xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision 1edb7116) 
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 #include "opt_inet.h"
29 #include "opt_inet6.h"
30 #include "opt_ipsec.h"
31 #include "opt_ratelimit.h"
32 #include "opt_kern_tls.h"
33 #if defined(INET) || defined(INET6)
34 #include <sys/param.h>
35 #include <sys/arb.h>
36 #include <sys/module.h>
37 #include <sys/kernel.h>
38 #ifdef TCP_HHOOK
39 #include <sys/hhook.h>
40 #endif
41 #include <sys/lock.h>
42 #include <sys/malloc.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/mbuf.h>
46 #include <sys/proc.h>		/* for proc0 declaration */
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/sysctl.h>
50 #include <sys/systm.h>
51 #ifdef STATS
52 #include <sys/qmath.h>
53 #include <sys/tree.h>
54 #include <sys/stats.h> /* Must come after qmath.h and tree.h */
55 #else
56 #include <sys/tree.h>
57 #endif
58 #include <sys/refcount.h>
59 #include <sys/queue.h>
60 #include <sys/tim_filter.h>
61 #include <sys/smp.h>
62 #include <sys/kthread.h>
63 #include <sys/kern_prefetch.h>
64 #include <sys/protosw.h>
65 #ifdef TCP_ACCOUNTING
66 #include <sys/sched.h>
67 #include <machine/cpu.h>
68 #endif
69 #include <vm/uma.h>
70 
71 #include <net/route.h>
72 #include <net/route/nhop.h>
73 #include <net/vnet.h>
74 
75 #define TCPSTATES		/* for logging */
76 
77 #include <netinet/in.h>
78 #include <netinet/in_kdtrace.h>
79 #include <netinet/in_pcb.h>
80 #include <netinet/ip.h>
81 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
82 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
83 #include <netinet/ip_var.h>
84 #include <netinet/ip6.h>
85 #include <netinet6/in6_pcb.h>
86 #include <netinet6/ip6_var.h>
87 #include <netinet/tcp.h>
88 #define	TCPOUTFLAGS
89 #include <netinet/tcp_fsm.h>
90 #include <netinet/tcp_seq.h>
91 #include <netinet/tcp_timer.h>
92 #include <netinet/tcp_var.h>
93 #include <netinet/tcp_log_buf.h>
94 #include <netinet/tcp_syncache.h>
95 #include <netinet/tcp_hpts.h>
96 #include <netinet/tcp_ratelimit.h>
97 #include <netinet/tcp_accounting.h>
98 #include <netinet/tcpip.h>
99 #include <netinet/cc/cc.h>
100 #include <netinet/cc/cc_newreno.h>
101 #include <netinet/tcp_fastopen.h>
102 #include <netinet/tcp_lro.h>
103 #ifdef NETFLIX_SHARED_CWND
104 #include <netinet/tcp_shared_cwnd.h>
105 #endif
106 #ifdef TCP_OFFLOAD
107 #include <netinet/tcp_offload.h>
108 #endif
109 #ifdef INET6
110 #include <netinet6/tcp6_var.h>
111 #endif
112 #include <netinet/tcp_ecn.h>
113 
114 #include <netipsec/ipsec_support.h>
115 
116 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
117 #include <netipsec/ipsec.h>
118 #include <netipsec/ipsec6.h>
119 #endif				/* IPSEC */
120 
121 #include <netinet/udp.h>
122 #include <netinet/udp_var.h>
123 #include <machine/in_cksum.h>
124 
125 #ifdef MAC
126 #include <security/mac/mac_framework.h>
127 #endif
128 #include "sack_filter.h"
129 #include "tcp_rack.h"
130 #include "tailq_hash.h"
131 #include "rack_bbr_common.h"
132 
133 uma_zone_t rack_zone;
134 uma_zone_t rack_pcb_zone;
135 
136 #ifndef TICKS2SBT
137 #define	TICKS2SBT(__t)	(tick_sbt * ((sbintime_t)(__t)))
138 #endif
139 
140 VNET_DECLARE(uint32_t, newreno_beta);
141 VNET_DECLARE(uint32_t, newreno_beta_ecn);
142 #define V_newreno_beta VNET(newreno_beta)
143 #define V_newreno_beta_ecn VNET(newreno_beta_ecn)
144 
145 
146 MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block");
147 MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options");
148 
149 struct sysctl_ctx_list rack_sysctl_ctx;
150 struct sysctl_oid *rack_sysctl_root;
151 
152 #define CUM_ACKED 1
153 #define SACKED 2
154 
155 /*
156  * The RACK module incorporates a number of
157  * TCP ideas that have been put out into the IETF
158  * over the last few years:
159  * - Matt Mathis's Rate Halving which slowly drops
160  *    the congestion window so that the ack clock can
161  *    be maintained during a recovery.
162  * - Yuchung Cheng's RACK TCP (for which its named) that
163  *    will stop us using the number of dup acks and instead
164  *    use time as the gage of when we retransmit.
165  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
166  *    of Dukkipati et.al.
167  * RACK depends on SACK, so if an endpoint arrives that
168  * cannot do SACK the state machine below will shuttle the
169  * connection back to using the "default" TCP stack that is
170  * in FreeBSD.
171  *
172  * To implement RACK the original TCP stack was first decomposed
173  * into a functional state machine with individual states
174  * for each of the possible TCP connection states. The do_segment
175  * functions role in life is to mandate the connection supports SACK
176  * initially and then assure that the RACK state matches the conenction
177  * state before calling the states do_segment function. Each
178  * state is simplified due to the fact that the original do_segment
179  * has been decomposed and we *know* what state we are in (no
180  * switches on the state) and all tests for SACK are gone. This
181  * greatly simplifies what each state does.
182  *
183  * TCP output is also over-written with a new version since it
184  * must maintain the new rack scoreboard.
185  *
186  */
187 static int32_t rack_tlp_thresh = 1;
188 static int32_t rack_tlp_limit = 2;	/* No more than 2 TLPs w-out new data */
189 static int32_t rack_tlp_use_greater = 1;
190 static int32_t rack_reorder_thresh = 2;
191 static int32_t rack_reorder_fade = 60000000;	/* 0 - never fade, def 60,000,000
192 						 * - 60 seconds */
193 static uint32_t rack_clamp_ss_upper = 110;
194 static uint32_t rack_clamp_ca_upper = 105;
195 static uint32_t rack_rxt_min_rnds = 10;	/* Min rounds if drastic rxt clamp is in place */
196 static uint32_t rack_unclamp_round_thresh = 100;	/* number of perfect rounds before we unclamp */
197 static uint32_t rack_unclamp_rxt_thresh = 5;	/* .5%  and under */
198 static uint64_t rack_rxt_clamp_thresh = 0;	/* Do we do the rxt clamp thing */
199 static int32_t rack_dnd_default = 0;		/* For rr_conf = 3, what is the default for dnd */
200 static int32_t rack_rxt_controls = 0;
201 static int32_t rack_fill_cw_state = 0;
202 static uint8_t rack_req_measurements = 1;
203 /* Attack threshold detections */
204 static uint32_t rack_highest_sack_thresh_seen = 0;
205 static uint32_t rack_highest_move_thresh_seen = 0;
206 static uint32_t rack_merge_out_sacks_on_attack = 0;
207 static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */
208 static int32_t rack_hw_pace_extra_slots = 0;	/* 2 extra MSS time betweens */
209 static int32_t rack_hw_rate_caps = 0; /* 1; */
210 static int32_t rack_hw_rate_cap_per = 0;	/* 0 -- off  */
211 static int32_t rack_hw_rate_min = 0; /* 1500000;*/
212 static int32_t rack_hw_rate_to_low = 0; /* 1200000; */
213 static int32_t rack_hw_up_only = 0;
214 static int32_t rack_stats_gets_ms_rtt = 1;
215 static int32_t rack_prr_addbackmax = 2;
216 static int32_t rack_do_hystart = 0;
217 static int32_t rack_apply_rtt_with_reduced_conf = 0;
218 static int32_t rack_hibeta_setting = 0;
219 static int32_t rack_default_pacing_divisor = 250;
220 static int32_t rack_uses_full_dgp_in_rec = 1;
221 static uint16_t rack_pacing_min_seg = 0;
222 
223 
224 static uint32_t sad_seg_size_per = 800;	/* 80.0 % */
225 static int32_t rack_pkt_delay = 1000;
226 static int32_t rack_send_a_lot_in_prr = 1;
227 static int32_t rack_min_to = 1000;	/* Number of microsecond  min timeout */
228 static int32_t rack_verbose_logging = 0;
229 static int32_t rack_ignore_data_after_close = 1;
230 static int32_t rack_enable_shared_cwnd = 1;
231 static int32_t rack_use_cmp_acks = 1;
232 static int32_t rack_use_fsb = 1;
233 static int32_t rack_use_rfo = 1;
234 static int32_t rack_use_rsm_rfo = 1;
235 static int32_t rack_max_abc_post_recovery = 2;
236 static int32_t rack_client_low_buf = 0;
237 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 */
238 static int32_t rack_bw_multipler = 2;		/* Limit on fill cw's jump up to be this x gp_est */
239 #ifdef TCP_ACCOUNTING
240 static int32_t rack_tcp_accounting = 0;
241 #endif
242 static int32_t rack_limits_scwnd = 1;
243 static int32_t rack_enable_mqueue_for_nonpaced = 0;
244 static int32_t rack_hybrid_allow_set_maxseg = 0;
245 static int32_t rack_disable_prr = 0;
246 static int32_t use_rack_rr = 1;
247 static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */
248 static int32_t rack_persist_min = 250000;	/* 250usec */
249 static int32_t rack_persist_max = 2000000;	/* 2 Second in usec's */
250 static int32_t rack_sack_not_required = 1;	/* set to one to allow non-sack to use rack */
251 static int32_t rack_default_init_window = 0;	/* Use system default */
252 static int32_t rack_limit_time_with_srtt = 0;
253 static int32_t rack_autosndbuf_inc = 20;	/* In percentage form */
254 static int32_t rack_enobuf_hw_boost_mult = 0;	/* How many times the hw rate we boost slot using time_between */
255 static int32_t rack_enobuf_hw_max = 12000;	/* 12 ms in usecs */
256 static int32_t rack_enobuf_hw_min = 10000;	/* 10 ms in usecs */
257 static int32_t rack_hw_rwnd_factor = 2;		/* How many max_segs the rwnd must be before we hold off sending */
258 static int32_t rack_hw_check_queue = 0;		/* Do we always pre-check queue depth of a hw queue */
259 static int32_t rack_full_buffer_discount = 10;
260 /*
261  * Currently regular tcp has a rto_min of 30ms
262  * the backoff goes 12 times so that ends up
263  * being a total of 122.850 seconds before a
264  * connection is killed.
265  */
266 static uint32_t rack_def_data_window = 20;
267 static uint32_t rack_goal_bdp = 2;
268 static uint32_t rack_min_srtts = 1;
269 static uint32_t rack_min_measure_usec = 0;
270 static int32_t rack_tlp_min = 10000;	/* 10ms */
271 static int32_t rack_rto_min = 30000;	/* 30,000 usec same as main freebsd */
272 static int32_t rack_rto_max = 4000000;	/* 4 seconds in usec's */
273 static const int32_t rack_free_cache = 2;
274 static int32_t rack_hptsi_segments = 40;
275 static int32_t rack_rate_sample_method = USE_RTT_LOW;
276 static int32_t rack_pace_every_seg = 0;
277 static int32_t rack_delayed_ack_time = 40000;	/* 40ms in usecs */
278 static int32_t rack_slot_reduction = 4;
279 static int32_t rack_wma_divisor = 8;		/* For WMA calculation */
280 static int32_t rack_cwnd_block_ends_measure = 0;
281 static int32_t rack_rwnd_block_ends_measure = 0;
282 static int32_t rack_def_profile = 0;
283 
284 static int32_t rack_lower_cwnd_at_tlp = 0;
285 static int32_t rack_limited_retran = 0;
286 static int32_t rack_always_send_oldest = 0;
287 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
288 
289 static uint16_t rack_per_of_gp_ss = 250;	/* 250 % slow-start */
290 static uint16_t rack_per_of_gp_ca = 200;	/* 200 % congestion-avoidance */
291 static uint16_t rack_per_of_gp_rec = 200;	/* 200 % of bw */
292 
293 /* Probertt */
294 static uint16_t rack_per_of_gp_probertt = 60;	/* 60% of bw */
295 static uint16_t rack_per_of_gp_lowthresh = 40;	/* 40% is bottom */
296 static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */
297 static uint16_t rack_atexit_prtt_hbp = 130;	/* Clamp to 130% on exit prtt if highly buffered path */
298 static uint16_t rack_atexit_prtt = 130;	/* Clamp to 100% on exit prtt if non highly buffered path */
299 
300 static uint32_t rack_max_drain_wait = 2;	/* How man gp srtt's before we give up draining */
301 static uint32_t rack_must_drain = 1;		/* How many GP srtt's we *must* wait */
302 static uint32_t rack_probertt_use_min_rtt_entry = 1;	/* Use the min to calculate the goal else gp_srtt */
303 static uint32_t rack_probertt_use_min_rtt_exit = 0;
304 static uint32_t rack_probe_rtt_sets_cwnd = 0;
305 static uint32_t rack_probe_rtt_safety_val = 2000000;	/* No more than 2 sec in probe-rtt */
306 static uint32_t rack_time_between_probertt = 9600000;	/* 9.6 sec in usecs */
307 static uint32_t rack_probertt_gpsrtt_cnt_mul = 0;	/* How many srtt periods does probe-rtt last top fraction */
308 static uint32_t rack_probertt_gpsrtt_cnt_div = 0;	/* How many srtt periods does probe-rtt last bottom fraction */
309 static uint32_t rack_min_probertt_hold = 40000;		/* Equal to delayed ack time */
310 static uint32_t rack_probertt_filter_life = 10000000;
311 static uint32_t rack_probertt_lower_within = 10;
312 static uint32_t rack_min_rtt_movement = 250000;	/* Must move at least 250ms (in microseconds)  to count as a lowering */
313 static int32_t rack_pace_one_seg = 0;		/* Shall we pace for less than 1.4Meg 1MSS at a time */
314 static int32_t rack_probertt_clear_is = 1;
315 static int32_t rack_max_drain_hbp = 1;		/* Extra drain times gpsrtt for highly buffered paths */
316 static int32_t rack_hbp_thresh = 3;		/* what is the divisor max_rtt/min_rtt to decided a hbp */
317 
318 /* Part of pacing */
319 static int32_t rack_max_per_above = 30;		/* When we go to increment stop if above 100+this% */
320 
321 /* Timely information:
322  *
323  * Here we have various control parameters on how
324  * timely may change the multiplier. rack_gain_p5_ub
325  * is associated with timely but not directly influencing
326  * the rate decision like the other variables. It controls
327  * the way fill-cw interacts with timely and caps how much
328  * timely can boost the fill-cw b/w.
329  *
330  * The other values are various boost/shrink numbers as well
331  * as potential caps when adjustments are made to the timely
332  * gain (returned by rack_get_output_gain(). Remember too that
333  * the gain returned can be overriden by other factors such as
334  * probeRTT as well as fixed-rate-pacing.
335  */
336 static int32_t rack_gain_p5_ub = 250;
337 static int32_t rack_gp_per_bw_mul_up = 2;	/* 2% */
338 static int32_t rack_gp_per_bw_mul_down = 4;	/* 4% */
339 static int32_t rack_gp_rtt_maxmul = 3;		/* 3 x maxmin */
340 static int32_t rack_gp_rtt_minmul = 1;		/* minrtt + (minrtt/mindiv) is lower rtt */
341 static int32_t rack_gp_rtt_mindiv = 4;		/* minrtt + (minrtt * minmul/mindiv) is lower rtt */
342 static int32_t rack_gp_decrease_per = 80;	/* Beta value of timely decrease (.8) = 80 */
343 static int32_t rack_gp_increase_per = 2;	/* 2% increase in multiplier */
344 static int32_t rack_per_lower_bound = 50;	/* Don't allow to drop below this multiplier */
345 static int32_t rack_per_upper_bound_ss = 0;	/* Don't allow SS to grow above this */
346 static int32_t rack_per_upper_bound_ca = 0;	/* Don't allow CA to grow above this */
347 static int32_t rack_do_dyn_mul = 0;		/* Are the rack gp multipliers dynamic */
348 static int32_t rack_gp_no_rec_chg = 1;		/* Prohibit recovery from reducing it's multiplier */
349 static int32_t rack_timely_dec_clear = 6;	/* Do we clear decrement count at a value (6)? */
350 static int32_t rack_timely_max_push_rise = 3;	/* One round of pushing */
351 static int32_t rack_timely_max_push_drop = 3;	/* Three round of pushing */
352 static int32_t rack_timely_min_segs = 4;	/* 4 segment minimum */
353 static int32_t rack_use_max_for_nobackoff = 0;
354 static int32_t rack_timely_int_timely_only = 0;	/* do interim timely's only use the timely algo (no b/w changes)? */
355 static int32_t rack_timely_no_stopping = 0;
356 static int32_t rack_down_raise_thresh = 100;
357 static int32_t rack_req_segs = 1;
358 static uint64_t rack_bw_rate_cap = 0;
359 
360 
361 /* Rack specific counters */
362 counter_u64_t rack_saw_enobuf;
363 counter_u64_t rack_saw_enobuf_hw;
364 counter_u64_t rack_saw_enetunreach;
365 counter_u64_t rack_persists_sends;
366 counter_u64_t rack_persists_acks;
367 counter_u64_t rack_persists_loss;
368 counter_u64_t rack_persists_lost_ends;
369 counter_u64_t rack_total_bytes;
370 #ifdef INVARIANTS
371 counter_u64_t rack_adjust_map_bw;
372 #endif
373 /* Tail loss probe counters */
374 counter_u64_t rack_tlp_tot;
375 counter_u64_t rack_tlp_newdata;
376 counter_u64_t rack_tlp_retran;
377 counter_u64_t rack_tlp_retran_bytes;
378 counter_u64_t rack_to_tot;
379 counter_u64_t rack_hot_alloc;
380 counter_u64_t rack_to_alloc;
381 counter_u64_t rack_to_alloc_hard;
382 counter_u64_t rack_to_alloc_emerg;
383 counter_u64_t rack_to_alloc_limited;
384 counter_u64_t rack_alloc_limited_conns;
385 counter_u64_t rack_split_limited;
386 counter_u64_t rack_rxt_clamps_cwnd;
387 counter_u64_t rack_rxt_clamps_cwnd_uniq;
388 
389 counter_u64_t rack_multi_single_eq;
390 counter_u64_t rack_proc_non_comp_ack;
391 
392 counter_u64_t rack_fto_send;
393 counter_u64_t rack_fto_rsm_send;
394 counter_u64_t rack_nfto_resend;
395 counter_u64_t rack_non_fto_send;
396 counter_u64_t rack_extended_rfo;
397 
398 counter_u64_t rack_sack_proc_all;
399 counter_u64_t rack_sack_proc_short;
400 counter_u64_t rack_sack_proc_restart;
401 counter_u64_t rack_sack_attacks_detected;
402 counter_u64_t rack_sack_attacks_reversed;
403 counter_u64_t rack_sack_attacks_suspect;
404 counter_u64_t rack_sack_used_next_merge;
405 counter_u64_t rack_sack_splits;
406 counter_u64_t rack_sack_used_prev_merge;
407 counter_u64_t rack_sack_skipped_acked;
408 counter_u64_t rack_ack_total;
409 counter_u64_t rack_express_sack;
410 counter_u64_t rack_sack_total;
411 counter_u64_t rack_move_none;
412 counter_u64_t rack_move_some;
413 
414 counter_u64_t rack_input_idle_reduces;
415 counter_u64_t rack_collapsed_win;
416 counter_u64_t rack_collapsed_win_seen;
417 counter_u64_t rack_collapsed_win_rxt;
418 counter_u64_t rack_collapsed_win_rxt_bytes;
419 counter_u64_t rack_try_scwnd;
420 counter_u64_t rack_hw_pace_init_fail;
421 counter_u64_t rack_hw_pace_lost;
422 
423 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
424 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
425 
426 
427 #define	RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2)))
428 
429 #define	RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do {	\
430 	(tv) = (value) + slop;	 \
431 	if ((u_long)(tv) < (u_long)(tvmin)) \
432 		(tv) = (tvmin); \
433 	if ((u_long)(tv) > (u_long)(tvmax)) \
434 		(tv) = (tvmax); \
435 } while (0)
436 
437 static void
438 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
439 
440 static int
441 rack_process_ack(struct mbuf *m, struct tcphdr *th,
442     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
443     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
444 static int
445 rack_process_data(struct mbuf *m, struct tcphdr *th,
446     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
447     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
448 static void
449 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
450    uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery);
451 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
452 static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack,
453     uint8_t limit_type);
454 static struct rack_sendmap *
455 rack_check_recovery_mode(struct tcpcb *tp,
456     uint32_t tsused);
457 static void
458 rack_cong_signal(struct tcpcb *tp,
459 		 uint32_t type, uint32_t ack, int );
460 static void rack_counter_destroy(void);
461 static int
462 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt);
463 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
464 static void
465 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override);
466 static void
467 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
468     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos);
469 static void rack_dtor(void *mem, int32_t size, void *arg);
470 static void
471 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
472     uint32_t flex1, uint32_t flex2,
473     uint32_t flex3, uint32_t flex4,
474     uint32_t flex5, uint32_t flex6,
475     uint16_t flex7, uint8_t mod);
476 
477 static void
478 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
479    uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line,
480    struct rack_sendmap *rsm, uint8_t quality);
481 static struct rack_sendmap *
482 rack_find_high_nonack(struct tcp_rack *rack,
483     struct rack_sendmap *rsm);
484 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
485 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
486 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
487 static int rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt);
488 static void
489 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
490 			    tcp_seq th_ack, int line, uint8_t quality);
491 static void
492 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm);
493 
494 static uint32_t
495 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss);
496 static int32_t rack_handoff_ok(struct tcpcb *tp);
497 static int32_t rack_init(struct tcpcb *tp, void **ptr);
498 static void rack_init_sysctls(void);
499 
500 static void
501 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
502     struct tcphdr *th, int entered_rec, int dup_ack_struck,
503     int *dsack_seen, int *sacks_seen);
504 static void
505 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
506     uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts,
507     struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls, int segsiz);
508 
509 static uint64_t rack_get_gp_est(struct tcp_rack *rack);
510 
511 static void
512 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
513     struct rack_sendmap *rsm);
514 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm);
515 static int32_t rack_output(struct tcpcb *tp);
516 
517 static uint32_t
518 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
519     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
520     uint32_t cts, int *no_extra, int *moved_two, uint32_t segsiz);
521 static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq);
522 static void rack_remxt_tmr(struct tcpcb *tp);
523 static int rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt);
524 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
525 static int32_t rack_stopall(struct tcpcb *tp);
526 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
527 static uint32_t
528 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
529     struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag, int segsiz);
530 static void
531 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
532     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag, int segsiz);
533 static int
534 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
535     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack);
536 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
537 static int
538 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
539     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
540     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
541 static int
542 rack_do_closing(struct mbuf *m, struct tcphdr *th,
543     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
544     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
545 static int
546 rack_do_established(struct mbuf *m, struct tcphdr *th,
547     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
548     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
549 static int
550 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
551     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
552     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos);
553 static int
554 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
555     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
556     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
557 static int
558 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
559     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
560     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
561 static int
562 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
563     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
564     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
565 static int
566 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
567     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
568     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
569 static int
570 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
571     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
572     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos);
573 static void rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts);
574 struct rack_sendmap *
575 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
576     uint32_t tsused);
577 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt,
578     uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt);
579 static void
580      tcp_rack_partialack(struct tcpcb *tp);
581 static int
582 rack_set_profile(struct tcp_rack *rack, int prof);
583 static void
584 rack_apply_deferred_options(struct tcp_rack *rack);
585 
586 int32_t rack_clear_counter=0;
587 
588 static uint64_t
589 rack_get_lt_bw(struct tcp_rack *rack)
590 {
591 	struct timeval tv;
592 	uint64_t tim, bytes;
593 
594 	tim = rack->r_ctl.lt_bw_time;
595 	bytes = rack->r_ctl.lt_bw_bytes;
596 	if (rack->lt_bw_up) {
597 		/* Include all the current bytes too */
598 		microuptime(&tv);
599 		bytes += (rack->rc_tp->snd_una - rack->r_ctl.lt_seq);
600 		tim += (tcp_tv_to_lusectick(&tv) - rack->r_ctl.lt_timemark);
601 	}
602 	if ((bytes != 0) && (tim != 0))
603 		return ((bytes * (uint64_t)1000000) / tim);
604 	else
605 		return (0);
606 }
607 
608 static void
609 rack_swap_beta_values(struct tcp_rack *rack, uint8_t flex8)
610 {
611 	struct sockopt sopt;
612 	struct cc_newreno_opts opt;
613 	struct newreno old;
614 	struct tcpcb *tp;
615 	int error, failed = 0;
616 
617 	tp = rack->rc_tp;
618 	if (tp->t_cc == NULL) {
619 		/* Tcb is leaving */
620 		return;
621 	}
622 	rack->rc_pacing_cc_set = 1;
623 	if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
624 		/* Not new-reno we can't play games with beta! */
625 		failed = 1;
626 		goto out;
627 
628 	}
629 	if (CC_ALGO(tp)->ctl_output == NULL)  {
630 		/* Huh, not using new-reno so no swaps.? */
631 		failed = 2;
632 		goto out;
633 	}
634 	/* Get the current values out */
635 	sopt.sopt_valsize = sizeof(struct cc_newreno_opts);
636 	sopt.sopt_dir = SOPT_GET;
637 	opt.name = CC_NEWRENO_BETA;
638 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
639 	if (error)  {
640 		failed = 3;
641 		goto out;
642 	}
643 	old.beta = opt.val;
644 	opt.name = CC_NEWRENO_BETA_ECN;
645 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
646 	if (error)  {
647 		failed = 4;
648 		goto out;
649 	}
650 	old.beta_ecn = opt.val;
651 
652 	/* Now lets set in the values we have stored */
653 	sopt.sopt_dir = SOPT_SET;
654 	opt.name = CC_NEWRENO_BETA;
655 	opt.val = rack->r_ctl.rc_saved_beta.beta;
656 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
657 	if (error)  {
658 		failed = 5;
659 		goto out;
660 	}
661 	opt.name = CC_NEWRENO_BETA_ECN;
662 	opt.val = rack->r_ctl.rc_saved_beta.beta_ecn;
663 	error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
664 	if (error) {
665 		failed = 6;
666 		goto out;
667 	}
668 	/* Save off the values for restoral */
669 	memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno));
670 out:
671 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
672 		union tcp_log_stackspecific log;
673 		struct timeval tv;
674 		struct newreno *ptr;
675 
676 		ptr = ((struct newreno *)tp->t_ccv.cc_data);
677 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
678 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
679 		log.u_bbr.flex1 = ptr->beta;
680 		log.u_bbr.flex2 = ptr->beta_ecn;
681 		log.u_bbr.flex3 = ptr->newreno_flags;
682 		log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta;
683 		log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn;
684 		log.u_bbr.flex6 = failed;
685 		log.u_bbr.flex7 = rack->gp_ready;
686 		log.u_bbr.flex7 <<= 1;
687 		log.u_bbr.flex7 |= rack->use_fixed_rate;
688 		log.u_bbr.flex7 <<= 1;
689 		log.u_bbr.flex7 |= rack->rc_pacing_cc_set;
690 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
691 		log.u_bbr.flex8 = flex8;
692 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, error,
693 			       0, &log, false, NULL, NULL, 0, &tv);
694 	}
695 }
696 
697 static void
698 rack_set_cc_pacing(struct tcp_rack *rack)
699 {
700 	if (rack->rc_pacing_cc_set)
701 		return;
702 	/*
703 	 * Use the swap utility placing in 3 for flex8 to id a
704 	 * set of a new set of values.
705 	 */
706 	rack->rc_pacing_cc_set = 1;
707 	rack_swap_beta_values(rack, 3);
708 }
709 
710 static void
711 rack_undo_cc_pacing(struct tcp_rack *rack)
712 {
713 	if (rack->rc_pacing_cc_set == 0)
714 		return;
715 	/*
716 	 * Use the swap utility placing in 4 for flex8 to id a
717 	 * restoral of the old values.
718 	 */
719 	rack->rc_pacing_cc_set = 0;
720 	rack_swap_beta_values(rack, 4);
721 }
722 
723 static void
724 rack_log_gpset(struct tcp_rack *rack, uint32_t seq_end, uint32_t ack_end_t,
725 	       uint32_t send_end_t, int line, uint8_t mode, struct rack_sendmap *rsm)
726 {
727 	if (tcp_bblogging_on(rack->rc_tp) && (rack_verbose_logging != 0)) {
728 		union tcp_log_stackspecific log;
729 		struct timeval tv;
730 
731 		memset(&log, 0, sizeof(log));
732 		log.u_bbr.flex1 = seq_end;
733 		log.u_bbr.flex2 = rack->rc_tp->gput_seq;
734 		log.u_bbr.flex3 = ack_end_t;
735 		log.u_bbr.flex4 = rack->rc_tp->gput_ts;
736 		log.u_bbr.flex5 = send_end_t;
737 		log.u_bbr.flex6 = rack->rc_tp->gput_ack;
738 		log.u_bbr.flex7 = mode;
739 		log.u_bbr.flex8 = 69;
740 		log.u_bbr.rttProp = rack->r_ctl.rc_gp_cumack_ts;
741 		log.u_bbr.delRate = rack->r_ctl.rc_gp_output_ts;
742 		log.u_bbr.pkts_out = line;
743 		log.u_bbr.cwnd_gain = rack->app_limited_needs_set;
744 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_app_limited_cnt;
745 		if (rsm != NULL) {
746 			log.u_bbr.applimited = rsm->r_start;
747 			log.u_bbr.delivered = rsm->r_end;
748 			log.u_bbr.epoch = rsm->r_flags;
749 		}
750 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
751 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
752 		    &rack->rc_inp->inp_socket->so_rcv,
753 		    &rack->rc_inp->inp_socket->so_snd,
754 		    BBR_LOG_HPTSI_CALC, 0,
755 		    0, &log, false, &tv);
756 	}
757 }
758 
759 static int
760 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
761 {
762 	uint32_t stat;
763 	int32_t error;
764 
765 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
766 	if (error || req->newptr == NULL)
767 		return error;
768 
769 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
770 	if (error)
771 		return (error);
772 	if (stat == 1) {
773 #ifdef INVARIANTS
774 		printf("Clearing RACK counters\n");
775 #endif
776 		counter_u64_zero(rack_tlp_tot);
777 		counter_u64_zero(rack_tlp_newdata);
778 		counter_u64_zero(rack_tlp_retran);
779 		counter_u64_zero(rack_tlp_retran_bytes);
780 		counter_u64_zero(rack_to_tot);
781 		counter_u64_zero(rack_saw_enobuf);
782 		counter_u64_zero(rack_saw_enobuf_hw);
783 		counter_u64_zero(rack_saw_enetunreach);
784 		counter_u64_zero(rack_persists_sends);
785 		counter_u64_zero(rack_total_bytes);
786 		counter_u64_zero(rack_persists_acks);
787 		counter_u64_zero(rack_persists_loss);
788 		counter_u64_zero(rack_persists_lost_ends);
789 #ifdef INVARIANTS
790 		counter_u64_zero(rack_adjust_map_bw);
791 #endif
792 		counter_u64_zero(rack_to_alloc_hard);
793 		counter_u64_zero(rack_to_alloc_emerg);
794 		counter_u64_zero(rack_sack_proc_all);
795 		counter_u64_zero(rack_fto_send);
796 		counter_u64_zero(rack_fto_rsm_send);
797 		counter_u64_zero(rack_extended_rfo);
798 		counter_u64_zero(rack_hw_pace_init_fail);
799 		counter_u64_zero(rack_hw_pace_lost);
800 		counter_u64_zero(rack_non_fto_send);
801 		counter_u64_zero(rack_nfto_resend);
802 		counter_u64_zero(rack_sack_proc_short);
803 		counter_u64_zero(rack_sack_proc_restart);
804 		counter_u64_zero(rack_to_alloc);
805 		counter_u64_zero(rack_to_alloc_limited);
806 		counter_u64_zero(rack_alloc_limited_conns);
807 		counter_u64_zero(rack_split_limited);
808 		counter_u64_zero(rack_rxt_clamps_cwnd);
809 		counter_u64_zero(rack_rxt_clamps_cwnd_uniq);
810 		counter_u64_zero(rack_multi_single_eq);
811 		counter_u64_zero(rack_proc_non_comp_ack);
812 		counter_u64_zero(rack_sack_attacks_detected);
813 		counter_u64_zero(rack_sack_attacks_reversed);
814 		counter_u64_zero(rack_sack_attacks_suspect);
815 		counter_u64_zero(rack_sack_used_next_merge);
816 		counter_u64_zero(rack_sack_used_prev_merge);
817 		counter_u64_zero(rack_sack_splits);
818 		counter_u64_zero(rack_sack_skipped_acked);
819 		counter_u64_zero(rack_ack_total);
820 		counter_u64_zero(rack_express_sack);
821 		counter_u64_zero(rack_sack_total);
822 		counter_u64_zero(rack_move_none);
823 		counter_u64_zero(rack_move_some);
824 		counter_u64_zero(rack_try_scwnd);
825 		counter_u64_zero(rack_collapsed_win);
826 		counter_u64_zero(rack_collapsed_win_rxt);
827 		counter_u64_zero(rack_collapsed_win_seen);
828 		counter_u64_zero(rack_collapsed_win_rxt_bytes);
829 	} else if (stat == 2) {
830 #ifdef INVARIANTS
831 		printf("Clearing RACK option array\n");
832 #endif
833 		COUNTER_ARRAY_ZERO(rack_opts_arry, RACK_OPTS_SIZE);
834 	} else if (stat == 3) {
835 		printf("Rack has no stats counters to clear (use 1 to clear all stats in sysctl node)\n");
836 	} else if (stat == 4) {
837 #ifdef INVARIANTS
838 		printf("Clearing RACK out size array\n");
839 #endif
840 		COUNTER_ARRAY_ZERO(rack_out_size, TCP_MSS_ACCT_SIZE);
841 	}
842 	rack_clear_counter = 0;
843 	return (0);
844 }
845 
846 static void
847 rack_init_sysctls(void)
848 {
849 	struct sysctl_oid *rack_counters;
850 	struct sysctl_oid *rack_attack;
851 	struct sysctl_oid *rack_pacing;
852 	struct sysctl_oid *rack_timely;
853 	struct sysctl_oid *rack_timers;
854 	struct sysctl_oid *rack_tlp;
855 	struct sysctl_oid *rack_misc;
856 	struct sysctl_oid *rack_features;
857 	struct sysctl_oid *rack_measure;
858 	struct sysctl_oid *rack_probertt;
859 	struct sysctl_oid *rack_hw_pacing;
860 
861 	rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
862 	    SYSCTL_CHILDREN(rack_sysctl_root),
863 	    OID_AUTO,
864 	    "sack_attack",
865 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
866 	    "Rack Sack Attack Counters and Controls");
867 	rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
868 	    SYSCTL_CHILDREN(rack_sysctl_root),
869 	    OID_AUTO,
870 	    "stats",
871 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
872 	    "Rack Counters");
873 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
874 	    SYSCTL_CHILDREN(rack_sysctl_root),
875 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
876 	    &rack_rate_sample_method , USE_RTT_LOW,
877 	    "What method should we use for rate sampling 0=high, 1=low ");
878 	/* Probe rtt related controls */
879 	rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
880 	    SYSCTL_CHILDREN(rack_sysctl_root),
881 	    OID_AUTO,
882 	    "probertt",
883 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
884 	    "ProbeRTT related Controls");
885 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
886 	    SYSCTL_CHILDREN(rack_probertt),
887 	    OID_AUTO, "exit_per_hpb", CTLFLAG_RW,
888 	    &rack_atexit_prtt_hbp, 130,
889 	    "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%");
890 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
891 	    SYSCTL_CHILDREN(rack_probertt),
892 	    OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW,
893 	    &rack_atexit_prtt, 130,
894 	    "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%");
895 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
896 	    SYSCTL_CHILDREN(rack_probertt),
897 	    OID_AUTO, "gp_per_mul", CTLFLAG_RW,
898 	    &rack_per_of_gp_probertt, 60,
899 	    "What percentage of goodput do we pace at in probertt");
900 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
901 	    SYSCTL_CHILDREN(rack_probertt),
902 	    OID_AUTO, "gp_per_reduce", CTLFLAG_RW,
903 	    &rack_per_of_gp_probertt_reduce, 10,
904 	    "What percentage of goodput do we reduce every gp_srtt");
905 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
906 	    SYSCTL_CHILDREN(rack_probertt),
907 	    OID_AUTO, "gp_per_low", CTLFLAG_RW,
908 	    &rack_per_of_gp_lowthresh, 40,
909 	    "What percentage of goodput do we allow the multiplier to fall to");
910 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
911 	    SYSCTL_CHILDREN(rack_probertt),
912 	    OID_AUTO, "time_between", CTLFLAG_RW,
913 	    & rack_time_between_probertt, 96000000,
914 	    "How many useconds between the lowest rtt falling must past before we enter probertt");
915 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
916 	    SYSCTL_CHILDREN(rack_probertt),
917 	    OID_AUTO, "safety", CTLFLAG_RW,
918 	    &rack_probe_rtt_safety_val, 2000000,
919 	    "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)");
920 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
921 	    SYSCTL_CHILDREN(rack_probertt),
922 	    OID_AUTO, "sets_cwnd", CTLFLAG_RW,
923 	    &rack_probe_rtt_sets_cwnd, 0,
924 	    "Do we set the cwnd too (if always_lower is on)");
925 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
926 	    SYSCTL_CHILDREN(rack_probertt),
927 	    OID_AUTO, "maxdrainsrtts", CTLFLAG_RW,
928 	    &rack_max_drain_wait, 2,
929 	    "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal");
930 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
931 	    SYSCTL_CHILDREN(rack_probertt),
932 	    OID_AUTO, "mustdrainsrtts", CTLFLAG_RW,
933 	    &rack_must_drain, 1,
934 	    "We must drain this many gp_srtt's waiting for flight to reach goal");
935 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
936 	    SYSCTL_CHILDREN(rack_probertt),
937 	    OID_AUTO, "goal_use_min_entry", CTLFLAG_RW,
938 	    &rack_probertt_use_min_rtt_entry, 1,
939 	    "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry");
940 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
941 	    SYSCTL_CHILDREN(rack_probertt),
942 	    OID_AUTO, "goal_use_min_exit", CTLFLAG_RW,
943 	    &rack_probertt_use_min_rtt_exit, 0,
944 	    "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt");
945 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
946 	    SYSCTL_CHILDREN(rack_probertt),
947 	    OID_AUTO, "length_div", CTLFLAG_RW,
948 	    &rack_probertt_gpsrtt_cnt_div, 0,
949 	    "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)");
950 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
951 	    SYSCTL_CHILDREN(rack_probertt),
952 	    OID_AUTO, "length_mul", CTLFLAG_RW,
953 	    &rack_probertt_gpsrtt_cnt_mul, 0,
954 	    "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)");
955 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
956 	    SYSCTL_CHILDREN(rack_probertt),
957 	    OID_AUTO, "holdtim_at_target", CTLFLAG_RW,
958 	    &rack_min_probertt_hold, 200000,
959 	    "What is the minimum time we hold probertt at target");
960 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
961 	    SYSCTL_CHILDREN(rack_probertt),
962 	    OID_AUTO, "filter_life", CTLFLAG_RW,
963 	    &rack_probertt_filter_life, 10000000,
964 	    "What is the time for the filters life in useconds");
965 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
966 	    SYSCTL_CHILDREN(rack_probertt),
967 	    OID_AUTO, "lower_within", CTLFLAG_RW,
968 	    &rack_probertt_lower_within, 10,
969 	    "If the rtt goes lower within this percentage of the time, go into probe-rtt");
970 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
971 	    SYSCTL_CHILDREN(rack_probertt),
972 	    OID_AUTO, "must_move", CTLFLAG_RW,
973 	    &rack_min_rtt_movement, 250,
974 	    "How much is the minimum movement in rtt to count as a drop for probertt purposes");
975 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
976 	    SYSCTL_CHILDREN(rack_probertt),
977 	    OID_AUTO, "clear_is_cnts", CTLFLAG_RW,
978 	    &rack_probertt_clear_is, 1,
979 	    "Do we clear I/S counts on exiting probe-rtt");
980 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
981 	    SYSCTL_CHILDREN(rack_probertt),
982 	    OID_AUTO, "hbp_extra_drain", CTLFLAG_RW,
983 	    &rack_max_drain_hbp, 1,
984 	    "How many extra drain gpsrtt's do we get in highly buffered paths");
985 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
986 	    SYSCTL_CHILDREN(rack_probertt),
987 	    OID_AUTO, "hbp_threshold", CTLFLAG_RW,
988 	    &rack_hbp_thresh, 3,
989 	    "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold");
990 	/* Pacing related sysctls */
991 	rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
992 	    SYSCTL_CHILDREN(rack_sysctl_root),
993 	    OID_AUTO,
994 	    "pacing",
995 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
996 	    "Pacing related Controls");
997 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
998 	    SYSCTL_CHILDREN(rack_pacing),
999 	    OID_AUTO, "fulldgpinrec", CTLFLAG_RW,
1000 	    &rack_uses_full_dgp_in_rec, 1,
1001 	    "Do we use all DGP features in recovery (fillcw, timely et.al.)?");
1002 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1003 	    SYSCTL_CHILDREN(rack_pacing),
1004 	    OID_AUTO, "fullbufdisc", CTLFLAG_RW,
1005 	    &rack_full_buffer_discount, 10,
1006 	    "What percentage b/w reduction over the GP estimate for a full buffer (default=0 off)?");
1007 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1008 	    SYSCTL_CHILDREN(rack_pacing),
1009 	    OID_AUTO, "fillcw", CTLFLAG_RW,
1010 	    &rack_fill_cw_state, 0,
1011 	    "Enable fillcw on new connections (default=0 off)?");
1012 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1013 	    SYSCTL_CHILDREN(rack_pacing),
1014 	    OID_AUTO, "min_burst", CTLFLAG_RW,
1015 	    &rack_pacing_min_seg, 0,
1016 	    "What is the min burst size for pacing (0 disables)?");
1017 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1018 	    SYSCTL_CHILDREN(rack_pacing),
1019 	    OID_AUTO, "divisor", CTLFLAG_RW,
1020 	    &rack_default_pacing_divisor, 4,
1021 	    "What is the default divisor given to the rl code?");
1022 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1023 	    SYSCTL_CHILDREN(rack_pacing),
1024 	    OID_AUTO, "fillcw_max_mult", CTLFLAG_RW,
1025 	    &rack_bw_multipler, 2,
1026 	    "What is the multiplier of the current gp_est that fillcw can increase the b/w too?");
1027 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1028 	    SYSCTL_CHILDREN(rack_pacing),
1029 	    OID_AUTO, "max_pace_over", CTLFLAG_RW,
1030 	    &rack_max_per_above, 30,
1031 	    "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)");
1032 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1033 	    SYSCTL_CHILDREN(rack_pacing),
1034 	    OID_AUTO, "allow1mss", CTLFLAG_RW,
1035 	    &rack_pace_one_seg, 0,
1036 	    "Do we allow low b/w pacing of 1MSS instead of two (1.2Meg and less)?");
1037 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1038 	    SYSCTL_CHILDREN(rack_pacing),
1039 	    OID_AUTO, "limit_wsrtt", CTLFLAG_RW,
1040 	    &rack_limit_time_with_srtt, 0,
1041 	    "Do we limit pacing time based on srtt");
1042 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1043 	    SYSCTL_CHILDREN(rack_pacing),
1044 	    OID_AUTO, "init_win", CTLFLAG_RW,
1045 	    &rack_default_init_window, 0,
1046 	    "Do we have a rack initial window 0 = system default");
1047 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1048 	    SYSCTL_CHILDREN(rack_pacing),
1049 	    OID_AUTO, "gp_per_ss", CTLFLAG_RW,
1050 	    &rack_per_of_gp_ss, 250,
1051 	    "If non zero, what percentage of goodput to pace at in slow start");
1052 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1053 	    SYSCTL_CHILDREN(rack_pacing),
1054 	    OID_AUTO, "gp_per_ca", CTLFLAG_RW,
1055 	    &rack_per_of_gp_ca, 150,
1056 	    "If non zero, what percentage of goodput to pace at in congestion avoidance");
1057 	SYSCTL_ADD_U16(&rack_sysctl_ctx,
1058 	    SYSCTL_CHILDREN(rack_pacing),
1059 	    OID_AUTO, "gp_per_rec", CTLFLAG_RW,
1060 	    &rack_per_of_gp_rec, 200,
1061 	    "If non zero, what percentage of goodput to pace at in recovery");
1062 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1063 	    SYSCTL_CHILDREN(rack_pacing),
1064 	    OID_AUTO, "pace_max_seg", CTLFLAG_RW,
1065 	    &rack_hptsi_segments, 40,
1066 	    "What size is the max for TSO segments in pacing and burst mitigation");
1067 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1068 	    SYSCTL_CHILDREN(rack_pacing),
1069 	    OID_AUTO, "burst_reduces", CTLFLAG_RW,
1070 	    &rack_slot_reduction, 4,
1071 	    "When doing only burst mitigation what is the reduce divisor");
1072 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1073 	    SYSCTL_CHILDREN(rack_sysctl_root),
1074 	    OID_AUTO, "use_pacing", CTLFLAG_RW,
1075 	    &rack_pace_every_seg, 0,
1076 	    "If set we use pacing, if clear we use only the original burst mitigation");
1077 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1078 	    SYSCTL_CHILDREN(rack_pacing),
1079 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1080 	    &rack_bw_rate_cap, 0,
1081 	    "If set we apply this value to the absolute rate cap used by pacing");
1082 	SYSCTL_ADD_U8(&rack_sysctl_ctx,
1083 	    SYSCTL_CHILDREN(rack_sysctl_root),
1084 	    OID_AUTO, "req_measure_cnt", CTLFLAG_RW,
1085 	    &rack_req_measurements, 1,
1086 	    "If doing dynamic pacing, how many measurements must be in before we start pacing?");
1087 	/* Hardware pacing */
1088 	rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1089 	    SYSCTL_CHILDREN(rack_sysctl_root),
1090 	    OID_AUTO,
1091 	    "hdwr_pacing",
1092 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1093 	    "Pacing related Controls");
1094 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1095 	    SYSCTL_CHILDREN(rack_hw_pacing),
1096 	    OID_AUTO, "rwnd_factor", CTLFLAG_RW,
1097 	    &rack_hw_rwnd_factor, 2,
1098 	    "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?");
1099 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1100 	    SYSCTL_CHILDREN(rack_hw_pacing),
1101 	    OID_AUTO, "precheck", CTLFLAG_RW,
1102 	    &rack_hw_check_queue, 0,
1103 	    "Do we always precheck the hdwr pacing queue to avoid ENOBUF's?");
1104 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1105 	    SYSCTL_CHILDREN(rack_hw_pacing),
1106 	    OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW,
1107 	    &rack_enobuf_hw_boost_mult, 0,
1108 	    "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?");
1109 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1110 	    SYSCTL_CHILDREN(rack_hw_pacing),
1111 	    OID_AUTO, "pace_enobuf_max", CTLFLAG_RW,
1112 	    &rack_enobuf_hw_max, 2,
1113 	    "What is the max boost the pacing time if we see a ENOBUFS?");
1114 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1115 	    SYSCTL_CHILDREN(rack_hw_pacing),
1116 	    OID_AUTO, "pace_enobuf_min", CTLFLAG_RW,
1117 	    &rack_enobuf_hw_min, 2,
1118 	    "What is the min boost the pacing time if we see a ENOBUFS?");
1119 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1120 	    SYSCTL_CHILDREN(rack_hw_pacing),
1121 	    OID_AUTO, "enable", CTLFLAG_RW,
1122 	    &rack_enable_hw_pacing, 0,
1123 	    "Should RACK attempt to use hw pacing?");
1124 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1125 	    SYSCTL_CHILDREN(rack_hw_pacing),
1126 	    OID_AUTO, "rate_cap", CTLFLAG_RW,
1127 	    &rack_hw_rate_caps, 0,
1128 	    "Does the highest hardware pacing rate cap the rate we will send at??");
1129 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1130 	    SYSCTL_CHILDREN(rack_hw_pacing),
1131 	    OID_AUTO, "uncap_per", CTLFLAG_RW,
1132 	    &rack_hw_rate_cap_per, 0,
1133 	    "If you go over b/w by this amount you will be uncapped (0 = never)");
1134 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1135 	    SYSCTL_CHILDREN(rack_hw_pacing),
1136 	    OID_AUTO, "rate_min", CTLFLAG_RW,
1137 	    &rack_hw_rate_min, 0,
1138 	    "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?");
1139 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1140 	    SYSCTL_CHILDREN(rack_hw_pacing),
1141 	    OID_AUTO, "rate_to_low", CTLFLAG_RW,
1142 	    &rack_hw_rate_to_low, 0,
1143 	    "If we fall below this rate, dis-engage hw pacing?");
1144 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1145 	    SYSCTL_CHILDREN(rack_hw_pacing),
1146 	    OID_AUTO, "up_only", CTLFLAG_RW,
1147 	    &rack_hw_up_only, 0,
1148 	    "Do we allow hw pacing to lower the rate selected?");
1149 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1150 	    SYSCTL_CHILDREN(rack_hw_pacing),
1151 	    OID_AUTO, "extra_mss_precise", CTLFLAG_RW,
1152 	    &rack_hw_pace_extra_slots, 0,
1153 	    "If the rates between software and hardware match precisely how many extra time_betweens do we get?");
1154 	rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1155 	    SYSCTL_CHILDREN(rack_sysctl_root),
1156 	    OID_AUTO,
1157 	    "timely",
1158 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1159 	    "Rack Timely RTT Controls");
1160 	/* Timely based GP dynmics */
1161 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1162 	    SYSCTL_CHILDREN(rack_timely),
1163 	    OID_AUTO, "upper", CTLFLAG_RW,
1164 	    &rack_gp_per_bw_mul_up, 2,
1165 	    "Rack timely upper range for equal b/w (in percentage)");
1166 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1167 	    SYSCTL_CHILDREN(rack_timely),
1168 	    OID_AUTO, "lower", CTLFLAG_RW,
1169 	    &rack_gp_per_bw_mul_down, 4,
1170 	    "Rack timely lower range for equal b/w (in percentage)");
1171 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1172 	    SYSCTL_CHILDREN(rack_timely),
1173 	    OID_AUTO, "rtt_max_mul", CTLFLAG_RW,
1174 	    &rack_gp_rtt_maxmul, 3,
1175 	    "Rack timely multiplier of lowest rtt for rtt_max");
1176 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1177 	    SYSCTL_CHILDREN(rack_timely),
1178 	    OID_AUTO, "rtt_min_div", CTLFLAG_RW,
1179 	    &rack_gp_rtt_mindiv, 4,
1180 	    "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt");
1181 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1182 	    SYSCTL_CHILDREN(rack_timely),
1183 	    OID_AUTO, "rtt_min_mul", CTLFLAG_RW,
1184 	    &rack_gp_rtt_minmul, 1,
1185 	    "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt");
1186 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1187 	    SYSCTL_CHILDREN(rack_timely),
1188 	    OID_AUTO, "decrease", CTLFLAG_RW,
1189 	    &rack_gp_decrease_per, 80,
1190 	    "Rack timely Beta value 80 = .8 (scaled by 100)");
1191 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1192 	    SYSCTL_CHILDREN(rack_timely),
1193 	    OID_AUTO, "increase", CTLFLAG_RW,
1194 	    &rack_gp_increase_per, 2,
1195 	    "Rack timely increase perentage of our GP multiplication factor");
1196 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1197 	    SYSCTL_CHILDREN(rack_timely),
1198 	    OID_AUTO, "lowerbound", CTLFLAG_RW,
1199 	    &rack_per_lower_bound, 50,
1200 	    "Rack timely lowest percentage we allow GP multiplier to fall to");
1201 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1202 	    SYSCTL_CHILDREN(rack_timely),
1203 	    OID_AUTO, "p5_upper", CTLFLAG_RW,
1204 	    &rack_gain_p5_ub, 250,
1205 	    "Profile 5 upper bound to timely gain");
1206 
1207 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1208 	    SYSCTL_CHILDREN(rack_timely),
1209 	    OID_AUTO, "upperboundss", CTLFLAG_RW,
1210 	    &rack_per_upper_bound_ss, 0,
1211 	    "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)");
1212 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1213 	    SYSCTL_CHILDREN(rack_timely),
1214 	    OID_AUTO, "upperboundca", CTLFLAG_RW,
1215 	    &rack_per_upper_bound_ca, 0,
1216 	    "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)");
1217 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1218 	    SYSCTL_CHILDREN(rack_timely),
1219 	    OID_AUTO, "dynamicgp", CTLFLAG_RW,
1220 	    &rack_do_dyn_mul, 0,
1221 	    "Rack timely do we enable dynmaic timely goodput by default");
1222 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1223 	    SYSCTL_CHILDREN(rack_timely),
1224 	    OID_AUTO, "no_rec_red", CTLFLAG_RW,
1225 	    &rack_gp_no_rec_chg, 1,
1226 	    "Rack timely do we prohibit the recovery multiplier from being lowered");
1227 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1228 	    SYSCTL_CHILDREN(rack_timely),
1229 	    OID_AUTO, "red_clear_cnt", CTLFLAG_RW,
1230 	    &rack_timely_dec_clear, 6,
1231 	    "Rack timely what threshold do we count to before another boost during b/w decent");
1232 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1233 	    SYSCTL_CHILDREN(rack_timely),
1234 	    OID_AUTO, "max_push_rise", CTLFLAG_RW,
1235 	    &rack_timely_max_push_rise, 3,
1236 	    "Rack timely how many times do we push up with b/w increase");
1237 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1238 	    SYSCTL_CHILDREN(rack_timely),
1239 	    OID_AUTO, "max_push_drop", CTLFLAG_RW,
1240 	    &rack_timely_max_push_drop, 3,
1241 	    "Rack timely how many times do we push back on b/w decent");
1242 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1243 	    SYSCTL_CHILDREN(rack_timely),
1244 	    OID_AUTO, "min_segs", CTLFLAG_RW,
1245 	    &rack_timely_min_segs, 4,
1246 	    "Rack timely when setting the cwnd what is the min num segments");
1247 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1248 	    SYSCTL_CHILDREN(rack_timely),
1249 	    OID_AUTO, "noback_max", CTLFLAG_RW,
1250 	    &rack_use_max_for_nobackoff, 0,
1251 	    "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min");
1252 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1253 	    SYSCTL_CHILDREN(rack_timely),
1254 	    OID_AUTO, "interim_timely_only", CTLFLAG_RW,
1255 	    &rack_timely_int_timely_only, 0,
1256 	    "Rack timely when doing interim timely's do we only do timely (no b/w consideration)");
1257 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1258 	    SYSCTL_CHILDREN(rack_timely),
1259 	    OID_AUTO, "nonstop", CTLFLAG_RW,
1260 	    &rack_timely_no_stopping, 0,
1261 	    "Rack timely don't stop increase");
1262 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1263 	    SYSCTL_CHILDREN(rack_timely),
1264 	    OID_AUTO, "dec_raise_thresh", CTLFLAG_RW,
1265 	    &rack_down_raise_thresh, 100,
1266 	    "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)");
1267 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1268 	    SYSCTL_CHILDREN(rack_timely),
1269 	    OID_AUTO, "bottom_drag_segs", CTLFLAG_RW,
1270 	    &rack_req_segs, 1,
1271 	    "Bottom dragging if not these many segments outstanding and room");
1272 
1273 	/* TLP and Rack related parameters */
1274 	rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1275 	    SYSCTL_CHILDREN(rack_sysctl_root),
1276 	    OID_AUTO,
1277 	    "tlp",
1278 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1279 	    "TLP and Rack related Controls");
1280 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1281 	    SYSCTL_CHILDREN(rack_tlp),
1282 	    OID_AUTO, "use_rrr", CTLFLAG_RW,
1283 	    &use_rack_rr, 1,
1284 	    "Do we use Rack Rapid Recovery");
1285 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1286 	    SYSCTL_CHILDREN(rack_tlp),
1287 	    OID_AUTO, "post_rec_labc", CTLFLAG_RW,
1288 	    &rack_max_abc_post_recovery, 2,
1289 	    "Since we do early recovery, do we override the l_abc to a value, if so what?");
1290 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1291 	    SYSCTL_CHILDREN(rack_tlp),
1292 	    OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW,
1293 	    &rack_non_rxt_use_cr, 0,
1294 	    "Do we use ss/ca rate if in recovery we are transmitting a new data chunk");
1295 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1296 	    SYSCTL_CHILDREN(rack_tlp),
1297 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
1298 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
1299 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
1300 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1301 	    SYSCTL_CHILDREN(rack_tlp),
1302 	    OID_AUTO, "limit", CTLFLAG_RW,
1303 	    &rack_tlp_limit, 2,
1304 	    "How many TLP's can be sent without sending new data");
1305 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1306 	    SYSCTL_CHILDREN(rack_tlp),
1307 	    OID_AUTO, "use_greater", CTLFLAG_RW,
1308 	    &rack_tlp_use_greater, 1,
1309 	    "Should we use the rack_rtt time if its greater than srtt");
1310 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1311 	    SYSCTL_CHILDREN(rack_tlp),
1312 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
1313 	    &rack_tlp_min, 10000,
1314 	    "TLP minimum timeout per the specification (in microseconds)");
1315 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1316 	    SYSCTL_CHILDREN(rack_tlp),
1317 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
1318 	    &rack_always_send_oldest, 0,
1319 	    "Should we always send the oldest TLP and RACK-TLP");
1320 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1321 	    SYSCTL_CHILDREN(rack_tlp),
1322 	    OID_AUTO, "rack_tlimit", CTLFLAG_RW,
1323 	    &rack_limited_retran, 0,
1324 	    "How many times can a rack timeout drive out sends");
1325 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1326 	    SYSCTL_CHILDREN(rack_tlp),
1327 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
1328 	    &rack_lower_cwnd_at_tlp, 0,
1329 	    "When a TLP completes a retran should we enter recovery");
1330 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1331 	    SYSCTL_CHILDREN(rack_tlp),
1332 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
1333 	    &rack_reorder_thresh, 2,
1334 	    "What factor for rack will be added when seeing reordering (shift right)");
1335 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1336 	    SYSCTL_CHILDREN(rack_tlp),
1337 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
1338 	    &rack_tlp_thresh, 1,
1339 	    "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
1340 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1341 	    SYSCTL_CHILDREN(rack_tlp),
1342 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
1343 	    &rack_reorder_fade, 60000000,
1344 	    "Does reorder detection fade, if so how many microseconds (0 means never)");
1345 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1346 	    SYSCTL_CHILDREN(rack_tlp),
1347 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
1348 	    &rack_pkt_delay, 1000,
1349 	    "Extra RACK time (in microseconds) besides reordering thresh");
1350 
1351 	/* Timer related controls */
1352 	rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1353 	    SYSCTL_CHILDREN(rack_sysctl_root),
1354 	    OID_AUTO,
1355 	    "timers",
1356 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1357 	    "Timer related controls");
1358 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1359 	    SYSCTL_CHILDREN(rack_timers),
1360 	    OID_AUTO, "persmin", CTLFLAG_RW,
1361 	    &rack_persist_min, 250000,
1362 	    "What is the minimum time in microseconds between persists");
1363 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1364 	    SYSCTL_CHILDREN(rack_timers),
1365 	    OID_AUTO, "persmax", CTLFLAG_RW,
1366 	    &rack_persist_max, 2000000,
1367 	    "What is the largest delay in microseconds between persists");
1368 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1369 	    SYSCTL_CHILDREN(rack_timers),
1370 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
1371 	    &rack_delayed_ack_time, 40000,
1372 	    "Delayed ack time (40ms in microseconds)");
1373 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1374 	    SYSCTL_CHILDREN(rack_timers),
1375 	    OID_AUTO, "minrto", CTLFLAG_RW,
1376 	    &rack_rto_min, 30000,
1377 	    "Minimum RTO in microseconds -- set with caution below 1000 due to TLP");
1378 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1379 	    SYSCTL_CHILDREN(rack_timers),
1380 	    OID_AUTO, "maxrto", CTLFLAG_RW,
1381 	    &rack_rto_max, 4000000,
1382 	    "Maximum RTO in microseconds -- should be at least as large as min_rto");
1383 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1384 	    SYSCTL_CHILDREN(rack_timers),
1385 	    OID_AUTO, "minto", CTLFLAG_RW,
1386 	    &rack_min_to, 1000,
1387 	    "Minimum rack timeout in microseconds");
1388 	/* Measure controls */
1389 	rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1390 	    SYSCTL_CHILDREN(rack_sysctl_root),
1391 	    OID_AUTO,
1392 	    "measure",
1393 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1394 	    "Measure related controls");
1395 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1396 	    SYSCTL_CHILDREN(rack_measure),
1397 	    OID_AUTO, "wma_divisor", CTLFLAG_RW,
1398 	    &rack_wma_divisor, 8,
1399 	    "When doing b/w calculation what is the  divisor for the WMA");
1400 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1401 	    SYSCTL_CHILDREN(rack_measure),
1402 	    OID_AUTO, "end_cwnd", CTLFLAG_RW,
1403 	    &rack_cwnd_block_ends_measure, 0,
1404 	    "Does a cwnd just-return end the measurement window (app limited)");
1405 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1406 	    SYSCTL_CHILDREN(rack_measure),
1407 	    OID_AUTO, "end_rwnd", CTLFLAG_RW,
1408 	    &rack_rwnd_block_ends_measure, 0,
1409 	    "Does an rwnd just-return end the measurement window (app limited -- not persists)");
1410 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1411 	    SYSCTL_CHILDREN(rack_measure),
1412 	    OID_AUTO, "min_target", CTLFLAG_RW,
1413 	    &rack_def_data_window, 20,
1414 	    "What is the minimum target window (in mss) for a GP measurements");
1415 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1416 	    SYSCTL_CHILDREN(rack_measure),
1417 	    OID_AUTO, "goal_bdp", CTLFLAG_RW,
1418 	    &rack_goal_bdp, 2,
1419 	    "What is the goal BDP to measure");
1420 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1421 	    SYSCTL_CHILDREN(rack_measure),
1422 	    OID_AUTO, "min_srtts", CTLFLAG_RW,
1423 	    &rack_min_srtts, 1,
1424 	    "What is the goal BDP to measure");
1425 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1426 	    SYSCTL_CHILDREN(rack_measure),
1427 	    OID_AUTO, "min_measure_tim", CTLFLAG_RW,
1428 	    &rack_min_measure_usec, 0,
1429 	    "What is the Minimum time time for a measurement if 0, this is off");
1430 	/* Features */
1431 	rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1432 	    SYSCTL_CHILDREN(rack_sysctl_root),
1433 	    OID_AUTO,
1434 	    "features",
1435 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1436 	    "Feature controls");
1437 	SYSCTL_ADD_U64(&rack_sysctl_ctx,
1438 	    SYSCTL_CHILDREN(rack_features),
1439 	    OID_AUTO, "rxt_clamp_thresh", CTLFLAG_RW,
1440 	    &rack_rxt_clamp_thresh, 0,
1441 	    "Bit encoded clamping setup bits CCCC CCCCC UUUU UULF PPPP PPPP PPPP PPPP");
1442 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1443 	    SYSCTL_CHILDREN(rack_features),
1444 	    OID_AUTO, "hybrid_set_maxseg", CTLFLAG_RW,
1445 	    &rack_hybrid_allow_set_maxseg, 0,
1446 	    "Should hybrid pacing allow the setmss command");
1447 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1448 	    SYSCTL_CHILDREN(rack_features),
1449 	    OID_AUTO, "cmpack", CTLFLAG_RW,
1450 	    &rack_use_cmp_acks, 1,
1451 	    "Should RACK have LRO send compressed acks");
1452 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1453 	    SYSCTL_CHILDREN(rack_features),
1454 	    OID_AUTO, "fsb", CTLFLAG_RW,
1455 	    &rack_use_fsb, 1,
1456 	    "Should RACK use the fast send block?");
1457 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1458 	    SYSCTL_CHILDREN(rack_features),
1459 	    OID_AUTO, "rfo", CTLFLAG_RW,
1460 	    &rack_use_rfo, 1,
1461 	    "Should RACK use rack_fast_output()?");
1462 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1463 	    SYSCTL_CHILDREN(rack_features),
1464 	    OID_AUTO, "rsmrfo", CTLFLAG_RW,
1465 	    &rack_use_rsm_rfo, 1,
1466 	    "Should RACK use rack_fast_rsm_output()?");
1467 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1468 	    SYSCTL_CHILDREN(rack_features),
1469 	    OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW,
1470 	    &rack_enable_mqueue_for_nonpaced, 0,
1471 	    "Should RACK use mbuf queuing for non-paced connections");
1472 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1473 	    SYSCTL_CHILDREN(rack_features),
1474 	    OID_AUTO, "hystartplusplus", CTLFLAG_RW,
1475 	    &rack_do_hystart, 0,
1476 	    "Should RACK enable HyStart++ on connections?");
1477 	/* Misc rack controls */
1478 	rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
1479 	    SYSCTL_CHILDREN(rack_sysctl_root),
1480 	    OID_AUTO,
1481 	    "misc",
1482 	    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1483 	    "Misc related controls");
1484 #ifdef TCP_ACCOUNTING
1485 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1486 	    SYSCTL_CHILDREN(rack_misc),
1487 	    OID_AUTO, "tcp_acct", CTLFLAG_RW,
1488 	    &rack_tcp_accounting, 0,
1489 	    "Should we turn on TCP accounting for all rack sessions?");
1490 #endif
1491 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1492 	    SYSCTL_CHILDREN(rack_misc),
1493 	    OID_AUTO, "dnd", CTLFLAG_RW,
1494 	    &rack_dnd_default, 0,
1495 	    "Do not disturb default for rack_rrr = 3");
1496 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1497 	    SYSCTL_CHILDREN(rack_misc),
1498 	    OID_AUTO, "sad_seg_per", CTLFLAG_RW,
1499 	    &sad_seg_size_per, 800,
1500 	    "Percentage of segment size needed in a sack 800 = 80.0?");
1501 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1502 	    SYSCTL_CHILDREN(rack_misc),
1503 	    OID_AUTO, "rxt_controls", CTLFLAG_RW,
1504 	    &rack_rxt_controls, 0,
1505 	    "Retransmit sending size controls (valid  values 0, 1, 2 default=1)?");
1506 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1507 	    SYSCTL_CHILDREN(rack_misc),
1508 	    OID_AUTO, "rack_hibeta", CTLFLAG_RW,
1509 	    &rack_hibeta_setting, 0,
1510 	    "Do we ue a high beta (80 instead of 50)?");
1511 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1512 	    SYSCTL_CHILDREN(rack_misc),
1513 	    OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW,
1514 	    &rack_apply_rtt_with_reduced_conf, 0,
1515 	    "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?");
1516 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1517 	    SYSCTL_CHILDREN(rack_misc),
1518 	    OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW,
1519 	    &rack_dsack_std_based, 3,
1520 	    "How do we process dsack with respect to rack timers, bit field, 3 is standards based?");
1521 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1522 	    SYSCTL_CHILDREN(rack_misc),
1523 	    OID_AUTO, "prr_addback_max", CTLFLAG_RW,
1524 	    &rack_prr_addbackmax, 2,
1525 	    "What is the maximum number of MSS we allow to be added back if prr can't send all its data?");
1526 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1527 	    SYSCTL_CHILDREN(rack_misc),
1528 	    OID_AUTO, "stats_gets_ms", CTLFLAG_RW,
1529 	    &rack_stats_gets_ms_rtt, 1,
1530 	    "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?");
1531 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1532 	    SYSCTL_CHILDREN(rack_misc),
1533 	    OID_AUTO, "clientlowbuf", CTLFLAG_RW,
1534 	    &rack_client_low_buf, 0,
1535 	    "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?");
1536 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1537 	    SYSCTL_CHILDREN(rack_misc),
1538 	    OID_AUTO, "defprofile", CTLFLAG_RW,
1539 	    &rack_def_profile, 0,
1540 	    "Should RACK use a default profile (0=no, num == profile num)?");
1541 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1542 	    SYSCTL_CHILDREN(rack_misc),
1543 	    OID_AUTO, "shared_cwnd", CTLFLAG_RW,
1544 	    &rack_enable_shared_cwnd, 1,
1545 	    "Should RACK try to use the shared cwnd on connections where allowed");
1546 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1547 	    SYSCTL_CHILDREN(rack_misc),
1548 	    OID_AUTO, "limits_on_scwnd", CTLFLAG_RW,
1549 	    &rack_limits_scwnd, 1,
1550 	    "Should RACK place low end time limits on the shared cwnd feature");
1551 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1552 	    SYSCTL_CHILDREN(rack_misc),
1553 	    OID_AUTO, "no_prr", CTLFLAG_RW,
1554 	    &rack_disable_prr, 0,
1555 	    "Should RACK not use prr and only pace (must have pacing on)");
1556 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1557 	    SYSCTL_CHILDREN(rack_misc),
1558 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
1559 	    &rack_verbose_logging, 0,
1560 	    "Should RACK black box logging be verbose");
1561 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1562 	    SYSCTL_CHILDREN(rack_misc),
1563 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
1564 	    &rack_ignore_data_after_close, 1,
1565 	    "Do we hold off sending a RST until all pending data is ack'd");
1566 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1567 	    SYSCTL_CHILDREN(rack_misc),
1568 	    OID_AUTO, "no_sack_needed", CTLFLAG_RW,
1569 	    &rack_sack_not_required, 1,
1570 	    "Do we allow rack to run on connections not supporting SACK");
1571 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1572 	    SYSCTL_CHILDREN(rack_misc),
1573 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
1574 	    &rack_send_a_lot_in_prr, 1,
1575 	    "Send a lot in prr");
1576 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
1577 	    SYSCTL_CHILDREN(rack_misc),
1578 	    OID_AUTO, "autoscale", CTLFLAG_RW,
1579 	    &rack_autosndbuf_inc, 20,
1580 	    "What percentage should rack scale up its snd buffer by?");
1581 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1582 	    SYSCTL_CHILDREN(rack_misc),
1583 	    OID_AUTO, "rnds_for_rxt_clamp", CTLFLAG_RW,
1584 	    &rack_rxt_min_rnds, 10,
1585 	    "Number of rounds needed between RTT clamps due to high loss rates");
1586 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1587 	    SYSCTL_CHILDREN(rack_misc),
1588 	    OID_AUTO, "rnds_for_unclamp", CTLFLAG_RW,
1589 	    &rack_unclamp_round_thresh, 100,
1590 	    "Number of rounds needed with no loss to unclamp");
1591 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1592 	    SYSCTL_CHILDREN(rack_misc),
1593 	    OID_AUTO, "rxt_threshs_for_unclamp", CTLFLAG_RW,
1594 	    &rack_unclamp_rxt_thresh, 5,
1595 	   "Percentage of retransmits we need to be under to unclamp (5 = .5 percent)\n");
1596 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1597 	    SYSCTL_CHILDREN(rack_misc),
1598 	    OID_AUTO, "clamp_ss_upper", CTLFLAG_RW,
1599 	    &rack_clamp_ss_upper, 110,
1600 	    "Clamp percentage ceiling in SS?");
1601 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1602 	    SYSCTL_CHILDREN(rack_misc),
1603 	    OID_AUTO, "clamp_ca_upper", CTLFLAG_RW,
1604 	    &rack_clamp_ca_upper, 110,
1605 	    "Clamp percentage ceiling in CA?");
1606 	/* Sack Attacker detection stuff */
1607 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1608 	    SYSCTL_CHILDREN(rack_attack),
1609 	    OID_AUTO, "merge_out", CTLFLAG_RW,
1610 	    &rack_merge_out_sacks_on_attack, 0,
1611 	    "Do we merge the sendmap when we decide we are being attacked?");
1612 
1613 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1614 	    SYSCTL_CHILDREN(rack_attack),
1615 	    OID_AUTO, "detect_highsackratio", CTLFLAG_RW,
1616 	    &rack_highest_sack_thresh_seen, 0,
1617 	    "Highest sack to ack ratio seen");
1618 	SYSCTL_ADD_U32(&rack_sysctl_ctx,
1619 	    SYSCTL_CHILDREN(rack_attack),
1620 	    OID_AUTO, "detect_highmoveratio", CTLFLAG_RW,
1621 	    &rack_highest_move_thresh_seen, 0,
1622 	    "Highest move to non-move ratio seen");
1623 	rack_ack_total = counter_u64_alloc(M_WAITOK);
1624 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1625 	    SYSCTL_CHILDREN(rack_attack),
1626 	    OID_AUTO, "acktotal", CTLFLAG_RD,
1627 	    &rack_ack_total,
1628 	    "Total number of Ack's");
1629 	rack_express_sack = counter_u64_alloc(M_WAITOK);
1630 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1631 	    SYSCTL_CHILDREN(rack_attack),
1632 	    OID_AUTO, "exp_sacktotal", CTLFLAG_RD,
1633 	    &rack_express_sack,
1634 	    "Total expresss number of Sack's");
1635 	rack_sack_total = counter_u64_alloc(M_WAITOK);
1636 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1637 	    SYSCTL_CHILDREN(rack_attack),
1638 	    OID_AUTO, "sacktotal", CTLFLAG_RD,
1639 	    &rack_sack_total,
1640 	    "Total number of SACKs");
1641 	rack_move_none = counter_u64_alloc(M_WAITOK);
1642 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1643 	    SYSCTL_CHILDREN(rack_attack),
1644 	    OID_AUTO, "move_none", CTLFLAG_RD,
1645 	    &rack_move_none,
1646 	    "Total number of SACK index reuse of positions under threshold");
1647 	rack_move_some = counter_u64_alloc(M_WAITOK);
1648 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1649 	    SYSCTL_CHILDREN(rack_attack),
1650 	    OID_AUTO, "move_some", CTLFLAG_RD,
1651 	    &rack_move_some,
1652 	    "Total number of SACK index reuse of positions over threshold");
1653 	rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK);
1654 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1655 	    SYSCTL_CHILDREN(rack_attack),
1656 	    OID_AUTO, "attacks", CTLFLAG_RD,
1657 	    &rack_sack_attacks_detected,
1658 	    "Total number of SACK attackers that had sack disabled");
1659 	rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK);
1660 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1661 	    SYSCTL_CHILDREN(rack_attack),
1662 	    OID_AUTO, "reversed", CTLFLAG_RD,
1663 	    &rack_sack_attacks_reversed,
1664 	    "Total number of SACK attackers that were later determined false positive");
1665 	rack_sack_attacks_suspect = counter_u64_alloc(M_WAITOK);
1666 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1667 	    SYSCTL_CHILDREN(rack_attack),
1668 	    OID_AUTO, "suspect", CTLFLAG_RD,
1669 	    &rack_sack_attacks_suspect,
1670 	    "Total number of SACKs that triggered early detection");
1671 
1672 	rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK);
1673 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1674 	    SYSCTL_CHILDREN(rack_attack),
1675 	    OID_AUTO, "nextmerge", CTLFLAG_RD,
1676 	    &rack_sack_used_next_merge,
1677 	    "Total number of times we used the next merge");
1678 	rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK);
1679 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1680 	    SYSCTL_CHILDREN(rack_attack),
1681 	    OID_AUTO, "prevmerge", CTLFLAG_RD,
1682 	    &rack_sack_used_prev_merge,
1683 	    "Total number of times we used the prev merge");
1684 	/* Counters */
1685 	rack_total_bytes = counter_u64_alloc(M_WAITOK);
1686 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1687 	    SYSCTL_CHILDREN(rack_counters),
1688 	    OID_AUTO, "totalbytes", CTLFLAG_RD,
1689 	    &rack_total_bytes,
1690 	    "Total number of bytes sent");
1691 	rack_fto_send = counter_u64_alloc(M_WAITOK);
1692 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1693 	    SYSCTL_CHILDREN(rack_counters),
1694 	    OID_AUTO, "fto_send", CTLFLAG_RD,
1695 	    &rack_fto_send, "Total number of rack_fast_output sends");
1696 	rack_fto_rsm_send = counter_u64_alloc(M_WAITOK);
1697 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1698 	    SYSCTL_CHILDREN(rack_counters),
1699 	    OID_AUTO, "fto_rsm_send", CTLFLAG_RD,
1700 	    &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends");
1701 	rack_nfto_resend = counter_u64_alloc(M_WAITOK);
1702 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1703 	    SYSCTL_CHILDREN(rack_counters),
1704 	    OID_AUTO, "nfto_resend", CTLFLAG_RD,
1705 	    &rack_nfto_resend, "Total number of rack_output retransmissions");
1706 	rack_non_fto_send = counter_u64_alloc(M_WAITOK);
1707 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1708 	    SYSCTL_CHILDREN(rack_counters),
1709 	    OID_AUTO, "nfto_send", CTLFLAG_RD,
1710 	    &rack_non_fto_send, "Total number of rack_output first sends");
1711 	rack_extended_rfo = counter_u64_alloc(M_WAITOK);
1712 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1713 	    SYSCTL_CHILDREN(rack_counters),
1714 	    OID_AUTO, "rfo_extended", CTLFLAG_RD,
1715 	    &rack_extended_rfo, "Total number of times we extended rfo");
1716 
1717 	rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK);
1718 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1719 	    SYSCTL_CHILDREN(rack_counters),
1720 	    OID_AUTO, "hwpace_init_fail", CTLFLAG_RD,
1721 	    &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing");
1722 	rack_hw_pace_lost = counter_u64_alloc(M_WAITOK);
1723 
1724 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1725 	    SYSCTL_CHILDREN(rack_counters),
1726 	    OID_AUTO, "hwpace_lost", CTLFLAG_RD,
1727 	    &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing");
1728 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
1729 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1730 	    SYSCTL_CHILDREN(rack_counters),
1731 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
1732 	    &rack_tlp_tot,
1733 	    "Total number of tail loss probe expirations");
1734 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
1735 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1736 	    SYSCTL_CHILDREN(rack_counters),
1737 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
1738 	    &rack_tlp_newdata,
1739 	    "Total number of tail loss probe sending new data");
1740 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
1741 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1742 	    SYSCTL_CHILDREN(rack_counters),
1743 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
1744 	    &rack_tlp_retran,
1745 	    "Total number of tail loss probe sending retransmitted data");
1746 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
1747 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1748 	    SYSCTL_CHILDREN(rack_counters),
1749 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
1750 	    &rack_tlp_retran_bytes,
1751 	    "Total bytes of tail loss probe sending retransmitted data");
1752 	rack_to_tot = counter_u64_alloc(M_WAITOK);
1753 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1754 	    SYSCTL_CHILDREN(rack_counters),
1755 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
1756 	    &rack_to_tot,
1757 	    "Total number of times the rack to expired");
1758 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
1759 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1760 	    SYSCTL_CHILDREN(rack_counters),
1761 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
1762 	    &rack_saw_enobuf,
1763 	    "Total number of times a sends returned enobuf for non-hdwr paced connections");
1764 	rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK);
1765 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1766 	    SYSCTL_CHILDREN(rack_counters),
1767 	    OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD,
1768 	    &rack_saw_enobuf_hw,
1769 	    "Total number of times a send returned enobuf for hdwr paced connections");
1770 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
1771 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1772 	    SYSCTL_CHILDREN(rack_counters),
1773 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
1774 	    &rack_saw_enetunreach,
1775 	    "Total number of times a send received a enetunreachable");
1776 	rack_hot_alloc = counter_u64_alloc(M_WAITOK);
1777 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1778 	    SYSCTL_CHILDREN(rack_counters),
1779 	    OID_AUTO, "alloc_hot", CTLFLAG_RD,
1780 	    &rack_hot_alloc,
1781 	    "Total allocations from the top of our list");
1782 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
1783 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1784 	    SYSCTL_CHILDREN(rack_counters),
1785 	    OID_AUTO, "allocs", CTLFLAG_RD,
1786 	    &rack_to_alloc,
1787 	    "Total allocations of tracking structures");
1788 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
1789 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1790 	    SYSCTL_CHILDREN(rack_counters),
1791 	    OID_AUTO, "allochard", CTLFLAG_RD,
1792 	    &rack_to_alloc_hard,
1793 	    "Total allocations done with sleeping the hard way");
1794 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
1795 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1796 	    SYSCTL_CHILDREN(rack_counters),
1797 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
1798 	    &rack_to_alloc_emerg,
1799 	    "Total allocations done from emergency cache");
1800 	rack_to_alloc_limited = counter_u64_alloc(M_WAITOK);
1801 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1802 	    SYSCTL_CHILDREN(rack_counters),
1803 	    OID_AUTO, "alloc_limited", CTLFLAG_RD,
1804 	    &rack_to_alloc_limited,
1805 	    "Total allocations dropped due to limit");
1806 	rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK);
1807 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1808 	    SYSCTL_CHILDREN(rack_counters),
1809 	    OID_AUTO, "alloc_limited_conns", CTLFLAG_RD,
1810 	    &rack_alloc_limited_conns,
1811 	    "Connections with allocations dropped due to limit");
1812 	rack_split_limited = counter_u64_alloc(M_WAITOK);
1813 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1814 	    SYSCTL_CHILDREN(rack_counters),
1815 	    OID_AUTO, "split_limited", CTLFLAG_RD,
1816 	    &rack_split_limited,
1817 	    "Split allocations dropped due to limit");
1818 	rack_rxt_clamps_cwnd = counter_u64_alloc(M_WAITOK);
1819 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1820 	    SYSCTL_CHILDREN(rack_counters),
1821 	    OID_AUTO, "rxt_clamps_cwnd", CTLFLAG_RD,
1822 	    &rack_rxt_clamps_cwnd,
1823 	    "Number of times that excessive rxt clamped the cwnd down");
1824 	rack_rxt_clamps_cwnd_uniq = counter_u64_alloc(M_WAITOK);
1825 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1826 	    SYSCTL_CHILDREN(rack_counters),
1827 	    OID_AUTO, "rxt_clamps_cwnd_uniq", CTLFLAG_RD,
1828 	    &rack_rxt_clamps_cwnd_uniq,
1829 	    "Number of connections that have had excessive rxt clamped the cwnd down");
1830 	rack_persists_sends = counter_u64_alloc(M_WAITOK);
1831 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1832 	    SYSCTL_CHILDREN(rack_counters),
1833 	    OID_AUTO, "persist_sends", CTLFLAG_RD,
1834 	    &rack_persists_sends,
1835 	    "Number of times we sent a persist probe");
1836 	rack_persists_acks = counter_u64_alloc(M_WAITOK);
1837 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1838 	    SYSCTL_CHILDREN(rack_counters),
1839 	    OID_AUTO, "persist_acks", CTLFLAG_RD,
1840 	    &rack_persists_acks,
1841 	    "Number of times a persist probe was acked");
1842 	rack_persists_loss = counter_u64_alloc(M_WAITOK);
1843 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1844 	    SYSCTL_CHILDREN(rack_counters),
1845 	    OID_AUTO, "persist_loss", CTLFLAG_RD,
1846 	    &rack_persists_loss,
1847 	    "Number of times we detected a lost persist probe (no ack)");
1848 	rack_persists_lost_ends = counter_u64_alloc(M_WAITOK);
1849 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1850 	    SYSCTL_CHILDREN(rack_counters),
1851 	    OID_AUTO, "persist_loss_ends", CTLFLAG_RD,
1852 	    &rack_persists_lost_ends,
1853 	    "Number of lost persist probe (no ack) that the run ended with a PERSIST abort");
1854 #ifdef INVARIANTS
1855 	rack_adjust_map_bw = counter_u64_alloc(M_WAITOK);
1856 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1857 	    SYSCTL_CHILDREN(rack_counters),
1858 	    OID_AUTO, "map_adjust_req", CTLFLAG_RD,
1859 	    &rack_adjust_map_bw,
1860 	    "Number of times we hit the case where the sb went up and down on a sendmap entry");
1861 #endif
1862 	rack_multi_single_eq = counter_u64_alloc(M_WAITOK);
1863 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1864 	    SYSCTL_CHILDREN(rack_counters),
1865 	    OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD,
1866 	    &rack_multi_single_eq,
1867 	    "Number of compressed acks total represented");
1868 	rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK);
1869 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1870 	    SYSCTL_CHILDREN(rack_counters),
1871 	    OID_AUTO, "cmp_ack_not", CTLFLAG_RD,
1872 	    &rack_proc_non_comp_ack,
1873 	    "Number of non compresseds acks that we processed");
1874 
1875 
1876 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
1877 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1878 	    SYSCTL_CHILDREN(rack_counters),
1879 	    OID_AUTO, "sack_long", CTLFLAG_RD,
1880 	    &rack_sack_proc_all,
1881 	    "Total times we had to walk whole list for sack processing");
1882 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
1883 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1884 	    SYSCTL_CHILDREN(rack_counters),
1885 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
1886 	    &rack_sack_proc_restart,
1887 	    "Total times we had to walk whole list due to a restart");
1888 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
1889 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1890 	    SYSCTL_CHILDREN(rack_counters),
1891 	    OID_AUTO, "sack_short", CTLFLAG_RD,
1892 	    &rack_sack_proc_short,
1893 	    "Total times we took shortcut for sack processing");
1894 	rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK);
1895 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1896 	    SYSCTL_CHILDREN(rack_attack),
1897 	    OID_AUTO, "skipacked", CTLFLAG_RD,
1898 	    &rack_sack_skipped_acked,
1899 	    "Total number of times we skipped previously sacked");
1900 	rack_sack_splits = counter_u64_alloc(M_WAITOK);
1901 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1902 	    SYSCTL_CHILDREN(rack_attack),
1903 	    OID_AUTO, "ofsplit", CTLFLAG_RD,
1904 	    &rack_sack_splits,
1905 	    "Total number of times we did the old fashion tree split");
1906 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
1907 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1908 	    SYSCTL_CHILDREN(rack_counters),
1909 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
1910 	    &rack_input_idle_reduces,
1911 	    "Total number of idle reductions on input");
1912 	rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK);
1913 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1914 	    SYSCTL_CHILDREN(rack_counters),
1915 	    OID_AUTO, "collapsed_win_seen", CTLFLAG_RD,
1916 	    &rack_collapsed_win_seen,
1917 	    "Total number of collapsed window events seen (where our window shrinks)");
1918 
1919 	rack_collapsed_win = counter_u64_alloc(M_WAITOK);
1920 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1921 	    SYSCTL_CHILDREN(rack_counters),
1922 	    OID_AUTO, "collapsed_win", CTLFLAG_RD,
1923 	    &rack_collapsed_win,
1924 	    "Total number of collapsed window events where we mark packets");
1925 	rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK);
1926 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1927 	    SYSCTL_CHILDREN(rack_counters),
1928 	    OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD,
1929 	    &rack_collapsed_win_rxt,
1930 	    "Total number of packets that were retransmitted");
1931 	rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK);
1932 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1933 	    SYSCTL_CHILDREN(rack_counters),
1934 	    OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD,
1935 	    &rack_collapsed_win_rxt_bytes,
1936 	    "Total number of bytes that were retransmitted");
1937 	rack_try_scwnd = counter_u64_alloc(M_WAITOK);
1938 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
1939 	    SYSCTL_CHILDREN(rack_counters),
1940 	    OID_AUTO, "tried_scwnd", CTLFLAG_RD,
1941 	    &rack_try_scwnd,
1942 	    "Total number of scwnd attempts");
1943 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
1944 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1945 	    OID_AUTO, "outsize", CTLFLAG_RD,
1946 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
1947 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
1948 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
1949 	    OID_AUTO, "opts", CTLFLAG_RD,
1950 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
1951 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
1952 	    SYSCTL_CHILDREN(rack_sysctl_root),
1953 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1954 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
1955 }
1956 
1957 static uint32_t
1958 rc_init_window(struct tcp_rack *rack)
1959 {
1960 	uint32_t win;
1961 
1962 	if (rack->rc_init_win == 0) {
1963 		/*
1964 		 * Nothing set by the user, use the system stack
1965 		 * default.
1966 		 */
1967 		return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)));
1968 	}
1969 	win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win;
1970 	return (win);
1971 }
1972 
1973 static uint64_t
1974 rack_get_fixed_pacing_bw(struct tcp_rack *rack)
1975 {
1976 	if (IN_FASTRECOVERY(rack->rc_tp->t_flags))
1977 		return (rack->r_ctl.rc_fixed_pacing_rate_rec);
1978 	else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
1979 		return (rack->r_ctl.rc_fixed_pacing_rate_ss);
1980 	else
1981 		return (rack->r_ctl.rc_fixed_pacing_rate_ca);
1982 }
1983 
1984 static void
1985 rack_log_hybrid_bw(struct tcp_rack *rack, uint32_t seq, uint64_t cbw, uint64_t tim,
1986 	uint64_t data, uint8_t mod, uint16_t aux,
1987 	struct tcp_sendfile_track *cur, int line)
1988 {
1989 #ifdef TCP_REQUEST_TRK
1990 	int do_log = 0;
1991 
1992 	/*
1993 	 * The rate cap one is noisy and only should come out when normal BB logging
1994 	 * is enabled, the other logs (not RATE_CAP and NOT CAP_CALC) only come out
1995 	 * once per chunk and make up the BBpoint that can be turned on by the client.
1996 	 */
1997 	if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
1998 		/*
1999 		 * The very noisy two need to only come out when
2000 		 * we have verbose logging on.
2001 		 */
2002 		if (rack_verbose_logging != 0)
2003 			do_log = tcp_bblogging_on(rack->rc_tp);
2004 		else
2005 			do_log = 0;
2006 	} else if (mod != HYBRID_LOG_BW_MEASURE) {
2007 		/*
2008 		 * All other less noisy logs here except the measure which
2009 		 * also needs to come out on the point and the log.
2010 		 */
2011 		do_log = tcp_bblogging_on(rack->rc_tp);
2012 	} else {
2013 		do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING);
2014 	}
2015 
2016 	if (do_log) {
2017 		union tcp_log_stackspecific log;
2018 		struct timeval tv;
2019 		uint64_t lt_bw;
2020 
2021 		/* Convert our ms to a microsecond */
2022 		memset(&log, 0, sizeof(log));
2023 
2024 		log.u_bbr.cwnd_gain = line;
2025 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2026 		log.u_bbr.rttProp = tim;
2027 		log.u_bbr.bw_inuse = cbw;
2028 		log.u_bbr.delRate = rack_get_gp_est(rack);
2029 		lt_bw = rack_get_lt_bw(rack);
2030 		log.u_bbr.flex1 = seq;
2031 		log.u_bbr.pacing_gain = aux;
2032 		/* lt_bw = < flex3 | flex2 > */
2033 		log.u_bbr.flex2 = (uint32_t)(lt_bw & 0x00000000ffffffff);
2034 		log.u_bbr.flex3 = (uint32_t)((lt_bw >> 32) & 0x00000000ffffffff);
2035 		/* Record the last obtained us rtt in inflight */
2036 		if (cur == NULL) {
2037 			/* Make sure we are looking at the right log if an overide comes in */
2038 			cur = rack->r_ctl.rc_last_sft;
2039 		}
2040 		if (rack->r_ctl.rack_rs.rs_flags != RACK_RTT_EMPTY)
2041 			log.u_bbr.inflight = rack->r_ctl.rack_rs.rs_us_rtt;
2042 		else {
2043 			/* Use the last known rtt i.e. the rack-rtt */
2044 			log.u_bbr.inflight = rack->rc_rack_rtt;
2045 		}
2046 		if (cur != NULL) {
2047 			uint64_t off;
2048 
2049 			log.u_bbr.cur_del_rate = cur->deadline;
2050 			if ((mod == HYBRID_LOG_RATE_CAP) || (mod == HYBRID_LOG_CAP_CALC)) {
2051 				/* start = < lost | pkt_epoch > */
2052 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2053 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2054 				log.u_bbr.flex6 = cur->start_seq;
2055 				log.u_bbr.pkts_out = cur->end_seq;
2056 			} else {
2057 				/* start = < lost | pkt_epoch > */
2058 				log.u_bbr.pkt_epoch = (uint32_t)(cur->start & 0x00000000ffffffff);
2059 				log.u_bbr.lost = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2060 				/* end = < pkts_out | flex6 > */
2061 				log.u_bbr.flex6 = (uint32_t)(cur->end & 0x00000000ffffffff);
2062 				log.u_bbr.pkts_out = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2063 			}
2064 			/* first_send = <lt_epoch | epoch> */
2065 			log.u_bbr.epoch = (uint32_t)(cur->first_send & 0x00000000ffffffff);
2066 			log.u_bbr.lt_epoch = (uint32_t)((cur->first_send >> 32) & 0x00000000ffffffff);
2067 			/* localtime = <delivered | applimited>*/
2068 			log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2069 			log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2070 #ifdef TCP_REQUEST_TRK
2071 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2072 			log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2073 #endif
2074 			log.u_bbr.flex4 = (uint32_t)(rack->rc_tp->t_sndbytes - cur->sent_at_fs);
2075 			log.u_bbr.flex5 = (uint32_t)(rack->rc_tp->t_snd_rxt_bytes - cur->rxt_at_fs);
2076 			log.u_bbr.flex7 = (uint16_t)cur->hybrid_flags;
2077 		} else {
2078 			log.u_bbr.flex7 = 0xffff;
2079 			log.u_bbr.cur_del_rate = 0xffffffffffffffff;
2080 		}
2081 		/*
2082 		 * Compose bbr_state to be a bit wise 0000ADHF
2083 		 * where A is the always_pace flag
2084 		 * where D is the dgp_on flag
2085 		 * where H is the hybrid_mode on flag
2086 		 * where F is the use_fixed_rate flag.
2087 		 */
2088 		log.u_bbr.bbr_state = rack->rc_always_pace;
2089 		log.u_bbr.bbr_state <<= 1;
2090 		log.u_bbr.bbr_state |= rack->dgp_on;
2091 		log.u_bbr.bbr_state <<= 1;
2092 		log.u_bbr.bbr_state |= rack->rc_hybrid_mode;
2093 		log.u_bbr.bbr_state <<= 1;
2094 		log.u_bbr.bbr_state |= rack->use_fixed_rate;
2095 		log.u_bbr.flex8 = mod;
2096 		tcp_log_event(rack->rc_tp, NULL,
2097 		    &rack->rc_inp->inp_socket->so_rcv,
2098 		    &rack->rc_inp->inp_socket->so_snd,
2099 		    TCP_HYBRID_PACING_LOG, 0,
2100 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2101 
2102 	}
2103 #endif
2104 }
2105 
2106 #ifdef TCP_REQUEST_TRK
2107 static void
2108 rack_log_hybrid_sends(struct tcp_rack *rack, struct tcp_sendfile_track *cur, int line)
2109 {
2110 	if (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING)) {
2111 		union tcp_log_stackspecific log;
2112 		struct timeval tv;
2113 		uint64_t off;
2114 
2115 		/* Convert our ms to a microsecond */
2116 		memset(&log, 0, sizeof(log));
2117 
2118 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2119 		log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
2120 		log.u_bbr.delRate = cur->sent_at_fs;
2121 		log.u_bbr.rttProp = rack->rc_tp->t_snd_rxt_bytes;
2122 		log.u_bbr.bw_inuse = cur->rxt_at_fs;
2123 		log.u_bbr.cwnd_gain = line;
2124 		off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
2125 		log.u_bbr.bbr_substate = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
2126 		/* start = < flex1 | flex2 > */
2127 		log.u_bbr.flex2 = (uint32_t)(cur->start & 0x00000000ffffffff);
2128 		log.u_bbr.flex1 = (uint32_t)((cur->start >> 32) & 0x00000000ffffffff);
2129 		/* end = < flex3 | flex4 > */
2130 		log.u_bbr.flex4 = (uint32_t)(cur->end & 0x00000000ffffffff);
2131 		log.u_bbr.flex3 = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff);
2132 
2133 		/* localtime = <delivered | applimited>*/
2134 		log.u_bbr.applimited = (uint32_t)(cur->localtime & 0x00000000ffffffff);
2135 		log.u_bbr.delivered = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
2136 		/* client timestamp = <lt_epoch | epoch>*/
2137 		log.u_bbr.epoch = (uint32_t)(cur->timestamp & 0x00000000ffffffff);
2138 		log.u_bbr.lt_epoch = (uint32_t)((cur->timestamp >> 32) & 0x00000000ffffffff);
2139 		/* now set all the flags in */
2140 		log.u_bbr.pkts_out = cur->hybrid_flags;
2141 		log.u_bbr.flex6 = cur->flags;
2142 		/*
2143 		 * Last send time  = <flex5 | pkt_epoch>  note we do not distinguish cases
2144 		 * where a false retransmit occurred so first_send  <-> lastsend may
2145 		 * include longer time then it actually took if we have a false rxt.
2146 		 */
2147 		log.u_bbr.pkt_epoch = (uint32_t)(rack->r_ctl.last_tmit_time_acked & 0x00000000ffffffff);
2148 		log.u_bbr.flex5 = (uint32_t)((rack->r_ctl.last_tmit_time_acked >> 32) & 0x00000000ffffffff);
2149 
2150 		log.u_bbr.flex8 = HYBRID_LOG_SENT_LOST;
2151 		tcp_log_event(rack->rc_tp, NULL,
2152 		    &rack->rc_inp->inp_socket->so_rcv,
2153 		    &rack->rc_inp->inp_socket->so_snd,
2154 		    TCP_HYBRID_PACING_LOG, 0,
2155 		    0, &log, false, NULL, __func__, __LINE__, &tv);
2156 	}
2157 }
2158 #endif
2159 
2160 static inline uint64_t
2161 rack_compensate_for_linerate(struct tcp_rack *rack, uint64_t bw)
2162 {
2163 	uint64_t ret_bw, ether;
2164 	uint64_t u_segsiz;
2165 
2166 	ether = rack->rc_tp->t_maxseg + sizeof(struct tcphdr);
2167 	if (rack->r_is_v6){
2168 #ifdef INET6
2169 		ether += sizeof(struct ip6_hdr);
2170 #endif
2171 		ether += 14;	/* eheader size 6+6+2 */
2172 	} else {
2173 #ifdef INET
2174 		ether += sizeof(struct ip);
2175 #endif
2176 		ether += 14;	/* eheader size 6+6+2 */
2177 	}
2178 	u_segsiz = (uint64_t)min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs);
2179 	ret_bw = bw;
2180 	ret_bw *= ether;
2181 	ret_bw /= u_segsiz;
2182 	return (ret_bw);
2183 }
2184 
2185 static void
2186 rack_rate_cap_bw(struct tcp_rack *rack, uint64_t *bw, int *capped)
2187 {
2188 #ifdef TCP_REQUEST_TRK
2189 	struct timeval tv;
2190 	uint64_t timenow, timeleft, lenleft, lengone, calcbw;
2191 #endif
2192 
2193 	if (rack->r_ctl.bw_rate_cap == 0)
2194 		return;
2195 #ifdef TCP_REQUEST_TRK
2196 	if (rack->rc_catch_up && rack->rc_hybrid_mode &&
2197 	    (rack->r_ctl.rc_last_sft != NULL)) {
2198 		/*
2199 		 * We have a dynamic cap. The original target
2200 		 * is in bw_rate_cap, but we need to look at
2201 		 * how long it is until we hit the deadline.
2202 		 */
2203 		struct tcp_sendfile_track *ent;
2204 
2205       		ent = rack->r_ctl.rc_last_sft;
2206 		microuptime(&tv);
2207 		timenow = tcp_tv_to_lusectick(&tv);
2208 		if (timenow >= ent->deadline) {
2209 			/* No time left we do DGP only */
2210 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2211 					   0, 0, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2212 			rack->r_ctl.bw_rate_cap = 0;
2213 			return;
2214 		}
2215 		/* We have the time */
2216 		timeleft = rack->r_ctl.rc_last_sft->deadline - timenow;
2217 		if (timeleft < HPTS_MSEC_IN_SEC) {
2218 			/* If there is less than a ms left just use DGPs rate */
2219 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2220 					   0, timeleft, 0, HYBRID_LOG_OUTOFTIME, 0, ent, __LINE__);
2221 			rack->r_ctl.bw_rate_cap = 0;
2222 			return;
2223 		}
2224 		/*
2225 		 * Now lets find the amount of data left to send.
2226 		 *
2227 		 * Now ideally we want to use the end_seq to figure out how much more
2228 		 * but it might not be possible (only if we have the TRACK_FG_COMP on the entry..
2229 		 */
2230 		if (ent->flags & TCP_TRK_TRACK_FLG_COMP) {
2231 			if (SEQ_GT(ent->end_seq, rack->rc_tp->snd_una))
2232 				lenleft = ent->end_seq - rack->rc_tp->snd_una;
2233 			else {
2234 				/* TSNH, we should catch it at the send */
2235 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2236 						   0, timeleft, 0, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2237 				rack->r_ctl.bw_rate_cap = 0;
2238 				return;
2239 			}
2240 		} else {
2241 			/*
2242 			 * The hard way, figure out how much is gone and then
2243 			 * take that away from the total the client asked for
2244 			 * (thats off by tls overhead if this is tls).
2245 			 */
2246 			if (SEQ_GT(rack->rc_tp->snd_una, ent->start_seq))
2247 				lengone = rack->rc_tp->snd_una - ent->start_seq;
2248 			else
2249 				lengone = 0;
2250 			if (lengone < (ent->end - ent->start))
2251 				lenleft = (ent->end - ent->start) - lengone;
2252 			else {
2253 				/* TSNH, we should catch it at the send */
2254 				rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2255 						   0, timeleft, lengone, HYBRID_LOG_CAPERROR, 0, ent, __LINE__);
2256 				rack->r_ctl.bw_rate_cap = 0;
2257 				return;
2258 			}
2259 		}
2260 		if (lenleft == 0) {
2261 			/* We have it all sent */
2262 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2263 					   0, timeleft, lenleft, HYBRID_LOG_ALLSENT, 0, ent, __LINE__);
2264 			if (rack->r_ctl.bw_rate_cap)
2265 				goto normal_ratecap;
2266 			else
2267 				return;
2268 		}
2269 		calcbw = lenleft * HPTS_USEC_IN_SEC;
2270 		calcbw /= timeleft;
2271 		/* Now we must compensate for IP/TCP overhead */
2272 		calcbw = rack_compensate_for_linerate(rack, calcbw);
2273 		/* Update the bit rate cap */
2274 		rack->r_ctl.bw_rate_cap = calcbw;
2275 		if ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2276 		    (rack_hybrid_allow_set_maxseg == 1) &&
2277 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2278 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2279 			uint32_t orig_max;
2280 
2281 			orig_max = rack->r_ctl.rc_pace_max_segs;
2282 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2283 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, calcbw, ctf_fixed_maxseg(rack->rc_tp));
2284 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2285 		}
2286 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2287 				   calcbw, timeleft, lenleft, HYBRID_LOG_CAP_CALC, 0, ent, __LINE__);
2288 		if ((calcbw > 0) && (*bw > calcbw)) {
2289 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2290 					   *bw, ent->deadline, lenleft, HYBRID_LOG_RATE_CAP, 0, ent, __LINE__);
2291 			*capped = 1;
2292 			*bw = calcbw;
2293 		}
2294 		return;
2295 	}
2296 normal_ratecap:
2297 #endif
2298 	if ((rack->r_ctl.bw_rate_cap > 0) && (*bw > rack->r_ctl.bw_rate_cap)) {
2299 #ifdef TCP_REQUEST_TRK
2300 		if (rack->rc_hybrid_mode &&
2301 		    rack->rc_catch_up &&
2302 		    (rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_S_MSS) &&
2303 		    (rack_hybrid_allow_set_maxseg == 1) &&
2304 		    ((rack->r_ctl.rc_last_sft->hybrid_flags & TCP_HYBRID_PACING_SETMSS) == 0)) {
2305 			/* Lets set in a smaller mss possibly here to match our rate-cap */
2306 			uint32_t orig_max;
2307 
2308 			orig_max = rack->r_ctl.rc_pace_max_segs;
2309 			rack->r_ctl.rc_last_sft->hybrid_flags |= TCP_HYBRID_PACING_SETMSS;
2310 			rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, rack->r_ctl.bw_rate_cap, ctf_fixed_maxseg(rack->rc_tp));
2311 			rack_log_type_pacing_sizes(rack->rc_tp, rack, rack->r_ctl.client_suggested_maxseg, orig_max, __LINE__, 5);
2312 		}
2313 #endif
2314 		*capped = 1;
2315 		*bw = rack->r_ctl.bw_rate_cap;
2316 		rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
2317 				   *bw, 0, 0,
2318 				   HYBRID_LOG_RATE_CAP, 1, NULL, __LINE__);
2319 	}
2320 }
2321 
2322 static uint64_t
2323 rack_get_gp_est(struct tcp_rack *rack)
2324 {
2325 	uint64_t bw, lt_bw, ret_bw;
2326 
2327 	if (rack->rc_gp_filled == 0) {
2328 		/*
2329 		 * We have yet no b/w measurement,
2330 		 * if we have a user set initial bw
2331 		 * return it. If we don't have that and
2332 		 * we have an srtt, use the tcp IW (10) to
2333 		 * calculate a fictional b/w over the SRTT
2334 		 * which is more or less a guess. Note
2335 		 * we don't use our IW from rack on purpose
2336 		 * so if we have like IW=30, we are not
2337 		 * calculating a "huge" b/w.
2338 		 */
2339 		uint64_t srtt;
2340 
2341 		lt_bw = rack_get_lt_bw(rack);
2342 		if (lt_bw) {
2343 			/*
2344 			 * No goodput bw but a long-term b/w does exist
2345 			 * lets use that.
2346 			 */
2347 			ret_bw = lt_bw;
2348 			goto compensate;
2349 		}
2350 		if (rack->r_ctl.init_rate)
2351 			return (rack->r_ctl.init_rate);
2352 
2353 		/* Ok lets come up with the IW guess, if we have a srtt */
2354 		if (rack->rc_tp->t_srtt == 0) {
2355 			/*
2356 			 * Go with old pacing method
2357 			 * i.e. burst mitigation only.
2358 			 */
2359 			return (0);
2360 		}
2361 		/* Ok lets get the initial TCP win (not racks) */
2362 		bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp));
2363 		srtt = (uint64_t)rack->rc_tp->t_srtt;
2364 		bw *= (uint64_t)USECS_IN_SECOND;
2365 		bw /= srtt;
2366 		ret_bw = bw;
2367 		goto compensate;
2368 
2369 	}
2370 	if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
2371 		/* Averaging is done, we can return the value */
2372 		bw = rack->r_ctl.gp_bw;
2373 	} else {
2374 		/* Still doing initial average must calculate */
2375 		bw = rack->r_ctl.gp_bw / max(rack->r_ctl.num_measurements, 1);
2376 	}
2377 	lt_bw = rack_get_lt_bw(rack);
2378 	if (lt_bw == 0) {
2379 		/* If we don't have one then equate it to the gp_bw */
2380 		lt_bw = rack->r_ctl.gp_bw;
2381 	}
2382 	if ((rack->r_cwnd_was_clamped == 1) && (rack->r_clamped_gets_lower > 0)){
2383 		/*  if clamped take the lowest */
2384 		if (lt_bw < bw)
2385 			ret_bw = lt_bw;
2386 		else
2387 			ret_bw = bw;
2388 	} else {
2389 		/* If not set for clamped to get lowest, take the highest */
2390 		if (lt_bw > bw)
2391 			ret_bw = lt_bw;
2392 		else
2393 			ret_bw = bw;
2394 	}
2395 	/*
2396 	 * Now lets compensate based on the TCP/IP overhead. Our
2397 	 * Goodput estimate does not include this so we must pace out
2398 	 * a bit faster since our pacing calculations do. The pacing
2399 	 * calculations use the base ETHERNET_SEGMENT_SIZE and the segsiz
2400 	 * we are using to do this, so we do that here in the opposite
2401 	 * direction as well. This means that if we are tunneled and the
2402 	 * segsiz is say 1200 bytes we will get quite a boost, but its
2403 	 * compensated for in the pacing time the opposite way.
2404 	 */
2405 compensate:
2406 	ret_bw = rack_compensate_for_linerate(rack, ret_bw);
2407 	return(ret_bw);
2408 }
2409 
2410 
2411 static uint64_t
2412 rack_get_bw(struct tcp_rack *rack)
2413 {
2414 	uint64_t bw;
2415 
2416 	if (rack->use_fixed_rate) {
2417 		/* Return the fixed pacing rate */
2418 		return (rack_get_fixed_pacing_bw(rack));
2419 	}
2420 	bw = rack_get_gp_est(rack);
2421 	return (bw);
2422 }
2423 
2424 static uint16_t
2425 rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm)
2426 {
2427 	if (rack->use_fixed_rate) {
2428 		return (100);
2429 	} else if (rack->in_probe_rtt && (rsm == NULL))
2430 		return (rack->r_ctl.rack_per_of_gp_probertt);
2431 	else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
2432 		  rack->r_ctl.rack_per_of_gp_rec)) {
2433 		if (rsm) {
2434 			/* a retransmission always use the recovery rate */
2435 			return (rack->r_ctl.rack_per_of_gp_rec);
2436 		} else if (rack->rack_rec_nonrxt_use_cr) {
2437 			/* Directed to use the configured rate */
2438 			goto configured_rate;
2439 		} else if (rack->rack_no_prr &&
2440 			   (rack->r_ctl.rack_per_of_gp_rec > 100)) {
2441 			/* No PRR, lets just use the b/w estimate only */
2442 			return (100);
2443 		} else {
2444 			/*
2445 			 * Here we may have a non-retransmit but we
2446 			 * have no overrides, so just use the recovery
2447 			 * rate (prr is in effect).
2448 			 */
2449 			return (rack->r_ctl.rack_per_of_gp_rec);
2450 		}
2451 	}
2452 configured_rate:
2453 	/* For the configured rate we look at our cwnd vs the ssthresh */
2454 	if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh)
2455 		return (rack->r_ctl.rack_per_of_gp_ss);
2456 	else
2457 		return (rack->r_ctl.rack_per_of_gp_ca);
2458 }
2459 
2460 static void
2461 rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6)
2462 {
2463 	/*
2464 	 * Types of logs (mod value)
2465 	 * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit.
2466 	 * 2 = a dsack round begins, persist is reset to 16.
2467 	 * 3 = a dsack round ends
2468 	 * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh
2469 	 * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack
2470 	 * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh.
2471 	 */
2472 	if (tcp_bblogging_on(rack->rc_tp)) {
2473 		union tcp_log_stackspecific log;
2474 		struct timeval tv;
2475 
2476 		memset(&log, 0, sizeof(log));
2477 		log.u_bbr.flex1 = rack->rc_rack_tmr_std_based;
2478 		log.u_bbr.flex1 <<= 1;
2479 		log.u_bbr.flex1 |= rack->rc_rack_use_dsack;
2480 		log.u_bbr.flex1 <<= 1;
2481 		log.u_bbr.flex1 |= rack->rc_dsack_round_seen;
2482 		log.u_bbr.flex2 = rack->r_ctl.dsack_round_end;
2483 		log.u_bbr.flex3 = rack->r_ctl.num_dsack;
2484 		log.u_bbr.flex4 = flex4;
2485 		log.u_bbr.flex5 = flex5;
2486 		log.u_bbr.flex6 = flex6;
2487 		log.u_bbr.flex7 = rack->r_ctl.dsack_persist;
2488 		log.u_bbr.flex8 = mod;
2489 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2490 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2491 		    &rack->rc_inp->inp_socket->so_rcv,
2492 		    &rack->rc_inp->inp_socket->so_snd,
2493 		    RACK_DSACK_HANDLING, 0,
2494 		    0, &log, false, &tv);
2495 	}
2496 }
2497 
2498 static void
2499 rack_log_hdwr_pacing(struct tcp_rack *rack,
2500 		     uint64_t rate, uint64_t hw_rate, int line,
2501 		     int error, uint16_t mod)
2502 {
2503 	if (tcp_bblogging_on(rack->rc_tp)) {
2504 		union tcp_log_stackspecific log;
2505 		struct timeval tv;
2506 		const struct ifnet *ifp;
2507 
2508 		memset(&log, 0, sizeof(log));
2509 		log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff);
2510 		log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff);
2511 		if (rack->r_ctl.crte) {
2512 			ifp = rack->r_ctl.crte->ptbl->rs_ifp;
2513 		} else if (rack->rc_inp->inp_route.ro_nh &&
2514 			   rack->rc_inp->inp_route.ro_nh->nh_ifp) {
2515 			ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp;
2516 		} else
2517 			ifp = NULL;
2518 		if (ifp) {
2519 			log.u_bbr.flex3 = (((uint64_t)ifp  >> 32) & 0x00000000ffffffff);
2520 			log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff);
2521 		}
2522 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2523 		log.u_bbr.bw_inuse = rate;
2524 		log.u_bbr.flex5 = line;
2525 		log.u_bbr.flex6 = error;
2526 		log.u_bbr.flex7 = mod;
2527 		log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs;
2528 		log.u_bbr.flex8 = rack->use_fixed_rate;
2529 		log.u_bbr.flex8 <<= 1;
2530 		log.u_bbr.flex8 |= rack->rack_hdrw_pacing;
2531 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
2532 		log.u_bbr.delRate = rack->r_ctl.crte_prev_rate;
2533 		if (rack->r_ctl.crte)
2534 			log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate;
2535 		else
2536 			log.u_bbr.cur_del_rate = 0;
2537 		log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req;
2538 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2539 		    &rack->rc_inp->inp_socket->so_rcv,
2540 		    &rack->rc_inp->inp_socket->so_snd,
2541 		    BBR_LOG_HDWR_PACE, 0,
2542 		    0, &log, false, &tv);
2543 	}
2544 }
2545 
2546 static uint64_t
2547 rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped)
2548 {
2549 	/*
2550 	 * We allow rack_per_of_gp_xx to dictate our bw rate we want.
2551 	 */
2552 	uint64_t bw_est, high_rate;
2553 	uint64_t gain;
2554 
2555 	if ((rack->r_pacing_discount == 0) ||
2556 	    (rack_full_buffer_discount == 0)) {
2557 		/*
2558 		 * No buffer level based discount from client buffer
2559 		 * level is enabled or the feature is disabled.
2560 		 */
2561 		gain = (uint64_t)rack_get_output_gain(rack, rsm);
2562 		bw_est = bw * gain;
2563 		bw_est /= (uint64_t)100;
2564 	} else {
2565 		/*
2566 		 * We have a discount in place apply it with
2567 		 * just a 100% gain (we get no boost if the buffer
2568 		 * is full).
2569 		 */
2570 		uint64_t discount;
2571 
2572 		discount = bw * (uint64_t)(rack_full_buffer_discount * rack->r_ctl.pacing_discount_amm);
2573 		discount /= 100;
2574 		/* What %% of the b/w do we discount */
2575 		bw_est = bw - discount;
2576 	}
2577 	/* Never fall below the minimum (def 64kbps) */
2578 	if (bw_est < RACK_MIN_BW)
2579 		bw_est = RACK_MIN_BW;
2580 	if (rack->r_rack_hw_rate_caps) {
2581 		/* Rate caps are in place */
2582 		if (rack->r_ctl.crte != NULL) {
2583 			/* We have a hdwr rate already */
2584 			high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
2585 			if (bw_est >= high_rate) {
2586 				/* We are capping bw at the highest rate table entry */
2587 				if (rack_hw_rate_cap_per &&
2588 				    (((high_rate * (100 + rack_hw_rate_cap_per)) / 100) < bw_est)) {
2589 					rack->r_rack_hw_rate_caps = 0;
2590 					goto done;
2591 				}
2592 				rack_log_hdwr_pacing(rack,
2593 						     bw_est, high_rate, __LINE__,
2594 						     0, 3);
2595 				bw_est = high_rate;
2596 				if (capped)
2597 					*capped = 1;
2598 			}
2599 		} else if ((rack->rack_hdrw_pacing == 0) &&
2600 			   (rack->rack_hdw_pace_ena) &&
2601 			   (rack->rack_attempt_hdwr_pace == 0) &&
2602 			   (rack->rc_inp->inp_route.ro_nh != NULL) &&
2603 			   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
2604 			/*
2605 			 * Special case, we have not yet attempted hardware
2606 			 * pacing, and yet we may, when we do, find out if we are
2607 			 * above the highest rate. We need to know the maxbw for the interface
2608 			 * in question (if it supports ratelimiting). We get back
2609 			 * a 0, if the interface is not found in the RL lists.
2610 			 */
2611 			high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
2612 			if (high_rate) {
2613 				/* Yep, we have a rate is it above this rate? */
2614 				if (bw_est > high_rate) {
2615 					bw_est = high_rate;
2616 					if (capped)
2617 						*capped = 1;
2618 				}
2619 			}
2620 		}
2621 	}
2622 done:
2623 	return (bw_est);
2624 }
2625 
2626 static void
2627 rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod)
2628 {
2629 	if (tcp_bblogging_on(rack->rc_tp)) {
2630 		union tcp_log_stackspecific log;
2631 		struct timeval tv;
2632 
2633 		if (rack->sack_attack_disable > 0)
2634 			goto log_anyway;
2635 		if ((mod != 1) && (rack_verbose_logging == 0))  {
2636 			/*
2637 			 * We get 3 values currently for mod
2638 			 * 1 - We are retransmitting and this tells the reason.
2639 			 * 2 - We are clearing a dup-ack count.
2640 			 * 3 - We are incrementing a dup-ack count.
2641 			 *
2642 			 * The clear/increment are only logged
2643 			 * if you have BBverbose on.
