb_offset) > avail) {
15452 			/*
15453 			 * We don't have that much in the SB, how much is
15454 			 * there?
15455 			 */
15456 			len = avail - sb_offset;
15457 		}
15458 	}
15459 	return (len);
15460 }
15461 
15462 static void
15463 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
15464 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
15465 	     int rsm_is_null, int optlen, int line, uint16_t mode)
15466 {
15467 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
15468 		union tcp_log_stackspecific log;
15469 		struct timeval tv;
15470 
15471 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15472 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
15473 		log.u_bbr.flex1 = error;
15474 		log.u_bbr.flex2 = flags;
15475 		log.u_bbr.flex3 = rsm_is_null;
15476 		log.u_bbr.flex4 = ipoptlen;
15477 		log.u_bbr.flex5 = tp->rcv_numsacks;
15478 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
15479 		log.u_bbr.flex7 = optlen;
15480 		log.u_bbr.flex8 = rack->r_fsb_inited;
15481 		log.u_bbr.applimited = rack->r_fast_output;
15482 		log.u_bbr.bw_inuse = rack_get_bw(rack);
15483 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
15484 		log.u_bbr.cwnd_gain = mode;
15485 		log.u_bbr.pkts_out = orig_len;
15486 		log.u_bbr.lt_epoch = len;
15487 		log.u_bbr.delivered = line;
15488 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15489 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15490 		tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
15491 			       len, &log, false, NULL, NULL, 0, &tv);
15492 	}
15493 }
15494 
15495 
15496 static struct mbuf *
15497 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
15498 		   struct rack_fast_send_blk *fsb,
15499 		   int32_t seglimit, int32_t segsize, int hw_tls)
15500 {
15501 #ifdef KERN_TLS
15502 	struct ktls_session *tls, *ntls;
15503 #ifdef INVARIANTS
15504 	struct mbuf *start;
15505 #endif
15506 #endif
15507 	struct mbuf *m, *n, **np, *smb;
15508 	struct mbuf *top;
15509 	int32_t off, soff;
15510 	int32_t len = *plen;
15511 	int32_t fragsize;
15512 	int32_t len_cp = 0;
15513 	uint32_t mlen, frags;
15514 
15515 	soff = off = the_off;
15516 	smb = m = the_m;
15517 	np = &top;
15518 	top = NULL;
15519 #ifdef KERN_TLS
15520 	if (hw_tls && (m->m_flags & M_EXTPG))
15521 		tls = m->m_epg_tls;
15522 	else
15523 		tls = NULL;
15524 #ifdef INVARIANTS
15525 	start = m;
15526 #endif
15527 #endif
15528 	while (len > 0) {
15529 		if (m == NULL) {
15530 			*plen = len_cp;
15531 			break;
15532 		}
15533 #ifdef KERN_TLS
15534 		if (hw_tls) {
15535 			if (m->m_flags & M_EXTPG)
15536 				ntls = m->m_epg_tls;
15537 			else
15538 				ntls = NULL;
15539 
15540 			/*
15541 			 * Avoid mixing TLS records with handshake
15542 			 * data or TLS records from different
15543 			 * sessions.
15544 			 */
15545 			if (tls != ntls) {
15546 				MPASS(m != start);
15547 				*plen = len_cp;
15548 				break;
15549 			}
15550 		}
15551 #endif
15552 		mlen = min(len, m->m_len - off);
15553 		if (seglimit) {
15554 			/*
15555 			 * For M_EXTPG mbufs, add 3 segments
15556 			 * + 1 in case we are crossing page boundaries
15557 			 * + 2 in case the TLS hdr/trailer are used
15558 			 * It is cheaper to just add the segments
15559 			 * than it is to take the cache miss to look
15560 			 * at the mbuf ext_pgs state in detail.
15561 			 */
15562 			if (m->m_flags & M_EXTPG) {
15563 				fragsize = min(segsize, PAGE_SIZE);
15564 				frags = 3;
15565 			} else {
15566 				fragsize = segsize;
15567 				frags = 0;
15568 			}
15569 
15570 			/* Break if we really can't fit anymore. */
15571 			if ((frags + 1) >= seglimit) {
15572 				*plen =	len_cp;
15573 				break;
15574 			}
15575 
15576 			/*
15577 			 * Reduce size if you can't copy the whole
15578 			 * mbuf. If we can't copy the whole mbuf, also
15579 			 * adjust len so the loop will end after this
15580 			 * mbuf.
15581 			 */
15582 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
15583 				mlen = (seglimit - frags - 1) * fragsize;
15584 				len = mlen;
15585 				*plen = len_cp + len;
15586 			}
15587 			frags += howmany(mlen, fragsize);
15588 			if (frags == 0)
15589 				frags++;
15590 			seglimit -= frags;
15591 			KASSERT(seglimit > 0,
15592 			    ("%s: seglimit went too low", __func__));
15593 		}
15594 		n = m_get(M_NOWAIT, m->m_type);
15595 		*np = n;
15596 		if (n == NULL)
15597 			goto nospace;
15598 		n->m_len = mlen;
15599 		soff += mlen;
15600 		len_cp += n->m_len;
15601 		if (m->m_flags & (M_EXT|M_EXTPG)) {
15602 			n->m_data = m->m_data + off;
15603 			mb_dupcl(n, m);
15604 		} else {
15605 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
15606 			    (u_int)n->m_len);
15607 		}
15608 		len -= n->m_len;
15609 		off = 0;
15610 		m = m->m_next;
15611 		np = &n->m_next;
15612 		if (len || (soff == smb->m_len)) {
15613 			/*
15614 			 * We have more so we move forward  or
15615 			 * we have consumed the entire mbuf and
15616 			 * len has fell to 0.
15617 			 */
15618 			soff = 0;
15619 			smb = m;
15620 		}
15621 
15622 	}
15623 	if (fsb != NULL) {
15624 		fsb->m = smb;
15625 		fsb->off = soff;
15626 		if (smb) {
15627 			/*
15628 			 * Save off the size of the mbuf. We do
15629 			 * this so that we can recognize when it
15630 			 * has been trimmed by sbcut() as acks
15631 			 * come in.
15632 			 */
15633 			fsb->o_m_len = smb->m_len;
15634 		} else {
15635 			/*
15636 			 * This is the case where the next mbuf went to NULL. This
15637 			 * means with this copy we have sent everything in the sb.
15638 			 * In theory we could clear the fast_output flag, but lets
15639 			 * not since its possible that we could get more added
15640 			 * and acks that call the extend function which would let
15641 			 * us send more.
15642 			 */
15643 			fsb->o_m_len = 0;
15644 		}
15645 	}
15646 	return (top);
15647 nospace:
15648 	if (top)
15649 		m_freem(top);
15650 	return (NULL);
15651 
15652 }
15653 
15654 /*
15655  * This is a copy of m_copym(), taking the TSO segment size/limit
15656  * constraints into account, and advancing the sndptr as it goes.
15657  */
15658 static struct mbuf *
15659 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
15660 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
15661 {
15662 	struct mbuf *m, *n;
15663 	int32_t soff;
15664 
15665 	soff = rack->r_ctl.fsb.off;
15666 	m = rack->r_ctl.fsb.m;
15667 	if (rack->r_ctl.fsb.o_m_len > m->m_len) {
15668 		/*
15669 		 * The mbuf had the front of it chopped off by an ack
15670 		 * we need to adjust the soff/off by that difference.
15671 		 */
15672 		uint32_t delta;
15673 
15674 		delta = rack->r_ctl.fsb.o_m_len - m->m_len;
15675 		soff -= delta;
15676 	} else if (rack->r_ctl.fsb.o_m_len < m->m_len) {
15677 		/*
15678 		 * The mbuf was expanded probably by
15679 		 * a m_compress. Just update o_m_len.
15680 		 */
15681 		rack->r_ctl.fsb.o_m_len = m->m_len;
15682 	}
15683 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
15684 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
15685 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
15686 				 __FUNCTION__,
15687 				 rack, *plen, m, m->m_len));
15688 	/* Save off the right location before we copy and advance */
15689 	*s_soff = soff;
15690 	*s_mb = rack->r_ctl.fsb.m;
15691 	n = rack_fo_base_copym(m, soff, plen,
15692 			       &rack->r_ctl.fsb,
15693 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
15694 	return (n);
15695 }
15696 
15697 static int
15698 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
15699 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
15700 {
15701 	/*
15702 	 * Enter the fast retransmit path. We are given that a sched_pin is
15703 	 * in place (if accounting is compliled in) and the cycle count taken
15704 	 * at the entry is in the ts_val. The concept her is that the rsm
15705 	 * now holds the mbuf offsets and such so we can directly transmit
15706 	 * without a lot of overhead, the len field is already set for
15707 	 * us to prohibit us from sending too much (usually its 1MSS).
15708 	 */
15709 	struct ip *ip = NULL;
15710 	struct udphdr *udp = NULL;
15711 	struct tcphdr *th = NULL;
15712 	struct mbuf *m = NULL;
15713 	struct inpcb *inp;
15714 	uint8_t *cpto;
15715 	struct tcp_log_buffer *lgb;
15716 #ifdef TCP_ACCOUNTING
15717 	uint64_t crtsc;
15718 	int cnt_thru = 1;
15719 #endif
15720 	struct tcpopt to;
15721 	u_char opt[TCP_MAXOLEN];
15722 	uint32_t hdrlen, optlen;
15723 	int32_t slot, segsiz, max_val, tso = 0, error, ulen = 0;
15724 	uint16_t flags;
15725 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
15726 	uint32_t if_hw_tsomaxsegsize;
15727 
15728 #ifdef INET6
15729 	struct ip6_hdr *ip6 = NULL;
15730 
15731 	if (rack->r_is_v6) {
15732 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
15733 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
15734 	} else
15735 #endif				/* INET6 */
15736 	{
15737 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
15738 		hdrlen = sizeof(struct tcpiphdr);
15739 	}
15740 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
15741 		goto failed;
15742 	}
15743 	if (doing_tlp) {
15744 		/* Its a TLP add the flag, it may already be there but be sure */
15745 		rsm->r_flags |= RACK_TLP;
15746 	} else {
15747 		/* If it was a TLP it is not not on this retransmit */
15748 		rsm->r_flags &= ~RACK_TLP;
15749 	}
15750 	startseq = rsm->r_start;
15751 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
15752 	inp = rack->rc_inp;
15753 	to.to_flags = 0;
15754 	flags = tcp_outflags[tp->t_state];
15755 	if (flags & (TH_SYN|TH_RST)) {
15756 		goto failed;
15757 	}
15758 	if (rsm->r_flags & RACK_HAS_FIN) {
15759 		/* We can't send a FIN here */
15760 		goto failed;
15761 	}
15762 	if (flags & TH_FIN) {
15763 		/* We never send a FIN */
15764 		flags &= ~TH_FIN;
15765 	}
15766 	if (tp->t_flags & TF_RCVD_TSTMP) {
15767 		to.to_tsval = ms_cts + tp->ts_offset;
15768 		to.to_tsecr = tp->ts_recent;
15769 		to.to_flags = TOF_TS;
15770 	}
15771 	optlen = tcp_addoptions(&to, opt);
15772 	hdrlen += optlen;
15773 	udp = rack->r_ctl.fsb.udp;
15774 	if (udp)
15775 		hdrlen += sizeof(struct udphdr);
15776 	if (rack->r_ctl.rc_pace_max_segs)
15777 		max_val = rack->r_ctl.rc_pace_max_segs;
15778 	else if (rack->rc_user_set_max_segs)
15779 		max_val = rack->rc_user_set_max_segs * segsiz;
15780 	else
15781 		max_val = len;
15782 	if ((tp->t_flags & TF_TSO) &&
15783 	    V_tcp_do_tso &&
15784 	    (len > segsiz) &&
15785 	    (tp->t_port == 0))
15786 		tso = 1;
15787 #ifdef INET6
15788 	if (MHLEN < hdrlen + max_linkhdr)
15789 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
15790 	else
15791 #endif
15792 		m = m_gethdr(M_NOWAIT, MT_DATA);
15793 	if (m == NULL)
15794 		goto failed;
15795 	m->m_data += max_linkhdr;
15796 	m->m_len = hdrlen;
15797 	th = rack->r_ctl.fsb.th;
15798 	/* Establish the len to send */
15799 	if (len > max_val)
15800 		len = max_val;
15801 	if ((tso) && (len + optlen > tp->t_maxseg)) {
15802 		uint32_t if_hw_tsomax;
15803 		int32_t max_len;
15804 
15805 		/* extract TSO information */
15806 		if_hw_tsomax = tp->t_tsomax;
15807 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
15808 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
15809 		/*
15810 		 * Check if we should limit by maximum payload
15811 		 * length:
15812 		 */
15813 		if (if_hw_tsomax != 0) {
15814 			/* compute maximum TSO length */
15815 			max_len = (if_hw_tsomax - hdrlen -
15816 				   max_linkhdr);
15817 			if (max_len <= 0) {
15818 				goto failed;
15819 			} else if (len > max_len) {
15820 				len = max_len;
15821 			}
15822 		}
15823 		if (len <= segsiz) {
15824 			/*
15825 			 * In case there are too many small fragments don't
15826 			 * use TSO:
15827 			 */
15828 			tso = 0;
15829 		}
15830 	} else {
15831 		tso = 0;
15832 	}
15833 	if ((tso == 0) && (len > segsiz))
15834 		len = segsiz;
15835 	if ((len == 0) ||
15836 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
15837 		goto failed;
15838 	}
15839 	th->th_seq = htonl(rsm->r_start);
15840 	th->th_ack = htonl(tp->rcv_nxt);
15841 	/*
15842 	 * The PUSH bit should only be applied
15843 	 * if the full retransmission is made. If
15844 	 * we are sending less than this is the
15845 	 * left hand edge and should not have
15846 	 * the PUSH bit.
15847 	 */
15848 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
15849 	    (len == (rsm->r_end - rsm->r_start)))
15850 		flags |= TH_PUSH;
15851 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
15852 	if (th->th_win == 0) {
15853 		tp->t_sndzerowin++;
15854 		tp->t_flags |= TF_RXWIN0SENT;
15855 	} else
15856 		tp->t_flags &= ~TF_RXWIN0SENT;
15857 	if (rsm->r_flags & RACK_TLP) {
15858 		/*
15859 		 * TLP should not count in retran count, but
15860 		 * in its own bin
15861 		 */
15862 		counter_u64_add(rack_tlp_retran, 1);
15863 		counter_u64_add(rack_tlp_retran_bytes, len);
15864 	} else {
15865 		tp->t_sndrexmitpack++;
15866 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
15867 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
15868 	}
15869 #ifdef STATS
15870 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
15871 				 len);
15872 #endif
15873 	if (rsm->m == NULL)
15874 		goto failed;
15875 	if (rsm->orig_m_len != rsm->m->m_len) {
15876 		/* Fix up the orig_m_len and possibly the mbuf offset */
15877 		rack_adjust_orig_mlen(rsm);
15878 	}
15879 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
15880 	if (len <= segsiz) {
15881 		/*
15882 		 * Must have ran out of mbufs for the copy
15883 		 * shorten it to no longer need tso. Lets
15884 		 * not put on sendalot since we are low on
15885 		 * mbufs.
15886 		 */
15887 		tso = 0;
15888 	}
15889 	if ((m->m_next == NULL) || (len <= 0)){
15890 		goto failed;
15891 	}
15892 	if (udp) {
15893 		if (rack->r_is_v6)
15894 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
15895 		else
15896 			ulen = hdrlen + len - sizeof(struct ip);
15897 		udp->uh_ulen = htons(ulen);
15898 	}
15899 	m->m_pkthdr.rcvif = (struct ifnet *)0;
15900 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
15901 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
15902 		int ect = tcp_ecn_output_established(tp, &flags, len, true);
15903 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
15904 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
15905 		    tp->t_flags2 &= ~TF2_ECN_SND_ECE;
15906 #ifdef INET6
15907 		if (rack->r_is_v6) {
15908 		    ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
15909 		    ip6->ip6_flow |= htonl(ect << 20);
15910 		}
15911 		else
15912 #endif
15913 		{
15914 		    ip->ip_tos &= ~IPTOS_ECN_MASK;
15915 		    ip->ip_tos |= ect;
15916 		}
15917 	}
15918 	tcp_set_flags(th, flags);
15919 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
15920 #ifdef INET6
15921 	if (rack->r_is_v6) {
15922 		if (tp->t_port) {
15923 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
15924 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15925 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
15926 			th->th_sum = htons(0);
15927 			UDPSTAT_INC(udps_opackets);
15928 		} else {
15929 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
15930 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15931 			th->th_sum = in6_cksum_pseudo(ip6,
15932 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
15933 						      0);
15934 		}
15935 	}
15936 #endif
15937 #if defined(INET6) && defined(INET)
15938 	else
15939 #endif
15940 #ifdef INET
15941 	{
15942 		if (tp->t_port) {
15943 			m->m_pkthdr.csum_flags = CSUM_UDP;
15944 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
15945 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
15946 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
15947 			th->th_sum = htons(0);
15948 			UDPSTAT_INC(udps_opackets);
15949 		} else {
15950 			m->m_pkthdr.csum_flags = CSUM_TCP;
15951 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
15952 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
15953 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
15954 									IPPROTO_TCP + len + optlen));
15955 		}
15956 		/* IP version must be set here for ipv4/ipv6 checking later */
15957 		KASSERT(ip->ip_v == IPVERSION,
15958 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
15959 	}
15960 #endif
15961 	if (tso) {
15962 		KASSERT(len > tp->t_maxseg - optlen,
15963 			("%s: len <= tso_segsz tp:%p", __func__, tp));
15964 		m->m_pkthdr.csum_flags |= CSUM_TSO;
15965 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
15966 	}
15967 #ifdef INET6
15968 	if (rack->r_is_v6) {
15969 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
15970 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
15971 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
15972 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15973 		else
15974 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15975 	}
15976 #endif
15977 #if defined(INET) && defined(INET6)
15978 	else
15979 #endif
15980 #ifdef INET
15981 	{
15982 		ip->ip_len = htons(m->m_pkthdr.len);
15983 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
15984 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
15985 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
15986 			if (tp->t_port == 0 || len < V_tcp_minmss) {
15987 				ip->ip_off |= htons(IP_DF);
15988 			}
15989 		} else {
15990 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
15991 		}
15992 	}
15993 #endif
15994 	/* Time to copy in our header */
15995 	cpto = mtod(m, uint8_t *);
15996 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
15997 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
15998 	if (optlen) {
15999 		bcopy(opt, th + 1, optlen);
16000 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
16001 	} else {
16002 		th->th_off = sizeof(struct tcphdr) >> 2;
16003 	}
16004 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
16005 		union tcp_log_stackspecific log;
16006 
16007 		if (rsm->r_flags & RACK_RWND_COLLAPSED) {
16008 			rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
16009 			counter_u64_add(rack_collapsed_win_rxt, 1);
16010 			counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
16011 		}
16012 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16013 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
16014 		if (rack->rack_no_prr)
16015 			log.u_bbr.flex1 = 0;
16016 		else
16017 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16018 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16019 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16020 		log.u_bbr.flex4 = max_val;
16021 		log.u_bbr.flex5 = 0;
16022 		/* Save off the early/late values */
16023 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16024 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16025 		log.u_bbr.bw_inuse = rack_get_bw(rack);
16026 		if (doing_tlp == 0)
16027 			log.u_bbr.flex8 = 1;
16028 		else
16029 			log.u_bbr.flex8 = 2;
16030 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16031 		log.u_bbr.flex7 = 55;
16032 		log.u_bbr.pkts_out = tp->t_maxseg;
16033 		log.u_bbr.timeStamp = cts;
16034 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16035 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16036 		log.u_bbr.delivered = 0;
16037 		lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16038 				     len, &log, false, NULL, NULL, 0, tv);
16039 	} else
16040 		lgb = NULL;
16041 #ifdef INET6
16042 	if (rack->r_is_v6) {
16043 		error = ip6_output(m, NULL,
16044 				   &inp->inp_route6,
16045 				   0, NULL, NULL, inp);
16046 	}
16047 #endif
16048 #if defined(INET) && defined(INET6)
16049 	else
16050 #endif
16051 #ifdef INET
16052 	{
16053 		error = ip_output(m, NULL,
16054 				  &inp->inp_route,
16055 				  0, 0, inp);
16056 	}
16057 #endif
16058 	m = NULL;
16059 	if (lgb) {
16060 		lgb->tlb_errno = error;
16061 		lgb = NULL;
16062 	}
16063 	if (error) {
16064 		goto failed;
16065 	}
16066 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
16067 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls);
16068 	if (doing_tlp && (rack->fast_rsm_hack == 0)) {
16069 		rack->rc_tlp_in_progress = 1;
16070 		rack->r_ctl.rc_tlp_cnt_out++;
16071 	}
16072 	if (error == 0) {
16073 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
16074 		if (doing_tlp) {
16075 			rack->rc_last_sent_tlp_past_cumack = 0;
16076 			rack->rc_last_sent_tlp_seq_valid = 1;
16077 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
16078 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
16079 		}
16080 	}
16081 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16082 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
16083 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
16084 		rack->r_ctl.retran_during_recovery += len;
16085 	{
16086 		int idx;
16087 
16088 		idx = (len / segsiz) + 3;
16089 		if (idx >= TCP_MSS_ACCT_ATIMER)
16090 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16091 		else
16092 			counter_u64_add(rack_out_size[idx], 1);
16093 	}
16094 	if (tp->t_rtttime == 0) {
16095 		tp->t_rtttime = ticks;
16096 		tp->t_rtseq = startseq;
16097 		KMOD_TCPSTAT_INC(tcps_segstimed);
16098 	}
16099 	counter_u64_add(rack_fto_rsm_send, 1);
16100 	if (error && (error == ENOBUFS)) {
16101 		if (rack->r_ctl.crte != NULL) {
16102 			rack_trace_point(rack, RACK_TP_HWENOBUF);
16103 		} else
16104 			rack_trace_point(rack, RACK_TP_ENOBUF);
16105 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
16106 		if (rack->rc_enobuf < 0x7f)
16107 			rack->rc_enobuf++;
16108 		if (slot < (10 * HPTS_USEC_IN_MSEC))
16109 			slot = 10 * HPTS_USEC_IN_MSEC;
16110 	} else
16111 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
16112 	if ((slot == 0) ||
16113 	    (rack->rc_always_pace == 0) ||
16114 	    (rack->r_rr_config == 1)) {
16115 		/*
16116 		 * We have no pacing set or we
16117 		 * are using old-style rack or
16118 		 * we are overridden to use the old 1ms pacing.
16119 		 */
16120 		slot = rack->r_ctl.rc_min_to;
16121 	}
16122 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
16123 #ifdef TCP_ACCOUNTING
16124 	crtsc = get_cyclecount();
16125 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16126 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16127 	}
16128 	counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16129 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16130 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16131 	}
16132 	counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16133 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16134 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
16135 	}
16136 	counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz));
16137 	sched_unpin();
16138 #endif
16139 	return (0);
16140 failed:
16141 	if (m)
16142 		m_free(m);
16143 	return (-1);
16144 }
16145 
16146 static void
16147 rack_sndbuf_autoscale(struct tcp_rack *rack)
16148 {
16149 	/*
16150 	 * Automatic sizing of send socket buffer.  Often the send buffer
16151 	 * size is not optimally adjusted to the actual network conditions
16152 	 * at hand (delay bandwidth product).  Setting the buffer size too
16153 	 * small limits throughput on links with high bandwidth and high
16154 	 * delay (eg. trans-continental/oceanic links).  Setting the
16155 	 * buffer size too big consumes too much real kernel memory,
16156 	 * especially with many connections on busy servers.
16157 	 *
16158 	 * The criteria to step up the send buffer one notch are:
16159 	 *  1. receive window of remote host is larger than send buffer
16160 	 *     (with a fudge factor of 5/4th);
16161 	 *  2. send buffer is filled to 7/8th with data (so we actually
16162 	 *     have data to make use of it);
16163 	 *  3. send buffer fill has not hit maximal automatic size;
16164 	 *  4. our send window (slow start and cogestion controlled) is
16165 	 *     larger than sent but unacknowledged data in send buffer.
