xref: /freebsd/sys/netinet/tcp_stacks/rack.c (revision e4c66ddabdb470bab319705c1834a4867c508a43)
1 /*-
2  * Copyright (c) 2016-2018
3  *	Netflix Inc.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  *
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_ipsec.h"
34 #include "opt_tcpdebug.h"
35 
36 #include <sys/param.h>
37 #include <sys/module.h>
38 #include <sys/kernel.h>
39 #ifdef TCP_HHOOK
40 #include <sys/hhook.h>
41 #endif
42 #include <sys/lock.h>
43 #include <sys/malloc.h>
44 #include <sys/lock.h>
45 #include <sys/mutex.h>
46 #include <sys/mbuf.h>
47 #include <sys/proc.h>		/* for proc0 declaration */
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
51 #include <sys/systm.h>
52 #ifdef NETFLIX_STATS
53 #include <sys/stats.h>
54 #endif
55 #include <sys/refcount.h>
56 #include <sys/queue.h>
57 #include <sys/smp.h>
58 #include <sys/kthread.h>
59 #include <sys/kern_prefetch.h>
60 
61 #include <vm/uma.h>
62 
63 #include <net/route.h>
64 #include <net/vnet.h>
65 
66 #define TCPSTATES		/* for logging */
67 
68 #include <netinet/in.h>
69 #include <netinet/in_kdtrace.h>
70 #include <netinet/in_pcb.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip_icmp.h>	/* required for icmp_var.h */
73 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM */
74 #include <netinet/ip_var.h>
75 #include <netinet/ip6.h>
76 #include <netinet6/in6_pcb.h>
77 #include <netinet6/ip6_var.h>
78 #include <netinet/tcp.h>
79 #define	TCPOUTFLAGS
80 #include <netinet/tcp_fsm.h>
81 #include <netinet/tcp_log_buf.h>
82 #include <netinet/tcp_seq.h>
83 #include <netinet/tcp_timer.h>
84 #include <netinet/tcp_var.h>
85 #include <netinet/tcp_hpts.h>
86 #include <netinet/tcpip.h>
87 #include <netinet/cc/cc.h>
88 #ifdef NETFLIX_CWV
89 #include <netinet/tcp_newcwv.h>
90 #endif
91 #include <netinet/tcp_fastopen.h>
92 #ifdef TCPDEBUG
93 #include <netinet/tcp_debug.h>
94 #endif				/* TCPDEBUG */
95 #ifdef TCP_OFFLOAD
96 #include <netinet/tcp_offload.h>
97 #endif
98 #ifdef INET6
99 #include <netinet6/tcp6_var.h>
100 #endif
101 
102 #include <netipsec/ipsec_support.h>
103 
104 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
105 #include <netipsec/ipsec.h>
106 #include <netipsec/ipsec6.h>
107 #endif				/* IPSEC */
108 
109 #include <netinet/udp.h>
110 #include <netinet/udp_var.h>
111 #include <machine/in_cksum.h>
112 
113 #ifdef MAC
114 #include <security/mac/mac_framework.h>
115 #endif
116 #include "sack_filter.h"
117 #include "tcp_rack.h"
118 #include "rack_bbr_common.h"
119 
120 uma_zone_t rack_zone;
121 uma_zone_t rack_pcb_zone;
122 
123 #ifndef TICKS2SBT
124 #define	TICKS2SBT(__t)	(tick_sbt * ((sbintime_t)(__t)))
125 #endif
126 
127 struct sysctl_ctx_list rack_sysctl_ctx;
128 struct sysctl_oid *rack_sysctl_root;
129 
130 #define CUM_ACKED 1
131 #define SACKED 2
132 
133 /*
134  * The RACK module incorporates a number of
135  * TCP ideas that have been put out into the IETF
136  * over the last few years:
137  * - Matt Mathis's Rate Halving which slowly drops
138  *    the congestion window so that the ack clock can
139  *    be maintained during a recovery.
140  * - Yuchung Cheng's RACK TCP (for which its named) that
141  *    will stop us using the number of dup acks and instead
142  *    use time as the gage of when we retransmit.
143  * - Reorder Detection of RFC4737 and the Tail-Loss probe draft
144  *    of Dukkipati et.al.
145  * RACK depends on SACK, so if an endpoint arrives that
146  * cannot do SACK the state machine below will shuttle the
147  * connection back to using the "default" TCP stack that is
148  * in FreeBSD.
149  *
150  * To implement RACK the original TCP stack was first decomposed
151  * into a functional state machine with individual states
152  * for each of the possible TCP connection states. The do_segement
153  * functions role in life is to mandate the connection supports SACK
154  * initially and then assure that the RACK state matches the conenction
155  * state before calling the states do_segment function. Each
156  * state is simplified due to the fact that the original do_segment
157  * has been decomposed and we *know* what state we are in (no
158  * switches on the state) and all tests for SACK are gone. This
159  * greatly simplifies what each state does.
160  *
161  * TCP output is also over-written with a new version since it
162  * must maintain the new rack scoreboard.
163  *
164  */
165 static int32_t rack_precache = 1;
166 static int32_t rack_tlp_thresh = 1;
167 static int32_t rack_reorder_thresh = 2;
168 static int32_t rack_reorder_fade = 60000;	/* 0 - never fade, def 60,000
169 						 * - 60 seconds */
170 static int32_t rack_pkt_delay = 1;
171 static int32_t rack_inc_var = 0;/* For TLP */
172 static int32_t rack_reduce_largest_on_idle = 0;
173 static int32_t rack_min_pace_time = 0;
174 static int32_t rack_min_pace_time_seg_req=6;
175 static int32_t rack_early_recovery = 1;
176 static int32_t rack_early_recovery_max_seg = 6;
177 static int32_t rack_send_a_lot_in_prr = 1;
178 static int32_t rack_min_to = 1;	/* Number of ms minimum timeout */
179 static int32_t rack_tlp_in_recovery = 1;	/* Can we do TLP in recovery? */
180 static int32_t rack_verbose_logging = 0;
181 static int32_t rack_ignore_data_after_close = 1;
182 /*
183  * Currently regular tcp has a rto_min of 30ms
184  * the backoff goes 12 times so that ends up
185  * being a total of 122.850 seconds before a
186  * connection is killed.
187  */
188 static int32_t rack_tlp_min = 10;
189 static int32_t rack_rto_min = 30;	/* 30ms same as main freebsd */
190 static int32_t rack_rto_max = 30000;	/* 30 seconds */
191 static const int32_t rack_free_cache = 2;
192 static int32_t rack_hptsi_segments = 40;
193 static int32_t rack_rate_sample_method = USE_RTT_LOW;
194 static int32_t rack_pace_every_seg = 1;
195 static int32_t rack_delayed_ack_time = 200;	/* 200ms */
196 static int32_t rack_slot_reduction = 4;
197 static int32_t rack_lower_cwnd_at_tlp = 0;
198 static int32_t rack_use_proportional_reduce = 0;
199 static int32_t rack_proportional_rate = 10;
200 static int32_t rack_tlp_max_resend = 2;
201 static int32_t rack_limited_retran = 0;
202 static int32_t rack_always_send_oldest = 0;
203 static int32_t rack_sack_block_limit = 128;
204 static int32_t rack_use_sack_filter = 1;
205 static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE;
206 
207 /* Rack specific counters */
208 counter_u64_t rack_badfr;
209 counter_u64_t rack_badfr_bytes;
210 counter_u64_t rack_rtm_prr_retran;
211 counter_u64_t rack_rtm_prr_newdata;
212 counter_u64_t rack_timestamp_mismatch;
213 counter_u64_t rack_reorder_seen;
214 counter_u64_t rack_paced_segments;
215 counter_u64_t rack_unpaced_segments;
216 counter_u64_t rack_saw_enobuf;
217 counter_u64_t rack_saw_enetunreach;
218 
219 /* Tail loss probe counters */
220 counter_u64_t rack_tlp_tot;
221 counter_u64_t rack_tlp_newdata;
222 counter_u64_t rack_tlp_retran;
223 counter_u64_t rack_tlp_retran_bytes;
224 counter_u64_t rack_tlp_retran_fail;
225 counter_u64_t rack_to_tot;
226 counter_u64_t rack_to_arm_rack;
227 counter_u64_t rack_to_arm_tlp;
228 counter_u64_t rack_to_alloc;
229 counter_u64_t rack_to_alloc_hard;
230 counter_u64_t rack_to_alloc_emerg;
231 
232 counter_u64_t rack_sack_proc_all;
233 counter_u64_t rack_sack_proc_short;
234 counter_u64_t rack_sack_proc_restart;
235 counter_u64_t rack_runt_sacks;
236 counter_u64_t rack_used_tlpmethod;
237 counter_u64_t rack_used_tlpmethod2;
238 counter_u64_t rack_enter_tlp_calc;
239 counter_u64_t rack_input_idle_reduces;
240 counter_u64_t rack_tlp_does_nada;
241 
242 /* Temp CPU counters */
243 counter_u64_t rack_find_high;
244 
245 counter_u64_t rack_progress_drops;
246 counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE];
247 counter_u64_t rack_opts_arry[RACK_OPTS_SIZE];
248 
249 static void
250 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line);
251 
252 static int
253 rack_process_ack(struct mbuf *m, struct tcphdr *th,
254     struct socket *so, struct tcpcb *tp, struct tcpopt *to,
255     uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val);
256 static int
257 rack_process_data(struct mbuf *m, struct tcphdr *th,
258     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
259     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
260 static void
261 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack,
262     struct tcphdr *th, uint16_t nsegs, uint16_t type, int32_t recovery);
263 static struct rack_sendmap *rack_alloc(struct tcp_rack *rack);
264 static struct rack_sendmap *
265 rack_check_recovery_mode(struct tcpcb *tp,
266     uint32_t tsused);
267 static void
268 rack_cong_signal(struct tcpcb *tp, struct tcphdr *th,
269     uint32_t type);
270 static void rack_counter_destroy(void);
271 static int
272 rack_ctloutput(struct socket *so, struct sockopt *sopt,
273     struct inpcb *inp, struct tcpcb *tp);
274 static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how);
275 static void
276 rack_do_segment(struct mbuf *m, struct tcphdr *th,
277     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
278     uint8_t iptos);
279 static void rack_dtor(void *mem, int32_t size, void *arg);
280 static void
281 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm,
282     uint32_t t, uint32_t cts);
283 static struct rack_sendmap *
284 rack_find_high_nonack(struct tcp_rack *rack,
285     struct rack_sendmap *rsm);
286 static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack);
287 static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm);
288 static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged);
289 static int
290 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
291     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
292 static int32_t rack_handoff_ok(struct tcpcb *tp);
293 static int32_t rack_init(struct tcpcb *tp);
294 static void rack_init_sysctls(void);
295 static void
296 rack_log_ack(struct tcpcb *tp, struct tcpopt *to,
297     struct tcphdr *th);
298 static void
299 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
300     uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts,
301     uint8_t pass, struct rack_sendmap *hintrsm);
302 static void
303 rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack,
304     struct rack_sendmap *rsm);
305 static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num);
306 static int32_t rack_output(struct tcpcb *tp);
307 static void
308 rack_hpts_do_segment(struct mbuf *m, struct tcphdr *th,
309     struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
310     uint8_t iptos, int32_t nxt_pkt, struct timeval *tv);
311 
312 static uint32_t
313 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack,
314     struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm,
315     uint32_t cts);
316 static void rack_post_recovery(struct tcpcb *tp, struct tcphdr *th);
317 static void rack_remxt_tmr(struct tcpcb *tp);
318 static int
319 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
320     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack);
321 static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack);
322 static int32_t rack_stopall(struct tcpcb *tp);
323 static void
324 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type,
325     uint32_t delta);
326 static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type);
327 static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line);
328 static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type);
329 static uint32_t
330 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
331     struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp);
332 static void
333 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
334     struct rack_sendmap *rsm, uint32_t ts);
335 static int
336 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
337     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type);
338 static int32_t tcp_addrack(module_t mod, int32_t type, void *data);
339 static void
340 rack_challenge_ack(struct mbuf *m, struct tcphdr *th,
341     struct tcpcb *tp, int32_t * ret_val);
342 static int
343 rack_do_close_wait(struct mbuf *m, struct tcphdr *th,
344     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
345     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
346 static int
347 rack_do_closing(struct mbuf *m, struct tcphdr *th,
348     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
349     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
350 static void
351 rack_do_drop(struct mbuf *m, struct tcpcb *tp);
352 static void
353 rack_do_dropafterack(struct mbuf *m, struct tcpcb *tp,
354     struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t * ret_val);
355 static void
356 rack_do_dropwithreset(struct mbuf *m, struct tcpcb *tp,
357 	struct tcphdr *th, int32_t rstreason, int32_t tlen);
358 static int
359 rack_do_established(struct mbuf *m, struct tcphdr *th,
360     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
361     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
362 static int
363 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th,
364     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
365     int32_t tlen, uint32_t tiwin, int32_t nxt_pkt);
366 static int
367 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th,
368     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
369     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
370 static int
371 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th,
372     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
373     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
374 static int
375 rack_do_lastack(struct mbuf *m, struct tcphdr *th,
376     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
377     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
378 static int
379 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th,
380     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
381     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
382 static int
383 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th,
384     struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen,
385     int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt);
386 static int
387 rack_drop_checks(struct tcpopt *to, struct mbuf *m,
388     struct tcphdr *th, struct tcpcb *tp, int32_t * tlenp, int32_t * thf,
389     int32_t * drop_hdrlen, int32_t * ret_val);
390 static int
391 rack_process_rst(struct mbuf *m, struct tcphdr *th,
392     struct socket *so, struct tcpcb *tp);
393 struct rack_sendmap *
394 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack,
395     uint32_t tsused);
396 static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt);
397 static void
398      tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th);
399 
400 static int
401 rack_ts_check(struct mbuf *m, struct tcphdr *th,
402     struct tcpcb *tp, int32_t tlen, int32_t thflags, int32_t * ret_val);
403 
404 int32_t rack_clear_counter=0;
405 
406 
407 static int
408 sysctl_rack_clear(SYSCTL_HANDLER_ARGS)
409 {
410 	uint32_t stat;
411 	int32_t error;
412 
413 	error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t));
414 	if (error || req->newptr == NULL)
415 		return error;
416 
417 	error = SYSCTL_IN(req, &stat, sizeof(uint32_t));
418 	if (error)
419 		return (error);
420 	if (stat == 1) {
421 #ifdef INVARIANTS
422 		printf("Clearing RACK counters\n");
423 #endif
424 		counter_u64_zero(rack_badfr);
425 		counter_u64_zero(rack_badfr_bytes);
426 		counter_u64_zero(rack_rtm_prr_retran);
427 		counter_u64_zero(rack_rtm_prr_newdata);
428 		counter_u64_zero(rack_timestamp_mismatch);
429 		counter_u64_zero(rack_reorder_seen);
430 		counter_u64_zero(rack_tlp_tot);
431 		counter_u64_zero(rack_tlp_newdata);
432 		counter_u64_zero(rack_tlp_retran);
433 		counter_u64_zero(rack_tlp_retran_bytes);
434 		counter_u64_zero(rack_tlp_retran_fail);
435 		counter_u64_zero(rack_to_tot);
436 		counter_u64_zero(rack_to_arm_rack);
437 		counter_u64_zero(rack_to_arm_tlp);
438 		counter_u64_zero(rack_paced_segments);
439 		counter_u64_zero(rack_unpaced_segments);
440 		counter_u64_zero(rack_saw_enobuf);
441 		counter_u64_zero(rack_saw_enetunreach);
442 		counter_u64_zero(rack_to_alloc_hard);
443 		counter_u64_zero(rack_to_alloc_emerg);
444 		counter_u64_zero(rack_sack_proc_all);
445 		counter_u64_zero(rack_sack_proc_short);
446 		counter_u64_zero(rack_sack_proc_restart);
447 		counter_u64_zero(rack_to_alloc);
448 		counter_u64_zero(rack_find_high);
449 		counter_u64_zero(rack_runt_sacks);
450 		counter_u64_zero(rack_used_tlpmethod);
451 		counter_u64_zero(rack_used_tlpmethod2);
452 		counter_u64_zero(rack_enter_tlp_calc);
453 		counter_u64_zero(rack_progress_drops);
454 		counter_u64_zero(rack_tlp_does_nada);
455 	}
456 	rack_clear_counter = 0;
457 	return (0);
458 }
459 
460 
461 
462 static void
463 rack_init_sysctls()
464 {
465 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
466 	    SYSCTL_CHILDREN(rack_sysctl_root),
467 	    OID_AUTO, "rate_sample_method", CTLFLAG_RW,
468 	    &rack_rate_sample_method , USE_RTT_LOW,
469 	    "What method should we use for rate sampling 0=high, 1=low ");
470 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
471 	    SYSCTL_CHILDREN(rack_sysctl_root),
472 	    OID_AUTO, "data_after_close", CTLFLAG_RW,
473 	    &rack_ignore_data_after_close, 0,
474 	    "Do we hold off sending a RST until all pending data is ack'd");
475 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
476 	    SYSCTL_CHILDREN(rack_sysctl_root),
477 	    OID_AUTO, "tlpmethod", CTLFLAG_RW,
478 	    &rack_tlp_threshold_use, TLP_USE_TWO_ONE,
479 	    "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2");
480 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
481 	    SYSCTL_CHILDREN(rack_sysctl_root),
482 	    OID_AUTO, "min_pace_time", CTLFLAG_RW,
483 	    &rack_min_pace_time, 0,
484 	    "Should we enforce a minimum pace time of 1ms");
485 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
486 	    SYSCTL_CHILDREN(rack_sysctl_root),
487 	    OID_AUTO, "min_pace_segs", CTLFLAG_RW,
488 	    &rack_min_pace_time_seg_req, 6,
489 	    "How many segments have to be in the len to enforce min-pace-time");
490 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
491 	    SYSCTL_CHILDREN(rack_sysctl_root),
492 	    OID_AUTO, "idle_reduce_high", CTLFLAG_RW,
493 	    &rack_reduce_largest_on_idle, 0,
494 	    "Should we reduce the largest cwnd seen to IW on idle reduction");
495 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
496 	    SYSCTL_CHILDREN(rack_sysctl_root),
497 	    OID_AUTO, "bb_verbose", CTLFLAG_RW,
498 	    &rack_verbose_logging, 0,
499 	    "Should RACK black box logging be verbose");
500 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
501 	    SYSCTL_CHILDREN(rack_sysctl_root),
502 	    OID_AUTO, "sackfiltering", CTLFLAG_RW,
503 	    &rack_use_sack_filter, 1,
504 	    "Do we use sack filtering?");
505 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
506 	    SYSCTL_CHILDREN(rack_sysctl_root),
507 	    OID_AUTO, "delayed_ack", CTLFLAG_RW,
508 	    &rack_delayed_ack_time, 200,
509 	    "Delayed ack time (200ms)");
510 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
511 	    SYSCTL_CHILDREN(rack_sysctl_root),
512 	    OID_AUTO, "tlpminto", CTLFLAG_RW,
513 	    &rack_tlp_min, 10,
514 	    "TLP minimum timeout per the specification (10ms)");
515 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
516 	    SYSCTL_CHILDREN(rack_sysctl_root),
517 	    OID_AUTO, "precache", CTLFLAG_RW,
518 	    &rack_precache, 0,
519 	    "Where should we precache the mcopy (0 is not at all)");
520 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
521 	    SYSCTL_CHILDREN(rack_sysctl_root),
522 	    OID_AUTO, "sblklimit", CTLFLAG_RW,
523 	    &rack_sack_block_limit, 128,
524 	    "When do we start paying attention to small sack blocks");
525 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
526 	    SYSCTL_CHILDREN(rack_sysctl_root),
527 	    OID_AUTO, "send_oldest", CTLFLAG_RW,
528 	    &rack_always_send_oldest, 1,
529 	    "Should we always send the oldest TLP and RACK-TLP");
530 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
531 	    SYSCTL_CHILDREN(rack_sysctl_root),
532 	    OID_AUTO, "rack_tlp_in_recovery", CTLFLAG_RW,
533 	    &rack_tlp_in_recovery, 1,
534 	    "Can we do a TLP during recovery?");
535 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
536 	    SYSCTL_CHILDREN(rack_sysctl_root),
537 	    OID_AUTO, "rack_tlimit", CTLFLAG_RW,
538 	    &rack_limited_retran, 0,
539 	    "How many times can a rack timeout drive out sends");
540 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
541 	    SYSCTL_CHILDREN(rack_sysctl_root),
542 	    OID_AUTO, "minrto", CTLFLAG_RW,
543 	    &rack_rto_min, 0,
544 	    "Minimum RTO in ms -- set with caution below 1000 due to TLP");
545 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
546 	    SYSCTL_CHILDREN(rack_sysctl_root),
547 	    OID_AUTO, "maxrto", CTLFLAG_RW,
548 	    &rack_rto_max, 0,
549 	    "Maxiumum RTO in ms -- should be at least as large as min_rto");
550 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
551 	    SYSCTL_CHILDREN(rack_sysctl_root),
552 	    OID_AUTO, "tlp_retry", CTLFLAG_RW,
553 	    &rack_tlp_max_resend, 2,
554 	    "How many times does TLP retry a single segment or multiple with no ACK");
555 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
556 	    SYSCTL_CHILDREN(rack_sysctl_root),
557 	    OID_AUTO, "recovery_loss_prop", CTLFLAG_RW,
558 	    &rack_use_proportional_reduce, 0,
559 	    "Should we proportionaly reduce cwnd based on the number of losses ");
560 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
561 	    SYSCTL_CHILDREN(rack_sysctl_root),
562 	    OID_AUTO, "recovery_prop", CTLFLAG_RW,
563 	    &rack_proportional_rate, 10,
564 	    "What percent reduction per loss");
565 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
566 	    SYSCTL_CHILDREN(rack_sysctl_root),
567 	    OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW,
568 	    &rack_lower_cwnd_at_tlp, 0,
569 	    "When a TLP completes a retran should we enter recovery?");
570 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
571 	    SYSCTL_CHILDREN(rack_sysctl_root),
572 	    OID_AUTO, "hptsi_reduces", CTLFLAG_RW,
573 	    &rack_slot_reduction, 4,
574 	    "When setting a slot should we reduce by divisor");
575 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
576 	    SYSCTL_CHILDREN(rack_sysctl_root),
577 	    OID_AUTO, "hptsi_every_seg", CTLFLAG_RW,
578 	    &rack_pace_every_seg, 1,
579 	    "Should we pace out every segment hptsi");
580 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
581 	    SYSCTL_CHILDREN(rack_sysctl_root),
582 	    OID_AUTO, "hptsi_seg_max", CTLFLAG_RW,
583 	    &rack_hptsi_segments, 6,
584 	    "Should we pace out only a limited size of segments");
585 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
586 	    SYSCTL_CHILDREN(rack_sysctl_root),
587 	    OID_AUTO, "prr_sendalot", CTLFLAG_RW,
588 	    &rack_send_a_lot_in_prr, 1,
589 	    "Send a lot in prr");
590 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
591 	    SYSCTL_CHILDREN(rack_sysctl_root),
592 	    OID_AUTO, "minto", CTLFLAG_RW,
593 	    &rack_min_to, 1,
594 	    "Minimum rack timeout in milliseconds");
595 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
596 	    SYSCTL_CHILDREN(rack_sysctl_root),
597 	    OID_AUTO, "earlyrecoveryseg", CTLFLAG_RW,
598 	    &rack_early_recovery_max_seg, 6,
599 	    "Max segments in early recovery");
600 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
601 	    SYSCTL_CHILDREN(rack_sysctl_root),
602 	    OID_AUTO, "earlyrecovery", CTLFLAG_RW,
603 	    &rack_early_recovery, 1,
604 	    "Do we do early recovery with rack");
605 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
606 	    SYSCTL_CHILDREN(rack_sysctl_root),
607 	    OID_AUTO, "reorder_thresh", CTLFLAG_RW,
608 	    &rack_reorder_thresh, 2,
609 	    "What factor for rack will be added when seeing reordering (shift right)");
610 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
611 	    SYSCTL_CHILDREN(rack_sysctl_root),
612 	    OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW,
613 	    &rack_tlp_thresh, 1,
614 	    "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)");
615 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
616 	    SYSCTL_CHILDREN(rack_sysctl_root),
617 	    OID_AUTO, "reorder_fade", CTLFLAG_RW,
618 	    &rack_reorder_fade, 0,
619 	    "Does reorder detection fade, if so how many ms (0 means never)");
620 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
621 	    SYSCTL_CHILDREN(rack_sysctl_root),
622 	    OID_AUTO, "pktdelay", CTLFLAG_RW,
623 	    &rack_pkt_delay, 1,
624 	    "Extra RACK time (in ms) besides reordering thresh");
625 	SYSCTL_ADD_S32(&rack_sysctl_ctx,
626 	    SYSCTL_CHILDREN(rack_sysctl_root),
627 	    OID_AUTO, "inc_var", CTLFLAG_RW,
628 	    &rack_inc_var, 0,
629 	    "Should rack add to the TLP timer the variance in rtt calculation");
630 	rack_badfr = counter_u64_alloc(M_WAITOK);
631 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
632 	    SYSCTL_CHILDREN(rack_sysctl_root),
633 	    OID_AUTO, "badfr", CTLFLAG_RD,
634 	    &rack_badfr, "Total number of bad FRs");
635 	rack_badfr_bytes = counter_u64_alloc(M_WAITOK);
636 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
637 	    SYSCTL_CHILDREN(rack_sysctl_root),
638 	    OID_AUTO, "badfr_bytes", CTLFLAG_RD,
639 	    &rack_badfr_bytes, "Total number of bad FRs");
640 	rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK);
641 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
642 	    SYSCTL_CHILDREN(rack_sysctl_root),
643 	    OID_AUTO, "prrsndret", CTLFLAG_RD,
644 	    &rack_rtm_prr_retran,
645 	    "Total number of prr based retransmits");
646 	rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK);
647 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
648 	    SYSCTL_CHILDREN(rack_sysctl_root),
649 	    OID_AUTO, "prrsndnew", CTLFLAG_RD,
650 	    &rack_rtm_prr_newdata,
651 	    "Total number of prr based new transmits");
652 	rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK);
653 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
654 	    SYSCTL_CHILDREN(rack_sysctl_root),
655 	    OID_AUTO, "tsnf", CTLFLAG_RD,
656 	    &rack_timestamp_mismatch,
657 	    "Total number of timestamps that we could not find the reported ts");
658 	rack_find_high = counter_u64_alloc(M_WAITOK);
659 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
660 	    SYSCTL_CHILDREN(rack_sysctl_root),
661 	    OID_AUTO, "findhigh", CTLFLAG_RD,
662 	    &rack_find_high,
663 	    "Total number of FIN causing find-high");
664 	rack_reorder_seen = counter_u64_alloc(M_WAITOK);
665 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
666 	    SYSCTL_CHILDREN(rack_sysctl_root),
667 	    OID_AUTO, "reordering", CTLFLAG_RD,
668 	    &rack_reorder_seen,
669 	    "Total number of times we added delay due to reordering");
670 	rack_tlp_tot = counter_u64_alloc(M_WAITOK);
671 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
672 	    SYSCTL_CHILDREN(rack_sysctl_root),
673 	    OID_AUTO, "tlp_to_total", CTLFLAG_RD,
674 	    &rack_tlp_tot,
675 	    "Total number of tail loss probe expirations");
676 	rack_tlp_newdata = counter_u64_alloc(M_WAITOK);
677 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
678 	    SYSCTL_CHILDREN(rack_sysctl_root),
679 	    OID_AUTO, "tlp_new", CTLFLAG_RD,
680 	    &rack_tlp_newdata,
681 	    "Total number of tail loss probe sending new data");
682 
683 	rack_tlp_retran = counter_u64_alloc(M_WAITOK);
684 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
685 	    SYSCTL_CHILDREN(rack_sysctl_root),
686 	    OID_AUTO, "tlp_retran", CTLFLAG_RD,
687 	    &rack_tlp_retran,
688 	    "Total number of tail loss probe sending retransmitted data");
689 	rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK);
690 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
691 	    SYSCTL_CHILDREN(rack_sysctl_root),
692 	    OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD,
693 	    &rack_tlp_retran_bytes,
694 	    "Total bytes of tail loss probe sending retransmitted data");
695 	rack_tlp_retran_fail = counter_u64_alloc(M_WAITOK);
696 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
697 	    SYSCTL_CHILDREN(rack_sysctl_root),
698 	    OID_AUTO, "tlp_retran_fail", CTLFLAG_RD,
699 	    &rack_tlp_retran_fail,
700 	    "Total number of tail loss probe sending retransmitted data that failed (wait for t3)");
701 	rack_to_tot = counter_u64_alloc(M_WAITOK);
702 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
703 	    SYSCTL_CHILDREN(rack_sysctl_root),
704 	    OID_AUTO, "rack_to_tot", CTLFLAG_RD,
705 	    &rack_to_tot,
706 	    "Total number of times the rack to expired?");
707 	rack_to_arm_rack = counter_u64_alloc(M_WAITOK);
708 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
709 	    SYSCTL_CHILDREN(rack_sysctl_root),
710 	    OID_AUTO, "arm_rack", CTLFLAG_RD,
711 	    &rack_to_arm_rack,
712 	    "Total number of times the rack timer armed?");
713 	rack_to_arm_tlp = counter_u64_alloc(M_WAITOK);
714 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
715 	    SYSCTL_CHILDREN(rack_sysctl_root),
716 	    OID_AUTO, "arm_tlp", CTLFLAG_RD,
717 	    &rack_to_arm_tlp,
718 	    "Total number of times the tlp timer armed?");
719 	rack_paced_segments = counter_u64_alloc(M_WAITOK);
720 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
721 	    SYSCTL_CHILDREN(rack_sysctl_root),
722 	    OID_AUTO, "paced", CTLFLAG_RD,
723 	    &rack_paced_segments,
724 	    "Total number of times a segment send caused hptsi");
725 	rack_unpaced_segments = counter_u64_alloc(M_WAITOK);
726 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
727 	    SYSCTL_CHILDREN(rack_sysctl_root),
728 	    OID_AUTO, "unpaced", CTLFLAG_RD,
729 	    &rack_unpaced_segments,
730 	    "Total number of times a segment did not cause hptsi");
731 	rack_saw_enobuf = counter_u64_alloc(M_WAITOK);
732 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
733 	    SYSCTL_CHILDREN(rack_sysctl_root),
734 	    OID_AUTO, "saw_enobufs", CTLFLAG_RD,
735 	    &rack_saw_enobuf,
736 	    "Total number of times a segment did not cause hptsi");
737 	rack_saw_enetunreach = counter_u64_alloc(M_WAITOK);
738 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
739 	    SYSCTL_CHILDREN(rack_sysctl_root),
740 	    OID_AUTO, "saw_enetunreach", CTLFLAG_RD,
741 	    &rack_saw_enetunreach,
742 	    "Total number of times a segment did not cause hptsi");
743 	rack_to_alloc = counter_u64_alloc(M_WAITOK);
744 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
745 	    SYSCTL_CHILDREN(rack_sysctl_root),
746 	    OID_AUTO, "allocs", CTLFLAG_RD,
747 	    &rack_to_alloc,
748 	    "Total allocations of tracking structures");
749 	rack_to_alloc_hard = counter_u64_alloc(M_WAITOK);
750 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
751 	    SYSCTL_CHILDREN(rack_sysctl_root),
752 	    OID_AUTO, "allochard", CTLFLAG_RD,
753 	    &rack_to_alloc_hard,
754 	    "Total allocations done with sleeping the hard way");
755 	rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK);
756 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
757 	    SYSCTL_CHILDREN(rack_sysctl_root),
758 	    OID_AUTO, "allocemerg", CTLFLAG_RD,
759 	    &rack_to_alloc_emerg,
760 	    "Total alocations done from emergency cache");
761 	rack_sack_proc_all = counter_u64_alloc(M_WAITOK);
762 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
763 	    SYSCTL_CHILDREN(rack_sysctl_root),
764 	    OID_AUTO, "sack_long", CTLFLAG_RD,
765 	    &rack_sack_proc_all,
766 	    "Total times we had to walk whole list for sack processing");
767 
768 	rack_sack_proc_restart = counter_u64_alloc(M_WAITOK);
769 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
770 	    SYSCTL_CHILDREN(rack_sysctl_root),
771 	    OID_AUTO, "sack_restart", CTLFLAG_RD,
772 	    &rack_sack_proc_restart,
773 	    "Total times we had to walk whole list due to a restart");
774 	rack_sack_proc_short = counter_u64_alloc(M_WAITOK);
775 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
776 	    SYSCTL_CHILDREN(rack_sysctl_root),
777 	    OID_AUTO, "sack_short", CTLFLAG_RD,
778 	    &rack_sack_proc_short,
779 	    "Total times we took shortcut for sack processing");
780 	rack_enter_tlp_calc = counter_u64_alloc(M_WAITOK);
781 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
782 	    SYSCTL_CHILDREN(rack_sysctl_root),
783 	    OID_AUTO, "tlp_calc_entered", CTLFLAG_RD,
784 	    &rack_enter_tlp_calc,
785 	    "Total times we called calc-tlp");
786 	rack_used_tlpmethod = counter_u64_alloc(M_WAITOK);
787 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
788 	    SYSCTL_CHILDREN(rack_sysctl_root),
789 	    OID_AUTO, "hit_tlp_method", CTLFLAG_RD,
790 	    &rack_used_tlpmethod,
791 	    "Total number of runt sacks");
792 	rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK);
793 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
794 	    SYSCTL_CHILDREN(rack_sysctl_root),
795 	    OID_AUTO, "hit_tlp_method2", CTLFLAG_RD,
796 	    &rack_used_tlpmethod2,
797 	    "Total number of runt sacks 2");
798 	rack_runt_sacks = counter_u64_alloc(M_WAITOK);
799 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
800 	    SYSCTL_CHILDREN(rack_sysctl_root),
801 	    OID_AUTO, "runtsacks", CTLFLAG_RD,
802 	    &rack_runt_sacks,
803 	    "Total number of runt sacks");
804 	rack_progress_drops = counter_u64_alloc(M_WAITOK);
805 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
806 	    SYSCTL_CHILDREN(rack_sysctl_root),
807 	    OID_AUTO, "prog_drops", CTLFLAG_RD,
808 	    &rack_progress_drops,
809 	    "Total number of progress drops");
810 	rack_input_idle_reduces = counter_u64_alloc(M_WAITOK);
811 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
812 	    SYSCTL_CHILDREN(rack_sysctl_root),
813 	    OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD,
814 	    &rack_input_idle_reduces,
815 	    "Total number of idle reductions on input");
816 	rack_tlp_does_nada = counter_u64_alloc(M_WAITOK);
817 	SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx,
818 	    SYSCTL_CHILDREN(rack_sysctl_root),
819 	    OID_AUTO, "tlp_nada", CTLFLAG_RD,
820 	    &rack_tlp_does_nada,
821 	    "Total number of nada tlp calls");
822 	COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK);
823 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
824 	    OID_AUTO, "outsize", CTLFLAG_RD,
825 	    rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes");
826 	COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK);
827 	SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root),
828 	    OID_AUTO, "opts", CTLFLAG_RD,
829 	    rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats");
830 	SYSCTL_ADD_PROC(&rack_sysctl_ctx,
831 	    SYSCTL_CHILDREN(rack_sysctl_root),
832 	    OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
833 	    &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters");
834 }
835 
836 static inline int32_t
837 rack_progress_timeout_check(struct tcpcb *tp)
838 {
839 	if (tp->t_maxunacktime && tp->t_acktime && TSTMP_GT(ticks, tp->t_acktime)) {
840 		if ((ticks - tp->t_acktime) >= tp->t_maxunacktime) {
841 			/*
842 			 * There is an assumption that the caller
843 			 * will drop the connection so we will
844 			 * increment the counters here.