2644 			 */
2645 			return;
2646 		}
2647 log_anyway:
2648 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2649 		log.u_bbr.flex1 = tsused;
2650 		log.u_bbr.flex2 = thresh;
2651 		log.u_bbr.flex3 = rsm->r_flags;
2652 		log.u_bbr.flex4 = rsm->r_dupack;
2653 		log.u_bbr.flex5 = rsm->r_start;
2654 		log.u_bbr.flex6 = rsm->r_end;
2655 		log.u_bbr.flex8 = mod;
2656 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2657 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2658 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2659 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2660 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2661 		log.u_bbr.pacing_gain = rack->r_must_retran;
2662 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2663 		    &rack->rc_inp->inp_socket->so_rcv,
2664 		    &rack->rc_inp->inp_socket->so_snd,
2665 		    BBR_LOG_SETTINGS_CHG, 0,
2666 		    0, &log, false, &tv);
2667 	}
2668 }
2669 
2670 static void
2671 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
2672 {
2673 	if (tcp_bblogging_on(rack->rc_tp)) {
2674 		union tcp_log_stackspecific log;
2675 		struct timeval tv;
2676 
2677 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2678 		log.u_bbr.flex1 = rack->rc_tp->t_srtt;
2679 		log.u_bbr.flex2 = to;
2680 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
2681 		log.u_bbr.flex4 = slot;
2682 		log.u_bbr.flex5 = rack->rc_tp->t_hpts_slot;
2683 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2684 		log.u_bbr.flex7 = rack->rc_in_persist;
2685 		log.u_bbr.flex8 = which;
2686 		if (rack->rack_no_prr)
2687 			log.u_bbr.pkts_out = 0;
2688 		else
2689 			log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
2690 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2691 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2692 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2693 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2694 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2695 		log.u_bbr.pacing_gain = rack->r_must_retran;
2696 		log.u_bbr.cwnd_gain = rack->rack_deferred_inited;
2697 		log.u_bbr.pkt_epoch = rack->rc_has_collapsed;
2698 		log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift;
2699 		log.u_bbr.lost = rack_rto_min;
2700 		log.u_bbr.epoch = rack->r_ctl.roundends;
2701 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2702 		    &rack->rc_inp->inp_socket->so_rcv,
2703 		    &rack->rc_inp->inp_socket->so_snd,
2704 		    BBR_LOG_TIMERSTAR, 0,
2705 		    0, &log, false, &tv);
2706 	}
2707 }
2708 
2709 static void
2710 rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm)
2711 {
2712 	if (tcp_bblogging_on(rack->rc_tp)) {
2713 		union tcp_log_stackspecific log;
2714 		struct timeval tv;
2715 
2716 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2717 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2718 		log.u_bbr.flex8 = to_num;
2719 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
2720 		log.u_bbr.flex2 = rack->rc_rack_rtt;
2721 		if (rsm == NULL)
2722 			log.u_bbr.flex3 = 0;
2723 		else
2724 			log.u_bbr.flex3 = rsm->r_end - rsm->r_start;
2725 		if (rack->rack_no_prr)
2726 			log.u_bbr.flex5 = 0;
2727 		else
2728 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
2729 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2730 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2731 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2732 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2733 		log.u_bbr.pacing_gain = rack->r_must_retran;
2734 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2735 		    &rack->rc_inp->inp_socket->so_rcv,
2736 		    &rack->rc_inp->inp_socket->so_snd,
2737 		    BBR_LOG_RTO, 0,
2738 		    0, &log, false, &tv);
2739 	}
2740 }
2741 
2742 static void
2743 rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack,
2744 		 struct rack_sendmap *prev,
2745 		 struct rack_sendmap *rsm,
2746 		 struct rack_sendmap *next,
2747 		 int flag, uint32_t th_ack, int line)
2748 {
2749 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2750 		union tcp_log_stackspecific log;
2751 		struct timeval tv;
2752 
2753 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2754 		log.u_bbr.flex8 = flag;
2755 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2756 		log.u_bbr.cur_del_rate = (uint64_t)prev;
2757 		log.u_bbr.delRate = (uint64_t)rsm;
2758 		log.u_bbr.rttProp = (uint64_t)next;
2759 		log.u_bbr.flex7 = 0;
2760 		if (prev) {
2761 			log.u_bbr.flex1 = prev->r_start;
2762 			log.u_bbr.flex2 = prev->r_end;
2763 			log.u_bbr.flex7 |= 0x4;
2764 		}
2765 		if (rsm) {
2766 			log.u_bbr.flex3 = rsm->r_start;
2767 			log.u_bbr.flex4 = rsm->r_end;
2768 			log.u_bbr.flex7 |= 0x2;
2769 		}
2770 		if (next) {
2771 			log.u_bbr.flex5 = next->r_start;
2772 			log.u_bbr.flex6 = next->r_end;
2773 			log.u_bbr.flex7 |= 0x1;
2774 		}
2775 		log.u_bbr.applimited = line;
2776 		log.u_bbr.pkts_out = th_ack;
2777 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2778 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2779 		if (rack->rack_no_prr)
2780 			log.u_bbr.lost = 0;
2781 		else
2782 			log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt;
2783 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2784 		    &rack->rc_inp->inp_socket->so_rcv,
2785 		    &rack->rc_inp->inp_socket->so_snd,
2786 		    TCP_LOG_MAPCHG, 0,
2787 		    0, &log, false, &tv);
2788 	}
2789 }
2790 
2791 static void
2792 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len,
2793 		 struct rack_sendmap *rsm, int conf)
2794 {
2795 	if (tcp_bblogging_on(tp)) {
2796 		union tcp_log_stackspecific log;
2797 		struct timeval tv;
2798 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2799 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2800 		log.u_bbr.flex1 = t;
2801 		log.u_bbr.flex2 = len;
2802 		log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt;
2803 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
2804 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
2805 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2806 		log.u_bbr.flex7 = conf;
2807 		log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot;
2808 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
2809 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2810 		log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt;
2811 		log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags;
2812 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2813 		if (rsm) {
2814 			log.u_bbr.pkt_epoch = rsm->r_start;
2815 			log.u_bbr.lost = rsm->r_end;
2816 			log.u_bbr.cwnd_gain = rsm->r_rtr_cnt;
2817 			/* We loose any upper of the 24 bits */
2818 			log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags;
2819 		} else {
2820 			/* Its a SYN */
2821 			log.u_bbr.pkt_epoch = rack->rc_tp->iss;
2822 			log.u_bbr.lost = 0;
2823 			log.u_bbr.cwnd_gain = 0;
2824 			log.u_bbr.pacing_gain = 0;
2825 		}
2826 		/* Write out general bits of interest rrs here */
2827 		log.u_bbr.use_lt_bw = rack->rc_highly_buffered;
2828 		log.u_bbr.use_lt_bw <<= 1;
2829 		log.u_bbr.use_lt_bw |= rack->forced_ack;
2830 		log.u_bbr.use_lt_bw <<= 1;
2831 		log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul;
2832 		log.u_bbr.use_lt_bw <<= 1;
2833 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
2834 		log.u_bbr.use_lt_bw <<= 1;
2835 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
2836 		log.u_bbr.use_lt_bw <<= 1;
2837 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
2838 		log.u_bbr.use_lt_bw <<= 1;
2839 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
2840 		log.u_bbr.use_lt_bw <<= 1;
2841 		log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom;
2842 		log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight;
2843 		log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts;
2844 		log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered;
2845 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts;
2846 		log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt;
2847 		log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
2848 		log.u_bbr.bw_inuse <<= 32;
2849 		if (rsm)
2850 			log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]);
2851 		TCP_LOG_EVENTP(tp, NULL,
2852 		    &rack->rc_inp->inp_socket->so_rcv,
2853 		    &rack->rc_inp->inp_socket->so_snd,
2854 		    BBR_LOG_BBRRTT, 0,
2855 		    0, &log, false, &tv);
2856 
2857 
2858 	}
2859 }
2860 
2861 static void
2862 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
2863 {
2864 	/*
2865 	 * Log the rtt sample we are
2866 	 * applying to the srtt algorithm in
2867 	 * useconds.
2868 	 */
2869 	if (tcp_bblogging_on(rack->rc_tp)) {
2870 		union tcp_log_stackspecific log;
2871 		struct timeval tv;
2872 
2873 		/* Convert our ms to a microsecond */
2874 		memset(&log, 0, sizeof(log));
2875 		log.u_bbr.flex1 = rtt;
2876 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
2877 		log.u_bbr.flex3 = rack->r_ctl.sack_count;
2878 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
2879 		log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra;
2880 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
2881 		log.u_bbr.flex7 = 1;
2882 		log.u_bbr.flex8 = rack->sack_attack_disable;
2883 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2884 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2885 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2886 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2887 		log.u_bbr.pacing_gain = rack->r_must_retran;
2888 		/*
2889 		 * We capture in delRate the upper 32 bits as
2890 		 * the confidence level we had declared, and the
2891 		 * lower 32 bits as the actual RTT using the arrival
2892 		 * timestamp.
2893 		 */
2894 		log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence;
2895 		log.u_bbr.delRate <<= 32;
2896 		log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt;
2897 		/* Lets capture all the things that make up t_rtxcur */
2898 		log.u_bbr.applimited = rack_rto_min;
2899 		log.u_bbr.epoch = rack_rto_max;
2900 		log.u_bbr.lt_epoch = rack->r_ctl.timer_slop;
2901 		log.u_bbr.lost = rack_rto_min;
2902 		log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop);
2903 		log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp);
2904 		log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec;
2905 		log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC;
2906 		log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec;
2907 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2908 		    &rack->rc_inp->inp_socket->so_rcv,
2909 		    &rack->rc_inp->inp_socket->so_snd,
2910 		    TCP_LOG_RTT, 0,
2911 		    0, &log, false, &tv);
2912 	}
2913 }
2914 
2915 static void
2916 rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where)
2917 {
2918 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2919 		union tcp_log_stackspecific log;
2920 		struct timeval tv;
2921 
2922 		/* Convert our ms to a microsecond */
2923 		memset(&log, 0, sizeof(log));
2924 		log.u_bbr.flex1 = rtt;
2925 		log.u_bbr.flex2 = send_time;
2926 		log.u_bbr.flex3 = ack_time;
2927 		log.u_bbr.flex4 = where;
2928 		log.u_bbr.flex7 = 2;
2929 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2930 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2931 		    &rack->rc_inp->inp_socket->so_rcv,
2932 		    &rack->rc_inp->inp_socket->so_snd,
2933 		    TCP_LOG_RTT, 0,
2934 		    0, &log, false, &tv);
2935 	}
2936 }
2937 
2938 
2939 static void
2940 rack_log_rtt_sendmap(struct tcp_rack *rack, uint32_t idx, uint64_t tsv, uint32_t tsecho)
2941 {
2942 	if (tcp_bblogging_on(rack->rc_tp)) {
2943 		union tcp_log_stackspecific log;
2944 		struct timeval tv;
2945 
2946 		/* Convert our ms to a microsecond */
2947 		memset(&log, 0, sizeof(log));
2948 		log.u_bbr.flex1 = idx;
2949 		log.u_bbr.flex2 = rack_ts_to_msec(tsv);
2950 		log.u_bbr.flex3 = tsecho;
2951 		log.u_bbr.flex7 = 3;
2952 		log.u_bbr.rttProp = tsv;
2953 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2954 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
2955 		    &rack->rc_inp->inp_socket->so_rcv,
2956 		    &rack->rc_inp->inp_socket->so_snd,
2957 		    TCP_LOG_RTT, 0,
2958 		    0, &log, false, &tv);
2959 	}
2960 }
2961 
2962 
2963 static inline void
2964 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
2965 {
2966 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2967 		union tcp_log_stackspecific log;
2968 		struct timeval tv;
2969 
2970 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2971 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2972 		log.u_bbr.flex1 = line;
2973 		log.u_bbr.flex2 = tick;
2974 		log.u_bbr.flex3 = tp->t_maxunacktime;
2975 		log.u_bbr.flex4 = tp->t_acktime;
2976 		log.u_bbr.flex8 = event;
2977 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
2978 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
2979 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
2980 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
2981 		log.u_bbr.pacing_gain = rack->r_must_retran;
2982 		TCP_LOG_EVENTP(tp, NULL,
2983 		    &rack->rc_inp->inp_socket->so_rcv,
2984 		    &rack->rc_inp->inp_socket->so_snd,
2985 		    BBR_LOG_PROGRESS, 0,
2986 		    0, &log, false, &tv);
2987 	}
2988 }
2989 
2990 static void
2991 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv, int line)
2992 {
2993 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
2994 		union tcp_log_stackspecific log;
2995 
2996 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
2997 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
2998 		log.u_bbr.flex1 = slot;
2999 		if (rack->rack_no_prr)
3000 			log.u_bbr.flex2 = 0;
3001 		else
3002 			log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt;
3003 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3004 		log.u_bbr.flex5 = rack->r_ctl.ack_during_sd;
3005 		log.u_bbr.flex6 = line;
3006 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
3007 		log.u_bbr.flex8 = rack->rc_in_persist;
3008 		log.u_bbr.timeStamp = cts;
3009 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3010 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3011 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3012 		log.u_bbr.pacing_gain = rack->r_must_retran;
3013 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3014 		    &rack->rc_inp->inp_socket->so_rcv,
3015 		    &rack->rc_inp->inp_socket->so_snd,
3016 		    BBR_LOG_BBRSND, 0,
3017 		    0, &log, false, tv);
3018 	}
3019 }
3020 
3021 static void
3022 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs)
3023 {
3024 	if (tcp_bblogging_on(rack->rc_tp)) {
3025 		union tcp_log_stackspecific log;
3026 		struct timeval tv;
3027 
3028 		memset(&log, 0, sizeof(log));
3029 		log.u_bbr.flex1 = did_out;
3030 		log.u_bbr.flex2 = nxt_pkt;
3031 		log.u_bbr.flex3 = way_out;
3032 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3033 		if (rack->rack_no_prr)
3034 			log.u_bbr.flex5 = 0;
3035 		else
3036 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3037 		log.u_bbr.flex6 = nsegs;
3038 		log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs;
3039 		log.u_bbr.flex7 = rack->rc_ack_can_sendout_data;	/* Do we have ack-can-send set */
3040 		log.u_bbr.flex7 <<= 1;
3041 		log.u_bbr.flex7 |= rack->r_fast_output;	/* is fast output primed */
3042 		log.u_bbr.flex7 <<= 1;
3043 		log.u_bbr.flex7 |= rack->r_wanted_output;	/* Do we want output */
3044 		log.u_bbr.flex8 = rack->rc_in_persist;
3045 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3046 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3047 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3048 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3049 		log.u_bbr.use_lt_bw <<= 1;
3050 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3051 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3052 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3053 		log.u_bbr.pacing_gain = rack->r_must_retran;
3054 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3055 		    &rack->rc_inp->inp_socket->so_rcv,
3056 		    &rack->rc_inp->inp_socket->so_snd,
3057 		    BBR_LOG_DOSEG_DONE, 0,
3058 		    0, &log, false, &tv);
3059 	}
3060 }
3061 
3062 static void
3063 rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm)
3064 {
3065 	if (tcp_bblogging_on(rack->rc_tp)) {
3066 		union tcp_log_stackspecific log;
3067 		struct timeval tv;
3068 
3069 		memset(&log, 0, sizeof(log));
3070 		log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs;
3071 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
3072 		log.u_bbr.flex4 = arg1;
3073 		log.u_bbr.flex5 = arg2;
3074 		log.u_bbr.flex7 = rack->r_ctl.rc_user_set_min_segs;
3075 		log.u_bbr.flex6 = arg3;
3076 		log.u_bbr.flex8 = frm;
3077 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3078 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3079 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3080 		log.u_bbr.applimited = rack->r_ctl.rc_sacked;
3081 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3082 		log.u_bbr.pacing_gain = rack->r_must_retran;
3083 		TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
3084 		    &tptosocket(tp)->so_snd,
3085 		    TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv);
3086 	}
3087 }
3088 
3089 static void
3090 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot,
3091 			  uint8_t hpts_calling, int reason, uint32_t cwnd_to_use)
3092 {
3093 	if (tcp_bblogging_on(rack->rc_tp)) {
3094 		union tcp_log_stackspecific log;
3095 		struct timeval tv;
3096 
3097 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3098 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3099 		log.u_bbr.flex1 = slot;
3100 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
3101 		log.u_bbr.flex4 = reason;
3102 		if (rack->rack_no_prr)
3103 			log.u_bbr.flex5 = 0;
3104 		else
3105 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3106 		log.u_bbr.flex7 = hpts_calling;
3107 		log.u_bbr.flex8 = rack->rc_in_persist;
3108 		log.u_bbr.lt_epoch = cwnd_to_use;
3109 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3110 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3111 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3112 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3113 		log.u_bbr.pacing_gain = rack->r_must_retran;
3114 		log.u_bbr.cwnd_gain = rack->rc_has_collapsed;
3115 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3116 		    &rack->rc_inp->inp_socket->so_rcv,
3117 		    &rack->rc_inp->inp_socket->so_snd,
3118 		    BBR_LOG_JUSTRET, 0,
3119 		    tlen, &log, false, &tv);
3120 	}
3121 }
3122 
3123 static void
3124 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts,
3125 		   struct timeval *tv, uint32_t flags_on_entry)
3126 {
3127 	if (tcp_bblogging_on(rack->rc_tp)) {
3128 		union tcp_log_stackspecific log;
3129 
3130 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3131 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
3132 		log.u_bbr.flex1 = line;
3133 		log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to;
3134 		log.u_bbr.flex3 = flags_on_entry;
3135 		log.u_bbr.flex4 = us_cts;
3136 		if (rack->rack_no_prr)
3137 			log.u_bbr.flex5 = 0;
3138 		else
3139 			log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt;
3140 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
3141 		log.u_bbr.flex7 = hpts_removed;
3142 		log.u_bbr.flex8 = 1;
3143 		log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags;
3144 		log.u_bbr.timeStamp = us_cts;
3145 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3146 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3147 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3148 		log.u_bbr.pacing_gain = rack->r_must_retran;
3149 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3150 		    &rack->rc_inp->inp_socket->so_rcv,
3151 		    &rack->rc_inp->inp_socket->so_snd,
3152 		    BBR_LOG_TIMERCANC, 0,
3153 		    0, &log, false, tv);
3154 	}
3155 }
3156 
3157 static void
3158 rack_log_alt_to_to_cancel(struct tcp_rack *rack,
3159 			  uint32_t flex1, uint32_t flex2,
3160 			  uint32_t flex3, uint32_t flex4,
3161 			  uint32_t flex5, uint32_t flex6,
3162 			  uint16_t flex7, uint8_t mod)
3163 {
3164 	if (tcp_bblogging_on(rack->rc_tp)) {
3165 		union tcp_log_stackspecific log;
3166 		struct timeval tv;
3167 
3168 		if (mod == 1) {
3169 			/* No you can't use 1, its for the real to cancel */
3170 			return;
3171 		}
3172 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3173 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3174 		log.u_bbr.flex1 = flex1;
3175 		log.u_bbr.flex2 = flex2;
3176 		log.u_bbr.flex3 = flex3;
3177 		log.u_bbr.flex4 = flex4;
3178 		log.u_bbr.flex5 = flex5;
3179 		log.u_bbr.flex6 = flex6;
3180 		log.u_bbr.flex7 = flex7;
3181 		log.u_bbr.flex8 = mod;
3182 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3183 		    &rack->rc_inp->inp_socket->so_rcv,
3184 		    &rack->rc_inp->inp_socket->so_snd,
3185 		    BBR_LOG_TIMERCANC, 0,
3186 		    0, &log, false, &tv);
3187 	}
3188 }
3189 
3190 static void
3191 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
3192 {
3193 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
3194 		union tcp_log_stackspecific log;
3195 		struct timeval tv;
3196 
3197 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3198 		log.u_bbr.flex1 = timers;
3199 		log.u_bbr.flex2 = ret;
3200 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
3201 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
3202 		log.u_bbr.flex5 = cts;
3203 		if (rack->rack_no_prr)
3204 			log.u_bbr.flex6 = 0;
3205 		else
3206 			log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt;
3207 		log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto;
3208 		log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto;
3209 		log.u_bbr.pacing_gain = rack->r_must_retran;
3210 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3211 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3212 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3213 		    &rack->rc_inp->inp_socket->so_rcv,
3214 		    &rack->rc_inp->inp_socket->so_snd,
3215 		    BBR_LOG_TO_PROCESS, 0,
3216 		    0, &log, false, &tv);
3217 	}
3218 }
3219 
3220 static void
3221 rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line)
3222 {
3223 	if (tcp_bblogging_on(rack->rc_tp)) {
3224 		union tcp_log_stackspecific log;
3225 		struct timeval tv;
3226 
3227 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3228 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_out;
3229 		log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs;
3230 		if (rack->rack_no_prr)
3231 			log.u_bbr.flex3 = 0;
3232 		else
3233 			log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt;
3234 		log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered;
3235 		log.u_bbr.flex5 = rack->r_ctl.rc_sacked;
3236 		log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt;
3237 		log.u_bbr.flex7 = line;
3238 		log.u_bbr.flex8 = frm;
3239 		log.u_bbr.pkts_out = orig_cwnd;
3240 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3241 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3242 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
3243 		log.u_bbr.use_lt_bw <<= 1;
3244 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
3245 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3246 		    &rack->rc_inp->inp_socket->so_rcv,
3247 		    &rack->rc_inp->inp_socket->so_snd,
3248 		    BBR_LOG_BBRUPD, 0,
3249 		    0, &log, false, &tv);
3250 	}
3251 }
3252 
3253 #ifdef TCP_SAD_DETECTION
3254 static void
3255 rack_log_sad(struct tcp_rack *rack, int event)
3256 {
3257 	if (tcp_bblogging_on(rack->rc_tp)) {
3258 		union tcp_log_stackspecific log;
3259 		struct timeval tv;
3260 
3261 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
3262 		log.u_bbr.flex1 = rack->r_ctl.sack_count;
3263 		log.u_bbr.flex2 = rack->r_ctl.ack_count;
3264 		log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra;
3265 		log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move;
3266 		log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced;
3267 		log.u_bbr.flex6 = tcp_sack_to_ack_thresh;
3268 		log.u_bbr.pkts_out = tcp_sack_to_move_thresh;
3269 		log.u_bbr.lt_epoch = (tcp_force_detection << 8);
3270 		log.u_bbr.lt_epoch |= rack->do_detection;
3271 		log.u_bbr.applimited = tcp_map_minimum;
3272 		log.u_bbr.flex7 = rack->sack_attack_disable;
3273 		log.u_bbr.flex8 = event;
3274 		log.u_bbr.bbr_state = rack->rc_suspicious;
3275 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3276 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3277 		log.u_bbr.delivered = tcp_sad_decay_val;
3278 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3279 		    &rack->rc_inp->inp_socket->so_rcv,
3280 		    &rack->rc_inp->inp_socket->so_snd,
3281 		    TCP_SAD_DETECT, 0,
3282 		    0, &log, false, &tv);
3283 	}
3284 }
3285 #endif
3286 
3287 static void
3288 rack_counter_destroy(void)
3289 {
3290 	counter_u64_free(rack_total_bytes);
3291 	counter_u64_free(rack_fto_send);
3292 	counter_u64_free(rack_fto_rsm_send);
3293 	counter_u64_free(rack_nfto_resend);
3294 	counter_u64_free(rack_hw_pace_init_fail);
3295 	counter_u64_free(rack_hw_pace_lost);
3296 	counter_u64_free(rack_non_fto_send);
3297 	counter_u64_free(rack_extended_rfo);
3298 	counter_u64_free(rack_ack_total);
3299 	counter_u64_free(rack_express_sack);
3300 	counter_u64_free(rack_sack_total);
3301 	counter_u64_free(rack_move_none);
3302 	counter_u64_free(rack_move_some);
3303 	counter_u64_free(rack_sack_attacks_detected);
3304 	counter_u64_free(rack_sack_attacks_reversed);
3305 	counter_u64_free(rack_sack_attacks_suspect);
3306 	counter_u64_free(rack_sack_used_next_merge);
3307 	counter_u64_free(rack_sack_used_prev_merge);
3308 	counter_u64_free(rack_tlp_tot);
3309 	counter_u64_free(rack_tlp_newdata);
3310 	counter_u64_free(rack_tlp_retran);
3311 	counter_u64_free(rack_tlp_retran_bytes);
3312 	counter_u64_free(rack_to_tot);
3313 	counter_u64_free(rack_saw_enobuf);
3314 	counter_u64_free(rack_saw_enobuf_hw);
3315 	counter_u64_free(rack_saw_enetunreach);
3316 	counter_u64_free(rack_hot_alloc);
3317 	counter_u64_free(rack_to_alloc);
3318 	counter_u64_free(rack_to_alloc_hard);
3319 	counter_u64_free(rack_to_alloc_emerg);
3320 	counter_u64_free(rack_to_alloc_limited);
3321 	counter_u64_free(rack_alloc_limited_conns);
3322 	counter_u64_free(rack_split_limited);
3323 	counter_u64_free(rack_multi_single_eq);
3324 	counter_u64_free(rack_rxt_clamps_cwnd);
3325 	counter_u64_free(rack_rxt_clamps_cwnd_uniq);
3326 	counter_u64_free(rack_proc_non_comp_ack);
3327 	counter_u64_free(rack_sack_proc_all);
3328 	counter_u64_free(rack_sack_proc_restart);
3329 	counter_u64_free(rack_sack_proc_short);
3330 	counter_u64_free(rack_sack_skipped_acked);
3331 	counter_u64_free(rack_sack_splits);
3332 	counter_u64_free(rack_input_idle_reduces);
3333 	counter_u64_free(rack_collapsed_win);
3334 	counter_u64_free(rack_collapsed_win_rxt);
3335 	counter_u64_free(rack_collapsed_win_rxt_bytes);
3336 	counter_u64_free(rack_collapsed_win_seen);
3337 	counter_u64_free(rack_try_scwnd);
3338 	counter_u64_free(rack_persists_sends);
3339 	counter_u64_free(rack_persists_acks);
3340 	counter_u64_free(rack_persists_loss);
3341 	counter_u64_free(rack_persists_lost_ends);
3342 #ifdef INVARIANTS
3343 	counter_u64_free(rack_adjust_map_bw);
3344 #endif
3345 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
3346 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
3347 }
3348 
3349 static struct rack_sendmap *
3350 rack_alloc(struct tcp_rack *rack)
3351 {
3352 	struct rack_sendmap *rsm;
3353 
3354 	/*
3355 	 * First get the top of the list it in
3356 	 * theory is the "hottest" rsm we have,
3357 	 * possibly just freed by ack processing.
3358 	 */
3359 	if (rack->rc_free_cnt > rack_free_cache) {
3360 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3361 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3362 		counter_u64_add(rack_hot_alloc, 1);
3363 		rack->rc_free_cnt--;
3364 		return (rsm);
3365 	}
3366 	/*
3367 	 * Once we get under our free cache we probably
3368 	 * no longer have a "hot" one available. Lets
3369 	 * get one from UMA.
3370 	 */
3371 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
3372 	if (rsm) {
3373 		rack->r_ctl.rc_num_maps_alloced++;
3374 		counter_u64_add(rack_to_alloc, 1);
3375 		return (rsm);
3376 	}
3377 	/*
3378 	 * Dig in to our aux rsm's (the last two) since
3379 	 * UMA failed to get us one.
3380 	 */
3381 	if (rack->rc_free_cnt) {
3382 		counter_u64_add(rack_to_alloc_emerg, 1);
3383 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
3384 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3385 		rack->rc_free_cnt--;
3386 		return (rsm);
3387 	}
3388 	return (NULL);
3389 }
3390 
3391 static struct rack_sendmap *
3392 rack_alloc_full_limit(struct tcp_rack *rack)
3393 {
3394 	if ((V_tcp_map_entries_limit > 0) &&
3395 	    (rack->do_detection == 0) &&
3396 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
3397 		counter_u64_add(rack_to_alloc_limited, 1);
3398 		if (!rack->alloc_limit_reported) {
3399 			rack->alloc_limit_reported = 1;
3400 			counter_u64_add(rack_alloc_limited_conns, 1);
3401 		}
3402 		return (NULL);
3403 	}
3404 	return (rack_alloc(rack));
3405 }
3406 
3407 /* wrapper to allocate a sendmap entry, subject to a specific limit */
3408 static struct rack_sendmap *
3409 rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type)
3410 {
3411 	struct rack_sendmap *rsm;
3412 
3413 	if (limit_type) {
3414 		/* currently there is only one limit type */
3415 		if (rack->r_ctl.rc_split_limit > 0 &&
3416 		    (rack->do_detection == 0) &&
3417 		    rack->r_ctl.rc_num_split_allocs >= rack->r_ctl.rc_split_limit) {
3418 			counter_u64_add(rack_split_limited, 1);
3419 			if (!rack->alloc_limit_reported) {
3420 				rack->alloc_limit_reported = 1;
3421 				counter_u64_add(rack_alloc_limited_conns, 1);
3422 			}
3423 			return (NULL);
3424 #ifdef TCP_SAD_DETECTION
3425 		} else if ((tcp_sad_limit != 0) &&
3426 			   (rack->do_detection == 1) &&
3427 			   (rack->r_ctl.rc_num_split_allocs >= tcp_sad_limit)) {
3428 			counter_u64_add(rack_split_limited, 1);
3429 			if (!rack->alloc_limit_reported) {
3430 				rack->alloc_limit_reported = 1;
3431 				counter_u64_add(rack_alloc_limited_conns, 1);
3432 			}
3433 			return (NULL);
3434 #endif
3435 		}
3436 	}
3437 
3438 	/* allocate and mark in the limit type, if set */
3439 	rsm = rack_alloc(rack);
3440 	if (rsm != NULL && limit_type) {
3441 		rsm->r_limit_type = limit_type;
3442 		rack->r_ctl.rc_num_split_allocs++;
3443 	}
3444 	return (rsm);
3445 }
3446 
3447 static void
3448 rack_free_trim(struct tcp_rack *rack)
3449 {
3450 	struct rack_sendmap *rsm;
3451 
3452 	/*
3453 	 * Free up all the tail entries until
3454 	 * we get our list down to the limit.
3455 	 */
3456 	while (rack->rc_free_cnt > rack_free_cache) {
3457 		rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head);
3458 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
3459 		rack->rc_free_cnt--;
3460 		rack->r_ctl.rc_num_maps_alloced--;
3461 		uma_zfree(rack_zone, rsm);
3462 	}
3463 }
3464 
3465 static void
3466 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
3467 {
3468 	if (rsm->r_flags & RACK_APP_LIMITED) {
3469 		if (rack->r_ctl.rc_app_limited_cnt > 0) {
3470 			rack->r_ctl.rc_app_limited_cnt--;
3471 		}
3472 	}
3473 	if (rsm->r_limit_type) {
3474 		/* currently there is only one limit type */
3475 		rack->r_ctl.rc_num_split_allocs--;
3476 	}
3477 	if (rsm == rack->r_ctl.rc_first_appl) {
3478 		if (rack->r_ctl.rc_app_limited_cnt == 0)
3479 			rack->r_ctl.rc_first_appl = NULL;
3480 		else
3481 			rack->r_ctl.rc_first_appl = tqhash_find(rack->r_ctl.tqh, rsm->r_nseq_appl);
3482 	}
3483 	if (rsm == rack->r_ctl.rc_resend)
3484 		rack->r_ctl.rc_resend = NULL;
3485 	if (rsm == rack->r_ctl.rc_end_appl)
3486 		rack->r_ctl.rc_end_appl = NULL;
3487 	if (rack->r_ctl.rc_tlpsend == rsm)
3488 		rack->r_ctl.rc_tlpsend = NULL;
3489 	if (rack->r_ctl.rc_sacklast == rsm)
3490 		rack->r_ctl.rc_sacklast = NULL;
3491 	memset(rsm, 0, sizeof(struct rack_sendmap));
3492 	/* Make sure we are not going to overrun our count limit of 0xff */
3493 	if ((rack->rc_free_cnt + 1) > 0xff) {
3494 		rack_free_trim(rack);
3495 	}
3496 	TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext);
3497 	rack->rc_free_cnt++;
3498 }
3499 
3500 static uint32_t
3501 rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack)
3502 {
3503 	uint64_t srtt, bw, len, tim;
3504 	uint32_t segsiz, def_len, minl;
3505 
3506 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3507 	def_len = rack_def_data_window * segsiz;
3508 	if (rack->rc_gp_filled == 0) {
3509 		/*
3510 		 * We have no measurement (IW is in flight?) so
3511 		 * we can only guess using our data_window sysctl
3512 		 * value (usually 20MSS).
3513 		 */
3514 		return (def_len);
3515 	}
3516 	/*
3517 	 * Now we have a number of factors to consider.
3518 	 *
3519 	 * 1) We have a desired BDP which is usually
3520 	 *    at least 2.
3521 	 * 2) We have a minimum number of rtt's usually 1 SRTT
3522 	 *    but we allow it too to be more.
3523 	 * 3) We want to make sure a measurement last N useconds (if
3524 	 *    we have set rack_min_measure_usec.
3525 	 *
3526 	 * We handle the first concern here by trying to create a data
3527 	 * window of max(rack_def_data_window, DesiredBDP). The
3528 	 * second concern we handle in not letting the measurement
3529 	 * window end normally until at least the required SRTT's
3530 	 * have gone by which is done further below in
3531 	 * rack_enough_for_measurement(). Finally the third concern
3532 	 * we also handle here by calculating how long that time
3533 	 * would take at the current BW and then return the
3534 	 * max of our first calculation and that length. Note
3535 	 * that if rack_min_measure_usec is 0, we don't deal
3536 	 * with concern 3. Also for both Concern 1 and 3 an
3537 	 * application limited period could end the measurement
3538 	 * earlier.
3539 	 *
3540 	 * So lets calculate the BDP with the "known" b/w using
3541 	 * the SRTT has our rtt and then multiply it by the
3542 	 * goal.
3543 	 */
3544 	bw = rack_get_bw(rack);
3545 	srtt = (uint64_t)tp->t_srtt;
3546 	len = bw * srtt;
3547 	len /= (uint64_t)HPTS_USEC_IN_SEC;
3548 	len *= max(1, rack_goal_bdp);
3549 	/* Now we need to round up to the nearest MSS */
3550 	len = roundup(len, segsiz);
3551 	if (rack_min_measure_usec) {
3552 		/* Now calculate our min length for this b/w */
3553 		tim = rack_min_measure_usec;
3554 		minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC;
3555 		if (minl == 0)
3556 			minl = 1;
3557 		minl = roundup(minl, segsiz);
3558 		if (len < minl)
3559 			len = minl;
3560 	}
3561 	/*
3562 	 * Now if we have a very small window we want
3563 	 * to attempt to get the window that is
3564 	 * as small as possible. This happens on
3565 	 * low b/w connections and we don't want to
3566 	 * span huge numbers of rtt's between measurements.
3567 	 *
3568 	 * We basically include 2 over our "MIN window" so
3569 	 * that the measurement can be shortened (possibly) by
3570 	 * an ack'ed packet.
3571 	 */
3572 	if (len < def_len)
3573 		return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz)));
3574 	else
3575 		return (max((uint32_t)len, def_len));
3576 
3577 }
3578 
3579 static int
3580 rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality)
3581 {
3582 	uint32_t tim, srtts, segsiz;
3583 
3584 	/*
3585 	 * Has enough time passed for the GP measurement to be valid?
3586 	 */
3587 	if (SEQ_LT(th_ack, tp->gput_seq)) {
3588 		/* Not enough bytes yet */
3589 		return (0);
3590 	}
3591 	if ((tp->snd_max == tp->snd_una) ||
3592 	    (th_ack == tp->snd_max)){
3593 		/*
3594 		 * All is acked quality of all acked is
3595 		 * usually low or medium, but we in theory could split
3596 		 * all acked into two cases, where you got
3597 		 * a signifigant amount of your window and
3598 		 * where you did not. For now we leave it
3599 		 * but it is something to contemplate in the
3600 		 * future. The danger here is that delayed ack
3601 		 * is effecting the last byte (which is a 50:50 chance).
3602 		 */
3603 		*quality = RACK_QUALITY_ALLACKED;
3604 		return (1);
3605 	}
3606 	if (SEQ_GEQ(th_ack,  tp->gput_ack)) {
3607 		/*
3608 		 * We obtained our entire window of data we wanted
3609 		 * no matter if we are in recovery or not then
3610 		 * its ok since expanding the window does not
3611 		 * make things fuzzy (or at least not as much).
3612 		 */
3613 		*quality = RACK_QUALITY_HIGH;
3614 		return (1);
3615 	}
3616 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
3617 	if (SEQ_LT(th_ack, tp->gput_ack) &&
3618 	    ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
3619 		/* Not enough bytes yet */
3620 		return (0);
3621 	}
3622 	if (rack->r_ctl.rc_first_appl &&
3623 	    (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) {
3624 		/*
3625 		 * We are up to the app limited send point
3626 		 * we have to measure irrespective of the time..
3627 		 */
3628 		*quality = RACK_QUALITY_APPLIMITED;
3629 		return (1);
3630 	}
3631 	/* Now what about time? */
3632 	srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts);
3633 	tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts;
3634 	if ((tim >= srtts) && (IN_RECOVERY(rack->rc_tp->t_flags) == 0)) {
3635 		/*
3636 		 * We do not allow a measurement if we are in recovery
3637 		 * that would shrink the goodput window we wanted.
3638 		 * This is to prevent cloudyness of when the last send
3639 		 * was actually made.
3640 		 */
3641 		*quality = RACK_QUALITY_HIGH;
3642 		return (1);
3643 	}
3644 	/* Nope not even a full SRTT has passed */
3645 	return (0);
3646 }
3647 
3648 static void
3649 rack_log_timely(struct tcp_rack *rack,
3650 		uint32_t logged, uint64_t cur_bw, uint64_t low_bnd,
3651 		uint64_t up_bnd, int line, uint8_t method)
3652 {
3653 	if (tcp_bblogging_on(rack->rc_tp)) {
3654 		union tcp_log_stackspecific log;
3655 		struct timeval tv;
3656 
3657 		memset(&log, 0, sizeof(log));
3658 		log.u_bbr.flex1 = logged;
3659 		log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt;
3660 		log.u_bbr.flex2 <<= 4;
3661 		log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt;
3662 		log.u_bbr.flex2 <<= 4;
3663 		log.u_bbr.flex2 |= rack->rc_gp_incr;
3664 		log.u_bbr.flex2 <<= 4;
3665 		log.u_bbr.flex2 |= rack->rc_gp_bwred;
3666 		log.u_bbr.flex3 = rack->rc_gp_incr;
3667 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
3668 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca;
3669 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec;
3670 		log.u_bbr.flex7 = rack->rc_gp_bwred;
3671 		log.u_bbr.flex8 = method;
3672 		log.u_bbr.cur_del_rate = cur_bw;
3673 		log.u_bbr.delRate = low_bnd;
3674 		log.u_bbr.bw_inuse = up_bnd;
3675 		log.u_bbr.rttProp = rack_get_bw(rack);
3676 		log.u_bbr.pkt_epoch = line;
3677 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
3678 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
3679 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
3680 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
3681 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
3682 		log.u_bbr.cwnd_gain = rack->rc_dragged_bottom;
3683 		log.u_bbr.cwnd_gain <<= 1;
3684 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec;
3685 		log.u_bbr.cwnd_gain <<= 1;
3686 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
3687 		log.u_bbr.cwnd_gain <<= 1;
3688 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
3689 		log.u_bbr.lost = rack->r_ctl.rc_loss_count;
3690 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
3691 		    &rack->rc_inp->inp_socket->so_rcv,
3692 		    &rack->rc_inp->inp_socket->so_snd,
3693 		    TCP_TIMELY_WORK, 0,
3694 		    0, &log, false, &tv);
3695 	}
3696 }
3697 
3698 static int
3699 rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult)
3700 {
3701 	/*
3702 	 * Before we increase we need to know if
3703 	 * the estimate just made was less than
3704 	 * our pacing goal (i.e. (cur_bw * mult) > last_bw_est)
3705 	 *
3706 	 * If we already are pacing at a fast enough
3707 	 * rate to push us faster there is no sense of
3708 	 * increasing.
3709 	 *
3710 	 * We first caculate our actual pacing rate (ss or ca multiplier
3711 	 * times our cur_bw).
3712 	 *
3713 	 * Then we take the last measured rate and multipy by our
3714 	 * maximum pacing overage to give us a max allowable rate.
3715 	 *
3716 	 * If our act_rate is smaller than our max_allowable rate
3717 	 * then we should increase. Else we should hold steady.
3718 	 *
3719 	 */
3720 	uint64_t act_rate, max_allow_rate;
3721 
3722 	if (rack_timely_no_stopping)
3723 		return (1);
3724 
3725 	if ((cur_bw == 0) || (last_bw_est == 0)) {
3726 		/*
3727 		 * Initial startup case or
3728 		 * everything is acked case.
3729 		 */
3730 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3731 				__LINE__, 9);
3732 		return (1);
3733 	}
3734 	if (mult <= 100) {
3735 		/*
3736 		 * We can always pace at or slightly above our rate.
3737 		 */
3738 		rack_log_timely(rack,  mult, cur_bw, 0, 0,
3739 				__LINE__, 9);
3740 		return (1);
3741 	}
3742 	act_rate = cur_bw * (uint64_t)mult;
3743 	act_rate /= 100;
3744 	max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100);
3745 	max_allow_rate /= 100;
3746 	if (act_rate < max_allow_rate) {
3747 		/*
3748 		 * Here the rate we are actually pacing at
3749 		 * is smaller than 10% above our last measurement.
3750 		 * This means we are pacing below what we would
3751 		 * like to try to achieve (plus some wiggle room).
3752 		 */
3753 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3754 				__LINE__, 9);
3755 		return (1);
3756 	} else {
3757 		/*
3758 		 * Here we are already pacing at least rack_max_per_above(10%)
3759 		 * what we are getting back. This indicates most likely
3760 		 * that we are being limited (cwnd/rwnd/app) and can't
3761 		 * get any more b/w. There is no sense of trying to
3762 		 * raise up the pacing rate its not speeding us up
3763 		 * and we already are pacing faster than we are getting.
3764 		 */
3765 		rack_log_timely(rack,  mult, cur_bw, act_rate, max_allow_rate,
3766 				__LINE__, 8);
3767 		return (0);
3768 	}
3769 }
3770 
3771 static void
3772 rack_validate_multipliers_at_or_above100(struct tcp_rack *rack)
3773 {
3774 	/*
3775 	 * When we drag bottom, we want to assure
3776 	 * that no multiplier is below 1.0, if so
3777 	 * we want to restore it to at least that.
3778 	 */
3779 	if (rack->r_ctl.rack_per_of_gp_rec  < 100) {
3780 		/* This is unlikely we usually do not touch recovery */
3781 		rack->r_ctl.rack_per_of_gp_rec = 100;
3782 	}
3783 	if (rack->r_ctl.rack_per_of_gp_ca < 100) {
3784 		rack->r_ctl.rack_per_of_gp_ca = 100;
3785 	}
3786 	if (rack->r_ctl.rack_per_of_gp_ss < 100) {
3787 		rack->r_ctl.rack_per_of_gp_ss = 100;
3788 	}
3789 }
3790 
3791 static void
3792 rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack)
3793 {
3794 	if (rack->r_ctl.rack_per_of_gp_ca > 100) {
3795 		rack->r_ctl.rack_per_of_gp_ca = 100;
3796 	}
3797 	if (rack->r_ctl.rack_per_of_gp_ss > 100) {
3798 		rack->r_ctl.rack_per_of_gp_ss = 100;
3799 	}
3800 }
3801 
3802 static void
3803 rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override)
3804 {
3805 	int32_t  calc, logged, plus;
3806 
3807 	logged = 0;
3808 
3809 	if (override) {
3810 		/*
3811 		 * override is passed when we are
3812 		 * loosing b/w and making one last
3813 		 * gasp at trying to not loose out
3814 		 * to a new-reno flow.
3815 		 */
3816 		goto extra_boost;
3817 	}
3818 	/* In classic timely we boost by 5x if we have 5 increases in a row, lets not */
3819 	if (rack->rc_gp_incr &&
3820 	    ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) {
3821 		/*
3822 		 * Reset and get 5 strokes more before the boost. Note
3823 		 * that the count is 0 based so we have to add one.
3824 		 */
3825 extra_boost:
3826 		plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST;
3827 		rack->rc_gp_timely_inc_cnt = 0;
3828 	} else
3829 		plus = (uint32_t)rack_gp_increase_per;
3830 	/* Must be at least 1% increase for true timely increases */
3831 	if ((plus < 1) &&
3832 	    ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0)))
3833 		plus = 1;
3834 	if (rack->rc_gp_saw_rec &&
3835 	    (rack->rc_gp_no_rec_chg == 0) &&
3836 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3837 				  rack->r_ctl.rack_per_of_gp_rec)) {
3838 		/* We have been in recovery ding it too */
3839 		calc = rack->r_ctl.rack_per_of_gp_rec + plus;
3840 		if (calc > 0xffff)
3841 			calc = 0xffff;
3842 		logged |= 1;
3843 		rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc;
3844 		if (rack->r_ctl.rack_per_upper_bound_ca &&
3845 		    (rack->rc_dragged_bottom == 0) &&
3846 		    (rack->r_ctl.rack_per_of_gp_rec > rack->r_ctl.rack_per_upper_bound_ca))
3847 			rack->r_ctl.rack_per_of_gp_rec = rack->r_ctl.rack_per_upper_bound_ca;
3848 	}
3849 	if (rack->rc_gp_saw_ca &&
3850 	    (rack->rc_gp_saw_ss == 0) &&
3851 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3852 				  rack->r_ctl.rack_per_of_gp_ca)) {
3853 		/* In CA */
3854 		calc = rack->r_ctl.rack_per_of_gp_ca + plus;
3855 		if (calc > 0xffff)
3856 			calc = 0xffff;
3857 		logged |= 2;
3858 		rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc;
3859 		if (rack->r_ctl.rack_per_upper_bound_ca &&
3860 		    (rack->rc_dragged_bottom == 0) &&
3861 		    (rack->r_ctl.rack_per_of_gp_ca > rack->r_ctl.rack_per_upper_bound_ca))
3862 			rack->r_ctl.rack_per_of_gp_ca = rack->r_ctl.rack_per_upper_bound_ca;
3863 	}
3864 	if (rack->rc_gp_saw_ss &&
3865 	    rack_bw_can_be_raised(rack, cur_bw, last_bw_est,
3866 				  rack->r_ctl.rack_per_of_gp_ss)) {
3867 		/* In SS */
3868 		calc = rack->r_ctl.rack_per_of_gp_ss + plus;
3869 		if (calc > 0xffff)
3870 			calc = 0xffff;
3871 		rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc;
3872 		if (rack->r_ctl.rack_per_upper_bound_ss &&
3873 		    (rack->rc_dragged_bottom == 0) &&
3874 		    (rack->r_ctl.rack_per_of_gp_ss > rack->r_ctl.rack_per_upper_bound_ss))
3875 			rack->r_ctl.rack_per_of_gp_ss = rack->r_ctl.rack_per_upper_bound_ss;
3876 		logged |= 4;
3877 	}
3878 	if (logged &&
3879 	    (rack->rc_gp_incr == 0)){
3880 		/* Go into increment mode */
3881 		rack->rc_gp_incr = 1;
3882 		rack->rc_gp_timely_inc_cnt = 0;
3883 	}
3884 	if (rack->rc_gp_incr &&
3885 	    logged &&
3886 	    (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) {
3887 		rack->rc_gp_timely_inc_cnt++;
3888 	}
3889 	rack_log_timely(rack,  logged, plus, 0, 0,
3890 			__LINE__, 1);
3891 }
3892 
3893 static uint32_t
3894 rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff)
3895 {
3896 	/*-
3897 	 * norm_grad = rtt_diff / minrtt;
3898 	 * new_per = curper * (1 - B * norm_grad)
3899 	 *
3900 	 * B = rack_gp_decrease_per (default 80%)
3901 	 * rtt_dif = input var current rtt-diff
3902 	 * curper = input var current percentage
3903 	 * minrtt = from rack filter
3904 	 *
3905 	 * In order to do the floating point calculations above we
3906 	 * do an integer conversion. The code looks confusing so let me
3907 	 * translate it into something that use more variables and
3908 	 * is clearer for us humans :)
3909 	 *
3910 	 * uint64_t norm_grad, inverse, reduce_by, final_result;
3911 	 * uint32_t perf;
3912 	 *
3913 	 * norm_grad = (((uint64_t)rtt_diff * 1000000) /
3914 	 *             (uint64_t)get_filter_small(&rack->r_ctl.rc_gp_min_rtt));
3915 	 * inverse = ((uint64_t)rack_gp_decrease * (uint64_t)1000000) * norm_grad;
3916 	 * inverse /= 1000000;
3917 	 * reduce_by = (1000000 - inverse);
3918 	 * final_result = (cur_per * reduce_by) / 1000000;
3919 	 * perf = (uint32_t)final_result;
3920 	 */
3921 	uint64_t perf;
3922 
3923 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3924 		    ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 *
3925 		     (((uint64_t)rtt_diff * (uint64_t)1000000)/
3926 		      (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/
3927 		     (uint64_t)1000000)) /
3928 		(uint64_t)1000000);
3929 	if (perf > curper) {
3930 		/* TSNH */
3931 		perf = curper - 1;
3932 	}
3933 	return ((uint32_t)perf);
3934 }
3935 
3936 static uint32_t
3937 rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt)
3938 {
3939 	/*
3940 	 *                                   highrttthresh
3941 	 * result = curper * (1 - (B * ( 1 -  ------          ))
3942 	 *                                     gp_srtt
3943 	 *
3944 	 * B = rack_gp_decrease_per (default .8 i.e. 80)
3945 	 * highrttthresh = filter_min * rack_gp_rtt_maxmul
3946 	 */
3947 	uint64_t perf;
3948 	uint32_t highrttthresh;
3949 
3950 	highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
3951 
3952 	perf = (((uint64_t)curper * ((uint64_t)1000000 -
3953 				     ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 -
3954 					((uint64_t)highrttthresh * (uint64_t)1000000) /
3955 						    (uint64_t)rtt)) / 100)) /(uint64_t)1000000);
3956 	if (tcp_bblogging_on(rack->rc_tp)) {
3957 		uint64_t log1;
3958 
3959 		log1 = rtt;
3960 		log1 <<= 32;
3961 		log1 |= highrttthresh;
3962 		rack_log_timely(rack,
3963 				rack_gp_decrease_per,
3964 				(uint64_t)curper,
3965 				log1,
3966 				perf,
3967 				__LINE__,
3968 				15);
3969 	}
3970 	return (perf);
3971 }
3972 
3973 static void
3974 rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff)
3975 {
3976 	uint64_t logvar, logvar2, logvar3;
3977 	uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val;
3978 
3979 	if (rack->rc_gp_incr) {
3980 		/* Turn off increment counting */
3981 		rack->rc_gp_incr = 0;
3982 		rack->rc_gp_timely_inc_cnt = 0;
3983 	}
3984 	ss_red = ca_red = rec_red = 0;
3985 	logged = 0;
3986 	/* Calculate the reduction value */
3987 	if (rtt_diff < 0) {
3988 		rtt_diff *= -1;
3989 	}
3990 	/* Must be at least 1% reduction */
3991 	if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) {
3992 		/* We have been in recovery ding it too */
3993 		if (timely_says == 2) {
3994 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt);
3995 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
3996 			if (alt < new_per)
3997 				val = alt;
3998 			else
3999 				val = new_per;
4000 		} else
4001 			 val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff);
4002 		if (rack->r_ctl.rack_per_of_gp_rec > val) {
4003 			rec_red = (rack->r_ctl.rack_per_of_gp_rec - val);
4004 			rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val;
4005 		} else {
4006 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4007 			rec_red = 0;
4008 		}
4009 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec)
4010 			rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound;
4011 		logged |= 1;
4012 	}
4013 	if (rack->rc_gp_saw_ss) {
4014 		/* Sent in SS */
4015 		if (timely_says == 2) {
4016 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt);
4017 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4018 			if (alt < new_per)
4019 				val = alt;
4020 			else
4021 				val = new_per;
4022 		} else
4023 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff);
4024 		if (rack->r_ctl.rack_per_of_gp_ss > new_per) {
4025 			ss_red = rack->r_ctl.rack_per_of_gp_ss - val;
4026 			rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val;
4027 		} else {
4028 			ss_red = new_per;
4029 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4030 			logvar = new_per;
4031 			logvar <<= 32;
4032 			logvar |= alt;
4033 			logvar2 = (uint32_t)rtt;
4034 			logvar2 <<= 32;
4035 			logvar2 |= (uint32_t)rtt_diff;
4036 			logvar3 = rack_gp_rtt_maxmul;
4037 			logvar3 <<= 32;
4038 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4039 			rack_log_timely(rack, timely_says,
4040 					logvar2, logvar3,
4041 					logvar, __LINE__, 10);
4042 		}
4043 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss)
4044 			rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound;
4045 		logged |= 4;
4046 	} else if (rack->rc_gp_saw_ca) {
4047 		/* Sent in CA */
4048 		if (timely_says == 2) {
4049 			new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt);
4050 			alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4051 			if (alt < new_per)
4052 				val = alt;
4053 			else
4054 				val = new_per;
4055 		} else
4056 			val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff);
4057 		if (rack->r_ctl.rack_per_of_gp_ca > val) {
4058 			ca_red = rack->r_ctl.rack_per_of_gp_ca - val;
4059 			rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val;
4060 		} else {
4061 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4062 			ca_red = 0;
4063 			logvar = new_per;
4064 			logvar <<= 32;
4065 			logvar |= alt;
4066 			logvar2 = (uint32_t)rtt;
4067 			logvar2 <<= 32;
4068 			logvar2 |= (uint32_t)rtt_diff;
4069 			logvar3 = rack_gp_rtt_maxmul;
4070 			logvar3 <<= 32;
4071 			logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4072 			rack_log_timely(rack, timely_says,
4073 					logvar2, logvar3,
4074 					logvar, __LINE__, 10);
4075 		}
4076 		if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca)
4077 			rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound;
4078 		logged |= 2;
4079 	}
4080 	if (rack->rc_gp_timely_dec_cnt < 0x7) {
4081 		rack->rc_gp_timely_dec_cnt++;
4082 		if (rack_timely_dec_clear &&
4083 		    (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear))
4084 			rack->rc_gp_timely_dec_cnt = 0;
4085 	}
4086 	logvar = ss_red;
4087 	logvar <<= 32;
4088 	logvar |= ca_red;
4089 	rack_log_timely(rack,  logged, rec_red, rack_per_lower_bound, logvar,
4090 			__LINE__, 2);
4091 }
4092 
4093 static void
4094 rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts,
4095 		     uint32_t rtt, uint32_t line, uint8_t reas)
4096 {
4097 	if (tcp_bblogging_on(rack->rc_tp)) {
4098 		union tcp_log_stackspecific log;
4099 		struct timeval tv;
4100 
4101 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
4102 		log.u_bbr.flex1 = line;
4103 		log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts;
4104 		log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts;
4105 		log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss;
4106 		log.u_bbr.flex5 = rtt;
4107 		log.u_bbr.flex6 = rack->rc_highly_buffered;
4108 		log.u_bbr.flex6 <<= 1;
4109 		log.u_bbr.flex6 |= rack->forced_ack;
4110 		log.u_bbr.flex6 <<= 1;
4111 		log.u_bbr.flex6 |= rack->rc_gp_dyn_mul;
4112 		log.u_bbr.flex6 <<= 1;
4113 		log.u_bbr.flex6 |= rack->in_probe_rtt;
4114 		log.u_bbr.flex6 <<= 1;
4115 		log.u_bbr.flex6 |= rack->measure_saw_probe_rtt;
4116 		log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt;
4117 		log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca;
4118 		log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec;
4119 		log.u_bbr.flex8 = reas;
4120 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
4121 		log.u_bbr.delRate = rack_get_bw(rack);
4122 		log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt;
4123 		log.u_bbr.cur_del_rate <<= 32;
4124 		log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt;
4125 		log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered;
4126 		log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff;
4127 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
4128 		log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt;
4129 		log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt;
4130 		log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts;
4131 		log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight;
4132 		log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4133 		log.u_bbr.rttProp = us_cts;
4134 		log.u_bbr.rttProp <<= 32;
4135 		log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt;
4136 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
4137 		    &rack->rc_inp->inp_socket->so_rcv,
4138 		    &rack->rc_inp->inp_socket->so_snd,
4139 		    BBR_LOG_RTT_SHRINKS, 0,
4140 		    0, &log, false, &rack->r_ctl.act_rcv_time);
4141 	}
4142 }
4143 
4144 static void
4145 rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt)
4146 {
4147 	uint64_t bwdp;
4148 
4149 	bwdp = rack_get_bw(rack);
4150 	bwdp *= (uint64_t)rtt;
4151 	bwdp /= (uint64_t)HPTS_USEC_IN_SEC;
4152 	rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz);
4153 	if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) {
4154 		/*
4155 		 * A window protocol must be able to have 4 packets
4156 		 * outstanding as the floor in order to function
4157 		 * (especially considering delayed ack :D).
4158 		 */
4159 		rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs);
4160 	}
4161 }
4162 
4163 static void
4164 rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts)
4165 {
4166 	/**
4167 	 * ProbeRTT is a bit different in rack_pacing than in
4168 	 * BBR. It is like BBR in that it uses the lowering of
4169 	 * the RTT as a signal that we saw something new and
4170 	 * counts from there for how long between. But it is
4171 	 * different in that its quite simple. It does not
4172 	 * play with the cwnd and wait until we get down
4173 	 * to N segments outstanding and hold that for
4174 	 * 200ms. Instead it just sets the pacing reduction
4175 	 * rate to a set percentage (70 by default) and hold
4176 	 * that for a number of recent GP Srtt's.
4177 	 */
4178 	uint32_t segsiz;
4179 
4180 	if (rack->rc_gp_dyn_mul == 0)
4181 		return;
4182 
4183 	if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) {
4184 		/* We are idle */
4185 		return;
4186 	}
4187 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4188 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4189 		/*
4190 		 * Stop the goodput now, the idea here is
4191 		 * that future measurements with in_probe_rtt
4192 		 * won't register if they are not greater so
4193 		 * we want to get what info (if any) is available
4194 		 * now.
4195 		 */
4196 		rack_do_goodput_measurement(rack->rc_tp, rack,
4197 					    rack->rc_tp->snd_una, __LINE__,
4198 					    RACK_QUALITY_PROBERTT);
4199 	}
4200 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4201 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4202 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4203 		     rack->r_ctl.rc_pace_min_segs);
4204 	rack->in_probe_rtt = 1;
4205 	rack->measure_saw_probe_rtt = 1;
4206 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4207 	rack->r_ctl.rc_time_probertt_starts = 0;
4208 	rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt;
4209 	if (rack_probertt_use_min_rtt_entry)
4210 		rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4211 	else
4212 		rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt);
4213 	rack_log_rtt_shrinks(rack,  us_cts,  get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4214 			     __LINE__, RACK_RTTS_ENTERPROBE);
4215 }
4216 
4217 static void
4218 rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts)
4219 {
4220 	struct rack_sendmap *rsm;
4221 	uint32_t segsiz;
4222 
4223 	segsiz = min(ctf_fixed_maxseg(rack->rc_tp),
4224 		     rack->r_ctl.rc_pace_min_segs);
4225 	rack->in_probe_rtt = 0;
4226 	if ((rack->rc_tp->t_flags & TF_GPUTINPROG) &&
4227 	    SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) {
4228 		/*
4229 		 * Stop the goodput now, the idea here is
4230 		 * that future measurements with in_probe_rtt
4231 		 * won't register if they are not greater so
4232 		 * we want to get what info (if any) is available
4233 		 * now.
4234 		 */
4235 		rack_do_goodput_measurement(rack->rc_tp, rack,
4236 					    rack->rc_tp->snd_una, __LINE__,
4237 					    RACK_QUALITY_PROBERTT);
4238 	} else if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
4239 		/*
4240 		 * We don't have enough data to make a measurement.
4241 		 * So lets just stop and start here after exiting
4242 		 * probe-rtt. We probably are not interested in
4243 		 * the results anyway.
4244 		 */
4245 		rack->rc_tp->t_flags &= ~TF_GPUTINPROG;
4246 	}
4247 	/*
4248 	 * Measurements through the current snd_max are going
4249 	 * to be limited by the slower pacing rate.
4250 	 *
4251 	 * We need to mark these as app-limited so we
4252 	 * don't collapse the b/w.
4253 	 */
4254 	rsm = tqhash_max(rack->r_ctl.tqh);
4255 	if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
4256 		if (rack->r_ctl.rc_app_limited_cnt == 0)
4257 			rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
4258 		else {
4259 			/*
4260 			 * Go out to the end app limited and mark
4261 			 * this new one as next and move the end_appl up
4262 			 * to this guy.
4263 			 */
4264 			if (rack->r_ctl.rc_end_appl)
4265 				rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
4266 			rack->r_ctl.rc_end_appl = rsm;
4267 		}
4268 		rsm->r_flags |= RACK_APP_LIMITED;
4269 		rack->r_ctl.rc_app_limited_cnt++;
4270 	}
4271 	/*
4272 	 * Now, we need to examine our pacing rate multipliers.
4273 	 * If its under 100%, we need to kick it back up to
4274 	 * 100%. We also don't let it be over our "max" above
4275 	 * the actual rate i.e. 100% + rack_clamp_atexit_prtt.
4276 	 * Note setting clamp_atexit_prtt to 0 has the effect
4277 	 * of setting CA/SS to 100% always at exit (which is
4278 	 * the default behavior).
4279 	 */
4280 	if (rack_probertt_clear_is) {
4281 		rack->rc_gp_incr = 0;
4282 		rack->rc_gp_bwred = 0;
4283 		rack->rc_gp_timely_inc_cnt = 0;
4284 		rack->rc_gp_timely_dec_cnt = 0;
4285 	}
4286 	/* Do we do any clamping at exit? */
4287 	if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) {
4288 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp;
4289 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp;
4290 	}
4291 	if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) {
4292 		rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt;
4293 		rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt;
4294 	}
4295 	/*
4296 	 * Lets set rtt_diff to 0, so that we will get a "boost"
4297 	 * after exiting.
4298 	 */
4299 	rack->r_ctl.rc_rtt_diff = 0;
4300 
4301 	/* Clear all flags so we start fresh */
4302 	rack->rc_tp->t_bytes_acked = 0;
4303 	rack->rc_tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
4304 	/*
4305 	 * If configured to, set the cwnd and ssthresh to
4306 	 * our targets.
4307 	 */
4308 	if (rack_probe_rtt_sets_cwnd) {
4309 		uint64_t ebdp;
4310 		uint32_t setto;
4311 
4312 		/* Set ssthresh so we get into CA once we hit our target */
4313 		if (rack_probertt_use_min_rtt_exit == 1) {
4314 			/* Set to min rtt */
4315 			rack_set_prtt_target(rack, segsiz,
4316 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt));
4317 		} else if (rack_probertt_use_min_rtt_exit == 2) {
4318 			/* Set to current gp rtt */
4319 			rack_set_prtt_target(rack, segsiz,
4320 					     rack->r_ctl.rc_gp_srtt);
4321 		} else if (rack_probertt_use_min_rtt_exit == 3) {
4322 			/* Set to entry gp rtt */
4323 			rack_set_prtt_target(rack, segsiz,
4324 					     rack->r_ctl.rc_entry_gp_rtt);
4325 		} else {
4326 			uint64_t sum;
4327 			uint32_t setval;
4328 
4329 			sum = rack->r_ctl.rc_entry_gp_rtt;
4330 			sum *= 10;
4331 			sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt));
4332 			if (sum >= 20) {
4333 				/*
4334 				 * A highly buffered path needs
4335 				 * cwnd space for timely to work.
4336 				 * Lets set things up as if
4337 				 * we are heading back here again.
4338 				 */
4339 				setval = rack->r_ctl.rc_entry_gp_rtt;
4340 			} else if (sum >= 15) {
4341 				/*
4342 				 * Lets take the smaller of the
4343 				 * two since we are just somewhat
4344 				 * buffered.
4345 				 */
4346 				setval = rack->r_ctl.rc_gp_srtt;
4347 				if (setval > rack->r_ctl.rc_entry_gp_rtt)
4348 					setval = rack->r_ctl.rc_entry_gp_rtt;
4349 			} else {
4350 				/*
4351 				 * Here we are not highly buffered
4352 				 * and should pick the min we can to
4353 				 * keep from causing loss.
4354 				 */
4355 				setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
4356 			}
4357 			rack_set_prtt_target(rack, segsiz,
4358 					     setval);
4359 		}
4360 		if (rack_probe_rtt_sets_cwnd > 1) {
4361 			/* There is a percentage here to boost */
4362 			ebdp = rack->r_ctl.rc_target_probertt_flight;
4363 			ebdp *= rack_probe_rtt_sets_cwnd;
4364 			ebdp /= 100;
4365 			setto = rack->r_ctl.rc_target_probertt_flight + ebdp;
4366 		} else
4367 			setto = rack->r_ctl.rc_target_probertt_flight;
4368 		rack->rc_tp->snd_cwnd = roundup(setto, segsiz);
4369 		if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) {
4370 			/* Enforce a min */
4371 			rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs;
4372 		}
4373 		/* If we set in the cwnd also set the ssthresh point so we are in CA */
4374 		rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1);
4375 	}
4376 	rack_log_rtt_shrinks(rack,  us_cts,
4377 			     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4378 			     __LINE__, RACK_RTTS_EXITPROBE);
4379 	/* Clear times last so log has all the info */
4380 	rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max;
4381 	rack->r_ctl.rc_time_probertt_entered = us_cts;
4382 	rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
4383 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
4384 }
4385 
4386 static void
4387 rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts)
4388 {
4389 	/* Check in on probe-rtt */
4390 	if (rack->rc_gp_filled == 0) {
4391 		/* We do not do p-rtt unless we have gp measurements */
4392 		return;
4393 	}
4394 	if (rack->in_probe_rtt) {
4395 		uint64_t no_overflow;
4396 		uint32_t endtime, must_stay;
4397 
4398 		if (rack->r_ctl.rc_went_idle_time &&
4399 		    ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) {
4400 			/*
4401 			 * We went idle during prtt, just exit now.
4402 			 */
4403 			rack_exit_probertt(rack, us_cts);
4404 		} else if (rack_probe_rtt_safety_val &&
4405 		    TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) &&
4406 		    ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) {
4407 			/*
4408 			 * Probe RTT safety value triggered!
4409 			 */
4410 			rack_log_rtt_shrinks(rack,  us_cts,
4411 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4412 					     __LINE__, RACK_RTTS_SAFETY);
4413 			rack_exit_probertt(rack, us_cts);
4414 		}
4415 		/* Calculate the max we will wait */
4416 		endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait);
4417 		if (rack->rc_highly_buffered)
4418 			endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp);
4419 		/* Calculate the min we must wait */
4420 		must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain);
4421 		if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) &&
4422 		    TSTMP_LT(us_cts, endtime)) {
4423 			uint32_t calc;
4424 			/* Do we lower more? */
4425 no_exit:
4426 			if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered))
4427 				calc = us_cts - rack->r_ctl.rc_time_probertt_entered;
4428 			else
4429 				calc = 0;
4430 			calc /= max(rack->r_ctl.rc_gp_srtt, 1);
4431 			if (calc) {
4432 				/* Maybe */
4433 				calc *= rack_per_of_gp_probertt_reduce;
4434 				rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc;
4435 				/* Limit it too */
4436 				if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh)
4437 					rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh;
4438 			}
4439 			/* We must reach target or the time set */
4440 			return;
4441 		}
4442 		if (rack->r_ctl.rc_time_probertt_starts == 0) {
4443 			if ((TSTMP_LT(us_cts, must_stay) &&
4444 			     rack->rc_highly_buffered) ||
4445 			     (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) >
4446 			      rack->r_ctl.rc_target_probertt_flight)) {
4447 				/* We are not past the must_stay time */
4448 				goto no_exit;
4449 			}
4450 			rack_log_rtt_shrinks(rack,  us_cts,
4451 					     get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4452 					     __LINE__, RACK_RTTS_REACHTARGET);
4453 			rack->r_ctl.rc_time_probertt_starts = us_cts;
4454 			if (rack->r_ctl.rc_time_probertt_starts == 0)
4455 				rack->r_ctl.rc_time_probertt_starts = 1;
4456 			/* Restore back to our rate we want to pace at in prtt */
4457 			rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
4458 		}
4459 		/*
4460 		 * Setup our end time, some number of gp_srtts plus 200ms.
4461 		 */
4462 		no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt *
4463 			       (uint64_t)rack_probertt_gpsrtt_cnt_mul);
4464 		if (rack_probertt_gpsrtt_cnt_div)
4465 			endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div);
4466 		else
4467 			endtime = 0;
4468 		endtime += rack_min_probertt_hold;
4469 		endtime += rack->r_ctl.rc_time_probertt_starts;
4470 		if (TSTMP_GEQ(us_cts,  endtime)) {
4471 			/* yes, exit probertt */
4472 			rack_exit_probertt(rack, us_cts);
4473 		}
4474 
4475 	} else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) {
4476 		/* Go into probertt, its been too long since we went lower */
4477 		rack_enter_probertt(rack, us_cts);
4478 	}
4479 }
4480 
4481 static void
4482 rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est,
4483 		       uint32_t rtt, int32_t rtt_diff)
4484 {
4485 	uint64_t cur_bw, up_bnd, low_bnd, subfr;
4486 	uint32_t losses;
4487 
4488 	if ((rack->rc_gp_dyn_mul == 0) ||
4489 	    (rack->use_fixed_rate) ||
4490 	    (rack->in_probe_rtt) ||
4491 	    (rack->rc_always_pace == 0)) {
4492 		/* No dynamic GP multiplier in play */
4493 		return;
4494 	}
4495 	losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start;
4496 	cur_bw = rack_get_bw(rack);
4497 	/* Calculate our up and down range */
4498 	up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up;
4499 	up_bnd /= 100;
4500 	up_bnd += rack->r_ctl.last_gp_comp_bw;
4501 
4502 	subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down;
4503 	subfr /= 100;
4504 	low_bnd = rack->r_ctl.last_gp_comp_bw - subfr;
4505 	if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) {
4506 		/*
4507 		 * This is the case where our RTT is above
4508 		 * the max target and we have been configured
4509 		 * to just do timely no bonus up stuff in that case.
4510 		 *
4511 		 * There are two configurations, set to 1, and we
4512 		 * just do timely if we are over our max. If its
4513 		 * set above 1 then we slam the multipliers down
4514 		 * to 100 and then decrement per timely.
4515 		 */
4516 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4517 				__LINE__, 3);
4518 		if (rack->r_ctl.rc_no_push_at_mrtt > 1)
4519 			rack_validate_multipliers_at_or_below_100(rack);
4520 		rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4521 	} else if ((timely_says != 0) && (last_bw_est < low_bnd) && !losses) {
4522 		/*
4523 		 * We are decreasing this is a bit complicated this
4524 		 * means we are loosing ground. This could be
4525 		 * because another flow entered and we are competing
4526 		 * for b/w with it. This will push the RTT up which
4527 		 * makes timely unusable unless we want to get shoved
4528 		 * into a corner and just be backed off (the age
4529 		 * old problem with delay based CC).
4530 		 *
4531 		 * On the other hand if it was a route change we
4532 		 * would like to stay somewhat contained and not
4533 		 * blow out the buffers.
4534 		 */
4535 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4536 				__LINE__, 3);
4537 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4538 		if (rack->rc_gp_bwred == 0) {
4539 			/* Go into reduction counting */
4540 			rack->rc_gp_bwred = 1;
4541 			rack->rc_gp_timely_dec_cnt = 0;
4542 		}
4543 		if (rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) {
4544 			/*
4545 			 * Push another time with a faster pacing
4546 			 * to try to gain back (we include override to
4547 			 * get a full raise factor).
4548 			 */
4549 			if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) ||
4550 			    (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) ||
4551 			    (timely_says == 0) ||
4552 			    (rack_down_raise_thresh == 0)) {
4553 				/*
4554 				 * Do an override up in b/w if we were
4555 				 * below the threshold or if the threshold
4556 				 * is zero we always do the raise.
4557 				 */
4558 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1);
4559 			} else {
4560 				/* Log it stays the same */
4561 				rack_log_timely(rack,  0, last_bw_est, low_bnd, 0,
4562 						__LINE__, 11);
4563 			}
4564 			rack->rc_gp_timely_dec_cnt++;
4565 			/* We are not incrementing really no-count */
4566 			rack->rc_gp_incr = 0;
4567 			rack->rc_gp_timely_inc_cnt = 0;
4568 		} else {
4569 			/*
4570 			 * Lets just use the RTT
4571 			 * information and give up
4572 			 * pushing.
4573 			 */
4574 			goto use_timely;
4575 		}
4576 	} else if ((timely_says != 2) &&
4577 		    !losses &&
4578 		    (last_bw_est > up_bnd)) {
4579 		/*
4580 		 * We are increasing b/w lets keep going, updating
4581 		 * our b/w and ignoring any timely input, unless
4582 		 * of course we are at our max raise (if there is one).
4583 		 */
4584 
4585 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4586 				__LINE__, 3);
4587 		rack->r_ctl.last_gp_comp_bw = cur_bw;
4588 		if (rack->rc_gp_saw_ss &&
4589 		    rack->r_ctl.rack_per_upper_bound_ss &&
4590 		     (rack->r_ctl.rack_per_of_gp_ss == rack->r_ctl.rack_per_upper_bound_ss)) {
4591 			    /*
4592 			     * In cases where we can't go higher
4593 			     * we should just use timely.
4594 			     */
4595 			    goto use_timely;
4596 		}
4597 		if (rack->rc_gp_saw_ca &&
4598 		    rack->r_ctl.rack_per_upper_bound_ca &&
4599 		    (rack->r_ctl.rack_per_of_gp_ca == rack->r_ctl.rack_per_upper_bound_ca)) {
4600 			    /*
4601 			     * In cases where we can't go higher
4602 			     * we should just use timely.
4603 			     */
4604 			    goto use_timely;
4605 		}
4606 		rack->rc_gp_bwred = 0;
4607 		rack->rc_gp_timely_dec_cnt = 0;
4608 		/* You get a set number of pushes if timely is trying to reduce */
4609 		if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) {
4610 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4611 		} else {
4612 			/* Log it stays the same */
4613 			rack_log_timely(rack,  0, last_bw_est, up_bnd, 0,
4614 			    __LINE__, 12);
4615 		}
4616 		return;
4617 	} else {
4618 		/*
4619 		 * We are staying between the lower and upper range bounds
4620 		 * so use timely to decide.
4621 		 */
4622 		rack_log_timely(rack,  timely_says, cur_bw, low_bnd, up_bnd,
4623 				__LINE__, 3);
4624 use_timely:
4625 		if (timely_says) {
4626 			rack->rc_gp_incr = 0;
4627 			rack->rc_gp_timely_inc_cnt = 0;
4628 			if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) &&
4629 			    !losses &&
4630 			    (last_bw_est < low_bnd)) {
4631 				/* We are loosing ground */
4632 				rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4633 				rack->rc_gp_timely_dec_cnt++;
4634 				/* We are not incrementing really no-count */
4635 				rack->rc_gp_incr = 0;
4636 				rack->rc_gp_timely_inc_cnt = 0;
4637 			} else
4638 				rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff);
4639 		} else {
4640 			rack->rc_gp_bwred = 0;
4641 			rack->rc_gp_timely_dec_cnt = 0;
4642 			rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0);
4643 		}
4644 	}
4645 }
4646 
4647 static int32_t
4648 rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt)
4649 {
4650 	int32_t timely_says;
4651 	uint64_t log_mult, log_rtt_a_diff;
4652 
4653 	log_rtt_a_diff = rtt;
4654 	log_rtt_a_diff <<= 32;
4655 	log_rtt_a_diff |= (uint32_t)rtt_diff;
4656 	if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) *
4657 		    rack_gp_rtt_maxmul)) {
4658 		/* Reduce the b/w multiplier */
4659 		timely_says = 2;
4660 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul;
4661 		log_mult <<= 32;
4662 		log_mult |= prev_rtt;
4663 		rack_log_timely(rack,  timely_says, log_mult,
4664 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4665 				log_rtt_a_diff, __LINE__, 4);
4666 	} else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4667 			   ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4668 			    max(rack_gp_rtt_mindiv , 1)))) {
4669 		/* Increase the b/w multiplier */
4670 		log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) +
4671 			((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) /
4672 			 max(rack_gp_rtt_mindiv , 1));
4673 		log_mult <<= 32;
4674 		log_mult |= prev_rtt;
4675 		timely_says = 0;
4676 		rack_log_timely(rack,  timely_says, log_mult ,
4677 				get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt),
4678 				log_rtt_a_diff, __LINE__, 5);
4679 	} else {
4680 		/*
4681 		 * Use a gradient to find it the timely gradient
4682 		 * is:
4683 		 * grad = rc_rtt_diff / min_rtt;
4684 		 *
4685 		 * anything below or equal to 0 will be
4686 		 * a increase indication. Anything above
4687 		 * zero is a decrease. Note we take care
4688 		 * of the actual gradient calculation
4689 		 * in the reduction (its not needed for
4690 		 * increase).
4691 		 */
4692 		log_mult = prev_rtt;
4693 		if (rtt_diff <= 0) {
4694 			/*
4695 			 * Rttdiff is less than zero, increase the
4696 			 * b/w multiplier (its 0 or negative)
4697 			 */
4698 			timely_says = 0;
4699 			rack_log_timely(rack,  timely_says, log_mult,
4700 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6);
4701 		} else {
4702 			/* Reduce the b/w multiplier */
4703 			timely_says = 1;
4704 			rack_log_timely(rack,  timely_says, log_mult,
4705 					get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7);
4706 		}
4707 	}
4708 	return (timely_says);
4709 }
4710 
4711 static __inline int
4712 rack_in_gp_window(struct tcpcb *tp, struct rack_sendmap *rsm)
4713 {
4714 	if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4715 	    SEQ_LEQ(rsm->r_end, tp->gput_ack)) {
4716 		/**
4717 		 * This covers the case that the
4718 		 * resent is completely inside
4719 		 * the gp range or up to it.
4720 		 *      |----------------|
4721 		 *      |-----| <or>
4722 		 *            |----|
4723 		 *            <or>   |---|
4724 		 */
4725 		return (1);
4726 	} else if (SEQ_LT(rsm->r_start, tp->gput_seq) &&
4727 		   SEQ_GT(rsm->r_end, tp->gput_seq)){
4728 		/**
4729 		 * This covers the case of
4730 		 *      |--------------|
4731 		 *  |-------->|
4732 		 */
4733 		return (1);
4734 	} else if (SEQ_GEQ(rsm->r_start, tp->gput_seq) &&
4735 		   SEQ_LT(rsm->r_start, tp->gput_ack) &&
4736 		   SEQ_GEQ(rsm->r_end, tp->gput_ack)) {
4737 
4738 		/**
4739 		 * This covers the case of
4740 		 *      |--------------|
4741 		 *              |-------->|
4742 		 */
4743 		return (1);
4744 	}
4745 	return (0);
4746 }
4747 
4748 static __inline void
4749 rack_mark_in_gp_win(struct tcpcb *tp, struct rack_sendmap *rsm)
4750 {
4751 
4752 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
4753 		return;
4754 	/*
4755 	 * We have a Goodput measurement in progress. Mark
4756 	 * the send if its within the window. If its not
4757 	 * in the window make sure it does not have the mark.
4758 	 */
4759 	if (rack_in_gp_window(tp, rsm))
4760 		rsm->r_flags |= RACK_IN_GP_WIN;
4761 	else
4762 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4763 }
4764 
4765 static __inline void
4766 rack_clear_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4767 {
4768 	/* A GP measurement is ending, clear all marks on the send map*/
4769 	struct rack_sendmap *rsm = NULL;
4770 
4771 	rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4772 	if (rsm == NULL) {
4773 		rsm = tqhash_min(rack->r_ctl.tqh);
4774 	}
4775 	/* Nothing left? */
4776 	while ((rsm != NULL) && (SEQ_GEQ(tp->gput_ack, rsm->r_start))){
4777 		rsm->r_flags &= ~RACK_IN_GP_WIN;
4778 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4779 	}
4780 }
4781 
4782 
4783 static __inline void
4784 rack_tend_gp_marks(struct tcpcb *tp, struct tcp_rack *rack)
4785 {
4786 	struct rack_sendmap *rsm = NULL;
4787 
4788 	if (tp->snd_una == tp->snd_max) {
4789 		/* Nothing outstanding yet, nothing to do here */
4790 		return;
4791 	}
4792 	if (SEQ_GT(tp->gput_seq, tp->snd_una)) {
4793 		/*
4794 		 * We are measuring ahead of some outstanding
4795 		 * data. We need to walk through up until we get
4796 		 * to gp_seq marking so that no rsm is set incorrectly
4797 		 * with RACK_IN_GP_WIN.
4798 		 */
4799 		rsm = tqhash_min(rack->r_ctl.tqh);
4800 		while (rsm != NULL) {
4801 			rack_mark_in_gp_win(tp, rsm);
4802 			if (SEQ_GEQ(rsm->r_end, tp->gput_seq))
4803 				break;
4804 			rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4805 		}
4806 	}
4807 	if (rsm == NULL) {
4808 		/*
4809 		 * Need to find the GP seq, if rsm is
4810 		 * set we stopped as we hit it.
4811 		 */
4812 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
4813 		if (rsm == NULL)
4814 			return;
4815 		rack_mark_in_gp_win(tp, rsm);
4816 	}
4817 	/*
4818 	 * Now we may need to mark already sent rsm, ahead of
4819 	 * gput_seq in the window since they may have been sent
4820 	 * *before* we started our measurment. The rsm, if non-null
4821 	 * has been marked (note if rsm would have been NULL we would have
4822 	 * returned in the previous block). So we go to the next, and continue
4823 	 * until we run out of entries or we exceed the gp_ack value.
4824 	 */
4825 	rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4826 	while (rsm) {
4827 		rack_mark_in_gp_win(tp, rsm);
4828 		if (SEQ_GT(rsm->r_end, tp->gput_ack))
4829 			break;
4830 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
4831 	}
4832 }
4833 
4834 static void
4835 rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack,
4836 			    tcp_seq th_ack, int line, uint8_t quality)
4837 {
4838 	uint64_t tim, bytes_ps, stim, utim;
4839 	uint32_t segsiz, bytes, reqbytes, us_cts;
4840 	int32_t gput, new_rtt_diff, timely_says;
4841 	uint64_t  resid_bw, subpart = 0, addpart = 0, srtt;
4842 	int did_add = 0;
4843 
4844 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
4845 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
4846 	if (TSTMP_GEQ(us_cts, tp->gput_ts))
4847 		tim = us_cts - tp->gput_ts;
4848 	else
4849 		tim = 0;
4850 	if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts)
4851 		stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts;
4852 	else
4853 		stim = 0;
4854 	/*
4855 	 * Use the larger of the send time or ack time. This prevents us
4856 	 * from being influenced by ack artifacts to come up with too
4857 	 * high of measurement. Note that since we are spanning over many more
4858 	 * bytes in most of our measurements hopefully that is less likely to
4859 	 * occur.
4860 	 */
4861 	if (tim > stim)
4862 		utim = max(tim, 1);
4863 	else
4864 		utim = max(stim, 1);
4865 	reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz));
4866 	rack_log_gpset(rack, th_ack, us_cts, rack->r_ctl.rc_gp_cumack_ts, __LINE__, 3, NULL);
4867 	if ((tim == 0) && (stim == 0)) {
4868 		/*
4869 		 * Invalid measurement time, maybe
4870 		 * all on one ack/one send?
4871 		 */
4872 		bytes = 0;
4873 		bytes_ps = 0;
4874 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4875 					   0, 0, 0, 10, __LINE__, NULL, quality);
4876 		goto skip_measurement;
4877 	}
4878 	if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) {
4879 		/* We never made a us_rtt measurement? */
4880 		bytes = 0;
4881 		bytes_ps = 0;
4882 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4883 					   0, 0, 0, 10, __LINE__, NULL, quality);
4884 		goto skip_measurement;
4885 	}
4886 	/*
4887 	 * Calculate the maximum possible b/w this connection
4888 	 * could have. We base our calculation on the lowest
4889 	 * rtt we have seen during the measurement and the
4890 	 * largest rwnd the client has given us in that time. This
4891 	 * forms a BDP that is the maximum that we could ever
4892 	 * get to the client. Anything larger is not valid.
4893 	 *
4894 	 * I originally had code here that rejected measurements
4895 	 * where the time was less than 1/2 the latest us_rtt.
4896 	 * But after thinking on that I realized its wrong since
4897 	 * say you had a 150Mbps or even 1Gbps link, and you
4898 	 * were a long way away.. example I am in Europe (100ms rtt)
4899 	 * talking to my 1Gbps link in S.C. Now measuring say 150,000
4900 	 * bytes my time would be 1.2ms, and yet my rtt would say
4901 	 * the measurement was invalid the time was < 50ms. The
4902 	 * same thing is true for 150Mb (8ms of time).
4903 	 *
4904 	 * A better way I realized is to look at what the maximum
4905 	 * the connection could possibly do. This is gated on
4906 	 * the lowest RTT we have seen and the highest rwnd.
4907 	 * We should in theory never exceed that, if we are
4908 	 * then something on the path is storing up packets
4909 	 * and then feeding them all at once to our endpoint
4910 	 * messing up our measurement.
4911 	 */
4912 	rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd;
4913 	rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC;
4914 	rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt;
4915 	if (SEQ_LT(th_ack, tp->gput_seq)) {
4916 		/* No measurement can be made */
4917 		bytes = 0;
4918 		bytes_ps = 0;
4919 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4920 					   0, 0, 0, 10, __LINE__, NULL, quality);
4921 		goto skip_measurement;
4922 	} else
4923 		bytes = (th_ack - tp->gput_seq);
4924 	bytes_ps = (uint64_t)bytes;
4925 	/*
4926 	 * Don't measure a b/w for pacing unless we have gotten at least
4927 	 * an initial windows worth of data in this measurement interval.
4928 	 *
4929 	 * Small numbers of bytes get badly influenced by delayed ack and
4930 	 * other artifacts. Note we take the initial window or our
4931 	 * defined minimum GP (defaulting to 10 which hopefully is the
4932 	 * IW).
4933 	 */
4934 	if (rack->rc_gp_filled == 0) {
4935 		/*
4936 		 * The initial estimate is special. We
4937 		 * have blasted out an IW worth of packets
4938 		 * without a real valid ack ts results. We
4939 		 * then setup the app_limited_needs_set flag,
4940 		 * this should get the first ack in (probably 2
4941 		 * MSS worth) to be recorded as the timestamp.
4942 		 * We thus allow a smaller number of bytes i.e.
4943 		 * IW - 2MSS.
4944 		 */
4945 		reqbytes -= (2 * segsiz);
4946 		/* Also lets fill previous for our first measurement to be neutral */
4947 		rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
4948 	}
4949 	if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) {
4950 		rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
4951 					   rack->r_ctl.rc_app_limited_cnt,
4952 					   0, 0, 10, __LINE__, NULL, quality);
4953 		goto skip_measurement;
4954 	}
4955 	/*
4956 	 * We now need to calculate the Timely like status so
4957 	 * we can update (possibly) the b/w multipliers.
4958 	 */
4959 	new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt;
4960 	if (rack->rc_gp_filled == 0) {
4961 		/* No previous reading */
4962 		rack->r_ctl.rc_rtt_diff = new_rtt_diff;
4963 	} else {
4964 		if (rack->measure_saw_probe_rtt == 0) {
4965 			/*
4966 			 * We don't want a probertt to be counted
4967 			 * since it will be negative incorrectly. We
4968 			 * expect to be reducing the RTT when we
4969 			 * pace at a slower rate.
4970 			 */
4971 			rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8);
4972 			rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8);
4973 		}
4974 	}
4975 	timely_says = rack_make_timely_judgement(rack,
4976 	    rack->r_ctl.rc_gp_srtt,
4977 	    rack->r_ctl.rc_rtt_diff,
4978 	    rack->r_ctl.rc_prev_gp_srtt
4979 	);
4980 	bytes_ps *= HPTS_USEC_IN_SEC;
4981 	bytes_ps /= utim;
4982 	if (bytes_ps > rack->r_ctl.last_max_bw) {
4983 		/*
4984 		 * Something is on path playing
4985 		 * since this b/w is not possible based
4986 		 * on our BDP (highest rwnd and lowest rtt
4987 		 * we saw in the measurement window).
4988 		 *
4989 		 * Another option here would be to
4990 		 * instead skip the measurement.
4991 		 */
4992 		rack_log_pacing_delay_calc(rack, bytes, reqbytes,
4993 					   bytes_ps, rack->r_ctl.last_max_bw, 0,
4994 					   11, __LINE__, NULL, quality);
4995 		bytes_ps = rack->r_ctl.last_max_bw;
4996 	}
4997 	/* We store gp for b/w in bytes per second */
4998 	if (rack->rc_gp_filled == 0) {
4999 		/* Initial measurement */
5000 		if (bytes_ps) {
5001 			rack->r_ctl.gp_bw = bytes_ps;
5002 			rack->rc_gp_filled = 1;
5003 			rack->r_ctl.num_measurements = 1;
5004 			rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
5005 		} else {
5006 			rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes,
5007 						   rack->r_ctl.rc_app_limited_cnt,
5008 						   0, 0, 10, __LINE__, NULL, quality);
5009 		}
5010 		if (tcp_in_hpts(rack->rc_tp) &&
5011 		    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
5012 			/*
5013 			 * Ok we can't trust the pacer in this case
5014 			 * where we transition from un-paced to paced.
5015 			 * Or for that matter when the burst mitigation
5016 			 * was making a wild guess and got it wrong.
5017 			 * Stop the pacer and clear up all the aggregate
5018 			 * delays etc.
5019 			 */
5020 			tcp_hpts_remove(rack->rc_tp);
5021 			rack->r_ctl.rc_hpts_flags = 0;
5022 			rack->r_ctl.rc_last_output_to = 0;
5023 		}
5024 		did_add = 2;
5025 	} else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) {
5026 		/* Still a small number run an average */
5027 		rack->r_ctl.gp_bw += bytes_ps;
5028 		addpart = rack->r_ctl.num_measurements;
5029 		rack->r_ctl.num_measurements++;
5030 		if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) {
5031 			/* We have collected enough to move forward */
5032 			rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements;
5033 		}
5034 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5035 		did_add = 3;
5036 	} else {
5037 		/*
5038 		 * We want to take 1/wma of the goodput and add in to 7/8th
5039 		 * of the old value weighted by the srtt. So if your measurement
5040 		 * period is say 2 SRTT's long you would get 1/4 as the
5041 		 * value, if it was like 1/2 SRTT then you would get 1/16th.
5042 		 *
5043 		 * But we must be careful not to take too much i.e. if the
5044 		 * srtt is say 20ms and the measurement is taken over
5045 		 * 400ms our weight would be 400/20 i.e. 20. On the
5046 		 * other hand if we get a measurement over 1ms with a
5047 		 * 10ms rtt we only want to take a much smaller portion.
5048 		 */
5049 		if (rack->r_ctl.num_measurements < 0xff) {
5050 			rack->r_ctl.num_measurements++;
5051 		}
5052 		srtt = (uint64_t)tp->t_srtt;
5053 		if (srtt == 0) {
5054 			/*
5055 			 * Strange why did t_srtt go back to zero?
5056 			 */
5057 			if (rack->r_ctl.rc_rack_min_rtt)
5058 				srtt = rack->r_ctl.rc_rack_min_rtt;
5059 			else
5060 				srtt = HPTS_USEC_IN_MSEC;
5061 		}
5062 		/*
5063 		 * XXXrrs: Note for reviewers, in playing with
5064 		 * dynamic pacing I discovered this GP calculation
5065 		 * as done originally leads to some undesired results.
5066 		 * Basically you can get longer measurements contributing
5067 		 * too much to the WMA. Thus I changed it if you are doing
5068 		 * dynamic adjustments to only do the aportioned adjustment
5069 		 * if we have a very small (time wise) measurement. Longer
5070 		 * measurements just get there weight (defaulting to 1/8)
5071 		 * add to the WMA. We may want to think about changing
5072 		 * this to always do that for both sides i.e. dynamic
5073 		 * and non-dynamic... but considering lots of folks
5074 		 * were playing with this I did not want to change the
5075 		 * calculation per.se. without your thoughts.. Lawerence?
5076 		 * Peter??
5077 		 */
5078 		if (rack->rc_gp_dyn_mul == 0) {
5079 			subpart = rack->r_ctl.gp_bw * utim;
5080 			subpart /= (srtt * 8);
5081 			if (subpart < (rack->r_ctl.gp_bw / 2)) {
5082 				/*
5083 				 * The b/w update takes no more
5084 				 * away then 1/2 our running total
5085 				 * so factor it in.
5086 				 */
5087 				addpart = bytes_ps * utim;
5088 				addpart /= (srtt * 8);
5089 			} else {
5090 				/*
5091 				 * Don't allow a single measurement
5092 				 * to account for more than 1/2 of the
5093 				 * WMA. This could happen on a retransmission
5094 				 * where utim becomes huge compared to
5095 				 * srtt (multiple retransmissions when using
5096 				 * the sending rate which factors in all the
5097 				 * transmissions from the first one).
5098 				 */
5099 				subpart = rack->r_ctl.gp_bw / 2;
5100 				addpart = bytes_ps / 2;
5101 			}
5102 			resid_bw = rack->r_ctl.gp_bw - subpart;
5103 			rack->r_ctl.gp_bw = resid_bw + addpart;
5104 			did_add = 1;
5105 		} else {
5106 			if ((utim / srtt) <= 1) {
5107 				/*
5108 				 * The b/w update was over a small period
5109 				 * of time. The idea here is to prevent a small
5110 				 * measurement time period from counting
5111 				 * too much. So we scale it based on the
5112 				 * time so it attributes less than 1/rack_wma_divisor
5113 				 * of its measurement.
5114 				 */
5115 				subpart = rack->r_ctl.gp_bw * utim;
5116 				subpart /= (srtt * rack_wma_divisor);
5117 				addpart = bytes_ps * utim;
5118 				addpart /= (srtt * rack_wma_divisor);
5119 			} else {
5120 				/*
5121 				 * The scaled measurement was long
5122 				 * enough so lets just add in the
5123 				 * portion of the measurement i.e. 1/rack_wma_divisor
5124 				 */
5125 				subpart = rack->r_ctl.gp_bw / rack_wma_divisor;
5126 				addpart = bytes_ps / rack_wma_divisor;
5127 			}
5128 			if ((rack->measure_saw_probe_rtt == 0) ||
5129 		            (bytes_ps > rack->r_ctl.gp_bw)) {
5130 				/*
5131 				 * For probe-rtt we only add it in
5132 				 * if its larger, all others we just
5133 				 * add in.
5134 				 */
5135 				did_add = 1;
5136 				resid_bw = rack->r_ctl.gp_bw - subpart;
5137 				rack->r_ctl.gp_bw = resid_bw + addpart;
5138 			}
5139 		}
5140 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
5141 	}
5142 	if ((rack->gp_ready == 0) &&
5143 	    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
5144 		/* We have enough measurements now */
5145 		rack->gp_ready = 1;
5146 		if (rack->dgp_on ||
5147 		    rack->rack_hibeta)
5148 			rack_set_cc_pacing(rack);
5149 		if (rack->defer_options)
5150 			rack_apply_deferred_options(rack);
5151 	}
5152 	rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim,
5153 				   rack_get_bw(rack), 22, did_add, NULL, quality);
5154 	/* We do not update any multipliers if we are in or have seen a probe-rtt */
5155 	if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set)
5156 		rack_update_multiplier(rack, timely_says, bytes_ps,
5157 				       rack->r_ctl.rc_gp_srtt,
5158 				       rack->r_ctl.rc_rtt_diff);
5159 	rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim,
5160 				   rack_get_bw(rack), 3, line, NULL, quality);
5161 	rack_log_pacing_delay_calc(rack,
5162 				   bytes, /* flex2 */
5163 				   tim, /* flex1 */
5164 				   bytes_ps, /* bw_inuse */
5165 				   rack->r_ctl.gp_bw, /* delRate */
5166 				   rack_get_lt_bw(rack), /* rttProp */
5167 				   20, line, NULL, 0);
5168 	/* reset the gp srtt and setup the new prev */
5169 	rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt;
5170 	/* Record the lost count for the next measurement */
5171 	rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count;
5172 skip_measurement:
5173 	/*
5174 	 * We restart our diffs based on the gpsrtt in the
5175 	 * measurement window.
5176 	 */
5177 	rack->rc_gp_rtt_set = 0;
5178 	rack->rc_gp_saw_rec = 0;
5179 	rack->rc_gp_saw_ca = 0;
5180 	rack->rc_gp_saw_ss = 0;
5181 	rack->rc_dragged_bottom = 0;
5182 
5183 	if (quality == RACK_QUALITY_HIGH) {
5184 		/*
5185 		 * Gput in the stats world is in kbps where bytes_ps is
5186 		 * bytes per second so we do ((x * 8)/ 1000).
5187 		 */
5188 		gput = (int32_t)((bytes_ps << 3) / (uint64_t)1000);
5189 #ifdef STATS
5190 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
5191 					 gput);
5192 		/*
5193 		 * XXXLAS: This is a temporary hack, and should be
5194 		 * chained off VOI_TCP_GPUT when stats(9) grows an
5195 		 * API to deal with chained VOIs.
5196 		 */
5197 		if (tp->t_stats_gput_prev > 0)
5198 			stats_voi_update_abs_s32(tp->t_stats,
5199 						 VOI_TCP_GPUT_ND,
5200 						 ((gput - tp->t_stats_gput_prev) * 100) /
5201 						 tp->t_stats_gput_prev);
5202 #endif
5203 		tp->t_stats_gput_prev = gput;
5204 	}
5205 	tp->t_flags &= ~TF_GPUTINPROG;
5206 	/*
5207 	 * Now are we app limited now and there is space from where we
5208 	 * were to where we want to go?
5209 	 *
5210 	 * We don't do the other case i.e. non-applimited here since
5211 	 * the next send will trigger us picking up the missing data.
5212 	 */
5213 	if (rack->r_ctl.rc_first_appl &&
5214 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
5215 	    rack->r_ctl.rc_app_limited_cnt &&
5216 	    (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) &&
5217 	    ((rack->r_ctl.rc_first_appl->r_end - th_ack) >
5218 	     max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) {
5219 		/*
5220 		 * Yep there is enough outstanding to make a measurement here.
5221 		 */
5222 		struct rack_sendmap *rsm;
5223 
5224 		rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
5225 		rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
5226 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
5227 		rack->app_limited_needs_set = 0;
5228 		tp->gput_seq = th_ack;
5229 		if (rack->in_probe_rtt)
5230 			rack->measure_saw_probe_rtt = 1;
5231 		else if ((rack->measure_saw_probe_rtt) &&
5232 			 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
5233 			rack->measure_saw_probe_rtt = 0;
5234 		if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) {
5235 			/* There is a full window to gain info from */
5236 			tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
5237 		} else {
5238 			/* We can only measure up to the applimited point */
5239 			tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack);
5240 			if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
5241 				/*
5242 				 * We don't have enough to make a measurement.
5243 				 */
5244 				tp->t_flags &= ~TF_GPUTINPROG;
5245 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
5246 							   0, 0, 0, 6, __LINE__, NULL, quality);
5247 				return;
5248 			}
5249 		}
5250 		if (tp->t_state >= TCPS_FIN_WAIT_1) {
5251 			/*
5252 			 * We will get no more data into the SB
5253 			 * this means we need to have the data available
5254 			 * before we start a measurement.
5255 			 */
5256 			if (sbavail(&tptosocket(tp)->so_snd) < (tp->gput_ack - tp->gput_seq)) {
5257 				/* Nope not enough data. */
5258 				return;
5259 			}
5260 		}
5261 		tp->t_flags |= TF_GPUTINPROG;
5262 		/*
5263 		 * Now we need to find the timestamp of the send at tp->gput_seq
5264 		 * for the send based measurement.
5265 		 */
5266 		rack->r_ctl.rc_gp_cumack_ts = 0;
5267 		rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
5268 		if (rsm) {
5269 			/* Ok send-based limit is set */
5270 			if (SEQ_LT(rsm->r_start, tp->gput_seq)) {
5271 				/*
5272 				 * Move back to include the earlier part
5273 				 * so our ack time lines up right (this may
5274 				 * make an overlapping measurement but thats
5275 				 * ok).
5276 				 */
5277 				tp->gput_seq = rsm->r_start;
5278 			}
5279 			if (rsm->r_flags & RACK_ACKED) {
5280 				struct rack_sendmap *nrsm;
5281 
5282 				tp->gput_ts = (uint32_t)rsm->r_ack_arrival;
5283 				tp->gput_seq = rsm->r_end;
5284 				nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
5285 				if (nrsm)
5286 					rsm = nrsm;
5287 				else {
5288 					rack->app_limited_needs_set = 1;
5289 				}
5290 			} else
5291 				rack->app_limited_needs_set = 1;
5292 			/* We always go from the first send */
5293 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
5294 		} else {
5295 			/*
5296 			 * If we don't find the rsm due to some
5297 			 * send-limit set the current time, which
5298 			 * basically disables the send-limit.
5299 			 */
5300 			struct timeval tv;
5301 
5302 			microuptime(&tv);
5303 			rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
5304 		}
5305 		rack_tend_gp_marks(tp, rack);
5306 		rack_log_pacing_delay_calc(rack,
5307 					   tp->gput_seq,
5308 					   tp->gput_ack,
5309 					   (uint64_t)rsm,
5310 					   tp->gput_ts,
5311 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
5312 					   9,
5313 					   __LINE__, rsm, quality);
5314 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
5315 	} else {
5316 		/*
5317 		 * To make sure proper timestamp merging occurs, we need to clear
5318 		 * all GP marks if we don't start a measurement.
5319 		 */
5320 		rack_clear_gp_marks(tp, rack);
5321 	}
5322 }
5323 
5324 /*
5325  * CC wrapper hook functions
5326  */
5327 static void
5328 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs,
5329     uint16_t type, int32_t recovery)
5330 {
5331 	uint32_t prior_cwnd, acked;
5332 	struct tcp_log_buffer *lgb = NULL;
5333 	uint8_t labc_to_use, quality;
5334 
5335 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5336 	tp->t_ccv.nsegs = nsegs;
5337 	acked = tp->t_ccv.bytes_this_ack = (th_ack - tp->snd_una);
5338 	if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
5339 		uint32_t max;
5340 
5341 		max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp);
5342 		if (tp->t_ccv.bytes_this_ack > max) {
5343 			tp->t_ccv.bytes_this_ack = max;
5344 		}
5345 	}
5346 #ifdef STATS
5347 	stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
5348 	    ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd);
5349 #endif
5350 	if ((th_ack == tp->snd_max) && rack->lt_bw_up) {
5351 		/* We will ack all, time
5352 		 * to end any lt_bw_up we
5353 		 * have running until something
5354 		 * new is sent.
5355 		 */
5356 		struct timeval tv;
5357 
5358 		rack->r_ctl.lt_bw_bytes += (tp->snd_max - rack->r_ctl.lt_seq);
5359 		rack->r_ctl.lt_seq = tp->snd_max;
5360 		(void)tcp_get_usecs(&tv);
5361 		rack->r_ctl.lt_bw_time += (tcp_tv_to_lusectick(&tv) - rack->r_ctl.lt_timemark);
5362 		rack->lt_bw_up = 0;
5363 	}
5364 	quality = RACK_QUALITY_NONE;
5365 	if ((tp->t_flags & TF_GPUTINPROG) &&
5366 	    rack_enough_for_measurement(tp, rack, th_ack, &quality)) {
5367 		/* Measure the Goodput */
5368 		rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality);
5369 	}
5370 	/* Which way our we limited, if not cwnd limited no advance in CA */
5371 	if (tp->snd_cwnd <= tp->snd_wnd)
5372 		tp->t_ccv.flags |= CCF_CWND_LIMITED;
5373 	else
5374 		tp->t_ccv.flags &= ~CCF_CWND_LIMITED;
5375 	if (tp->snd_cwnd > tp->snd_ssthresh) {
5376 		tp->t_bytes_acked += min(tp->t_ccv.bytes_this_ack,
5377 			 nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp));
5378 		/* For the setting of a window past use the actual scwnd we are using */
5379 		if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) {
5380 			tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use;
5381 			tp->t_ccv.flags |= CCF_ABC_SENTAWND;
5382 		}
5383 	} else {
5384 		tp->t_ccv.flags &= ~CCF_ABC_SENTAWND;
5385 		tp->t_bytes_acked = 0;
5386 	}
5387 	prior_cwnd = tp->snd_cwnd;
5388 	if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec ||
5389 	    (rack_client_low_buf && rack->client_bufferlvl &&
5390 	    (rack->client_bufferlvl < rack_client_low_buf)))
5391 		labc_to_use = rack->rc_labc;
5392 	else
5393 		labc_to_use = rack_max_abc_post_recovery;
5394 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5395 		union tcp_log_stackspecific log;
5396 		struct timeval tv;
5397 
5398 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5399 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5400 		log.u_bbr.flex1 = th_ack;
5401 		log.u_bbr.flex2 = tp->t_ccv.flags;
5402 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5403 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
5404 		log.u_bbr.flex5 = labc_to_use;
5405 		log.u_bbr.flex6 = prior_cwnd;
5406 		log.u_bbr.flex7 = V_tcp_do_newsack;
5407 		log.u_bbr.flex8 = 1;
5408 		lgb = tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5409 				     0, &log, false, NULL, __func__, __LINE__,&tv);
5410 	}
5411 	if (CC_ALGO(tp)->ack_received != NULL) {
5412 		/* XXXLAS: Find a way to live without this */
5413 		tp->t_ccv.curack = th_ack;
5414 		tp->t_ccv.labc = labc_to_use;
5415 		tp->t_ccv.flags |= CCF_USE_LOCAL_ABC;
5416 		CC_ALGO(tp)->ack_received(&tp->t_ccv, type);
5417 	}
5418 	if (lgb) {
5419 		lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd;
5420 	}
5421 	if (rack->r_must_retran) {
5422 		if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) {
5423 			/*
5424 			 * We now are beyond the rxt point so lets disable
5425 			 * the flag.
5426 			 */
5427 			rack->r_ctl.rc_out_at_rto = 0;
5428 			rack->r_must_retran = 0;
5429 		} else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) {
5430 			/*
5431 			 * Only decrement the rc_out_at_rto if the cwnd advances
5432 			 * at least a whole segment. Otherwise next time the peer
5433 			 * acks, we won't be able to send this generaly happens
5434 			 * when we are in Congestion Avoidance.
5435 			 */
5436 			if (acked <= rack->r_ctl.rc_out_at_rto){
5437 				rack->r_ctl.rc_out_at_rto -= acked;
5438 			} else {
5439 				rack->r_ctl.rc_out_at_rto = 0;
5440 			}
5441 		}
5442 	}
5443 #ifdef STATS
5444 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use);
5445 #endif
5446 	if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) {
5447 		rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use;
5448 	}
5449 }
5450 
5451 static void
5452 tcp_rack_partialack(struct tcpcb *tp)
5453 {
5454 	struct tcp_rack *rack;
5455 
5456 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5457 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5458 	/*
5459 	 * If we are doing PRR and have enough
5460 	 * room to send <or> we are pacing and prr
5461 	 * is disabled we will want to see if we
5462 	 * can send data (by setting r_wanted_output to
5463 	 * true).
5464 	 */
5465 	if ((rack->r_ctl.rc_prr_sndcnt > 0) ||
5466 	    rack->rack_no_prr)
5467 		rack->r_wanted_output = 1;
5468 }
5469 
5470 static inline void
5471 rack_set_most_aggr(struct tcp_rack *rack)
5472 {
5473 	rack->r_fill_less_agg = 0;
5474 	/* Once the cwnd as been clamped we don't do fill_cw */
5475 	if (rack->r_cwnd_was_clamped == 0)
5476 		rack->rc_pace_to_cwnd = 1;
5477 	rack->r_pacing_discount = 0;
5478 }
5479 
5480 static inline void
5481 rack_limit_fillcw(struct tcp_rack *rack)
5482 {
5483 	rack->r_fill_less_agg = 1;
5484 	/* Once the cwnd as been clamped we don't do fill_cw */
5485 	if (rack->r_cwnd_was_clamped == 0)
5486 		rack->rc_pace_to_cwnd = 1;
5487 	rack->r_pacing_discount = 0;
5488 }
5489 
5490 static inline void
5491 rack_disable_fillcw(struct tcp_rack *rack)
5492 {
5493 	rack->r_fill_less_agg = 1;
5494 	rack->rc_pace_to_cwnd = 0;
5495 	rack->r_pacing_discount = 0;
5496 }
5497 
5498 static void
5499 rack_client_buffer_level_set(struct tcp_rack *rack)
5500 {
5501 	/*
5502 	 * Only if DGP is on do we do anything that
5503 	 * changes stack behavior. If DGP is off all
5504 	 * we will do is issue a BB log (if BB logging is
5505 	 * on) and return.
5506 	 */
5507 	if (rack->dgp_on == 0) {
5508 		rack_log_pacing_delay_calc(rack, 0, rack->client_bufferlvl,
5509 					   0, 0, 0, 30, __LINE__, NULL, 0);
5510 		return;
5511 	}
5512 	if (IN_RECOVERY(rack->rc_tp->t_flags) && rack->r_ctl.full_dgp_in_rec) {
5513 		goto set_most_agg;
5514 	}
5515 	/*
5516 	 * We are in DGP so what setting should we
5517 	 * apply based on where the client is?
5518 	 */
5519 	switch(rack->r_ctl.rc_dgp_bl_agg) {
5520 	default:
5521 	case DGP_LEVEL0:
5522 set_most_agg:
5523 		rack_set_most_aggr(rack);
5524 		break;
5525 	case DGP_LEVEL1:
5526 		if (rack->client_bufferlvl == 4)
5527 			rack_limit_fillcw(rack);
5528 		else if (rack->client_bufferlvl == 5)
5529 			rack_disable_fillcw(rack);
5530 		else
5531 			rack_set_most_aggr(rack);
5532 		break;
5533 	case DGP_LEVEL2:
5534 		if (rack->client_bufferlvl == 3)
5535 			rack_limit_fillcw(rack);
5536 		else if (rack->client_bufferlvl == 4)
5537 			rack_disable_fillcw(rack);
5538 		else if (rack->client_bufferlvl == 5) {
5539 			rack_disable_fillcw(rack);
5540 			rack->r_pacing_discount = 1;
5541 			rack->r_ctl.pacing_discount_amm = 1;
5542 		} else
5543 			rack_set_most_aggr(rack);
5544 		break;
5545 	case DGP_LEVEL3:
5546 		if (rack->client_bufferlvl == 2)
5547 			rack_limit_fillcw(rack);
5548 		else if (rack->client_bufferlvl == 3)
5549 			rack_disable_fillcw(rack);
5550 		else if (rack->client_bufferlvl == 4) {
5551 			rack_disable_fillcw(rack);
5552 			rack->r_pacing_discount = 1;
5553 			rack->r_ctl.pacing_discount_amm = 1;
5554 		} else if (rack->client_bufferlvl == 5) {
5555 			rack_disable_fillcw(rack);
5556 			rack->r_pacing_discount = 1;
5557 			rack->r_ctl.pacing_discount_amm = 2;
5558 		} else
5559 			rack_set_most_aggr(rack);
5560 		break;
5561 	}
5562 	rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_dgp_bl_agg, rack->client_bufferlvl, 0,
5563 				   0, 0, 30, __LINE__, NULL, 0);
5564 }
5565 
5566 static void
5567 do_rack_check_for_unclamp(struct tcpcb *tp, struct tcp_rack *rack)
5568 {
5569 	/*
5570 	 * Can we unclamp. We unclamp if more than
5571 	 * N rounds have transpired with no loss.
5572 	 */
5573 	uint64_t snds, rxts, rxt_per;
5574 	uint32_t rnds;
5575 
5576 	rnds = rack->r_ctl.current_round - rack->r_ctl.last_rnd_rxt_clamped;
5577 	if ((rack_unclamp_round_thresh > 0) &&
5578 	    (rnds >= rack_unclamp_round_thresh)) {
5579 		snds = tp->t_sndbytes - rack->r_ctl.last_sndbytes;
5580 		KASSERT ((snds > 0), ("rack:%p tp:%p snds:%ju is 0", rack, tp,
5581 		    (uintmax_t)snds));
5582 		rxts = tp->t_snd_rxt_bytes - rack->r_ctl.last_snd_rxt_bytes;
5583 		rxt_per = rxts * 1000;
5584 		rxt_per /= snds;
5585 		if ((uint32_t)rxt_per <= rack_unclamp_rxt_thresh) {
5586 			/* Unclamp */
5587 			if (tcp_bblogging_on(rack->rc_tp)) {
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.timeStamp = tcp_get_usecs(&tv);
5593 				log.u_bbr.flex3 = rnds;
5594 				log.u_bbr.flex4 = rack_unclamp_round_thresh;
5595 				log.u_bbr.flex5 = (uint32_t)rxt_per;
5596 				log.u_bbr.flex8 = 6;
5597 				log.u_bbr.pkt_epoch = rack->r_ctl.rc_pace_max_segs;
5598 				log.u_bbr.bbr_state = rack->rc_pace_to_cwnd;
5599 				log.u_bbr.delivered = rack->r_ctl.num_of_clamps_applied;
5600 				log.u_bbr.applimited = rack->r_ctl.max_clamps;
5601 				log.u_bbr.epoch = rack->r_ctl.clamp_options;
5602 				log.u_bbr.cur_del_rate = rxts;
5603 				log.u_bbr.bw_inuse = rack_get_lt_bw(rack);
5604 				log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
5605 				log.u_bbr.lt_epoch = (uint32_t)((rack->r_ctl.gp_bw >> 32) & 0x00000000ffffffff);
5606 				log.u_bbr.pkts_out = (uint32_t)(rack->r_ctl.gp_bw & 0x00000000ffffffff);
5607 				tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5608 					      0, &log, false, NULL, NULL, 0, &tv);
5609 			}
5610 			rack->r_ctl.num_of_clamps_applied = 0;
5611 			rack->r_cwnd_was_clamped = 0;
5612 			rack->excess_rxt_on = 1;
5613 			if (rack->r_ctl.clamp_options) {
5614 				/*
5615 				 * We only allow fillcw to be toggled
5616 				 * if you are setting a max seg too.
5617 				 */
5618 				if (rack->r_ctl.clamp_options & 0x1) {
5619 					if ((rack->rc_pace_to_cwnd == 0) && (rack->dgp_on == 0)) {
5620 						/* turn on fill cw  for non-dgp*/
5621 						rack->rc_pace_to_cwnd = 0;
5622 					} else if ((rack->dgp_on == 1) && (rack->rc_pace_to_cwnd == 1)) {
5623 						/* For DGP we want it off */
5624 						rack->rc_pace_to_cwnd = 1;
5625 					}
5626 				}
5627 			}
5628 			if (rack->dgp_on) {
5629 				/* Reset all multipliers to 100.0 so just the measured bw */
5630 				/* Crash any per boosts down to 100% */
5631 				rack->r_ctl.rack_per_of_gp_rec = 100;
5632 				rack->r_ctl.rack_per_of_gp_ss = 100;
5633 				rack->r_ctl.rack_per_of_gp_ca = 100;
5634 				/* Set in an upper bound for ss/ca % increase */
5635 				rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
5636 				rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
5637 			}
5638 		}
5639 	}
5640 }
5641 
5642 static void
5643 do_rack_excess_rxt(struct tcpcb *tp, struct tcp_rack *rack)
5644 {
5645 	/*
5646 	 * Rack excess rxt accounting is turned on. If we
5647 	 * are above a threshold of rxt's in at least N
5648 	 * rounds, then back off the cwnd and ssthresh
5649 	 * to fit into the long-term b/w.
5650 	 */
5651 	uint64_t snds, rxts, rxt_per, lt_bw, bdp;
5652 	uint32_t rnds, new_cwnd, new_ssthresh, rtt, shared_cwnd_was_enabled = 0;
5653 
5654 	/* Is it shut off by 0 rounds? */
5655 	if (rack_rxt_min_rnds == 0)
5656 		return;
5657 	if ((rack->r_ctl.max_clamps > 0) &&
5658 	    (rack->r_ctl.num_of_clamps_applied >= rack->r_ctl.max_clamps)) {
5659 		/*
5660 		 * The idea, if max_clamps is set, is that if clamping it
5661 		 * N times did not work again, then there is no sense
5662 		 * clamping it again. The link is just a lossy link and
5663 		 * our clamps are doing no good. Turn it off so we don't come
5664 		 * back here again.
5665 		 */
5666 		rack->excess_rxt_on = 0;
5667 		rack->r_cwnd_was_clamped = 0;
5668 		rack->r_ctl.num_of_clamps_applied = 0;
5669 		return;
5670 	}
5671 	snds = tp->t_sndbytes - rack->r_ctl.last_sndbytes;
5672 	rxts = tp->t_snd_rxt_bytes - rack->r_ctl.last_snd_rxt_bytes;
5673 	rnds = rack->r_ctl.current_round - rack->r_ctl.last_rnd_rxt_clamped;
5674 	/* Has enough rounds progressed for us to re-measure? */
5675 	if ((rnds >= rack_rxt_min_rnds) &&
5676 	    (rack->r_ctl.rxt_threshold > 0)){
5677 		rxt_per = rxts * 1000;
5678 		rxt_per /= snds;
5679 		if (rxt_per >= rack->r_ctl.rxt_threshold) {
5680 			/*
5681 			 * Action required:
5682 			 *  We are above our excess retransmit level, lets
5683 			 *  cut down the cwnd and ssthresh to match the long-term
5684 			 *  b/w we are getting.
5685 			 */
5686 			/* First disable scwnd if enabled */
5687 #ifdef NETFLIX_SHARED_CWND
5688 			rack->rack_enable_scwnd = 0;
5689 			if (rack->r_ctl.rc_scw) {
5690 				uint32_t limit;
5691 
5692 				shared_cwnd_was_enabled = 1;
5693 				if (rack->r_limit_scw)
5694 					limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
5695 				else
5696 					limit = 0;
5697 				tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
5698 							  rack->r_ctl.rc_scw_index,
5699 							  limit);
5700 				rack->r_ctl.rc_scw = NULL;
5701 			}
5702 
5703 #endif
5704 			/* Calculate what the cwnd and ssthresh should be */
5705 			tcp_trace_point(rack->rc_tp, TCP_TP_EXCESS_RXT);
5706 			lt_bw = rack_get_lt_bw(rack);
5707 			if (lt_bw == 0) {
5708 				/*
5709 				 * No lt_bw, lets chop things to one MSS
5710 				 * and the ssthresh to the iwnd.
5711 				 */
5712 reset_to_iw:
5713 				new_cwnd = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
5714 				new_ssthresh = tcp_compute_initwnd(tcp_maxseg(tp));
5715 			} else {
5716 				rtt = rack->rc_rack_rtt;
5717 				if (rtt == 0) {
5718 					/* If we have no rack_rtt drop to the IW situation */
5719 					goto reset_to_iw;
5720 				}
5721 				bdp = lt_bw * (uint64_t)rtt;
5722 				bdp /= HPTS_USEC_IN_SEC;
5723 				new_cwnd = (uint32_t)bdp;
5724 				new_ssthresh = new_cwnd - 1;
5725 				if (new_cwnd < ctf_fixed_maxseg(tp)) {
5726 					/* Rock bottom, goto IW settings  */
5727 					goto reset_to_iw;
5728 				}
5729 			}
5730 			rack->r_cwnd_was_clamped = 1;
5731 			rack->r_ctl.num_of_clamps_applied++;
5732 			/* Reset the counter fromn now */
5733 			tp->t_bytes_acked = 0;
5734 			/*
5735 			 * Now what about options?