16166 	 *
16167 	 * Note that the rack version moves things much faster since
16168 	 * we want to avoid hitting cache lines in the rack_fast_output()
16169 	 * path so this is called much less often and thus moves
16170 	 * the SB forward by a percentage.
16171 	 */
16172 	struct socket *so;
16173 	struct tcpcb *tp;
16174 	uint32_t sendwin, scaleup;
16175 
16176 	tp = rack->rc_tp;
16177 	so = rack->rc_inp->inp_socket;
16178 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
16179 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
16180 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
16181 		    sbused(&so->so_snd) >=
16182 		    (so->so_snd.sb_hiwat / 8 * 7) &&
16183 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
16184 		    sendwin >= (sbused(&so->so_snd) -
16185 		    (tp->snd_nxt - tp->snd_una))) {
16186 			if (rack_autosndbuf_inc)
16187 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
16188 			else
16189 				scaleup = V_tcp_autosndbuf_inc;
16190 			if (scaleup < V_tcp_autosndbuf_inc)
16191 				scaleup = V_tcp_autosndbuf_inc;
16192 			scaleup += so->so_snd.sb_hiwat;
16193 			if (scaleup > V_tcp_autosndbuf_max)
16194 				scaleup = V_tcp_autosndbuf_max;
16195 			if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
16196 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
16197 		}
16198 	}
16199 }
16200 
16201 static int
16202 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
16203 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
16204 {
16205 	/*
16206 	 * Enter to do fast output. We are given that the sched_pin is
16207 	 * in place (if accounting is compiled in) and the cycle count taken
16208 	 * at entry is in place in ts_val. The idea here is that
16209 	 * we know how many more bytes needs to be sent (presumably either
16210 	 * during pacing or to fill the cwnd and that was greater than
16211 	 * the max-burst). We have how much to send and all the info we
16212 	 * need to just send.
16213 	 */
16214 	struct ip *ip = NULL;
16215 	struct udphdr *udp = NULL;
16216 	struct tcphdr *th = NULL;
16217 	struct mbuf *m, *s_mb;
16218 	struct inpcb *inp;
16219 	uint8_t *cpto;
16220 	struct tcp_log_buffer *lgb;
16221 #ifdef TCP_ACCOUNTING
16222 	uint64_t crtsc;
16223 #endif
16224 	struct tcpopt to;
16225 	u_char opt[TCP_MAXOLEN];
16226 	uint32_t hdrlen, optlen;
16227 #ifdef TCP_ACCOUNTING
16228 	int cnt_thru = 1;
16229 #endif
16230 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
16231 	uint16_t flags;
16232 	uint32_t s_soff;
16233 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
16234 	uint32_t if_hw_tsomaxsegsize;
16235 	uint16_t add_flag = RACK_SENT_FP;
16236 #ifdef INET6
16237 	struct ip6_hdr *ip6 = NULL;
16238 
16239 	if (rack->r_is_v6) {
16240 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
16241 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
16242 	} else
16243 #endif				/* INET6 */
16244 	{
16245 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
16246 		hdrlen = sizeof(struct tcpiphdr);
16247 	}
16248 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
16249 		m = NULL;
16250 		goto failed;
16251 	}
16252 	startseq = tp->snd_max;
16253 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16254 	inp = rack->rc_inp;
16255 	len = rack->r_ctl.fsb.left_to_send;
16256 	to.to_flags = 0;
16257 	flags = rack->r_ctl.fsb.tcp_flags;
16258 	if (tp->t_flags & TF_RCVD_TSTMP) {
16259 		to.to_tsval = ms_cts + tp->ts_offset;
16260 		to.to_tsecr = tp->ts_recent;
16261 		to.to_flags = TOF_TS;
16262 	}
16263 	optlen = tcp_addoptions(&to, opt);
16264 	hdrlen += optlen;
16265 	udp = rack->r_ctl.fsb.udp;
16266 	if (udp)
16267 		hdrlen += sizeof(struct udphdr);
16268 	if (rack->r_ctl.rc_pace_max_segs)
16269 		max_val = rack->r_ctl.rc_pace_max_segs;
16270 	else if (rack->rc_user_set_max_segs)
16271 		max_val = rack->rc_user_set_max_segs * segsiz;
16272 	else
16273 		max_val = len;
16274 	if ((tp->t_flags & TF_TSO) &&
16275 	    V_tcp_do_tso &&
16276 	    (len > segsiz) &&
16277 	    (tp->t_port == 0))
16278 		tso = 1;
16279 again:
16280 #ifdef INET6
16281 	if (MHLEN < hdrlen + max_linkhdr)
16282 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
16283 	else
16284 #endif
16285 		m = m_gethdr(M_NOWAIT, MT_DATA);
16286 	if (m == NULL)
16287 		goto failed;
16288 	m->m_data += max_linkhdr;
16289 	m->m_len = hdrlen;
16290 	th = rack->r_ctl.fsb.th;
16291 	/* Establish the len to send */
16292 	if (len > max_val)
16293 		len = max_val;
16294 	if ((tso) && (len + optlen > tp->t_maxseg)) {
16295 		uint32_t if_hw_tsomax;
16296 		int32_t max_len;
16297 
16298 		/* extract TSO information */
16299 		if_hw_tsomax = tp->t_tsomax;
16300 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
16301 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
16302 		/*
16303 		 * Check if we should limit by maximum payload
16304 		 * length:
16305 		 */
16306 		if (if_hw_tsomax != 0) {
16307 			/* compute maximum TSO length */
16308 			max_len = (if_hw_tsomax - hdrlen -
16309 				   max_linkhdr);
16310 			if (max_len <= 0) {
16311 				goto failed;
16312 			} else if (len > max_len) {
16313 				len = max_len;
16314 			}
16315 		}
16316 		if (len <= segsiz) {
16317 			/*
16318 			 * In case there are too many small fragments don't
16319 			 * use TSO:
16320 			 */
16321 			tso = 0;
16322 		}
16323 	} else {
16324 		tso = 0;
16325 	}
16326 	if ((tso == 0) && (len > segsiz))
16327 		len = segsiz;
16328 	if ((len == 0) ||
16329 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
16330 		goto failed;
16331 	}
16332 	sb_offset = tp->snd_max - tp->snd_una;
16333 	th->th_seq = htonl(tp->snd_max);
16334 	th->th_ack = htonl(tp->rcv_nxt);
16335 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
16336 	if (th->th_win == 0) {
16337 		tp->t_sndzerowin++;
16338 		tp->t_flags |= TF_RXWIN0SENT;
16339 	} else
16340 		tp->t_flags &= ~TF_RXWIN0SENT;
16341 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
16342 	KMOD_TCPSTAT_INC(tcps_sndpack);
16343 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
16344 #ifdef STATS
16345 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
16346 				 len);
16347 #endif
16348 	if (rack->r_ctl.fsb.m == NULL)
16349 		goto failed;
16350 
16351 	/* s_mb and s_soff are saved for rack_log_output */
16352 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
16353 				    &s_mb, &s_soff);
16354 	if (len <= segsiz) {
16355 		/*
16356 		 * Must have ran out of mbufs for the copy
16357 		 * shorten it to no longer need tso. Lets
16358 		 * not put on sendalot since we are low on
16359 		 * mbufs.
16360 		 */
16361 		tso = 0;
16362 	}
16363 	if (rack->r_ctl.fsb.rfo_apply_push &&
16364 	    (len == rack->r_ctl.fsb.left_to_send)) {
16365 		flags |= TH_PUSH;
16366 		add_flag |= RACK_HAD_PUSH;
16367 	}
16368 	if ((m->m_next == NULL) || (len <= 0)){
16369 		goto failed;
16370 	}
16371 	if (udp) {
16372 		if (rack->r_is_v6)
16373 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
16374 		else
16375 			ulen = hdrlen + len - sizeof(struct ip);
16376 		udp->uh_ulen = htons(ulen);
16377 	}
16378 	m->m_pkthdr.rcvif = (struct ifnet *)0;
16379 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
16380 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
16381 		int ect = tcp_ecn_output_established(tp, &flags, len, false);
16382 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
16383 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
16384 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
16385 #ifdef INET6
16386 		if (rack->r_is_v6) {
16387 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
16388 			ip6->ip6_flow |= htonl(ect << 20);
16389 		}
16390 		else
16391 #endif
16392 		{
16393 			ip->ip_tos &= ~IPTOS_ECN_MASK;
16394 			ip->ip_tos |= ect;
16395 		}
16396 	}
16397 	tcp_set_flags(th, flags);
16398 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
16399 #ifdef INET6
16400 	if (rack->r_is_v6) {
16401 		if (tp->t_port) {
16402 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
16403 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16404 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
16405 			th->th_sum = htons(0);
16406 			UDPSTAT_INC(udps_opackets);
16407 		} else {
16408 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
16409 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16410 			th->th_sum = in6_cksum_pseudo(ip6,
16411 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
16412 						      0);
16413 		}
16414 	}
16415 #endif
16416 #if defined(INET6) && defined(INET)
16417 	else
16418 #endif
16419 #ifdef INET
16420 	{
16421 		if (tp->t_port) {
16422 			m->m_pkthdr.csum_flags = CSUM_UDP;
16423 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
16424 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
16425 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
16426 			th->th_sum = htons(0);
16427 			UDPSTAT_INC(udps_opackets);
16428 		} else {
16429 			m->m_pkthdr.csum_flags = CSUM_TCP;
16430 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
16431 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
16432 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
16433 									IPPROTO_TCP + len + optlen));
16434 		}
16435 		/* IP version must be set here for ipv4/ipv6 checking later */
16436 		KASSERT(ip->ip_v == IPVERSION,
16437 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
16438 	}
16439 #endif
16440 	if (tso) {
16441 		KASSERT(len > tp->t_maxseg - optlen,
16442 			("%s: len <= tso_segsz tp:%p", __func__, tp));
16443 		m->m_pkthdr.csum_flags |= CSUM_TSO;
16444 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
16445 	}
16446 #ifdef INET6
16447 	if (rack->r_is_v6) {
16448 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
16449 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
16450 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
16451 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16452 		else
16453 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16454 	}
16455 #endif
16456 #if defined(INET) && defined(INET6)
16457 	else
16458 #endif
16459 #ifdef INET
16460 	{
16461 		ip->ip_len = htons(m->m_pkthdr.len);
16462 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
16463 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
16464 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
16465 			if (tp->t_port == 0 || len < V_tcp_minmss) {
16466 				ip->ip_off |= htons(IP_DF);
16467 			}
16468 		} else {
16469 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
16470 		}
16471 	}
16472 #endif
16473 	/* Time to copy in our header */
16474 	cpto = mtod(m, uint8_t *);
16475 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
16476 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
16477 	if (optlen) {
16478 		bcopy(opt, th + 1, optlen);
16479 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
16480 	} else {
16481 		th->th_off = sizeof(struct tcphdr) >> 2;
16482 	}
16483 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
16484 		union tcp_log_stackspecific log;
16485 
16486 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16487 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
16488 		if (rack->rack_no_prr)
16489 			log.u_bbr.flex1 = 0;
16490 		else
16491 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16492 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
16493 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
16494 		log.u_bbr.flex4 = max_val;
16495 		log.u_bbr.flex5 = 0;
16496 		/* Save off the early/late values */
16497 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
16498 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
16499 		log.u_bbr.bw_inuse = rack_get_bw(rack);
16500 		log.u_bbr.flex8 = 0;
16501 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
16502 		log.u_bbr.flex7 = 44;
16503 		log.u_bbr.pkts_out = tp->t_maxseg;
16504 		log.u_bbr.timeStamp = cts;
16505 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16506 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
16507 		log.u_bbr.delivered = 0;
16508 		lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
16509 				     len, &log, false, NULL, NULL, 0, tv);
16510 	} else
16511 		lgb = NULL;
16512 #ifdef INET6
16513 	if (rack->r_is_v6) {
16514 		error = ip6_output(m, NULL,
16515 				   &inp->inp_route6,
16516 				   0, NULL, NULL, inp);
16517 	}
16518 #endif
16519 #if defined(INET) && defined(INET6)
16520 	else
16521 #endif
16522 #ifdef INET
16523 	{
16524 		error = ip_output(m, NULL,
16525 				  &inp->inp_route,
16526 				  0, 0, inp);
16527 	}
16528 #endif
16529 	if (lgb) {
16530 		lgb->tlb_errno = error;
16531 		lgb = NULL;
16532 	}
16533 	if (error) {
16534 		*send_err = error;
16535 		m = NULL;
16536 		goto failed;
16537 	}
16538 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
16539 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls);
16540 	m = NULL;
16541 	if (tp->snd_una == tp->snd_max) {
16542 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
16543 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
16544 		tp->t_acktime = ticks;
16545 	}
16546 	if (error == 0)
16547 		tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
16548 
16549 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
16550 	tot_len += len;
16551 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
16552 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
16553 	tp->snd_max += len;
16554 	tp->snd_nxt = tp->snd_max;
16555 	{
16556 		int idx;
16557 
16558 		idx = (len / segsiz) + 3;
16559 		if (idx >= TCP_MSS_ACCT_ATIMER)
16560 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
16561 		else
16562 			counter_u64_add(rack_out_size[idx], 1);
16563 	}
16564 	if (len <= rack->r_ctl.fsb.left_to_send)
16565 		rack->r_ctl.fsb.left_to_send -= len;
16566 	else
16567 		rack->r_ctl.fsb.left_to_send = 0;
16568 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
16569 		rack->r_fast_output = 0;
16570 		rack->r_ctl.fsb.left_to_send = 0;
16571 		/* At the end of fast_output scale up the sb */
16572 		SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
16573 		rack_sndbuf_autoscale(rack);
16574 		SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
16575 	}
16576 	if (tp->t_rtttime == 0) {
16577 		tp->t_rtttime = ticks;
16578 		tp->t_rtseq = startseq;
16579 		KMOD_TCPSTAT_INC(tcps_segstimed);
16580 	}
16581 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
16582 	    (max_val > len) &&
16583 	    (tso == 0)) {
16584 		max_val -= len;
16585 		len = segsiz;
16586 		th = rack->r_ctl.fsb.th;
16587 #ifdef TCP_ACCOUNTING
16588 		cnt_thru++;
16589 #endif
16590 		goto again;
16591 	}
16592 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
16593 	counter_u64_add(rack_fto_send, 1);
16594 	slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
16595 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
16596 #ifdef TCP_ACCOUNTING
16597 	crtsc = get_cyclecount();
16598 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16599 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
16600 	}
16601 	counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru);
16602 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16603 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
16604 	}
16605 	counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
16606 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16607 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
16608 	}
16609 	counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz));
16610 	sched_unpin();
16611 #endif
16612 	return (0);
16613 failed:
16614 	if (m)
16615 		m_free(m);
16616 	rack->r_fast_output = 0;
16617 	return (-1);
16618 }
16619 
16620 static struct rack_sendmap *
16621 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
16622 {
16623 	struct rack_sendmap *rsm = NULL;
16624 	struct rack_sendmap fe;
16625 	int thresh;
16626 
16627 restart:
16628 	fe.r_start = rack->r_ctl.last_collapse_point;
16629 	rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe);
16630 	if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
16631 		/* Nothing, strange turn off validity  */
16632 		rack->r_collapse_point_valid = 0;
16633 		return (NULL);
16634 	}
16635 	/* Can we send it yet? */
16636 	if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
16637 		/*
16638 		 * Receiver window has not grown enough for
16639 		 * the segment to be put on the wire.
16640 		 */
16641 		return (NULL);
16642 	}
16643 	if (rsm->r_flags & RACK_ACKED) {
16644 		/*
16645 		 * It has been sacked, lets move to the
16646 		 * next one if possible.
16647 		 */
16648 		rack->r_ctl.last_collapse_point = rsm->r_end;
16649 		/* Are we done? */
16650 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
16651 			    rack->r_ctl.high_collapse_point)) {
16652 			rack->r_collapse_point_valid = 0;
16653 			return (NULL);
16654 		}
16655 		goto restart;
16656 	}
16657 	/* Now has it been long enough ? */
16658 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts);
16659 	if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
16660 		rack_log_collapse(rack, rsm->r_start,
16661 				  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
16662 				  thresh, __LINE__, 6, rsm->r_flags, rsm);
16663 		return (rsm);
16664 	}
16665 	/* Not enough time */
16666 	rack_log_collapse(rack, rsm->r_start,
16667 			  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
16668 			  thresh, __LINE__, 7, rsm->r_flags, rsm);
16669 	return (NULL);
16670 }
16671 
16672 static int
16673 rack_output(struct tcpcb *tp)
16674 {
16675 	struct socket *so;
16676 	uint32_t recwin;
16677 	uint32_t sb_offset, s_moff = 0;
16678 	int32_t len, error = 0;
16679 	uint16_t flags;
16680 	struct mbuf *m, *s_mb = NULL;
16681 	struct mbuf *mb;
16682 	uint32_t if_hw_tsomaxsegcount = 0;
16683 	uint32_t if_hw_tsomaxsegsize;
16684 	int32_t segsiz, minseg;
16685 	long tot_len_this_send = 0;
16686 #ifdef INET
16687 	struct ip *ip = NULL;
16688 #endif
16689 	struct udphdr *udp = NULL;
16690 	struct tcp_rack *rack;
16691 	struct tcphdr *th;
16692 	uint8_t pass = 0;
16693 	uint8_t mark = 0;
16694 	uint8_t wanted_cookie = 0;
16695 	u_char opt[TCP_MAXOLEN];
16696 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
16697 	uint32_t rack_seq;
16698 
16699 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
16700 	unsigned ipsec_optlen = 0;
16701 
16702 #endif
16703 	int32_t idle, sendalot;
16704 	int32_t sub_from_prr = 0;
16705 	volatile int32_t sack_rxmit;
16706 	struct rack_sendmap *rsm = NULL;
16707 	int32_t tso, mtu;
16708 	struct tcpopt to;
16709 	int32_t slot = 0;
16710 	int32_t sup_rack = 0;
16711 	uint32_t cts, ms_cts, delayed, early;
16712 	uint16_t add_flag = RACK_SENT_SP;
16713 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
16714 	uint8_t hpts_calling,  doing_tlp = 0;
16715 	uint32_t cwnd_to_use, pace_max_seg;
16716 	int32_t do_a_prefetch = 0;
16717 	int32_t prefetch_rsm = 0;
16718 	int32_t orig_len = 0;
16719 	struct timeval tv;
16720 	int32_t prefetch_so_done = 0;
16721 	struct tcp_log_buffer *lgb;
16722 	struct inpcb *inp = tptoinpcb(tp);
16723 	struct sockbuf *sb;
16724 	uint64_t ts_val = 0;
16725 #ifdef TCP_ACCOUNTING
16726 	uint64_t crtsc;
16727 #endif
16728 #ifdef INET6
16729 	struct ip6_hdr *ip6 = NULL;
16730 	int32_t isipv6;
16731 #endif
16732 	bool hw_tls = false;
16733 
16734 	NET_EPOCH_ASSERT();
16735 	INP_WLOCK_ASSERT(inp);
16736 
16737 	/* setup and take the cache hits here */
16738 	rack = (struct tcp_rack *)tp->t_fb_ptr;
16739 #ifdef TCP_ACCOUNTING
16740 	sched_pin();
16741 	ts_val = get_cyclecount();
16742 #endif
16743 	hpts_calling = inp->inp_hpts_calls;
16744 #ifdef TCP_OFFLOAD
16745 	if (tp->t_flags & TF_TOE) {
16746 #ifdef TCP_ACCOUNTING
16747 		sched_unpin();
16748 #endif
16749 		return (tcp_offload_output(tp));
16750 	}
16751 #endif
16752 	/*
16753 	 * For TFO connections in SYN_RECEIVED, only allow the initial
16754 	 * SYN|ACK and those sent by the retransmit timer.
16755 	 */
16756 	if (IS_FASTOPEN(tp->t_flags) &&
16757 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
16758 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
16759 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
16760 #ifdef TCP_ACCOUNTING
16761 		sched_unpin();
16762 #endif
16763 		return (0);
16764 	}
16765 #ifdef INET6
16766 	if (rack->r_state) {
16767 		/* Use the cache line loaded if possible */
16768 		isipv6 = rack->r_is_v6;
16769 	} else {
16770 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
16771 	}
16772 #endif
16773 	early = 0;
16774 	cts = tcp_get_usecs(&tv);
16775 	ms_cts = tcp_tv_to_mssectick(&tv);
16776 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
16777 	    tcp_in_hpts(rack->rc_inp)) {
16778 		/*
16779 		 * We are on the hpts for some timer but not hptsi output.
16780 		 * Remove from the hpts unconditionally.
16781 		 */
16782 		rack_timer_cancel(tp, rack, cts, __LINE__);
16783 	}
16784 	/* Are we pacing and late? */
16785 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16786 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
16787 		/* We are delayed */
16788 		delayed = cts - rack->r_ctl.rc_last_output_to;
16789 	} else {
16790 		delayed = 0;
16791 	}
16792 	/* Do the timers, which may override the pacer */
16793 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
16794 		int retval;
16795 
16796 		retval = rack_process_timers(tp, rack, cts, hpts_calling,
16797 		    &doing_tlp);
16798 		if (retval != 0) {
16799 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
16800 #ifdef TCP_ACCOUNTING
16801 			sched_unpin();
16802 #endif
16803 			/*
16804 			 * If timers want tcp_drop(), then pass error out,
16805 			 * otherwise suppress it.
16806 			 */
16807 			return (retval < 0 ? retval : 0);
16808 		}
16809 	}
16810 	if (rack->rc_in_persist) {
16811 		if (tcp_in_hpts(rack->rc_inp) == 0) {
16812 			/* Timer is not running */
16813 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16814 		}
16815 #ifdef TCP_ACCOUNTING
16816 		sched_unpin();
16817 #endif
16818 		return (0);
16819 	}
16820 	if ((rack->rc_ack_required == 1) &&
16821 	    (rack->r_timer_override == 0)){
16822 		/* A timeout occurred and no ack has arrived */
16823 		if (tcp_in_hpts(rack->rc_inp) == 0) {
16824 			/* Timer is not running */
16825 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
16826 		}
16827 #ifdef TCP_ACCOUNTING
16828 		sched_unpin();
16829 #endif
16830 		return (0);
16831 	}
16832 	if ((rack->r_timer_override) ||
16833 	    (rack->rc_ack_can_sendout_data) ||
16834 	    (delayed) ||
16835 	    (tp->t_state < TCPS_ESTABLISHED)) {
16836 		rack->rc_ack_can_sendout_data = 0;
16837 		if (tcp_in_hpts(rack->rc_inp))
16838 			tcp_hpts_remove(rack->rc_inp);
16839 	} else if (tcp_in_hpts(rack->rc_inp)) {
16840 		/*
16841 		 * On the hpts you can't pass even if ACKNOW is on, we will
16842 		 * when the hpts fires.
16843 		 */
16844 #ifdef TCP_ACCOUNTING
16845 		crtsc = get_cyclecount();
16846 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16847 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
16848 		}
16849 		counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val));
16850 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16851 			tp->tcp_cnt_counters[SND_BLOCKED]++;
16852 		}
16853 		counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1);
16854 		sched_unpin();
16855 #endif
16856 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
16857 		return (0);
16858 	}
16859 	rack->rc_inp->inp_hpts_calls = 0;
16860 	/* Finish out both pacing early and late accounting */
16861 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16862 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
16863 		early = rack->r_ctl.rc_last_output_to - cts;
16864 	} else
16865 		early = 0;
16866 	if (delayed) {
16867 		rack->r_ctl.rc_agg_delayed += delayed;
16868 		rack->r_late = 1;
16869 	} else if (early) {
16870 		rack->r_ctl.rc_agg_early += early;
16871 		rack->r_early = 1;
16872 	}
16873 	/* Now that early/late accounting is done turn off the flag */
16874 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
16875 	rack->r_wanted_output = 0;
16876 	rack->r_timer_override = 0;
16877 	if ((tp->t_state != rack->r_state) &&
16878 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
16879 		rack_set_state(tp, rack);
16880 	}
16881 	if ((rack->r_fast_output) &&
16882 	    (doing_tlp == 0) &&
16883 	    (tp->rcv_numsacks == 0)) {
16884 		int ret;
16885 
16886 		error = 0;
16887 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
16888 		if (ret >= 0)
16889 			return(ret);
16890 		else if (error) {
16891 			inp = rack->rc_inp;
16892 			so = inp->inp_socket;
16893 			sb = &so->so_snd;
16894 			goto nomore;
16895 		}
16896 	}
16897 	inp = rack->rc_inp;
16898 	/*
16899 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
16900 	 * only allow the initial SYN or SYN|ACK and those sent
16901 	 * by the retransmit timer.