845 			 */
846 			struct tcp_rack *rack;
847 			rack = (struct tcp_rack *)tp->t_fb_ptr;
848 			counter_u64_add(rack_progress_drops, 1);
849 #ifdef NETFLIX_STATS
850 			TCPSTAT_INC(tcps_progdrops);
851 #endif
852 			rack_log_progress_event(rack, tp, ticks, PROGRESS_DROP, __LINE__);
853 			return (1);
854 		}
855 	}
856 	return (0);
857 }
858 
859 
860 static void
861 rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which)
862 {
863 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
864 		union tcp_log_stackspecific log;
865 
866 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
867 		log.u_bbr.flex1 = TICKS_2_MSEC(rack->rc_tp->t_srtt >> TCP_RTT_SHIFT);
868 		log.u_bbr.flex2 = to;
869 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
870 		log.u_bbr.flex4 = slot;
871 		log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot;
872 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
873 		log.u_bbr.flex8 = which;
874 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
875 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
876 		TCP_LOG_EVENT(rack->rc_tp, NULL,
877 		    &rack->rc_inp->inp_socket->so_rcv,
878 		    &rack->rc_inp->inp_socket->so_snd,
879 		    BBR_LOG_TIMERSTAR, 0,
880 		    0, &log, false);
881 	}
882 }
883 
884 static void
885 rack_log_to_event(struct tcp_rack *rack, int32_t to_num)
886 {
887 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
888 		union tcp_log_stackspecific log;
889 
890 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
891 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
892 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
893 		log.u_bbr.flex8 = to_num;
894 		log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt;
895 		log.u_bbr.flex2 = rack->rc_rack_rtt;
896 		TCP_LOG_EVENT(rack->rc_tp, NULL,
897 		    &rack->rc_inp->inp_socket->so_rcv,
898 		    &rack->rc_inp->inp_socket->so_snd,
899 		    BBR_LOG_RTO, 0,
900 		    0, &log, false);
901 	}
902 }
903 
904 static void
905 rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, int32_t t,
906     uint32_t o_srtt, uint32_t o_var)
907 {
908 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
909 		union tcp_log_stackspecific log;
910 
911 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
912 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
913 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
914 		log.u_bbr.flex1 = t;
915 		log.u_bbr.flex2 = o_srtt;
916 		log.u_bbr.flex3 = o_var;
917 		log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest;
918 		log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest;
919 		log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_rtt_cnt;
920 		log.u_bbr.rttProp = rack->r_ctl.rack_rs.rs_rtt_tot;
921 		log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method;
922 		TCP_LOG_EVENT(tp, NULL,
923 		    &rack->rc_inp->inp_socket->so_rcv,
924 		    &rack->rc_inp->inp_socket->so_snd,
925 		    BBR_LOG_BBRRTT, 0,
926 		    0, &log, false);
927 	}
928 }
929 
930 static void
931 rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt)
932 {
933 	/*
934 	 * Log the rtt sample we are
935 	 * applying to the srtt algorithm in
936 	 * useconds.
937 	 */
938 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
939 		union tcp_log_stackspecific log;
940 		struct timeval tv;
941 
942 		/* Convert our ms to a microsecond */
943 		log.u_bbr.flex1 = rtt * 1000;
944 		log.u_bbr.timeStamp = tcp_get_usecs(&tv);
945 		TCP_LOG_EVENTP(rack->rc_tp, NULL,
946 		    &rack->rc_inp->inp_socket->so_rcv,
947 		    &rack->rc_inp->inp_socket->so_snd,
948 		    TCP_LOG_RTT, 0,
949 		    0, &log, false, &tv);
950 	}
951 }
952 
953 
954 static inline void
955 rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick,  int event, int line)
956 {
957 	if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) {
958 		union tcp_log_stackspecific log;
959 
960 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
961 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
962 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
963 		log.u_bbr.flex1 = line;
964 		log.u_bbr.flex2 = tick;
965 		log.u_bbr.flex3 = tp->t_maxunacktime;
966 		log.u_bbr.flex4 = tp->t_acktime;
967 		log.u_bbr.flex8 = event;
968 		TCP_LOG_EVENT(tp, NULL,
969 		    &rack->rc_inp->inp_socket->so_rcv,
970 		    &rack->rc_inp->inp_socket->so_snd,
971 		    BBR_LOG_PROGRESS, 0,
972 		    0, &log, false);
973 	}
974 }
975 
976 static void
977 rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts)
978 {
979 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
980 		union tcp_log_stackspecific log;
981 
982 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
983 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
984 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
985 		log.u_bbr.flex1 = slot;
986 		log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags);
987 		log.u_bbr.flex8 = rack->rc_in_persist;
988 		TCP_LOG_EVENT(rack->rc_tp, NULL,
989 		    &rack->rc_inp->inp_socket->so_rcv,
990 		    &rack->rc_inp->inp_socket->so_snd,
991 		    BBR_LOG_BBRSND, 0,
992 		    0, &log, false);
993 	}
994 }
995 
996 static void
997 rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out)
998 {
999 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1000 		union tcp_log_stackspecific log;
1001 		log.u_bbr.flex1 = did_out;
1002 		log.u_bbr.flex2 = nxt_pkt;
1003 		log.u_bbr.flex3 = way_out;
1004 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
1005 		log.u_bbr.flex7 = rack->r_wanted_output;
1006 		log.u_bbr.flex8 = rack->rc_in_persist;
1007 		TCP_LOG_EVENT(rack->rc_tp, NULL,
1008 		    &rack->rc_inp->inp_socket->so_rcv,
1009 		    &rack->rc_inp->inp_socket->so_snd,
1010 		    BBR_LOG_DOSEG_DONE, 0,
1011 		    0, &log, false);
1012 	}
1013 }
1014 
1015 
1016 static void
1017 rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot, uint8_t hpts_calling)
1018 {
1019 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1020 		union tcp_log_stackspecific log;
1021 
1022 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1023 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1024 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1025 		log.u_bbr.flex1 = slot;
1026 		log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags;
1027 		log.u_bbr.flex7 = hpts_calling;
1028 		log.u_bbr.flex8 = rack->rc_in_persist;
1029 		TCP_LOG_EVENT(rack->rc_tp, NULL,
1030 		    &rack->rc_inp->inp_socket->so_rcv,
1031 		    &rack->rc_inp->inp_socket->so_snd,
1032 		    BBR_LOG_JUSTRET, 0,
1033 		    tlen, &log, false);
1034 	}
1035 }
1036 
1037 static void
1038 rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line)
1039 {
1040 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1041 		union tcp_log_stackspecific log;
1042 
1043 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1044 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
1045 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
1046 		log.u_bbr.flex1 = line;
1047 		log.u_bbr.flex2 = 0;
1048 		log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags;
1049 		log.u_bbr.flex4 = 0;
1050 		log.u_bbr.flex6 = rack->rc_tp->t_rxtcur;
1051 		log.u_bbr.flex8 = hpts_removed;
1052 		TCP_LOG_EVENT(rack->rc_tp, NULL,
1053 		    &rack->rc_inp->inp_socket->so_rcv,
1054 		    &rack->rc_inp->inp_socket->so_snd,
1055 		    BBR_LOG_TIMERCANC, 0,
1056 		    0, &log, false);
1057 	}
1058 }
1059 
1060 static void
1061 rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers)
1062 {
1063 	if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) {
1064 		union tcp_log_stackspecific log;
1065 
1066 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
1067 		log.u_bbr.flex1 = timers;
1068 		log.u_bbr.flex2 = ret;
1069 		log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp;
1070 		log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags;
1071 		log.u_bbr.flex5 = cts;
1072 		TCP_LOG_EVENT(rack->rc_tp, NULL,
1073 		    &rack->rc_inp->inp_socket->so_rcv,
1074 		    &rack->rc_inp->inp_socket->so_snd,
1075 		    BBR_LOG_TO_PROCESS, 0,
1076 		    0, &log, false);
1077 	}
1078 }
1079 
1080 static void
1081 rack_counter_destroy()
1082 {
1083 	counter_u64_free(rack_badfr);
1084 	counter_u64_free(rack_badfr_bytes);
1085 	counter_u64_free(rack_rtm_prr_retran);
1086 	counter_u64_free(rack_rtm_prr_newdata);
1087 	counter_u64_free(rack_timestamp_mismatch);
1088 	counter_u64_free(rack_reorder_seen);
1089 	counter_u64_free(rack_tlp_tot);
1090 	counter_u64_free(rack_tlp_newdata);
1091 	counter_u64_free(rack_tlp_retran);
1092 	counter_u64_free(rack_tlp_retran_bytes);
1093 	counter_u64_free(rack_tlp_retran_fail);
1094 	counter_u64_free(rack_to_tot);
1095 	counter_u64_free(rack_to_arm_rack);
1096 	counter_u64_free(rack_to_arm_tlp);
1097 	counter_u64_free(rack_paced_segments);
1098 	counter_u64_free(rack_unpaced_segments);
1099 	counter_u64_free(rack_saw_enobuf);
1100 	counter_u64_free(rack_saw_enetunreach);
1101 	counter_u64_free(rack_to_alloc_hard);
1102 	counter_u64_free(rack_to_alloc_emerg);
1103 	counter_u64_free(rack_sack_proc_all);
1104 	counter_u64_free(rack_sack_proc_short);
1105 	counter_u64_free(rack_sack_proc_restart);
1106 	counter_u64_free(rack_to_alloc);
1107 	counter_u64_free(rack_find_high);
1108 	counter_u64_free(rack_runt_sacks);
1109 	counter_u64_free(rack_enter_tlp_calc);
1110 	counter_u64_free(rack_used_tlpmethod);
1111 	counter_u64_free(rack_used_tlpmethod2);
1112 	counter_u64_free(rack_progress_drops);
1113 	counter_u64_free(rack_input_idle_reduces);
1114 	counter_u64_free(rack_tlp_does_nada);
1115 	COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE);
1116 	COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE);
1117 }
1118 
1119 static struct rack_sendmap *
1120 rack_alloc(struct tcp_rack *rack)
1121 {
1122 	struct rack_sendmap *rsm;
1123 
1124 	counter_u64_add(rack_to_alloc, 1);
1125 	rack->r_ctl.rc_num_maps_alloced++;
1126 	rsm = uma_zalloc(rack_zone, M_NOWAIT);
1127 	if (rsm) {
1128 		return (rsm);
1129 	}
1130 	if (rack->rc_free_cnt) {
1131 		counter_u64_add(rack_to_alloc_emerg, 1);
1132 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
1133 		TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_next);
1134 		rack->rc_free_cnt--;
1135 		return (rsm);
1136 	}
1137 	return (NULL);
1138 }
1139 
1140 static void
1141 rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm)
1142 {
1143 	rack->r_ctl.rc_num_maps_alloced--;
1144 	if (rack->r_ctl.rc_tlpsend == rsm)
1145 		rack->r_ctl.rc_tlpsend = NULL;
1146 	if (rack->r_ctl.rc_next == rsm)
1147 		rack->r_ctl.rc_next = NULL;
1148 	if (rack->r_ctl.rc_sacklast == rsm)
1149 		rack->r_ctl.rc_sacklast = NULL;
1150 	if (rack->rc_free_cnt < rack_free_cache) {
1151 		memset(rsm, 0, sizeof(struct rack_sendmap));
1152 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_next);
1153 		rack->rc_free_cnt++;
1154 		return;
1155 	}
1156 	uma_zfree(rack_zone, rsm);
1157 }
1158 
1159 /*
1160  * CC wrapper hook functions
1161  */
1162 static void
1163 rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, struct tcphdr *th, uint16_t nsegs,
1164     uint16_t type, int32_t recovery)
1165 {
1166 #ifdef NETFLIX_STATS
1167 	int32_t gput;
1168 #endif
1169 #ifdef NETFLIX_CWV
1170 	u_long old_cwnd = tp->snd_cwnd;
1171 #endif
1172 
1173 	INP_WLOCK_ASSERT(tp->t_inpcb);
1174 	tp->ccv->nsegs = nsegs;
1175 	tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th);
1176 	if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) {
1177 		uint32_t max;
1178 
1179 		max = rack->r_ctl.rc_early_recovery_segs * tp->t_maxseg;
1180 		if (tp->ccv->bytes_this_ack > max) {
1181 			tp->ccv->bytes_this_ack = max;
1182 		}
1183 	}
1184 	if (tp->snd_cwnd <= tp->snd_wnd)
1185 		tp->ccv->flags |= CCF_CWND_LIMITED;
1186 	else
1187 		tp->ccv->flags &= ~CCF_CWND_LIMITED;
1188 
1189 	if (type == CC_ACK) {
1190 #ifdef NETFLIX_STATS
1191 		stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF,
1192 		    ((int32_t) tp->snd_cwnd) - tp->snd_wnd);
1193 		if ((tp->t_flags & TF_GPUTINPROG) &&
1194 		    SEQ_GEQ(th->th_ack, tp->gput_ack)) {
1195 			gput = (((int64_t) (th->th_ack - tp->gput_seq)) << 3) /
1196 			    max(1, tcp_ts_getticks() - tp->gput_ts);
1197 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT,
1198 			    gput);
1199 			/*
1200 			 * XXXLAS: This is a temporary hack, and should be
1201 			 * chained off VOI_TCP_GPUT when stats(9) grows an
1202 			 * API to deal with chained VOIs.
1203 			 */
1204 			if (tp->t_stats_gput_prev > 0)
1205 				stats_voi_update_abs_s32(tp->t_stats,
1206 				    VOI_TCP_GPUT_ND,
1207 				    ((gput - tp->t_stats_gput_prev) * 100) /
1208 				    tp->t_stats_gput_prev);
1209 			tp->t_flags &= ~TF_GPUTINPROG;
1210 			tp->t_stats_gput_prev = gput;
1211 #ifdef NETFLIX_CWV
1212 			if (tp->t_maxpeakrate) {
1213 				/*
1214 				 * We update t_peakrate_thr. This gives us roughly
1215 				 * one update per round trip time.
1216 				 */
1217 				tcp_update_peakrate_thr(tp);
1218 			}
1219 #endif
1220 		}
1221 #endif
1222 		if (tp->snd_cwnd > tp->snd_ssthresh) {
1223 			tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
1224 			    nsegs * V_tcp_abc_l_var * tp->t_maxseg);
1225 			if (tp->t_bytes_acked >= tp->snd_cwnd) {
1226 				tp->t_bytes_acked -= tp->snd_cwnd;
1227 				tp->ccv->flags |= CCF_ABC_SENTAWND;
1228 			}
1229 		} else {
1230 			tp->ccv->flags &= ~CCF_ABC_SENTAWND;
1231 			tp->t_bytes_acked = 0;
1232 		}
1233 	}
1234 	if (CC_ALGO(tp)->ack_received != NULL) {
1235 		/* XXXLAS: Find a way to live without this */
1236 		tp->ccv->curack = th->th_ack;
1237 		CC_ALGO(tp)->ack_received(tp->ccv, type);
1238 	}
1239 #ifdef NETFLIX_STATS
1240 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, tp->snd_cwnd);
1241 #endif
1242 	if (rack->r_ctl.rc_rack_largest_cwnd < tp->snd_cwnd) {
1243 		rack->r_ctl.rc_rack_largest_cwnd = tp->snd_cwnd;
1244 	}
1245 #ifdef NETFLIX_CWV
1246 	if (tp->cwv_enabled) {
1247 		/*
1248 		 * Per RFC 7661: The behaviour in the non-validated phase is
1249 		 * specified as: o  A sender determines whether to increase
1250 		 * the cwnd based upon whether it is cwnd-limited (see
1251 		 * Section 4.5.3): * A sender that is cwnd-limited MAY use
1252 		 * the standard TCP method to increase cwnd (i.e., the
1253 		 * standard method permits a TCP sender that fully utilises
1254 		 * the cwnd to increase the cwnd each time it receives an
1255 		 * ACK). * A sender that is not cwnd-limited MUST NOT
1256 		 * increase the cwnd when ACK packets are received in this
1257 		 * phase (i.e., needs to avoid growing the cwnd when it has
1258 		 * not recently sent using the current size of cwnd).
1259 		 */
1260 		if ((tp->snd_cwnd > old_cwnd) &&
1261 		    (tp->cwv_cwnd_valid == 0) &&
1262 		    (!(tp->ccv->flags & CCF_CWND_LIMITED))) {
1263 			tp->snd_cwnd = old_cwnd;
1264 		}
1265 		/* Try to update pipeAck and NCWV state */
1266 		if (TCPS_HAVEESTABLISHED(tp->t_state) &&
1267 		    !IN_RECOVERY(tp->t_flags)) {
1268 			uint32_t data = sbavail(&(tp->t_inpcb->inp_socket->so_snd));
1269 
1270 			tcp_newcwv_update_pipeack(tp, data);
1271 		}
1272 	}
1273 	/* we enforce max peak rate if it is set. */
1274 	if (tp->t_peakrate_thr && tp->snd_cwnd > tp->t_peakrate_thr) {
1275 		tp->snd_cwnd = tp->t_peakrate_thr;
1276 	}
1277 #endif
1278 }
1279 
1280 static void
1281 tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th)
1282 {
1283 	struct tcp_rack *rack;
1284 
1285 	rack = (struct tcp_rack *)tp->t_fb_ptr;
1286 	INP_WLOCK_ASSERT(tp->t_inpcb);
1287 	if (rack->r_ctl.rc_prr_sndcnt > 0)
1288 		rack->r_wanted_output++;
1289 }
1290 
1291 static void
1292 rack_post_recovery(struct tcpcb *tp, struct tcphdr *th)
1293 {
1294 	struct tcp_rack *rack;
1295 
1296 	INP_WLOCK_ASSERT(tp->t_inpcb);
1297 	rack = (struct tcp_rack *)tp->t_fb_ptr;
1298 	if (CC_ALGO(tp)->post_recovery != NULL) {
1299 		tp->ccv->curack = th->th_ack;
1300 		CC_ALGO(tp)->post_recovery(tp->ccv);
1301 	}
1302 	/*
1303 	 * Here we can in theory adjust cwnd to be based on the number of
1304 	 * losses in the window (rack->r_ctl.rc_loss_count). This is done
1305 	 * based on the rack_use_proportional flag.
1306 	 */
1307 	if (rack->r_ctl.rc_prop_reduce && rack->r_ctl.rc_prop_rate) {
1308 		int32_t reduce;
1309 
1310 		reduce = (rack->r_ctl.rc_loss_count * rack->r_ctl.rc_prop_rate);
1311 		if (reduce > 50) {
1312 			reduce = 50;
1313 		}
1314 		tp->snd_cwnd -= ((reduce * tp->snd_cwnd) / 100);
1315 	} else {
1316 		if (tp->snd_cwnd > tp->snd_ssthresh) {
1317 			/* Drop us down to the ssthresh (1/2 cwnd at loss) */
1318 			tp->snd_cwnd = tp->snd_ssthresh;
1319 		}
1320 	}
1321 	if (rack->r_ctl.rc_prr_sndcnt > 0) {
1322 		/* Suck the next prr cnt back into cwnd */
1323 		tp->snd_cwnd += rack->r_ctl.rc_prr_sndcnt;
1324 		rack->r_ctl.rc_prr_sndcnt = 0;
1325 	}
1326 	EXIT_RECOVERY(tp->t_flags);
1327 
1328 
1329 #ifdef NETFLIX_CWV
1330 	if (tp->cwv_enabled) {
1331 		if ((tp->cwv_cwnd_valid == 0) &&
1332 		    (tp->snd_cwv.in_recovery))
1333 			tcp_newcwv_end_recovery(tp);
1334 	}
1335 #endif
1336 }
1337 
1338 static void
1339 rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type)
1340 {
1341 	struct tcp_rack *rack;
1342 
1343 	INP_WLOCK_ASSERT(tp->t_inpcb);
1344 
1345 	rack = (struct tcp_rack *)tp->t_fb_ptr;
1346 	switch (type) {
1347 	case CC_NDUPACK:
1348 /*		rack->r_ctl.rc_ssthresh_set = 1;*/
1349 		if (!IN_FASTRECOVERY(tp->t_flags)) {
1350 			rack->r_ctl.rc_tlp_rtx_out = 0;
1351 			rack->r_ctl.rc_prr_delivered = 0;
1352 			rack->r_ctl.rc_prr_out = 0;
1353 			rack->r_ctl.rc_loss_count = 0;
1354 			rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg;
1355 			rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
1356 			tp->snd_recover = tp->snd_max;
1357 			if (tp->t_flags & TF_ECN_PERMIT)
1358 				tp->t_flags |= TF_ECN_SND_CWR;
1359 		}
1360 		break;
1361 	case CC_ECN:
1362 		if (!IN_CONGRECOVERY(tp->t_flags)) {
1363 			TCPSTAT_INC(tcps_ecn_rcwnd);
1364 			tp->snd_recover = tp->snd_max;
1365 			if (tp->t_flags & TF_ECN_PERMIT)
1366 				tp->t_flags |= TF_ECN_SND_CWR;
1367 		}
1368 		break;
1369 	case CC_RTO:
1370 		tp->t_dupacks = 0;
1371 		tp->t_bytes_acked = 0;
1372 		EXIT_RECOVERY(tp->t_flags);
1373 		tp->snd_ssthresh = max(2, min(tp->snd_wnd, tp->snd_cwnd) / 2 /
1374 		    tp->t_maxseg) * tp->t_maxseg;
1375 		tp->snd_cwnd = tp->t_maxseg;
1376 		break;
1377 	case CC_RTO_ERR:
1378 		TCPSTAT_INC(tcps_sndrexmitbad);
1379 		/* RTO was unnecessary, so reset everything. */
1380 		tp->snd_cwnd = tp->snd_cwnd_prev;
1381 		tp->snd_ssthresh = tp->snd_ssthresh_prev;
1382 		tp->snd_recover = tp->snd_recover_prev;
1383 		if (tp->t_flags & TF_WASFRECOVERY)
1384 			ENTER_FASTRECOVERY(tp->t_flags);
1385 		if (tp->t_flags & TF_WASCRECOVERY)
1386 			ENTER_CONGRECOVERY(tp->t_flags);
1387 		tp->snd_nxt = tp->snd_max;
1388 		tp->t_badrxtwin = 0;
1389 		break;
1390 	}
1391 
1392 	if (CC_ALGO(tp)->cong_signal != NULL) {
1393 		if (th != NULL)
1394 			tp->ccv->curack = th->th_ack;
1395 		CC_ALGO(tp)->cong_signal(tp->ccv, type);
1396 	}
1397 #ifdef NETFLIX_CWV
1398 	if (tp->cwv_enabled) {
1399 		if (tp->snd_cwv.in_recovery == 0 && IN_RECOVERY(tp->t_flags)) {
1400 			tcp_newcwv_enter_recovery(tp);
1401 		}
1402 		if (type == CC_RTO) {
1403 			tcp_newcwv_reset(tp);
1404 		}
1405 	}
1406 #endif
1407 }
1408 
1409 
1410 
1411 static inline void
1412 rack_cc_after_idle(struct tcpcb *tp, int reduce_largest)
1413 {
1414 	uint32_t i_cwnd;
1415 
1416 	INP_WLOCK_ASSERT(tp->t_inpcb);
1417 
1418 #ifdef NETFLIX_STATS
1419 	TCPSTAT_INC(tcps_idle_restarts);
1420 	if (tp->t_state == TCPS_ESTABLISHED)
1421 		TCPSTAT_INC(tcps_idle_estrestarts);
1422 #endif
1423 	if (CC_ALGO(tp)->after_idle != NULL)
1424 		CC_ALGO(tp)->after_idle(tp->ccv);
1425 
1426 	if (tp->snd_cwnd == 1)
1427 		i_cwnd = tp->t_maxseg;		/* SYN(-ACK) lost */
1428 	else if (V_tcp_initcwnd_segments)
1429 		i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg),
1430 		    max(2 * tp->t_maxseg, V_tcp_initcwnd_segments * 1460));
1431 	else if (V_tcp_do_rfc3390)
1432 		i_cwnd = min(4 * tp->t_maxseg,
1433 		    max(2 * tp->t_maxseg, 4380));
1434 	else {
1435 		/* Per RFC5681 Section 3.1 */
1436 		if (tp->t_maxseg > 2190)
1437 			i_cwnd = 2 * tp->t_maxseg;
1438 		else if (tp->t_maxseg > 1095)
1439 			i_cwnd = 3 * tp->t_maxseg;
1440 		else
1441 			i_cwnd = 4 * tp->t_maxseg;
1442 	}
1443 	if (reduce_largest) {
1444 		/*
1445 		 * Do we reduce the largest cwnd to make
1446 		 * rack play nice on restart hptsi wise?
1447 		 */
1448 		if (((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rack_largest_cwnd  > i_cwnd)
1449 			((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rack_largest_cwnd = i_cwnd;
1450 	}
1451 	/*
1452 	 * Being idle is no differnt than the initial window. If the cc
1453 	 * clamps it down below the initial window raise it to the initial
1454 	 * window.
1455 	 */
1456 	if (tp->snd_cwnd < i_cwnd) {
1457 		tp->snd_cwnd = i_cwnd;
1458 	}
1459 }
1460 
1461 
1462 /*
1463  * Indicate whether this ack should be delayed.  We can delay the ack if
1464  * following conditions are met:
1465  *	- There is no delayed ack timer in progress.
1466  *	- Our last ack wasn't a 0-sized window. We never want to delay
1467  *	  the ack that opens up a 0-sized window.
1468  *	- LRO wasn't used for this segment. We make sure by checking that the
1469  *	  segment size is not larger than the MSS.
1470  *	- Delayed acks are enabled or this is a half-synchronized T/TCP
1471  *	  connection.
1472  */
1473 #define DELAY_ACK(tp, tlen)			 \
1474 	(((tp->t_flags & TF_RXWIN0SENT) == 0) && \
1475 	((tp->t_flags & TF_DELACK) == 0) && 	 \
1476 	(tlen <= tp->t_maxseg) &&		 \
1477 	(tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN)))
1478 
1479 static inline void
1480 rack_calc_rwin(struct socket *so, struct tcpcb *tp)
1481 {
1482 	int32_t win;
1483 
1484 	/*
1485 	 * Calculate amount of space in receive window, and then do TCP
1486 	 * input processing. Receive window is amount of space in rcv queue,
1487 	 * but not less than advertised window.
1488 	 */
1489 	win = sbspace(&so->so_rcv);
1490 	if (win < 0)
1491 		win = 0;
1492 	tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt));
1493 }
1494 
1495 static void
1496 rack_do_drop(struct mbuf *m, struct tcpcb *tp)
1497 {
1498 	/*
1499 	 * Drop space held by incoming segment and return.
1500 	 */
1501 	if (tp != NULL)
1502 		INP_WUNLOCK(tp->t_inpcb);
1503 	if (m)
1504 		m_freem(m);
1505 }
1506 
1507 static void
1508 rack_do_dropwithreset(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th,
1509     int32_t rstreason, int32_t tlen)
1510 {
1511 	if (tp != NULL) {
1512 		tcp_dropwithreset(m, th, tp, tlen, rstreason);
1513 		INP_WUNLOCK(tp->t_inpcb);
1514 	} else
1515 		tcp_dropwithreset(m, th, NULL, tlen, rstreason);
1516 }
1517 
1518 /*
1519  * The value in ret_val informs the caller
1520  * if we dropped the tcb (and lock) or not.
1521  * 1 = we dropped it, 0 = the TCB is still locked
1522  * and valid.
1523  */
1524 static void
1525 rack_do_dropafterack(struct mbuf *m, struct tcpcb *tp, struct tcphdr *th, int32_t thflags, int32_t tlen, int32_t * ret_val)
1526 {
1527 	/*
1528 	 * Generate an ACK dropping incoming segment if it occupies sequence
1529 	 * space, where the ACK reflects our state.
1530 	 *
1531 	 * We can now skip the test for the RST flag since all paths to this
1532 	 * code happen after packets containing RST have been dropped.
1533 	 *
1534 	 * In the SYN-RECEIVED state, don't send an ACK unless the segment
1535 	 * we received passes the SYN-RECEIVED ACK test. If it fails send a
1536 	 * RST.  This breaks the loop in the "LAND" DoS attack, and also
1537 	 * prevents an ACK storm between two listening ports that have been
1538 	 * sent forged SYN segments, each with the source address of the
1539 	 * other.