5736 			 * We look at the bottom  8 bits:
5737 			 * F = fill cw bit (toggle it if set)
5738 			 * S = Segment bits
5739 			 * M = set max segment bit
5740 			 *
5741 			 * SSSS SSMF
5742 			 */
5743 			if (rack->r_ctl.clamp_options) {
5744 				if (rack->r_ctl.clamp_options & 0x1) {
5745 					if ((rack->rc_pace_to_cwnd == 0) && (rack->dgp_on == 0)) {
5746 						/* turn on fill cw  for non-dgp*/
5747 						rack->rc_pace_to_cwnd = 1;
5748 					} else if ((rack->dgp_on == 1) && (rack->rc_pace_to_cwnd == 1)) {
5749 						/* For DGP we want it off */
5750 						rack->rc_pace_to_cwnd = 0;
5751 					}
5752 				}
5753 			}
5754 			if (rack->dgp_on) {
5755 				/* Reset all multipliers to 100.0 so just the measured bw */
5756 				/* Crash any per boosts down to 100% */
5757 				rack->r_ctl.rack_per_of_gp_rec = 100;
5758 				rack->r_ctl.rack_per_of_gp_ss = 100;
5759 				rack->r_ctl.rack_per_of_gp_ca = 100;
5760 				/* Set in an upper bound for ss/ca % increase */
5761 				rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_clamp_ss_upper;
5762 				rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_clamp_ca_upper;
5763 				/* Now move to the lt_bw */
5764 				rack->r_ctl.gp_bw = lt_bw;
5765 				rack->rc_gp_filled = 1;
5766 				rack->r_ctl.num_measurements = RACK_REQ_AVG;
5767 			}
5768 			if (tcp_bblogging_on(rack->rc_tp)) {
5769 				union tcp_log_stackspecific log;
5770 				struct timeval tv;
5771 
5772 				memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5773 				log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5774 				log.u_bbr.flex1 = new_cwnd;
5775 				log.u_bbr.flex2 = new_ssthresh;
5776 				log.u_bbr.flex3 = rnds;
5777 				log.u_bbr.flex4 = rack_rxt_min_rnds;
5778 				log.u_bbr.flex5 = rtt;
5779 				log.u_bbr.flex6 = shared_cwnd_was_enabled;
5780 				log.u_bbr.flex8 = 5;
5781 				log.u_bbr.pkt_epoch = rack->r_ctl.rc_pace_max_segs;
5782 				log.u_bbr.bbr_state = rack->rc_pace_to_cwnd;
5783 				log.u_bbr.delivered = rack->r_ctl.num_of_clamps_applied;
5784 				log.u_bbr.applimited = rack->r_ctl.max_clamps;
5785 				log.u_bbr.epoch = rack->r_ctl.clamp_options;
5786 				log.u_bbr.cur_del_rate = rxts;
5787 				log.u_bbr.delRate = snds;
5788 				log.u_bbr.rttProp = rack->r_ctl.rxt_threshold;
5789 				log.u_bbr.bw_inuse = lt_bw;
5790 				log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
5791 				log.u_bbr.lt_epoch = (uint32_t)((rack->r_ctl.gp_bw >> 32) & 0x00000000ffffffff);
5792 				log.u_bbr.pkts_out = (uint32_t)(rack->r_ctl.gp_bw & 0x00000000ffffffff);
5793 				tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5794 					       0, &log, false, NULL, NULL, 0, &tv);
5795 			}
5796 			/* Update our point where we did it */
5797 			if (rack->r_ctl.already_had_a_excess == 0) {
5798 				rack->r_ctl.already_had_a_excess = 1;
5799 				counter_u64_add(rack_rxt_clamps_cwnd_uniq, 1);
5800 			}
5801 			counter_u64_add(rack_rxt_clamps_cwnd, 1);
5802 			rack->r_ctl.last_sndbytes = tp->t_sndbytes;
5803 			rack->r_ctl.last_snd_rxt_bytes = tp->t_snd_rxt_bytes;
5804 			rack->r_ctl.last_rnd_rxt_clamped = rack->r_ctl.current_round;
5805 			if (new_cwnd < tp->snd_cwnd)
5806 				tp->snd_cwnd = new_cwnd;
5807 			if (new_ssthresh < tp->snd_ssthresh)
5808 				tp->snd_ssthresh = new_ssthresh;
5809 		}
5810 	}
5811 }
5812 
5813 static void
5814 rack_post_recovery(struct tcpcb *tp, uint32_t th_ack)
5815 {
5816 	struct tcp_rack *rack;
5817 	uint32_t orig_cwnd;
5818 
5819 	orig_cwnd = tp->snd_cwnd;
5820 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5821 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5822 	/* only alert CC if we alerted when we entered */
5823 	if (CC_ALGO(tp)->post_recovery != NULL) {
5824 		tp->t_ccv.curack = th_ack;
5825 		CC_ALGO(tp)->post_recovery(&tp->t_ccv);
5826 		if (tp->snd_cwnd < tp->snd_ssthresh) {
5827 			/*
5828 			 * Rack has burst control and pacing
5829 			 * so lets not set this any lower than
5830 			 * snd_ssthresh per RFC-6582 (option 2).
5831 			 */
5832 			tp->snd_cwnd = tp->snd_ssthresh;
5833 		}
5834 	}
5835 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
5836 		union tcp_log_stackspecific log;
5837 		struct timeval tv;
5838 
5839 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
5840 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
5841 		log.u_bbr.flex1 = th_ack;
5842 		log.u_bbr.flex2 = tp->t_ccv.flags;
5843 		log.u_bbr.flex3 = tp->t_ccv.bytes_this_ack;
5844 		log.u_bbr.flex4 = tp->t_ccv.nsegs;
5845 		log.u_bbr.flex5 = V_tcp_abc_l_var;
5846 		log.u_bbr.flex6 = orig_cwnd;
5847 		log.u_bbr.flex7 = V_tcp_do_newsack;
5848 		log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt;
5849 		log.u_bbr.flex8 = 2;
5850 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0,
5851 			       0, &log, false, NULL, __func__, __LINE__, &tv);
5852 	}
5853 	if ((rack->rack_no_prr == 0) &&
5854 	    (rack->no_prr_addback == 0) &&
5855 	    (rack->r_ctl.rc_prr_sndcnt > 0)) {
5856 		/*
5857 		 * Suck the next prr cnt back into cwnd, but
5858 		 * only do that if we are not application limited.
5859 		 */
5860 		if (ctf_outstanding(tp) <= sbavail(&tptosocket(tp)->so_snd)) {
5861 			/*
5862 			 * We are allowed to add back to the cwnd the amount we did
5863 			 * not get out if:
5864 			 * a) no_prr_addback is off.
5865 			 * b) we are not app limited
5866 			 * c) we are doing prr
5867 			 * <and>
5868 			 * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none).
5869 			 */
5870 			tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax),
5871 					    rack->r_ctl.rc_prr_sndcnt);
5872 		}
5873 		rack->r_ctl.rc_prr_sndcnt = 0;
5874 		rack_log_to_prr(rack, 1, 0, __LINE__);
5875 	}
5876 	rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
5877 	tp->snd_recover = tp->snd_una;
5878 	if (rack->r_ctl.dsack_persist) {
5879 		rack->r_ctl.dsack_persist--;
5880 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
5881 			rack->r_ctl.num_dsack = 0;
5882 		}
5883 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
5884 	}
5885 	EXIT_RECOVERY(tp->t_flags);
5886 	if (rack->r_ctl.full_dgp_in_rec)
5887 		rack_client_buffer_level_set(rack);
5888 }
5889 
5890 static void
5891 rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line)
5892 {
5893 	struct tcp_rack *rack;
5894 	uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd;
5895 
5896 	INP_WLOCK_ASSERT(tptoinpcb(tp));
5897 #ifdef STATS
5898 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type);
5899 #endif
5900 	if (IN_RECOVERY(tp->t_flags) == 0) {
5901 		in_rec_at_entry = 0;
5902 		ssthresh_enter = tp->snd_ssthresh;
5903 		cwnd_enter = tp->snd_cwnd;
5904 	} else
5905 		in_rec_at_entry = 1;
5906 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5907 	switch (type) {
5908 	case CC_NDUPACK:
5909 		tp->t_flags &= ~TF_WASFRECOVERY;
5910 		tp->t_flags &= ~TF_WASCRECOVERY;
5911 		if (!IN_FASTRECOVERY(tp->t_flags)) {
5912 			if (rack->dgp_on && rack->r_cwnd_was_clamped) {
5913 				/* Reset the gains so that on exit we will be softer longer */
5914 				rack->r_ctl.rack_per_of_gp_rec = 100;
5915 				rack->r_ctl.rack_per_of_gp_ss = 98;
5916 				rack->r_ctl.rack_per_of_gp_ca = 98;
5917 			}
5918 			rack->r_ctl.rc_prr_delivered = 0;
5919 			rack->r_ctl.rc_prr_out = 0;
5920 			rack->r_fast_output = 0;
5921 			if (rack->rack_no_prr == 0) {
5922 				rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
5923 				rack_log_to_prr(rack, 2, in_rec_at_entry, line);
5924 			}
5925 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
5926 			tp->snd_recover = tp->snd_max;
5927 			if (tp->t_flags2 & TF2_ECN_PERMIT)
5928 				tp->t_flags2 |= TF2_ECN_SND_CWR;
5929 		}
5930 		break;
5931 	case CC_ECN:
5932 		if (!IN_CONGRECOVERY(tp->t_flags) ||
5933 		    /*
5934 		     * Allow ECN reaction on ACK to CWR, if
5935 		     * that data segment was also CE marked.
5936 		     */
5937 		    SEQ_GEQ(ack, tp->snd_recover)) {
5938 			EXIT_CONGRECOVERY(tp->t_flags);
5939 			KMOD_TCPSTAT_INC(tcps_ecn_rcwnd);
5940 			rack->r_fast_output = 0;
5941 			tp->snd_recover = tp->snd_max + 1;
5942 			if (tp->t_flags2 & TF2_ECN_PERMIT)
5943 				tp->t_flags2 |= TF2_ECN_SND_CWR;
5944 		}
5945 		break;
5946 	case CC_RTO:
5947 		tp->t_dupacks = 0;
5948 		tp->t_bytes_acked = 0;
5949 		rack->r_fast_output = 0;
5950 		EXIT_RECOVERY(tp->t_flags);
5951 		if (tp->t_rxtshift == 1) {
5952 			tp->snd_ssthresh = max(2,
5953 			    min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 /
5954 			    ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp);
5955 		}
5956 		orig_cwnd = tp->snd_cwnd;
5957 		tp->snd_cwnd = ctf_fixed_maxseg(tp);
5958 		rack_log_to_prr(rack, 16, orig_cwnd, line);
5959 		if (tp->t_flags2 & TF2_ECN_PERMIT)
5960 			tp->t_flags2 |= TF2_ECN_SND_CWR;
5961 		break;
5962 	case CC_RTO_ERR:
5963 		KMOD_TCPSTAT_INC(tcps_sndrexmitbad);
5964 		/* RTO was unnecessary, so reset everything. */
5965 		tp->snd_cwnd = tp->snd_cwnd_prev;
5966 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
5967 		tp->snd_recover = tp->snd_recover_prev;
5968 		if (tp->t_flags & TF_WASFRECOVERY) {
5969 			ENTER_FASTRECOVERY(tp->t_flags);
5970 			tp->t_flags &= ~TF_WASFRECOVERY;
5971 		}
5972 		if (tp->t_flags & TF_WASCRECOVERY) {
5973 			ENTER_CONGRECOVERY(tp->t_flags);
5974 			tp->t_flags &= ~TF_WASCRECOVERY;
5975 		}
5976 		tp->snd_nxt = tp->snd_max;
5977 		tp->t_badrxtwin = 0;
5978 		break;
5979 	}
5980 	if ((CC_ALGO(tp)->cong_signal != NULL)  &&
5981 	    (type != CC_RTO)){
5982 		tp->t_ccv.curack = ack;
5983 		CC_ALGO(tp)->cong_signal(&tp->t_ccv, type);
5984 	}
5985 	if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) {
5986 		rack_log_to_prr(rack, 15, cwnd_enter, line);
5987 		if (rack->r_ctl.full_dgp_in_rec)
5988 			rack_client_buffer_level_set(rack);
5989 		rack->r_ctl.dsack_byte_cnt = 0;
5990 		rack->r_ctl.retran_during_recovery = 0;
5991 		rack->r_ctl.rc_cwnd_at_erec = cwnd_enter;
5992 		rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter;
5993 		rack->r_ent_rec_ns = 1;
5994 	}
5995 }
5996 
5997 static inline void
5998 rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp)
5999 {
6000 	uint32_t i_cwnd;
6001 
6002 	INP_WLOCK_ASSERT(tptoinpcb(tp));
6003 
6004 	if (CC_ALGO(tp)->after_idle != NULL)
6005 		CC_ALGO(tp)->after_idle(&tp->t_ccv);
6006 
6007 	if (tp->snd_cwnd == 1)
6008 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
6009 	else
6010 		i_cwnd = rc_init_window(rack);
6011 
6012 	/*
6013 	 * Being idle is no different than the initial window. If the cc
6014 	 * clamps it down below the initial window raise it to the initial
6015 	 * window.
6016 	 */
6017 	if (tp->snd_cwnd < i_cwnd) {
6018 		tp->snd_cwnd = i_cwnd;
6019 	}
6020 }
6021 
6022 /*
6023  * Indicate whether this ack should be delayed.  We can delay the ack if
6024  * following conditions are met:
6025  *	- There is no delayed ack timer in progress.
6026  *	- Our last ack wasn't a 0-sized window. We never want to delay
6027  *	  the ack that opens up a 0-sized window.
6028  *	- LRO wasn't used for this segment. We make sure by checking that the
6029  *	  segment size is not larger than the MSS.
6030  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
6031  *	  connection.
6032  */
6033 #define DELAY_ACK(tp, tlen)			 \
6034 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
6035 	((tp->t_flags & TF_DELACK) == 0) &&	 \
6036 	(tlen <= tp->t_maxseg) &&		 \
6037 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
6038 
6039 static struct rack_sendmap *
6040 rack_find_lowest_rsm(struct tcp_rack *rack)
6041 {
6042 	struct rack_sendmap *rsm;
6043 
6044 	/*
6045 	 * Walk the time-order transmitted list looking for an rsm that is
6046 	 * not acked. This will be the one that was sent the longest time
6047 	 * ago that is still outstanding.
6048 	 */
6049 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
6050 		if (rsm->r_flags & RACK_ACKED) {
6051 			continue;
6052 		}
6053 		goto finish;
6054 	}
6055 finish:
6056 	return (rsm);
6057 }
6058 
6059 static struct rack_sendmap *
6060 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
6061 {
6062 	struct rack_sendmap *prsm;
6063 
6064 	/*
6065 	 * Walk the sequence order list backward until we hit and arrive at
6066 	 * the highest seq not acked. In theory when this is called it
6067 	 * should be the last segment (which it was not).
6068 	 */
6069 	prsm = rsm;
6070 
6071 	TQHASH_FOREACH_REVERSE_FROM(prsm, rack->r_ctl.tqh) {
6072 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
6073 			continue;
6074 		}
6075 		return (prsm);
6076 	}
6077 	return (NULL);
6078 }
6079 
6080 static uint32_t
6081 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
6082 {
6083 	int32_t lro;
6084 	uint32_t thresh;
6085 
6086 	/*
6087 	 * lro is the flag we use to determine if we have seen reordering.
6088 	 * If it gets set we have seen reordering. The reorder logic either
6089 	 * works in one of two ways:
6090 	 *
6091 	 * If reorder-fade is configured, then we track the last time we saw
6092 	 * re-ordering occur. If we reach the point where enough time as
6093 	 * passed we no longer consider reordering has occuring.
6094 	 *
6095 	 * Or if reorder-face is 0, then once we see reordering we consider
6096 	 * the connection to alway be subject to reordering and just set lro
6097 	 * to 1.
6098 	 *
6099 	 * In the end if lro is non-zero we add the extra time for
6100 	 * reordering in.
6101 	 */
6102 	if (srtt == 0)
6103 		srtt = 1;
6104 	if (rack->r_ctl.rc_reorder_ts) {
6105 		if (rack->r_ctl.rc_reorder_fade) {
6106 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
6107 				lro = cts - rack->r_ctl.rc_reorder_ts;
6108 				if (lro == 0) {
6109 					/*
6110 					 * No time as passed since the last
6111 					 * reorder, mark it as reordering.
6112 					 */
6113 					lro = 1;
6114 				}
6115 			} else {
6116 				/* Negative time? */
6117 				lro = 0;
6118 			}
6119 			if (lro > rack->r_ctl.rc_reorder_fade) {
6120 				/* Turn off reordering seen too */
6121 				rack->r_ctl.rc_reorder_ts = 0;
6122 				lro = 0;
6123 			}
6124 		} else {
6125 			/* Reodering does not fade */
6126 			lro = 1;
6127 		}
6128 	} else {
6129 		lro = 0;
6130 	}
6131 	if (rack->rc_rack_tmr_std_based == 0) {
6132 		thresh = srtt + rack->r_ctl.rc_pkt_delay;
6133 	} else {
6134 		/* Standards based pkt-delay is 1/4 srtt */
6135 		thresh = srtt +  (srtt >> 2);
6136 	}
6137 	if (lro && (rack->rc_rack_tmr_std_based == 0)) {
6138 		/* It must be set, if not you get 1/4 rtt */
6139 		if (rack->r_ctl.rc_reorder_shift)
6140 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
6141 		else
6142 			thresh += (srtt >> 2);
6143 	}
6144 	if (rack->rc_rack_use_dsack &&
6145 	    lro &&
6146 	    (rack->r_ctl.num_dsack > 0)) {
6147 		/*
6148 		 * We only increase the reordering window if we
6149 		 * have seen reordering <and> we have a DSACK count.
6150 		 */
6151 		thresh += rack->r_ctl.num_dsack * (srtt >> 2);
6152 		rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh);
6153 	}
6154 	/* SRTT * 2 is the ceiling */
6155 	if (thresh > (srtt * 2)) {
6156 		thresh = srtt * 2;
6157 	}
6158 	/* And we don't want it above the RTO max either */
6159 	if (thresh > rack_rto_max) {
6160 		thresh = rack_rto_max;
6161 	}
6162 	rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh);
6163 	return (thresh);
6164 }
6165 
6166 static uint32_t
6167 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
6168 		     struct rack_sendmap *rsm, uint32_t srtt)
6169 {
6170 	struct rack_sendmap *prsm;
6171 	uint32_t thresh, len;
6172 	int segsiz;
6173 
6174 	if (srtt == 0)
6175 		srtt = 1;
6176 	if (rack->r_ctl.rc_tlp_threshold)
6177 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
6178 	else
6179 		thresh = (srtt * 2);
6180 
6181 	/* Get the previous sent packet, if any */
6182 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
6183 	len = rsm->r_end - rsm->r_start;
6184 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
6185 		/* Exactly like the ID */
6186 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) {
6187 			uint32_t alt_thresh;
6188 			/*
6189 			 * Compensate for delayed-ack with the d-ack time.
6190 			 */
6191 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6192 			if (alt_thresh > thresh)
6193 				thresh = alt_thresh;
6194 		}
6195 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
6196 		/* 2.1 behavior */
6197 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
6198 		if (prsm && (len <= segsiz)) {
6199 			/*
6200 			 * Two packets outstanding, thresh should be (2*srtt) +
6201 			 * possible inter-packet delay (if any).
6202 			 */
6203 			uint32_t inter_gap = 0;
6204 			int idx, nidx;
6205 
6206 			idx = rsm->r_rtr_cnt - 1;
6207 			nidx = prsm->r_rtr_cnt - 1;
6208 			if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) {
6209 				/* Yes it was sent later (or at the same time) */
6210 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
6211 			}
6212 			thresh += inter_gap;
6213 		} else if (len <= segsiz) {
6214 			/*
6215 			 * Possibly compensate for delayed-ack.
6216 			 */
6217 			uint32_t alt_thresh;
6218 
6219 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6220 			if (alt_thresh > thresh)
6221 				thresh = alt_thresh;
6222 		}
6223 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
6224 		/* 2.2 behavior */
6225 		if (len <= segsiz) {
6226 			uint32_t alt_thresh;
6227 			/*
6228 			 * Compensate for delayed-ack with the d-ack time.
6229 			 */
6230 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
6231 			if (alt_thresh > thresh)
6232 				thresh = alt_thresh;
6233 		}
6234 	}
6235 	/* Not above an RTO */
6236 	if (thresh > tp->t_rxtcur) {
6237 		thresh = tp->t_rxtcur;
6238 	}
6239 	/* Not above a RTO max */
6240 	if (thresh > rack_rto_max) {
6241 		thresh = rack_rto_max;
6242 	}
6243 	/* Apply user supplied min TLP */
6244 	if (thresh < rack_tlp_min) {
6245 		thresh = rack_tlp_min;
6246 	}
6247 	return (thresh);
6248 }
6249 
6250 static uint32_t
6251 rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack)
6252 {
6253 	/*
6254 	 * We want the rack_rtt which is the
6255 	 * last rtt we measured. However if that
6256 	 * does not exist we fallback to the srtt (which
6257 	 * we probably will never do) and then as a last
6258 	 * resort we use RACK_INITIAL_RTO if no srtt is
6259 	 * yet set.
6260 	 */
6261 	if (rack->rc_rack_rtt)
6262 		return (rack->rc_rack_rtt);
6263 	else if (tp->t_srtt == 0)
6264 		return (RACK_INITIAL_RTO);
6265 	return (tp->t_srtt);
6266 }
6267 
6268 static struct rack_sendmap *
6269 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
6270 {
6271 	/*
6272 	 * Check to see that we don't need to fall into recovery. We will
6273 	 * need to do so if our oldest transmit is past the time we should
6274 	 * have had an ack.
6275 	 */
6276 	struct tcp_rack *rack;
6277 	struct rack_sendmap *rsm;
6278 	int32_t idx;
6279 	uint32_t srtt, thresh;
6280 
6281 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6282 	if (tqhash_empty(rack->r_ctl.tqh)) {
6283 		return (NULL);
6284 	}
6285 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6286 	if (rsm == NULL)
6287 		return (NULL);
6288 
6289 
6290 	if (rsm->r_flags & RACK_ACKED) {
6291 		rsm = rack_find_lowest_rsm(rack);
6292 		if (rsm == NULL)
6293 			return (NULL);
6294 	}
6295 	idx = rsm->r_rtr_cnt - 1;
6296 	srtt = rack_grab_rtt(tp, rack);
6297 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
6298 	if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) {
6299 		return (NULL);
6300 	}
6301 	if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) {
6302 		return (NULL);
6303 	}
6304 	/* Ok if we reach here we are over-due and this guy can be sent */
6305 	rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
6306 	return (rsm);
6307 }
6308 
6309 static uint32_t
6310 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
6311 {
6312 	int32_t t;
6313 	int32_t tt;
6314 	uint32_t ret_val;
6315 
6316 	t = (tp->t_srtt + (tp->t_rttvar << 2));
6317 	RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
6318  	    rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop);
6319 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
6320 	ret_val = (uint32_t)tt;
6321 	return (ret_val);
6322 }
6323 
6324 static uint32_t
6325 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack)
6326 {
6327 	/*
6328 	 * Start the FR timer, we do this based on getting the first one in
6329 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
6330 	 * events we need to stop the running timer (if its running) before
6331 	 * starting the new one.
6332 	 */
6333 	uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse;
6334 	uint32_t srtt_cur;
6335 	int32_t idx;
6336 	int32_t is_tlp_timer = 0;
6337 	struct rack_sendmap *rsm;
6338 
6339 	if (rack->t_timers_stopped) {
6340 		/* All timers have been stopped none are to run */
6341 		return (0);
6342 	}
6343 	if (rack->rc_in_persist) {
6344 		/* We can't start any timer in persists */
6345 		return (rack_get_persists_timer_val(tp, rack));
6346 	}
6347 	rack->rc_on_min_to = 0;
6348 	if ((tp->t_state < TCPS_ESTABLISHED) ||
6349 	    (rack->sack_attack_disable > 0) ||
6350 	    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
6351 		goto activate_rxt;
6352 	}
6353 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6354 	if ((rsm == NULL) || sup_rack) {
6355 		/* Nothing on the send map or no rack */
6356 activate_rxt:
6357 		time_since_sent = 0;
6358 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6359 		if (rsm) {
6360 			/*
6361 			 * Should we discount the RTX timer any?
6362 			 *
6363 			 * We want to discount it the smallest amount.
6364 			 * If a timer (Rack/TLP or RXT) has gone off more
6365 			 * recently thats the discount we want to use (now - timer time).
6366 			 * If the retransmit of the oldest packet was more recent then
6367 			 * we want to use that (now - oldest-packet-last_transmit_time).
6368 			 *
6369 			 */
6370 			idx = rsm->r_rtr_cnt - 1;
6371 			if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx])))
6372 				tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6373 			else
6374 				tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6375 			if (TSTMP_GT(cts, tstmp_touse))
6376 			    time_since_sent = cts - tstmp_touse;
6377 		}
6378 		if (SEQ_LT(tp->snd_una, tp->snd_max) ||
6379 		    sbavail(&tptosocket(tp)->so_snd)) {
6380 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
6381 			to = tp->t_rxtcur;
6382 			if (to > time_since_sent)
6383 				to -= time_since_sent;
6384 			else
6385 				to = rack->r_ctl.rc_min_to;
6386 			if (to == 0)
6387 				to = 1;
6388 			/* Special case for KEEPINIT */
6389 			if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
6390 			    (TP_KEEPINIT(tp) != 0) &&
6391 			    rsm) {
6392 				/*
6393 				 * We have to put a ceiling on the rxt timer
6394 				 * of the keep-init timeout.
6395 				 */
6396 				uint32_t max_time, red;
6397 
6398 				max_time = TICKS_2_USEC(TP_KEEPINIT(tp));
6399 				if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) {
6400 					red = (cts - (uint32_t)rsm->r_tim_lastsent[0]);
6401 					if (red < max_time)
6402 						max_time -= red;
6403 					else
6404 						max_time = 1;
6405 				}
6406 				/* Reduce timeout to the keep value if needed */
6407 				if (max_time < to)
6408 					to = max_time;
6409 			}
6410 			return (to);
6411 		}
6412 		return (0);
6413 	}
6414 	if (rsm->r_flags & RACK_ACKED) {
6415 		rsm = rack_find_lowest_rsm(rack);
6416 		if (rsm == NULL) {
6417 			/* No lowest? */
6418 			goto activate_rxt;
6419 		}
6420 	}
6421 	if (rack->sack_attack_disable) {
6422 		/*
6423 		 * We don't want to do
6424 		 * any TLP's if you are an attacker.
6425 		 * Though if you are doing what
6426 		 * is expected you may still have
6427 		 * SACK-PASSED marks.
6428 		 */
6429 		goto activate_rxt;
6430 	}
6431 	/* Convert from ms to usecs */
6432 	if ((rsm->r_flags & RACK_SACK_PASSED) ||
6433 	    (rsm->r_flags & RACK_RWND_COLLAPSED) ||
6434 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
6435 		if ((tp->t_flags & TF_SENTFIN) &&
6436 		    ((tp->snd_max - tp->snd_una) == 1) &&
6437 		    (rsm->r_flags & RACK_HAS_FIN)) {
6438 			/*
6439 			 * We don't start a rack timer if all we have is a
6440 			 * FIN outstanding.
6441 			 */
6442 			goto activate_rxt;
6443 		}
6444 		if ((rack->use_rack_rr == 0) &&
6445 		    (IN_FASTRECOVERY(tp->t_flags)) &&
6446 		    (rack->rack_no_prr == 0) &&
6447 		     (rack->r_ctl.rc_prr_sndcnt  < ctf_fixed_maxseg(tp))) {
6448 			/*
6449 			 * We are not cheating, in recovery  and
6450 			 * not enough ack's to yet get our next
6451 			 * retransmission out.
6452 			 *
6453 			 * Note that classified attackers do not
6454 			 * get to use the rack-cheat.
6455 			 */
6456 			goto activate_tlp;
6457 		}
6458 		srtt = rack_grab_rtt(tp, rack);
6459 		thresh = rack_calc_thresh_rack(rack, srtt, cts);
6460 		idx = rsm->r_rtr_cnt - 1;
6461 		exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh;
6462 		if (SEQ_GEQ(exp, cts)) {
6463 			to = exp - cts;
6464 			if (to < rack->r_ctl.rc_min_to) {
6465 				to = rack->r_ctl.rc_min_to;
6466 				if (rack->r_rr_config == 3)
6467 					rack->rc_on_min_to = 1;
6468 			}
6469 		} else {
6470 			to = rack->r_ctl.rc_min_to;
6471 			if (rack->r_rr_config == 3)
6472 				rack->rc_on_min_to = 1;
6473 		}
6474 	} else {
6475 		/* Ok we need to do a TLP not RACK */
6476 activate_tlp:
6477 		if ((rack->rc_tlp_in_progress != 0) &&
6478 		    (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) {
6479 			/*
6480 			 * The previous send was a TLP and we have sent
6481 			 * N TLP's without sending new data.
6482 			 */
6483 			goto activate_rxt;
6484 		}
6485 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
6486 		if (rsm == NULL) {
6487 			/* We found no rsm to TLP with. */
6488 			goto activate_rxt;
6489 		}
6490 		if (rsm->r_flags & RACK_HAS_FIN) {
6491 			/* If its a FIN we dont do TLP */
6492 			rsm = NULL;
6493 			goto activate_rxt;
6494 		}
6495 		idx = rsm->r_rtr_cnt - 1;
6496 		time_since_sent = 0;
6497 		if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time))
6498 			tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx];
6499 		else
6500 			tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time;
6501 		if (TSTMP_GT(cts, tstmp_touse))
6502 		    time_since_sent = cts - tstmp_touse;
6503 		is_tlp_timer = 1;
6504 		if (tp->t_srtt) {
6505 			if ((rack->rc_srtt_measure_made == 0) &&
6506 			    (tp->t_srtt == 1)) {
6507 				/*
6508 				 * If another stack as run and set srtt to 1,
6509 				 * then the srtt was 0, so lets use the initial.
6510 				 */
6511 				srtt = RACK_INITIAL_RTO;
6512 			} else {
6513 				srtt_cur = tp->t_srtt;
6514 				srtt = srtt_cur;
6515 			}
6516 		} else
6517 			srtt = RACK_INITIAL_RTO;
6518 		/*
6519 		 * If the SRTT is not keeping up and the
6520 		 * rack RTT has spiked we want to use
6521 		 * the last RTT not the smoothed one.
6522 		 */
6523 		if (rack_tlp_use_greater &&
6524 		    tp->t_srtt &&
6525 		    (srtt < rack_grab_rtt(tp, rack))) {
6526 			srtt = rack_grab_rtt(tp, rack);
6527 		}
6528 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
6529 		if (thresh > time_since_sent) {
6530 			to = thresh - time_since_sent;
6531 		} else {
6532 			to = rack->r_ctl.rc_min_to;
6533 			rack_log_alt_to_to_cancel(rack,
6534 						  thresh,		/* flex1 */
6535 						  time_since_sent,	/* flex2 */
6536 						  tstmp_touse,		/* flex3 */
6537 						  rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */
6538 						  (uint32_t)rsm->r_tim_lastsent[idx],
6539 						  srtt,
6540 						  idx, 99);
6541 		}
6542 		if (to < rack_tlp_min) {
6543 			to = rack_tlp_min;
6544 		}
6545 		if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) {
6546 			/*
6547 			 * If the TLP time works out to larger than the max
6548 			 * RTO lets not do TLP.. just RTO.
6549 			 */
6550 			goto activate_rxt;
6551 		}
6552 	}
6553 	if (is_tlp_timer == 0) {
6554 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
6555 	} else {
6556 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
6557 	}
6558 	if (to == 0)
6559 		to = 1;
6560 	return (to);
6561 }
6562 
6563 static void
6564 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, tcp_seq snd_una)
6565 {
6566 	struct timeval tv;
6567 
6568 	if (rack->rc_in_persist == 0) {
6569 		if (tp->t_flags & TF_GPUTINPROG) {
6570 			/*
6571 			 * Stop the goodput now, the calling of the
6572 			 * measurement function clears the flag.
6573 			 */
6574 			rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__,
6575 						    RACK_QUALITY_PERSIST);
6576 		}
6577 #ifdef NETFLIX_SHARED_CWND
6578 		if (rack->r_ctl.rc_scw) {
6579 			tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6580 			rack->rack_scwnd_is_idle = 1;
6581 		}
6582 #endif
6583 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(&tv);
6584 		if (rack->lt_bw_up) {
6585 			/* Suspend our LT BW measurement */
6586 			uint64_t tmark;
6587 
6588 			rack->r_ctl.lt_bw_bytes += (snd_una - rack->r_ctl.lt_seq);
6589 			rack->r_ctl.lt_seq = snd_una;
6590 			tmark = tcp_tv_to_lusectick(&tv);
6591 			rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
6592 			rack->r_ctl.lt_timemark = tmark;
6593 			rack->lt_bw_up = 0;
6594 			rack->r_persist_lt_bw_off = 1;
6595 		}
6596 		if (rack->r_ctl.rc_went_idle_time == 0)
6597 			rack->r_ctl.rc_went_idle_time = 1;
6598 		rack_timer_cancel(tp, rack, cts, __LINE__);
6599 		rack->r_ctl.persist_lost_ends = 0;
6600 		rack->probe_not_answered = 0;
6601 		rack->forced_ack = 0;
6602 		tp->t_rxtshift = 0;
6603 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6604 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6605 		rack->rc_in_persist = 1;
6606 	}
6607 }
6608 
6609 static void
6610 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
6611 {
6612 	struct timeval tv;
6613 	uint32_t t_time;
6614 
6615 	if (tcp_in_hpts(rack->rc_tp)) {
6616 		tcp_hpts_remove(rack->rc_tp);
6617 		rack->r_ctl.rc_hpts_flags = 0;
6618 	}
6619 #ifdef NETFLIX_SHARED_CWND
6620 	if (rack->r_ctl.rc_scw) {
6621 		tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
6622 		rack->rack_scwnd_is_idle = 0;
6623 	}
6624 #endif
6625 	t_time = tcp_get_usecs(&tv);
6626 	if (rack->rc_gp_dyn_mul &&
6627 	    (rack->use_fixed_rate == 0) &&
6628 	    (rack->rc_always_pace)) {
6629 		/*
6630 		 * Do we count this as if a probe-rtt just
6631 		 * finished?
6632 		 */
6633 		uint32_t time_idle, idle_min;
6634 
6635 		time_idle = t_time - rack->r_ctl.rc_went_idle_time;
6636 		idle_min = rack_min_probertt_hold;
6637 		if (rack_probertt_gpsrtt_cnt_div) {
6638 			uint64_t extra;
6639 			extra = (uint64_t)rack->r_ctl.rc_gp_srtt *
6640 				(uint64_t)rack_probertt_gpsrtt_cnt_mul;
6641 			extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div;
6642 			idle_min += (uint32_t)extra;
6643 		}
6644 		if (time_idle >= idle_min) {
6645 			/* Yes, we count it as a probe-rtt. */
6646 			uint32_t us_cts;
6647 
6648 			us_cts = tcp_get_usecs(NULL);
6649 			if (rack->in_probe_rtt == 0) {
6650 				rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
6651 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
6652 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
6653 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
6654 			} else {
6655 				rack_exit_probertt(rack, us_cts);
6656 			}
6657 		}
6658 	}
6659 	if (rack->r_persist_lt_bw_off) {
6660 		/* Continue where we left off */
6661 		rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(&tv);
6662 		rack->lt_bw_up = 1;
6663 		rack->r_persist_lt_bw_off = 0;
6664 	}
6665 	rack->rc_in_persist = 0;
6666 	rack->r_ctl.rc_went_idle_time = 0;
6667 	tp->t_rxtshift = 0;
6668 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
6669 	   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
6670 	rack->r_ctl.rc_agg_delayed = 0;
6671 	rack->r_early = 0;
6672 	rack->r_late = 0;
6673 	rack->r_ctl.rc_agg_early = 0;
6674 }
6675 
6676 static void
6677 rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts,
6678 		   struct hpts_diag *diag, struct timeval *tv)
6679 {
6680 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6681 		union tcp_log_stackspecific log;
6682 
6683 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6684 		log.u_bbr.flex1 = diag->p_nxt_slot;
6685 		log.u_bbr.flex2 = diag->p_cur_slot;
6686 		log.u_bbr.flex3 = diag->slot_req;
6687 		log.u_bbr.flex4 = diag->inp_hptsslot;
6688 		log.u_bbr.flex5 = diag->slot_remaining;
6689 		log.u_bbr.flex6 = diag->need_new_to;
6690 		log.u_bbr.flex7 = diag->p_hpts_active;
6691 		log.u_bbr.flex8 = diag->p_on_min_sleep;
6692 		/* Hijack other fields as needed */
6693 		log.u_bbr.epoch = diag->have_slept;
6694 		log.u_bbr.lt_epoch = diag->yet_to_sleep;
6695 		log.u_bbr.pkts_out = diag->co_ret;
6696 		log.u_bbr.applimited = diag->hpts_sleep_time;
6697 		log.u_bbr.delivered = diag->p_prev_slot;
6698 		log.u_bbr.inflight = diag->p_runningslot;
6699 		log.u_bbr.bw_inuse = diag->wheel_slot;
6700 		log.u_bbr.rttProp = diag->wheel_cts;
6701 		log.u_bbr.timeStamp = cts;
6702 		log.u_bbr.delRate = diag->maxslots;
6703 		log.u_bbr.cur_del_rate = diag->p_curtick;
6704 		log.u_bbr.cur_del_rate <<= 32;
6705 		log.u_bbr.cur_del_rate |= diag->p_lasttick;
6706 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
6707 		    &rack->rc_inp->inp_socket->so_rcv,
6708 		    &rack->rc_inp->inp_socket->so_snd,
6709 		    BBR_LOG_HPTSDIAG, 0,
6710 		    0, &log, false, tv);
6711 	}
6712 
6713 }
6714 
6715 static void
6716 rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type)
6717 {
6718 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
6719 		union tcp_log_stackspecific log;
6720 		struct timeval tv;
6721 
6722 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6723 		log.u_bbr.flex1 = sb->sb_flags;
6724 		log.u_bbr.flex2 = len;
6725 		log.u_bbr.flex3 = sb->sb_state;
6726 		log.u_bbr.flex8 = type;
6727 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
6728 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
6729 		    &rack->rc_inp->inp_socket->so_rcv,
6730 		    &rack->rc_inp->inp_socket->so_snd,
6731 		    TCP_LOG_SB_WAKE, 0,
6732 		    len, &log, false, &tv);
6733 	}
6734 }
6735 
6736 static void
6737 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts,
6738       int32_t slot, uint32_t tot_len_this_send, int sup_rack)
6739 {
6740 	struct hpts_diag diag;
6741 	struct inpcb *inp = tptoinpcb(tp);
6742 	struct timeval tv;
6743 	uint32_t delayed_ack = 0;
6744 	uint32_t hpts_timeout;
6745 	uint32_t entry_slot = slot;
6746 	uint8_t stopped;
6747 	uint32_t left = 0;
6748 	uint32_t us_cts;
6749 
6750 	if ((tp->t_state == TCPS_CLOSED) ||
6751 	    (tp->t_state == TCPS_LISTEN)) {
6752 		return;
6753 	}
6754 	if (tcp_in_hpts(tp)) {
6755 		/* Already on the pacer */
6756 		return;
6757 	}
6758 	stopped = rack->rc_tmr_stopped;
6759 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
6760 		left = rack->r_ctl.rc_timer_exp - cts;
6761 	}
6762 	rack->r_ctl.rc_timer_exp = 0;
6763 	rack->r_ctl.rc_hpts_flags = 0;
6764 	us_cts = tcp_get_usecs(&tv);
6765 	/* Now early/late accounting */
6766 	rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0);
6767 	if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) {
6768 		/*
6769 		 * We have a early carry over set,
6770 		 * we can always add more time so we
6771 		 * can always make this compensation.
6772 		 *
6773 		 * Note if ack's are allowed to wake us do not
6774 		 * penalize the next timer for being awoke
6775 		 * by an ack aka the rc_agg_early (non-paced mode).
6776 		 */
6777 		slot += rack->r_ctl.rc_agg_early;
6778 		rack->r_early = 0;
6779 		rack->r_ctl.rc_agg_early = 0;
6780 	}
6781 	if (rack->r_late) {
6782 		/*
6783 		 * This is harder, we can
6784 		 * compensate some but it
6785 		 * really depends on what
6786 		 * the current pacing time is.
6787 		 */
6788 		if (rack->r_ctl.rc_agg_delayed >= slot) {
6789 			/*
6790 			 * We can't compensate for it all.
6791 			 * And we have to have some time
6792 			 * on the clock. We always have a min
6793 			 * 10 slots (10 x 10 i.e. 100 usecs).
6794 			 */
6795 			if (slot <= HPTS_TICKS_PER_SLOT) {
6796 				/* We gain delay */
6797 				rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot);
6798 				slot = HPTS_TICKS_PER_SLOT;
6799 			} else {
6800 				/* We take off some */
6801 				rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT);
6802 				slot = HPTS_TICKS_PER_SLOT;
6803 			}
6804 		} else {
6805 			slot -= rack->r_ctl.rc_agg_delayed;
6806 			rack->r_ctl.rc_agg_delayed = 0;
6807 			/* Make sure we have 100 useconds at minimum */
6808 			if (slot < HPTS_TICKS_PER_SLOT) {
6809 				rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot;
6810 				slot = HPTS_TICKS_PER_SLOT;
6811 			}
6812 			if (rack->r_ctl.rc_agg_delayed == 0)
6813 				rack->r_late = 0;
6814 		}
6815 	}
6816 	hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack);
6817 #ifdef TCP_SAD_DETECTION
6818 	if (rack->sack_attack_disable &&
6819 	    (rack->r_ctl.ack_during_sd > 0) &&
6820 	    (slot < tcp_sad_pacing_interval)) {
6821 		/*
6822 		 * We have a potential attacker on
6823 		 * the line. We have possibly some
6824 		 * (or now) pacing time set. We want to
6825 		 * slow down the processing of sacks by some
6826 		 * amount (if it is an attacker). Set the default
6827 		 * slot for attackers in place (unless the original
6828 		 * interval is longer). Its stored in
6829 		 * micro-seconds, so lets convert to msecs.
6830 		 */
6831 		slot = tcp_sad_pacing_interval;
6832 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
6833 		rack->r_ctl.ack_during_sd = 0;
6834 	}
6835 #endif
6836 	if (tp->t_flags & TF_DELACK) {
6837 		delayed_ack = TICKS_2_USEC(tcp_delacktime);
6838 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
6839 	}
6840 	if (delayed_ack && ((hpts_timeout == 0) ||
6841 			    (delayed_ack < hpts_timeout)))
6842 		hpts_timeout = delayed_ack;
6843 	else
6844 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
6845 	/*
6846 	 * If no timers are going to run and we will fall off the hptsi
6847 	 * wheel, we resort to a keep-alive timer if its configured.
6848 	 */
6849 	if ((hpts_timeout == 0) &&
6850 	    (slot == 0)) {
6851 		if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
6852 		    (tp->t_state <= TCPS_CLOSING)) {
6853 			/*
6854 			 * Ok we have no timer (persists, rack, tlp, rxt  or
6855 			 * del-ack), we don't have segments being paced. So
6856 			 * all that is left is the keepalive timer.
6857 			 */
6858 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
6859 				/* Get the established keep-alive time */
6860 				hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp));
6861 			} else {
6862 				/*
6863 				 * Get the initial setup keep-alive time,
6864 				 * note that this is probably not going to
6865 				 * happen, since rack will be running a rxt timer
6866 				 * if a SYN of some sort is outstanding. It is
6867 				 * actually handled in rack_timeout_rxt().
6868 				 */
6869 				hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp));
6870 			}
6871 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
6872 			if (rack->in_probe_rtt) {
6873 				/*
6874 				 * We want to instead not wake up a long time from
6875 				 * now but to wake up about the time we would
6876 				 * exit probe-rtt and initiate a keep-alive ack.
6877 				 * This will get us out of probe-rtt and update
6878 				 * our min-rtt.
6879 				 */
6880 				hpts_timeout = rack_min_probertt_hold;
6881 			}
6882 		}
6883 	}
6884 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
6885 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
6886 		/*
6887 		 * RACK, TLP, persists and RXT timers all are restartable
6888 		 * based on actions input .. i.e we received a packet (ack
6889 		 * or sack) and that changes things (rw, or snd_una etc).
6890 		 * Thus we can restart them with a new value. For
6891 		 * keep-alive, delayed_ack we keep track of what was left
6892 		 * and restart the timer with a smaller value.
6893 		 */
6894 		if (left < hpts_timeout)
6895 			hpts_timeout = left;
6896 	}
6897 	if (hpts_timeout) {
6898 		/*
6899 		 * Hack alert for now we can't time-out over 2,147,483
6900 		 * seconds (a bit more than 596 hours), which is probably ok
6901 		 * :).
6902 		 */
6903 		if (hpts_timeout > 0x7ffffffe)
6904 			hpts_timeout = 0x7ffffffe;
6905 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
6906 	}
6907 	rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0);
6908 	if ((rack->gp_ready == 0) &&
6909 	    (rack->use_fixed_rate == 0) &&
6910 	    (hpts_timeout < slot) &&
6911 	    (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) {
6912 		/*
6913 		 * We have no good estimate yet for the
6914 		 * old clunky burst mitigation or the
6915 		 * real pacing. And the tlp or rxt is smaller
6916 		 * than the pacing calculation. Lets not
6917 		 * pace that long since we know the calculation
6918 		 * so far is not accurate.
6919 		 */
6920 		slot = hpts_timeout;
6921 	}
6922 	/**
6923 	 * Turn off all the flags for queuing by default. The
6924 	 * flags have important meanings to what happens when
6925 	 * LRO interacts with the transport. Most likely (by default now)
6926 	 * mbuf_queueing and ack compression are on. So the transport
6927 	 * has a couple of flags that control what happens (if those
6928 	 * are not on then these flags won't have any effect since it
6929 	 * won't go through the queuing LRO path).
6930 	 *
6931 	 * TF2_MBUF_QUEUE_READY - This flags says that I am busy
6932 	 *                        pacing output, so don't disturb. But
6933 	 *                        it also means LRO can wake me if there
6934 	 *                        is a SACK arrival.
6935 	 *
6936 	 * TF2_DONT_SACK_QUEUE - This flag is used in conjunction
6937 	 *                       with the above flag (QUEUE_READY) and
6938 	 *                       when present it says don't even wake me
6939 	 *                       if a SACK arrives.
6940 	 *
6941 	 * The idea behind these flags is that if we are pacing we
6942 	 * set the MBUF_QUEUE_READY and only get woken up if
6943 	 * a SACK arrives (which could change things) or if
6944 	 * our pacing timer expires. If, however, we have a rack
6945 	 * timer running, then we don't even want a sack to wake
6946 	 * us since the rack timer has to expire before we can send.
6947 	 *
6948 	 * Other cases should usually have none of the flags set
6949 	 * so LRO can call into us.
6950 	 */
6951 	tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE|TF2_MBUF_QUEUE_READY);
6952 	if (slot) {
6953 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
6954 		rack->r_ctl.rc_last_output_to = us_cts + slot;
6955 		/*
6956 		 * A pacing timer (slot) is being set, in
6957 		 * such a case we cannot send (we are blocked by
6958 		 * the timer). So lets tell LRO that it should not
6959 		 * wake us unless there is a SACK. Note this only
6960 		 * will be effective if mbuf queueing is on or
6961 		 * compressed acks are being processed.
6962 		 */
6963 		tp->t_flags2 |= TF2_MBUF_QUEUE_READY;
6964 		/*
6965 		 * But wait if we have a Rack timer running
6966 		 * even a SACK should not disturb us (with
6967 		 * the exception of r_rr_config 3).
6968 		 */
6969 		if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) ||
6970 		    (IN_RECOVERY(tp->t_flags))) {
6971 			if (rack->r_rr_config != 3)
6972 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6973 			else if (rack->rc_pace_dnd) {
6974 				/*
6975 				 * When DND is on, we only let a sack
6976 				 * interrupt us if we are not in recovery.
6977 				 *
6978 				 * If DND is off, then we never hit here
6979 				 * and let all sacks wake us up.
6980 				 *
6981 				 */
6982 				tp->t_flags2 |= TF2_DONT_SACK_QUEUE;
6983 			}
6984 		}
6985 		/* For sack attackers we want to ignore sack */
6986 		if (rack->sack_attack_disable == 1) {
6987 			tp->t_flags2 |= (TF2_DONT_SACK_QUEUE |
6988 			    TF2_MBUF_QUEUE_READY);
6989 		} else if (rack->rc_ack_can_sendout_data) {
6990 			/*
6991 			 * Ahh but wait, this is that special case
6992 			 * where the pacing timer can be disturbed
6993 			 * backout the changes (used for non-paced
6994 			 * burst limiting).
6995 			 */
6996 			tp->t_flags2 &= ~(TF2_DONT_SACK_QUEUE |
6997 			    TF2_MBUF_QUEUE_READY);
6998 		}
6999 		if ((rack->use_rack_rr) &&
7000 		    (rack->r_rr_config < 2) &&
7001 		    ((hpts_timeout) && (hpts_timeout < slot))) {
7002 			/*
7003 			 * Arrange for the hpts to kick back in after the
7004 			 * t-o if the t-o does not cause a send.
7005 			 */
7006 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
7007 						   __LINE__, &diag);
7008 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7009 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
7010 		} else {
7011 			(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(slot),
7012 						   __LINE__, &diag);
7013 			rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7014 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
7015 		}
7016 	} else if (hpts_timeout) {
7017 		/*
7018 		 * With respect to t_flags2(?) here, lets let any new acks wake
7019 		 * us up here. Since we are not pacing (no pacing timer), output
7020 		 * can happen so we should let it. If its a Rack timer, then any inbound
7021 		 * packet probably won't change the sending (we will be blocked)
7022 		 * but it may change the prr stats so letting it in (the set defaults
7023 		 * at the start of this block) are good enough.
7024 		 */
7025 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7026 		(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(hpts_timeout),
7027 					   __LINE__, &diag);
7028 		rack_log_hpts_diag(rack, us_cts, &diag, &tv);
7029 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
7030 	} else {
7031 		/* No timer starting */
7032 #ifdef INVARIANTS
7033 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
7034 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
7035 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
7036 		}
7037 #endif
7038 	}
7039 	rack->rc_tmr_stopped = 0;
7040 	if (slot)
7041 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv, __LINE__);
7042 }
7043 
7044 /*
7045  * RACK Timer, here we simply do logging and house keeping.
7046  * the normal rack_output() function will call the
7047  * appropriate thing to check if we need to do a RACK retransmit.
7048  * We return 1, saying don't proceed with rack_output only
7049  * when all timers have been stopped (destroyed PCB?).
7050  */
7051 static int
7052 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7053 {
7054 	/*
7055 	 * This timer simply provides an internal trigger to send out data.
7056 	 * The check_recovery_mode call will see if there are needed
7057 	 * retransmissions, if so we will enter fast-recovery. The output
7058 	 * call may or may not do the same thing depending on sysctl
7059 	 * settings.
7060 	 */
7061 	struct rack_sendmap *rsm;
7062 
7063 	counter_u64_add(rack_to_tot, 1);
7064 	if (rack->r_state && (rack->r_state != tp->t_state))
7065 		rack_set_state(tp, rack);
7066 	rack->rc_on_min_to = 0;
7067 	rsm = rack_check_recovery_mode(tp, cts);
7068 	rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm);
7069 	if (rsm) {
7070 		rack->r_ctl.rc_resend = rsm;
7071 		rack->r_timer_override = 1;
7072 		if (rack->use_rack_rr) {
7073 			/*
7074 			 * Don't accumulate extra pacing delay
7075 			 * we are allowing the rack timer to
7076 			 * over-ride pacing i.e. rrr takes precedence
7077 			 * if the pacing interval is longer than the rrr
7078 			 * time (in other words we get the min pacing
7079 			 * time versus rrr pacing time).
7080 			 */
7081 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
7082 		}
7083 	}
7084 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
7085 	if (rsm == NULL) {
7086 		/* restart a timer and return 1 */
7087 		rack_start_hpts_timer(rack, tp, cts,
7088 				      0, 0, 0);
7089 		return (1);
7090 	}
7091 	return (0);
7092 }
7093 
7094 
7095 
7096 static void
7097 rack_adjust_orig_mlen(struct rack_sendmap *rsm)
7098 {
7099 
7100 	if ((M_TRAILINGROOM(rsm->m) != rsm->orig_t_space)) {
7101 		/*
7102 		 * The trailing space changed, mbufs can grow
7103 		 * at the tail but they can't shrink from
7104 		 * it, KASSERT that. Adjust the orig_m_len to
7105 		 * compensate for this change.
7106 		 */
7107 		KASSERT((rsm->orig_t_space > M_TRAILINGROOM(rsm->m)),
7108 			("mbuf:%p rsm:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
7109 			 rsm->m,
7110 			 rsm,
7111 			 (intmax_t)M_TRAILINGROOM(rsm->m),
7112 			 rsm->orig_t_space,
7113 			 rsm->orig_m_len,
7114 			 rsm->m->m_len));
7115 		rsm->orig_m_len += (rsm->orig_t_space - M_TRAILINGROOM(rsm->m));
7116 		rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7117 	}
7118 	if (rsm->m->m_len < rsm->orig_m_len) {
7119 		/*
7120 		 * Mbuf shrank, trimmed off the top by an ack, our
7121 		 * offset changes.
7122 		 */
7123 		KASSERT((rsm->soff >= (rsm->orig_m_len - rsm->m->m_len)),
7124 			("mbuf:%p len:%u rsm:%p oml:%u soff:%u\n",
7125 			 rsm->m, rsm->m->m_len,
7126 			 rsm, rsm->orig_m_len,
7127 			 rsm->soff));
7128 		if (rsm->soff >= (rsm->orig_m_len - rsm->m->m_len))
7129 			rsm->soff -= (rsm->orig_m_len - rsm->m->m_len);
7130 		else
7131 			rsm->soff = 0;
7132 		rsm->orig_m_len = rsm->m->m_len;
7133 #ifdef INVARIANTS
7134 	} else if (rsm->m->m_len > rsm->orig_m_len) {
7135 		panic("rsm:%p m:%p m_len grew outside of t_space compensation",
7136 		      rsm, rsm->m);
7137 #endif
7138 	}
7139 }
7140 
7141 static void
7142 rack_setup_offset_for_rsm(struct tcp_rack *rack, struct rack_sendmap *src_rsm, struct rack_sendmap *rsm)
7143 {
7144 	struct mbuf *m;
7145 	uint32_t soff;
7146 
7147 	if (src_rsm->m &&
7148 	    ((src_rsm->orig_m_len != src_rsm->m->m_len) ||
7149 	     (M_TRAILINGROOM(src_rsm->m) != src_rsm->orig_t_space))) {
7150 		/* Fix up the orig_m_len and possibly the mbuf offset */
7151 		rack_adjust_orig_mlen(src_rsm);
7152 	}
7153 	m = src_rsm->m;
7154 	soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start);
7155 	while (soff >= m->m_len) {
7156 		/* Move out past this mbuf */
7157 		soff -= m->m_len;
7158 		m = m->m_next;
7159 		KASSERT((m != NULL),
7160 			("rsm:%p nrsm:%p hit at soff:%u null m",
7161 			 src_rsm, rsm, soff));
7162 		if (m == NULL) {
7163 			/* This should *not* happen which is why there is a kassert */
7164 			src_rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7165 					       (src_rsm->r_start - rack->rc_tp->snd_una),
7166 					       &src_rsm->soff);
7167 			src_rsm->orig_m_len = src_rsm->m->m_len;
7168 			src_rsm->orig_t_space = M_TRAILINGROOM(src_rsm->m);
7169 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
7170 					   (rsm->r_start - rack->rc_tp->snd_una),
7171 					   &rsm->soff);
7172 			rsm->orig_m_len = rsm->m->m_len;
7173 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7174 			return;
7175 		}
7176 	}
7177 	rsm->m = m;
7178 	rsm->soff = soff;
7179 	rsm->orig_m_len = m->m_len;
7180 	rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
7181 }
7182 
7183 static __inline void
7184 rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm,
7185 	       struct rack_sendmap *rsm, uint32_t start)
7186 {
7187 	int idx;
7188 
7189 	nrsm->r_start = start;
7190 	nrsm->r_end = rsm->r_end;
7191 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
7192 	nrsm->r_flags = rsm->r_flags;
7193 	nrsm->r_dupack = rsm->r_dupack;
7194 	nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed;
7195 	nrsm->r_rtr_bytes = 0;
7196 	nrsm->r_fas = rsm->r_fas;
7197 	nrsm->r_bas = rsm->r_bas;
7198 	rsm->r_end = nrsm->r_start;
7199 	nrsm->r_just_ret = rsm->r_just_ret;
7200 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
7201 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
7202 	}
7203 	/* Now if we have SYN flag we keep it on the left edge */
7204 	if (nrsm->r_flags & RACK_HAS_SYN)
7205 		nrsm->r_flags &= ~RACK_HAS_SYN;
7206 	/* Now if we have a FIN flag we keep it on the right edge */
7207 	if (rsm->r_flags & RACK_HAS_FIN)
7208 		rsm->r_flags &= ~RACK_HAS_FIN;
7209 	/* Push bit must go to the right edge as well */
7210 	if (rsm->r_flags & RACK_HAD_PUSH)
7211 		rsm->r_flags &= ~RACK_HAD_PUSH;
7212 	/* Clone over the state of the hw_tls flag */
7213 	nrsm->r_hw_tls = rsm->r_hw_tls;
7214 	/*
7215 	 * Now we need to find nrsm's new location in the mbuf chain
7216 	 * we basically calculate a new offset, which is soff +
7217 	 * how much is left in original rsm. Then we walk out the mbuf
7218 	 * chain to find the righ position, it may be the same mbuf
7219 	 * or maybe not.
7220 	 */
7221 	KASSERT(((rsm->m != NULL) ||
7222 		 (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))),
7223 		("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack));
7224 	if (rsm->m)
7225 		rack_setup_offset_for_rsm(rack, rsm, nrsm);
7226 }
7227 
7228 static struct rack_sendmap *
7229 rack_merge_rsm(struct tcp_rack *rack,
7230 	       struct rack_sendmap *l_rsm,
7231 	       struct rack_sendmap *r_rsm)
7232 {
7233 	/*
7234 	 * We are merging two ack'd RSM's,
7235 	 * the l_rsm is on the left (lower seq
7236 	 * values) and the r_rsm is on the right
7237 	 * (higher seq value). The simplest way
7238 	 * to merge these is to move the right
7239 	 * one into the left. I don't think there
7240 	 * is any reason we need to try to find
7241 	 * the oldest (or last oldest retransmitted).
7242 	 */
7243 	rack_log_map_chg(rack->rc_tp, rack, NULL,
7244 			 l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__);
7245 	l_rsm->r_end = r_rsm->r_end;
7246 	if (l_rsm->r_dupack < r_rsm->r_dupack)
7247 		l_rsm->r_dupack = r_rsm->r_dupack;
7248 	if (r_rsm->r_rtr_bytes)
7249 		l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes;
7250 	if (r_rsm->r_in_tmap) {
7251 		/* This really should not happen */
7252 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext);
7253 		r_rsm->r_in_tmap = 0;
7254 	}
7255 
7256 	/* Now the flags */
7257 	if (r_rsm->r_flags & RACK_HAS_FIN)
7258 		l_rsm->r_flags |= RACK_HAS_FIN;
7259 	if (r_rsm->r_flags & RACK_TLP)
7260 		l_rsm->r_flags |= RACK_TLP;
7261 	if (r_rsm->r_flags & RACK_RWND_COLLAPSED)
7262 		l_rsm->r_flags |= RACK_RWND_COLLAPSED;
7263 	if ((r_rsm->r_flags & RACK_APP_LIMITED)  &&
7264 	    ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) {
7265 		/*
7266 		 * If both are app-limited then let the
7267 		 * free lower the count. If right is app
7268 		 * limited and left is not, transfer.
7269 		 */
7270 		l_rsm->r_flags |= RACK_APP_LIMITED;
7271 		r_rsm->r_flags &= ~RACK_APP_LIMITED;
7272 		if (r_rsm == rack->r_ctl.rc_first_appl)
7273 			rack->r_ctl.rc_first_appl = l_rsm;
7274 	}
7275 	tqhash_remove(rack->r_ctl.tqh, r_rsm, REMOVE_TYPE_MERGE);
7276 	/*
7277 	 * We keep the largest value, which is the newest
7278 	 * send. We do this in case a segment that is
7279 	 * joined together and not part of a GP estimate
7280 	 * later gets expanded into the GP estimate.
7281 	 *
7282 	 * We prohibit the merging of unlike kinds i.e.
7283 	 * all pieces that are in the GP estimate can be
7284 	 * merged and all pieces that are not in a GP estimate
7285 	 * can be merged, but not disimilar pieces. Combine
7286 	 * this with taking the highest here and we should
7287 	 * be ok unless of course the client reneges. Then
7288 	 * all bets are off.
7289 	 */
7290 	if(l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] <
7291 	   r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)]) {
7292 		l_rsm->r_tim_lastsent[(l_rsm->r_rtr_cnt-1)] = r_rsm->r_tim_lastsent[(r_rsm->r_rtr_cnt-1)];
7293 	}
7294 	/*
7295 	 * When merging two RSM's we also need to consider the ack time and keep
7296 	 * newest. If the ack gets merged into a measurement then that is the
7297 	 * one we will want to be using.
7298 	 */
7299 	if(l_rsm->r_ack_arrival	 < r_rsm->r_ack_arrival)
7300 		l_rsm->r_ack_arrival = r_rsm->r_ack_arrival;
7301 
7302 	if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) {
7303 		/* Transfer the split limit to the map we free */
7304 		r_rsm->r_limit_type = l_rsm->r_limit_type;
7305 		l_rsm->r_limit_type = 0;
7306 	}
7307 	rack_free(rack, r_rsm);
7308 	l_rsm->r_flags |= RACK_MERGED;
7309 	return (l_rsm);
7310 }
7311 
7312 /*
7313  * TLP Timer, here we simply setup what segment we want to
7314  * have the TLP expire on, the normal rack_output() will then
7315  * send it out.
7316  *
7317  * We return 1, saying don't proceed with rack_output only
7318  * when all timers have been stopped (destroyed PCB?).
7319  */
7320 static int
7321 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp)
7322 {
7323 	/*
7324 	 * Tail Loss Probe.
7325 	 */
7326 	struct rack_sendmap *rsm = NULL;
7327 	int insret __diagused;
7328 	struct socket *so = tptosocket(tp);
7329 	uint32_t amm;
7330 	uint32_t out, avail;
7331 	int collapsed_win = 0;
7332 
7333 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
7334 		/* Its not time yet */
7335 		return (0);
7336 	}
7337 	if (ctf_progress_timeout_check(tp, true)) {
7338 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7339 		return (-ETIMEDOUT);	/* tcp_drop() */
7340 	}
7341 	/*
7342 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
7343 	 * need to figure out how to force a full MSS segment out.
7344 	 */
7345 	rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL);
7346 	rack->r_ctl.retran_during_recovery = 0;
7347 	rack->r_ctl.dsack_byte_cnt = 0;
7348 	counter_u64_add(rack_tlp_tot, 1);
7349 	if (rack->r_state && (rack->r_state != tp->t_state))
7350 		rack_set_state(tp, rack);
7351 	avail = sbavail(&so->so_snd);
7352 	out = tp->snd_max - tp->snd_una;
7353 	if ((out > tp->snd_wnd) || rack->rc_has_collapsed) {
7354 		/* special case, we need a retransmission */
7355 		collapsed_win = 1;
7356 		goto need_retran;
7357 	}
7358 	if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) {
7359 		rack->r_ctl.dsack_persist--;
7360 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7361 			rack->r_ctl.num_dsack = 0;
7362 		}
7363 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7364 	}
7365 	if ((tp->t_flags & TF_GPUTINPROG) &&
7366 	    (rack->r_ctl.rc_tlp_cnt_out == 1)) {
7367 		/*
7368 		 * If this is the second in a row
7369 		 * TLP and we are doing a measurement
7370 		 * its time to abandon the measurement.
7371 		 * Something is likely broken on
7372 		 * the clients network and measuring a
7373 		 * broken network does us no good.
7374 		 */
7375 		tp->t_flags &= ~TF_GPUTINPROG;
7376 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7377 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
7378 					   tp->gput_seq,
7379 					   0, 0, 18, __LINE__, NULL, 0);
7380 	}
7381 	/*
7382 	 * Check our send oldest always settings, and if
7383 	 * there is an oldest to send jump to the need_retran.
7384 	 */
7385 	if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0))
7386 		goto need_retran;
7387 
7388 	if (avail > out) {
7389 		/* New data is available */
7390 		amm = avail - out;
7391 		if (amm > ctf_fixed_maxseg(tp)) {
7392 			amm = ctf_fixed_maxseg(tp);
7393 			if ((amm + out) > tp->snd_wnd) {
7394 				/* We are rwnd limited */
7395 				goto need_retran;
7396 			}
7397 		} else if (amm < ctf_fixed_maxseg(tp)) {
7398 			/* not enough to fill a MTU */
7399 			goto need_retran;
7400 		}
7401 		if (IN_FASTRECOVERY(tp->t_flags)) {
7402 			/* Unlikely */
7403 			if (rack->rack_no_prr == 0) {
7404 				if (out + amm <= tp->snd_wnd) {
7405 					rack->r_ctl.rc_prr_sndcnt = amm;
7406 					rack->r_ctl.rc_tlp_new_data = amm;
7407 					rack_log_to_prr(rack, 4, 0, __LINE__);
7408 				}
7409 			} else
7410 				goto need_retran;
7411 		} else {
7412 			/* Set the send-new override */
7413 			if (out + amm <= tp->snd_wnd)
7414 				rack->r_ctl.rc_tlp_new_data = amm;
7415 			else
7416 				goto need_retran;
7417 		}
7418 		rack->r_ctl.rc_tlpsend = NULL;
7419 		counter_u64_add(rack_tlp_newdata, 1);
7420 		goto send;
7421 	}
7422 need_retran:
7423 	/*
7424 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
7425 	 * optionally the first un-acked segment.
7426 	 */
7427 	if (collapsed_win == 0) {
7428 		if (rack_always_send_oldest)
7429 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
7430 		else {
7431 			rsm = tqhash_max(rack->r_ctl.tqh);
7432 			if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
7433 				rsm = rack_find_high_nonack(rack, rsm);
7434 			}
7435 		}
7436 		if (rsm == NULL) {
7437 #ifdef TCP_BLACKBOX
7438 			tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
7439 #endif
7440 			goto out;
7441 		}
7442 	} else {
7443 		/*
7444 		 * We had a collapsed window, lets find
7445 		 * the point before the collapse.
7446 		 */
7447 		if (SEQ_GT((rack->r_ctl.last_collapse_point - 1), rack->rc_tp->snd_una))
7448 			rsm = tqhash_find(rack->r_ctl.tqh, (rack->r_ctl.last_collapse_point - 1));
7449 		else {
7450 			rsm = tqhash_min(rack->r_ctl.tqh);
7451 		}
7452 		if (rsm == NULL) {
7453 			/* Huh */
7454 			goto out;
7455 		}
7456 	}
7457 	if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) {
7458 		/*
7459 		 * We need to split this the last segment in two.
7460 		 */
7461 		struct rack_sendmap *nrsm;
7462 
7463 		nrsm = rack_alloc_full_limit(rack);
7464 		if (nrsm == NULL) {
7465 			/*
7466 			 * No memory to split, we will just exit and punt
7467 			 * off to the RXT timer.
7468 			 */
7469 			goto out;
7470 		}
7471 		rack_clone_rsm(rack, nrsm, rsm,
7472 			       (rsm->r_end - ctf_fixed_maxseg(tp)));
7473 		rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
7474 #ifndef INVARIANTS
7475 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
7476 #else
7477 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
7478 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
7479 			      nrsm, insret, rack, rsm);
7480 		}
7481 #endif
7482 		if (rsm->r_in_tmap) {
7483 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
7484 			nrsm->r_in_tmap = 1;
7485 		}
7486 		rsm = nrsm;
7487 	}
7488 	rack->r_ctl.rc_tlpsend = rsm;
7489 send:
7490 	/* Make sure output path knows we are doing a TLP */
7491 	*doing_tlp = 1;
7492 	rack->r_timer_override = 1;
7493 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7494 	return (0);
7495 out:
7496 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
7497 	return (0);
7498 }
7499 
7500 /*
7501  * Delayed ack Timer, here we simply need to setup the
7502  * ACK_NOW flag and remove the DELACK flag. From there
7503  * the output routine will send the ack out.
7504  *
7505  * We only return 1, saying don't proceed, if all timers
7506  * are stopped (destroyed PCB?).
7507  */
7508 static int
7509 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7510 {
7511 
7512 	rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL);
7513 	tp->t_flags &= ~TF_DELACK;
7514 	tp->t_flags |= TF_ACKNOW;
7515 	KMOD_TCPSTAT_INC(tcps_delack);
7516 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
7517 	return (0);
7518 }
7519 
7520 /*
7521  * Persists timer, here we simply send the
7522  * same thing as a keepalive will.
7523  * the one byte send.
7524  *
7525  * We only return 1, saying don't proceed, if all timers
7526  * are stopped (destroyed PCB?).
7527  */
7528 static int
7529 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7530 {
7531 	struct tcptemp *t_template;
7532 	int32_t retval = 1;
7533 
7534 	if (rack->rc_in_persist == 0)
7535 		return (0);
7536 	if (ctf_progress_timeout_check(tp, false)) {
7537 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7538 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7539 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7540 		return (-ETIMEDOUT);	/* tcp_drop() */
7541 	}
7542 	/*
7543 	 * Persistence timer into zero window. Force a byte to be output, if
7544 	 * possible.
7545 	 */
7546 	KMOD_TCPSTAT_INC(tcps_persisttimeo);
7547 	/*
7548 	 * Hack: if the peer is dead/unreachable, we do not time out if the
7549 	 * window is closed.  After a full backoff, drop the connection if
7550 	 * the idle time (no responses to probes) reaches the maximum
7551 	 * backoff that we would use if retransmitting.
7552 	 */
7553 	if (tp->t_rxtshift >= V_tcp_retries &&
7554 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
7555 	     TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) {
7556 		KMOD_TCPSTAT_INC(tcps_persistdrop);
7557 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7558 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7559 		retval = -ETIMEDOUT;	/* tcp_drop() */
7560 		goto out;
7561 	}
7562 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
7563 	    tp->snd_una == tp->snd_max)
7564 		rack_exit_persist(tp, rack, cts);
7565 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
7566 	/*
7567 	 * If the user has closed the socket then drop a persisting
7568 	 * connection after a much reduced timeout.
7569 	 */
7570 	if (tp->t_state > TCPS_CLOSE_WAIT &&
7571 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
7572 		KMOD_TCPSTAT_INC(tcps_persistdrop);
7573 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
7574 		counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends);
7575 		retval = -ETIMEDOUT;	/* tcp_drop() */
7576 		goto out;
7577 	}
7578 	t_template = tcpip_maketemplate(rack->rc_inp);
7579 	if (t_template) {
7580 		/* only set it if we were answered */
7581 		if (rack->forced_ack == 0) {
7582 			rack->forced_ack = 1;
7583 			rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7584 		} else {
7585 			rack->probe_not_answered = 1;
7586 			counter_u64_add(rack_persists_loss, 1);
7587 			rack->r_ctl.persist_lost_ends++;
7588 		}
7589 		counter_u64_add(rack_persists_sends, 1);
7590 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
7591 		tcp_respond(tp, t_template->tt_ipgen,
7592 			    &t_template->tt_t, (struct mbuf *)NULL,
7593 			    tp->rcv_nxt, tp->snd_una - 1, 0);
7594 		/* This sends an ack */
7595 		if (tp->t_flags & TF_DELACK)
7596 			tp->t_flags &= ~TF_DELACK;
7597 		free(t_template, M_TEMP);
7598 	}
7599 	if (tp->t_rxtshift < V_tcp_retries)
7600 		tp->t_rxtshift++;
7601 out:
7602 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL);
7603 	rack_start_hpts_timer(rack, tp, cts,
7604 			      0, 0, 0);
7605 	return (retval);
7606 }
7607 
7608 /*
7609  * If a keepalive goes off, we had no other timers
7610  * happening. We always return 1 here since this
7611  * routine either drops the connection or sends
7612  * out a segment with respond.
7613  */
7614 static int
7615 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7616 {
7617 	struct tcptemp *t_template;
7618 	struct inpcb *inp = tptoinpcb(tp);
7619 
7620 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
7621 	rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL);
7622 	/*
7623 	 * Keep-alive timer went off; send something or drop connection if
7624 	 * idle for too long.
7625 	 */
7626 	KMOD_TCPSTAT_INC(tcps_keeptimeo);
7627 	if (tp->t_state < TCPS_ESTABLISHED)
7628 		goto dropit;
7629 	if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
7630 	    tp->t_state <= TCPS_CLOSING) {
7631 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
7632 			goto dropit;
7633 		/*
7634 		 * Send a packet designed to force a response if the peer is
7635 		 * up and reachable: either an ACK if the connection is
7636 		 * still alive, or an RST if the peer has closed the
7637 		 * connection due to timeout or reboot. Using sequence
7638 		 * number tp->snd_una-1 causes the transmitted zero-length
7639 		 * segment to lie outside the receive window; by the
7640 		 * protocol spec, this requires the correspondent TCP to
7641 		 * respond.
7642 		 */
7643 		KMOD_TCPSTAT_INC(tcps_keepprobe);
7644 		t_template = tcpip_maketemplate(inp);
7645 		if (t_template) {
7646 			if (rack->forced_ack == 0) {
7647 				rack->forced_ack = 1;
7648 				rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL);
7649 			} else {
7650 				rack->probe_not_answered = 1;
7651 			}
7652 			tcp_respond(tp, t_template->tt_ipgen,
7653 			    &t_template->tt_t, (struct mbuf *)NULL,
7654 			    tp->rcv_nxt, tp->snd_una - 1, 0);
7655 			free(t_template, M_TEMP);
7656 		}
7657 	}
7658 	rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
7659 	return (1);
7660 dropit:
7661 	KMOD_TCPSTAT_INC(tcps_keepdrops);
7662 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7663 	return (-ETIMEDOUT);	/* tcp_drop() */
7664 }
7665 
7666 /*
7667  * Retransmit helper function, clear up all the ack
7668  * flags and take care of important book keeping.
7669  */
7670 static void
7671 rack_remxt_tmr(struct tcpcb *tp)
7672 {
7673 	/*
7674 	 * The retransmit timer went off, all sack'd blocks must be
7675 	 * un-acked.
7676 	 */
7677 	struct rack_sendmap *rsm, *trsm = NULL;
7678 	struct tcp_rack *rack;
7679 
7680 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7681 	rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__);
7682 	rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL);
7683 	if (rack->r_state && (rack->r_state != tp->t_state))
7684 		rack_set_state(tp, rack);
7685 	/*
7686 	 * Ideally we would like to be able to
7687 	 * mark SACK-PASS on anything not acked here.
7688 	 *
7689 	 * However, if we do that we would burst out
7690 	 * all that data 1ms apart. This would be unwise,
7691 	 * so for now we will just let the normal rxt timer
7692 	 * and tlp timer take care of it.
7693 	 *
7694 	 * Also we really need to stick them back in sequence
7695 	 * order. This way we send in the proper order and any
7696 	 * sacks that come floating in will "re-ack" the data.
7697 	 * To do this we zap the tmap with an INIT and then
7698 	 * walk through and place every rsm in the tail queue
7699 	 * hash table back in its seq ordered place.
7700 	 */
7701 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
7702 
7703 	TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
7704 		rsm->r_dupack = 0;
7705 		if (rack_verbose_logging)
7706 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
7707 		/* We must re-add it back to the tlist */
7708 		if (trsm == NULL) {
7709 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
7710 		} else {
7711 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
7712 		}
7713 		rsm->r_in_tmap = 1;
7714 		trsm = rsm;
7715 		if (rsm->r_flags & RACK_ACKED)
7716 			rsm->r_flags |= RACK_WAS_ACKED;
7717 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
7718 		rsm->r_flags |= RACK_MUST_RXT;
7719 	}
7720 	/* Clear the count (we just un-acked them) */
7721 	rack->r_ctl.rc_last_timeout_snduna = tp->snd_una;
7722 	rack->r_ctl.rc_sacked = 0;
7723 	rack->r_ctl.rc_sacklast = NULL;
7724 	rack->r_ctl.rc_agg_delayed = 0;
7725 	rack->r_early = 0;
7726 	rack->r_ctl.rc_agg_early = 0;
7727 	rack->r_late = 0;
7728 	/* Clear the tlp rtx mark */
7729 	rack->r_ctl.rc_resend = tqhash_min(rack->r_ctl.tqh);
7730 	if (rack->r_ctl.rc_resend != NULL)
7731 		rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT;
7732 	rack->r_ctl.rc_prr_sndcnt = 0;
7733 	rack_log_to_prr(rack, 6, 0, __LINE__);
7734 	rack->r_timer_override = 1;
7735 	if ((((tp->t_flags & TF_SACK_PERMIT) == 0)
7736 #ifdef TCP_SAD_DETECTION
7737 	    || (rack->sack_attack_disable != 0)
7738 #endif
7739 		    ) && ((tp->t_flags & TF_SENTFIN) == 0)) {
7740 		/*
7741 		 * For non-sack customers new data
7742 		 * needs to go out as retransmits until
7743 		 * we retransmit up to snd_max.
7744 		 */
7745 		rack->r_must_retran = 1;
7746 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp,
7747 						rack->r_ctl.rc_sacked);
7748 	}
7749 	rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
7750 }
7751 
7752 static void
7753 rack_convert_rtts(struct tcpcb *tp)
7754 {
7755 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
7756 	tp->t_rxtcur = RACK_REXMTVAL(tp);
7757 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
7758 		tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop);
7759 	}
7760 	if (tp->t_rxtcur > rack_rto_max) {
7761 		tp->t_rxtcur = rack_rto_max;
7762 	}
7763 }
7764 
7765 static void
7766 rack_cc_conn_init(struct tcpcb *tp)
7767 {
7768 	struct tcp_rack *rack;
7769 	uint32_t srtt;
7770 
7771 	rack = (struct tcp_rack *)tp->t_fb_ptr;
7772 	srtt = tp->t_srtt;
7773 	cc_conn_init(tp);
7774 	/*
7775 	 * Now convert to rack's internal format,
7776 	 * if required.
7777 	 */
7778 	if ((srtt == 0) && (tp->t_srtt != 0))
7779 		rack_convert_rtts(tp);
7780 	/*
7781 	 * We want a chance to stay in slowstart as
7782 	 * we create a connection. TCP spec says that
7783 	 * initially ssthresh is infinite. For our
7784 	 * purposes that is the snd_wnd.
7785 	 */
7786 	if (tp->snd_ssthresh < tp->snd_wnd) {
7787 		tp->snd_ssthresh = tp->snd_wnd;
7788 	}
7789 	/*
7790 	 * We also want to assure a IW worth of
7791 	 * data can get inflight.
7792 	 */
7793 	if (rc_init_window(rack) < tp->snd_cwnd)
7794 		tp->snd_cwnd = rc_init_window(rack);
7795 }
7796 
7797 /*
7798  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
7799  * we will setup to retransmit the lowest seq number outstanding.
7800  */
7801 static int
7802 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
7803 {
7804 	struct inpcb *inp = tptoinpcb(tp);
7805 	int32_t rexmt;
7806 	int32_t retval = 0;
7807 	bool isipv6;
7808 
7809 	if ((tp->t_flags & TF_GPUTINPROG) &&
7810 	    (tp->t_rxtshift)) {
7811 		/*
7812 		 * We have had a second timeout
7813 		 * measurements on successive rxt's are not profitable.
7814 		 * It is unlikely to be of any use (the network is
7815 		 * broken or the client went away).
7816 		 */
7817 		tp->t_flags &= ~TF_GPUTINPROG;
7818 		rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
7819 					   rack->r_ctl.rc_gp_srtt /*flex1*/,
7820 					   tp->gput_seq,
7821 					   0, 0, 18, __LINE__, NULL, 0);
7822 	}
7823 	if (ctf_progress_timeout_check(tp, false)) {
7824 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7825 		rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
7826 		return (-ETIMEDOUT);	/* tcp_drop() */
7827 	}
7828 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
7829 	rack->r_ctl.retran_during_recovery = 0;
7830 	rack->rc_ack_required = 1;
7831 	rack->r_ctl.dsack_byte_cnt = 0;
7832 	if (IN_FASTRECOVERY(tp->t_flags))
7833 		tp->t_flags |= TF_WASFRECOVERY;
7834 	else
7835 		tp->t_flags &= ~TF_WASFRECOVERY;
7836 	if (IN_CONGRECOVERY(tp->t_flags))
7837 		tp->t_flags |= TF_WASCRECOVERY;
7838 	else
7839 		tp->t_flags &= ~TF_WASCRECOVERY;
7840 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
7841 	    (tp->snd_una == tp->snd_max)) {
7842 		/* Nothing outstanding .. nothing to do */
7843 		return (0);
7844 	}
7845 	if (rack->r_ctl.dsack_persist) {
7846 		rack->r_ctl.dsack_persist--;
7847 		if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) {
7848 			rack->r_ctl.num_dsack = 0;
7849 		}
7850 		rack_log_dsack_event(rack, 1, __LINE__, 0, 0);
7851 	}
7852 	/*
7853 	 * Rack can only run one timer  at a time, so we cannot
7854 	 * run a KEEPINIT (gating SYN sending) and a retransmit
7855 	 * timer for the SYN. So if we are in a front state and
7856 	 * have a KEEPINIT timer we need to check the first transmit
7857 	 * against now to see if we have exceeded the KEEPINIT time
7858 	 * (if one is set).
7859 	 */
7860 	if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) &&
7861 	    (TP_KEEPINIT(tp) != 0)) {
7862 		struct rack_sendmap *rsm;
7863 
7864 		rsm = tqhash_min(rack->r_ctl.tqh);
7865 		if (rsm) {
7866 			/* Ok we have something outstanding to test keepinit with */
7867 			if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) &&
7868 			    ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) {
7869 				/* We have exceeded the KEEPINIT time */
7870 				tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
7871 				goto drop_it;
7872 			}
7873 		}
7874 	}
7875 	/*
7876 	 * Retransmission timer went off.  Message has not been acked within
7877 	 * retransmit interval.  Back off to a longer retransmit interval
7878 	 * and retransmit one segment.
7879 	 */
7880 	rack_remxt_tmr(tp);
7881 	if ((rack->r_ctl.rc_resend == NULL) ||
7882 	    ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) {
7883 		/*
7884 		 * If the rwnd collapsed on
7885 		 * the one we are retransmitting
7886 		 * it does not count against the
7887 		 * rxt count.
7888 		 */
7889 		tp->t_rxtshift++;
7890 	}
7891 	if (tp->t_rxtshift > V_tcp_retries) {
7892 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
7893 drop_it:
7894 		tp->t_rxtshift = V_tcp_retries;
7895 		KMOD_TCPSTAT_INC(tcps_timeoutdrop);
7896 		/* XXXGL: previously t_softerror was casted to uint16_t */
7897 		MPASS(tp->t_softerror >= 0);
7898 		retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT;
7899 		goto out;	/* tcp_drop() */
7900 	}
7901 	if (tp->t_state == TCPS_SYN_SENT) {
7902 		/*
7903 		 * If the SYN was retransmitted, indicate CWND to be limited
7904 		 * to 1 segment in cc_conn_init().
7905 		 */
7906 		tp->snd_cwnd = 1;
7907 	} else if (tp->t_rxtshift == 1) {
7908 		/*
7909 		 * first retransmit; record ssthresh and cwnd so they can be
7910 		 * recovered if this turns out to be a "bad" retransmit. A
7911 		 * retransmit is considered "bad" if an ACK for this segment
7912 		 * is received within RTT/2 interval; the assumption here is
7913 		 * that the ACK was already in flight.  See "On Estimating
7914 		 * End-to-End Network Path Properties" by Allman and Paxson
7915 		 * for more details.
7916 		 */
7917 		tp->snd_cwnd_prev = tp->snd_cwnd;
7918 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
7919 		tp->snd_recover_prev = tp->snd_recover;
7920 		tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2);
7921 		tp->t_flags |= TF_PREVVALID;
7922 	} else if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
7923 		tp->t_flags &= ~TF_PREVVALID;
7924 	KMOD_TCPSTAT_INC(tcps_rexmttimeo);
7925 	if ((tp->t_state == TCPS_SYN_SENT) ||
7926 	    (tp->t_state == TCPS_SYN_RECEIVED))
7927 		rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift];
7928 	else
7929 		rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift];
7930 
7931 	RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt,
7932 	   max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop);
7933 	/*
7934 	 * We enter the path for PLMTUD if connection is established or, if
7935 	 * connection is FIN_WAIT_1 status, reason for the last is that if
7936 	 * amount of data we send is very small, we could send it in couple
7937 	 * of packets and process straight to FIN. In that case we won't
7938 	 * catch ESTABLISHED state.
7939 	 */
7940 #ifdef INET6
7941 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
7942 #else
7943 	isipv6 = false;
7944 #endif
7945 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
7946 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
7947 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
7948 	    ((tp->t_state == TCPS_ESTABLISHED) ||
7949 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
7950 		/*
7951 		 * Idea here is that at each stage of mtu probe (usually,
7952 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
7953 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
7954 		 * should take care of that.
7955 		 */
7956 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
7957 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
7958 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
7959 		    tp->t_rxtshift % 2 == 0)) {
7960 			/*
7961 			 * Enter Path MTU Black-hole Detection mechanism: -
7962 			 * Disable Path MTU Discovery (IP "DF" bit). -
7963 			 * Reduce MTU to lower value than what we negotiated
7964 			 * with peer.
7965 			 */
7966 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
7967 				/* Record that we may have found a black hole. */
7968 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
7969 				/* Keep track of previous MSS. */
7970 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
7971 			}
7972 
7973 			/*
7974 			 * Reduce the MSS to blackhole value or to the
7975 			 * default in an attempt to retransmit.
7976 			 */
7977 #ifdef INET6
7978 			if (isipv6 &&
7979 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
7980 				/* Use the sysctl tuneable blackhole MSS. */
7981 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
7982 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
7983 			} else if (isipv6) {
7984 				/* Use the default MSS. */
7985 				tp->t_maxseg = V_tcp_v6mssdflt;
7986 				/*
7987 				 * Disable Path MTU Discovery when we switch
7988 				 * to minmss.
7989 				 */
7990 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
7991 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
7992 			}
7993 #endif
7994 #if defined(INET6) && defined(INET)
7995 			else
7996 #endif
7997 #ifdef INET
7998 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
7999 				/* Use the sysctl tuneable blackhole MSS. */
8000 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
8001 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated);
8002 			} else {
8003 				/* Use the default MSS. */
8004 				tp->t_maxseg = V_tcp_mssdflt;
8005 				/*
8006 				 * Disable Path MTU Discovery when we switch
8007 				 * to minmss.
8008 				 */
8009 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
8010 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
8011 			}
8012 #endif
8013 		} else {
8014 			/*
8015 			 * If further retransmissions are still unsuccessful
8016 			 * with a lowered MTU, maybe this isn't a blackhole
8017 			 * and we restore the previous MSS and blackhole
8018 			 * detection flags. The limit '6' is determined by
8019 			 * giving each probe stage (1448, 1188, 524) 2
8020 			 * chances to recover.
8021 			 */
8022 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
8023 			    (tp->t_rxtshift >= 6)) {
8024 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
8025 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
8026 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
8027 				KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed);
8028 			}
8029 		}
8030 	}
8031 	/*
8032 	 * Disable RFC1323 and SACK if we haven't got any response to
8033 	 * our third SYN to work-around some broken terminal servers
8034 	 * (most of which have hopefully been retired) that have bad VJ
8035 	 * header compression code which trashes TCP segments containing
8036 	 * unknown-to-them TCP options.
8037 	 */
8038 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
8039 	    (tp->t_rxtshift == 3))
8040 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
8041 	/*
8042 	 * If we backed off this far, our srtt estimate is probably bogus.
8043 	 * Clobber it so we'll take the next rtt measurement as our srtt;
8044 	 * move the current srtt into rttvar to keep the current retransmit
8045 	 * times until then.
8046 	 */
8047 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
8048 #ifdef INET6
8049 		if ((inp->inp_vflag & INP_IPV6) != 0)
8050 			in6_losing(inp);
8051 		else
8052 #endif
8053 			in_losing(inp);
8054 		tp->t_rttvar += tp->t_srtt;
8055 		tp->t_srtt = 0;
8056 	}
8057 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
8058 	tp->snd_recover = tp->snd_max;
8059 	tp->t_flags |= TF_ACKNOW;
8060 	tp->t_rtttime = 0;
8061 	rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__);
8062 out:
8063 	return (retval);
8064 }
8065 
8066 static int
8067 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp)
8068 {
8069 	int32_t ret = 0;
8070 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
8071 
8072 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
8073 	    (tp->t_flags & TF_GPUTINPROG)) {
8074 		/*
8075 		 * We have a goodput in progress
8076 		 * and we have entered a late state.
8077 		 * Do we have enough data in the sb
8078 		 * to handle the GPUT request?
8079 		 */
8080 		uint32_t bytes;
8081 
8082 		bytes = tp->gput_ack - tp->gput_seq;
8083 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
8084 			bytes += tp->gput_seq - tp->snd_una;
8085 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
8086 			/*
8087 			 * There are not enough bytes in the socket
8088 			 * buffer that have been sent to cover this
8089 			 * measurement. Cancel it.
8090 			 */
8091 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
8092 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
8093 						   tp->gput_seq,
8094 						   0, 0, 18, __LINE__, NULL, 0);
8095 			tp->t_flags &= ~TF_GPUTINPROG;
8096 		}
8097 	}
8098 	if (timers == 0) {
8099 		return (0);
8100 	}
8101 	if (tp->t_state == TCPS_LISTEN) {
8102 		/* no timers on listen sockets */
8103 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
8104 			return (0);
8105 		return (1);
8106 	}
8107 	if ((timers & PACE_TMR_RACK) &&
8108 	    rack->rc_on_min_to) {
8109 		/*
8110 		 * For the rack timer when we
8111 		 * are on a min-timeout (which means rrr_conf = 3)
8112 		 * we don't want to check the timer. It may
8113 		 * be going off for a pace and thats ok we
8114 		 * want to send the retransmit (if its ready).
8115 		 *
8116 		 * If its on a normal rack timer (non-min) then
8117 		 * we will check if its expired.
8118 		 */
8119 		goto skip_time_check;
8120 	}
8121 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
8122 		uint32_t left;
8123 
8124 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
8125 			ret = -1;
8126 			rack_log_to_processing(rack, cts, ret, 0);
8127 			return (0);
8128 		}
8129 		if (hpts_calling == 0) {
8130 			/*
8131 			 * A user send or queued mbuf (sack) has called us? We
8132 			 * return 0 and let the pacing guards
8133 			 * deal with it if they should or
8134 			 * should not cause a send.
8135 			 */
8136 			ret = -2;
8137 			rack_log_to_processing(rack, cts, ret, 0);
8138 			return (0);
8139 		}
8140 		/*
8141 		 * Ok our timer went off early and we are not paced false
8142 		 * alarm, go back to sleep. We make sure we don't have
8143 		 * no-sack wakeup on since we no longer have a PKT_OUTPUT
8144 		 * flag in place.
8145 		 */
8146 		rack->rc_tp->t_flags2 &= ~TF2_DONT_SACK_QUEUE;
8147 		ret = -3;
8148 		left = rack->r_ctl.rc_timer_exp - cts;
8149 		tcp_hpts_insert(tp, HPTS_MS_TO_SLOTS(left));
8150 		rack_log_to_processing(rack, cts, ret, left);
8151 		return (1);
8152 	}
8153 skip_time_check:
8154 	rack->rc_tmr_stopped = 0;
8155 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
8156 	if (timers & PACE_TMR_DELACK) {
8157 		ret = rack_timeout_delack(tp, rack, cts);
8158 	} else if (timers & PACE_TMR_RACK) {
8159 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8160 		rack->r_fast_output = 0;
8161 		ret = rack_timeout_rack(tp, rack, cts);
8162 	} else if (timers & PACE_TMR_TLP) {
8163 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8164 		ret = rack_timeout_tlp(tp, rack, cts, doing_tlp);
8165 	} else if (timers & PACE_TMR_RXT) {
8166 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
8167 		rack->r_fast_output = 0;
8168 		ret = rack_timeout_rxt(tp, rack, cts);
8169 	} else if (timers & PACE_TMR_PERSIT) {
8170 		ret = rack_timeout_persist(tp, rack, cts);
8171 	} else if (timers & PACE_TMR_KEEP) {
8172 		ret = rack_timeout_keepalive(tp, rack, cts);
8173 	}
8174 	rack_log_to_processing(rack, cts, ret, timers);
8175 	return (ret);
8176 }
8177 
8178 static void
8179 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
8180 {
8181 	struct timeval tv;
8182 	uint32_t us_cts, flags_on_entry;
8183 	uint8_t hpts_removed = 0;
8184 
8185 	flags_on_entry = rack->r_ctl.rc_hpts_flags;
8186 	us_cts = tcp_get_usecs(&tv);
8187 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
8188 	    ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) ||
8189 	     ((tp->snd_max - tp->snd_una) == 0))) {
8190 		tcp_hpts_remove(rack->rc_tp);
8191 		hpts_removed = 1;
8192 		/* If we were not delayed cancel out the flag. */
8193 		if ((tp->snd_max - tp->snd_una) == 0)
8194 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
8195 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8196 	}
8197 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
8198 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
8199 		if (tcp_in_hpts(rack->rc_tp) &&
8200 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
8201 			/*
8202 			 * Canceling timer's when we have no output being
8203 			 * paced. We also must remove ourselves from the
8204 			 * hpts.
8205 			 */
8206 			tcp_hpts_remove(rack->rc_tp);
8207 			hpts_removed = 1;
8208 		}
8209 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
8210 	}
8211 	if (hpts_removed == 0)
8212 		rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry);
8213 }
8214 
8215 static int
8216 rack_stopall(struct tcpcb *tp)
8217 {
8218 	struct tcp_rack *rack;
8219 
8220 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8221 	rack->t_timers_stopped = 1;
8222 
8223 	tcp_hpts_remove(tp);
8224 
8225 	return (0);
8226 }
8227 
8228 static void
8229 rack_stop_all_timers(struct tcpcb *tp, struct tcp_rack *rack)
8230 {
8231 	/*
8232 	 * Assure no timers are running.
8233 	 */
8234 	if (tcp_timer_active(tp, TT_PERSIST)) {
8235 		/* We enter in persists, set the flag appropriately */
8236 		rack->rc_in_persist = 1;
8237 	}
8238 	if (tcp_in_hpts(rack->rc_tp)) {
8239 		tcp_hpts_remove(rack->rc_tp);
8240 	}
8241 }
8242 
8243 static void
8244 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
8245     struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag, int segsiz)
8246 {
8247 	int32_t idx;
8248 
8249 	rsm->r_rtr_cnt++;
8250 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8251 	rsm->r_dupack = 0;
8252 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
8253 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
8254 		rsm->r_flags |= RACK_OVERMAX;
8255 	}
8256 	if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) {
8257 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
8258 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
8259 	}
8260 	idx = rsm->r_rtr_cnt - 1;
8261 	rsm->r_tim_lastsent[idx] = ts;
8262 	/*
8263 	 * Here we don't add in the len of send, since its already
8264 	 * in snduna <->snd_max.
8265 	 */
8266 	rsm->r_fas = ctf_flight_size(rack->rc_tp,
8267 				     rack->r_ctl.rc_sacked);
8268 	if (rsm->r_flags & RACK_ACKED) {
8269 		/* Problably MTU discovery messing with us */
8270 		rsm->r_flags &= ~RACK_ACKED;
8271 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
8272 	}
8273 	if (rsm->r_in_tmap) {
8274 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8275 		rsm->r_in_tmap = 0;
8276 	}
8277 	/* Lets make sure it really is in or not the GP window */
8278 	rack_mark_in_gp_win(tp, rsm);
8279 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8280 	rsm->r_in_tmap = 1;
8281 	rsm->r_bas = (uint8_t)(((rsm->r_end - rsm->r_start) + segsiz - 1) / segsiz);
8282 	/* Take off the must retransmit flag, if its on */
8283 	if (rsm->r_flags & RACK_MUST_RXT) {
8284 		if (rack->r_must_retran)
8285 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
8286 		if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
8287 			/*
8288 			 * We have retransmitted all we need. Clear
8289 			 * any must retransmit flags.
8290 			 */
8291 			rack->r_must_retran = 0;
8292 			rack->r_ctl.rc_out_at_rto = 0;
8293 		}
8294 		rsm->r_flags &= ~RACK_MUST_RXT;
8295 	}
8296 	/* Remove any collapsed flag */
8297 	rsm->r_flags &= ~RACK_RWND_COLLAPSED;
8298 	if (rsm->r_flags & RACK_SACK_PASSED) {
8299 		/* We have retransmitted due to the SACK pass */
8300 		rsm->r_flags &= ~RACK_SACK_PASSED;
8301 		rsm->r_flags |= RACK_WAS_SACKPASS;
8302 	}
8303 }
8304 
8305 static uint32_t
8306 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
8307     struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag, int segsiz)
8308 {
8309 	/*
8310 	 * We (re-)transmitted starting at rsm->r_start for some length
8311 	 * (possibly less than r_end.
8312 	 */
8313 	struct rack_sendmap *nrsm;
8314 	int insret __diagused;
8315 	uint32_t c_end;
8316 	int32_t len;
8317 
8318 	len = *lenp;
8319 	c_end = rsm->r_start + len;
8320 	if (SEQ_GEQ(c_end, rsm->r_end)) {
8321 		/*
8322 		 * We retransmitted the whole piece or more than the whole
8323 		 * slopping into the next rsm.
8324 		 */
8325 		rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8326 		if (c_end == rsm->r_end) {
8327 			*lenp = 0;
8328 			return (0);
8329 		} else {
8330 			int32_t act_len;
8331 
8332 			/* Hangs over the end return whats left */
8333 			act_len = rsm->r_end - rsm->r_start;
8334 			*lenp = (len - act_len);
8335 			return (rsm->r_end);
8336 		}
8337 		/* We don't get out of this block. */
8338 	}
8339 	/*
8340 	 * Here we retransmitted less than the whole thing which means we
8341 	 * have to split this into what was transmitted and what was not.
8342 	 */
8343 	nrsm = rack_alloc_full_limit(rack);
8344 	if (nrsm == NULL) {
8345 		/*
8346 		 * We can't get memory, so lets not proceed.
8347 		 */
8348 		*lenp = 0;
8349 		return (0);
8350 	}
8351 	/*
8352 	 * So here we are going to take the original rsm and make it what we
8353 	 * retransmitted. nrsm will be the tail portion we did not
8354 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
8355 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
8356 	 * 1, 6 and the new piece will be 6, 11.
8357 	 */
8358 	rack_clone_rsm(rack, nrsm, rsm, c_end);
8359 	nrsm->r_dupack = 0;
8360 	rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
8361 #ifndef INVARIANTS
8362 	(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8363 #else
8364 	if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8365 		panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8366 		      nrsm, insret, rack, rsm);
8367 	}
8368 #endif
8369 	if (rsm->r_in_tmap) {
8370 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8371 		nrsm->r_in_tmap = 1;
8372 	}
8373 	rsm->r_flags &= (~RACK_HAS_FIN);
8374 	rack_update_rsm(tp, rack, rsm, ts, add_flag, segsiz);
8375 	/* Log a split of rsm into rsm and nrsm */
8376 	rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8377 	*lenp = 0;
8378 	return (0);
8379 }
8380 
8381 static void
8382 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
8383 		uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts,
8384 		struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb,
8385 		uint32_t s_moff, int hw_tls, int segsiz)
8386 {
8387 	struct tcp_rack *rack;
8388 	struct rack_sendmap *rsm, *nrsm;
8389 	int insret __diagused;
8390 
8391 	register uint32_t snd_max, snd_una;
8392 
8393 	/*
8394 	 * Add to the RACK log of packets in flight or retransmitted. If
8395 	 * there is a TS option we will use the TS echoed, if not we will
8396 	 * grab a TS.
8397 	 *
8398 	 * Retransmissions will increment the count and move the ts to its
8399 	 * proper place. Note that if options do not include TS's then we
8400 	 * won't be able to effectively use the ACK for an RTT on a retran.
8401 	 *
8402 	 * Notes about r_start and r_end. Lets consider a send starting at
8403 	 * sequence 1 for 10 bytes. In such an example the r_start would be
8404 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
8405 	 * This means that r_end is actually the first sequence for the next
8406 	 * slot (11).
8407 	 *
8408 	 */
8409 	/*
8410 	 * If err is set what do we do XXXrrs? should we not add the thing?
8411 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
8412 	 * i.e. proceed with add ** do this for now.
8413 	 */
8414 	INP_WLOCK_ASSERT(tptoinpcb(tp));
8415 	if (err)
8416 		/*
8417 		 * We don't log errors -- we could but snd_max does not
8418 		 * advance in this case either.
8419 		 */
8420 		return;
8421 
8422 	if (th_flags & TH_RST) {
8423 		/*
8424 		 * We don't log resets and we return immediately from
8425 		 * sending
8426 		 */
8427 		return;
8428 	}
8429 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8430 	snd_una = tp->snd_una;
8431 	snd_max = tp->snd_max;
8432 	if (th_flags & (TH_SYN | TH_FIN)) {
8433 		/*
8434 		 * The call to rack_log_output is made before bumping
8435 		 * snd_max. This means we can record one extra byte on a SYN
8436 		 * or FIN if seq_out is adding more on and a FIN is present
8437 		 * (and we are not resending).
8438 		 */
8439 		if ((th_flags & TH_SYN) && (seq_out == tp->iss))
8440 			len++;
8441 		if (th_flags & TH_FIN)
8442 			len++;
8443 		if (SEQ_LT(snd_max, tp->snd_nxt)) {
8444 			/*
8445 			 * The add/update as not been done for the FIN/SYN
8446 			 * yet.
8447 			 */
8448 			snd_max = tp->snd_nxt;
8449 		}
8450 	}
8451 	if (SEQ_LEQ((seq_out + len), snd_una)) {
8452 		/* Are sending an old segment to induce an ack (keep-alive)? */
8453 		return;
8454 	}
8455 	if (SEQ_LT(seq_out, snd_una)) {
8456 		/* huh? should we panic? */
8457 		uint32_t end;
8458 
8459 		end = seq_out + len;
8460 		seq_out = snd_una;
8461 		if (SEQ_GEQ(end, seq_out))
8462 			len = end - seq_out;
8463 		else
8464 			len = 0;
8465 	}
8466 	if (len == 0) {
8467 		/* We don't log zero window probes */
8468 		return;
8469 	}
8470 	if (IN_FASTRECOVERY(tp->t_flags)) {
8471 		rack->r_ctl.rc_prr_out += len;
8472 	}
8473 	/* First question is it a retransmission or new? */
8474 	if (seq_out == snd_max) {
8475 		/* Its new */
8476 		rack_chk_req_and_hybrid_on_out(rack, seq_out, len, cts);
8477 again:
8478 		rsm = rack_alloc(rack);
8479 		if (rsm == NULL) {
8480 			/*
8481 			 * Hmm out of memory and the tcb got destroyed while
8482 			 * we tried to wait.
8483 			 */
8484 			return;
8485 		}
8486 		if (th_flags & TH_FIN) {
8487 			rsm->r_flags = RACK_HAS_FIN|add_flag;
8488 		} else {
8489 			rsm->r_flags = add_flag;
8490 		}
8491 		if (hw_tls)
8492 			rsm->r_hw_tls = 1;
8493 		rsm->r_tim_lastsent[0] = cts;
8494 		rsm->r_rtr_cnt = 1;
8495 		rsm->r_rtr_bytes = 0;
8496 		if (th_flags & TH_SYN) {
8497 			/* The data space is one beyond snd_una */
8498 			rsm->r_flags |= RACK_HAS_SYN;
8499 		}
8500 		rsm->r_start = seq_out;
8501 		rsm->r_end = rsm->r_start + len;
8502 		rack_mark_in_gp_win(tp, rsm);
8503 		rsm->r_dupack = 0;
8504 		/*
8505 		 * save off the mbuf location that
8506 		 * sndmbuf_noadv returned (which is
8507 		 * where we started copying from)..
8508 		 */
8509 		rsm->m = s_mb;
8510 		rsm->soff = s_moff;
8511 		/*
8512 		 * Here we do add in the len of send, since its not yet
8513 		 * reflected in in snduna <->snd_max
8514 		 */
8515 		rsm->r_fas = (ctf_flight_size(rack->rc_tp,
8516 					      rack->r_ctl.rc_sacked) +
8517 			      (rsm->r_end - rsm->r_start));
8518 		/* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */
8519 		if (rsm->m) {
8520 			if (rsm->m->m_len <= rsm->soff) {
8521 				/*
8522 				 * XXXrrs Question, will this happen?
8523 				 *
8524 				 * If sbsndptr is set at the correct place
8525 				 * then s_moff should always be somewhere
8526 				 * within rsm->m. But if the sbsndptr was
8527 				 * off then that won't be true. If it occurs
8528 				 * we need to walkout to the correct location.
8529 				 */
8530 				struct mbuf *lm;
8531 
8532 				lm = rsm->m;
8533 				while (lm->m_len <= rsm->soff) {
8534 					rsm->soff -= lm->m_len;
8535 					lm = lm->m_next;
8536 					KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u",
8537 							     __func__, rack, s_moff, s_mb, rsm->soff));
8538 				}
8539 				rsm->m = lm;
8540 			}
8541 			rsm->orig_m_len = rsm->m->m_len;
8542 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
8543 		} else {
8544 			rsm->orig_m_len = 0;
8545 			rsm->orig_t_space = 0;
8546 		}
8547 		rsm->r_bas = (uint8_t)((len + segsiz - 1) / segsiz);
8548 		rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
8549 		/* Log a new rsm */
8550 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__);
8551 #ifndef INVARIANTS
8552 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
8553 #else
8554 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
8555 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8556 			      nrsm, insret, rack, rsm);
8557 		}
8558 #endif
8559 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
8560 		rsm->r_in_tmap = 1;
8561 		/*
8562 		 * Special case detection, is there just a single
8563 		 * packet outstanding when we are not in recovery?
8564 		 *
8565 		 * If this is true mark it so.
8566 		 */
8567 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
8568 		    (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) {
8569 			struct rack_sendmap *prsm;
8570 
8571 			prsm = tqhash_prev(rack->r_ctl.tqh, rsm);
8572 			if (prsm)
8573 				prsm->r_one_out_nr = 1;
8574 		}
8575 		return;
8576 	}
8577 	/*
8578 	 * If we reach here its a retransmission and we need to find it.
8579 	 */
8580 more:
8581 	if (hintrsm && (hintrsm->r_start == seq_out)) {
8582 		rsm = hintrsm;
8583 		hintrsm = NULL;
8584 	} else {
8585 		/* No hints sorry */
8586 		rsm = NULL;
8587 	}
8588 	if ((rsm) && (rsm->r_start == seq_out)) {
8589 		seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8590 		if (len == 0) {
8591 			return;
8592 		} else {
8593 			goto more;
8594 		}
8595 	}
8596 	/* Ok it was not the last pointer go through it the hard way. */
8597 refind:
8598 	rsm = tqhash_find(rack->r_ctl.tqh, seq_out);
8599 	if (rsm) {
8600 		if (rsm->r_start == seq_out) {
8601 			seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag, segsiz);
8602 			if (len == 0) {
8603 				return;
8604 			} else {
8605 				goto refind;
8606 			}
8607 		}
8608 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
8609 			/* Transmitted within this piece */
8610 			/*
8611 			 * Ok we must split off the front and then let the
8612 			 * update do the rest
8613 			 */
8614 			nrsm = rack_alloc_full_limit(rack);
8615 			if (nrsm == NULL) {
8616 				rack_update_rsm(tp, rack, rsm, cts, add_flag, segsiz);
8617 				return;
8618 			}
8619 			/*
8620 			 * copy rsm to nrsm and then trim the front of rsm
8621 			 * to not include this part.
8622 			 */
8623 			rack_clone_rsm(rack, nrsm, rsm, seq_out);
8624 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__);
8625 #ifndef INVARIANTS
8626 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
8627 #else
8628 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
8629 				panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
8630 				      nrsm, insret, rack, rsm);
8631 			}
8632 #endif
8633 			if (rsm->r_in_tmap) {
8634 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
8635 				nrsm->r_in_tmap = 1;
8636 			}
8637 			rsm->r_flags &= (~RACK_HAS_FIN);
8638 			seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag, segsiz);
8639 			if (len == 0) {
8640 				return;
8641 			} else if (len > 0)
8642 				goto refind;
8643 		}
8644 	}
8645 	/*
8646 	 * Hmm not found in map did they retransmit both old and on into the
8647 	 * new?
8648 	 */
8649 	if (seq_out == tp->snd_max) {
8650 		goto again;
8651 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
8652 #ifdef INVARIANTS
8653 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
8654 		       seq_out, len, tp->snd_una, tp->snd_max);
8655 		printf("Starting Dump of all rack entries\n");
8656 		TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
8657 			printf("rsm:%p start:%u end:%u\n",
8658 			       rsm, rsm->r_start, rsm->r_end);
8659 		}
8660 		printf("Dump complete\n");
8661 		panic("seq_out not found rack:%p tp:%p",
8662 		      rack, tp);
8663 #endif
8664 	} else {
8665 #ifdef INVARIANTS
8666 		/*
8667 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
8668 		 * flag)
8669 		 */
8670 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
8671 		      seq_out, len, tp->snd_max, tp);
8672 #endif
8673 	}
8674 }
8675 
8676 /*
8677  * Record one of the RTT updates from an ack into
8678  * our sample structure.
8679  */
8680 
8681 static void
8682 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt,
8683 		    int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt)
8684 {
8685 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8686 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
8687 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
8688 	}
8689 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8690 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
8691 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
8692 	}
8693 	if (rack->rc_tp->t_flags & TF_GPUTINPROG) {
8694 	    if (us_rtt < rack->r_ctl.rc_gp_lowrtt)
8695 		rack->r_ctl.rc_gp_lowrtt = us_rtt;
8696 	    if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd)
8697 		    rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
8698 	}
8699 	if ((confidence == 1) &&
8700 	    ((rsm == NULL) ||
8701 	     (rsm->r_just_ret) ||
8702 	     (rsm->r_one_out_nr &&
8703 	      len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) {
8704 		/*
8705 		 * If the rsm had a just return
8706 		 * hit it then we can't trust the
8707 		 * rtt measurement for buffer deterimination
8708 		 * Note that a confidence of 2, indicates
8709 		 * SACK'd which overrides the r_just_ret or
8710 		 * the r_one_out_nr. If it was a CUM-ACK and
8711 		 * we had only two outstanding, but get an
8712 		 * ack for only 1. Then that also lowers our
8713 		 * confidence.
8714 		 */
8715 		confidence = 0;
8716 	}
8717 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
8718 	    (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) {
8719 		if (rack->r_ctl.rack_rs.confidence == 0) {
8720 			/*
8721 			 * We take anything with no current confidence
8722 			 * saved.
8723 			 */
8724 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8725 			rack->r_ctl.rack_rs.confidence = confidence;
8726 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8727 		} else if (confidence != 0) {
8728 			/*
8729 			 * Once we have a confident number,
8730 			 * we can update it with a smaller
8731 			 * value since this confident number
8732 			 * may include the DSACK time until
8733 			 * the next segment (the second one) arrived.
8734 			 */
8735 			rack->r_ctl.rack_rs.rs_us_rtt = us_rtt;
8736 			rack->r_ctl.rack_rs.confidence = confidence;
8737 			rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt;
8738 		}
8739 	}
8740 	rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence);
8741 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
8742 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
8743 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
8744 }
8745 
8746 /*
8747  * Collect new round-trip time estimate
8748  * and update averages and current timeout.
8749  */
8750 static void
8751 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
8752 {
8753 	int32_t delta;
8754 	int32_t rtt;
8755 
8756 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
8757 		/* No valid sample */
8758 		return;
8759 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
8760 		/* We are to use the lowest RTT seen in a single ack */
8761 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
8762 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
8763 		/* We are to use the highest RTT seen in a single ack */
8764 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
8765 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
8766 		/* We are to use the average RTT seen in a single ack */
8767 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
8768 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
8769 	} else {
8770 #ifdef INVARIANTS
8771 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
8772 #endif
8773 		return;
8774 	}
8775 	if (rtt == 0)
8776 		rtt = 1;
8777 	if (rack->rc_gp_rtt_set == 0) {
8778 		/*
8779 		 * With no RTT we have to accept
8780 		 * even one we are not confident of.
8781 		 */
8782 		rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt;
8783 		rack->rc_gp_rtt_set = 1;
8784 	} else if (rack->r_ctl.rack_rs.confidence) {
8785 		/* update the running gp srtt */
8786 		rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8);
8787 		rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8;
8788 	}
8789 	if (rack->r_ctl.rack_rs.confidence) {
8790 		/*
8791 		 * record the low and high for highly buffered path computation,
8792 		 * we only do this if we are confident (not a retransmission).
8793 		 */
8794 		if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) {
8795 			rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8796 		}
8797 		if (rack->rc_highly_buffered == 0) {
8798 			/*
8799 			 * Currently once we declare a path has
8800 			 * highly buffered there is no going
8801 			 * back, which may be a problem...
8802 			 */
8803 			if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) {
8804 				rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt,
8805 						     rack->r_ctl.rc_highest_us_rtt,
8806 						     rack->r_ctl.rc_lowest_us_rtt,
8807 						     RACK_RTTS_SEEHBP);
8808 				rack->rc_highly_buffered = 1;
8809 			}
8810 		}
8811 	}
8812 	if ((rack->r_ctl.rack_rs.confidence) ||
8813 	    (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) {
8814 		/*
8815 		 * If we are highly confident of it <or> it was
8816 		 * never retransmitted we accept it as the last us_rtt.
8817 		 */
8818 		rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8819 		/* The lowest rtt can be set if its was not retransmited */
8820 		if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) {
8821 			rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt;
8822 			if (rack->r_ctl.rc_lowest_us_rtt == 0)
8823 				rack->r_ctl.rc_lowest_us_rtt = 1;
8824 		}
8825 	}
8826 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8827 	if (tp->t_srtt != 0) {
8828 		/*
8829 		 * We keep a simple srtt in microseconds, like our rtt
8830 		 * measurement. We don't need to do any tricks with shifting
8831 		 * etc. Instead we just add in 1/8th of the new measurement
8832 		 * and subtract out 1/8 of the old srtt. We do the same with
8833 		 * the variance after finding the absolute value of the
8834 		 * difference between this sample and the current srtt.
8835 		 */
8836 		delta = tp->t_srtt - rtt;
8837 		/* Take off 1/8th of the current sRTT */
8838 		tp->t_srtt -= (tp->t_srtt >> 3);
8839 		/* Add in 1/8th of the new RTT just measured */
8840 		tp->t_srtt += (rtt >> 3);
8841 		if (tp->t_srtt <= 0)
8842 			tp->t_srtt = 1;
8843 		/* Now lets make the absolute value of the variance */
8844 		if (delta < 0)
8845 			delta = -delta;
8846 		/* Subtract out 1/8th */
8847 		tp->t_rttvar -= (tp->t_rttvar >> 3);
8848 		/* Add in 1/8th of the new variance we just saw */
8849 		tp->t_rttvar += (delta >> 3);
8850 		if (tp->t_rttvar <= 0)
8851 			tp->t_rttvar = 1;
8852 	} else {
8853 		/*
8854 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
8855 		 * variance to half the rtt (so our first retransmit happens
8856 		 * at 3*rtt).
8857 		 */
8858 		tp->t_srtt = rtt;
8859 		tp->t_rttvar = rtt >> 1;
8860 	}
8861 	rack->rc_srtt_measure_made = 1;
8862 	KMOD_TCPSTAT_INC(tcps_rttupdated);
8863 	if (tp->t_rttupdated < UCHAR_MAX)
8864 		tp->t_rttupdated++;
8865 #ifdef STATS
8866 	if (rack_stats_gets_ms_rtt == 0) {
8867 		/* Send in the microsecond rtt used for rxt timeout purposes */
8868 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
8869 	} else if (rack_stats_gets_ms_rtt == 1) {
8870 		/* Send in the millisecond rtt used for rxt timeout purposes */
8871 		int32_t ms_rtt;
8872 
8873 		/* Round up */
8874 		ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8875 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8876 	} else if (rack_stats_gets_ms_rtt == 2) {
8877 		/* Send in the millisecond rtt has close to the path RTT as we can get  */
8878 		int32_t ms_rtt;
8879 
8880 		/* Round up */
8881 		ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC;
8882 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt));
8883 	}  else {
8884 		/* Send in the microsecond rtt has close to the path RTT as we can get  */
8885 		stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8886 	}
8887 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_PATHRTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt));
8888 #endif
8889 	/*
8890 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
8891 	 * way we do the smoothing, srtt and rttvar will each average +1/2
8892 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
8893 	 * tick of rounding and 1 extra tick because of +-1/2 tick
8894 	 * uncertainty in the firing of the timer.  The bias will give us
8895 	 * exactly the 1.5 tick we need.  But, because the bias is
8896 	 * statistical, we have to test that we don't drop below the minimum
8897 	 * feasible timer (which is 2 ticks).
8898 	 */
8899 	tp->t_rxtshift = 0;
8900 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
8901 		      max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop);
8902 	rack_log_rtt_sample(rack, rtt);
8903 	tp->t_softerror = 0;
8904 }
8905 
8906 
8907 static void
8908 rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts)
8909 {
8910 	/*
8911 	 * Apply to filter the inbound us-rtt at us_cts.
8912 	 */
8913 	uint32_t old_rtt;
8914 
8915 	old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt);
8916 	apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt,
8917 			       us_rtt, us_cts);
8918 	if (old_rtt > us_rtt) {
8919 		/* We just hit a new lower rtt time */
8920 		rack_log_rtt_shrinks(rack,  us_cts,  old_rtt,
8921 				     __LINE__, RACK_RTTS_NEWRTT);
8922 		/*
8923 		 * Only count it if its lower than what we saw within our
8924 		 * calculated range.
8925 		 */
8926 		if ((old_rtt - us_rtt) > rack_min_rtt_movement) {
8927 			if (rack_probertt_lower_within &&
8928 			    rack->rc_gp_dyn_mul &&
8929 			    (rack->use_fixed_rate == 0) &&
8930 			    (rack->rc_always_pace)) {
8931 				/*
8932 				 * We are seeing a new lower rtt very close
8933 				 * to the time that we would have entered probe-rtt.
8934 				 * This is probably due to the fact that a peer flow
8935 				 * has entered probe-rtt. Lets go in now too.