16902 	 */
16903 	if (IS_FASTOPEN(tp->t_flags) &&
16904 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
16905 	     (tp->t_state == TCPS_SYN_SENT)) &&
16906 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
16907 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
16908 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16909 		so = inp->inp_socket;
16910 		sb = &so->so_snd;
16911 		goto just_return_nolock;
16912 	}
16913 	/*
16914 	 * Determine length of data that should be transmitted, and flags
16915 	 * that will be used. If there is some data or critical controls
16916 	 * (SYN, RST) to send, then transmit; otherwise, investigate
16917 	 * further.
16918 	 */
16919 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
16920 	if (tp->t_idle_reduce) {
16921 		if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
16922 			rack_cc_after_idle(rack, tp);
16923 	}
16924 	tp->t_flags &= ~TF_LASTIDLE;
16925 	if (idle) {
16926 		if (tp->t_flags & TF_MORETOCOME) {
16927 			tp->t_flags |= TF_LASTIDLE;
16928 			idle = 0;
16929 		}
16930 	}
16931 	if ((tp->snd_una == tp->snd_max) &&
16932 	    rack->r_ctl.rc_went_idle_time &&
16933 	    TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
16934 		idle = cts - rack->r_ctl.rc_went_idle_time;
16935 		if (idle > rack_min_probertt_hold) {
16936 			/* Count as a probe rtt */
16937 			if (rack->in_probe_rtt == 0) {
16938 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
16939 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
16940 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
16941 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
16942 			} else {
16943 				rack_exit_probertt(rack, cts);
16944 			}
16945 		}
16946 		idle = 0;
16947 	}
16948 	if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED))
16949 		rack_init_fsb_block(tp, rack);
16950 again:
16951 	/*
16952 	 * If we've recently taken a timeout, snd_max will be greater than
16953 	 * snd_nxt.  There may be SACK information that allows us to avoid
16954 	 * resending already delivered data.  Adjust snd_nxt accordingly.
16955 	 */
16956 	sendalot = 0;
16957 	cts = tcp_get_usecs(&tv);
16958 	ms_cts = tcp_tv_to_mssectick(&tv);
16959 	tso = 0;
16960 	mtu = 0;
16961 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
16962 	minseg = segsiz;
16963 	if (rack->r_ctl.rc_pace_max_segs == 0)
16964 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
16965 	else
16966 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
16967 	sb_offset = tp->snd_max - tp->snd_una;
16968 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
16969 	flags = tcp_outflags[tp->t_state];
16970 	while (rack->rc_free_cnt < rack_free_cache) {
16971 		rsm = rack_alloc(rack);
16972 		if (rsm == NULL) {
16973 			if (inp->inp_hpts_calls)
16974 				/* Retry in a ms */
16975 				slot = (1 * HPTS_USEC_IN_MSEC);
16976 			so = inp->inp_socket;
16977 			sb = &so->so_snd;
16978 			goto just_return_nolock;
16979 		}
16980 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
16981 		rack->rc_free_cnt++;
16982 		rsm = NULL;
16983 	}
16984 	if (inp->inp_hpts_calls)
16985 		inp->inp_hpts_calls = 0;
16986 	sack_rxmit = 0;
16987 	len = 0;
16988 	rsm = NULL;
16989 	if (flags & TH_RST) {
16990 		SOCKBUF_LOCK(&inp->inp_socket->so_snd);
16991 		so = inp->inp_socket;
16992 		sb = &so->so_snd;
16993 		goto send;
16994 	}
16995 	if (rack->r_ctl.rc_resend) {
16996 		/* Retransmit timer */
16997 		rsm = rack->r_ctl.rc_resend;
16998 		rack->r_ctl.rc_resend = NULL;
16999 		len = rsm->r_end - rsm->r_start;
17000 		sack_rxmit = 1;
17001 		sendalot = 0;
17002 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
17003 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
17004 			 __func__, __LINE__,
17005 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
17006 		sb_offset = rsm->r_start - tp->snd_una;
17007 		if (len >= segsiz)
17008 			len = segsiz;
17009 	} else if (rack->r_collapse_point_valid &&
17010 		   ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
17011 		/*
17012 		 * If an RSM is returned then enough time has passed
17013 		 * for us to retransmit it. Move up the collapse point,
17014 		 * since this rsm has its chance to retransmit now.
17015 		 */
17016 		rack_trace_point(rack, RACK_TP_COLLAPSED_RXT);
17017 		rack->r_ctl.last_collapse_point = rsm->r_end;
17018 		/* Are we done? */
17019 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
17020 			    rack->r_ctl.high_collapse_point))
17021 			rack->r_collapse_point_valid = 0;
17022 		sack_rxmit = 1;
17023 		/* We are not doing a TLP */
17024 		doing_tlp = 0;
17025 		len = rsm->r_end - rsm->r_start;
17026 		sb_offset = rsm->r_start - tp->snd_una;
17027 		sendalot = 0;
17028 		if ((rack->full_size_rxt == 0) &&
17029 		    (rack->shape_rxt_to_pacing_min == 0) &&
17030 		    (len >= segsiz))
17031 			len = segsiz;
17032 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
17033 		/* We have a retransmit that takes precedence */
17034 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
17035 		    ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
17036 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
17037 			/* Enter recovery if not induced by a time-out */
17038 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
17039 		}
17040 #ifdef INVARIANTS
17041 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
17042 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
17043 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
17044 		}
17045 #endif
17046 		len = rsm->r_end - rsm->r_start;
17047 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
17048 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
17049 			 __func__, __LINE__,
17050 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
17051 		sb_offset = rsm->r_start - tp->snd_una;
17052 		sendalot = 0;
17053 		if (len >= segsiz)
17054 			len = segsiz;
17055 		if (len > 0) {
17056 			sack_rxmit = 1;
17057 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
17058 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
17059 			    min(len, segsiz));
17060 		}
17061 	} else if (rack->r_ctl.rc_tlpsend) {
17062 		/* Tail loss probe */
17063 		long cwin;
17064 		long tlen;
17065 
17066 		/*
17067 		 * Check if we can do a TLP with a RACK'd packet
17068 		 * this can happen if we are not doing the rack
17069 		 * cheat and we skipped to a TLP and it
17070 		 * went off.
17071 		 */
17072 		rsm = rack->r_ctl.rc_tlpsend;
17073 		/* We are doing a TLP make sure the flag is preent */
17074 		rsm->r_flags |= RACK_TLP;
17075 		rack->r_ctl.rc_tlpsend = NULL;
17076 		sack_rxmit = 1;
17077 		tlen = rsm->r_end - rsm->r_start;
17078 		if (tlen > segsiz)
17079 			tlen = segsiz;
17080 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
17081 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
17082 			 __func__, __LINE__,
17083 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
17084 		sb_offset = rsm->r_start - tp->snd_una;
17085 		cwin = min(tp->snd_wnd, tlen);
17086 		len = cwin;
17087 	}
17088 	if (rack->r_must_retran &&
17089 	    (doing_tlp == 0) &&
17090 	    (SEQ_GT(tp->snd_max, tp->snd_una)) &&
17091 	    (rsm == NULL)) {
17092 		/*
17093 		 * There are two different ways that we
17094 		 * can get into this block:
17095 		 * a) This is a non-sack connection, we had a time-out
17096 		 *    and thus r_must_retran was set and everything
17097 		 *    left outstanding as been marked for retransmit.
17098 		 * b) The MTU of the path shrank, so that everything
17099 		 *    was marked to be retransmitted with the smaller
17100 		 *    mtu and r_must_retran was set.
17101 		 *
17102 		 * This means that we expect the sendmap (outstanding)
17103 		 * to all be marked must. We can use the tmap to
17104 		 * look at them.
17105 		 *
17106 		 */
17107 		int sendwin, flight;
17108 
17109 		sendwin = min(tp->snd_wnd, tp->snd_cwnd);
17110 		flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
17111 		if (flight >= sendwin) {
17112 			/*
17113 			 * We can't send yet.
17114 			 */
17115 			so = inp->inp_socket;
17116 			sb = &so->so_snd;
17117 			goto just_return_nolock;
17118 		}
17119 		/*
17120 		 * This is the case a/b mentioned above. All
17121 		 * outstanding/not-acked should be marked.
17122 		 * We can use the tmap to find them.
17123 		 */
17124 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17125 		if (rsm == NULL) {
17126 			/* TSNH */
17127 			rack->r_must_retran = 0;
17128 			rack->r_ctl.rc_out_at_rto = 0;
17129 			so = inp->inp_socket;
17130 			sb = &so->so_snd;
17131 			goto just_return_nolock;
17132 		}
17133 		if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
17134 			/*
17135 			 * The first one does not have the flag, did we collapse
17136 			 * further up in our list?
17137 			 */
17138 			rack->r_must_retran = 0;
17139 			rack->r_ctl.rc_out_at_rto = 0;
17140 			rsm = NULL;
17141 			sack_rxmit = 0;
17142 		} else {
17143 			sack_rxmit = 1;
17144 			len = rsm->r_end - rsm->r_start;
17145 			sb_offset = rsm->r_start - tp->snd_una;
17146 			sendalot = 0;
17147 			if ((rack->full_size_rxt == 0) &&
17148 			    (rack->shape_rxt_to_pacing_min == 0) &&
17149 			    (len >= segsiz))
17150 				len = segsiz;
17151 			/*
17152 			 * Delay removing the flag RACK_MUST_RXT so
17153 			 * that the fastpath for retransmit will
17154 			 * work with this rsm.
17155 			 */
17156 		}
17157 	}
17158 	/*
17159 	 * Enforce a connection sendmap count limit if set
17160 	 * as long as we are not retransmiting.
17161 	 */
17162 	if ((rsm == NULL) &&
17163 	    (rack->do_detection == 0) &&
17164 	    (V_tcp_map_entries_limit > 0) &&
17165 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
17166 		counter_u64_add(rack_to_alloc_limited, 1);
17167 		if (!rack->alloc_limit_reported) {
17168 			rack->alloc_limit_reported = 1;
17169 			counter_u64_add(rack_alloc_limited_conns, 1);
17170 		}
17171 		so = inp->inp_socket;
17172 		sb = &so->so_snd;
17173 		goto just_return_nolock;
17174 	}
17175 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
17176 		/* we are retransmitting the fin */
17177 		len--;
17178 		if (len) {
17179 			/*
17180 			 * When retransmitting data do *not* include the
17181 			 * FIN. This could happen from a TLP probe.
17182 			 */
17183 			flags &= ~TH_FIN;
17184 		}
17185 	}
17186 	if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo &&
17187 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
17188 		int ret;
17189 
17190 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
17191 		if (ret == 0)
17192 			return (0);
17193 	}
17194 	so = inp->inp_socket;
17195 	sb = &so->so_snd;
17196 	if (do_a_prefetch == 0) {
17197 		kern_prefetch(sb, &do_a_prefetch);
17198 		do_a_prefetch = 1;
17199 	}
17200 #ifdef NETFLIX_SHARED_CWND
17201 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
17202 	    rack->rack_enable_scwnd) {
17203 		/* We are doing cwnd sharing */
17204 		if (rack->gp_ready &&
17205 		    (rack->rack_attempted_scwnd == 0) &&
17206 		    (rack->r_ctl.rc_scw == NULL) &&
17207 		    tp->t_lib) {
17208 			/* The pcbid is in, lets make an attempt */
17209 			counter_u64_add(rack_try_scwnd, 1);
17210 			rack->rack_attempted_scwnd = 1;
17211 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
17212 								   &rack->r_ctl.rc_scw_index,
17213 								   segsiz);
17214 		}
17215 		if (rack->r_ctl.rc_scw &&
17216 		    (rack->rack_scwnd_is_idle == 1) &&
17217 		    sbavail(&so->so_snd)) {
17218 			/* we are no longer out of data */
17219 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17220 			rack->rack_scwnd_is_idle = 0;
17221 		}
17222 		if (rack->r_ctl.rc_scw) {
17223 			/* First lets update and get the cwnd */
17224 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
17225 								    rack->r_ctl.rc_scw_index,
17226 								    tp->snd_cwnd, tp->snd_wnd, segsiz);
17227 		}
17228 	}
17229 #endif
17230 	/*
17231 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
17232 	 * state flags.
17233 	 */
17234 	if (tp->t_flags & TF_NEEDFIN)
17235 		flags |= TH_FIN;
17236 	if (tp->t_flags & TF_NEEDSYN)
17237 		flags |= TH_SYN;
17238 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
17239 		void *end_rsm;
17240 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
17241 		if (end_rsm)
17242 			kern_prefetch(end_rsm, &prefetch_rsm);
17243 		prefetch_rsm = 1;
17244 	}
17245 	SOCKBUF_LOCK(sb);
17246 	/*
17247 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
17248 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
17249 	 * negative length.  This can also occur when TCP opens up its
17250 	 * congestion window while receiving additional duplicate acks after
17251 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
17252 	 * the fast-retransmit.
17253 	 *
17254 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
17255 	 * set to snd_una, the sb_offset will be 0, and the length may wind
17256 	 * up 0.
17257 	 *
17258 	 * If sack_rxmit is true we are retransmitting from the scoreboard
17259 	 * in which case len is already set.
17260 	 */
17261 	if ((sack_rxmit == 0) &&
17262 	    (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
17263 		uint32_t avail;
17264 
17265 		avail = sbavail(sb);
17266 		if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
17267 			sb_offset = tp->snd_nxt - tp->snd_una;
17268 		else
17269 			sb_offset = 0;
17270 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
17271 			if (rack->r_ctl.rc_tlp_new_data) {
17272 				/* TLP is forcing out new data */
17273 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
17274 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
17275 				}
17276 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
17277 					if (tp->snd_wnd > sb_offset)
17278 						len = tp->snd_wnd - sb_offset;
17279 					else
17280 						len = 0;
17281 				} else {
17282 					len = rack->r_ctl.rc_tlp_new_data;
17283 				}
17284 				rack->r_ctl.rc_tlp_new_data = 0;
17285 			}  else {
17286 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
17287 			}
17288 			if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) {
17289 				/*
17290 				 * For prr=off, we need to send only 1 MSS
17291 				 * at a time. We do this because another sack could
17292 				 * be arriving that causes us to send retransmits and
17293 				 * we don't want to be on a long pace due to a larger send
17294 				 * that keeps us from sending out the retransmit.
17295 				 */
17296 				len = segsiz;
17297 			}
17298 		} else {
17299 			uint32_t outstanding;
17300 			/*
17301 			 * We are inside of a Fast recovery episode, this
17302 			 * is caused by a SACK or 3 dup acks. At this point
17303 			 * we have sent all the retransmissions and we rely
17304 			 * on PRR to dictate what we will send in the form of
17305 			 * new data.
17306 			 */
17307 
17308 			outstanding = tp->snd_max - tp->snd_una;
17309 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
17310 				if (tp->snd_wnd > outstanding) {
17311 					len = tp->snd_wnd - outstanding;
17312 					/* Check to see if we have the data */
17313 					if ((sb_offset + len) > avail) {
17314 						/* It does not all fit */
17315 						if (avail > sb_offset)
17316 							len = avail - sb_offset;
17317 						else
17318 							len = 0;
17319 					}
17320 				} else {
17321 					len = 0;
17322 				}
17323 			} else if (avail > sb_offset) {
17324 				len = avail - sb_offset;
17325 			} else {
17326 				len = 0;
17327 			}
17328 			if (len > 0) {
17329 				if (len > rack->r_ctl.rc_prr_sndcnt) {
17330 					len = rack->r_ctl.rc_prr_sndcnt;
17331 				}
17332 				if (len > 0) {
17333 					sub_from_prr = 1;
17334 				}
17335 			}
17336 			if (len > segsiz) {
17337 				/*
17338 				 * We should never send more than a MSS when
17339 				 * retransmitting or sending new data in prr
17340 				 * mode unless the override flag is on. Most
17341 				 * likely the PRR algorithm is not going to
17342 				 * let us send a lot as well :-)
17343 				 */
17344 				if (rack->r_ctl.rc_prr_sendalot == 0) {
17345 					len = segsiz;
17346 				}
17347 			} else if (len < segsiz) {
17348 				/*
17349 				 * Do we send any? The idea here is if the
17350 				 * send empty's the socket buffer we want to
17351 				 * do it. However if not then lets just wait
17352 				 * for our prr_sndcnt to get bigger.
17353 				 */
17354 				long leftinsb;
17355 
17356 				leftinsb = sbavail(sb) - sb_offset;
17357 				if (leftinsb > len) {
17358 					/* This send does not empty the sb */
17359 					len = 0;
17360 				}
17361 			}
17362 		}
17363 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
17364 		/*
17365 		 * If you have not established
17366 		 * and are not doing FAST OPEN
17367 		 * no data please.
17368 		 */
17369 		if ((sack_rxmit == 0) &&
17370 		    (!IS_FASTOPEN(tp->t_flags))){
17371 			len = 0;
17372 			sb_offset = 0;
17373 		}
17374 	}
17375 	if (prefetch_so_done == 0) {
17376 		kern_prefetch(so, &prefetch_so_done);
17377 		prefetch_so_done = 1;
17378 	}
17379 	/*
17380 	 * Lop off SYN bit if it has already been sent.  However, if this is
17381 	 * SYN-SENT state and if segment contains data and if we don't know
17382 	 * that foreign host supports TAO, suppress sending segment.
17383 	 */
17384 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
17385 	    ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
17386 		/*
17387 		 * When sending additional segments following a TFO SYN|ACK,
17388 		 * do not include the SYN bit.
17389 		 */
17390 		if (IS_FASTOPEN(tp->t_flags) &&
17391 		    (tp->t_state == TCPS_SYN_RECEIVED))
17392 			flags &= ~TH_SYN;
17393 	}
17394 	/*
17395 	 * Be careful not to send data and/or FIN on SYN segments. This
17396 	 * measure is needed to prevent interoperability problems with not
17397 	 * fully conformant TCP implementations.
17398 	 */
17399 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
17400 		len = 0;
17401 		flags &= ~TH_FIN;
17402 	}
17403 	/*
17404 	 * On TFO sockets, ensure no data is sent in the following cases:
17405 	 *
17406 	 *  - When retransmitting SYN|ACK on a passively-created socket
17407 	 *
17408 	 *  - When retransmitting SYN on an actively created socket
17409 	 *
17410 	 *  - When sending a zero-length cookie (cookie request) on an
17411 	 *    actively created socket
17412 	 *
17413 	 *  - When the socket is in the CLOSED state (RST is being sent)
17414 	 */
17415 	if (IS_FASTOPEN(tp->t_flags) &&
17416 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
17417 	     ((tp->t_state == TCPS_SYN_SENT) &&
17418 	      (tp->t_tfo_client_cookie_len == 0)) ||
17419 	     (flags & TH_RST))) {
17420 		sack_rxmit = 0;
17421 		len = 0;
17422 	}
17423 	/* Without fast-open there should never be data sent on a SYN */
17424 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
17425 		tp->snd_nxt = tp->iss;
17426 		len = 0;
17427 	}
17428 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
17429 		/* We only send 1 MSS if we have a DSACK block */
17430 		add_flag |= RACK_SENT_W_DSACK;
17431 		len = segsiz;
17432 	}
17433 	orig_len = len;
17434 	if (len <= 0) {
17435 		/*
17436 		 * If FIN has been sent but not acked, but we haven't been
17437 		 * called to retransmit, len will be < 0.  Otherwise, window
17438 		 * shrank after we sent into it.  If window shrank to 0,
17439 		 * cancel pending retransmit, pull snd_nxt back to (closed)
17440 		 * window, and set the persist timer if it isn't already
17441 		 * going.  If the window didn't close completely, just wait
17442 		 * for an ACK.
17443 		 *
17444 		 * We also do a general check here to ensure that we will
17445 		 * set the persist timer when we have data to send, but a
17446 		 * 0-byte window. This makes sure the persist timer is set
17447 		 * even if the packet hits one of the "goto send" lines
17448 		 * below.
17449 		 */
17450 		len = 0;
17451 		if ((tp->snd_wnd == 0) &&
17452 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17453 		    (tp->snd_una == tp->snd_max) &&
17454 		    (sb_offset < (int)sbavail(sb))) {
17455 			rack_enter_persist(tp, rack, cts);
17456 		}
17457 	} else if ((rsm == NULL) &&
17458 		   (doing_tlp == 0) &&
17459 		   (len < pace_max_seg)) {
17460 		/*
17461 		 * We are not sending a maximum sized segment for
17462 		 * some reason. Should we not send anything (think
17463 		 * sws or persists)?
17464 		 */
17465 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17466 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
17467 		    (len < minseg) &&
17468 		    (len < (int)(sbavail(sb) - sb_offset))) {
17469 			/*
17470 			 * Here the rwnd is less than
17471 			 * the minimum pacing size, this is not a retransmit,
17472 			 * we are established and
17473 			 * the send is not the last in the socket buffer
17474 			 * we send nothing, and we may enter persists
17475 			 * if nothing is outstanding.
17476 			 */
17477 			len = 0;
17478 			if (tp->snd_max == tp->snd_una) {
17479 				/*
17480 				 * Nothing out we can
17481 				 * go into persists.
17482 				 */
17483 				rack_enter_persist(tp, rack, cts);
17484 			}
17485 		     } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
17486 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17487 			   (len < (int)(sbavail(sb) - sb_offset)) &&
17488 			   (len < minseg)) {
17489 			/*
17490 			 * Here we are not retransmitting, and
17491 			 * the cwnd is not so small that we could
17492 			 * not send at least a min size (rxt timer
17493 			 * not having gone off), We have 2 segments or
17494 			 * more already in flight, its not the tail end
17495 			 * of the socket buffer  and the cwnd is blocking
17496 			 * us from sending out a minimum pacing segment size.
17497 			 * Lets not send anything.
17498 			 */
17499 			len = 0;
17500 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
17501 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
17502 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
17503 			   (len < (int)(sbavail(sb) - sb_offset)) &&
17504 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
17505 			/*
17506 			 * Here we have a send window but we have
17507 			 * filled it up and we can't send another pacing segment.
17508 			 * We also have in flight more than 2 segments
17509 			 * and we are not completing the sb i.e. we allow
17510 			 * the last bytes of the sb to go out even if
17511 			 * its not a full pacing segment.
17512 			 */
17513 			len = 0;
17514 		} else if ((rack->r_ctl.crte != NULL) &&
17515 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
17516 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
17517 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
17518 			   (len < (int)(sbavail(sb) - sb_offset))) {
17519 			/*
17520 			 * Here we are doing hardware pacing, this is not a TLP,
17521 			 * we are not sending a pace max segment size, there is rwnd
17522 			 * room to send at least N pace_max_seg, the cwnd is greater
17523 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
17524 			 * more segments in flight and its not the tail of the socket buffer.
17525 			 *
17526 			 * We don't want to send instead we need to get more ack's in to
17527 			 * allow us to send a full pacing segment. Normally, if we are pacing
17528 			 * about the right speed, we should have finished our pacing
17529 			 * send as most of the acks have come back if we are at the
17530 			 * right rate. This is a bit fuzzy since return path delay
17531 			 * can delay the acks, which is why we want to make sure we
17532 			 * have cwnd space to have a bit more than a max pace segments in flight.
17533 			 *
17534 			 * If we have not gotten our acks back we are pacing at too high a
17535 			 * rate delaying will not hurt and will bring our GP estimate down by
17536 			 * injecting the delay. If we don't do this we will send
17537 			 * 2 MSS out in response to the acks being clocked in which
17538 			 * defeats the point of hw-pacing (i.e. to help us get
17539 			 * larger TSO's out).