1540 	 */
1541 	struct tcp_rack *rack;
1542 
1543 	if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) &&
1544 	    (SEQ_GT(tp->snd_una, th->th_ack) ||
1545 	    SEQ_GT(th->th_ack, tp->snd_max))) {
1546 		*ret_val = 1;
1547 		rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
1548 		return;
1549 	} else
1550 		*ret_val = 0;
1551 	rack = (struct tcp_rack *)tp->t_fb_ptr;
1552 	rack->r_wanted_output++;
1553 	tp->t_flags |= TF_ACKNOW;
1554 	if (m)
1555 		m_freem(m);
1556 }
1557 
1558 
1559 static int
1560 rack_process_rst(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp)
1561 {
1562 	/*
1563 	 * RFC5961 Section 3.2
1564 	 *
1565 	 * - RST drops connection only if SEG.SEQ == RCV.NXT. - If RST is in
1566 	 * window, we send challenge ACK.
1567 	 *
1568 	 * Note: to take into account delayed ACKs, we should test against
1569 	 * last_ack_sent instead of rcv_nxt. Note 2: we handle special case
1570 	 * of closed window, not covered by the RFC.
1571 	 */
1572 	int dropped = 0;
1573 
1574 	if ((SEQ_GEQ(th->th_seq, (tp->last_ack_sent - 1)) &&
1575 	    SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) ||
1576 	    (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) {
1577 
1578 		INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1579 		KASSERT(tp->t_state != TCPS_SYN_SENT,
1580 		    ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p",
1581 		    __func__, th, tp));
1582 
1583 		if (V_tcp_insecure_rst ||
1584 		    (tp->last_ack_sent == th->th_seq) ||
1585 		    (tp->rcv_nxt == th->th_seq) ||
1586 		    ((tp->last_ack_sent - 1) == th->th_seq)) {
1587 			TCPSTAT_INC(tcps_drops);
1588 			/* Drop the connection. */
1589 			switch (tp->t_state) {
1590 			case TCPS_SYN_RECEIVED:
1591 				so->so_error = ECONNREFUSED;
1592 				goto close;
1593 			case TCPS_ESTABLISHED:
1594 			case TCPS_FIN_WAIT_1:
1595 			case TCPS_FIN_WAIT_2:
1596 			case TCPS_CLOSE_WAIT:
1597 			case TCPS_CLOSING:
1598 			case TCPS_LAST_ACK:
1599 				so->so_error = ECONNRESET;
1600 		close:
1601 				tcp_state_change(tp, TCPS_CLOSED);
1602 				/* FALLTHROUGH */
1603 			default:
1604 				tp = tcp_close(tp);
1605 			}
1606 			dropped = 1;
1607 			rack_do_drop(m, tp);
1608 		} else {
1609 			TCPSTAT_INC(tcps_badrst);
1610 			/* Send challenge ACK. */
1611 			tcp_respond(tp, mtod(m, void *), th, m,
1612 			    tp->rcv_nxt, tp->snd_nxt, TH_ACK);
1613 			tp->last_ack_sent = tp->rcv_nxt;
1614 		}
1615 	} else {
1616 		m_freem(m);
1617 	}
1618 	return (dropped);
1619 }
1620 
1621 /*
1622  * The value in ret_val informs the caller
1623  * if we dropped the tcb (and lock) or not.
1624  * 1 = we dropped it, 0 = the TCB is still locked
1625  * and valid.
1626  */
1627 static void
1628 rack_challenge_ack(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * ret_val)
1629 {
1630 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
1631 
1632 	TCPSTAT_INC(tcps_badsyn);
1633 	if (V_tcp_insecure_syn &&
1634 	    SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
1635 	    SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) {
1636 		tp = tcp_drop(tp, ECONNRESET);
1637 		*ret_val = 1;
1638 		rack_do_drop(m, tp);
1639 	} else {
1640 		/* Send challenge ACK. */
1641 		tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt,
1642 		    tp->snd_nxt, TH_ACK);
1643 		tp->last_ack_sent = tp->rcv_nxt;
1644 		m = NULL;
1645 		*ret_val = 0;
1646 		rack_do_drop(m, NULL);
1647 	}
1648 }
1649 
1650 /*
1651  * rack_ts_check returns 1 for you should not proceed. It places
1652  * in ret_val what should be returned 1/0 by the caller. The 1 indicates
1653  * that the TCB is unlocked and probably dropped. The 0 indicates the
1654  * TCB is still valid and locked.
1655  */
1656 static int
1657 rack_ts_check(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t tlen, int32_t thflags, int32_t * ret_val)
1658 {
1659 
1660 	/* Check to see if ts_recent is over 24 days old.  */
1661 	if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) {
1662 		/*
1663 		 * Invalidate ts_recent.  If this segment updates ts_recent,
1664 		 * the age will be reset later and ts_recent will get a
1665 		 * valid value.  If it does not, setting ts_recent to zero
1666 		 * will at least satisfy the requirement that zero be placed
1667 		 * in the timestamp echo reply when ts_recent isn't valid.
1668 		 * The age isn't reset until we get a valid ts_recent
1669 		 * because we don't want out-of-order segments to be dropped
1670 		 * when ts_recent is old.
1671 		 */
1672 		tp->ts_recent = 0;
1673 	} else {
1674 		TCPSTAT_INC(tcps_rcvduppack);
1675 		TCPSTAT_ADD(tcps_rcvdupbyte, tlen);
1676 		TCPSTAT_INC(tcps_pawsdrop);
1677 		*ret_val = 0;
1678 		if (tlen) {
1679 			rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
1680 		} else {
1681 			rack_do_drop(m, NULL);
1682 		}
1683 		return (1);
1684 	}
1685 	return (0);
1686 }
1687 
1688 /*
1689  * rack_drop_checks returns 1 for you should not proceed. It places
1690  * in ret_val what should be returned 1/0 by the caller. The 1 indicates
1691  * that the TCB is unlocked and probably dropped. The 0 indicates the
1692  * TCB is still valid and locked.
1693  */
1694 static int
1695 rack_drop_checks(struct tcpopt *to, struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int32_t * tlenp,  int32_t * thf, int32_t * drop_hdrlen, int32_t * ret_val)
1696 {
1697 	int32_t todrop;
1698 	int32_t thflags;
1699 	int32_t tlen;
1700 
1701 	thflags = *thf;
1702 	tlen = *tlenp;
1703 	todrop = tp->rcv_nxt - th->th_seq;
1704 	if (todrop > 0) {
1705 		if (thflags & TH_SYN) {
1706 			thflags &= ~TH_SYN;
1707 			th->th_seq++;
1708 			if (th->th_urp > 1)
1709 				th->th_urp--;
1710 			else
1711 				thflags &= ~TH_URG;
1712 			todrop--;
1713 		}
1714 		/*
1715 		 * Following if statement from Stevens, vol. 2, p. 960.
1716 		 */
1717 		if (todrop > tlen
1718 		    || (todrop == tlen && (thflags & TH_FIN) == 0)) {
1719 			/*
1720 			 * Any valid FIN must be to the left of the window.
1721 			 * At this point the FIN must be a duplicate or out
1722 			 * of sequence; drop it.
1723 			 */
1724 			thflags &= ~TH_FIN;
1725 			/*
1726 			 * Send an ACK to resynchronize and drop any data.
1727 			 * But keep on processing for RST or ACK.
1728 			 */
1729 			tp->t_flags |= TF_ACKNOW;
1730 			todrop = tlen;
1731 			TCPSTAT_INC(tcps_rcvduppack);
1732 			TCPSTAT_ADD(tcps_rcvdupbyte, todrop);
1733 		} else {
1734 			TCPSTAT_INC(tcps_rcvpartduppack);
1735 			TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop);
1736 		}
1737 		*drop_hdrlen += todrop;	/* drop from the top afterwards */
1738 		th->th_seq += todrop;
1739 		tlen -= todrop;
1740 		if (th->th_urp > todrop)
1741 			th->th_urp -= todrop;
1742 		else {
1743 			thflags &= ~TH_URG;
1744 			th->th_urp = 0;
1745 		}
1746 	}
1747 	/*
1748 	 * If segment ends after window, drop trailing data (and PUSH and
1749 	 * FIN); if nothing left, just ACK.
1750 	 */
1751 	todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd);
1752 	if (todrop > 0) {
1753 		TCPSTAT_INC(tcps_rcvpackafterwin);
1754 		if (todrop >= tlen) {
1755 			TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen);
1756 			/*
1757 			 * If window is closed can only take segments at
1758 			 * window edge, and have to drop data and PUSH from
1759 			 * incoming segments.  Continue processing, but
1760 			 * remember to ack.  Otherwise, drop segment and
1761 			 * ack.
1762 			 */
1763 			if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) {
1764 				tp->t_flags |= TF_ACKNOW;
1765 				TCPSTAT_INC(tcps_rcvwinprobe);
1766 			} else {
1767 				rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
1768 				return (1);
1769 			}
1770 		} else
1771 			TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
1772 		m_adj(m, -todrop);
1773 		tlen -= todrop;
1774 		thflags &= ~(TH_PUSH | TH_FIN);
1775 	}
1776 	*thf = thflags;
1777 	*tlenp = tlen;
1778 	return (0);
1779 }
1780 
1781 static struct rack_sendmap *
1782 rack_find_lowest_rsm(struct tcp_rack *rack)
1783 {
1784 	struct rack_sendmap *rsm;
1785 
1786 	/*
1787 	 * Walk the time-order transmitted list looking for an rsm that is
1788 	 * not acked. This will be the one that was sent the longest time
1789 	 * ago that is still outstanding.
1790 	 */
1791 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) {
1792 		if (rsm->r_flags & RACK_ACKED) {
1793 			continue;
1794 		}
1795 		goto finish;
1796 	}
1797 finish:
1798 	return (rsm);
1799 }
1800 
1801 static struct rack_sendmap *
1802 rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm)
1803 {
1804 	struct rack_sendmap *prsm;
1805 
1806 	/*
1807 	 * Walk the sequence order list backward until we hit and arrive at
1808 	 * the highest seq not acked. In theory when this is called it
1809 	 * should be the last segment (which it was not).
1810 	 */
1811 	counter_u64_add(rack_find_high, 1);
1812 	prsm = rsm;
1813 	TAILQ_FOREACH_REVERSE_FROM(prsm, &rack->r_ctl.rc_map, rack_head, r_next) {
1814 		if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) {
1815 			continue;
1816 		}
1817 		return (prsm);
1818 	}
1819 	return (NULL);
1820 }
1821 
1822 
1823 static uint32_t
1824 rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts)
1825 {
1826 	int32_t lro;
1827 	uint32_t thresh;
1828 
1829 	/*
1830 	 * lro is the flag we use to determine if we have seen reordering.
1831 	 * If it gets set we have seen reordering. The reorder logic either
1832 	 * works in one of two ways:
1833 	 *
1834 	 * If reorder-fade is configured, then we track the last time we saw
1835 	 * re-ordering occur. If we reach the point where enough time as
1836 	 * passed we no longer consider reordering has occuring.
1837 	 *
1838 	 * Or if reorder-face is 0, then once we see reordering we consider
1839 	 * the connection to alway be subject to reordering and just set lro
1840 	 * to 1.
1841 	 *
1842 	 * In the end if lro is non-zero we add the extra time for
1843 	 * reordering in.
1844 	 */
1845 	if (srtt == 0)
1846 		srtt = 1;
1847 	if (rack->r_ctl.rc_reorder_ts) {
1848 		if (rack->r_ctl.rc_reorder_fade) {
1849 			if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) {
1850 				lro = cts - rack->r_ctl.rc_reorder_ts;
1851 				if (lro == 0) {
1852 					/*
1853 					 * No time as passed since the last
1854 					 * reorder, mark it as reordering.
1855 					 */
1856 					lro = 1;
1857 				}
1858 			} else {
1859 				/* Negative time? */
1860 				lro = 0;
1861 			}
1862 			if (lro > rack->r_ctl.rc_reorder_fade) {
1863 				/* Turn off reordering seen too */
1864 				rack->r_ctl.rc_reorder_ts = 0;
1865 				lro = 0;
1866 			}
1867 		} else {
1868 			/* Reodering does not fade */
1869 			lro = 1;
1870 		}
1871 	} else {
1872 		lro = 0;
1873 	}
1874 	thresh = srtt + rack->r_ctl.rc_pkt_delay;
1875 	if (lro) {
1876 		/* It must be set, if not you get 1/4 rtt */
1877 		if (rack->r_ctl.rc_reorder_shift)
1878 			thresh += (srtt >> rack->r_ctl.rc_reorder_shift);
1879 		else
1880 			thresh += (srtt >> 2);
1881 	} else {
1882 		thresh += 1;
1883 	}
1884 	/* We don't let the rack timeout be above a RTO */
1885 
1886 	if (thresh > TICKS_2_MSEC(rack->rc_tp->t_rxtcur)) {
1887 		thresh = TICKS_2_MSEC(rack->rc_tp->t_rxtcur);
1888 	}
1889 	/* And we don't want it above the RTO max either */
1890 	if (thresh > rack_rto_max) {
1891 		thresh = rack_rto_max;
1892 	}
1893 	return (thresh);
1894 }
1895 
1896 static uint32_t
1897 rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack,
1898 		     struct rack_sendmap *rsm, uint32_t srtt)
1899 {
1900 	struct rack_sendmap *prsm;
1901 	uint32_t thresh, len;
1902 	int maxseg;
1903 
1904 	if (srtt == 0)
1905 		srtt = 1;
1906 	if (rack->r_ctl.rc_tlp_threshold)
1907 		thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold);
1908 	else
1909 		thresh = (srtt * 2);
1910 
1911 	/* Get the previous sent packet, if any  */
1912 	maxseg = tcp_maxseg(tp);
1913 	counter_u64_add(rack_enter_tlp_calc, 1);
1914 	len = rsm->r_end - rsm->r_start;
1915 	if (rack->rack_tlp_threshold_use == TLP_USE_ID) {
1916 		/* Exactly like the ID */
1917 		if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= maxseg) {
1918 			uint32_t alt_thresh;
1919 			/*
1920 			 * Compensate for delayed-ack with the d-ack time.
1921 			 */
1922 			counter_u64_add(rack_used_tlpmethod, 1);
1923 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
1924 			if (alt_thresh > thresh)
1925 				thresh = alt_thresh;
1926 		}
1927 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) {
1928 		/* 2.1 behavior */
1929 		prsm = TAILQ_PREV(rsm, rack_head, r_tnext);
1930 		if (prsm && (len <= maxseg)) {
1931 			/*
1932 			 * Two packets outstanding, thresh should be (2*srtt) +
1933 			 * possible inter-packet delay (if any).
1934 			 */
1935 			uint32_t inter_gap = 0;
1936 			int idx, nidx;
1937 
1938 			counter_u64_add(rack_used_tlpmethod, 1);
1939 			idx = rsm->r_rtr_cnt - 1;
1940 			nidx = prsm->r_rtr_cnt - 1;
1941 			if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) {
1942 				/* Yes it was sent later (or at the same time) */
1943 				inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx];
1944 			}
1945 			thresh += inter_gap;
1946 		} else 	if (len <= maxseg) {
1947 			/*
1948 			 * Possibly compensate for delayed-ack.
1949 			 */
1950 			uint32_t alt_thresh;
1951 
1952 			counter_u64_add(rack_used_tlpmethod2, 1);
1953 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
1954 			if (alt_thresh > thresh)
1955 				thresh = alt_thresh;
1956 		}
1957 	} else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) {
1958 		/* 2.2 behavior */
1959 		if (len <= maxseg) {
1960 			uint32_t alt_thresh;
1961 			/*
1962 			 * Compensate for delayed-ack with the d-ack time.
1963 			 */
1964 			counter_u64_add(rack_used_tlpmethod, 1);
1965 			alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time;
1966 			if (alt_thresh > thresh)
1967 				thresh = alt_thresh;
1968 		}
1969 	}
1970  	/* Not above an RTO */
1971 	if (thresh > TICKS_2_MSEC(tp->t_rxtcur)) {
1972 		thresh = TICKS_2_MSEC(tp->t_rxtcur);
1973 	}
1974 	/* Not above a RTO max */
1975 	if (thresh > rack_rto_max) {
1976 		thresh = rack_rto_max;
1977 	}
1978 	/* Apply user supplied min TLP */
1979 	if (thresh < rack_tlp_min) {
1980 		thresh = rack_tlp_min;
1981 	}
1982 	return (thresh);
1983 }
1984 
1985 static struct rack_sendmap *
1986 rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused)
1987 {
1988 	/*
1989 	 * Check to see that we don't need to fall into recovery. We will
1990 	 * need to do so if our oldest transmit is past the time we should
1991 	 * have had an ack.
1992 	 */
1993 	struct tcp_rack *rack;
1994 	struct rack_sendmap *rsm;
1995 	int32_t idx;
1996 	uint32_t srtt_cur, srtt, thresh;
1997 
1998 	rack = (struct tcp_rack *)tp->t_fb_ptr;
1999 	if (TAILQ_EMPTY(&rack->r_ctl.rc_map)) {
2000 		return (NULL);
2001 	}
2002 	srtt_cur = tp->t_srtt >> TCP_RTT_SHIFT;
2003 	srtt = TICKS_2_MSEC(srtt_cur);
2004 	if (rack->rc_rack_rtt && (srtt > rack->rc_rack_rtt))
2005 		srtt = rack->rc_rack_rtt;
2006 
2007 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
2008 	if (rsm == NULL)
2009 		return (NULL);
2010 
2011 	if (rsm->r_flags & RACK_ACKED) {
2012 		rsm = rack_find_lowest_rsm(rack);
2013 		if (rsm == NULL)
2014 			return (NULL);
2015 	}
2016 	idx = rsm->r_rtr_cnt - 1;
2017 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
2018 	if (tsused < rsm->r_tim_lastsent[idx]) {
2019 		return (NULL);
2020 	}
2021 	if ((tsused - rsm->r_tim_lastsent[idx]) < thresh) {
2022 		return (NULL);
2023 	}
2024 	/* Ok if we reach here we are over-due */
2025 	rack->r_ctl.rc_rsm_start = rsm->r_start;
2026 	rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
2027 	rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
2028 	rack_cong_signal(tp, NULL, CC_NDUPACK);
2029 	return (rsm);
2030 }
2031 
2032 static uint32_t
2033 rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack)
2034 {
2035 	int32_t t;
2036 	int32_t tt;
2037 	uint32_t ret_val;
2038 
2039 	t = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT) + ((tp->t_rttvar * 4) >> TCP_RTT_SHIFT));
2040 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
2041 	    tcp_persmin, tcp_persmax);
2042 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
2043 		tp->t_rxtshift++;
2044 	rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT;
2045 	ret_val = (uint32_t)tt;
2046 	return (ret_val);
2047 }
2048 
2049 static uint32_t
2050 rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
2051 {
2052 	/*
2053 	 * Start the FR timer, we do this based on getting the first one in
2054 	 * the rc_tmap. Note that if its NULL we must stop the timer. in all
2055 	 * events we need to stop the running timer (if its running) before
2056 	 * starting the new one.
2057 	 */
2058 	uint32_t thresh, exp, to, srtt, time_since_sent;
2059 	uint32_t srtt_cur;
2060 	int32_t idx;
2061 	int32_t is_tlp_timer = 0;
2062 	struct rack_sendmap *rsm;
2063 
2064 	if (rack->t_timers_stopped) {
2065 		/* All timers have been stopped none are to run */
2066 		return (0);
2067 	}
2068 	if (rack->rc_in_persist) {
2069 		/* We can't start any timer in persists */
2070 		return (rack_get_persists_timer_val(tp, rack));
2071 	}
2072 	if (tp->t_state < TCPS_ESTABLISHED)
2073 		goto activate_rxt;
2074 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
2075 	if (rsm == NULL) {
2076 		/* Nothing on the send map */
2077 activate_rxt:
2078 		if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) {
2079 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT;
2080 			to = TICKS_2_MSEC(tp->t_rxtcur);
2081 			if (to == 0)
2082 				to = 1;
2083 			return (to);
2084 		}
2085 		return (0);
2086 	}
2087 	if (rsm->r_flags & RACK_ACKED) {
2088 		rsm = rack_find_lowest_rsm(rack);
2089 		if (rsm == NULL) {
2090 			/* No lowest? */
2091 			goto activate_rxt;
2092 		}
2093 	}
2094 	/* Convert from ms to usecs */
2095 	if (rsm->r_flags & RACK_SACK_PASSED) {
2096 		if ((tp->t_flags & TF_SENTFIN) &&
2097 		    ((tp->snd_max - tp->snd_una) == 1) &&
2098 		    (rsm->r_flags & RACK_HAS_FIN)) {
2099 			/*
2100 			 * We don't start a rack timer if all we have is a
2101 			 * FIN outstanding.
2102 			 */
2103 			goto activate_rxt;
2104 		}
2105 		if (tp->t_srtt) {
2106 			srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT);
2107 			srtt = TICKS_2_MSEC(srtt_cur);
2108 		} else
2109 			srtt = RACK_INITIAL_RTO;
2110 
2111 		thresh = rack_calc_thresh_rack(rack, srtt, cts);
2112 		idx = rsm->r_rtr_cnt - 1;
2113 		exp = rsm->r_tim_lastsent[idx] + thresh;
2114 		if (SEQ_GEQ(exp, cts)) {
2115 			to = exp - cts;
2116 			if (to < rack->r_ctl.rc_min_to) {
2117 				to = rack->r_ctl.rc_min_to;
2118 			}
2119 		} else {
2120 			to = rack->r_ctl.rc_min_to;
2121 		}
2122 	} else {
2123 		/* Ok we need to do a TLP not RACK */
2124 		if ((rack->rc_tlp_in_progress != 0) ||
2125 		    (rack->r_ctl.rc_tlp_rtx_out != 0)) {
2126 			/*
2127 			 * The previous send was a TLP or a tlp_rtx is in
2128 			 * process.
2129 			 */
2130 			goto activate_rxt;
2131 		}
2132 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
2133 		if (rsm == NULL) {
2134 			/* We found no rsm to TLP with. */
2135 			goto activate_rxt;
2136 		}
2137 		if (rsm->r_flags & RACK_HAS_FIN) {
2138 			/* If its a FIN we dont do TLP */
2139 			rsm = NULL;
2140 			goto activate_rxt;
2141 		}
2142 		idx = rsm->r_rtr_cnt - 1;
2143 		if (TSTMP_GT(cts,  rsm->r_tim_lastsent[idx]))
2144 			time_since_sent = cts - rsm->r_tim_lastsent[idx];
2145 		else
2146 			time_since_sent = 0;
2147 		is_tlp_timer = 1;
2148 		if (tp->t_srtt) {
2149 			srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT);
2150 			srtt = TICKS_2_MSEC(srtt_cur);
2151 		} else
2152 			srtt = RACK_INITIAL_RTO;
2153 		thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt);
2154 		if (thresh > time_since_sent)
2155 			to = thresh - time_since_sent;
2156 		else
2157 			to = rack->r_ctl.rc_min_to;
2158 		if (to > TCPTV_REXMTMAX) {
2159 			/*
2160 			 * If the TLP time works out to larger than the max
2161 			 * RTO lets not do TLP.. just RTO.
2162 			 */
2163 			goto activate_rxt;
2164 		}
2165 		if (rsm->r_start != rack->r_ctl.rc_last_tlp_seq) {
2166 			/*
2167 			 * The tail is no longer the last one I did a probe
2168 			 * on
2169 			 */
2170 			rack->r_ctl.rc_tlp_seg_send_cnt = 0;
2171 			rack->r_ctl.rc_last_tlp_seq = rsm->r_start;
2172 		}
2173 	}
2174 	if (is_tlp_timer == 0) {
2175 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK;
2176 	} else {
2177 		if ((rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) ||
2178 		    (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) {
2179 			/*
2180 			 * We have exceeded how many times we can retran the
2181 			 * current TLP timer, switch to the RTO timer.
2182 			 */
2183 			goto activate_rxt;
2184 		} else {
2185 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP;
2186 		}
2187 	}
2188 	if (to == 0)
2189 		to = 1;
2190 	return (to);
2191 }
2192 
2193 static void
2194 rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
2195 {
2196 	if (rack->rc_in_persist == 0) {
2197 		if (((tp->t_flags & TF_SENTFIN) == 0) &&
2198 		    (tp->snd_max - tp->snd_una) >= sbavail(&rack->rc_inp->inp_socket->so_snd))
2199 			/* Must need to send more data to enter persist */
2200 			return;
2201 		rack->r_ctl.rc_went_idle_time = cts;
2202 		rack_timer_cancel(tp, rack, cts, __LINE__);
2203 		tp->t_rxtshift = 0;
2204 		rack->rc_in_persist = 1;
2205 	}
2206 }
2207 
2208 static void
2209 rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack)
2210 {
2211 	if (rack->rc_inp->inp_in_hpts)  {
2212 		tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
2213 		rack->r_ctl.rc_hpts_flags  = 0;
2214 	}
2215 	rack->rc_in_persist = 0;
2216 	rack->r_ctl.rc_went_idle_time = 0;
2217 	tp->t_flags &= ~TF_FORCEDATA;
2218 	tp->t_rxtshift = 0;
2219 }
2220 
2221 static void
2222 rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts, int32_t line,
2223     int32_t slot, uint32_t tot_len_this_send, int32_t frm_out_sbavail)
2224 {
2225 	struct inpcb *inp;
2226 	uint32_t delayed_ack = 0;
2227 	uint32_t hpts_timeout;
2228 	uint8_t stopped;
2229 	uint32_t left = 0;
2230 
2231 	inp = tp->t_inpcb;
2232 	if (inp->inp_in_hpts) {
2233 		/* A previous call is already set up */
2234 		return;
2235 	}
2236 	if (tp->t_state == TCPS_CLOSED) {
2237 		return;
2238 	}
2239 	stopped = rack->rc_tmr_stopped;
2240 	if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) {
2241 		left = rack->r_ctl.rc_timer_exp - cts;
2242 	}
2243 	rack->r_ctl.rc_timer_exp = 0;
2244 	if (rack->rc_inp->inp_in_hpts == 0) {
2245 		rack->r_ctl.rc_hpts_flags = 0;
2246 	}
2247 	if (slot) {
2248 		/* We are hptsi too */
2249 		rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT;
2250 	} else if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
2251 		/*
2252 		 * We are still left on the hpts when the to goes
2253 		 * it will be for output.
2254 		 */
2255 		if (TSTMP_GT(cts, rack->r_ctl.rc_last_output_to))
2256 			slot = cts - rack->r_ctl.rc_last_output_to;
2257 		else
2258 			slot = 1;
2259 	}
2260 	if ((tp->snd_wnd == 0) && TCPS_HAVEESTABLISHED(tp->t_state)) {
2261 		/* No send window.. we must enter persist */
2262 		rack_enter_persist(tp, rack, cts);
2263 	} else if ((frm_out_sbavail &&
2264 		    (frm_out_sbavail > (tp->snd_max - tp->snd_una)) &&
2265 		    (tp->snd_wnd < tp->t_maxseg)) &&
2266 	    TCPS_HAVEESTABLISHED(tp->t_state)) {
2267 		/*
2268 		 * If we have no window or we can't send a segment (and have
2269 		 * data to send.. we cheat here and frm_out_sbavail is
2270 		 * passed in with the sbavail(sb) only from bbr_output) and
2271 		 * we are established, then we must enter persits (if not
2272 		 * already in persits).
2273 		 */
2274 		rack_enter_persist(tp, rack, cts);
2275 	}
2276 	hpts_timeout = rack_timer_start(tp, rack, cts);
2277 	if (tp->t_flags & TF_DELACK) {
2278 		delayed_ack = tcp_delacktime;
2279 		rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK;
2280 	}
2281 	if (delayed_ack && ((hpts_timeout == 0) ||
2282 			    (delayed_ack < hpts_timeout)))
2283 		hpts_timeout = delayed_ack;
2284 	else
2285 		rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
2286 	/*
2287 	 * If no timers are going to run and we will fall off the hptsi
2288 	 * wheel, we resort to a keep-alive timer if its configured.
2289 	 */
2290 	if ((hpts_timeout == 0) &&
2291 	    (slot == 0)) {
2292 		if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
2293 		    (tp->t_state <= TCPS_CLOSING)) {
2294 			/*
2295 			 * Ok we have no timer (persists, rack, tlp, rxt  or
2296 			 * del-ack), we don't have segments being paced. So
2297 			 * all that is left is the keepalive timer.
2298 			 */
2299 			if (TCPS_HAVEESTABLISHED(tp->t_state)) {
2300 				/* Get the established keep-alive time */
2301 				hpts_timeout = TP_KEEPIDLE(tp);
2302 			} else {
2303 				/* Get the initial setup keep-alive time */
2304 				hpts_timeout = TP_KEEPINIT(tp);
2305 			}
2306 			rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP;
2307 		}
2308 	}
2309 	if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) ==
2310 	    (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) {
2311 		/*
2312 		 * RACK, TLP, persists and RXT timers all are restartable
2313 		 * based on actions input .. i.e we received a packet (ack
2314 		 * or sack) and that changes things (rw, or snd_una etc).
2315 		 * Thus we can restart them with a new value. For
2316 		 * keep-alive, delayed_ack we keep track of what was left
2317 		 * and restart the timer with a smaller value.
2318 		 */
2319 		if (left < hpts_timeout)
2320 			hpts_timeout = left;
2321 	}
2322 	if (hpts_timeout) {
2323 		/*
2324 		 * Hack alert for now we can't time-out over 2,147,483
2325 		 * seconds (a bit more than 596 hours), which is probably ok
2326 		 * :).
2327 		 */
2328 		if (hpts_timeout > 0x7ffffffe)
2329 			hpts_timeout = 0x7ffffffe;
2330 		rack->r_ctl.rc_timer_exp = cts + hpts_timeout;
2331 	}
2332 	if (slot) {
2333 		rack->r_ctl.rc_last_output_to = cts + slot;
2334 		if ((hpts_timeout == 0) || (hpts_timeout > slot)) {
2335 			if (rack->rc_inp->inp_in_hpts == 0)
2336 				tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(slot));
2337 			rack_log_to_start(rack, cts, hpts_timeout, slot, 1);
2338 		} else {
2339 			/*
2340 			 * Arrange for the hpts to kick back in after the
2341 			 * t-o if the t-o does not cause a send.
2342 			 */
2343 			if (rack->rc_inp->inp_in_hpts == 0)
2344 				tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout));
2345 			rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
2346 		}
2347 	} else if (hpts_timeout) {
2348 		if (rack->rc_inp->inp_in_hpts == 0)
2349 			tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout));
2350 		rack_log_to_start(rack, cts, hpts_timeout, slot, 0);
2351 	} else {
2352 		/* No timer starting */
2353 #ifdef INVARIANTS
2354 		if (SEQ_GT(tp->snd_max, tp->snd_una)) {
2355 			panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?",
2356 			    tp, rack, tot_len_this_send, cts, slot, hpts_timeout);
2357 		}
2358 #endif
2359 	}
2360 	rack->rc_tmr_stopped = 0;
2361 	if (slot)
2362 		rack_log_type_bbrsnd(rack, tot_len_this_send, slot, cts);
2363 }
2364 
2365 /*
2366  * RACK Timer, here we simply do logging and house keeping.
2367  * the normal rack_output() function will call the
2368  * appropriate thing to check if we need to do a RACK retransmit.
2369  * We return 1, saying don't proceed with rack_output only
2370  * when all timers have been stopped (destroyed PCB?).
2371  */
2372 static int
2373 rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
2374 {
2375 	/*
2376 	 * This timer simply provides an internal trigger to send out data.
2377 	 * The check_recovery_mode call will see if there are needed
2378 	 * retransmissions, if so we will enter fast-recovery. The output
2379 	 * call may or may not do the same thing depending on sysctl
2380 	 * settings.
2381 	 */
2382 	struct rack_sendmap *rsm;
2383 	int32_t recovery;
2384 
2385 	if (tp->t_timers->tt_flags & TT_STOPPED) {
2386 		return (1);
2387 	}
2388 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
2389 		/* Its not time yet */
2390 		return (0);
2391 	}
2392 	rack_log_to_event(rack, RACK_TO_FRM_RACK);
2393 	recovery = IN_RECOVERY(tp->t_flags);
2394 	counter_u64_add(rack_to_tot, 1);
2395 	if (rack->r_state && (rack->r_state != tp->t_state))
2396 		rack_set_state(tp, rack);
2397 	rsm = rack_check_recovery_mode(tp, cts);
2398 	if (rsm) {
2399 		uint32_t rtt;
2400 
2401 		rtt = rack->rc_rack_rtt;
2402 		if (rtt == 0)
2403 			rtt = 1;
2404 		if ((recovery == 0) &&
2405 		    (rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg)) {
2406 			/*
2407 			 * The rack-timeout that enter's us into recovery
2408 			 * will force out one MSS and set us up so that we
2409 			 * can do one more send in 2*rtt (transitioning the
2410 			 * rack timeout into a rack-tlp).