8936 				 */
8937 				uint32_t val;
8938 
8939 				val = rack_probertt_lower_within * rack_time_between_probertt;
8940 				val /= 100;
8941 				if ((rack->in_probe_rtt == 0)  &&
8942 				    ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val)))	{
8943 					rack_enter_probertt(rack, us_cts);
8944 				}
8945 			}
8946 			rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
8947 		}
8948 	}
8949 }
8950 
8951 static int
8952 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
8953     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack)
8954 {
8955 	uint32_t us_rtt;
8956 	int32_t i, all;
8957 	uint32_t t, len_acked;
8958 
8959 	if ((rsm->r_flags & RACK_ACKED) ||
8960 	    (rsm->r_flags & RACK_WAS_ACKED))
8961 		/* Already done */
8962 		return (0);
8963 	if (rsm->r_no_rtt_allowed) {
8964 		/* Not allowed */
8965 		return (0);
8966 	}
8967 	if (ack_type == CUM_ACKED) {
8968 		if (SEQ_GT(th_ack, rsm->r_end)) {
8969 			len_acked = rsm->r_end - rsm->r_start;
8970 			all = 1;
8971 		} else {
8972 			len_acked = th_ack - rsm->r_start;
8973 			all = 0;
8974 		}
8975 	} else {
8976 		len_acked = rsm->r_end - rsm->r_start;
8977 		all = 0;
8978 	}
8979 	if (rsm->r_rtr_cnt == 1) {
8980 
8981 		t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
8982 		if ((int)t <= 0)
8983 			t = 1;
8984 		if (!tp->t_rttlow || tp->t_rttlow > t)
8985 			tp->t_rttlow = t;
8986 		if (!rack->r_ctl.rc_rack_min_rtt ||
8987 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
8988 			rack->r_ctl.rc_rack_min_rtt = t;
8989 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
8990 				rack->r_ctl.rc_rack_min_rtt = 1;
8991 			}
8992 		}
8993 		if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]))
8994 			us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8995 		else
8996 			us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
8997 		if (us_rtt == 0)
8998 			us_rtt = 1;
8999 		if (CC_ALGO(tp)->rttsample != NULL) {
9000 			/* Kick the RTT to the CC */
9001 			CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
9002 		}
9003 		rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
9004 		if (ack_type == SACKED) {
9005 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1);
9006 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt);
9007 		} else {
9008 			/*
9009 			 * We need to setup what our confidence
9010 			 * is in this ack.
9011 			 *
9012 			 * If the rsm was app limited and it is
9013 			 * less than a mss in length (the end
9014 			 * of the send) then we have a gap. If we
9015 			 * were app limited but say we were sending
9016 			 * multiple MSS's then we are more confident
9017 			 * int it.
9018 			 *
9019 			 * When we are not app-limited then we see if
9020 			 * the rsm is being included in the current
9021 			 * measurement, we tell this by the app_limited_needs_set
9022 			 * flag.
9023 			 *
9024 			 * Note that being cwnd blocked is not applimited
9025 			 * as well as the pacing delay between packets which
9026 			 * are sending only 1 or 2 MSS's also will show up
9027 			 * in the RTT. We probably need to examine this algorithm
9028 			 * a bit more and enhance it to account for the delay
9029 			 * between rsm's. We could do that by saving off the
9030 			 * pacing delay of each rsm (in an rsm) and then
9031 			 * factoring that in somehow though for now I am
9032 			 * not sure how :)
9033 			 */
9034 			int calc_conf = 0;
9035 
9036 			if (rsm->r_flags & RACK_APP_LIMITED) {
9037 				if (all && (len_acked <= ctf_fixed_maxseg(tp)))
9038 					calc_conf = 0;
9039 				else
9040 					calc_conf = 1;
9041 			} else if (rack->app_limited_needs_set == 0) {
9042 				calc_conf = 1;
9043 			} else {
9044 				calc_conf = 0;
9045 			}
9046 			rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2);
9047 			tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt,
9048 					    calc_conf, rsm, rsm->r_rtr_cnt);
9049 		}
9050 		if ((rsm->r_flags & RACK_TLP) &&
9051 		    (!IN_FASTRECOVERY(tp->t_flags))) {
9052 			/* Segment was a TLP and our retrans matched */
9053 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
9054 				rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
9055 			}
9056 		}
9057 		if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9058 		    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9059 			    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9060 			/* New more recent rack_tmit_time */
9061 			rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9062 			if (rack->r_ctl.rc_rack_tmit_time == 0)
9063 				rack->r_ctl.rc_rack_tmit_time = 1;
9064 			rack->rc_rack_rtt = t;
9065 		}
9066 		return (1);
9067 	}
9068 	/*
9069 	 * We clear the soft/rxtshift since we got an ack.
9070 	 * There is no assurance we will call the commit() function
9071 	 * so we need to clear these to avoid incorrect handling.
9072 	 */
9073 	tp->t_rxtshift = 0;
9074 	RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
9075 		      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
9076 	tp->t_softerror = 0;
9077 	if (to && (to->to_flags & TOF_TS) &&
9078 	    (ack_type == CUM_ACKED) &&
9079 	    (to->to_tsecr) &&
9080 	    ((rsm->r_flags & RACK_OVERMAX) == 0)) {
9081 		/*
9082 		 * Now which timestamp does it match? In this block the ACK
9083 		 * must be coming from a previous transmission.
9084 		 */
9085 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
9086 			if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) {
9087 				t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9088 				if ((int)t <= 0)
9089 					t = 1;
9090 				if (CC_ALGO(tp)->rttsample != NULL) {
9091 					/*
9092 					 * Kick the RTT to the CC, here
9093 					 * we lie a bit in that we know the
9094 					 * retransmission is correct even though
9095 					 * we retransmitted. This is because
9096 					 * we match the timestamps.
9097 					 */
9098 					if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i]))
9099 						us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i];
9100 					else
9101 						us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i];
9102 					CC_ALGO(tp)->rttsample(&tp->t_ccv, us_rtt, 1, rsm->r_fas);
9103 				}
9104 				if ((i + 1) < rsm->r_rtr_cnt) {
9105 					/*
9106 					 * The peer ack'd from our previous
9107 					 * transmission. We have a spurious
9108 					 * retransmission and thus we dont
9109 					 * want to update our rack_rtt.
9110 					 *
9111 					 * Hmm should there be a CC revert here?
9112 					 *
9113 					 */
9114 					return (0);
9115 				}
9116 				if (!tp->t_rttlow || tp->t_rttlow > t)
9117 					tp->t_rttlow = t;
9118 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9119 					rack->r_ctl.rc_rack_min_rtt = t;
9120 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
9121 						rack->r_ctl.rc_rack_min_rtt = 1;
9122 					}
9123 				}
9124 				if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9125 				    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9126 					    (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]))) {
9127 					/* New more recent rack_tmit_time */
9128 					rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
9129 					if (rack->r_ctl.rc_rack_tmit_time == 0)
9130 						rack->r_ctl.rc_rack_tmit_time = 1;
9131 					rack->rc_rack_rtt = t;
9132 				}
9133 				rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3);
9134 				tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm,
9135 						    rsm->r_rtr_cnt);
9136 				return (1);
9137 			}
9138 		}
9139 		/* If we are logging log out the sendmap */
9140 		if (tcp_bblogging_on(rack->rc_tp)) {
9141 			for (i = 0; i < rsm->r_rtr_cnt; i++) {
9142 				rack_log_rtt_sendmap(rack, i, rsm->r_tim_lastsent[i], to->to_tsecr);
9143 			}
9144 		}
9145 		goto ts_not_found;
9146 	} else {
9147 		/*
9148 		 * Ok its a SACK block that we retransmitted. or a windows
9149 		 * machine without timestamps. We can tell nothing from the
9150 		 * time-stamp since its not there or the time the peer last
9151 		 * received a segment that moved forward its cum-ack point.
9152 		 */
9153 ts_not_found:
9154 		i = rsm->r_rtr_cnt - 1;
9155 		t = cts - (uint32_t)rsm->r_tim_lastsent[i];
9156 		if ((int)t <= 0)
9157 			t = 1;
9158 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9159 			/*
9160 			 * We retransmitted and the ack came back in less
9161 			 * than the smallest rtt we have observed. We most
9162 			 * likely did an improper retransmit as outlined in
9163 			 * 6.2 Step 2 point 2 in the rack-draft so we
9164 			 * don't want to update our rack_rtt. We in
9165 			 * theory (in future) might want to think about reverting our
9166 			 * cwnd state but we won't for now.
9167 			 */
9168 			return (0);
9169 		} else if (rack->r_ctl.rc_rack_min_rtt) {
9170 			/*
9171 			 * We retransmitted it and the retransmit did the
9172 			 * job.
9173 			 */
9174 			if (!rack->r_ctl.rc_rack_min_rtt ||
9175 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
9176 				rack->r_ctl.rc_rack_min_rtt = t;
9177 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
9178 					rack->r_ctl.rc_rack_min_rtt = 1;
9179 				}
9180 			}
9181 			if ((rack->r_ctl.rc_rack_tmit_time == 0) ||
9182 			    (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
9183 				    (uint32_t)rsm->r_tim_lastsent[i]))) {
9184 				/* New more recent rack_tmit_time */
9185 				rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i];
9186 				if (rack->r_ctl.rc_rack_tmit_time == 0)
9187 					rack->r_ctl.rc_rack_tmit_time = 1;
9188 				rack->rc_rack_rtt = t;
9189 			}
9190 			return (1);
9191 		}
9192 	}
9193 	return (0);
9194 }
9195 
9196 /*
9197  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
9198  */
9199 static void
9200 rack_log_sack_passed(struct tcpcb *tp,
9201     struct tcp_rack *rack, struct rack_sendmap *rsm)
9202 {
9203 	struct rack_sendmap *nrsm;
9204 
9205 	nrsm = rsm;
9206 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
9207 	    rack_head, r_tnext) {
9208 		if (nrsm == rsm) {
9209 			/* Skip original segment he is acked */
9210 			continue;
9211 		}
9212 		if (nrsm->r_flags & RACK_ACKED) {
9213 			/*
9214 			 * Skip ack'd segments, though we
9215 			 * should not see these, since tmap
9216 			 * should not have ack'd segments.
9217 			 */
9218 			continue;
9219 		}
9220 		if (nrsm->r_flags & RACK_RWND_COLLAPSED) {
9221 			/*
9222 			 * If the peer dropped the rwnd on
9223 			 * these then we don't worry about them.
9224 			 */
9225 			continue;
9226 		}
9227 		if (nrsm->r_flags & RACK_SACK_PASSED) {
9228 			/*
9229 			 * We found one that is already marked
9230 			 * passed, we have been here before and
9231 			 * so all others below this are marked.
9232 			 */
9233 			break;
9234 		}
9235 		nrsm->r_flags |= RACK_SACK_PASSED;
9236 		nrsm->r_flags &= ~RACK_WAS_SACKPASS;
9237 	}
9238 }
9239 
9240 static void
9241 rack_need_set_test(struct tcpcb *tp,
9242 		   struct tcp_rack *rack,
9243 		   struct rack_sendmap *rsm,
9244 		   tcp_seq th_ack,
9245 		   int line,
9246 		   int use_which)
9247 {
9248 	struct rack_sendmap *s_rsm;
9249 
9250 	if ((tp->t_flags & TF_GPUTINPROG) &&
9251 	    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9252 		/*
9253 		 * We were app limited, and this ack
9254 		 * butts up or goes beyond the point where we want
9255 		 * to start our next measurement. We need
9256 		 * to record the new gput_ts as here and
9257 		 * possibly update the start sequence.
9258 		 */
9259 		uint32_t seq, ts;
9260 
9261 		if (rsm->r_rtr_cnt > 1) {
9262 			/*
9263 			 * This is a retransmit, can we
9264 			 * really make any assessment at this
9265 			 * point?  We are not really sure of
9266 			 * the timestamp, is it this or the
9267 			 * previous transmission?
9268 			 *
9269 			 * Lets wait for something better that
9270 			 * is not retransmitted.
9271 			 */
9272 			return;
9273 		}
9274 		seq = tp->gput_seq;
9275 		ts = tp->gput_ts;
9276 		rack->app_limited_needs_set = 0;
9277 		tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
9278 		/* Do we start at a new end? */
9279 		if ((use_which == RACK_USE_BEG) &&
9280 		    SEQ_GEQ(rsm->r_start, tp->gput_seq)) {
9281 			/*
9282 			 * When we get an ACK that just eats
9283 			 * up some of the rsm, we set RACK_USE_BEG
9284 			 * since whats at r_start (i.e. th_ack)
9285 			 * is left unacked and thats where the
9286 			 * measurement now starts.
9287 			 */
9288 			tp->gput_seq = rsm->r_start;
9289 		}
9290 		if ((use_which == RACK_USE_END) &&
9291 		    SEQ_GEQ(rsm->r_end, tp->gput_seq)) {
9292 			/*
9293 			 * We use the end when the cumack
9294 			 * is moving forward and completely
9295 			 * deleting the rsm passed so basically
9296 			 * r_end holds th_ack.
9297 			 *
9298 			 * For SACK's we also want to use the end
9299 			 * since this piece just got sacked and
9300 			 * we want to target anything after that
9301 			 * in our measurement.
9302 			 */
9303 			tp->gput_seq = rsm->r_end;
9304 		}
9305 		if (use_which == RACK_USE_END_OR_THACK) {
9306 			/*
9307 			 * special case for ack moving forward,
9308 			 * not a sack, we need to move all the
9309 			 * way up to where this ack cum-ack moves
9310 			 * to.
9311 			 */
9312 			if (SEQ_GT(th_ack, rsm->r_end))
9313 				tp->gput_seq = th_ack;
9314 			else
9315 				tp->gput_seq = rsm->r_end;
9316 		}
9317 		if (SEQ_LT(tp->gput_seq, tp->snd_max))
9318 			s_rsm = tqhash_find(rack->r_ctl.tqh, tp->gput_seq);
9319 		else
9320 			s_rsm = NULL;
9321 		/*
9322 		 * Pick up the correct send time if we can the rsm passed in
9323 		 * may be equal to s_rsm if the RACK_USE_BEG was set. For the other
9324 		 * two cases (RACK_USE_THACK or RACK_USE_END) most likely we will
9325 		 * find a different seq i.e. the next send up.
9326 		 *
9327 		 * If that has not been sent, s_rsm will be NULL and we must
9328 		 * arrange it so this function will get called again by setting
9329 		 * app_limited_needs_set.
9330 		 */
9331 		if (s_rsm)
9332 			rack->r_ctl.rc_gp_output_ts = s_rsm->r_tim_lastsent[0];
9333 		else {
9334 			/* If we hit here we have to have *not* sent tp->gput_seq */
9335 			rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[0];
9336 			/* Set it up so we will go through here again */
9337 			rack->app_limited_needs_set = 1;
9338 		}
9339 		if (SEQ_GT(tp->gput_seq, tp->gput_ack)) {
9340 			/*
9341 			 * We moved beyond this guy's range, re-calculate
9342 			 * the new end point.
9343 			 */
9344 			if (rack->rc_gp_filled == 0) {
9345 				tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
9346 			} else {
9347 				tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
9348 			}
9349 		}
9350 		/*
9351 		 * We are moving the goal post, we may be able to clear the
9352 		 * measure_saw_probe_rtt flag.
9353 		 */
9354 		if ((rack->in_probe_rtt == 0) &&
9355 		    (rack->measure_saw_probe_rtt) &&
9356 		    (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
9357 			rack->measure_saw_probe_rtt = 0;
9358 		rack_log_pacing_delay_calc(rack, ts, tp->gput_ts,
9359 					   seq, tp->gput_seq,
9360 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9361 					    (uint64_t)rack->r_ctl.rc_gp_output_ts),
9362 					   5, line, NULL, 0);
9363 		if (rack->rc_gp_filled &&
9364 		    ((tp->gput_ack - tp->gput_seq) <
9365 		     max(rc_init_window(rack), (MIN_GP_WIN *
9366 						ctf_fixed_maxseg(tp))))) {
9367 			uint32_t ideal_amount;
9368 
9369 			ideal_amount = rack_get_measure_window(tp, rack);
9370 			if (ideal_amount > sbavail(&tptosocket(tp)->so_snd)) {
9371 				/*
9372 				 * There is no sense of continuing this measurement
9373 				 * because its too small to gain us anything we
9374 				 * trust. Skip it and that way we can start a new
9375 				 * measurement quicker.
9376 				 */
9377 				tp->t_flags &= ~TF_GPUTINPROG;
9378 				rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq,
9379 							   0, 0,
9380 							   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) |
9381 							    (uint64_t)rack->r_ctl.rc_gp_output_ts),
9382 							   6, __LINE__, NULL, 0);
9383 			} else {
9384 				/*
9385 				 * Reset the window further out.
9386 				 */
9387 				tp->gput_ack = tp->gput_seq + ideal_amount;
9388 			}
9389 		}
9390 		rack_tend_gp_marks(tp, rack);
9391 		rack_log_gpset(rack, tp->gput_ack, 0, 0, line, 2, rsm);
9392 	}
9393 }
9394 
9395 static inline int
9396 is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm)
9397 {
9398 	if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) {
9399 		/* Behind our TLP definition or right at */
9400 		return (0);
9401 	}
9402 	if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) {
9403 		/* The start is beyond or right at our end of TLP definition */
9404 		return (0);
9405 	}
9406 	/* It has to be a sub-part of the original TLP recorded */
9407 	return (1);
9408 }
9409 
9410 
9411 
9412 static uint32_t
9413 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
9414 		   struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts,
9415 		   int *no_extra,
9416 		   int *moved_two, uint32_t segsiz)
9417 {
9418 	uint32_t start, end, changed = 0;
9419 	struct rack_sendmap stack_map;
9420 	struct rack_sendmap *rsm, *nrsm, *prev, *next;
9421 	int insret __diagused;
9422 	int32_t used_ref = 1;
9423 	int moved = 0;
9424 #ifdef TCP_SAD_DETECTION
9425 	int allow_segsiz;
9426 	int first_time_through = 1;
9427 #endif
9428 	int noextra = 0;
9429 	int can_use_hookery = 0;
9430 
9431 	start = sack->start;
9432 	end = sack->end;
9433 	rsm = *prsm;
9434 
9435 #ifdef TCP_SAD_DETECTION
9436 	/*
9437 	 * There are a strange number of proxys and meddle boxes in the world
9438 	 * that seem to cut up segments on different boundaries. This gets us
9439 	 * smaller sacks that are still ok in terms of it being an attacker.
9440 	 * We use the base segsiz to calculate an allowable smallness but
9441 	 * also enforce a min on the segsiz in case it is an attacker playing
9442 	 * games with MSS. So basically if the sack arrives and it is
9443 	 * larger than a worse case 960 bytes, we don't classify the guy
9444 	 * as supicious.
9445 	 */
9446 	allow_segsiz = max(segsiz, 1200) * sad_seg_size_per;
9447 	allow_segsiz /= 1000;
9448 #endif
9449 do_rest_ofb:
9450 	if ((rsm == NULL) ||
9451 	    (SEQ_LT(end, rsm->r_start)) ||
9452 	    (SEQ_GEQ(start, rsm->r_end)) ||
9453 	    (SEQ_LT(start, rsm->r_start))) {
9454 		/*
9455 		 * We are not in the right spot,
9456 		 * find the correct spot in the tree.
9457 		 */
9458 		used_ref = 0;
9459 		rsm = tqhash_find(rack->r_ctl.tqh, start);
9460 		moved++;
9461 	}
9462 	if (rsm == NULL) {
9463 		/* TSNH */
9464 		goto out;
9465 	}
9466 #ifdef TCP_SAD_DETECTION
9467 	/* Now we must check for suspicous activity */
9468 	if ((first_time_through == 1) &&
9469 	    ((end - start) < min((rsm->r_end - rsm->r_start), allow_segsiz)) &&
9470 	    ((rsm->r_flags & RACK_PMTU_CHG) == 0) &&
9471 	    ((rsm->r_flags & RACK_TLP) == 0)) {
9472 		/*
9473 		 * Its less than a full MSS or the segment being acked
9474 		 * this should only happen if the rsm in question had the
9475 		 * r_just_ret flag set <and> the end matches the end of
9476 		 * the rsm block.
9477 		 *
9478 		 * Note we do not look at segments that have had TLP's on
9479 		 * them since we can get un-reported rwnd collapses that
9480 		 * basically we TLP on and then we get back a sack block
9481 		 * that goes from the start to only a small way.
9482 		 *
9483 		 */
9484 		int loss, ok;
9485 
9486 		ok = 0;
9487 		if (SEQ_GEQ(end, rsm->r_end)) {
9488 			if (rsm->r_just_ret == 1) {
9489 				/* This was at the end of a send which is ok */
9490 				ok = 1;
9491 			} else {
9492 				/* A bit harder was it the end of our segment */
9493 				int segs, len;
9494 
9495 				len = (rsm->r_end - rsm->r_start);
9496 				segs = len / segsiz;
9497 				segs *= segsiz;
9498 				if ((segs + (rsm->r_end - start)) == len) {
9499 					/*
9500 					 * So this last bit was the
9501 					 * end of our send if we cut it
9502 					 * up into segsiz pieces so its ok.
9503 					 */
9504 					ok = 1;
9505 				}
9506 			}
9507 		}
9508 		if (ok == 0) {
9509 			/*
9510 			 * This guy is doing something suspicious
9511 			 * lets start detection.
9512 			 */
9513 			if (rack->rc_suspicious == 0) {
9514 				tcp_trace_point(rack->rc_tp, TCP_TP_SAD_SUSPECT);
9515 				counter_u64_add(rack_sack_attacks_suspect, 1);
9516 				rack->rc_suspicious = 1;
9517 				rack_log_sad(rack, 4);
9518 				if (tcp_bblogging_on(rack->rc_tp)) {
9519 					union tcp_log_stackspecific log;
9520 					struct timeval tv;
9521 
9522 					memset(&log.u_bbr, 0, sizeof(log.u_bbr));
9523 					log.u_bbr.flex1 = end;
9524 					log.u_bbr.flex2 = start;
9525 					log.u_bbr.flex3 = rsm->r_end;
9526 					log.u_bbr.flex4 = rsm->r_start;
9527 					log.u_bbr.flex5 = segsiz;
9528 					log.u_bbr.flex6 = rsm->r_fas;
9529 					log.u_bbr.flex7 = rsm->r_bas;
9530 					log.u_bbr.flex8 = 5;
9531 					log.u_bbr.pkts_out = rsm->r_flags;
9532 					log.u_bbr.bbr_state = rack->rc_suspicious;
9533 					log.u_bbr.bbr_substate = rsm->r_just_ret;
9534 					log.u_bbr.timeStamp = tcp_get_usecs(&tv);
9535 					log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
9536 					TCP_LOG_EVENTP(rack->rc_tp, NULL,
9537 						       &rack->rc_inp->inp_socket->so_rcv,
9538 						       &rack->rc_inp->inp_socket->so_snd,
9539 						       TCP_SAD_DETECTION, 0,
9540 						       0, &log, false, &tv);
9541 				}
9542 			}
9543 			/* You loose some ack count every time you sack
9544 			 * a small bit that is not butting to the end of
9545 			 * what we have sent. This is because we never
9546 			 * send small bits unless its the end of the sb.
9547 			 * Anyone sending a sack that is not at the end
9548 			 * is thus very very suspicious.
9549 			 */
9550 			loss = (segsiz/2) / (end - start);
9551 			if (loss < rack->r_ctl.ack_count)
9552 				rack->r_ctl.ack_count -= loss;
9553 			else
9554 				rack->r_ctl.ack_count = 0;
9555 		}
9556 	}
9557 	first_time_through = 0;
9558 #endif
9559 	/* Ok we have an ACK for some piece of this rsm */
9560 	if (rsm->r_start != start) {
9561 		if ((rsm->r_flags & RACK_ACKED) == 0) {
9562 			/*
9563 			 * Before any splitting or hookery is
9564 			 * done is it a TLP of interest i.e. rxt?
9565 			 */
9566 			if ((rsm->r_flags & RACK_TLP) &&
9567 			    (rsm->r_rtr_cnt > 1)) {
9568 				/*
9569 				 * We are splitting a rxt TLP, check
9570 				 * if we need to save off the start/end
9571 				 */
9572 				if (rack->rc_last_tlp_acked_set &&
9573 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9574 					/*
9575 					 * We already turned this on since we are inside
9576 					 * the previous one was a partially sack now we
9577 					 * are getting another one (maybe all of it).
9578 					 *
9579 					 */
9580 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9581 					/*
9582 					 * Lets make sure we have all of it though.
9583 					 */
9584 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9585 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9586 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9587 								     rack->r_ctl.last_tlp_acked_end);
9588 					}
9589 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9590 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9591 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9592 								     rack->r_ctl.last_tlp_acked_end);
9593 					}
9594 				} else {
9595 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9596 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9597 					rack->rc_last_tlp_past_cumack = 0;
9598 					rack->rc_last_tlp_acked_set = 1;
9599 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9600 				}
9601 			}
9602 			/**
9603 			 * Need to split this in two pieces the before and after,
9604 			 * the before remains in the map, the after must be
9605 			 * added. In other words we have:
9606 			 * rsm        |--------------|
9607 			 * sackblk        |------->
9608 			 * rsm will become
9609 			 *     rsm    |---|
9610 			 * and nrsm will be  the sacked piece
9611 			 *     nrsm       |----------|
9612 			 *
9613 			 * But before we start down that path lets
9614 			 * see if the sack spans over on top of
9615 			 * the next guy and it is already sacked.
9616 			 *
9617 			 */
9618 			/*
9619 			 * Hookery can only be used if the two entries
9620 			 * are in the same bucket and neither one of
9621 			 * them staddle the bucket line.
9622 			 */
9623 			next = tqhash_next(rack->r_ctl.tqh, rsm);
9624 			if (next &&
9625 			    (rsm->bindex == next->bindex) &&
9626 			    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9627 			    ((next->r_flags & RACK_STRADDLE) == 0) &&
9628 			    (rsm->r_flags & RACK_IN_GP_WIN) &&
9629 			    (next->r_flags & RACK_IN_GP_WIN))
9630 				can_use_hookery = 1;
9631 			else if (next &&
9632 				 (rsm->bindex == next->bindex) &&
9633 				 ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9634 				 ((next->r_flags & RACK_STRADDLE) == 0) &&
9635 				 ((rsm->r_flags & RACK_IN_GP_WIN) == 0) &&
9636 				 ((next->r_flags & RACK_IN_GP_WIN) == 0))
9637 				can_use_hookery = 1;
9638 			else
9639 				can_use_hookery = 0;
9640 			if (next && can_use_hookery &&
9641 			    (next->r_flags & RACK_ACKED) &&
9642 			    SEQ_GEQ(end, next->r_start)) {
9643 				/**
9644 				 * So the next one is already acked, and
9645 				 * we can thus by hookery use our stack_map
9646 				 * to reflect the piece being sacked and
9647 				 * then adjust the two tree entries moving
9648 				 * the start and ends around. So we start like:
9649 				 *  rsm     |------------|             (not-acked)
9650 				 *  next                 |-----------| (acked)
9651 				 *  sackblk        |-------->
9652 				 *  We want to end like so:
9653 				 *  rsm     |------|                   (not-acked)
9654 				 *  next           |-----------------| (acked)
9655 				 *  nrsm           |-----|
9656 				 * Where nrsm is a temporary stack piece we
9657 				 * use to update all the gizmos.
9658 				 */
9659 				/* Copy up our fudge block */
9660 				noextra++;
9661 				nrsm = &stack_map;
9662 				memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
9663 				/* Now adjust our tree blocks */
9664 				rsm->r_end = start;
9665 				next->r_start = start;
9666  				rsm->r_flags |= RACK_SHUFFLED;
9667 				next->r_flags |= RACK_SHUFFLED;
9668 				/* Now we must adjust back where next->m is */
9669 				rack_setup_offset_for_rsm(rack, rsm, next);
9670 				/*
9671 				 * Which timestamp do we keep? It is rather
9672 				 * important in GP measurements to have the
9673 				 * accurate end of the send window.
9674 				 *
9675 				 * We keep the largest value, which is the newest
9676 				 * send. We do this in case a segment that is
9677 				 * joined together and not part of a GP estimate
9678 				 * later gets expanded into the GP estimate.
9679 				 *
9680 				 * We prohibit the merging of unlike kinds i.e.
9681 				 * all pieces that are in the GP estimate can be
9682 				 * merged and all pieces that are not in a GP estimate
9683 				 * can be merged, but not disimilar pieces. Combine
9684 				 * this with taking the highest here and we should
9685 				 * be ok unless of course the client reneges. Then
9686 				 * all bets are off.
9687 				 */
9688 				if (next->r_tim_lastsent[(next->r_rtr_cnt-1)] <
9689 				    nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)])
9690 					next->r_tim_lastsent[(next->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)];
9691 				/*
9692 				 * And we must keep the newest ack arrival time.
9693 				 */
9694 				if (next->r_ack_arrival <
9695 				    rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
9696 					next->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9697 
9698 
9699 				/* We don't need to adjust rsm, it did not change */
9700 				/* Clear out the dup ack count of the remainder */
9701 				rsm->r_dupack = 0;
9702 				rsm->r_just_ret = 0;
9703 				rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
9704 				/* Now lets make sure our fudge block is right */
9705 				nrsm->r_start = start;
9706 				/* Now lets update all the stats and such */
9707 				rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
9708 				if (rack->app_limited_needs_set)
9709 					rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
9710 				changed += (nrsm->r_end - nrsm->r_start);
9711 				/* You get a count for acking a whole segment or more */
9712 				if ((nrsm->r_end - nrsm->r_start) >= segsiz)
9713 					rack->r_ctl.ack_count += ((nrsm->r_end - nrsm->r_start) / segsiz);
9714 				rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
9715 				if (nrsm->r_flags & RACK_SACK_PASSED) {
9716 					rack->r_ctl.rc_reorder_ts = cts;
9717 					if (rack->r_ctl.rc_reorder_ts == 0)
9718 						rack->r_ctl.rc_reorder_ts = 1;
9719 				}
9720 				/*
9721 				 * Now we want to go up from rsm (the
9722 				 * one left un-acked) to the next one
9723 				 * in the tmap. We do this so when
9724 				 * we walk backwards we include marking
9725 				 * sack-passed on rsm (The one passed in
9726 				 * is skipped since it is generally called
9727 				 * on something sacked before removing it
9728 				 * from the tmap).
9729 				 */
9730 				if (rsm->r_in_tmap) {
9731 					nrsm = TAILQ_NEXT(rsm, r_tnext);
9732 					/*
9733 					 * Now that we have the next
9734 					 * one walk backwards from there.
9735 					 */
9736 					if (nrsm && nrsm->r_in_tmap)
9737 						rack_log_sack_passed(tp, rack, nrsm);
9738 				}
9739 				/* Now are we done? */
9740 				if (SEQ_LT(end, next->r_end) ||
9741 				    (end == next->r_end)) {
9742 					/* Done with block */
9743 					goto out;
9744 				}
9745 				rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__);
9746 				counter_u64_add(rack_sack_used_next_merge, 1);
9747 				/* Postion for the next block */
9748 				start = next->r_end;
9749 				rsm = tqhash_next(rack->r_ctl.tqh, next);
9750 				if (rsm == NULL)
9751 					goto out;
9752 			} else {
9753 				/**
9754 				 * We can't use any hookery here, so we
9755 				 * need to split the map. We enter like
9756 				 * so:
9757 				 *  rsm      |--------|
9758 				 *  sackblk       |----->
9759 				 * We will add the new block nrsm and
9760 				 * that will be the new portion, and then
9761 				 * fall through after reseting rsm. So we
9762 				 * split and look like this:
9763 				 *  rsm      |----|
9764 				 *  sackblk       |----->
9765 				 *  nrsm          |---|
9766 				 * We then fall through reseting
9767 				 * rsm to nrsm, so the next block
9768 				 * picks it up.
9769 				 */
9770 				nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
9771 				if (nrsm == NULL) {
9772 					/*
9773 					 * failed XXXrrs what can we do but loose the sack
9774 					 * info?
9775 					 */
9776 					goto out;
9777 				}
9778 				counter_u64_add(rack_sack_splits, 1);
9779 				rack_clone_rsm(rack, nrsm, rsm, start);
9780 				moved++;
9781 				rsm->r_just_ret = 0;
9782 #ifndef INVARIANTS
9783 				(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
9784 #else
9785 				if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
9786 					panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
9787 					      nrsm, insret, rack, rsm);
9788 				}
9789 #endif
9790 				if (rsm->r_in_tmap) {
9791 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
9792 					nrsm->r_in_tmap = 1;
9793 				}
9794 				rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__);
9795 				rsm->r_flags &= (~RACK_HAS_FIN);
9796 				/* Position us to point to the new nrsm that starts the sack blk */
9797 				rsm = nrsm;
9798 			}
9799 		} else {
9800 			/* Already sacked this piece */
9801 			counter_u64_add(rack_sack_skipped_acked, 1);
9802 			moved++;
9803 			if (end == rsm->r_end) {
9804 				/* Done with block */
9805 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9806 				goto out;
9807 			} else if (SEQ_LT(end, rsm->r_end)) {
9808 				/* A partial sack to a already sacked block */
9809 				moved++;
9810 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9811 				goto out;
9812 			} else {
9813 				/*
9814 				 * The end goes beyond this guy
9815 				 * reposition the start to the
9816 				 * next block.
9817 				 */
9818 				start = rsm->r_end;
9819 				rsm = tqhash_next(rack->r_ctl.tqh, rsm);
9820 				if (rsm == NULL)
9821 					goto out;
9822 			}
9823 		}
9824 	}
9825 	if (SEQ_GEQ(end, rsm->r_end)) {
9826 		/**
9827 		 * The end of this block is either beyond this guy or right
9828 		 * at this guy. I.e.:
9829 		 *  rsm ---                 |-----|
9830 		 *  end                     |-----|
9831 		 *  <or>
9832 		 *  end                     |---------|
9833 		 */
9834 		if ((rsm->r_flags & RACK_ACKED) == 0) {
9835 			/*
9836 			 * Is it a TLP of interest?
9837 			 */
9838 			if ((rsm->r_flags & RACK_TLP) &&
9839 			    (rsm->r_rtr_cnt > 1)) {
9840 				/*
9841 				 * We are splitting a rxt TLP, check
9842 				 * if we need to save off the start/end
9843 				 */
9844 				if (rack->rc_last_tlp_acked_set &&
9845 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9846 					/*
9847 					 * We already turned this on since we are inside
9848 					 * the previous one was a partially sack now we
9849 					 * are getting another one (maybe all of it).
9850 					 */
9851 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9852 					/*
9853 					 * Lets make sure we have all of it though.
9854 					 */
9855 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9856 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9857 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9858 								     rack->r_ctl.last_tlp_acked_end);
9859 					}
9860 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9861 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9862 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9863 								     rack->r_ctl.last_tlp_acked_end);
9864 					}
9865 				} else {
9866 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9867 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9868 					rack->rc_last_tlp_past_cumack = 0;
9869 					rack->rc_last_tlp_acked_set = 1;
9870 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9871 				}
9872 			}
9873 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
9874 			changed += (rsm->r_end - rsm->r_start);
9875 			/* You get a count for acking a whole segment or more */
9876 			if ((rsm->r_end - rsm->r_start) >= segsiz)
9877 				rack->r_ctl.ack_count += ((rsm->r_end - rsm->r_start) / segsiz);
9878 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
9879 			if (rsm->r_in_tmap) /* should be true */
9880 				rack_log_sack_passed(tp, rack, rsm);
9881 			/* Is Reordering occuring? */
9882 			if (rsm->r_flags & RACK_SACK_PASSED) {
9883 				rsm->r_flags &= ~RACK_SACK_PASSED;
9884 				rack->r_ctl.rc_reorder_ts = cts;
9885 				if (rack->r_ctl.rc_reorder_ts == 0)
9886 					rack->r_ctl.rc_reorder_ts = 1;
9887 			}
9888 			if (rack->app_limited_needs_set)
9889 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
9890 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
9891 			rsm->r_flags |= RACK_ACKED;
9892 			if (rsm->r_in_tmap) {
9893 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
9894 				rsm->r_in_tmap = 0;
9895 			}
9896 			rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__);
9897 		} else {
9898 			counter_u64_add(rack_sack_skipped_acked, 1);
9899 			moved++;
9900 		}
9901 		if (end == rsm->r_end) {
9902 			/* This block only - done, setup for next */
9903 			goto out;
9904 		}
9905 		/*
9906 		 * There is more not coverend by this rsm move on
9907 		 * to the next block in the tail queue hash table.
9908 		 */
9909 		nrsm = tqhash_next(rack->r_ctl.tqh, rsm);
9910 		start = rsm->r_end;
9911 		rsm = nrsm;
9912 		if (rsm == NULL)
9913 			goto out;
9914 		goto do_rest_ofb;
9915 	}
9916 	/**
9917 	 * The end of this sack block is smaller than
9918 	 * our rsm i.e.:
9919 	 *  rsm ---                 |-----|
9920 	 *  end                     |--|
9921 	 */
9922 	if ((rsm->r_flags & RACK_ACKED) == 0) {
9923 		/*
9924 		 * Is it a TLP of interest?
9925 		 */
9926 		if ((rsm->r_flags & RACK_TLP) &&
9927 		    (rsm->r_rtr_cnt > 1)) {
9928 			/*
9929 			 * We are splitting a rxt TLP, check
9930 			 * if we need to save off the start/end
9931 			 */
9932 			if (rack->rc_last_tlp_acked_set &&
9933 			    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
9934 				/*
9935 				 * We already turned this on since we are inside
9936 				 * the previous one was a partially sack now we
9937 				 * are getting another one (maybe all of it).
9938 				 */
9939 				rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
9940 				/*
9941 				 * Lets make sure we have all of it though.
9942 				 */
9943 				if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
9944 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9945 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9946 							     rack->r_ctl.last_tlp_acked_end);
9947 				}
9948 				if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
9949 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9950 					rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
9951 							     rack->r_ctl.last_tlp_acked_end);
9952 				}
9953 			} else {
9954 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
9955 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
9956 				rack->rc_last_tlp_past_cumack = 0;
9957 				rack->rc_last_tlp_acked_set = 1;
9958 				rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
9959 			}
9960 		}
9961 		/*
9962 		 * Hookery can only be used if the two entries
9963 		 * are in the same bucket and neither one of
9964 		 * them staddle the bucket line.
9965 		 */
9966 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
9967 		if (prev &&
9968 		    (rsm->bindex == prev->bindex) &&
9969 		    ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9970 		    ((prev->r_flags & RACK_STRADDLE) == 0) &&
9971 		    (rsm->r_flags & RACK_IN_GP_WIN) &&
9972 		    (prev->r_flags & RACK_IN_GP_WIN))
9973 			can_use_hookery = 1;
9974 		else if (prev &&
9975 			 (rsm->bindex == prev->bindex) &&
9976 			 ((rsm->r_flags & RACK_STRADDLE) == 0) &&
9977 			 ((prev->r_flags & RACK_STRADDLE) == 0) &&
9978 			 ((rsm->r_flags & RACK_IN_GP_WIN) == 0) &&
9979 			 ((prev->r_flags & RACK_IN_GP_WIN) == 0))
9980 			can_use_hookery = 1;
9981 		else
9982 			can_use_hookery = 0;
9983 
9984 		if (prev && can_use_hookery &&
9985 		    (prev->r_flags & RACK_ACKED)) {
9986 			/**
9987 			 * Goal, we want the right remainder of rsm to shrink
9988 			 * in place and span from (rsm->r_start = end) to rsm->r_end.
9989 			 * We want to expand prev to go all the way
9990 			 * to prev->r_end <- end.
9991 			 * so in the tree we have before:
9992 			 *   prev     |--------|         (acked)
9993 			 *   rsm               |-------| (non-acked)
9994 			 *   sackblk           |-|
9995 			 * We churn it so we end up with
9996 			 *   prev     |----------|       (acked)
9997 			 *   rsm                 |-----| (non-acked)
9998 			 *   nrsm              |-| (temporary)
9999 			 *
10000 			 * Note if either prev/rsm is a TLP we don't
10001 			 * do this.
10002 			 */
10003 			noextra++;
10004 			nrsm = &stack_map;
10005 			memcpy(nrsm, rsm, sizeof(struct rack_sendmap));
10006 			prev->r_end = end;
10007 			rsm->r_start = end;
10008 			rsm->r_flags |= RACK_SHUFFLED;
10009 			prev->r_flags |= RACK_SHUFFLED;
10010 			/* Now adjust nrsm (stack copy) to be
10011 			 * the one that is the small
10012 			 * piece that was "sacked".
10013 			 */
10014 			nrsm->r_end = end;
10015 			rsm->r_dupack = 0;
10016 			/*
10017 			 * Which timestamp do we keep? It is rather
10018 			 * important in GP measurements to have the
10019 			 * accurate end of the send window.
10020 			 *
10021 			 * We keep the largest value, which is the newest
10022 			 * send. We do this in case a segment that is
10023 			 * joined together and not part of a GP estimate
10024 			 * later gets expanded into the GP estimate.
10025 			 *
10026 			 * We prohibit the merging of unlike kinds i.e.
10027 			 * all pieces that are in the GP estimate can be
10028 			 * merged and all pieces that are not in a GP estimate
10029 			 * can be merged, but not disimilar pieces. Combine
10030 			 * this with taking the highest here and we should
10031 			 * be ok unless of course the client reneges. Then
10032 			 * all bets are off.
10033 			 */
10034 			if(prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] <
10035 			   nrsm->r_tim_lastsent[(nrsm->r_rtr_cnt-1)]) {
10036 				prev->r_tim_lastsent[(prev->r_rtr_cnt-1)] = nrsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)];
10037 			}
10038 			/*
10039 			 * And we must keep the newest ack arrival time.
10040 			 */
10041 
10042 			if(prev->r_ack_arrival <
10043 			   rack_to_usec_ts(&rack->r_ctl.act_rcv_time))
10044 				prev->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10045 
10046 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10047 			/*
10048 			 * Now that the rsm has had its start moved forward
10049 			 * lets go ahead and get its new place in the world.
10050 			 */
10051 			rack_setup_offset_for_rsm(rack, prev, rsm);
10052 			/*
10053 			 * Now nrsm is our new little piece
10054 			 * that is acked (which was merged
10055 			 * to prev). Update the rtt and changed
10056 			 * based on that. Also check for reordering.
10057 			 */
10058 			rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0);
10059 			if (rack->app_limited_needs_set)
10060 				rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END);
10061 			changed += (nrsm->r_end - nrsm->r_start);
10062 			/* You get a count for acking a whole segment or more */
10063 			if ((nrsm->r_end - nrsm->r_start) >= segsiz)
10064 				rack->r_ctl.ack_count += ((nrsm->r_end - nrsm->r_start) / segsiz);
10065 
10066 			rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start);
10067 			if (nrsm->r_flags & RACK_SACK_PASSED) {
10068 				rack->r_ctl.rc_reorder_ts = cts;
10069 				if (rack->r_ctl.rc_reorder_ts == 0)
10070 					rack->r_ctl.rc_reorder_ts = 1;
10071 			}
10072 			rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__);
10073 			rsm = prev;
10074 			counter_u64_add(rack_sack_used_prev_merge, 1);
10075 		} else {
10076 			/**
10077 			 * This is the case where our previous
10078 			 * block is not acked either, so we must
10079 			 * split the block in two.
10080 			 */
10081 			nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
10082 			if (nrsm == NULL) {
10083 				/* failed rrs what can we do but loose the sack info? */
10084 				goto out;
10085 			}
10086 			if ((rsm->r_flags & RACK_TLP) &&
10087 			    (rsm->r_rtr_cnt > 1)) {
10088 				/*
10089 				 * We are splitting a rxt TLP, check
10090 				 * if we need to save off the start/end
10091 				 */
10092 				if (rack->rc_last_tlp_acked_set &&
10093 				    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10094 					/*
10095 					 * We already turned this on since this block is inside
10096 					 * the previous one was a partially sack now we
10097 					 * are getting another one (maybe all of it).
10098 					 */
10099 					rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10100 					/*
10101 					 * Lets make sure we have all of it though.
10102 					 */
10103 					if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10104 						rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10105 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10106 								     rack->r_ctl.last_tlp_acked_end);
10107 					}
10108 					if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10109 						rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10110 						rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10111 								     rack->r_ctl.last_tlp_acked_end);
10112 					}
10113 				} else {
10114 					rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10115 					rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10116 					rack->rc_last_tlp_acked_set = 1;
10117 					rack->rc_last_tlp_past_cumack = 0;
10118 					rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10119 				}
10120 			}
10121 			/**
10122 			 * In this case nrsm becomes
10123 			 * nrsm->r_start = end;
10124 			 * nrsm->r_end = rsm->r_end;
10125 			 * which is un-acked.
10126 			 * <and>
10127 			 * rsm->r_end = nrsm->r_start;
10128 			 * i.e. the remaining un-acked
10129 			 * piece is left on the left
10130 			 * hand side.
10131 			 *
10132 			 * So we start like this
10133 			 * rsm      |----------| (not acked)
10134 			 * sackblk  |---|
10135 			 * build it so we have
10136 			 * rsm      |---|         (acked)
10137 			 * nrsm         |------|  (not acked)
10138 			 */
10139 			counter_u64_add(rack_sack_splits, 1);
10140 			rack_clone_rsm(rack, nrsm, rsm, end);
10141 			moved++;
10142 			rsm->r_flags &= (~RACK_HAS_FIN);
10143 			rsm->r_just_ret = 0;
10144 #ifndef INVARIANTS
10145 			(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
10146 #else
10147 			if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
10148 				panic("Insert in tailq_hash of %p fails ret:% rack:%p rsm:%p",
10149 				      nrsm, insret, rack, rsm);
10150 			}
10151 #endif
10152 			if (rsm->r_in_tmap) {
10153 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
10154 				nrsm->r_in_tmap = 1;
10155 			}
10156 			nrsm->r_dupack = 0;
10157 			rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2);
10158 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0);
10159 			changed += (rsm->r_end - rsm->r_start);
10160 			/* You get a count for acking a whole segment or more */
10161 			if ((rsm->r_end - rsm->r_start) >= segsiz)
10162 				rack->r_ctl.ack_count += ((rsm->r_end - rsm->r_start) / segsiz);
10163 
10164 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
10165 			if (rsm->r_in_tmap) /* should be true */
10166 				rack_log_sack_passed(tp, rack, rsm);
10167 			/* Is Reordering occuring? */
10168 			if (rsm->r_flags & RACK_SACK_PASSED) {
10169 				rsm->r_flags &= ~RACK_SACK_PASSED;
10170 				rack->r_ctl.rc_reorder_ts = cts;
10171 				if (rack->r_ctl.rc_reorder_ts == 0)
10172 					rack->r_ctl.rc_reorder_ts = 1;
10173 			}
10174 			if (rack->app_limited_needs_set)
10175 				rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END);
10176 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10177 			rsm->r_flags |= RACK_ACKED;
10178 			rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__);
10179 			if (rsm->r_in_tmap) {
10180 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10181 				rsm->r_in_tmap = 0;
10182 			}
10183 		}
10184 	} else if (start != end){
10185 		/*
10186 		 * The block was already acked.
10187 		 */
10188 		counter_u64_add(rack_sack_skipped_acked, 1);
10189 		moved++;
10190 	}
10191 out:
10192 	if (rsm &&
10193 	    ((rsm->r_flags & RACK_TLP) == 0) &&
10194 	    (rsm->r_flags & RACK_ACKED)) {
10195 		/*
10196 		 * Now can we merge where we worked
10197 		 * with either the previous or
10198 		 * next block?
10199 		 */
10200 		next = tqhash_next(rack->r_ctl.tqh, rsm);
10201 		while (next) {
10202 			if (next->r_flags & RACK_TLP)
10203 				break;
10204 			/* Only allow merges between ones in or out of GP window */
10205 			if ((next->r_flags & RACK_IN_GP_WIN) &&
10206 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10207 				break;
10208 			}
10209 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10210 			    ((next->r_flags & RACK_IN_GP_WIN) == 0)) {
10211 				break;
10212 			}
10213 			if (rsm->bindex != next->bindex)
10214 				break;
10215 			if (rsm->r_flags & RACK_STRADDLE)
10216 				break;
10217 			if (next->r_flags & RACK_STRADDLE)
10218 				break;
10219 			if (next->r_flags & RACK_ACKED) {
10220 				/* yep this and next can be merged */
10221 				rsm = rack_merge_rsm(rack, rsm, next);
10222 				noextra++;
10223 				next = tqhash_next(rack->r_ctl.tqh, rsm);
10224 			} else
10225 				break;
10226 		}
10227 		/* Now what about the previous? */
10228 		prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10229 		while (prev) {
10230 			if (prev->r_flags & RACK_TLP)
10231 				break;
10232 			/* Only allow merges between ones in or out of GP window */
10233 			if ((prev->r_flags & RACK_IN_GP_WIN) &&
10234 			    ((rsm->r_flags & RACK_IN_GP_WIN) == 0)) {
10235 				break;
10236 			}
10237 			if ((rsm->r_flags & RACK_IN_GP_WIN) &&
10238 			    ((prev->r_flags & RACK_IN_GP_WIN) == 0)) {
10239 				break;
10240 			}
10241 			if (rsm->bindex != prev->bindex)
10242 				break;
10243 			if (rsm->r_flags & RACK_STRADDLE)
10244 				break;
10245 			if (prev->r_flags & RACK_STRADDLE)
10246 				break;
10247 			if (prev->r_flags & RACK_ACKED) {
10248 				/* yep the previous and this can be merged */
10249 				rsm = rack_merge_rsm(rack, prev, rsm);
10250 				noextra++;
10251 				prev = tqhash_prev(rack->r_ctl.tqh, rsm);
10252 			} else
10253 				break;
10254 		}
10255 	}
10256 	if (used_ref == 0) {
10257 		counter_u64_add(rack_sack_proc_all, 1);
10258 	} else {
10259 		counter_u64_add(rack_sack_proc_short, 1);
10260 	}
10261 	/* Save off the next one for quick reference. */
10262 	nrsm = tqhash_find(rack->r_ctl.tqh, end);
10263 	*prsm = rack->r_ctl.rc_sacklast = nrsm;
10264 	/* Pass back the moved. */
10265 	*moved_two = moved;
10266 	*no_extra = noextra;
10267 	return (changed);
10268 }
10269 
10270 static void inline
10271 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
10272 {
10273 	struct rack_sendmap *tmap;
10274 
10275 	tmap = NULL;
10276 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
10277 		/* Its no longer sacked, mark it so */
10278 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10279 #ifdef INVARIANTS
10280 		if (rsm->r_in_tmap) {
10281 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
10282 			      rack, rsm, rsm->r_flags);
10283 		}
10284 #endif
10285 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
10286 		/* Rebuild it into our tmap */
10287 		if (tmap == NULL) {
10288 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10289 			tmap = rsm;
10290 		} else {
10291 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
10292 			tmap = rsm;
10293 		}
10294 		tmap->r_in_tmap = 1;
10295 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
10296 	}
10297 	/*
10298 	 * Now lets possibly clear the sack filter so we start
10299 	 * recognizing sacks that cover this area.
10300 	 */
10301 	sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
10302 
10303 }
10304 
10305 static void
10306 rack_do_decay(struct tcp_rack *rack)
10307 {
10308 	struct timeval res;
10309 
10310 #define	timersub(tvp, uvp, vvp)						\
10311 	do {								\
10312 		(vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec;		\
10313 		(vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec;	\
10314 		if ((vvp)->tv_usec < 0) {				\
10315 			(vvp)->tv_sec--;				\
10316 			(vvp)->tv_usec += 1000000;			\
10317 		}							\
10318 	} while (0)
10319 
10320 	timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res);
10321 #undef timersub
10322 
10323 	rack->r_ctl.input_pkt++;
10324 	if ((rack->rc_in_persist) ||
10325 	    (res.tv_sec >= 1) ||
10326 	    (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) {
10327 		/*
10328 		 * Check for decay of non-SAD,
10329 		 * we want all SAD detection metrics to
10330 		 * decay 1/4 per second (or more) passed.
10331 		 * Current default is 800 so it decays
10332 		 * 80% every second.
10333 		 */
10334 #ifdef TCP_SAD_DETECTION
10335 		uint32_t pkt_delta;
10336 
10337 		pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt;
10338 #endif
10339 		/* Update our saved tracking values */
10340 		rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt;
10341 		rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
10342 		/* Now do we escape without decay? */
10343 #ifdef TCP_SAD_DETECTION
10344 		if (rack->rc_in_persist ||
10345 		    (rack->rc_tp->snd_max == rack->rc_tp->snd_una) ||
10346 		    (pkt_delta < tcp_sad_low_pps)){
10347 			/*
10348 			 * We don't decay idle connections
10349 			 * or ones that have a low input pps.
10350 			 */
10351 			return;
10352 		}
10353 		/* Decay the counters */
10354 		rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count,
10355 							tcp_sad_decay_val);
10356 		rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count,
10357 							 tcp_sad_decay_val);
10358 		rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra,
10359 							       tcp_sad_decay_val);
10360 		rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move,
10361 								tcp_sad_decay_val);
10362 #endif
10363 	}
10364 }
10365 
10366 static void inline
10367 rack_rsm_sender_update(struct tcp_rack *rack, struct tcpcb *tp, struct rack_sendmap *rsm, uint8_t from)
10368 {
10369 	/*
10370 	 * We look at advancing the end send time for our GP
10371 	 * measurement tracking only as the cumulative acknowledgment
10372 	 * moves forward. You might wonder about this, why not
10373 	 * at every transmission or retransmission within the
10374 	 * GP window update the rc_gp_cumack_ts? Well its rather
10375 	 * nuanced but basically the GP window *may* expand (as
10376 	 * it does below) or worse and harder to track it may shrink.
10377 	 *
10378 	 * This last makes it impossible to track at the time of
10379 	 * the send, since you may set forward your rc_gp_cumack_ts
10380 	 * when you send, because that send *is* in your currently
10381 	 * "guessed" window, but then it shrinks. Now which was
10382 	 * the send time of the last bytes in the window, by the
10383 	 * time you ask that question that part of the sendmap
10384 	 * is freed. So you don't know and you will have too
10385 	 * long of send window. Instead by updating the time
10386 	 * marker only when the cumack advances this assures us
10387 	 * that we will have only the sends in the window of our
10388 	 * GP measurement.
10389 	 *
10390 	 * Another complication from this is the
10391 	 * merging of sendmap entries. During SACK processing this
10392 	 * can happen to conserve the sendmap size. That breaks
10393 	 * everything down in tracking the send window of the GP
10394 	 * estimate. So to prevent that and keep it working with
10395 	 * a tiny bit more limited merging, we only allow like
10396 	 * types to be merged. I.e. if two sends are in the GP window
10397 	 * then its ok to merge them together. If two sends are not
10398 	 * in the GP window its ok to merge them together too. Though
10399 	 * one send in and one send out cannot be merged. We combine
10400 	 * this with never allowing the shrinking of the GP window when
10401 	 * we are in recovery so that we can properly calculate the
10402 	 * sending times.
10403 	 *
10404 	 * This all of course seems complicated, because it is.. :)
10405 	 *
10406 	 * The cum-ack is being advanced upon the sendmap.
10407 	 * If we are not doing a GP estimate don't
10408 	 * proceed.
10409 	 */
10410 	uint64_t ts;
10411 
10412 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
10413 		return;
10414 	/*
10415 	 * If this sendmap entry is going
10416 	 * beyond the measurement window we had picked,
10417 	 * expand the measurement window by that much.
10418 	 */
10419 	if (SEQ_GT(rsm->r_end, tp->gput_ack)) {
10420 		tp->gput_ack = rsm->r_end;
10421 	}
10422 	/*
10423 	 * If we have not setup a ack, then we
10424 	 * have no idea if the newly acked pieces
10425 	 * will be "in our seq measurement range". If
10426 	 * it is when we clear the app_limited_needs_set
10427 	 * flag the timestamp will be updated.
10428 	 */
10429 	if (rack->app_limited_needs_set)
10430 		return;
10431 	/*
10432 	 * Finally, we grab out the latest timestamp
10433 	 * that this packet was sent and then see
10434 	 * if:
10435 	 *  a) The packet touches are newly defined GP range.
10436 	 *  b) The time is greater than (newer) than the
10437 	 *     one we currently have. If so we update
10438 	 *     our sending end time window.
10439 	 *
10440 	 * Note we *do not* do this at send time. The reason
10441 	 * is that if you do you *may* pick up a newer timestamp
10442 	 * for a range you are not going to measure. We project
10443 	 * out how far and then sometimes modify that to be
10444 	 * smaller. If that occurs then you will have a send
10445 	 * that does not belong to the range included.
10446 	 */
10447 	if ((ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) <=
10448 	    rack->r_ctl.rc_gp_cumack_ts)
10449 		return;
10450 	if (rack_in_gp_window(tp, rsm)) {
10451 		rack->r_ctl.rc_gp_cumack_ts = ts;
10452 		rack_log_gpset(rack, tp->gput_ack, (uint32_t)ts, rsm->r_end,
10453 			       __LINE__, from, rsm);
10454 	}
10455 }
10456 
10457 static void
10458 rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to, uint64_t acktime)
10459 {
10460 	struct rack_sendmap *rsm;
10461 	/*
10462 	 * The ACK point is advancing to th_ack, we must drop off
10463 	 * the packets in the rack log and calculate any eligble
10464 	 * RTT's.
10465 	 */
10466 
10467 	rack->r_wanted_output = 1;
10468 	if (SEQ_GT(th_ack, tp->snd_una))
10469 		rack->r_ctl.last_cumack_advance = acktime;
10470 
10471 	/* Tend any TLP that has been marked for 1/2 the seq space (its old)  */
10472 	if ((rack->rc_last_tlp_acked_set == 1)&&
10473 	    (rack->rc_last_tlp_past_cumack == 1) &&
10474 	    (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) {
10475 		/*
10476 		 * We have reached the point where our last rack
10477 		 * tlp retransmit sequence is ahead of the cum-ack.
10478 		 * This can only happen when the cum-ack moves all
10479 		 * the way around (its been a full 2^^31+1 bytes
10480 		 * or more since we sent a retransmitted TLP). Lets
10481 		 * turn off the valid flag since its not really valid.
10482 		 *
10483 		 * Note since sack's also turn on this event we have
10484 		 * a complication, we have to wait to age it out until
10485 		 * the cum-ack is by the TLP before checking which is
10486 		 * what the next else clause does.
10487 		 */
10488 		rack_log_dsack_event(rack, 9, __LINE__,
10489 				     rack->r_ctl.last_tlp_acked_start,
10490 				     rack->r_ctl.last_tlp_acked_end);
10491 		rack->rc_last_tlp_acked_set = 0;
10492 		rack->rc_last_tlp_past_cumack = 0;
10493 	} else if ((rack->rc_last_tlp_acked_set == 1) &&
10494 		   (rack->rc_last_tlp_past_cumack == 0) &&
10495 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) {
10496 		/*
10497 		 * It is safe to start aging TLP's out.
10498 		 */
10499 		rack->rc_last_tlp_past_cumack = 1;
10500 	}
10501 	/* We do the same for the tlp send seq as well */
10502 	if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10503 	    (rack->rc_last_sent_tlp_past_cumack == 1) &&
10504 	    (SEQ_GT(rack->r_ctl.last_sent_tlp_seq,  th_ack))) {
10505 		rack_log_dsack_event(rack, 9, __LINE__,
10506 				     rack->r_ctl.last_sent_tlp_seq,
10507 				     (rack->r_ctl.last_sent_tlp_seq +
10508 				      rack->r_ctl.last_sent_tlp_len));
10509 		rack->rc_last_sent_tlp_seq_valid = 0;
10510 		rack->rc_last_sent_tlp_past_cumack = 0;
10511 	} else if ((rack->rc_last_sent_tlp_seq_valid == 1) &&
10512 		   (rack->rc_last_sent_tlp_past_cumack == 0) &&
10513 		   (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) {
10514 		/*
10515 		 * It is safe to start aging TLP's send.
10516 		 */
10517 		rack->rc_last_sent_tlp_past_cumack = 1;
10518 	}
10519 more:
10520 	rsm = tqhash_min(rack->r_ctl.tqh);
10521 	if (rsm == NULL) {
10522 		if ((th_ack - 1) == tp->iss) {
10523 			/*
10524 			 * For the SYN incoming case we will not
10525 			 * have called tcp_output for the sending of
10526 			 * the SYN, so there will be no map. All
10527 			 * other cases should probably be a panic.
10528 			 */
10529 			return;
10530 		}
10531 		if (tp->t_flags & TF_SENTFIN) {
10532 			/* if we sent a FIN we often will not have map */
10533 			return;
10534 		}
10535 #ifdef INVARIANTS
10536 		panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n",
10537 		      tp,
10538 		      tp->t_state, th_ack, rack,
10539 		      tp->snd_una, tp->snd_max, tp->snd_nxt);
10540 #endif
10541 		return;
10542 	}
10543 	if (SEQ_LT(th_ack, rsm->r_start)) {
10544 		/* Huh map is missing this */
10545 #ifdef INVARIANTS
10546 		printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
10547 		       rsm->r_start,
10548 		       th_ack, tp->t_state, rack->r_state);
10549 #endif
10550 		return;
10551 	}
10552 	rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack);
10553 
10554 	/* Now was it a retransmitted TLP? */
10555 	if ((rsm->r_flags & RACK_TLP) &&
10556 	    (rsm->r_rtr_cnt > 1)) {
10557 		/*
10558 		 * Yes, this rsm was a TLP and retransmitted, remember that
10559 		 * since if a DSACK comes back on this we don't want
10560 		 * to think of it as a reordered segment. This may
10561 		 * get updated again with possibly even other TLPs
10562 		 * in flight, but thats ok. Only when we don't send
10563 		 * a retransmitted TLP for 1/2 the sequences space
10564 		 * will it get turned off (above).
10565 		 */
10566 		if (rack->rc_last_tlp_acked_set &&
10567 		    (is_rsm_inside_declared_tlp_block(rack, rsm))) {
10568 			/*
10569 			 * We already turned this on since the end matches,
10570 			 * the previous one was a partially ack now we
10571 			 * are getting another one (maybe all of it).
10572 			 */
10573 			rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end);
10574 			/*
10575 			 * Lets make sure we have all of it though.
10576 			 */
10577 			if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) {
10578 				rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10579 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10580 						     rack->r_ctl.last_tlp_acked_end);
10581 			}
10582 			if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) {
10583 				rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10584 				rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start,
10585 						     rack->r_ctl.last_tlp_acked_end);
10586 			}
10587 		} else {
10588 			rack->rc_last_tlp_past_cumack = 1;
10589 			rack->r_ctl.last_tlp_acked_start = rsm->r_start;
10590 			rack->r_ctl.last_tlp_acked_end = rsm->r_end;
10591 			rack->rc_last_tlp_acked_set = 1;
10592 			rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end);
10593 		}
10594 	}
10595 	/* Now do we consume the whole thing? */
10596 	rack->r_ctl.last_tmit_time_acked = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
10597 	if (SEQ_GEQ(th_ack, rsm->r_end)) {
10598 		/* Its all consumed. */
10599 		uint32_t left;
10600 		uint8_t newly_acked;
10601 
10602 		rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__);
10603 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
10604 		rsm->r_rtr_bytes = 0;
10605 		/*
10606 		 * Record the time of highest cumack sent if its in our measurement
10607 		 * window and possibly bump out the end.
10608 		 */
10609 		rack_rsm_sender_update(rack, tp, rsm, 4);
10610 		tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
10611 		if (rsm->r_in_tmap) {
10612 			TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10613 			rsm->r_in_tmap = 0;
10614 		}
10615 		newly_acked = 1;
10616 		if (rsm->r_flags & RACK_ACKED) {
10617 			/*
10618 			 * It was acked on the scoreboard -- remove
10619 			 * it from total
10620 			 */
10621 			rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
10622 			newly_acked = 0;
10623 		} else if (rsm->r_flags & RACK_SACK_PASSED) {
10624 			/*
10625 			 * There are segments ACKED on the
10626 			 * scoreboard further up. We are seeing
10627 			 * reordering.
10628 			 */
10629 			rsm->r_flags &= ~RACK_SACK_PASSED;
10630 			rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
10631 			rsm->r_flags |= RACK_ACKED;
10632 			rack->r_ctl.rc_reorder_ts = cts;
10633 			if (rack->r_ctl.rc_reorder_ts == 0)
10634 				rack->r_ctl.rc_reorder_ts = 1;
10635 			if (rack->r_ent_rec_ns) {
10636 				/*
10637 				 * We have sent no more, and we saw an sack
10638 				 * then ack arrive.
10639 				 */
10640 				rack->r_might_revert = 1;
10641 			}
10642 		}
10643 		if ((rsm->r_flags & RACK_TO_REXT) &&
10644 		    (tp->t_flags & TF_RCVD_TSTMP) &&
10645 		    (to->to_flags & TOF_TS) &&
10646 		    (to->to_tsecr != 0) &&
10647 		    (tp->t_flags & TF_PREVVALID)) {
10648 			/*
10649 			 * We can use the timestamp to see
10650 			 * if this retransmission was from the
10651 			 * first transmit. If so we made a mistake.
10652 			 */
10653 			tp->t_flags &= ~TF_PREVVALID;
10654 			if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) {
10655 				/* The first transmit is what this ack is for */
10656 				rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__);
10657 			}
10658 		}
10659 		left = th_ack - rsm->r_end;
10660 		if (rack->app_limited_needs_set && newly_acked)
10661 			rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK);
10662 		/* Free back to zone */
10663 		rack_free(rack, rsm);
10664 		if (left) {
10665 			goto more;
10666 		}
10667 		/* Check for reneging */
10668 		rsm = tqhash_min(rack->r_ctl.tqh);
10669 		if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
10670 			/*
10671 			 * The peer has moved snd_una up to
10672 			 * the edge of this send, i.e. one
10673 			 * that it had previously acked. The only
10674 			 * way that can be true if the peer threw
10675 			 * away data (space issues) that it had
10676 			 * previously sacked (else it would have
10677 			 * given us snd_una up to (rsm->r_end).
10678 			 * We need to undo the acked markings here.
10679 			 *
10680 			 * Note we have to look to make sure th_ack is
10681 			 * our rsm->r_start in case we get an old ack
10682 			 * where th_ack is behind snd_una.
10683 			 */
10684 			rack_peer_reneges(rack, rsm, th_ack);
10685 		}
10686 		return;
10687 	}
10688 	if (rsm->r_flags & RACK_ACKED) {
10689 		/*
10690 		 * It was acked on the scoreboard -- remove it from
10691 		 * total for the part being cum-acked.
10692 		 */
10693 		rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
10694 	}
10695 	/*
10696 	 * Clear the dup ack count for
10697 	 * the piece that remains.
10698 	 */
10699 	rsm->r_dupack = 0;
10700 	rack_log_retran_reason(rack, rsm, __LINE__, 0, 2);
10701 	if (rsm->r_rtr_bytes) {
10702 		/*
10703 		 * It was retransmitted adjust the
10704 		 * sack holes for what was acked.
10705 		 */
10706 		int ack_am;
10707 
10708 		ack_am = (th_ack - rsm->r_start);
10709 		if (ack_am >= rsm->r_rtr_bytes) {
10710 			rack->r_ctl.rc_holes_rxt -= ack_am;
10711 			rsm->r_rtr_bytes -= ack_am;
10712 		}
10713 	}
10714 	/*
10715 	 * Update where the piece starts and record
10716 	 * the time of send of highest cumack sent if
10717 	 * its in our GP range.
10718 	 */
10719 	rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__);
10720 	/* Now we need to move our offset forward too */
10721 	if (rsm->m &&
10722 	    ((rsm->orig_m_len != rsm->m->m_len) ||
10723 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
10724 		/* Fix up the orig_m_len and possibly the mbuf offset */
10725 		rack_adjust_orig_mlen(rsm);
10726 	}
10727 	rsm->soff += (th_ack - rsm->r_start);
10728 	rack_rsm_sender_update(rack, tp, rsm, 5);
10729 	/* The trim will move th_ack into r_start for us */
10730 	tqhash_trim(rack->r_ctl.tqh, th_ack);
10731 	/* Now do we need to move the mbuf fwd too? */
10732 	{
10733 		struct mbuf *m;
10734 		uint32_t soff;
10735 
10736 		m = rsm->m;
10737 		soff = rsm->soff;
10738 		if (m) {
10739 			while (soff >= m->m_len) {
10740 				soff -= m->m_len;
10741 				KASSERT((m->m_next != NULL),
10742 					(" rsm:%p  off:%u soff:%u m:%p",
10743 					 rsm, rsm->soff, soff, m));
10744 				m = m->m_next;
10745 				if (m == NULL) {
10746 					/*
10747 					 * This is a fall-back that prevents a panic. In reality
10748 					 * we should be able to walk the mbuf's and find our place.
10749 					 * At this point snd_una has not been updated with the sbcut() yet
10750 					 * but tqhash_trim did update rsm->r_start so the offset calcuation
10751 					 * should work fine. This is undesirable since we will take cache
10752 					 * hits to access the socket buffer. And even more puzzling is that
10753 					 * it happens occasionally. It should not :(
10754 					 */
10755 					m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
10756 						      (rsm->r_start - tp->snd_una),
10757 						      &soff);
10758 					break;
10759 				}
10760 			}
10761 			/*
10762 			 * Now save in our updated values.
10763 			 */
10764 			rsm->m = m;
10765 			rsm->soff = soff;
10766 			rsm->orig_m_len = rsm->m->m_len;
10767 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
10768 		}
10769 	}
10770 	if (rack->app_limited_needs_set &&
10771 	    SEQ_GEQ(th_ack, tp->gput_seq))
10772 		rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG);
10773 }
10774 
10775 static void
10776 rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack)
10777 {
10778 	struct rack_sendmap *rsm;
10779 	int sack_pass_fnd = 0;
10780 
10781 	if (rack->r_might_revert) {
10782 		/*
10783 		 * Ok we have reordering, have not sent anything, we
10784 		 * might want to revert the congestion state if nothing
10785 		 * further has SACK_PASSED on it. Lets check.
10786 		 *
10787 		 * We also get here when we have DSACKs come in for
10788 		 * all the data that we FR'd. Note that a rxt or tlp
10789 		 * timer clears this from happening.
10790 		 */
10791 
10792 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
10793 			if (rsm->r_flags & RACK_SACK_PASSED) {
10794 				sack_pass_fnd = 1;
10795 				break;
10796 			}
10797 		}
10798 		if (sack_pass_fnd == 0) {
10799 			/*
10800 			 * We went into recovery
10801 			 * incorrectly due to reordering!
10802 			 */
10803 			int orig_cwnd;
10804 
10805 			rack->r_ent_rec_ns = 0;
10806 			orig_cwnd = tp->snd_cwnd;
10807 			tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec;
10808 			tp->snd_recover = tp->snd_una;
10809 			rack_log_to_prr(rack, 14, orig_cwnd, __LINE__);
10810 			EXIT_RECOVERY(tp->t_flags);
10811 		}
10812 		rack->r_might_revert = 0;
10813 	}
10814 }
10815 
10816 #ifdef TCP_SAD_DETECTION
10817 
10818 static void
10819 rack_merge_out_sacks(struct tcp_rack *rack)
10820 {
10821 	struct rack_sendmap *cur, *next, *rsm, *trsm = NULL;
10822 
10823 	cur = tqhash_min(rack->r_ctl.tqh);
10824 	while(cur) {
10825 		next = tqhash_next(rack->r_ctl.tqh, cur);
10826 		/*
10827 		 * The idea is to go through all and merge back
10828 		 * together the pieces sent together,
10829 		 */
10830 		if ((next != NULL) &&
10831 		    (cur->r_tim_lastsent[0] == next->r_tim_lastsent[0])) {
10832 			rack_merge_rsm(rack, cur, next);
10833 		} else {
10834 			cur = next;
10835 		}
10836 	}
10837 	/*
10838 	 * now treat it like a rxt event, everything is outstanding
10839 	 * and sent nothing acvked and dupacks are all zero. If this
10840 	 * is not an attacker it will have to dupack its way through
10841 	 * it all.
10842 	 */
10843 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
10844 	TQHASH_FOREACH(rsm, rack->r_ctl.tqh)  {
10845 		rsm->r_dupack = 0;
10846 		/* We must re-add it back to the tlist */
10847 		if (trsm == NULL) {
10848 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
10849 		} else {
10850 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
10851 		}
10852 		rsm->r_in_tmap = 1;
10853 		trsm = rsm;
10854 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED);
10855 	}
10856 	sack_filter_clear(&rack->r_ctl.rack_sf, rack->rc_tp->snd_una);
10857 }
10858 
10859 static void
10860 rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t bytes_this_ack, uint32_t segsiz)
10861 {
10862 	int do_detection = 0;
10863 
10864 	if (rack->sack_attack_disable || rack->rc_suspicious) {
10865 		/*
10866 		 * If we have been disabled we must detect
10867 		 * to possibly reverse it. Or if the guy has
10868 		 * sent in suspicious sacks we want to do detection too.
10869 		 */
10870 		do_detection = 1;
10871 
10872 	} else if  ((rack->do_detection || tcp_force_detection) &&
10873 		    (tcp_sack_to_ack_thresh > 0) &&
10874 		    (tcp_sack_to_move_thresh > 0) &&
10875 		    (rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum)) {
10876 		/*
10877 		 * We only detect here if:
10878 		 * 1) System wide forcing is on <or> do_detection is on
10879 		 *   <and>
10880 		 * 2) We have thresholds for move and ack (set one to 0 and we are off)
10881 		 *   <and>
10882 		 * 3) We have maps allocated larger than our min (500).
10883 		 */
10884 		do_detection = 1;
10885 	}
10886 	if (do_detection > 0) {
10887 		/*
10888 		 * We have thresholds set to find
10889 		 * possible attackers and disable sack.
10890 		 * Check them.
10891 		 */
10892 		uint64_t ackratio, moveratio, movetotal;
10893 
10894 		/* Log detecting */
10895 		rack_log_sad(rack, 1);
10896 		/* Do we establish a ack ratio */
10897 		if ((rack->r_ctl.sack_count > tcp_map_minimum)  ||
10898 		    (rack->rc_suspicious == 1) ||
10899 		    (rack->sack_attack_disable > 0)) {
10900 			ackratio = (uint64_t)(rack->r_ctl.sack_count);
10901 			ackratio *= (uint64_t)(1000);
10902 			if (rack->r_ctl.ack_count)
10903 				ackratio /= (uint64_t)(rack->r_ctl.ack_count);
10904 			else {
10905 				/* We can hit this due to ack totals degregation (via small sacks) */
10906 				ackratio = 1000;
10907 			}
10908 		} else {
10909 			/*
10910 			 * No ack ratio needed if we have not
10911 			 * seen more sacks then the number of map entries.
10912 			 * The exception to that is if we have disabled sack then
10913 			 * we need to find a ratio.
10914 			 */
10915 			ackratio = 0;
10916 		}
10917 
10918 		if ((rack->sack_attack_disable == 0) &&
10919 		    (ackratio > rack_highest_sack_thresh_seen))
10920 			rack_highest_sack_thresh_seen = (uint32_t)ackratio;
10921 		/* Do we establish a move ratio? */
10922 		if ((rack->r_ctl.sack_moved_extra > tcp_map_minimum) ||
10923 		    (rack->rc_suspicious == 1) ||
10924 		    (rack->sack_attack_disable > 0)) {
10925 			/*
10926 			 * We need to have more sack moves than maps
10927 			 * allocated to have a move ratio considered.
10928 			 */
10929 			movetotal = rack->r_ctl.sack_moved_extra;
10930 			movetotal += rack->r_ctl.sack_noextra_move;
10931 			moveratio = rack->r_ctl.sack_moved_extra;
10932 			moveratio *= (uint64_t)1000;
10933 			if (movetotal)
10934 				moveratio /= movetotal;
10935 			else {
10936 				/* No moves, thats pretty good */
10937 				moveratio = 0;
10938 			}
10939 		} else {
10940 			/*
10941 			 * Not enough moves have occured to consider
10942 			 * if we are out of whack in that ratio.
10943 			 * The exception to that is if we have disabled sack then
10944 			 * we need to find a ratio.
10945 			 */
10946 			moveratio = 0;
10947 		}
10948 		if ((rack->sack_attack_disable == 0) &&
10949 		    (moveratio > rack_highest_move_thresh_seen))
10950 			rack_highest_move_thresh_seen = (uint32_t)moveratio;
10951 		/* Now the tests */
10952 		if (rack->sack_attack_disable == 0) {
10953 			/* Not disabled, do we need to disable? */
10954 			if ((ackratio > tcp_sack_to_ack_thresh) &&
10955 			    (moveratio > tcp_sack_to_move_thresh)) {
10956 				/* Disable sack processing */
10957 				tcp_trace_point(rack->rc_tp, TCP_TP_SAD_TRIGGERED);
10958 				rack->sack_attack_disable = 1;
10959 				/* set it so we have the built in delay */
10960 				rack->r_ctl.ack_during_sd = 1;
10961 				if (rack_merge_out_sacks_on_attack)
10962 					rack_merge_out_sacks(rack);
10963 				counter_u64_add(rack_sack_attacks_detected, 1);
10964 				tcp_trace_point(rack->rc_tp, TCP_TP_SAD_TRIGGERED);
10965 				/* Clamp the cwnd at flight size */
10966 				rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd;
10967 				rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
10968 				rack_log_sad(rack, 2);
10969 			}
10970 		} else {
10971 			/* We are sack-disabled check for false positives */
10972 			if ((ackratio <= tcp_restoral_thresh) ||
10973 			    ((rack_merge_out_sacks_on_attack == 0) &&
10974 			     (rack->rc_suspicious == 0) &&
10975 			     (rack->r_ctl.rc_num_maps_alloced <= (tcp_map_minimum/2)))) {
10976 				rack->sack_attack_disable = 0;
10977 				rack_log_sad(rack, 3);
10978 				/* Restart counting */
10979 				rack->r_ctl.sack_count = 0;
10980 				rack->r_ctl.sack_moved_extra = 0;
10981 				rack->r_ctl.sack_noextra_move = 1;
10982 				rack->rc_suspicious = 0;
10983 				rack->r_ctl.ack_count = max(1,
10984 							    (bytes_this_ack / segsiz));
10985 
10986 				counter_u64_add(rack_sack_attacks_reversed, 1);
10987 				/* Restore the cwnd */
10988 				if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd)
10989 					rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd;
10990 			}
10991 		}
10992 	}
10993 }
10994 #endif
10995 
10996 static int
10997 rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end)
10998 {
10999 
11000 	uint32_t am, l_end;
11001 	int was_tlp = 0;
11002 
11003 	if (SEQ_GT(end, start))
11004 		am = end - start;
11005 	else
11006 		am = 0;
11007 	if ((rack->rc_last_tlp_acked_set ) &&
11008 	    (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) &&
11009 	    (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) {
11010 		/*
11011 		 * The DSACK is because of a TLP which we don't
11012 		 * do anything with the reordering window over since
11013 		 * it was not reordering that caused the DSACK but
11014 		 * our previous retransmit TLP.
11015 		 */
11016 		rack_log_dsack_event(rack, 7, __LINE__, start, end);
11017 		was_tlp = 1;
11018 		goto skip_dsack_round;
11019 	}
11020 	if (rack->rc_last_sent_tlp_seq_valid) {
11021 		l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len;
11022 		if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) &&
11023 		    (SEQ_LEQ(end, l_end))) {
11024 			/*
11025 			 * This dsack is from the last sent TLP, ignore it
11026 			 * for reordering purposes.
11027 			 */
11028 			rack_log_dsack_event(rack, 7, __LINE__, start, end);
11029 			was_tlp = 1;
11030 			goto skip_dsack_round;
11031 		}
11032 	}
11033 	if (rack->rc_dsack_round_seen == 0) {
11034 		rack->rc_dsack_round_seen = 1;
11035 		rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max;
11036 		rack->r_ctl.num_dsack++;
11037 		rack->r_ctl.dsack_persist = 16;	/* 16 is from the standard */
11038 		rack_log_dsack_event(rack, 2, __LINE__, 0, 0);
11039 	}
11040 skip_dsack_round:
11041 	/*
11042 	 * We keep track of how many DSACK blocks we get
11043 	 * after a recovery incident.
11044 	 */
11045 	rack->r_ctl.dsack_byte_cnt += am;
11046 	if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) &&
11047 	    rack->r_ctl.retran_during_recovery &&
11048 	    (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) {
11049 		/*
11050 		 * False recovery most likely culprit is reordering. If
11051 		 * nothing else is missing we need to revert.
11052 		 */
11053 		rack->r_might_revert = 1;
11054 		rack_handle_might_revert(rack->rc_tp, rack);
11055 		rack->r_might_revert = 0;
11056 		rack->r_ctl.retran_during_recovery = 0;
11057 		rack->r_ctl.dsack_byte_cnt = 0;
11058 	}
11059 	return (was_tlp);
11060 }
11061 
11062 static uint32_t
11063 do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una)
11064 {
11065 	return (((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt);
11066 }
11067 
11068 static int32_t
11069 rack_compute_pipe(struct tcpcb *tp)
11070 {
11071 	return ((int32_t)do_rack_compute_pipe(tp,
11072 					      (struct tcp_rack *)tp->t_fb_ptr,
11073 					      tp->snd_una));
11074 }
11075 
11076 static void
11077 rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack)
11078 {
11079 	/* Deal with changed and PRR here (in recovery only) */
11080 	uint32_t pipe, snd_una;
11081 
11082 	rack->r_ctl.rc_prr_delivered += changed;
11083 
11084 	if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) {
11085 		/*
11086 		 * It is all outstanding, we are application limited
11087 		 * and thus we don't need more room to send anything.
11088 		 * Note we use tp->snd_una here and not th_ack because
11089 		 * the data as yet not been cut from the sb.
11090 		 */
11091 		rack->r_ctl.rc_prr_sndcnt = 0;
11092 		return;
11093 	}
11094 	/* Compute prr_sndcnt */
11095 	if (SEQ_GT(tp->snd_una, th_ack)) {
11096 		snd_una = tp->snd_una;
11097 	} else {
11098 		snd_una = th_ack;
11099 	}
11100 	pipe = do_rack_compute_pipe(tp, rack, snd_una);
11101 	if (pipe > tp->snd_ssthresh) {
11102 		long sndcnt;
11103 
11104 		sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
11105 		if (rack->r_ctl.rc_prr_recovery_fs > 0)
11106 			sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
11107 		else {
11108 			rack->r_ctl.rc_prr_sndcnt = 0;
11109 			rack_log_to_prr(rack, 9, 0, __LINE__);
11110 			sndcnt = 0;
11111 		}
11112 		sndcnt++;
11113 		if (sndcnt > (long)rack->r_ctl.rc_prr_out)
11114 			sndcnt -= rack->r_ctl.rc_prr_out;
11115 		else
11116 			sndcnt = 0;
11117 		rack->r_ctl.rc_prr_sndcnt = sndcnt;
11118 		rack_log_to_prr(rack, 10, 0, __LINE__);
11119 	} else {
11120 		uint32_t limit;
11121 
11122 		if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
11123 			limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
11124 		else
11125 			limit = 0;
11126 		if (changed > limit)
11127 			limit = changed;
11128 		limit += ctf_fixed_maxseg(tp);
11129 		if (tp->snd_ssthresh > pipe) {
11130 			rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
11131 			rack_log_to_prr(rack, 11, 0, __LINE__);
11132 		} else {
11133 			rack->r_ctl.rc_prr_sndcnt = min(0, limit);
11134 			rack_log_to_prr(rack, 12, 0, __LINE__);
11135 		}
11136 	}
11137 }
11138 
11139 static void
11140 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck,
11141 	     int *dsack_seen, int *sacks_seen)
11142 {
11143 	uint32_t changed;
11144 	struct tcp_rack *rack;
11145 	struct rack_sendmap *rsm;
11146 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
11147 	register uint32_t th_ack;
11148 	int32_t i, j, k, num_sack_blks = 0;
11149 	uint32_t cts, acked, ack_point;
11150 	int loop_start = 0, moved_two = 0, no_extra = 0;
11151 	uint32_t tsused;
11152 	uint32_t segsiz, o_cnt;
11153 
11154 
11155 	INP_WLOCK_ASSERT(tptoinpcb(tp));
11156 	if (tcp_get_flags(th) & TH_RST) {
11157 		/* We don't log resets */
11158 		return;
11159 	}
11160 	rack = (struct tcp_rack *)tp->t_fb_ptr;
11161 	cts = tcp_get_usecs(NULL);
11162 	rsm = tqhash_min(rack->r_ctl.tqh);
11163 	changed = 0;
11164 	th_ack = th->th_ack;
11165 	if (rack->sack_attack_disable == 0)
11166 		rack_do_decay(rack);
11167 	segsiz = ctf_fixed_maxseg(rack->rc_tp);
11168 	if (BYTES_THIS_ACK(tp, th) >=  segsiz) {
11169 		/*
11170 		 * You only get credit for
11171 		 * MSS and greater (and you get extra
11172 		 * credit for larger cum-ack moves).
11173 		 */
11174 		int ac;
11175 
11176 		ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp);
11177 		rack->r_ctl.ack_count += ac;
11178 		counter_u64_add(rack_ack_total, ac);
11179 	}
11180 	if (rack->r_ctl.ack_count > 0xfff00000) {
11181 		/*
11182 		 * reduce the number to keep us under
11183 		 * a uint32_t.
11184 		 */
11185 		rack->r_ctl.ack_count /= 2;
11186 		rack->r_ctl.sack_count /= 2;
11187 	}
11188 	if (SEQ_GT(th_ack, tp->snd_una)) {
11189 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
11190 		tp->t_acktime = ticks;
11191 	}
11192 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
11193 		changed = th_ack - rsm->r_start;
11194 	if (changed) {
11195 		rack_process_to_cumack(tp, rack, th_ack, cts, to,
11196 				       tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
11197 	}
11198 	if ((to->to_flags & TOF_SACK) == 0) {
11199 		/* We are done nothing left and no sack. */
11200 		rack_handle_might_revert(tp, rack);
11201 		/*
11202 		 * For cases where we struck a dup-ack
11203 		 * with no SACK, add to the changes so
11204 		 * PRR will work right.
11205 		 */
11206 		if (dup_ack_struck && (changed == 0)) {
11207 			changed += ctf_fixed_maxseg(rack->rc_tp);
11208 		}
11209 		goto out;
11210 	}
11211 	/* Sack block processing */
11212 	if (SEQ_GT(th_ack, tp->snd_una))
11213 		ack_point = th_ack;
11214 	else
11215 		ack_point = tp->snd_una;
11216 	for (i = 0; i < to->to_nsacks; i++) {
11217 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
11218 		      &sack, sizeof(sack));
11219 		sack.start = ntohl(sack.start);
11220 		sack.end = ntohl(sack.end);
11221 		if (SEQ_GT(sack.end, sack.start) &&
11222 		    SEQ_GT(sack.start, ack_point) &&
11223 		    SEQ_LT(sack.start, tp->snd_max) &&
11224 		    SEQ_GT(sack.end, ack_point) &&
11225 		    SEQ_LEQ(sack.end, tp->snd_max)) {
11226 			sack_blocks[num_sack_blks] = sack;
11227 			num_sack_blks++;
11228 		} else if (SEQ_LEQ(sack.start, th_ack) &&
11229 			   SEQ_LEQ(sack.end, th_ack)) {
11230 			int was_tlp;
11231 
11232 			if (dsack_seen != NULL)
11233 				*dsack_seen = 1;
11234 			was_tlp = rack_note_dsack(rack, sack.start, sack.end);
11235 			/*
11236 			 * Its a D-SACK block.
11237 			 */
11238 			tcp_record_dsack(tp, sack.start, sack.end, was_tlp);
11239 		}
11240 	}
11241 	if (rack->rc_dsack_round_seen) {
11242 		/* Is the dsack roound over? */
11243 		if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) {
11244 			/* Yes it is */
11245 			rack->rc_dsack_round_seen = 0;
11246 			rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
11247 		}
11248 	}
11249 	/*
11250 	 * Sort the SACK blocks so we can update the rack scoreboard with
11251 	 * just one pass.
11252 	 */
11253 	o_cnt = num_sack_blks;
11254 	num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks,
11255 					 num_sack_blks, th->th_ack);
11256 	ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks);
11257 	if (sacks_seen != NULL)
11258 		*sacks_seen = num_sack_blks;
11259 	if (num_sack_blks == 0) {
11260 		/* Nothing to sack, but we need to update counts */
11261 		if ((o_cnt == 1) &&
11262 		    (*dsack_seen != 1))
11263 			rack->r_ctl.sack_count++;
11264 		else if (o_cnt > 1)
11265 			rack->r_ctl.sack_count++;
11266 		goto out_with_totals;
11267 	}
11268 	if (rack->sack_attack_disable) {
11269 		/*
11270 		 * An attacker disablement is in place, for
11271 		 * every sack block that is not at least a full MSS
11272 		 * count up sack_count.
11273 		 */
11274 		for (i = 0; i < num_sack_blks; i++) {
11275 			if ((sack_blocks[i].end - sack_blocks[i].start) < segsiz) {
11276 				rack->r_ctl.sack_count++;
11277 			}
11278 			if (rack->r_ctl.sack_count > 0xfff00000) {
11279 				/*
11280 				 * reduce the number to keep us under
11281 				 * a uint32_t.
11282 				 */
11283 				rack->r_ctl.ack_count /= 2;
11284 				rack->r_ctl.sack_count /= 2;
11285 			}
11286 		}
11287 		goto out;
11288 	}
11289 	/* Its a sack of some sort */
11290 	rack->r_ctl.sack_count += num_sack_blks;
11291 	if (rack->r_ctl.sack_count > 0xfff00000) {
11292 		/*
11293 		 * reduce the number to keep us under
11294 		 * a uint32_t.
11295 		 */
11296 		rack->r_ctl.ack_count /= 2;
11297 		rack->r_ctl.sack_count /= 2;
11298 	}
11299 	if (num_sack_blks < 2) {
11300 		/* Only one, we don't need to sort */
11301 		goto do_sack_work;
11302 	}
11303 	/* Sort the sacks */
11304 	for (i = 0; i < num_sack_blks; i++) {
11305 		for (j = i + 1; j < num_sack_blks; j++) {
11306 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
11307 				sack = sack_blocks[i];
11308 				sack_blocks[i] = sack_blocks[j];
11309 				sack_blocks[j] = sack;
11310 			}
11311 		}
11312 	}
11313 	/*
11314 	 * Now are any of the sack block ends the same (yes some
11315 	 * implementations send these)?
11316 	 */
11317 again:
11318 	if (num_sack_blks == 0)
11319 		goto out_with_totals;
11320 	if (num_sack_blks > 1) {
11321 		for (i = 0; i < num_sack_blks; i++) {
11322 			for (j = i + 1; j < num_sack_blks; j++) {
11323 				if (sack_blocks[i].end == sack_blocks[j].end) {
11324 					/*
11325 					 * Ok these two have the same end we
11326 					 * want the smallest end and then
11327 					 * throw away the larger and start
11328 					 * again.
11329 					 */
11330 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
11331 						/*
11332 						 * The second block covers
11333 						 * more area use that
11334 						 */
11335 						sack_blocks[i].start = sack_blocks[j].start;
11336 					}
11337 					/*
11338 					 * Now collapse out the dup-sack and
11339 					 * lower the count
11340 					 */
11341 					for (k = (j + 1); k < num_sack_blks; k++) {
11342 						sack_blocks[j].start = sack_blocks[k].start;
11343 						sack_blocks[j].end = sack_blocks[k].end;
11344 						j++;
11345 					}
11346 					num_sack_blks--;
11347 					goto again;
11348 				}
11349 			}
11350 		}
11351 	}
11352 do_sack_work:
11353 	/*
11354 	 * First lets look to see if
11355 	 * we have retransmitted and
11356 	 * can use the transmit next?
11357 	 */
11358 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11359 	if (rsm &&
11360 	    SEQ_GT(sack_blocks[0].end, rsm->r_start) &&
11361 	    SEQ_LT(sack_blocks[0].start, rsm->r_end)) {
11362 		/*
11363 		 * We probably did the FR and the next
11364 		 * SACK in continues as we would expect.
11365 		 */
11366 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &no_extra, &moved_two, segsiz);
11367 		if (acked) {
11368 			rack->r_wanted_output = 1;
11369 			changed += acked;
11370 		}
11371 		if (num_sack_blks == 1) {
11372 			/*
11373 			 * This is what we would expect from
11374 			 * a normal implementation to happen
11375 			 * after we have retransmitted the FR,
11376 			 * i.e the sack-filter pushes down
11377 			 * to 1 block and the next to be retransmitted
11378 			 * is the sequence in the sack block (has more
11379 			 * are acked). Count this as ACK'd data to boost
11380 			 * up the chances of recovering any false positives.
11381 			 */
11382 			rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp));
11383 			counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp)));
11384 			counter_u64_add(rack_express_sack, 1);
11385 			if (rack->r_ctl.ack_count > 0xfff00000) {
11386 				/*
11387 				 * reduce the number to keep us under
11388 				 * a uint32_t.
11389 				 */
11390 				rack->r_ctl.ack_count /= 2;
11391 				rack->r_ctl.sack_count /= 2;
11392 			}
11393 			if (moved_two) {
11394 				/*
11395 				 * If we did not get a SACK for at least a MSS and
11396 				 * had to move at all, or if we moved more than our
11397 				 * threshold, it counts against the "extra" move.
11398 				 */
11399 				rack->r_ctl.sack_moved_extra += moved_two;
11400 				rack->r_ctl.sack_noextra_move += no_extra;
11401 				counter_u64_add(rack_move_some, 1);
11402 			} else {
11403 				/*
11404 				 * else we did not have to move
11405 				 * any more than we would expect.
11406 				 */
11407 				rack->r_ctl.sack_noextra_move += no_extra;
11408 				rack->r_ctl.sack_noextra_move++;
11409 				counter_u64_add(rack_move_none, 1);
11410 			}
11411 			if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
11412 			    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
11413 				rack->r_ctl.sack_moved_extra /= 2;
11414 				rack->r_ctl.sack_noextra_move /= 2;
11415 			}
11416 			goto out_with_totals;
11417 		} else {
11418 			/*
11419 			 * Start the loop through the
11420 			 * rest of blocks, past the first block.
11421 			 */
11422 			loop_start = 1;
11423 		}
11424 	}
11425 	counter_u64_add(rack_sack_total, 1);
11426 	rsm = rack->r_ctl.rc_sacklast;
11427 	for (i = loop_start; i < num_sack_blks; i++) {
11428 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &no_extra, &moved_two, segsiz);
11429 		if (acked) {
11430 			rack->r_wanted_output = 1;
11431 			changed += acked;
11432 		}
11433 		if (moved_two) {
11434 			/*
11435 			 * If we did not get a SACK for at least a MSS and
11436 			 * had to move at all, or if we moved more than our
11437 			 * threshold, it counts against the "extra" move.
11438 			 */
11439 			rack->r_ctl.sack_moved_extra += moved_two;
11440 			rack->r_ctl.sack_noextra_move += no_extra;
11441 			counter_u64_add(rack_move_some, 1);
11442 		} else {
11443 			/*
11444 			 * else we did not have to move
11445 			 * any more than we would expect.
11446 			 */
11447 			rack->r_ctl.sack_noextra_move += no_extra;
11448 			rack->r_ctl.sack_noextra_move++;
11449 			counter_u64_add(rack_move_none, 1);
11450 		}
11451 		if ((rack->r_ctl.sack_moved_extra > 0xfff00000) ||
11452 		    (rack->r_ctl.sack_noextra_move > 0xfff00000)) {
11453 			rack->r_ctl.sack_moved_extra /= 2;
11454 			rack->r_ctl.sack_noextra_move /= 2;
11455 		}
11456 		if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) {
11457 			/*
11458 			 * If the SACK was not a full MSS then
11459 			 * we add to sack_count the number of
11460 			 * MSS's (or possibly more than
11461 			 * a MSS if its a TSO send) we had to skip by.
11462 			 */
11463 			rack->r_ctl.sack_count += moved_two;
11464 			if (rack->r_ctl.sack_count > 0xfff00000) {
11465 				rack->r_ctl.ack_count /= 2;
11466 				rack->r_ctl.sack_count /= 2;
11467 			}
11468 			counter_u64_add(rack_sack_total, moved_two);
11469 		}
11470 		/*
11471 		 * Now we need to setup for the next
11472 		 * round. First we make sure we won't
11473 		 * exceed the size of our uint32_t on
11474 		 * the various counts, and then clear out
11475 		 * moved_two.
11476 		 */
11477 		moved_two = 0;
11478 		no_extra = 0;
11479 	}
11480 out_with_totals:
11481 	if (num_sack_blks > 1) {
11482 		/*
11483 		 * You get an extra stroke if
11484 		 * you have more than one sack-blk, this
11485 		 * could be where we are skipping forward
11486 		 * and the sack-filter is still working, or
11487 		 * it could be an attacker constantly
11488 		 * moving us.
11489 		 */
11490 		rack->r_ctl.sack_moved_extra++;
11491 		counter_u64_add(rack_move_some, 1);
11492 	}
11493 out:
11494 #ifdef TCP_SAD_DETECTION
11495 	rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp));
11496 #endif
11497 	if (changed) {
11498 		/* Something changed cancel the rack timer */
11499 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
11500 	}
11501 	tsused = tcp_get_usecs(NULL);
11502 	rsm = tcp_rack_output(tp, rack, tsused);
11503 	if ((!IN_FASTRECOVERY(tp->t_flags)) &&
11504 	    rsm &&
11505 	    ((rsm->r_flags & RACK_MUST_RXT) == 0)) {
11506 		/* Enter recovery */
11507 		entered_recovery = 1;
11508 		rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
11509 		/*
11510 		 * When we enter recovery we need to assure we send
11511 		 * one packet.
11512 		 */
11513 		if (rack->rack_no_prr == 0) {
11514 			rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp);
11515 			rack_log_to_prr(rack, 8, 0, __LINE__);
11516 		}
11517 		rack->r_timer_override = 1;
11518 		rack->r_early = 0;
11519 		rack->r_ctl.rc_agg_early = 0;
11520 	} else if (IN_FASTRECOVERY(tp->t_flags) &&
11521 		   rsm &&
11522 		   (rack->r_rr_config == 3)) {
11523 		/*
11524 		 * Assure we can output and we get no
11525 		 * remembered pace time except the retransmit.
11526 		 */
11527 		rack->r_timer_override = 1;
11528 		rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
11529 		rack->r_ctl.rc_resend = rsm;
11530 	}
11531 	if (IN_FASTRECOVERY(tp->t_flags) &&
11532 	    (rack->rack_no_prr == 0) &&
11533 	    (entered_recovery == 0)) {
11534 		rack_update_prr(tp, rack, changed, th_ack);
11535 		if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) &&
11536 		     ((tcp_in_hpts(rack->rc_tp) == 0) &&
11537 		      ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) {
11538 			/*
11539 			 * If you are pacing output you don't want
11540 			 * to override.
11541 			 */
11542 			rack->r_early = 0;
11543 			rack->r_ctl.rc_agg_early = 0;
11544 			rack->r_timer_override = 1;
11545 		}
11546 	}
11547 }
11548 
11549 static void
11550 rack_strike_dupack(struct tcp_rack *rack)
11551 {
11552 	struct rack_sendmap *rsm;
11553 
11554 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
11555 	while (rsm) {
11556 		/*
11557 		 * We need to skip anything already set
11558 		 * to be retransmitted.
11559 		 */
11560 		if ((rsm->r_dupack >= DUP_ACK_THRESHOLD)  ||
11561 		    (rsm->r_flags & RACK_MUST_RXT)) {
11562 			rsm = TAILQ_NEXT(rsm, r_tnext);
11563 			continue;
11564 		}
11565 		break;
11566 	}
11567 	if (rsm && (rsm->r_dupack < 0xff)) {
11568 		rsm->r_dupack++;
11569 		if (rsm->r_dupack >= DUP_ACK_THRESHOLD) {
11570 			struct timeval tv;
11571 			uint32_t cts;
11572 			/*
11573 			 * Here we see if we need to retransmit. For
11574 			 * a SACK type connection if enough time has passed
11575 			 * we will get a return of the rsm. For a non-sack
11576 			 * connection we will get the rsm returned if the
11577 			 * dupack value is 3 or more.
11578 			 */
11579 			cts = tcp_get_usecs(&tv);
11580 			rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts);
11581 			if (rack->r_ctl.rc_resend != NULL) {
11582 				if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) {
11583 					rack_cong_signal(rack->rc_tp, CC_NDUPACK,
11584 							 rack->rc_tp->snd_una, __LINE__);
11585 				}
11586 				rack->r_wanted_output = 1;
11587 				rack->r_timer_override = 1;
11588 				rack_log_retran_reason(rack, rsm, __LINE__, 1, 3);
11589 			}
11590 		} else {
11591 			rack_log_retran_reason(rack, rsm, __LINE__, 0, 3);
11592 		}
11593 	}
11594 }
11595 
11596 static void
11597 rack_check_bottom_drag(struct tcpcb *tp,
11598 		       struct tcp_rack *rack,
11599 		       struct socket *so)
11600 {
11601 	uint32_t segsiz, minseg;
11602 
11603 	segsiz = ctf_fixed_maxseg(tp);
11604 	minseg = segsiz;
11605 	if (tp->snd_max == tp->snd_una) {
11606 		/*
11607 		 * We are doing dynamic pacing and we are way
11608 		 * under. Basically everything got acked while
11609 		 * we were still waiting on the pacer to expire.
11610 		 *
11611 		 * This means we need to boost the b/w in
11612 		 * addition to any earlier boosting of
11613 		 * the multiplier.
11614 		 */
11615 		uint64_t lt_bw;
11616 
11617 		lt_bw = rack_get_lt_bw(rack);
11618 		rack->rc_dragged_bottom = 1;
11619 		rack_validate_multipliers_at_or_above100(rack);
11620 		if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) &&
11621 		    (lt_bw > 0)) {
11622 			/*
11623 			 * Lets use the long-term b/w we have
11624 			 * been getting as a base.
11625 			 */
11626 			if (rack->rc_gp_filled == 0) {
11627 				if (lt_bw > ONE_POINT_TWO_MEG) {
11628 					/*
11629 					 * If we have no measurement
11630 					 * don't let us set in more than
11631 					 * 1.2Mbps. If we are still too
11632 					 * low after pacing with this we
11633 					 * will hopefully have a max b/w
11634 					 * available to sanity check things.
11635 					 */
11636 					lt_bw = ONE_POINT_TWO_MEG;
11637 				}
11638 				rack->r_ctl.rc_rtt_diff = 0;
11639 				rack->r_ctl.gp_bw = lt_bw;
11640 				rack->rc_gp_filled = 1;
11641 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11642 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
11643 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11644 			} else if (lt_bw > rack->r_ctl.gp_bw) {
11645 				rack->r_ctl.rc_rtt_diff = 0;
11646 				if (rack->r_ctl.num_measurements < RACK_REQ_AVG)
11647 					rack->r_ctl.num_measurements = RACK_REQ_AVG;
11648 				rack->r_ctl.gp_bw = lt_bw;
11649 				rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
11650 			} else
11651 				rack_increase_bw_mul(rack, -1, 0, 0, 1);
11652 			if ((rack->gp_ready == 0) &&
11653 			    (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) {
11654 				/* We have enough measurements now */
11655 				rack->gp_ready = 1;
11656 				if (rack->dgp_on ||
11657 				    rack->rack_hibeta)
11658 					rack_set_cc_pacing(rack);
11659 				if (rack->defer_options)
11660 					rack_apply_deferred_options(rack);
11661 			}
11662 		} else {
11663 			/*
11664 			 * zero rtt possibly?, settle for just an old increase.
11665 			 */
11666 			rack_increase_bw_mul(rack, -1, 0, 0, 1);
11667 		}
11668 	} else if ((IN_FASTRECOVERY(tp->t_flags) == 0) &&
11669 		   (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)),
11670 					       minseg)) &&
11671 		   (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11672 		   (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) &&
11673 		   (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <=
11674 		    (segsiz * rack_req_segs))) {
11675 		/*
11676 		 * We are doing dynamic GP pacing and
11677 		 * we have everything except 1MSS or less
11678 		 * bytes left out. We are still pacing away.
11679 		 * And there is data that could be sent, This
11680 		 * means we are inserting delayed ack time in
11681 		 * our measurements because we are pacing too slow.
11682 		 */
11683 		rack_validate_multipliers_at_or_above100(rack);
11684 		rack->rc_dragged_bottom = 1;
11685 		rack_increase_bw_mul(rack, -1, 0, 0, 1);
11686 	}
11687 }
11688 
11689 #ifdef TCP_REQUEST_TRK
11690 static void
11691 rack_log_hybrid(struct tcp_rack *rack, uint32_t seq,
11692 		struct tcp_sendfile_track *cur, uint8_t mod, int line, int err)
11693 {
11694 	int do_log;
11695 
11696 	do_log = tcp_bblogging_on(rack->rc_tp);
11697 	if (do_log == 0) {
11698 		if ((do_log = tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) )== 0)
11699 			return;
11700 		/* We only allow the three below with point logging on */
11701 		if ((mod != HYBRID_LOG_RULES_APP) &&
11702 		    (mod != HYBRID_LOG_RULES_SET) &&
11703 		    (mod != HYBRID_LOG_REQ_COMP))
11704 			return;
11705 
11706 	}
11707 	if (do_log) {
11708 		union tcp_log_stackspecific log;
11709 		struct timeval tv;
11710 
11711 		/* Convert our ms to a microsecond */
11712 		memset(&log, 0, sizeof(log));
11713 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
11714 		log.u_bbr.flex1 = seq;
11715 		log.u_bbr.cwnd_gain = line;
11716 		if (cur != NULL) {
11717 			uint64_t off;
11718 
11719 			log.u_bbr.flex2 = cur->start_seq;
11720 			log.u_bbr.flex3 = cur->end_seq;
11721 			log.u_bbr.flex4 = (uint32_t)((cur->localtime >> 32) & 0x00000000ffffffff);
11722 			log.u_bbr.flex5 = (uint32_t)(cur->localtime & 0x00000000ffffffff);
11723 			log.u_bbr.flex6 = cur->flags;
11724 			log.u_bbr.pkts_out = cur->hybrid_flags;
11725 			log.u_bbr.rttProp = cur->timestamp;
11726 			log.u_bbr.cur_del_rate = cur->cspr;
11727 			log.u_bbr.bw_inuse = cur->start;
11728 			log.u_bbr.applimited = (uint32_t)(cur->end & 0x00000000ffffffff);
11729 			log.u_bbr.delivered = (uint32_t)((cur->end >> 32) & 0x00000000ffffffff) ;
11730 			log.u_bbr.epoch = (uint32_t)(cur->deadline & 0x00000000ffffffff);
11731 			log.u_bbr.lt_epoch = (uint32_t)((cur->deadline >> 32) & 0x00000000ffffffff) ;
11732 			log.u_bbr.bbr_state = 1;
11733 #ifdef TCP_REQUEST_TRK
11734 			off = (uint64_t)(cur) - (uint64_t)(&rack->rc_tp->t_tcpreq_info[0]);
11735 			log.u_bbr.use_lt_bw = (uint8_t)(off / sizeof(struct tcp_sendfile_track));
11736 #endif
11737 		} else {
11738 			log.u_bbr.flex2 = err;
11739 		}
11740 		/*
11741 		 * Fill in flex7 to be CHD (catchup|hybrid|DGP)
11742 		 */
11743 		log.u_bbr.flex7 = rack->rc_catch_up;
11744 		log.u_bbr.flex7 <<= 1;
11745 		log.u_bbr.flex7 |= rack->rc_hybrid_mode;
11746 		log.u_bbr.flex7 <<= 1;
11747 		log.u_bbr.flex7 |= rack->dgp_on;
11748 		log.u_bbr.flex8 = mod;
11749 		log.u_bbr.delRate = rack->r_ctl.bw_rate_cap;
11750 		log.u_bbr.bbr_substate = rack->r_ctl.client_suggested_maxseg;
11751 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
11752 		log.u_bbr.pkt_epoch = rack->rc_tp->tcp_hybrid_start;
11753 		log.u_bbr.lost = rack->rc_tp->tcp_hybrid_error;
11754 		log.u_bbr.pacing_gain = (uint16_t)rack->rc_tp->tcp_hybrid_stop;
11755 		tcp_log_event(rack->rc_tp, NULL,
11756 		    &rack->rc_inp->inp_socket->so_rcv,
11757 		    &rack->rc_inp->inp_socket->so_snd,
11758 		    TCP_HYBRID_PACING_LOG, 0,
11759 	            0, &log, false, NULL, __func__, __LINE__, &tv);
11760 	}
11761 }
11762 #endif
11763 
11764 #ifdef TCP_REQUEST_TRK
11765 static void
11766 rack_set_dgp_hybrid_mode(struct tcp_rack *rack, tcp_seq seq, uint32_t len)
11767 {
11768 	struct tcp_sendfile_track *rc_cur;
11769 	struct tcpcb *tp;
11770 	int err = 0;
11771 
11772 	rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, seq);
11773 	if (rc_cur == NULL) {
11774 		/* If not in the beginning what about the end piece */
11775 		if (rack->rc_hybrid_mode)
11776 			rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11777 		rc_cur = tcp_req_find_req_for_seq(rack->rc_tp, (seq + len - 1));
11778 	} else {
11779 		err = 12345;
11780 	}
11781 	/* If we find no parameters we are in straight DGP mode */
11782 	if(rc_cur == NULL) {
11783 		/* None found for this seq, just DGP for now */
11784 		rack->r_ctl.client_suggested_maxseg = 0;
11785 		rack->rc_catch_up = 0;
11786 		rack->r_ctl.bw_rate_cap = 0;
11787 		if (rack->rc_hybrid_mode)
11788 			rack_log_hybrid(rack, (seq + len - 1), NULL, HYBRID_LOG_NO_RANGE, __LINE__, err);
11789 		if (rack->r_ctl.rc_last_sft) {
11790 			rack->r_ctl.rc_last_sft = NULL;
11791 		}
11792 		return;
11793 	}
11794 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_WASSET) == 0) {
11795 		/* This entry was never setup for hybrid pacing on/off etc */
11796 		return;
11797 	}
11798 	/*
11799 	 * Ok if we have a new entry *or* have never
11800 	 * set up an entry we need to proceed. If
11801 	 * we have already set it up this entry we
11802 	 * just continue along with what we already
11803 	 * setup.
11804 	 */
11805 	tp = rack->rc_tp;
11806 	if ((rack->r_ctl.rc_last_sft != NULL) &&
11807 	    (rack->r_ctl.rc_last_sft == rc_cur)) {
11808 		/* Its already in place */
11809 		if (rack->rc_hybrid_mode)
11810 			rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_ISSAME, __LINE__, 0);
11811 		return;
11812 	}
11813 	if (rack->rc_hybrid_mode == 0) {
11814 		rack->r_ctl.rc_last_sft = rc_cur;
11815 		rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11816 		return;
11817 	}
11818 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CSPR) && rc_cur->cspr){
11819 		/* Compensate for all the header overhead's */
11820 		rack->r_ctl.bw_rate_cap	= rack_compensate_for_linerate(rack, rc_cur->cspr);
11821 	} else
11822 		rack->r_ctl.bw_rate_cap = 0;
11823 	if (rc_cur->hybrid_flags & TCP_HYBRID_PACING_H_MS)
11824 		rack->r_ctl.client_suggested_maxseg = rc_cur->hint_maxseg;
11825 	else
11826 		rack->r_ctl.client_suggested_maxseg = 0;
11827 	if ((rc_cur->hybrid_flags & TCP_HYBRID_PACING_CU) &&
11828 	    (rc_cur->cspr > 0)) {
11829 		uint64_t len;
11830 
11831 		rack->rc_catch_up = 1;
11832 		/*
11833 		 * Calculate the deadline time, first set the
11834 		 * time to when the request arrived.
11835 		 */
11836 		rc_cur->deadline = rc_cur->localtime;
11837 		/*
11838 		 * Next calculate the length and compensate for
11839 		 * TLS if need be.
11840 		 */
11841 		len = rc_cur->end - rc_cur->start;
11842 		if (tp->t_inpcb.inp_socket->so_snd.sb_tls_info) {
11843 			/*
11844 			 * This session is doing TLS. Take a swag guess
11845 			 * at the overhead.
11846 			 */
11847 			len += tcp_estimate_tls_overhead(tp->t_inpcb.inp_socket, len);
11848 		}
11849 		/*
11850 		 * Now considering the size, and the cspr, what is the time that
11851 		 * would be required at the cspr rate. Here we use the raw
11852 		 * cspr value since the client only looks at the raw data. We
11853 		 * do use len which includes TLS overhead, but not the TCP/IP etc.
11854 		 * That will get made up for in the CU pacing rate set.
11855 		 */
11856 		len *= HPTS_USEC_IN_SEC;
11857 		len /= rc_cur->cspr;
11858 		rc_cur->deadline += len;
11859 	} else {
11860 		rack->rc_catch_up = 0;
11861 		rc_cur->deadline = 0;
11862 	}
11863 	if (rack->r_ctl.client_suggested_maxseg != 0) {
11864 		/*
11865 		 * We need to reset the max pace segs if we have a
11866 		 * client_suggested_maxseg.
11867 		 */
11868 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
11869 	}
11870 	rack_log_hybrid(rack, seq, rc_cur, HYBRID_LOG_RULES_APP, __LINE__, 0);
11871 	/* Remember it for next time and for CU mode */
11872 	rack->r_ctl.rc_last_sft = rc_cur;
11873 }
11874 #endif
11875 
11876 static void
11877 rack_chk_req_and_hybrid_on_out(struct tcp_rack *rack, tcp_seq seq, uint32_t len, uint64_t cts)
11878 {
11879 #ifdef TCP_REQUEST_TRK
11880 	struct tcp_sendfile_track *ent;
11881 
11882 	ent = rack->r_ctl.rc_last_sft;
11883 	if ((ent == NULL) ||
11884 	    (ent->flags == TCP_TRK_TRACK_FLG_EMPTY) ||
11885 	    (SEQ_GEQ(seq, ent->end_seq))) {
11886 		/* Time to update the track. */
11887 		rack_set_dgp_hybrid_mode(rack, seq, len);
11888 		ent = rack->r_ctl.rc_last_sft;
11889 	}
11890 	/* Out of all */
11891 	if (ent == NULL) {
11892 		return;
11893 	}
11894 	if (SEQ_LT(ent->end_seq, (seq + len))) {
11895 		/*
11896 		 * This is the case where our end_seq guess
11897 		 * was wrong. This is usually due to TLS having
11898 		 * more bytes then our guess. It could also be the
11899 		 * case that the client sent in two requests closely
11900 		 * and the SB is full of both so we are sending part
11901 		 * of each (end|beg). In such a case lets move this
11902 		 * guys end to match the end of this send. That
11903 		 * way it will complete when all of it is acked.
11904 		 */
11905 		ent->end_seq = (seq + len);
11906 		if (rack->rc_hybrid_mode)
11907 			rack_log_hybrid_bw(rack, seq, len, 0, 0, HYBRID_LOG_EXTEND, 0, ent, __LINE__);
11908 	}
11909 	/* Now validate we have set the send time of this one */
11910 	if ((ent->flags & TCP_TRK_TRACK_FLG_FSND) == 0) {
11911 		ent->flags |= TCP_TRK_TRACK_FLG_FSND;
11912 		ent->first_send = cts;
11913 		ent->sent_at_fs = rack->rc_tp->t_sndbytes;
11914 		ent->rxt_at_fs = rack->rc_tp->t_snd_rxt_bytes;
11915 	}
11916 #endif
11917 }
11918 
11919 static void
11920 rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount)
11921 {
11922 	/*
11923 	 * The fast output path is enabled and we
11924 	 * have moved the cumack forward. Lets see if
11925 	 * we can expand forward the fast path length by
11926 	 * that amount. What we would ideally like to
11927 	 * do is increase the number of bytes in the
11928 	 * fast path block (left_to_send) by the
11929 	 * acked amount. However we have to gate that
11930 	 * by two factors:
11931 	 * 1) The amount outstanding and the rwnd of the peer
11932 	 *    (i.e. we don't want to exceed the rwnd of the peer).
11933 	 *    <and>
11934 	 * 2) The amount of data left in the socket buffer (i.e.
11935 	 *    we can't send beyond what is in the buffer).
11936 	 *
11937 	 * Note that this does not take into account any increase
11938 	 * in the cwnd. We will only extend the fast path by
11939 	 * what was acked.
11940 	 */
11941 	uint32_t new_total, gating_val;
11942 
11943 	new_total = acked_amount + rack->r_ctl.fsb.left_to_send;
11944 	gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)),
11945 			 (tp->snd_wnd - (tp->snd_max - tp->snd_una)));
11946 	if (new_total <= gating_val) {
11947 		/* We can increase left_to_send by the acked amount */
11948 		counter_u64_add(rack_extended_rfo, 1);
11949 		rack->r_ctl.fsb.left_to_send = new_total;
11950 		KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))),
11951 			("rack:%p left_to_send:%u sbavail:%u out:%u",
11952 			 rack, rack->r_ctl.fsb.left_to_send,
11953 			 sbavail(&rack->rc_inp->inp_socket->so_snd),
11954 			 (tp->snd_max - tp->snd_una)));
11955 
11956 	}
11957 }
11958 
11959 static void
11960 rack_adjust_sendmap_head(struct tcp_rack *rack, struct sockbuf *sb)
11961 {
11962 	/*
11963 	 * Here any sendmap entry that points to the
11964 	 * beginning mbuf must be adjusted to the correct
11965 	 * offset. This must be called with:
11966 	 * 1) The socket buffer locked
11967 	 * 2) snd_una adjusted to its new position.
11968 	 *
11969 	 * Note that (2) implies rack_ack_received has also
11970 	 * been called and all the sbcut's have been done.
11971 	 *
11972 	 * We grab the first mbuf in the socket buffer and
11973 	 * then go through the front of the sendmap, recalculating
11974 	 * the stored offset for any sendmap entry that has
11975 	 * that mbuf. We must use the sb functions to do this
11976 	 * since its possible an add was done has well as
11977 	 * the subtraction we may have just completed. This should
11978 	 * not be a penalty though, since we just referenced the sb
11979 	 * to go in and trim off the mbufs that we freed (of course
11980 	 * there will be a penalty for the sendmap references though).
11981 	 *
11982 	 * Note also with INVARIANT on, we validate with a KASSERT
11983 	 * that the first sendmap entry has a soff of 0.
11984 	 *
11985 	 */
11986 	struct mbuf *m;
11987 	struct rack_sendmap *rsm;
11988 	tcp_seq snd_una;
11989 #ifdef INVARIANTS
11990 	int first_processed = 0;
11991 #endif
11992 
11993 	snd_una = rack->rc_tp->snd_una;
11994 	SOCKBUF_LOCK_ASSERT(sb);
11995 	m = sb->sb_mb;
11996 	rsm = tqhash_min(rack->r_ctl.tqh);
11997 	if ((rsm == NULL) || (m == NULL)) {
11998 		/* Nothing outstanding */
11999 		return;
12000 	}
12001 	/* The very first RSM's mbuf must point to the head mbuf in the sb */
12002 	KASSERT((rsm->m == m),
12003 		("Rack:%p sb:%p rsm:%p -- first rsm mbuf not aligned to sb",
12004 		 rack, sb, rsm));
12005 	while (rsm->m && (rsm->m == m)) {
12006 		/* one to adjust */
12007 #ifdef INVARIANTS
12008 		struct mbuf *tm;
12009 		uint32_t soff;
12010 
12011 		tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff);
12012 		if ((rsm->orig_m_len != m->m_len) ||
12013 		    (rsm->orig_t_space != M_TRAILINGROOM(m))){
12014 			rack_adjust_orig_mlen(rsm);
12015 		}
12016 		if (first_processed == 0) {
12017 			KASSERT((rsm->soff == 0),
12018 				("Rack:%p rsm:%p -- rsm at head but soff not zero",
12019 				 rack, rsm));
12020 			first_processed = 1;
12021 		}
12022 		if ((rsm->soff != soff) || (rsm->m != tm)) {
12023 			/*
12024 			 * This is not a fatal error, we anticipate it
12025 			 * might happen (the else code), so we count it here
12026 			 * so that under invariant we can see that it really
12027 			 * does happen.