17540 			 */
17541 			len = 0;
17542 
17543 		}
17544 
17545 	}
17546 	/* len will be >= 0 after this point. */
17547 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
17548 	rack_sndbuf_autoscale(rack);
17549 	/*
17550 	 * Decide if we can use TCP Segmentation Offloading (if supported by
17551 	 * hardware).
17552 	 *
17553 	 * TSO may only be used if we are in a pure bulk sending state.  The
17554 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
17555 	 * options prevent using TSO.  With TSO the TCP header is the same
17556 	 * (except for the sequence number) for all generated packets.  This
17557 	 * makes it impossible to transmit any options which vary per
17558 	 * generated segment or packet.
17559 	 *
17560 	 * IPv4 handling has a clear separation of ip options and ip header
17561 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
17562 	 * the right thing below to provide length of just ip options and thus
17563 	 * checking for ipoptlen is enough to decide if ip options are present.
17564 	 */
17565 	ipoptlen = 0;
17566 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17567 	/*
17568 	 * Pre-calculate here as we save another lookup into the darknesses
17569 	 * of IPsec that way and can actually decide if TSO is ok.
17570 	 */
17571 #ifdef INET6
17572 	if (isipv6 && IPSEC_ENABLED(ipv6))
17573 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
17574 #ifdef INET
17575 	else
17576 #endif
17577 #endif				/* INET6 */
17578 #ifdef INET
17579 		if (IPSEC_ENABLED(ipv4))
17580 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
17581 #endif				/* INET */
17582 #endif
17583 
17584 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
17585 	ipoptlen += ipsec_optlen;
17586 #endif
17587 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
17588 	    (tp->t_port == 0) &&
17589 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
17590 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
17591 	    ipoptlen == 0)
17592 		tso = 1;
17593 	{
17594 		uint32_t outstanding __unused;
17595 
17596 		outstanding = tp->snd_max - tp->snd_una;
17597 		if (tp->t_flags & TF_SENTFIN) {
17598 			/*
17599 			 * If we sent a fin, snd_max is 1 higher than
17600 			 * snd_una
17601 			 */
17602 			outstanding--;
17603 		}
17604 		if (sack_rxmit) {
17605 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
17606 				flags &= ~TH_FIN;
17607 		} else {
17608 			if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
17609 				   sbused(sb)))
17610 				flags &= ~TH_FIN;
17611 		}
17612 	}
17613 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
17614 	    (long)TCP_MAXWIN << tp->rcv_scale);
17615 
17616 	/*
17617 	 * Sender silly window avoidance.   We transmit under the following
17618 	 * conditions when len is non-zero:
17619 	 *
17620 	 * - We have a full segment (or more with TSO) - This is the last
17621 	 * buffer in a write()/send() and we are either idle or running
17622 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
17623 	 * then 1/2 the maximum send window's worth of data (receiver may be
17624 	 * limited the window size) - we need to retransmit
17625 	 */
17626 	if (len) {
17627 		if (len >= segsiz) {
17628 			goto send;
17629 		}
17630 		/*
17631 		 * NOTE! on localhost connections an 'ack' from the remote
17632 		 * end may occur synchronously with the output and cause us
17633 		 * to flush a buffer queued with moretocome.  XXX
17634 		 *
17635 		 */
17636 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
17637 		    (idle || (tp->t_flags & TF_NODELAY)) &&
17638 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17639 		    (tp->t_flags & TF_NOPUSH) == 0) {
17640 			pass = 2;
17641 			goto send;
17642 		}
17643 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
17644 			pass = 22;
17645 			goto send;
17646 		}
17647 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
17648 			pass = 4;
17649 			goto send;
17650 		}
17651 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {	/* retransmit case */
17652 			pass = 5;
17653 			goto send;
17654 		}
17655 		if (sack_rxmit) {
17656 			pass = 6;
17657 			goto send;
17658 		}
17659 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
17660 		    (ctf_outstanding(tp) < (segsiz * 2))) {
17661 			/*
17662 			 * We have less than two MSS outstanding (delayed ack)
17663 			 * and our rwnd will not let us send a full sized
17664 			 * MSS. Lets go ahead and let this small segment
17665 			 * out because we want to try to have at least two
17666 			 * packets inflight to not be caught by delayed ack.
17667 			 */
17668 			pass = 12;
17669 			goto send;
17670 		}
17671 	}
17672 	/*
17673 	 * Sending of standalone window updates.
17674 	 *
17675 	 * Window updates are important when we close our window due to a
17676 	 * full socket buffer and are opening it again after the application
17677 	 * reads data from it.  Once the window has opened again and the
17678 	 * remote end starts to send again the ACK clock takes over and
17679 	 * provides the most current window information.
17680 	 *
17681 	 * We must avoid the silly window syndrome whereas every read from
17682 	 * the receive buffer, no matter how small, causes a window update
17683 	 * to be sent.  We also should avoid sending a flurry of window
17684 	 * updates when the socket buffer had queued a lot of data and the
17685 	 * application is doing small reads.
17686 	 *
17687 	 * Prevent a flurry of pointless window updates by only sending an
17688 	 * update when we can increase the advertized window by more than
17689 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
17690 	 * full or is very small be more aggressive and send an update
17691 	 * whenever we can increase by two mss sized segments. In all other
17692 	 * situations the ACK's to new incoming data will carry further
17693 	 * window increases.
17694 	 *
17695 	 * Don't send an independent window update if a delayed ACK is
17696 	 * pending (it will get piggy-backed on it) or the remote side
17697 	 * already has done a half-close and won't send more data.  Skip
17698 	 * this if the connection is in T/TCP half-open state.
17699 	 */
17700 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
17701 	    !(tp->t_flags & TF_DELACK) &&
17702 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
17703 		/*
17704 		 * "adv" is the amount we could increase the window, taking
17705 		 * into account that we are limited by TCP_MAXWIN <<
17706 		 * tp->rcv_scale.
17707 		 */
17708 		int32_t adv;
17709 		int oldwin;
17710 
17711 		adv = recwin;
17712 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
17713 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
17714 			if (adv > oldwin)
17715 			    adv -= oldwin;
17716 			else {
17717 				/* We can't increase the window */
17718 				adv = 0;
17719 			}
17720 		} else
17721 			oldwin = 0;
17722 
17723 		/*
17724 		 * If the new window size ends up being the same as or less
17725 		 * than the old size when it is scaled, then don't force
17726 		 * a window update.
17727 		 */
17728 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
17729 			goto dontupdate;
17730 
17731 		if (adv >= (int32_t)(2 * segsiz) &&
17732 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
17733 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
17734 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
17735 			pass = 7;
17736 			goto send;
17737 		}
17738 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
17739 			pass = 23;
17740 			goto send;
17741 		}
17742 	}
17743 dontupdate:
17744 
17745 	/*
17746 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
17747 	 * is also a catch-all for the retransmit timer timeout case.
17748 	 */
17749 	if (tp->t_flags & TF_ACKNOW) {
17750 		pass = 8;
17751 		goto send;
17752 	}
17753 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
17754 		pass = 9;
17755 		goto send;
17756 	}
17757 	/*
17758 	 * If our state indicates that FIN should be sent and we have not
17759 	 * yet done so, then we need to send.
17760 	 */
17761 	if ((flags & TH_FIN) &&
17762 	    (tp->snd_nxt == tp->snd_una)) {
17763 		pass = 11;
17764 		goto send;
17765 	}
17766 	/*
17767 	 * No reason to send a segment, just return.
17768 	 */
17769 just_return:
17770 	SOCKBUF_UNLOCK(sb);
17771 just_return_nolock:
17772 	{
17773 		int app_limited = CTF_JR_SENT_DATA;
17774 
17775 		if (tot_len_this_send > 0) {
17776 			/* Make sure snd_nxt is up to max */
17777 			rack->r_ctl.fsb.recwin = recwin;
17778 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
17779 			if ((error == 0) &&
17780 			    rack_use_rfo &&
17781 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
17782 			    (ipoptlen == 0) &&
17783 			    (tp->snd_nxt == tp->snd_max) &&
17784 			    (tp->rcv_numsacks == 0) &&
17785 			    rack->r_fsb_inited &&
17786 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
17787 			    (rack->r_must_retran == 0) &&
17788 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
17789 			    (len > 0) && (orig_len > 0) &&
17790 			    (orig_len > len) &&
17791 			    ((orig_len - len) >= segsiz) &&
17792 			    ((optlen == 0) ||
17793 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
17794 				/* We can send at least one more MSS using our fsb */
17795 
17796 				rack->r_fast_output = 1;
17797 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
17798 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
17799 				rack->r_ctl.fsb.tcp_flags = flags;
17800 				rack->r_ctl.fsb.left_to_send = orig_len - len;
17801 				if (hw_tls)
17802 					rack->r_ctl.fsb.hw_tls = 1;
17803 				else
17804 					rack->r_ctl.fsb.hw_tls = 0;
17805 				KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
17806 					("rack:%p left_to_send:%u sbavail:%u out:%u",
17807 					rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
17808 					 (tp->snd_max - tp->snd_una)));
17809 				if (rack->r_ctl.fsb.left_to_send < segsiz)
17810 					rack->r_fast_output = 0;
17811 				else {
17812 					if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
17813 						rack->r_ctl.fsb.rfo_apply_push = 1;
17814 					else
17815 						rack->r_ctl.fsb.rfo_apply_push = 0;
17816 				}
17817 			} else
17818 				rack->r_fast_output = 0;
17819 
17820 
17821 			rack_log_fsb(rack, tp, so, flags,
17822 				     ipoptlen, orig_len, len, 0,
17823 				     1, optlen, __LINE__, 1);
17824 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
17825 				tp->snd_nxt = tp->snd_max;
17826 		} else {
17827 			int end_window = 0;
17828 			uint32_t seq = tp->gput_ack;
17829 
17830 			rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17831 			if (rsm) {
17832 				/*
17833 				 * Mark the last sent that we just-returned (hinting
17834 				 * that delayed ack may play a role in any rtt measurement).
17835 				 */
17836 				rsm->r_just_ret = 1;
17837 			}
17838 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
17839 			rack->r_ctl.rc_agg_delayed = 0;
17840 			rack->r_early = 0;
17841 			rack->r_late = 0;
17842 			rack->r_ctl.rc_agg_early = 0;
17843 			if ((ctf_outstanding(tp) +
17844 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
17845 				 minseg)) >= tp->snd_wnd) {
17846 				/* We are limited by the rwnd */
17847 				app_limited = CTF_JR_RWND_LIMITED;
17848 				if (IN_FASTRECOVERY(tp->t_flags))
17849 				    rack->r_ctl.rc_prr_sndcnt = 0;
17850 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
17851 				/* We are limited by whats available -- app limited */
17852 				app_limited = CTF_JR_APP_LIMITED;
17853 				if (IN_FASTRECOVERY(tp->t_flags))
17854 				    rack->r_ctl.rc_prr_sndcnt = 0;
17855 			} else if ((idle == 0) &&
17856 				   ((tp->t_flags & TF_NODELAY) == 0) &&
17857 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
17858 				   (len < segsiz)) {
17859 				/*
17860 				 * No delay is not on and the
17861 				 * user is sending less than 1MSS. This
17862 				 * brings out SWS avoidance so we
17863 				 * don't send. Another app-limited case.
17864 				 */
17865 				app_limited = CTF_JR_APP_LIMITED;
17866 			} else if (tp->t_flags & TF_NOPUSH) {
17867 				/*
17868 				 * The user has requested no push of
17869 				 * the last segment and we are
17870 				 * at the last segment. Another app
17871 				 * limited case.
17872 				 */
17873 				app_limited = CTF_JR_APP_LIMITED;
17874 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
17875 				/* Its the cwnd */
17876 				app_limited = CTF_JR_CWND_LIMITED;
17877 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
17878 				   (rack->rack_no_prr == 0) &&
17879 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
17880 				app_limited = CTF_JR_PRR;
17881 			} else {
17882 				/* Now why here are we not sending? */
17883 #ifdef NOW
17884 #ifdef INVARIANTS
17885 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
17886 #endif
17887 #endif
17888 				app_limited = CTF_JR_ASSESSING;
17889 			}
17890 			/*
17891 			 * App limited in some fashion, for our pacing GP
17892 			 * measurements we don't want any gap (even cwnd).
17893 			 * Close  down the measurement window.
17894 			 */
17895 			if (rack_cwnd_block_ends_measure &&
17896 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
17897 			     (app_limited == CTF_JR_PRR))) {
17898 				/*
17899 				 * The reason we are not sending is
17900 				 * the cwnd (or prr). We have been configured
17901 				 * to end the measurement window in
17902 				 * this case.
17903 				 */
17904 				end_window = 1;
17905 			} else if (rack_rwnd_block_ends_measure &&
17906 				   (app_limited == CTF_JR_RWND_LIMITED)) {
17907 				/*
17908 				 * We are rwnd limited and have been
17909 				 * configured to end the measurement
17910 				 * window in this case.
17911 				 */
17912 				end_window = 1;
17913 			} else if (app_limited == CTF_JR_APP_LIMITED) {
17914 				/*
17915 				 * A true application limited period, we have
17916 				 * ran out of data.
17917 				 */
17918 				end_window = 1;
17919 			} else if (app_limited == CTF_JR_ASSESSING) {
17920 				/*
17921 				 * In the assessing case we hit the end of
17922 				 * the if/else and had no known reason
17923 				 * This will panic us under invariants..
17924 				 *
17925 				 * If we get this out in logs we need to
17926 				 * investagate which reason we missed.
17927 				 */
17928 				end_window = 1;
17929 			}
17930 			if (end_window) {
17931 				uint8_t log = 0;
17932 
17933 				/* Adjust the Gput measurement */
17934 				if ((tp->t_flags & TF_GPUTINPROG) &&
17935 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
17936 					tp->gput_ack = tp->snd_max;
17937 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
17938 						/*
17939 						 * There is not enough to measure.
17940 						 */
17941 						tp->t_flags &= ~TF_GPUTINPROG;
17942 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
17943 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
17944 									   tp->gput_seq,
17945 									   0, 0, 18, __LINE__, NULL, 0);
17946 					} else
17947 						log = 1;
17948 				}
17949 				/* Mark the last packet has app limited */
17950 				rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree);
17951 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
17952 					if (rack->r_ctl.rc_app_limited_cnt == 0)
17953 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
17954 					else {
17955 						/*
17956 						 * Go out to the end app limited and mark
17957 						 * this new one as next and move the end_appl up
17958 						 * to this guy.
17959 						 */
17960 						if (rack->r_ctl.rc_end_appl)
17961 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
17962 						rack->r_ctl.rc_end_appl = rsm;
17963 					}
17964 					rsm->r_flags |= RACK_APP_LIMITED;
17965 					rack->r_ctl.rc_app_limited_cnt++;
17966 				}
17967 				if (log)
17968 					rack_log_pacing_delay_calc(rack,
17969 								   rack->r_ctl.rc_app_limited_cnt, seq,
17970 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
17971 			}
17972 		}
17973 		/* Check if we need to go into persists or not */
17974 		if ((tp->snd_max == tp->snd_una) &&
17975 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
17976 		    sbavail(sb) &&
17977 		    (sbavail(sb) > tp->snd_wnd) &&
17978 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
17979 			/* Yes lets make sure to move to persist before timer-start */
17980 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime);
17981 		}
17982 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
17983 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
17984 	}
17985 #ifdef NETFLIX_SHARED_CWND
17986 	if ((sbavail(sb) == 0) &&
17987 	    rack->r_ctl.rc_scw) {
17988 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
17989 		rack->rack_scwnd_is_idle = 1;
17990 	}
17991 #endif
17992 #ifdef TCP_ACCOUNTING
17993 	if (tot_len_this_send > 0) {
17994 		crtsc = get_cyclecount();
17995 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17996 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
17997 		}
17998 		counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
17999 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18000 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
18001 		}
18002 		counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val));
18003 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18004 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
18005 		}
18006 		counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz));
18007 	} else {
18008 		crtsc = get_cyclecount();
18009 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18010 			tp->tcp_cnt_counters[SND_LIMITED]++;
18011 		}
18012 		counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1);
18013 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18014 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
18015 		}
18016 		counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val));
18017 	}
18018 	sched_unpin();
18019 #endif
18020 	return (0);
18021 
18022 send:
18023 	if (rsm || sack_rxmit)
18024 		counter_u64_add(rack_nfto_resend, 1);
18025 	else
18026 		counter_u64_add(rack_non_fto_send, 1);
18027 	if ((flags & TH_FIN) &&
18028 	    sbavail(sb)) {
18029 		/*
18030 		 * We do not transmit a FIN
18031 		 * with data outstanding. We
18032 		 * need to make it so all data
18033 		 * is acked first.
18034 		 */
18035 		flags &= ~TH_FIN;
18036 	}
18037 	/* Enforce stack imposed max seg size if we have one */
18038 	if (rack->r_ctl.rc_pace_max_segs &&
18039 	    (len > rack->r_ctl.rc_pace_max_segs)) {
18040 		mark = 1;
18041 		len = rack->r_ctl.rc_pace_max_segs;
18042 	}
18043 	SOCKBUF_LOCK_ASSERT(sb);
18044 	if (len > 0) {
18045 		if (len >= segsiz)
18046 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
18047 		else
18048 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
18049 	}
18050 	/*
18051 	 * Before ESTABLISHED, force sending of initial options unless TCP
18052 	 * set not to do any options. NOTE: we assume that the IP/TCP header
18053 	 * plus TCP options always fit in a single mbuf, leaving room for a
18054 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
18055 	 * + optlen <= MCLBYTES
18056 	 */
18057 	optlen = 0;
18058 #ifdef INET6
18059 	if (isipv6)
18060 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
18061 	else
18062 #endif
18063 		hdrlen = sizeof(struct tcpiphdr);
18064 
18065 	/*
18066 	 * Compute options for segment. We only have to care about SYN and
18067 	 * established connection segments.  Options for SYN-ACK segments
18068 	 * are handled in TCP syncache.
18069 	 */
18070 	to.to_flags = 0;
18071 	if ((tp->t_flags & TF_NOOPT) == 0) {
18072 		/* Maximum segment size. */
18073 		if (flags & TH_SYN) {
18074 			tp->snd_nxt = tp->iss;
18075 			to.to_mss = tcp_mssopt(&inp->inp_inc);
18076 			if (tp->t_port)
18077 				to.to_mss -= V_tcp_udp_tunneling_overhead;
18078 			to.to_flags |= TOF_MSS;
18079 
18080 			/*
18081 			 * On SYN or SYN|ACK transmits on TFO connections,
18082 			 * only include the TFO option if it is not a
18083 			 * retransmit, as the presence of the TFO option may
18084 			 * have caused the original SYN or SYN|ACK to have
18085 			 * been dropped by a middlebox.
18086 			 */
18087 			if (IS_FASTOPEN(tp->t_flags) &&
18088 			    (tp->t_rxtshift == 0)) {
18089 				if (tp->t_state == TCPS_SYN_RECEIVED) {
18090 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
18091 					to.to_tfo_cookie =
18092 						(u_int8_t *)&tp->t_tfo_cookie.server;
18093 					to.to_flags |= TOF_FASTOPEN;
18094 					wanted_cookie = 1;
18095 				} else if (tp->t_state == TCPS_SYN_SENT) {
18096 					to.to_tfo_len =
18097 						tp->t_tfo_client_cookie_len;
18098 					to.to_tfo_cookie =
18099 						tp->t_tfo_cookie.client;
18100 					to.to_flags |= TOF_FASTOPEN;
18101 					wanted_cookie = 1;
18102 					/*
18103 					 * If we wind up having more data to
18104 					 * send with the SYN than can fit in
18105 					 * one segment, don't send any more
18106 					 * until the SYN|ACK comes back from
18107 					 * the other end.
18108 					 */
18109 					sendalot = 0;
18110 				}
18111 			}
18112 		}
18113 		/* Window scaling. */
18114 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
18115 			to.to_wscale = tp->request_r_scale;
18116 			to.to_flags |= TOF_SCALE;
18117 		}
18118 		/* Timestamps. */
18119 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
18120 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
18121 			to.to_tsval = ms_cts + tp->ts_offset;
18122 			to.to_tsecr = tp->ts_recent;
18123 			to.to_flags |= TOF_TS;
18124 		}
18125 		/* Set receive buffer autosizing timestamp. */
18126 		if (tp->rfbuf_ts == 0 &&
18127 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
18128 			tp->rfbuf_ts = tcp_ts_getticks();
18129 		/* Selective ACK's. */
18130 		if (tp->t_flags & TF_SACK_PERMIT) {
18131 			if (flags & TH_SYN)
18132 				to.to_flags |= TOF_SACKPERM;
18133 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18134 				 tp->rcv_numsacks > 0) {
18135 				to.to_flags |= TOF_SACK;
18136 				to.to_nsacks = tp->rcv_numsacks;
18137 				to.to_sacks = (u_char *)tp->sackblks;
18138 			}
18139 		}
18140 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18141 		/* TCP-MD5 (RFC2385). */
18142 		if (tp->t_flags & TF_SIGNATURE)
18143 			to.to_flags |= TOF_SIGNATURE;
18144 #endif				/* TCP_SIGNATURE */
18145 
18146 		/* Processing the options. */
18147 		hdrlen += optlen = tcp_addoptions(&to, opt);
18148 		/*
18149 		 * If we wanted a TFO option to be added, but it was unable
18150 		 * to fit, ensure no data is sent.
18151 		 */
18152 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
18153 		    !(to.to_flags & TOF_FASTOPEN))
18154 			len = 0;
18155 	}
18156 	if (tp->t_port) {
18157 		if (V_tcp_udp_tunneling_port == 0) {
18158 			/* The port was removed?? */
18159 			SOCKBUF_UNLOCK(&so->so_snd);
18160 #ifdef TCP_ACCOUNTING
18161 			crtsc = get_cyclecount();
18162 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18163 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
18164 			}
18165 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
18166 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
18167 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
18168 			}
18169 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
18170 			sched_unpin();
18171 #endif
18172 			return (EHOSTUNREACH);
18173 		}
18174 		hdrlen += sizeof(struct udphdr);
18175 	}
18176 #ifdef INET6
18177 	if (isipv6)
18178 		ipoptlen = ip6_optlen(inp);
18179 	else
18180 #endif
18181 		if (inp->inp_options)
18182 			ipoptlen = inp->inp_options->m_len -
18183 				offsetof(struct ipoption, ipopt_list);
18184 		else
18185 			ipoptlen = 0;
18186 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18187 	ipoptlen += ipsec_optlen;
18188 #endif
18189 
18190 	/*
18191 	 * Adjust data length if insertion of options will bump the packet
18192 	 * length beyond the t_maxseg length. Clear the FIN bit because we
18193 	 * cut off the tail of the segment.
18194 	 */
18195 	if (len + optlen + ipoptlen > tp->t_maxseg) {
18196 		if (tso) {
18197 			uint32_t if_hw_tsomax;
18198 			uint32_t moff;
18199 			int32_t max_len;
18200 
18201 			/* extract TSO information */
18202 			if_hw_tsomax = tp->t_tsomax;
18203 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18204 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18205 			KASSERT(ipoptlen == 0,
18206 				("%s: TSO can't do IP options", __func__));
18207 
18208 			/*
18209 			 * Check if we should limit by maximum payload
18210 			 * length:
18211 			 */
18212 			if (if_hw_tsomax != 0) {
18213 				/* compute maximum TSO length */
18214 				max_len = (if_hw_tsomax - hdrlen -
18215 					   max_linkhdr);
18216 				if (max_len <= 0) {
18217 					len = 0;
18218 				} else if (len > max_len) {
18219 					sendalot = 1;
18220 					len = max_len;
18221 					mark = 2;
18222 				}
18223 			}
18224 			/*
18225 			 * Prevent the last segment from being fractional
18226 			 * unless the send sockbuf can be emptied:
18227 			 */
18228 			max_len = (tp->t_maxseg - optlen);
18229 			if ((sb_offset + len) < sbavail(sb)) {
18230 				moff = len % (u_int)max_len;
18231 				if (moff != 0) {
18232 					mark = 3;
18233 					len -= moff;
18234 				}
18235 			}
18236 			/*
18237 			 * In case there are too many small fragments don't
18238 			 * use TSO:
18239 			 */
18240 			if (len <= segsiz) {
18241 				mark = 4;
18242 				tso = 0;
18243 			}
18244 			/*
18245 			 * Send the FIN in a separate segment after the bulk
18246 			 * sending is done. We don't trust the TSO
18247 			 * implementations to clear the FIN flag on all but
18248 			 * the last segment.