2411 			 */
2412 			rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg;
2413 		} else if ((rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg) &&
2414 		    ((rsm->r_end - rsm->r_start) > rack->r_ctl.rc_prr_sndcnt)) {
2415 			/*
2416 			 * When a rack timer goes, we have to send at
2417 			 * least one segment. They will be paced a min of 1ms
2418 			 * apart via the next rack timer (or further
2419 			 * if the rack timer dictates it).
2420 			 */
2421 			rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg;
2422 		}
2423 	} else {
2424 		/* This is a case that should happen rarely if ever */
2425 		counter_u64_add(rack_tlp_does_nada, 1);
2426 #ifdef TCP_BLACKBOX
2427 		tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
2428 #endif
2429 		rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
2430 	}
2431 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK;
2432 	return (0);
2433 }
2434 
2435 /*
2436  * TLP Timer, here we simply setup what segment we want to
2437  * have the TLP expire on, the normal rack_output() will then
2438  * send it out.
2439  *
2440  * We return 1, saying don't proceed with rack_output only
2441  * when all timers have been stopped (destroyed PCB?).
2442  */
2443 static int
2444 rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
2445 {
2446 	/*
2447 	 * Tail Loss Probe.
2448 	 */
2449 	struct rack_sendmap *rsm = NULL;
2450 	struct socket *so;
2451 	uint32_t amm, old_prr_snd = 0;
2452 	uint32_t out, avail;
2453 
2454 	if (tp->t_timers->tt_flags & TT_STOPPED) {
2455 		return (1);
2456 	}
2457 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
2458 		/* Its not time yet */
2459 		return (0);
2460 	}
2461 	if (rack_progress_timeout_check(tp)) {
2462 		tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
2463 		return (1);
2464 	}
2465 	/*
2466 	 * A TLP timer has expired. We have been idle for 2 rtts. So we now
2467 	 * need to figure out how to force a full MSS segment out.
2468 	 */
2469 	rack_log_to_event(rack, RACK_TO_FRM_TLP);
2470 	counter_u64_add(rack_tlp_tot, 1);
2471 	if (rack->r_state && (rack->r_state != tp->t_state))
2472 		rack_set_state(tp, rack);
2473 	so = tp->t_inpcb->inp_socket;
2474 	avail = sbavail(&so->so_snd);
2475 	out = tp->snd_max - tp->snd_una;
2476 	rack->rc_timer_up = 1;
2477 	/*
2478 	 * If we are in recovery we can jazz out a segment if new data is
2479 	 * present simply by setting rc_prr_sndcnt to a segment.
2480 	 */
2481 	if ((avail > out) &&
2482 	    ((rack_always_send_oldest == 0) || (TAILQ_EMPTY(&rack->r_ctl.rc_tmap)))) {
2483 		/* New data is available */
2484 		amm = avail - out;
2485 		if (amm > tp->t_maxseg) {
2486 			amm = tp->t_maxseg;
2487 		} else if ((amm < tp->t_maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) {
2488 			/* not enough to fill a MTU and no-delay is off */
2489 			goto need_retran;
2490 		}
2491 		if (IN_RECOVERY(tp->t_flags)) {
2492 			/* Unlikely */
2493 			old_prr_snd = rack->r_ctl.rc_prr_sndcnt;
2494 			if (out + amm <= tp->snd_wnd)
2495 				rack->r_ctl.rc_prr_sndcnt = amm;
2496 			else
2497 				goto need_retran;
2498 		} else {
2499 			/* Set the send-new override */
2500 			if (out + amm <= tp->snd_wnd)
2501 				rack->r_ctl.rc_tlp_new_data = amm;
2502 			else
2503 				goto need_retran;
2504 		}
2505 		rack->r_ctl.rc_tlp_seg_send_cnt = 0;
2506 		rack->r_ctl.rc_last_tlp_seq = tp->snd_max;
2507 		rack->r_ctl.rc_tlpsend = NULL;
2508 		counter_u64_add(rack_tlp_newdata, 1);
2509 		goto send;
2510 	}
2511 need_retran:
2512 	/*
2513 	 * Ok we need to arrange the last un-acked segment to be re-sent, or
2514 	 * optionally the first un-acked segment.
2515 	 */
2516 	if (rack_always_send_oldest)
2517 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
2518 	else {
2519 		rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next);
2520 		if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) {
2521 			rsm = rack_find_high_nonack(rack, rsm);
2522 		}
2523 	}
2524 	if (rsm == NULL) {
2525 		counter_u64_add(rack_tlp_does_nada, 1);
2526 #ifdef TCP_BLACKBOX
2527 		tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true);
2528 #endif
2529 		goto out;
2530 	}
2531 	if ((rsm->r_end - rsm->r_start) > tp->t_maxseg) {
2532 		/*
2533 		 * We need to split this the last segment in two.
2534 		 */
2535 		int32_t idx;
2536 		struct rack_sendmap *nrsm;
2537 
2538 		nrsm = rack_alloc(rack);
2539 		if (nrsm == NULL) {
2540 			/*
2541 			 * No memory to split, we will just exit and punt
2542 			 * off to the RXT timer.
2543 			 */
2544 			counter_u64_add(rack_tlp_does_nada, 1);
2545 			goto out;
2546 		}
2547 		nrsm->r_start = (rsm->r_end - tp->t_maxseg);
2548 		nrsm->r_end = rsm->r_end;
2549 		nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
2550 		nrsm->r_flags = rsm->r_flags;
2551 		nrsm->r_sndcnt = rsm->r_sndcnt;
2552 		nrsm->r_rtr_bytes = 0;
2553 		rsm->r_end = nrsm->r_start;
2554 		for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
2555 			nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
2556 		}
2557 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next);
2558 		if (rsm->r_in_tmap) {
2559 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
2560 			nrsm->r_in_tmap = 1;
2561 		}
2562 		rsm->r_flags &= (~RACK_HAS_FIN);
2563 		rsm = nrsm;
2564 	}
2565 	rack->r_ctl.rc_tlpsend = rsm;
2566 	rack->r_ctl.rc_tlp_rtx_out = 1;
2567 	if (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) {
2568 		rack->r_ctl.rc_tlp_seg_send_cnt++;
2569 		tp->t_rxtshift++;
2570 	} else {
2571 		rack->r_ctl.rc_last_tlp_seq = rsm->r_start;
2572 		rack->r_ctl.rc_tlp_seg_send_cnt = 1;
2573 	}
2574 send:
2575 	rack->r_ctl.rc_tlp_send_cnt++;
2576 	if (rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) {
2577 		/*
2578 		 * Can't [re]/transmit a segment we have not heard from the
2579 		 * peer in max times. We need the retransmit timer to take
2580 		 * over.
2581 		 */
2582 restore:
2583 		rack->r_ctl.rc_tlpsend = NULL;
2584 		if (rsm)
2585 			rsm->r_flags &= ~RACK_TLP;
2586 		rack->r_ctl.rc_prr_sndcnt = old_prr_snd;
2587 		counter_u64_add(rack_tlp_retran_fail, 1);
2588 		goto out;
2589 	} else if (rsm) {
2590 		rsm->r_flags |= RACK_TLP;
2591 	}
2592 	if (rsm && (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) &&
2593 	    (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) {
2594 		/*
2595 		 * We don't want to send a single segment more than the max
2596 		 * either.
2597 		 */
2598 		goto restore;
2599 	}
2600 	rack->r_timer_override = 1;
2601 	rack->r_tlp_running = 1;
2602 	rack->rc_tlp_in_progress = 1;
2603 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
2604 	return (0);
2605 out:
2606 	rack->rc_timer_up = 0;
2607 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP;
2608 	return (0);
2609 }
2610 
2611 /*
2612  * Delayed ack Timer, here we simply need to setup the
2613  * ACK_NOW flag and remove the DELACK flag. From there
2614  * the output routine will send the ack out.
2615  *
2616  * We only return 1, saying don't proceed, if all timers
2617  * are stopped (destroyed PCB?).
2618  */
2619 static int
2620 rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
2621 {
2622 	if (tp->t_timers->tt_flags & TT_STOPPED) {
2623 		return (1);
2624 	}
2625 	rack_log_to_event(rack, RACK_TO_FRM_DELACK);
2626 	tp->t_flags &= ~TF_DELACK;
2627 	tp->t_flags |= TF_ACKNOW;
2628 	TCPSTAT_INC(tcps_delack);
2629 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK;
2630 	return (0);
2631 }
2632 
2633 /*
2634  * Persists timer, here we simply need to setup the
2635  * FORCE-DATA flag the output routine will send
2636  * the one byte send.
2637  *
2638  * We only return 1, saying don't proceed, if all timers
2639  * are stopped (destroyed PCB?).
2640  */
2641 static int
2642 rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
2643 {
2644 	struct inpcb *inp;
2645 	int32_t retval = 0;
2646 
2647 	inp = tp->t_inpcb;
2648 
2649 	if (tp->t_timers->tt_flags & TT_STOPPED) {
2650 		return (1);
2651 	}
2652 	if (rack->rc_in_persist == 0)
2653 		return (0);
2654 	if (rack_progress_timeout_check(tp)) {
2655 		tcp_set_inp_to_drop(inp, ETIMEDOUT);
2656 		return (1);
2657 	}
2658 	KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp));
2659 	/*
2660 	 * Persistence timer into zero window. Force a byte to be output, if
2661 	 * possible.
2662 	 */
2663 	TCPSTAT_INC(tcps_persisttimeo);
2664 	/*
2665 	 * Hack: if the peer is dead/unreachable, we do not time out if the
2666 	 * window is closed.  After a full backoff, drop the connection if
2667 	 * the idle time (no responses to probes) reaches the maximum
2668 	 * backoff that we would use if retransmitting.
2669 	 */
2670 	if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
2671 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
2672 	    ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) {
2673 		TCPSTAT_INC(tcps_persistdrop);
2674 		retval = 1;
2675 		tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
2676 		goto out;
2677 	}
2678 	if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) &&
2679 	    tp->snd_una == tp->snd_max)
2680 		rack_exit_persist(tp, rack);
2681 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT;
2682 	/*
2683 	 * If the user has closed the socket then drop a persisting
2684 	 * connection after a much reduced timeout.
2685 	 */
2686 	if (tp->t_state > TCPS_CLOSE_WAIT &&
2687 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
2688 		retval = 1;
2689 		TCPSTAT_INC(tcps_persistdrop);
2690 		tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
2691 		goto out;
2692 	}
2693 	tp->t_flags |= TF_FORCEDATA;
2694 out:
2695 	rack_log_to_event(rack, RACK_TO_FRM_PERSIST);
2696 	return (retval);
2697 }
2698 
2699 /*
2700  * If a keepalive goes off, we had no other timers
2701  * happening. We always return 1 here since this
2702  * routine either drops the connection or sends
2703  * out a segment with respond.
2704  */
2705 static int
2706 rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
2707 {
2708 	struct tcptemp *t_template;
2709 	struct inpcb *inp;
2710 
2711 	if (tp->t_timers->tt_flags & TT_STOPPED) {
2712 		return (1);
2713 	}
2714 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP;
2715 	inp = tp->t_inpcb;
2716 	rack_log_to_event(rack, RACK_TO_FRM_KEEP);
2717 	/*
2718 	 * Keep-alive timer went off; send something or drop connection if
2719 	 * idle for too long.
2720 	 */
2721 	TCPSTAT_INC(tcps_keeptimeo);
2722 	if (tp->t_state < TCPS_ESTABLISHED)
2723 		goto dropit;
2724 	if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
2725 	    tp->t_state <= TCPS_CLOSING) {
2726 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
2727 			goto dropit;
2728 		/*
2729 		 * Send a packet designed to force a response if the peer is
2730 		 * up and reachable: either an ACK if the connection is
2731 		 * still alive, or an RST if the peer has closed the
2732 		 * connection due to timeout or reboot. Using sequence
2733 		 * number tp->snd_una-1 causes the transmitted zero-length
2734 		 * segment to lie outside the receive window; by the
2735 		 * protocol spec, this requires the correspondent TCP to
2736 		 * respond.
2737 		 */
2738 		TCPSTAT_INC(tcps_keepprobe);
2739 		t_template = tcpip_maketemplate(inp);
2740 		if (t_template) {
2741 			tcp_respond(tp, t_template->tt_ipgen,
2742 			    &t_template->tt_t, (struct mbuf *)NULL,
2743 			    tp->rcv_nxt, tp->snd_una - 1, 0);
2744 			free(t_template, M_TEMP);
2745 		}
2746 	}
2747 	rack_start_hpts_timer(rack, tp, cts, __LINE__, 0, 0, 0);
2748 	return (1);
2749 dropit:
2750 	TCPSTAT_INC(tcps_keepdrops);
2751 	tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT);
2752 	return (1);
2753 }
2754 
2755 /*
2756  * Retransmit helper function, clear up all the ack
2757  * flags and take care of important book keeping.
2758  */
2759 static void
2760 rack_remxt_tmr(struct tcpcb *tp)
2761 {
2762 	/*
2763 	 * The retransmit timer went off, all sack'd blocks must be
2764 	 * un-acked.
2765 	 */
2766 	struct rack_sendmap *rsm, *trsm = NULL;
2767 	struct tcp_rack *rack;
2768 	int32_t cnt = 0;
2769 
2770 	rack = (struct tcp_rack *)tp->t_fb_ptr;
2771 	rack_timer_cancel(tp, rack, tcp_ts_getticks(), __LINE__);
2772 	rack_log_to_event(rack, RACK_TO_FRM_TMR);
2773 	if (rack->r_state && (rack->r_state != tp->t_state))
2774 		rack_set_state(tp, rack);
2775 	/*
2776 	 * Ideally we would like to be able to
2777 	 * mark SACK-PASS on anything not acked here.
2778 	 * However, if we do that we would burst out
2779 	 * all that data 1ms apart. This would be unwise,
2780 	 * so for now we will just let the normal rxt timer
2781 	 * and tlp timer take care of it.
2782 	 */
2783 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) {
2784 		if (rsm->r_flags & RACK_ACKED) {
2785 			cnt++;
2786 			rsm->r_sndcnt = 0;
2787 			if (rsm->r_in_tmap == 0) {
2788 				/* We must re-add it back to the tlist */
2789 				if (trsm == NULL) {
2790 					TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
2791 				} else {
2792 					TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext);
2793 				}
2794 				rsm->r_in_tmap = 1;
2795 				trsm = rsm;
2796 			}
2797 		}
2798 		rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS);
2799 	}
2800 	/* Clear the count (we just un-acked them) */
2801 	rack->r_ctl.rc_sacked = 0;
2802 	/* Clear the tlp rtx mark */
2803 	rack->r_ctl.rc_tlp_rtx_out = 0;
2804 	rack->r_ctl.rc_tlp_seg_send_cnt = 0;
2805 	rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_map);
2806 	/* Setup so we send one segment */
2807 	if (rack->r_ctl.rc_prr_sndcnt < tp->t_maxseg)
2808 		rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg;
2809 	rack->r_timer_override = 1;
2810 }
2811 
2812 /*
2813  * Re-transmit timeout! If we drop the PCB we will return 1, otherwise
2814  * we will setup to retransmit the lowest seq number outstanding.
2815  */
2816 static int
2817 rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts)
2818 {
2819 	int32_t rexmt;
2820 	struct inpcb *inp;
2821 	int32_t retval = 0;
2822 
2823 	inp = tp->t_inpcb;
2824 	if (tp->t_timers->tt_flags & TT_STOPPED) {
2825 		return (1);
2826 	}
2827 	if (rack_progress_timeout_check(tp)) {
2828 		tcp_set_inp_to_drop(inp, ETIMEDOUT);
2829 		return (1);
2830 	}
2831 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT;
2832 	if (TCPS_HAVEESTABLISHED(tp->t_state) &&
2833 	    (tp->snd_una == tp->snd_max)) {
2834 		/* Nothing outstanding .. nothing to do */
2835 		return (0);
2836 	}
2837 	/*
2838 	 * Retransmission timer went off.  Message has not been acked within
2839 	 * retransmit interval.  Back off to a longer retransmit interval
2840 	 * and retransmit one segment.
2841 	 */
2842 	if (++tp->t_rxtshift > TCP_MAXRXTSHIFT) {
2843 		tp->t_rxtshift = TCP_MAXRXTSHIFT;
2844 		TCPSTAT_INC(tcps_timeoutdrop);
2845 		retval = 1;
2846 		tcp_set_inp_to_drop(rack->rc_inp,
2847 		    (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT));
2848 		goto out;
2849 	}
2850 	rack_remxt_tmr(tp);
2851 	if (tp->t_state == TCPS_SYN_SENT) {
2852 		/*
2853 		 * If the SYN was retransmitted, indicate CWND to be limited
2854 		 * to 1 segment in cc_conn_init().
2855 		 */
2856 		tp->snd_cwnd = 1;
2857 	} else if (tp->t_rxtshift == 1) {
2858 		/*
2859 		 * first retransmit; record ssthresh and cwnd so they can be
2860 		 * recovered if this turns out to be a "bad" retransmit. A
2861 		 * retransmit is considered "bad" if an ACK for this segment
2862 		 * is received within RTT/2 interval; the assumption here is
2863 		 * that the ACK was already in flight.  See "On Estimating
2864 		 * End-to-End Network Path Properties" by Allman and Paxson
2865 		 * for more details.
2866 		 */
2867 		tp->snd_cwnd_prev = tp->snd_cwnd;
2868 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
2869 		tp->snd_recover_prev = tp->snd_recover;
2870 		if (IN_FASTRECOVERY(tp->t_flags))
2871 			tp->t_flags |= TF_WASFRECOVERY;
2872 		else
2873 			tp->t_flags &= ~TF_WASFRECOVERY;
2874 		if (IN_CONGRECOVERY(tp->t_flags))
2875 			tp->t_flags |= TF_WASCRECOVERY;
2876 		else
2877 			tp->t_flags &= ~TF_WASCRECOVERY;
2878 		tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
2879 		tp->t_flags |= TF_PREVVALID;
2880 	} else
2881 		tp->t_flags &= ~TF_PREVVALID;
2882 	TCPSTAT_INC(tcps_rexmttimeo);
2883 	if ((tp->t_state == TCPS_SYN_SENT) ||
2884 	    (tp->t_state == TCPS_SYN_RECEIVED))
2885 		rexmt = MSEC_2_TICKS(RACK_INITIAL_RTO * tcp_syn_backoff[tp->t_rxtshift]);
2886 	else
2887 		rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
2888 	TCPT_RANGESET(tp->t_rxtcur, rexmt,
2889 	   max(MSEC_2_TICKS(rack_rto_min), rexmt),
2890 	   MSEC_2_TICKS(rack_rto_max));
2891 	/*
2892 	 * We enter the path for PLMTUD if connection is established or, if
2893 	 * connection is FIN_WAIT_1 status, reason for the last is that if
2894 	 * amount of data we send is very small, we could send it in couple
2895 	 * of packets and process straight to FIN. In that case we won't
2896 	 * catch ESTABLISHED state.
2897 	 */
2898 	if (V_tcp_pmtud_blackhole_detect && (((tp->t_state == TCPS_ESTABLISHED))
2899 	    || (tp->t_state == TCPS_FIN_WAIT_1))) {
2900 #ifdef INET6
2901 		int32_t isipv6;
2902 #endif
2903 
2904 		/*
2905 		 * Idea here is that at each stage of mtu probe (usually,
2906 		 * 1448 -> 1188 -> 524) should be given 2 chances to recover
2907 		 * before further clamping down. 'tp->t_rxtshift % 2 == 0'
2908 		 * should take care of that.
2909 		 */
2910 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) ==
2911 		    (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) &&
2912 		    (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 &&
2913 		    tp->t_rxtshift % 2 == 0)) {
2914 			/*
2915 			 * Enter Path MTU Black-hole Detection mechanism: -
2916 			 * Disable Path MTU Discovery (IP "DF" bit). -
2917 			 * Reduce MTU to lower value than what we negotiated
2918 			 * with peer.
2919 			 */
2920 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
2921 				/* Record that we may have found a black hole. */
2922 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
2923 				/* Keep track of previous MSS. */
2924 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
2925 			}
2926 
2927 			/*
2928 			 * Reduce the MSS to blackhole value or to the
2929 			 * default in an attempt to retransmit.
2930 			 */
2931 #ifdef INET6
2932 			isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? 1 : 0;
2933 			if (isipv6 &&
2934 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) {
2935 				/* Use the sysctl tuneable blackhole MSS. */
2936 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
2937 				TCPSTAT_INC(tcps_pmtud_blackhole_activated);
2938 			} else if (isipv6) {
2939 				/* Use the default MSS. */
2940 				tp->t_maxseg = V_tcp_v6mssdflt;
2941 				/*
2942 				 * Disable Path MTU Discovery when we switch
2943 				 * to minmss.
2944 				 */
2945 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
2946 				TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
2947 			}
2948 #endif
2949 #if defined(INET6) && defined(INET)
2950 			else
2951 #endif
2952 #ifdef INET
2953 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) {
2954 				/* Use the sysctl tuneable blackhole MSS. */
2955 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
2956 				TCPSTAT_INC(tcps_pmtud_blackhole_activated);
2957 			} else {
2958 				/* Use the default MSS. */
2959 				tp->t_maxseg = V_tcp_mssdflt;
2960 				/*
2961 				 * Disable Path MTU Discovery when we switch
2962 				 * to minmss.
2963 				 */
2964 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
2965 				TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
2966 			}
2967 #endif
2968 		} else {
2969 			/*
2970 			 * If further retransmissions are still unsuccessful
2971 			 * with a lowered MTU, maybe this isn't a blackhole
2972 			 * and we restore the previous MSS and blackhole
2973 			 * detection flags. The limit '6' is determined by
2974 			 * giving each probe stage (1448, 1188, 524) 2
2975 			 * chances to recover.
2976 			 */
2977 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
2978 			    (tp->t_rxtshift >= 6)) {
2979 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
2980 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
2981 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
2982 				TCPSTAT_INC(tcps_pmtud_blackhole_failed);
2983 			}
2984 		}
2985 	}
2986 	/*
2987 	 * Disable RFC1323 and SACK if we haven't got any response to our
2988 	 * third SYN to work-around some broken terminal servers (most of
2989 	 * which have hopefully been retired) that have bad VJ header
2990 	 * compression code which trashes TCP segments containing
2991 	 * unknown-to-them TCP options.
2992 	 */
2993 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
2994 	    (tp->t_rxtshift == 3))
2995 		tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT);
2996 	/*
2997 	 * If we backed off this far, our srtt estimate is probably bogus.
2998 	 * Clobber it so we'll take the next rtt measurement as our srtt;
2999 	 * move the current srtt into rttvar to keep the current retransmit
3000 	 * times until then.
3001 	 */
3002 	if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
3003 #ifdef INET6
3004 		if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
3005 			in6_losing(tp->t_inpcb);
3006 		else
3007 #endif
3008 			in_losing(tp->t_inpcb);
3009 		tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT);
3010 		tp->t_srtt = 0;
3011 	}
3012 	if (rack_use_sack_filter)
3013 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
3014 	tp->snd_recover = tp->snd_max;
3015 	tp->t_flags |= TF_ACKNOW;
3016 	tp->t_rtttime = 0;
3017 	rack_cong_signal(tp, NULL, CC_RTO);
3018 out:
3019 	return (retval);
3020 }
3021 
3022 static int
3023 rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling)
3024 {
3025 	int32_t ret = 0;
3026 	int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK);
3027 
3028 	if (timers == 0) {
3029 		return (0);
3030 	}
3031 	if (tp->t_state == TCPS_LISTEN) {
3032 		/* no timers on listen sockets */
3033 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)
3034 			return (0);
3035 		return (1);
3036 	}
3037 	if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) {
3038 		uint32_t left;
3039 
3040 		if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) {
3041 			ret = -1;
3042 			rack_log_to_processing(rack, cts, ret, 0);
3043 			return (0);
3044 		}
3045 		if (hpts_calling == 0) {
3046 			ret = -2;
3047 			rack_log_to_processing(rack, cts, ret, 0);
3048 			return (0);
3049 		}
3050 		/*
3051 		 * Ok our timer went off early and we are not paced false
3052 		 * alarm, go back to sleep.
3053 		 */
3054 		ret = -3;
3055 		left = rack->r_ctl.rc_timer_exp - cts;
3056 		tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left));
3057 		rack_log_to_processing(rack, cts, ret, left);
3058 		rack->rc_last_pto_set = 0;
3059 		return (1);
3060 	}
3061 	rack->rc_tmr_stopped = 0;
3062 	rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK;
3063 	if (timers & PACE_TMR_DELACK) {
3064 		ret = rack_timeout_delack(tp, rack, cts);
3065 	} else if (timers & PACE_TMR_RACK) {
3066 		ret = rack_timeout_rack(tp, rack, cts);
3067 	} else if (timers & PACE_TMR_TLP) {
3068 		ret = rack_timeout_tlp(tp, rack, cts);
3069 	} else if (timers & PACE_TMR_RXT) {
3070 		ret = rack_timeout_rxt(tp, rack, cts);
3071 	} else if (timers & PACE_TMR_PERSIT) {
3072 		ret = rack_timeout_persist(tp, rack, cts);
3073 	} else if (timers & PACE_TMR_KEEP) {
3074 		ret = rack_timeout_keepalive(tp, rack, cts);
3075 	}
3076 	rack_log_to_processing(rack, cts, ret, timers);
3077 	return (ret);
3078 }
3079 
3080 static void
3081 rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line)
3082 {
3083 	uint8_t hpts_removed = 0;
3084 
3085 	if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) &&
3086 	    TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) {
3087 		tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
3088 		hpts_removed = 1;
3089 	}
3090 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
3091 		rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
3092 		if (rack->rc_inp->inp_in_hpts &&
3093 		    ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) {
3094 			/*
3095 			 * Canceling timer's when we have no output being
3096 			 * paced. We also must remove ourselves from the
3097 			 * hpts.
3098 			 */
3099 			tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
3100 			hpts_removed = 1;
3101 		}
3102 		rack_log_to_cancel(rack, hpts_removed, line);
3103 		rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK);
3104 	}
3105 }
3106 
3107 static void
3108 rack_timer_stop(struct tcpcb *tp, uint32_t timer_type)
3109 {
3110 	return;
3111 }
3112 
3113 static int
3114 rack_stopall(struct tcpcb *tp)
3115 {
3116 	struct tcp_rack *rack;
3117 	rack = (struct tcp_rack *)tp->t_fb_ptr;
3118 	rack->t_timers_stopped = 1;
3119 	return (0);
3120 }
3121 
3122 static void
3123 rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta)
3124 {
3125 	return;
3126 }
3127 
3128 static int
3129 rack_timer_active(struct tcpcb *tp, uint32_t timer_type)
3130 {
3131 	return (0);
3132 }
3133 
3134 static void
3135 rack_stop_all_timers(struct tcpcb *tp)
3136 {
3137 	struct tcp_rack *rack;
3138 
3139 	/*
3140 	 * Assure no timers are running.
3141 	 */
3142 	if (tcp_timer_active(tp, TT_PERSIST)) {
3143 		/* We enter in persists, set the flag appropriately */
3144 		rack = (struct tcp_rack *)tp->t_fb_ptr;
3145 		rack->rc_in_persist = 1;
3146 	}
3147 	tcp_timer_suspend(tp, TT_PERSIST);
3148 	tcp_timer_suspend(tp, TT_REXMT);
3149 	tcp_timer_suspend(tp, TT_KEEP);
3150 	tcp_timer_suspend(tp, TT_DELACK);
3151 }
3152 
3153 static void
3154 rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack,
3155     struct rack_sendmap *rsm, uint32_t ts)
3156 {
3157 	int32_t idx;
3158 
3159 	rsm->r_rtr_cnt++;
3160 	rsm->r_sndcnt++;
3161 	if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) {
3162 		rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS;
3163 		rsm->r_flags |= RACK_OVERMAX;
3164 	}
3165 	if ((rsm->r_rtr_cnt > 1) && (rack->r_tlp_running == 0)) {
3166 		rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start);
3167 		rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start);
3168 	}
3169 	idx = rsm->r_rtr_cnt - 1;
3170 	rsm->r_tim_lastsent[idx] = ts;
3171 	if (rsm->r_flags & RACK_ACKED) {
3172 		/* Problably MTU discovery messing with us */
3173 		rsm->r_flags &= ~RACK_ACKED;
3174 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
3175 	}
3176 	if (rsm->r_in_tmap) {
3177 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
3178 	}
3179 	TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
3180 	rsm->r_in_tmap = 1;
3181 	if (rsm->r_flags & RACK_SACK_PASSED) {
3182 		/* We have retransmitted due to the SACK pass */
3183 		rsm->r_flags &= ~RACK_SACK_PASSED;
3184 		rsm->r_flags |= RACK_WAS_SACKPASS;
3185 	}
3186 	/* Update memory for next rtr */
3187 	rack->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
3188 }
3189 
3190 
3191 static uint32_t
3192 rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack,
3193     struct rack_sendmap *rsm, uint32_t ts, int32_t * lenp)
3194 {
3195 	/*
3196 	 * We (re-)transmitted starting at rsm->r_start for some length
3197 	 * (possibly less than r_end.
3198 	 */
3199 	struct rack_sendmap *nrsm;
3200 	uint32_t c_end;
3201 	int32_t len;
3202 	int32_t idx;
3203 
3204 	len = *lenp;
3205 	c_end = rsm->r_start + len;
3206 	if (SEQ_GEQ(c_end, rsm->r_end)) {
3207 		/*
3208 		 * We retransmitted the whole piece or more than the whole
3209 		 * slopping into the next rsm.
3210 		 */
3211 		rack_update_rsm(tp, rack, rsm, ts);
3212 		if (c_end == rsm->r_end) {
3213 			*lenp = 0;
3214 			return (0);
3215 		} else {
3216 			int32_t act_len;
3217 
3218 			/* Hangs over the end return whats left */
3219 			act_len = rsm->r_end - rsm->r_start;
3220 			*lenp = (len - act_len);
3221 			return (rsm->r_end);
3222 		}
3223 		/* We don't get out of this block. */
3224 	}
3225 	/*
3226 	 * Here we retransmitted less than the whole thing which means we
3227 	 * have to split this into what was transmitted and what was not.
3228 	 */
3229 	nrsm = rack_alloc(rack);
3230 	if (nrsm == NULL) {
3231 		/*
3232 		 * We can't get memory, so lets not proceed.
3233 		 */
3234 		*lenp = 0;
3235 		return (0);
3236 	}
3237 	/*
3238 	 * So here we are going to take the original rsm and make it what we
3239 	 * retransmitted. nrsm will be the tail portion we did not
3240 	 * retransmit. For example say the chunk was 1, 11 (10 bytes). And
3241 	 * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to
3242 	 * 1, 6 and the new piece will be 6, 11.
3243 	 */
3244 	nrsm->r_start = c_end;
3245 	nrsm->r_end = rsm->r_end;
3246 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
3247 	nrsm->r_flags = rsm->r_flags;
3248 	nrsm->r_sndcnt = rsm->r_sndcnt;
3249 	nrsm->r_rtr_bytes = 0;
3250 	rsm->r_end = c_end;
3251 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
3252 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
3253 	}
3254 	TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next);
3255 	if (rsm->r_in_tmap) {
3256 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
3257 		nrsm->r_in_tmap = 1;
3258 	}
3259 	rsm->r_flags &= (~RACK_HAS_FIN);
3260 	rack_update_rsm(tp, rack, rsm, ts);
3261 	*lenp = 0;
3262 	return (0);
3263 }
3264 
3265 
3266 static void
3267 rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len,
3268     uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts,
3269     uint8_t pass, struct rack_sendmap *hintrsm)
3270 {
3271 	struct tcp_rack *rack;
3272 	struct rack_sendmap *rsm, *nrsm;
3273 	register uint32_t snd_max, snd_una;
3274 	int32_t idx;
3275 
3276 	/*
3277 	 * Add to the RACK log of packets in flight or retransmitted. If
3278 	 * there is a TS option we will use the TS echoed, if not we will
3279 	 * grab a TS.