12028 			 */
12029 			counter_u64_add(rack_adjust_map_bw, 1);
12030 		}
12031 		rsm->m = tm;
12032 		rsm->soff = soff;
12033 		if (tm) {
12034 			rsm->orig_m_len = rsm->m->m_len;
12035 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
12036 		} else {
12037 			rsm->orig_m_len = 0;
12038 			rsm->orig_t_space = 0;
12039 		}
12040 #else
12041 		rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff);
12042 		if (rsm->m) {
12043 			rsm->orig_m_len = rsm->m->m_len;
12044 			rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
12045 		} else {
12046 			rsm->orig_m_len = 0;
12047 			rsm->orig_t_space = 0;
12048 		}
12049 #endif
12050 		rsm = tqhash_next(rack->r_ctl.tqh, rsm);
12051 		if (rsm == NULL)
12052 			break;
12053 	}
12054 }
12055 
12056 #ifdef TCP_REQUEST_TRK
12057 static inline void
12058 rack_req_check_for_comp(struct tcp_rack *rack, tcp_seq th_ack)
12059 {
12060 	struct tcp_sendfile_track *ent;
12061 	int i;
12062 
12063 	if ((rack->rc_hybrid_mode == 0) &&
12064 	    (tcp_bblogging_point_on(rack->rc_tp, TCP_BBPOINT_REQ_LEVEL_LOGGING) == 0)) {
12065 		/*
12066 		 * Just do normal completions hybrid pacing is not on
12067 		 * and CLDL is off as well.
12068 		 */
12069 		tcp_req_check_for_comp(rack->rc_tp, th_ack);
12070 		return;
12071 	}
12072 	/*
12073 	 * Originally I was just going to find the th_ack associated
12074 	 * with an entry. But then I realized a large strech ack could
12075 	 * in theory ack two or more requests at once. So instead we
12076 	 * need to find all entries that are completed by th_ack not
12077 	 * just a single entry and do our logging.
12078 	 */
12079 	ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
12080 	while (ent != NULL) {
12081 		/*
12082 		 * We may be doing hybrid pacing or CLDL and need more details possibly
12083 		 * so we do it manually instead of calling
12084 		 * tcp_req_check_for_comp()
12085 		 */
12086 		uint64_t laa, tim, data, cbw, ftim;
12087 
12088 		/* Ok this ack frees it */
12089 		rack_log_hybrid(rack, th_ack,
12090 				ent, HYBRID_LOG_REQ_COMP, __LINE__, 0);
12091 		rack_log_hybrid_sends(rack, ent, __LINE__);
12092 		/* calculate the time based on the ack arrival */
12093 		data = ent->end - ent->start;
12094 		laa = tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time);
12095 		if (ent->flags & TCP_TRK_TRACK_FLG_FSND) {
12096 			if (ent->first_send > ent->localtime)
12097 				ftim = ent->first_send;
12098 			else
12099 				ftim = ent->localtime;
12100 		} else {
12101 			/* TSNH */
12102 			ftim = ent->localtime;
12103 		}
12104 		if (laa > ent->localtime)
12105 			tim = laa - ftim;
12106 		else
12107 			tim = 0;
12108 		cbw = data * HPTS_USEC_IN_SEC;
12109 		if (tim > 0)
12110 			cbw /= tim;
12111 		else
12112 			cbw = 0;
12113 		rack_log_hybrid_bw(rack, th_ack, cbw, tim, data, HYBRID_LOG_BW_MEASURE, 0, ent, __LINE__);
12114 		/*
12115 		 * Check to see if we are freeing what we are pointing to send wise
12116 		 * if so be sure to NULL the pointer so we know we are no longer
12117 		 * set to anything.
12118 		 */
12119 		if (ent == rack->r_ctl.rc_last_sft)
12120 			rack->r_ctl.rc_last_sft = NULL;
12121 		/* Generate the log that the tcp_netflix call would have */
12122 		tcp_req_log_req_info(rack->rc_tp, ent,
12123 				      i, TCP_TRK_REQ_LOG_FREED, 0, 0);
12124 		/* Free it and see if there is another one */
12125 		tcp_req_free_a_slot(rack->rc_tp, ent);
12126 		ent = tcp_req_find_a_req_that_is_completed_by(rack->rc_tp, th_ack, &i);
12127 	}
12128 }
12129 #endif
12130 
12131 
12132 /*
12133  * Return value of 1, we do not need to call rack_process_data().
12134  * return value of 0, rack_process_data can be called.
12135  * For ret_val if its 0 the TCP is locked, if its non-zero
12136  * its unlocked and probably unsafe to touch the TCB.
12137  */
12138 static int
12139 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12140     struct tcpcb *tp, struct tcpopt *to,
12141     uint32_t tiwin, int32_t tlen,
12142     int32_t * ofia, int32_t thflags, int32_t *ret_val)
12143 {
12144 	int32_t ourfinisacked = 0;
12145 	int32_t nsegs, acked_amount;
12146 	int32_t acked;
12147 	struct mbuf *mfree;
12148 	struct tcp_rack *rack;
12149 	int32_t under_pacing = 0;
12150 	int32_t recovery = 0;
12151 
12152 	INP_WLOCK_ASSERT(tptoinpcb(tp));
12153 
12154 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12155 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
12156 		__ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val,
12157 				      &rack->r_ctl.challenge_ack_ts,
12158 				      &rack->r_ctl.challenge_ack_cnt);
12159 		rack->r_wanted_output = 1;
12160 		return (1);
12161 	}
12162 	if (rack->gp_ready &&
12163 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
12164 		under_pacing = 1;
12165 	}
12166 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
12167 		int in_rec, dup_ack_struck = 0;
12168 		int dsack_seen = 0, sacks_seen = 0;
12169 
12170 		in_rec = IN_FASTRECOVERY(tp->t_flags);
12171 		if (rack->rc_in_persist) {
12172 			tp->t_rxtshift = 0;
12173 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12174 				      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12175 		}
12176 
12177 		if ((th->th_ack == tp->snd_una) &&
12178 		    (tiwin == tp->snd_wnd) &&
12179 		    ((to->to_flags & TOF_SACK) == 0)) {
12180 			rack_strike_dupack(rack);
12181 			dup_ack_struck = 1;
12182 		}
12183 		rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)),
12184 			     dup_ack_struck, &dsack_seen, &sacks_seen);
12185 		if ((rack->sack_attack_disable > 0) &&
12186 		    (th->th_ack == tp->snd_una) &&
12187 		    (tiwin == tp->snd_wnd) &&
12188 		    (dsack_seen == 0) &&
12189 		    (sacks_seen > 0)) {
12190 			/*
12191 			 * If sacks have been disabled we may
12192 			 * want to strike a dup-ack "ignoring" the
12193 			 * sack as long as the sack was not a "dsack". Note
12194 			 * that if no sack is sent (TOF_SACK is off) then the
12195 			 * normal dsack code above rack_log_ack() would have
12196 			 * already struck. So this is just to catch the case
12197 			 * were we are ignoring sacks from this guy due to
12198 			 * it being a suspected attacker.
12199 			 */
12200 			rack_strike_dupack(rack);
12201 		}
12202 
12203 	}
12204 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
12205 		/*
12206 		 * Old ack, behind (or duplicate to) the last one rcv'd
12207 		 * Note: We mark reordering is occuring if its
12208 		 * less than and we have not closed our window.
12209 		 */
12210 		if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) {
12211 			rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
12212 			if (rack->r_ctl.rc_reorder_ts == 0)
12213 				rack->r_ctl.rc_reorder_ts = 1;
12214 		}
12215 		return (0);
12216 	}
12217 	/*
12218 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
12219 	 * something we sent.
12220 	 */
12221 	if (tp->t_flags & TF_NEEDSYN) {
12222 		/*
12223 		 * T/TCP: Connection was half-synchronized, and our SYN has
12224 		 * been ACK'd (so connection is now fully synchronized).  Go
12225 		 * to non-starred state, increment snd_una for ACK of SYN,
12226 		 * and check if we can do window scaling.
12227 		 */
12228 		tp->t_flags &= ~TF_NEEDSYN;
12229 		tp->snd_una++;
12230 		/* Do window scaling? */
12231 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
12232 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
12233 			tp->rcv_scale = tp->request_r_scale;
12234 			/* Send window already scaled. */
12235 		}
12236 	}
12237 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12238 
12239 	acked = BYTES_THIS_ACK(tp, th);
12240 	if (acked) {
12241 		/*
12242 		 * Any time we move the cum-ack forward clear
12243 		 * keep-alive tied probe-not-answered. The
12244 		 * persists clears its own on entry.
12245 		 */
12246 		rack->probe_not_answered = 0;
12247 	}
12248 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
12249 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
12250 	/*
12251 	 * If we just performed our first retransmit, and the ACK arrives
12252 	 * within our recovery window, then it was a mistake to do the
12253 	 * retransmit in the first place.  Recover our original cwnd and
12254 	 * ssthresh, and proceed to transmit where we left off.
12255 	 */
12256 	if ((tp->t_flags & TF_PREVVALID) &&
12257 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
12258 		tp->t_flags &= ~TF_PREVVALID;
12259 		if (tp->t_rxtshift == 1 &&
12260 		    (int)(ticks - tp->t_badrxtwin) < 0)
12261 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
12262 	}
12263 	if (acked) {
12264 		/* assure we are not backed off */
12265 		tp->t_rxtshift = 0;
12266 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
12267 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
12268 		rack->rc_tlp_in_progress = 0;
12269 		rack->r_ctl.rc_tlp_cnt_out = 0;
12270 		/*
12271 		 * If it is the RXT timer we want to
12272 		 * stop it, so we can restart a TLP.
12273 		 */
12274 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
12275 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12276 #ifdef TCP_REQUEST_TRK
12277 		rack_req_check_for_comp(rack, th->th_ack);
12278 #endif
12279 	}
12280 	/*
12281 	 * If we have a timestamp reply, update smoothed round trip time. If
12282 	 * no timestamp is present but transmit timer is running and timed
12283 	 * sequence number was acked, update smoothed round trip time. Since
12284 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
12285 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
12286 	 * timer.
12287 	 *
12288 	 * Some boxes send broken timestamp replies during the SYN+ACK
12289 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
12290 	 * and blow up the retransmit timer.
12291 	 */
12292 	/*
12293 	 * If all outstanding data is acked, stop retransmit timer and
12294 	 * remember to restart (more output or persist). If there is more
12295 	 * data to be acked, restart retransmit timer, using current
12296 	 * (possibly backed-off) value.
12297 	 */
12298 	if (acked == 0) {
12299 		if (ofia)
12300 			*ofia = ourfinisacked;
12301 		return (0);
12302 	}
12303 	if (IN_RECOVERY(tp->t_flags)) {
12304 		if (SEQ_LT(th->th_ack, tp->snd_recover) &&
12305 		    (SEQ_LT(th->th_ack, tp->snd_max))) {
12306 			tcp_rack_partialack(tp);
12307 		} else {
12308 			rack_post_recovery(tp, th->th_ack);
12309 			recovery = 1;
12310 		}
12311 	}
12312 	/*
12313 	 * Let the congestion control algorithm update congestion control
12314 	 * related information. This typically means increasing the
12315 	 * congestion window.
12316 	 */
12317 	rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery);
12318 	SOCKBUF_LOCK(&so->so_snd);
12319 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
12320 	tp->snd_wnd -= acked_amount;
12321 	mfree = sbcut_locked(&so->so_snd, acked_amount);
12322 	if ((sbused(&so->so_snd) == 0) &&
12323 	    (acked > acked_amount) &&
12324 	    (tp->t_state >= TCPS_FIN_WAIT_1) &&
12325 	    (tp->t_flags & TF_SENTFIN)) {
12326 		/*
12327 		 * We must be sure our fin
12328 		 * was sent and acked (we can be
12329 		 * in FIN_WAIT_1 without having
12330 		 * sent the fin).
12331 		 */
12332 		ourfinisacked = 1;
12333 	}
12334 	tp->snd_una = th->th_ack;
12335 	/* wakeups? */
12336 	if (acked_amount && sbavail(&so->so_snd))
12337 		rack_adjust_sendmap_head(rack, &so->so_snd);
12338 	rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
12339 	/* NB: sowwakeup_locked() does an implicit unlock. */
12340 	sowwakeup_locked(so);
12341 	/* now check the rxt clamps */
12342 	if ((recovery == 1) &&
12343 	    (rack->excess_rxt_on) &&
12344 	    (rack->r_cwnd_was_clamped == 0))  {
12345 		do_rack_excess_rxt(tp, rack);
12346 	} else if (rack->r_cwnd_was_clamped)
12347 		do_rack_check_for_unclamp(tp, rack);
12348 	m_freem(mfree);
12349 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
12350 		tp->snd_recover = tp->snd_una;
12351 
12352 	if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
12353 		tp->snd_nxt = tp->snd_max;
12354 	}
12355 	if (under_pacing &&
12356 	    (rack->use_fixed_rate == 0) &&
12357 	    (rack->in_probe_rtt == 0) &&
12358 	    rack->rc_gp_dyn_mul &&
12359 	    rack->rc_always_pace) {
12360 		/* Check if we are dragging bottom */
12361 		rack_check_bottom_drag(tp, rack, so);
12362 	}
12363 	if (tp->snd_una == tp->snd_max) {
12364 		/* Nothing left outstanding */
12365 		tp->t_flags &= ~TF_PREVVALID;
12366 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
12367 		rack->r_ctl.retran_during_recovery = 0;
12368 		rack->r_ctl.dsack_byte_cnt = 0;
12369 		if (rack->r_ctl.rc_went_idle_time == 0)
12370 			rack->r_ctl.rc_went_idle_time = 1;
12371 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
12372 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
12373 			tp->t_acktime = 0;
12374 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
12375 		rack->rc_suspicious = 0;
12376 		/* Set need output so persist might get set */
12377 		rack->r_wanted_output = 1;
12378 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12379 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
12380 		    (sbavail(&so->so_snd) == 0) &&
12381 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
12382 			/*
12383 			 * The socket was gone and the
12384 			 * peer sent data (now or in the past), time to
12385 			 * reset him.
12386 			 */
12387 			*ret_val = 1;
12388 			/* tcp_close will kill the inp pre-log the Reset */
12389 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
12390 			tp = tcp_close(tp);
12391 			ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
12392 			return (1);
12393 		}
12394 	}
12395 	if (ofia)
12396 		*ofia = ourfinisacked;
12397 	return (0);
12398 }
12399 
12400 
12401 static void
12402 rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line,
12403 		  int dir, uint32_t flags, struct rack_sendmap *rsm)
12404 {
12405 	if (tcp_bblogging_on(rack->rc_tp)) {
12406 		union tcp_log_stackspecific log;
12407 		struct timeval tv;
12408 
12409 		memset(&log, 0, sizeof(log));
12410 		log.u_bbr.flex1 = cnt;
12411 		log.u_bbr.flex2 = split;
12412 		log.u_bbr.flex3 = out;
12413 		log.u_bbr.flex4 = line;
12414 		log.u_bbr.flex5 = rack->r_must_retran;
12415 		log.u_bbr.flex6 = flags;
12416 		log.u_bbr.flex7 = rack->rc_has_collapsed;
12417 		log.u_bbr.flex8 = dir;	/*
12418 					 * 1 is collapsed, 0 is uncollapsed,
12419 					 * 2 is log of a rsm being marked, 3 is a split.
12420 					 */
12421 		if (rsm == NULL)
12422 			log.u_bbr.rttProp = 0;
12423 		else
12424 			log.u_bbr.rttProp = (uint64_t)rsm;
12425 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
12426 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
12427 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
12428 		    &rack->rc_inp->inp_socket->so_rcv,
12429 		    &rack->rc_inp->inp_socket->so_snd,
12430 		    TCP_RACK_LOG_COLLAPSE, 0,
12431 		    0, &log, false, &tv);
12432 	}
12433 }
12434 
12435 static void
12436 rack_collapsed_window(struct tcp_rack *rack, uint32_t out, tcp_seq th_ack, int line)
12437 {
12438 	/*
12439 	 * Here all we do is mark the collapsed point and set the flag.
12440 	 * This may happen again and again, but there is no
12441 	 * sense splitting our map until we know where the
12442 	 * peer finally lands in the collapse.
12443 	 */
12444 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12445 	if ((rack->rc_has_collapsed == 0) ||
12446 	    (rack->r_ctl.last_collapse_point != (th_ack + rack->rc_tp->snd_wnd)))
12447 		counter_u64_add(rack_collapsed_win_seen, 1);
12448 	rack->r_ctl.last_collapse_point = th_ack + rack->rc_tp->snd_wnd;
12449 	rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max;
12450 	rack->rc_has_collapsed = 1;
12451 	rack->r_collapse_point_valid = 1;
12452 	rack_log_collapse(rack, 0, th_ack, rack->r_ctl.last_collapse_point, line, 1, 0, NULL);
12453 }
12454 
12455 static void
12456 rack_un_collapse_window(struct tcp_rack *rack, int line)
12457 {
12458 	struct rack_sendmap *nrsm, *rsm;
12459 	int cnt = 0, split = 0;
12460 	int insret __diagused;
12461 
12462 
12463 	tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_WND);
12464 	rack->rc_has_collapsed = 0;
12465 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
12466 	if (rsm == NULL) {
12467 		/* Nothing to do maybe the peer ack'ed it all */
12468 		rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12469 		return;
12470 	}
12471 	/* Now do we need to split this one? */
12472 	if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) {
12473 		rack_log_collapse(rack, rsm->r_start, rsm->r_end,
12474 				  rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm);
12475 		nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT);
12476 		if (nrsm == NULL) {
12477 			/* We can't get a rsm, mark all? */
12478 			nrsm = rsm;
12479 			goto no_split;
12480 		}
12481 		/* Clone it */
12482 		split = 1;
12483 		rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point);
12484 #ifndef INVARIANTS
12485 		(void)tqhash_insert(rack->r_ctl.tqh, nrsm);
12486 #else
12487 		if ((insret = tqhash_insert(rack->r_ctl.tqh, nrsm)) != 0) {
12488 			panic("Insert in tailq_hash of %p fails ret:%d rack:%p rsm:%p",
12489 			      nrsm, insret, rack, rsm);
12490 		}
12491 #endif
12492 		rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT,
12493 				 rack->r_ctl.last_collapse_point, __LINE__);
12494 		if (rsm->r_in_tmap) {
12495 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
12496 			nrsm->r_in_tmap = 1;
12497 		}
12498 		/*
12499 		 * Set in the new RSM as the
12500 		 * collapsed starting point
12501 		 */
12502 		rsm = nrsm;
12503 	}
12504 
12505 no_split:
12506 	TQHASH_FOREACH_FROM(nrsm, rack->r_ctl.tqh, rsm)  {
12507 		cnt++;
12508 		nrsm->r_flags |= RACK_RWND_COLLAPSED;
12509 		rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm);
12510 		cnt++;
12511 	}
12512 	if (cnt) {
12513 		counter_u64_add(rack_collapsed_win, 1);
12514 	}
12515 	rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL);
12516 }
12517 
12518 static void
12519 rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack,
12520 			int32_t tlen, int32_t tfo_syn)
12521 {
12522 	if (DELAY_ACK(tp, tlen) || tfo_syn) {
12523 		rack_timer_cancel(tp, rack,
12524 				  rack->r_ctl.rc_rcvtime, __LINE__);
12525 		tp->t_flags |= TF_DELACK;
12526 	} else {
12527 		rack->r_wanted_output = 1;
12528 		tp->t_flags |= TF_ACKNOW;
12529 	}
12530 }
12531 
12532 static void
12533 rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack)
12534 {
12535 	/*
12536 	 * If fast output is in progress, lets validate that
12537 	 * the new window did not shrink on us and make it
12538 	 * so fast output should end.
12539 	 */
12540 	if (rack->r_fast_output) {
12541 		uint32_t out;
12542 
12543 		/*
12544 		 * Calculate what we will send if left as is
12545 		 * and compare that to our send window.
12546 		 */
12547 		out = ctf_outstanding(tp);
12548 		if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) {
12549 			/* ok we have an issue */
12550 			if (out >= tp->snd_wnd) {
12551 				/* Turn off fast output the window is met or collapsed */
12552 				rack->r_fast_output = 0;
12553 			} else {
12554 				/* we have some room left */
12555 				rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out;
12556 				if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) {
12557 					/* If not at least 1 full segment never mind */
12558 					rack->r_fast_output = 0;
12559 				}
12560 			}
12561 		}
12562 	}
12563 }
12564 
12565 
12566 /*
12567  * Return value of 1, the TCB is unlocked and most
12568  * likely gone, return value of 0, the TCP is still
12569  * locked.
12570  */
12571 static int
12572 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
12573     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
12574     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
12575 {
12576 	/*
12577 	 * Update window information. Don't look at window if no ACK: TAC's
12578 	 * send garbage on first SYN.
12579 	 */
12580 	int32_t nsegs;
12581 	int32_t tfo_syn;
12582 	struct tcp_rack *rack;
12583 
12584 	INP_WLOCK_ASSERT(tptoinpcb(tp));
12585 
12586 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12587 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12588 	if ((thflags & TH_ACK) &&
12589 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
12590 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
12591 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
12592 		/* keep track of pure window updates */
12593 		if (tlen == 0 &&
12594 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
12595 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
12596 		tp->snd_wnd = tiwin;
12597 		rack_validate_fo_sendwin_up(tp, rack);
12598 		tp->snd_wl1 = th->th_seq;
12599 		tp->snd_wl2 = th->th_ack;
12600 		if (tp->snd_wnd > tp->max_sndwnd)
12601 			tp->max_sndwnd = tp->snd_wnd;
12602 		rack->r_wanted_output = 1;
12603 	} else if (thflags & TH_ACK) {
12604 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
12605 			tp->snd_wnd = tiwin;
12606 			rack_validate_fo_sendwin_up(tp, rack);
12607 			tp->snd_wl1 = th->th_seq;
12608 			tp->snd_wl2 = th->th_ack;
12609 		}
12610 	}
12611 	if (tp->snd_wnd < ctf_outstanding(tp))
12612 		/* The peer collapsed the window */
12613 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
12614 	else if (rack->rc_has_collapsed)
12615 		rack_un_collapse_window(rack, __LINE__);
12616 	if ((rack->r_collapse_point_valid) &&
12617 	    (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point)))
12618 		rack->r_collapse_point_valid = 0;
12619 	/* Was persist timer active and now we have window space? */
12620 	if ((rack->rc_in_persist != 0) &&
12621 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
12622 				rack->r_ctl.rc_pace_min_segs))) {
12623 		rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime);
12624 		tp->snd_nxt = tp->snd_max;
12625 		/* Make sure we output to start the timer */
12626 		rack->r_wanted_output = 1;
12627 	}
12628 	/* Do we enter persists? */
12629 	if ((rack->rc_in_persist == 0) &&
12630 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
12631 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
12632 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
12633 	    sbavail(&tptosocket(tp)->so_snd) &&
12634 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
12635 		/*
12636 		 * Here the rwnd is less than
12637 		 * the pacing size, we are established,
12638 		 * nothing is outstanding, and there is
12639 		 * data to send. Enter persists.
12640 		 */
12641 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
12642 	}
12643 	if (tp->t_flags2 & TF2_DROP_AF_DATA) {
12644 		m_freem(m);
12645 		return (0);
12646 	}
12647 	/*
12648 	 * don't process the URG bit, ignore them drag
12649 	 * along the up.
12650 	 */
12651 	tp->rcv_up = tp->rcv_nxt;
12652 
12653 	/*
12654 	 * Process the segment text, merging it into the TCP sequencing
12655 	 * queue, and arranging for acknowledgment of receipt if necessary.
12656 	 * This process logically involves adjusting tp->rcv_wnd as data is
12657 	 * presented to the user (this happens in tcp_usrreq.c, case
12658 	 * PRU_RCVD).  If a FIN has already been received on this connection
12659 	 * then we just ignore the text.
12660 	 */
12661 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
12662 		   IS_FASTOPEN(tp->t_flags));
12663 	if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) &&
12664 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12665 		tcp_seq save_start = th->th_seq;
12666 		tcp_seq save_rnxt  = tp->rcv_nxt;
12667 		int     save_tlen  = tlen;
12668 
12669 		m_adj(m, drop_hdrlen);	/* delayed header drop */
12670 		/*
12671 		 * Insert segment which includes th into TCP reassembly
12672 		 * queue with control block tp.  Set thflags to whether
12673 		 * reassembly now includes a segment with FIN.  This handles
12674 		 * the common case inline (segment is the next to be
12675 		 * received on an established connection, and the queue is
12676 		 * empty), avoiding linkage into and removal from the queue
12677 		 * and repetition of various conversions. Set DELACK for
12678 		 * segments received in order, but ack immediately when
12679 		 * segments are out of order (so fast retransmit can work).
12680 		 */
12681 		if (th->th_seq == tp->rcv_nxt &&
12682 		    SEGQ_EMPTY(tp) &&
12683 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
12684 		    tfo_syn)) {
12685 #ifdef NETFLIX_SB_LIMITS
12686 			u_int mcnt, appended;
12687 
12688 			if (so->so_rcv.sb_shlim) {
12689 				mcnt = m_memcnt(m);
12690 				appended = 0;
12691 				if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12692 				    CFO_NOSLEEP, NULL) == false) {
12693 					counter_u64_add(tcp_sb_shlim_fails, 1);
12694 					m_freem(m);
12695 					return (0);
12696 				}
12697 			}
12698 #endif
12699 			rack_handle_delayed_ack(tp, rack, tlen, tfo_syn);
12700 			tp->rcv_nxt += tlen;
12701 			if (tlen &&
12702 			    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12703 			    (tp->t_fbyte_in == 0)) {
12704 				tp->t_fbyte_in = ticks;
12705 				if (tp->t_fbyte_in == 0)
12706 					tp->t_fbyte_in = 1;
12707 				if (tp->t_fbyte_out && tp->t_fbyte_in)
12708 					tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12709 			}
12710 			thflags = tcp_get_flags(th) & TH_FIN;
12711 			KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12712 			KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12713 			SOCKBUF_LOCK(&so->so_rcv);
12714 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12715 				m_freem(m);
12716 			} else
12717 #ifdef NETFLIX_SB_LIMITS
12718 				appended =
12719 #endif
12720 					sbappendstream_locked(&so->so_rcv, m, 0);
12721 
12722 			rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12723 			/* NB: sorwakeup_locked() does an implicit unlock. */
12724 			sorwakeup_locked(so);
12725 #ifdef NETFLIX_SB_LIMITS
12726 			if (so->so_rcv.sb_shlim && appended != mcnt)
12727 				counter_fo_release(so->so_rcv.sb_shlim,
12728 				    mcnt - appended);
12729 #endif
12730 		} else {
12731 			/*
12732 			 * XXX: Due to the header drop above "th" is
12733 			 * theoretically invalid by now.  Fortunately
12734 			 * m_adj() doesn't actually frees any mbufs when
12735 			 * trimming from the head.
12736 			 */
12737 			tcp_seq temp = save_start;
12738 
12739 			thflags = tcp_reass(tp, th, &temp, &tlen, m);
12740 			tp->t_flags |= TF_ACKNOW;
12741 			if (tp->t_flags & TF_WAKESOR) {
12742 				tp->t_flags &= ~TF_WAKESOR;
12743 				/* NB: sorwakeup_locked() does an implicit unlock. */
12744 				sorwakeup_locked(so);
12745 			}
12746 		}
12747 		if ((tp->t_flags & TF_SACK_PERMIT) &&
12748 		    (save_tlen > 0) &&
12749 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
12750 			if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) {
12751 				/*
12752 				 * DSACK actually handled in the fastpath
12753 				 * above.
12754 				 */
12755 				tcp_update_sack_list(tp, save_start,
12756 				    save_start + save_tlen);
12757 			} else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) {
12758 				if ((tp->rcv_numsacks >= 1) &&
12759 				    (tp->sackblks[0].end == save_start)) {
12760 					/*
12761 					 * Partial overlap, recorded at todrop
12762 					 * above.
12763 					 */
12764 					tcp_update_sack_list(tp,
12765 					    tp->sackblks[0].start,
12766 					    tp->sackblks[0].end);
12767 				} else {
12768 					tcp_update_dsack_list(tp, save_start,
12769 					    save_start + save_tlen);
12770 				}
12771 			} else if (tlen >= save_tlen) {
12772 				/* Update of sackblks. */
12773 				tcp_update_dsack_list(tp, save_start,
12774 				    save_start + save_tlen);
12775 			} else if (tlen > 0) {
12776 				tcp_update_dsack_list(tp, save_start,
12777 				    save_start + tlen);
12778 			}
12779 		}
12780 	} else {
12781 		m_freem(m);
12782 		thflags &= ~TH_FIN;
12783 	}
12784 
12785 	/*
12786 	 * If FIN is received ACK the FIN and let the user know that the
12787 	 * connection is closing.
12788 	 */
12789 	if (thflags & TH_FIN) {
12790 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
12791 			/* The socket upcall is handled by socantrcvmore. */
12792 			socantrcvmore(so);
12793 			/*
12794 			 * If connection is half-synchronized (ie NEEDSYN
12795 			 * flag on) then delay ACK, so it may be piggybacked
12796 			 * when SYN is sent. Otherwise, since we received a
12797 			 * FIN then no more input can be expected, send ACK
12798 			 * now.
12799 			 */
12800 			if (tp->t_flags & TF_NEEDSYN) {
12801 				rack_timer_cancel(tp, rack,
12802 				    rack->r_ctl.rc_rcvtime, __LINE__);
12803 				tp->t_flags |= TF_DELACK;
12804 			} else {
12805 				tp->t_flags |= TF_ACKNOW;
12806 			}
12807 			tp->rcv_nxt++;
12808 		}
12809 		switch (tp->t_state) {
12810 			/*
12811 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
12812 			 * CLOSE_WAIT state.
12813 			 */
12814 		case TCPS_SYN_RECEIVED:
12815 			tp->t_starttime = ticks;
12816 			/* FALLTHROUGH */
12817 		case TCPS_ESTABLISHED:
12818 			rack_timer_cancel(tp, rack,
12819 			    rack->r_ctl.rc_rcvtime, __LINE__);
12820 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
12821 			break;
12822 
12823 			/*
12824 			 * If still in FIN_WAIT_1 STATE FIN has not been
12825 			 * acked so enter the CLOSING state.
12826 			 */
12827 		case TCPS_FIN_WAIT_1:
12828 			rack_timer_cancel(tp, rack,
12829 			    rack->r_ctl.rc_rcvtime, __LINE__);
12830 			tcp_state_change(tp, TCPS_CLOSING);
12831 			break;
12832 
12833 			/*
12834 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
12835 			 * starting the time-wait timer, turning off the
12836 			 * other standard timers.
12837 			 */
12838 		case TCPS_FIN_WAIT_2:
12839 			rack_timer_cancel(tp, rack,
12840 			    rack->r_ctl.rc_rcvtime, __LINE__);
12841 			tcp_twstart(tp);
12842 			return (1);
12843 		}
12844 	}
12845 	/*
12846 	 * Return any desired output.
12847 	 */
12848 	if ((tp->t_flags & TF_ACKNOW) ||
12849 	    (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
12850 		rack->r_wanted_output = 1;
12851 	}
12852 	return (0);
12853 }
12854 
12855 /*
12856  * Here nothing is really faster, its just that we
12857  * have broken out the fast-data path also just like
12858  * the fast-ack.
12859  */
12860 static int
12861 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
12862     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12863     uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos)
12864 {
12865 	int32_t nsegs;
12866 	int32_t newsize = 0;	/* automatic sockbuf scaling */
12867 	struct tcp_rack *rack;
12868 #ifdef NETFLIX_SB_LIMITS
12869 	u_int mcnt, appended;
12870 #endif
12871 
12872 	/*
12873 	 * If last ACK falls within this segment's sequence numbers, record
12874 	 * the timestamp. NOTE that the test is modified according to the
12875 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
12876 	 */
12877 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
12878 		return (0);
12879 	}
12880 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
12881 		return (0);
12882 	}
12883 	if (tiwin && tiwin != tp->snd_wnd) {
12884 		return (0);
12885 	}
12886 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
12887 		return (0);
12888 	}
12889 	if (__predict_false((to->to_flags & TOF_TS) &&
12890 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
12891 		return (0);
12892 	}
12893 	if (__predict_false((th->th_ack != tp->snd_una))) {
12894 		return (0);
12895 	}
12896 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
12897 		return (0);
12898 	}
12899 	if ((to->to_flags & TOF_TS) != 0 &&
12900 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
12901 		tp->ts_recent_age = tcp_ts_getticks();
12902 		tp->ts_recent = to->to_tsval;
12903 	}
12904 	rack = (struct tcp_rack *)tp->t_fb_ptr;
12905 	/*
12906 	 * This is a pure, in-sequence data packet with nothing on the
12907 	 * reassembly queue and we have enough buffer space to take it.
12908 	 */
12909 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
12910 
12911 #ifdef NETFLIX_SB_LIMITS
12912 	if (so->so_rcv.sb_shlim) {
12913 		mcnt = m_memcnt(m);
12914 		appended = 0;
12915 		if (counter_fo_get(so->so_rcv.sb_shlim, mcnt,
12916 		    CFO_NOSLEEP, NULL) == false) {
12917 			counter_u64_add(tcp_sb_shlim_fails, 1);
12918 			m_freem(m);
12919 			return (1);
12920 		}
12921 	}
12922 #endif
12923 	/* Clean receiver SACK report if present */
12924 	if (tp->rcv_numsacks)
12925 		tcp_clean_sackreport(tp);
12926 	KMOD_TCPSTAT_INC(tcps_preddat);
12927 	tp->rcv_nxt += tlen;
12928 	if (tlen &&
12929 	    ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) &&
12930 	    (tp->t_fbyte_in == 0)) {
12931 		tp->t_fbyte_in = ticks;
12932 		if (tp->t_fbyte_in == 0)
12933 			tp->t_fbyte_in = 1;
12934 		if (tp->t_fbyte_out && tp->t_fbyte_in)
12935 			tp->t_flags2 |= TF2_FBYTES_COMPLETE;
12936 	}
12937 	/*
12938 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
12939 	 */
12940 	tp->snd_wl1 = th->th_seq;
12941 	/*
12942 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
12943 	 */
12944 	tp->rcv_up = tp->rcv_nxt;
12945 	KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs);
12946 	KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen);
12947 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
12948 
12949 	/* Add data to socket buffer. */
12950 	SOCKBUF_LOCK(&so->so_rcv);
12951 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
12952 		m_freem(m);
12953 	} else {
12954 		/*
12955 		 * Set new socket buffer size. Give up when limit is
12956 		 * reached.
12957 		 */
12958 		if (newsize)
12959 			if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
12960 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
12961 		m_adj(m, drop_hdrlen);	/* delayed header drop */
12962 #ifdef NETFLIX_SB_LIMITS
12963 		appended =
12964 #endif
12965 			sbappendstream_locked(&so->so_rcv, m, 0);
12966 		ctf_calc_rwin(so, tp);
12967 	}
12968 	rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1);
12969 	/* NB: sorwakeup_locked() does an implicit unlock. */
12970 	sorwakeup_locked(so);
12971 #ifdef NETFLIX_SB_LIMITS
12972 	if (so->so_rcv.sb_shlim && mcnt != appended)
12973 		counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended);
12974 #endif
12975 	rack_handle_delayed_ack(tp, rack, tlen, 0);
12976 	if (tp->snd_una == tp->snd_max)
12977 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
12978 	return (1);
12979 }
12980 
12981 /*
12982  * This subfunction is used to try to highly optimize the
12983  * fast path. We again allow window updates that are
12984  * in sequence to remain in the fast-path. We also add
12985  * in the __predict's to attempt to help the compiler.
12986  * Note that if we return a 0, then we can *not* process
12987  * it and the caller should push the packet into the
12988  * slow-path.
12989  */
12990 static int
12991 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
12992     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
12993     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
12994 {
12995 	int32_t acked;
12996 	int32_t nsegs;
12997 	int32_t under_pacing = 0;
12998 	struct tcp_rack *rack;
12999 
13000 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
13001 		/* Old ack, behind (or duplicate to) the last one rcv'd */
13002 		return (0);
13003 	}
13004 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
13005 		/* Above what we have sent? */
13006 		return (0);
13007 	}
13008 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
13009 		/* We are retransmitting */
13010 		return (0);
13011 	}
13012 	if (__predict_false(tiwin == 0)) {
13013 		/* zero window */
13014 		return (0);
13015 	}
13016 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
13017 		/* We need a SYN or a FIN, unlikely.. */
13018 		return (0);
13019 	}
13020 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
13021 		/* Timestamp is behind .. old ack with seq wrap? */
13022 		return (0);
13023 	}
13024 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
13025 		/* Still recovering */
13026 		return (0);
13027 	}
13028 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13029 	if (rack->r_ctl.rc_sacked) {
13030 		/* We have sack holes on our scoreboard */
13031 		return (0);
13032 	}
13033 	/* Ok if we reach here, we can process a fast-ack */
13034 	if (rack->gp_ready &&
13035 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
13036 		under_pacing = 1;
13037 	}
13038 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
13039 	rack_log_ack(tp, to, th, 0, 0, NULL, NULL);
13040 	/* Did the window get updated? */
13041 	if (tiwin != tp->snd_wnd) {
13042 		tp->snd_wnd = tiwin;
13043 		rack_validate_fo_sendwin_up(tp, rack);
13044 		tp->snd_wl1 = th->th_seq;
13045 		if (tp->snd_wnd > tp->max_sndwnd)
13046 			tp->max_sndwnd = tp->snd_wnd;
13047 	}
13048 	/* Do we exit persists? */
13049 	if ((rack->rc_in_persist != 0) &&
13050 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
13051 			       rack->r_ctl.rc_pace_min_segs))) {
13052 		rack_exit_persist(tp, rack, cts);
13053 	}
13054 	/* Do we enter persists? */
13055 	if ((rack->rc_in_persist == 0) &&
13056 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
13057 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
13058 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
13059 	    sbavail(&tptosocket(tp)->so_snd) &&
13060 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
13061 		/*
13062 		 * Here the rwnd is less than
13063 		 * the pacing size, we are established,
13064 		 * nothing is outstanding, and there is
13065 		 * data to send. Enter persists.
13066 		 */
13067 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, th->th_ack);
13068 	}
13069 	/*
13070 	 * If last ACK falls within this segment's sequence numbers, record
13071 	 * the timestamp. NOTE that the test is modified according to the
13072 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
13073 	 */
13074 	if ((to->to_flags & TOF_TS) != 0 &&
13075 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
13076 		tp->ts_recent_age = tcp_ts_getticks();
13077 		tp->ts_recent = to->to_tsval;
13078 	}
13079 	/*
13080 	 * This is a pure ack for outstanding data.
13081 	 */
13082 	KMOD_TCPSTAT_INC(tcps_predack);
13083 
13084 	/*
13085 	 * "bad retransmit" recovery.
13086 	 */
13087 	if ((tp->t_flags & TF_PREVVALID) &&
13088 	    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
13089 		tp->t_flags &= ~TF_PREVVALID;
13090 		if (tp->t_rxtshift == 1 &&
13091 		    (int)(ticks - tp->t_badrxtwin) < 0)
13092 			rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__);
13093 	}
13094 	/*
13095 	 * Recalculate the transmit timer / rtt.
13096 	 *
13097 	 * Some boxes send broken timestamp replies during the SYN+ACK
13098 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
13099 	 * and blow up the retransmit timer.
13100 	 */
13101 	acked = BYTES_THIS_ACK(tp, th);
13102 
13103 #ifdef TCP_HHOOK
13104 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
13105 	hhook_run_tcp_est_in(tp, th, to);
13106 #endif
13107 	KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs);
13108 	KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
13109 	if (acked) {
13110 		struct mbuf *mfree;
13111 
13112 		rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0);
13113 		SOCKBUF_LOCK(&so->so_snd);
13114 		mfree = sbcut_locked(&so->so_snd, acked);
13115 		tp->snd_una = th->th_ack;
13116 		/* Note we want to hold the sb lock through the sendmap adjust */
13117 		rack_adjust_sendmap_head(rack, &so->so_snd);
13118 		/* Wake up the socket if we have room to write more */
13119 		rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
13120 		sowwakeup_locked(so);
13121 		m_freem(mfree);
13122 		tp->t_rxtshift = 0;
13123 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
13124 			      rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
13125 		rack->rc_tlp_in_progress = 0;
13126 		rack->r_ctl.rc_tlp_cnt_out = 0;
13127 		/*
13128 		 * If it is the RXT timer we want to
13129 		 * stop it, so we can restart a TLP.
13130 		 */
13131 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
13132 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13133 
13134 #ifdef TCP_REQUEST_TRK
13135 		rack_req_check_for_comp(rack, th->th_ack);
13136 #endif
13137 	}
13138 	/*
13139 	 * Let the congestion control algorithm update congestion control
13140 	 * related information. This typically means increasing the
13141 	 * congestion window.
13142 	 */
13143 	if (tp->snd_wnd < ctf_outstanding(tp)) {
13144 		/* The peer collapsed the window */
13145 		rack_collapsed_window(rack, ctf_outstanding(tp), th->th_ack, __LINE__);
13146 	} else if (rack->rc_has_collapsed)
13147 		rack_un_collapse_window(rack, __LINE__);
13148 	if ((rack->r_collapse_point_valid) &&
13149 	    (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point)))
13150 		rack->r_collapse_point_valid = 0;
13151 	/*
13152 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
13153 	 */
13154 	tp->snd_wl2 = th->th_ack;
13155 	tp->t_dupacks = 0;
13156 	m_freem(m);
13157 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
13158 
13159 	/*
13160 	 * If all outstanding data are acked, stop retransmit timer,
13161 	 * otherwise restart timer using current (possibly backed-off)
13162 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
13163 	 * If data are ready to send, let tcp_output decide between more
13164 	 * output or persist.
13165 	 */
13166 	if (under_pacing &&
13167 	    (rack->use_fixed_rate == 0) &&
13168 	    (rack->in_probe_rtt == 0) &&
13169 	    rack->rc_gp_dyn_mul &&
13170 	    rack->rc_always_pace) {
13171 		/* Check if we are dragging bottom */
13172 		rack_check_bottom_drag(tp, rack, so);
13173 	}
13174 	if (tp->snd_una == tp->snd_max) {
13175 		tp->t_flags &= ~TF_PREVVALID;
13176 		rack->r_ctl.retran_during_recovery = 0;
13177 		rack->rc_suspicious = 0;
13178 		rack->r_ctl.dsack_byte_cnt = 0;
13179 		rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
13180 		if (rack->r_ctl.rc_went_idle_time == 0)
13181 			rack->r_ctl.rc_went_idle_time = 1;
13182 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
13183 		if (sbavail(&tptosocket(tp)->so_snd) == 0)
13184 			tp->t_acktime = 0;
13185 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
13186 	}
13187 	if (acked && rack->r_fast_output)
13188 		rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked);
13189 	if (sbavail(&so->so_snd)) {
13190 		rack->r_wanted_output = 1;
13191 	}
13192 	return (1);
13193 }
13194 
13195 /*
13196  * Return value of 1, the TCB is unlocked and most
13197  * likely gone, return value of 0, the TCP is still
13198  * locked.
13199  */
13200 static int
13201 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
13202     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13203     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13204 {
13205 	int32_t ret_val = 0;
13206 	int32_t todrop;
13207 	int32_t ourfinisacked = 0;
13208 	struct tcp_rack *rack;
13209 
13210 	INP_WLOCK_ASSERT(tptoinpcb(tp));
13211 
13212 	ctf_calc_rwin(so, tp);
13213 	/*
13214 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
13215 	 * SYN, drop the input. if seg contains a RST, then drop the
13216 	 * connection. if seg does not contain SYN, then drop it. Otherwise
13217 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
13218 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
13219 	 * contains an ECE and ECN support is enabled, the stream is ECN
13220 	 * capable. if SYN has been acked change to ESTABLISHED else
13221 	 * SYN_RCVD state arrange for segment to be acked (eventually)
13222 	 * continue processing rest of data/controls.
13223 	 */
13224 	if ((thflags & TH_ACK) &&
13225 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
13226 	    SEQ_GT(th->th_ack, tp->snd_max))) {
13227 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13228 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13229 		return (1);
13230 	}
13231 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
13232 		TCP_PROBE5(connect__refused, NULL, tp,
13233 		    mtod(m, const char *), tp, th);
13234 		tp = tcp_drop(tp, ECONNREFUSED);
13235 		ctf_do_drop(m, tp);
13236 		return (1);
13237 	}
13238 	if (thflags & TH_RST) {
13239 		ctf_do_drop(m, tp);
13240 		return (1);
13241 	}
13242 	if (!(thflags & TH_SYN)) {
13243 		ctf_do_drop(m, tp);
13244 		return (1);
13245 	}
13246 	tp->irs = th->th_seq;
13247 	tcp_rcvseqinit(tp);
13248 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13249 	if (thflags & TH_ACK) {
13250 		int tfo_partial = 0;
13251 
13252 		KMOD_TCPSTAT_INC(tcps_connects);
13253 		soisconnected(so);
13254 #ifdef MAC
13255 		mac_socketpeer_set_from_mbuf(m, so);
13256 #endif
13257 		/* Do window scaling on this connection? */
13258 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13259 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13260 			tp->rcv_scale = tp->request_r_scale;
13261 		}
13262 		tp->rcv_adv += min(tp->rcv_wnd,
13263 		    TCP_MAXWIN << tp->rcv_scale);
13264 		/*
13265 		 * If not all the data that was sent in the TFO SYN
13266 		 * has been acked, resend the remainder right away.
13267 		 */
13268 		if (IS_FASTOPEN(tp->t_flags) &&
13269 		    (tp->snd_una != tp->snd_max)) {
13270 			tp->snd_nxt = th->th_ack;
13271 			tfo_partial = 1;
13272 		}
13273 		/*
13274 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
13275 		 * will be turned on later.
13276 		 */
13277 		if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) {
13278 			rack_timer_cancel(tp, rack,
13279 					  rack->r_ctl.rc_rcvtime, __LINE__);
13280 			tp->t_flags |= TF_DELACK;
13281 		} else {
13282 			rack->r_wanted_output = 1;
13283 			tp->t_flags |= TF_ACKNOW;
13284 		}
13285 
13286 		tcp_ecn_input_syn_sent(tp, thflags, iptos);
13287 
13288 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
13289 			/*
13290 			 * We advance snd_una for the
13291 			 * fast open case. If th_ack is
13292 			 * acknowledging data beyond
13293 			 * snd_una we can't just call
13294 			 * ack-processing since the
13295 			 * data stream in our send-map
13296 			 * will start at snd_una + 1 (one
13297 			 * beyond the SYN). If its just
13298 			 * equal we don't need to do that
13299 			 * and there is no send_map.
13300 			 */
13301 			tp->snd_una++;
13302 		}
13303 		/*
13304 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
13305 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
13306 		 */
13307 		tp->t_starttime = ticks;
13308 		if (tp->t_flags & TF_NEEDFIN) {
13309 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
13310 			tp->t_flags &= ~TF_NEEDFIN;
13311 			thflags &= ~TH_SYN;
13312 		} else {
13313 			tcp_state_change(tp, TCPS_ESTABLISHED);
13314 			TCP_PROBE5(connect__established, NULL, tp,
13315 			    mtod(m, const char *), tp, th);
13316 			rack_cc_conn_init(tp);
13317 		}
13318 	} else {
13319 		/*
13320 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
13321 		 * open.  If segment contains CC option and there is a
13322 		 * cached CC, apply TAO test. If it succeeds, connection is *
13323 		 * half-synchronized. Otherwise, do 3-way handshake:
13324 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
13325 		 * there was no CC option, clear cached CC value.
13326 		 */
13327 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN);
13328 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
13329 	}
13330 	/*
13331 	 * Advance th->th_seq to correspond to first data byte. If data,
13332 	 * trim to stay within window, dropping FIN if necessary.
13333 	 */
13334 	th->th_seq++;
13335 	if (tlen > tp->rcv_wnd) {
13336 		todrop = tlen - tp->rcv_wnd;
13337 		m_adj(m, -todrop);
13338 		tlen = tp->rcv_wnd;
13339 		thflags &= ~TH_FIN;
13340 		KMOD_TCPSTAT_INC(tcps_rcvpackafterwin);
13341 		KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
13342 	}
13343 	tp->snd_wl1 = th->th_seq - 1;
13344 	tp->rcv_up = th->th_seq;
13345 	/*
13346 	 * Client side of transaction: already sent SYN and data. If the
13347 	 * remote host used T/TCP to validate the SYN, our data will be
13348 	 * ACK'd; if so, enter normal data segment processing in the middle
13349 	 * of step 5, ack processing. Otherwise, goto step 6.
13350 	 */
13351 	if (thflags & TH_ACK) {
13352 		/* For syn-sent we need to possibly update the rtt */
13353 		if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13354 			uint32_t t, mcts;
13355 
13356 			mcts = tcp_ts_getticks();
13357 			t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13358 			if (!tp->t_rttlow || tp->t_rttlow > t)
13359 				tp->t_rttlow = t;
13360 			rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4);
13361 			tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13362 			tcp_rack_xmit_timer_commit(rack, tp);
13363 		}
13364 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
13365 			return (ret_val);
13366 		/* We may have changed to FIN_WAIT_1 above */
13367 		if (tp->t_state == TCPS_FIN_WAIT_1) {
13368 			/*
13369 			 * In FIN_WAIT_1 STATE in addition to the processing
13370 			 * for the ESTABLISHED state if our FIN is now
13371 			 * acknowledged then enter FIN_WAIT_2.
13372 			 */
13373 			if (ourfinisacked) {
13374 				/*
13375 				 * If we can't receive any more data, then
13376 				 * closing user can proceed. Starting the
13377 				 * timer is contrary to the specification,
13378 				 * but if we don't get a FIN we'll hang
13379 				 * forever.
13380 				 *
13381 				 * XXXjl: we should release the tp also, and
13382 				 * use a compressed state.
13383 				 */
13384 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13385 					soisdisconnected(so);
13386 					tcp_timer_activate(tp, TT_2MSL,
13387 					    (tcp_fast_finwait2_recycle ?
13388 					    tcp_finwait2_timeout :
13389 					    TP_MAXIDLE(tp)));
13390 				}
13391 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
13392 			}
13393 		}
13394 	}
13395 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13396 	   tiwin, thflags, nxt_pkt));
13397 }
13398 
13399 /*
13400  * Return value of 1, the TCB is unlocked and most
13401  * likely gone, return value of 0, the TCP is still
13402  * locked.
13403  */
13404 static int
13405 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
13406     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13407     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13408 {
13409 	struct tcp_rack *rack;
13410 	int32_t ret_val = 0;
13411 	int32_t ourfinisacked = 0;
13412 
13413 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13414 	ctf_calc_rwin(so, tp);
13415 	if ((thflags & TH_RST) ||
13416 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13417 		return (__ctf_process_rst(m, th, so, tp,
13418 					  &rack->r_ctl.challenge_ack_ts,
13419 					  &rack->r_ctl.challenge_ack_cnt));
13420 	if ((thflags & TH_ACK) &&
13421 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
13422 	    SEQ_GT(th->th_ack, tp->snd_max))) {
13423 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13424 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13425 		return (1);
13426 	}
13427 	if (IS_FASTOPEN(tp->t_flags)) {
13428 		/*
13429 		 * When a TFO connection is in SYN_RECEIVED, the
13430 		 * only valid packets are the initial SYN, a
13431 		 * retransmit/copy of the initial SYN (possibly with
13432 		 * a subset of the original data), a valid ACK, a
13433 		 * FIN, or a RST.
13434 		 */
13435 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
13436 			tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13437 			ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13438 			return (1);
13439 		} else if (thflags & TH_SYN) {
13440 			/* non-initial SYN is ignored */
13441 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
13442 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
13443 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
13444 				ctf_do_drop(m, NULL);
13445 				return (0);
13446 			}
13447 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
13448 			ctf_do_drop(m, NULL);
13449 			return (0);
13450 		}
13451 	}
13452 
13453 	/*
13454 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13455 	 * it's less than ts_recent, drop it.
13456 	 */
13457 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13458 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13459 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13460 			return (ret_val);
13461 	}
13462 	/*
13463 	 * In the SYN-RECEIVED state, validate that the packet belongs to
13464 	 * this connection before trimming the data to fit the receive
13465 	 * window.  Check the sequence number versus IRS since we know the
13466 	 * sequence numbers haven't wrapped.  This is a partial fix for the
13467 	 * "LAND" DoS attack.
13468 	 */
13469 	if (SEQ_LT(th->th_seq, tp->irs)) {
13470 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
13471 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13472 		return (1);
13473 	}
13474 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13475 			      &rack->r_ctl.challenge_ack_ts,
13476 			      &rack->r_ctl.challenge_ack_cnt)) {
13477 		return (ret_val);
13478 	}
13479 	/*
13480 	 * If last ACK falls within this segment's sequence numbers, record
13481 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13482 	 * from the latest proposal of the tcplw@cray.com list (Braden
13483 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13484 	 * with our earlier PAWS tests, so this check should be solely
13485 	 * predicated on the sequence space of this segment. 3) That we
13486 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13487 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13488 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13489 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13490 	 * p.869. In such cases, we can still calculate the RTT correctly
13491 	 * when RCV.NXT == Last.ACK.Sent.
13492 	 */
13493 	if ((to->to_flags & TOF_TS) != 0 &&
13494 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13495 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13496 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13497 		tp->ts_recent_age = tcp_ts_getticks();
13498 		tp->ts_recent = to->to_tsval;
13499 	}
13500 	tp->snd_wnd = tiwin;
13501 	rack_validate_fo_sendwin_up(tp, rack);
13502 	/*
13503 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13504 	 * is on (half-synchronized state), then queue data for later
13505 	 * processing; else drop segment and return.
13506 	 */
13507 	if ((thflags & TH_ACK) == 0) {
13508 		if (IS_FASTOPEN(tp->t_flags)) {
13509 			rack_cc_conn_init(tp);
13510 		}
13511 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13512 		    tiwin, thflags, nxt_pkt));
13513 	}
13514 	KMOD_TCPSTAT_INC(tcps_connects);
13515 	if (tp->t_flags & TF_SONOTCONN) {
13516 		tp->t_flags &= ~TF_SONOTCONN;
13517 		soisconnected(so);
13518 	}
13519 	/* Do window scaling? */
13520 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
13521 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
13522 		tp->rcv_scale = tp->request_r_scale;
13523 	}
13524 	/*
13525 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
13526 	 * FIN-WAIT-1
13527 	 */
13528 	tp->t_starttime = ticks;
13529 	if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
13530 		tcp_fastopen_decrement_counter(tp->t_tfo_pending);
13531 		tp->t_tfo_pending = NULL;
13532 	}
13533 	if (tp->t_flags & TF_NEEDFIN) {
13534 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
13535 		tp->t_flags &= ~TF_NEEDFIN;
13536 	} else {
13537 		tcp_state_change(tp, TCPS_ESTABLISHED);
13538 		TCP_PROBE5(accept__established, NULL, tp,
13539 		    mtod(m, const char *), tp, th);
13540 		/*
13541 		 * TFO connections call cc_conn_init() during SYN
13542 		 * processing.  Calling it again here for such connections
13543 		 * is not harmless as it would undo the snd_cwnd reduction
13544 		 * that occurs when a TFO SYN|ACK is retransmitted.
13545 		 */
13546 		if (!IS_FASTOPEN(tp->t_flags))
13547 			rack_cc_conn_init(tp);
13548 	}
13549 	/*
13550 	 * Account for the ACK of our SYN prior to
13551 	 * regular ACK processing below, except for
13552 	 * simultaneous SYN, which is handled later.
13553 	 */
13554 	if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN))
13555 		tp->snd_una++;
13556 	/*
13557 	 * If segment contains data or ACK, will call tcp_reass() later; if
13558 	 * not, do so now to pass queued data to user.
13559 	 */
13560 	if (tlen == 0 && (thflags & TH_FIN) == 0) {
13561 		(void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0,
13562 		    (struct mbuf *)0);
13563 		if (tp->t_flags & TF_WAKESOR) {
13564 			tp->t_flags &= ~TF_WAKESOR;
13565 			/* NB: sorwakeup_locked() does an implicit unlock. */
13566 			sorwakeup_locked(so);
13567 		}
13568 	}
13569 	tp->snd_wl1 = th->th_seq - 1;
13570 	/* For syn-recv we need to possibly update the rtt */
13571 	if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) {
13572 		uint32_t t, mcts;
13573 
13574 		mcts = tcp_ts_getticks();
13575 		t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC;
13576 		if (!tp->t_rttlow || tp->t_rttlow > t)
13577 			tp->t_rttlow = t;
13578 		rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5);
13579 		tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2);
13580 		tcp_rack_xmit_timer_commit(rack, tp);
13581 	}
13582 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
13583 		return (ret_val);
13584 	}
13585 	if (tp->t_state == TCPS_FIN_WAIT_1) {
13586 		/* We could have went to FIN_WAIT_1 (or EST) above */
13587 		/*
13588 		 * In FIN_WAIT_1 STATE in addition to the processing for the
13589 		 * ESTABLISHED state if our FIN is now acknowledged then
13590 		 * enter FIN_WAIT_2.
13591 		 */
13592 		if (ourfinisacked) {
13593 			/*
13594 			 * If we can't receive any more data, then closing
13595 			 * user can proceed. Starting the timer is contrary
13596 			 * to the specification, but if we don't get a FIN
13597 			 * we'll hang forever.
13598 			 *
13599 			 * XXXjl: we should release the tp also, and use a
13600 			 * compressed state.
13601 			 */
13602 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13603 				soisdisconnected(so);
13604 				tcp_timer_activate(tp, TT_2MSL,
13605 				    (tcp_fast_finwait2_recycle ?
13606 				    tcp_finwait2_timeout :
13607 				    TP_MAXIDLE(tp)));
13608 			}
13609 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
13610 		}
13611 	}
13612 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13613 	    tiwin, thflags, nxt_pkt));
13614 }
13615 
13616 /*
13617  * Return value of 1, the TCB is unlocked and most
13618  * likely gone, return value of 0, the TCP is still
13619  * locked.
13620  */
13621 static int
13622 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
13623     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13624     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13625 {
13626 	int32_t ret_val = 0;
13627 	struct tcp_rack *rack;
13628 
13629 	/*
13630 	 * Header prediction: check for the two common cases of a
13631 	 * uni-directional data xfer.  If the packet has no control flags,
13632 	 * is in-sequence, the window didn't change and we're not
13633 	 * retransmitting, it's a candidate.  If the length is zero and the
13634 	 * ack moved forward, we're the sender side of the xfer.  Just free
13635 	 * the data acked & wake any higher level process that was blocked
13636 	 * waiting for space.  If the length is non-zero and the ack didn't
13637 	 * move, we're the receiver side.  If we're getting packets in-order
13638 	 * (the reassembly queue is empty), add the data toc The socket
13639 	 * buffer and note that we need a delayed ack. Make sure that the
13640 	 * hidden state-flags are also off. Since we check for
13641 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
13642 	 */
13643 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13644 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
13645 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) &&
13646 	    __predict_true(SEGQ_EMPTY(tp)) &&
13647 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
13648 		if (tlen == 0) {
13649 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
13650 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
13651 				return (0);
13652 			}
13653 		} else {
13654 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
13655 			    tiwin, nxt_pkt, iptos)) {
13656 				return (0);
13657 			}
13658 		}
13659 	}
13660 	ctf_calc_rwin(so, tp);
13661 
13662 	if ((thflags & TH_RST) ||
13663 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13664 		return (__ctf_process_rst(m, th, so, tp,
13665 					  &rack->r_ctl.challenge_ack_ts,
13666 					  &rack->r_ctl.challenge_ack_cnt));
13667 
13668 	/*
13669 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13670 	 * synchronized state.
13671 	 */
13672 	if (thflags & TH_SYN) {
13673 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13674 		return (ret_val);
13675 	}
13676 	/*
13677 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13678 	 * it's less than ts_recent, drop it.
13679 	 */
13680 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13681 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13682 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13683 			return (ret_val);
13684 	}
13685 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13686 			      &rack->r_ctl.challenge_ack_ts,
13687 			      &rack->r_ctl.challenge_ack_cnt)) {
13688 		return (ret_val);
13689 	}
13690 	/*
13691 	 * If last ACK falls within this segment's sequence numbers, record
13692 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13693 	 * from the latest proposal of the tcplw@cray.com list (Braden
13694 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13695 	 * with our earlier PAWS tests, so this check should be solely
13696 	 * predicated on the sequence space of this segment. 3) That we
13697 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13698 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13699 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13700 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13701 	 * p.869. In such cases, we can still calculate the RTT correctly
13702 	 * when RCV.NXT == Last.ACK.Sent.
13703 	 */
13704 	if ((to->to_flags & TOF_TS) != 0 &&
13705 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13706 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13707 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13708 		tp->ts_recent_age = tcp_ts_getticks();
13709 		tp->ts_recent = to->to_tsval;
13710 	}
13711 	/*
13712 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13713 	 * is on (half-synchronized state), then queue data for later
13714 	 * processing; else drop segment and return.
13715 	 */
13716 	if ((thflags & TH_ACK) == 0) {
13717 		if (tp->t_flags & TF_NEEDSYN) {
13718 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13719 			    tiwin, thflags, nxt_pkt));
13720 
13721 		} else if (tp->t_flags & TF_ACKNOW) {
13722 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13723 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13724 			return (ret_val);
13725 		} else {
13726 			ctf_do_drop(m, NULL);
13727 			return (0);
13728 		}
13729 	}
13730 	/*
13731 	 * Ack processing.
13732 	 */
13733 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
13734 		return (ret_val);
13735 	}
13736 	if (sbavail(&so->so_snd)) {
13737 		if (ctf_progress_timeout_check(tp, true)) {
13738 			rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__);
13739 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13740 			return (1);
13741 		}
13742 	}
13743 	/* State changes only happen in rack_process_data() */
13744 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13745 	    tiwin, thflags, nxt_pkt));
13746 }
13747 
13748 /*
13749  * Return value of 1, the TCB is unlocked and most
13750  * likely gone, return value of 0, the TCP is still
13751  * locked.
13752  */
13753 static int
13754 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
13755     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13756     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13757 {
13758 	int32_t ret_val = 0;
13759 	struct tcp_rack *rack;
13760 
13761 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13762 	ctf_calc_rwin(so, tp);
13763 	if ((thflags & TH_RST) ||
13764 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13765 		return (__ctf_process_rst(m, th, so, tp,
13766 					  &rack->r_ctl.challenge_ack_ts,
13767 					  &rack->r_ctl.challenge_ack_cnt));
13768 	/*
13769 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13770 	 * synchronized state.
13771 	 */
13772 	if (thflags & TH_SYN) {
13773 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13774 		return (ret_val);
13775 	}
13776 	/*
13777 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13778 	 * it's less than ts_recent, drop it.
13779 	 */
13780 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13781 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13782 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13783 			return (ret_val);
13784 	}
13785 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13786 			      &rack->r_ctl.challenge_ack_ts,
13787 			      &rack->r_ctl.challenge_ack_cnt)) {
13788 		return (ret_val);
13789 	}
13790 	/*
13791 	 * If last ACK falls within this segment's sequence numbers, record
13792 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13793 	 * from the latest proposal of the tcplw@cray.com list (Braden
13794 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13795 	 * with our earlier PAWS tests, so this check should be solely
13796 	 * predicated on the sequence space of this segment. 3) That we
13797 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13798 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13799 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13800 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13801 	 * p.869. In such cases, we can still calculate the RTT correctly
13802 	 * when RCV.NXT == Last.ACK.Sent.
13803 	 */
13804 	if ((to->to_flags & TOF_TS) != 0 &&
13805 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13806 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13807 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13808 		tp->ts_recent_age = tcp_ts_getticks();
13809 		tp->ts_recent = to->to_tsval;
13810 	}
13811 	/*
13812 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13813 	 * is on (half-synchronized state), then queue data for later
13814 	 * processing; else drop segment and return.
13815 	 */
13816 	if ((thflags & TH_ACK) == 0) {
13817 		if (tp->t_flags & TF_NEEDSYN) {
13818 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13819 			    tiwin, thflags, nxt_pkt));
13820 
13821 		} else if (tp->t_flags & TF_ACKNOW) {
13822 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13823 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13824 			return (ret_val);
13825 		} else {
13826 			ctf_do_drop(m, NULL);
13827 			return (0);
13828 		}
13829 	}
13830 	/*
13831 	 * Ack processing.
13832 	 */
13833 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
13834 		return (ret_val);
13835 	}
13836 	if (sbavail(&so->so_snd)) {
13837 		if (ctf_progress_timeout_check(tp, true)) {
13838 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13839 						tp, tick, PROGRESS_DROP, __LINE__);
13840 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13841 			return (1);
13842 		}
13843 	}
13844 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13845 	    tiwin, thflags, nxt_pkt));
13846 }
13847 
13848 static int
13849 rack_check_data_after_close(struct mbuf *m,
13850     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
13851 {
13852 	struct tcp_rack *rack;
13853 
13854 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13855 	if (rack->rc_allow_data_af_clo == 0) {
13856 	close_now:
13857 		tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13858 		/* tcp_close will kill the inp pre-log the Reset */
13859 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
13860 		tp = tcp_close(tp);
13861 		KMOD_TCPSTAT_INC(tcps_rcvafterclose);
13862 		ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
13863 		return (1);
13864 	}
13865 	if (sbavail(&so->so_snd) == 0)
13866 		goto close_now;
13867 	/* Ok we allow data that is ignored and a followup reset */
13868 	tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE);
13869 	tp->rcv_nxt = th->th_seq + *tlen;
13870 	tp->t_flags2 |= TF2_DROP_AF_DATA;
13871 	rack->r_wanted_output = 1;
13872 	*tlen = 0;
13873 	return (0);
13874 }
13875 
13876 /*
13877  * Return value of 1, the TCB is unlocked and most
13878  * likely gone, return value of 0, the TCP is still
13879  * locked.
13880  */
13881 static int
13882 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
13883     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
13884     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
13885 {
13886 	int32_t ret_val = 0;
13887 	int32_t ourfinisacked = 0;
13888 	struct tcp_rack *rack;
13889 
13890 	rack = (struct tcp_rack *)tp->t_fb_ptr;
13891 	ctf_calc_rwin(so, tp);
13892 
13893 	if ((thflags & TH_RST) ||
13894 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
13895 		return (__ctf_process_rst(m, th, so, tp,
13896 					  &rack->r_ctl.challenge_ack_ts,
13897 					  &rack->r_ctl.challenge_ack_cnt));
13898 	/*
13899 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
13900 	 * synchronized state.
13901 	 */
13902 	if (thflags & TH_SYN) {
13903 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
13904 		return (ret_val);
13905 	}
13906 	/*
13907 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
13908 	 * it's less than ts_recent, drop it.
13909 	 */
13910 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
13911 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
13912 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
13913 			return (ret_val);
13914 	}
13915 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
13916 			      &rack->r_ctl.challenge_ack_ts,
13917 			      &rack->r_ctl.challenge_ack_cnt)) {
13918 		return (ret_val);
13919 	}
13920 	/*
13921 	 * If new data are received on a connection after the user processes
13922 	 * are gone, then RST the other end.
13923 	 */
13924 	if ((tp->t_flags & TF_CLOSED) && tlen &&
13925 	    rack_check_data_after_close(m, tp, &tlen, th, so))
13926 		return (1);
13927 	/*
13928 	 * If last ACK falls within this segment's sequence numbers, record
13929 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
13930 	 * from the latest proposal of the tcplw@cray.com list (Braden
13931 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
13932 	 * with our earlier PAWS tests, so this check should be solely
13933 	 * predicated on the sequence space of this segment. 3) That we
13934 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
13935 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
13936 	 * SEG.Len, This modified check allows us to overcome RFC1323's
13937 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
13938 	 * p.869. In such cases, we can still calculate the RTT correctly
13939 	 * when RCV.NXT == Last.ACK.Sent.
13940 	 */
13941 	if ((to->to_flags & TOF_TS) != 0 &&
13942 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
13943 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
13944 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
13945 		tp->ts_recent_age = tcp_ts_getticks();
13946 		tp->ts_recent = to->to_tsval;
13947 	}
13948 	/*
13949 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
13950 	 * is on (half-synchronized state), then queue data for later
13951 	 * processing; else drop segment and return.
13952 	 */
13953 	if ((thflags & TH_ACK) == 0) {
13954 		if (tp->t_flags & TF_NEEDSYN) {
13955 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
13956 			    tiwin, thflags, nxt_pkt));
13957 		} else if (tp->t_flags & TF_ACKNOW) {
13958 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
13959 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
13960 			return (ret_val);
13961 		} else {
13962 			ctf_do_drop(m, NULL);
13963 			return (0);
13964 		}
13965 	}
13966 	/*
13967 	 * Ack processing.
13968 	 */
13969 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
13970 		return (ret_val);
13971 	}
13972 	if (ourfinisacked) {
13973 		/*
13974 		 * If we can't receive any more data, then closing user can
13975 		 * proceed. Starting the timer is contrary to the
13976 		 * specification, but if we don't get a FIN we'll hang
13977 		 * forever.
13978 		 *
13979 		 * XXXjl: we should release the tp also, and use a
13980 		 * compressed state.
13981 		 */
13982 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
13983 			soisdisconnected(so);
13984 			tcp_timer_activate(tp, TT_2MSL,
13985 			    (tcp_fast_finwait2_recycle ?
13986 			    tcp_finwait2_timeout :
13987 			    TP_MAXIDLE(tp)));
13988 		}
13989 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
13990 	}
13991 	if (sbavail(&so->so_snd)) {
13992 		if (ctf_progress_timeout_check(tp, true)) {
13993 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
13994 						tp, tick, PROGRESS_DROP, __LINE__);
13995 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
13996 			return (1);
13997 		}
13998 	}
13999 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14000 	    tiwin, thflags, nxt_pkt));
14001 }
14002 
14003 /*
14004  * Return value of 1, the TCB is unlocked and most
14005  * likely gone, return value of 0, the TCP is still
14006  * locked.
14007  */
14008 static int
14009 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
14010     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14011     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14012 {
14013 	int32_t ret_val = 0;
14014 	int32_t ourfinisacked = 0;
14015 	struct tcp_rack *rack;
14016 
14017 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14018 	ctf_calc_rwin(so, tp);
14019 
14020 	if ((thflags & TH_RST) ||
14021 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14022 		return (__ctf_process_rst(m, th, so, tp,
14023 					  &rack->r_ctl.challenge_ack_ts,
14024 					  &rack->r_ctl.challenge_ack_cnt));
14025 	/*
14026 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14027 	 * synchronized state.
14028 	 */
14029 	if (thflags & TH_SYN) {
14030 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14031 		return (ret_val);
14032 	}
14033 	/*
14034 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14035 	 * it's less than ts_recent, drop it.
14036 	 */
14037 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14038 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14039 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14040 			return (ret_val);
14041 	}
14042 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14043 			      &rack->r_ctl.challenge_ack_ts,
14044 			      &rack->r_ctl.challenge_ack_cnt)) {
14045 		return (ret_val);
14046 	}
14047 	/*
14048 	 * If new data are received on a connection after the user processes
14049 	 * are gone, then RST the other end.
14050 	 */
14051 	if ((tp->t_flags & TF_CLOSED) && tlen &&
14052 	    rack_check_data_after_close(m, tp, &tlen, th, so))
14053 		return (1);
14054 	/*
14055 	 * If last ACK falls within this segment's sequence numbers, record
14056 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14057 	 * from the latest proposal of the tcplw@cray.com list (Braden
14058 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14059 	 * with our earlier PAWS tests, so this check should be solely
14060 	 * predicated on the sequence space of this segment. 3) That we
14061 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14062 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14063 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14064 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14065 	 * p.869. In such cases, we can still calculate the RTT correctly
14066 	 * when RCV.NXT == Last.ACK.Sent.
14067 	 */
14068 	if ((to->to_flags & TOF_TS) != 0 &&
14069 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14070 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14071 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14072 		tp->ts_recent_age = tcp_ts_getticks();
14073 		tp->ts_recent = to->to_tsval;
14074 	}
14075 	/*
14076 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14077 	 * is on (half-synchronized state), then queue data for later
14078 	 * processing; else drop segment and return.
14079 	 */
14080 	if ((thflags & TH_ACK) == 0) {
14081 		if (tp->t_flags & TF_NEEDSYN) {
14082 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14083 			    tiwin, thflags, nxt_pkt));
14084 		} else if (tp->t_flags & TF_ACKNOW) {
14085 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14086 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14087 			return (ret_val);
14088 		} else {
14089 			ctf_do_drop(m, NULL);
14090 			return (0);
14091 		}
14092 	}
14093 	/*
14094 	 * Ack processing.
14095 	 */
14096 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
14097 		return (ret_val);
14098 	}
14099 	if (ourfinisacked) {
14100 		tcp_twstart(tp);
14101 		m_freem(m);
14102 		return (1);
14103 	}
14104 	if (sbavail(&so->so_snd)) {
14105 		if (ctf_progress_timeout_check(tp, true)) {
14106 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14107 						tp, tick, PROGRESS_DROP, __LINE__);
14108 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14109 			return (1);
14110 		}
14111 	}
14112 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14113 	    tiwin, thflags, nxt_pkt));
14114 }
14115 
14116 /*
14117  * Return value of 1, the TCB is unlocked and most
14118  * likely gone, return value of 0, the TCP is still
14119  * locked.
14120  */
14121 static int
14122 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
14123     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14124     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14125 {
14126 	int32_t ret_val = 0;
14127 	int32_t ourfinisacked = 0;
14128 	struct tcp_rack *rack;
14129 
14130 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14131 	ctf_calc_rwin(so, tp);
14132 
14133 	if ((thflags & TH_RST) ||
14134 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14135 		return (__ctf_process_rst(m, th, so, tp,
14136 					  &rack->r_ctl.challenge_ack_ts,
14137 					  &rack->r_ctl.challenge_ack_cnt));
14138 	/*
14139 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14140 	 * synchronized state.
14141 	 */
14142 	if (thflags & TH_SYN) {
14143 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14144 		return (ret_val);
14145 	}
14146 	/*
14147 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14148 	 * it's less than ts_recent, drop it.
14149 	 */
14150 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14151 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14152 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14153 			return (ret_val);
14154 	}
14155 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14156 			      &rack->r_ctl.challenge_ack_ts,
14157 			      &rack->r_ctl.challenge_ack_cnt)) {
14158 		return (ret_val);
14159 	}
14160 	/*
14161 	 * If new data are received on a connection after the user processes
14162 	 * are gone, then RST the other end.
14163 	 */
14164 	if ((tp->t_flags & TF_CLOSED) && tlen &&
14165 	    rack_check_data_after_close(m, tp, &tlen, th, so))
14166 		return (1);
14167 	/*
14168 	 * If last ACK falls within this segment's sequence numbers, record
14169 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14170 	 * from the latest proposal of the tcplw@cray.com list (Braden
14171 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14172 	 * with our earlier PAWS tests, so this check should be solely
14173 	 * predicated on the sequence space of this segment. 3) That we
14174 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14175 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14176 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14177 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14178 	 * p.869. In such cases, we can still calculate the RTT correctly
14179 	 * when RCV.NXT == Last.ACK.Sent.
14180 	 */
14181 	if ((to->to_flags & TOF_TS) != 0 &&
14182 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14183 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14184 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14185 		tp->ts_recent_age = tcp_ts_getticks();
14186 		tp->ts_recent = to->to_tsval;
14187 	}
14188 	/*
14189 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14190 	 * is on (half-synchronized state), then queue data for later
14191 	 * processing; else drop segment and return.
14192 	 */
14193 	if ((thflags & TH_ACK) == 0) {
14194 		if (tp->t_flags & TF_NEEDSYN) {
14195 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14196 			    tiwin, thflags, nxt_pkt));
14197 		} else if (tp->t_flags & TF_ACKNOW) {
14198 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14199 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14200 			return (ret_val);
14201 		} else {
14202 			ctf_do_drop(m, NULL);
14203 			return (0);
14204 		}
14205 	}
14206 	/*
14207 	 * case TCPS_LAST_ACK: Ack processing.
14208 	 */
14209 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
14210 		return (ret_val);
14211 	}
14212 	if (ourfinisacked) {
14213 		tp = tcp_close(tp);
14214 		ctf_do_drop(m, tp);
14215 		return (1);
14216 	}
14217 	if (sbavail(&so->so_snd)) {
14218 		if (ctf_progress_timeout_check(tp, true)) {
14219 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14220 						tp, tick, PROGRESS_DROP, __LINE__);
14221 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14222 			return (1);
14223 		}
14224 	}
14225 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14226 	    tiwin, thflags, nxt_pkt));
14227 }
14228 
14229 /*
14230  * Return value of 1, the TCB is unlocked and most
14231  * likely gone, return value of 0, the TCP is still
14232  * locked.
14233  */
14234 static int
14235 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
14236     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
14237     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos)
14238 {
14239 	int32_t ret_val = 0;
14240 	int32_t ourfinisacked = 0;
14241 	struct tcp_rack *rack;
14242 
14243 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14244 	ctf_calc_rwin(so, tp);
14245 
14246 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
14247 	if ((thflags & TH_RST) ||
14248 	    (tp->t_fin_is_rst && (thflags & TH_FIN)))
14249 		return (__ctf_process_rst(m, th, so, tp,
14250 					  &rack->r_ctl.challenge_ack_ts,
14251 					  &rack->r_ctl.challenge_ack_cnt));
14252 	/*
14253 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
14254 	 * synchronized state.
14255 	 */
14256 	if (thflags & TH_SYN) {
14257 		ctf_challenge_ack(m, th, tp, iptos, &ret_val);
14258 		return (ret_val);
14259 	}
14260 	/*
14261 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
14262 	 * it's less than ts_recent, drop it.
14263 	 */
14264 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
14265 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
14266 		if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val))
14267 			return (ret_val);
14268 	}
14269 	if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val,
14270 			      &rack->r_ctl.challenge_ack_ts,
14271 			      &rack->r_ctl.challenge_ack_cnt)) {
14272 		return (ret_val);
14273 	}
14274 	/*
14275 	 * If new data are received on a connection after the user processes
14276 	 * are gone, then RST the other end.
14277 	 */
14278 	if ((tp->t_flags & TF_CLOSED) && tlen &&
14279 	    rack_check_data_after_close(m, tp, &tlen, th, so))
14280 		return (1);
14281 	/*
14282 	 * If last ACK falls within this segment's sequence numbers, record
14283 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
14284 	 * from the latest proposal of the tcplw@cray.com list (Braden
14285 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
14286 	 * with our earlier PAWS tests, so this check should be solely
14287 	 * predicated on the sequence space of this segment. 3) That we
14288 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
14289 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
14290 	 * SEG.Len, This modified check allows us to overcome RFC1323's
14291 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
14292 	 * p.869. In such cases, we can still calculate the RTT correctly
14293 	 * when RCV.NXT == Last.ACK.Sent.
14294 	 */
14295 	if ((to->to_flags & TOF_TS) != 0 &&
14296 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
14297 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
14298 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
14299 		tp->ts_recent_age = tcp_ts_getticks();
14300 		tp->ts_recent = to->to_tsval;
14301 	}
14302 	/*
14303 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
14304 	 * is on (half-synchronized state), then queue data for later
14305 	 * processing; else drop segment and return.
14306 	 */
14307 	if ((thflags & TH_ACK) == 0) {
14308 		if (tp->t_flags & TF_NEEDSYN) {
14309 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14310 			    tiwin, thflags, nxt_pkt));
14311 		} else if (tp->t_flags & TF_ACKNOW) {
14312 			ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
14313 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1;
14314 			return (ret_val);
14315 		} else {
14316 			ctf_do_drop(m, NULL);
14317 			return (0);
14318 		}
14319 	}
14320 	/*
14321 	 * Ack processing.
14322 	 */
14323 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
14324 		return (ret_val);
14325 	}
14326 	if (sbavail(&so->so_snd)) {
14327 		if (ctf_progress_timeout_check(tp, true)) {
14328 			rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
14329 						tp, tick, PROGRESS_DROP, __LINE__);
14330 			ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
14331 			return (1);
14332 		}
14333 	}
14334 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
14335 	    tiwin, thflags, nxt_pkt));
14336 }
14337 
14338 static void inline
14339 rack_clear_rate_sample(struct tcp_rack *rack)
14340 {
14341 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
14342 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
14343 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
14344 }
14345 
14346 static void
14347 rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override)
14348 {
14349 	uint64_t bw_est, rate_wanted;
14350 	int chged = 0;
14351 	uint32_t user_max, orig_min, orig_max;
14352 
14353 #ifdef TCP_REQUEST_TRK
14354 	if (rack->rc_hybrid_mode &&
14355 	    (rack->r_ctl.rc_pace_max_segs != 0) &&
14356 	    (rack_hybrid_allow_set_maxseg == 1) &&
14357 	    (rack->r_ctl.rc_last_sft != NULL)) {
14358 		rack->r_ctl.rc_last_sft->hybrid_flags &= ~TCP_HYBRID_PACING_SETMSS;
14359 		return;
14360 	}
14361 #endif
14362 	orig_min = rack->r_ctl.rc_pace_min_segs;
14363 	orig_max = rack->r_ctl.rc_pace_max_segs;
14364 	user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs;
14365 	if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs)
14366 		chged = 1;
14367 	rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp);
14368 	if (rack->use_fixed_rate || rack->rc_force_max_seg) {
14369 		if (user_max != rack->r_ctl.rc_pace_max_segs)
14370 			chged = 1;
14371 	}
14372 	if (rack->rc_force_max_seg) {
14373 		rack->r_ctl.rc_pace_max_segs = user_max;
14374 	} else if (rack->use_fixed_rate) {
14375 		bw_est = rack_get_bw(rack);
14376 		if ((rack->r_ctl.crte == NULL) ||
14377 		    (bw_est != rack->r_ctl.crte->rate)) {
14378 			rack->r_ctl.rc_pace_max_segs = user_max;
14379 		} else {
14380 			/* We are pacing right at the hardware rate */
14381 			uint32_t segsiz, pace_one;
14382 
14383 			if (rack_pace_one_seg ||
14384 			    (rack->r_ctl.rc_user_set_min_segs == 1))
14385 				pace_one = 1;
14386 			else
14387 				pace_one = 0;
14388 			segsiz = min(ctf_fixed_maxseg(tp),
14389 				     rack->r_ctl.rc_pace_min_segs);
14390 			rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(
14391 				tp, bw_est, segsiz, pace_one,
14392 				rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
14393 		}
14394 	} else if (rack->rc_always_pace) {
14395 		if (rack->r_ctl.gp_bw ||
14396 		    rack->r_ctl.init_rate) {
14397 			/* We have a rate of some sort set */
14398 			uint32_t  orig;
14399 
14400 			bw_est = rack_get_bw(rack);
14401 			orig = rack->r_ctl.rc_pace_max_segs;
14402 			if (fill_override)
14403 				rate_wanted = *fill_override;
14404 			else
14405 				rate_wanted = rack_get_gp_est(rack);
14406 			if (rate_wanted) {
14407 				/* We have something */
14408 				rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack,
14409 										   rate_wanted,
14410 										   ctf_fixed_maxseg(rack->rc_tp));
14411 			} else
14412 				rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs;
14413 			if (orig != rack->r_ctl.rc_pace_max_segs)
14414 				chged = 1;
14415 		} else if ((rack->r_ctl.gp_bw == 0) &&
14416 			   (rack->r_ctl.rc_pace_max_segs == 0)) {
14417 			/*
14418 			 * If we have nothing limit us to bursting
14419 			 * out IW sized pieces.
14420 			 */
14421 			chged = 1;
14422 			rack->r_ctl.rc_pace_max_segs = rc_init_window(rack);
14423 		}
14424 	}
14425 	if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) {
14426 		chged = 1;
14427 		rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES;
14428 	}
14429 	if (chged)
14430 		rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2);
14431 }
14432 
14433 
14434 static void
14435 rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack, int32_t flags)
14436 {
14437 #ifdef INET6
14438 	struct ip6_hdr *ip6 = NULL;
14439 #endif
14440 #ifdef INET
14441 	struct ip *ip = NULL;
14442 #endif
14443 	struct udphdr *udp = NULL;
14444 
14445 	/* Ok lets fill in the fast block, it can only be used with no IP options! */
14446 #ifdef INET6
14447 	if (rack->r_is_v6) {
14448 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
14449 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
14450 		if (tp->t_port) {
14451 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14452 			udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
14453 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14454 			udp->uh_dport = tp->t_port;
14455 			rack->r_ctl.fsb.udp = udp;
14456 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14457 		} else
14458 		{
14459 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1);
14460 			rack->r_ctl.fsb.udp = NULL;
14461 		}
14462 		tcpip_fillheaders(rack->rc_inp,
14463 				  tp->t_port,
14464 				  ip6, rack->r_ctl.fsb.th);
14465 		rack->r_ctl.fsb.hoplimit = in6_selecthlim(rack->rc_inp, NULL);
14466 	} else
14467 #endif				/* INET6 */
14468 #ifdef INET
14469 	{
14470 		rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr);
14471 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
14472 		if (tp->t_port) {
14473 			rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr);
14474 			udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
14475 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
14476 			udp->uh_dport = tp->t_port;
14477 			rack->r_ctl.fsb.udp = udp;
14478 			rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1);
14479 		} else
14480 		{
14481 			rack->r_ctl.fsb.udp = NULL;
14482 			rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1);
14483 		}
14484 		tcpip_fillheaders(rack->rc_inp,
14485 				  tp->t_port,
14486 				  ip, rack->r_ctl.fsb.th);
14487 		rack->r_ctl.fsb.hoplimit = tptoinpcb(tp)->inp_ip_ttl;
14488 	}
14489 #endif
14490 	rack->r_ctl.fsb.recwin = lmin(lmax(sbspace(&tptosocket(tp)->so_rcv), 0),
14491 	    (long)TCP_MAXWIN << tp->rcv_scale);
14492 	rack->r_fsb_inited = 1;
14493 }
14494 
14495 static int
14496 rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack)
14497 {
14498 	/*
14499 	 * Allocate the larger of spaces V6 if available else just
14500 	 * V4 and include udphdr (overbook)
14501 	 */
14502 #ifdef INET6
14503 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr);
14504 #else
14505 	rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr);
14506 #endif
14507 	rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len,
14508 					    M_TCPFSB, M_NOWAIT|M_ZERO);
14509 	if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) {
14510 		return (ENOMEM);
14511 	}
14512 	rack->r_fsb_inited = 0;
14513 	return (0);
14514 }
14515 
14516 static void
14517 rack_log_hystart_event(struct tcp_rack *rack, uint32_t high_seq, uint8_t mod)
14518 {
14519 	/*
14520 	 * Types of logs (mod value)
14521 	 * 20 - Initial round setup
14522 	 * 21 - Rack declares a new round.
14523 	 */
14524 	struct tcpcb *tp;
14525 
14526 	tp = rack->rc_tp;
14527 	if (tcp_bblogging_on(tp)) {
14528 		union tcp_log_stackspecific log;
14529 		struct timeval tv;
14530 
14531 		memset(&log, 0, sizeof(log));
14532 		log.u_bbr.flex1 = rack->r_ctl.current_round;
14533 		log.u_bbr.flex2 = rack->r_ctl.roundends;
14534 		log.u_bbr.flex3 = high_seq;
14535 		log.u_bbr.flex4 = tp->snd_max;
14536 		log.u_bbr.flex8 = mod;
14537 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14538 		log.u_bbr.cur_del_rate = rack->rc_tp->t_sndbytes;
14539 		log.u_bbr.delRate = rack->rc_tp->t_snd_rxt_bytes;
14540 		TCP_LOG_EVENTP(tp, NULL,
14541 		    &tptosocket(tp)->so_rcv,
14542 		    &tptosocket(tp)->so_snd,
14543 		    TCP_HYSTART, 0,
14544 		    0, &log, false, &tv);
14545 	}
14546 }
14547 
14548 static void
14549 rack_deferred_init(struct tcpcb *tp, struct tcp_rack *rack)
14550 {
14551 	rack->rack_deferred_inited = 1;
14552 	rack->r_ctl.roundends = tp->snd_max;
14553 	rack->r_ctl.rc_high_rwnd = tp->snd_wnd;
14554 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
14555 }
14556 
14557 static void
14558 rack_init_retransmit_value(struct tcp_rack *rack, int ctl)
14559 {
14560 	/* Retransmit bit controls.
14561 	 *
14562 	 * The setting of these values control one of
14563 	 * three settings you can have and dictate
14564 	 * how rack does retransmissions. Note this
14565 	 * is in *any* mode i.e. pacing on or off DGP
14566 	 * fixed rate pacing, or just bursting rack.
14567 	 *
14568 	 * 1 - Use full sized retransmits i.e. limit
14569 	 *     the size to whatever the pace_max_segments
14570 	 *     size is.
14571 	 *
14572 	 * 2 - Use pacer min granularity as a guide to
14573 	 *     the size combined with the current calculated
14574 	 *     goodput b/w measurement. So for example if
14575 	 *     the goodput is measured at 20Mbps we would
14576 	 *     calculate 8125 (pacer minimum 250usec in
14577 	 *     that b/w) and then round it up to the next
14578 	 *     MSS i.e. for 1448 mss 6 MSS or 8688 bytes.
14579 	 *
14580 	 * 0 - The rack default 1 MSS (anything not 0/1/2
14581 	 *     fall here too if we are setting via rack_init()).
14582 	 *
14583 	 */
14584 	if (ctl == 1) {
14585 		rack->full_size_rxt = 1;
14586 		rack->shape_rxt_to_pacing_min  = 0;
14587 	} else if (ctl == 2) {
14588 		rack->full_size_rxt = 0;
14589 		rack->shape_rxt_to_pacing_min  = 1;
14590 	} else {
14591 		rack->full_size_rxt = 0;
14592 		rack->shape_rxt_to_pacing_min  = 0;
14593 	}
14594 }
14595 
14596 static void
14597 rack_log_chg_info(struct tcpcb *tp, struct tcp_rack *rack, uint8_t mod,
14598 		  uint32_t flex1,
14599 		  uint32_t flex2,
14600 		  uint32_t flex3)
14601 {
14602 	if (tcp_bblogging_on(rack->rc_tp)) {
14603 		union tcp_log_stackspecific log;
14604 		struct timeval tv;
14605 
14606 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
14607 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
14608 		log.u_bbr.flex8 = mod;
14609 		log.u_bbr.flex1 = flex1;
14610 		log.u_bbr.flex2 = flex2;
14611 		log.u_bbr.flex3 = flex3;
14612 		tcp_log_event(tp, NULL, NULL, NULL, TCP_CHG_QUERY, 0,
14613 			       0, &log, false, NULL, __func__, __LINE__, &tv);
14614 	}
14615 }
14616 
14617 static int
14618 rack_chg_query(struct tcpcb *tp, struct tcp_query_resp *reqr)
14619 {
14620 	struct tcp_rack *rack;
14621 	struct rack_sendmap *rsm;
14622 	int i;
14623 
14624 
14625 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14626 	switch (reqr->req) {
14627 	case TCP_QUERY_SENDMAP:
14628 		if ((reqr->req_param == tp->snd_max) ||
14629 		    (tp->snd_max == tp->snd_una)){
14630 			/* Unlikely */
14631 			return (0);
14632 		}
14633 		rsm = tqhash_find(rack->r_ctl.tqh, reqr->req_param);
14634 		if (rsm == NULL) {
14635 			/* Can't find that seq -- unlikely */
14636 			return (0);
14637 		}
14638 		reqr->sendmap_start = rsm->r_start;
14639 		reqr->sendmap_end = rsm->r_end;
14640 		reqr->sendmap_send_cnt = rsm->r_rtr_cnt;
14641 		reqr->sendmap_fas = rsm->r_fas;
14642 		if (reqr->sendmap_send_cnt > SNDMAP_NRTX)
14643 			reqr->sendmap_send_cnt = SNDMAP_NRTX;
14644 		for(i=0; i<reqr->sendmap_send_cnt; i++)
14645 			reqr->sendmap_time[i] = rsm->r_tim_lastsent[i];
14646 		reqr->sendmap_ack_arrival = rsm->r_ack_arrival;
14647 		reqr->sendmap_flags = rsm->r_flags & SNDMAP_MASK;
14648 		reqr->sendmap_r_rtr_bytes = rsm->r_rtr_bytes;
14649 		reqr->sendmap_dupacks = rsm->r_dupack;
14650 		rack_log_chg_info(tp, rack, 1,
14651 				  rsm->r_start,
14652 				  rsm->r_end,
14653 				  rsm->r_flags);
14654 		return(1);
14655 		break;
14656 	case TCP_QUERY_TIMERS_UP:
14657 		if (rack->r_ctl.rc_hpts_flags == 0) {
14658 			/* no timers up */
14659 			return (0);
14660 		}
14661 		reqr->timer_hpts_flags = rack->r_ctl.rc_hpts_flags;
14662 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14663 			reqr->timer_pacing_to = rack->r_ctl.rc_last_output_to;
14664 		}
14665 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14666 			reqr->timer_timer_exp = rack->r_ctl.rc_timer_exp;
14667 		}
14668 		rack_log_chg_info(tp, rack, 2,
14669 				  rack->r_ctl.rc_hpts_flags,
14670 				  rack->r_ctl.rc_last_output_to,
14671 				  rack->r_ctl.rc_timer_exp);
14672 		return (1);
14673 		break;
14674 	case TCP_QUERY_RACK_TIMES:
14675 		/* Reordering items */
14676 		reqr->rack_num_dsacks = rack->r_ctl.num_dsack;
14677 		reqr->rack_reorder_ts = rack->r_ctl.rc_reorder_ts;
14678 		/* Timerstamps and timers */
14679 		reqr->rack_rxt_last_time = rack->r_ctl.rc_tlp_rxt_last_time;
14680 		reqr->rack_min_rtt = rack->r_ctl.rc_rack_min_rtt;
14681 		reqr->rack_rtt = rack->rc_rack_rtt;
14682 		reqr->rack_tmit_time = rack->r_ctl.rc_rack_tmit_time;
14683 		reqr->rack_srtt_measured = rack->rc_srtt_measure_made;
14684 		/* PRR data */
14685 		reqr->rack_sacked = rack->r_ctl.rc_sacked;
14686 		reqr->rack_holes_rxt = rack->r_ctl.rc_holes_rxt;
14687 		reqr->rack_prr_delivered = rack->r_ctl.rc_prr_delivered;
14688 		reqr->rack_prr_recovery_fs = rack->r_ctl.rc_prr_recovery_fs;
14689 		reqr->rack_prr_sndcnt = rack->r_ctl.rc_prr_sndcnt;
14690 		reqr->rack_prr_out = rack->r_ctl.rc_prr_out;
14691 		/* TLP and persists info */
14692 		reqr->rack_tlp_out = rack->rc_tlp_in_progress;
14693 		reqr->rack_tlp_cnt_out = rack->r_ctl.rc_tlp_cnt_out;
14694 		if (rack->rc_in_persist) {
14695 			reqr->rack_time_went_idle = rack->r_ctl.rc_went_idle_time;
14696 			reqr->rack_in_persist = 1;
14697 		} else {
14698 			reqr->rack_time_went_idle = 0;
14699 			reqr->rack_in_persist = 0;
14700 		}
14701 		if (rack->r_wanted_output)
14702 			reqr->rack_wanted_output = 1;
14703 		else
14704 			reqr->rack_wanted_output = 0;
14705 		return (1);
14706 		break;
14707 	default:
14708 		return (-EINVAL);
14709 	}
14710 }
14711 
14712 static void
14713 rack_switch_failed(struct tcpcb *tp)
14714 {
14715 	/*
14716 	 * This method gets called if a stack switch was
14717 	 * attempted and it failed. We are left
14718 	 * but our hpts timers were stopped and we
14719 	 * need to validate time units and t_flags2.
14720 	 */
14721 	struct tcp_rack *rack;
14722 	struct timeval tv;
14723 	uint32_t cts;
14724 	uint32_t toval;
14725 	struct hpts_diag diag;
14726 
14727 	rack = (struct tcp_rack *)tp->t_fb_ptr;
14728 	tcp_change_time_units(tp, TCP_TMR_GRANULARITY_USEC);
14729 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
14730 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
14731 	else
14732 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
14733 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
14734 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
14735 	if (tp->t_in_hpts > IHPTS_NONE) {
14736 		/* Strange */
14737 		return;
14738 	}
14739 	cts = tcp_get_usecs(&tv);
14740 	if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
14741 		if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
14742 			toval = rack->r_ctl.rc_last_output_to - cts;
14743 		} else {
14744 			/* one slot please */
14745 			toval = HPTS_TICKS_PER_SLOT;
14746 		}
14747 	} else if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
14748 		if (TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
14749 			toval = rack->r_ctl.rc_timer_exp - cts;
14750 		} else {
14751 			/* one slot please */
14752 			toval = HPTS_TICKS_PER_SLOT;
14753 		}
14754 	} else
14755 		toval = HPTS_TICKS_PER_SLOT;
14756 	(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(toval),
14757 				   __LINE__, &diag);
14758 	rack_log_hpts_diag(rack, cts, &diag, &tv);
14759 }
14760 
14761 static int
14762 rack_init_outstanding(struct tcpcb *tp, struct tcp_rack *rack, uint32_t us_cts, void *ptr)
14763 {
14764 	struct rack_sendmap *rsm, *ersm;
14765 	int insret __diagused;
14766 	/*
14767 	 * When initing outstanding, we must be quite careful
14768 	 * to not refer to tp->t_fb_ptr. This has the old rack
14769 	 * pointer in it, not the "new" one (when we are doing
14770 	 * a stack switch).
14771 	 */
14772 
14773 
14774 	if (tp->t_fb->tfb_chg_query == NULL) {
14775 		/* Create a send map for the current outstanding data */
14776 
14777 		rsm = rack_alloc(rack);
14778 		if (rsm == NULL) {
14779 			uma_zfree(rack_pcb_zone, ptr);
14780 			return (ENOMEM);
14781 		}
14782 		rsm->r_no_rtt_allowed = 1;
14783 		rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time);
14784 		rsm->r_rtr_cnt = 1;
14785 		rsm->r_rtr_bytes = 0;
14786 		if (tp->t_flags & TF_SENTFIN)
14787 			rsm->r_flags |= RACK_HAS_FIN;
14788 		rsm->r_end = tp->snd_max;
14789 		if (tp->snd_una == tp->iss) {
14790 			/* The data space is one beyond snd_una */
14791 			rsm->r_flags |= RACK_HAS_SYN;
14792 			rsm->r_start = tp->iss;
14793 			rsm->r_end = rsm->r_start + (tp->snd_max - tp->snd_una);
14794 		} else
14795 			rsm->r_start = tp->snd_una;
14796 		rsm->r_dupack = 0;
14797 		if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) {
14798 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff);
14799 			if (rsm->m) {
14800 				rsm->orig_m_len = rsm->m->m_len;
14801 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14802 			} else {
14803 				rsm->orig_m_len = 0;
14804 				rsm->orig_t_space = 0;
14805 			}
14806 		} else {
14807 			/*
14808 			 * This can happen if we have a stand-alone FIN or
14809 			 *  SYN.
14810 			 */
14811 			rsm->m = NULL;
14812 			rsm->orig_m_len = 0;
14813 			rsm->orig_t_space = 0;
14814 			rsm->soff = 0;
14815 		}
14816 #ifdef INVARIANTS
14817 		if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14818 			panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
14819 			      insret, rack, rsm);
14820 		}
14821 #else
14822 		(void)tqhash_insert(rack->r_ctl.tqh, rsm);
14823 #endif
14824 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14825 		rsm->r_in_tmap = 1;
14826 	} else {
14827 		/* We have a query mechanism, lets use it */
14828 		struct tcp_query_resp qr;
14829 		int i;
14830 		tcp_seq at;
14831 
14832 		at = tp->snd_una;
14833 		while (at != tp->snd_max) {
14834 			memset(&qr, 0, sizeof(qr));
14835 			qr.req = TCP_QUERY_SENDMAP;
14836 			qr.req_param = at;
14837 			if ((*tp->t_fb->tfb_chg_query)(tp, &qr) == 0)
14838 				break;
14839 			/* Move forward */
14840 			at = qr.sendmap_end;
14841 			/* Now lets build the entry for this one */
14842 			rsm = rack_alloc(rack);
14843 			if (rsm == NULL) {
14844 				uma_zfree(rack_pcb_zone, ptr);
14845 				return (ENOMEM);
14846 			}
14847 			memset(rsm, 0, sizeof(struct rack_sendmap));
14848 			/* Now configure the rsm and insert it */
14849 			rsm->r_dupack = qr.sendmap_dupacks;
14850 			rsm->r_start = qr.sendmap_start;
14851 			rsm->r_end = qr.sendmap_end;
14852 			if (qr.sendmap_fas)
14853 				rsm->r_fas = qr.sendmap_end;
14854 			else
14855 				rsm->r_fas = rsm->r_start - tp->snd_una;
14856 			/*
14857 			 * We have carefully aligned the bits
14858 			 * so that all we have to do is copy over
14859 			 * the bits with the mask.
14860 			 */
14861 			rsm->r_flags = qr.sendmap_flags & SNDMAP_MASK;
14862 			rsm->r_rtr_bytes = qr.sendmap_r_rtr_bytes;
14863 			rsm->r_rtr_cnt = qr.sendmap_send_cnt;
14864 			rsm->r_ack_arrival = qr.sendmap_ack_arrival;
14865 			for (i=0 ; i<rsm->r_rtr_cnt; i++)
14866 				rsm->r_tim_lastsent[i]	= qr.sendmap_time[i];
14867 			rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd,
14868 					   (rsm->r_start - tp->snd_una), &rsm->soff);
14869 			if (rsm->m) {
14870 				rsm->orig_m_len = rsm->m->m_len;
14871 				rsm->orig_t_space = M_TRAILINGROOM(rsm->m);
14872 			} else {
14873 				rsm->orig_m_len = 0;
14874 				rsm->orig_t_space = 0;
14875 			}
14876 #ifdef INVARIANTS
14877 			if ((insret = tqhash_insert(rack->r_ctl.tqh, rsm)) != 0) {
14878 				panic("Insert in tailq_hash fails ret:%d rack:%p rsm:%p",
14879 				      insret, rack, rsm);
14880 			}
14881 #else
14882 			(void)tqhash_insert(rack->r_ctl.tqh, rsm);
14883 #endif
14884 			if ((rsm->r_flags & RACK_ACKED) == 0)  {
14885 				TAILQ_FOREACH(ersm, &rack->r_ctl.rc_tmap, r_tnext) {
14886 					if (ersm->r_tim_lastsent[(ersm->r_rtr_cnt-1)] >
14887 					    rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]) {
14888 						/*
14889 						 * If the existing ersm was sent at
14890 						 * a later time than the new one, then
14891 						 * the new one should appear ahead of this
14892 						 * ersm.
14893 						 */
14894 						rsm->r_in_tmap = 1;
14895 						TAILQ_INSERT_BEFORE(ersm, rsm, r_tnext);
14896 						break;
14897 					}
14898 				}
14899 				if (rsm->r_in_tmap == 0) {
14900 					/*
14901 					 * Not found so shove it on the tail.
14902 					 */
14903 					TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
14904 					rsm->r_in_tmap = 1;
14905 				}
14906  			} else {
14907 				if ((rack->r_ctl.rc_sacklast == NULL) ||
14908 				    (SEQ_GT(rsm->r_end, rack->r_ctl.rc_sacklast->r_end))) {
14909 					rack->r_ctl.rc_sacklast = rsm;
14910 				}
14911 			}
14912 			rack_log_chg_info(tp, rack, 3,
14913 					  rsm->r_start,
14914 					  rsm->r_end,
14915 					  rsm->r_flags);
14916 		}
14917 	}
14918 	return (0);
14919 }
14920 
14921 static void
14922 rack_translate_clamp_value(struct tcp_rack *rack, uint32_t optval)
14923 {
14924 	/*
14925 	 * P = percent bits
14926 	 * F = fill cw bit -- Toggle fillcw if this bit is set.
14927 	 * S = Segment bits
14928 	 * M = set max segment bit
14929 	 * U = Unclamined
14930 	 * C = If set to non-zero override the max number of clamps.
14931 	 * L = Bit to indicate if clamped gets lower.
14932 	 *
14933 	 * CCCC CCCCC UUUU UULF PPPP PPPP PPPP PPPP
14934 	 *
14935 	 * The lowest 3 nibbles is the perentage .1 - 6553.5%
14936 	 * where 10.1 = 101, max 6553.5
14937 	 * The upper 16 bits  holds some options.
14938 	 * The F bit will turn on fill-cw on if you are
14939 	 * not pacing, it will turn it off if dgp is on.
14940 	 * The L bit will change it so when clamped we get
14941 	 * the min(gp, lt-bw) for dgp.
14942 	 */
14943 	uint16_t per;
14944 
14945 	rack->r_ctl.saved_rxt_clamp_val = optval;
14946 	per = optval & 0x0000ffff;
14947 	rack->r_ctl.rxt_threshold = (uint64_t)(per & 0xffff);
14948 	if (optval > 0) {
14949 		uint16_t clamp_opt;
14950 
14951 		rack->excess_rxt_on = 1;
14952 		clamp_opt = ((optval & 0xffff0000) >> 16);
14953 		rack->r_ctl.clamp_options = clamp_opt & 0x00ff;
14954 		if (clamp_opt & 0xff00) {
14955 			/* A max clamps is also present */
14956 			rack->r_ctl.max_clamps = (clamp_opt >> 8);
14957 		} else {
14958 			/* No specified clamps means no limit */
14959 			rack->r_ctl.max_clamps = 0;
14960 		}
14961 		if (rack->r_ctl.clamp_options & 0x0002) {
14962 			rack->r_clamped_gets_lower  = 1;
14963 		} else {
14964 			rack->r_clamped_gets_lower  = 0;
14965 		}
14966 	} else {
14967 		/* Turn it off back to default */
14968 		rack->excess_rxt_on = 0;
14969 		rack->r_clamped_gets_lower  = 0;
14970 	}
14971 
14972 }
14973 
14974 
14975 static int32_t
14976 rack_init(struct tcpcb *tp, void **ptr)
14977 {
14978 	struct inpcb *inp = tptoinpcb(tp);
14979 	struct tcp_rack *rack = NULL;
14980 	uint32_t iwin, snt, us_cts;
14981 	int err, no_query;
14982 
14983 	tcp_hpts_init(tp);
14984 
14985 	/*
14986 	 * First are we the initial or are we a switched stack?
14987 	 * If we are initing via tcp_newtcppcb the ptr passed
14988 	 * will be tp->t_fb_ptr. If its a stack switch that
14989 	 * has a previous stack we can query it will be a local
14990 	 * var that will in the end be set into t_fb_ptr.
14991 	 */
14992 	if (ptr == &tp->t_fb_ptr)
14993 		no_query = 1;
14994 	else
14995 		no_query = 0;
14996 	*ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
14997 	if (*ptr == NULL) {
14998 		/*
14999 		 * We need to allocate memory but cant. The INP and INP_INFO
15000 		 * locks and they are recursive (happens during setup. So a
15001 		 * scheme to drop the locks fails :(
15002 		 *
15003 		 */
15004 		return(ENOMEM);
15005 	}
15006 	memset(*ptr, 0, sizeof(struct tcp_rack));
15007 	rack = (struct tcp_rack *)*ptr;
15008 	rack->r_ctl.tqh = malloc(sizeof(struct tailq_hash), M_TCPFSB, M_NOWAIT);
15009 	if (rack->r_ctl.tqh == NULL) {
15010 		uma_zfree(rack_pcb_zone, rack);
15011 		return(ENOMEM);
15012 	}
15013 	tqhash_init(rack->r_ctl.tqh);
15014 	TAILQ_INIT(&rack->r_ctl.rc_free);
15015 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
15016 	rack->rc_tp = tp;
15017 	rack->rc_inp = inp;
15018 	/* Set the flag */
15019 	rack->r_is_v6 = (inp->inp_vflag & INP_IPV6) != 0;
15020 	/* Probably not needed but lets be sure */
15021 	rack_clear_rate_sample(rack);
15022 	/*
15023 	 * Save off the default values, socket options will poke
15024 	 * at these if pacing is not on or we have not yet
15025 	 * reached where pacing is on (gp_ready/fixed enabled).
15026 	 * When they get set into the CC module (when gp_ready
15027 	 * is enabled or we enable fixed) then we will set these
15028 	 * values into the CC and place in here the old values
15029 	 * so we have a restoral. Then we will set the flag
15030 	 * rc_pacing_cc_set. That way whenever we turn off pacing
15031 	 * or switch off this stack, we will know to go restore
15032 	 * the saved values.
15033 	 *
15034 	 * We specifically put into the beta the ecn value for pacing.
15035 	 */
15036 	rack->rc_new_rnd_needed = 1;
15037 	rack->r_ctl.rc_split_limit = V_tcp_map_split_limit;
15038 	/* We want abe like behavior as well */
15039 	rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED;
15040 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
15041 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
15042 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
15043 	if (rack_rxt_clamp_thresh) {
15044 		rack_translate_clamp_value(rack, rack_rxt_clamp_thresh);
15045 		rack->excess_rxt_on = 1;
15046 	}
15047 	if (rack_uses_full_dgp_in_rec)
15048 		rack->r_ctl.full_dgp_in_rec = 1;
15049 	if (rack_fill_cw_state)
15050 		rack->rc_pace_to_cwnd = 1;
15051 	if (rack_pacing_min_seg)
15052 		rack->r_ctl.rc_user_set_min_segs = rack_pacing_min_seg;
15053 	if (use_rack_rr)
15054 		rack->use_rack_rr = 1;
15055 	if (rack_dnd_default) {
15056 		rack->rc_pace_dnd = 1;
15057 	}
15058 	if (V_tcp_delack_enabled)
15059 		tp->t_delayed_ack = 1;
15060 	else
15061 		tp->t_delayed_ack = 0;
15062 #ifdef TCP_ACCOUNTING
15063 	if (rack_tcp_accounting) {
15064 		tp->t_flags2 |= TF2_TCP_ACCOUNTING;
15065 	}
15066 #endif
15067 	rack->r_ctl.rack_per_upper_bound_ss = (uint8_t)rack_per_upper_bound_ss;
15068 	rack->r_ctl.rack_per_upper_bound_ca = (uint8_t)rack_per_upper_bound_ca;
15069 	if (rack_enable_shared_cwnd)
15070 		rack->rack_enable_scwnd = 1;
15071 	rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
15072 	rack->rc_user_set_max_segs = rack_hptsi_segments;
15073 	rack->rc_force_max_seg = 0;
15074 	TAILQ_INIT(&rack->r_ctl.opt_list);
15075 	rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn;
15076 	rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn;
15077 	if (rack_hibeta_setting) {
15078 		rack->rack_hibeta = 1;
15079 		if ((rack_hibeta_setting >= 50) &&
15080 		    (rack_hibeta_setting <= 100)) {
15081 			rack->r_ctl.rc_saved_beta.beta = rack_hibeta_setting;
15082 			rack->r_ctl.saved_hibeta = rack_hibeta_setting;
15083 		}
15084 	} else {
15085 		rack->r_ctl.saved_hibeta = 50;
15086 	}
15087 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
15088 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
15089 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
15090 	rack->r_ctl.rc_lowest_us_rtt = 0xffffffff;
15091 	rack->r_ctl.rc_highest_us_rtt = 0;
15092 	rack->r_ctl.bw_rate_cap = rack_bw_rate_cap;
15093 	rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop);
15094 	if (rack_use_cmp_acks)
15095 		rack->r_use_cmp_ack = 1;
15096 	if (rack_disable_prr)
15097 		rack->rack_no_prr = 1;
15098 	if (rack_gp_no_rec_chg)
15099 		rack->rc_gp_no_rec_chg = 1;
15100 	if (rack_pace_every_seg && tcp_can_enable_pacing()) {
15101 		rack->rc_always_pace = 1;
15102 		if (rack->rack_hibeta)
15103 			rack_set_cc_pacing(rack);
15104 	} else
15105 		rack->rc_always_pace = 0;
15106 	if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack)
15107 		rack->r_mbuf_queue = 1;
15108 	else
15109 		rack->r_mbuf_queue = 0;
15110 	rack_set_pace_segments(tp, rack, __LINE__, NULL);
15111 	if (rack_limits_scwnd)
15112 		rack->r_limit_scw = 1;
15113 	else
15114 		rack->r_limit_scw = 0;
15115 	rack_init_retransmit_value(rack, rack_rxt_controls);
15116 	rack->rc_labc = V_tcp_abc_l_var;
15117 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
15118 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
15119 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
15120 	rack->r_ctl.rc_min_to = rack_min_to;
15121 	microuptime(&rack->r_ctl.act_rcv_time);
15122 	rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time;
15123 	rack->rc_init_win = rack_default_init_window;
15124 	rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss;
15125 	if (rack_hw_up_only)
15126 		rack->r_up_only = 1;
15127 	if (rack_do_dyn_mul) {
15128 		/* When dynamic adjustment is on CA needs to start at 100% */
15129 		rack->rc_gp_dyn_mul = 1;
15130 		if (rack_do_dyn_mul >= 100)
15131 			rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
15132 	} else
15133 		rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
15134 	rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec;
15135 	rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt;
15136 	rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time);
15137 	setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN,
15138 				rack_probertt_filter_life);
15139 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
15140 	rack->r_ctl.rc_lower_rtt_us_cts = us_cts;
15141 	rack->r_ctl.rc_time_of_last_probertt = us_cts;
15142 	rack->r_ctl.challenge_ack_ts = tcp_ts_getticks();
15143 	rack->r_ctl.rc_time_probertt_starts = 0;
15144 	if (rack_dsack_std_based & 0x1) {
15145 		/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
15146 		rack->rc_rack_tmr_std_based = 1;
15147 	}
15148 	if (rack_dsack_std_based & 0x2) {
15149 		/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
15150 		rack->rc_rack_use_dsack = 1;
15151 	}
15152 	/* We require at least one measurement, even if the sysctl is 0 */
15153 	if (rack_req_measurements)
15154 		rack->r_ctl.req_measurements = rack_req_measurements;
15155 	else
15156 		rack->r_ctl.req_measurements = 1;
15157 	if (rack_enable_hw_pacing)
15158 		rack->rack_hdw_pace_ena = 1;
15159 	if (rack_hw_rate_caps)
15160 		rack->r_rack_hw_rate_caps = 1;
15161 #ifdef TCP_SAD_DETECTION
15162 	rack->do_detection = 1;
15163 #else
15164 	rack->do_detection = 0;
15165 #endif
15166 	if (rack_non_rxt_use_cr)
15167 		rack->rack_rec_nonrxt_use_cr = 1;
15168 	/* Lets setup the fsb block */
15169 	err = rack_init_fsb(tp, rack);
15170 	if (err) {
15171 		uma_zfree(rack_pcb_zone, *ptr);
15172 		*ptr = NULL;
15173 		return (err);
15174 	}
15175 	if (rack_do_hystart) {
15176 		tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
15177 		if (rack_do_hystart > 1)
15178 			tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
15179 		if (rack_do_hystart > 2)
15180 			tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
15181 	}
15182 	/* Log what we will do with queries */
15183 	rack_log_chg_info(tp, rack, 7,
15184 			  no_query, 0, 0);
15185 	if (rack_def_profile)
15186 		rack_set_profile(rack, rack_def_profile);
15187 	/* Cancel the GP measurement in progress */
15188 	tp->t_flags &= ~TF_GPUTINPROG;
15189 	if ((tp->t_state != TCPS_CLOSED) &&
15190 	    (tp->t_state != TCPS_TIME_WAIT)) {
15191 		/*
15192 		 * We are already open, we may
15193 		 * need to adjust a few things.
15194 		 */
15195 		if (SEQ_GT(tp->snd_max, tp->iss))
15196 			snt = tp->snd_max - tp->iss;
15197 		else
15198 			snt = 0;
15199 		iwin = rc_init_window(rack);
15200 		if ((snt < iwin) &&
15201 		    (no_query == 1)) {
15202 			/* We are not past the initial window
15203 			 * on the first init (i.e. a stack switch
15204 			 * has not yet occured) so we need to make
15205 			 * sure cwnd and ssthresh is correct.
15206 			 */
15207 			if (tp->snd_cwnd < iwin)
15208 				tp->snd_cwnd = iwin;
15209 			/*
15210 			 * If we are within the initial window
15211 			 * we want ssthresh to be unlimited. Setting
15212 			 * it to the rwnd (which the default stack does
15213 			 * and older racks) is not really a good idea
15214 			 * since we want to be in SS and grow both the
15215 			 * cwnd and the rwnd (via dynamic rwnd growth). If
15216 			 * we set it to the rwnd then as the peer grows its
15217 			 * rwnd we will be stuck in CA and never hit SS.
15218 			 *
15219 			 * Its far better to raise it up high (this takes the
15220 			 * risk that there as been a loss already, probably
15221 			 * we should have an indicator in all stacks of loss
15222 			 * but we don't), but considering the normal use this
15223 			 * is a risk worth taking. The consequences of not
15224 			 * hitting SS are far worse than going one more time
15225 			 * into it early on (before we have sent even a IW).
15226 			 * It is highly unlikely that we will have had a loss
15227 			 * before getting the IW out.
15228 			 */
15229 			tp->snd_ssthresh = 0xffffffff;
15230 		}
15231 		/*
15232 		 * Any init based on sequence numbers
15233 		 * should be done in the deferred init path
15234 		 * since we can be CLOSED and not have them
15235 		 * inited when rack_init() is called. We
15236 		 * are not closed so lets call it.
15237 		 */
15238 		rack_deferred_init(tp, rack);
15239 	}
15240 	if ((tp->t_state != TCPS_CLOSED) &&
15241 	    (tp->t_state != TCPS_TIME_WAIT) &&
15242 	    (no_query == 0) &&
15243 	    (tp->snd_una != tp->snd_max))  {
15244 		err = rack_init_outstanding(tp, rack, us_cts, *ptr);
15245 		if (err) {
15246 			*ptr = NULL;
15247 			return(err);
15248 		}
15249 	}
15250 	rack_stop_all_timers(tp, rack);
15251 	/* Setup all the t_flags2 */
15252 	if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
15253 		tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
15254 	else
15255 		tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
15256 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15257 		tp->t_flags2 |= TF2_MBUF_ACKCMP;
15258 	/*
15259 	 * Timers in Rack are kept in microseconds so lets
15260 	 * convert any initial incoming variables
15261 	 * from ticks into usecs. Note that we
15262 	 * also change the values of t_srtt and t_rttvar, if
15263 	 * they are non-zero. They are kept with a 5
15264 	 * bit decimal so we have to carefully convert
15265 	 * these to get the full precision.