18249 			 */
18250 			if (tp->t_flags & TF_NEEDFIN) {
18251 				sendalot = 4;
18252 			}
18253 		} else {
18254 			mark = 5;
18255 			if (optlen + ipoptlen >= tp->t_maxseg) {
18256 				/*
18257 				 * Since we don't have enough space to put
18258 				 * the IP header chain and the TCP header in
18259 				 * one packet as required by RFC 7112, don't
18260 				 * send it. Also ensure that at least one
18261 				 * byte of the payload can be put into the
18262 				 * TCP segment.
18263 				 */
18264 				SOCKBUF_UNLOCK(&so->so_snd);
18265 				error = EMSGSIZE;
18266 				sack_rxmit = 0;
18267 				goto out;
18268 			}
18269 			len = tp->t_maxseg - optlen - ipoptlen;
18270 			sendalot = 5;
18271 		}
18272 	} else {
18273 		tso = 0;
18274 		mark = 6;
18275 	}
18276 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
18277 		("%s: len > IP_MAXPACKET", __func__));
18278 #ifdef DIAGNOSTIC
18279 #ifdef INET6
18280 	if (max_linkhdr + hdrlen > MCLBYTES)
18281 #else
18282 		if (max_linkhdr + hdrlen > MHLEN)
18283 #endif
18284 			panic("tcphdr too big");
18285 #endif
18286 
18287 	/*
18288 	 * This KASSERT is here to catch edge cases at a well defined place.
18289 	 * Before, those had triggered (random) panic conditions further
18290 	 * down.
18291 	 */
18292 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
18293 	if ((len == 0) &&
18294 	    (flags & TH_FIN) &&
18295 	    (sbused(sb))) {
18296 		/*
18297 		 * We have outstanding data, don't send a fin by itself!.
18298 		 */
18299 		goto just_return;
18300 	}
18301 	/*
18302 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
18303 	 * and initialize the header from the template for sends on this
18304 	 * connection.
18305 	 */
18306 	hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0;
18307 	if (len) {
18308 		uint32_t max_val;
18309 		uint32_t moff;
18310 
18311 		if (rack->r_ctl.rc_pace_max_segs)
18312 			max_val = rack->r_ctl.rc_pace_max_segs;
18313 		else if (rack->rc_user_set_max_segs)
18314 			max_val = rack->rc_user_set_max_segs * segsiz;
18315 		else
18316 			max_val = len;
18317 		/*
18318 		 * We allow a limit on sending with hptsi.
18319 		 */
18320 		if (len > max_val) {
18321 			mark = 7;
18322 			len = max_val;
18323 		}
18324 #ifdef INET6
18325 		if (MHLEN < hdrlen + max_linkhdr)
18326 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18327 		else
18328 #endif
18329 			m = m_gethdr(M_NOWAIT, MT_DATA);
18330 
18331 		if (m == NULL) {
18332 			SOCKBUF_UNLOCK(sb);
18333 			error = ENOBUFS;
18334 			sack_rxmit = 0;
18335 			goto out;
18336 		}
18337 		m->m_data += max_linkhdr;
18338 		m->m_len = hdrlen;
18339 
18340 		/*
18341 		 * Start the m_copy functions from the closest mbuf to the
18342 		 * sb_offset in the socket buffer chain.
18343 		 */
18344 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
18345 		s_mb = mb;
18346 		s_moff = moff;
18347 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
18348 			m_copydata(mb, moff, (int)len,
18349 				   mtod(m, caddr_t)+hdrlen);
18350 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18351 				sbsndptr_adv(sb, mb, len);
18352 			m->m_len += len;
18353 		} else {
18354 			struct sockbuf *msb;
18355 
18356 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
18357 				msb = NULL;
18358 			else
18359 				msb = sb;
18360 			m->m_next = tcp_m_copym(
18361 				mb, moff, &len,
18362 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
18363 				((rsm == NULL) ? hw_tls : 0)
18364 #ifdef NETFLIX_COPY_ARGS
18365 				, &s_mb, &s_moff
18366 #endif
18367 				);
18368 			if (len <= (tp->t_maxseg - optlen)) {
18369 				/*
18370 				 * Must have ran out of mbufs for the copy
18371 				 * shorten it to no longer need tso. Lets
18372 				 * not put on sendalot since we are low on
18373 				 * mbufs.
18374 				 */
18375 				tso = 0;
18376 			}
18377 			if (m->m_next == NULL) {
18378 				SOCKBUF_UNLOCK(sb);
18379 				(void)m_free(m);
18380 				error = ENOBUFS;
18381 				sack_rxmit = 0;
18382 				goto out;
18383 			}
18384 		}
18385 		if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
18386 			if (rsm && (rsm->r_flags & RACK_TLP)) {
18387 				/*
18388 				 * TLP should not count in retran count, but
18389 				 * in its own bin
18390 				 */
18391 				counter_u64_add(rack_tlp_retran, 1);
18392 				counter_u64_add(rack_tlp_retran_bytes, len);
18393 			} else {
18394 				tp->t_sndrexmitpack++;
18395 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18396 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18397 			}
18398 #ifdef STATS
18399 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18400 						 len);
18401 #endif
18402 		} else {
18403 			KMOD_TCPSTAT_INC(tcps_sndpack);
18404 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
18405 #ifdef STATS
18406 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
18407 						 len);
18408 #endif
18409 		}
18410 		/*
18411 		 * If we're sending everything we've got, set PUSH. (This
18412 		 * will keep happy those implementations which only give
18413 		 * data to the user when a buffer fills or a PUSH comes in.)
18414 		 */
18415 		if (sb_offset + len == sbused(sb) &&
18416 		    sbused(sb) &&
18417 		    !(flags & TH_SYN)) {
18418 			flags |= TH_PUSH;
18419 			add_flag |= RACK_HAD_PUSH;
18420 		}
18421 
18422 		SOCKBUF_UNLOCK(sb);
18423 	} else {
18424 		SOCKBUF_UNLOCK(sb);
18425 		if (tp->t_flags & TF_ACKNOW)
18426 			KMOD_TCPSTAT_INC(tcps_sndacks);
18427 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
18428 			KMOD_TCPSTAT_INC(tcps_sndctrl);
18429 		else
18430 			KMOD_TCPSTAT_INC(tcps_sndwinup);
18431 
18432 		m = m_gethdr(M_NOWAIT, MT_DATA);
18433 		if (m == NULL) {
18434 			error = ENOBUFS;
18435 			sack_rxmit = 0;
18436 			goto out;
18437 		}
18438 #ifdef INET6
18439 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
18440 		    MHLEN >= hdrlen) {
18441 			M_ALIGN(m, hdrlen);
18442 		} else
18443 #endif
18444 			m->m_data += max_linkhdr;
18445 		m->m_len = hdrlen;
18446 	}
18447 	SOCKBUF_UNLOCK_ASSERT(sb);
18448 	m->m_pkthdr.rcvif = (struct ifnet *)0;
18449 #ifdef MAC
18450 	mac_inpcb_create_mbuf(inp, m);
18451 #endif
18452 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
18453 #ifdef INET6
18454 		if (isipv6)
18455 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18456 		else
18457 #endif				/* INET6 */
18458 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18459 		th = rack->r_ctl.fsb.th;
18460 		udp = rack->r_ctl.fsb.udp;
18461 		if (udp) {
18462 #ifdef INET6
18463 			if (isipv6)
18464 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
18465 			else
18466 #endif				/* INET6 */
18467 				ulen = hdrlen + len - sizeof(struct ip);
18468 			udp->uh_ulen = htons(ulen);
18469 		}
18470 	} else {
18471 #ifdef INET6
18472 		if (isipv6) {
18473 			ip6 = mtod(m, struct ip6_hdr *);
18474 			if (tp->t_port) {
18475 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
18476 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18477 				udp->uh_dport = tp->t_port;
18478 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
18479 				udp->uh_ulen = htons(ulen);
18480 				th = (struct tcphdr *)(udp + 1);
18481 			} else
18482 				th = (struct tcphdr *)(ip6 + 1);
18483 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
18484 		} else
18485 #endif				/* INET6 */
18486 		{
18487 			ip = mtod(m, struct ip *);
18488 			if (tp->t_port) {
18489 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
18490 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
18491 				udp->uh_dport = tp->t_port;
18492 				ulen = hdrlen + len - sizeof(struct ip);
18493 				udp->uh_ulen = htons(ulen);
18494 				th = (struct tcphdr *)(udp + 1);
18495 			} else
18496 				th = (struct tcphdr *)(ip + 1);
18497 			tcpip_fillheaders(inp, tp->t_port, ip, th);
18498 		}
18499 	}
18500 	/*
18501 	 * Fill in fields, remembering maximum advertised window for use in
18502 	 * delaying messages about window sizes. If resending a FIN, be sure
18503 	 * not to use a new sequence number.
18504 	 */
18505 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
18506 	    tp->snd_nxt == tp->snd_max)
18507 		tp->snd_nxt--;
18508 	/*
18509 	 * If we are starting a connection, send ECN setup SYN packet. If we
18510 	 * are on a retransmit, we may resend those bits a number of times
18511 	 * as per RFC 3168.
18512 	 */
18513 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
18514 		flags |= tcp_ecn_output_syn_sent(tp);
18515 	}
18516 	/* Also handle parallel SYN for ECN */
18517 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
18518 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18519 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
18520 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18521 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
18522 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18523 #ifdef INET6
18524 		if (isipv6) {
18525 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18526 			ip6->ip6_flow |= htonl(ect << 20);
18527 		}
18528 		else
18529 #endif
18530 		{
18531 			ip->ip_tos &= ~IPTOS_ECN_MASK;
18532 			ip->ip_tos |= ect;
18533 		}
18534 	}
18535 	/*
18536 	 * If we are doing retransmissions, then snd_nxt will not reflect
18537 	 * the first unsent octet.  For ACK only packets, we do not want the
18538 	 * sequence number of the retransmitted packet, we want the sequence
18539 	 * number of the next unsent octet.  So, if there is no data (and no
18540 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
18541 	 * ti_seq.  But if we are in persist state, snd_max might reflect
18542 	 * one byte beyond the right edge of the window, so use snd_nxt in
18543 	 * that case, since we know we aren't doing a retransmission.
18544 	 * (retransmit and persist are mutually exclusive...)
18545 	 */
18546 	if (sack_rxmit == 0) {
18547 		if (len || (flags & (TH_SYN | TH_FIN))) {
18548 			th->th_seq = htonl(tp->snd_nxt);
18549 			rack_seq = tp->snd_nxt;
18550 		} else {
18551 			th->th_seq = htonl(tp->snd_max);
18552 			rack_seq = tp->snd_max;
18553 		}
18554 	} else {
18555 		th->th_seq = htonl(rsm->r_start);
18556 		rack_seq = rsm->r_start;
18557 	}
18558 	th->th_ack = htonl(tp->rcv_nxt);
18559 	tcp_set_flags(th, flags);
18560 	/*
18561 	 * Calculate receive window.  Don't shrink window, but avoid silly
18562 	 * window syndrome.
18563 	 * If a RST segment is sent, advertise a window of zero.
18564 	 */
18565 	if (flags & TH_RST) {
18566 		recwin = 0;
18567 	} else {
18568 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
18569 		    recwin < (long)segsiz) {
18570 			recwin = 0;
18571 		}
18572 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
18573 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
18574 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
18575 	}
18576 
18577 	/*
18578 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
18579 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
18580 	 * handled in syncache.
18581 	 */
18582 	if (flags & TH_SYN)
18583 		th->th_win = htons((u_short)
18584 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
18585 	else {
18586 		/* Avoid shrinking window with window scaling. */
18587 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
18588 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
18589 	}
18590 	/*
18591 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
18592 	 * window.  This may cause the remote transmitter to stall.  This
18593 	 * flag tells soreceive() to disable delayed acknowledgements when
18594 	 * draining the buffer.  This can occur if the receiver is
18595 	 * attempting to read more data than can be buffered prior to
18596 	 * transmitting on the connection.
18597 	 */
18598 	if (th->th_win == 0) {
18599 		tp->t_sndzerowin++;
18600 		tp->t_flags |= TF_RXWIN0SENT;
18601 	} else
18602 		tp->t_flags &= ~TF_RXWIN0SENT;
18603 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
18604 	/* Now are we using fsb?, if so copy the template data to the mbuf */
18605 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
18606 		uint8_t *cpto;
18607 
18608 		cpto = mtod(m, uint8_t *);
18609 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
18610 		/*
18611 		 * We have just copied in:
18612 		 * IP/IP6
18613 		 * <optional udphdr>
18614 		 * tcphdr (no options)
18615 		 *
18616 		 * We need to grab the correct pointers into the mbuf
18617 		 * for both the tcp header, and possibly the udp header (if tunneling).
18618 		 * We do this by using the offset in the copy buffer and adding it
18619 		 * to the mbuf base pointer (cpto).
18620 		 */
18621 #ifdef INET6
18622 		if (isipv6)
18623 			ip6 = mtod(m, struct ip6_hdr *);
18624 		else
18625 #endif				/* INET6 */
18626 			ip = mtod(m, struct ip *);
18627 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
18628 		/* If we have a udp header lets set it into the mbuf as well */
18629 		if (udp)
18630 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
18631 	}
18632 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18633 	if (to.to_flags & TOF_SIGNATURE) {
18634 		/*
18635 		 * Calculate MD5 signature and put it into the place
18636 		 * determined before.
18637 		 * NOTE: since TCP options buffer doesn't point into
18638 		 * mbuf's data, calculate offset and use it.
18639 		 */
18640 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18641 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18642 			/*
18643 			 * Do not send segment if the calculation of MD5
18644 			 * digest has failed.
18645 			 */
18646 			goto out;
18647 		}
18648 	}
18649 #endif
18650 	if (optlen) {
18651 		bcopy(opt, th + 1, optlen);
18652 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
18653 	}
18654 	/*
18655 	 * Put TCP length in extended header, and then checksum extended
18656 	 * header and data.
18657 	 */
18658 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
18659 #ifdef INET6
18660 	if (isipv6) {
18661 		/*
18662 		 * ip6_plen is not need to be filled now, and will be filled
18663 		 * in ip6_output.
18664 		 */
18665 		if (tp->t_port) {
18666 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18667 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18668 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18669 			th->th_sum = htons(0);
18670 			UDPSTAT_INC(udps_opackets);
18671 		} else {
18672 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18673 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18674 			th->th_sum = in6_cksum_pseudo(ip6,
18675 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18676 						      0);
18677 		}
18678 	}
18679 #endif
18680 #if defined(INET6) && defined(INET)
18681 	else
18682 #endif
18683 #ifdef INET
18684 	{
18685 		if (tp->t_port) {
18686 			m->m_pkthdr.csum_flags = CSUM_UDP;
18687 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18688 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18689 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18690 			th->th_sum = htons(0);
18691 			UDPSTAT_INC(udps_opackets);
18692 		} else {
18693 			m->m_pkthdr.csum_flags = CSUM_TCP;
18694 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18695 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
18696 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18697 									IPPROTO_TCP + len + optlen));
18698 		}
18699 		/* IP version must be set here for ipv4/ipv6 checking later */
18700 		KASSERT(ip->ip_v == IPVERSION,
18701 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
18702 	}
18703 #endif
18704 	/*
18705 	 * Enable TSO and specify the size of the segments. The TCP pseudo
18706 	 * header checksum is always provided. XXX: Fixme: This is currently
18707 	 * not the case for IPv6.
18708 	 */
18709 	if (tso) {
18710 		KASSERT(len > tp->t_maxseg - optlen,
18711 			("%s: len <= tso_segsz", __func__));
18712 		m->m_pkthdr.csum_flags |= CSUM_TSO;
18713 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
18714 	}
18715 	KASSERT(len + hdrlen == m_length(m, NULL),
18716 		("%s: mbuf chain different than expected: %d + %u != %u",
18717 		 __func__, len, hdrlen, m_length(m, NULL)));
18718 
18719 #ifdef TCP_HHOOK
18720 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
18721 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
18722 #endif
18723 	/* We're getting ready to send; log now. */
18724 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
18725 		union tcp_log_stackspecific log;
18726 
18727 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18728 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp);
18729 		if (rack->rack_no_prr)
18730 			log.u_bbr.flex1 = 0;
18731 		else
18732 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
18733 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
18734 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
18735 		log.u_bbr.flex4 = orig_len;
18736 		/* Save off the early/late values */
18737 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18738 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
18739 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18740 		log.u_bbr.flex8 = 0;
18741 		if (rsm) {
18742 			if (rsm->r_flags & RACK_RWND_COLLAPSED) {
18743 				rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
18744 				counter_u64_add(rack_collapsed_win_rxt, 1);
18745 				counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
18746 			}
18747 			if (doing_tlp)
18748 				log.u_bbr.flex8 = 2;
18749 			else
18750 				log.u_bbr.flex8 = 1;
18751 		} else {
18752 			if (doing_tlp)
18753 				log.u_bbr.flex8 = 3;
18754 			else
18755 				log.u_bbr.flex8 = 0;
18756 		}
18757 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
18758 		log.u_bbr.flex7 = mark;
18759 		log.u_bbr.flex7 <<= 8;
18760 		log.u_bbr.flex7 |= pass;
18761 		log.u_bbr.pkts_out = tp->t_maxseg;
18762 		log.u_bbr.timeStamp = cts;
18763 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18764 		log.u_bbr.lt_epoch = cwnd_to_use;
18765 		log.u_bbr.delivered = sendalot;
18766 		lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
18767 				     len, &log, false, NULL, NULL, 0, &tv);
18768 	} else
18769 		lgb = NULL;
18770 
18771 	/*
18772 	 * Fill in IP length and desired time to live and send to IP level.
18773 	 * There should be a better way to handle ttl and tos; we could keep
18774 	 * them in the template, but need a way to checksum without them.
18775 	 */
18776 	/*
18777 	 * m->m_pkthdr.len should have been set before cksum calcuration,
18778 	 * because in6_cksum() need it.
18779 	 */
18780 #ifdef INET6
18781 	if (isipv6) {
18782 		/*
18783 		 * we separately set hoplimit for every segment, since the
18784 		 * user might want to change the value via setsockopt. Also,
18785 		 * desired default hop limit might be changed via Neighbor
18786 		 * Discovery.
18787 		 */
18788 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
18789 
18790 		/*
18791 		 * Set the packet size here for the benefit of DTrace
18792 		 * probes. ip6_output() will set it properly; it's supposed
18793 		 * to include the option header lengths as well.
18794 		 */
18795 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18796 
18797 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18798 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18799 		else
18800 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18801 
18802 		if (tp->t_state == TCPS_SYN_SENT)
18803 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
18804 
18805 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
18806 		/* TODO: IPv6 IP6TOS_ECT bit on */
18807 		error = ip6_output(m,
18808 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18809 				   inp->in6p_outputopts,
18810 #else
18811 				   NULL,
18812 #endif
18813 				   &inp->inp_route6,
18814 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
18815 				   NULL, NULL, inp);
18816 
18817 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
18818 			mtu = inp->inp_route6.ro_nh->nh_mtu;
18819 	}
18820 #endif				/* INET6 */
18821 #if defined(INET) && defined(INET6)
18822 	else
18823 #endif
18824 #ifdef INET
18825 	{
18826 		ip->ip_len = htons(m->m_pkthdr.len);
18827 #ifdef INET6
18828 		if (inp->inp_vflag & INP_IPV6PROTO)
18829 			ip->ip_ttl = in6_selecthlim(inp, NULL);
18830 #endif				/* INET6 */
18831 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
18832 		/*
18833 		 * If we do path MTU discovery, then we set DF on every
18834 		 * packet. This might not be the best thing to do according
18835 		 * to RFC3390 Section 2. However the tcp hostcache migitates
18836 		 * the problem so it affects only the first tcp connection
18837 		 * with a host.
18838 		 *
18839 		 * NB: Don't set DF on small MTU/MSS to have a safe
18840 		 * fallback.
18841 		 */
18842 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
18843 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18844 			if (tp->t_port == 0 || len < V_tcp_minmss) {
18845 				ip->ip_off |= htons(IP_DF);
18846 			}
18847 		} else {
18848 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18849 		}
18850 
18851 		if (tp->t_state == TCPS_SYN_SENT)
18852 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
18853 
18854 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
18855 
18856 		error = ip_output(m,
18857 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
18858 				  inp->inp_options,
18859 #else
18860 				  NULL,
18861 #endif
18862 				  &inp->inp_route,
18863 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
18864 				  inp);
18865 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
18866 			mtu = inp->inp_route.ro_nh->nh_mtu;
18867 	}
18868 #endif				/* INET */
18869 
18870 out:
18871 	if (lgb) {
18872 		lgb->tlb_errno = error;
18873 		lgb = NULL;
18874 	}
18875 	/*
18876 	 * In transmit state, time the transmission and arrange for the
18877 	 * retransmit.  In persist state, just set snd_max.
18878 	 */
18879 	if (error == 0) {
18880 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
18881 		if (rsm && doing_tlp) {
18882 			rack->rc_last_sent_tlp_past_cumack = 0;
18883 			rack->rc_last_sent_tlp_seq_valid = 1;
18884 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
18885 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
18886 		}
18887 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
18888 		if (rsm && (doing_tlp == 0)) {
18889 			/* Set we retransmitted */
18890 			rack->rc_gp_saw_rec = 1;
18891 		} else {
18892 			if (cwnd_to_use > tp->snd_ssthresh) {
18893 				/* Set we sent in CA */
18894 				rack->rc_gp_saw_ca = 1;
18895 			} else {
18896 				/* Set we sent in SS */
18897 				rack->rc_gp_saw_ss = 1;
18898 			}
18899 		}
18900 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
18901 		    (tp->t_flags & TF_SACK_PERMIT) &&
18902 		    tp->rcv_numsacks > 0)
18903 			tcp_clean_dsack_blocks(tp);
18904 		tot_len_this_send += len;
18905 		if (len == 0)
18906 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
18907 		else if (len == 1) {
18908 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
18909 		} else if (len > 1) {
18910 			int idx;
18911 
18912 			idx = (len / segsiz) + 3;
18913 			if (idx >= TCP_MSS_ACCT_ATIMER)
18914 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
18915 			else
18916 				counter_u64_add(rack_out_size[idx], 1);
18917 		}
18918 	}
18919 	if ((rack->rack_no_prr == 0) &&
18920 	    sub_from_prr &&
18921 	    (error == 0)) {
18922 		if (rack->r_ctl.rc_prr_sndcnt >= len)
18923 			rack->r_ctl.rc_prr_sndcnt -= len;
18924 		else
18925 			rack->r_ctl.rc_prr_sndcnt = 0;
18926 	}
18927 	sub_from_prr = 0;
18928 	if (doing_tlp) {
18929 		/* Make sure the TLP is added */
18930 		add_flag |= RACK_TLP;
18931 	} else if (rsm) {
18932 		/* If its a resend without TLP then it must not have the flag */
18933 		rsm->r_flags &= ~RACK_TLP;
18934 	}
18935 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
18936 			rack_to_usec_ts(&tv),
18937 			rsm, add_flag, s_mb, s_moff, hw_tls);
18938 
18939 
18940 	if ((error == 0) &&
18941 	    (len > 0) &&
18942 	    (tp->snd_una == tp->snd_max))
18943 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
18944 	{
18945 		tcp_seq startseq = tp->snd_nxt;
18946 
18947 		/* Track our lost count */
18948 		if (rsm && (doing_tlp == 0))
18949 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
18950 		/*
18951 		 * Advance snd_nxt over sequence space of this segment.
18952 		 */
18953 		if (error)
18954 			/* We don't log or do anything with errors */
18955 			goto nomore;
18956 		if (doing_tlp == 0) {
18957 			if (rsm == NULL) {
18958 				/*
18959 				 * Not a retransmission of some
18960 				 * sort, new data is going out so
18961 				 * clear our TLP count and flag.