3280 	 *
3281 	 * Retransmissions will increment the count and move the ts to its
3282 	 * proper place. Note that if options do not include TS's then we
3283 	 * won't be able to effectively use the ACK for an RTT on a retran.
3284 	 *
3285 	 * Notes about r_start and r_end. Lets consider a send starting at
3286 	 * sequence 1 for 10 bytes. In such an example the r_start would be
3287 	 * 1 (starting sequence) but the r_end would be r_start+len i.e. 11.
3288 	 * This means that r_end is actually the first sequence for the next
3289 	 * slot (11).
3290 	 *
3291 	 */
3292 	/*
3293 	 * If err is set what do we do XXXrrs? should we not add the thing?
3294 	 * -- i.e. return if err != 0 or should we pretend we sent it? --
3295 	 * i.e. proceed with add ** do this for now.
3296 	 */
3297 	INP_WLOCK_ASSERT(tp->t_inpcb);
3298 	if (err)
3299 		/*
3300 		 * We don't log errors -- we could but snd_max does not
3301 		 * advance in this case either.
3302 		 */
3303 		return;
3304 
3305 	if (th_flags & TH_RST) {
3306 		/*
3307 		 * We don't log resets and we return immediately from
3308 		 * sending
3309 		 */
3310 		return;
3311 	}
3312 	rack = (struct tcp_rack *)tp->t_fb_ptr;
3313 	snd_una = tp->snd_una;
3314 	if (SEQ_LEQ((seq_out + len), snd_una)) {
3315 		/* Are sending an old segment to induce an ack (keep-alive)? */
3316 		return;
3317 	}
3318 	if (SEQ_LT(seq_out, snd_una)) {
3319 		/* huh? should we panic? */
3320 		uint32_t end;
3321 
3322 		end = seq_out + len;
3323 		seq_out = snd_una;
3324 		len = end - seq_out;
3325 	}
3326 	snd_max = tp->snd_max;
3327 	if (th_flags & (TH_SYN | TH_FIN)) {
3328 		/*
3329 		 * The call to rack_log_output is made before bumping
3330 		 * snd_max. This means we can record one extra byte on a SYN
3331 		 * or FIN if seq_out is adding more on and a FIN is present
3332 		 * (and we are not resending).
3333 		 */
3334 		if (th_flags & TH_SYN)
3335 			len++;
3336 		if (th_flags & TH_FIN)
3337 			len++;
3338 		if (SEQ_LT(snd_max, tp->snd_nxt)) {
3339 			/*
3340 			 * The add/update as not been done for the FIN/SYN
3341 			 * yet.
3342 			 */
3343 			snd_max = tp->snd_nxt;
3344 		}
3345 	}
3346 	if (len == 0) {
3347 		/* We don't log zero window probes */
3348 		return;
3349 	}
3350 	rack->r_ctl.rc_time_last_sent = ts;
3351 	if (IN_RECOVERY(tp->t_flags)) {
3352 		rack->r_ctl.rc_prr_out += len;
3353 	}
3354 	/* First question is it a retransmission? */
3355 	if (seq_out == snd_max) {
3356 again:
3357 		rsm = rack_alloc(rack);
3358 		if (rsm == NULL) {
3359 			/*
3360 			 * Hmm out of memory and the tcb got destroyed while
3361 			 * we tried to wait.
3362 			 */
3363 #ifdef INVARIANTS
3364 			panic("Out of memory when we should not be rack:%p", rack);
3365 #endif
3366 			return;
3367 		}
3368 		if (th_flags & TH_FIN) {
3369 			rsm->r_flags = RACK_HAS_FIN;
3370 		} else {
3371 			rsm->r_flags = 0;
3372 		}
3373 		rsm->r_tim_lastsent[0] = ts;
3374 		rsm->r_rtr_cnt = 1;
3375 		rsm->r_rtr_bytes = 0;
3376 		if (th_flags & TH_SYN) {
3377 			/* The data space is one beyond snd_una */
3378 			rsm->r_start = seq_out + 1;
3379 			rsm->r_end = rsm->r_start + (len - 1);
3380 		} else {
3381 			/* Normal case */
3382 			rsm->r_start = seq_out;
3383 			rsm->r_end = rsm->r_start + len;
3384 		}
3385 		rsm->r_sndcnt = 0;
3386 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_map, rsm, r_next);
3387 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
3388 		rsm->r_in_tmap = 1;
3389 		return;
3390 	}
3391 	/*
3392 	 * If we reach here its a retransmission and we need to find it.
3393 	 */
3394 more:
3395 	if (hintrsm && (hintrsm->r_start == seq_out)) {
3396 		rsm = hintrsm;
3397 		hintrsm = NULL;
3398 	} else if (rack->r_ctl.rc_next) {
3399 		/* We have a hint from a previous run */
3400 		rsm = rack->r_ctl.rc_next;
3401 	} else {
3402 		/* No hints sorry */
3403 		rsm = NULL;
3404 	}
3405 	if ((rsm) && (rsm->r_start == seq_out)) {
3406 		/*
3407 		 * We used rc_next or hintrsm  to retransmit, hopefully the
3408 		 * likely case.
3409 		 */
3410 		seq_out = rack_update_entry(tp, rack, rsm, ts, &len);
3411 		if (len == 0) {
3412 			return;
3413 		} else {
3414 			goto more;
3415 		}
3416 	}
3417 	/* Ok it was not the last pointer go through it the hard way. */
3418 	TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) {
3419 		if (rsm->r_start == seq_out) {
3420 			seq_out = rack_update_entry(tp, rack, rsm, ts, &len);
3421 			rack->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next);
3422 			if (len == 0) {
3423 				return;
3424 			} else {
3425 				continue;
3426 			}
3427 		}
3428 		if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) {
3429 			/* Transmitted within this piece */
3430 			/*
3431 			 * Ok we must split off the front and then let the
3432 			 * update do the rest
3433 			 */
3434 			nrsm = rack_alloc(rack);
3435 			if (nrsm == NULL) {
3436 #ifdef INVARIANTS
3437 				panic("Ran out of memory that was preallocated? rack:%p", rack);
3438 #endif
3439 				rack_update_rsm(tp, rack, rsm, ts);
3440 				return;
3441 			}
3442 			/*
3443 			 * copy rsm to nrsm and then trim the front of rsm
3444 			 * to not include this part.
3445 			 */
3446 			nrsm->r_start = seq_out;
3447 			nrsm->r_end = rsm->r_end;
3448 			nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
3449 			nrsm->r_flags = rsm->r_flags;
3450 			nrsm->r_sndcnt = rsm->r_sndcnt;
3451 			nrsm->r_rtr_bytes = 0;
3452 			for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
3453 				nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
3454 			}
3455 			rsm->r_end = nrsm->r_start;
3456 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next);
3457 			if (rsm->r_in_tmap) {
3458 				TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
3459 				nrsm->r_in_tmap = 1;
3460 			}
3461 			rsm->r_flags &= (~RACK_HAS_FIN);
3462 			seq_out = rack_update_entry(tp, rack, nrsm, ts, &len);
3463 			if (len == 0) {
3464 				return;
3465 			}
3466 		}
3467 	}
3468 	/*
3469 	 * Hmm not found in map did they retransmit both old and on into the
3470 	 * new?
3471 	 */
3472 	if (seq_out == tp->snd_max) {
3473 		goto again;
3474 	} else if (SEQ_LT(seq_out, tp->snd_max)) {
3475 #ifdef INVARIANTS
3476 		printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n",
3477 		    seq_out, len, tp->snd_una, tp->snd_max);
3478 		printf("Starting Dump of all rack entries\n");
3479 		TAILQ_FOREACH(rsm, &rack->r_ctl.rc_map, r_next) {
3480 			printf("rsm:%p start:%u end:%u\n",
3481 			    rsm, rsm->r_start, rsm->r_end);
3482 		}
3483 		printf("Dump complete\n");
3484 		panic("seq_out not found rack:%p tp:%p",
3485 		    rack, tp);
3486 #endif
3487 	} else {
3488 #ifdef INVARIANTS
3489 		/*
3490 		 * Hmm beyond sndmax? (only if we are using the new rtt-pack
3491 		 * flag)
3492 		 */
3493 		panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p",
3494 		    seq_out, len, tp->snd_max, tp);
3495 #endif
3496 	}
3497 }
3498 
3499 /*
3500  * Record one of the RTT updates from an ack into
3501  * our sample structure.
3502  */
3503 static void
3504 tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt)
3505 {
3506 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
3507 	    (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) {
3508 		rack->r_ctl.rack_rs.rs_rtt_lowest = rtt;
3509 	}
3510 	if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) ||
3511 	    (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) {
3512 		rack->r_ctl.rack_rs.rs_rtt_highest = rtt;
3513 	}
3514 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID;
3515 	rack->r_ctl.rack_rs.rs_rtt_tot += rtt;
3516 	rack->r_ctl.rack_rs.rs_rtt_cnt++;
3517 }
3518 
3519 /*
3520  * Collect new round-trip time estimate
3521  * and update averages and current timeout.
3522  */
3523 static void
3524 tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp)
3525 {
3526 	int32_t delta;
3527 	uint32_t o_srtt, o_var;
3528 	int32_t rtt;
3529 
3530 	if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY)
3531 		/* No valid sample */
3532 		return;
3533 	if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) {
3534 		/* We are to use the lowest RTT seen in a single ack */
3535 		rtt = rack->r_ctl.rack_rs.rs_rtt_lowest;
3536 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) {
3537 		/* We are to use the highest RTT seen in a single ack */
3538 		rtt = rack->r_ctl.rack_rs.rs_rtt_highest;
3539 	} else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) {
3540 		/* We are to use the average RTT seen in a single ack */
3541 		rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot /
3542 				(uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt);
3543 	} else {
3544 #ifdef INVARIANTS
3545 		panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method);
3546 #endif
3547 		return;
3548 	}
3549 	if (rtt == 0)
3550 		rtt = 1;
3551 	rack_log_rtt_sample(rack, rtt);
3552 	o_srtt = tp->t_srtt;
3553 	o_var = tp->t_rttvar;
3554 	rack = (struct tcp_rack *)tp->t_fb_ptr;
3555 	if (tp->t_srtt != 0) {
3556 		/*
3557 		 * srtt is stored as fixed point with 5 bits after the
3558 		 * binary point (i.e., scaled by 8).  The following magic is
3559 		 * equivalent to the smoothing algorithm in rfc793 with an
3560 		 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
3561 		 * Adjust rtt to origin 0.
3562 		 */
3563 		delta = ((rtt - 1) << TCP_DELTA_SHIFT)
3564 		    - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
3565 
3566 		tp->t_srtt += delta;
3567 		if (tp->t_srtt <= 0)
3568 			tp->t_srtt = 1;
3569 
3570 		/*
3571 		 * We accumulate a smoothed rtt variance (actually, a
3572 		 * smoothed mean difference), then set the retransmit timer
3573 		 * to smoothed rtt + 4 times the smoothed variance. rttvar
3574 		 * is stored as fixed point with 4 bits after the binary
3575 		 * point (scaled by 16).  The following is equivalent to
3576 		 * rfc793 smoothing with an alpha of .75 (rttvar =
3577 		 * rttvar*3/4 + |delta| / 4).  This replaces rfc793's
3578 		 * wired-in beta.
3579 		 */
3580 		if (delta < 0)
3581 			delta = -delta;
3582 		delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
3583 		tp->t_rttvar += delta;
3584 		if (tp->t_rttvar <= 0)
3585 			tp->t_rttvar = 1;
3586 		if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
3587 			tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
3588 	} else {
3589 		/*
3590 		 * No rtt measurement yet - use the unsmoothed rtt. Set the
3591 		 * variance to half the rtt (so our first retransmit happens
3592 		 * at 3*rtt).
3593 		 */
3594 		tp->t_srtt = rtt << TCP_RTT_SHIFT;
3595 		tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1);
3596 		tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
3597 	}
3598 	TCPSTAT_INC(tcps_rttupdated);
3599 	rack_log_rtt_upd(tp, rack, rtt, o_srtt, o_var);
3600 	tp->t_rttupdated++;
3601 #ifdef NETFLIX_STATS
3602 	stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt));
3603 #endif
3604 	tp->t_rxtshift = 0;
3605 
3606 	/*
3607 	 * the retransmit should happen at rtt + 4 * rttvar. Because of the
3608 	 * way we do the smoothing, srtt and rttvar will each average +1/2
3609 	 * tick of bias.  When we compute the retransmit timer, we want 1/2
3610 	 * tick of rounding and 1 extra tick because of +-1/2 tick
3611 	 * uncertainty in the firing of the timer.  The bias will give us
3612 	 * exactly the 1.5 tick we need.  But, because the bias is
3613 	 * statistical, we have to test that we don't drop below the minimum
3614 	 * feasible timer (which is 2 ticks).
3615 	 */
3616 	TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
3617 	   max(MSEC_2_TICKS(rack_rto_min), rtt + 2), MSEC_2_TICKS(rack_rto_max));
3618 	tp->t_softerror = 0;
3619 }
3620 
3621 static void
3622 rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm,
3623     uint32_t t, uint32_t cts)
3624 {
3625 	/*
3626 	 * For this RSM, we acknowledged the data from a previous
3627 	 * transmission, not the last one we made. This means we did a false
3628 	 * retransmit.
3629 	 */
3630 	struct tcp_rack *rack;
3631 
3632 	if (rsm->r_flags & RACK_HAS_FIN) {
3633 		/*
3634 		 * The sending of the FIN often is multiple sent when we
3635 		 * have everything outstanding ack'd. We ignore this case
3636 		 * since its over now.
3637 		 */
3638 		return;
3639 	}
3640 	if (rsm->r_flags & RACK_TLP) {
3641 		/*
3642 		 * We expect TLP's to have this occur.
3643 		 */
3644 		return;
3645 	}
3646 	rack = (struct tcp_rack *)tp->t_fb_ptr;
3647 	/* should we undo cc changes and exit recovery? */
3648 	if (IN_RECOVERY(tp->t_flags)) {
3649 		if (rack->r_ctl.rc_rsm_start == rsm->r_start) {
3650 			/*
3651 			 * Undo what we ratched down and exit recovery if
3652 			 * possible
3653 			 */
3654 			EXIT_RECOVERY(tp->t_flags);
3655 			tp->snd_recover = tp->snd_una;
3656 			if (rack->r_ctl.rc_cwnd_at > tp->snd_cwnd)
3657 				tp->snd_cwnd = rack->r_ctl.rc_cwnd_at;
3658 			if (rack->r_ctl.rc_ssthresh_at > tp->snd_ssthresh)
3659 				tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at;
3660 		}
3661 	}
3662 	if (rsm->r_flags & RACK_WAS_SACKPASS) {
3663 		/*
3664 		 * We retransmitted based on a sack and the earlier
3665 		 * retransmission ack'd it - re-ordering is occuring.
3666 		 */
3667 		counter_u64_add(rack_reorder_seen, 1);
3668 		rack->r_ctl.rc_reorder_ts = cts;
3669 	}
3670 	counter_u64_add(rack_badfr, 1);
3671 	counter_u64_add(rack_badfr_bytes, (rsm->r_end - rsm->r_start));
3672 }
3673 
3674 
3675 static int
3676 rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack,
3677     struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type)
3678 {
3679 	int32_t i;
3680 	uint32_t t;
3681 
3682 	if (rsm->r_flags & RACK_ACKED)
3683 		/* Already done */
3684 		return (0);
3685 
3686 
3687 	if ((rsm->r_rtr_cnt == 1) ||
3688 	    ((ack_type == CUM_ACKED) &&
3689 	    (to->to_flags & TOF_TS) &&
3690 	    (to->to_tsecr) &&
3691 	    (rsm->r_tim_lastsent[rsm->r_rtr_cnt - 1] == to->to_tsecr))
3692 	    ) {
3693 		/*
3694 		 * We will only find a matching timestamp if its cum-acked.
3695 		 * But if its only one retransmission its for-sure matching
3696 		 * :-)
3697 		 */
3698 		t = cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
3699 		if ((int)t <= 0)
3700 			t = 1;
3701 		if (!tp->t_rttlow || tp->t_rttlow > t)
3702 			tp->t_rttlow = t;
3703 		if (!rack->r_ctl.rc_rack_min_rtt ||
3704 		    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
3705 			rack->r_ctl.rc_rack_min_rtt = t;
3706 			if (rack->r_ctl.rc_rack_min_rtt == 0) {
3707 				rack->r_ctl.rc_rack_min_rtt = 1;
3708 			}
3709 		}
3710 		tcp_rack_xmit_timer(rack, TCP_TS_TO_TICKS(t) + 1);
3711 		if ((rsm->r_flags & RACK_TLP) &&
3712 		    (!IN_RECOVERY(tp->t_flags))) {
3713 			/* Segment was a TLP and our retrans matched */
3714 			if (rack->r_ctl.rc_tlp_cwnd_reduce) {
3715 				rack->r_ctl.rc_rsm_start = tp->snd_max;
3716 				rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
3717 				rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
3718 				rack_cong_signal(tp, NULL, CC_NDUPACK);
3719 				/*
3720 				 * When we enter recovery we need to assure
3721 				 * we send one packet.
3722 				 */
3723 				rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg;
3724 			} else
3725 				rack->r_ctl.rc_tlp_rtx_out = 0;
3726 		}
3727 		if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
3728 			/* New more recent rack_tmit_time */
3729 			rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
3730 			rack->rc_rack_rtt = t;
3731 		}
3732 		return (1);
3733 	}
3734 	/*
3735 	 * We clear the soft/rxtshift since we got an ack.
3736 	 * There is no assurance we will call the commit() function
3737 	 * so we need to clear these to avoid incorrect handling.
3738 	 */
3739 	tp->t_rxtshift = 0;
3740 	tp->t_softerror = 0;
3741 	if ((to->to_flags & TOF_TS) &&
3742 	    (ack_type == CUM_ACKED) &&
3743 	    (to->to_tsecr) &&
3744 	    ((rsm->r_flags & (RACK_DEFERRED | RACK_OVERMAX)) == 0)) {
3745 		/*
3746 		 * Now which timestamp does it match? In this block the ACK
3747 		 * must be coming from a previous transmission.
3748 		 */
3749 		for (i = 0; i < rsm->r_rtr_cnt; i++) {
3750 			if (rsm->r_tim_lastsent[i] == to->to_tsecr) {
3751 				t = cts - rsm->r_tim_lastsent[i];
3752 				if ((int)t <= 0)
3753 					t = 1;
3754 				if ((i + 1) < rsm->r_rtr_cnt) {
3755 					/* Likely */
3756 					rack_earlier_retran(tp, rsm, t, cts);
3757 				}
3758 				if (!tp->t_rttlow || tp->t_rttlow > t)
3759 					tp->t_rttlow = t;
3760 				if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
3761 					rack->r_ctl.rc_rack_min_rtt = t;
3762 					if (rack->r_ctl.rc_rack_min_rtt == 0) {
3763 						rack->r_ctl.rc_rack_min_rtt = 1;
3764 					}
3765 				}
3766                                 /*
3767 				 * Note the following calls to
3768 				 * tcp_rack_xmit_timer() are being commented
3769 				 * out for now. They give us no more accuracy
3770 				 * and often lead to a wrong choice. We have
3771 				 * enough samples that have not been
3772 				 * retransmitted. I leave the commented out
3773 				 * code in here in case in the future we
3774 				 * decide to add it back (though I can't forsee
3775 				 * doing that). That way we will easily see
3776 				 * where they need to be placed.
3777 				 */
3778 				if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time,
3779 				    rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) {
3780 					/* New more recent rack_tmit_time */
3781 					rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)];
3782 					rack->rc_rack_rtt = t;
3783 				}
3784 				return (1);
3785 			}
3786 		}
3787 		goto ts_not_found;
3788 	} else {
3789 		/*
3790 		 * Ok its a SACK block that we retransmitted. or a windows
3791 		 * machine without timestamps. We can tell nothing from the
3792 		 * time-stamp since its not there or the time the peer last
3793 		 * recieved a segment that moved forward its cum-ack point.
3794 		 */
3795 ts_not_found:
3796 		i = rsm->r_rtr_cnt - 1;
3797 		t = cts - rsm->r_tim_lastsent[i];
3798 		if ((int)t <= 0)
3799 			t = 1;
3800 		if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
3801 			/*
3802 			 * We retransmitted and the ack came back in less
3803 			 * than the smallest rtt we have observed. We most
3804 			 * likey did an improper retransmit as outlined in
3805 			 * 4.2 Step 3 point 2 in the rack-draft.
3806 			 */
3807 			i = rsm->r_rtr_cnt - 2;
3808 			t = cts - rsm->r_tim_lastsent[i];
3809 			rack_earlier_retran(tp, rsm, t, cts);
3810 		} else if (rack->r_ctl.rc_rack_min_rtt) {
3811 			/*
3812 			 * We retransmitted it and the retransmit did the
3813 			 * job.
3814 			 */
3815 			if (!rack->r_ctl.rc_rack_min_rtt ||
3816 			    SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) {
3817 				rack->r_ctl.rc_rack_min_rtt = t;
3818 				if (rack->r_ctl.rc_rack_min_rtt == 0) {
3819 					rack->r_ctl.rc_rack_min_rtt = 1;
3820 				}
3821 			}
3822 			if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[i])) {
3823 				/* New more recent rack_tmit_time */
3824 				rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[i];
3825 				rack->rc_rack_rtt = t;
3826 			}
3827 			return (1);
3828 		}
3829 	}
3830 	return (0);
3831 }
3832 
3833 /*
3834  * Mark the SACK_PASSED flag on all entries prior to rsm send wise.
3835  */
3836 static void
3837 rack_log_sack_passed(struct tcpcb *tp,
3838     struct tcp_rack *rack, struct rack_sendmap *rsm)
3839 {
3840 	struct rack_sendmap *nrsm;
3841 	uint32_t ts;
3842 	int32_t idx;
3843 
3844 	idx = rsm->r_rtr_cnt - 1;
3845 	ts = rsm->r_tim_lastsent[idx];
3846 	nrsm = rsm;
3847 	TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap,
3848 	    rack_head, r_tnext) {
3849 		if (nrsm == rsm) {
3850 			/* Skip orginal segment he is acked */
3851 			continue;
3852 		}
3853 		if (nrsm->r_flags & RACK_ACKED) {
3854 			/* Skip ack'd segments */
3855 			continue;
3856 		}
3857 		idx = nrsm->r_rtr_cnt - 1;
3858 		if (ts == nrsm->r_tim_lastsent[idx]) {
3859 			/*
3860 			 * For this case lets use seq no, if we sent in a
3861 			 * big block (TSO) we would have a bunch of segments
3862 			 * sent at the same time.
3863 			 *
3864 			 * We would only get a report if its SEQ is earlier.
3865 			 * If we have done multiple retransmits the times
3866 			 * would not be equal.
3867 			 */
3868 			if (SEQ_LT(nrsm->r_start, rsm->r_start)) {
3869 				nrsm->r_flags |= RACK_SACK_PASSED;
3870 				nrsm->r_flags &= ~RACK_WAS_SACKPASS;
3871 			}
3872 		} else {
3873 			/*
3874 			 * Here they were sent at different times, not a big
3875 			 * block. Since we transmitted this one later and
3876 			 * see it sack'd then this must also be missing (or
3877 			 * we would have gotten a sack block for it)
3878 			 */
3879 			nrsm->r_flags |= RACK_SACK_PASSED;
3880 			nrsm->r_flags &= ~RACK_WAS_SACKPASS;
3881 		}
3882 	}
3883 }
3884 
3885 static uint32_t
3886 rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack,
3887     struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts)
3888 {
3889 	int32_t idx;
3890 	int32_t times = 0;
3891 	uint32_t start, end, changed = 0;
3892 	struct rack_sendmap *rsm, *nrsm;
3893 	int32_t used_ref = 1;
3894 
3895 	start = sack->start;
3896 	end = sack->end;
3897 	rsm = *prsm;
3898 	if (rsm && SEQ_LT(start, rsm->r_start)) {
3899 		TAILQ_FOREACH_REVERSE_FROM(rsm, &rack->r_ctl.rc_map, rack_head, r_next) {
3900 			if (SEQ_GEQ(start, rsm->r_start) &&
3901 			    SEQ_LT(start, rsm->r_end)) {
3902 				goto do_rest_ofb;
3903 			}
3904 		}
3905 	}
3906 	if (rsm == NULL) {
3907 start_at_beginning:
3908 		rsm = NULL;
3909 		used_ref = 0;
3910 	}
3911 	/* First lets locate the block where this guy is */
3912 	TAILQ_FOREACH_FROM(rsm, &rack->r_ctl.rc_map, r_next) {
3913 		if (SEQ_GEQ(start, rsm->r_start) &&
3914 		    SEQ_LT(start, rsm->r_end)) {
3915 			break;
3916 		}
3917 	}
3918 do_rest_ofb:
3919 	if (rsm == NULL) {
3920 		/*
3921 		 * This happens when we get duplicate sack blocks with the
3922 		 * same end. For example SACK 4: 100 SACK 3: 100 The sort
3923 		 * will not change there location so we would just start at
3924 		 * the end of the first one and get lost.
3925 		 */
3926 		if (tp->t_flags & TF_SENTFIN) {
3927 			/*
3928 			 * Check to see if we have not logged the FIN that
3929 			 * went out.
3930 			 */
3931 			nrsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next);
3932 			if (nrsm && (nrsm->r_end + 1) == tp->snd_max) {
3933 				/*
3934 				 * Ok we did not get the FIN logged.
3935 				 */
3936 				nrsm->r_end++;
3937 				rsm = nrsm;
3938 				goto do_rest_ofb;
3939 			}
3940 		}
3941 		if (times == 1) {
3942 #ifdef INVARIANTS
3943 			panic("tp:%p rack:%p sack:%p to:%p prsm:%p",
3944 			    tp, rack, sack, to, prsm);
3945 #else
3946 			goto out;
3947 #endif
3948 		}
3949 		times++;
3950 		counter_u64_add(rack_sack_proc_restart, 1);
3951 		goto start_at_beginning;
3952 	}
3953 	/* Ok we have an ACK for some piece of rsm */
3954 	if (rsm->r_start != start) {
3955 		/*
3956 		 * Need to split this in two pieces the before and after.
3957 		 */
3958 		nrsm = rack_alloc(rack);
3959 		if (nrsm == NULL) {
3960 			/*
3961 			 * failed XXXrrs what can we do but loose the sack
3962 			 * info?
3963 			 */
3964 			goto out;
3965 		}
3966 		nrsm->r_start = start;
3967 		nrsm->r_rtr_bytes = 0;
3968 		nrsm->r_end = rsm->r_end;
3969 		nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
3970 		nrsm->r_flags = rsm->r_flags;
3971 		nrsm->r_sndcnt = rsm->r_sndcnt;
3972 		for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
3973 			nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
3974 		}
3975 		rsm->r_end = nrsm->r_start;
3976 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next);
3977 		if (rsm->r_in_tmap) {
3978 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
3979 			nrsm->r_in_tmap = 1;
3980 		}
3981 		rsm->r_flags &= (~RACK_HAS_FIN);
3982 		rsm = nrsm;
3983 	}
3984 	if (SEQ_GEQ(end, rsm->r_end)) {
3985 		/*
3986 		 * The end of this block is either beyond this guy or right
3987 		 * at this guy.
3988 		 */
3989 
3990 		if ((rsm->r_flags & RACK_ACKED) == 0) {
3991 			rack_update_rtt(tp, rack, rsm, to, cts, SACKED);
3992 			changed += (rsm->r_end - rsm->r_start);
3993 			rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
3994 			rack_log_sack_passed(tp, rack, rsm);
3995 			/* Is Reordering occuring? */
3996 			if (rsm->r_flags & RACK_SACK_PASSED) {
3997 				counter_u64_add(rack_reorder_seen, 1);
3998 				rack->r_ctl.rc_reorder_ts = cts;
3999 			}
4000 			rsm->r_flags |= RACK_ACKED;
4001 			rsm->r_flags &= ~RACK_TLP;
4002 			if (rsm->r_in_tmap) {
4003 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
4004 				rsm->r_in_tmap = 0;
4005 			}
4006 		}
4007 		if (end == rsm->r_end) {
4008 			/* This block only - done */
4009 			goto out;
4010 		}
4011 		/* There is more not coverend by this rsm move on */
4012 		start = rsm->r_end;
4013 		nrsm = TAILQ_NEXT(rsm, r_next);
4014 		rsm = nrsm;
4015 		times = 0;
4016 		goto do_rest_ofb;
4017 	}
4018 	/* Ok we need to split off this one at the tail */
4019 	nrsm = rack_alloc(rack);
4020 	if (nrsm == NULL) {
4021 		/* failed rrs what can we do but loose the sack info? */
4022 		goto out;
4023 	}
4024 	/* Clone it */
4025 	nrsm->r_start = end;
4026 	nrsm->r_end = rsm->r_end;
4027 	nrsm->r_rtr_bytes = 0;
4028 	nrsm->r_rtr_cnt = rsm->r_rtr_cnt;
4029 	nrsm->r_flags = rsm->r_flags;
4030 	nrsm->r_sndcnt = rsm->r_sndcnt;
4031 	for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) {
4032 		nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx];
4033 	}
4034 	/* The sack block does not cover this guy fully */
4035 	rsm->r_flags &= (~RACK_HAS_FIN);
4036 	rsm->r_end = end;
4037 	TAILQ_INSERT_AFTER(&rack->r_ctl.rc_map, rsm, nrsm, r_next);
4038 	if (rsm->r_in_tmap) {
4039 		TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext);
4040 		nrsm->r_in_tmap = 1;
4041 	}
4042 	if (rsm->r_flags & RACK_ACKED) {
4043 		/* Been here done that */
4044 		goto out;
4045 	}
4046 	rack_update_rtt(tp, rack, rsm, to, cts, SACKED);
4047 	changed += (rsm->r_end - rsm->r_start);
4048 	rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start);
4049 	rack_log_sack_passed(tp, rack, rsm);
4050 	/* Is Reordering occuring? */
4051 	if (rsm->r_flags & RACK_SACK_PASSED) {
4052 		counter_u64_add(rack_reorder_seen, 1);
4053 		rack->r_ctl.rc_reorder_ts = cts;
4054 	}
4055 	rsm->r_flags |= RACK_ACKED;
4056 	rsm->r_flags &= ~RACK_TLP;
4057 	if (rsm->r_in_tmap) {
4058 		TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
4059 		rsm->r_in_tmap = 0;
4060 	}
4061 out:
4062 	if (used_ref == 0) {
4063 		counter_u64_add(rack_sack_proc_all, 1);
4064 	} else {
4065 		counter_u64_add(rack_sack_proc_short, 1);
4066 	}
4067 	/* Save off where we last were */
4068 	if (rsm)
4069 		rack->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next);
4070 	else
4071 		rack->r_ctl.rc_sacklast = NULL;
4072 	*prsm = rsm;
4073 	return (changed);
4074 }
4075 
4076 static void inline
4077 rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack)
4078 {
4079 	struct rack_sendmap *tmap;
4080 
4081 	tmap = NULL;
4082 	while (rsm && (rsm->r_flags & RACK_ACKED)) {
4083 		/* Its no longer sacked, mark it so */
4084 		rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
4085 #ifdef INVARIANTS
4086 		if (rsm->r_in_tmap) {
4087 			panic("rack:%p rsm:%p flags:0x%x in tmap?",
4088 			      rack, rsm, rsm->r_flags);
4089 		}
4090 #endif
4091 		rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS);
4092 		/* Rebuild it into our tmap */
4093 		if (tmap == NULL) {
4094 			TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext);
4095 			tmap = rsm;
4096 		} else {
4097 			TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext);
4098 			tmap = rsm;
4099 		}
4100 		tmap->r_in_tmap = 1;
4101 		rsm = TAILQ_NEXT(rsm, r_next);
4102 	}
4103 	/*
4104 	 * Now lets possibly clear the sack filter so we start
4105 	 * recognizing sacks that cover this area.