15266 	 */
15267 	rack_convert_rtts(tp);
15268 	rack_log_hystart_event(rack, rack->r_ctl.roundends, 20);
15269 	if ((tptoinpcb(tp)->inp_flags & INP_DROPPED) == 0) {
15270 		/* We do not start any timers on DROPPED connections */
15271 		if (tp->t_fb->tfb_chg_query == NULL) {
15272 			rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15273 		} else {
15274 			struct tcp_query_resp qr;
15275 			int ret;
15276 
15277 			memset(&qr, 0, sizeof(qr));
15278 
15279 			/* Get the misc time stamps and such for rack */
15280 			qr.req = TCP_QUERY_RACK_TIMES;
15281 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
15282 			if (ret == 1) {
15283 				rack->r_ctl.rc_reorder_ts = qr.rack_reorder_ts;
15284 				rack->r_ctl.num_dsack  = qr.rack_num_dsacks;
15285 				rack->r_ctl.rc_tlp_rxt_last_time = qr.rack_rxt_last_time;
15286 				rack->r_ctl.rc_rack_min_rtt = qr.rack_min_rtt;
15287 				rack->rc_rack_rtt = qr.rack_rtt;
15288 				rack->r_ctl.rc_rack_tmit_time = qr.rack_tmit_time;
15289 				rack->r_ctl.rc_sacked = qr.rack_sacked;
15290 				rack->r_ctl.rc_holes_rxt = qr.rack_holes_rxt;
15291 				rack->r_ctl.rc_prr_delivered = qr.rack_prr_delivered;
15292 				rack->r_ctl.rc_prr_recovery_fs = qr.rack_prr_recovery_fs;
15293 				rack->r_ctl.rc_prr_sndcnt = qr.rack_prr_sndcnt;
15294 				rack->r_ctl.rc_prr_out = qr.rack_prr_out;
15295 				if (qr.rack_tlp_out) {
15296 					rack->rc_tlp_in_progress = 1;
15297 					rack->r_ctl.rc_tlp_cnt_out = qr.rack_tlp_cnt_out;
15298 				} else {
15299 					rack->rc_tlp_in_progress = 0;
15300 					rack->r_ctl.rc_tlp_cnt_out = 0;
15301 				}
15302 				if (qr.rack_srtt_measured)
15303 					rack->rc_srtt_measure_made = 1;
15304 				if (qr.rack_in_persist == 1) {
15305 					rack->r_ctl.rc_went_idle_time = qr.rack_time_went_idle;
15306 #ifdef NETFLIX_SHARED_CWND
15307 					if (rack->r_ctl.rc_scw) {
15308 						tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
15309 						rack->rack_scwnd_is_idle = 1;
15310 					}
15311 #endif
15312 					rack->r_ctl.persist_lost_ends = 0;
15313 					rack->probe_not_answered = 0;
15314 					rack->forced_ack = 0;
15315 					tp->t_rxtshift = 0;
15316 					rack->rc_in_persist = 1;
15317 					RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
15318 							   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
15319 				}
15320 				if (qr.rack_wanted_output)
15321 					rack->r_wanted_output = 1;
15322 				rack_log_chg_info(tp, rack, 6,
15323 						  qr.rack_min_rtt,
15324 						  qr.rack_rtt,
15325 						  qr.rack_reorder_ts);
15326 			}
15327 			/* Get the old stack timers */
15328 			qr.req_param = 0;
15329 			qr.req = TCP_QUERY_TIMERS_UP;
15330 			ret = (*tp->t_fb->tfb_chg_query)(tp, &qr);
15331 			if (ret) {
15332 				/*
15333 				 * non-zero return means we have a timer('s)
15334 				 * to start. Zero means no timer (no keepalive
15335 				 * I suppose).
15336 				 */
15337 				uint32_t tov = 0;
15338 
15339 				rack->r_ctl.rc_hpts_flags = qr.timer_hpts_flags;
15340 				if (qr.timer_hpts_flags & PACE_PKT_OUTPUT) {
15341 					rack->r_ctl.rc_last_output_to = qr.timer_pacing_to;
15342 					if (TSTMP_GT(qr.timer_pacing_to, us_cts))
15343 						tov = qr.timer_pacing_to - us_cts;
15344 					else
15345 						tov = HPTS_TICKS_PER_SLOT;
15346 				}
15347 				if (qr.timer_hpts_flags & PACE_TMR_MASK) {
15348 					rack->r_ctl.rc_timer_exp = qr.timer_timer_exp;
15349 					if (tov == 0) {
15350 						if (TSTMP_GT(qr.timer_timer_exp, us_cts))
15351 							tov = qr.timer_timer_exp - us_cts;
15352 						else
15353 							tov = HPTS_TICKS_PER_SLOT;
15354 					}
15355 				}
15356 				rack_log_chg_info(tp, rack, 4,
15357 						  rack->r_ctl.rc_hpts_flags,
15358 						  rack->r_ctl.rc_last_output_to,
15359 						  rack->r_ctl.rc_timer_exp);
15360 				if (tov) {
15361 					struct hpts_diag diag;
15362 
15363 					(void)tcp_hpts_insert_diag(tp, HPTS_USEC_TO_SLOTS(tov),
15364 								   __LINE__, &diag);
15365 					rack_log_hpts_diag(rack, us_cts, &diag, &rack->r_ctl.act_rcv_time);
15366 				}
15367 			}
15368 		}
15369 		rack_log_rtt_shrinks(rack,  us_cts,  tp->t_rxtcur,
15370 				     __LINE__, RACK_RTTS_INIT);
15371 	}
15372 	return (0);
15373 }
15374 
15375 static int
15376 rack_handoff_ok(struct tcpcb *tp)
15377 {
15378 	if ((tp->t_state == TCPS_CLOSED) ||
15379 	    (tp->t_state == TCPS_LISTEN)) {
15380 		/* Sure no problem though it may not stick */
15381 		return (0);
15382 	}
15383 	if ((tp->t_state == TCPS_SYN_SENT) ||
15384 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
15385 		/*
15386 		 * We really don't know if you support sack,
15387 		 * you have to get to ESTAB or beyond to tell.
15388 		 */
15389 		return (EAGAIN);
15390 	}
15391 	if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) {
15392 		/*
15393 		 * Rack will only send a FIN after all data is acknowledged.
15394 		 * So in this case we have more data outstanding. We can't
15395 		 * switch stacks until either all data and only the FIN
15396 		 * is left (in which case rack_init() now knows how
15397 		 * to deal with that) <or> all is acknowledged and we
15398 		 * are only left with incoming data, though why you
15399 		 * would want to switch to rack after all data is acknowledged
15400 		 * I have no idea (rrs)!
15401 		 */
15402 		return (EAGAIN);
15403 	}
15404 	if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){
15405 		return (0);
15406 	}
15407 	/*
15408 	 * If we reach here we don't do SACK on this connection so we can
15409 	 * never do rack.
15410 	 */
15411 	return (EINVAL);
15412 }
15413 
15414 static void
15415 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
15416 {
15417 
15418 	if (tp->t_fb_ptr) {
15419 		uint32_t cnt_free = 0;
15420 		struct tcp_rack *rack;
15421 		struct rack_sendmap *rsm;
15422 
15423 		tcp_handle_orphaned_packets(tp);
15424 		tp->t_flags &= ~TF_FORCEDATA;
15425 		rack = (struct tcp_rack *)tp->t_fb_ptr;
15426 		rack_log_pacing_delay_calc(rack,
15427 					   0,
15428 					   0,
15429 					   0,
15430 					   rack_get_gp_est(rack), /* delRate */
15431 					   rack_get_lt_bw(rack), /* rttProp */
15432 					   20, __LINE__, NULL, 0);
15433 #ifdef NETFLIX_SHARED_CWND
15434 		if (rack->r_ctl.rc_scw) {
15435 			uint32_t limit;
15436 
15437 			if (rack->r_limit_scw)
15438 				limit = max(1, rack->r_ctl.rc_lowest_us_rtt);
15439 			else
15440 				limit = 0;
15441 			tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw,
15442 						  rack->r_ctl.rc_scw_index,
15443 						  limit);
15444 			rack->r_ctl.rc_scw = NULL;
15445 		}
15446 #endif
15447 		if (rack->r_ctl.fsb.tcp_ip_hdr) {
15448 			free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB);
15449 			rack->r_ctl.fsb.tcp_ip_hdr = NULL;
15450 			rack->r_ctl.fsb.th = NULL;
15451 		}
15452 		if (rack->rc_always_pace) {
15453 			tcp_decrement_paced_conn();
15454 			rack_undo_cc_pacing(rack);
15455 			rack->rc_always_pace = 0;
15456 		}
15457 		/* Clean up any options if they were not applied */
15458 		while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) {
15459 			struct deferred_opt_list *dol;
15460 
15461 			dol = TAILQ_FIRST(&rack->r_ctl.opt_list);
15462 			TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
15463 			free(dol, M_TCPDO);
15464 		}
15465 		/* rack does not use force data but other stacks may clear it */
15466 		if (rack->r_ctl.crte != NULL) {
15467 			tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
15468 			rack->rack_hdrw_pacing = 0;
15469 			rack->r_ctl.crte = NULL;
15470 		}
15471 #ifdef TCP_BLACKBOX
15472 		tcp_log_flowend(tp);
15473 #endif
15474 		/*
15475 		 * Lets take a different approach to purging just
15476 		 * get each one and free it like a cum-ack would and
15477 		 * not use a foreach loop.
15478 		 */
15479 		rsm = tqhash_min(rack->r_ctl.tqh);
15480 		while (rsm) {
15481 			tqhash_remove(rack->r_ctl.tqh, rsm, REMOVE_TYPE_CUMACK);
15482 			rack->r_ctl.rc_num_maps_alloced--;
15483 			uma_zfree(rack_zone, rsm);
15484 			rsm = tqhash_min(rack->r_ctl.tqh);
15485 		}
15486 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15487 		while (rsm) {
15488 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext);
15489 			rack->r_ctl.rc_num_maps_alloced--;
15490 			rack->rc_free_cnt--;
15491 			cnt_free++;
15492 			uma_zfree(rack_zone, rsm);
15493 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15494 		}
15495 		if ((rack->r_ctl.rc_num_maps_alloced > 0) &&
15496 		    (tcp_bblogging_on(tp))) {
15497 			union tcp_log_stackspecific log;
15498 			struct timeval tv;
15499 
15500 			memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15501 			log.u_bbr.flex8 = 10;
15502 			log.u_bbr.flex1 = rack->r_ctl.rc_num_maps_alloced;
15503 			log.u_bbr.flex2 = rack->rc_free_cnt;
15504 			log.u_bbr.flex3 = cnt_free;
15505 			log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
15506 			rsm = tqhash_min(rack->r_ctl.tqh);
15507 			log.u_bbr.delRate = (uint64_t)rsm;
15508 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
15509 			log.u_bbr.cur_del_rate = (uint64_t)rsm;
15510 			log.u_bbr.timeStamp = tcp_get_usecs(&tv);
15511 			log.u_bbr.pkt_epoch = __LINE__;
15512 			(void)tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
15513 					     0, &log, false, NULL, NULL, 0, &tv);
15514 		}
15515 		KASSERT((rack->r_ctl.rc_num_maps_alloced == 0),
15516 			("rack:%p num_aloc:%u after freeing all?",
15517 			 rack,
15518 			 rack->r_ctl.rc_num_maps_alloced));
15519 		rack->rc_free_cnt = 0;
15520 		free(rack->r_ctl.tqh, M_TCPFSB);
15521 		rack->r_ctl.tqh = NULL;
15522 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
15523 		tp->t_fb_ptr = NULL;
15524 	}
15525 	/* Make sure snd_nxt is correctly set */
15526 	tp->snd_nxt = tp->snd_max;
15527 }
15528 
15529 static void
15530 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
15531 {
15532 	if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) {
15533 		rack->r_is_v6 = (tptoinpcb(tp)->inp_vflag & INP_IPV6) != 0;
15534 	}
15535 	switch (tp->t_state) {
15536 	case TCPS_SYN_SENT:
15537 		rack->r_state = TCPS_SYN_SENT;
15538 		rack->r_substate = rack_do_syn_sent;
15539 		break;
15540 	case TCPS_SYN_RECEIVED:
15541 		rack->r_state = TCPS_SYN_RECEIVED;
15542 		rack->r_substate = rack_do_syn_recv;
15543 		break;
15544 	case TCPS_ESTABLISHED:
15545 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15546 		rack->r_state = TCPS_ESTABLISHED;
15547 		rack->r_substate = rack_do_established;
15548 		break;
15549 	case TCPS_CLOSE_WAIT:
15550 		rack->r_state = TCPS_CLOSE_WAIT;
15551 		rack->r_substate = rack_do_close_wait;
15552 		break;
15553 	case TCPS_FIN_WAIT_1:
15554 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15555 		rack->r_state = TCPS_FIN_WAIT_1;
15556 		rack->r_substate = rack_do_fin_wait_1;
15557 		break;
15558 	case TCPS_CLOSING:
15559 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15560 		rack->r_state = TCPS_CLOSING;
15561 		rack->r_substate = rack_do_closing;
15562 		break;
15563 	case TCPS_LAST_ACK:
15564 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
15565 		rack->r_state = TCPS_LAST_ACK;
15566 		rack->r_substate = rack_do_lastack;
15567 		break;
15568 	case TCPS_FIN_WAIT_2:
15569 		rack->r_state = TCPS_FIN_WAIT_2;
15570 		rack->r_substate = rack_do_fin_wait_2;
15571 		break;
15572 	case TCPS_LISTEN:
15573 	case TCPS_CLOSED:
15574 	case TCPS_TIME_WAIT:
15575 	default:
15576 		break;
15577 	};
15578 	if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
15579 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
15580 
15581 }
15582 
15583 static void
15584 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
15585 {
15586 	/*
15587 	 * We received an ack, and then did not
15588 	 * call send or were bounced out due to the
15589 	 * hpts was running. Now a timer is up as well, is
15590 	 * it the right timer?
15591 	 */
15592 	struct rack_sendmap *rsm;
15593 	int tmr_up;
15594 
15595 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
15596 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
15597 		return;
15598 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
15599 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
15600 	    (tmr_up == PACE_TMR_RXT)) {
15601 		/* Should be an RXT */
15602 		return;
15603 	}
15604 	if (rsm == NULL) {
15605 		/* Nothing outstanding? */
15606 		if (tp->t_flags & TF_DELACK) {
15607 			if (tmr_up == PACE_TMR_DELACK)
15608 				/* We are supposed to have delayed ack up and we do */
15609 				return;
15610 		} else if (sbavail(&tptosocket(tp)->so_snd) && (tmr_up == PACE_TMR_RXT)) {
15611 			/*
15612 			 * if we hit enobufs then we would expect the possibility
15613 			 * of nothing outstanding and the RXT up (and the hptsi timer).
15614 			 */
15615 			return;
15616 		} else if (((V_tcp_always_keepalive ||
15617 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
15618 			    (tp->t_state <= TCPS_CLOSING)) &&
15619 			   (tmr_up == PACE_TMR_KEEP) &&
15620 			   (tp->snd_max == tp->snd_una)) {
15621 			/* We should have keep alive up and we do */
15622 			return;
15623 		}
15624 	}
15625 	if (SEQ_GT(tp->snd_max, tp->snd_una) &&
15626 		   ((tmr_up == PACE_TMR_TLP) ||
15627 		    (tmr_up == PACE_TMR_RACK) ||
15628 		    (tmr_up == PACE_TMR_RXT))) {
15629 		/*
15630 		 * Either a Rack, TLP or RXT is fine if  we
15631 		 * have outstanding data.
15632 		 */
15633 		return;
15634 	} else if (tmr_up == PACE_TMR_DELACK) {
15635 		/*
15636 		 * If the delayed ack was going to go off
15637 		 * before the rtx/tlp/rack timer were going to
15638 		 * expire, then that would be the timer in control.
15639 		 * Note we don't check the time here trusting the
15640 		 * code is correct.
15641 		 */
15642 		return;
15643 	}
15644 	/*
15645 	 * Ok the timer originally started is not what we want now.
15646 	 * We will force the hpts to be stopped if any, and restart
15647 	 * with the slot set to what was in the saved slot.
15648 	 */
15649 	if (tcp_in_hpts(rack->rc_tp)) {
15650 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
15651 			uint32_t us_cts;
15652 
15653 			us_cts = tcp_get_usecs(NULL);
15654 			if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
15655 				rack->r_early = 1;
15656 				rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
15657 			}
15658 			rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
15659 		}
15660 		tcp_hpts_remove(rack->rc_tp);
15661 	}
15662 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
15663 	rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
15664 }
15665 
15666 
15667 static void
15668 rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts)
15669 {
15670 	if ((SEQ_LT(tp->snd_wl1, seq) ||
15671 	    (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) ||
15672 	    (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) {
15673 		/* keep track of pure window updates */
15674 		if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd))
15675 			KMOD_TCPSTAT_INC(tcps_rcvwinupd);
15676 		tp->snd_wnd = tiwin;
15677 		rack_validate_fo_sendwin_up(tp, rack);
15678 		tp->snd_wl1 = seq;
15679 		tp->snd_wl2 = ack;
15680 		if (tp->snd_wnd > tp->max_sndwnd)
15681 			tp->max_sndwnd = tp->snd_wnd;
15682 	    rack->r_wanted_output = 1;
15683 	} else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) {
15684 		tp->snd_wnd = tiwin;
15685 		rack_validate_fo_sendwin_up(tp, rack);
15686 		tp->snd_wl1 = seq;
15687 		tp->snd_wl2 = ack;
15688 	} else {
15689 		/* Not a valid win update */
15690 		return;
15691 	}
15692 	if (tp->snd_wnd > tp->max_sndwnd)
15693 		tp->max_sndwnd = tp->snd_wnd;
15694 	/* Do we exit persists? */
15695 	if ((rack->rc_in_persist != 0) &&
15696 	    (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2),
15697 				rack->r_ctl.rc_pace_min_segs))) {
15698 		rack_exit_persist(tp, rack, cts);
15699 	}
15700 	/* Do we enter persists? */
15701 	if ((rack->rc_in_persist == 0) &&
15702 	    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) &&
15703 	    TCPS_HAVEESTABLISHED(tp->t_state) &&
15704 	    ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) &&
15705 	    sbavail(&tptosocket(tp)->so_snd) &&
15706 	    (sbavail(&tptosocket(tp)->so_snd) > tp->snd_wnd)) {
15707 		/*
15708 		 * Here the rwnd is less than
15709 		 * the pacing size, we are established,
15710 		 * nothing is outstanding, and there is
15711 		 * data to send. Enter persists.
15712 		 */
15713 		rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, ack);
15714 	}
15715 }
15716 
15717 static void
15718 rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq)
15719 {
15720 
15721 	if (tcp_bblogging_on(rack->rc_tp)) {
15722 		struct inpcb *inp = tptoinpcb(tp);
15723 		union tcp_log_stackspecific log;
15724 		struct timeval ltv;
15725 		char tcp_hdr_buf[60];
15726 		struct tcphdr *th;
15727 		struct timespec ts;
15728 		uint32_t orig_snd_una;
15729 		uint8_t xx = 0;
15730 
15731 #ifdef TCP_REQUEST_TRK
15732 		struct tcp_sendfile_track *tcp_req;
15733 
15734 		if (SEQ_GT(ae->ack, tp->snd_una)) {
15735 			tcp_req = tcp_req_find_req_for_seq(tp, (ae->ack-1));
15736 		} else {
15737 			tcp_req = tcp_req_find_req_for_seq(tp, ae->ack);
15738 		}
15739 #endif
15740 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
15741 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
15742 		if (rack->rack_no_prr == 0)
15743 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
15744 		else
15745 			log.u_bbr.flex1 = 0;
15746 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
15747 		log.u_bbr.use_lt_bw <<= 1;
15748 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
15749 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
15750 		log.u_bbr.bbr_state = rack->rc_free_cnt;
15751 		log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
15752 		log.u_bbr.pkts_out = tp->t_maxseg;
15753 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
15754 		log.u_bbr.flex7 = 1;
15755 		log.u_bbr.lost = ae->flags;
15756 		log.u_bbr.cwnd_gain = ackval;
15757 		log.u_bbr.pacing_gain = 0x2;
15758 		if (ae->flags & TSTMP_HDWR) {
15759 			/* Record the hardware timestamp if present */
15760 			log.u_bbr.flex3 = M_TSTMP;
15761 			ts.tv_sec = ae->timestamp / 1000000000;
15762 			ts.tv_nsec = ae->timestamp % 1000000000;
15763 			ltv.tv_sec = ts.tv_sec;
15764 			ltv.tv_usec = ts.tv_nsec / 1000;
15765 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
15766 		} else if (ae->flags & TSTMP_LRO) {
15767 			/* Record the LRO the arrival timestamp */
15768 			log.u_bbr.flex3 = M_TSTMP_LRO;
15769 			ts.tv_sec = ae->timestamp / 1000000000;
15770 			ts.tv_nsec = ae->timestamp % 1000000000;
15771 			ltv.tv_sec = ts.tv_sec;
15772 			ltv.tv_usec = ts.tv_nsec / 1000;
15773 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
15774 		}
15775 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
15776 		/* Log the rcv time */
15777 		log.u_bbr.delRate = ae->timestamp;
15778 #ifdef TCP_REQUEST_TRK
15779 		log.u_bbr.applimited = tp->t_tcpreq_closed;
15780 		log.u_bbr.applimited <<= 8;
15781 		log.u_bbr.applimited |= tp->t_tcpreq_open;
15782 		log.u_bbr.applimited <<= 8;
15783 		log.u_bbr.applimited |= tp->t_tcpreq_req;
15784 		if (tcp_req) {
15785 			/* Copy out any client req info */
15786 			/* seconds */
15787 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
15788 			/* useconds */
15789 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
15790 			log.u_bbr.rttProp = tcp_req->timestamp;
15791 			log.u_bbr.cur_del_rate = tcp_req->start;
15792 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
15793 				log.u_bbr.flex8 |= 1;
15794 			} else {
15795 				log.u_bbr.flex8 |= 2;
15796 				log.u_bbr.bw_inuse = tcp_req->end;
15797 			}
15798 			log.u_bbr.flex6 = tcp_req->start_seq;
15799 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
15800 				log.u_bbr.flex8 |= 4;
15801 				log.u_bbr.epoch = tcp_req->end_seq;
15802 			}
15803 		}
15804 #endif
15805 		memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf));
15806 		th = (struct tcphdr *)tcp_hdr_buf;
15807 		th->th_seq = ae->seq;
15808 		th->th_ack = ae->ack;
15809 		th->th_win = ae->win;
15810 		/* Now fill in the ports */
15811 		th->th_sport = inp->inp_fport;
15812 		th->th_dport = inp->inp_lport;
15813 		tcp_set_flags(th, ae->flags);
15814 		/* Now do we have a timestamp option? */
15815 		if (ae->flags & HAS_TSTMP) {
15816 			u_char *cp;
15817 			uint32_t val;
15818 
15819 			th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2);
15820 			cp = (u_char *)(th + 1);
15821 			*cp = TCPOPT_NOP;
15822 			cp++;
15823 			*cp = TCPOPT_NOP;
15824 			cp++;
15825 			*cp = TCPOPT_TIMESTAMP;
15826 			cp++;
15827 			*cp = TCPOLEN_TIMESTAMP;
15828 			cp++;
15829 			val = htonl(ae->ts_value);
15830 			bcopy((char *)&val,
15831 			      (char *)cp, sizeof(uint32_t));
15832 			val = htonl(ae->ts_echo);
15833 			bcopy((char *)&val,
15834 			      (char *)(cp + 4), sizeof(uint32_t));
15835 		} else
15836 			th->th_off = (sizeof(struct tcphdr) >> 2);
15837 
15838 		/*
15839 		 * For sane logging we need to play a little trick.
15840 		 * If the ack were fully processed we would have moved
15841 		 * snd_una to high_seq, but since compressed acks are
15842 		 * processed in two phases, at this point (logging) snd_una
15843 		 * won't be advanced. So we would see multiple acks showing
15844 		 * the advancement. We can prevent that by "pretending" that
15845 		 * snd_una was advanced and then un-advancing it so that the
15846 		 * logging code has the right value for tlb_snd_una.
15847 		 */
15848 		if (tp->snd_una != high_seq) {
15849 			orig_snd_una = tp->snd_una;
15850 			tp->snd_una = high_seq;
15851 			xx = 1;
15852 		} else
15853 			xx = 0;
15854 		TCP_LOG_EVENTP(tp, th,
15855 			       &tptosocket(tp)->so_rcv,
15856 			       &tptosocket(tp)->so_snd, TCP_LOG_IN, 0,
15857 			       0, &log, true, &ltv);
15858 		if (xx) {
15859 			tp->snd_una = orig_snd_una;
15860 		}
15861 	}
15862 
15863 }
15864 
15865 static void
15866 rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts)
15867 {
15868 	uint32_t us_rtt;
15869 	/*
15870 	 * A persist or keep-alive was forced out, update our
15871 	 * min rtt time. Note now worry about lost responses.
15872 	 * When a subsequent keep-alive or persist times out
15873 	 * and forced_ack is still on, then the last probe
15874 	 * was not responded to. In such cases we have a
15875 	 * sysctl that controls the behavior. Either we apply
15876 	 * the rtt but with reduced confidence (0). Or we just
15877 	 * plain don't apply the rtt estimate. Having data flow
15878 	 * will clear the probe_not_answered flag i.e. cum-ack
15879 	 * move forward <or> exiting and reentering persists.
15880 	 */
15881 
15882 	rack->forced_ack = 0;
15883 	rack->rc_tp->t_rxtshift = 0;
15884 	if ((rack->rc_in_persist &&
15885 	     (tiwin == rack->rc_tp->snd_wnd)) ||
15886 	    (rack->rc_in_persist == 0)) {
15887 		/*
15888 		 * In persists only apply the RTT update if this is
15889 		 * a response to our window probe. And that
15890 		 * means the rwnd sent must match the current
15891 		 * snd_wnd. If it does not, then we got a
15892 		 * window update ack instead. For keepalive
15893 		 * we allow the answer no matter what the window.
15894 		 *
15895 		 * Note that if the probe_not_answered is set then
15896 		 * the forced_ack_ts is the oldest one i.e. the first
15897 		 * probe sent that might have been lost. This assures
15898 		 * us that if we do calculate an RTT it is longer not
15899 		 * some short thing.
15900 		 */
15901 		if (rack->rc_in_persist)
15902 			counter_u64_add(rack_persists_acks, 1);
15903 		us_rtt = us_cts - rack->r_ctl.forced_ack_ts;
15904 		if (us_rtt == 0)
15905 			us_rtt = 1;
15906 		if (rack->probe_not_answered == 0) {
15907 			rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15908 			tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1);
15909 		} else {
15910 			/* We have a retransmitted probe here too */
15911 			if (rack_apply_rtt_with_reduced_conf) {
15912 				rack_apply_updated_usrtt(rack, us_rtt, us_cts);
15913 				tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1);
15914 			}
15915 		}
15916 	}
15917 }
15918 
15919 static int
15920 rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv)
15921 {
15922 	/*
15923 	 * Handle a "special" compressed ack mbuf. Each incoming
15924 	 * ack has only four possible dispositions:
15925 	 *
15926 	 * A) It moves the cum-ack forward
15927 	 * B) It is behind the cum-ack.
15928 	 * C) It is a window-update ack.
15929 	 * D) It is a dup-ack.
15930 	 *
15931 	 * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES
15932 	 * in the incoming mbuf. We also need to still pay attention
15933 	 * to nxt_pkt since there may be another packet after this
15934 	 * one.
15935 	 */
15936 #ifdef TCP_ACCOUNTING
15937 	uint64_t ts_val;
15938 	uint64_t rdstc;
15939 #endif
15940 	int segsiz;
15941 	struct timespec ts;
15942 	struct tcp_rack *rack;
15943 	struct tcp_ackent *ae;
15944 	uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack;
15945 	int cnt, i, did_out, ourfinisacked = 0;
15946 	struct tcpopt to_holder, *to = NULL;
15947 #ifdef TCP_ACCOUNTING
15948 	int win_up_req = 0;
15949 #endif
15950 	int nsegs = 0;
15951 	int under_pacing = 0;
15952 	int recovery = 0;
15953 #ifdef TCP_ACCOUNTING
15954 	sched_pin();
15955 #endif
15956 	rack = (struct tcp_rack *)tp->t_fb_ptr;
15957 	if (rack->gp_ready &&
15958 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT))
15959 		under_pacing = 1;
15960 
15961 	if (rack->r_state != tp->t_state)
15962 		rack_set_state(tp, rack);
15963 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
15964 	    (tp->t_flags & TF_GPUTINPROG)) {
15965 		/*
15966 		 * We have a goodput in progress
15967 		 * and we have entered a late state.
15968 		 * Do we have enough data in the sb
15969 		 * to handle the GPUT request?
15970 		 */
15971 		uint32_t bytes;
15972 
15973 		bytes = tp->gput_ack - tp->gput_seq;
15974 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
15975 			bytes += tp->gput_seq - tp->snd_una;
15976 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
15977 			/*
15978 			 * There are not enough bytes in the socket
15979 			 * buffer that have been sent to cover this
15980 			 * measurement. Cancel it.
15981 			 */
15982 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
15983 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
15984 						   tp->gput_seq,
15985 						   0, 0, 18, __LINE__, NULL, 0);
15986 			tp->t_flags &= ~TF_GPUTINPROG;
15987 		}
15988 	}
15989 	to = &to_holder;
15990 	to->to_flags = 0;
15991 	KASSERT((m->m_len >= sizeof(struct tcp_ackent)),
15992 		("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len));
15993 	cnt = m->m_len / sizeof(struct tcp_ackent);
15994 	counter_u64_add(rack_multi_single_eq, cnt);
15995 	high_seq = tp->snd_una;
15996 	the_win = tp->snd_wnd;
15997 	win_seq = tp->snd_wl1;
15998 	win_upd_ack = tp->snd_wl2;
15999 	cts = tcp_tv_to_usectick(tv);
16000 	ms_cts = tcp_tv_to_mssectick(tv);
16001 	rack->r_ctl.rc_rcvtime = cts;
16002 	segsiz = ctf_fixed_maxseg(tp);
16003 	if ((rack->rc_gp_dyn_mul) &&
16004 	    (rack->use_fixed_rate == 0) &&
16005 	    (rack->rc_always_pace)) {
16006 		/* Check in on probertt */
16007 		rack_check_probe_rtt(rack, cts);
16008 	}
16009 	for (i = 0; i < cnt; i++) {
16010 #ifdef TCP_ACCOUNTING
16011 		ts_val = get_cyclecount();
16012 #endif
16013 		rack_clear_rate_sample(rack);
16014 		ae = ((mtod(m, struct tcp_ackent *)) + i);
16015 		if (ae->flags & TH_FIN)
16016 			rack_log_pacing_delay_calc(rack,
16017 						   0,
16018 						   0,
16019 						   0,
16020 						   rack_get_gp_est(rack), /* delRate */
16021 						   rack_get_lt_bw(rack), /* rttProp */
16022 						   20, __LINE__, NULL, 0);
16023 		/* Setup the window */
16024 		tiwin = ae->win << tp->snd_scale;
16025 		if (tiwin > rack->r_ctl.rc_high_rwnd)
16026 			rack->r_ctl.rc_high_rwnd = tiwin;
16027 		/* figure out the type of ack */
16028 		if (SEQ_LT(ae->ack, high_seq)) {
16029 			/* Case B*/
16030 			ae->ack_val_set = ACK_BEHIND;
16031 		} else if (SEQ_GT(ae->ack, high_seq)) {
16032 			/* Case A */
16033 			ae->ack_val_set = ACK_CUMACK;
16034 		} else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){
16035 			/* Case D */
16036 			ae->ack_val_set = ACK_DUPACK;
16037 		} else {
16038 			/* Case C */
16039 			ae->ack_val_set = ACK_RWND;
16040 		}
16041 		if (rack->sack_attack_disable > 0) {
16042 			rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
16043 			rack->r_ctl.ack_during_sd++;
16044 		}
16045 		rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq);
16046 		/* Validate timestamp */
16047 		if (ae->flags & HAS_TSTMP) {
16048 			/* Setup for a timestamp */
16049 			to->to_flags = TOF_TS;
16050 			ae->ts_echo -= tp->ts_offset;
16051 			to->to_tsecr = ae->ts_echo;
16052 			to->to_tsval = ae->ts_value;
16053 			/*
16054 			 * If echoed timestamp is later than the current time, fall back to
16055 			 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
16056 			 * were used when this connection was established.
16057 			 */
16058 			if (TSTMP_GT(ae->ts_echo, ms_cts))
16059 				to->to_tsecr = 0;
16060 			if (tp->ts_recent &&
16061 			    TSTMP_LT(ae->ts_value, tp->ts_recent)) {
16062 				if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) {
16063 #ifdef TCP_ACCOUNTING
16064 					rdstc = get_cyclecount();
16065 					if (rdstc > ts_val) {
16066 						if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16067 							tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
16068 						}
16069 					}
16070 #endif
16071 					continue;
16072 				}
16073 			}
16074 			if (SEQ_LEQ(ae->seq, tp->last_ack_sent) &&
16075 			    SEQ_LEQ(tp->last_ack_sent, ae->seq)) {
16076 				tp->ts_recent_age = tcp_ts_getticks();
16077 				tp->ts_recent = ae->ts_value;
16078 			}
16079 		} else {
16080 			/* Setup for a no options */
16081 			to->to_flags = 0;
16082 		}
16083 		/* Update the rcv time and perform idle reduction possibly */
16084 		if  (tp->t_idle_reduce &&
16085 		     (tp->snd_max == tp->snd_una) &&
16086 		     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
16087 			counter_u64_add(rack_input_idle_reduces, 1);
16088 			rack_cc_after_idle(rack, tp);
16089 		}
16090 		tp->t_rcvtime = ticks;
16091 		/* Now what about ECN of a chain of pure ACKs? */
16092 		if (tcp_ecn_input_segment(tp, ae->flags, 0,
16093 			tcp_packets_this_ack(tp, ae->ack),
16094 			ae->codepoint))
16095 			rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__);
16096 #ifdef TCP_ACCOUNTING
16097 		/* Count for the specific type of ack in */
16098 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16099 			tp->tcp_cnt_counters[ae->ack_val_set]++;
16100 		}
16101 #endif
16102 		/*
16103 		 * Note how we could move up these in the determination
16104 		 * above, but we don't so that way the timestamp checks (and ECN)
16105 		 * is done first before we do any processing on the ACK.
16106 		 * The non-compressed path through the code has this
16107 		 * weakness (noted by @jtl) that it actually does some
16108 		 * processing before verifying the timestamp information.
16109 		 * We don't take that path here which is why we set
16110 		 * the ack_val_set first, do the timestamp and ecn
16111 		 * processing, and then look at what we have setup.
16112 		 */
16113 		if (ae->ack_val_set == ACK_BEHIND) {
16114 			/*
16115 			 * Case B flag reordering, if window is not closed
16116 			 * or it could be a keep-alive or persists
16117 			 */
16118 			if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) {
16119 				rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
16120 				if (rack->r_ctl.rc_reorder_ts == 0)
16121 					rack->r_ctl.rc_reorder_ts = 1;
16122 			}
16123 		} else if (ae->ack_val_set == ACK_DUPACK) {
16124 			/* Case D */
16125 			rack_strike_dupack(rack);
16126 		} else if (ae->ack_val_set == ACK_RWND) {
16127 			/* Case C */
16128 			if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
16129 				ts.tv_sec = ae->timestamp / 1000000000;
16130 				ts.tv_nsec = ae->timestamp % 1000000000;
16131 				rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16132 				rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16133 			} else {
16134 				rack->r_ctl.act_rcv_time = *tv;
16135 			}
16136 			if (rack->forced_ack) {
16137 				rack_handle_probe_response(rack, tiwin,
16138 							   tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time));
16139 			}
16140 #ifdef TCP_ACCOUNTING
16141 			win_up_req = 1;
16142 #endif
16143 			win_upd_ack = ae->ack;
16144 			win_seq = ae->seq;
16145 			the_win = tiwin;
16146 			rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
16147 		} else {
16148 			/* Case A */
16149 			if (SEQ_GT(ae->ack, tp->snd_max)) {
16150 				/*
16151 				 * We just send an ack since the incoming
16152 				 * ack is beyond the largest seq we sent.
16153 				 */
16154 				if ((tp->t_flags & TF_ACKNOW) == 0) {
16155 					ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt);
16156 					if (tp->t_flags && TF_ACKNOW)
16157 						rack->r_wanted_output = 1;
16158 				}
16159 			} else {
16160 				nsegs++;
16161 				/* If the window changed setup to update */
16162 				if (tiwin != tp->snd_wnd) {
16163 					win_upd_ack = ae->ack;
16164 					win_seq = ae->seq;
16165 					the_win = tiwin;
16166 					rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts);
16167 				}
16168 #ifdef TCP_ACCOUNTING
16169 				/* Account for the acks */
16170 				if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16171 					tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz);
16172 				}
16173 #endif
16174 				high_seq = ae->ack;
16175 				if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp))
16176 					rack_log_hystart_event(rack, high_seq, 8);
16177 				/* Setup our act_rcv_time */
16178 				if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) {
16179 					ts.tv_sec = ae->timestamp / 1000000000;
16180 					ts.tv_nsec = ae->timestamp % 1000000000;
16181 					rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16182 					rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16183 				} else {
16184 					rack->r_ctl.act_rcv_time = *tv;
16185 				}
16186 				rack_process_to_cumack(tp, rack, ae->ack, cts, to,
16187 						       tcp_tv_to_lusectick(&rack->r_ctl.act_rcv_time));
16188 #ifdef TCP_REQUEST_TRK
16189 				rack_req_check_for_comp(rack, high_seq);
16190 #endif
16191 				if (rack->rc_dsack_round_seen) {
16192 					/* Is the dsack round over? */
16193 					if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) {
16194 						/* Yes it is */
16195 						rack->rc_dsack_round_seen = 0;
16196 						rack_log_dsack_event(rack, 3, __LINE__, 0, 0);
16197 					}
16198 				}
16199 			}
16200 		}
16201 		/* And lets be sure to commit the rtt measurements for this ack */
16202 		tcp_rack_xmit_timer_commit(rack, tp);
16203 #ifdef TCP_ACCOUNTING
16204 		rdstc = get_cyclecount();
16205 		if (rdstc > ts_val) {
16206 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16207 				tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val);
16208 				if (ae->ack_val_set == ACK_CUMACK)
16209 					tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val);
16210 			}
16211 		}
16212 #endif
16213 	}
16214 #ifdef TCP_ACCOUNTING
16215 	ts_val = get_cyclecount();
16216 #endif
16217 	/* Tend to any collapsed window */
16218 	if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) {
16219 		/* The peer collapsed the window */
16220 		rack_collapsed_window(rack, (tp->snd_max - high_seq), high_seq, __LINE__);
16221 	} else if (rack->rc_has_collapsed)
16222 		rack_un_collapse_window(rack, __LINE__);
16223 	if ((rack->r_collapse_point_valid) &&
16224 	    (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point)))
16225 		rack->r_collapse_point_valid = 0;
16226 	acked_amount = acked = (high_seq - tp->snd_una);
16227 	if (acked) {
16228 		/*
16229 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
16230 		 * causes issues when we are just going app limited. Lets
16231 		 * instead use SEQ_GT <or> where its equal but more data
16232 		 * is outstanding.
16233 		 *
16234 		 * Also make sure we are on the last ack of a series. We
16235 		 * have to have all the ack's processed in queue to know
16236 		 * if there is something left outstanding.
16237 		 *
16238 		 */
16239 		if (SEQ_GEQ(high_seq, rack->r_ctl.roundends) &&
16240 		    (rack->rc_new_rnd_needed == 0) &&
16241 		    (nxt_pkt == 0)) {
16242 			rack_log_hystart_event(rack, high_seq, 21);
16243 			rack->r_ctl.current_round++;
16244 			/* Force the next send to setup the next round */
16245 			rack->rc_new_rnd_needed = 1;
16246 			if (CC_ALGO(tp)->newround != NULL) {
16247 				CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
16248 			}
16249 		}
16250 		/*
16251 		 * Clear the probe not answered flag
16252 		 * since cum-ack moved forward.
16253 		 */
16254 		rack->probe_not_answered = 0;
16255 		if (rack->sack_attack_disable == 0)
16256 			rack_do_decay(rack);
16257 		if (acked >= segsiz) {
16258 			/*
16259 			 * You only get credit for
16260 			 * MSS and greater (and you get extra
16261 			 * credit for larger cum-ack moves).
16262 			 */
16263 			int ac;
16264 
16265 			ac = acked / segsiz;
16266 			rack->r_ctl.ack_count += ac;
16267 			counter_u64_add(rack_ack_total, ac);
16268 		}
16269 		if (rack->r_ctl.ack_count > 0xfff00000) {
16270 			/*
16271 			 * reduce the number to keep us under
16272 			 * a uint32_t.
16273 			 */
16274 			rack->r_ctl.ack_count /= 2;
16275 			rack->r_ctl.sack_count /= 2;
16276 		}
16277 		if (tp->t_flags & TF_NEEDSYN) {
16278 			/*
16279 			 * T/TCP: Connection was half-synchronized, and our SYN has
16280 			 * been ACK'd (so connection is now fully synchronized).  Go
16281 			 * to non-starred state, increment snd_una for ACK of SYN,
16282 			 * and check if we can do window scaling.
16283 			 */
16284 			tp->t_flags &= ~TF_NEEDSYN;
16285 			tp->snd_una++;
16286 			acked_amount = acked = (high_seq - tp->snd_una);
16287 		}
16288 		if (acked > sbavail(&so->so_snd))
16289 			acked_amount = sbavail(&so->so_snd);
16290 #ifdef TCP_SAD_DETECTION
16291 		/*
16292 		 * We only care on a cum-ack move if we are in a sack-disabled
16293 		 * state. We have already added in to the ack_count, and we never
16294 		 * would disable on a cum-ack move, so we only care to do the
16295 		 * detection if it may "undo" it, i.e. we were in disabled already.
16296 		 */
16297 		if (rack->sack_attack_disable)
16298 			rack_do_detection(tp, rack, acked_amount, segsiz);
16299 #endif
16300 		if (IN_FASTRECOVERY(tp->t_flags) &&
16301 		    (rack->rack_no_prr == 0))
16302 			rack_update_prr(tp, rack, acked_amount, high_seq);
16303 		if (IN_RECOVERY(tp->t_flags)) {
16304 			if (SEQ_LT(high_seq, tp->snd_recover) &&
16305 			    (SEQ_LT(high_seq, tp->snd_max))) {
16306 				tcp_rack_partialack(tp);
16307 			} else {
16308 				rack_post_recovery(tp, high_seq);
16309 				recovery = 1;
16310 			}
16311 		}
16312 		/* Handle the rack-log-ack part (sendmap) */
16313 		if ((sbused(&so->so_snd) == 0) &&
16314 		    (acked > acked_amount) &&
16315 		    (tp->t_state >= TCPS_FIN_WAIT_1) &&
16316 		    (tp->t_flags & TF_SENTFIN)) {
16317 			/*
16318 			 * We must be sure our fin
16319 			 * was sent and acked (we can be
16320 			 * in FIN_WAIT_1 without having
16321 			 * sent the fin).
16322 			 */
16323 			ourfinisacked = 1;
16324 			/*
16325 			 * Lets make sure snd_una is updated
16326 			 * since most likely acked_amount = 0 (it
16327 			 * should be).
16328 			 */
16329 			tp->snd_una = high_seq;
16330 		}
16331 		/* Did we make a RTO error? */
16332 		if ((tp->t_flags & TF_PREVVALID) &&
16333 		    ((tp->t_flags & TF_RCVD_TSTMP) == 0)) {
16334 			tp->t_flags &= ~TF_PREVVALID;
16335 			if (tp->t_rxtshift == 1 &&
16336 			    (int)(ticks - tp->t_badrxtwin) < 0)
16337 				rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__);
16338 		}
16339 		/* Handle the data in the socket buffer */
16340 		KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1);
16341 		KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked);
16342 		if (acked_amount > 0) {
16343 			struct mbuf *mfree;
16344 
16345 			rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery);
16346 			SOCKBUF_LOCK(&so->so_snd);
16347 			mfree = sbcut_locked(&so->so_snd, acked_amount);
16348 			tp->snd_una = high_seq;
16349 			/* Note we want to hold the sb lock through the sendmap adjust */
16350 			rack_adjust_sendmap_head(rack, &so->so_snd);
16351 			/* Wake up the socket if we have room to write more */
16352 			rack_log_wakeup(tp,rack, &so->so_snd, acked, 2);
16353 			sowwakeup_locked(so);
16354 			if ((recovery == 1) &&
16355 			    (rack->excess_rxt_on) &&
16356 			    (rack->r_cwnd_was_clamped == 0)) {
16357 				do_rack_excess_rxt(tp, rack);
16358 			} else if (rack->r_cwnd_was_clamped)
16359 				do_rack_check_for_unclamp(tp, rack);
16360 			m_freem(mfree);
16361 		}
16362 		/* update progress */
16363 		tp->t_acktime = ticks;
16364 		rack_log_progress_event(rack, tp, tp->t_acktime,
16365 					PROGRESS_UPDATE, __LINE__);
16366 		/* Clear out shifts and such */
16367 		tp->t_rxtshift = 0;
16368 		RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
16369 				   rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop);
16370 		rack->rc_tlp_in_progress = 0;
16371 		rack->r_ctl.rc_tlp_cnt_out = 0;
16372 		/* Send recover and snd_nxt must be dragged along */
16373 		if (SEQ_GT(tp->snd_una, tp->snd_recover))
16374 			tp->snd_recover = tp->snd_una;
16375 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))
16376 			tp->snd_nxt = tp->snd_max;
16377 		/*
16378 		 * If the RXT timer is running we want to
16379 		 * stop it, so we can restart a TLP (or new RXT).
16380 		 */
16381 		if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT)
16382 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16383 		tp->snd_wl2 = high_seq;
16384 		tp->t_dupacks = 0;
16385 		if (under_pacing &&
16386 		    (rack->use_fixed_rate == 0) &&
16387 		    (rack->in_probe_rtt == 0) &&
16388 		    rack->rc_gp_dyn_mul &&
16389 		    rack->rc_always_pace) {
16390 			/* Check if we are dragging bottom */
16391 			rack_check_bottom_drag(tp, rack, so);
16392 		}
16393 		if (tp->snd_una == tp->snd_max) {
16394 			tp->t_flags &= ~TF_PREVVALID;
16395 			rack->r_ctl.retran_during_recovery = 0;
16396 			rack->rc_suspicious = 0;
16397 			rack->r_ctl.dsack_byte_cnt = 0;
16398 			rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL);
16399 			if (rack->r_ctl.rc_went_idle_time == 0)
16400 				rack->r_ctl.rc_went_idle_time = 1;
16401 			rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
16402 			if (sbavail(&tptosocket(tp)->so_snd) == 0)
16403 				tp->t_acktime = 0;
16404 			/* Set so we might enter persists... */
16405 			rack->r_wanted_output = 1;
16406 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16407 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
16408 			if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16409 			    (sbavail(&so->so_snd) == 0) &&
16410 			    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
16411 				/*
16412 				 * The socket was gone and the
16413 				 * peer sent data (not now in the past), time to
16414 				 * reset him.
16415 				 */
16416 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
16417 				/* tcp_close will kill the inp pre-log the Reset */
16418 				tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
16419 #ifdef TCP_ACCOUNTING
16420 				rdstc = get_cyclecount();
16421 				if (rdstc > ts_val) {
16422 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16423 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16424 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16425 					}
16426 				}
16427 #endif
16428 				m_freem(m);
16429 				tp = tcp_close(tp);
16430 				if (tp == NULL) {
16431 #ifdef TCP_ACCOUNTING
16432 					sched_unpin();
16433 #endif
16434 					return (1);
16435 				}
16436 				/*
16437 				 * We would normally do drop-with-reset which would
16438 				 * send back a reset. We can't since we don't have
16439 				 * all the needed bits. Instead lets arrange for
16440 				 * a call to tcp_output(). That way since we
16441 				 * are in the closed state we will generate a reset.
16442 				 *
16443 				 * Note if tcp_accounting is on we don't unpin since
16444 				 * we do that after the goto label.
16445 				 */
16446 				goto send_out_a_rst;
16447 			}
16448 			if ((sbused(&so->so_snd) == 0) &&
16449 			    (tp->t_state >= TCPS_FIN_WAIT_1) &&
16450 			    (tp->t_flags & TF_SENTFIN)) {
16451 				/*
16452 				 * If we can't receive any more data, then closing user can
16453 				 * proceed. Starting the timer is contrary to the
16454 				 * specification, but if we don't get a FIN we'll hang
16455 				 * forever.
16456 				 *
16457 				 */
16458 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16459 					soisdisconnected(so);
16460 					tcp_timer_activate(tp, TT_2MSL,
16461 							   (tcp_fast_finwait2_recycle ?
16462 							    tcp_finwait2_timeout :
16463 							    TP_MAXIDLE(tp)));
16464 				}
16465 				if (ourfinisacked == 0) {
16466 					/*
16467 					 * We don't change to fin-wait-2 if we have our fin acked
16468 					 * which means we are probably in TCPS_CLOSING.
16469 					 */
16470 					tcp_state_change(tp, TCPS_FIN_WAIT_2);
16471 				}
16472 			}
16473 		}
16474 		/* Wake up the socket if we have room to write more */
16475 		if (sbavail(&so->so_snd)) {
16476 			rack->r_wanted_output = 1;
16477 			if (ctf_progress_timeout_check(tp, true)) {
16478 				rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr,
16479 							tp, tick, PROGRESS_DROP, __LINE__);
16480 				/*
16481 				 * We cheat here and don't send a RST, we should send one
16482 				 * when the pacer drops the connection.
16483 				 */
16484 #ifdef TCP_ACCOUNTING
16485 				rdstc = get_cyclecount();
16486 				if (rdstc > ts_val) {
16487 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16488 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16489 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16490 					}
16491 				}
16492 				sched_unpin();
16493 #endif
16494 				(void)tcp_drop(tp, ETIMEDOUT);
16495 				m_freem(m);
16496 				return (1);
16497 			}
16498 		}
16499 		if (ourfinisacked) {
16500 			switch(tp->t_state) {
16501 			case TCPS_CLOSING:
16502 #ifdef TCP_ACCOUNTING
16503 				rdstc = get_cyclecount();
16504 				if (rdstc > ts_val) {
16505 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16506 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16507 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16508 					}
16509 				}
16510 				sched_unpin();
16511 #endif
16512 				tcp_twstart(tp);
16513 				m_freem(m);
16514 				return (1);
16515 				break;
16516 			case TCPS_LAST_ACK:
16517 #ifdef TCP_ACCOUNTING
16518 				rdstc = get_cyclecount();
16519 				if (rdstc > ts_val) {
16520 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16521 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16522 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16523 					}
16524 				}
16525 				sched_unpin();
16526 #endif
16527 				tp = tcp_close(tp);
16528 				ctf_do_drop(m, tp);
16529 				return (1);
16530 				break;
16531 			case TCPS_FIN_WAIT_1:
16532 #ifdef TCP_ACCOUNTING
16533 				rdstc = get_cyclecount();
16534 				if (rdstc > ts_val) {
16535 					if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16536 						tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16537 						tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16538 					}
16539 				}
16540 #endif
16541 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
16542 					soisdisconnected(so);
16543 					tcp_timer_activate(tp, TT_2MSL,
16544 							   (tcp_fast_finwait2_recycle ?
16545 							    tcp_finwait2_timeout :
16546 							    TP_MAXIDLE(tp)));
16547 				}
16548 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
16549 				break;
16550 			default:
16551 				break;
16552 			}
16553 		}
16554 		if (rack->r_fast_output) {
16555 			/*
16556 			 * We re doing fast output.. can we expand that?
16557 			 */
16558 			rack_gain_for_fastoutput(rack, tp, so, acked_amount);
16559 		}
16560 #ifdef TCP_ACCOUNTING
16561 		rdstc = get_cyclecount();
16562 		if (rdstc > ts_val) {
16563 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16564 				tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val);
16565 				tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val);
16566 			}
16567 		}
16568 
16569 	} else if (win_up_req) {
16570 		rdstc = get_cyclecount();
16571 		if (rdstc > ts_val) {
16572 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
16573 				tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val);
16574 			}
16575 		}
16576 #endif
16577 	}
16578 	/* Now is there a next packet, if so we are done */
16579 	m_freem(m);
16580 	did_out = 0;
16581 	if (nxt_pkt) {
16582 #ifdef TCP_ACCOUNTING
16583 		sched_unpin();
16584 #endif
16585 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs);
16586 		return (0);
16587 	}
16588 	rack_handle_might_revert(tp, rack);
16589 	ctf_calc_rwin(so, tp);
16590 	if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
16591 	send_out_a_rst:
16592 		if (tcp_output(tp) < 0) {
16593 #ifdef TCP_ACCOUNTING
16594 			sched_unpin();
16595 #endif
16596 			return (1);
16597 		}
16598 		did_out = 1;
16599 	}
16600 	if (tp->t_flags2 & TF2_HPTS_CALLS)
16601 		tp->t_flags2 &= ~TF2_HPTS_CALLS;
16602 	rack_free_trim(rack);
16603 #ifdef TCP_ACCOUNTING
16604 	sched_unpin();
16605 #endif
16606 	rack_timer_audit(tp, rack, &so->so_snd);
16607 	rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs);
16608 	return (0);
16609 }
16610 
16611 #define	TCP_LRO_TS_OPTION \
16612     ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
16613 	  (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
16614 
16615 static int
16616 rack_do_segment_nounlock(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
16617     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt,
16618     struct timeval *tv)
16619 {
16620 	struct inpcb *inp = tptoinpcb(tp);
16621 	struct socket *so = tptosocket(tp);
16622 #ifdef TCP_ACCOUNTING
16623 	uint64_t ts_val;
16624 #endif
16625 	int32_t thflags, retval, did_out = 0;
16626 	int32_t way_out = 0;
16627 	/*
16628 	 * cts - is the current time from tv (caller gets ts) in microseconds.
16629 	 * ms_cts - is the current time from tv in milliseconds.
16630 	 * us_cts - is the time that LRO or hardware actually got the packet in microseconds.
16631 	 */
16632 	uint32_t cts, us_cts, ms_cts;
16633 	uint32_t tiwin, high_seq;
16634 	struct timespec ts;
16635 	struct tcpopt to;
16636 	struct tcp_rack *rack;
16637 	struct rack_sendmap *rsm;
16638 	int32_t prev_state = 0;
16639 	int no_output = 0;
16640 	int slot_remaining = 0;
16641 #ifdef TCP_ACCOUNTING
16642 	int ack_val_set = 0xf;
16643 #endif
16644 	int nsegs;
16645 
16646 	NET_EPOCH_ASSERT();
16647 	INP_WLOCK_ASSERT(inp);
16648 
16649 	/*
16650 	 * tv passed from common code is from either M_TSTMP_LRO or
16651 	 * tcp_get_usecs() if no LRO m_pkthdr timestamp is present.
16652 	 */
16653 	rack = (struct tcp_rack *)tp->t_fb_ptr;
16654 	if (rack->rack_deferred_inited == 0) {
16655 		/*
16656 		 * If we are the connecting socket we will
16657 		 * hit rack_init() when no sequence numbers
16658 		 * are setup. This makes it so we must defer
16659 		 * some initialization. Call that now.
16660 		 */
16661 		rack_deferred_init(tp, rack);
16662 	}
16663 	/*
16664 	 * Check to see if we need to skip any output plans. This
16665 	 * can happen in the non-LRO path where we are pacing and
16666 	 * must process the ack coming in but need to defer sending
16667 	 * anything becase a pacing timer is running.
16668 	 */
16669 	us_cts = tcp_tv_to_usectick(tv);
16670 	if (m->m_flags & M_ACKCMP) {
16671 		/*
16672 		 * All compressed ack's are ack's by definition so
16673 		 * remove any ack required flag and then do the processing.
16674 		 */
16675 		rack->rc_ack_required = 0;
16676 		return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv));
16677 	}
16678 	thflags = tcp_get_flags(th);
16679 	if ((rack->rc_always_pace == 1) &&
16680 	    (rack->rc_ack_can_sendout_data == 0) &&
16681 	    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
16682 	    (TSTMP_LT(us_cts, rack->r_ctl.rc_last_output_to))) {
16683 		/*
16684 		 * Ok conditions are right for queuing the packets
16685 		 * but we do have to check the flags in the inp, it
16686 		 * could be, if a sack is present, we want to be awoken and
16687 		 * so should process the packets.
16688 		 */
16689 		slot_remaining = rack->r_ctl.rc_last_output_to - us_cts;
16690 		if (rack->rc_tp->t_flags2 & TF2_DONT_SACK_QUEUE) {
16691 			no_output = 1;
16692 		} else {
16693 			/*
16694 			 * If there is no options, or just a
16695 			 * timestamp option, we will want to queue
16696 			 * the packets. This is the same that LRO does
16697 			 * and will need to change with accurate ECN.
16698 			 */
16699 			uint32_t *ts_ptr;
16700 			int optlen;
16701 
16702 			optlen = (th->th_off << 2) - sizeof(struct tcphdr);
16703 			ts_ptr = (uint32_t *)(th + 1);
16704 			if ((optlen == 0) ||
16705 			    ((optlen == TCPOLEN_TSTAMP_APPA) &&
16706 			     (*ts_ptr == TCP_LRO_TS_OPTION)))
16707 				no_output = 1;
16708 		}
16709 		if ((no_output == 1) && (slot_remaining < tcp_min_hptsi_time)) {
16710 			/*
16711 			 * It is unrealistic to think we can pace in less than
16712 			 * the minimum granularity of the pacer (def:250usec). So
16713 			 * if we have less than that time remaining we should go
16714 			 * ahead and allow output to be "early". We will attempt to
16715 			 * make up for it in any pacing time we try to apply on
16716 			 * the outbound packet.
16717 			 */
16718 			no_output = 0;
16719 		}
16720 	}
16721 	/*
16722 	 * If there is a RST or FIN lets dump out the bw
16723 	 * with a FIN the connection may go on but we
16724 	 * may not.
16725 	 */
16726 	if ((thflags & TH_FIN) || (thflags & TH_RST))
16727 		rack_log_pacing_delay_calc(rack,
16728 					   rack->r_ctl.gp_bw,
16729 					   0,
16730 					   0,
16731 					   rack_get_gp_est(rack), /* delRate */
16732 					   rack_get_lt_bw(rack), /* rttProp */
16733 					   20, __LINE__, NULL, 0);
16734 	if (m->m_flags & M_ACKCMP) {
16735 		panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp);
16736 	}
16737 	cts = tcp_tv_to_usectick(tv);
16738 	ms_cts =  tcp_tv_to_mssectick(tv);
16739 	nsegs = m->m_pkthdr.lro_nsegs;
16740 	counter_u64_add(rack_proc_non_comp_ack, 1);
16741 #ifdef TCP_ACCOUNTING
16742 	sched_pin();
16743 	if (thflags & TH_ACK)
16744 		ts_val = get_cyclecount();
16745 #endif
16746 	if ((m->m_flags & M_TSTMP) ||
16747 	    (m->m_flags & M_TSTMP_LRO)) {
16748 		mbuf_tstmp2timespec(m, &ts);
16749 		rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec;
16750 		rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000;
16751 	} else
16752 		rack->r_ctl.act_rcv_time = *tv;
16753 	kern_prefetch(rack, &prev_state);
16754 	prev_state = 0;
16755 	/*
16756 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
16757 	 * the scale is zero.
16758 	 */
16759 	tiwin = th->th_win << tp->snd_scale;
16760 #ifdef TCP_ACCOUNTING
16761 	if (thflags & TH_ACK) {
16762 		/*
16763 		 * We have a tradeoff here. We can either do what we are
16764 		 * doing i.e. pinning to this CPU and then doing the accounting
16765 		 * <or> we could do a critical enter, setup the rdtsc and cpu
16766 		 * as in below, and then validate we are on the same CPU on
16767 		 * exit. I have choosen to not do the critical enter since
16768 		 * that often will gain you a context switch, and instead lock
16769 		 * us (line above this if) to the same CPU with sched_pin(). This
16770 		 * means we may be context switched out for a higher priority
16771 		 * interupt but we won't be moved to another CPU.
16772 		 *
16773 		 * If this occurs (which it won't very often since we most likely
16774 		 * are running this code in interupt context and only a higher
16775 		 * priority will bump us ... clock?) we will falsely add in
16776 		 * to the time the interupt processing time plus the ack processing
16777 		 * time. This is ok since its a rare event.
16778 		 */
16779 		ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin,
16780 						    ctf_fixed_maxseg(tp));
16781 	}
16782 #endif
16783 	/*
16784 	 * Parse options on any incoming segment.
16785 	 */
16786 	memset(&to, 0, sizeof(to));
16787 	tcp_dooptions(&to, (u_char *)(th + 1),
16788 	    (th->th_off << 2) - sizeof(struct tcphdr),
16789 	    (thflags & TH_SYN) ? TO_SYN : 0);
16790 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
16791 	    __func__));
16792 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
16793 	    __func__));
16794 
16795 	if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
16796 	    (tp->t_flags & TF_GPUTINPROG)) {
16797 		/*
16798 		 * We have a goodput in progress
16799 		 * and we have entered a late state.
16800 		 * Do we have enough data in the sb
16801 		 * to handle the GPUT request?
16802 		 */
16803 		uint32_t bytes;
16804 
16805 		bytes = tp->gput_ack - tp->gput_seq;
16806 		if (SEQ_GT(tp->gput_seq, tp->snd_una))
16807 			bytes += tp->gput_seq - tp->snd_una;
16808 		if (bytes > sbavail(&tptosocket(tp)->so_snd)) {
16809 			/*
16810 			 * There are not enough bytes in the socket
16811 			 * buffer that have been sent to cover this
16812 			 * measurement. Cancel it.
16813 			 */
16814 			rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
16815 						   rack->r_ctl.rc_gp_srtt /*flex1*/,
16816 						   tp->gput_seq,
16817 						   0, 0, 18, __LINE__, NULL, 0);
16818 			tp->t_flags &= ~TF_GPUTINPROG;
16819 		}
16820 	}
16821 	high_seq = th->th_ack;
16822 	if (tcp_bblogging_on(rack->rc_tp)) {
16823 		union tcp_log_stackspecific log;
16824 		struct timeval ltv;
16825 #ifdef TCP_REQUEST_TRK
16826 		struct tcp_sendfile_track *tcp_req;
16827 
16828 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
16829 			tcp_req = tcp_req_find_req_for_seq(tp, (th->th_ack-1));
16830 		} else {
16831 			tcp_req = tcp_req_find_req_for_seq(tp, th->th_ack);
16832 		}
16833 #endif
16834 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
16835 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
16836 		if (rack->rack_no_prr == 0)
16837 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
16838 		else
16839 			log.u_bbr.flex1 = 0;
16840 		log.u_bbr.use_lt_bw = rack->r_ent_rec_ns;
16841 		log.u_bbr.use_lt_bw <<= 1;
16842 		log.u_bbr.use_lt_bw |= rack->r_might_revert;
16843 		log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced;
16844 		log.u_bbr.bbr_state = rack->rc_free_cnt;
16845 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
16846 		log.u_bbr.pkts_out = rack->rc_tp->t_maxseg;
16847 		log.u_bbr.flex3 = m->m_flags;
16848 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
16849 		log.u_bbr.lost = thflags;
16850 		log.u_bbr.pacing_gain = 0x1;
16851 #ifdef TCP_ACCOUNTING
16852 		log.u_bbr.cwnd_gain = ack_val_set;
16853 #endif
16854 		log.u_bbr.flex7 = 2;
16855 		if (m->m_flags & M_TSTMP) {
16856 			/* Record the hardware timestamp if present */
16857 			mbuf_tstmp2timespec(m, &ts);
16858 			ltv.tv_sec = ts.tv_sec;
16859 			ltv.tv_usec = ts.tv_nsec / 1000;
16860 			log.u_bbr.lt_epoch = tcp_tv_to_usectick(&ltv);
16861 		} else if (m->m_flags & M_TSTMP_LRO) {
16862 			/* Record the LRO the arrival timestamp */
16863 			mbuf_tstmp2timespec(m, &ts);
16864 			ltv.tv_sec = ts.tv_sec;
16865 			ltv.tv_usec = ts.tv_nsec / 1000;
16866 			log.u_bbr.flex5 = tcp_tv_to_usectick(&ltv);
16867 		}
16868 		log.u_bbr.timeStamp = tcp_get_usecs(&ltv);
16869 		/* Log the rcv time */
16870 		log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp;
16871 #ifdef TCP_REQUEST_TRK
16872 		log.u_bbr.applimited = tp->t_tcpreq_closed;
16873 		log.u_bbr.applimited <<= 8;
16874 		log.u_bbr.applimited |= tp->t_tcpreq_open;
16875 		log.u_bbr.applimited <<= 8;
16876 		log.u_bbr.applimited |= tp->t_tcpreq_req;
16877 		if (tcp_req) {
16878 			/* Copy out any client req info */
16879 			/* seconds */
16880 			log.u_bbr.pkt_epoch = (tcp_req->localtime / HPTS_USEC_IN_SEC);
16881 			/* useconds */
16882 			log.u_bbr.delivered = (tcp_req->localtime % HPTS_USEC_IN_SEC);
16883 			log.u_bbr.rttProp = tcp_req->timestamp;
16884 			log.u_bbr.cur_del_rate = tcp_req->start;
16885 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_OPEN) {
16886 				log.u_bbr.flex8 |= 1;
16887 			} else {
16888 				log.u_bbr.flex8 |= 2;
16889 				log.u_bbr.bw_inuse = tcp_req->end;
16890 			}
16891 			log.u_bbr.flex6 = tcp_req->start_seq;
16892 			if (tcp_req->flags & TCP_TRK_TRACK_FLG_COMP) {
16893 				log.u_bbr.flex8 |= 4;
16894 				log.u_bbr.epoch = tcp_req->end_seq;
16895 			}
16896 		}
16897 #endif
16898 		TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
16899 		    tlen, &log, true, &ltv);
16900 	}
16901 	/* Remove ack required flag if set, we have one  */
16902 	if (thflags & TH_ACK)
16903 		rack->rc_ack_required = 0;
16904 	if (rack->sack_attack_disable > 0) {
16905 		rack->r_ctl.ack_during_sd++;
16906 		rack_log_type_bbrsnd(rack, 0, 0, cts, tv, __LINE__);
16907 	}
16908 	if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) {
16909 		way_out = 4;
16910 		retval = 0;
16911 		m_freem(m);
16912 		goto done_with_input;
16913 	}
16914 	/*
16915 	 * If a segment with the ACK-bit set arrives in the SYN-SENT state
16916 	 * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9.
16917 	 */
16918 	if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) &&
16919 	    (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) {
16920 		tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT);
16921 		ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
16922 #ifdef TCP_ACCOUNTING
16923 		sched_unpin();
16924 #endif
16925 		return (1);
16926 	}
16927 	/*
16928 	 * If timestamps were negotiated during SYN/ACK and a
16929 	 * segment without a timestamp is received, silently drop
16930 	 * the segment, unless it is a RST segment or missing timestamps are
16931 	 * tolerated.
16932 	 * See section 3.2 of RFC 7323.
16933 	 */
16934 	if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) &&
16935 	    ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) {
16936 		way_out = 5;
16937 		retval = 0;
16938 		m_freem(m);
16939 		goto done_with_input;
16940 	}
16941 
16942 	/*
16943 	 * Segment received on connection. Reset idle time and keep-alive
16944 	 * timer. XXX: This should be done after segment validation to
16945 	 * ignore broken/spoofed segs.
16946 	 */
16947 	if  (tp->t_idle_reduce &&
16948 	     (tp->snd_max == tp->snd_una) &&
16949 	     (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) {
16950 		counter_u64_add(rack_input_idle_reduces, 1);
16951 		rack_cc_after_idle(rack, tp);
16952 	}
16953 	tp->t_rcvtime = ticks;
16954 #ifdef STATS
16955 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
16956 #endif
16957 	if (tiwin > rack->r_ctl.rc_high_rwnd)
16958 		rack->r_ctl.rc_high_rwnd = tiwin;
16959 	/*
16960 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
16961 	 * this to occur after we've validated the segment.
16962 	 */
16963 	if (tcp_ecn_input_segment(tp, thflags, tlen,
16964 	    tcp_packets_this_ack(tp, th->th_ack),
16965 	    iptos))
16966 		rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__);
16967 
16968 	/*
16969 	 * If echoed timestamp is later than the current time, fall back to
16970 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
16971 	 * were used when this connection was established.
16972 	 */
16973 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
16974 		to.to_tsecr -= tp->ts_offset;
16975 		if (TSTMP_GT(to.to_tsecr, ms_cts))
16976 			to.to_tsecr = 0;
16977 	}
16978 
16979 	/*
16980 	 * If its the first time in we need to take care of options and
16981 	 * verify we can do SACK for rack!
16982 	 */
16983 	if (rack->r_state == 0) {
16984 		/* Should be init'd by rack_init() */
16985 		KASSERT(rack->rc_inp != NULL,
16986 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
16987 		if (rack->rc_inp == NULL) {
16988 			rack->rc_inp = inp;
16989 		}
16990 
16991 		/*
16992 		 * Process options only when we get SYN/ACK back. The SYN
16993 		 * case for incoming connections is handled in tcp_syncache.
16994 		 * According to RFC1323 the window field in a SYN (i.e., a
16995 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
16996 		 * this is traditional behavior, may need to be cleaned up.
16997 		 */
16998 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
16999 			/* Handle parallel SYN for ECN */
17000 			tcp_ecn_input_parallel_syn(tp, thflags, iptos);
17001 			if ((to.to_flags & TOF_SCALE) &&
17002 			    (tp->t_flags & TF_REQ_SCALE)) {
17003 				tp->t_flags |= TF_RCVD_SCALE;
17004 				tp->snd_scale = to.to_wscale;
17005 			} else
17006 				tp->t_flags &= ~TF_REQ_SCALE;
17007 			/*
17008 			 * Initial send window.  It will be updated with the
17009 			 * next incoming segment to the scaled value.
17010 			 */
17011 			tp->snd_wnd = th->th_win;
17012 			rack_validate_fo_sendwin_up(tp, rack);
17013 			if ((to.to_flags & TOF_TS) &&
17014 			    (tp->t_flags & TF_REQ_TSTMP)) {
17015 				tp->t_flags |= TF_RCVD_TSTMP;
17016 				tp->ts_recent = to.to_tsval;
17017 				tp->ts_recent_age = cts;
17018 			} else
17019 				tp->t_flags &= ~TF_REQ_TSTMP;
17020 			if (to.to_flags & TOF_MSS) {
17021 				tcp_mss(tp, to.to_mss);
17022 			}
17023 			if ((tp->t_flags & TF_SACK_PERMIT) &&
17024 			    (to.to_flags & TOF_SACKPERM) == 0)
17025 				tp->t_flags &= ~TF_SACK_PERMIT;
17026 			if (IS_FASTOPEN(tp->t_flags)) {
17027 				if (to.to_flags & TOF_FASTOPEN) {
17028 					uint16_t mss;
17029 
17030 					if (to.to_flags & TOF_MSS)
17031 						mss = to.to_mss;
17032 					else
17033 						if ((inp->inp_vflag & INP_IPV6) != 0)
17034 							mss = TCP6_MSS;
17035 						else
17036 							mss = TCP_MSS;
17037 					tcp_fastopen_update_cache(tp, mss,
17038 					    to.to_tfo_len, to.to_tfo_cookie);
17039 				} else
17040 					tcp_fastopen_disable_path(tp);
17041 			}
17042 		}
17043 		/*
17044 		 * At this point we are at the initial call. Here we decide
17045 		 * if we are doing RACK or not. We do this by seeing if
17046 		 * TF_SACK_PERMIT is set and the sack-not-required is clear.
17047 		 * The code now does do dup-ack counting so if you don't
17048 		 * switch back you won't get rack & TLP, but you will still
17049 		 * get this stack.
17050 		 */
17051 
17052 		if ((rack_sack_not_required == 0) &&
17053 		    ((tp->t_flags & TF_SACK_PERMIT) == 0)) {
17054 			tcp_switch_back_to_default(tp);
17055 			(*tp->t_fb->tfb_tcp_do_segment)(tp, m, th, drop_hdrlen,
17056 			    tlen, iptos);
17057 #ifdef TCP_ACCOUNTING
17058 			sched_unpin();
17059 #endif
17060 			return (1);
17061 		}
17062 		tcp_set_hpts(tp);
17063 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
17064 	}
17065 	if (thflags & TH_FIN)
17066 		tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN);
17067 	us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time);
17068 	if ((rack->rc_gp_dyn_mul) &&
17069 	    (rack->use_fixed_rate == 0) &&
17070 	    (rack->rc_always_pace)) {
17071 		/* Check in on probertt */
17072 		rack_check_probe_rtt(rack, us_cts);
17073 	}
17074 	rack_clear_rate_sample(rack);
17075 	if ((rack->forced_ack) &&
17076 	    ((tcp_get_flags(th) & TH_RST) == 0)) {
17077 		rack_handle_probe_response(rack, tiwin, us_cts);
17078 	}
17079 	/*
17080 	 * This is the one exception case where we set the rack state
17081 	 * always. All other times (timers etc) we must have a rack-state
17082 	 * set (so we assure we have done the checks above for SACK).
17083 	 */
17084 	rack->r_ctl.rc_rcvtime = cts;
17085 	if (rack->r_state != tp->t_state)
17086 		rack_set_state(tp, rack);
17087 	if (SEQ_GT(th->th_ack, tp->snd_una) &&
17088 	    (rsm = tqhash_min(rack->r_ctl.tqh)) != NULL)
17089 		kern_prefetch(rsm, &prev_state);
17090 	prev_state = rack->r_state;
17091 	if ((thflags & TH_RST) &&
17092 	    ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
17093 	      SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
17094 	     (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq))) {
17095 		/* The connection will be killed by a reset check the tracepoint */
17096 		tcp_trace_point(rack->rc_tp, TCP_TP_RESET_RCV);
17097 	}
17098 	retval = (*rack->r_substate) (m, th, so,
17099 	    tp, &to, drop_hdrlen,
17100 	    tlen, tiwin, thflags, nxt_pkt, iptos);
17101 	if (retval == 0) {
17102 		/*
17103 		 * If retval is 1 the tcb is unlocked and most likely the tp
17104 		 * is gone.
17105 		 */
17106 		INP_WLOCK_ASSERT(inp);
17107 		if ((rack->rc_gp_dyn_mul) &&
17108 		    (rack->rc_always_pace) &&
17109 		    (rack->use_fixed_rate == 0) &&
17110 		    rack->in_probe_rtt &&
17111 		    (rack->r_ctl.rc_time_probertt_starts == 0)) {
17112 			/*
17113 			 * If we are going for target, lets recheck before
17114 			 * we output.
17115 			 */
17116 			rack_check_probe_rtt(rack, us_cts);
17117 		}
17118 		if (rack->set_pacing_done_a_iw == 0) {
17119 			/* How much has been acked? */
17120 			if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) {
17121 				/* We have enough to set in the pacing segment size */
17122 				rack->set_pacing_done_a_iw = 1;
17123 				rack_set_pace_segments(tp, rack, __LINE__, NULL);
17124 			}
17125 		}
17126 		tcp_rack_xmit_timer_commit(rack, tp);
17127 #ifdef TCP_ACCOUNTING
17128 		/*
17129 		 * If we set the ack_val_se to what ack processing we are doing
17130 		 * we also want to track how many cycles we burned. Note
17131 		 * the bits after tcp_output we let be "free". This is because
17132 		 * we are also tracking the tcp_output times as well. Note the
17133 		 * use of 0xf here since we only have 11 counter (0 - 0xa) and
17134 		 * 0xf cannot be returned and is what we initialize it too to
17135 		 * indicate we are not doing the tabulations.
17136 		 */
17137 		if (ack_val_set != 0xf) {
17138 			uint64_t crtsc;
17139 
17140 			crtsc = get_cyclecount();
17141 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
17142 				tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val);
17143 			}
17144 		}
17145 #endif
17146 		if ((nxt_pkt == 0) && (no_output == 0)) {
17147 			if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) {
17148 do_output_now:
17149 				if (tcp_output(tp) < 0) {
17150 #ifdef TCP_ACCOUNTING
17151 					sched_unpin();
17152 #endif
17153 					return (1);
17154 				}
17155 				did_out = 1;
17156 			}
17157 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
17158 			rack_free_trim(rack);
17159 		} else if ((no_output == 1) &&
17160 			   (nxt_pkt == 0)  &&
17161 			   (tcp_in_hpts(rack->rc_tp) == 0)) {
17162 			/*
17163 			 * We are not in hpts and we had a pacing timer up. Use
17164 			 * the remaining time (slot_remaining) to restart the timer.
17165 			 */
17166 			KASSERT ((slot_remaining != 0), ("slot remaining is zero for rack:%p tp:%p", rack, tp));
17167 			rack_start_hpts_timer(rack, tp, cts, slot_remaining, 0, 0);
17168 			rack_free_trim(rack);
17169 		}
17170 		/* Clear the flag, it may have been cleared by output but we may not have  */
17171 		if ((nxt_pkt == 0) && (tp->t_flags2 & TF2_HPTS_CALLS))
17172 			tp->t_flags2 &= ~TF2_HPTS_CALLS;
17173 		/* Update any rounds needed */
17174 		if (rack_verbose_logging &&  tcp_bblogging_on(rack->rc_tp))
17175 			rack_log_hystart_event(rack, high_seq, 8);
17176 		/*
17177 		 * The draft (v3) calls for us to use SEQ_GEQ, but that
17178 		 * causes issues when we are just going app limited. Lets
17179 		 * instead use SEQ_GT <or> where its equal but more data
17180 		 * is outstanding.
17181 		 *
17182 		 * Also make sure we are on the last ack of a series. We
17183 		 * have to have all the ack's processed in queue to know
17184 		 * if there is something left outstanding.
17185 		 */
17186 		if (SEQ_GEQ(tp->snd_una, rack->r_ctl.roundends) &&
17187 		    (rack->rc_new_rnd_needed == 0) &&
17188 		    (nxt_pkt == 0)) {
17189 			rack_log_hystart_event(rack, tp->snd_una, 21);
17190 			rack->r_ctl.current_round++;
17191 			/* Force the next send to setup the next round */
17192 			rack->rc_new_rnd_needed = 1;
17193 			if (CC_ALGO(tp)->newround != NULL) {
17194 				CC_ALGO(tp)->newround(&tp->t_ccv, rack->r_ctl.current_round);
17195 			}
17196 		}
17197 		if ((nxt_pkt == 0) &&
17198 		    ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
17199 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
17200 		     (tp->t_flags & TF_DELACK) ||
17201 		     ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
17202 		      (tp->t_state <= TCPS_CLOSING)))) {
17203 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
17204 			if ((tp->snd_max == tp->snd_una) &&
17205 			    ((tp->t_flags & TF_DELACK) == 0) &&
17206 			    (tcp_in_hpts(rack->rc_tp)) &&
17207 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
17208 				/* keep alive not needed if we are hptsi output yet */
17209 				;
17210 			} else {
17211 				int late = 0;
17212 				if (tcp_in_hpts(tp)) {
17213 					if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
17214 						us_cts = tcp_get_usecs(NULL);
17215 						if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) {
17216 							rack->r_early = 1;
17217 							rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts);
17218 						} else
17219 							late = 1;
17220 						rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
17221 					}
17222 					tcp_hpts_remove(tp);
17223 				}
17224 				if (late && (did_out == 0)) {
17225 					/*
17226 					 * We are late in the sending
17227 					 * and we did not call the output
17228 					 * (this probably should not happen).
17229 					 */
17230 					goto do_output_now;
17231 				}
17232 				rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0);
17233 			}
17234 			way_out = 1;
17235 		} else if (nxt_pkt == 0) {
17236 			/* Do we have the correct timer running? */
17237 			rack_timer_audit(tp, rack, &so->so_snd);
17238 			way_out = 2;
17239 		}
17240 	done_with_input:
17241 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs));
17242 		if (did_out)
17243 			rack->r_wanted_output = 0;
17244 	}
17245 #ifdef TCP_ACCOUNTING
17246 	sched_unpin();
17247 #endif
17248 	return (retval);
17249 }
17250 
17251 static void
17252 rack_do_segment(struct tcpcb *tp, struct mbuf *m, struct tcphdr *th,
17253     int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
17254 {
17255 	struct timeval tv;
17256 
17257 	/* First lets see if we have old packets */
17258 	if (!STAILQ_EMPTY(&tp->t_inqueue)) {
17259 		if (ctf_do_queued_segments(tp, 1)) {
17260 			m_freem(m);
17261 			return;
17262 		}
17263 	}
17264 	if (m->m_flags & M_TSTMP_LRO) {
17265 		mbuf_tstmp2timeval(m, &tv);
17266 	} else {
17267 		/* Should not be should we kassert instead? */
17268 		tcp_get_usecs(&tv);
17269 	}
17270 	if (rack_do_segment_nounlock(tp, m, th, drop_hdrlen, tlen, iptos, 0,
17271 	    &tv) == 0) {
17272 		INP_WUNLOCK(tptoinpcb(tp));
17273 	}
17274 }
17275 
17276 struct rack_sendmap *
17277 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
17278 {
17279 	struct rack_sendmap *rsm = NULL;
17280 	int32_t idx;
17281 	uint32_t srtt = 0, thresh = 0, ts_low = 0;
17282 	int no_sack = 0;
17283 
17284 	/* Return the next guy to be re-transmitted */
17285 	if (tqhash_empty(rack->r_ctl.tqh)) {
17286 		return (NULL);
17287 	}
17288 	if (tp->t_flags & TF_SENTFIN) {
17289 		/* retran the end FIN? */
17290 		return (NULL);
17291 	}
17292 	/* ok lets look at this one */
17293 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
17294 	if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) {
17295 		return (rsm);
17296 	}
17297 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
17298 		goto check_it;
17299 	}
17300 	rsm = rack_find_lowest_rsm(rack);
17301 	if (rsm == NULL) {
17302 		return (NULL);
17303 	}
17304 check_it:
17305 	if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) ||
17306 	    (rack->sack_attack_disable > 0)) {
17307 		no_sack = 1;
17308 	}
17309 	if ((no_sack > 0) &&
17310 	    (rsm->r_dupack >= DUP_ACK_THRESHOLD)) {
17311 		/*
17312 		 * No sack so we automatically do the 3 strikes and
17313 		 * retransmit (no rack timer would be started).
17314 		 */
17315 		return (rsm);
17316 	}
17317 	if (rsm->r_flags & RACK_ACKED) {
17318 		return (NULL);
17319 	}
17320 	if (((rsm->r_flags & RACK_SACK_PASSED) == 0) &&
17321 	    (rsm->r_dupack < DUP_ACK_THRESHOLD)) {
17322 		/* Its not yet ready */
17323 		return (NULL);
17324 	}
17325 	srtt = rack_grab_rtt(tp, rack);
17326 	idx = rsm->r_rtr_cnt - 1;
17327 	ts_low = (uint32_t)rsm->r_tim_lastsent[idx];
17328 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
17329 	if ((tsused == ts_low) ||
17330 	    (TSTMP_LT(tsused, ts_low))) {
17331 		/* No time since sending */
17332 		return (NULL);
17333 	}
17334 	if ((tsused - ts_low) < thresh) {
17335 		/* It has not been long enough yet */
17336 		return (NULL);
17337 	}
17338 	if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) ||
17339 	    ((rsm->r_flags & RACK_SACK_PASSED) &&
17340 	     (rack->sack_attack_disable == 0))) {
17341 		/*
17342 		 * We have passed the dup-ack threshold <or>
17343 		 * a SACK has indicated this is missing.
17344 		 * Note that if you are a declared attacker
17345 		 * it is only the dup-ack threshold that
17346 		 * will cause retransmits.
17347 		 */
17348 		/* log retransmit reason */
17349 		rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1);
17350 		rack->r_fast_output = 0;
17351 		return (rsm);
17352 	}
17353 	return (NULL);
17354 }
17355 
17356 static void
17357 rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot,
17358 			   uint64_t bw_est, uint64_t bw, uint64_t len_time, int method,
17359 			   int line, struct rack_sendmap *rsm, uint8_t quality)
17360 {
17361 	if (tcp_bblogging_on(rack->rc_tp)) {
17362 		union tcp_log_stackspecific log;
17363 		struct timeval tv;
17364 
17365 		if (rack_verbose_logging == 0) {
17366 			/*
17367 			 * We are not verbose screen out all but
17368 			 * ones we always want.
17369 			 */
17370 			if ((method != 2) &&
17371 			    (method != 3) &&
17372 			    (method != 7) &&
17373 			    (method != 14) &&
17374 			    (method != 20)) {
17375 				return;
17376 			}
17377 		}
17378 		memset(&log, 0, sizeof(log));
17379 		log.u_bbr.flex1 = slot;
17380 		log.u_bbr.flex2 = len;
17381 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs;
17382 		log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs;
17383 		log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss;
17384 		log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca;
17385 		log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data;
17386 		log.u_bbr.use_lt_bw <<= 1;
17387 		log.u_bbr.use_lt_bw |= rack->r_late;
17388 		log.u_bbr.use_lt_bw <<= 1;
17389 		log.u_bbr.use_lt_bw |= rack->r_early;
17390 		log.u_bbr.use_lt_bw <<= 1;
17391 		log.u_bbr.use_lt_bw |= rack->app_limited_needs_set;
17392 		log.u_bbr.use_lt_bw <<= 1;
17393 		log.u_bbr.use_lt_bw |= rack->rc_gp_filled;
17394 		log.u_bbr.use_lt_bw <<= 1;
17395 		log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt;
17396 		log.u_bbr.use_lt_bw <<= 1;
17397 		log.u_bbr.use_lt_bw |= rack->in_probe_rtt;
17398 		log.u_bbr.use_lt_bw <<= 1;
17399 		log.u_bbr.use_lt_bw |= rack->gp_ready;
17400 		log.u_bbr.pkt_epoch = line;
17401 		log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed;
17402 		log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early;
17403 		log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec;
17404 		log.u_bbr.bw_inuse = bw_est;
17405 		log.u_bbr.delRate = bw;
17406 		if (rack->r_ctl.gp_bw == 0)
17407 			log.u_bbr.cur_del_rate = 0;
17408 		else
17409 			log.u_bbr.cur_del_rate = rack_get_bw(rack);
17410 		log.u_bbr.rttProp = len_time;
17411 		log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt;
17412 		log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit;
17413 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
17414 		if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) {
17415 			/* We are in slow start */
17416 			log.u_bbr.flex7 = 1;
17417 		} else {
17418 			/* we are on congestion avoidance */
17419 			log.u_bbr.flex7 = 0;
17420 		}
17421 		log.u_bbr.flex8 = method;
17422 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
17423 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
17424 		log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec;
17425 		log.u_bbr.cwnd_gain <<= 1;
17426 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss;
17427 		log.u_bbr.cwnd_gain <<= 1;
17428 		log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca;
17429 		log.u_bbr.bbr_substate = quality;
17430 		log.u_bbr.bbr_state = rack->dgp_on;
17431 		log.u_bbr.bbr_state <<= 1;
17432 		log.u_bbr.bbr_state |= rack->r_fill_less_agg;
17433 		log.u_bbr.bbr_state <<= 1;
17434 		log.u_bbr.bbr_state |= rack->rc_pace_to_cwnd;
17435 		log.u_bbr.bbr_state <<= 2;
17436 		log.u_bbr.bbr_state |= rack->r_pacing_discount;
17437 		log.u_bbr.flex7 = ((rack->r_ctl.pacing_discount_amm << 1) | log.u_bbr.flex7);
17438 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
17439 		    &rack->rc_inp->inp_socket->so_rcv,
17440 		    &rack->rc_inp->inp_socket->so_snd,
17441 		    BBR_LOG_HPTSI_CALC, 0,
17442 		    0, &log, false, &tv);
17443 	}
17444 }
17445 
17446 static uint32_t
17447 rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss)
17448 {
17449 	uint32_t new_tso, user_max, pace_one;
17450 
17451 	user_max = rack->rc_user_set_max_segs * mss;
17452 	if (rack->rc_force_max_seg) {
17453 		return (user_max);
17454 	}
17455 	if (rack->use_fixed_rate &&
17456 	    ((rack->r_ctl.crte == NULL) ||
17457 	     (bw != rack->r_ctl.crte->rate))) {
17458 		/* Use the user mss since we are not exactly matched */
17459 		return (user_max);
17460 	}
17461 	if (rack_pace_one_seg ||
17462 	    (rack->r_ctl.rc_user_set_min_segs == 1))
17463 		pace_one = 1;
17464 	else
17465 		pace_one = 0;
17466 
17467 	new_tso = tcp_get_pacing_burst_size_w_divisor(rack->rc_tp, bw, mss,
17468 		     pace_one, rack->r_ctl.crte, NULL, rack->r_ctl.pace_len_divisor);
17469 	if (new_tso > user_max)
17470 		new_tso = user_max;
17471 	if (rack->rc_hybrid_mode && rack->r_ctl.client_suggested_maxseg) {
17472 		if (((uint32_t)rack->r_ctl.client_suggested_maxseg * mss) > new_tso)
17473 			new_tso = (uint32_t)rack->r_ctl.client_suggested_maxseg * mss;
17474 	}
17475 	if (rack->r_ctl.rc_user_set_min_segs &&
17476 	    ((rack->r_ctl.rc_user_set_min_segs * mss) > new_tso))
17477 	    new_tso = rack->r_ctl.rc_user_set_min_segs * mss;
17478 	return (new_tso);
17479 }
17480 
17481 static uint64_t
17482 rack_arrive_at_discounted_rate(struct tcp_rack *rack, uint64_t window_input, uint32_t *rate_set, uint32_t *gain_b)
17483 {
17484 	uint64_t reduced_win;
17485 	uint32_t gain;
17486 
17487 	if (window_input < rc_init_window(rack)) {
17488 		/*
17489 		 * The cwnd is collapsed to
17490 		 * nearly zero, maybe because of a time-out?
17491 		 * Lets drop back to the lt-bw.
17492 		 */
17493 		reduced_win = rack_get_lt_bw(rack);
17494 		/* Set the flag so the caller knows its a rate and not a reduced window */
17495 		*rate_set = 1;
17496 		gain = 100;
17497 	} else if  (IN_RECOVERY(rack->rc_tp->t_flags)) {
17498 		/*
17499 		 * If we are in recover our cwnd needs to be less for
17500 		 * our pacing consideration.
17501 		 */
17502 		if (rack->rack_hibeta == 0) {
17503 			reduced_win = window_input / 2;
17504 			gain = 50;
17505 		} else {
17506 			reduced_win = window_input * rack->r_ctl.saved_hibeta;
17507 			reduced_win /= 100;
17508 			gain = rack->r_ctl.saved_hibeta;
17509 		}
17510 	} else {
17511 		/*
17512 		 * Apply Timely factor to increase/decrease the
17513 		 * amount we are pacing at.
17514 		 */
17515 		gain = rack_get_output_gain(rack, NULL);
17516 		if (gain > rack_gain_p5_ub) {
17517 			gain = rack_gain_p5_ub;
17518 		}
17519 		reduced_win = window_input * gain;
17520 		reduced_win /= 100;
17521 	}
17522 	if (gain_b != NULL)
17523 		*gain_b = gain;
17524 	/*
17525 	 * What is being returned here is a trimmed down
17526 	 * window values in all cases where rate_set is left
17527 	 * at 0. In one case we actually return the rate (lt_bw).
17528 	 * the "reduced_win" is returned as a slimmed down cwnd that
17529 	 * is then calculated by the caller into a rate when rate_set
17530 	 * is 0.
17531 	 */
17532 	return (reduced_win);
17533 }
17534 
17535 static int32_t
17536 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)
17537 {
17538 	uint64_t lentim, fill_bw;
17539 
17540 	/* Lets first see if we are full, if so continue with normal rate */
17541 	rack->r_via_fill_cw = 0;
17542 	if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use)
17543 		return (slot);
17544 	if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd)
17545 		return (slot);
17546 	if (rack->r_ctl.rc_last_us_rtt == 0)
17547 		return (slot);
17548 	if (rack->rc_pace_fill_if_rttin_range &&
17549 	    (rack->r_ctl.rc_last_us_rtt >=
17550 	     (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) {
17551 		/* The rtt is huge, N * smallest, lets not fill */
17552 		return (slot);
17553 	}
17554 	/*
17555 	 * first lets calculate the b/w based on the last us-rtt
17556 	 * and the the smallest send window.
17557 	 */
17558 	fill_bw = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17559 	if (rack->rc_fillcw_apply_discount) {
17560 		uint32_t rate_set = 0;
17561 
17562 		fill_bw = rack_arrive_at_discounted_rate(rack, fill_bw, &rate_set, NULL);
17563 		if (rate_set) {
17564 			goto at_lt_bw;
17565 		}
17566 	}
17567 	/* Take the rwnd if its smaller */
17568 	if (fill_bw > rack->rc_tp->snd_wnd)
17569 		fill_bw = rack->rc_tp->snd_wnd;
17570 	/* Now lets make it into a b/w */
17571 	fill_bw *= (uint64_t)HPTS_USEC_IN_SEC;
17572 	fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17573 at_lt_bw:
17574 	if (rack->r_fill_less_agg) {
17575 		/*
17576 		 * We want the average of the rate_wanted
17577 		 * and our fill-cw calculated bw. We also want
17578 		 * to cap any increase to be no more than
17579 		 * X times the lt_bw (where X is the rack_bw_multipler).
17580 		 */
17581 		uint64_t lt_bw, rate;
17582 
17583 		lt_bw = rack_get_lt_bw(rack);
17584 		if (lt_bw > *rate_wanted)
17585 			rate = lt_bw;
17586 		else
17587 			rate = *rate_wanted;
17588 		fill_bw += rate;
17589 		fill_bw /= 2;
17590 		if (rack_bw_multipler && (fill_bw > (rate * rack_bw_multipler))) {
17591 			fill_bw = rate * rack_bw_multipler;
17592 		}
17593 	}
17594 	/* We are below the min b/w */
17595 	if (non_paced)
17596 		*rate_wanted = fill_bw;
17597 	if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted))
17598 		return (slot);
17599 	rack->r_via_fill_cw = 1;
17600 	if (rack->r_rack_hw_rate_caps &&
17601 	    (rack->r_ctl.crte != NULL)) {
17602 		uint64_t high_rate;
17603 
17604 		high_rate = tcp_hw_highest_rate(rack->r_ctl.crte);
17605 		if (fill_bw > high_rate) {
17606 			/* We are capping bw at the highest rate table entry */
17607 			if (*rate_wanted > high_rate) {
17608 				/* The original rate was also capped */
17609 				rack->r_via_fill_cw = 0;
17610 			}
17611 			rack_log_hdwr_pacing(rack,
17612 					     fill_bw, high_rate, __LINE__,
17613 					     0, 3);
17614 			fill_bw = high_rate;
17615 			if (capped)
17616 				*capped = 1;
17617 		}
17618 	} else if ((rack->r_ctl.crte == NULL) &&
17619 		   (rack->rack_hdrw_pacing == 0) &&
17620 		   (rack->rack_hdw_pace_ena) &&
17621 		   rack->r_rack_hw_rate_caps &&
17622 		   (rack->rack_attempt_hdwr_pace == 0) &&
17623 		   (rack->rc_inp->inp_route.ro_nh != NULL) &&
17624 		   (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17625 		/*
17626 		 * Ok we may have a first attempt that is greater than our top rate
17627 		 * lets check.
17628 		 */
17629 		uint64_t high_rate;
17630 
17631 		high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp);
17632 		if (high_rate) {
17633 			if (fill_bw > high_rate) {
17634 				fill_bw = high_rate;
17635 				if (capped)
17636 					*capped = 1;
17637 			}
17638 		}
17639 	}
17640 	if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) {
17641 		if (rack->rc_hybrid_mode)
17642 			rack_log_hybrid_bw(rack, rack->rc_tp->snd_max,
17643 					   fill_bw, 0, 0, HYBRID_LOG_RATE_CAP, 2, NULL, __LINE__);
17644 		fill_bw = rack->r_ctl.bw_rate_cap;
17645 	}
17646 	/*
17647 	 * Ok fill_bw holds our mythical b/w to fill the cwnd
17648 	 * in an rtt (unless it was capped), what does that
17649 	 * time wise equate too?
17650 	 */
17651 	lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC;
17652 	lentim /= fill_bw;
17653 	*rate_wanted = fill_bw;
17654 	if (non_paced || (lentim < slot)) {
17655 		rack_log_pacing_delay_calc(rack, len, slot, fill_bw,
17656 					   0, lentim, 12, __LINE__, NULL, 0);
17657 		return ((int32_t)lentim);
17658 	} else
17659 		return (slot);
17660 }
17661 
17662 static int32_t
17663 rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz)
17664 {
17665 	uint64_t srtt;
17666 	int32_t slot = 0;
17667 	int32_t minslot = 0;
17668 	int can_start_hw_pacing = 1;
17669 	int err;
17670 	int pace_one;
17671 
17672 	if (rack_pace_one_seg ||
17673 	    (rack->r_ctl.rc_user_set_min_segs == 1))
17674 		pace_one = 1;
17675 	else
17676 		pace_one = 0;
17677 	if (rack->rc_always_pace == 0) {
17678 		/*
17679 		 * We use the most optimistic possible cwnd/srtt for
17680 		 * sending calculations. This will make our
17681 		 * calculation anticipate getting more through
17682 		 * quicker then possible. But thats ok we don't want
17683 		 * the peer to have a gap in data sending.
17684 		 */
17685 		uint64_t cwnd, tr_perms = 0;
17686 		int32_t reduce = 0;
17687 
17688 	old_method:
17689 		/*
17690 		 * We keep no precise pacing with the old method
17691 		 * instead we use the pacer to mitigate bursts.
17692 		 */
17693 		if (rack->r_ctl.rc_rack_min_rtt)
17694 			srtt = rack->r_ctl.rc_rack_min_rtt;
17695 		else
17696 			srtt = max(tp->t_srtt, 1);
17697 		if (rack->r_ctl.rc_rack_largest_cwnd)
17698 			cwnd = rack->r_ctl.rc_rack_largest_cwnd;
17699 		else
17700 			cwnd = rack->r_ctl.cwnd_to_use;
17701 		/* Inflate cwnd by 1000 so srtt of usecs is in ms */
17702 		tr_perms = (cwnd * 1000) / srtt;
17703 		if (tr_perms == 0) {
17704 			tr_perms = ctf_fixed_maxseg(tp);
17705 		}
17706 		/*
17707 		 * Calculate how long this will take to drain, if
17708 		 * the calculation comes out to zero, thats ok we
17709 		 * will use send_a_lot to possibly spin around for
17710 		 * more increasing tot_len_this_send to the point
17711 		 * that its going to require a pace, or we hit the
17712 		 * cwnd. Which in that case we are just waiting for
17713 		 * a ACK.
17714 		 */
17715 		slot = len / tr_perms;
17716 		/* Now do we reduce the time so we don't run dry? */
17717 		if (slot && rack_slot_reduction) {
17718 			reduce = (slot / rack_slot_reduction);
17719 			if (reduce < slot) {
17720 				slot -= reduce;
17721 			} else
17722 				slot = 0;
17723 		}
17724 		slot *= HPTS_USEC_IN_MSEC;
17725 		if (rack->rc_pace_to_cwnd) {
17726 			uint64_t rate_wanted = 0;
17727 
17728 			slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1);
17729 			rack->rc_ack_can_sendout_data = 1;
17730 			rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0);
17731 		} else
17732 			rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0);
17733 		/*******************************************************/
17734 		/* RRS: We insert non-paced call to stats here for len */
17735 		/*******************************************************/
17736 	} else {
17737 		uint64_t bw_est, res, lentim, rate_wanted;
17738 		uint32_t segs, oh;
17739 		int capped = 0;
17740 		int prev_fill;
17741 
17742 		if ((rack->r_rr_config == 1) && rsm) {
17743 			return (rack->r_ctl.rc_min_to);
17744 		}
17745 		if (rack->use_fixed_rate) {
17746 			rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack);
17747 		} else if ((rack->r_ctl.init_rate == 0) &&
17748 			   (rack->r_ctl.gp_bw == 0)) {
17749 			/* no way to yet do an estimate */
17750 			bw_est = rate_wanted = 0;
17751 		} else if (rack->dgp_on) {
17752 			bw_est = rack_get_bw(rack);
17753 			rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped);
17754 		} else {
17755 			uint32_t gain, rate_set = 0;
17756 
17757 			rate_wanted = min(rack->rc_tp->snd_cwnd, rack->r_ctl.cwnd_to_use);
17758 			rate_wanted = rack_arrive_at_discounted_rate(rack, rate_wanted, &rate_set, &gain);
17759 			if (rate_set == 0) {
17760 				if (rate_wanted > rack->rc_tp->snd_wnd)
17761 					rate_wanted = rack->rc_tp->snd_wnd;
17762 				/* Now lets make it into a b/w */
17763 				rate_wanted *= (uint64_t)HPTS_USEC_IN_SEC;
17764 				rate_wanted /= (uint64_t)rack->r_ctl.rc_last_us_rtt;
17765 			}
17766 			bw_est = rate_wanted;
17767 			rack_log_pacing_delay_calc(rack, rack->rc_tp->snd_cwnd,
17768 						   rack->r_ctl.cwnd_to_use,
17769 						   rate_wanted, bw_est,
17770 						   rack->r_ctl.rc_last_us_rtt,
17771 						   88, __LINE__, NULL, gain);
17772 		}
17773 		if ((bw_est == 0) || (rate_wanted == 0) ||
17774 		    ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) {
17775 			/*
17776 			 * No way yet to make a b/w estimate or
17777 			 * our raise is set incorrectly.
17778 			 */
17779 			goto old_method;
17780 		}
17781 		rack_rate_cap_bw(rack, &rate_wanted, &capped);
17782 		/* We need to account for all the overheads */
17783 		segs = (len + segsiz - 1) / segsiz;
17784 		/*
17785 		 * We need the diff between 1514 bytes (e-mtu with e-hdr)
17786 		 * and how much data we put in each packet. Yes this
17787 		 * means we may be off if we are larger than 1500 bytes
17788 		 * or smaller. But this just makes us more conservative.
17789 		 */
17790 
17791 		oh =  (tp->t_maxseg - segsiz) + sizeof(struct tcphdr);
17792 		if (rack->r_is_v6) {
17793 #ifdef INET6
17794 			oh += sizeof(struct ip6_hdr);
17795 #endif
17796 		} else {
17797 #ifdef INET
17798 			oh += sizeof(struct ip);
17799 #endif
17800 		}
17801 		/* We add a fixed 14 for the ethernet header */
17802 		oh += 14;
17803 		segs *= oh;
17804 		lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC;
17805 		res = lentim / rate_wanted;
17806 		slot = (uint32_t)res;
17807 		if (rack_hw_rate_min &&
17808 		    (rate_wanted < rack_hw_rate_min)) {
17809 			can_start_hw_pacing = 0;
17810 			if (rack->r_ctl.crte) {
17811 				/*
17812 				 * Ok we need to release it, we
17813 				 * have fallen too low.
17814 				 */
17815 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17816 				rack->r_ctl.crte = NULL;
17817 				rack->rack_attempt_hdwr_pace = 0;
17818 				rack->rack_hdrw_pacing = 0;
17819 			}
17820 		}
17821 		if (rack->r_ctl.crte &&
17822 		    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17823 			/*
17824 			 * We want more than the hardware can give us,
17825 			 * don't start any hw pacing.
17826 			 */
17827 			can_start_hw_pacing = 0;
17828 			if (rack->r_rack_hw_rate_caps == 0) {
17829 				/*
17830 				 * Ok we need to release it, we
17831 				 * want more than the card can give us and
17832 				 * no rate cap is in place. Set it up so
17833 				 * when we want less we can retry.
17834 				 */
17835 				tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17836 				rack->r_ctl.crte = NULL;
17837 				rack->rack_attempt_hdwr_pace = 0;
17838 				rack->rack_hdrw_pacing = 0;
17839 			}
17840 		}
17841 		if ((rack->r_ctl.crte != NULL) && (rack->rc_inp->inp_snd_tag == NULL)) {
17842 			/*
17843 			 * We lost our rate somehow, this can happen
17844 			 * if the interface changed underneath us.
17845 			 */
17846 			tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17847 			rack->r_ctl.crte = NULL;
17848 			/* Lets re-allow attempting to setup pacing */
17849 			rack->rack_hdrw_pacing = 0;
17850 			rack->rack_attempt_hdwr_pace = 0;
17851 			rack_log_hdwr_pacing(rack,
17852 					     rate_wanted, bw_est, __LINE__,
17853 					     0, 6);
17854 		}
17855 		prev_fill = rack->r_via_fill_cw;
17856 		if ((rack->rc_pace_to_cwnd) &&
17857 		    (capped == 0) &&
17858 		    (rack->dgp_on == 1) &&
17859 		    (rack->use_fixed_rate == 0) &&
17860 		    (rack->in_probe_rtt == 0) &&
17861 		    (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) {
17862 			/*
17863 			 * We want to pace at our rate *or* faster to
17864 			 * fill the cwnd to the max if its not full.
17865 			 */
17866 			slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0);
17867 			/* Re-check to make sure we are not exceeding our max b/w */
17868 			if ((rack->r_ctl.crte != NULL) &&
17869 			    (tcp_hw_highest_rate(rack->r_ctl.crte) < rate_wanted)) {
17870 				/*
17871 				 * We want more than the hardware can give us,
17872 				 * don't start any hw pacing.
17873 				 */
17874 				can_start_hw_pacing = 0;
17875 				if (rack->r_rack_hw_rate_caps == 0) {
17876 					/*
17877 					 * Ok we need to release it, we
17878 					 * want more than the card can give us and
17879 					 * no rate cap is in place. Set it up so
17880 					 * when we want less we can retry.
17881 					 */
17882 					tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17883 					rack->r_ctl.crte = NULL;
17884 					rack->rack_attempt_hdwr_pace = 0;
17885 					rack->rack_hdrw_pacing = 0;
17886 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
17887 				}
17888 			}
17889 		}
17890 		if ((rack->rc_inp->inp_route.ro_nh != NULL) &&
17891 		    (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) {
17892 			if ((rack->rack_hdw_pace_ena) &&
17893 			    (can_start_hw_pacing > 0) &&
17894 			    (rack->rack_hdrw_pacing == 0) &&
17895 			    (rack->rack_attempt_hdwr_pace == 0)) {
17896 				/*
17897 				 * Lets attempt to turn on hardware pacing
17898 				 * if we can.
17899 				 */
17900 				rack->rack_attempt_hdwr_pace = 1;
17901 				rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp,
17902 								       rack->rc_inp->inp_route.ro_nh->nh_ifp,
17903 								       rate_wanted,
17904 								       RS_PACING_GEQ,
17905 								       &err, &rack->r_ctl.crte_prev_rate);
17906 				if (rack->r_ctl.crte) {
17907 					rack->rack_hdrw_pacing = 1;
17908 					rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted, segsiz,
17909 													   pace_one, rack->r_ctl.crte,
17910 													   NULL, rack->r_ctl.pace_len_divisor);
17911 					rack_log_hdwr_pacing(rack,
17912 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17913 							     err, 0);
17914 					rack->r_ctl.last_hw_bw_req = rate_wanted;
17915 				} else {
17916 					counter_u64_add(rack_hw_pace_init_fail, 1);
17917 				}
17918 			} else if (rack->rack_hdrw_pacing &&
17919 				   (rack->r_ctl.last_hw_bw_req != rate_wanted)) {
17920 				/* Do we need to adjust our rate? */
17921 				const struct tcp_hwrate_limit_table *nrte;
17922 
17923 				if (rack->r_up_only &&
17924 				    (rate_wanted < rack->r_ctl.crte->rate)) {
17925 					/**
17926 					 * We have four possible states here
17927 					 * having to do with the previous time
17928 					 * and this time.
17929 					 *   previous  |  this-time
17930 					 * A)     0      |     0   -- fill_cw not in the picture
17931 					 * B)     1      |     0   -- we were doing a fill-cw but now are not
17932 					 * C)     1      |     1   -- all rates from fill_cw
17933 					 * D)     0      |     1   -- we were doing non-fill and now we are filling
17934 					 *
17935 					 * For case A, C and D we don't allow a drop. But for
17936 					 * case B where we now our on our steady rate we do
17937 					 * allow a drop.
17938 					 *
17939 					 */
17940 					if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0)))
17941 						goto done_w_hdwr;
17942 				}
17943 				if ((rate_wanted > rack->r_ctl.crte->rate) ||
17944 				    (rate_wanted <= rack->r_ctl.crte_prev_rate)) {
17945 					if (rack_hw_rate_to_low &&
17946 					    (bw_est < rack_hw_rate_to_low)) {
17947 						/*
17948 						 * The pacing rate is too low for hardware, but
17949 						 * do allow hardware pacing to be restarted.
17950 						 */
17951 						rack_log_hdwr_pacing(rack,
17952 								     bw_est, rack->r_ctl.crte->rate, __LINE__,
17953 								     0, 5);
17954 						tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp);
17955 						rack->r_ctl.crte = NULL;
17956 						rack->rack_attempt_hdwr_pace = 0;
17957 						rack->rack_hdrw_pacing = 0;
17958 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17959 						goto done_w_hdwr;
17960 					}
17961 					nrte = tcp_chg_pacing_rate(rack->r_ctl.crte,
17962 								   rack->rc_tp,
17963 								   rack->rc_inp->inp_route.ro_nh->nh_ifp,
17964 								   rate_wanted,
17965 								   RS_PACING_GEQ,
17966 								   &err, &rack->r_ctl.crte_prev_rate);
17967 					if (nrte == NULL) {
17968 						/*
17969 						 * Lost the rate, lets drop hardware pacing
17970 						 * period.
17971 						 */
17972 						rack->rack_hdrw_pacing = 0;
17973 						rack->r_ctl.crte = NULL;
17974 						rack_log_hdwr_pacing(rack,
17975 								     rate_wanted, 0, __LINE__,
17976 								     err, 1);
17977 						rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17978 						counter_u64_add(rack_hw_pace_lost, 1);
17979 					} else if (nrte != rack->r_ctl.crte) {
17980 						rack->r_ctl.crte = nrte;
17981 						rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size_w_divisor(tp, rate_wanted,
17982 														   segsiz, pace_one, rack->r_ctl.crte,
17983 														   NULL, rack->r_ctl.pace_len_divisor);
17984 						rack_log_hdwr_pacing(rack,
17985 								     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17986 								     err, 2);
17987 						rack->r_ctl.last_hw_bw_req = rate_wanted;
17988 					}
17989 				} else {
17990 					/* We just need to adjust the segment size */
17991 					rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted);
17992 					rack_log_hdwr_pacing(rack,
17993 							     rate_wanted, rack->r_ctl.crte->rate, __LINE__,
17994 							     0, 4);
17995 					rack->r_ctl.last_hw_bw_req = rate_wanted;
17996 				}
17997 			}
17998 		}
17999 		if (minslot && (minslot > slot)) {
18000 			rack_log_pacing_delay_calc(rack, minslot, slot, rack->r_ctl.crte->rate, bw_est, lentim,
18001 						   98, __LINE__, NULL, 0);
18002 			slot = minslot;
18003 		}
18004 	done_w_hdwr:
18005 		if (rack_limit_time_with_srtt &&
18006 		    (rack->use_fixed_rate == 0) &&
18007 		    (rack->rack_hdrw_pacing == 0)) {
18008 			/*
18009 			 * Sanity check, we do not allow the pacing delay
18010 			 * to be longer than the SRTT of the path. If it is
18011 			 * a slow path, then adding a packet should increase
18012 			 * the RTT and compensate for this i.e. the srtt will
18013 			 * be greater so the allowed pacing time will be greater.
18014 			 *
18015 			 * Note this restriction is not for where a peak rate
18016 			 * is set, we are doing fixed pacing or hardware pacing.
18017 			 */
18018 			if (rack->rc_tp->t_srtt)
18019 				srtt = rack->rc_tp->t_srtt;
18020 			else
18021 				srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC;	/* its in ms convert */
18022 			if (srtt < (uint64_t)slot) {
18023 				rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0);
18024 				slot = srtt;
18025 			}
18026 		}
18027 		/*******************************************************************/
18028 		/* RRS: We insert paced call to stats here for len and rate_wanted */
18029 		/*******************************************************************/
18030 		rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0);
18031 	}
18032 	if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) {
18033 		/*
18034 		 * If this rate is seeing enobufs when it
18035 		 * goes to send then either the nic is out
18036 		 * of gas or we are mis-estimating the time
18037 		 * somehow and not letting the queue empty
18038 		 * completely. Lets add to the pacing time.
18039 		 */
18040 		int hw_boost_delay;
18041 
18042 		hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult;
18043 		if (hw_boost_delay > rack_enobuf_hw_max)
18044 			hw_boost_delay = rack_enobuf_hw_max;
18045 		else if (hw_boost_delay < rack_enobuf_hw_min)
18046 			hw_boost_delay = rack_enobuf_hw_min;
18047 		slot += hw_boost_delay;
18048 	}
18049 	return (slot);
18050 }
18051 
18052 static void
18053 rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack,
18054     tcp_seq startseq, uint32_t sb_offset)
18055 {
18056 	struct rack_sendmap *my_rsm = NULL;
18057 
18058 	if (tp->t_state < TCPS_ESTABLISHED) {
18059 		/*
18060 		 * We don't start any measurements if we are
18061 		 * not at least established.
18062 		 */
18063 		return;
18064 	}
18065 	if (tp->t_state >= TCPS_FIN_WAIT_1) {
18066 		/*
18067 		 * We will get no more data into the SB
18068 		 * this means we need to have the data available
18069 		 * before we start a measurement.
18070 		 */
18071 
18072 		if (sbavail(&tptosocket(tp)->so_snd) <
18073 		    max(rc_init_window(rack),
18074 			(MIN_GP_WIN * ctf_fixed_maxseg(tp)))) {
18075 			/* Nope not enough data */
18076 			return;
18077 		}
18078 	}
18079 	tp->t_flags |= TF_GPUTINPROG;
18080 	rack->r_ctl.rc_gp_cumack_ts = 0;
18081 	rack->r_ctl.rc_gp_lowrtt = 0xffffffff;
18082 	rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd;
18083 	tp->gput_seq = startseq;
18084 	rack->app_limited_needs_set = 0;
18085 	if (rack->in_probe_rtt)
18086 		rack->measure_saw_probe_rtt = 1;
18087 	else if ((rack->measure_saw_probe_rtt) &&
18088 		 (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit)))
18089 		rack->measure_saw_probe_rtt = 0;
18090 	if (rack->rc_gp_filled)
18091 		tp->gput_ts = rack->r_ctl.last_cumack_advance;
18092 	else {
18093 		/* Special case initial measurement */
18094 		struct timeval tv;
18095 
18096 		tp->gput_ts = tcp_get_usecs(&tv);
18097 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18098 	}
18099 	/*
18100 	 * We take a guess out into the future,
18101 	 * if we have no measurement and no
18102 	 * initial rate, we measure the first
18103 	 * initial-windows worth of data to
18104 	 * speed up getting some GP measurement and
18105 	 * thus start pacing.
18106 	 */
18107 	if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) {
18108 		rack->app_limited_needs_set = 1;
18109 		tp->gput_ack = startseq + max(rc_init_window(rack),
18110 					      (MIN_GP_WIN * ctf_fixed_maxseg(tp)));
18111 		rack_log_pacing_delay_calc(rack,
18112 					   tp->gput_seq,
18113 					   tp->gput_ack,
18114 					   0,
18115 					   tp->gput_ts,
18116 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18117 					   9,
18118 					   __LINE__, NULL, 0);
18119 		rack_tend_gp_marks(tp, rack);
18120 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18121 		return;
18122 	}
18123 	if (sb_offset) {
18124 		/*
18125 		 * We are out somewhere in the sb
18126 		 * can we use the already outstanding data?
18127 		 */
18128 
18129 		if (rack->r_ctl.rc_app_limited_cnt == 0) {
18130 			/*
18131 			 * Yes first one is good and in this case
18132 			 * the tp->gput_ts is correctly set based on
18133 			 * the last ack that arrived (no need to
18134 			 * set things up when an ack comes in).
18135 			 */
18136 			my_rsm = tqhash_min(rack->r_ctl.tqh);
18137 			if ((my_rsm == NULL) ||
18138 			    (my_rsm->r_rtr_cnt != 1)) {
18139 				/* retransmission? */
18140 				goto use_latest;
18141 			}
18142 		} else {
18143 			if (rack->r_ctl.rc_first_appl == NULL) {
18144 				/*
18145 				 * If rc_first_appl is NULL
18146 				 * then the cnt should be 0.
18147 				 * This is probably an error, maybe
18148 				 * a KASSERT would be approprate.
18149 				 */
18150 				goto use_latest;
18151 			}
18152 			/*
18153 			 * If we have a marker pointer to the last one that is
18154 			 * app limited we can use that, but we need to set
18155 			 * things up so that when it gets ack'ed we record
18156 			 * the ack time (if its not already acked).
18157 			 */
18158 			rack->app_limited_needs_set = 1;
18159 			/*
18160 			 * We want to get to the rsm that is either
18161 			 * next with space i.e. over 1 MSS or the one
18162 			 * after that (after the app-limited).
18163 			 */
18164 			my_rsm = tqhash_next(rack->r_ctl.tqh, rack->r_ctl.rc_first_appl);
18165 			if (my_rsm) {
18166 				if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp))
18167 					/* Have to use the next one */
18168 					my_rsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
18169 				else {
18170 					/* Use after the first MSS of it is acked */
18171 					tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp);
18172 					goto start_set;
18173 				}
18174 			}
18175 			if ((my_rsm == NULL) ||
18176 			    (my_rsm->r_rtr_cnt != 1)) {
18177 				/*
18178 				 * Either its a retransmit or
18179 				 * the last is the app-limited one.
18180 				 */
18181 				goto use_latest;
18182 			}
18183 		}
18184 		tp->gput_seq = my_rsm->r_start;
18185 start_set:
18186 		if (my_rsm->r_flags & RACK_ACKED) {
18187 			/*
18188 			 * This one has been acked use the arrival ack time
18189 			 */
18190 			struct rack_sendmap *nrsm;
18191 
18192 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
18193 			rack->app_limited_needs_set = 0;
18194 			/*
18195 			 * Ok in this path we need to use the r_end now
18196 			 * since this guy is the starting ack.
18197 			 */
18198 			tp->gput_seq = my_rsm->r_end;
18199 			/*
18200 			 * We also need to adjust up the sendtime
18201 			 * to the send of the next data after my_rsm.
18202 			 */
18203 			nrsm = tqhash_next(rack->r_ctl.tqh, my_rsm);
18204 			if (nrsm != NULL)
18205 				my_rsm = nrsm;
18206 			else {
18207 				/*
18208 				 * The next as not been sent, thats the
18209 				 * case for using the latest.
18210 				 */
18211 				goto use_latest;
18212 			}
18213 		}
18214 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18215 		tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack);
18216 		rack->r_ctl.rc_gp_cumack_ts = 0;
18217 		rack_log_pacing_delay_calc(rack,
18218 					   tp->gput_seq,
18219 					   tp->gput_ack,
18220 					   (uint64_t)my_rsm,
18221 					   tp->gput_ts,
18222 					   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18223 					   9,
18224 					   __LINE__, my_rsm, 0);
18225 		/* Now lets make sure all are marked as they should be */
18226 		rack_tend_gp_marks(tp, rack);
18227 		rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18228 		return;
18229 	}
18230 
18231 use_latest:
18232 	/*
18233 	 * We don't know how long we may have been
18234 	 * idle or if this is the first-send. Lets
18235 	 * setup the flag so we will trim off
18236 	 * the first ack'd data so we get a true
18237 	 * measurement.