18962 				 */
18963 				rack->rc_tlp_in_progress = 0;
18964 				rack->r_ctl.rc_tlp_cnt_out = 0;
18965 			}
18966 		} else {
18967 			/*
18968 			 * We have just sent a TLP, mark that it is true
18969 			 * and make sure our in progress is set so we
18970 			 * continue to check the count.
18971 			 */
18972 			rack->rc_tlp_in_progress = 1;
18973 			rack->r_ctl.rc_tlp_cnt_out++;
18974 		}
18975 		if (flags & (TH_SYN | TH_FIN)) {
18976 			if (flags & TH_SYN)
18977 				tp->snd_nxt++;
18978 			if (flags & TH_FIN) {
18979 				tp->snd_nxt++;
18980 				tp->t_flags |= TF_SENTFIN;
18981 			}
18982 		}
18983 		/* In the ENOBUFS case we do *not* update snd_max */
18984 		if (sack_rxmit)
18985 			goto nomore;
18986 
18987 		tp->snd_nxt += len;
18988 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
18989 			if (tp->snd_una == tp->snd_max) {
18990 				/*
18991 				 * Update the time we just added data since
18992 				 * none was outstanding.
18993 				 */
18994 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
18995 				tp->t_acktime = ticks;
18996 			}
18997 			tp->snd_max = tp->snd_nxt;
18998 			/*
18999 			 * Time this transmission if not a retransmission and
19000 			 * not currently timing anything.
19001 			 * This is only relevant in case of switching back to
19002 			 * the base stack.
19003 			 */
19004 			if (tp->t_rtttime == 0) {
19005 				tp->t_rtttime = ticks;
19006 				tp->t_rtseq = startseq;
19007 				KMOD_TCPSTAT_INC(tcps_segstimed);
19008 			}
19009 			if (len &&
19010 			    ((tp->t_flags & TF_GPUTINPROG) == 0))
19011 				rack_start_gp_measurement(tp, rack, startseq, sb_offset);
19012 		}
19013 		/*
19014 		 * If we are doing FO we need to update the mbuf position and subtract
19015 		 * this happens when the peer sends us duplicate information and
19016 		 * we thus want to send a DSACK.
19017 		 *
19018 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
19019 		 * turned off? If not then we are going to echo multiple DSACK blocks
19020 		 * out (with the TSO), which we should not be doing.
19021 		 */
19022 		if (rack->r_fast_output && len) {
19023 			if (rack->r_ctl.fsb.left_to_send > len)
19024 				rack->r_ctl.fsb.left_to_send -= len;
19025 			else
19026 				rack->r_ctl.fsb.left_to_send = 0;
19027 			if (rack->r_ctl.fsb.left_to_send < segsiz)
19028 				rack->r_fast_output = 0;
19029 			if (rack->r_fast_output) {
19030 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19031 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19032 			}
19033 		}
19034 	}
19035 nomore:
19036 	if (error) {
19037 		rack->r_ctl.rc_agg_delayed = 0;
19038 		rack->r_early = 0;
19039 		rack->r_late = 0;
19040 		rack->r_ctl.rc_agg_early = 0;
19041 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
19042 		/*
19043 		 * Failures do not advance the seq counter above. For the
19044 		 * case of ENOBUFS we will fall out and retry in 1ms with
19045 		 * the hpts. Everything else will just have to retransmit
19046 		 * with the timer.
19047 		 *
19048 		 * In any case, we do not want to loop around for another
19049 		 * send without a good reason.
19050 		 */
19051 		sendalot = 0;
19052 		switch (error) {
19053 		case EPERM:
19054 			tp->t_softerror = error;
19055 #ifdef TCP_ACCOUNTING
19056 			crtsc = get_cyclecount();
19057 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19058 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19059 			}
19060 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19061 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19062 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19063 			}
19064 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19065 			sched_unpin();
19066 #endif
19067 			return (error);
19068 		case ENOBUFS:
19069 			/*
19070 			 * Pace us right away to retry in a some
19071 			 * time
19072 			 */
19073 			if (rack->r_ctl.crte != NULL) {
19074 				rack_trace_point(rack, RACK_TP_HWENOBUF);
19075 			} else
19076 				rack_trace_point(rack, RACK_TP_ENOBUF);
19077 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
19078 			if (rack->rc_enobuf < 0x7f)
19079 				rack->rc_enobuf++;
19080 			if (slot < (10 * HPTS_USEC_IN_MSEC))
19081 				slot = 10 * HPTS_USEC_IN_MSEC;
19082 			if (rack->r_ctl.crte != NULL) {
19083 				counter_u64_add(rack_saw_enobuf_hw, 1);
19084 				tcp_rl_log_enobuf(rack->r_ctl.crte);
19085 			}
19086 			counter_u64_add(rack_saw_enobuf, 1);
19087 			goto enobufs;
19088 		case EMSGSIZE:
19089 			/*
19090 			 * For some reason the interface we used initially
19091 			 * to send segments changed to another or lowered
19092 			 * its MTU. If TSO was active we either got an
19093 			 * interface without TSO capabilits or TSO was
19094 			 * turned off. If we obtained mtu from ip_output()
19095 			 * then update it and try again.
19096 			 */
19097 			if (tso)
19098 				tp->t_flags &= ~TF_TSO;
19099 			if (mtu != 0) {
19100 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
19101 				goto again;
19102 			}
19103 			slot = 10 * HPTS_USEC_IN_MSEC;
19104 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19105 #ifdef TCP_ACCOUNTING
19106 			crtsc = get_cyclecount();
19107 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19108 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19109 			}
19110 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19111 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19112 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19113 			}
19114 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19115 			sched_unpin();
19116 #endif
19117 			return (error);
19118 		case ENETUNREACH:
19119 			counter_u64_add(rack_saw_enetunreach, 1);
19120 		case EHOSTDOWN:
19121 		case EHOSTUNREACH:
19122 		case ENETDOWN:
19123 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
19124 				tp->t_softerror = error;
19125 			}
19126 			/* FALLTHROUGH */
19127 		default:
19128 			slot = 10 * HPTS_USEC_IN_MSEC;
19129 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
19130 #ifdef TCP_ACCOUNTING
19131 			crtsc = get_cyclecount();
19132 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19133 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
19134 			}
19135 			counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1);
19136 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19137 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
19138 			}
19139 			counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val));
19140 			sched_unpin();
19141 #endif
19142 			return (error);
19143 		}
19144 	} else {
19145 		rack->rc_enobuf = 0;
19146 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
19147 			rack->r_ctl.retran_during_recovery += len;
19148 	}
19149 	KMOD_TCPSTAT_INC(tcps_sndtotal);
19150 
19151 	/*
19152 	 * Data sent (as far as we can tell). If this advertises a larger
19153 	 * window than any other segment, then remember the size of the
19154 	 * advertised window. Any pending ACK has now been sent.
19155 	 */
19156 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
19157 		tp->rcv_adv = tp->rcv_nxt + recwin;
19158 
19159 	tp->last_ack_sent = tp->rcv_nxt;
19160 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19161 enobufs:
19162 	if (sendalot) {
19163 		/* Do we need to turn off sendalot? */
19164 		if (rack->r_ctl.rc_pace_max_segs &&
19165 		    (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
19166 			/* We hit our max. */
19167 			sendalot = 0;
19168 		} else if ((rack->rc_user_set_max_segs) &&
19169 			   (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
19170 			/* We hit the user defined max */
19171 			sendalot = 0;
19172 		}
19173 	}
19174 	if ((error == 0) && (flags & TH_FIN))
19175 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
19176 	if (flags & TH_RST) {
19177 		/*
19178 		 * We don't send again after sending a RST.
19179 		 */
19180 		slot = 0;
19181 		sendalot = 0;
19182 		if (error == 0)
19183 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
19184 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
19185 		/*
19186 		 * Get our pacing rate, if an error
19187 		 * occurred in sending (ENOBUF) we would
19188 		 * hit the else if with slot preset. Other
19189 		 * errors return.
19190 		 */
19191 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
19192 	}
19193 	if (rsm &&
19194 	    (rsm->r_flags & RACK_HAS_SYN) == 0 &&
19195 	    rack->use_rack_rr) {
19196 		/* Its a retransmit and we use the rack cheat? */
19197 		if ((slot == 0) ||
19198 		    (rack->rc_always_pace == 0) ||
19199 		    (rack->r_rr_config == 1)) {
19200 			/*
19201 			 * We have no pacing set or we
19202 			 * are using old-style rack or
19203 			 * we are overridden to use the old 1ms pacing.
19204 			 */
19205 			slot = rack->r_ctl.rc_min_to;
19206 		}
19207 	}
19208 	/* We have sent clear the flag */
19209 	rack->r_ent_rec_ns = 0;
19210 	if (rack->r_must_retran) {
19211 		if (rsm) {
19212 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
19213 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
19214 				/*
19215 				 * We have retransmitted all.
19216 				 */
19217 				rack->r_must_retran = 0;
19218 				rack->r_ctl.rc_out_at_rto = 0;
19219 			}
19220 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19221 			/*
19222 			 * Sending new data will also kill
19223 			 * the loop.
19224 			 */
19225 			rack->r_must_retran = 0;
19226 			rack->r_ctl.rc_out_at_rto = 0;
19227 		}
19228 	}
19229 	rack->r_ctl.fsb.recwin = recwin;
19230 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
19231 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
19232 		/*
19233 		 * We hit an RTO and now have past snd_max at the RTO
19234 		 * clear all the WAS flags.
19235 		 */
19236 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
19237 	}
19238 	if (slot) {
19239 		/* set the rack tcb into the slot N */
19240 		if ((error == 0) &&
19241 		    rack_use_rfo &&
19242 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
19243 		    (rsm == NULL) &&
19244 		    (tp->snd_nxt == tp->snd_max) &&
19245 		    (ipoptlen == 0) &&
19246 		    (tp->rcv_numsacks == 0) &&
19247 		    rack->r_fsb_inited &&
19248 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
19249 		    (rack->r_must_retran == 0) &&
19250 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
19251 		    (len > 0) && (orig_len > 0) &&
19252 		    (orig_len > len) &&
19253 		    ((orig_len - len) >= segsiz) &&
19254 		    ((optlen == 0) ||
19255 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19256 			/* We can send at least one more MSS using our fsb */
19257 
19258 			rack->r_fast_output = 1;
19259 			rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19260 			rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19261 			rack->r_ctl.fsb.tcp_flags = flags;
19262 			rack->r_ctl.fsb.left_to_send = orig_len - len;
19263 			if (hw_tls)
19264 				rack->r_ctl.fsb.hw_tls = 1;
19265 			else
19266 				rack->r_ctl.fsb.hw_tls = 0;
19267 			KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19268 				("rack:%p left_to_send:%u sbavail:%u out:%u",
19269 				 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19270 				 (tp->snd_max - tp->snd_una)));
19271 			if (rack->r_ctl.fsb.left_to_send < segsiz)
19272 				rack->r_fast_output = 0;
19273 			else {
19274 				if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19275 					rack->r_ctl.fsb.rfo_apply_push = 1;
19276 				else
19277 					rack->r_ctl.fsb.rfo_apply_push = 0;
19278 			}
19279 		} else
19280 			rack->r_fast_output = 0;
19281 		rack_log_fsb(rack, tp, so, flags,
19282 			     ipoptlen, orig_len, len, error,
19283 			     (rsm == NULL), optlen, __LINE__, 2);
19284 	} else if (sendalot) {
19285 		int ret;
19286 
19287 		sack_rxmit = 0;
19288 		if ((error == 0) &&
19289 		    rack_use_rfo &&
19290 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
19291 		    (rsm == NULL) &&
19292 		    (ipoptlen == 0) &&
19293 		    (tp->rcv_numsacks == 0) &&
19294 		    (tp->snd_nxt == tp->snd_max) &&
19295 		    (rack->r_must_retran == 0) &&
19296 		    rack->r_fsb_inited &&
19297 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
19298 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
19299 		    (len > 0) && (orig_len > 0) &&
19300 		    (orig_len > len) &&
19301 		    ((orig_len - len) >= segsiz) &&
19302 		    ((optlen == 0) ||
19303 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
19304 			/* we can use fast_output for more */
19305 
19306 			rack->r_fast_output = 1;
19307 			rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19308 			rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19309 			rack->r_ctl.fsb.tcp_flags = flags;
19310 			rack->r_ctl.fsb.left_to_send = orig_len - len;
19311 			if (hw_tls)
19312 				rack->r_ctl.fsb.hw_tls = 1;
19313 			else
19314 				rack->r_ctl.fsb.hw_tls = 0;
19315 			KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19316 				("rack:%p left_to_send:%u sbavail:%u out:%u",
19317 				 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19318 				 (tp->snd_max - tp->snd_una)));
19319 			if (rack->r_ctl.fsb.left_to_send < segsiz) {
19320 				rack->r_fast_output = 0;
19321 			}
19322 			if (rack->r_fast_output) {
19323 				if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19324 					rack->r_ctl.fsb.rfo_apply_push = 1;
19325 				else
19326 					rack->r_ctl.fsb.rfo_apply_push = 0;
19327 				rack_log_fsb(rack, tp, so, flags,
19328 					     ipoptlen, orig_len, len, error,
19329 					     (rsm == NULL), optlen, __LINE__, 3);
19330 				error = 0;
19331 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
19332 				if (ret >= 0)
19333 					return (ret);
19334 			        else if (error)
19335 					goto nomore;
19336 
19337 			}
19338 		}
19339 		goto again;
19340 	}
19341 	/* Assure when we leave that snd_nxt will point to top */
19342 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
19343 		tp->snd_nxt = tp->snd_max;
19344 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
19345 #ifdef TCP_ACCOUNTING
19346 	crtsc = get_cyclecount() - ts_val;
19347 	if (tot_len_this_send) {
19348 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19349 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
19350 		}
19351 		counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1);
19352 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19353 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
19354 		}
19355 		counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc);
19356 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19357 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
19358 		}
19359 		counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz));
19360 	} else {
19361 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19362 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
19363 		}
19364 		counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1);
19365 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19366 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
19367 		}
19368 		counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc);
19369 	}
19370 	sched_unpin();
19371 #endif
19372 	if (error == ENOBUFS)
19373 		error = 0;
19374 	return (error);
19375 }
19376 
19377 static void
19378 rack_update_seg(struct tcp_rack *rack)
19379 {
19380 	uint32_t orig_val;
19381 
19382 	orig_val = rack->r_ctl.rc_pace_max_segs;
19383 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
19384 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
19385 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
19386 }
19387 
19388 static void
19389 rack_mtu_change(struct tcpcb *tp)
19390 {
19391 	/*
19392 	 * The MSS may have changed
19393 	 */
19394 	struct tcp_rack *rack;
19395 	struct rack_sendmap *rsm;
19396 
19397 	rack = (struct tcp_rack *)tp->t_fb_ptr;
19398 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
19399 		/*
19400 		 * The MTU has changed we need to resend everything
19401 		 * since all we have sent is lost. We first fix
19402 		 * up the mtu though.
19403 		 */
19404 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
19405 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
19406 		rack_remxt_tmr(tp);
19407 		rack->r_fast_output = 0;
19408 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
19409 						rack->r_ctl.rc_sacked);
19410 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
19411 		rack->r_must_retran = 1;
19412 		/* Mark all inflight to needing to be rxt'd */
19413 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
19414 			rsm->r_flags |= RACK_MUST_RXT;
19415 		}
19416 	}
19417 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
19418 	/* We don't use snd_nxt to retransmit */
19419 	tp->snd_nxt = tp->snd_max;
19420 }
19421 
19422 static int
19423 rack_set_profile(struct tcp_rack *rack, int prof)
19424 {
19425 	int err = EINVAL;
19426 	if (prof == 1) {
19427 		/* pace_always=1 */
19428 		if (rack->rc_always_pace == 0) {
19429 			if (tcp_can_enable_pacing() == 0)
19430 				return (EBUSY);
19431 		}
19432 		rack->rc_always_pace = 1;
19433 		if (rack->use_fixed_rate || rack->gp_ready)
19434 			rack_set_cc_pacing(rack);
19435 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19436 		rack->rack_attempt_hdwr_pace = 0;
19437 		/* cmpack=1 */
19438 		if (rack_use_cmp_acks)
19439 			rack->r_use_cmp_ack = 1;
19440 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19441 		    rack->r_use_cmp_ack)
19442 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19443 		/* scwnd=1 */
19444 		rack->rack_enable_scwnd = 1;
19445 		/* dynamic=100 */
19446 		rack->rc_gp_dyn_mul = 1;
19447 		/* gp_inc_ca */
19448 		rack->r_ctl.rack_per_of_gp_ca = 100;
19449 		/* rrr_conf=3 */
19450 		rack->r_rr_config = 3;
19451 		/* npush=2 */
19452 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19453 		/* fillcw=1 */
19454 		rack->rc_pace_to_cwnd = 1;
19455 		rack->rc_pace_fill_if_rttin_range = 0;
19456 		rack->rtt_limit_mul = 0;
19457 		/* noprr=1 */
19458 		rack->rack_no_prr = 1;
19459 		/* lscwnd=1 */
19460 		rack->r_limit_scw = 1;
19461 		/* gp_inc_rec */
19462 		rack->r_ctl.rack_per_of_gp_rec = 90;
19463 		err = 0;
19464 
19465 	} else if (prof == 3) {
19466 		/* Same as profile one execept fill_cw becomes 2 (less aggressive set) */
19467 		/* pace_always=1 */
19468 		if (rack->rc_always_pace == 0) {
19469 			if (tcp_can_enable_pacing() == 0)
19470 				return (EBUSY);
19471 		}
19472 		rack->rc_always_pace = 1;
19473 		if (rack->use_fixed_rate || rack->gp_ready)
19474 			rack_set_cc_pacing(rack);
19475 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19476 		rack->rack_attempt_hdwr_pace = 0;
19477 		/* cmpack=1 */
19478 		if (rack_use_cmp_acks)
19479 			rack->r_use_cmp_ack = 1;
19480 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
19481 		    rack->r_use_cmp_ack)
19482 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19483 		/* scwnd=1 */
19484 		rack->rack_enable_scwnd = 1;
19485 		/* dynamic=100 */
19486 		rack->rc_gp_dyn_mul = 1;
19487 		/* gp_inc_ca */
19488 		rack->r_ctl.rack_per_of_gp_ca = 100;
19489 		/* rrr_conf=3 */
19490 		rack->r_rr_config = 3;
19491 		/* npush=2 */
19492 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19493 		/* fillcw=2 */
19494 		rack->rc_pace_to_cwnd = 1;
19495 		rack->r_fill_less_agg = 1;
19496 		rack->rc_pace_fill_if_rttin_range = 0;
19497 		rack->rtt_limit_mul = 0;
19498 		/* noprr=1 */
19499 		rack->rack_no_prr = 1;
19500 		/* lscwnd=1 */
19501 		rack->r_limit_scw = 1;
19502 		/* gp_inc_rec */
19503 		rack->r_ctl.rack_per_of_gp_rec = 90;
19504 		err = 0;
19505 
19506 
19507 	} else if (prof == 2) {
19508 		/* cmpack=1 */
19509 		if (rack->rc_always_pace == 0) {
19510 			if (tcp_can_enable_pacing() == 0)
19511 				return (EBUSY);
19512 		}
19513 		rack->rc_always_pace = 1;
19514 		if (rack->use_fixed_rate || rack->gp_ready)
19515 			rack_set_cc_pacing(rack);
19516 		rack->r_use_cmp_ack = 1;
19517 		if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19518 			rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19519 		/* pace_always=1 */
19520 		rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19521 		/* scwnd=1 */
19522 		rack->rack_enable_scwnd = 1;
19523 		/* dynamic=100 */
19524 		rack->rc_gp_dyn_mul = 1;
19525 		rack->r_ctl.rack_per_of_gp_ca = 100;
19526 		/* rrr_conf=3 */
19527 		rack->r_rr_config = 3;
19528 		/* npush=2 */
19529 		rack->r_ctl.rc_no_push_at_mrtt = 2;
19530 		/* fillcw=1 */
19531 		rack->rc_pace_to_cwnd = 1;
19532 		rack->rc_pace_fill_if_rttin_range = 0;
19533 		rack->rtt_limit_mul = 0;
19534 		/* noprr=1 */
19535 		rack->rack_no_prr = 1;
19536 		/* lscwnd=0 */
19537 		rack->r_limit_scw = 0;
19538 		err = 0;
19539 	} else if (prof == 0) {
19540 		/* This changes things back to the default settings */
19541 		err = 0;
19542 		if (rack->rc_always_pace) {
19543 			tcp_decrement_paced_conn();
19544 			rack_undo_cc_pacing(rack);
19545 			rack->rc_always_pace = 0;
19546 		}
19547 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
19548 			rack->rc_always_pace = 1;
19549 			if (rack->use_fixed_rate || rack->gp_ready)
19550 				rack_set_cc_pacing(rack);
19551 		} else
19552 			rack->rc_always_pace = 0;
19553 		if (rack_dsack_std_based & 0x1) {
19554 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
19555 			rack->rc_rack_tmr_std_based = 1;
19556 		}
19557 		if (rack_dsack_std_based & 0x2) {
19558 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
19559 			rack->rc_rack_use_dsack = 1;
19560 		}
19561 		if (rack_use_cmp_acks)
19562 			rack->r_use_cmp_ack = 1;
19563 		else
19564 			rack->r_use_cmp_ack = 0;
19565 		if (rack_disable_prr)
19566 			rack->rack_no_prr = 1;
19567 		else
19568 			rack->rack_no_prr = 0;
19569 		if (rack_gp_no_rec_chg)
19570 			rack->rc_gp_no_rec_chg = 1;
19571 		else
19572 			rack->rc_gp_no_rec_chg = 0;
19573 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
19574 			rack->r_mbuf_queue = 1;
19575 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
19576 				rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP;
19577 			rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19578 		} else {
19579 			rack->r_mbuf_queue = 0;
19580 			rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19581 		}
19582 		if (rack_enable_shared_cwnd)
19583 			rack->rack_enable_scwnd = 1;
19584 		else
19585 			rack->rack_enable_scwnd = 0;
19586 		if (rack_do_dyn_mul) {
19587 			/* When dynamic adjustment is on CA needs to start at 100% */
19588 			rack->rc_gp_dyn_mul = 1;
19589 			if (rack_do_dyn_mul >= 100)
19590 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
19591 		} else {
19592 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
19593 			rack->rc_gp_dyn_mul = 0;
19594 		}
19595 		rack->r_rr_config = 0;
19596 		rack->r_ctl.rc_no_push_at_mrtt = 0;
19597 		rack->rc_pace_to_cwnd = 0;
19598 		rack->rc_pace_fill_if_rttin_range = 0;
19599 		rack->rtt_limit_mul = 0;
19600 
19601 		if (rack_enable_hw_pacing)
19602 			rack->rack_hdw_pace_ena = 1;
19603 		else
19604 			rack->rack_hdw_pace_ena = 0;
19605 		if (rack_disable_prr)
19606 			rack->rack_no_prr = 1;
19607 		else
19608 			rack->rack_no_prr = 0;
19609 		if (rack_limits_scwnd)
19610 			rack->r_limit_scw  = 1;
19611 		else
19612 			rack->r_limit_scw  = 0;
19613 		err = 0;
19614 	}
19615 	return (err);
19616 }
19617 
19618 static int
19619 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
19620 {
19621 	struct deferred_opt_list *dol;
19622 
19623 	dol = malloc(sizeof(struct deferred_opt_list),
19624 		     M_TCPFSB, M_NOWAIT|M_ZERO);
19625 	if (dol == NULL) {
19626 		/*
19627 		 * No space yikes -- fail out..