4106 	 */
4107 	if (rack_use_sack_filter)
4108 		sack_filter_clear(&rack->r_ctl.rack_sf, th_ack);
4109 
4110 }
4111 
4112 static void
4113 rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th)
4114 {
4115 	uint32_t changed, last_seq, entered_recovery = 0;
4116 	struct tcp_rack *rack;
4117 	struct rack_sendmap *rsm;
4118 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1];
4119 	register uint32_t th_ack;
4120 	int32_t i, j, k, num_sack_blks = 0;
4121 	uint32_t cts, acked, ack_point, sack_changed = 0;
4122 
4123 	INP_WLOCK_ASSERT(tp->t_inpcb);
4124 	if (th->th_flags & TH_RST) {
4125 		/* We don't log resets */
4126 		return;
4127 	}
4128 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4129 	cts = tcp_ts_getticks();
4130 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_map);
4131 	changed = 0;
4132 	th_ack = th->th_ack;
4133 
4134 	if (SEQ_GT(th_ack, tp->snd_una)) {
4135 		rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__);
4136 		tp->t_acktime = ticks;
4137 	}
4138 	if (rsm && SEQ_GT(th_ack, rsm->r_start))
4139 		changed = th_ack - rsm->r_start;
4140 	if (changed) {
4141 		/*
4142 		 * The ACK point is advancing to th_ack, we must drop off
4143 		 * the packets in the rack log and calculate any eligble
4144 		 * RTT's.
4145 		 */
4146 		rack->r_wanted_output++;
4147 more:
4148 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_map);
4149 		if (rsm == NULL) {
4150 			if ((th_ack - 1) == tp->iss) {
4151 				/*
4152 				 * For the SYN incoming case we will not
4153 				 * have called tcp_output for the sending of
4154 				 * the SYN, so there will be no map. All
4155 				 * other cases should probably be a panic.
4156 				 */
4157 				goto proc_sack;
4158 			}
4159 			if (tp->t_flags & TF_SENTFIN) {
4160 				/* if we send a FIN we will not hav a map */
4161 				goto proc_sack;
4162 			}
4163 #ifdef INVARIANTS
4164 			panic("No rack map tp:%p for th:%p state:%d rack:%p snd_una:%u snd_max:%u snd_nxt:%u chg:%d\n",
4165 			    tp,
4166 			    th, tp->t_state, rack,
4167 			    tp->snd_una, tp->snd_max, tp->snd_nxt, changed);
4168 #endif
4169 			goto proc_sack;
4170 		}
4171 		if (SEQ_LT(th_ack, rsm->r_start)) {
4172 			/* Huh map is missing this */
4173 #ifdef INVARIANTS
4174 			printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n",
4175 			    rsm->r_start,
4176 			    th_ack, tp->t_state, rack->r_state);
4177 #endif
4178 			goto proc_sack;
4179 		}
4180 		rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED);
4181 		/* Now do we consume the whole thing? */
4182 		if (SEQ_GEQ(th_ack, rsm->r_end)) {
4183 			/* Its all consumed. */
4184 			uint32_t left;
4185 
4186 			rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
4187 			rsm->r_rtr_bytes = 0;
4188 			TAILQ_REMOVE(&rack->r_ctl.rc_map, rsm, r_next);
4189 			if (rsm->r_in_tmap) {
4190 				TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext);
4191 				rsm->r_in_tmap = 0;
4192 			}
4193 			if (rack->r_ctl.rc_next == rsm) {
4194 				/* scoot along the marker */
4195 				rack->r_ctl.rc_next = TAILQ_FIRST(&rack->r_ctl.rc_map);
4196 			}
4197 			if (rsm->r_flags & RACK_ACKED) {
4198 				/*
4199 				 * It was acked on the scoreboard -- remove
4200 				 * it from total
4201 				 */
4202 				rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start);
4203 			} else if (rsm->r_flags & RACK_SACK_PASSED) {
4204 				/*
4205 				 * There are acked segments ACKED on the
4206 				 * scoreboard further up. We are seeing
4207 				 * reordering.
4208 				 */
4209 				counter_u64_add(rack_reorder_seen, 1);
4210 				rsm->r_flags |= RACK_ACKED;
4211 				rack->r_ctl.rc_reorder_ts = cts;
4212 			}
4213 			left = th_ack - rsm->r_end;
4214 			if (rsm->r_rtr_cnt > 1) {
4215 				/*
4216 				 * Technically we should make r_rtr_cnt be
4217 				 * monotonicly increasing and just mod it to
4218 				 * the timestamp it is replacing.. that way
4219 				 * we would have the last 3 retransmits. Now
4220 				 * rc_loss_count will be wrong if we
4221 				 * retransmit something more than 2 times in
4222 				 * recovery :(
4223 				 */
4224 				rack->r_ctl.rc_loss_count += (rsm->r_rtr_cnt - 1);
4225 			}
4226 			/* Free back to zone */
4227 			rack_free(rack, rsm);
4228 			if (left) {
4229 				goto more;
4230 			}
4231 			goto proc_sack;
4232 		}
4233 		if (rsm->r_flags & RACK_ACKED) {
4234 			/*
4235 			 * It was acked on the scoreboard -- remove it from
4236 			 * total for the part being cum-acked.
4237 			 */
4238 			rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start);
4239 		}
4240 		rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes;
4241 		rsm->r_rtr_bytes = 0;
4242 		rsm->r_start = th_ack;
4243 	}
4244 proc_sack:
4245 	/* Check for reneging */
4246 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_map);
4247 	if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) {
4248 		/*
4249 		 * The peer has moved snd_una up to
4250 		 * the edge of this send, i.e. one
4251 		 * that it had previously acked. The only
4252 		 * way that can be true if the peer threw
4253 		 * away data (space issues) that it had
4254 		 * previously sacked (else it would have
4255 		 * given us snd_una up to (rsm->r_end).
4256 		 * We need to undo the acked markings here.
4257 		 *
4258 		 * Note we have to look to make sure th_ack is
4259 		 * our rsm->r_start in case we get an old ack
4260 		 * where th_ack is behind snd_una.
4261 		 */
4262 		rack_peer_reneges(rack, rsm, th->th_ack);
4263 	}
4264 	if ((to->to_flags & TOF_SACK) == 0) {
4265 		/* We are done nothing left to log */
4266 		goto out;
4267 	}
4268 	rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_map, rack_sendmap, r_next);
4269 	if (rsm) {
4270 		last_seq = rsm->r_end;
4271 	} else {
4272 		last_seq = tp->snd_max;
4273 	}
4274 	/* Sack block processing */
4275 	if (SEQ_GT(th_ack, tp->snd_una))
4276 		ack_point = th_ack;
4277 	else
4278 		ack_point = tp->snd_una;
4279 	for (i = 0; i < to->to_nsacks; i++) {
4280 		bcopy((to->to_sacks + i * TCPOLEN_SACK),
4281 		    &sack, sizeof(sack));
4282 		sack.start = ntohl(sack.start);
4283 		sack.end = ntohl(sack.end);
4284 		if (SEQ_GT(sack.end, sack.start) &&
4285 		    SEQ_GT(sack.start, ack_point) &&
4286 		    SEQ_LT(sack.start, tp->snd_max) &&
4287 		    SEQ_GT(sack.end, ack_point) &&
4288 		    SEQ_LEQ(sack.end, tp->snd_max)) {
4289 			if ((rack->r_ctl.rc_num_maps_alloced > rack_sack_block_limit) &&
4290 			    (SEQ_LT(sack.end, last_seq)) &&
4291 			    ((sack.end - sack.start) < (tp->t_maxseg / 8))) {
4292 				/*
4293 				 * Not the last piece and its smaller than
4294 				 * 1/8th of a MSS. We ignore this.
4295 				 */
4296 				counter_u64_add(rack_runt_sacks, 1);
4297 				continue;
4298 			}
4299 			sack_blocks[num_sack_blks] = sack;
4300 			num_sack_blks++;
4301 #ifdef NETFLIX_STATS
4302 		} else if (SEQ_LEQ(sack.start, th_ack) &&
4303 			   SEQ_LEQ(sack.end, th_ack)) {
4304 			/*
4305 			 * Its a D-SACK block.
4306 			 */
4307 			tcp_record_dsack(sack.start, sack.end);
4308 #endif
4309 		}
4310 
4311 	}
4312 	if (num_sack_blks == 0)
4313 		goto out;
4314 	/*
4315 	 * Sort the SACK blocks so we can update the rack scoreboard with
4316 	 * just one pass.
4317 	 */
4318 	if (rack_use_sack_filter) {
4319 		num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks, num_sack_blks, th->th_ack);
4320 	}
4321 	if (num_sack_blks < 2) {
4322 		goto do_sack_work;
4323 	}
4324 	/* Sort the sacks */
4325 	for (i = 0; i < num_sack_blks; i++) {
4326 		for (j = i + 1; j < num_sack_blks; j++) {
4327 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
4328 				sack = sack_blocks[i];
4329 				sack_blocks[i] = sack_blocks[j];
4330 				sack_blocks[j] = sack;
4331 			}
4332 		}
4333 	}
4334 	/*
4335 	 * Now are any of the sack block ends the same (yes some
4336 	 * implememtations send these)?
4337 	 */
4338 again:
4339 	if (num_sack_blks > 1) {
4340 		for (i = 0; i < num_sack_blks; i++) {
4341 			for (j = i + 1; j < num_sack_blks; j++) {
4342 				if (sack_blocks[i].end == sack_blocks[j].end) {
4343 					/*
4344 					 * Ok these two have the same end we
4345 					 * want the smallest end and then
4346 					 * throw away the larger and start
4347 					 * again.
4348 					 */
4349 					if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) {
4350 						/*
4351 						 * The second block covers
4352 						 * more area use that
4353 						 */
4354 						sack_blocks[i].start = sack_blocks[j].start;
4355 					}
4356 					/*
4357 					 * Now collapse out the dup-sack and
4358 					 * lower the count
4359 					 */
4360 					for (k = (j + 1); k < num_sack_blks; k++) {
4361 						sack_blocks[j].start = sack_blocks[k].start;
4362 						sack_blocks[j].end = sack_blocks[k].end;
4363 						j++;
4364 					}
4365 					num_sack_blks--;
4366 					goto again;
4367 				}
4368 			}
4369 		}
4370 	}
4371 do_sack_work:
4372 	rsm = rack->r_ctl.rc_sacklast;
4373 	for (i = 0; i < num_sack_blks; i++) {
4374 		acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts);
4375 		if (acked) {
4376 			rack->r_wanted_output++;
4377 			changed += acked;
4378 			sack_changed += acked;
4379 		}
4380 	}
4381 out:
4382 	if (changed) {
4383 		/* Something changed cancel the rack timer */
4384 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
4385 	}
4386 	if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) {
4387 		/*
4388 		 * Ok we have a high probability that we need to go in to
4389 		 * recovery since we have data sack'd
4390 		 */
4391 		struct rack_sendmap *rsm;
4392 		uint32_t tsused;
4393 
4394 		tsused = tcp_ts_getticks();
4395 		rsm = tcp_rack_output(tp, rack, tsused);
4396 		if (rsm) {
4397 			/* Enter recovery */
4398 			rack->r_ctl.rc_rsm_start = rsm->r_start;
4399 			rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
4400 			rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
4401 			entered_recovery = 1;
4402 			rack_cong_signal(tp, NULL, CC_NDUPACK);
4403 			/*
4404 			 * When we enter recovery we need to assure we send
4405 			 * one packet.
4406 			 */
4407 			rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg;
4408 			rack->r_timer_override = 1;
4409 		}
4410 	}
4411 	if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) {
4412 		/* Deal with changed an PRR here (in recovery only) */
4413 		uint32_t pipe, snd_una;
4414 
4415 		rack->r_ctl.rc_prr_delivered += changed;
4416 		/* Compute prr_sndcnt */
4417 		if (SEQ_GT(tp->snd_una, th_ack)) {
4418 			snd_una = tp->snd_una;
4419 		} else {
4420 			snd_una = th_ack;
4421 		}
4422 		pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt;
4423 		if (pipe > tp->snd_ssthresh) {
4424 			long sndcnt;
4425 
4426 			sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh;
4427 			if (rack->r_ctl.rc_prr_recovery_fs > 0)
4428 				sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs;
4429 			else {
4430 				rack->r_ctl.rc_prr_sndcnt = 0;
4431 				sndcnt = 0;
4432 			}
4433 			sndcnt++;
4434 			if (sndcnt > (long)rack->r_ctl.rc_prr_out)
4435 				sndcnt -= rack->r_ctl.rc_prr_out;
4436 			else
4437 				sndcnt = 0;
4438 			rack->r_ctl.rc_prr_sndcnt = sndcnt;
4439 		} else {
4440 			uint32_t limit;
4441 
4442 			if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out)
4443 				limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out);
4444 			else
4445 				limit = 0;
4446 			if (changed > limit)
4447 				limit = changed;
4448 			limit += tp->t_maxseg;
4449 			if (tp->snd_ssthresh > pipe) {
4450 				rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit);
4451 			} else {
4452 				rack->r_ctl.rc_prr_sndcnt = min(0, limit);
4453 			}
4454 		}
4455 		if (rack->r_ctl.rc_prr_sndcnt >= tp->t_maxseg) {
4456 			rack->r_timer_override = 1;
4457 		}
4458 	}
4459 }
4460 
4461 /*
4462  * Return value of 1, we do not need to call rack_process_data().
4463  * return value of 0, rack_process_data can be called.
4464  * For ret_val if its 0 the TCP is locked, if its non-zero
4465  * its unlocked and probably unsafe to touch the TCB.
4466  */
4467 static int
4468 rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so,
4469     struct tcpcb *tp, struct tcpopt *to,
4470     uint32_t tiwin, int32_t tlen,
4471     int32_t * ofia, int32_t thflags, int32_t * ret_val)
4472 {
4473 	int32_t ourfinisacked = 0;
4474 	int32_t nsegs, acked_amount;
4475 	int32_t acked;
4476 	struct mbuf *mfree;
4477 	struct tcp_rack *rack;
4478 	int32_t recovery = 0;
4479 
4480 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4481 	if (SEQ_GT(th->th_ack, tp->snd_max)) {
4482 		rack_do_dropafterack(m, tp, th, thflags, tlen, ret_val);
4483 		return (1);
4484 	}
4485 	if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) {
4486 		rack_log_ack(tp, to, th);
4487 	}
4488 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
4489 		/*
4490 		 * Old ack, behind (or duplicate to) the last one rcv'd
4491 		 * Note: Should mark reordering is occuring! We should also
4492 		 * look for sack blocks arriving e.g. ack 1, 4-4 then ack 1,
4493 		 * 3-3, 4-4 would be reording. As well as ack 1, 3-3 <no
4494 		 * retran and> ack 3
4495 		 */
4496 		return (0);
4497 	}
4498 	/*
4499 	 * If we reach this point, ACK is not a duplicate, i.e., it ACKs
4500 	 * something we sent.
4501 	 */
4502 	if (tp->t_flags & TF_NEEDSYN) {
4503 		/*
4504 		 * T/TCP: Connection was half-synchronized, and our SYN has
4505 		 * been ACK'd (so connection is now fully synchronized).  Go
4506 		 * to non-starred state, increment snd_una for ACK of SYN,
4507 		 * and check if we can do window scaling.
4508 		 */
4509 		tp->t_flags &= ~TF_NEEDSYN;
4510 		tp->snd_una++;
4511 		/* Do window scaling? */
4512 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
4513 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
4514 			tp->rcv_scale = tp->request_r_scale;
4515 			/* Send window already scaled. */
4516 		}
4517 	}
4518 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
4519 	INP_WLOCK_ASSERT(tp->t_inpcb);
4520 
4521 	acked = BYTES_THIS_ACK(tp, th);
4522 	TCPSTAT_ADD(tcps_rcvackpack, nsegs);
4523 	TCPSTAT_ADD(tcps_rcvackbyte, acked);
4524 
4525 	/*
4526 	 * If we just performed our first retransmit, and the ACK arrives
4527 	 * within our recovery window, then it was a mistake to do the
4528 	 * retransmit in the first place.  Recover our original cwnd and
4529 	 * ssthresh, and proceed to transmit where we left off.
4530 	 */
4531 	if (tp->t_flags & TF_PREVVALID) {
4532 		tp->t_flags &= ~TF_PREVVALID;
4533 		if (tp->t_rxtshift == 1 &&
4534 		    (int)(ticks - tp->t_badrxtwin) < 0)
4535 			rack_cong_signal(tp, th, CC_RTO_ERR);
4536 	}
4537 	/*
4538 	 * If we have a timestamp reply, update smoothed round trip time. If
4539 	 * no timestamp is present but transmit timer is running and timed
4540 	 * sequence number was acked, update smoothed round trip time. Since
4541 	 * we now have an rtt measurement, cancel the timer backoff (cf.,
4542 	 * Phil Karn's retransmit alg.). Recompute the initial retransmit
4543 	 * timer.
4544 	 *
4545 	 * Some boxes send broken timestamp replies during the SYN+ACK
4546 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
4547 	 * and blow up the retransmit timer.
4548 	 */
4549 	/*
4550 	 * If all outstanding data is acked, stop retransmit timer and
4551 	 * remember to restart (more output or persist). If there is more
4552 	 * data to be acked, restart retransmit timer, using current
4553 	 * (possibly backed-off) value.
4554 	 */
4555 	if (th->th_ack == tp->snd_max) {
4556 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
4557 		rack->r_wanted_output++;
4558 	}
4559 	/*
4560 	 * If no data (only SYN) was ACK'd, skip rest of ACK processing.
4561 	 */
4562 	if (acked == 0) {
4563 		if (ofia)
4564 			*ofia = ourfinisacked;
4565 		return (0);
4566 	}
4567 	if (rack->r_ctl.rc_early_recovery) {
4568 		if (IN_FASTRECOVERY(tp->t_flags)) {
4569 			if (SEQ_LT(th->th_ack, tp->snd_recover)) {
4570 				tcp_rack_partialack(tp, th);
4571 			} else {
4572 				rack_post_recovery(tp, th);
4573 				recovery = 1;
4574 			}
4575 		}
4576 	}
4577 	/*
4578 	 * Let the congestion control algorithm update congestion control
4579 	 * related information. This typically means increasing the
4580 	 * congestion window.
4581 	 */
4582 	rack_ack_received(tp, rack, th, nsegs, CC_ACK, recovery);
4583 	SOCKBUF_LOCK(&so->so_snd);
4584 	acked_amount = min(acked, (int)sbavail(&so->so_snd));
4585 	tp->snd_wnd -= acked_amount;
4586 	mfree = sbcut_locked(&so->so_snd, acked_amount);
4587 	if ((sbused(&so->so_snd) == 0) &&
4588 	    (acked > acked_amount) &&
4589 	    (tp->t_state >= TCPS_FIN_WAIT_1)) {
4590 		ourfinisacked = 1;
4591 	}
4592 	/* NB: sowwakeup_locked() does an implicit unlock. */
4593 	sowwakeup_locked(so);
4594 	m_freem(mfree);
4595 	if (rack->r_ctl.rc_early_recovery == 0) {
4596 		if (IN_FASTRECOVERY(tp->t_flags)) {
4597 			if (SEQ_LT(th->th_ack, tp->snd_recover)) {
4598 				tcp_rack_partialack(tp, th);
4599 			} else {
4600 				rack_post_recovery(tp, th);
4601 			}
4602 		}
4603 	}
4604 	tp->snd_una = th->th_ack;
4605 	if (SEQ_GT(tp->snd_una, tp->snd_recover))
4606 		tp->snd_recover = tp->snd_una;
4607 
4608 	if (SEQ_LT(tp->snd_nxt, tp->snd_una)) {
4609 		tp->snd_nxt = tp->snd_una;
4610 	}
4611 	if (tp->snd_una == tp->snd_max) {
4612 		/* Nothing left outstanding */
4613 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
4614 		tp->t_acktime = 0;
4615 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
4616 		/* Set need output so persist might get set */
4617 		rack->r_wanted_output++;
4618 		if (rack_use_sack_filter)
4619 			sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
4620 		if ((tp->t_state >= TCPS_FIN_WAIT_1) &&
4621 		    (sbavail(&so->so_snd) == 0) &&
4622 		    (tp->t_flags2 & TF2_DROP_AF_DATA)) {
4623 			/*
4624 			 * The socket was gone and the
4625 			 * peer sent data, time to
4626 			 * reset him.
4627 			 */
4628 			*ret_val = 1;
4629 			tp = tcp_close(tp);
4630 			rack_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen);
4631 			return (1);
4632 		}
4633 	}
4634 	if (ofia)
4635 		*ofia = ourfinisacked;
4636 	return (0);
4637 }
4638 
4639 
4640 /*
4641  * Return value of 1, the TCB is unlocked and most
4642  * likely gone, return value of 0, the TCP is still
4643  * locked.
4644  */
4645 static int
4646 rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so,
4647     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen,
4648     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
4649 {
4650 	/*
4651 	 * Update window information. Don't look at window if no ACK: TAC's
4652 	 * send garbage on first SYN.
4653 	 */
4654 	int32_t nsegs;
4655 	int32_t tfo_syn;
4656 	struct tcp_rack *rack;
4657 
4658 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4659 	INP_WLOCK_ASSERT(tp->t_inpcb);
4660 
4661 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
4662 	if ((thflags & TH_ACK) &&
4663 	    (SEQ_LT(tp->snd_wl1, th->th_seq) ||
4664 	    (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) ||
4665 	    (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) {
4666 		/* keep track of pure window updates */
4667 		if (tlen == 0 &&
4668 		    tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)
4669 			TCPSTAT_INC(tcps_rcvwinupd);
4670 		tp->snd_wnd = tiwin;
4671 		tp->snd_wl1 = th->th_seq;
4672 		tp->snd_wl2 = th->th_ack;
4673 		if (tp->snd_wnd > tp->max_sndwnd)
4674 			tp->max_sndwnd = tp->snd_wnd;
4675 		rack->r_wanted_output++;
4676 	} else if (thflags & TH_ACK) {
4677 		if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) {
4678 			tp->snd_wnd = tiwin;
4679 			tp->snd_wl1 = th->th_seq;
4680 			tp->snd_wl2 = th->th_ack;
4681 		}
4682 	}
4683 	/* Was persist timer active and now we have window space? */
4684 	if ((rack->rc_in_persist != 0) && tp->snd_wnd) {
4685 		rack_exit_persist(tp, rack);
4686 		tp->snd_nxt = tp->snd_max;
4687 		/* Make sure we output to start the timer */
4688 		rack->r_wanted_output++;
4689 	}
4690 	/*
4691 	 * Process segments with URG.
4692 	 */
4693 	if ((thflags & TH_URG) && th->th_urp &&
4694 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
4695 		/*
4696 		 * This is a kludge, but if we receive and accept random
4697 		 * urgent pointers, we'll crash in soreceive.  It's hard to
4698 		 * imagine someone actually wanting to send this much urgent
4699 		 * data.
4700 		 */
4701 		SOCKBUF_LOCK(&so->so_rcv);
4702 		if (th->th_urp + sbavail(&so->so_rcv) > sb_max) {
4703 			th->th_urp = 0;	/* XXX */
4704 			thflags &= ~TH_URG;	/* XXX */
4705 			SOCKBUF_UNLOCK(&so->so_rcv);	/* XXX */
4706 			goto dodata;	/* XXX */
4707 		}
4708 		/*
4709 		 * If this segment advances the known urgent pointer, then
4710 		 * mark the data stream.  This should not happen in
4711 		 * CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since a
4712 		 * FIN has been received from the remote side. In these
4713 		 * states we ignore the URG.
4714 		 *
4715 		 * According to RFC961 (Assigned Protocols), the urgent
4716 		 * pointer points to the last octet of urgent data.  We
4717 		 * continue, however, to consider it to indicate the first
4718 		 * octet of data past the urgent section as the original
4719 		 * spec states (in one of two places).
4720 		 */
4721 		if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) {
4722 			tp->rcv_up = th->th_seq + th->th_urp;
4723 			so->so_oobmark = sbavail(&so->so_rcv) +
4724 			    (tp->rcv_up - tp->rcv_nxt) - 1;
4725 			if (so->so_oobmark == 0)
4726 				so->so_rcv.sb_state |= SBS_RCVATMARK;
4727 			sohasoutofband(so);
4728 			tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA);
4729 		}
4730 		SOCKBUF_UNLOCK(&so->so_rcv);
4731 		/*
4732 		 * Remove out of band data so doesn't get presented to user.
4733 		 * This can happen independent of advancing the URG pointer,
4734 		 * but if two URG's are pending at once, some out-of-band
4735 		 * data may creep in... ick.
4736 		 */
4737 		if (th->th_urp <= (uint32_t) tlen &&
4738 		    !(so->so_options & SO_OOBINLINE)) {
4739 			/* hdr drop is delayed */
4740 			tcp_pulloutofband(so, th, m, drop_hdrlen);
4741 		}
4742 	} else {
4743 		/*
4744 		 * If no out of band data is expected, pull receive urgent
4745 		 * pointer along with the receive window.
4746 		 */
4747 		if (SEQ_GT(tp->rcv_nxt, tp->rcv_up))
4748 			tp->rcv_up = tp->rcv_nxt;
4749 	}
4750 dodata:				/* XXX */
4751 	INP_WLOCK_ASSERT(tp->t_inpcb);
4752 
4753 	/*
4754 	 * Process the segment text, merging it into the TCP sequencing
4755 	 * queue, and arranging for acknowledgment of receipt if necessary.
4756 	 * This process logically involves adjusting tp->rcv_wnd as data is
4757 	 * presented to the user (this happens in tcp_usrreq.c, case
4758 	 * PRU_RCVD).  If a FIN has already been received on this connection
4759 	 * then we just ignore the text.
4760 	 */
4761 	tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) &&
4762 		   IS_FASTOPEN(tp->t_flags));
4763 	if ((tlen || (thflags & TH_FIN) || tfo_syn) &&
4764 	    TCPS_HAVERCVDFIN(tp->t_state) == 0) {
4765 		tcp_seq save_start = th->th_seq;
4766 
4767 		m_adj(m, drop_hdrlen);	/* delayed header drop */
4768 		/*
4769 		 * Insert segment which includes th into TCP reassembly
4770 		 * queue with control block tp.  Set thflags to whether
4771 		 * reassembly now includes a segment with FIN.  This handles
4772 		 * the common case inline (segment is the next to be
4773 		 * received on an established connection, and the queue is
4774 		 * empty), avoiding linkage into and removal from the queue
4775 		 * and repetition of various conversions. Set DELACK for
4776 		 * segments received in order, but ack immediately when
4777 		 * segments are out of order (so fast retransmit can work).
4778 		 */
4779 		if (th->th_seq == tp->rcv_nxt &&
4780 		    LIST_EMPTY(&tp->t_segq) &&
4781 		    (TCPS_HAVEESTABLISHED(tp->t_state) ||
4782 		    tfo_syn)) {
4783 			if (DELAY_ACK(tp, tlen) || tfo_syn) {
4784 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
4785 				tp->t_flags |= TF_DELACK;
4786 			} else {
4787 				rack->r_wanted_output++;
4788 				tp->t_flags |= TF_ACKNOW;
4789 			}
4790 			tp->rcv_nxt += tlen;
4791 			thflags = th->th_flags & TH_FIN;
4792 			TCPSTAT_ADD(tcps_rcvpack, nsegs);
4793 			TCPSTAT_ADD(tcps_rcvbyte, tlen);
4794 			SOCKBUF_LOCK(&so->so_rcv);
4795 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
4796 				m_freem(m);
4797 			else
4798 				sbappendstream_locked(&so->so_rcv, m, 0);
4799 			/* NB: sorwakeup_locked() does an implicit unlock. */
4800 			sorwakeup_locked(so);
4801 		} else {
4802 			/*
4803 			 * XXX: Due to the header drop above "th" is
4804 			 * theoretically invalid by now.  Fortunately
4805 			 * m_adj() doesn't actually frees any mbufs when
4806 			 * trimming from the head.
4807 			 */
4808 			thflags = tcp_reass(tp, th, &tlen, m);
4809 			tp->t_flags |= TF_ACKNOW;
4810 		}
4811 		if (tlen > 0)
4812 			tcp_update_sack_list(tp, save_start, save_start + tlen);
4813 	} else {
4814 		m_freem(m);
4815 		thflags &= ~TH_FIN;
4816 	}
4817 
4818 	/*
4819 	 * If FIN is received ACK the FIN and let the user know that the
4820 	 * connection is closing.
4821 	 */
4822 	if (thflags & TH_FIN) {
4823 		if (TCPS_HAVERCVDFIN(tp->t_state) == 0) {
4824 			socantrcvmore(so);
4825 			/*
4826 			 * If connection is half-synchronized (ie NEEDSYN
4827 			 * flag on) then delay ACK, so it may be piggybacked
4828 			 * when SYN is sent. Otherwise, since we received a
4829 			 * FIN then no more input can be expected, send ACK
4830 			 * now.
4831 			 */
4832 			if (tp->t_flags & TF_NEEDSYN) {
4833 				rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
4834 				tp->t_flags |= TF_DELACK;
4835 			} else {
4836 				tp->t_flags |= TF_ACKNOW;
4837 			}
4838 			tp->rcv_nxt++;
4839 		}
4840 		switch (tp->t_state) {
4841 
4842 			/*
4843 			 * In SYN_RECEIVED and ESTABLISHED STATES enter the
4844 			 * CLOSE_WAIT state.
4845 			 */
4846 		case TCPS_SYN_RECEIVED:
4847 			tp->t_starttime = ticks;
4848 			/* FALLTHROUGH */
4849 		case TCPS_ESTABLISHED:
4850 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
4851 			tcp_state_change(tp, TCPS_CLOSE_WAIT);
4852 			break;
4853 
4854 			/*
4855 			 * If still in FIN_WAIT_1 STATE FIN has not been
4856 			 * acked so enter the CLOSING state.
4857 			 */
4858 		case TCPS_FIN_WAIT_1:
4859 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
4860 			tcp_state_change(tp, TCPS_CLOSING);
4861 			break;
4862 
4863 			/*
4864 			 * In FIN_WAIT_2 state enter the TIME_WAIT state,
4865 			 * starting the time-wait timer, turning off the
4866 			 * other standard timers.
4867 			 */
4868 		case TCPS_FIN_WAIT_2:
4869 			rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
4870 			INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
4871 			tcp_twstart(tp);
4872 			return (1);
4873 		}
4874 	}
4875 	/*
4876 	 * Return any desired output.
4877 	 */
4878 	if ((tp->t_flags & TF_ACKNOW) || (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) {
4879 		rack->r_wanted_output++;
4880 	}
4881 	INP_WLOCK_ASSERT(tp->t_inpcb);
4882 	return (0);
4883 }
4884 
4885 /*
4886  * Here nothing is really faster, its just that we
4887  * have broken out the fast-data path also just like
4888  * the fast-ack.
4889  */
4890 static int
4891 rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so,
4892     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
4893     uint32_t tiwin, int32_t nxt_pkt)
4894 {
4895 	int32_t nsegs;
4896 	int32_t newsize = 0;	/* automatic sockbuf scaling */
4897 	struct tcp_rack *rack;
4898 #ifdef TCPDEBUG
4899 	/*
4900 	 * The size of tcp_saveipgen must be the size of the max ip header,
4901 	 * now IPv6.
4902 	 */
4903 	u_char tcp_saveipgen[IP6_HDR_LEN];
4904 	struct tcphdr tcp_savetcp;
4905 	short ostate = 0;
4906 
4907 #endif
4908 	/*
4909 	 * If last ACK falls within this segment's sequence numbers, record
4910 	 * the timestamp. NOTE that the test is modified according to the
4911 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
4912 	 */
4913 	if (__predict_false(th->th_seq != tp->rcv_nxt)) {
4914 		return (0);
4915 	}
4916 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
4917 		return (0);
4918 	}
4919 	if (tiwin && tiwin != tp->snd_wnd) {
4920 		return (0);
4921 	}
4922 	if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) {
4923 		return (0);
4924 	}
4925 	if (__predict_false((to->to_flags & TOF_TS) &&
4926 	    (TSTMP_LT(to->to_tsval, tp->ts_recent)))) {
4927 		return (0);
4928 	}
4929 	if (__predict_false((th->th_ack != tp->snd_una))) {
4930 		return (0);
4931 	}
4932 	if (__predict_false(tlen > sbspace(&so->so_rcv))) {
4933 		return (0);
4934 	}
4935 	if ((to->to_flags & TOF_TS) != 0 &&
4936 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
4937 		tp->ts_recent_age = tcp_ts_getticks();
4938 		tp->ts_recent = to->to_tsval;
4939 	}
4940 	rack = (struct tcp_rack *)tp->t_fb_ptr;
4941 	/*
4942 	 * This is a pure, in-sequence data packet with nothing on the
4943 	 * reassembly queue and we have enough buffer space to take it.