18238 	 */
18239 	rack->app_limited_needs_set = 1;
18240 	tp->gput_ack = startseq + rack_get_measure_window(tp, rack);
18241 	rack->r_ctl.rc_gp_cumack_ts = 0;
18242 	/* Find this guy so we can pull the send time */
18243 	my_rsm = tqhash_find(rack->r_ctl.tqh, startseq);
18244 	if (my_rsm) {
18245 		rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[0];
18246 		if (my_rsm->r_flags & RACK_ACKED) {
18247 			/*
18248 			 * Unlikely since its probably what was
18249 			 * just transmitted (but I am paranoid).
18250 			 */
18251 			tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival;
18252 			rack->app_limited_needs_set = 0;
18253 		}
18254 		if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) {
18255 			/* This also is unlikely */
18256 			tp->gput_seq = my_rsm->r_start;
18257 		}
18258 	} else {
18259 		/*
18260 		 * TSNH unless we have some send-map limit,
18261 		 * and even at that it should not be hitting
18262 		 * that limit (we should have stopped sending).
18263 		 */
18264 		struct timeval tv;
18265 
18266 		microuptime(&tv);
18267 		rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv);
18268 	}
18269 	rack_tend_gp_marks(tp, rack);
18270 	rack_log_pacing_delay_calc(rack,
18271 				   tp->gput_seq,
18272 				   tp->gput_ack,
18273 				   (uint64_t)my_rsm,
18274 				   tp->gput_ts,
18275 				   (((uint64_t)rack->r_ctl.rc_app_limited_cnt << 32) | (uint64_t)rack->r_ctl.rc_gp_output_ts),
18276 				   9, __LINE__, NULL, 0);
18277 	rack_log_gpset(rack, tp->gput_ack, 0, 0, __LINE__, 1, NULL);
18278 }
18279 
18280 static inline uint32_t
18281 rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack,  uint32_t cwnd_to_use,
18282     uint32_t avail, int32_t sb_offset)
18283 {
18284 	uint32_t len;
18285 	uint32_t sendwin;
18286 
18287 	if (tp->snd_wnd > cwnd_to_use)
18288 		sendwin = cwnd_to_use;
18289 	else
18290 		sendwin = tp->snd_wnd;
18291 	if (ctf_outstanding(tp) >= tp->snd_wnd) {
18292 		/* We never want to go over our peers rcv-window */
18293 		len = 0;
18294 	} else {
18295 		uint32_t flight;
18296 
18297 		flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked);
18298 		if (flight >= sendwin) {
18299 			/*
18300 			 * We have in flight what we are allowed by cwnd (if
18301 			 * it was rwnd blocking it would have hit above out
18302 			 * >= tp->snd_wnd).
18303 			 */
18304 			return (0);
18305 		}
18306 		len = sendwin - flight;
18307 		if ((len + ctf_outstanding(tp)) > tp->snd_wnd) {
18308 			/* We would send too much (beyond the rwnd) */
18309 			len = tp->snd_wnd - ctf_outstanding(tp);
18310 		}
18311 		if ((len + sb_offset) > avail) {
18312 			/*
18313 			 * We don't have that much in the SB, how much is
18314 			 * there?
18315 			 */
18316 			len = avail - sb_offset;
18317 		}
18318 	}
18319 	return (len);
18320 }
18321 
18322 static void
18323 rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags,
18324 	     unsigned ipoptlen, int32_t orig_len, int32_t len, int error,
18325 	     int rsm_is_null, int optlen, int line, uint16_t mode)
18326 {
18327 	if (rack_verbose_logging && tcp_bblogging_on(rack->rc_tp)) {
18328 		union tcp_log_stackspecific log;
18329 		struct timeval tv;
18330 
18331 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18332 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18333 		log.u_bbr.flex1 = error;
18334 		log.u_bbr.flex2 = flags;
18335 		log.u_bbr.flex3 = rsm_is_null;
18336 		log.u_bbr.flex4 = ipoptlen;
18337 		log.u_bbr.flex5 = tp->rcv_numsacks;
18338 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
18339 		log.u_bbr.flex7 = optlen;
18340 		log.u_bbr.flex8 = rack->r_fsb_inited;
18341 		log.u_bbr.applimited = rack->r_fast_output;
18342 		log.u_bbr.bw_inuse = rack_get_bw(rack);
18343 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
18344 		log.u_bbr.cwnd_gain = mode;
18345 		log.u_bbr.pkts_out = orig_len;
18346 		log.u_bbr.lt_epoch = len;
18347 		log.u_bbr.delivered = line;
18348 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
18349 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18350 		tcp_log_event(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0,
18351 			       len, &log, false, NULL, __func__, __LINE__, &tv);
18352 	}
18353 }
18354 
18355 
18356 static struct mbuf *
18357 rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen,
18358 		   struct rack_fast_send_blk *fsb,
18359 		   int32_t seglimit, int32_t segsize, int hw_tls)
18360 {
18361 #ifdef KERN_TLS
18362 	struct ktls_session *tls, *ntls;
18363 #ifdef INVARIANTS
18364 	struct mbuf *start;
18365 #endif
18366 #endif
18367 	struct mbuf *m, *n, **np, *smb;
18368 	struct mbuf *top;
18369 	int32_t off, soff;
18370 	int32_t len = *plen;
18371 	int32_t fragsize;
18372 	int32_t len_cp = 0;
18373 	uint32_t mlen, frags;
18374 
18375 	soff = off = the_off;
18376 	smb = m = the_m;
18377 	np = &top;
18378 	top = NULL;
18379 #ifdef KERN_TLS
18380 	if (hw_tls && (m->m_flags & M_EXTPG))
18381 		tls = m->m_epg_tls;
18382 	else
18383 		tls = NULL;
18384 #ifdef INVARIANTS
18385 	start = m;
18386 #endif
18387 #endif
18388 	while (len > 0) {
18389 		if (m == NULL) {
18390 			*plen = len_cp;
18391 			break;
18392 		}
18393 #ifdef KERN_TLS
18394 		if (hw_tls) {
18395 			if (m->m_flags & M_EXTPG)
18396 				ntls = m->m_epg_tls;
18397 			else
18398 				ntls = NULL;
18399 
18400 			/*
18401 			 * Avoid mixing TLS records with handshake
18402 			 * data or TLS records from different
18403 			 * sessions.
18404 			 */
18405 			if (tls != ntls) {
18406 				MPASS(m != start);
18407 				*plen = len_cp;
18408 				break;
18409 			}
18410 		}
18411 #endif
18412 		mlen = min(len, m->m_len - off);
18413 		if (seglimit) {
18414 			/*
18415 			 * For M_EXTPG mbufs, add 3 segments
18416 			 * + 1 in case we are crossing page boundaries
18417 			 * + 2 in case the TLS hdr/trailer are used
18418 			 * It is cheaper to just add the segments
18419 			 * than it is to take the cache miss to look
18420 			 * at the mbuf ext_pgs state in detail.
18421 			 */
18422 			if (m->m_flags & M_EXTPG) {
18423 				fragsize = min(segsize, PAGE_SIZE);
18424 				frags = 3;
18425 			} else {
18426 				fragsize = segsize;
18427 				frags = 0;
18428 			}
18429 
18430 			/* Break if we really can't fit anymore. */
18431 			if ((frags + 1) >= seglimit) {
18432 				*plen =	len_cp;
18433 				break;
18434 			}
18435 
18436 			/*
18437 			 * Reduce size if you can't copy the whole
18438 			 * mbuf. If we can't copy the whole mbuf, also
18439 			 * adjust len so the loop will end after this
18440 			 * mbuf.
18441 			 */
18442 			if ((frags + howmany(mlen, fragsize)) >= seglimit) {
18443 				mlen = (seglimit - frags - 1) * fragsize;
18444 				len = mlen;
18445 				*plen = len_cp + len;
18446 			}
18447 			frags += howmany(mlen, fragsize);
18448 			if (frags == 0)
18449 				frags++;
18450 			seglimit -= frags;
18451 			KASSERT(seglimit > 0,
18452 			    ("%s: seglimit went too low", __func__));
18453 		}
18454 		n = m_get(M_NOWAIT, m->m_type);
18455 		*np = n;
18456 		if (n == NULL)
18457 			goto nospace;
18458 		n->m_len = mlen;
18459 		soff += mlen;
18460 		len_cp += n->m_len;
18461 		if (m->m_flags & (M_EXT | M_EXTPG)) {
18462 			n->m_data = m->m_data + off;
18463 			mb_dupcl(n, m);
18464 		} else {
18465 			bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
18466 			    (u_int)n->m_len);
18467 		}
18468 		len -= n->m_len;
18469 		off = 0;
18470 		m = m->m_next;
18471 		np = &n->m_next;
18472 		if (len || (soff == smb->m_len)) {
18473 			/*
18474 			 * We have more so we move forward  or
18475 			 * we have consumed the entire mbuf and
18476 			 * len has fell to 0.
18477 			 */
18478 			soff = 0;
18479 			smb = m;
18480 		}
18481 
18482 	}
18483 	if (fsb != NULL) {
18484 		fsb->m = smb;
18485 		fsb->off = soff;
18486 		if (smb) {
18487 			/*
18488 			 * Save off the size of the mbuf. We do
18489 			 * this so that we can recognize when it
18490 			 * has been trimmed by sbcut() as acks
18491 			 * come in.
18492 			 */
18493 			fsb->o_m_len = smb->m_len;
18494 			fsb->o_t_len = M_TRAILINGROOM(smb);
18495 		} else {
18496 			/*
18497 			 * This is the case where the next mbuf went to NULL. This
18498 			 * means with this copy we have sent everything in the sb.
18499 			 * In theory we could clear the fast_output flag, but lets
18500 			 * not since its possible that we could get more added
18501 			 * and acks that call the extend function which would let
18502 			 * us send more.
18503 			 */
18504 			fsb->o_m_len = 0;
18505 			fsb->o_t_len = 0;
18506 		}
18507 	}
18508 	return (top);
18509 nospace:
18510 	if (top)
18511 		m_freem(top);
18512 	return (NULL);
18513 
18514 }
18515 
18516 /*
18517  * This is a copy of m_copym(), taking the TSO segment size/limit
18518  * constraints into account, and advancing the sndptr as it goes.
18519  */
18520 static struct mbuf *
18521 rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen,
18522 		int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff)
18523 {
18524 	struct mbuf *m, *n;
18525 	int32_t soff;
18526 
18527 	m = rack->r_ctl.fsb.m;
18528 	if (M_TRAILINGROOM(m) != rack->r_ctl.fsb.o_t_len) {
18529 		/*
18530 		 * The trailing space changed, mbufs can grow
18531 		 * at the tail but they can't shrink from
18532 		 * it, KASSERT that. Adjust the orig_m_len to
18533 		 * compensate for this change.
18534 		 */
18535 		KASSERT((rack->r_ctl.fsb.o_t_len > M_TRAILINGROOM(m)),
18536 			("mbuf:%p rack:%p trailing_space:%jd ots:%u oml:%u mlen:%u\n",
18537 			 m,
18538 			 rack,
18539 			 (intmax_t)M_TRAILINGROOM(m),
18540 			 rack->r_ctl.fsb.o_t_len,
18541 			 rack->r_ctl.fsb.o_m_len,
18542 			 m->m_len));
18543 		rack->r_ctl.fsb.o_m_len += (rack->r_ctl.fsb.o_t_len - M_TRAILINGROOM(m));
18544 		rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(m);
18545 	}
18546 	if (m->m_len < rack->r_ctl.fsb.o_m_len) {
18547 		/*
18548 		 * Mbuf shrank, trimmed off the top by an ack, our
18549 		 * offset changes.
18550 		 */
18551 		KASSERT((rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len - m->m_len)),
18552 			("mbuf:%p len:%u rack:%p oml:%u soff:%u\n",
18553 			 m, m->m_len,
18554 			 rack, rack->r_ctl.fsb.o_m_len,
18555 			 rack->r_ctl.fsb.off));
18556 
18557 		if (rack->r_ctl.fsb.off >= (rack->r_ctl.fsb.o_m_len- m->m_len))
18558 			rack->r_ctl.fsb.off -= (rack->r_ctl.fsb.o_m_len - m->m_len);
18559 		else
18560 			rack->r_ctl.fsb.off = 0;
18561 		rack->r_ctl.fsb.o_m_len = m->m_len;
18562 #ifdef INVARIANTS
18563 	} else if (m->m_len > rack->r_ctl.fsb.o_m_len) {
18564 		panic("rack:%p m:%p m_len grew outside of t_space compensation",
18565 		      rack, m);
18566 #endif
18567 	}
18568 	soff = rack->r_ctl.fsb.off;
18569 	KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff));
18570 	KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen));
18571 	KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?",
18572 				 __FUNCTION__,
18573 				 rack, *plen, m, m->m_len));
18574 	/* Save off the right location before we copy and advance */
18575 	*s_soff = soff;
18576 	*s_mb = rack->r_ctl.fsb.m;
18577 	n = rack_fo_base_copym(m, soff, plen,
18578 			       &rack->r_ctl.fsb,
18579 			       seglimit, segsize, rack->r_ctl.fsb.hw_tls);
18580 	return (n);
18581 }
18582 
18583 /* Log the buffer level */
18584 static void
18585 rack_log_queue_level(struct tcpcb *tp, struct tcp_rack *rack,
18586 		     int len, struct timeval *tv,
18587 		     uint32_t cts)
18588 {
18589 	uint32_t p_rate = 0, p_queue = 0, err = 0;
18590 	union tcp_log_stackspecific log;
18591 
18592 #ifdef RATELIMIT
18593 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18594 	err = in_pcbquery_txrtlmt(rack->rc_inp,	&p_rate);
18595 #endif
18596 	memset(&log.u_bbr, 0, sizeof(log.u_bbr));
18597 	log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18598 	log.u_bbr.flex1 = p_rate;
18599 	log.u_bbr.flex2 = p_queue;
18600 	log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18601 	log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18602 	log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18603 	log.u_bbr.flex7 = 99;
18604 	log.u_bbr.flex8 = 0;
18605 	log.u_bbr.pkts_out = err;
18606 	log.u_bbr.delRate = rack->r_ctl.crte->rate;
18607 	log.u_bbr.timeStamp = cts;
18608 	log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18609 	tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18610 		       len, &log, false, NULL, __func__, __LINE__, tv);
18611 
18612 }
18613 
18614 static uint32_t
18615 rack_check_queue_level(struct tcp_rack *rack, struct tcpcb *tp,
18616 		       struct timeval *tv, uint32_t cts, int len, uint32_t segsiz)
18617 {
18618 	uint64_t lentime = 0;
18619 #ifdef RATELIMIT
18620 	uint32_t p_rate = 0, p_queue = 0, err;
18621 	union tcp_log_stackspecific log;
18622 	uint64_t bw;
18623 
18624 	err = in_pcbquery_txrlevel(rack->rc_inp, &p_queue);
18625 	/* Failed or queue is zero */
18626 	if (err || (p_queue == 0)) {
18627 		lentime = 0;
18628 		goto out;
18629 	}
18630 	err = in_pcbquery_txrtlmt(rack->rc_inp, &p_rate);
18631 	if (err) {
18632 		lentime = 0;
18633 		goto out;
18634 	}
18635 	/*
18636 	 * If we reach here we have some bytes in
18637 	 * the queue. The number returned is a value
18638 	 * between 0 and 0xffff where ffff is full
18639 	 * and 0 is empty. So how best to make this into
18640 	 * something usable?
18641 	 *
18642 	 * The "safer" way is lets take the b/w gotten
18643 	 * from the query (which should be our b/w rate)
18644 	 * and pretend that a full send (our rc_pace_max_segs)
18645 	 * is outstanding. We factor it so its as if a full
18646 	 * number of our MSS segment is terms of full
18647 	 * ethernet segments are outstanding.
18648 	 */
18649 	bw = p_rate / 8;
18650 	if (bw) {
18651 		lentime = (rack->r_ctl.rc_pace_max_segs / segsiz);
18652 		lentime *= ETHERNET_SEGMENT_SIZE;
18653 		lentime *= (uint64_t)HPTS_USEC_IN_SEC;
18654 		lentime /= bw;
18655 	} else {
18656 		/* TSNH -- KASSERT? */
18657 		lentime = 0;
18658 	}
18659 out:
18660 	if (tcp_bblogging_on(tp)) {
18661 		memset(&log, 0, sizeof(log));
18662 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
18663 		log.u_bbr.flex1 = p_rate;
18664 		log.u_bbr.flex2 = p_queue;
18665 		log.u_bbr.flex4 = (uint32_t)rack->r_ctl.crte->using;
18666 		log.u_bbr.flex5 = (uint32_t)rack->r_ctl.crte->rs_num_enobufs;
18667 		log.u_bbr.flex6 = rack->r_ctl.crte->time_between;
18668 		log.u_bbr.flex7 = 99;
18669 		log.u_bbr.flex8 = 0;
18670 		log.u_bbr.pkts_out = err;
18671 		log.u_bbr.delRate = rack->r_ctl.crte->rate;
18672 		log.u_bbr.cur_del_rate = lentime;
18673 		log.u_bbr.timeStamp = cts;
18674 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
18675 		tcp_log_event(tp, NULL, NULL, NULL, BBR_LOG_HDWR_PACE, 0,
18676 			       len, &log, false, NULL, __func__, __LINE__,tv);
18677 	}
18678 #endif
18679 	return ((uint32_t)lentime);
18680 }
18681 
18682 static int
18683 rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm,
18684 		     uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp)
18685 {
18686 	/*
18687 	 * Enter the fast retransmit path. We are given that a sched_pin is
18688 	 * in place (if accounting is compliled in) and the cycle count taken
18689 	 * at the entry is in the ts_val. The concept her is that the rsm
18690 	 * now holds the mbuf offsets and such so we can directly transmit
18691 	 * without a lot of overhead, the len field is already set for
18692 	 * us to prohibit us from sending too much (usually its 1MSS).
18693 	 */
18694 	struct ip *ip = NULL;
18695 	struct udphdr *udp = NULL;
18696 	struct tcphdr *th = NULL;
18697 	struct mbuf *m = NULL;
18698 	struct inpcb *inp;
18699 	uint8_t *cpto;
18700 	struct tcp_log_buffer *lgb;
18701 #ifdef TCP_ACCOUNTING
18702 	uint64_t crtsc;
18703 	int cnt_thru = 1;
18704 #endif
18705 	struct tcpopt to;
18706 	u_char opt[TCP_MAXOLEN];
18707 	uint32_t hdrlen, optlen;
18708 	int32_t slot, segsiz, max_val, tso = 0, error = 0, ulen = 0;
18709 	uint16_t flags;
18710 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
18711 	uint32_t if_hw_tsomaxsegsize;
18712 	int32_t ip_sendflag = IP_NO_SND_TAG_RL;
18713 
18714 #ifdef INET6
18715 	struct ip6_hdr *ip6 = NULL;
18716 
18717 	if (rack->r_is_v6) {
18718 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
18719 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
18720 	} else
18721 #endif				/* INET6 */
18722 	{
18723 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
18724 		hdrlen = sizeof(struct tcpiphdr);
18725 	}
18726 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
18727 		goto failed;
18728 	}
18729 	if (doing_tlp) {
18730 		/* Its a TLP add the flag, it may already be there but be sure */
18731 		rsm->r_flags |= RACK_TLP;
18732 	} else {
18733 		/* If it was a TLP it is not not on this retransmit */
18734 		rsm->r_flags &= ~RACK_TLP;
18735 	}
18736 	startseq = rsm->r_start;
18737 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
18738 	inp = rack->rc_inp;
18739 	to.to_flags = 0;
18740 	flags = tcp_outflags[tp->t_state];
18741 	if (flags & (TH_SYN|TH_RST)) {
18742 		goto failed;
18743 	}
18744 	if (rsm->r_flags & RACK_HAS_FIN) {
18745 		/* We can't send a FIN here */
18746 		goto failed;
18747 	}
18748 	if (flags & TH_FIN) {
18749 		/* We never send a FIN */
18750 		flags &= ~TH_FIN;
18751 	}
18752 	if (tp->t_flags & TF_RCVD_TSTMP) {
18753 		to.to_tsval = ms_cts + tp->ts_offset;
18754 		to.to_tsecr = tp->ts_recent;
18755 		to.to_flags = TOF_TS;
18756 	}
18757 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18758 	/* TCP-MD5 (RFC2385). */
18759 	if (tp->t_flags & TF_SIGNATURE)
18760 		to.to_flags |= TOF_SIGNATURE;
18761 #endif
18762 	optlen = tcp_addoptions(&to, opt);
18763 	hdrlen += optlen;
18764 	udp = rack->r_ctl.fsb.udp;
18765 	if (udp)
18766 		hdrlen += sizeof(struct udphdr);
18767 	if (rack->r_ctl.rc_pace_max_segs)
18768 		max_val = rack->r_ctl.rc_pace_max_segs;
18769 	else if (rack->rc_user_set_max_segs)
18770 		max_val = rack->rc_user_set_max_segs * segsiz;
18771 	else
18772 		max_val = len;
18773 	if ((tp->t_flags & TF_TSO) &&
18774 	    V_tcp_do_tso &&
18775 	    (len > segsiz) &&
18776 	    (tp->t_port == 0))
18777 		tso = 1;
18778 #ifdef INET6
18779 	if (MHLEN < hdrlen + max_linkhdr)
18780 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
18781 	else
18782 #endif
18783 		m = m_gethdr(M_NOWAIT, MT_DATA);
18784 	if (m == NULL)
18785 		goto failed;
18786 	m->m_data += max_linkhdr;
18787 	m->m_len = hdrlen;
18788 	th = rack->r_ctl.fsb.th;
18789 	/* Establish the len to send */
18790 	if (len > max_val)
18791 		len = max_val;
18792 	if ((tso) && (len + optlen > segsiz)) {
18793 		uint32_t if_hw_tsomax;
18794 		int32_t max_len;
18795 
18796 		/* extract TSO information */
18797 		if_hw_tsomax = tp->t_tsomax;
18798 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
18799 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
18800 		/*
18801 		 * Check if we should limit by maximum payload
18802 		 * length:
18803 		 */
18804 		if (if_hw_tsomax != 0) {
18805 			/* compute maximum TSO length */
18806 			max_len = (if_hw_tsomax - hdrlen -
18807 				   max_linkhdr);
18808 			if (max_len <= 0) {
18809 				goto failed;
18810 			} else if (len > max_len) {
18811 				len = max_len;
18812 			}
18813 		}
18814 		if (len <= segsiz) {
18815 			/*
18816 			 * In case there are too many small fragments don't
18817 			 * use TSO:
18818 			 */
18819 			tso = 0;
18820 		}
18821 	} else {
18822 		tso = 0;
18823 	}
18824 	if ((tso == 0) && (len > segsiz))
18825 		len = segsiz;
18826 	(void)tcp_get_usecs(tv);
18827 	if ((len == 0) ||
18828 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
18829 		goto failed;
18830 	}
18831 	th->th_seq = htonl(rsm->r_start);
18832 	th->th_ack = htonl(tp->rcv_nxt);
18833 	/*
18834 	 * The PUSH bit should only be applied
18835 	 * if the full retransmission is made. If
18836 	 * we are sending less than this is the
18837 	 * left hand edge and should not have
18838 	 * the PUSH bit.
18839 	 */
18840 	if ((rsm->r_flags & RACK_HAD_PUSH) &&
18841 	    (len == (rsm->r_end - rsm->r_start)))
18842 		flags |= TH_PUSH;
18843 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
18844 	if (th->th_win == 0) {
18845 		tp->t_sndzerowin++;
18846 		tp->t_flags |= TF_RXWIN0SENT;
18847 	} else
18848 		tp->t_flags &= ~TF_RXWIN0SENT;
18849 	if (rsm->r_flags & RACK_TLP) {
18850 		/*
18851 		 * TLP should not count in retran count, but
18852 		 * in its own bin
18853 		 */
18854 		counter_u64_add(rack_tlp_retran, 1);
18855 		counter_u64_add(rack_tlp_retran_bytes, len);
18856 	} else {
18857 		tp->t_sndrexmitpack++;
18858 		KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
18859 		KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
18860 	}
18861 #ifdef STATS
18862 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
18863 				 len);
18864 #endif
18865 	if (rsm->m == NULL)
18866 		goto failed;
18867 	if (rsm->m &&
18868 	    ((rsm->orig_m_len != rsm->m->m_len) ||
18869 	     (M_TRAILINGROOM(rsm->m) != rsm->orig_t_space))) {
18870 		/* Fix up the orig_m_len and possibly the mbuf offset */
18871 		rack_adjust_orig_mlen(rsm);
18872 	}
18873 	m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls);
18874 	if (len <= segsiz) {
18875 		/*
18876 		 * Must have ran out of mbufs for the copy
18877 		 * shorten it to no longer need tso. Lets
18878 		 * not put on sendalot since we are low on
18879 		 * mbufs.
18880 		 */
18881 		tso = 0;
18882 	}
18883 	if ((m->m_next == NULL) || (len <= 0)){
18884 		goto failed;
18885 	}
18886 	if (udp) {
18887 		if (rack->r_is_v6)
18888 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
18889 		else
18890 			ulen = hdrlen + len - sizeof(struct ip);
18891 		udp->uh_ulen = htons(ulen);
18892 	}
18893 	m->m_pkthdr.rcvif = (struct ifnet *)0;
18894 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
18895 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
18896 		int ect = tcp_ecn_output_established(tp, &flags, len, true);
18897 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
18898 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
18899 		    tp->t_flags2 &= ~TF2_ECN_SND_ECE;
18900 #ifdef INET6
18901 		if (rack->r_is_v6) {
18902 		    ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
18903 		    ip6->ip6_flow |= htonl(ect << 20);
18904 		}
18905 		else
18906 #endif
18907 		{
18908 		    ip->ip_tos &= ~IPTOS_ECN_MASK;
18909 		    ip->ip_tos |= ect;
18910 		}
18911 	}
18912 	if (rack->r_ctl.crte != NULL) {
18913 		/* See if we can send via the hw queue */
18914 		slot = rack_check_queue_level(rack, tp, tv, cts, len, segsiz);
18915 		/* If there is nothing in queue (no pacing time) we can send via the hw queue */
18916 		if (slot == 0)
18917 			ip_sendflag = 0;
18918 	}
18919 	tcp_set_flags(th, flags);
18920 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
18921 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
18922 	if (to.to_flags & TOF_SIGNATURE) {
18923 		/*
18924 		 * Calculate MD5 signature and put it into the place
18925 		 * determined before.
18926 		 * NOTE: since TCP options buffer doesn't point into
18927 		 * mbuf's data, calculate offset and use it.
18928 		 */
18929 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
18930 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
18931 			/*
18932 			 * Do not send segment if the calculation of MD5
18933 			 * digest has failed.
18934 			 */
18935 			goto failed;
18936 		}
18937 	}
18938 #endif
18939 #ifdef INET6
18940 	if (rack->r_is_v6) {
18941 		if (tp->t_port) {
18942 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
18943 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18944 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
18945 			th->th_sum = htons(0);
18946 			UDPSTAT_INC(udps_opackets);
18947 		} else {
18948 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
18949 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18950 			th->th_sum = in6_cksum_pseudo(ip6,
18951 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
18952 						      0);
18953 		}
18954 	}
18955 #endif
18956 #if defined(INET6) && defined(INET)
18957 	else
18958 #endif
18959 #ifdef INET
18960 	{
18961 		if (tp->t_port) {
18962 			m->m_pkthdr.csum_flags = CSUM_UDP;
18963 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
18964 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
18965 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
18966 			th->th_sum = htons(0);
18967 			UDPSTAT_INC(udps_opackets);
18968 		} else {
18969 			m->m_pkthdr.csum_flags = CSUM_TCP;
18970 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
18971 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
18972 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
18973 									IPPROTO_TCP + len + optlen));
18974 		}
18975 		/* IP version must be set here for ipv4/ipv6 checking later */
18976 		KASSERT(ip->ip_v == IPVERSION,
18977 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
18978 	}
18979 #endif
18980 	if (tso) {
18981 		/*
18982 		 * Here we use segsiz since we have no added options besides
18983 		 * any standard timestamp options (no DSACKs or SACKS are sent
18984 		 * via either fast-path).
18985 		 */
18986 		KASSERT(len > segsiz,
18987 			("%s: len <= tso_segsz tp:%p", __func__, tp));
18988 		m->m_pkthdr.csum_flags |= CSUM_TSO;
18989 		m->m_pkthdr.tso_segsz = segsiz;
18990 	}
18991 #ifdef INET6
18992 	if (rack->r_is_v6) {
18993 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
18994 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
18995 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
18996 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
18997 		else
18998 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
18999 	}
19000 #endif
19001 #if defined(INET) && defined(INET6)
19002 	else
19003 #endif
19004 #ifdef INET
19005 	{
19006 		ip->ip_len = htons(m->m_pkthdr.len);
19007 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
19008 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
19009 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19010 			if (tp->t_port == 0 || len < V_tcp_minmss) {
19011 				ip->ip_off |= htons(IP_DF);
19012 			}
19013 		} else {
19014 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19015 		}
19016 	}
19017 #endif
19018 	if (doing_tlp == 0) {
19019 		/* Set we retransmitted */
19020 		rack->rc_gp_saw_rec = 1;
19021 	} else {
19022 		/* Its a TLP set ca or ss */
19023 		if (tp->snd_cwnd > tp->snd_ssthresh) {
19024 			/* Set we sent in CA */
19025 			rack->rc_gp_saw_ca = 1;
19026 		} else {
19027 			/* Set we sent in SS */
19028 			rack->rc_gp_saw_ss = 1;
19029 		}
19030 	}
19031 	/* Time to copy in our header */
19032 	cpto = mtod(m, uint8_t *);
19033 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
19034 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
19035 	if (optlen) {
19036 		bcopy(opt, th + 1, optlen);
19037 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
19038 	} else {
19039 		th->th_off = sizeof(struct tcphdr) >> 2;
19040 	}
19041 	if (tcp_bblogging_on(rack->rc_tp)) {
19042 		union tcp_log_stackspecific log;
19043 
19044 		if (rsm->r_flags & RACK_RWND_COLLAPSED) {
19045 			rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
19046 			counter_u64_add(rack_collapsed_win_rxt, 1);
19047 			counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
19048 		}
19049 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19050 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19051 		if (rack->rack_no_prr)
19052 			log.u_bbr.flex1 = 0;
19053 		else
19054 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
19055 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
19056 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
19057 		log.u_bbr.flex4 = max_val;
19058 		/* Save off the early/late values */
19059 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19060 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
19061 		log.u_bbr.bw_inuse = rack_get_bw(rack);
19062 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
19063 		if (doing_tlp == 0)
19064 			log.u_bbr.flex8 = 1;
19065 		else
19066 			log.u_bbr.flex8 = 2;
19067 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19068 		log.u_bbr.flex7 = 55;
19069 		log.u_bbr.pkts_out = tp->t_maxseg;
19070 		log.u_bbr.timeStamp = cts;
19071 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19072 		if (rsm && (rsm->r_rtr_cnt > 0)) {
19073 			/*
19074 			 * When we have a retransmit we want to log the
19075 			 * burst at send and flight at send from before.
19076 			 */
19077 			log.u_bbr.flex5 = rsm->r_fas;
19078 			log.u_bbr.bbr_substate = rsm->r_bas;
19079 		} else {
19080 			/*
19081 			 * This is currently unlikely until we do the
19082 			 * packet pair probes but I will add it for completeness.
19083 			 */
19084 			log.u_bbr.flex5 = log.u_bbr.inflight;
19085 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
19086 		}
19087 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
19088 		log.u_bbr.delivered = 0;
19089 		log.u_bbr.rttProp = (uint64_t)rsm;
19090 		log.u_bbr.delRate = rsm->r_flags;
19091 		log.u_bbr.delRate <<= 31;
19092 		log.u_bbr.delRate |= rack->r_must_retran;
19093 		log.u_bbr.delRate <<= 1;
19094 		log.u_bbr.delRate |= 1;
19095 		log.u_bbr.pkt_epoch = __LINE__;
19096 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
19097 				     len, &log, false, NULL, __func__, __LINE__, tv);
19098 	} else
19099 		lgb = NULL;
19100 	if ((rack->r_ctl.crte != NULL) &&
19101 	    tcp_bblogging_on(tp)) {
19102 		rack_log_queue_level(tp, rack, len, tv, cts);
19103 	}
19104 #ifdef INET6
19105 	if (rack->r_is_v6) {
19106 		error = ip6_output(m, inp->in6p_outputopts,
19107 				   &inp->inp_route6,
19108 				   ip_sendflag, NULL, NULL, inp);
19109 	}
19110 	else
19111 #endif
19112 #ifdef INET
19113 	{
19114 		error = ip_output(m, NULL,
19115 				  &inp->inp_route,
19116 				  ip_sendflag, 0, inp);
19117 	}
19118 #endif
19119 	m = NULL;
19120 	if (lgb) {
19121 		lgb->tlb_errno = error;
19122 		lgb = NULL;
19123 	}
19124 	/* Move snd_nxt to snd_max so we don't have false retransmissions */
19125 	tp->snd_nxt = tp->snd_max;
19126 	if (error) {
19127 		goto failed;
19128 	} else if (rack->rc_hw_nobuf && (ip_sendflag != IP_NO_SND_TAG_RL)) {
19129 		rack->rc_hw_nobuf = 0;
19130 		rack->r_ctl.rc_agg_delayed = 0;
19131 		rack->r_early = 0;
19132 		rack->r_late = 0;
19133 		rack->r_ctl.rc_agg_early = 0;
19134 	}
19135 
19136 	rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv),
19137 			rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls, segsiz);
19138 	if (doing_tlp) {
19139 		rack->rc_tlp_in_progress = 1;
19140 		rack->r_ctl.rc_tlp_cnt_out++;
19141 	}
19142 	if (error == 0) {
19143 		counter_u64_add(rack_total_bytes, len);
19144 		tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls);
19145 		if (doing_tlp) {
19146 			rack->rc_last_sent_tlp_past_cumack = 0;
19147 			rack->rc_last_sent_tlp_seq_valid = 1;
19148 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
19149 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
19150 		}
19151 		if (rack->r_ctl.rc_prr_sndcnt >= len)
19152 			rack->r_ctl.rc_prr_sndcnt -= len;
19153 		else
19154 			rack->r_ctl.rc_prr_sndcnt = 0;
19155 	}
19156 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19157 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
19158 	if (IN_FASTRECOVERY(tp->t_flags) && rsm)
19159 		rack->r_ctl.retran_during_recovery += len;
19160 	{
19161 		int idx;
19162 
19163 		idx = (len / segsiz) + 3;
19164 		if (idx >= TCP_MSS_ACCT_ATIMER)
19165 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
19166 		else
19167 			counter_u64_add(rack_out_size[idx], 1);
19168 	}
19169 	if (tp->t_rtttime == 0) {
19170 		tp->t_rtttime = ticks;
19171 		tp->t_rtseq = startseq;
19172 		KMOD_TCPSTAT_INC(tcps_segstimed);
19173 	}
19174 	counter_u64_add(rack_fto_rsm_send, 1);
19175 	if (error && (error == ENOBUFS)) {
19176 		if (rack->r_ctl.crte != NULL) {
19177 			tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
19178 			if (tcp_bblogging_on(rack->rc_tp))
19179 				rack_log_queue_level(tp, rack, len, tv, cts);
19180 		} else
19181 			tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
19182 		slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
19183 		if (rack->rc_enobuf < 0x7f)
19184 			rack->rc_enobuf++;
19185 		if (slot < (10 * HPTS_USEC_IN_MSEC))
19186 			slot = 10 * HPTS_USEC_IN_MSEC;
19187 		if (rack->r_ctl.crte != NULL) {
19188 			counter_u64_add(rack_saw_enobuf_hw, 1);
19189 			tcp_rl_log_enobuf(rack->r_ctl.crte);
19190 		}
19191 		counter_u64_add(rack_saw_enobuf, 1);
19192 	} else
19193 		slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz);
19194 	if ((slot == 0) ||
19195 	    (rack->rc_always_pace == 0) ||
19196 	    (rack->r_rr_config == 1)) {
19197 		/*
19198 		 * We have no pacing set or we
19199 		 * are using old-style rack or
19200 		 * we are overridden to use the old 1ms pacing.
19201 		 */
19202 		slot = rack->r_ctl.rc_min_to;
19203 	}
19204 	rack_start_hpts_timer(rack, tp, cts, slot, len, 0);
19205 #ifdef TCP_ACCOUNTING
19206 	crtsc = get_cyclecount();
19207 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19208 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19209 	}
19210 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19211 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19212 	}
19213 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19214 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz);
19215 	}
19216 	sched_unpin();
19217 #endif
19218 	return (0);
19219 failed:
19220 	if (m)
19221 		m_free(m);
19222 	return (-1);
19223 }
19224 
19225 static void
19226 rack_sndbuf_autoscale(struct tcp_rack *rack)
19227 {
19228 	/*
19229 	 * Automatic sizing of send socket buffer.  Often the send buffer
19230 	 * size is not optimally adjusted to the actual network conditions
19231 	 * at hand (delay bandwidth product).  Setting the buffer size too
19232 	 * small limits throughput on links with high bandwidth and high
19233 	 * delay (eg. trans-continental/oceanic links).  Setting the
19234 	 * buffer size too big consumes too much real kernel memory,
19235 	 * especially with many connections on busy servers.
19236 	 *
19237 	 * The criteria to step up the send buffer one notch are:
19238 	 *  1. receive window of remote host is larger than send buffer
19239 	 *     (with a fudge factor of 5/4th);
19240 	 *  2. send buffer is filled to 7/8th with data (so we actually
19241 	 *     have data to make use of it);
19242 	 *  3. send buffer fill has not hit maximal automatic size;
19243 	 *  4. our send window (slow start and cogestion controlled) is
19244 	 *     larger than sent but unacknowledged data in send buffer.
19245 	 *
19246 	 * Note that the rack version moves things much faster since
19247 	 * we want to avoid hitting cache lines in the rack_fast_output()
19248 	 * path so this is called much less often and thus moves
19249 	 * the SB forward by a percentage.
19250 	 */
19251 	struct socket *so;
19252 	struct tcpcb *tp;
19253 	uint32_t sendwin, scaleup;
19254 
19255 	tp = rack->rc_tp;
19256 	so = rack->rc_inp->inp_socket;
19257 	sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd);
19258 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
19259 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
19260 		    sbused(&so->so_snd) >=
19261 		    (so->so_snd.sb_hiwat / 8 * 7) &&
19262 		    sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
19263 		    sendwin >= (sbused(&so->so_snd) -
19264 		    (tp->snd_nxt - tp->snd_una))) {
19265 			if (rack_autosndbuf_inc)
19266 				scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100;
19267 			else
19268 				scaleup = V_tcp_autosndbuf_inc;
19269 			if (scaleup < V_tcp_autosndbuf_inc)
19270 				scaleup = V_tcp_autosndbuf_inc;
19271 			scaleup += so->so_snd.sb_hiwat;
19272 			if (scaleup > V_tcp_autosndbuf_max)
19273 				scaleup = V_tcp_autosndbuf_max;
19274 			if (!sbreserve_locked(so, SO_SND, scaleup, curthread))
19275 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
19276 		}
19277 	}
19278 }
19279 
19280 static int
19281 rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val,
19282 		 uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err)
19283 {
19284 	/*
19285 	 * Enter to do fast output. We are given that the sched_pin is
19286 	 * in place (if accounting is compiled in) and the cycle count taken
19287 	 * at entry is in place in ts_val. The idea here is that
19288 	 * we know how many more bytes needs to be sent (presumably either
19289 	 * during pacing or to fill the cwnd and that was greater than
19290 	 * the max-burst). We have how much to send and all the info we
19291 	 * need to just send.
19292 	 */
19293 #ifdef INET
19294 	struct ip *ip = NULL;
19295 #endif
19296 	struct udphdr *udp = NULL;
19297 	struct tcphdr *th = NULL;
19298 	struct mbuf *m, *s_mb;
19299 	struct inpcb *inp;
19300 	uint8_t *cpto;
19301 	struct tcp_log_buffer *lgb;
19302 #ifdef TCP_ACCOUNTING
19303 	uint64_t crtsc;
19304 #endif
19305 	struct tcpopt to;
19306 	u_char opt[TCP_MAXOLEN];
19307 	uint32_t hdrlen, optlen;
19308 #ifdef TCP_ACCOUNTING
19309 	int cnt_thru = 1;
19310 #endif
19311 	int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0;
19312 	uint16_t flags;
19313 	uint32_t s_soff;
19314 	uint32_t if_hw_tsomaxsegcount = 0, startseq;
19315 	uint32_t if_hw_tsomaxsegsize;
19316 	uint16_t add_flag = RACK_SENT_FP;
19317 #ifdef INET6
19318 	struct ip6_hdr *ip6 = NULL;
19319 
19320 	if (rack->r_is_v6) {
19321 		ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
19322 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
19323 	} else
19324 #endif				/* INET6 */
19325 	{
19326 #ifdef INET
19327 		ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
19328 		hdrlen = sizeof(struct tcpiphdr);
19329 #endif
19330 	}
19331 	if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) {
19332 		m = NULL;
19333 		goto failed;
19334 	}
19335 	rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
19336 	startseq = tp->snd_max;
19337 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
19338 	inp = rack->rc_inp;
19339 	len = rack->r_ctl.fsb.left_to_send;
19340 	to.to_flags = 0;
19341 	flags = rack->r_ctl.fsb.tcp_flags;
19342 	if (tp->t_flags & TF_RCVD_TSTMP) {
19343 		to.to_tsval = ms_cts + tp->ts_offset;
19344 		to.to_tsecr = tp->ts_recent;
19345 		to.to_flags = TOF_TS;
19346 	}
19347 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19348 	/* TCP-MD5 (RFC2385). */
19349 	if (tp->t_flags & TF_SIGNATURE)
19350 		to.to_flags |= TOF_SIGNATURE;
19351 #endif
19352 	optlen = tcp_addoptions(&to, opt);
19353 	hdrlen += optlen;
19354 	udp = rack->r_ctl.fsb.udp;
19355 	if (udp)
19356 		hdrlen += sizeof(struct udphdr);
19357 	if (rack->r_ctl.rc_pace_max_segs)
19358 		max_val = rack->r_ctl.rc_pace_max_segs;
19359 	else if (rack->rc_user_set_max_segs)
19360 		max_val = rack->rc_user_set_max_segs * segsiz;
19361 	else
19362 		max_val = len;
19363 	if ((tp->t_flags & TF_TSO) &&
19364 	    V_tcp_do_tso &&
19365 	    (len > segsiz) &&
19366 	    (tp->t_port == 0))
19367 		tso = 1;
19368 again:
19369 #ifdef INET6
19370 	if (MHLEN < hdrlen + max_linkhdr)
19371 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
19372 	else
19373 #endif
19374 		m = m_gethdr(M_NOWAIT, MT_DATA);
19375 	if (m == NULL)
19376 		goto failed;
19377 	m->m_data += max_linkhdr;
19378 	m->m_len = hdrlen;
19379 	th = rack->r_ctl.fsb.th;
19380 	/* Establish the len to send */
19381 	if (len > max_val)
19382 		len = max_val;
19383 	if ((tso) && (len + optlen > segsiz)) {
19384 		uint32_t if_hw_tsomax;
19385 		int32_t max_len;
19386 
19387 		/* extract TSO information */
19388 		if_hw_tsomax = tp->t_tsomax;
19389 		if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
19390 		if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
19391 		/*
19392 		 * Check if we should limit by maximum payload
19393 		 * length:
19394 		 */
19395 		if (if_hw_tsomax != 0) {
19396 			/* compute maximum TSO length */
19397 			max_len = (if_hw_tsomax - hdrlen -
19398 				   max_linkhdr);
19399 			if (max_len <= 0) {
19400 				goto failed;
19401 			} else if (len > max_len) {
19402 				len = max_len;
19403 			}
19404 		}
19405 		if (len <= segsiz) {
19406 			/*
19407 			 * In case there are too many small fragments don't
19408 			 * use TSO:
19409 			 */
19410 			tso = 0;
19411 		}
19412 	} else {
19413 		tso = 0;
19414 	}
19415 	if ((tso == 0) && (len > segsiz))
19416 		len = segsiz;
19417 	(void)tcp_get_usecs(tv);
19418 	if ((len == 0) ||
19419 	    (len <= MHLEN - hdrlen - max_linkhdr)) {
19420 		goto failed;
19421 	}
19422 	sb_offset = tp->snd_max - tp->snd_una;
19423 	th->th_seq = htonl(tp->snd_max);
19424 	th->th_ack = htonl(tp->rcv_nxt);
19425 	th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale));
19426 	if (th->th_win == 0) {
19427 		tp->t_sndzerowin++;
19428 		tp->t_flags |= TF_RXWIN0SENT;
19429 	} else
19430 		tp->t_flags &= ~TF_RXWIN0SENT;
19431 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
19432 	KMOD_TCPSTAT_INC(tcps_sndpack);
19433 	KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
19434 #ifdef STATS
19435 	stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
19436 				 len);
19437 #endif
19438 	if (rack->r_ctl.fsb.m == NULL)
19439 		goto failed;
19440 
19441 	/* s_mb and s_soff are saved for rack_log_output */
19442 	m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize,
19443 				    &s_mb, &s_soff);
19444 	if (len <= segsiz) {
19445 		/*
19446 		 * Must have ran out of mbufs for the copy
19447 		 * shorten it to no longer need tso. Lets
19448 		 * not put on sendalot since we are low on
19449 		 * mbufs.
19450 		 */
19451 		tso = 0;
19452 	}
19453 	if (rack->r_ctl.fsb.rfo_apply_push &&
19454 	    (len == rack->r_ctl.fsb.left_to_send)) {
19455 		tcp_set_flags(th, flags | TH_PUSH);
19456 		add_flag |= RACK_HAD_PUSH;
19457 	}
19458 	if ((m->m_next == NULL) || (len <= 0)){
19459 		goto failed;
19460 	}
19461 	if (udp) {
19462 		if (rack->r_is_v6)
19463 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
19464 		else
19465 			ulen = hdrlen + len - sizeof(struct ip);
19466 		udp->uh_ulen = htons(ulen);
19467 	}
19468 	m->m_pkthdr.rcvif = (struct ifnet *)0;
19469 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
19470 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
19471 		int ect = tcp_ecn_output_established(tp, &flags, len, false);
19472 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
19473 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
19474 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
19475 #ifdef INET6
19476 		if (rack->r_is_v6) {
19477 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
19478 			ip6->ip6_flow |= htonl(ect << 20);
19479 		}
19480 		else
19481 #endif
19482 		{
19483 #ifdef INET
19484 			ip->ip_tos &= ~IPTOS_ECN_MASK;
19485 			ip->ip_tos |= ect;
19486 #endif
19487 		}
19488 	}
19489 	tcp_set_flags(th, flags);
19490 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
19491 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
19492 	if (to.to_flags & TOF_SIGNATURE) {
19493 		/*
19494 		 * Calculate MD5 signature and put it into the place
19495 		 * determined before.
19496 		 * NOTE: since TCP options buffer doesn't point into
19497 		 * mbuf's data, calculate offset and use it.
19498 		 */
19499 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
19500 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
19501 			/*
19502 			 * Do not send segment if the calculation of MD5
19503 			 * digest has failed.
19504 			 */
19505 			goto failed;
19506 		}
19507 	}
19508 #endif
19509 #ifdef INET6
19510 	if (rack->r_is_v6) {
19511 		if (tp->t_port) {
19512 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
19513 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19514 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
19515 			th->th_sum = htons(0);
19516 			UDPSTAT_INC(udps_opackets);
19517 		} else {
19518 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
19519 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19520 			th->th_sum = in6_cksum_pseudo(ip6,
19521 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
19522 						      0);
19523 		}
19524 	}
19525 #endif
19526 #if defined(INET6) && defined(INET)
19527 	else
19528 #endif
19529 #ifdef INET
19530 	{
19531 		if (tp->t_port) {
19532 			m->m_pkthdr.csum_flags = CSUM_UDP;
19533 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
19534 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
19535 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
19536 			th->th_sum = htons(0);
19537 			UDPSTAT_INC(udps_opackets);
19538 		} else {
19539 			m->m_pkthdr.csum_flags = CSUM_TCP;
19540 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
19541 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
19542 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
19543 									IPPROTO_TCP + len + optlen));
19544 		}
19545 		/* IP version must be set here for ipv4/ipv6 checking later */
19546 		KASSERT(ip->ip_v == IPVERSION,
19547 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
19548 	}
19549 #endif
19550 	if (tso) {
19551 		/*
19552 		 * Here we use segsiz since we have no added options besides
19553 		 * any standard timestamp options (no DSACKs or SACKS are sent
19554 		 * via either fast-path).
19555 		 */
19556 		KASSERT(len > segsiz,
19557 			("%s: len <= tso_segsz tp:%p", __func__, tp));
19558 		m->m_pkthdr.csum_flags |= CSUM_TSO;
19559 		m->m_pkthdr.tso_segsz = segsiz;
19560 	}
19561 #ifdef INET6
19562 	if (rack->r_is_v6) {
19563 		ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit;
19564 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
19565 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
19566 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19567 		else
19568 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19569 	}
19570 #endif
19571 #if defined(INET) && defined(INET6)
19572 	else
19573 #endif
19574 #ifdef INET
19575 	{
19576 		ip->ip_len = htons(m->m_pkthdr.len);
19577 		ip->ip_ttl = rack->r_ctl.fsb.hoplimit;
19578 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
19579 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
19580 			if (tp->t_port == 0 || len < V_tcp_minmss) {
19581 				ip->ip_off |= htons(IP_DF);
19582 			}
19583 		} else {
19584 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
19585 		}
19586 	}
19587 #endif
19588 	if (tp->snd_cwnd > tp->snd_ssthresh) {
19589 		/* Set we sent in CA */
19590 		rack->rc_gp_saw_ca = 1;
19591 	} else {
19592 		/* Set we sent in SS */
19593 		rack->rc_gp_saw_ss = 1;
19594 	}
19595 	/* Time to copy in our header */
19596 	cpto = mtod(m, uint8_t *);
19597 	memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
19598 	th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
19599 	if (optlen) {
19600 		bcopy(opt, th + 1, optlen);
19601 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
19602 	} else {
19603 		th->th_off = sizeof(struct tcphdr) >> 2;
19604 	}
19605 	if ((rack->r_ctl.crte != NULL) &&
19606 	    tcp_bblogging_on(tp)) {
19607 		rack_log_queue_level(tp, rack, len, tv, cts);
19608 	}
19609 	if (tcp_bblogging_on(rack->rc_tp)) {
19610 		union tcp_log_stackspecific log;
19611 
19612 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
19613 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
19614 		if (rack->rack_no_prr)
19615 			log.u_bbr.flex1 = 0;
19616 		else
19617 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
19618 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
19619 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
19620 		log.u_bbr.flex4 = max_val;
19621 		/* Save off the early/late values */
19622 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
19623 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
19624 		log.u_bbr.bw_inuse = rack_get_bw(rack);
19625 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
19626 		log.u_bbr.flex8 = 0;
19627 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL);
19628 		log.u_bbr.flex7 = 44;
19629 		log.u_bbr.pkts_out = tp->t_maxseg;
19630 		log.u_bbr.timeStamp = cts;
19631 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
19632 		log.u_bbr.flex5 = log.u_bbr.inflight;
19633 		log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use;
19634 		log.u_bbr.delivered = 0;
19635 		log.u_bbr.rttProp = 0;
19636 		log.u_bbr.delRate = rack->r_must_retran;
19637 		log.u_bbr.delRate <<= 1;
19638 		log.u_bbr.pkt_epoch = __LINE__;
19639 		/* For fast output no retrans so just inflight and how many mss we send */
19640 		log.u_bbr.flex5 = log.u_bbr.inflight;
19641 		log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
19642 		lgb = tcp_log_event(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK,
19643 				     len, &log, false, NULL, __func__, __LINE__, tv);
19644 	} else
19645 		lgb = NULL;
19646 #ifdef INET6
19647 	if (rack->r_is_v6) {
19648 		error = ip6_output(m, inp->in6p_outputopts,
19649 				   &inp->inp_route6,
19650 				   0, NULL, NULL, inp);
19651 	}
19652 #endif
19653 #if defined(INET) && defined(INET6)
19654 	else
19655 #endif
19656 #ifdef INET
19657 	{
19658 		error = ip_output(m, NULL,
19659 				  &inp->inp_route,
19660 				  0, 0, inp);
19661 	}
19662 #endif
19663 	if (lgb) {
19664 		lgb->tlb_errno = error;
19665 		lgb = NULL;
19666 	}
19667 	if (error) {
19668 		*send_err = error;
19669 		m = NULL;
19670 		goto failed;
19671 	} else if (rack->rc_hw_nobuf) {
19672 		rack->rc_hw_nobuf = 0;
19673 		rack->r_ctl.rc_agg_delayed = 0;
19674 		rack->r_early = 0;
19675 		rack->r_late = 0;
19676 		rack->r_ctl.rc_agg_early = 0;
19677 	}
19678 	if ((error == 0) && (rack->lt_bw_up == 0)) {
19679 		/* Unlikely */
19680 		rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(tv);
19681 		rack->r_ctl.lt_seq = tp->snd_una;
19682 		rack->lt_bw_up = 1;
19683 	}
19684 	rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv),
19685 			NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls, segsiz);
19686 	m = NULL;
19687 	if (tp->snd_una == tp->snd_max) {
19688 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
19689 		rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
19690 		tp->t_acktime = ticks;
19691 	}
19692 	counter_u64_add(rack_total_bytes, len);
19693 	tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls);
19694 
19695 	rack->forced_ack = 0;	/* If we send something zap the FA flag */
19696 	tot_len += len;
19697 	if ((tp->t_flags & TF_GPUTINPROG) == 0)
19698 		rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset);
19699 	tp->snd_max += len;
19700 	tp->snd_nxt = tp->snd_max;
19701 	if (rack->rc_new_rnd_needed) {
19702 		/*
19703 		 * Update the rnd to start ticking not
19704 		 * that from a time perspective all of
19705 		 * the preceding idle time is "in the round"
19706 		 */
19707 		rack->rc_new_rnd_needed = 0;
19708 		rack->r_ctl.roundends = tp->snd_max;
19709 	}
19710 	{
19711 		int idx;
19712 
19713 		idx = (len / segsiz) + 3;
19714 		if (idx >= TCP_MSS_ACCT_ATIMER)
19715 			counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
19716 		else
19717 			counter_u64_add(rack_out_size[idx], 1);
19718 	}
19719 	if (len <= rack->r_ctl.fsb.left_to_send)
19720 		rack->r_ctl.fsb.left_to_send -= len;
19721 	else
19722 		rack->r_ctl.fsb.left_to_send = 0;
19723 	if (rack->r_ctl.fsb.left_to_send < segsiz) {
19724 		rack->r_fast_output = 0;
19725 		rack->r_ctl.fsb.left_to_send = 0;
19726 		/* At the end of fast_output scale up the sb */
19727 		SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd);
19728 		rack_sndbuf_autoscale(rack);
19729 		SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd);
19730 	}
19731 	if (tp->t_rtttime == 0) {
19732 		tp->t_rtttime = ticks;
19733 		tp->t_rtseq = startseq;
19734 		KMOD_TCPSTAT_INC(tcps_segstimed);
19735 	}
19736 	if ((rack->r_ctl.fsb.left_to_send >= segsiz) &&
19737 	    (max_val > len) &&
19738 	    (tso == 0)) {
19739 		max_val -= len;
19740 		len = segsiz;
19741 		th = rack->r_ctl.fsb.th;
19742 #ifdef TCP_ACCOUNTING
19743 		cnt_thru++;
19744 #endif
19745 		goto again;
19746 	}
19747 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
19748 	counter_u64_add(rack_fto_send, 1);
19749 	slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz);
19750 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0);
19751 #ifdef TCP_ACCOUNTING
19752 	crtsc = get_cyclecount();
19753 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19754 		tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru;
19755 	}
19756 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19757 		tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
19758 	}
19759 	if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
19760 		tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz);
19761 	}
19762 	sched_unpin();
19763 #endif
19764 	return (0);
19765 failed:
19766 	if (m)
19767 		m_free(m);
19768 	rack->r_fast_output = 0;
19769 	return (-1);
19770 }
19771 
19772 static inline void
19773 rack_setup_fast_output(struct tcpcb *tp, struct tcp_rack *rack,
19774 		       struct sockbuf *sb,
19775 		       int len, int orig_len, int segsiz, uint32_t pace_max_seg,
19776 		       bool hw_tls,
19777 		       uint16_t flags)
19778 {
19779 	rack->r_fast_output = 1;
19780 	rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
19781 	rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
19782 	rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
19783 	rack->r_ctl.fsb.tcp_flags = flags;
19784 	rack->r_ctl.fsb.left_to_send = orig_len - len;
19785 	if (rack->r_ctl.fsb.left_to_send < pace_max_seg) {
19786 		/* Less than a full sized pace, lets not  */
19787 		rack->r_fast_output = 0;
19788 		return;
19789 	} else {
19790 		/* Round down to the nearest pace_max_seg */
19791 		rack->r_ctl.fsb.left_to_send = rounddown(rack->r_ctl.fsb.left_to_send, pace_max_seg);
19792 	}
19793 	if (hw_tls)
19794 		rack->r_ctl.fsb.hw_tls = 1;
19795 	else
19796 		rack->r_ctl.fsb.hw_tls = 0;
19797 	KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))),
19798 		("rack:%p left_to_send:%u sbavail:%u out:%u",
19799 		 rack, rack->r_ctl.fsb.left_to_send, sbavail(sb),
19800 		 (tp->snd_max - tp->snd_una)));
19801 	if (rack->r_ctl.fsb.left_to_send < segsiz)
19802 		rack->r_fast_output = 0;
19803 	else {
19804 		if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una)))
19805 			rack->r_ctl.fsb.rfo_apply_push = 1;
19806 		else
19807 			rack->r_ctl.fsb.rfo_apply_push = 0;
19808 	}
19809 }
19810 
19811 static uint32_t
19812 rack_get_hpts_pacing_min_for_bw(struct tcp_rack *rack, int32_t segsiz)
19813 {
19814 	uint64_t min_time;
19815 	uint32_t maxlen;
19816 
19817 	min_time = (uint64_t)get_hpts_min_sleep_time();
19818 	maxlen = (uint32_t)((rack->r_ctl.gp_bw * min_time) / (uint64_t)HPTS_USEC_IN_SEC);
19819 	maxlen = roundup(maxlen, segsiz);
19820 	return (maxlen);
19821 }
19822 
19823 static struct rack_sendmap *
19824 rack_check_collapsed(struct tcp_rack *rack, uint32_t cts)
19825 {
19826 	struct rack_sendmap *rsm = NULL;
19827 	int thresh;
19828 
19829 restart:
19830 	rsm = tqhash_find(rack->r_ctl.tqh, rack->r_ctl.last_collapse_point);
19831 	if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) {
19832 		/* Nothing, strange turn off validity  */
19833 		rack->r_collapse_point_valid = 0;
19834 		return (NULL);
19835 	}
19836 	/* Can we send it yet? */
19837 	if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) {
19838 		/*
19839 		 * Receiver window has not grown enough for
19840 		 * the segment to be put on the wire.
19841 		 */
19842 		return (NULL);
19843 	}
19844 	if (rsm->r_flags & RACK_ACKED) {
19845 		/*
19846 		 * It has been sacked, lets move to the
19847 		 * next one if possible.
19848 		 */
19849 		rack->r_ctl.last_collapse_point = rsm->r_end;
19850 		/* Are we done? */
19851 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
19852 			    rack->r_ctl.high_collapse_point)) {
19853 			rack->r_collapse_point_valid = 0;
19854 			return (NULL);
19855 		}
19856 		goto restart;
19857 	}
19858 	/* Now has it been long enough ? */
19859 	thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts);
19860 	if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) {
19861 		rack_log_collapse(rack, rsm->r_start,
19862 				  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19863 				  thresh, __LINE__, 6, rsm->r_flags, rsm);
19864 		return (rsm);
19865 	}
19866 	/* Not enough time */
19867 	rack_log_collapse(rack, rsm->r_start,
19868 			  (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])),
19869 			  thresh, __LINE__, 7, rsm->r_flags, rsm);
19870 	return (NULL);
19871 }
19872 
19873 static inline void
19874 rack_validate_sizes(struct tcp_rack *rack, int32_t *len, int32_t segsiz, uint32_t pace_max_seg)
19875 {
19876 	if ((rack->full_size_rxt == 0) &&
19877 	    (rack->shape_rxt_to_pacing_min == 0) &&
19878 	    (*len >= segsiz)) {
19879 		*len = segsiz;
19880 	} else if (rack->shape_rxt_to_pacing_min &&
19881 		 rack->gp_ready) {
19882 		/* We use pacing min as shaping len req */
19883 		uint32_t maxlen;
19884 
19885 		maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
19886 		if (*len > maxlen)
19887 			*len = maxlen;
19888 	} else {
19889 		/*
19890 		 * The else is full_size_rxt is on so send it all
19891 		 * note we do need to check this for exceeding
19892 		 * our max segment size due to the fact that
19893 		 * we do sometimes merge chunks together i.e.
19894 		 * we cannot just assume that we will never have
19895 		 * a chunk greater than pace_max_seg
19896 		 */
19897 		if (*len > pace_max_seg)
19898 			*len = pace_max_seg;
19899 	}
19900 }
19901 
19902 static int
19903 rack_output(struct tcpcb *tp)
19904 {
19905 	struct socket *so;
19906 	uint32_t recwin;
19907 	uint32_t sb_offset, s_moff = 0;
19908 	int32_t len, error = 0;
19909 	uint16_t flags;
19910 	struct mbuf *m, *s_mb = NULL;
19911 	struct mbuf *mb;
19912 	uint32_t if_hw_tsomaxsegcount = 0;
19913 	uint32_t if_hw_tsomaxsegsize;
19914 	int32_t segsiz, minseg;
19915 	long tot_len_this_send = 0;
19916 #ifdef INET
19917 	struct ip *ip = NULL;
19918 #endif
19919 	struct udphdr *udp = NULL;
19920 	struct tcp_rack *rack;
19921 	struct tcphdr *th;
19922 	uint8_t pass = 0;
19923 	uint8_t mark = 0;
19924 	uint8_t check_done = 0;
19925 	uint8_t wanted_cookie = 0;
19926 	u_char opt[TCP_MAXOLEN];
19927 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
19928 	uint32_t rack_seq;
19929 
19930 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
19931 	unsigned ipsec_optlen = 0;
19932 
19933 #endif
19934 	int32_t idle, sendalot;
19935 	int32_t sub_from_prr = 0;
19936 	volatile int32_t sack_rxmit;
19937 	struct rack_sendmap *rsm = NULL;
19938 	int32_t tso, mtu;
19939 	struct tcpopt to;
19940 	int32_t slot = 0;
19941 	int32_t sup_rack = 0;
19942 	uint32_t cts, ms_cts, delayed, early;
19943 	uint16_t add_flag = RACK_SENT_SP;
19944 	/* The doing_tlp flag will be set by the actual rack_timeout_tlp() */
19945 	uint8_t doing_tlp = 0;
19946 	uint32_t cwnd_to_use, pace_max_seg;
19947 	int32_t do_a_prefetch = 0;
19948 	int32_t prefetch_rsm = 0;
19949 	int32_t orig_len = 0;
19950 	struct timeval tv;
19951 	int32_t prefetch_so_done = 0;
19952 	struct tcp_log_buffer *lgb;
19953 	struct inpcb *inp = tptoinpcb(tp);
19954 	struct sockbuf *sb;
19955 	uint64_t ts_val = 0;
19956 #ifdef TCP_ACCOUNTING
19957 	uint64_t crtsc;
19958 #endif
19959 #ifdef INET6
19960 	struct ip6_hdr *ip6 = NULL;
19961 	int32_t isipv6;
19962 #endif
19963 	bool hpts_calling, hw_tls = false;
19964 
19965 	NET_EPOCH_ASSERT();
19966 	INP_WLOCK_ASSERT(inp);
19967 
19968 	/* setup and take the cache hits here */
19969 	rack = (struct tcp_rack *)tp->t_fb_ptr;
19970 #ifdef TCP_ACCOUNTING
19971 	sched_pin();
19972 	ts_val = get_cyclecount();
19973 #endif
19974 	hpts_calling = !!(tp->t_flags2 & TF2_HPTS_CALLS);
19975 	tp->t_flags2 &= ~TF2_HPTS_CALLS;
19976 #ifdef TCP_OFFLOAD
19977 	if (tp->t_flags & TF_TOE) {
19978 #ifdef TCP_ACCOUNTING
19979 		sched_unpin();
19980 #endif
19981 		return (tcp_offload_output(tp));
19982 	}
19983 #endif
19984 	if (rack->rack_deferred_inited == 0) {
19985 		/*
19986 		 * If we are the connecting socket we will
19987 		 * hit rack_init() when no sequence numbers
19988 		 * are setup. This makes it so we must defer
19989 		 * some initialization. Call that now.
19990 		 */
19991 		rack_deferred_init(tp, rack);
19992 	}
19993 	/*
19994 	 * For TFO connections in SYN_RECEIVED, only allow the initial
19995 	 * SYN|ACK and those sent by the retransmit timer.
19996 	 */
19997 	if (IS_FASTOPEN(tp->t_flags) &&
19998 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
19999 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
20000 	    (rack->r_ctl.rc_resend == NULL)) {         /* not a retransmit */
20001 #ifdef TCP_ACCOUNTING
20002 		sched_unpin();
20003 #endif
20004 		return (0);
20005 	}
20006 #ifdef INET6
20007 	if (rack->r_state) {
20008 		/* Use the cache line loaded if possible */
20009 		isipv6 = rack->r_is_v6;
20010 	} else {
20011 		isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0;
20012 	}
20013 #endif
20014 	early = 0;
20015 	cts = tcp_get_usecs(&tv);
20016 	ms_cts = tcp_tv_to_mssectick(&tv);
20017 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
20018 	    tcp_in_hpts(rack->rc_tp)) {
20019 		/*
20020 		 * We are on the hpts for some timer but not hptsi output.
20021 		 * Remove from the hpts unconditionally.
20022 		 */
20023 		rack_timer_cancel(tp, rack, cts, __LINE__);
20024 	}
20025 	/* Are we pacing and late? */
20026 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
20027 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
20028 		/* We are delayed */
20029 		delayed = cts - rack->r_ctl.rc_last_output_to;
20030 	} else {
20031 		delayed = 0;
20032 	}
20033 	/* Do the timers, which may override the pacer */
20034 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
20035 		int retval;
20036 
20037 		retval = rack_process_timers(tp, rack, cts, hpts_calling,
20038 					     &doing_tlp);
20039 		if (retval != 0) {
20040 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
20041 #ifdef TCP_ACCOUNTING
20042 			sched_unpin();
20043 #endif
20044 			/*
20045 			 * If timers want tcp_drop(), then pass error out,
20046 			 * otherwise suppress it.
20047 			 */
20048 			return (retval < 0 ? retval : 0);
20049 		}
20050 	}
20051 	if (rack->rc_in_persist) {
20052 		if (tcp_in_hpts(rack->rc_tp) == 0) {
20053 			/* Timer is not running */
20054 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
20055 		}
20056 #ifdef TCP_ACCOUNTING
20057 		sched_unpin();
20058 #endif
20059 		return (0);
20060 	}
20061 	if ((rack->rc_ack_required == 1) &&
20062 	    (rack->r_timer_override == 0)){
20063 		/* A timeout occurred and no ack has arrived */
20064 		if (tcp_in_hpts(rack->rc_tp) == 0) {
20065 			/* Timer is not running */
20066 			rack_start_hpts_timer(rack, tp, cts, 0, 0, 0);
20067 		}
20068 #ifdef TCP_ACCOUNTING
20069 		sched_unpin();
20070 #endif
20071 		return (0);
20072 	}
20073 	if ((rack->r_timer_override) ||
20074 	    (rack->rc_ack_can_sendout_data) ||
20075 	    (delayed) ||
20076 	    (tp->t_state < TCPS_ESTABLISHED)) {
20077 		rack->rc_ack_can_sendout_data = 0;
20078 		if (tcp_in_hpts(rack->rc_tp))
20079 			tcp_hpts_remove(rack->rc_tp);
20080 	} else if (tcp_in_hpts(rack->rc_tp)) {
20081 		/*
20082 		 * On the hpts you can't pass even if ACKNOW is on, we will
20083 		 * when the hpts fires.
20084 		 */
20085 #ifdef TCP_ACCOUNTING
20086 		crtsc = get_cyclecount();
20087 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20088 			tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val);
20089 		}
20090 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
20091 			tp->tcp_cnt_counters[SND_BLOCKED]++;
20092 		}
20093 		sched_unpin();
20094 #endif
20095 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
20096 		return (0);
20097 	}
20098 	/* Finish out both pacing early and late accounting */
20099 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
20100 	    TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) {
20101 		early = rack->r_ctl.rc_last_output_to - cts;
20102 	} else
20103 		early = 0;
20104 	if (delayed) {
20105 		rack->r_ctl.rc_agg_delayed += delayed;
20106 		rack->r_late = 1;
20107 	} else if (early) {
20108 		rack->r_ctl.rc_agg_early += early;
20109 		rack->r_early = 1;
20110 	}
20111 	/* Now that early/late accounting is done turn off the flag */
20112 	rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT;
20113 	rack->r_wanted_output = 0;
20114 	rack->r_timer_override = 0;
20115 	if ((tp->t_state != rack->r_state) &&
20116 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
20117 		rack_set_state(tp, rack);
20118 	}
20119 	if ((rack->r_fast_output) &&
20120 	    (doing_tlp == 0) &&
20121 	    (tp->rcv_numsacks == 0)) {
20122 		int ret;
20123 
20124 		error = 0;
20125 		ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
20126 		if (ret >= 0)
20127 			return(ret);
20128 		else if (error) {
20129 			inp = rack->rc_inp;
20130 			so = inp->inp_socket;
20131 			sb = &so->so_snd;
20132 			goto nomore;
20133 		}
20134 	}
20135 	inp = rack->rc_inp;
20136 	/*
20137 	 * For TFO connections in SYN_SENT or SYN_RECEIVED,
20138 	 * only allow the initial SYN or SYN|ACK and those sent
20139 	 * by the retransmit timer.
20140 	 */
20141 	if (IS_FASTOPEN(tp->t_flags) &&
20142 	    ((tp->t_state == TCPS_SYN_RECEIVED) ||
20143 	     (tp->t_state == TCPS_SYN_SENT)) &&
20144 	    SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
20145 	    (tp->t_rxtshift == 0)) {              /* not a retransmit */
20146 		cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
20147 		so = inp->inp_socket;
20148 		sb = &so->so_snd;
20149 		goto just_return_nolock;
20150 	}
20151 	/*
20152 	 * Determine length of data that should be transmitted, and flags
20153 	 * that will be used. If there is some data or critical controls
20154 	 * (SYN, RST) to send, then transmit; otherwise, investigate
20155 	 * further.
20156 	 */
20157 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
20158 	if (tp->t_idle_reduce) {
20159 		if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur))
20160 			rack_cc_after_idle(rack, tp);
20161 	}
20162 	tp->t_flags &= ~TF_LASTIDLE;
20163 	if (idle) {
20164 		if (tp->t_flags & TF_MORETOCOME) {
20165 			tp->t_flags |= TF_LASTIDLE;
20166 			idle = 0;
20167 		}
20168 	}
20169 	if ((tp->snd_una == tp->snd_max) &&
20170 	    rack->r_ctl.rc_went_idle_time &&
20171 	    TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) {
20172 		idle = cts - rack->r_ctl.rc_went_idle_time;
20173 		if (idle > rack_min_probertt_hold) {
20174 			/* Count as a probe rtt */
20175 			if (rack->in_probe_rtt == 0) {
20176 				rack->r_ctl.rc_lower_rtt_us_cts = cts;
20177 				rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts;
20178 				rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts;
20179 				rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts;
20180 			} else {
20181 				rack_exit_probertt(rack, cts);
20182 			}
20183 		}
20184 		idle = 0;
20185 	}
20186 	if (rack_use_fsb &&
20187 	    (rack->r_ctl.fsb.tcp_ip_hdr) &&
20188 	    (rack->r_fsb_inited == 0) &&
20189 	    (rack->r_state != TCPS_CLOSED))
20190 		rack_init_fsb_block(tp, rack, tcp_outflags[tp->t_state]);
20191 again:
20192 	/*
20193 	 * If we've recently taken a timeout, snd_max will be greater than
20194 	 * snd_nxt.  There may be SACK information that allows us to avoid
20195 	 * resending already delivered data.  Adjust snd_nxt accordingly.
20196 	 */
20197 	sendalot = 0;
20198 	cts = tcp_get_usecs(&tv);
20199 	ms_cts = tcp_tv_to_mssectick(&tv);
20200 	tso = 0;
20201 	mtu = 0;
20202 	segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs);
20203 	minseg = segsiz;
20204 	if (rack->r_ctl.rc_pace_max_segs == 0)
20205 		pace_max_seg = rack->rc_user_set_max_segs * segsiz;
20206 	else
20207 		pace_max_seg = rack->r_ctl.rc_pace_max_segs;
20208 	sb_offset = tp->snd_max - tp->snd_una;
20209 	cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd;
20210 	flags = tcp_outflags[tp->t_state];
20211 	while (rack->rc_free_cnt < rack_free_cache) {
20212 		rsm = rack_alloc(rack);
20213 		if (rsm == NULL) {
20214 			if (hpts_calling)
20215 				/* Retry in a ms */
20216 				slot = (1 * HPTS_USEC_IN_MSEC);
20217 			so = inp->inp_socket;
20218 			sb = &so->so_snd;
20219 			goto just_return_nolock;
20220 		}
20221 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext);
20222 		rack->rc_free_cnt++;
20223 		rsm = NULL;
20224 	}
20225 	sack_rxmit = 0;
20226 	len = 0;
20227 	rsm = NULL;
20228 	if (flags & TH_RST) {
20229 		SOCKBUF_LOCK(&inp->inp_socket->so_snd);
20230 		so = inp->inp_socket;
20231 		sb = &so->so_snd;
20232 		goto send;
20233 	}
20234 	if (rack->r_ctl.rc_resend) {
20235 		/* Retransmit timer */
20236 		rsm = rack->r_ctl.rc_resend;
20237 		rack->r_ctl.rc_resend = NULL;
20238 		len = rsm->r_end - rsm->r_start;
20239 		sack_rxmit = 1;
20240 		sendalot = 0;
20241 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20242 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20243 			 __func__, __LINE__,
20244 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20245 		sb_offset = rsm->r_start - tp->snd_una;
20246 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20247 	} else if (rack->r_collapse_point_valid &&
20248 		   ((rsm = rack_check_collapsed(rack, cts)) != NULL)) {
20249 		/*
20250 		 * If an RSM is returned then enough time has passed
20251 		 * for us to retransmit it. Move up the collapse point,
20252 		 * since this rsm has its chance to retransmit now.
20253 		 */
20254 		tcp_trace_point(rack->rc_tp, TCP_TP_COLLAPSED_RXT);
20255 		rack->r_ctl.last_collapse_point = rsm->r_end;
20256 		/* Are we done? */
20257 		if (SEQ_GEQ(rack->r_ctl.last_collapse_point,
20258 			    rack->r_ctl.high_collapse_point))
20259 			rack->r_collapse_point_valid = 0;
20260 		sack_rxmit = 1;
20261 		/* We are not doing a TLP */
20262 		doing_tlp = 0;
20263 		len = rsm->r_end - rsm->r_start;
20264 		sb_offset = rsm->r_start - tp->snd_una;
20265 		sendalot = 0;
20266 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20267 	} else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) {
20268 		/* We have a retransmit that takes precedence */
20269 		if ((!IN_FASTRECOVERY(tp->t_flags)) &&
20270 		    ((rsm->r_flags & RACK_MUST_RXT) == 0) &&
20271 		    ((tp->t_flags & TF_WASFRECOVERY) == 0)) {
20272 			/* Enter recovery if not induced by a time-out */
20273 			rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__);
20274 		}
20275 #ifdef INVARIANTS
20276 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
20277 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
20278 			      tp, rack, rsm, rsm->r_start, tp->snd_una);
20279 		}
20280 #endif
20281 		len = rsm->r_end - rsm->r_start;
20282 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20283 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20284 			 __func__, __LINE__,
20285 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20286 		sb_offset = rsm->r_start - tp->snd_una;
20287 		sendalot = 0;
20288 		rack_validate_sizes(rack, &len, segsiz, pace_max_seg);
20289 		if (len > 0) {
20290 			sack_rxmit = 1;
20291 			KMOD_TCPSTAT_INC(tcps_sack_rexmits);
20292 			KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes,
20293 					 min(len, segsiz));
20294 		}
20295 	} else if (rack->r_ctl.rc_tlpsend) {
20296 		/* Tail loss probe */
20297 		long cwin;
20298 		long tlen;
20299 
20300 		/*
20301 		 * Check if we can do a TLP with a RACK'd packet
20302 		 * this can happen if we are not doing the rack
20303 		 * cheat and we skipped to a TLP and it
20304 		 * went off.
20305 		 */
20306 		rsm = rack->r_ctl.rc_tlpsend;
20307 		/* We are doing a TLP make sure the flag is preent */
20308 		rsm->r_flags |= RACK_TLP;
20309 		rack->r_ctl.rc_tlpsend = NULL;
20310 		sack_rxmit = 1;
20311 		tlen = rsm->r_end - rsm->r_start;
20312 		if (tlen > segsiz)
20313 			tlen = segsiz;
20314 		KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start),
20315 			("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p",
20316 			 __func__, __LINE__,
20317 			 rsm->r_start, tp->snd_una, tp, rack, rsm));
20318 		sb_offset = rsm->r_start - tp->snd_una;
20319 		cwin = min(tp->snd_wnd, tlen);
20320 		len = cwin;
20321 	}
20322 	if (rack->r_must_retran &&
20323 	    (doing_tlp == 0) &&
20324 	    (SEQ_GT(tp->snd_max, tp->snd_una)) &&
20325 	    (rsm == NULL)) {
20326 		/*
20327 		 * There are two different ways that we
20328 		 * can get into this block:
20329 		 * a) This is a non-sack connection, we had a time-out
20330 		 *    and thus r_must_retran was set and everything
20331 		 *    left outstanding as been marked for retransmit.
20332 		 * b) The MTU of the path shrank, so that everything
20333 		 *    was marked to be retransmitted with the smaller
20334 		 *    mtu and r_must_retran was set.
20335 		 *
20336 		 * This means that we expect the sendmap (outstanding)
20337 		 * to all be marked must. We can use the tmap to
20338 		 * look at them.
20339 		 *
20340 		 */
20341 		int sendwin, flight;
20342 
20343 		sendwin = min(tp->snd_wnd, tp->snd_cwnd);
20344 		flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto);
20345 		if (flight >= sendwin) {
20346 			/*
20347 			 * We can't send yet.
20348 			 */
20349 			so = inp->inp_socket;
20350 			sb = &so->so_snd;
20351 			goto just_return_nolock;
20352 		}
20353 		/*
20354 		 * This is the case a/b mentioned above. All
20355 		 * outstanding/not-acked should be marked.
20356 		 * We can use the tmap to find them.
20357 		 */
20358 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
20359 		if (rsm == NULL) {
20360 			/* TSNH */
20361 			rack->r_must_retran = 0;
20362 			rack->r_ctl.rc_out_at_rto = 0;
20363 			so = inp->inp_socket;
20364 			sb = &so->so_snd;
20365 			goto just_return_nolock;
20366 		}
20367 		if ((rsm->r_flags & RACK_MUST_RXT) == 0) {
20368 			/*
20369 			 * The first one does not have the flag, did we collapse
20370 			 * further up in our list?
20371 			 */
20372 			rack->r_must_retran = 0;
20373 			rack->r_ctl.rc_out_at_rto = 0;
20374 			rsm = NULL;
20375 			sack_rxmit = 0;
20376 		} else {
20377 			sack_rxmit = 1;
20378 			len = rsm->r_end - rsm->r_start;
20379 			sb_offset = rsm->r_start - tp->snd_una;
20380 			sendalot = 0;
20381 			if ((rack->full_size_rxt == 0) &&
20382 			    (rack->shape_rxt_to_pacing_min == 0) &&
20383 			    (len >= segsiz))
20384 				len = segsiz;
20385 			else if (rack->shape_rxt_to_pacing_min &&
20386 				 rack->gp_ready) {
20387 				/* We use pacing min as shaping len req */
20388 				uint32_t maxlen;
20389 
20390 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20391 				if (len > maxlen)
20392 					len = maxlen;
20393 			}
20394 			/*
20395 			 * Delay removing the flag RACK_MUST_RXT so
20396 			 * that the fastpath for retransmit will
20397 			 * work with this rsm.
20398 			 */
20399 		}
20400 	}
20401 	/*
20402 	 * Enforce a connection sendmap count limit if set
20403 	 * as long as we are not retransmiting.
20404 	 */
20405 	if ((rsm == NULL) &&
20406 	    (rack->do_detection == 0) &&
20407 	    (V_tcp_map_entries_limit > 0) &&
20408 	    (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) {
20409 		counter_u64_add(rack_to_alloc_limited, 1);
20410 		if (!rack->alloc_limit_reported) {
20411 			rack->alloc_limit_reported = 1;
20412 			counter_u64_add(rack_alloc_limited_conns, 1);
20413 		}
20414 		so = inp->inp_socket;
20415 		sb = &so->so_snd;
20416 		goto just_return_nolock;
20417 	}
20418 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
20419 		/* we are retransmitting the fin */
20420 		len--;
20421 		if (len) {
20422 			/*
20423 			 * When retransmitting data do *not* include the
20424 			 * FIN. This could happen from a TLP probe.
20425 			 */
20426 			flags &= ~TH_FIN;
20427 		}
20428 	}
20429 	if (rsm && rack->r_fsb_inited &&
20430 	    rack_use_rsm_rfo &&
20431 	    ((rsm->r_flags & RACK_HAS_FIN) == 0)) {
20432 		int ret;
20433 
20434 		ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp);
20435 		if (ret == 0)
20436 			return (0);
20437 	}
20438 	so = inp->inp_socket;
20439 	sb = &so->so_snd;
20440 	if (do_a_prefetch == 0) {
20441 		kern_prefetch(sb, &do_a_prefetch);
20442 		do_a_prefetch = 1;
20443 	}
20444 #ifdef NETFLIX_SHARED_CWND
20445 	if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) &&
20446 	    rack->rack_enable_scwnd) {
20447 		/* We are doing cwnd sharing */
20448 		if (rack->gp_ready &&
20449 		    (rack->rack_attempted_scwnd == 0) &&
20450 		    (rack->r_ctl.rc_scw == NULL) &&
20451 		    tp->t_lib) {
20452 			/* The pcbid is in, lets make an attempt */
20453 			counter_u64_add(rack_try_scwnd, 1);
20454 			rack->rack_attempted_scwnd = 1;
20455 			rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp,
20456 								   &rack->r_ctl.rc_scw_index,
20457 								   segsiz);
20458 		}
20459 		if (rack->r_ctl.rc_scw &&
20460 		    (rack->rack_scwnd_is_idle == 1) &&
20461 		    sbavail(&so->so_snd)) {
20462 			/* we are no longer out of data */
20463 			tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
20464 			rack->rack_scwnd_is_idle = 0;
20465 		}
20466 		if (rack->r_ctl.rc_scw) {
20467 			/* First lets update and get the cwnd */
20468 			rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw,
20469 										       rack->r_ctl.rc_scw_index,
20470 										       tp->snd_cwnd, tp->snd_wnd, segsiz);
20471 		}
20472 	}
20473 #endif
20474 	/*
20475 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
20476 	 * state flags.
20477 	 */
20478 	if (tp->t_flags & TF_NEEDFIN)
20479 		flags |= TH_FIN;
20480 	if (tp->t_flags & TF_NEEDSYN)
20481 		flags |= TH_SYN;
20482 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
20483 		void *end_rsm;
20484 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
20485 		if (end_rsm)
20486 			kern_prefetch(end_rsm, &prefetch_rsm);
20487 		prefetch_rsm = 1;
20488 	}
20489 	SOCKBUF_LOCK(sb);
20490 	/*
20491 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
20492 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
20493 	 * negative length.  This can also occur when TCP opens up its
20494 	 * congestion window while receiving additional duplicate acks after
20495 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
20496 	 * the fast-retransmit.
20497 	 *
20498 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
20499 	 * set to snd_una, the sb_offset will be 0, and the length may wind
20500 	 * up 0.
20501 	 *
20502 	 * If sack_rxmit is true we are retransmitting from the scoreboard
20503 	 * in which case len is already set.
20504 	 */
20505 	if ((sack_rxmit == 0) &&
20506 	    (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) {
20507 		uint32_t avail;
20508 
20509 		avail = sbavail(sb);
20510 		if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
20511 			sb_offset = tp->snd_nxt - tp->snd_una;
20512 		else
20513 			sb_offset = 0;
20514 		if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) {
20515 			if (rack->r_ctl.rc_tlp_new_data) {
20516 				/* TLP is forcing out new data */
20517 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
20518 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
20519 				}
20520 				if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) {
20521 					if (tp->snd_wnd > sb_offset)
20522 						len = tp->snd_wnd - sb_offset;
20523 					else
20524 						len = 0;
20525 				} else {
20526 					len = rack->r_ctl.rc_tlp_new_data;
20527 				}
20528 				rack->r_ctl.rc_tlp_new_data = 0;
20529 			}  else {
20530 				len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset);
20531 			}
20532 			if ((rack->r_ctl.crte == NULL) &&
20533 			    IN_FASTRECOVERY(tp->t_flags) &&
20534 			    (rack->full_size_rxt == 0) &&
20535 			    (rack->shape_rxt_to_pacing_min == 0) &&
20536 			    (len > segsiz)) {
20537 				/*
20538 				 * For prr=off, we need to send only 1 MSS
20539 				 * at a time. We do this because another sack could
20540 				 * be arriving that causes us to send retransmits and
20541 				 * we don't want to be on a long pace due to a larger send
20542 				 * that keeps us from sending out the retransmit.
20543 				 */
20544 				len = segsiz;
20545 			} else if (rack->shape_rxt_to_pacing_min &&
20546 				   rack->gp_ready) {
20547 				/* We use pacing min as shaping len req */
20548 				uint32_t maxlen;
20549 
20550 				maxlen = rack_get_hpts_pacing_min_for_bw(rack, segsiz);
20551 				if (len > maxlen)
20552 					len = maxlen;
20553 			}/* The else is full_size_rxt is on so send it all */
20554 		} else {
20555 			uint32_t outstanding;
20556 			/*
20557 			 * We are inside of a Fast recovery episode, this
20558 			 * is caused by a SACK or 3 dup acks. At this point
20559 			 * we have sent all the retransmissions and we rely
20560 			 * on PRR to dictate what we will send in the form of
20561 			 * new data.
20562 			 */
20563 
20564 			outstanding = tp->snd_max - tp->snd_una;
20565 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) {
20566 				if (tp->snd_wnd > outstanding) {
20567 					len = tp->snd_wnd - outstanding;
20568 					/* Check to see if we have the data */
20569 					if ((sb_offset + len) > avail) {
20570 						/* It does not all fit */
20571 						if (avail > sb_offset)
20572 							len = avail - sb_offset;
20573 						else
20574 							len = 0;
20575 					}
20576 				} else {
20577 					len = 0;
20578 				}
20579 			} else if (avail > sb_offset) {
20580 				len = avail - sb_offset;
20581 			} else {
20582 				len = 0;
20583 			}
20584 			if (len > 0) {
20585 				if (len > rack->r_ctl.rc_prr_sndcnt) {
20586 					len = rack->r_ctl.rc_prr_sndcnt;
20587 				}
20588 				if (len > 0) {
20589 					sub_from_prr = 1;
20590 				}
20591 			}
20592 			if (len > segsiz) {
20593 				/*
20594 				 * We should never send more than a MSS when
20595 				 * retransmitting or sending new data in prr
20596 				 * mode unless the override flag is on. Most
20597 				 * likely the PRR algorithm is not going to
20598 				 * let us send a lot as well :-)
20599 				 */
20600 				if (rack->r_ctl.rc_prr_sendalot == 0) {
20601 					len = segsiz;
20602 				}
20603 			} else if (len < segsiz) {
20604 				/*
20605 				 * Do we send any? The idea here is if the
20606 				 * send empty's the socket buffer we want to
20607 				 * do it. However if not then lets just wait
20608 				 * for our prr_sndcnt to get bigger.
20609 				 */
20610 				long leftinsb;
20611 
20612 				leftinsb = sbavail(sb) - sb_offset;
20613 				if (leftinsb > len) {
20614 					/* This send does not empty the sb */
20615 					len = 0;
20616 				}
20617 			}
20618 		}
20619 	} else if (!TCPS_HAVEESTABLISHED(tp->t_state)) {
20620 		/*
20621 		 * If you have not established
20622 		 * and are not doing FAST OPEN
20623 		 * no data please.
20624 		 */
20625 		if ((sack_rxmit == 0) &&
20626 		    (!IS_FASTOPEN(tp->t_flags))){
20627 			len = 0;
20628 			sb_offset = 0;
20629 		}
20630 	}
20631 	if (prefetch_so_done == 0) {
20632 		kern_prefetch(so, &prefetch_so_done);
20633 		prefetch_so_done = 1;
20634 	}
20635 	/*
20636 	 * Lop off SYN bit if it has already been sent.  However, if this is
20637 	 * SYN-SENT state and if segment contains data and if we don't know
20638 	 * that foreign host supports TAO, suppress sending segment.
20639 	 */
20640 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
20641 	    ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
20642 		/*
20643 		 * When sending additional segments following a TFO SYN|ACK,
20644 		 * do not include the SYN bit.
20645 		 */
20646 		if (IS_FASTOPEN(tp->t_flags) &&
20647 		    (tp->t_state == TCPS_SYN_RECEIVED))
20648 			flags &= ~TH_SYN;
20649 	}
20650 	/*
20651 	 * Be careful not to send data and/or FIN on SYN segments. This
20652 	 * measure is needed to prevent interoperability problems with not
20653 	 * fully conformant TCP implementations.
20654 	 */
20655 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
20656 		len = 0;
20657 		flags &= ~TH_FIN;
20658 	}
20659 	/*
20660 	 * On TFO sockets, ensure no data is sent in the following cases:
20661 	 *
20662 	 *  - When retransmitting SYN|ACK on a passively-created socket
20663 	 *
20664 	 *  - When retransmitting SYN on an actively created socket
20665 	 *
20666 	 *  - When sending a zero-length cookie (cookie request) on an
20667 	 *    actively created socket
20668 	 *
20669 	 *  - When the socket is in the CLOSED state (RST is being sent)
20670 	 */
20671 	if (IS_FASTOPEN(tp->t_flags) &&
20672 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
20673 	     ((tp->t_state == TCPS_SYN_SENT) &&
20674 	      (tp->t_tfo_client_cookie_len == 0)) ||
20675 	     (flags & TH_RST))) {
20676 		sack_rxmit = 0;
20677 		len = 0;
20678 	}
20679 	/* Without fast-open there should never be data sent on a SYN */
20680 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) {
20681 		tp->snd_nxt = tp->iss;
20682 		len = 0;
20683 	}
20684 	if ((len > segsiz) && (tcp_dsack_block_exists(tp))) {
20685 		/* We only send 1 MSS if we have a DSACK block */
20686 		add_flag |= RACK_SENT_W_DSACK;
20687 		len = segsiz;
20688 	}
20689 	orig_len = len;
20690 	if (len <= 0) {
20691 		/*
20692 		 * If FIN has been sent but not acked, but we haven't been
20693 		 * called to retransmit, len will be < 0.  Otherwise, window
20694 		 * shrank after we sent into it.  If window shrank to 0,
20695 		 * cancel pending retransmit, pull snd_nxt back to (closed)
20696 		 * window, and set the persist timer if it isn't already
20697 		 * going.  If the window didn't close completely, just wait
20698 		 * for an ACK.
20699 		 *
20700 		 * We also do a general check here to ensure that we will
20701 		 * set the persist timer when we have data to send, but a
20702 		 * 0-byte window. This makes sure the persist timer is set
20703 		 * even if the packet hits one of the "goto send" lines
20704 		 * below.
20705 		 */
20706 		len = 0;
20707 		if ((tp->snd_wnd == 0) &&
20708 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20709 		    (tp->snd_una == tp->snd_max) &&
20710 		    (sb_offset < (int)sbavail(sb))) {
20711 			rack_enter_persist(tp, rack, cts, tp->snd_una);
20712 		}
20713 	} else if ((rsm == NULL) &&
20714 		   (doing_tlp == 0) &&
20715 		   (len < pace_max_seg)) {
20716 		/*
20717 		 * We are not sending a maximum sized segment for
20718 		 * some reason. Should we not send anything (think
20719 		 * sws or persists)?
20720 		 */
20721 		if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20722 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
20723 		    (len < minseg) &&
20724 		    (len < (int)(sbavail(sb) - sb_offset))) {
20725 			/*
20726 			 * Here the rwnd is less than
20727 			 * the minimum pacing size, this is not a retransmit,
20728 			 * we are established and
20729 			 * the send is not the last in the socket buffer
20730 			 * we send nothing, and we may enter persists
20731 			 * if nothing is outstanding.
20732 			 */
20733 			len = 0;
20734 			if (tp->snd_max == tp->snd_una) {
20735 				/*
20736 				 * Nothing out we can
20737 				 * go into persists.
20738 				 */
20739 				rack_enter_persist(tp, rack, cts, tp->snd_una);
20740 			}
20741 		} else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) &&
20742 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20743 			   (len < (int)(sbavail(sb) - sb_offset)) &&
20744 			   (len < minseg)) {
20745 			/*
20746 			 * Here we are not retransmitting, and
20747 			 * the cwnd is not so small that we could
20748 			 * not send at least a min size (rxt timer
20749 			 * not having gone off), We have 2 segments or
20750 			 * more already in flight, its not the tail end
20751 			 * of the socket buffer  and the cwnd is blocking
20752 			 * us from sending out a minimum pacing segment size.
20753 			 * Lets not send anything.
20754 			 */
20755 			len = 0;
20756 		} else if (((tp->snd_wnd - ctf_outstanding(tp)) <
20757 			    min((rack->r_ctl.rc_high_rwnd/2), minseg)) &&
20758 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) &&
20759 			   (len < (int)(sbavail(sb) - sb_offset)) &&
20760 			   (TCPS_HAVEESTABLISHED(tp->t_state))) {
20761 			/*
20762 			 * Here we have a send window but we have
20763 			 * filled it up and we can't send another pacing segment.
20764 			 * We also have in flight more than 2 segments
20765 			 * and we are not completing the sb i.e. we allow
20766 			 * the last bytes of the sb to go out even if
20767 			 * its not a full pacing segment.
20768 			 */
20769 			len = 0;
20770 		} else if ((rack->r_ctl.crte != NULL) &&
20771 			   (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) &&
20772 			   (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) &&
20773 			   (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) &&
20774 			   (len < (int)(sbavail(sb) - sb_offset))) {
20775 			/*
20776 			 * Here we are doing hardware pacing, this is not a TLP,
20777 			 * we are not sending a pace max segment size, there is rwnd
20778 			 * room to send at least N pace_max_seg, the cwnd is greater
20779 			 * than or equal to a full pacing segments plus 4 mss and we have 2 or
20780 			 * more segments in flight and its not the tail of the socket buffer.
20781 			 *
20782 			 * We don't want to send instead we need to get more ack's in to
20783 			 * allow us to send a full pacing segment. Normally, if we are pacing
20784 			 * about the right speed, we should have finished our pacing
20785 			 * send as most of the acks have come back if we are at the
20786 			 * right rate. This is a bit fuzzy since return path delay
20787 			 * can delay the acks, which is why we want to make sure we
20788 			 * have cwnd space to have a bit more than a max pace segments in flight.
20789 			 *
20790 			 * If we have not gotten our acks back we are pacing at too high a
20791 			 * rate delaying will not hurt and will bring our GP estimate down by
20792 			 * injecting the delay. If we don't do this we will send
20793 			 * 2 MSS out in response to the acks being clocked in which
20794 			 * defeats the point of hw-pacing (i.e. to help us get
20795 			 * larger TSO's out).
20796 			 */
20797 			len = 0;
20798 		}
20799 
20800 	}
20801 	/* len will be >= 0 after this point. */
20802 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
20803 	rack_sndbuf_autoscale(rack);
20804 	/*
20805 	 * Decide if we can use TCP Segmentation Offloading (if supported by
20806 	 * hardware).
20807 	 *
20808 	 * TSO may only be used if we are in a pure bulk sending state.  The
20809 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
20810 	 * options prevent using TSO.  With TSO the TCP header is the same
20811 	 * (except for the sequence number) for all generated packets.  This
20812 	 * makes it impossible to transmit any options which vary per
20813 	 * generated segment or packet.
20814 	 *
20815 	 * IPv4 handling has a clear separation of ip options and ip header
20816 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
20817 	 * the right thing below to provide length of just ip options and thus
20818 	 * checking for ipoptlen is enough to decide if ip options are present.
20819 	 */
20820 	ipoptlen = 0;
20821 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20822 	/*
20823 	 * Pre-calculate here as we save another lookup into the darknesses
20824 	 * of IPsec that way and can actually decide if TSO is ok.
20825 	 */
20826 #ifdef INET6
20827 	if (isipv6 && IPSEC_ENABLED(ipv6))
20828 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
20829 #ifdef INET
20830 	else
20831 #endif
20832 #endif				/* INET6 */
20833 #ifdef INET
20834 		if (IPSEC_ENABLED(ipv4))
20835 			ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
20836 #endif				/* INET */
20837 #endif
20838 
20839 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
20840 	ipoptlen += ipsec_optlen;
20841 #endif
20842 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz &&
20843 	    (tp->t_port == 0) &&
20844 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
20845 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
20846 	    ipoptlen == 0)
20847 		tso = 1;
20848 	{
20849 		uint32_t outstanding __unused;
20850 
20851 		outstanding = tp->snd_max - tp->snd_una;
20852 		if (tp->t_flags & TF_SENTFIN) {
20853 			/*
20854 			 * If we sent a fin, snd_max is 1 higher than
20855 			 * snd_una
20856 			 */
20857 			outstanding--;
20858 		}
20859 		if (sack_rxmit) {
20860 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
20861 				flags &= ~TH_FIN;
20862 		} else {
20863 			if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
20864 				   sbused(sb)))
20865 				flags &= ~TH_FIN;
20866 		}
20867 	}
20868 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
20869 		      (long)TCP_MAXWIN << tp->rcv_scale);
20870 
20871 	/*
20872 	 * Sender silly window avoidance.   We transmit under the following
20873 	 * conditions when len is non-zero:
20874 	 *
20875 	 * - We have a full segment (or more with TSO) - This is the last
20876 	 * buffer in a write()/send() and we are either idle or running
20877 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
20878 	 * then 1/2 the maximum send window's worth of data (receiver may be
20879 	 * limited the window size) - we need to retransmit
20880 	 */
20881 	if (len) {
20882 		if (len >= segsiz) {
20883 			goto send;
20884 		}
20885 		/*
20886 		 * NOTE! on localhost connections an 'ack' from the remote
20887 		 * end may occur synchronously with the output and cause us
20888 		 * to flush a buffer queued with moretocome.  XXX
20889 		 *
20890 		 */
20891 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
20892 		    (idle || (tp->t_flags & TF_NODELAY)) &&
20893 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
20894 		    (tp->t_flags & TF_NOPUSH) == 0) {
20895 			pass = 2;
20896 			goto send;
20897 		}
20898 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
20899 			pass = 22;
20900 			goto send;
20901 		}
20902 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
20903 			pass = 4;
20904 			goto send;
20905 		}
20906 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {	/* retransmit case */
20907 			pass = 5;
20908 			goto send;
20909 		}
20910 		if (sack_rxmit) {
20911 			pass = 6;
20912 			goto send;
20913 		}
20914 		if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) &&
20915 		    (ctf_outstanding(tp) < (segsiz * 2))) {
20916 			/*
20917 			 * We have less than two MSS outstanding (delayed ack)
20918 			 * and our rwnd will not let us send a full sized
20919 			 * MSS. Lets go ahead and let this small segment
20920 			 * out because we want to try to have at least two
20921 			 * packets inflight to not be caught by delayed ack.
20922 			 */
20923 			pass = 12;
20924 			goto send;
20925 		}
20926 	}
20927 	/*
20928 	 * Sending of standalone window updates.
20929 	 *
20930 	 * Window updates are important when we close our window due to a
20931 	 * full socket buffer and are opening it again after the application
20932 	 * reads data from it.  Once the window has opened again and the
20933 	 * remote end starts to send again the ACK clock takes over and
20934 	 * provides the most current window information.
20935 	 *
20936 	 * We must avoid the silly window syndrome whereas every read from
20937 	 * the receive buffer, no matter how small, causes a window update
20938 	 * to be sent.  We also should avoid sending a flurry of window
20939 	 * updates when the socket buffer had queued a lot of data and the
20940 	 * application is doing small reads.
20941 	 *
20942 	 * Prevent a flurry of pointless window updates by only sending an
20943 	 * update when we can increase the advertized window by more than
20944 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
20945 	 * full or is very small be more aggressive and send an update
20946 	 * whenever we can increase by two mss sized segments. In all other
20947 	 * situations the ACK's to new incoming data will carry further
20948 	 * window increases.
20949 	 *
20950 	 * Don't send an independent window update if a delayed ACK is
20951 	 * pending (it will get piggy-backed on it) or the remote side
20952 	 * already has done a half-close and won't send more data.  Skip
20953 	 * this if the connection is in T/TCP half-open state.
20954 	 */
20955 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
20956 	    !(tp->t_flags & TF_DELACK) &&
20957 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
20958 		/*
20959 		 * "adv" is the amount we could increase the window, taking
20960 		 * into account that we are limited by TCP_MAXWIN <<
20961 		 * tp->rcv_scale.
20962 		 */
20963 		int32_t adv;
20964 		int oldwin;
20965 
20966 		adv = recwin;
20967 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
20968 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
20969 			if (adv > oldwin)
20970 				adv -= oldwin;
20971 			else {
20972 				/* We can't increase the window */
20973 				adv = 0;
20974 			}
20975 		} else
20976 			oldwin = 0;
20977 
20978 		/*
20979 		 * If the new window size ends up being the same as or less
20980 		 * than the old size when it is scaled, then don't force
20981 		 * a window update.
20982 		 */
20983 		if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
20984 			goto dontupdate;
20985 
20986 		if (adv >= (int32_t)(2 * segsiz) &&
20987 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
20988 		     recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
20989 		     so->so_rcv.sb_hiwat <= 8 * segsiz)) {
20990 			pass = 7;
20991 			goto send;
20992 		}
20993 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) {
20994 			pass = 23;
20995 			goto send;
20996 		}
20997 	}
20998 dontupdate:
20999 
21000 	/*
21001 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
21002 	 * is also a catch-all for the retransmit timer timeout case.
21003 	 */
21004 	if (tp->t_flags & TF_ACKNOW) {
21005 		pass = 8;
21006 		goto send;
21007 	}
21008 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
21009 		pass = 9;
21010 		goto send;
21011 	}
21012 	/*
21013 	 * If our state indicates that FIN should be sent and we have not
21014 	 * yet done so, then we need to send.
21015 	 */
21016 	if ((flags & TH_FIN) &&
21017 	    (tp->snd_nxt == tp->snd_una)) {
21018 		pass = 11;
21019 		goto send;
21020 	}
21021 	/*
21022 	 * No reason to send a segment, just return.
21023 	 */
21024 just_return:
21025 	SOCKBUF_UNLOCK(sb);
21026 just_return_nolock:
21027 	{
21028 		int app_limited = CTF_JR_SENT_DATA;
21029 
21030 		if (tot_len_this_send > 0) {
21031 			/* Make sure snd_nxt is up to max */
21032 			rack->r_ctl.fsb.recwin = recwin;
21033 			slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz);
21034 			if ((error == 0) &&
21035 			    rack_use_rfo &&
21036 			    ((flags & (TH_SYN|TH_FIN)) == 0) &&
21037 			    (ipoptlen == 0) &&
21038 			    (tp->snd_nxt == tp->snd_max) &&
21039 			    (tp->rcv_numsacks == 0) &&
21040 			    rack->r_fsb_inited &&
21041 			    TCPS_HAVEESTABLISHED(tp->t_state) &&
21042 			    ((IN_RECOVERY(tp->t_flags)) == 0) &&
21043 			    (rack->r_must_retran == 0) &&
21044 			    ((tp->t_flags & TF_NEEDFIN) == 0) &&
21045 			    (len > 0) && (orig_len > 0) &&
21046 			    (orig_len > len) &&
21047 			    ((orig_len - len) >= segsiz) &&
21048 			    ((optlen == 0) ||
21049 			     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
21050 				/* We can send at least one more MSS using our fsb */
21051 				rack_setup_fast_output(tp, rack, sb, len, orig_len,
21052 						       segsiz, pace_max_seg, hw_tls, flags);
21053 			} else
21054 				rack->r_fast_output = 0;
21055 
21056 
21057 			rack_log_fsb(rack, tp, so, flags,
21058 				     ipoptlen, orig_len, len, 0,
21059 				     1, optlen, __LINE__, 1);
21060 			if (SEQ_GT(tp->snd_max, tp->snd_nxt))
21061 				tp->snd_nxt = tp->snd_max;
21062 		} else {
21063 			int end_window = 0;
21064 			uint32_t seq = tp->gput_ack;
21065 
21066 			rsm = tqhash_max(rack->r_ctl.tqh);
21067 			if (rsm) {
21068 				/*
21069 				 * Mark the last sent that we just-returned (hinting
21070 				 * that delayed ack may play a role in any rtt measurement).
21071 				 */
21072 				rsm->r_just_ret = 1;
21073 			}
21074 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
21075 			rack->r_ctl.rc_agg_delayed = 0;
21076 			rack->r_early = 0;
21077 			rack->r_late = 0;
21078 			rack->r_ctl.rc_agg_early = 0;
21079 			if ((ctf_outstanding(tp) +
21080 			     min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)),
21081 				 minseg)) >= tp->snd_wnd) {
21082 				/* We are limited by the rwnd */
21083 				app_limited = CTF_JR_RWND_LIMITED;
21084 				if (IN_FASTRECOVERY(tp->t_flags))
21085 					rack->r_ctl.rc_prr_sndcnt = 0;
21086 			} else if (ctf_outstanding(tp) >= sbavail(sb)) {
21087 				/* We are limited by whats available -- app limited */
21088 				app_limited = CTF_JR_APP_LIMITED;
21089 				if (IN_FASTRECOVERY(tp->t_flags))
21090 					rack->r_ctl.rc_prr_sndcnt = 0;
21091 			} else if ((idle == 0) &&
21092 				   ((tp->t_flags & TF_NODELAY) == 0) &&
21093 				   ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) &&
21094 				   (len < segsiz)) {
21095 				/*
21096 				 * No delay is not on and the
21097 				 * user is sending less than 1MSS. This
21098 				 * brings out SWS avoidance so we
21099 				 * don't send. Another app-limited case.
21100 				 */
21101 				app_limited = CTF_JR_APP_LIMITED;
21102 			} else if (tp->t_flags & TF_NOPUSH) {
21103 				/*
21104 				 * The user has requested no push of
21105 				 * the last segment and we are
21106 				 * at the last segment. Another app
21107 				 * limited case.
21108 				 */
21109 				app_limited = CTF_JR_APP_LIMITED;
21110 			} else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) {
21111 				/* Its the cwnd */
21112 				app_limited = CTF_JR_CWND_LIMITED;
21113 			} else if (IN_FASTRECOVERY(tp->t_flags) &&
21114 				   (rack->rack_no_prr == 0) &&
21115 				   (rack->r_ctl.rc_prr_sndcnt < segsiz)) {
21116 				app_limited = CTF_JR_PRR;
21117 			} else {
21118 				/* Now why here are we not sending? */
21119 #ifdef NOW
21120 #ifdef INVARIANTS
21121 				panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use);
21122 #endif
21123 #endif
21124 				app_limited = CTF_JR_ASSESSING;
21125 			}
21126 			/*
21127 			 * App limited in some fashion, for our pacing GP
21128 			 * measurements we don't want any gap (even cwnd).
21129 			 * Close  down the measurement window.
21130 			 */
21131 			if (rack_cwnd_block_ends_measure &&
21132 			    ((app_limited == CTF_JR_CWND_LIMITED) ||
21133 			     (app_limited == CTF_JR_PRR))) {
21134 				/*
21135 				 * The reason we are not sending is
21136 				 * the cwnd (or prr). We have been configured
21137 				 * to end the measurement window in
21138 				 * this case.
21139 				 */
21140 				end_window = 1;
21141 			} else if (rack_rwnd_block_ends_measure &&
21142 				   (app_limited == CTF_JR_RWND_LIMITED)) {
21143 				/*
21144 				 * We are rwnd limited and have been
21145 				 * configured to end the measurement
21146 				 * window in this case.
21147 				 */
21148 				end_window = 1;
21149 			} else if (app_limited == CTF_JR_APP_LIMITED) {
21150 				/*
21151 				 * A true application limited period, we have
21152 				 * ran out of data.
21153 				 */
21154 				end_window = 1;
21155 			} else if (app_limited == CTF_JR_ASSESSING) {
21156 				/*
21157 				 * In the assessing case we hit the end of
21158 				 * the if/else and had no known reason
21159 				 * This will panic us under invariants..
21160 				 *
21161 				 * If we get this out in logs we need to
21162 				 * investagate which reason we missed.
21163 				 */
21164 				end_window = 1;
21165 			}
21166 			if (end_window) {
21167 				uint8_t log = 0;
21168 
21169 				/* Adjust the Gput measurement */
21170 				if ((tp->t_flags & TF_GPUTINPROG) &&
21171 				    SEQ_GT(tp->gput_ack, tp->snd_max)) {
21172 					tp->gput_ack = tp->snd_max;
21173 					if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) {
21174 						/*
21175 						 * There is not enough to measure.
21176 						 */
21177 						tp->t_flags &= ~TF_GPUTINPROG;
21178 						rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/,
21179 									   rack->r_ctl.rc_gp_srtt /*flex1*/,
21180 									   tp->gput_seq,
21181 									   0, 0, 18, __LINE__, NULL, 0);
21182 					} else
21183 						log = 1;
21184 				}
21185 				/* Mark the last packet has app limited */
21186 				rsm = tqhash_max(rack->r_ctl.tqh);
21187 				if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) {
21188 					if (rack->r_ctl.rc_app_limited_cnt == 0)
21189 						rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm;
21190 					else {
21191 						/*
21192 						 * Go out to the end app limited and mark
21193 						 * this new one as next and move the end_appl up
21194 						 * to this guy.
21195 						 */
21196 						if (rack->r_ctl.rc_end_appl)
21197 							rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start;
21198 						rack->r_ctl.rc_end_appl = rsm;
21199 					}
21200 					rsm->r_flags |= RACK_APP_LIMITED;
21201 					rack->r_ctl.rc_app_limited_cnt++;
21202 				}
21203 				if (log)
21204 					rack_log_pacing_delay_calc(rack,
21205 								   rack->r_ctl.rc_app_limited_cnt, seq,
21206 								   tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0);
21207 			}
21208 		}
21209 		/* Check if we need to go into persists or not */
21210 		if ((tp->snd_max == tp->snd_una) &&
21211 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
21212 		    sbavail(sb) &&
21213 		    (sbavail(sb) > tp->snd_wnd) &&
21214 		    (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) {
21215 			/* Yes lets make sure to move to persist before timer-start */
21216 			rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime, tp->snd_una);
21217 		}
21218 		rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack);
21219 		rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use);
21220 	}
21221 #ifdef NETFLIX_SHARED_CWND
21222 	if ((sbavail(sb) == 0) &&
21223 	    rack->r_ctl.rc_scw) {
21224 		tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index);
21225 		rack->rack_scwnd_is_idle = 1;
21226 	}
21227 #endif
21228 #ifdef TCP_ACCOUNTING
21229 	if (tot_len_this_send > 0) {
21230 		crtsc = get_cyclecount();
21231 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21232 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
21233 		}
21234 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21235 			tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val);
21236 		}
21237 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21238 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz);
21239 		}
21240 	} else {
21241 		crtsc = get_cyclecount();
21242 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21243 			tp->tcp_cnt_counters[SND_LIMITED]++;
21244 		}
21245 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21246 			tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val);
21247 		}
21248 	}
21249 	sched_unpin();
21250 #endif
21251 	return (0);
21252 
21253 send:
21254 	if ((rack->r_ctl.crte != NULL) &&
21255 	    (rsm == NULL) &&
21256 	    ((rack->rc_hw_nobuf == 1) ||
21257 	     (rack_hw_check_queue && (check_done == 0)))) {
21258 		/*
21259 		 * We only want to do this once with the hw_check_queue,
21260 		 * for the enobuf case we would only do it once if
21261 		 * we come around to again, the flag will be clear.
21262 		 */
21263 		check_done = 1;
21264 		slot = rack_check_queue_level(rack, tp, &tv, cts, len, segsiz);
21265 		if (slot) {
21266 			rack->r_ctl.rc_agg_delayed = 0;
21267 			rack->r_ctl.rc_agg_early = 0;
21268 			rack->r_early = 0;
21269 			rack->r_late = 0;
21270 			SOCKBUF_UNLOCK(&so->so_snd);
21271 			goto skip_all_send;
21272 		}
21273 	}
21274 	if (rsm || sack_rxmit)
21275 		counter_u64_add(rack_nfto_resend, 1);
21276 	else
21277 		counter_u64_add(rack_non_fto_send, 1);
21278 	if ((flags & TH_FIN) &&
21279 	    sbavail(sb)) {
21280 		/*
21281 		 * We do not transmit a FIN
21282 		 * with data outstanding. We
21283 		 * need to make it so all data
21284 		 * is acked first.
21285 		 */
21286 		flags &= ~TH_FIN;
21287 	}
21288 	/* Enforce stack imposed max seg size if we have one */
21289 	if (rack->r_ctl.rc_pace_max_segs &&
21290 	    (len > rack->r_ctl.rc_pace_max_segs)) {
21291 		mark = 1;
21292 		len = rack->r_ctl.rc_pace_max_segs;
21293 	}
21294 	SOCKBUF_LOCK_ASSERT(sb);
21295 	if (len > 0) {
21296 		if (len >= segsiz)
21297 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
21298 		else
21299 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
21300 	}
21301 	/*
21302 	 * Before ESTABLISHED, force sending of initial options unless TCP
21303 	 * set not to do any options. NOTE: we assume that the IP/TCP header
21304 	 * plus TCP options always fit in a single mbuf, leaving room for a
21305 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
21306 	 * + optlen <= MCLBYTES
21307 	 */
21308 	optlen = 0;
21309 #ifdef INET6
21310 	if (isipv6)
21311 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
21312 	else
21313 #endif
21314 		hdrlen = sizeof(struct tcpiphdr);
21315 
21316 	/*
21317 	 * Compute options for segment. We only have to care about SYN and
21318 	 * established connection segments.  Options for SYN-ACK segments
21319 	 * are handled in TCP syncache.
21320 	 */
21321 	to.to_flags = 0;
21322 	if ((tp->t_flags & TF_NOOPT) == 0) {
21323 		/* Maximum segment size. */
21324 		if (flags & TH_SYN) {
21325 			tp->snd_nxt = tp->iss;
21326 			to.to_mss = tcp_mssopt(&inp->inp_inc);
21327 			if (tp->t_port)
21328 				to.to_mss -= V_tcp_udp_tunneling_overhead;
21329 			to.to_flags |= TOF_MSS;
21330 
21331 			/*
21332 			 * On SYN or SYN|ACK transmits on TFO connections,
21333 			 * only include the TFO option if it is not a
21334 			 * retransmit, as the presence of the TFO option may
21335 			 * have caused the original SYN or SYN|ACK to have
21336 			 * been dropped by a middlebox.
21337 			 */
21338 			if (IS_FASTOPEN(tp->t_flags) &&
21339 			    (tp->t_rxtshift == 0)) {
21340 				if (tp->t_state == TCPS_SYN_RECEIVED) {
21341 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
21342 					to.to_tfo_cookie =
21343 						(u_int8_t *)&tp->t_tfo_cookie.server;
21344 					to.to_flags |= TOF_FASTOPEN;
21345 					wanted_cookie = 1;
21346 				} else if (tp->t_state == TCPS_SYN_SENT) {
21347 					to.to_tfo_len =
21348 						tp->t_tfo_client_cookie_len;
21349 					to.to_tfo_cookie =
21350 						tp->t_tfo_cookie.client;
21351 					to.to_flags |= TOF_FASTOPEN;
21352 					wanted_cookie = 1;
21353 					/*
21354 					 * If we wind up having more data to
21355 					 * send with the SYN than can fit in
21356 					 * one segment, don't send any more
21357 					 * until the SYN|ACK comes back from
21358 					 * the other end.
21359 					 */
21360 					sendalot = 0;
21361 				}
21362 			}
21363 		}
21364 		/* Window scaling. */
21365 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
21366 			to.to_wscale = tp->request_r_scale;
21367 			to.to_flags |= TOF_SCALE;
21368 		}
21369 		/* Timestamps. */
21370 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
21371 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
21372 			to.to_tsval = ms_cts + tp->ts_offset;
21373 			to.to_tsecr = tp->ts_recent;
21374 			to.to_flags |= TOF_TS;
21375 		}
21376 		/* Set receive buffer autosizing timestamp. */
21377 		if (tp->rfbuf_ts == 0 &&
21378 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
21379 			tp->rfbuf_ts = tcp_ts_getticks();
21380 		/* Selective ACK's. */
21381 		if (tp->t_flags & TF_SACK_PERMIT) {
21382 			if (flags & TH_SYN)
21383 				to.to_flags |= TOF_SACKPERM;
21384 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
21385 				 tp->rcv_numsacks > 0) {
21386 				to.to_flags |= TOF_SACK;
21387 				to.to_nsacks = tp->rcv_numsacks;
21388 				to.to_sacks = (u_char *)tp->sackblks;
21389 			}
21390 		}
21391 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21392 		/* TCP-MD5 (RFC2385). */
21393 		if (tp->t_flags & TF_SIGNATURE)
21394 			to.to_flags |= TOF_SIGNATURE;
21395 #endif
21396 
21397 		/* Processing the options. */
21398 		hdrlen += optlen = tcp_addoptions(&to, opt);
21399 		/*
21400 		 * If we wanted a TFO option to be added, but it was unable
21401 		 * to fit, ensure no data is sent.