19628 		 */
19629 		return (0);
19630 	}
19631 	dol->optname = sopt_name;
19632 	dol->optval = loptval;
19633 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
19634 	return (1);
19635 }
19636 
19637 static int
19638 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
19639 		    uint32_t optval, uint64_t loptval)
19640 {
19641 	struct epoch_tracker et;
19642 	struct sockopt sopt;
19643 	struct cc_newreno_opts opt;
19644 	struct inpcb *inp = tptoinpcb(tp);
19645 	uint64_t val;
19646 	int error = 0;
19647 	uint16_t ca, ss;
19648 
19649 	switch (sopt_name) {
19650 
19651 	case TCP_RACK_DSACK_OPT:
19652 		RACK_OPTS_INC(tcp_rack_dsack_opt);
19653 		if (optval & 0x1) {
19654 			rack->rc_rack_tmr_std_based = 1;
19655 		} else {
19656 			rack->rc_rack_tmr_std_based = 0;
19657 		}
19658 		if (optval & 0x2) {
19659 			rack->rc_rack_use_dsack = 1;
19660 		} else {
19661 			rack->rc_rack_use_dsack = 0;
19662 		}
19663 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
19664 		break;
19665 	case TCP_RACK_PACING_BETA:
19666 		RACK_OPTS_INC(tcp_rack_beta);
19667 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
19668 			/* This only works for newreno. */
19669 			error = EINVAL;
19670 			break;
19671 		}
19672 		if (rack->rc_pacing_cc_set) {
19673 			/*
19674 			 * Set them into the real CC module
19675 			 * whats in the rack pcb is the old values
19676 			 * to be used on restoral/
19677 			 */
19678 			sopt.sopt_dir = SOPT_SET;
19679 			opt.name = CC_NEWRENO_BETA;
19680 			opt.val = optval;
19681 			if (CC_ALGO(tp)->ctl_output != NULL)
19682 				error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
19683 			else {
19684 				error = ENOENT;
19685 				break;
19686 			}
19687 		} else {
19688 			/*
19689 			 * Not pacing yet so set it into our local
19690 			 * rack pcb storage.
19691 			 */
19692 			rack->r_ctl.rc_saved_beta.beta = optval;
19693 		}
19694 		break;
19695 	case TCP_RACK_TIMER_SLOP:
19696 		RACK_OPTS_INC(tcp_rack_timer_slop);
19697 		rack->r_ctl.timer_slop = optval;
19698 		if (rack->rc_tp->t_srtt) {
19699 			/*
19700 			 * If we have an SRTT lets update t_rxtcur
19701 			 * to have the new slop.
19702 			 */
19703 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
19704 					   rack_rto_min, rack_rto_max,
19705 					   rack->r_ctl.timer_slop);
19706 		}
19707 		break;
19708 	case TCP_RACK_PACING_BETA_ECN:
19709 		RACK_OPTS_INC(tcp_rack_beta_ecn);
19710 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) {
19711 			/* This only works for newreno. */
19712 			error = EINVAL;
19713 			break;
19714 		}
19715 		if (rack->rc_pacing_cc_set) {
19716 			/*
19717 			 * Set them into the real CC module
19718 			 * whats in the rack pcb is the old values
19719 			 * to be used on restoral/
19720 			 */
19721 			sopt.sopt_dir = SOPT_SET;
19722 			opt.name = CC_NEWRENO_BETA_ECN;
19723 			opt.val = optval;
19724 			if (CC_ALGO(tp)->ctl_output != NULL)
19725 				error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt);
19726 			else
19727 				error = ENOENT;
19728 		} else {
19729 			/*
19730 			 * Not pacing yet so set it into our local
19731 			 * rack pcb storage.
19732 			 */
19733 			rack->r_ctl.rc_saved_beta.beta_ecn = optval;
19734 			rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
19735 		}
19736 		break;
19737 	case TCP_DEFER_OPTIONS:
19738 		RACK_OPTS_INC(tcp_defer_opt);
19739 		if (optval) {
19740 			if (rack->gp_ready) {
19741 				/* Too late */
19742 				error = EINVAL;
19743 				break;
19744 			}
19745 			rack->defer_options = 1;
19746 		} else
19747 			rack->defer_options = 0;
19748 		break;
19749 	case TCP_RACK_MEASURE_CNT:
19750 		RACK_OPTS_INC(tcp_rack_measure_cnt);
19751 		if (optval && (optval <= 0xff)) {
19752 			rack->r_ctl.req_measurements = optval;
19753 		} else
19754 			error = EINVAL;
19755 		break;
19756 	case TCP_REC_ABC_VAL:
19757 		RACK_OPTS_INC(tcp_rec_abc_val);
19758 		if (optval > 0)
19759 			rack->r_use_labc_for_rec = 1;
19760 		else
19761 			rack->r_use_labc_for_rec = 0;
19762 		break;
19763 	case TCP_RACK_ABC_VAL:
19764 		RACK_OPTS_INC(tcp_rack_abc_val);
19765 		if ((optval > 0) && (optval < 255))
19766 			rack->rc_labc = optval;
19767 		else
19768 			error = EINVAL;
19769 		break;
19770 	case TCP_HDWR_UP_ONLY:
19771 		RACK_OPTS_INC(tcp_pacing_up_only);
19772 		if (optval)
19773 			rack->r_up_only = 1;
19774 		else
19775 			rack->r_up_only = 0;
19776 		break;
19777 	case TCP_PACING_RATE_CAP:
19778 		RACK_OPTS_INC(tcp_pacing_rate_cap);
19779 		rack->r_ctl.bw_rate_cap = loptval;
19780 		break;
19781 	case TCP_RACK_PROFILE:
19782 		RACK_OPTS_INC(tcp_profile);
19783 		error = rack_set_profile(rack, optval);
19784 		break;
19785 	case TCP_USE_CMP_ACKS:
19786 		RACK_OPTS_INC(tcp_use_cmp_acks);
19787 		if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) {
19788 			/* You can't turn it off once its on! */
19789 			error = EINVAL;
19790 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
19791 			rack->r_use_cmp_ack = 1;
19792 			rack->r_mbuf_queue = 1;
19793 			inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19794 		}
19795 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
19796 			inp->inp_flags2 |= INP_MBUF_ACKCMP;
19797 		break;
19798 	case TCP_SHARED_CWND_TIME_LIMIT:
19799 		RACK_OPTS_INC(tcp_lscwnd);
19800 		if (optval)
19801 			rack->r_limit_scw = 1;
19802 		else
19803 			rack->r_limit_scw = 0;
19804 		break;
19805  	case TCP_RACK_PACE_TO_FILL:
19806 		RACK_OPTS_INC(tcp_fillcw);
19807 		if (optval == 0)
19808 			rack->rc_pace_to_cwnd = 0;
19809 		else {
19810 			rack->rc_pace_to_cwnd = 1;
19811 			if (optval > 1)
19812 				rack->r_fill_less_agg = 1;
19813 		}
19814 		if ((optval >= rack_gp_rtt_maxmul) &&
19815 		    rack_gp_rtt_maxmul &&
19816 		    (optval < 0xf)) {
19817 			rack->rc_pace_fill_if_rttin_range = 1;
19818 			rack->rtt_limit_mul = optval;
19819 		} else {
19820 			rack->rc_pace_fill_if_rttin_range = 0;
19821 			rack->rtt_limit_mul = 0;
19822 		}
19823 		break;
19824 	case TCP_RACK_NO_PUSH_AT_MAX:
19825 		RACK_OPTS_INC(tcp_npush);
19826 		if (optval == 0)
19827 			rack->r_ctl.rc_no_push_at_mrtt = 0;
19828 		else if (optval < 0xff)
19829 			rack->r_ctl.rc_no_push_at_mrtt = optval;
19830 		else
19831 			error = EINVAL;
19832 		break;
19833 	case TCP_SHARED_CWND_ENABLE:
19834 		RACK_OPTS_INC(tcp_rack_scwnd);
19835 		if (optval == 0)
19836 			rack->rack_enable_scwnd = 0;
19837 		else
19838 			rack->rack_enable_scwnd = 1;
19839 		break;
19840 	case TCP_RACK_MBUF_QUEUE:
19841 		/* Now do we use the LRO mbuf-queue feature */
19842 		RACK_OPTS_INC(tcp_rack_mbufq);
19843 		if (optval || rack->r_use_cmp_ack)
19844 			rack->r_mbuf_queue = 1;
19845 		else
19846 			rack->r_mbuf_queue = 0;
19847 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19848 			inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19849 		else
19850 			inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19851 		break;
19852 	case TCP_RACK_NONRXT_CFG_RATE:
19853 		RACK_OPTS_INC(tcp_rack_cfg_rate);
19854 		if (optval == 0)
19855 			rack->rack_rec_nonrxt_use_cr = 0;
19856 		else
19857 			rack->rack_rec_nonrxt_use_cr = 1;
19858 		break;
19859 	case TCP_NO_PRR:
19860 		RACK_OPTS_INC(tcp_rack_noprr);
19861 		if (optval == 0)
19862 			rack->rack_no_prr = 0;
19863 		else if (optval == 1)
19864 			rack->rack_no_prr = 1;
19865 		else if (optval == 2)
19866 			rack->no_prr_addback = 1;
19867 		else
19868 			error = EINVAL;
19869 		break;
19870 	case TCP_TIMELY_DYN_ADJ:
19871 		RACK_OPTS_INC(tcp_timely_dyn);
19872 		if (optval == 0)
19873 			rack->rc_gp_dyn_mul = 0;
19874 		else {
19875 			rack->rc_gp_dyn_mul = 1;
19876 			if (optval >= 100) {
19877 				/*
19878 				 * If the user sets something 100 or more
19879 				 * its the gp_ca value.
19880 				 */
19881 				rack->r_ctl.rack_per_of_gp_ca  = optval;
19882 			}
19883 		}
19884 		break;
19885 	case TCP_RACK_DO_DETECTION:
19886 		RACK_OPTS_INC(tcp_rack_do_detection);
19887 		if (optval == 0)
19888 			rack->do_detection = 0;
19889 		else
19890 			rack->do_detection = 1;
19891 		break;
19892 	case TCP_RACK_TLP_USE:
19893 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
19894 			error = EINVAL;
19895 			break;
19896 		}
19897 		RACK_OPTS_INC(tcp_tlp_use);
19898 		rack->rack_tlp_threshold_use = optval;
19899 		break;
19900 	case TCP_RACK_TLP_REDUCE:
19901 		/* RACK TLP cwnd reduction (bool) */
19902 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
19903 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
19904 		break;
19905 	/*  Pacing related ones */
19906 	case TCP_RACK_PACE_ALWAYS:
19907 		/*
19908 		 * zero is old rack method, 1 is new
19909 		 * method using a pacing rate.
19910 		 */
19911 		RACK_OPTS_INC(tcp_rack_pace_always);
19912 		if (optval > 0) {
19913 			if (rack->rc_always_pace) {
19914 				error = EALREADY;
19915 				break;
19916 			} else if (tcp_can_enable_pacing()) {
19917 				rack->rc_always_pace = 1;
19918 				if (rack->use_fixed_rate || rack->gp_ready)
19919 					rack_set_cc_pacing(rack);
19920 			}
19921 			else {
19922 				error = ENOSPC;
19923 				break;
19924 			}
19925 		} else {
19926 			if (rack->rc_always_pace) {
19927 				tcp_decrement_paced_conn();
19928 				rack->rc_always_pace = 0;
19929 				rack_undo_cc_pacing(rack);
19930 			}
19931 		}
19932 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
19933 			inp->inp_flags2 |= INP_SUPPORTS_MBUFQ;
19934 		else
19935 			inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ;
19936 		/* A rate may be set irate or other, if so set seg size */
19937 		rack_update_seg(rack);
19938 		break;
19939 	case TCP_BBR_RACK_INIT_RATE:
19940 		RACK_OPTS_INC(tcp_initial_rate);
19941 		val = optval;
19942 		/* Change from kbits per second to bytes per second */
19943 		val *= 1000;
19944 		val /= 8;
19945 		rack->r_ctl.init_rate = val;
19946 		if (rack->rc_init_win != rack_default_init_window) {
19947 			uint32_t win, snt;
19948 
19949 			/*
19950 			 * Options don't always get applied
19951 			 * in the order you think. So in order
19952 			 * to assure we update a cwnd we need
19953 			 * to check and see if we are still
19954 			 * where we should raise the cwnd.
19955 			 */
19956 			win = rc_init_window(rack);
19957 			if (SEQ_GT(tp->snd_max, tp->iss))
19958 				snt = tp->snd_max - tp->iss;
19959 			else
19960 				snt = 0;
19961 			if ((snt < win) &&
19962 			    (tp->snd_cwnd < win))
19963 				tp->snd_cwnd = win;
19964 		}
19965 		if (rack->rc_always_pace)
19966 			rack_update_seg(rack);
19967 		break;
19968 	case TCP_BBR_IWINTSO:
19969 		RACK_OPTS_INC(tcp_initial_win);
19970 		if (optval && (optval <= 0xff)) {
19971 			uint32_t win, snt;
19972 
19973 			rack->rc_init_win = optval;
19974 			win = rc_init_window(rack);
19975 			if (SEQ_GT(tp->snd_max, tp->iss))
19976 				snt = tp->snd_max - tp->iss;
19977 			else
19978 				snt = 0;
19979 			if ((snt < win) &&
19980 			    (tp->t_srtt |
19981 #ifdef NETFLIX_PEAKRATE
19982 			     tp->t_maxpeakrate |
19983 #endif
19984 			     rack->r_ctl.init_rate)) {
19985 				/*
19986 				 * We are not past the initial window
19987 				 * and we have some bases for pacing,
19988 				 * so we need to possibly adjust up
19989 				 * the cwnd. Note even if we don't set
19990 				 * the cwnd, its still ok to raise the rc_init_win
19991 				 * which can be used coming out of idle when we
19992 				 * would have a rate.
19993 				 */
19994 				if (tp->snd_cwnd < win)
19995 					tp->snd_cwnd = win;
19996 			}
19997 			if (rack->rc_always_pace)
19998 				rack_update_seg(rack);
19999 		} else
20000 			error = EINVAL;
20001 		break;
20002 	case TCP_RACK_FORCE_MSEG:
20003 		RACK_OPTS_INC(tcp_rack_force_max_seg);
20004 		if (optval)
20005 			rack->rc_force_max_seg = 1;
20006 		else
20007 			rack->rc_force_max_seg = 0;
20008 		break;
20009 	case TCP_RACK_PACE_MAX_SEG:
20010 		/* Max segments size in a pace in bytes */
20011 		RACK_OPTS_INC(tcp_rack_max_seg);
20012 		rack->rc_user_set_max_segs = optval;
20013 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
20014 		break;
20015 	case TCP_RACK_PACE_RATE_REC:
20016 		/* Set the fixed pacing rate in Bytes per second ca */
20017 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
20018 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
20019 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
20020 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
20021 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
20022 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
20023 		rack->use_fixed_rate = 1;
20024 		if (rack->rc_always_pace)
20025 			rack_set_cc_pacing(rack);
20026 		rack_log_pacing_delay_calc(rack,
20027 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
20028 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
20029 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
20030 					   __LINE__, NULL,0);
20031 		break;
20032 
20033 	case TCP_RACK_PACE_RATE_SS:
20034 		/* Set the fixed pacing rate in Bytes per second ca */
20035 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
20036 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
20037 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
20038 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
20039 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
20040 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
20041 		rack->use_fixed_rate = 1;
20042 		if (rack->rc_always_pace)
20043 			rack_set_cc_pacing(rack);
20044 		rack_log_pacing_delay_calc(rack,
20045 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
20046 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
20047 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
20048 					   __LINE__, NULL, 0);
20049 		break;
20050 
20051 	case TCP_RACK_PACE_RATE_CA:
20052 		/* Set the fixed pacing rate in Bytes per second ca */
20053 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
20054 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
20055 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
20056 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
20057 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
20058 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
20059 		rack->use_fixed_rate = 1;
20060 		if (rack->rc_always_pace)
20061 			rack_set_cc_pacing(rack);
20062 		rack_log_pacing_delay_calc(rack,
20063 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
20064 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
20065 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
20066 					   __LINE__, NULL, 0);
20067 		break;
20068 	case TCP_RACK_GP_INCREASE_REC:
20069 		RACK_OPTS_INC(tcp_gp_inc_rec);
20070 		rack->r_ctl.rack_per_of_gp_rec = optval;
20071 		rack_log_pacing_delay_calc(rack,
20072 					   rack->r_ctl.rack_per_of_gp_ss,
20073 					   rack->r_ctl.rack_per_of_gp_ca,
20074 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20075 					   __LINE__, NULL, 0);
20076 		break;
20077 	case TCP_RACK_GP_INCREASE_CA:
20078 		RACK_OPTS_INC(tcp_gp_inc_ca);
20079 		ca = optval;
20080 		if (ca < 100) {
20081 			/*
20082 			 * We don't allow any reduction
20083 			 * over the GP b/w.
20084 			 */
20085 			error = EINVAL;
20086 			break;
20087 		}
20088 		rack->r_ctl.rack_per_of_gp_ca = ca;
20089 		rack_log_pacing_delay_calc(rack,
20090 					   rack->r_ctl.rack_per_of_gp_ss,
20091 					   rack->r_ctl.rack_per_of_gp_ca,
20092 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20093 					   __LINE__, NULL, 0);
20094 		break;
20095 	case TCP_RACK_GP_INCREASE_SS:
20096 		RACK_OPTS_INC(tcp_gp_inc_ss);
20097 		ss = optval;
20098 		if (ss < 100) {
20099 			/*
20100 			 * We don't allow any reduction
20101 			 * over the GP b/w.
20102 			 */
20103 			error = EINVAL;
20104 			break;
20105 		}
20106 		rack->r_ctl.rack_per_of_gp_ss = ss;
20107 		rack_log_pacing_delay_calc(rack,
20108 					   rack->r_ctl.rack_per_of_gp_ss,
20109 					   rack->r_ctl.rack_per_of_gp_ca,
20110 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
20111 					   __LINE__, NULL, 0);
20112 		break;
20113 	case TCP_RACK_RR_CONF:
20114 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
20115 		if (optval && optval <= 3)
20116 			rack->r_rr_config = optval;
20117 		else
20118 			rack->r_rr_config = 0;
20119 		break;
20120 	case TCP_HDWR_RATE_CAP:
20121 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
20122 		if (optval) {
20123 			if (rack->r_rack_hw_rate_caps == 0)
20124 				rack->r_rack_hw_rate_caps = 1;
20125 			else
20126 				error = EALREADY;
20127 		} else {
20128 			rack->r_rack_hw_rate_caps = 0;
20129 		}
20130 		break;
20131 	case TCP_BBR_HDWR_PACE:
20132 		RACK_OPTS_INC(tcp_hdwr_pacing);
20133 		if (optval){
20134 			if (rack->rack_hdrw_pacing == 0) {
20135 				rack->rack_hdw_pace_ena = 1;
20136 				rack->rack_attempt_hdwr_pace = 0;
20137 			} else
20138 				error = EALREADY;
20139 		} else {
20140 			rack->rack_hdw_pace_ena = 0;
20141 #ifdef RATELIMIT
20142 			if (rack->r_ctl.crte != NULL) {
20143 				rack->rack_hdrw_pacing = 0;
20144 				rack->rack_attempt_hdwr_pace = 0;
20145 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
20146 				rack->r_ctl.crte = NULL;
20147 			}
20148 #endif
20149 		}
20150 		break;
20151 	/*  End Pacing related ones */
20152 	case TCP_RACK_PRR_SENDALOT:
20153 		/* Allow PRR to send more than one seg */
20154 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
20155 		rack->r_ctl.rc_prr_sendalot = optval;
20156 		break;
20157 	case TCP_RACK_MIN_TO:
20158 		/* Minimum time between rack t-o's in ms */
20159 		RACK_OPTS_INC(tcp_rack_min_to);
20160 		rack->r_ctl.rc_min_to = optval;
20161 		break;
20162 	case TCP_RACK_EARLY_SEG:
20163 		/* If early recovery max segments */
20164 		RACK_OPTS_INC(tcp_rack_early_seg);
20165 		rack->r_ctl.rc_early_recovery_segs = optval;
20166 		break;
20167 	case TCP_RACK_ENABLE_HYSTART:
20168 	{
20169 		if (optval) {
20170 			tp->ccv->flags |= CCF_HYSTART_ALLOWED;
20171 			if (rack_do_hystart > RACK_HYSTART_ON)
20172 				tp->ccv->flags |= CCF_HYSTART_CAN_SH_CWND;
20173 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
20174 				tp->ccv->flags |= CCF_HYSTART_CONS_SSTH;
20175 		} else {
20176 			tp->ccv->flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
20177 		}
20178 	}
20179 	break;
20180 	case TCP_RACK_REORD_THRESH:
20181 		/* RACK reorder threshold (shift amount) */
20182 		RACK_OPTS_INC(tcp_rack_reord_thresh);
20183 		if ((optval > 0) && (optval < 31))
20184 			rack->r_ctl.rc_reorder_shift = optval;
20185 		else
20186 			error = EINVAL;
20187 		break;
20188 	case TCP_RACK_REORD_FADE:
20189 		/* Does reordering fade after ms time */
20190 		RACK_OPTS_INC(tcp_rack_reord_fade);
20191 		rack->r_ctl.rc_reorder_fade = optval;
20192 		break;
20193 	case TCP_RACK_TLP_THRESH:
20194 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
20195 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
20196 		if (optval)
20197 			rack->r_ctl.rc_tlp_threshold = optval;
20198 		else
20199 			error = EINVAL;
20200 		break;
20201 	case TCP_BBR_USE_RACK_RR:
20202 		RACK_OPTS_INC(tcp_rack_rr);
20203 		if (optval)
20204 			rack->use_rack_rr = 1;
20205 		else
20206 			rack->use_rack_rr = 0;
20207 		break;
20208 	case TCP_FAST_RSM_HACK:
20209 		RACK_OPTS_INC(tcp_rack_fastrsm_hack);
20210 		if (optval)
20211 			rack->fast_rsm_hack = 1;
20212 		else
20213 			rack->fast_rsm_hack = 0;
20214 		break;
20215 	case TCP_RACK_PKT_DELAY:
20216 		/* RACK added ms i.e. rack-rtt + reord + N */
20217 		RACK_OPTS_INC(tcp_rack_pkt_delay);
20218 		rack->r_ctl.rc_pkt_delay = optval;
20219 		break;
20220 	case TCP_DELACK:
20221 		RACK_OPTS_INC(tcp_rack_delayed_ack);
20222 		if (optval == 0)
20223 			tp->t_delayed_ack = 0;
20224 		else
20225 			tp->t_delayed_ack = 1;
20226 		if (tp->t_flags & TF_DELACK) {
20227 			tp->t_flags &= ~TF_DELACK;
20228 			tp->t_flags |= TF_ACKNOW;
20229 			NET_EPOCH_ENTER(et);
20230 			rack_output(tp);
20231 			NET_EPOCH_EXIT(et);
20232 		}
20233 		break;
20234 
20235 	case TCP_BBR_RACK_RTT_USE:
20236 		RACK_OPTS_INC(tcp_rack_rtt_use);
20237 		if ((optval != USE_RTT_HIGH) &&
20238 		    (optval != USE_RTT_LOW) &&
20239 		    (optval != USE_RTT_AVG))
20240 			error = EINVAL;
20241 		else
20242 			rack->r_ctl.rc_rate_sample_method = optval;
20243 		break;
20244 	case TCP_DATA_AFTER_CLOSE:
20245 		RACK_OPTS_INC(tcp_data_after_close);
20246 		if (optval)
20247 			rack->rc_allow_data_af_clo = 1;
20248 		else
20249 			rack->rc_allow_data_af_clo = 0;
20250 		break;
20251 	default:
20252 		break;
20253 	}
20254 #ifdef NETFLIX_STATS
20255 	tcp_log_socket_option(tp, sopt_name, optval, error);
20256 #endif
20257 	return (error);
20258 }
20259 
20260 
20261 static void
20262 rack_apply_deferred_options(struct tcp_rack *rack)
20263 {
20264 	struct deferred_opt_list *dol, *sdol;
20265 	uint32_t s_optval;
20266 
20267 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
20268 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
20269 		/* Disadvantage of deferal is you loose the error return */
20270 		s_optval = (uint32_t)dol->optval;
20271 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval);
20272 		free(dol, M_TCPDO);
20273 	}
20274 }
20275 
20276 static void
20277 rack_hw_tls_change(struct tcpcb *tp, int chg)
20278 {
20279 	/*
20280 	 * HW tls state has changed.. fix all
20281 	 * rsm's in flight.