4944 	 */
4945 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
4946 
4947 
4948 	/* Clean receiver SACK report if present */
4949 	if (tp->rcv_numsacks)
4950 		tcp_clean_sackreport(tp);
4951 	TCPSTAT_INC(tcps_preddat);
4952 	tp->rcv_nxt += tlen;
4953 	/*
4954 	 * Pull snd_wl1 up to prevent seq wrap relative to th_seq.
4955 	 */
4956 	tp->snd_wl1 = th->th_seq;
4957 	/*
4958 	 * Pull rcv_up up to prevent seq wrap relative to rcv_nxt.
4959 	 */
4960 	tp->rcv_up = tp->rcv_nxt;
4961 	TCPSTAT_ADD(tcps_rcvpack, nsegs);
4962 	TCPSTAT_ADD(tcps_rcvbyte, tlen);
4963 #ifdef TCPDEBUG
4964 	if (so->so_options & SO_DEBUG)
4965 		tcp_trace(TA_INPUT, ostate, tp,
4966 		    (void *)tcp_saveipgen, &tcp_savetcp, 0);
4967 #endif
4968 	newsize = tcp_autorcvbuf(m, th, so, tp, tlen);
4969 
4970 	/* Add data to socket buffer. */
4971 	SOCKBUF_LOCK(&so->so_rcv);
4972 	if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
4973 		m_freem(m);
4974 	} else {
4975 		/*
4976 		 * Set new socket buffer size. Give up when limit is
4977 		 * reached.
4978 		 */
4979 		if (newsize)
4980 			if (!sbreserve_locked(&so->so_rcv,
4981 			    newsize, so, NULL))
4982 				so->so_rcv.sb_flags &= ~SB_AUTOSIZE;
4983 		m_adj(m, drop_hdrlen);	/* delayed header drop */
4984 		sbappendstream_locked(&so->so_rcv, m, 0);
4985 		rack_calc_rwin(so, tp);
4986 	}
4987 	/* NB: sorwakeup_locked() does an implicit unlock. */
4988 	sorwakeup_locked(so);
4989 	if (DELAY_ACK(tp, tlen)) {
4990 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
4991 		tp->t_flags |= TF_DELACK;
4992 	} else {
4993 		tp->t_flags |= TF_ACKNOW;
4994 		rack->r_wanted_output++;
4995 	}
4996 	if ((tp->snd_una == tp->snd_max) && rack_use_sack_filter)
4997 		sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una);
4998 	return (1);
4999 }
5000 
5001 /*
5002  * This subfunction is used to try to highly optimize the
5003  * fast path. We again allow window updates that are
5004  * in sequence to remain in the fast-path. We also add
5005  * in the __predict's to attempt to help the compiler.
5006  * Note that if we return a 0, then we can *not* process
5007  * it and the caller should push the packet into the
5008  * slow-path.
5009  */
5010 static int
5011 rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
5012     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
5013     uint32_t tiwin, int32_t nxt_pkt, uint32_t cts)
5014 {
5015 	int32_t acked;
5016 	int32_t nsegs;
5017 
5018 #ifdef TCPDEBUG
5019 	/*
5020 	 * The size of tcp_saveipgen must be the size of the max ip header,
5021 	 * now IPv6.
5022 	 */
5023 	u_char tcp_saveipgen[IP6_HDR_LEN];
5024 	struct tcphdr tcp_savetcp;
5025 	short ostate = 0;
5026 
5027 #endif
5028 	struct tcp_rack *rack;
5029 
5030 	if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) {
5031 		/* Old ack, behind (or duplicate to) the last one rcv'd */
5032 		return (0);
5033 	}
5034 	if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) {
5035 		/* Above what we have sent? */
5036 		return (0);
5037 	}
5038 	if (__predict_false(tp->snd_nxt != tp->snd_max)) {
5039 		/* We are retransmitting */
5040 		return (0);
5041 	}
5042 	if (__predict_false(tiwin == 0)) {
5043 		/* zero window */
5044 		return (0);
5045 	}
5046 	if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) {
5047 		/* We need a SYN or a FIN, unlikely.. */
5048 		return (0);
5049 	}
5050 	if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) {
5051 		/* Timestamp is behind .. old ack with seq wrap? */
5052 		return (0);
5053 	}
5054 	if (__predict_false(IN_RECOVERY(tp->t_flags))) {
5055 		/* Still recovering */
5056 		return (0);
5057 	}
5058 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5059 	if (rack->r_ctl.rc_sacked) {
5060 		/* We have sack holes on our scoreboard */
5061 		return (0);
5062 	}
5063 	/* Ok if we reach here, we can process a fast-ack */
5064 	nsegs = max(1, m->m_pkthdr.lro_nsegs);
5065 	rack_log_ack(tp, to, th);
5066 	/* Did the window get updated? */
5067 	if (tiwin != tp->snd_wnd) {
5068 		tp->snd_wnd = tiwin;
5069 		tp->snd_wl1 = th->th_seq;
5070 		if (tp->snd_wnd > tp->max_sndwnd)
5071 			tp->max_sndwnd = tp->snd_wnd;
5072 	}
5073 	if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= tp->t_maxseg)) {
5074 		rack_exit_persist(tp, rack);
5075 	}
5076 	/*
5077 	 * If last ACK falls within this segment's sequence numbers, record
5078 	 * the timestamp. NOTE that the test is modified according to the
5079 	 * latest proposal of the tcplw@cray.com list (Braden 1993/04/26).
5080 	 */
5081 	if ((to->to_flags & TOF_TS) != 0 &&
5082 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
5083 		tp->ts_recent_age = tcp_ts_getticks();
5084 		tp->ts_recent = to->to_tsval;
5085 	}
5086 	/*
5087 	 * This is a pure ack for outstanding data.
5088 	 */
5089 	TCPSTAT_INC(tcps_predack);
5090 
5091 	/*
5092 	 * "bad retransmit" recovery.
5093 	 */
5094 	if (tp->t_flags & TF_PREVVALID) {
5095 		tp->t_flags &= ~TF_PREVVALID;
5096 		if (tp->t_rxtshift == 1 &&
5097 		    (int)(ticks - tp->t_badrxtwin) < 0)
5098 			rack_cong_signal(tp, th, CC_RTO_ERR);
5099 	}
5100 	/*
5101 	 * Recalculate the transmit timer / rtt.
5102 	 *
5103 	 * Some boxes send broken timestamp replies during the SYN+ACK
5104 	 * phase, ignore timestamps of 0 or we could calculate a huge RTT
5105 	 * and blow up the retransmit timer.
5106 	 */
5107 	acked = BYTES_THIS_ACK(tp, th);
5108 
5109 #ifdef TCP_HHOOK
5110 	/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
5111 	hhook_run_tcp_est_in(tp, th, to);
5112 #endif
5113 
5114 	TCPSTAT_ADD(tcps_rcvackpack, nsegs);
5115 	TCPSTAT_ADD(tcps_rcvackbyte, acked);
5116 	sbdrop(&so->so_snd, acked);
5117 	/*
5118 	 * Let the congestion control algorithm update congestion control
5119 	 * related information. This typically means increasing the
5120 	 * congestion window.
5121 	 */
5122 	rack_ack_received(tp, rack, th, nsegs, CC_ACK, 0);
5123 
5124 	tp->snd_una = th->th_ack;
5125 	/*
5126 	 * Pull snd_wl2 up to prevent seq wrap relative to th_ack.
5127 	 */
5128 	tp->snd_wl2 = th->th_ack;
5129 	tp->t_dupacks = 0;
5130 	m_freem(m);
5131 	/* ND6_HINT(tp);	 *//* Some progress has been made. */
5132 
5133 	/*
5134 	 * If all outstanding data are acked, stop retransmit timer,
5135 	 * otherwise restart timer using current (possibly backed-off)
5136 	 * value. If process is waiting for space, wakeup/selwakeup/signal.
5137 	 * If data are ready to send, let tcp_output decide between more
5138 	 * output or persist.
5139 	 */
5140 #ifdef TCPDEBUG
5141 	if (so->so_options & SO_DEBUG)
5142 		tcp_trace(TA_INPUT, ostate, tp,
5143 		    (void *)tcp_saveipgen,
5144 		    &tcp_savetcp, 0);
5145 #endif
5146 	if (tp->snd_una == tp->snd_max) {
5147 		rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__);
5148 		tp->t_acktime = 0;
5149 		rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
5150 	}
5151 	/* Wake up the socket if we have room to write more */
5152 	sowwakeup(so);
5153 	if (sbavail(&so->so_snd)) {
5154 		rack->r_wanted_output++;
5155 	}
5156 	return (1);
5157 }
5158 
5159 /*
5160  * Return value of 1, the TCB is unlocked and most
5161  * likely gone, return value of 0, the TCP is still
5162  * locked.
5163  */
5164 static int
5165 rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so,
5166     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
5167     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
5168 {
5169 	int32_t ret_val = 0;
5170 	int32_t todrop;
5171 	int32_t ourfinisacked = 0;
5172 
5173 	rack_calc_rwin(so, tp);
5174 	/*
5175 	 * If the state is SYN_SENT: if seg contains an ACK, but not for our
5176 	 * SYN, drop the input. if seg contains a RST, then drop the
5177 	 * connection. if seg does not contain SYN, then drop it. Otherwise
5178 	 * this is an acceptable SYN segment initialize tp->rcv_nxt and
5179 	 * tp->irs if seg contains ack then advance tp->snd_una if seg
5180 	 * contains an ECE and ECN support is enabled, the stream is ECN
5181 	 * capable. if SYN has been acked change to ESTABLISHED else
5182 	 * SYN_RCVD state arrange for segment to be acked (eventually)
5183 	 * continue processing rest of data/controls, beginning with URG
5184 	 */
5185 	if ((thflags & TH_ACK) &&
5186 	    (SEQ_LEQ(th->th_ack, tp->iss) ||
5187 	    SEQ_GT(th->th_ack, tp->snd_max))) {
5188 		rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
5189 		return (1);
5190 	}
5191 	if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) {
5192 		TCP_PROBE5(connect__refused, NULL, tp,
5193 		    mtod(m, const char *), tp, th);
5194 		tp = tcp_drop(tp, ECONNREFUSED);
5195 		rack_do_drop(m, tp);
5196 		return (1);
5197 	}
5198 	if (thflags & TH_RST) {
5199 		rack_do_drop(m, tp);
5200 		return (1);
5201 	}
5202 	if (!(thflags & TH_SYN)) {
5203 		rack_do_drop(m, tp);
5204 		return (1);
5205 	}
5206 	tp->irs = th->th_seq;
5207 	tcp_rcvseqinit(tp);
5208 	if (thflags & TH_ACK) {
5209 		int tfo_partial = 0;
5210 
5211 		TCPSTAT_INC(tcps_connects);
5212 		soisconnected(so);
5213 #ifdef MAC
5214 		mac_socketpeer_set_from_mbuf(m, so);
5215 #endif
5216 		/* Do window scaling on this connection? */
5217 		if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
5218 		    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
5219 			tp->rcv_scale = tp->request_r_scale;
5220 		}
5221 		tp->rcv_adv += min(tp->rcv_wnd,
5222 		    TCP_MAXWIN << tp->rcv_scale);
5223 		/*
5224 		 * If not all the data that was sent in the TFO SYN
5225 		 * has been acked, resend the remainder right away.
5226 		 */
5227 		if (IS_FASTOPEN(tp->t_flags) &&
5228 		    (tp->snd_una != tp->snd_max)) {
5229 			tp->snd_nxt = th->th_ack;
5230 			tfo_partial = 1;
5231 		}
5232 		/*
5233 		 * If there's data, delay ACK; if there's also a FIN ACKNOW
5234 		 * will be turned on later.
5235 		 */
5236 		if (DELAY_ACK(tp, tlen) && tlen != 0 && (tfo_partial == 0)) {
5237 			rack_timer_cancel(tp, (struct tcp_rack *)tp->t_fb_ptr,
5238 					  ((struct tcp_rack *)tp->t_fb_ptr)->r_ctl.rc_rcvtime, __LINE__);
5239 			tp->t_flags |= TF_DELACK;
5240 		} else {
5241 			((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output++;
5242 			tp->t_flags |= TF_ACKNOW;
5243 		}
5244 
5245 		if ((thflags & TH_ECE) && V_tcp_do_ecn) {
5246 			tp->t_flags |= TF_ECN_PERMIT;
5247 			TCPSTAT_INC(tcps_ecn_shs);
5248 		}
5249 		if (SEQ_GT(th->th_ack, tp->snd_una)) {
5250 			/*
5251 			 * We advance snd_una for the
5252 			 * fast open case. If th_ack is
5253 			 * acknowledging data beyond
5254 			 * snd_una we can't just call
5255 			 * ack-processing since the
5256 			 * data stream in our send-map
5257 			 * will start at snd_una + 1 (one
5258 			 * beyond the SYN). If its just
5259 			 * equal we don't need to do that
5260 			 * and there is no send_map.
5261 			 */
5262 			tp->snd_una++;
5263 		}
5264 		/*
5265 		 * Received <SYN,ACK> in SYN_SENT[*] state. Transitions:
5266 		 * SYN_SENT  --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1
5267 		 */
5268 		tp->t_starttime = ticks;
5269 		if (tp->t_flags & TF_NEEDFIN) {
5270 			tcp_state_change(tp, TCPS_FIN_WAIT_1);
5271 			tp->t_flags &= ~TF_NEEDFIN;
5272 			thflags &= ~TH_SYN;
5273 		} else {
5274 			tcp_state_change(tp, TCPS_ESTABLISHED);
5275 			TCP_PROBE5(connect__established, NULL, tp,
5276 			    mtod(m, const char *), tp, th);
5277 			cc_conn_init(tp);
5278 		}
5279 	} else {
5280 		/*
5281 		 * Received initial SYN in SYN-SENT[*] state => simultaneous
5282 		 * open.  If segment contains CC option and there is a
5283 		 * cached CC, apply TAO test. If it succeeds, connection is *
5284 		 * half-synchronized. Otherwise, do 3-way handshake:
5285 		 * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If
5286 		 * there was no CC option, clear cached CC value.
5287 		 */
5288 		tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
5289 		tcp_state_change(tp, TCPS_SYN_RECEIVED);
5290 	}
5291 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
5292 	INP_WLOCK_ASSERT(tp->t_inpcb);
5293 	/*
5294 	 * Advance th->th_seq to correspond to first data byte. If data,
5295 	 * trim to stay within window, dropping FIN if necessary.
5296 	 */
5297 	th->th_seq++;
5298 	if (tlen > tp->rcv_wnd) {
5299 		todrop = tlen - tp->rcv_wnd;
5300 		m_adj(m, -todrop);
5301 		tlen = tp->rcv_wnd;
5302 		thflags &= ~TH_FIN;
5303 		TCPSTAT_INC(tcps_rcvpackafterwin);
5304 		TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop);
5305 	}
5306 	tp->snd_wl1 = th->th_seq - 1;
5307 	tp->rcv_up = th->th_seq;
5308 	/*
5309 	 * Client side of transaction: already sent SYN and data. If the
5310 	 * remote host used T/TCP to validate the SYN, our data will be
5311 	 * ACK'd; if so, enter normal data segment processing in the middle
5312 	 * of step 5, ack processing. Otherwise, goto step 6.
5313 	 */
5314 	if (thflags & TH_ACK) {
5315 		if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val))
5316 			return (ret_val);
5317 		/* We may have changed to FIN_WAIT_1 above */
5318 		if (tp->t_state == TCPS_FIN_WAIT_1) {
5319 			/*
5320 			 * In FIN_WAIT_1 STATE in addition to the processing
5321 			 * for the ESTABLISHED state if our FIN is now
5322 			 * acknowledged then enter FIN_WAIT_2.
5323 			 */
5324 			if (ourfinisacked) {
5325 				/*
5326 				 * If we can't receive any more data, then
5327 				 * closing user can proceed. Starting the
5328 				 * timer is contrary to the specification,
5329 				 * but if we don't get a FIN we'll hang
5330 				 * forever.
5331 				 *
5332 				 * XXXjl: we should release the tp also, and
5333 				 * use a compressed state.
5334 				 */
5335 				if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
5336 					soisdisconnected(so);
5337 					tcp_timer_activate(tp, TT_2MSL,
5338 					    (tcp_fast_finwait2_recycle ?
5339 					    tcp_finwait2_timeout :
5340 					    TP_MAXIDLE(tp)));
5341 				}
5342 				tcp_state_change(tp, TCPS_FIN_WAIT_2);
5343 			}
5344 		}
5345 	}
5346 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
5347 	   tiwin, thflags, nxt_pkt));
5348 }
5349 
5350 /*
5351  * Return value of 1, the TCB is unlocked and most
5352  * likely gone, return value of 0, the TCP is still
5353  * locked.
5354  */
5355 static int
5356 rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so,
5357     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
5358     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
5359 {
5360 	int32_t ret_val = 0;
5361 	int32_t ourfinisacked = 0;
5362 
5363 	rack_calc_rwin(so, tp);
5364 
5365 	if ((thflags & TH_ACK) &&
5366 	    (SEQ_LEQ(th->th_ack, tp->snd_una) ||
5367 	    SEQ_GT(th->th_ack, tp->snd_max))) {
5368 		rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
5369 		return (1);
5370 	}
5371 	if (IS_FASTOPEN(tp->t_flags)) {
5372 		/*
5373 		 * When a TFO connection is in SYN_RECEIVED, the
5374 		 * only valid packets are the initial SYN, a
5375 		 * retransmit/copy of the initial SYN (possibly with
5376 		 * a subset of the original data), a valid ACK, a
5377 		 * FIN, or a RST.
5378 		 */
5379 		if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) {
5380 			rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
5381 			return (1);
5382 		} else if (thflags & TH_SYN) {
5383 			/* non-initial SYN is ignored */
5384 			struct tcp_rack *rack;
5385 
5386 			rack = (struct tcp_rack *)tp->t_fb_ptr;
5387 			if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) ||
5388 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) ||
5389 			    (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) {
5390 				rack_do_drop(m, NULL);
5391 				return (0);
5392 			}
5393 		} else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) {
5394 			rack_do_drop(m, NULL);
5395 			return (0);
5396 		}
5397 	}
5398 	if (thflags & TH_RST)
5399 		return (rack_process_rst(m, th, so, tp));
5400 	/*
5401 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
5402 	 * synchronized state.
5403 	 */
5404 	if (thflags & TH_SYN) {
5405 		rack_challenge_ack(m, th, tp, &ret_val);
5406 		return (ret_val);
5407 	}
5408 	/*
5409 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
5410 	 * it's less than ts_recent, drop it.
5411 	 */
5412 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
5413 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
5414 		if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val))
5415 			return (ret_val);
5416 	}
5417 	/*
5418 	 * In the SYN-RECEIVED state, validate that the packet belongs to
5419 	 * this connection before trimming the data to fit the receive
5420 	 * window.  Check the sequence number versus IRS since we know the
5421 	 * sequence numbers haven't wrapped.  This is a partial fix for the
5422 	 * "LAND" DoS attack.
5423 	 */
5424 	if (SEQ_LT(th->th_seq, tp->irs)) {
5425 		rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
5426 		return (1);
5427 	}
5428 	if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
5429 		return (ret_val);
5430 	}
5431 	/*
5432 	 * If last ACK falls within this segment's sequence numbers, record
5433 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
5434 	 * from the latest proposal of the tcplw@cray.com list (Braden
5435 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
5436 	 * with our earlier PAWS tests, so this check should be solely
5437 	 * predicated on the sequence space of this segment. 3) That we
5438 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
5439 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
5440 	 * SEG.Len, This modified check allows us to overcome RFC1323's
5441 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
5442 	 * p.869. In such cases, we can still calculate the RTT correctly
5443 	 * when RCV.NXT == Last.ACK.Sent.
5444 	 */
5445 	if ((to->to_flags & TOF_TS) != 0 &&
5446 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
5447 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
5448 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
5449 		tp->ts_recent_age = tcp_ts_getticks();
5450 		tp->ts_recent = to->to_tsval;
5451 	}
5452 	/*
5453 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
5454 	 * is on (half-synchronized state), then queue data for later
5455 	 * processing; else drop segment and return.
5456 	 */
5457 	if ((thflags & TH_ACK) == 0) {
5458 		if (IS_FASTOPEN(tp->t_flags)) {
5459 			tp->snd_wnd = tiwin;
5460 			cc_conn_init(tp);
5461 		}
5462 		return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
5463 		    tiwin, thflags, nxt_pkt));
5464 	}
5465 	TCPSTAT_INC(tcps_connects);
5466 	soisconnected(so);
5467 	/* Do window scaling? */
5468 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
5469 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
5470 		tp->rcv_scale = tp->request_r_scale;
5471 		tp->snd_wnd = tiwin;
5472 	}
5473 	/*
5474 	 * Make transitions: SYN-RECEIVED  -> ESTABLISHED SYN-RECEIVED* ->
5475 	 * FIN-WAIT-1
5476 	 */
5477 	tp->t_starttime = ticks;
5478 	if (tp->t_flags & TF_NEEDFIN) {
5479 		tcp_state_change(tp, TCPS_FIN_WAIT_1);
5480 		tp->t_flags &= ~TF_NEEDFIN;
5481 	} else {
5482 		tcp_state_change(tp, TCPS_ESTABLISHED);
5483 		TCP_PROBE5(accept__established, NULL, tp,
5484 		    mtod(m, const char *), tp, th);
5485 		if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) {
5486 			tcp_fastopen_decrement_counter(tp->t_tfo_pending);
5487 			tp->t_tfo_pending = NULL;
5488 
5489 			/*
5490 			 * Account for the ACK of our SYN prior to regular
5491 			 * ACK processing below.
5492 			 */
5493 			tp->snd_una++;
5494 		}
5495 		/*
5496 		 * TFO connections call cc_conn_init() during SYN
5497 		 * processing.  Calling it again here for such connections
5498 		 * is not harmless as it would undo the snd_cwnd reduction
5499 		 * that occurs when a TFO SYN|ACK is retransmitted.
5500 		 */
5501 		if (!IS_FASTOPEN(tp->t_flags))
5502 			cc_conn_init(tp);
5503 	}
5504 	/*
5505 	 * If segment contains data or ACK, will call tcp_reass() later; if
5506 	 * not, do so now to pass queued data to user.
5507 	 */
5508 	if (tlen == 0 && (thflags & TH_FIN) == 0)
5509 		(void)tcp_reass(tp, (struct tcphdr *)0, 0,
5510 		    (struct mbuf *)0);
5511 	tp->snd_wl1 = th->th_seq - 1;
5512 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
5513 		return (ret_val);
5514 	}
5515 	if (tp->t_state == TCPS_FIN_WAIT_1) {
5516 		/* We could have went to FIN_WAIT_1 (or EST) above */
5517 		/*
5518 		 * In FIN_WAIT_1 STATE in addition to the processing for the
5519 		 * ESTABLISHED state if our FIN is now acknowledged then
5520 		 * enter FIN_WAIT_2.
5521 		 */
5522 		if (ourfinisacked) {
5523 			/*
5524 			 * If we can't receive any more data, then closing
5525 			 * user can proceed. Starting the timer is contrary
5526 			 * to the specification, but if we don't get a FIN
5527 			 * we'll hang forever.
5528 			 *
5529 			 * XXXjl: we should release the tp also, and use a
5530 			 * compressed state.
5531 			 */
5532 			if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
5533 				soisdisconnected(so);
5534 				tcp_timer_activate(tp, TT_2MSL,
5535 				    (tcp_fast_finwait2_recycle ?
5536 				    tcp_finwait2_timeout :
5537 				    TP_MAXIDLE(tp)));
5538 			}
5539 			tcp_state_change(tp, TCPS_FIN_WAIT_2);
5540 		}
5541 	}
5542 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
5543 	    tiwin, thflags, nxt_pkt));
5544 }
5545 
5546 /*
5547  * Return value of 1, the TCB is unlocked and most
5548  * likely gone, return value of 0, the TCP is still
5549  * locked.
5550  */
5551 static int
5552 rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so,
5553     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
5554     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
5555 {
5556 	int32_t ret_val = 0;
5557 
5558 	/*
5559 	 * Header prediction: check for the two common cases of a
5560 	 * uni-directional data xfer.  If the packet has no control flags,
5561 	 * is in-sequence, the window didn't change and we're not
5562 	 * retransmitting, it's a candidate.  If the length is zero and the
5563 	 * ack moved forward, we're the sender side of the xfer.  Just free
5564 	 * the data acked & wake any higher level process that was blocked
5565 	 * waiting for space.  If the length is non-zero and the ack didn't
5566 	 * move, we're the receiver side.  If we're getting packets in-order
5567 	 * (the reassembly queue is empty), add the data toc The socket
5568 	 * buffer and note that we need a delayed ack. Make sure that the
5569 	 * hidden state-flags are also off. Since we check for
5570 	 * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN.
5571 	 */
5572 	if (__predict_true(((to->to_flags & TOF_SACK) == 0)) &&
5573 	    __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) &&
5574 	    __predict_true(LIST_EMPTY(&tp->t_segq)) &&
5575 	    __predict_true(th->th_seq == tp->rcv_nxt)) {
5576 		struct tcp_rack *rack;
5577 
5578 		rack = (struct tcp_rack *)tp->t_fb_ptr;
5579 		if (tlen == 0) {
5580 			if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen,
5581 			    tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) {
5582 				return (0);
5583 			}
5584 		} else {
5585 			if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen,
5586 			    tiwin, nxt_pkt)) {
5587 				return (0);
5588 			}
5589 		}
5590 	}
5591 	rack_calc_rwin(so, tp);
5592 
5593 	if (thflags & TH_RST)
5594 		return (rack_process_rst(m, th, so, tp));
5595 
5596 	/*
5597 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
5598 	 * synchronized state.
5599 	 */
5600 	if (thflags & TH_SYN) {
5601 		rack_challenge_ack(m, th, tp, &ret_val);
5602 		return (ret_val);
5603 	}
5604 	/*
5605 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
5606 	 * it's less than ts_recent, drop it.
5607 	 */
5608 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
5609 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
5610 		if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val))
5611 			return (ret_val);
5612 	}
5613 	if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
5614 		return (ret_val);
5615 	}
5616 	/*
5617 	 * If last ACK falls within this segment's sequence numbers, record
5618 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
5619 	 * from the latest proposal of the tcplw@cray.com list (Braden
5620 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
5621 	 * with our earlier PAWS tests, so this check should be solely
5622 	 * predicated on the sequence space of this segment. 3) That we
5623 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
5624 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
5625 	 * SEG.Len, This modified check allows us to overcome RFC1323's
5626 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
5627 	 * p.869. In such cases, we can still calculate the RTT correctly
5628 	 * when RCV.NXT == Last.ACK.Sent.
5629 	 */
5630 	if ((to->to_flags & TOF_TS) != 0 &&
5631 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
5632 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
5633 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
5634 		tp->ts_recent_age = tcp_ts_getticks();
5635 		tp->ts_recent = to->to_tsval;
5636 	}
5637 	/*
5638 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
5639 	 * is on (half-synchronized state), then queue data for later
5640 	 * processing; else drop segment and return.
5641 	 */
5642 	if ((thflags & TH_ACK) == 0) {
5643 		if (tp->t_flags & TF_NEEDSYN) {
5644 
5645 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
5646 			    tiwin, thflags, nxt_pkt));
5647 
5648 		} else if (tp->t_flags & TF_ACKNOW) {
5649 			rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
5650 			return (ret_val);
5651 		} else {
5652 			rack_do_drop(m, NULL);
5653 			return (0);
5654 		}
5655 	}
5656 	/*
5657 	 * Ack processing.
5658 	 */
5659 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
5660 		return (ret_val);
5661 	}
5662 	if (sbavail(&so->so_snd)) {
5663 		if (rack_progress_timeout_check(tp)) {
5664 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
5665 			rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
5666 			return (1);
5667 		}
5668 	}
5669 	/* State changes only happen in rack_process_data() */
5670 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
5671 	    tiwin, thflags, nxt_pkt));
5672 }
5673 
5674 /*
5675  * Return value of 1, the TCB is unlocked and most
5676  * likely gone, return value of 0, the TCP is still
5677  * locked.
5678  */
5679 static int
5680 rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so,
5681     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
5682     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
5683 {
5684 	int32_t ret_val = 0;
5685 
5686 	rack_calc_rwin(so, tp);
5687 	if (thflags & TH_RST)
5688 		return (rack_process_rst(m, th, so, tp));
5689 	/*
5690 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
5691 	 * synchronized state.
5692 	 */
5693 	if (thflags & TH_SYN) {
5694 		rack_challenge_ack(m, th, tp, &ret_val);
5695 		return (ret_val);
5696 	}
5697 	/*
5698 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
5699 	 * it's less than ts_recent, drop it.
5700 	 */
5701 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
5702 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
5703 		if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val))
5704 			return (ret_val);
5705 	}
5706 	if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
5707 		return (ret_val);
5708 	}
5709 	/*
5710 	 * If last ACK falls within this segment's sequence numbers, record
5711 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
5712 	 * from the latest proposal of the tcplw@cray.com list (Braden
5713 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
5714 	 * with our earlier PAWS tests, so this check should be solely
5715 	 * predicated on the sequence space of this segment. 3) That we
5716 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
5717 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
5718 	 * SEG.Len, This modified check allows us to overcome RFC1323's
5719 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
5720 	 * p.869. In such cases, we can still calculate the RTT correctly
5721 	 * when RCV.NXT == Last.ACK.Sent.
5722 	 */
5723 	if ((to->to_flags & TOF_TS) != 0 &&
5724 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
5725 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
5726 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
5727 		tp->ts_recent_age = tcp_ts_getticks();
5728 		tp->ts_recent = to->to_tsval;
5729 	}
5730 	/*
5731 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
5732 	 * is on (half-synchronized state), then queue data for later
5733 	 * processing; else drop segment and return.
5734 	 */
5735 	if ((thflags & TH_ACK) == 0) {
5736 		if (tp->t_flags & TF_NEEDSYN) {
5737 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
5738 			    tiwin, thflags, nxt_pkt));
5739 
5740 		} else if (tp->t_flags & TF_ACKNOW) {
5741 			rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
5742 			return (ret_val);
5743 		} else {
5744 			rack_do_drop(m, NULL);
5745 			return (0);
5746 		}
5747 	}
5748 	/*
5749 	 * Ack processing.
5750 	 */
5751 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) {
5752 		return (ret_val);
5753 	}
5754 	if (sbavail(&so->so_snd)) {
5755 		if (rack_progress_timeout_check(tp)) {
5756 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
5757 			rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
5758 			return (1);
5759 		}
5760 	}
5761 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
5762 	    tiwin, thflags, nxt_pkt));
5763 }
5764 
5765 static int
5766 rack_check_data_after_close(struct mbuf *m,
5767     struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so)
5768 {
5769 	struct tcp_rack *rack;
5770 
5771 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
5772 	rack = (struct tcp_rack *)tp->t_fb_ptr;
5773 	if (rack->rc_allow_data_af_clo == 0) {
5774 	close_now:
5775 		tp = tcp_close(tp);
5776 		TCPSTAT_INC(tcps_rcvafterclose);
5777 		rack_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen));
5778 		return (1);
5779 	}
5780 	if (sbavail(&so->so_snd) == 0)
5781 		goto close_now;
5782 	/* Ok we allow data that is ignored and a followup reset */
5783 	tp->rcv_nxt = th->th_seq + *tlen;
5784 	tp->t_flags2 |= TF2_DROP_AF_DATA;
5785 	rack->r_wanted_output = 1;
5786 	*tlen = 0;
5787 	return (0);
5788 }
5789 
5790 /*
5791  * Return value of 1, the TCB is unlocked and most
5792  * likely gone, return value of 0, the TCP is still
5793  * locked.
5794  */
5795 static int
5796 rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so,
5797     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
5798     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
5799 {
5800 	int32_t ret_val = 0;
5801 	int32_t ourfinisacked = 0;
5802 
5803 	rack_calc_rwin(so, tp);
5804 
5805 	if (thflags & TH_RST)
5806 		return (rack_process_rst(m, th, so, tp));
5807 	/*
5808 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
5809 	 * synchronized state.