21402 		 */
21403 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
21404 		    !(to.to_flags & TOF_FASTOPEN))
21405 			len = 0;
21406 	}
21407 	if (tp->t_port) {
21408 		if (V_tcp_udp_tunneling_port == 0) {
21409 			/* The port was removed?? */
21410 			SOCKBUF_UNLOCK(&so->so_snd);
21411 #ifdef TCP_ACCOUNTING
21412 			crtsc = get_cyclecount();
21413 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21414 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
21415 			}
21416 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
21417 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
21418 			}
21419 			sched_unpin();
21420 #endif
21421 			return (EHOSTUNREACH);
21422 		}
21423 		hdrlen += sizeof(struct udphdr);
21424 	}
21425 #ifdef INET6
21426 	if (isipv6)
21427 		ipoptlen = ip6_optlen(inp);
21428 	else
21429 #endif
21430 		if (inp->inp_options)
21431 			ipoptlen = inp->inp_options->m_len -
21432 				offsetof(struct ipoption, ipopt_list);
21433 		else
21434 			ipoptlen = 0;
21435 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
21436 	ipoptlen += ipsec_optlen;
21437 #endif
21438 
21439 	/*
21440 	 * Adjust data length if insertion of options will bump the packet
21441 	 * length beyond the t_maxseg length. Clear the FIN bit because we
21442 	 * cut off the tail of the segment.
21443 	 */
21444 	if (len + optlen + ipoptlen > tp->t_maxseg) {
21445 		if (tso) {
21446 			uint32_t if_hw_tsomax;
21447 			uint32_t moff;
21448 			int32_t max_len;
21449 
21450 			/* extract TSO information */
21451 			if_hw_tsomax = tp->t_tsomax;
21452 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
21453 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
21454 			KASSERT(ipoptlen == 0,
21455 				("%s: TSO can't do IP options", __func__));
21456 
21457 			/*
21458 			 * Check if we should limit by maximum payload
21459 			 * length:
21460 			 */
21461 			if (if_hw_tsomax != 0) {
21462 				/* compute maximum TSO length */
21463 				max_len = (if_hw_tsomax - hdrlen -
21464 					   max_linkhdr);
21465 				if (max_len <= 0) {
21466 					len = 0;
21467 				} else if (len > max_len) {
21468 					sendalot = 1;
21469 					len = max_len;
21470 					mark = 2;
21471 				}
21472 			}
21473 			/*
21474 			 * Prevent the last segment from being fractional
21475 			 * unless the send sockbuf can be emptied:
21476 			 */
21477 			max_len = (tp->t_maxseg - optlen);
21478 			if ((sb_offset + len) < sbavail(sb)) {
21479 				moff = len % (u_int)max_len;
21480 				if (moff != 0) {
21481 					mark = 3;
21482 					len -= moff;
21483 				}
21484 			}
21485 			/*
21486 			 * In case there are too many small fragments don't
21487 			 * use TSO:
21488 			 */
21489 			if (len <= max_len) {
21490 				mark = 4;
21491 				tso = 0;
21492 			}
21493 			/*
21494 			 * Send the FIN in a separate segment after the bulk
21495 			 * sending is done. We don't trust the TSO
21496 			 * implementations to clear the FIN flag on all but
21497 			 * the last segment.
21498 			 */
21499 			if (tp->t_flags & TF_NEEDFIN) {
21500 				sendalot = 4;
21501 			}
21502 		} else {
21503 			mark = 5;
21504 			if (optlen + ipoptlen >= tp->t_maxseg) {
21505 				/*
21506 				 * Since we don't have enough space to put
21507 				 * the IP header chain and the TCP header in
21508 				 * one packet as required by RFC 7112, don't
21509 				 * send it. Also ensure that at least one
21510 				 * byte of the payload can be put into the
21511 				 * TCP segment.
21512 				 */
21513 				SOCKBUF_UNLOCK(&so->so_snd);
21514 				error = EMSGSIZE;
21515 				sack_rxmit = 0;
21516 				goto out;
21517 			}
21518 			len = tp->t_maxseg - optlen - ipoptlen;
21519 			sendalot = 5;
21520 		}
21521 	} else {
21522 		tso = 0;
21523 		mark = 6;
21524 	}
21525 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
21526 		("%s: len > IP_MAXPACKET", __func__));
21527 #ifdef DIAGNOSTIC
21528 #ifdef INET6
21529 	if (max_linkhdr + hdrlen > MCLBYTES)
21530 #else
21531 		if (max_linkhdr + hdrlen > MHLEN)
21532 #endif
21533 			panic("tcphdr too big");
21534 #endif
21535 
21536 	/*
21537 	 * This KASSERT is here to catch edge cases at a well defined place.
21538 	 * Before, those had triggered (random) panic conditions further
21539 	 * down.
21540 	 */
21541 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
21542 	if ((len == 0) &&
21543 	    (flags & TH_FIN) &&
21544 	    (sbused(sb))) {
21545 		/*
21546 		 * We have outstanding data, don't send a fin by itself!.
21547 		 */
21548 		goto just_return;
21549 	}
21550 	/*
21551 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
21552 	 * and initialize the header from the template for sends on this
21553 	 * connection.
21554 	 */
21555 	hw_tls = tp->t_nic_ktls_xmit != 0;
21556 	if (len) {
21557 		uint32_t max_val;
21558 		uint32_t moff;
21559 
21560 		if (rack->r_ctl.rc_pace_max_segs)
21561 			max_val = rack->r_ctl.rc_pace_max_segs;
21562 		else if (rack->rc_user_set_max_segs)
21563 			max_val = rack->rc_user_set_max_segs * segsiz;
21564 		else
21565 			max_val = len;
21566 		/*
21567 		 * We allow a limit on sending with hptsi.
21568 		 */
21569 		if (len > max_val) {
21570 			mark = 7;
21571 			len = max_val;
21572 		}
21573 #ifdef INET6
21574 		if (MHLEN < hdrlen + max_linkhdr)
21575 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
21576 		else
21577 #endif
21578 			m = m_gethdr(M_NOWAIT, MT_DATA);
21579 
21580 		if (m == NULL) {
21581 			SOCKBUF_UNLOCK(sb);
21582 			error = ENOBUFS;
21583 			sack_rxmit = 0;
21584 			goto out;
21585 		}
21586 		m->m_data += max_linkhdr;
21587 		m->m_len = hdrlen;
21588 
21589 		/*
21590 		 * Start the m_copy functions from the closest mbuf to the
21591 		 * sb_offset in the socket buffer chain.
21592 		 */
21593 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
21594 		s_mb = mb;
21595 		s_moff = moff;
21596 		if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
21597 			m_copydata(mb, moff, (int)len,
21598 				   mtod(m, caddr_t)+hdrlen);
21599 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
21600 				sbsndptr_adv(sb, mb, len);
21601 			m->m_len += len;
21602 		} else {
21603 			struct sockbuf *msb;
21604 
21605 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
21606 				msb = NULL;
21607 			else
21608 				msb = sb;
21609 			m->m_next = tcp_m_copym(
21610 				mb, moff, &len,
21611 				if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb,
21612 				((rsm == NULL) ? hw_tls : 0)
21613 #ifdef NETFLIX_COPY_ARGS
21614 				, &s_mb, &s_moff
21615 #endif
21616 				);
21617 			if (len <= (tp->t_maxseg - optlen)) {
21618 				/*
21619 				 * Must have ran out of mbufs for the copy
21620 				 * shorten it to no longer need tso. Lets
21621 				 * not put on sendalot since we are low on
21622 				 * mbufs.
21623 				 */
21624 				tso = 0;
21625 			}
21626 			if (m->m_next == NULL) {
21627 				SOCKBUF_UNLOCK(sb);
21628 				(void)m_free(m);
21629 				error = ENOBUFS;
21630 				sack_rxmit = 0;
21631 				goto out;
21632 			}
21633 		}
21634 		if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
21635 			if (rsm && (rsm->r_flags & RACK_TLP)) {
21636 				/*
21637 				 * TLP should not count in retran count, but
21638 				 * in its own bin
21639 				 */
21640 				counter_u64_add(rack_tlp_retran, 1);
21641 				counter_u64_add(rack_tlp_retran_bytes, len);
21642 			} else {
21643 				tp->t_sndrexmitpack++;
21644 				KMOD_TCPSTAT_INC(tcps_sndrexmitpack);
21645 				KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len);
21646 			}
21647 #ifdef STATS
21648 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
21649 						 len);
21650 #endif
21651 		} else {
21652 			KMOD_TCPSTAT_INC(tcps_sndpack);
21653 			KMOD_TCPSTAT_ADD(tcps_sndbyte, len);
21654 #ifdef STATS
21655 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
21656 						 len);
21657 #endif
21658 		}
21659 		/*
21660 		 * If we're sending everything we've got, set PUSH. (This
21661 		 * will keep happy those implementations which only give
21662 		 * data to the user when a buffer fills or a PUSH comes in.)
21663 		 */
21664 		if (sb_offset + len == sbused(sb) &&
21665 		    sbused(sb) &&
21666 		    !(flags & TH_SYN)) {
21667 			flags |= TH_PUSH;
21668 			add_flag |= RACK_HAD_PUSH;
21669 		}
21670 
21671 		SOCKBUF_UNLOCK(sb);
21672 	} else {
21673 		SOCKBUF_UNLOCK(sb);
21674 		if (tp->t_flags & TF_ACKNOW)
21675 			KMOD_TCPSTAT_INC(tcps_sndacks);
21676 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
21677 			KMOD_TCPSTAT_INC(tcps_sndctrl);
21678 		else
21679 			KMOD_TCPSTAT_INC(tcps_sndwinup);
21680 
21681 		m = m_gethdr(M_NOWAIT, MT_DATA);
21682 		if (m == NULL) {
21683 			error = ENOBUFS;
21684 			sack_rxmit = 0;
21685 			goto out;
21686 		}
21687 #ifdef INET6
21688 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
21689 		    MHLEN >= hdrlen) {
21690 			M_ALIGN(m, hdrlen);
21691 		} else
21692 #endif
21693 			m->m_data += max_linkhdr;
21694 		m->m_len = hdrlen;
21695 	}
21696 	SOCKBUF_UNLOCK_ASSERT(sb);
21697 	m->m_pkthdr.rcvif = (struct ifnet *)0;
21698 #ifdef MAC
21699 	mac_inpcb_create_mbuf(inp, m);
21700 #endif
21701 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) &&  rack->r_fsb_inited) {
21702 #ifdef INET6
21703 		if (isipv6)
21704 			ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr;
21705 		else
21706 #endif				/* INET6 */
21707 #ifdef INET
21708 			ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
21709 #endif
21710 		th = rack->r_ctl.fsb.th;
21711 		udp = rack->r_ctl.fsb.udp;
21712 		if (udp) {
21713 #ifdef INET6
21714 			if (isipv6)
21715 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
21716 			else
21717 #endif				/* INET6 */
21718 				ulen = hdrlen + len - sizeof(struct ip);
21719 			udp->uh_ulen = htons(ulen);
21720 		}
21721 	} else {
21722 #ifdef INET6
21723 		if (isipv6) {
21724 			ip6 = mtod(m, struct ip6_hdr *);
21725 			if (tp->t_port) {
21726 				udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
21727 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21728 				udp->uh_dport = tp->t_port;
21729 				ulen = hdrlen + len - sizeof(struct ip6_hdr);
21730 				udp->uh_ulen = htons(ulen);
21731 				th = (struct tcphdr *)(udp + 1);
21732 			} else
21733 				th = (struct tcphdr *)(ip6 + 1);
21734 			tcpip_fillheaders(inp, tp->t_port, ip6, th);
21735 		} else
21736 #endif				/* INET6 */
21737 		{
21738 #ifdef INET
21739 			ip = mtod(m, struct ip *);
21740 			if (tp->t_port) {
21741 				udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
21742 				udp->uh_sport = htons(V_tcp_udp_tunneling_port);
21743 				udp->uh_dport = tp->t_port;
21744 				ulen = hdrlen + len - sizeof(struct ip);
21745 				udp->uh_ulen = htons(ulen);
21746 				th = (struct tcphdr *)(udp + 1);
21747 			} else
21748 				th = (struct tcphdr *)(ip + 1);
21749 			tcpip_fillheaders(inp, tp->t_port, ip, th);
21750 #endif
21751 		}
21752 	}
21753 	/*
21754 	 * Fill in fields, remembering maximum advertised window for use in
21755 	 * delaying messages about window sizes. If resending a FIN, be sure
21756 	 * not to use a new sequence number.
21757 	 */
21758 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
21759 	    tp->snd_nxt == tp->snd_max)
21760 		tp->snd_nxt--;
21761 	/*
21762 	 * If we are starting a connection, send ECN setup SYN packet. If we
21763 	 * are on a retransmit, we may resend those bits a number of times
21764 	 * as per RFC 3168.
21765 	 */
21766 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
21767 		flags |= tcp_ecn_output_syn_sent(tp);
21768 	}
21769 	/* Also handle parallel SYN for ECN */
21770 	if (TCPS_HAVERCVDSYN(tp->t_state) &&
21771 	    (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
21772 		int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
21773 		if ((tp->t_state == TCPS_SYN_RECEIVED) &&
21774 		    (tp->t_flags2 & TF2_ECN_SND_ECE))
21775 			tp->t_flags2 &= ~TF2_ECN_SND_ECE;
21776 #ifdef INET6
21777 		if (isipv6) {
21778 			ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20);
21779 			ip6->ip6_flow |= htonl(ect << 20);
21780 		}
21781 		else
21782 #endif
21783 		{
21784 #ifdef INET
21785 			ip->ip_tos &= ~IPTOS_ECN_MASK;
21786 			ip->ip_tos |= ect;
21787 #endif
21788 		}
21789 	}
21790 	/*
21791 	 * If we are doing retransmissions, then snd_nxt will not reflect
21792 	 * the first unsent octet.  For ACK only packets, we do not want the
21793 	 * sequence number of the retransmitted packet, we want the sequence
21794 	 * number of the next unsent octet.  So, if there is no data (and no
21795 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
21796 	 * ti_seq.  But if we are in persist state, snd_max might reflect
21797 	 * one byte beyond the right edge of the window, so use snd_nxt in
21798 	 * that case, since we know we aren't doing a retransmission.
21799 	 * (retransmit and persist are mutually exclusive...)
21800 	 */
21801 	if (sack_rxmit == 0) {
21802 		if (len || (flags & (TH_SYN | TH_FIN))) {
21803 			th->th_seq = htonl(tp->snd_nxt);
21804 			rack_seq = tp->snd_nxt;
21805 		} else {
21806 			th->th_seq = htonl(tp->snd_max);
21807 			rack_seq = tp->snd_max;
21808 		}
21809 	} else {
21810 		th->th_seq = htonl(rsm->r_start);
21811 		rack_seq = rsm->r_start;
21812 	}
21813 	th->th_ack = htonl(tp->rcv_nxt);
21814 	tcp_set_flags(th, flags);
21815 	/*
21816 	 * Calculate receive window.  Don't shrink window, but avoid silly
21817 	 * window syndrome.
21818 	 * If a RST segment is sent, advertise a window of zero.
21819 	 */
21820 	if (flags & TH_RST) {
21821 		recwin = 0;
21822 	} else {
21823 		if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
21824 		    recwin < (long)segsiz) {
21825 			recwin = 0;
21826 		}
21827 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
21828 		    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
21829 			recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
21830 	}
21831 
21832 	/*
21833 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
21834 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
21835 	 * handled in syncache.
21836 	 */
21837 	if (flags & TH_SYN)
21838 		th->th_win = htons((u_short)
21839 				   (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
21840 	else {
21841 		/* Avoid shrinking window with window scaling. */
21842 		recwin = roundup2(recwin, 1 << tp->rcv_scale);
21843 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
21844 	}
21845 	/*
21846 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
21847 	 * window.  This may cause the remote transmitter to stall.  This
21848 	 * flag tells soreceive() to disable delayed acknowledgements when
21849 	 * draining the buffer.  This can occur if the receiver is
21850 	 * attempting to read more data than can be buffered prior to
21851 	 * transmitting on the connection.
21852 	 */
21853 	if (th->th_win == 0) {
21854 		tp->t_sndzerowin++;
21855 		tp->t_flags |= TF_RXWIN0SENT;
21856 	} else
21857 		tp->t_flags &= ~TF_RXWIN0SENT;
21858 	tp->snd_up = tp->snd_una;	/* drag it along, its deprecated */
21859 	/* Now are we using fsb?, if so copy the template data to the mbuf */
21860 	if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) {
21861 		uint8_t *cpto;
21862 
21863 		cpto = mtod(m, uint8_t *);
21864 		memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len);
21865 		/*
21866 		 * We have just copied in:
21867 		 * IP/IP6
21868 		 * <optional udphdr>
21869 		 * tcphdr (no options)
21870 		 *
21871 		 * We need to grab the correct pointers into the mbuf
21872 		 * for both the tcp header, and possibly the udp header (if tunneling).
21873 		 * We do this by using the offset in the copy buffer and adding it
21874 		 * to the mbuf base pointer (cpto).
21875 		 */
21876 #ifdef INET6
21877 		if (isipv6)
21878 			ip6 = mtod(m, struct ip6_hdr *);
21879 		else
21880 #endif				/* INET6 */
21881 #ifdef INET
21882 			ip = mtod(m, struct ip *);
21883 #endif
21884 		th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr));
21885 		/* If we have a udp header lets set it into the mbuf as well */
21886 		if (udp)
21887 			udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr));
21888 	}
21889 	if (optlen) {
21890 		bcopy(opt, th + 1, optlen);
21891 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
21892 	}
21893 	/*
21894 	 * Put TCP length in extended header, and then checksum extended
21895 	 * header and data.
21896 	 */
21897 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
21898 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
21899 	if (to.to_flags & TOF_SIGNATURE) {
21900 		/*
21901 		 * Calculate MD5 signature and put it into the place
21902 		 * determined before.
21903 		 * NOTE: since TCP options buffer doesn't point into
21904 		 * mbuf's data, calculate offset and use it.
21905 		 */
21906 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
21907 						       (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
21908 			/*
21909 			 * Do not send segment if the calculation of MD5
21910 			 * digest has failed.
21911 			 */
21912 			goto out;
21913 		}
21914 	}
21915 #endif
21916 #ifdef INET6
21917 	if (isipv6) {
21918 		/*
21919 		 * ip6_plen is not need to be filled now, and will be filled
21920 		 * in ip6_output.
21921 		 */
21922 		if (tp->t_port) {
21923 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
21924 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21925 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
21926 			th->th_sum = htons(0);
21927 			UDPSTAT_INC(udps_opackets);
21928 		} else {
21929 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
21930 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21931 			th->th_sum = in6_cksum_pseudo(ip6,
21932 						      sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
21933 						      0);
21934 		}
21935 	}
21936 #endif
21937 #if defined(INET6) && defined(INET)
21938 	else
21939 #endif
21940 #ifdef INET
21941 	{
21942 		if (tp->t_port) {
21943 			m->m_pkthdr.csum_flags = CSUM_UDP;
21944 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
21945 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
21946 						ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
21947 			th->th_sum = htons(0);
21948 			UDPSTAT_INC(udps_opackets);
21949 		} else {
21950 			m->m_pkthdr.csum_flags = CSUM_TCP;
21951 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
21952 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
21953 					       ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
21954 									IPPROTO_TCP + len + optlen));
21955 		}
21956 		/* IP version must be set here for ipv4/ipv6 checking later */
21957 		KASSERT(ip->ip_v == IPVERSION,
21958 			("%s: IP version incorrect: %d", __func__, ip->ip_v));
21959 	}
21960 #endif
21961 	/*
21962 	 * Enable TSO and specify the size of the segments. The TCP pseudo
21963 	 * header checksum is always provided. XXX: Fixme: This is currently
21964 	 * not the case for IPv6.
21965 	 */
21966 	if (tso) {
21967 		/*
21968 		 * Here we must use t_maxseg and the optlen since
21969 		 * the optlen may include SACK's (or DSACK).
21970 		 */
21971 		KASSERT(len > tp->t_maxseg - optlen,
21972 			("%s: len <= tso_segsz", __func__));
21973 		m->m_pkthdr.csum_flags |= CSUM_TSO;
21974 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
21975 	}
21976 	KASSERT(len + hdrlen == m_length(m, NULL),
21977 		("%s: mbuf chain different than expected: %d + %u != %u",
21978 		 __func__, len, hdrlen, m_length(m, NULL)));
21979 
21980 #ifdef TCP_HHOOK
21981 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
21982 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
21983 #endif
21984 	if ((rack->r_ctl.crte != NULL) &&
21985 	    (rack->rc_hw_nobuf == 0) &&
21986 	    tcp_bblogging_on(tp)) {
21987 		rack_log_queue_level(tp, rack, len, &tv, cts);
21988 	}
21989 	/* We're getting ready to send; log now. */
21990 	if (tcp_bblogging_on(rack->rc_tp)) {
21991 		union tcp_log_stackspecific log;
21992 
21993 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
21994 		log.u_bbr.inhpts = tcp_in_hpts(rack->rc_tp);
21995 		if (rack->rack_no_prr)
21996 			log.u_bbr.flex1 = 0;
21997 		else
21998 			log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
21999 		log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs;
22000 		log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs;
22001 		log.u_bbr.flex4 = orig_len;
22002 		/* Save off the early/late values */
22003 		log.u_bbr.flex6 = rack->r_ctl.rc_agg_early;
22004 		log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed;
22005 		log.u_bbr.bw_inuse = rack_get_bw(rack);
22006 		log.u_bbr.cur_del_rate = rack->r_ctl.gp_bw;
22007 		log.u_bbr.flex8 = 0;
22008 		if (rsm) {
22009 			if (rsm->r_flags & RACK_RWND_COLLAPSED) {
22010 				rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm);
22011 				counter_u64_add(rack_collapsed_win_rxt, 1);
22012 				counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start));
22013 			}
22014 			if (doing_tlp)
22015 				log.u_bbr.flex8 = 2;
22016 			else
22017 				log.u_bbr.flex8 = 1;
22018 		} else {
22019 			if (doing_tlp)
22020 				log.u_bbr.flex8 = 3;
22021 		}
22022 		log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm);
22023 		log.u_bbr.flex7 = mark;
22024 		log.u_bbr.flex7 <<= 8;
22025 		log.u_bbr.flex7 |= pass;
22026 		log.u_bbr.pkts_out = tp->t_maxseg;
22027 		log.u_bbr.timeStamp = cts;
22028 		log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked);
22029 		if (rsm && (rsm->r_rtr_cnt > 0)) {
22030 			/*
22031 			 * When we have a retransmit we want to log the
22032 			 * burst at send and flight at send from before.
22033 			 */
22034 			log.u_bbr.flex5 = rsm->r_fas;
22035 			log.u_bbr.bbr_substate = rsm->r_bas;
22036 		} else {
22037 			/*
22038 			 * New transmits we log in flex5 the inflight again as
22039 			 * well as the number of segments in our send in the
22040 			 * substate field.
22041 			 */
22042 			log.u_bbr.flex5 = log.u_bbr.inflight;
22043 			log.u_bbr.bbr_substate = (uint8_t)((len + segsiz - 1)/segsiz);
22044 		}
22045 		log.u_bbr.lt_epoch = cwnd_to_use;
22046 		log.u_bbr.delivered = sendalot;
22047 		log.u_bbr.rttProp = (uint64_t)rsm;
22048 		log.u_bbr.pkt_epoch = __LINE__;
22049 		if (rsm) {
22050 			log.u_bbr.delRate = rsm->r_flags;
22051 			log.u_bbr.delRate <<= 31;
22052 			log.u_bbr.delRate |= rack->r_must_retran;
22053 			log.u_bbr.delRate <<= 1;
22054 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
22055 		} else {
22056 			log.u_bbr.delRate = rack->r_must_retran;
22057 			log.u_bbr.delRate <<= 1;
22058 			log.u_bbr.delRate |= (sack_rxmit & 0x00000001);
22059 		}
22060 		lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
22061 				    len, &log, false, NULL, __func__, __LINE__, &tv);
22062 	} else
22063 		lgb = NULL;
22064 
22065 	/*
22066 	 * Fill in IP length and desired time to live and send to IP level.
22067 	 * There should be a better way to handle ttl and tos; we could keep
22068 	 * them in the template, but need a way to checksum without them.
22069 	 */
22070 	/*
22071 	 * m->m_pkthdr.len should have been set before cksum calcuration,
22072 	 * because in6_cksum() need it.
22073 	 */
22074 #ifdef INET6
22075 	if (isipv6) {
22076 		/*
22077 		 * we separately set hoplimit for every segment, since the
22078 		 * user might want to change the value via setsockopt. Also,
22079 		 * desired default hop limit might be changed via Neighbor
22080 		 * Discovery.
22081 		 */
22082 		rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL);
22083 
22084 		/*
22085 		 * Set the packet size here for the benefit of DTrace
22086 		 * probes. ip6_output() will set it properly; it's supposed
22087 		 * to include the option header lengths as well.
22088 		 */
22089 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
22090 
22091 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
22092 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22093 		else
22094 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22095 
22096 		if (tp->t_state == TCPS_SYN_SENT)
22097 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
22098 
22099 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
22100 		/* TODO: IPv6 IP6TOS_ECT bit on */
22101 		error = ip6_output(m,
22102 				   inp->in6p_outputopts,
22103 				   &inp->inp_route6,
22104 				   ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0),
22105 				   NULL, NULL, inp);
22106 
22107 		if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
22108 			mtu = inp->inp_route6.ro_nh->nh_mtu;
22109 	}
22110 #endif				/* INET6 */
22111 #if defined(INET) && defined(INET6)
22112 	else
22113 #endif
22114 #ifdef INET
22115 	{
22116 		ip->ip_len = htons(m->m_pkthdr.len);
22117 #ifdef INET6
22118 		if (inp->inp_vflag & INP_IPV6PROTO)
22119 			ip->ip_ttl = in6_selecthlim(inp, NULL);
22120 #endif				/* INET6 */
22121 		rack->r_ctl.fsb.hoplimit = ip->ip_ttl;
22122 		/*
22123 		 * If we do path MTU discovery, then we set DF on every
22124 		 * packet. This might not be the best thing to do according
22125 		 * to RFC3390 Section 2. However the tcp hostcache migitates
22126 		 * the problem so it affects only the first tcp connection
22127 		 * with a host.
22128 		 *
22129 		 * NB: Don't set DF on small MTU/MSS to have a safe
22130 		 * fallback.
22131 		 */
22132 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
22133 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
22134 			if (tp->t_port == 0 || len < V_tcp_minmss) {
22135 				ip->ip_off |= htons(IP_DF);
22136 			}
22137 		} else {
22138 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
22139 		}
22140 
22141 		if (tp->t_state == TCPS_SYN_SENT)
22142 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
22143 
22144 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
22145 
22146 		error = ip_output(m,
22147 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
22148 				  inp->inp_options,
22149 #else
22150 				  NULL,
22151 #endif
22152 				  &inp->inp_route,
22153 				  ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0,
22154 				  inp);
22155 		if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
22156 			mtu = inp->inp_route.ro_nh->nh_mtu;
22157 	}
22158 #endif				/* INET */
22159 
22160 out:
22161 	if (lgb) {
22162 		lgb->tlb_errno = error;
22163 		lgb = NULL;
22164 	}
22165 	/*
22166 	 * In transmit state, time the transmission and arrange for the
22167 	 * retransmit.  In persist state, just set snd_max.
22168 	 */
22169 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error,
22170 			rack_to_usec_ts(&tv),
22171 			rsm, add_flag, s_mb, s_moff, hw_tls, segsiz);
22172 	if (error == 0) {
22173 		if (rsm == NULL) {
22174 			if (rack->lt_bw_up == 0) {
22175 				rack->r_ctl.lt_timemark = tcp_tv_to_lusectick(&tv);
22176 				rack->r_ctl.lt_seq = tp->snd_una;
22177 				rack->lt_bw_up = 1;
22178 			} else if (((rack_seq + len) - rack->r_ctl.lt_seq) > 0x7fffffff) {
22179 				/*
22180 				 * Need to record what we have since we are
22181 				 * approaching seq wrap.
22182 				 */
22183 				uint64_t tmark;
22184 
22185 				rack->r_ctl.lt_bw_bytes += (tp->snd_una - rack->r_ctl.lt_seq);
22186 				rack->r_ctl.lt_seq = tp->snd_una;
22187 				tmark = tcp_tv_to_lusectick(&tv);
22188 				rack->r_ctl.lt_bw_time += (tmark - rack->r_ctl.lt_timemark);
22189 				rack->r_ctl.lt_timemark = tmark;
22190 			}
22191 		}
22192 		rack->forced_ack = 0;	/* If we send something zap the FA flag */
22193 		counter_u64_add(rack_total_bytes, len);
22194 		tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls);
22195 		if (rsm && doing_tlp) {
22196 			rack->rc_last_sent_tlp_past_cumack = 0;
22197 			rack->rc_last_sent_tlp_seq_valid = 1;
22198 			rack->r_ctl.last_sent_tlp_seq = rsm->r_start;
22199 			rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start;
22200 		}
22201 		if (rack->rc_hw_nobuf) {
22202 			rack->rc_hw_nobuf = 0;
22203 			rack->r_ctl.rc_agg_delayed = 0;
22204 			rack->r_early = 0;
22205 			rack->r_late = 0;
22206 			rack->r_ctl.rc_agg_early = 0;
22207 		}
22208 		if (rsm && (doing_tlp == 0)) {
22209 			/* Set we retransmitted */
22210 			rack->rc_gp_saw_rec = 1;
22211 		} else {
22212 			if (cwnd_to_use > tp->snd_ssthresh) {
22213 				/* Set we sent in CA */
22214 				rack->rc_gp_saw_ca = 1;
22215 			} else {
22216 				/* Set we sent in SS */
22217 				rack->rc_gp_saw_ss = 1;
22218 			}
22219 		}
22220 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
22221 		    (tp->t_flags & TF_SACK_PERMIT) &&
22222 		    tp->rcv_numsacks > 0)
22223 			tcp_clean_dsack_blocks(tp);
22224 		tot_len_this_send += len;
22225 		if (len == 0) {
22226 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
22227 		} else {
22228 			int idx;
22229 
22230 			idx = (len / segsiz) + 3;
22231 			if (idx >= TCP_MSS_ACCT_ATIMER)
22232 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
22233 			else
22234 				counter_u64_add(rack_out_size[idx], 1);
22235 		}
22236 	}
22237 	if ((rack->rack_no_prr == 0) &&
22238 	    sub_from_prr &&
22239 	    (error == 0)) {
22240 		if (rack->r_ctl.rc_prr_sndcnt >= len)
22241 			rack->r_ctl.rc_prr_sndcnt -= len;
22242 		else
22243 			rack->r_ctl.rc_prr_sndcnt = 0;
22244 	}
22245 	sub_from_prr = 0;
22246 	if (doing_tlp) {
22247 		/* Make sure the TLP is added */
22248 		add_flag |= RACK_TLP;
22249 	} else if (rsm) {
22250 		/* If its a resend without TLP then it must not have the flag */
22251 		rsm->r_flags &= ~RACK_TLP;
22252 	}
22253 
22254 
22255 	if ((error == 0) &&
22256 	    (len > 0) &&
22257 	    (tp->snd_una == tp->snd_max))
22258 		rack->r_ctl.rc_tlp_rxt_last_time = cts;
22259 	{
22260 		tcp_seq startseq = tp->snd_nxt;
22261 
22262 		/* Track our lost count */
22263 		if (rsm && (doing_tlp == 0))
22264 			rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start;
22265 		/*
22266 		 * Advance snd_nxt over sequence space of this segment.
22267 		 */
22268 		if (error)
22269 			/* We don't log or do anything with errors */
22270 			goto nomore;
22271 		if (doing_tlp == 0) {
22272 			if (rsm == NULL) {
22273 				/*
22274 				 * Not a retransmission of some
22275 				 * sort, new data is going out so
22276 				 * clear our TLP count and flag.
22277 				 */
22278 				rack->rc_tlp_in_progress = 0;
22279 				rack->r_ctl.rc_tlp_cnt_out = 0;
22280 			}
22281 		} else {
22282 			/*
22283 			 * We have just sent a TLP, mark that it is true
22284 			 * and make sure our in progress is set so we
22285 			 * continue to check the count.
22286 			 */
22287 			rack->rc_tlp_in_progress = 1;
22288 			rack->r_ctl.rc_tlp_cnt_out++;
22289 		}
22290 		if (flags & (TH_SYN | TH_FIN)) {
22291 			if (flags & TH_SYN)
22292 				tp->snd_nxt++;
22293 			if (flags & TH_FIN) {
22294 				tp->snd_nxt++;
22295 				tp->t_flags |= TF_SENTFIN;
22296 			}
22297 		}
22298 		/* In the ENOBUFS case we do *not* update snd_max */
22299 		if (sack_rxmit)
22300 			goto nomore;
22301 
22302 		tp->snd_nxt += len;
22303 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
22304 			if (tp->snd_una == tp->snd_max) {
22305 				/*
22306 				 * Update the time we just added data since
22307 				 * none was outstanding.
22308 				 */
22309 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
22310 				tp->t_acktime = ticks;
22311 			}
22312 			tp->snd_max = tp->snd_nxt;
22313 			if (rack->rc_new_rnd_needed) {
22314 				/*
22315 				 * Update the rnd to start ticking not
22316 				 * that from a time perspective all of
22317 				 * the preceding idle time is "in the round"
22318 				 */
22319 				rack->rc_new_rnd_needed = 0;
22320 				rack->r_ctl.roundends = tp->snd_max;
22321 			}
22322 			/*
22323 			 * Time this transmission if not a retransmission and
22324 			 * not currently timing anything.
22325 			 * This is only relevant in case of switching back to
22326 			 * the base stack.
22327 			 */
22328 			if (tp->t_rtttime == 0) {
22329 				tp->t_rtttime = ticks;
22330 				tp->t_rtseq = startseq;
22331 				KMOD_TCPSTAT_INC(tcps_segstimed);
22332 			}
22333 			if (len &&
22334 			    ((tp->t_flags & TF_GPUTINPROG) == 0))
22335 				rack_start_gp_measurement(tp, rack, startseq, sb_offset);
22336 		}
22337 		/*
22338 		 * If we are doing FO we need to update the mbuf position and subtract
22339 		 * this happens when the peer sends us duplicate information and
22340 		 * we thus want to send a DSACK.
22341 		 *
22342 		 * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO
22343 		 * turned off? If not then we are going to echo multiple DSACK blocks
22344 		 * out (with the TSO), which we should not be doing.
22345 		 */
22346 		if (rack->r_fast_output && len) {
22347 			if (rack->r_ctl.fsb.left_to_send > len)
22348 				rack->r_ctl.fsb.left_to_send -= len;
22349 			else
22350 				rack->r_ctl.fsb.left_to_send = 0;
22351 			if (rack->r_ctl.fsb.left_to_send < segsiz)
22352 				rack->r_fast_output = 0;
22353 			if (rack->r_fast_output) {
22354 				rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off);
22355 				rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len;
22356 				rack->r_ctl.fsb.o_t_len = M_TRAILINGROOM(rack->r_ctl.fsb.m);
22357 			}
22358 		}
22359 	}
22360 nomore:
22361 	if (error) {
22362 		rack->r_ctl.rc_agg_delayed = 0;
22363 		rack->r_early = 0;
22364 		rack->r_late = 0;
22365 		rack->r_ctl.rc_agg_early = 0;
22366 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
22367 		/*
22368 		 * Failures do not advance the seq counter above. For the
22369 		 * case of ENOBUFS we will fall out and retry in 1ms with
22370 		 * the hpts. Everything else will just have to retransmit
22371 		 * with the timer.
22372 		 *
22373 		 * In any case, we do not want to loop around for another
22374 		 * send without a good reason.
22375 		 */
22376 		sendalot = 0;
22377 		switch (error) {
22378 		case EPERM:
22379 		case EACCES:
22380 			tp->t_softerror = error;
22381 #ifdef TCP_ACCOUNTING
22382 			crtsc = get_cyclecount();
22383 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22384 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22385 			}
22386 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22387 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22388 			}
22389 			sched_unpin();
22390 #endif
22391 			return (error);
22392 		case ENOBUFS:
22393 			/*
22394 			 * Pace us right away to retry in a some
22395 			 * time
22396 			 */
22397 			if (rack->r_ctl.crte != NULL) {
22398 				tcp_trace_point(rack->rc_tp, TCP_TP_HWENOBUF);
22399 				if (tcp_bblogging_on(rack->rc_tp))
22400 					rack_log_queue_level(tp, rack, len, &tv, cts);
22401 			} else
22402 				tcp_trace_point(rack->rc_tp, TCP_TP_ENOBUF);
22403 			slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC);
22404 			if (rack->rc_enobuf < 0x7f)
22405 				rack->rc_enobuf++;
22406 			if (slot < (10 * HPTS_USEC_IN_MSEC))
22407 				slot = 10 * HPTS_USEC_IN_MSEC;
22408 			if (rack->r_ctl.crte != NULL) {
22409 				counter_u64_add(rack_saw_enobuf_hw, 1);
22410 				tcp_rl_log_enobuf(rack->r_ctl.crte);
22411 			}
22412 			counter_u64_add(rack_saw_enobuf, 1);
22413 			goto enobufs;
22414 		case EMSGSIZE:
22415 			/*
22416 			 * For some reason the interface we used initially
22417 			 * to send segments changed to another or lowered
22418 			 * its MTU. If TSO was active we either got an
22419 			 * interface without TSO capabilits or TSO was
22420 			 * turned off. If we obtained mtu from ip_output()
22421 			 * then update it and try again.
22422 			 */
22423 			if (tso)
22424 				tp->t_flags &= ~TF_TSO;
22425 			if (mtu != 0) {
22426 				int saved_mtu;
22427 
22428 				saved_mtu = tp->t_maxseg;
22429 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
22430 				if (saved_mtu > tp->t_maxseg) {
22431 					goto again;
22432 				}
22433 			}
22434 			slot = 10 * HPTS_USEC_IN_MSEC;
22435 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22436 #ifdef TCP_ACCOUNTING
22437 			crtsc = get_cyclecount();
22438 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22439 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22440 			}
22441 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22442 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22443 			}
22444 			sched_unpin();
22445 #endif
22446 			return (error);
22447 		case ENETUNREACH:
22448 			counter_u64_add(rack_saw_enetunreach, 1);
22449 		case EHOSTDOWN:
22450 		case EHOSTUNREACH:
22451 		case ENETDOWN:
22452 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
22453 				tp->t_softerror = error;
22454 			}
22455 			/* FALLTHROUGH */
22456 		default:
22457 			slot = 10 * HPTS_USEC_IN_MSEC;
22458 			rack_start_hpts_timer(rack, tp, cts, slot, 0, 0);
22459 #ifdef TCP_ACCOUNTING
22460 			crtsc = get_cyclecount();
22461 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22462 				tp->tcp_cnt_counters[SND_OUT_FAIL]++;
22463 			}
22464 			if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22465 				tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val);
22466 			}
22467 			sched_unpin();
22468 #endif
22469 			return (error);
22470 		}
22471 	} else {
22472 		rack->rc_enobuf = 0;
22473 		if (IN_FASTRECOVERY(tp->t_flags) && rsm)
22474 			rack->r_ctl.retran_during_recovery += len;
22475 	}
22476 	KMOD_TCPSTAT_INC(tcps_sndtotal);
22477 
22478 	/*
22479 	 * Data sent (as far as we can tell). If this advertises a larger
22480 	 * window than any other segment, then remember the size of the
22481 	 * advertised window. Any pending ACK has now been sent.
22482 	 */
22483 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
22484 		tp->rcv_adv = tp->rcv_nxt + recwin;
22485 
22486 	tp->last_ack_sent = tp->rcv_nxt;
22487 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
22488 enobufs:
22489 	if (sendalot) {
22490 		/* Do we need to turn off sendalot? */
22491 		if (rack->r_ctl.rc_pace_max_segs &&
22492 		    (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) {
22493 			/* We hit our max. */
22494 			sendalot = 0;
22495 		} else if ((rack->rc_user_set_max_segs) &&
22496 			   (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) {
22497 			/* We hit the user defined max */
22498 			sendalot = 0;
22499 		}
22500 	}
22501 	if ((error == 0) && (flags & TH_FIN))
22502 		tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN);
22503 	if (flags & TH_RST) {
22504 		/*
22505 		 * We don't send again after sending a RST.
22506 		 */
22507 		slot = 0;
22508 		sendalot = 0;
22509 		if (error == 0)
22510 			tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST);
22511 	} else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) {
22512 		/*
22513 		 * Get our pacing rate, if an error
22514 		 * occurred in sending (ENOBUF) we would
22515 		 * hit the else if with slot preset. Other
22516 		 * errors return.
22517 		 */
22518 		slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz);
22519 	}
22520 	if (rsm &&
22521 	    (rsm->r_flags & RACK_HAS_SYN) == 0 &&
22522 	    rack->use_rack_rr) {
22523 		/* Its a retransmit and we use the rack cheat? */
22524 		if ((slot == 0) ||
22525 		    (rack->rc_always_pace == 0) ||
22526 		    (rack->r_rr_config == 1)) {
22527 			/*
22528 			 * We have no pacing set or we
22529 			 * are using old-style rack or
22530 			 * we are overridden to use the old 1ms pacing.
22531 			 */
22532 			slot = rack->r_ctl.rc_min_to;
22533 		}
22534 	}
22535 	/* We have sent clear the flag */
22536 	rack->r_ent_rec_ns = 0;
22537 	if (rack->r_must_retran) {
22538 		if (rsm) {
22539 			rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start);
22540 			if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) {
22541 				/*
22542 				 * We have retransmitted all.
22543 				 */
22544 				rack->r_must_retran = 0;
22545 				rack->r_ctl.rc_out_at_rto = 0;
22546 			}
22547 		} else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22548 			/*
22549 			 * Sending new data will also kill
22550 			 * the loop.
22551 			 */
22552 			rack->r_must_retran = 0;
22553 			rack->r_ctl.rc_out_at_rto = 0;
22554 		}
22555 	}
22556 	rack->r_ctl.fsb.recwin = recwin;
22557 	if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) &&
22558 	    SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) {
22559 		/*
22560 		 * We hit an RTO and now have past snd_max at the RTO
22561 		 * clear all the WAS flags.
22562 		 */
22563 		tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY);
22564 	}
22565 	if (slot) {
22566 		/* set the rack tcb into the slot N */
22567 		if ((error == 0) &&
22568 		    rack_use_rfo &&
22569 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22570 		    (rsm == NULL) &&
22571 		    (tp->snd_nxt == tp->snd_max) &&
22572 		    (ipoptlen == 0) &&
22573 		    (tp->rcv_numsacks == 0) &&
22574 		    rack->r_fsb_inited &&
22575 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
22576 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
22577 		    (rack->r_must_retran == 0) &&
22578 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
22579 		    (len > 0) && (orig_len > 0) &&
22580 		    (orig_len > len) &&
22581 		    ((orig_len - len) >= segsiz) &&
22582 		    ((optlen == 0) ||
22583 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22584 			/* We can send at least one more MSS using our fsb */
22585 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
22586 					       segsiz, pace_max_seg, hw_tls, flags);
22587 		} else
22588 			rack->r_fast_output = 0;
22589 		rack_log_fsb(rack, tp, so, flags,
22590 			     ipoptlen, orig_len, len, error,
22591 			     (rsm == NULL), optlen, __LINE__, 2);
22592 	} else if (sendalot) {
22593 		int ret;
22594 
22595 		sack_rxmit = 0;
22596 		if ((error == 0) &&
22597 		    rack_use_rfo &&
22598 		    ((flags & (TH_SYN|TH_FIN)) == 0) &&
22599 		    (rsm == NULL) &&
22600 		    (ipoptlen == 0) &&
22601 		    (tp->rcv_numsacks == 0) &&
22602 		    (tp->snd_nxt == tp->snd_max) &&
22603 		    (rack->r_must_retran == 0) &&
22604 		    rack->r_fsb_inited &&
22605 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
22606 		    ((IN_RECOVERY(tp->t_flags)) == 0) &&
22607 		    ((tp->t_flags & TF_NEEDFIN) == 0) &&
22608 		    (len > 0) && (orig_len > 0) &&
22609 		    (orig_len > len) &&
22610 		    ((orig_len - len) >= segsiz) &&
22611 		    ((optlen == 0) ||
22612 		     ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) {
22613 			/* we can use fast_output for more */
22614 			rack_setup_fast_output(tp, rack, sb, len, orig_len,
22615 					       segsiz, pace_max_seg, hw_tls, flags);
22616 			if (rack->r_fast_output) {
22617 				error = 0;
22618 				ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error);
22619 				if (ret >= 0)
22620 					return (ret);
22621 			        else if (error)
22622 					goto nomore;
22623 
22624 			}
22625 		}
22626 		goto again;
22627 	}
22628 	/* Assure when we leave that snd_nxt will point to top */
22629 skip_all_send:
22630 	if (SEQ_GT(tp->snd_max, tp->snd_nxt))
22631 		tp->snd_nxt = tp->snd_max;
22632 	rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0);
22633 #ifdef TCP_ACCOUNTING
22634 	crtsc = get_cyclecount() - ts_val;
22635 	if (tot_len_this_send) {
22636 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22637 			tp->tcp_cnt_counters[SND_OUT_DATA]++;
22638 		}
22639 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22640 			tp->tcp_proc_time[SND_OUT_DATA] += crtsc;
22641 		}
22642 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22643 			tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz);
22644 		}
22645 	} else {
22646 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22647 			tp->tcp_cnt_counters[SND_OUT_ACK]++;
22648 		}
22649 		if (tp->t_flags2 & TF2_TCP_ACCOUNTING) {
22650 			tp->tcp_proc_time[SND_OUT_ACK] += crtsc;
22651 		}
22652 	}
22653 	sched_unpin();
22654 #endif
22655 	if (error == ENOBUFS)
22656 		error = 0;
22657 	return (error);
22658 }
22659 
22660 static void
22661 rack_update_seg(struct tcp_rack *rack)
22662 {
22663 	uint32_t orig_val;
22664 
22665 	orig_val = rack->r_ctl.rc_pace_max_segs;
22666 	rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL);
22667 	if (orig_val != rack->r_ctl.rc_pace_max_segs)
22668 		rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0);
22669 }
22670 
22671 static void
22672 rack_mtu_change(struct tcpcb *tp)
22673 {
22674 	/*
22675 	 * The MSS may have changed
22676 	 */
22677 	struct tcp_rack *rack;
22678 	struct rack_sendmap *rsm;
22679 
22680 	rack = (struct tcp_rack *)tp->t_fb_ptr;
22681 	if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) {
22682 		/*
22683 		 * The MTU has changed we need to resend everything
22684 		 * since all we have sent is lost. We first fix
22685 		 * up the mtu though.
22686 		 */
22687 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
22688 		/* We treat this like a full retransmit timeout without the cwnd adjustment */
22689 		rack_remxt_tmr(tp);
22690 		rack->r_fast_output = 0;
22691 		rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp,
22692 						rack->r_ctl.rc_sacked);
22693 		rack->r_ctl.rc_snd_max_at_rto = tp->snd_max;
22694 		rack->r_must_retran = 1;
22695 		/* Mark all inflight to needing to be rxt'd */
22696 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
22697 			rsm->r_flags |= (RACK_MUST_RXT|RACK_PMTU_CHG);
22698 		}
22699 	}
22700 	sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
22701 	/* We don't use snd_nxt to retransmit */
22702 	tp->snd_nxt = tp->snd_max;
22703 }
22704 
22705 static int
22706 rack_set_dgp(struct tcp_rack *rack)
22707 {
22708 	/* pace_always=1 */
22709 	if (rack->rc_always_pace == 0) {
22710 		if (tcp_can_enable_pacing() == 0)
22711 			return (EBUSY);
22712 	}
22713 	rack->rc_fillcw_apply_discount = 0;
22714 	rack->dgp_on = 1;
22715 	rack->rc_always_pace = 1;
22716 	rack->use_fixed_rate = 0;
22717 	if (rack->gp_ready)
22718 		rack_set_cc_pacing(rack);
22719 	rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22720 	rack->rack_attempt_hdwr_pace = 0;
22721 	/* rxt settings */
22722 	rack->full_size_rxt = 1;
22723 	rack->shape_rxt_to_pacing_min  = 0;
22724 	/* cmpack=1 */
22725 	rack->r_use_cmp_ack = 1;
22726 	if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) &&
22727 	    rack->r_use_cmp_ack)
22728 		rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22729 	/* scwnd=1 */
22730 	rack->rack_enable_scwnd = 1;
22731 	/* dynamic=100 */
22732 	rack->rc_gp_dyn_mul = 1;
22733 	/* gp_inc_ca */
22734 	rack->r_ctl.rack_per_of_gp_ca = 100;
22735 	/* rrr_conf=3 */
22736 	rack->r_rr_config = 3;
22737 	/* npush=2 */
22738 	rack->r_ctl.rc_no_push_at_mrtt = 2;
22739 	/* fillcw=1 */
22740 	if (rack->r_cwnd_was_clamped == 0) {
22741 		rack->rc_pace_to_cwnd = 1;
22742 	} else {
22743 		rack->rc_pace_to_cwnd = 0;
22744 		/* Reset all multipliers to 100.0 so just the measured bw */
22745 		rack->r_ctl.rack_per_of_gp_ss = 100;
22746 		rack->r_ctl.rack_per_of_gp_ca = 100;
22747 	}
22748 	rack->rc_pace_fill_if_rttin_range = 0;
22749 	rack->rtt_limit_mul = 0;
22750 	/* noprr=1 */
22751 	rack->rack_no_prr = 1;
22752 	/* lscwnd=1 */
22753 	rack->r_limit_scw = 1;
22754 	/* gp_inc_rec */
22755 	rack->r_ctl.rack_per_of_gp_rec = 90;
22756 	rack_client_buffer_level_set(rack);
22757 	return (0);
22758 }
22759 
22760 
22761 
22762 static int
22763 rack_set_profile(struct tcp_rack *rack, int prof)
22764 {
22765 	int err = EINVAL;
22766 	if (prof == 1) {
22767 		/*
22768 		 * Profile 1 is "standard" DGP. It ignores
22769 		 * client buffer level.
22770 		 */
22771 		rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL0;
22772 		err = rack_set_dgp(rack);
22773 		if (err)
22774 			return (err);
22775 	} else if (prof == 2) {
22776 		/*
22777 		 * Profile 2 is DGP. Less aggressive with
22778 		 * respect to client buffer level.
22779 		 */
22780 		rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL1;
22781 		err = rack_set_dgp(rack);
22782 		if (err)
22783 			return (err);
22784 	} else if (prof == 3) {
22785 		/*
22786 		 * Profile 3 is DGP. Even Less aggressive with
22787 		 * respect to client buffer level.
22788 		 */
22789 		rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL2;
22790 		err = rack_set_dgp(rack);
22791 		if (err)
22792 			return (err);
22793 	} else if (prof == 4) {
22794 		/*
22795 		 * Profile 4 is DGP with the most responsiveness
22796 		 * to client buffer level.
22797 		 */
22798 		rack->r_ctl.rc_dgp_bl_agg = DGP_LEVEL3;
22799 		err = rack_set_dgp(rack);
22800 		if (err)
22801 			return (err);
22802 	} else if (prof == 5) {
22803 		err = rack_set_dgp(rack);
22804 		if (err)
22805 			return (err);
22806 		/*
22807 		 * By turning DGP off we change the rate
22808 		 * picked to be only the one the cwnd and rtt
22809 		 * get us.
22810 		 */
22811 		rack->dgp_on = 0;
22812 	} else if (prof == 6) {
22813 		err = rack_set_dgp(rack);
22814 		if (err)
22815 			return (err);
22816 		/*
22817 		 * Profile 6 tweaks DGP so that it will apply to
22818 		 * fill-cw the same settings that profile5 does
22819 		 * to replace DGP. It gets then the max(dgp-rate, fillcw(discounted).
22820 		 */
22821 		rack->rc_fillcw_apply_discount = 1;
22822 	} else if (prof == 0) {
22823 		/* This changes things back to the default settings */
22824 		rack->dgp_on = 0;
22825 		rack->rc_hybrid_mode = 0;
22826 		err = 0;
22827 		if (rack_fill_cw_state)
22828 			rack->rc_pace_to_cwnd = 1;
22829 		else
22830 			rack->rc_pace_to_cwnd = 0;
22831 		if (rack->rc_always_pace) {
22832 			tcp_decrement_paced_conn();
22833 			rack_undo_cc_pacing(rack);
22834 			rack->rc_always_pace = 0;
22835 		}
22836 		if (rack_pace_every_seg && tcp_can_enable_pacing()) {
22837 			rack->rc_always_pace = 1;
22838 			if (rack->rack_hibeta)
22839 				rack_set_cc_pacing(rack);
22840 		} else
22841 			rack->rc_always_pace = 0;
22842 		if (rack_dsack_std_based & 0x1) {
22843 			/* Basically this means all rack timers are at least (srtt + 1/4 srtt) */
22844 			rack->rc_rack_tmr_std_based = 1;
22845 		}
22846 		if (rack_dsack_std_based & 0x2) {
22847 			/* Basically this means  rack timers are extended based on dsack by up to (2 * srtt) */
22848 			rack->rc_rack_use_dsack = 1;
22849 		}
22850 		if (rack_use_cmp_acks)
22851 			rack->r_use_cmp_ack = 1;
22852 		else
22853 			rack->r_use_cmp_ack = 0;
22854 		if (rack_disable_prr)
22855 			rack->rack_no_prr = 1;
22856 		else
22857 			rack->rack_no_prr = 0;
22858 		if (rack_gp_no_rec_chg)
22859 			rack->rc_gp_no_rec_chg = 1;
22860 		else
22861 			rack->rc_gp_no_rec_chg = 0;
22862 		if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) {
22863 			rack->r_mbuf_queue = 1;
22864 			if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state))
22865 				rack->rc_tp->t_flags2 |= TF2_MBUF_ACKCMP;
22866 			rack->rc_tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
22867 		} else {
22868 			rack->r_mbuf_queue = 0;
22869 			rack->rc_tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
22870 		}
22871 		if (rack_enable_shared_cwnd)
22872 			rack->rack_enable_scwnd = 1;
22873 		else
22874 			rack->rack_enable_scwnd = 0;
22875 		if (rack_do_dyn_mul) {
22876 			/* When dynamic adjustment is on CA needs to start at 100% */
22877 			rack->rc_gp_dyn_mul = 1;
22878 			if (rack_do_dyn_mul >= 100)
22879 				rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul;
22880 		} else {
22881 			rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca;
22882 			rack->rc_gp_dyn_mul = 0;
22883 		}
22884 		rack->r_rr_config = 0;
22885 		rack->r_ctl.rc_no_push_at_mrtt = 0;
22886 		rack->rc_pace_to_cwnd = 0;
22887 		rack->rc_pace_fill_if_rttin_range = 0;
22888 		rack->rtt_limit_mul = 0;
22889 
22890 		if (rack_enable_hw_pacing)
22891 			rack->rack_hdw_pace_ena = 1;
22892 		else
22893 			rack->rack_hdw_pace_ena = 0;
22894 		if (rack_disable_prr)
22895 			rack->rack_no_prr = 1;
22896 		else
22897 			rack->rack_no_prr = 0;
22898 		if (rack_limits_scwnd)
22899 			rack->r_limit_scw  = 1;
22900 		else
22901 			rack->r_limit_scw  = 0;
22902 		rack_init_retransmit_value(rack, rack_rxt_controls);
22903 		err = 0;
22904 	}
22905 	return (err);
22906 }
22907 
22908 static int
22909 rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval)
22910 {
22911 	struct deferred_opt_list *dol;
22912 
22913 	dol = malloc(sizeof(struct deferred_opt_list),
22914 		     M_TCPFSB, M_NOWAIT|M_ZERO);
22915 	if (dol == NULL) {
22916 		/*
22917 		 * No space yikes -- fail out..
22918 		 */
22919 		return (0);
22920 	}
22921 	dol->optname = sopt_name;
22922 	dol->optval = loptval;
22923 	TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next);
22924 	return (1);
22925 }
22926 
22927 static int
22928 process_hybrid_pacing(struct tcp_rack *rack, struct tcp_hybrid_req *hybrid)
22929 {
22930 #ifdef TCP_REQUEST_TRK
22931 	struct tcp_sendfile_track *sft;
22932 	struct timeval tv;
22933 	tcp_seq seq;
22934 	int err;
22935 
22936 	microuptime(&tv);
22937 
22938 	/*
22939 	 * If BB logging is not on we need to look at the DTL flag.
22940 	 * If its on already then those reasons override the DTL input.
22941 	 * We do this with any request, you can turn DTL on, but it does
22942 	 * not turn off at least from hybrid pacing requests.
22943 	 */
22944 	if (tcp_bblogging_on(rack->rc_tp) == 0) {
22945 		if (hybrid->hybrid_flags & TCP_HYBRID_PACING_DTL) {
22946 			/* Turn on BB point logging  */
22947 			tcp_set_bblog_state(rack->rc_tp, TCP_LOG_VIA_BBPOINTS,
22948 					    TCP_BBPOINT_REQ_LEVEL_LOGGING);
22949 		}
22950 	}
22951 	/* Make sure no fixed rate is on */
22952 	rack->use_fixed_rate = 0;
22953 	rack->r_ctl.rc_fixed_pacing_rate_rec = 0;
22954 	rack->r_ctl.rc_fixed_pacing_rate_ca = 0;
22955 	rack->r_ctl.rc_fixed_pacing_rate_ss = 0;
22956 	/* Now allocate or find our entry that will have these settings */
22957 	sft = tcp_req_alloc_req_full(rack->rc_tp, &hybrid->req, tcp_tv_to_lusectick(&tv), 0);
22958 	if (sft == NULL) {
22959 		rack->rc_tp->tcp_hybrid_error++;
22960 		/* no space, where would it have gone? */
22961 		seq = rack->rc_tp->snd_una + rack->rc_tp->t_inpcb.inp_socket->so_snd.sb_ccc;
22962 		rack_log_hybrid(rack, seq, NULL, HYBRID_LOG_NO_ROOM, __LINE__, 0);
22963 		return (ENOSPC);
22964 	}
22965 	/* The seq will be snd_una + everything in the buffer */
22966 	seq = sft->start_seq;
22967 	if ((hybrid->hybrid_flags & TCP_HYBRID_PACING_ENABLE) == 0) {
22968 		/* Disabling hybrid pacing */
22969 		if (rack->rc_hybrid_mode) {
22970 			rack_set_profile(rack, 0);
22971 			rack->rc_tp->tcp_hybrid_stop++;
22972 		}
22973 		rack_log_hybrid(rack, seq, sft, HYBRID_LOG_TURNED_OFF, __LINE__, 0);
22974 		return (0);
22975 	}
22976 	if (rack->dgp_on == 0) {
22977 		/*
22978 		 * If we have not yet turned DGP on, do so
22979 		 * now setting pure DGP mode, no buffer level
22980 		 * response.
22981 		 */
22982 		if ((err = rack_set_profile(rack, 1)) != 0){
22983 			/* Failed to turn pacing on */
22984 			rack->rc_tp->tcp_hybrid_error++;
22985 			rack_log_hybrid(rack, seq, sft, HYBRID_LOG_NO_PACING, __LINE__, 0);
22986 			return (err);
22987 		}
22988 	}
22989 	/* Now set in our flags */
22990 	sft->hybrid_flags = hybrid->hybrid_flags | TCP_HYBRID_PACING_WASSET;
22991 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_CSPR)
22992 		sft->cspr = hybrid->cspr;
22993 	else
22994 		sft->cspr = 0;
22995 	if (hybrid->hybrid_flags & TCP_HYBRID_PACING_H_MS)
22996 		sft->hint_maxseg = hybrid->hint_maxseg;
22997 	else
22998 		sft->hint_maxseg = 0;
22999 	rack->rc_hybrid_mode = 1;
23000 	rack->rc_tp->tcp_hybrid_start++;
23001 	rack_log_hybrid(rack, seq, sft, HYBRID_LOG_RULES_SET, __LINE__,0);
23002 	return (0);
23003 #else
23004 	return (ENOTSUP);
23005 #endif
23006 }
23007 
23008 static int
23009 rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name,
23010 		    uint32_t optval, uint64_t loptval, struct tcp_hybrid_req *hybrid)
23011 
23012 {
23013 	struct epoch_tracker et;
23014 	struct sockopt sopt;
23015 	struct cc_newreno_opts opt;
23016 	uint64_t val;
23017 	int error = 0;
23018 	uint16_t ca, ss;
23019 
23020 	switch (sopt_name) {
23021 	case TCP_RACK_SET_RXT_OPTIONS:
23022 		if ((optval >= 0) && (optval <= 2)) {
23023 			rack_init_retransmit_value(rack, optval);
23024 		} else {
23025 			/*
23026 			 * You must send in 0, 1 or 2 all else is
23027 			 * invalid.
23028 			 */
23029 			error = EINVAL;
23030 		}
23031 		break;
23032 	case TCP_RACK_DSACK_OPT:
23033 		RACK_OPTS_INC(tcp_rack_dsack_opt);
23034 		if (optval & 0x1) {
23035 			rack->rc_rack_tmr_std_based = 1;
23036 		} else {
23037 			rack->rc_rack_tmr_std_based = 0;
23038 		}
23039 		if (optval & 0x2) {
23040 			rack->rc_rack_use_dsack = 1;
23041 		} else {
23042 			rack->rc_rack_use_dsack = 0;
23043 		}
23044 		rack_log_dsack_event(rack, 5, __LINE__, 0, 0);
23045 		break;
23046 	case TCP_RACK_PACING_DIVISOR:
23047 		RACK_OPTS_INC(tcp_rack_pacing_divisor);
23048 		if (optval == 0) {
23049 			rack->r_ctl.pace_len_divisor = rack_default_pacing_divisor;
23050 		} else {
23051 			if (optval < RL_MIN_DIVISOR)
23052 				rack->r_ctl.pace_len_divisor = RL_MIN_DIVISOR;
23053 			else
23054 				rack->r_ctl.pace_len_divisor = optval;
23055 		}
23056 		break;
23057 	case TCP_RACK_HI_BETA:
23058 		RACK_OPTS_INC(tcp_rack_hi_beta);
23059 		if (optval > 0) {
23060 			rack->rack_hibeta = 1;
23061 			if ((optval >= 50) &&
23062 			    (optval <= 100)) {
23063 				/*
23064 				 * User wants to set a custom beta.
23065 				 */
23066 				rack->r_ctl.saved_hibeta = optval;
23067 				if (rack->rc_pacing_cc_set)
23068 					rack_undo_cc_pacing(rack);
23069 				rack->r_ctl.rc_saved_beta.beta = optval;
23070 			}
23071 			if (rack->rc_pacing_cc_set == 0)
23072 				rack_set_cc_pacing(rack);
23073 		} else {
23074 			rack->rack_hibeta = 0;
23075 			if (rack->rc_pacing_cc_set)
23076 				rack_undo_cc_pacing(rack);
23077 		}
23078 		break;
23079 	case TCP_RACK_PACING_BETA:
23080 		RACK_OPTS_INC(tcp_rack_beta);
23081 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
23082 			/* This only works for newreno. */
23083 			error = EINVAL;
23084 			break;
23085 		}
23086 		if (rack->rc_pacing_cc_set) {
23087 			/*
23088 			 * Set them into the real CC module
23089 			 * whats in the rack pcb is the old values
23090 			 * to be used on restoral/
23091 			 */
23092 			sopt.sopt_dir = SOPT_SET;
23093 			opt.name = CC_NEWRENO_BETA;
23094 			opt.val = optval;
23095 			if (CC_ALGO(tp)->ctl_output != NULL)
23096 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
23097 			else {
23098 				error = ENOENT;
23099 				break;
23100 			}
23101 		} else {
23102 			/*
23103 			 * Not pacing yet so set it into our local
23104 			 * rack pcb storage.
23105 			 */
23106 			rack->r_ctl.rc_saved_beta.beta = optval;
23107 		}
23108 		break;
23109 	case TCP_RACK_TIMER_SLOP:
23110 		RACK_OPTS_INC(tcp_rack_timer_slop);
23111 		rack->r_ctl.timer_slop = optval;
23112 		if (rack->rc_tp->t_srtt) {
23113 			/*
23114 			 * If we have an SRTT lets update t_rxtcur
23115 			 * to have the new slop.
23116 			 */
23117 			RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp),
23118 					   rack_rto_min, rack_rto_max,
23119 					   rack->r_ctl.timer_slop);
23120 		}
23121 		break;
23122 	case TCP_RACK_PACING_BETA_ECN:
23123 		RACK_OPTS_INC(tcp_rack_beta_ecn);
23124 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0) {
23125 			/* This only works for newreno. */
23126 			error = EINVAL;
23127 			break;
23128 		}
23129 		if (rack->rc_pacing_cc_set) {
23130 			/*
23131 			 * Set them into the real CC module
23132 			 * whats in the rack pcb is the old values
23133 			 * to be used on restoral/
23134 			 */
23135 			sopt.sopt_dir = SOPT_SET;
23136 			opt.name = CC_NEWRENO_BETA_ECN;
23137 			opt.val = optval;
23138 			if (CC_ALGO(tp)->ctl_output != NULL)
23139 				error = CC_ALGO(tp)->ctl_output(&tp->t_ccv, &sopt, &opt);
23140 			else
23141 				error = ENOENT;
23142 		} else {
23143 			/*
23144 			 * Not pacing yet so set it into our local
23145 			 * rack pcb storage.
23146 			 */
23147 			rack->r_ctl.rc_saved_beta.beta_ecn = optval;
23148 			rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED;
23149 		}
23150 		break;
23151 	case TCP_DEFER_OPTIONS:
23152 		RACK_OPTS_INC(tcp_defer_opt);
23153 		if (optval) {
23154 			if (rack->gp_ready) {
23155 				/* Too late */
23156 				error = EINVAL;
23157 				break;
23158 			}
23159 			rack->defer_options = 1;
23160 		} else
23161 			rack->defer_options = 0;
23162 		break;
23163 	case TCP_RACK_MEASURE_CNT:
23164 		RACK_OPTS_INC(tcp_rack_measure_cnt);
23165 		if (optval && (optval <= 0xff)) {
23166 			rack->r_ctl.req_measurements = optval;
23167 		} else
23168 			error = EINVAL;
23169 		break;
23170 	case TCP_REC_ABC_VAL:
23171 		RACK_OPTS_INC(tcp_rec_abc_val);
23172 		if (optval > 0)
23173 			rack->r_use_labc_for_rec = 1;
23174 		else
23175 			rack->r_use_labc_for_rec = 0;
23176 		break;
23177 	case TCP_RACK_ABC_VAL:
23178 		RACK_OPTS_INC(tcp_rack_abc_val);
23179 		if ((optval > 0) && (optval < 255))
23180 			rack->rc_labc = optval;
23181 		else
23182 			error = EINVAL;
23183 		break;
23184 	case TCP_HDWR_UP_ONLY:
23185 		RACK_OPTS_INC(tcp_pacing_up_only);
23186 		if (optval)
23187 			rack->r_up_only = 1;
23188 		else
23189 			rack->r_up_only = 0;
23190 		break;
23191 	case TCP_PACING_RATE_CAP:
23192 		RACK_OPTS_INC(tcp_pacing_rate_cap);
23193 		rack->r_ctl.bw_rate_cap = loptval;
23194 		break;
23195 	case TCP_HYBRID_PACING:
23196 		if (hybrid == NULL) {
23197 			error = EINVAL;
23198 			break;
23199 		}
23200 		error = process_hybrid_pacing(rack, hybrid);
23201 		break;
23202 	case TCP_RACK_PROFILE:
23203 		RACK_OPTS_INC(tcp_profile);
23204 		error = rack_set_profile(rack, optval);
23205 		break;
23206 	case TCP_USE_CMP_ACKS:
23207 		RACK_OPTS_INC(tcp_use_cmp_acks);
23208 		if ((optval == 0) && (tp->t_flags2 & TF2_MBUF_ACKCMP)) {
23209 			/* You can't turn it off once its on! */
23210 			error = EINVAL;
23211 		} else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) {
23212 			rack->r_use_cmp_ack = 1;
23213 			rack->r_mbuf_queue = 1;
23214 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23215 		}
23216 		if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state))
23217 			tp->t_flags2 |= TF2_MBUF_ACKCMP;
23218 		break;
23219 	case TCP_SHARED_CWND_TIME_LIMIT:
23220 		RACK_OPTS_INC(tcp_lscwnd);
23221 		if (optval)
23222 			rack->r_limit_scw = 1;
23223 		else
23224 			rack->r_limit_scw = 0;
23225 		break;
23226 	case TCP_RACK_DGP_IN_REC:
23227 		RACK_OPTS_INC(tcp_dgp_in_rec);
23228 		if (optval)
23229 			rack->r_ctl.full_dgp_in_rec = 1;
23230 		else
23231 			rack->r_ctl.full_dgp_in_rec = 0;
23232 		break;
23233 	case TCP_RXT_CLAMP:
23234 		RACK_OPTS_INC(tcp_rxt_clamp);
23235 		rack_translate_clamp_value(rack, optval);
23236 		break;
23237  	case TCP_RACK_PACE_TO_FILL:
23238 		RACK_OPTS_INC(tcp_fillcw);
23239 		if (optval == 0)
23240 			rack->rc_pace_to_cwnd = 0;
23241 		else {
23242 			rack->rc_pace_to_cwnd = 1;
23243 			if (optval > 1)
23244 				rack->r_fill_less_agg = 1;
23245 		}
23246 		if ((optval >= rack_gp_rtt_maxmul) &&
23247 		    rack_gp_rtt_maxmul &&
23248 		    (optval < 0xf)) {
23249 			rack->rc_pace_fill_if_rttin_range = 1;
23250 			rack->rtt_limit_mul = optval;
23251 		} else {
23252 			rack->rc_pace_fill_if_rttin_range = 0;
23253 			rack->rtt_limit_mul = 0;
23254 		}
23255 		break;
23256 	case TCP_RACK_NO_PUSH_AT_MAX:
23257 		RACK_OPTS_INC(tcp_npush);
23258 		if (optval == 0)
23259 			rack->r_ctl.rc_no_push_at_mrtt = 0;
23260 		else if (optval < 0xff)
23261 			rack->r_ctl.rc_no_push_at_mrtt = optval;
23262 		else
23263 			error = EINVAL;
23264 		break;
23265 	case TCP_SHARED_CWND_ENABLE:
23266 		RACK_OPTS_INC(tcp_rack_scwnd);
23267 		if (optval == 0)
23268 			rack->rack_enable_scwnd = 0;
23269 		else
23270 			rack->rack_enable_scwnd = 1;
23271 		break;
23272 	case TCP_RACK_MBUF_QUEUE:
23273 		/* Now do we use the LRO mbuf-queue feature */
23274 		RACK_OPTS_INC(tcp_rack_mbufq);
23275 		if (optval || rack->r_use_cmp_ack)
23276 			rack->r_mbuf_queue = 1;
23277 		else
23278 			rack->r_mbuf_queue = 0;
23279 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23280 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23281 		else
23282 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23283 		break;
23284 	case TCP_RACK_NONRXT_CFG_RATE:
23285 		RACK_OPTS_INC(tcp_rack_cfg_rate);
23286 		if (optval == 0)
23287 			rack->rack_rec_nonrxt_use_cr = 0;
23288 		else
23289 			rack->rack_rec_nonrxt_use_cr = 1;
23290 		break;
23291 	case TCP_NO_PRR:
23292 		RACK_OPTS_INC(tcp_rack_noprr);
23293 		if (optval == 0)
23294 			rack->rack_no_prr = 0;
23295 		else if (optval == 1)
23296 			rack->rack_no_prr = 1;
23297 		else if (optval == 2)
23298 			rack->no_prr_addback = 1;
23299 		else
23300 			error = EINVAL;
23301 		break;
23302 	case TCP_TIMELY_DYN_ADJ:
23303 		RACK_OPTS_INC(tcp_timely_dyn);
23304 		if (optval == 0)
23305 			rack->rc_gp_dyn_mul = 0;
23306 		else {
23307 			rack->rc_gp_dyn_mul = 1;
23308 			if (optval >= 100) {
23309 				/*
23310 				 * If the user sets something 100 or more
23311 				 * its the gp_ca value.
23312 				 */
23313 				rack->r_ctl.rack_per_of_gp_ca  = optval;
23314 			}
23315 		}
23316 		break;
23317 	case TCP_RACK_DO_DETECTION:
23318 		RACK_OPTS_INC(tcp_rack_do_detection);
23319 		if (optval == 0)
23320 			rack->do_detection = 0;
23321 		else
23322 			rack->do_detection = 1;
23323 		break;
23324 	case TCP_RACK_TLP_USE:
23325 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
23326 			error = EINVAL;
23327 			break;
23328 		}
23329 		RACK_OPTS_INC(tcp_tlp_use);
23330 		rack->rack_tlp_threshold_use = optval;
23331 		break;
23332 	case TCP_RACK_TLP_REDUCE:
23333 		/* RACK TLP cwnd reduction (bool) */
23334 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
23335 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
23336 		break;
23337 		/*  Pacing related ones */
23338 	case TCP_RACK_PACE_ALWAYS:
23339 		/*
23340 		 * zero is old rack method, 1 is new
23341 		 * method using a pacing rate.
23342 		 */
23343 		RACK_OPTS_INC(tcp_rack_pace_always);
23344 		if (optval > 0) {
23345 			if (rack->rc_always_pace) {
23346 				error = EALREADY;
23347 				break;
23348 			} else if (tcp_can_enable_pacing()) {
23349 				rack->rc_always_pace = 1;
23350 				if (rack->rack_hibeta)
23351 					rack_set_cc_pacing(rack);
23352 			}
23353 			else {
23354 				error = ENOSPC;
23355 				break;
23356 			}
23357 		} else {
23358 			if (rack->rc_always_pace) {
23359 				tcp_decrement_paced_conn();
23360 				rack->rc_always_pace = 0;
23361 				rack_undo_cc_pacing(rack);
23362 			}
23363 		}
23364 		if  (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack)
23365 			tp->t_flags2 |= TF2_SUPPORTS_MBUFQ;
23366 		else
23367 			tp->t_flags2 &= ~TF2_SUPPORTS_MBUFQ;
23368 		/* A rate may be set irate or other, if so set seg size */
23369 		rack_update_seg(rack);
23370 		break;
23371 	case TCP_BBR_RACK_INIT_RATE:
23372 		RACK_OPTS_INC(tcp_initial_rate);
23373 		val = optval;
23374 		/* Change from kbits per second to bytes per second */
23375 		val *= 1000;
23376 		val /= 8;
23377 		rack->r_ctl.init_rate = val;
23378 		if (rack->rc_init_win != rack_default_init_window) {
23379 			uint32_t win, snt;
23380 
23381 			/*
23382 			 * Options don't always get applied
23383 			 * in the order you think. So in order
23384 			 * to assure we update a cwnd we need
23385 			 * to check and see if we are still
23386 			 * where we should raise the cwnd.
23387 			 */
23388 			win = rc_init_window(rack);
23389 			if (SEQ_GT(tp->snd_max, tp->iss))
23390 				snt = tp->snd_max - tp->iss;
23391 			else
23392 				snt = 0;
23393 			if ((snt < win) &&
23394 			    (tp->snd_cwnd < win))
23395 				tp->snd_cwnd = win;
23396 		}
23397 		if (rack->rc_always_pace)
23398 			rack_update_seg(rack);
23399 		break;
23400 	case TCP_BBR_IWINTSO:
23401 		RACK_OPTS_INC(tcp_initial_win);
23402 		if (optval && (optval <= 0xff)) {
23403 			uint32_t win, snt;
23404 
23405 			rack->rc_init_win = optval;
23406 			win = rc_init_window(rack);
23407 			if (SEQ_GT(tp->snd_max, tp->iss))
23408 				snt = tp->snd_max - tp->iss;
23409 			else
23410 				snt = 0;
23411 			if ((snt < win) &&
23412 			    (tp->t_srtt |
23413 			     rack->r_ctl.init_rate)) {
23414 				/*
23415 				 * We are not past the initial window
23416 				 * and we have some bases for pacing,
23417 				 * so we need to possibly adjust up
23418 				 * the cwnd. Note even if we don't set
23419 				 * the cwnd, its still ok to raise the rc_init_win
23420 				 * which can be used coming out of idle when we
23421 				 * would have a rate.
23422 				 */
23423 				if (tp->snd_cwnd < win)
23424 					tp->snd_cwnd = win;
23425 			}
23426 			if (rack->rc_always_pace)
23427 				rack_update_seg(rack);
23428 		} else
23429 			error = EINVAL;
23430 		break;
23431 	case TCP_RACK_FORCE_MSEG:
23432 		RACK_OPTS_INC(tcp_rack_force_max_seg);
23433 		if (optval)
23434 			rack->rc_force_max_seg = 1;
23435 		else
23436 			rack->rc_force_max_seg = 0;
23437 		break;
23438 	case TCP_RACK_PACE_MIN_SEG:
23439 		RACK_OPTS_INC(tcp_rack_min_seg);
23440 		rack->r_ctl.rc_user_set_min_segs = (0x0000ffff & optval);
23441 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
23442 		break;
23443 	case TCP_RACK_PACE_MAX_SEG:
23444 		/* Max segments size in a pace in bytes */
23445 		RACK_OPTS_INC(tcp_rack_max_seg);
23446 		if (optval <= MAX_USER_SET_SEG)
23447 			rack->rc_user_set_max_segs = optval;
23448 		else
23449 			rack->rc_user_set_max_segs = MAX_USER_SET_SEG;
23450 		rack_set_pace_segments(tp, rack, __LINE__, NULL);
23451 		break;
23452 	case TCP_RACK_PACE_RATE_REC:
23453 		/* Set the fixed pacing rate in Bytes per second ca */
23454 		RACK_OPTS_INC(tcp_rack_pace_rate_rec);
23455 		rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23456 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23457 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23458 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23459 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23460 		rack->use_fixed_rate = 1;
23461 		if (rack->rack_hibeta)
23462 			rack_set_cc_pacing(rack);
23463 		rack_log_pacing_delay_calc(rack,
23464 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23465 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23466 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23467 					   __LINE__, NULL,0);
23468 		break;
23469 
23470 	case TCP_RACK_PACE_RATE_SS:
23471 		/* Set the fixed pacing rate in Bytes per second ca */
23472 		RACK_OPTS_INC(tcp_rack_pace_rate_ss);
23473 		rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23474 		if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0)
23475 			rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23476 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23477 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23478 		rack->use_fixed_rate = 1;
23479 		if (rack->rack_hibeta)
23480 			rack_set_cc_pacing(rack);
23481 		rack_log_pacing_delay_calc(rack,
23482 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23483 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23484 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23485 					   __LINE__, NULL, 0);
23486 		break;
23487 
23488 	case TCP_RACK_PACE_RATE_CA:
23489 		/* Set the fixed pacing rate in Bytes per second ca */
23490 		RACK_OPTS_INC(tcp_rack_pace_rate_ca);
23491 		rack->r_ctl.rc_fixed_pacing_rate_ca = optval;
23492 		if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0)
23493 			rack->r_ctl.rc_fixed_pacing_rate_ss = optval;
23494 		if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0)
23495 			rack->r_ctl.rc_fixed_pacing_rate_rec = optval;
23496 		rack->use_fixed_rate = 1;
23497 		if (rack->rack_hibeta)
23498 			rack_set_cc_pacing(rack);
23499 		rack_log_pacing_delay_calc(rack,
23500 					   rack->r_ctl.rc_fixed_pacing_rate_ss,
23501 					   rack->r_ctl.rc_fixed_pacing_rate_ca,
23502 					   rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8,
23503 					   __LINE__, NULL, 0);
23504 		break;
23505 	case TCP_RACK_GP_INCREASE_REC:
23506 		RACK_OPTS_INC(tcp_gp_inc_rec);
23507 		rack->r_ctl.rack_per_of_gp_rec = optval;
23508 		rack_log_pacing_delay_calc(rack,
23509 					   rack->r_ctl.rack_per_of_gp_ss,
23510 					   rack->r_ctl.rack_per_of_gp_ca,
23511 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23512 					   __LINE__, NULL, 0);
23513 		break;
23514 	case TCP_RACK_GP_INCREASE_CA:
23515 		RACK_OPTS_INC(tcp_gp_inc_ca);
23516 		ca = optval;
23517 		if (ca < 100) {
23518 			/*
23519 			 * We don't allow any reduction
23520 			 * over the GP b/w.
23521 			 */
23522 			error = EINVAL;
23523 			break;
23524 		}
23525 		rack->r_ctl.rack_per_of_gp_ca = ca;
23526 		rack_log_pacing_delay_calc(rack,
23527 					   rack->r_ctl.rack_per_of_gp_ss,
23528 					   rack->r_ctl.rack_per_of_gp_ca,
23529 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23530 					   __LINE__, NULL, 0);
23531 		break;
23532 	case TCP_RACK_GP_INCREASE_SS:
23533 		RACK_OPTS_INC(tcp_gp_inc_ss);
23534 		ss = optval;
23535 		if (ss < 100) {
23536 			/*
23537 			 * We don't allow any reduction
23538 			 * over the GP b/w.
23539 			 */
23540 			error = EINVAL;
23541 			break;
23542 		}
23543 		rack->r_ctl.rack_per_of_gp_ss = ss;
23544 		rack_log_pacing_delay_calc(rack,
23545 					   rack->r_ctl.rack_per_of_gp_ss,
23546 					   rack->r_ctl.rack_per_of_gp_ca,
23547 					   rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1,
23548 					   __LINE__, NULL, 0);
23549 		break;
23550 	case TCP_RACK_RR_CONF:
23551 		RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate);
23552 		if (optval && optval <= 3)
23553 			rack->r_rr_config = optval;
23554 		else
23555 			rack->r_rr_config = 0;
23556 		break;
23557 	case TCP_PACING_DND:			/*  URL:dnd */
23558 		if (optval > 0)
23559 			rack->rc_pace_dnd = 1;
23560 		else
23561 			rack->rc_pace_dnd = 0;
23562 		break;
23563 	case TCP_HDWR_RATE_CAP:
23564 		RACK_OPTS_INC(tcp_hdwr_rate_cap);
23565 		if (optval) {
23566 			if (rack->r_rack_hw_rate_caps == 0)
23567 				rack->r_rack_hw_rate_caps = 1;
23568 			else
23569 				error = EALREADY;
23570 		} else {
23571 			rack->r_rack_hw_rate_caps = 0;
23572 		}
23573 		break;
23574 	case TCP_RACK_SPLIT_LIMIT:
23575 		RACK_OPTS_INC(tcp_split_limit);
23576 		rack->r_ctl.rc_split_limit = optval;
23577 		break;
23578 	case TCP_BBR_HDWR_PACE:
23579 		RACK_OPTS_INC(tcp_hdwr_pacing);
23580 		if (optval){
23581 			if (rack->rack_hdrw_pacing == 0) {
23582 				rack->rack_hdw_pace_ena = 1;
23583 				rack->rack_attempt_hdwr_pace = 0;
23584 			} else
23585 				error = EALREADY;
23586 		} else {
23587 			rack->rack_hdw_pace_ena = 0;
23588 #ifdef RATELIMIT
23589 			if (rack->r_ctl.crte != NULL) {
23590 				rack->rack_hdrw_pacing = 0;
23591 				rack->rack_attempt_hdwr_pace = 0;
23592 				tcp_rel_pacing_rate(rack->r_ctl.crte, tp);
23593 				rack->r_ctl.crte = NULL;
23594 			}
23595 #endif
23596 		}
23597 		break;
23598 		/*  End Pacing related ones */
23599 	case TCP_RACK_PRR_SENDALOT:
23600 		/* Allow PRR to send more than one seg */
23601 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
23602 		rack->r_ctl.rc_prr_sendalot = optval;
23603 		break;
23604 	case TCP_RACK_MIN_TO:
23605 		/* Minimum time between rack t-o's in ms */
23606 		RACK_OPTS_INC(tcp_rack_min_to);
23607 		rack->r_ctl.rc_min_to = optval;
23608 		break;
23609 	case TCP_RACK_EARLY_SEG:
23610 		/* If early recovery max segments */
23611 		RACK_OPTS_INC(tcp_rack_early_seg);
23612 		rack->r_ctl.rc_early_recovery_segs = optval;
23613 		break;
23614 	case TCP_RACK_ENABLE_HYSTART:
23615 	{
23616 		if (optval) {
23617 			tp->t_ccv.flags |= CCF_HYSTART_ALLOWED;
23618 			if (rack_do_hystart > RACK_HYSTART_ON)
23619 				tp->t_ccv.flags |= CCF_HYSTART_CAN_SH_CWND;
23620 			if (rack_do_hystart > RACK_HYSTART_ON_W_SC)
23621 				tp->t_ccv.flags |= CCF_HYSTART_CONS_SSTH;
23622 		} else {
23623 			tp->t_ccv.flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH);
23624 		}
23625 	}
23626 	break;
23627 	case TCP_RACK_REORD_THRESH:
23628 		/* RACK reorder threshold (shift amount) */
23629 		RACK_OPTS_INC(tcp_rack_reord_thresh);
23630 		if ((optval > 0) && (optval < 31))
23631 			rack->r_ctl.rc_reorder_shift = optval;
23632 		else
23633 			error = EINVAL;
23634 		break;
23635 	case TCP_RACK_REORD_FADE:
23636 		/* Does reordering fade after ms time */
23637 		RACK_OPTS_INC(tcp_rack_reord_fade);
23638 		rack->r_ctl.rc_reorder_fade = optval;
23639 		break;
23640 	case TCP_RACK_TLP_THRESH:
23641 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
23642 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
23643 		if (optval)
23644 			rack->r_ctl.rc_tlp_threshold = optval;
23645 		else
23646 			error = EINVAL;
23647 		break;
23648 	case TCP_BBR_USE_RACK_RR:
23649 		RACK_OPTS_INC(tcp_rack_rr);
23650 		if (optval)
23651 			rack->use_rack_rr = 1;
23652 		else
23653 			rack->use_rack_rr = 0;
23654 		break;
23655 	case TCP_RACK_PKT_DELAY:
23656 		/* RACK added ms i.e. rack-rtt + reord + N */
23657 		RACK_OPTS_INC(tcp_rack_pkt_delay);
23658 		rack->r_ctl.rc_pkt_delay = optval;
23659 		break;
23660 	case TCP_DELACK:
23661 		RACK_OPTS_INC(tcp_rack_delayed_ack);
23662 		if (optval == 0)
23663 			tp->t_delayed_ack = 0;
23664 		else
23665 			tp->t_delayed_ack = 1;
23666 		if (tp->t_flags & TF_DELACK) {
23667 			tp->t_flags &= ~TF_DELACK;
23668 			tp->t_flags |= TF_ACKNOW;
23669 			NET_EPOCH_ENTER(et);
23670 			rack_output(tp);
23671 			NET_EPOCH_EXIT(et);
23672 		}
23673 		break;
23674 
23675 	case TCP_BBR_RACK_RTT_USE:
23676 		RACK_OPTS_INC(tcp_rack_rtt_use);
23677 		if ((optval != USE_RTT_HIGH) &&
23678 		    (optval != USE_RTT_LOW) &&
23679 		    (optval != USE_RTT_AVG))
23680 			error = EINVAL;
23681 		else
23682 			rack->r_ctl.rc_rate_sample_method = optval;
23683 		break;
23684 	case TCP_DATA_AFTER_CLOSE:
23685 		RACK_OPTS_INC(tcp_data_after_close);
23686 		if (optval)
23687 			rack->rc_allow_data_af_clo = 1;
23688 		else
23689 			rack->rc_allow_data_af_clo = 0;
23690 		break;
23691 	default:
23692 		break;
23693 	}
23694 	tcp_log_socket_option(tp, sopt_name, optval, error);
23695 	return (error);
23696 }
23697 
23698 
23699 static void
23700 rack_apply_deferred_options(struct tcp_rack *rack)
23701 {
23702 	struct deferred_opt_list *dol, *sdol;
23703 	uint32_t s_optval;
23704 
23705 	TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) {
23706 		TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next);
23707 		/* Disadvantage of deferal is you loose the error return */
23708 		s_optval = (uint32_t)dol->optval;
23709 		(void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval, NULL);
23710 		free(dol, M_TCPDO);
23711 	}
23712 }
23713 
23714 static void
23715 rack_hw_tls_change(struct tcpcb *tp, int chg)
23716 {
23717 	/* Update HW tls state */
23718 	struct tcp_rack *rack;
23719 
23720 	rack = (struct tcp_rack *)tp->t_fb_ptr;
23721 	if (chg)
23722 		rack->r_ctl.fsb.hw_tls = 1;
23723 	else
23724 		rack->r_ctl.fsb.hw_tls = 0;
23725 }
23726 
23727 static int
23728 rack_pru_options(struct tcpcb *tp, int flags)
23729 {
23730 	if (flags & PRUS_OOB)
23731 		return (EOPNOTSUPP);
23732 	return (0);
23733 }
23734 
23735 static bool
23736 rack_wake_check(struct tcpcb *tp)
23737 {
23738 	struct tcp_rack *rack;
23739 	struct timeval tv;
23740 	uint32_t cts;
23741 
23742 	rack = (struct tcp_rack *)tp->t_fb_ptr;
23743 	if (rack->r_ctl.rc_hpts_flags) {
23744 		cts = tcp_get_usecs(&tv);
23745 		if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == PACE_PKT_OUTPUT){
23746 			/*
23747 			 * Pacing timer is up, check if we are ready.
23748 			 */
23749 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to))
23750 				return (true);
23751 		} else if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) != 0) {
23752 			/*
23753 			 * A timer is up, check if we are ready.
23754 			 */
23755 			if (TSTMP_GEQ(cts, rack->r_ctl.rc_timer_exp))
23756 				return (true);
23757 		}
23758 	}
23759 	return (false);
23760 }
23761 
23762 static struct tcp_function_block __tcp_rack = {
23763 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
23764 	.tfb_tcp_output = rack_output,
23765 	.tfb_do_queued_segments = ctf_do_queued_segments,
23766 	.tfb_do_segment_nounlock = rack_do_segment_nounlock,
23767 	.tfb_tcp_do_segment = rack_do_segment,
23768 	.tfb_tcp_ctloutput = rack_ctloutput,
23769 	.tfb_tcp_fb_init = rack_init,
23770 	.tfb_tcp_fb_fini = rack_fini,
23771 	.tfb_tcp_timer_stop_all = rack_stopall,
23772 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
23773 	.tfb_tcp_handoff_ok = rack_handoff_ok,
23774 	.tfb_tcp_mtu_chg = rack_mtu_change,
23775 	.tfb_pru_options = rack_pru_options,
23776 	.tfb_hwtls_change = rack_hw_tls_change,
23777 	.tfb_chg_query = rack_chg_query,
23778 	.tfb_switch_failed = rack_switch_failed,
23779 	.tfb_early_wake_check = rack_wake_check,
23780 	.tfb_compute_pipe = rack_compute_pipe,
23781 	.tfb_flags = TCP_FUNC_OUTPUT_CANDROP,
23782 };
23783 
23784 /*
23785  * rack_ctloutput() must drop the inpcb lock before performing copyin on
23786  * socket option arguments.  When it re-acquires the lock after the copy, it
23787  * has to revalidate that the connection is still valid for the socket
23788  * option.
23789  */
23790 static int
23791 rack_set_sockopt(struct tcpcb *tp, struct sockopt *sopt)
23792 {
23793 	struct inpcb *inp = tptoinpcb(tp);
23794 #ifdef INET
23795 	struct ip *ip;
23796 #endif
23797 	struct tcp_rack *rack;
23798 	struct tcp_hybrid_req hybrid;
23799 	uint64_t loptval;
23800 	int32_t error = 0, optval;
23801 
23802 	rack = (struct tcp_rack *)tp->t_fb_ptr;
23803 	if (rack == NULL) {
23804 		INP_WUNLOCK(inp);
23805 		return (EINVAL);
23806 	}
23807 #ifdef INET
23808 	ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr;
23809 #endif
23810 
23811 	switch (sopt->sopt_level) {
23812 #ifdef INET6
23813 	case IPPROTO_IPV6:
23814 		MPASS(inp->inp_vflag & INP_IPV6PROTO);
23815 		switch (sopt->sopt_name) {
23816 		case IPV6_USE_MIN_MTU:
23817 			tcp6_use_min_mtu(tp);
23818 			break;
23819 		}
23820 		INP_WUNLOCK(inp);
23821 		return (0);
23822 #endif
23823 #ifdef INET
23824 	case IPPROTO_IP:
23825 		switch (sopt->sopt_name) {
23826 		case IP_TOS:
23827 			/*
23828 			 * The DSCP codepoint has changed, update the fsb.
23829 			 */
23830 			ip->ip_tos = rack->rc_inp->inp_ip_tos;
23831 			break;
23832 		case IP_TTL:
23833 			/*
23834 			 * The TTL has changed, update the fsb.
23835 			 */
23836 			ip->ip_ttl = rack->rc_inp->inp_ip_ttl;
23837 			break;
23838 		}
23839 		INP_WUNLOCK(inp);
23840 		return (0);
23841 #endif
23842 #ifdef SO_PEERPRIO
23843 	case SOL_SOCKET:
23844 		switch (sopt->sopt_name) {
23845 		case SO_PEERPRIO:			/*  SC-URL:bs */
23846 			/* Already read in and sanity checked in sosetopt(). */
23847 			if (inp->inp_socket) {
23848 				rack->client_bufferlvl = inp->inp_socket->so_peerprio;
23849 				rack_client_buffer_level_set(rack);
23850 			}
23851 			break;
23852 		}
23853 		INP_WUNLOCK(inp);
23854 		return (0);
23855 #endif
23856 	case IPPROTO_TCP:
23857 		switch (sopt->sopt_name) {
23858 		case TCP_RACK_TLP_REDUCE:		/*  URL:tlp_reduce */
23859 		/*  Pacing related ones */
23860 		case TCP_RACK_PACE_ALWAYS:		/*  URL:pace_always */
23861 		case TCP_BBR_RACK_INIT_RATE:		/*  URL:irate */
23862 		case TCP_BBR_IWINTSO:			/*  URL:tso_iwin */
23863 		case TCP_RACK_PACE_MIN_SEG:		/*  URL:pace_min_seg */
23864 		case TCP_RACK_PACE_MAX_SEG:		/*  URL:pace_max_seg */
23865 		case TCP_RACK_FORCE_MSEG:		/*  URL:force_max_seg */
23866 		case TCP_RACK_PACE_RATE_CA:		/*  URL:pr_ca */
23867 		case TCP_RACK_PACE_RATE_SS:		/*  URL:pr_ss*/
23868 		case TCP_RACK_PACE_RATE_REC:		/*  URL:pr_rec */
23869 		case TCP_RACK_GP_INCREASE_CA:		/*  URL:gp_inc_ca */
23870 		case TCP_RACK_GP_INCREASE_SS:		/*  URL:gp_inc_ss */
23871 		case TCP_RACK_GP_INCREASE_REC:		/*  URL:gp_inc_rec */
23872 		case TCP_RACK_RR_CONF:			/*  URL:rrr_conf */
23873 		case TCP_BBR_HDWR_PACE:			/*  URL:hdwrpace */
23874 		case TCP_HDWR_RATE_CAP:			/*  URL:hdwrcap boolean */
23875 		case TCP_PACING_RATE_CAP:		/*  URL:cap  -- used by side-channel */
23876 		case TCP_HDWR_UP_ONLY:			/*  URL:uponly -- hardware pacing  boolean */
23877 		case TCP_RACK_PACING_BETA:		/*  URL:pacing_beta */
23878 		case TCP_RACK_PACING_BETA_ECN:		/*  URL:pacing_beta_ecn */
23879 		case TCP_RACK_PACE_TO_FILL:		/*  URL:fillcw */
23880 		case TCP_RACK_DGP_IN_REC:		/*  URL:dgpinrec */
23881 			/* End pacing related */
23882 		case TCP_RXT_CLAMP:			/*  URL:rxtclamp */
23883 		case TCP_DELACK:			/*  URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */
23884 		case TCP_RACK_PRR_SENDALOT:		/*  URL:prr_sendalot */
23885 		case TCP_RACK_MIN_TO:			/*  URL:min_to */
23886 		case TCP_RACK_EARLY_SEG:		/*  URL:early_seg */
23887 		case TCP_RACK_REORD_THRESH:		/*  URL:reord_thresh */
23888 		case TCP_RACK_REORD_FADE:		/*  URL:reord_fade */
23889 		case TCP_RACK_TLP_THRESH:		/*  URL:tlp_thresh */
23890 		case TCP_RACK_PKT_DELAY:		/*  URL:pkt_delay */
23891 		case TCP_RACK_TLP_USE:			/*  URL:tlp_use */
23892 		case TCP_BBR_RACK_RTT_USE:		/*  URL:rttuse */
23893 		case TCP_BBR_USE_RACK_RR:		/*  URL:rackrr */
23894 		case TCP_RACK_DO_DETECTION:		/*  URL:detect */
23895 		case TCP_NO_PRR:			/*  URL:noprr */
23896 		case TCP_TIMELY_DYN_ADJ:      		/*  URL:dynamic */
23897 		case TCP_DATA_AFTER_CLOSE:		/*  no URL */
23898 		case TCP_RACK_NONRXT_CFG_RATE:		/*  URL:nonrxtcr */
23899 		case TCP_SHARED_CWND_ENABLE:		/*  URL:scwnd */
23900 		case TCP_RACK_MBUF_QUEUE:		/*  URL:mqueue */
23901 		case TCP_RACK_NO_PUSH_AT_MAX:		/*  URL:npush */
23902 		case TCP_SHARED_CWND_TIME_LIMIT:	/*  URL:lscwnd */
23903 		case TCP_RACK_PROFILE:			/*  URL:profile */
23904 		case TCP_HYBRID_PACING:			/*  URL:hybrid */
23905 		case TCP_USE_CMP_ACKS:			/*  URL:cmpack */
23906 		case TCP_RACK_ABC_VAL:			/*  URL:labc */
23907 		case TCP_REC_ABC_VAL:			/*  URL:reclabc */
23908 		case TCP_RACK_MEASURE_CNT:		/*  URL:measurecnt */
23909 		case TCP_DEFER_OPTIONS:			/*  URL:defer */
23910 		case TCP_RACK_DSACK_OPT:		/*  URL:dsack */
23911 		case TCP_RACK_TIMER_SLOP:		/*  URL:timer_slop */
23912 		case TCP_RACK_ENABLE_HYSTART:		/*  URL:hystart */
23913 		case TCP_RACK_SET_RXT_OPTIONS:		/*  URL:rxtsz */
23914 		case TCP_RACK_HI_BETA:			/*  URL:hibeta */
23915 		case TCP_RACK_SPLIT_LIMIT:		/*  URL:split */
23916 		case TCP_RACK_PACING_DIVISOR:		/*  URL:divisor */
23917 		case TCP_PACING_DND:			/*  URL:dnd */
23918 			goto process_opt;
23919 			break;
23920 		default:
23921 			/* Filter off all unknown options to the base stack */
23922 			return (tcp_default_ctloutput(tp, sopt));
23923 			break;
23924 		}
23925 
23926 	default:
23927 		INP_WUNLOCK(inp);
23928 		return (0);
23929 	}
23930 process_opt:
23931 	INP_WUNLOCK(inp);
23932 	if (sopt->sopt_name == TCP_PACING_RATE_CAP) {
23933 		error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval));
23934 		/*
23935 		 * We truncate it down to 32 bits for the socket-option trace this
23936 		 * means rates > 34Gbps won't show right, but thats probably ok.
23937 		 */
23938 		optval = (uint32_t)loptval;
23939 	} else if (sopt->sopt_name == TCP_HYBRID_PACING) {
23940 		error = sooptcopyin(sopt, &hybrid, sizeof(hybrid), sizeof(hybrid));
23941 	} else {
23942 		error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
23943 		/* Save it in 64 bit form too */
23944 		loptval = optval;
23945 	}
23946 	if (error)
23947 		return (error);
23948 	INP_WLOCK(inp);
23949 	if (tp->t_fb != &__tcp_rack) {
23950 		INP_WUNLOCK(inp);
23951 		return (ENOPROTOOPT);
23952 	}
23953 	if (rack->defer_options && (rack->gp_ready == 0) &&
23954 	    (sopt->sopt_name != TCP_DEFER_OPTIONS) &&
23955 	    (sopt->sopt_name != TCP_HYBRID_PACING) &&
23956 	    (sopt->sopt_name != TCP_RACK_PACING_BETA) &&
23957 	    (sopt->sopt_name != TCP_RACK_SET_RXT_OPTIONS) &&
23958 	    (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) &&
23959 	    (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) {
23960 		/* Options are beind deferred */
23961 		if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) {
23962 			INP_WUNLOCK(inp);
23963 			return (0);
23964 		} else {
23965 			/* No memory to defer, fail */
23966 			INP_WUNLOCK(inp);
23967 			return (ENOMEM);
23968 		}
23969 	}
23970 	error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval, &hybrid);
23971 	INP_WUNLOCK(inp);
23972 	return (error);
23973 }
23974 
23975 static void
23976 rack_fill_info(struct tcpcb *tp, struct tcp_info *ti)
23977 {
23978 
23979 	INP_WLOCK_ASSERT(tptoinpcb(tp));
23980 	bzero(ti, sizeof(*ti));
23981 
23982 	ti->tcpi_state = tp->t_state;
23983 	if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP))
23984 		ti->tcpi_options |= TCPI_OPT_TIMESTAMPS;
23985 	if (tp->t_flags & TF_SACK_PERMIT)
23986 		ti->tcpi_options |= TCPI_OPT_SACK;
23987 	if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) {
23988 		ti->tcpi_options |= TCPI_OPT_WSCALE;
23989 		ti->tcpi_snd_wscale = tp->snd_scale;
23990 		ti->tcpi_rcv_wscale = tp->rcv_scale;
23991 	}
23992 	if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))
23993 		ti->tcpi_options |= TCPI_OPT_ECN;
23994 	if (tp->t_flags & TF_FASTOPEN)
23995 		ti->tcpi_options |= TCPI_OPT_TFO;
23996 	/* still kept in ticks is t_rcvtime */
23997 	ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick;
23998 	/* Since we hold everything in precise useconds this is easy */
23999 	ti->tcpi_rtt = tp->t_srtt;
24000 	ti->tcpi_rttvar = tp->t_rttvar;
24001 	ti->tcpi_rto = tp->t_rxtcur;
24002 	ti->tcpi_snd_ssthresh = tp->snd_ssthresh;
24003 	ti->tcpi_snd_cwnd = tp->snd_cwnd;
24004 	/*
24005 	 * FreeBSD-specific extension fields for tcp_info.
24006 	 */
24007 	ti->tcpi_rcv_space = tp->rcv_wnd;
24008 	ti->tcpi_rcv_nxt = tp->rcv_nxt;
24009 	ti->tcpi_snd_wnd = tp->snd_wnd;
24010 	ti->tcpi_snd_bwnd = 0;		/* Unused, kept for compat. */
24011 	ti->tcpi_snd_nxt = tp->snd_nxt;
24012 	ti->tcpi_snd_mss = tp->t_maxseg;
24013 	ti->tcpi_rcv_mss = tp->t_maxseg;
24014 	ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack;
24015 	ti->tcpi_rcv_ooopack = tp->t_rcvoopack;
24016 	ti->tcpi_snd_zerowin = tp->t_sndzerowin;
24017 	ti->tcpi_total_tlp = tp->t_sndtlppack;
24018 	ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte;
24019 #ifdef NETFLIX_STATS
24020 	memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo));
24021 #endif
24022 #ifdef TCP_OFFLOAD
24023 	if (tp->t_flags & TF_TOE) {
24024 		ti->tcpi_options |= TCPI_OPT_TOE;
24025 		tcp_offload_tcp_info(tp, ti);
24026 	}
24027 #endif
24028 }
24029 
24030 static int
24031 rack_get_sockopt(struct tcpcb *tp, struct sockopt *sopt)
24032 {
24033 	struct inpcb *inp = tptoinpcb(tp);
24034 	struct tcp_rack *rack;
24035 	int32_t error, optval;
24036 	uint64_t val, loptval;
24037 	struct	tcp_info ti;
24038 	/*
24039 	 * Because all our options are either boolean or an int, we can just
24040 	 * pull everything into optval and then unlock and copy. If we ever
24041 	 * add a option that is not a int, then this will have quite an
24042 	 * impact to this routine.
24043 	 */
24044 	error = 0;
24045 	rack = (struct tcp_rack *)tp->t_fb_ptr;
24046 	if (rack == NULL) {
24047 		INP_WUNLOCK(inp);
24048 		return (EINVAL);
24049 	}
24050 	switch (sopt->sopt_name) {
24051 	case TCP_INFO:
24052 		/* First get the info filled */
24053 		rack_fill_info(tp, &ti);
24054 		/* Fix up the rtt related fields if needed */
24055 		INP_WUNLOCK(inp);
24056 		error = sooptcopyout(sopt, &ti, sizeof ti);
24057 		return (error);
24058 	/*
24059 	 * Beta is the congestion control value for NewReno that influences how
24060 	 * much of a backoff happens when loss is detected. It is normally set
24061 	 * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value
24062 	 * when you exit recovery.
24063 	 */
24064 	case TCP_RACK_PACING_BETA:
24065 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24066 			error = EINVAL;
24067 		else if (rack->rc_pacing_cc_set == 0)
24068 			optval = rack->r_ctl.rc_saved_beta.beta;
24069 		else {
24070 			/*
24071 			 * Reach out into the CC data and report back what
24072 			 * I have previously set. Yeah it looks hackish but
24073 			 * we don't want to report the saved values.
24074 			 */
24075 			if (tp->t_ccv.cc_data)
24076 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta;
24077 			else
24078 				error = EINVAL;
24079 		}
24080 		break;
24081 		/*
24082 		 * Beta_ecn is the congestion control value for NewReno that influences how
24083 		 * much of a backoff happens when a ECN mark is detected. It is normally set
24084 		 * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when
24085 		 * you exit recovery. Note that classic ECN has a beta of 50, it is only
24086 		 * ABE Ecn that uses this "less" value, but we do too with pacing :)
24087 		 */
24088 
24089 	case TCP_RACK_PACING_BETA_ECN:
24090 		if (strcmp(tp->t_cc->name, CCALGONAME_NEWRENO) != 0)
24091 			error = EINVAL;
24092 		else if (rack->rc_pacing_cc_set == 0)
24093 			optval = rack->r_ctl.rc_saved_beta.beta_ecn;
24094 		else {
24095 			/*
24096 			 * Reach out into the CC data and report back what
24097 			 * I have previously set. Yeah it looks hackish but
24098 			 * we don't want to report the saved values.
24099 			 */
24100 			if (tp->t_ccv.cc_data)
24101 				optval = ((struct newreno *)tp->t_ccv.cc_data)->beta_ecn;
24102 			else
24103 				error = EINVAL;
24104 		}
24105 		break;
24106 	case TCP_RACK_DSACK_OPT:
24107 		optval = 0;
24108 		if (rack->rc_rack_tmr_std_based) {
24109 			optval |= 1;
24110 		}
24111 		if (rack->rc_rack_use_dsack) {
24112 			optval |= 2;
24113 		}
24114 		break;
24115  	case TCP_RACK_ENABLE_HYSTART:
24116 	{
24117 		if (tp->t_ccv.flags & CCF_HYSTART_ALLOWED) {
24118 			optval = RACK_HYSTART_ON;
24119 			if (tp->t_ccv.flags & CCF_HYSTART_CAN_SH_CWND)
24120 				optval = RACK_HYSTART_ON_W_SC;
24121 			if (tp->t_ccv.flags & CCF_HYSTART_CONS_SSTH)
24122 				optval = RACK_HYSTART_ON_W_SC_C;
24123 		} else {
24124 			optval = RACK_HYSTART_OFF;
24125 		}
24126 	}
24127 	break;
24128 	case TCP_RACK_DGP_IN_REC:
24129 		optval = rack->r_ctl.full_dgp_in_rec;
24130 		break;
24131 	case TCP_RACK_HI_BETA:
24132 		optval = rack->rack_hibeta;
24133 		break;
24134 	case TCP_RXT_CLAMP:
24135 		optval = rack->r_ctl.saved_rxt_clamp_val;
24136 		break;
24137 	case TCP_DEFER_OPTIONS:
24138 		optval = rack->defer_options;
24139 		break;
24140 	case TCP_RACK_MEASURE_CNT:
24141 		optval = rack->r_ctl.req_measurements;
24142 		break;
24143 	case TCP_REC_ABC_VAL:
24144 		optval = rack->r_use_labc_for_rec;
24145 		break;
24146 	case TCP_RACK_ABC_VAL:
24147 		optval = rack->rc_labc;
24148 		break;
24149 	case TCP_HDWR_UP_ONLY:
24150 		optval= rack->r_up_only;
24151 		break;
24152 	case TCP_PACING_RATE_CAP:
24153 		loptval = rack->r_ctl.bw_rate_cap;
24154 		break;
24155 	case TCP_RACK_PROFILE:
24156 		/* You cannot retrieve a profile, its write only */
24157 		error = EINVAL;
24158 		break;
24159 	case TCP_HYBRID_PACING:
24160 		/* You cannot retrieve hybrid pacing information, its write only */
24161 		error = EINVAL;
24162 		break;
24163 	case TCP_USE_CMP_ACKS:
24164 		optval = rack->r_use_cmp_ack;
24165 		break;
24166 	case TCP_RACK_PACE_TO_FILL:
24167 		optval = rack->rc_pace_to_cwnd;
24168 		if (optval && rack->r_fill_less_agg)
24169 			optval++;
24170 		break;
24171 	case TCP_RACK_NO_PUSH_AT_MAX:
24172 		optval = rack->r_ctl.rc_no_push_at_mrtt;
24173 		break;
24174 	case TCP_SHARED_CWND_ENABLE:
24175 		optval = rack->rack_enable_scwnd;
24176 		break;
24177 	case TCP_RACK_NONRXT_CFG_RATE:
24178 		optval = rack->rack_rec_nonrxt_use_cr;
24179 		break;
24180 	case TCP_NO_PRR:
24181 		if (rack->rack_no_prr  == 1)
24182 			optval = 1;
24183 		else if (rack->no_prr_addback == 1)
24184 			optval = 2;
24185 		else
24186 			optval = 0;
24187 		break;
24188 	case TCP_RACK_DO_DETECTION:
24189 		optval = rack->do_detection;
24190 		break;
24191 	case TCP_RACK_MBUF_QUEUE:
24192 		/* Now do we use the LRO mbuf-queue feature */
24193 		optval = rack->r_mbuf_queue;
24194 		break;
24195 	case TCP_TIMELY_DYN_ADJ:
24196 		optval = rack->rc_gp_dyn_mul;
24197 		break;
24198 	case TCP_BBR_IWINTSO:
24199 		optval = rack->rc_init_win;
24200 		break;
24201 	case TCP_RACK_TLP_REDUCE:
24202 		/* RACK TLP cwnd reduction (bool) */
24203 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
24204 		break;
24205 	case TCP_BBR_RACK_INIT_RATE:
24206 		val = rack->r_ctl.init_rate;
24207 		/* convert to kbits per sec */
24208 		val *= 8;
24209 		val /= 1000;
24210 		optval = (uint32_t)val;
24211 		break;
24212 	case TCP_RACK_FORCE_MSEG:
24213 		optval = rack->rc_force_max_seg;
24214 		break;
24215 	case TCP_RACK_PACE_MIN_SEG:
24216 		optval = rack->r_ctl.rc_user_set_min_segs;
24217 		break;
24218 	case TCP_RACK_PACE_MAX_SEG:
24219 		/* Max segments in a pace */
24220 		optval = rack->rc_user_set_max_segs;
24221 		break;
24222 	case TCP_RACK_PACE_ALWAYS:
24223 		/* Use the always pace method */
24224 		optval = rack->rc_always_pace;
24225 		break;
24226 	case TCP_RACK_PRR_SENDALOT:
24227 		/* Allow PRR to send more than one seg */
24228 		optval = rack->r_ctl.rc_prr_sendalot;
24229 		break;
24230 	case TCP_RACK_MIN_TO:
24231 		/* Minimum time between rack t-o's in ms */
24232 		optval = rack->r_ctl.rc_min_to;
24233 		break;
24234 	case TCP_RACK_SPLIT_LIMIT:
24235 		optval = rack->r_ctl.rc_split_limit;
24236 		break;
24237 	case TCP_RACK_EARLY_SEG:
24238 		/* If early recovery max segments */
24239 		optval = rack->r_ctl.rc_early_recovery_segs;
24240 		break;
24241 	case TCP_RACK_REORD_THRESH:
24242 		/* RACK reorder threshold (shift amount) */
24243 		optval = rack->r_ctl.rc_reorder_shift;
24244 		break;
24245 	case TCP_RACK_REORD_FADE:
24246 		/* Does reordering fade after ms time */
24247 		optval = rack->r_ctl.rc_reorder_fade;
24248 		break;
24249 	case TCP_BBR_USE_RACK_RR:
24250 		/* Do we use the rack cheat for rxt */
24251 		optval = rack->use_rack_rr;
24252 		break;
24253 	case TCP_RACK_RR_CONF:
24254 		optval = rack->r_rr_config;
24255 		break;
24256 	case TCP_HDWR_RATE_CAP:
24257 		optval = rack->r_rack_hw_rate_caps;
24258 		break;
24259 	case TCP_BBR_HDWR_PACE:
24260 		optval = rack->rack_hdw_pace_ena;
24261 		break;
24262 	case TCP_RACK_TLP_THRESH:
24263 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
24264 		optval = rack->r_ctl.rc_tlp_threshold;
24265 		break;
24266 	case TCP_RACK_PKT_DELAY:
24267 		/* RACK added ms i.e. rack-rtt + reord + N */
24268 		optval = rack->r_ctl.rc_pkt_delay;
24269 		break;
24270 	case TCP_RACK_TLP_USE:
24271 		optval = rack->rack_tlp_threshold_use;
24272 		break;
24273 	case TCP_PACING_DND:
24274 		optval = rack->rc_pace_dnd;
24275 		break;
24276 	case TCP_RACK_PACE_RATE_CA:
24277 		optval = rack->r_ctl.rc_fixed_pacing_rate_ca;
24278 		break;
24279 	case TCP_RACK_PACE_RATE_SS:
24280 		optval = rack->r_ctl.rc_fixed_pacing_rate_ss;
24281 		break;
24282 	case TCP_RACK_PACE_RATE_REC:
24283 		optval = rack->r_ctl.rc_fixed_pacing_rate_rec;
24284 		break;
24285 	case TCP_RACK_GP_INCREASE_SS:
24286 		optval = rack->r_ctl.rack_per_of_gp_ca;
24287 		break;
24288 	case TCP_RACK_GP_INCREASE_CA:
24289 		optval = rack->r_ctl.rack_per_of_gp_ss;
24290 		break;
24291 	case TCP_RACK_PACING_DIVISOR:
24292 		optval = rack->r_ctl.pace_len_divisor;
24293 		break;
24294 	case TCP_BBR_RACK_RTT_USE:
24295 		optval = rack->r_ctl.rc_rate_sample_method;
24296 		break;
24297 	case TCP_DELACK:
24298 		optval = tp->t_delayed_ack;
24299 		break;
24300 	case TCP_DATA_AFTER_CLOSE:
24301 		optval = rack->rc_allow_data_af_clo;
24302 		break;
24303 	case TCP_SHARED_CWND_TIME_LIMIT:
24304 		optval = rack->r_limit_scw;
24305 		break;
24306 	case TCP_RACK_TIMER_SLOP:
24307 		optval = rack->r_ctl.timer_slop;
24308 		break;
24309 	default:
24310 		return (tcp_default_ctloutput(tp, sopt));
24311 		break;
24312 	}
24313 	INP_WUNLOCK(inp);
24314 	if (error == 0) {
24315 		if (TCP_PACING_RATE_CAP)
24316 			error = sooptcopyout(sopt, &loptval, sizeof loptval);
24317 		else
24318 			error = sooptcopyout(sopt, &optval, sizeof optval);
24319 	}
24320 	return (error);
24321 }
24322 
24323 static int
24324 rack_ctloutput(struct tcpcb *tp, struct sockopt *sopt)
24325 {
24326 	if (sopt->sopt_dir == SOPT_SET) {
24327 		return (rack_set_sockopt(tp, sopt));
24328 	} else if (sopt->sopt_dir == SOPT_GET) {
24329 		return (rack_get_sockopt(tp, sopt));
24330 	} else {
24331 		panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir);
24332 	}
24333 }
24334 
24335 static const char *rack_stack_names[] = {
24336 	__XSTRING(STACKNAME),
24337 #ifdef STACKALIAS
24338 	__XSTRING(STACKALIAS),
24339 #endif
24340 };
24341 
24342 static int
24343 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
24344 {
24345 	memset(mem, 0, size);
24346 	return (0);
24347 }
24348 
24349 static void
24350 rack_dtor(void *mem, int32_t size, void *arg)
24351 {
24352 
24353 }
24354 
24355 static bool rack_mod_inited = false;
24356 
24357 static int
24358 tcp_addrack(module_t mod, int32_t type, void *data)
24359 {
24360 	int32_t err = 0;
24361 	int num_stacks;
24362 
24363 	switch (type) {
24364 	case MOD_LOAD:
24365 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
24366 		    sizeof(struct rack_sendmap),
24367 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
24368 
24369 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
24370 		    sizeof(struct tcp_rack),
24371 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
24372 
24373 		sysctl_ctx_init(&rack_sysctl_ctx);
24374 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
24375 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
24376 		    OID_AUTO,
24377 #ifdef STACKALIAS
24378 		    __XSTRING(STACKALIAS),
24379 #else
24380 		    __XSTRING(STACKNAME),
24381 #endif
24382 		    CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
24383 		    "");
24384 		if (rack_sysctl_root == NULL) {
24385 			printf("Failed to add sysctl node\n");
24386 			err = EFAULT;
24387 			goto free_uma;
24388 		}
24389 		rack_init_sysctls();
24390 		num_stacks = nitems(rack_stack_names);
24391 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
24392 		    rack_stack_names, &num_stacks);
24393 		if (err) {
24394 			printf("Failed to register %s stack name for "
24395 			    "%s module\n", rack_stack_names[num_stacks],
24396 			    __XSTRING(MODNAME));
24397 			sysctl_ctx_free(&rack_sysctl_ctx);
24398 free_uma:
24399 			uma_zdestroy(rack_zone);
24400 			uma_zdestroy(rack_pcb_zone);
24401 			rack_counter_destroy();
24402 			printf("Failed to register rack module -- err:%d\n", err);
24403 			return (err);
24404 		}
24405 		tcp_lro_reg_mbufq();
24406 		rack_mod_inited = true;
24407 		break;
24408 	case MOD_QUIESCE:
24409 		err = deregister_tcp_functions(&__tcp_rack, true, false);
24410 		break;
24411 	case MOD_UNLOAD:
24412 		err = deregister_tcp_functions(&__tcp_rack, false, true);
24413 		if (err == EBUSY)
24414 			break;
24415 		if (rack_mod_inited) {
24416 			uma_zdestroy(rack_zone);
24417 			uma_zdestroy(rack_pcb_zone);
24418 			sysctl_ctx_free(&rack_sysctl_ctx);
24419 			rack_counter_destroy();
24420 			rack_mod_inited = false;
24421 		}
24422 		tcp_lro_dereg_mbufq();
24423 		err = 0;
24424 		break;
24425 	default:
24426 		return (EOPNOTSUPP);
24427 	}
24428 	return (err);
24429 }
24430 
24431 static moduledata_t tcp_rack = {
24432 	.name = __XSTRING(MODNAME),
24433 	.evhand = tcp_addrack,
24434 	.priv = 0
24435 };
24436 
24437 MODULE_VERSION(MODNAME, 1);
24438 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
24439 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
24440 
24441 #endif /* #if !defined(INET) && !defined(INET6) */
24442