20282 	 */
20283 	struct tcp_rack *rack;
20284 	struct rack_sendmap *rsm;
20285 
20286 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20287 	RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) {
20288 		if (chg)
20289 			rsm->r_hw_tls = 1;
20290 		else
20291 			rsm->r_hw_tls = 0;
20292 	}
20293 	if (chg)
20294 		rack->r_ctl.fsb.hw_tls = 1;
20295 	else
20296 		rack->r_ctl.fsb.hw_tls = 0;
20297 }
20298 
20299 static int
20300 rack_pru_options(struct tcpcb *tp, int flags)
20301 {
20302 	if (flags & PRUS_OOB)
20303 		return (EOPNOTSUPP);
20304 	return (0);
20305 }
20306 
20307 static struct tcp_function_block __tcp_rack = {
20308 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
20309 	.tfb_tcp_output = rack_output,
20310 	.tfb_do_queued_segments = ctf_do_queued_segments,
20311 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
20312 	.tfb_tcp_do_segment = rack_do_segment,
20313 	.tfb_tcp_ctloutput = rack_ctloutput,
20314 	.tfb_tcp_fb_init = rack_init,
20315 	.tfb_tcp_fb_fini = rack_fini,
20316 	.tfb_tcp_timer_stop_all = rack_stopall,
20317 	.tfb_tcp_timer_activate = rack_timer_activate,
20318 	.tfb_tcp_timer_active = rack_timer_active,
20319 	.tfb_tcp_timer_stop = rack_timer_stop,
20320 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
20321 	.tfb_tcp_handoff_ok = rack_handoff_ok,
20322 	.tfb_tcp_mtu_chg = rack_mtu_change,
20323 	.tfb_pru_options = rack_pru_options,
20324 	.tfb_hwtls_change = rack_hw_tls_change,
20325 	.tfb_compute_pipe = rack_compute_pipe,
20326 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
20327 };
20328 
20329 /*
20330  * rack_ctloutput() must drop the inpcb lock before performing copyin on
20331  * socket option arguments.  When it re-acquires the lock after the copy, it
20332  * has to revalidate that the connection is still valid for the socket
20333  * option.
20334  */
20335 static int
20336 rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt)
20337 {
20338 #ifdef INET6
20339 	struct ip6_hdr *ip6;
20340 #endif
20341 #ifdef INET
20342 	struct ip *ip;
20343 #endif
20344 	struct tcpcb *tp;
20345 	struct tcp_rack *rack;
20346 	uint64_t loptval;
20347 	int32_t error = 0, optval;
20348 
20349 	tp = intotcpcb(inp);
20350 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20351 	if (rack == NULL) {
20352 		INP_WUNLOCK(inp);
20353 		return (EINVAL);
20354 	}
20355 #ifdef INET6
20356 	ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
20357 #endif
20358 #ifdef INET
20359 	ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
20360 #endif
20361 
20362 	switch (sopt->sopt_level) {
20363 #ifdef INET6
20364 	case IPPROTO_IPV6:
20365 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
20366 		switch (sopt->sopt_name) {
20367 		case IPV6_USE_MIN_MTU:
20368 			tcp6_use_min_mtu(tp);
20369 			break;
20370 		case IPV6_TCLASS:
20371 			/*
20372 			 * The DSCP codepoint has changed, update the fsb.
20373 			 */
20374 			ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
20375 			    (rack->rc_inp->inp_flow & IPV6_FLOWINFO_MASK);
20376 			break;
20377 		}
20378 		INP_WUNLOCK(inp);
20379 		return (0);
20380 #endif
20381 #ifdef INET
20382 	case IPPROTO_IP:
20383 		switch (sopt->sopt_name) {
20384 		case IP_TOS:
20385 			/*
20386 			 * The DSCP codepoint has changed, update the fsb.
20387 			 */
20388 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
20389 			break;
20390 		case IP_TTL:
20391 			/*
20392 			 * The TTL has changed, update the fsb.
20393 			 */
20394 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
20395 			break;
20396 		}
20397 		INP_WUNLOCK(inp);
20398 		return (0);
20399 #endif
20400 	}
20401 
20402 	switch (sopt->sopt_name) {
20403 	case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
20404 	/*  Pacing related ones */
20405 	case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
20406 	case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
20407 	case TCP_BBR_IWINTSO:			/*  URL:tso_iwin */
20408 	case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
20409 	case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
20410 	case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
20411 	case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
20412 	case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
20413 	case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
20414 	case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
20415 	case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
20416 	case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
20417 	case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
20418 	case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
20419 	case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
20420 	case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
20421        /* End pacing related */
20422 	case TCP_FAST_RSM_HACK:			/*  URL:frsm_hack */
20423 	case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
20424 	case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
20425 	case TCP_RACK_MIN_TO:			/*  URL:min_to */
20426 	case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
20427 	case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
20428 	case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
20429 	case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
20430 	case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
20431 	case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
20432 	case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
20433 	case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
20434 	case TCP_RACK_DO_DETECTION:		/*  URL:detect */
20435 	case TCP_NO_PRR:			/*  URL:noprr */
20436 	case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
20437 	case TCP_DATA_AFTER_CLOSE:		/*  no URL */
20438 	case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
20439 	case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
20440 	case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
20441 	case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
20442 	case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
20443 	case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
20444 	case TCP_RACK_PROFILE:			/*  URL:profile */
20445 	case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
20446 	case TCP_RACK_ABC_VAL:			/*  URL:labc */
20447 	case TCP_REC_ABC_VAL:			/*  URL:reclabc */
20448 	case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
20449 	case TCP_DEFER_OPTIONS:			/*  URL:defer */
20450 	case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
20451 	case TCP_RACK_PACING_BETA:		/*  URL:pacing_beta */
20452 	case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
20453 	case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
20454 	case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
20455 		break;
20456 	default:
20457 		/* Filter off all unknown options to the base stack */
20458 		return (tcp_default_ctloutput(inp, sopt));
20459 		break;
20460 	}
20461 	INP_WUNLOCK(inp);
20462 	if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
20463 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
20464 		/*
20465 		 * We truncate it down to 32 bits for the socket-option trace this
20466 		 * means rates > 34Gbps won't show right, but thats probably ok.
20467 		 */
20468 		optval = (uint32_t)loptval;
20469 	} else {
20470 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
20471 		/* Save it in 64 bit form too */
20472 		loptval = optval;
20473 	}
20474 	if (error)
20475 		return (error);
20476 	INP_WLOCK(inp);
20477 	if (inp->inp_flags & INP_DROPPED) {
20478 		INP_WUNLOCK(inp);
20479 		return (ECONNRESET);
20480 	}
20481 	if (tp->t_fb != &__tcp_rack) {
20482 		INP_WUNLOCK(inp);
20483 		return (ENOPROTOOPT);
20484 	}
20485 	if (rack->defer_options && (rack->gp_ready == 0) &&
20486 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
20487 	    (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
20488 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
20489 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
20490 		/* Options are beind deferred */
20491 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
20492 			INP_WUNLOCK(inp);
20493 			return (0);
20494 		} else {
20495 			/* No memory to defer, fail */
20496 			INP_WUNLOCK(inp);
20497 			return (ENOMEM);
20498 		}
20499 	}
20500 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval);
20501 	INP_WUNLOCK(inp);
20502 	return (error);
20503 }
20504 
20505 static void
20506 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
20507 {
20508 
20509 	INP_WLOCK_ASSERT(tptoinpcb(tp));
20510 	bzero(ti, sizeof(*ti));
20511 
20512 	ti->tcpi_state = tp->t_state;
20513 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
20514 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
20515 	if (tp->t_flags & TF_SACK_PERMIT)
20516 		ti->tcpi_options |= TCPI_OPT_SACK;
20517 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
20518 		ti->tcpi_options |= TCPI_OPT_WSCALE;
20519 		ti->tcpi_snd_wscale = tp->snd_scale;
20520 		ti->tcpi_rcv_wscale = tp->rcv_scale;
20521 	}
20522 	if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
20523 		ti->tcpi_options |= TCPI_OPT_ECN;
20524 	if (tp->t_flags & TF_FASTOPEN)
20525 		ti->tcpi_options |= TCPI_OPT_TFO;
20526 	/* still kept in ticks is t_rcvtime */
20527 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
20528 	/* Since we hold everything in precise useconds this is easy */
20529 	ti->tcpi_rtt = tp->t_srtt;
20530 	ti->tcpi_rttvar = tp->t_rttvar;
20531 	ti->tcpi_rto = tp->t_rxtcur;
20532 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
20533 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
20534 	/*
20535 	 * FreeBSD-specific extension fields for tcp_info.
20536 	 */
20537 	ti->tcpi_rcv_space = tp->rcv_wnd;
20538 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
20539 	ti->tcpi_snd_wnd = tp->snd_wnd;
20540 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
20541 	ti->tcpi_snd_nxt = tp->snd_nxt;
20542 	ti->tcpi_snd_mss = tp->t_maxseg;
20543 	ti->tcpi_rcv_mss = tp->t_maxseg;
20544 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
20545 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
20546 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
20547 #ifdef NETFLIX_STATS
20548 	ti->tcpi_total_tlp = tp->t_sndtlppack;
20549 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
20550 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
20551 #endif
20552 #ifdef TCP_OFFLOAD
20553 	if (tp->t_flags & TF_TOE) {
20554 		ti->tcpi_options |= TCPI_OPT_TOE;
20555 		tcp_offload_tcp_info(tp, ti);
20556 	}
20557 #endif
20558 }
20559 
20560 static int
20561 rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt)
20562 {
20563 	struct tcpcb *tp;
20564 	struct tcp_rack *rack;
20565 	int32_t error, optval;
20566 	uint64_t val, loptval;
20567 	struct	tcp_info ti;
20568 	/*
20569 	 * Because all our options are either boolean or an int, we can just
20570 	 * pull everything into optval and then unlock and copy. If we ever
20571 	 * add a option that is not a int, then this will have quite an
20572 	 * impact to this routine.
20573 	 */
20574 	error = 0;
20575 	tp = intotcpcb(inp);
20576 	rack = (struct tcp_rack *)tp->t_fb_ptr;
20577 	if (rack == NULL) {
20578 		INP_WUNLOCK(inp);
20579 		return (EINVAL);
20580 	}
20581 	switch (sopt->sopt_name) {
20582 	case TCP_INFO:
20583 		/* First get the info filled */
20584 		rack_fill_info(tp, &ti);
20585 		/* Fix up the rtt related fields if needed */
20586 		INP_WUNLOCK(inp);
20587 		error = sooptcopyout(sopt, &ti, sizeof ti);
20588 		return (error);
20589 	/*
20590 	 * Beta is the congestion control value for NewReno that influences how
20591 	 * much of a backoff happens when loss is detected. It is normally set
20592 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
20593 	 * when you exit recovery.
20594 	 */
20595 	case TCP_RACK_PACING_BETA:
20596 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
20597 			error = EINVAL;
20598 		else if (rack->rc_pacing_cc_set == 0)
20599 			optval = rack->r_ctl.rc_saved_beta.beta;
20600 		else {
20601 			/*
20602 			 * Reach out into the CC data and report back what
20603 			 * I have previously set. Yeah it looks hackish but
20604 			 * we don't want to report the saved values.
20605 			 */
20606 			if (tp->ccv->cc_data)
20607 				optval = ((struct newreno *)tp->ccv->cc_data)->beta;
20608 			else
20609 				error = EINVAL;
20610 		}
20611 		break;
20612 		/*
20613 		 * Beta_ecn is the congestion control value for NewReno that influences how
20614 		 * much of a backoff happens when a ECN mark is detected. It is normally set
20615 		 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
20616 		 * you exit recovery. Note that classic ECN has a beta of 50, it is only
20617 		 * ABE Ecn that uses this "less" value, but we do too with pacing :)
20618 		 */
20619 
20620 	case TCP_RACK_PACING_BETA_ECN:
20621 		if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0)
20622 			error = EINVAL;
20623 		else if (rack->rc_pacing_cc_set == 0)
20624 			optval = rack->r_ctl.rc_saved_beta.beta_ecn;
20625 		else {
20626 			/*
20627 			 * Reach out into the CC data and report back what
20628 			 * I have previously set. Yeah it looks hackish but
20629 			 * we don't want to report the saved values.
20630 			 */
20631 			if (tp->ccv->cc_data)
20632 				optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn;
20633 			else
20634 				error = EINVAL;
20635 		}
20636 		break;
20637 	case TCP_RACK_DSACK_OPT:
20638 		optval = 0;
20639 		if (rack->rc_rack_tmr_std_based) {
20640 			optval |= 1;
20641 		}
20642 		if (rack->rc_rack_use_dsack) {
20643 			optval |= 2;
20644 		}
20645 		break;
20646  	case TCP_RACK_ENABLE_HYSTART:
20647 	{
20648 		if (tp->ccv->flags & CCF_HYSTART_ALLOWED) {
20649 			optval = RACK_HYSTART_ON;
20650 			if (tp->ccv->flags & CCF_HYSTART_CAN_SH_CWND)
20651 				optval = RACK_HYSTART_ON_W_SC;
20652 			if (tp->ccv->flags & CCF_HYSTART_CONS_SSTH)
20653 				optval = RACK_HYSTART_ON_W_SC_C;
20654 		} else {
20655 			optval = RACK_HYSTART_OFF;
20656 		}
20657 	}
20658 	break;
20659 	case TCP_FAST_RSM_HACK:
20660 		optval = rack->fast_rsm_hack;
20661 		break;
20662 	case TCP_DEFER_OPTIONS:
20663 		optval = rack->defer_options;
20664 		break;
20665 	case TCP_RACK_MEASURE_CNT:
20666 		optval = rack->r_ctl.req_measurements;
20667 		break;
20668 	case TCP_REC_ABC_VAL:
20669 		optval = rack->r_use_labc_for_rec;
20670 		break;
20671 	case TCP_RACK_ABC_VAL:
20672 		optval = rack->rc_labc;
20673 		break;
20674 	case TCP_HDWR_UP_ONLY:
20675 		optval= rack->r_up_only;
20676 		break;
20677 	case TCP_PACING_RATE_CAP:
20678 		loptval = rack->r_ctl.bw_rate_cap;
20679 		break;
20680 	case TCP_RACK_PROFILE:
20681 		/* You cannot retrieve a profile, its write only */
20682 		error = EINVAL;
20683 		break;
20684 	case TCP_USE_CMP_ACKS:
20685 		optval = rack->r_use_cmp_ack;
20686 		break;
20687 	case TCP_RACK_PACE_TO_FILL:
20688 		optval = rack->rc_pace_to_cwnd;
20689 		if (optval && rack->r_fill_less_agg)
20690 			optval++;
20691 		break;
20692 	case TCP_RACK_NO_PUSH_AT_MAX:
20693 		optval = rack->r_ctl.rc_no_push_at_mrtt;
20694 		break;
20695 	case TCP_SHARED_CWND_ENABLE:
20696 		optval = rack->rack_enable_scwnd;
20697 		break;
20698 	case TCP_RACK_NONRXT_CFG_RATE:
20699 		optval = rack->rack_rec_nonrxt_use_cr;
20700 		break;
20701 	case TCP_NO_PRR:
20702 		if (rack->rack_no_prr  == 1)
20703 			optval = 1;
20704 		else if (rack->no_prr_addback == 1)
20705 			optval = 2;
20706 		else
20707 			optval = 0;
20708 		break;
20709 	case TCP_RACK_DO_DETECTION:
20710 		optval = rack->do_detection;
20711 		break;
20712 	case TCP_RACK_MBUF_QUEUE:
20713 		/* Now do we use the LRO mbuf-queue feature */
20714 		optval = rack->r_mbuf_queue;
20715 		break;
20716 	case TCP_TIMELY_DYN_ADJ:
20717 		optval = rack->rc_gp_dyn_mul;
20718 		break;
20719 	case TCP_BBR_IWINTSO:
20720 		optval = rack->rc_init_win;
20721 		break;
20722 	case TCP_RACK_TLP_REDUCE:
20723 		/* RACK TLP cwnd reduction (bool) */
20724 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
20725 		break;
20726 	case TCP_BBR_RACK_INIT_RATE:
20727 		val = rack->r_ctl.init_rate;
20728 		/* convert to kbits per sec */
20729 		val *= 8;
20730 		val /= 1000;
20731 		optval = (uint32_t)val;
20732 		break;
20733 	case TCP_RACK_FORCE_MSEG:
20734 		optval = rack->rc_force_max_seg;
20735 		break;
20736 	case TCP_RACK_PACE_MAX_SEG:
20737 		/* Max segments in a pace */
20738 		optval = rack->rc_user_set_max_segs;
20739 		break;
20740 	case TCP_RACK_PACE_ALWAYS:
20741 		/* Use the always pace method */
20742 		optval = rack->rc_always_pace;
20743 		break;
20744 	case TCP_RACK_PRR_SENDALOT:
20745 		/* Allow PRR to send more than one seg */
20746 		optval = rack->r_ctl.rc_prr_sendalot;
20747 		break;
20748 	case TCP_RACK_MIN_TO:
20749 		/* Minimum time between rack t-o's in ms */
20750 		optval = rack->r_ctl.rc_min_to;
20751 		break;
20752 	case TCP_RACK_EARLY_SEG:
20753 		/* If early recovery max segments */
20754 		optval = rack->r_ctl.rc_early_recovery_segs;
20755 		break;
20756 	case TCP_RACK_REORD_THRESH:
20757 		/* RACK reorder threshold (shift amount) */
20758 		optval = rack->r_ctl.rc_reorder_shift;
20759 		break;
20760 	case TCP_RACK_REORD_FADE:
20761 		/* Does reordering fade after ms time */
20762 		optval = rack->r_ctl.rc_reorder_fade;
20763 		break;
20764 	case TCP_BBR_USE_RACK_RR:
20765 		/* Do we use the rack cheat for rxt */
20766 		optval = rack->use_rack_rr;
20767 		break;
20768 	case TCP_RACK_RR_CONF:
20769 		optval = rack->r_rr_config;
20770 		break;
20771 	case TCP_HDWR_RATE_CAP:
20772 		optval = rack->r_rack_hw_rate_caps;
20773 		break;
20774 	case TCP_BBR_HDWR_PACE:
20775 		optval = rack->rack_hdw_pace_ena;
20776 		break;
20777 	case TCP_RACK_TLP_THRESH:
20778 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
20779 		optval = rack->r_ctl.rc_tlp_threshold;
20780 		break;
20781 	case TCP_RACK_PKT_DELAY:
20782 		/* RACK added ms i.e. rack-rtt + reord + N */
20783 		optval = rack->r_ctl.rc_pkt_delay;
20784 		break;
20785 	case TCP_RACK_TLP_USE:
20786 		optval = rack->rack_tlp_threshold_use;
20787 		break;
20788 	case TCP_RACK_PACE_RATE_CA:
20789 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
20790 		break;
20791 	case TCP_RACK_PACE_RATE_SS:
20792 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
20793 		break;
20794 	case TCP_RACK_PACE_RATE_REC:
20795 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
20796 		break;
20797 	case TCP_RACK_GP_INCREASE_SS:
20798 		optval = rack->r_ctl.rack_per_of_gp_ca;
20799 		break;
20800 	case TCP_RACK_GP_INCREASE_CA:
20801 		optval = rack->r_ctl.rack_per_of_gp_ss;
20802 		break;
20803 	case TCP_BBR_RACK_RTT_USE:
20804 		optval = rack->r_ctl.rc_rate_sample_method;
20805 		break;
20806 	case TCP_DELACK:
20807 		optval = tp->t_delayed_ack;
20808 		break;
20809 	case TCP_DATA_AFTER_CLOSE:
20810 		optval = rack->rc_allow_data_af_clo;
20811 		break;
20812 	case TCP_SHARED_CWND_TIME_LIMIT:
20813 		optval = rack->r_limit_scw;
20814 		break;
20815 	case TCP_RACK_TIMER_SLOP:
20816 		optval = rack->r_ctl.timer_slop;
20817 		break;
20818 	default:
20819 		return (tcp_default_ctloutput(inp, sopt));
20820 		break;
20821 	}
20822 	INP_WUNLOCK(inp);
20823 	if (error == 0) {
20824 		if (TCP_PACING_RATE_CAP)
20825 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
20826 		else
20827 			error = sooptcopyout(sopt, &optval, sizeof optval);
20828 	}
20829 	return (error);
20830 }
20831 
20832 static int
20833 rack_ctloutput(struct inpcb *inp, struct sockopt *sopt)
20834 {
20835 	if (sopt->sopt_dir == SOPT_SET) {
20836 		return (rack_set_sockopt(inp, sopt));
20837 	} else if (sopt->sopt_dir == SOPT_GET) {
20838 		return (rack_get_sockopt(inp, sopt));
20839 	} else {
20840 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
20841 	}
20842 }
20843 
20844 static const char *rack_stack_names[] = {
20845 	__XSTRING(STACKNAME),
20846 #ifdef STACKALIAS
20847 	__XSTRING(STACKALIAS),
20848 #endif
20849 };
20850 
20851 static int
20852 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
20853 {
20854 	memset(mem, 0, size);
20855 	return (0);
20856 }
20857 
20858 static void
20859 rack_dtor(void *mem, int32_t size, void *arg)
20860 {
20861 
20862 }
20863 
20864 static bool rack_mod_inited = false;
20865 
20866 static int
20867 tcp_addrack(module_t mod, int32_t type, void *data)
20868 {
20869 	int32_t err = 0;
20870 	int num_stacks;
20871 
20872 	switch (type) {
20873 	case MOD_LOAD:
20874 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
20875 		    sizeof(struct rack_sendmap),
20876 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
20877 
20878 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
20879 		    sizeof(struct tcp_rack),
20880 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
20881 
20882 		sysctl_ctx_init(&rack_sysctl_ctx);
20883 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
20884 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
20885 		    OID_AUTO,
20886 #ifdef STACKALIAS
20887 		    __XSTRING(STACKALIAS),
20888 #else
20889 		    __XSTRING(STACKNAME),
20890 #endif
20891 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
20892 		    "");
20893 		if (rack_sysctl_root == NULL) {
20894 			printf("Failed to add sysctl node\n");
20895 			err = EFAULT;
20896 			goto free_uma;
20897 		}
20898 		rack_init_sysctls();
20899 		num_stacks = nitems(rack_stack_names);
20900 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
20901 		    rack_stack_names, &num_stacks);
20902 		if (err) {
20903 			printf("Failed to register %s stack name for "
20904 			    "%s module\n", rack_stack_names[num_stacks],
20905 			    __XSTRING(MODNAME));
20906 			sysctl_ctx_free(&rack_sysctl_ctx);
20907 free_uma:
20908 			uma_zdestroy(rack_zone);
20909 			uma_zdestroy(rack_pcb_zone);
20910 			rack_counter_destroy();
20911 			printf("Failed to register rack module -- err:%d\n", err);
20912 			return (err);
20913 		}
20914 		tcp_lro_reg_mbufq();
20915 		rack_mod_inited = true;
20916 		break;
20917 	case MOD_QUIESCE:
20918 		err = deregister_tcp_functions(&__tcp_rack, true, false);
20919 		break;
20920 	case MOD_UNLOAD:
20921 		err = deregister_tcp_functions(&__tcp_rack, false, true);
20922 		if (err == EBUSY)
20923 			break;
20924 		if (rack_mod_inited) {
20925 			uma_zdestroy(rack_zone);
20926 			uma_zdestroy(rack_pcb_zone);
20927 			sysctl_ctx_free(&rack_sysctl_ctx);
20928 			rack_counter_destroy();
20929 			rack_mod_inited = false;
20930 		}
20931 		tcp_lro_dereg_mbufq();
20932 		err = 0;
20933 		break;
20934 	default:
20935 		return (EOPNOTSUPP);
20936 	}
20937 	return (err);
20938 }
20939 
20940 static moduledata_t tcp_rack = {
20941 	.name = __XSTRING(MODNAME),
20942 	.evhand = tcp_addrack,
20943 	.priv = 0
20944 };
20945 
20946 MODULE_VERSION(MODNAME, 1);
20947 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
20948 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
20949