5810 	 */
5811 	if (thflags & TH_SYN) {
5812 		rack_challenge_ack(m, th, tp, &ret_val);
5813 		return (ret_val);
5814 	}
5815 	/*
5816 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
5817 	 * it's less than ts_recent, drop it.
5818 	 */
5819 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
5820 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
5821 		if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val))
5822 			return (ret_val);
5823 	}
5824 	if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
5825 		return (ret_val);
5826 	}
5827 	/*
5828 	 * If new data are received on a connection after the user processes
5829 	 * are gone, then RST the other end.
5830 	 */
5831 	if ((so->so_state & SS_NOFDREF) && tlen) {
5832 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
5833 			return (1);
5834 	}
5835 	/*
5836 	 * If last ACK falls within this segment's sequence numbers, record
5837 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
5838 	 * from the latest proposal of the tcplw@cray.com list (Braden
5839 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
5840 	 * with our earlier PAWS tests, so this check should be solely
5841 	 * predicated on the sequence space of this segment. 3) That we
5842 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
5843 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
5844 	 * SEG.Len, This modified check allows us to overcome RFC1323's
5845 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
5846 	 * p.869. In such cases, we can still calculate the RTT correctly
5847 	 * when RCV.NXT == Last.ACK.Sent.
5848 	 */
5849 	if ((to->to_flags & TOF_TS) != 0 &&
5850 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
5851 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
5852 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
5853 		tp->ts_recent_age = tcp_ts_getticks();
5854 		tp->ts_recent = to->to_tsval;
5855 	}
5856 	/*
5857 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
5858 	 * is on (half-synchronized state), then queue data for later
5859 	 * processing; else drop segment and return.
5860 	 */
5861 	if ((thflags & TH_ACK) == 0) {
5862 		if (tp->t_flags & TF_NEEDSYN) {
5863 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
5864 			    tiwin, thflags, nxt_pkt));
5865 		} else if (tp->t_flags & TF_ACKNOW) {
5866 			rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
5867 			return (ret_val);
5868 		} else {
5869 			rack_do_drop(m, NULL);
5870 			return (0);
5871 		}
5872 	}
5873 	/*
5874 	 * Ack processing.
5875 	 */
5876 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
5877 		return (ret_val);
5878 	}
5879 	if (ourfinisacked) {
5880 		/*
5881 		 * If we can't receive any more data, then closing user can
5882 		 * proceed. Starting the timer is contrary to the
5883 		 * specification, but if we don't get a FIN we'll hang
5884 		 * forever.
5885 		 *
5886 		 * XXXjl: we should release the tp also, and use a
5887 		 * compressed state.
5888 		 */
5889 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
5890 			soisdisconnected(so);
5891 			tcp_timer_activate(tp, TT_2MSL,
5892 			    (tcp_fast_finwait2_recycle ?
5893 			    tcp_finwait2_timeout :
5894 			    TP_MAXIDLE(tp)));
5895 		}
5896 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
5897 	}
5898 	if (sbavail(&so->so_snd)) {
5899 		if (rack_progress_timeout_check(tp)) {
5900 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
5901 			rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
5902 			return (1);
5903 		}
5904 	}
5905 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
5906 	    tiwin, thflags, nxt_pkt));
5907 }
5908 
5909 /*
5910  * Return value of 1, the TCB is unlocked and most
5911  * likely gone, return value of 0, the TCP is still
5912  * locked.
5913  */
5914 static int
5915 rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so,
5916     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
5917     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
5918 {
5919 	int32_t ret_val = 0;
5920 	int32_t ourfinisacked = 0;
5921 
5922 	rack_calc_rwin(so, tp);
5923 
5924 	if (thflags & TH_RST)
5925 		return (rack_process_rst(m, th, so, tp));
5926 	/*
5927 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
5928 	 * synchronized state.
5929 	 */
5930 	if (thflags & TH_SYN) {
5931 		rack_challenge_ack(m, th, tp, &ret_val);
5932 		return (ret_val);
5933 	}
5934 	/*
5935 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
5936 	 * it's less than ts_recent, drop it.
5937 	 */
5938 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
5939 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
5940 		if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val))
5941 			return (ret_val);
5942 	}
5943 	if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
5944 		return (ret_val);
5945 	}
5946 	/*
5947 	 * If new data are received on a connection after the user processes
5948 	 * are gone, then RST the other end.
5949 	 */
5950 	if ((so->so_state & SS_NOFDREF) && tlen) {
5951 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
5952 			return (1);
5953 	}
5954 	/*
5955 	 * If last ACK falls within this segment's sequence numbers, record
5956 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
5957 	 * from the latest proposal of the tcplw@cray.com list (Braden
5958 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
5959 	 * with our earlier PAWS tests, so this check should be solely
5960 	 * predicated on the sequence space of this segment. 3) That we
5961 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
5962 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
5963 	 * SEG.Len, This modified check allows us to overcome RFC1323's
5964 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
5965 	 * p.869. In such cases, we can still calculate the RTT correctly
5966 	 * when RCV.NXT == Last.ACK.Sent.
5967 	 */
5968 	if ((to->to_flags & TOF_TS) != 0 &&
5969 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
5970 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
5971 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
5972 		tp->ts_recent_age = tcp_ts_getticks();
5973 		tp->ts_recent = to->to_tsval;
5974 	}
5975 	/*
5976 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
5977 	 * is on (half-synchronized state), then queue data for later
5978 	 * processing; else drop segment and return.
5979 	 */
5980 	if ((thflags & TH_ACK) == 0) {
5981 		if (tp->t_flags & TF_NEEDSYN) {
5982 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
5983 			    tiwin, thflags, nxt_pkt));
5984 		} else if (tp->t_flags & TF_ACKNOW) {
5985 			rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
5986 			return (ret_val);
5987 		} else {
5988 			rack_do_drop(m, NULL);
5989 			return (0);
5990 		}
5991 	}
5992 	/*
5993 	 * Ack processing.
5994 	 */
5995 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
5996 		return (ret_val);
5997 	}
5998 	if (ourfinisacked) {
5999 		INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
6000 		tcp_twstart(tp);
6001 		m_freem(m);
6002 		return (1);
6003 	}
6004 	if (sbavail(&so->so_snd)) {
6005 		if (rack_progress_timeout_check(tp)) {
6006 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
6007 			rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
6008 			return (1);
6009 		}
6010 	}
6011 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
6012 	    tiwin, thflags, nxt_pkt));
6013 }
6014 
6015 /*
6016  * Return value of 1, the TCB is unlocked and most
6017  * likely gone, return value of 0, the TCP is still
6018  * locked.
6019  */
6020 static int
6021 rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so,
6022     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
6023     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
6024 {
6025 	int32_t ret_val = 0;
6026 	int32_t ourfinisacked = 0;
6027 
6028 	rack_calc_rwin(so, tp);
6029 
6030 	if (thflags & TH_RST)
6031 		return (rack_process_rst(m, th, so, tp));
6032 	/*
6033 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
6034 	 * synchronized state.
6035 	 */
6036 	if (thflags & TH_SYN) {
6037 		rack_challenge_ack(m, th, tp, &ret_val);
6038 		return (ret_val);
6039 	}
6040 	/*
6041 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
6042 	 * it's less than ts_recent, drop it.
6043 	 */
6044 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
6045 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
6046 		if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val))
6047 			return (ret_val);
6048 	}
6049 	if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
6050 		return (ret_val);
6051 	}
6052 	/*
6053 	 * If new data are received on a connection after the user processes
6054 	 * are gone, then RST the other end.
6055 	 */
6056 	if ((so->so_state & SS_NOFDREF) && tlen) {
6057 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
6058 			return (1);
6059 	}
6060 	/*
6061 	 * If last ACK falls within this segment's sequence numbers, record
6062 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
6063 	 * from the latest proposal of the tcplw@cray.com list (Braden
6064 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
6065 	 * with our earlier PAWS tests, so this check should be solely
6066 	 * predicated on the sequence space of this segment. 3) That we
6067 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
6068 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
6069 	 * SEG.Len, This modified check allows us to overcome RFC1323's
6070 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
6071 	 * p.869. In such cases, we can still calculate the RTT correctly
6072 	 * when RCV.NXT == Last.ACK.Sent.
6073 	 */
6074 	if ((to->to_flags & TOF_TS) != 0 &&
6075 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
6076 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
6077 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
6078 		tp->ts_recent_age = tcp_ts_getticks();
6079 		tp->ts_recent = to->to_tsval;
6080 	}
6081 	/*
6082 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
6083 	 * is on (half-synchronized state), then queue data for later
6084 	 * processing; else drop segment and return.
6085 	 */
6086 	if ((thflags & TH_ACK) == 0) {
6087 		if (tp->t_flags & TF_NEEDSYN) {
6088 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
6089 			    tiwin, thflags, nxt_pkt));
6090 		} else if (tp->t_flags & TF_ACKNOW) {
6091 			rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
6092 			return (ret_val);
6093 		} else {
6094 			rack_do_drop(m, NULL);
6095 			return (0);
6096 		}
6097 	}
6098 	/*
6099 	 * case TCPS_LAST_ACK: Ack processing.
6100 	 */
6101 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
6102 		return (ret_val);
6103 	}
6104 	if (ourfinisacked) {
6105 		INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
6106 		tp = tcp_close(tp);
6107 		rack_do_drop(m, tp);
6108 		return (1);
6109 	}
6110 	if (sbavail(&so->so_snd)) {
6111 		if (rack_progress_timeout_check(tp)) {
6112 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
6113 			rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
6114 			return (1);
6115 		}
6116 	}
6117 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
6118 	    tiwin, thflags, nxt_pkt));
6119 }
6120 
6121 
6122 /*
6123  * Return value of 1, the TCB is unlocked and most
6124  * likely gone, return value of 0, the TCP is still
6125  * locked.
6126  */
6127 static int
6128 rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so,
6129     struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen,
6130     uint32_t tiwin, int32_t thflags, int32_t nxt_pkt)
6131 {
6132 	int32_t ret_val = 0;
6133 	int32_t ourfinisacked = 0;
6134 
6135 	rack_calc_rwin(so, tp);
6136 
6137 	/* Reset receive buffer auto scaling when not in bulk receive mode. */
6138 	if (thflags & TH_RST)
6139 		return (rack_process_rst(m, th, so, tp));
6140 	/*
6141 	 * RFC5961 Section 4.2 Send challenge ACK for any SYN in
6142 	 * synchronized state.
6143 	 */
6144 	if (thflags & TH_SYN) {
6145 		rack_challenge_ack(m, th, tp, &ret_val);
6146 		return (ret_val);
6147 	}
6148 	/*
6149 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment and
6150 	 * it's less than ts_recent, drop it.
6151 	 */
6152 	if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent &&
6153 	    TSTMP_LT(to->to_tsval, tp->ts_recent)) {
6154 		if (rack_ts_check(m, th, tp, tlen, thflags, &ret_val))
6155 			return (ret_val);
6156 	}
6157 	if (rack_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) {
6158 		return (ret_val);
6159 	}
6160 	/*
6161 	 * If new data are received on a connection after the user processes
6162 	 * are gone, then RST the other end.
6163 	 */
6164 	if ((so->so_state & SS_NOFDREF) &&
6165 	    tlen) {
6166 		if (rack_check_data_after_close(m, tp, &tlen, th, so))
6167 			return (1);
6168 	}
6169 	/*
6170 	 * If last ACK falls within this segment's sequence numbers, record
6171 	 * its timestamp. NOTE: 1) That the test incorporates suggestions
6172 	 * from the latest proposal of the tcplw@cray.com list (Braden
6173 	 * 1993/04/26). 2) That updating only on newer timestamps interferes
6174 	 * with our earlier PAWS tests, so this check should be solely
6175 	 * predicated on the sequence space of this segment. 3) That we
6176 	 * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ
6177 	 * + SEG.Len  instead of RFC1323's Last.ACK.Sent < SEG.SEQ +
6178 	 * SEG.Len, This modified check allows us to overcome RFC1323's
6179 	 * limitations as described in Stevens TCP/IP Illustrated Vol. 2
6180 	 * p.869. In such cases, we can still calculate the RTT correctly
6181 	 * when RCV.NXT == Last.ACK.Sent.
6182 	 */
6183 	if ((to->to_flags & TOF_TS) != 0 &&
6184 	    SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
6185 	    SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen +
6186 	    ((thflags & (TH_SYN | TH_FIN)) != 0))) {
6187 		tp->ts_recent_age = tcp_ts_getticks();
6188 		tp->ts_recent = to->to_tsval;
6189 	}
6190 	/*
6191 	 * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN flag
6192 	 * is on (half-synchronized state), then queue data for later
6193 	 * processing; else drop segment and return.
6194 	 */
6195 	if ((thflags & TH_ACK) == 0) {
6196 		if (tp->t_flags & TF_NEEDSYN) {
6197 			return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
6198 			    tiwin, thflags, nxt_pkt));
6199 		} else if (tp->t_flags & TF_ACKNOW) {
6200 			rack_do_dropafterack(m, tp, th, thflags, tlen, &ret_val);
6201 			return (ret_val);
6202 		} else {
6203 			rack_do_drop(m, NULL);
6204 			return (0);
6205 		}
6206 	}
6207 	/*
6208 	 * Ack processing.
6209 	 */
6210 	if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) {
6211 		return (ret_val);
6212 	}
6213 	if (sbavail(&so->so_snd)) {
6214 		if (rack_progress_timeout_check(tp)) {
6215 			tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT);
6216 			rack_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen);
6217 			return (1);
6218 		}
6219 	}
6220 	return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen,
6221 	    tiwin, thflags, nxt_pkt));
6222 }
6223 
6224 
6225 static void inline
6226 rack_clear_rate_sample(struct tcp_rack *rack)
6227 {
6228 	rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY;
6229 	rack->r_ctl.rack_rs.rs_rtt_cnt = 0;
6230 	rack->r_ctl.rack_rs.rs_rtt_tot = 0;
6231 }
6232 
6233 static int
6234 rack_init(struct tcpcb *tp)
6235 {
6236 	struct tcp_rack *rack = NULL;
6237 
6238 	tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT);
6239 	if (tp->t_fb_ptr == NULL) {
6240 		/*
6241 		 * We need to allocate memory but cant. The INP and INP_INFO
6242 		 * locks and they are recusive (happens during setup. So a
6243 		 * scheme to drop the locks fails :(
6244 		 *
6245 		 */
6246 		return (ENOMEM);
6247 	}
6248 	memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack));
6249 
6250 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6251 	TAILQ_INIT(&rack->r_ctl.rc_map);
6252 	TAILQ_INIT(&rack->r_ctl.rc_free);
6253 	TAILQ_INIT(&rack->r_ctl.rc_tmap);
6254 	rack->rc_tp = tp;
6255 	if (tp->t_inpcb) {
6256 		rack->rc_inp = tp->t_inpcb;
6257 	}
6258 	/* Probably not needed but lets be sure */
6259 	rack_clear_rate_sample(rack);
6260 	rack->r_cpu = 0;
6261 	rack->r_ctl.rc_reorder_fade = rack_reorder_fade;
6262 	rack->rc_allow_data_af_clo = rack_ignore_data_after_close;
6263 	rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh;
6264 	rack->rc_pace_reduce = rack_slot_reduction;
6265 	if (V_tcp_delack_enabled)
6266 		tp->t_delayed_ack = 1;
6267 	else
6268 		tp->t_delayed_ack = 0;
6269 	rack->rc_pace_max_segs = rack_hptsi_segments;
6270 	rack->r_ctl.rc_early_recovery_segs = rack_early_recovery_max_seg;
6271 	rack->r_ctl.rc_reorder_shift = rack_reorder_thresh;
6272 	rack->r_ctl.rc_pkt_delay = rack_pkt_delay;
6273 	rack->r_ctl.rc_prop_reduce = rack_use_proportional_reduce;
6274 	rack->r_idle_reduce_largest  = rack_reduce_largest_on_idle;
6275 	rack->r_enforce_min_pace = rack_min_pace_time;
6276 	rack->r_min_pace_seg_thresh = rack_min_pace_time_seg_req;
6277 	rack->r_ctl.rc_prop_rate = rack_proportional_rate;
6278 	rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp;
6279 	rack->r_ctl.rc_early_recovery = rack_early_recovery;
6280 	rack->rc_always_pace = rack_pace_every_seg;
6281 	rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method;
6282 	rack->rack_tlp_threshold_use = rack_tlp_threshold_use;
6283 	rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr;
6284 	rack->r_ctl.rc_min_to = rack_min_to;
6285 	rack->r_ctl.rc_prr_inc_var = rack_inc_var;
6286 	rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0);
6287 	if (tp->snd_una != tp->snd_max) {
6288 		/* Create a send map for the current outstanding data */
6289 		struct rack_sendmap *rsm;
6290 
6291 		rsm = rack_alloc(rack);
6292 		if (rsm == NULL) {
6293 			uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
6294 			tp->t_fb_ptr = NULL;
6295 			return (ENOMEM);
6296 		}
6297 		rsm->r_flags = RACK_OVERMAX;
6298 		rsm->r_tim_lastsent[0] = tcp_ts_getticks();
6299 		rsm->r_rtr_cnt = 1;
6300 		rsm->r_rtr_bytes = 0;
6301 		rsm->r_start = tp->snd_una;
6302 		rsm->r_end = tp->snd_max;
6303 		rsm->r_sndcnt = 0;
6304 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_map, rsm, r_next);
6305 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext);
6306 		rsm->r_in_tmap = 1;
6307 	}
6308 	return (0);
6309 }
6310 
6311 static int
6312 rack_handoff_ok(struct tcpcb *tp)
6313 {
6314 	if ((tp->t_state == TCPS_CLOSED) ||
6315 	    (tp->t_state == TCPS_LISTEN)) {
6316 		/* Sure no problem though it may not stick */
6317 		return (0);
6318 	}
6319 	if ((tp->t_state == TCPS_SYN_SENT) ||
6320 	    (tp->t_state == TCPS_SYN_RECEIVED)) {
6321 		/*
6322 		 * We really don't know you have to get to ESTAB or beyond
6323 		 * to tell.
6324 		 */
6325 		return (EAGAIN);
6326 	}
6327 	if (tp->t_flags & TF_SACK_PERMIT) {
6328 		return (0);
6329 	}
6330 	/*
6331 	 * If we reach here we don't do SACK on this connection so we can
6332 	 * never do rack.
6333 	 */
6334 	return (EINVAL);
6335 }
6336 
6337 static void
6338 rack_fini(struct tcpcb *tp, int32_t tcb_is_purged)
6339 {
6340 	if (tp->t_fb_ptr) {
6341 		struct tcp_rack *rack;
6342 		struct rack_sendmap *rsm;
6343 
6344 		rack = (struct tcp_rack *)tp->t_fb_ptr;
6345 #ifdef TCP_BLACKBOX
6346 		tcp_log_flowend(tp);
6347 #endif
6348 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_map);
6349 		while (rsm) {
6350 			TAILQ_REMOVE(&rack->r_ctl.rc_map, rsm, r_next);
6351 			uma_zfree(rack_zone, rsm);
6352 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_map);
6353 		}
6354 		rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
6355 		while (rsm) {
6356 			TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_next);
6357 			uma_zfree(rack_zone, rsm);
6358 			rsm = TAILQ_FIRST(&rack->r_ctl.rc_free);
6359 		}
6360 		rack->rc_free_cnt = 0;
6361 		uma_zfree(rack_pcb_zone, tp->t_fb_ptr);
6362 		tp->t_fb_ptr = NULL;
6363 	}
6364 }
6365 
6366 static void
6367 rack_set_state(struct tcpcb *tp, struct tcp_rack *rack)
6368 {
6369 	switch (tp->t_state) {
6370 	case TCPS_SYN_SENT:
6371 		rack->r_state = TCPS_SYN_SENT;
6372 		rack->r_substate = rack_do_syn_sent;
6373 		break;
6374 	case TCPS_SYN_RECEIVED:
6375 		rack->r_state = TCPS_SYN_RECEIVED;
6376 		rack->r_substate = rack_do_syn_recv;
6377 		break;
6378 	case TCPS_ESTABLISHED:
6379 		rack->r_state = TCPS_ESTABLISHED;
6380 		rack->r_substate = rack_do_established;
6381 		break;
6382 	case TCPS_CLOSE_WAIT:
6383 		rack->r_state = TCPS_CLOSE_WAIT;
6384 		rack->r_substate = rack_do_close_wait;
6385 		break;
6386 	case TCPS_FIN_WAIT_1:
6387 		rack->r_state = TCPS_FIN_WAIT_1;
6388 		rack->r_substate = rack_do_fin_wait_1;
6389 		break;
6390 	case TCPS_CLOSING:
6391 		rack->r_state = TCPS_CLOSING;
6392 		rack->r_substate = rack_do_closing;
6393 		break;
6394 	case TCPS_LAST_ACK:
6395 		rack->r_state = TCPS_LAST_ACK;
6396 		rack->r_substate = rack_do_lastack;
6397 		break;
6398 	case TCPS_FIN_WAIT_2:
6399 		rack->r_state = TCPS_FIN_WAIT_2;
6400 		rack->r_substate = rack_do_fin_wait_2;
6401 		break;
6402 	case TCPS_LISTEN:
6403 	case TCPS_CLOSED:
6404 	case TCPS_TIME_WAIT:
6405 	default:
6406 #ifdef INVARIANTS
6407 		panic("tcp tp:%p state:%d sees impossible state?", tp, tp->t_state);
6408 #endif
6409 		break;
6410 	};
6411 }
6412 
6413 
6414 static void
6415 rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb)
6416 {
6417 	/*
6418 	 * We received an ack, and then did not
6419 	 * call send or were bounced out due to the
6420 	 * hpts was running. Now a timer is up as well, is
6421 	 * it the right timer?
6422 	 */
6423 	struct rack_sendmap *rsm;
6424 	int tmr_up;
6425 
6426 	tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK;
6427 	if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT))
6428 		return;
6429 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6430 	if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) &&
6431 	    (tmr_up == PACE_TMR_RXT)) {
6432 		/* Should be an RXT */
6433 		return;
6434 	}
6435 	if (rsm == NULL) {
6436 		/* Nothing outstanding? */
6437 		if (tp->t_flags & TF_DELACK) {
6438 			if (tmr_up == PACE_TMR_DELACK)
6439 				/* We are supposed to have delayed ack up and we do */
6440 				return;
6441 		} else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) {
6442 			/*
6443 			 * if we hit enobufs then we would expect the possiblity
6444 			 * of nothing outstanding and the RXT up (and the hptsi timer).
6445 			 */
6446 			return;
6447 		} else if (((tcp_always_keepalive ||
6448 			     rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
6449 			    (tp->t_state <= TCPS_CLOSING)) &&
6450 			   (tmr_up == PACE_TMR_KEEP) &&
6451 			   (tp->snd_max == tp->snd_una)) {
6452 			/* We should have keep alive up and we do */
6453 			return;
6454 		}
6455 	}
6456 	if (rsm && (rsm->r_flags & RACK_SACK_PASSED)) {
6457 		if ((tp->t_flags & TF_SENTFIN) &&
6458 		    ((tp->snd_max - tp->snd_una) == 1) &&
6459 		    (rsm->r_flags & RACK_HAS_FIN)) {
6460 			/* needs to be a RXT */
6461 			if (tmr_up == PACE_TMR_RXT)
6462 				return;
6463 		} else if (tmr_up == PACE_TMR_RACK)
6464 			return;
6465 	} else if (SEQ_GT(tp->snd_max,tp->snd_una) &&
6466 		   ((tmr_up == PACE_TMR_TLP) ||
6467 		    (tmr_up == PACE_TMR_RXT))) {
6468 		/*
6469 		 * Either a TLP or RXT is fine if no sack-passed
6470 		 * is in place and data is outstanding.
6471 		 */
6472 		return;
6473 	} else if (tmr_up == PACE_TMR_DELACK) {
6474 		/*
6475 		 * If the delayed ack was going to go off
6476 		 * before the rtx/tlp/rack timer were going to
6477 		 * expire, then that would be the timer in control.
6478 		 * Note we don't check the time here trusting the
6479 		 * code is correct.
6480 		 */
6481 		return;
6482 	}
6483 	/*
6484 	 * Ok the timer originally started is not what we want now.
6485 	 * We will force the hpts to be stopped if any, and restart
6486 	 * with the slot set to what was in the saved slot.
6487 	 */
6488 	rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__);
6489 	rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0);
6490 }
6491 
6492 static void
6493 rack_hpts_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
6494     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos,
6495     int32_t nxt_pkt, struct timeval *tv)
6496 {
6497 	int32_t thflags, retval, did_out = 0;
6498 	int32_t way_out = 0;
6499 	uint32_t cts;
6500 	uint32_t tiwin;
6501 	struct tcpopt to;
6502 	struct tcp_rack *rack;
6503 	struct rack_sendmap *rsm;
6504 	int32_t prev_state = 0;
6505 
6506 	cts = tcp_tv_to_mssectick(tv);
6507 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6508 
6509 	kern_prefetch(rack, &prev_state);
6510 	prev_state = 0;
6511 	thflags = th->th_flags;
6512 	/*
6513 	 * If this is either a state-changing packet or current state isn't
6514 	 * established, we require a read lock on tcbinfo.  Otherwise, we
6515 	 * allow the tcbinfo to be in either locked or unlocked, as the
6516 	 * caller may have unnecessarily acquired a lock due to a race.
6517 	 */
6518 	if ((thflags & (TH_SYN | TH_FIN | TH_RST)) != 0 ||
6519 	    tp->t_state != TCPS_ESTABLISHED) {
6520 		INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
6521 	}
6522 	INP_WLOCK_ASSERT(tp->t_inpcb);
6523 	KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN",
6524 	    __func__));
6525 	KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT",
6526 	    __func__));
6527 	{
6528 		union tcp_log_stackspecific log;
6529 
6530 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
6531 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
6532 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
6533 		TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0,
6534 		    tlen, &log, true);
6535 	}
6536 	/*
6537 	 * Segment received on connection. Reset idle time and keep-alive
6538 	 * timer. XXX: This should be done after segment validation to
6539 	 * ignore broken/spoofed segs.
6540 	 */
6541 	if  (tp->t_idle_reduce && (tp->snd_max == tp->snd_una)) {
6542 #ifdef NETFLIX_CWV
6543 		if ((tp->cwv_enabled) &&
6544 		    ((tp->cwv_cwnd_valid == 0) &&
6545 		     TCPS_HAVEESTABLISHED(tp->t_state) &&
6546 		     (tp->snd_cwnd > tp->snd_cwv.init_cwnd))) {
6547 			tcp_newcwv_nvp_closedown(tp);
6548 		} else
6549 #endif
6550 		       if ((ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
6551 			counter_u64_add(rack_input_idle_reduces, 1);
6552 			rack_cc_after_idle(tp,
6553 			    (rack->r_idle_reduce_largest ? 1 :0));
6554 		}
6555 	}
6556 	rack->r_ctl.rc_rcvtime = cts;
6557 	tp->t_rcvtime = ticks;
6558 
6559 #ifdef NETFLIX_CWV
6560 	if (tp->cwv_enabled) {
6561 		if ((tp->cwv_cwnd_valid == 0) &&
6562 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
6563 		    (tp->snd_cwnd > tp->snd_cwv.init_cwnd))
6564 			tcp_newcwv_nvp_closedown(tp);
6565 	}
6566 #endif
6567 	/*
6568 	 * Unscale the window into a 32-bit value. For the SYN_SENT state
6569 	 * the scale is zero.
6570 	 */
6571 	tiwin = th->th_win << tp->snd_scale;
6572 #ifdef NETFLIX_STATS
6573 	stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin);
6574 #endif
6575 	/*
6576 	 * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move
6577 	 * this to occur after we've validated the segment.
6578 	 */
6579 	if (tp->t_flags & TF_ECN_PERMIT) {
6580 		if (thflags & TH_CWR)
6581 			tp->t_flags &= ~TF_ECN_SND_ECE;
6582 		switch (iptos & IPTOS_ECN_MASK) {
6583 		case IPTOS_ECN_CE:
6584 			tp->t_flags |= TF_ECN_SND_ECE;
6585 			TCPSTAT_INC(tcps_ecn_ce);
6586 			break;
6587 		case IPTOS_ECN_ECT0:
6588 			TCPSTAT_INC(tcps_ecn_ect0);
6589 			break;
6590 		case IPTOS_ECN_ECT1:
6591 			TCPSTAT_INC(tcps_ecn_ect1);
6592 			break;
6593 		}
6594 		/* Congestion experienced. */
6595 		if (thflags & TH_ECE) {
6596 			rack_cong_signal(tp, th, CC_ECN);
6597 		}
6598 	}
6599 	/*
6600 	 * Parse options on any incoming segment.
6601 	 */
6602 	tcp_dooptions(&to, (u_char *)(th + 1),
6603 	    (th->th_off << 2) - sizeof(struct tcphdr),
6604 	    (thflags & TH_SYN) ? TO_SYN : 0);
6605 
6606 	/*
6607 	 * If echoed timestamp is later than the current time, fall back to
6608 	 * non RFC1323 RTT calculation.  Normalize timestamp if syncookies
6609 	 * were used when this connection was established.
6610 	 */
6611 	if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) {
6612 		to.to_tsecr -= tp->ts_offset;
6613 		if (TSTMP_GT(to.to_tsecr, cts))
6614 			to.to_tsecr = 0;
6615 	}
6616 	/*
6617 	 * If its the first time in we need to take care of options and
6618 	 * verify we can do SACK for rack!
6619 	 */
6620 	if (rack->r_state == 0) {
6621 		/* Should be init'd by rack_init() */
6622 		KASSERT(rack->rc_inp != NULL,
6623 		    ("%s: rack->rc_inp unexpectedly NULL", __func__));
6624 		if (rack->rc_inp == NULL) {
6625 			rack->rc_inp = tp->t_inpcb;
6626 		}
6627 
6628 		/*
6629 		 * Process options only when we get SYN/ACK back. The SYN
6630 		 * case for incoming connections is handled in tcp_syncache.
6631 		 * According to RFC1323 the window field in a SYN (i.e., a
6632 		 * <SYN> or <SYN,ACK>) segment itself is never scaled. XXX
6633 		 * this is traditional behavior, may need to be cleaned up.
6634 		 */
6635 		rack->r_cpu = inp_to_cpuid(tp->t_inpcb);
6636 		if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
6637 			if ((to.to_flags & TOF_SCALE) &&
6638 			    (tp->t_flags & TF_REQ_SCALE)) {
6639 				tp->t_flags |= TF_RCVD_SCALE;
6640 				tp->snd_scale = to.to_wscale;
6641 			}
6642 			/*
6643 			 * Initial send window.  It will be updated with the
6644 			 * next incoming segment to the scaled value.
6645 			 */
6646 			tp->snd_wnd = th->th_win;
6647 			if (to.to_flags & TOF_TS) {
6648 				tp->t_flags |= TF_RCVD_TSTMP;
6649 				tp->ts_recent = to.to_tsval;
6650 				tp->ts_recent_age = cts;
6651 			}
6652 			if (to.to_flags & TOF_MSS)
6653 				tcp_mss(tp, to.to_mss);
6654 			if ((tp->t_flags & TF_SACK_PERMIT) &&
6655 			    (to.to_flags & TOF_SACKPERM) == 0)
6656 				tp->t_flags &= ~TF_SACK_PERMIT;
6657 			if (IS_FASTOPEN(tp->t_flags)) {
6658 				if (to.to_flags & TOF_FASTOPEN) {
6659 					uint16_t mss;
6660 
6661 					if (to.to_flags & TOF_MSS)
6662 						mss = to.to_mss;
6663 					else
6664 						if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
6665 							mss = TCP6_MSS;
6666 						else
6667 							mss = TCP_MSS;
6668 					tcp_fastopen_update_cache(tp, mss,
6669 					    to.to_tfo_len, to.to_tfo_cookie);
6670 				} else
6671 					tcp_fastopen_disable_path(tp);
6672 			}
6673 		}
6674 		/*
6675 		 * At this point we are at the initial call. Here we decide
6676 		 * if we are doing RACK or not. We do this by seeing if
6677 		 * TF_SACK_PERMIT is set, if not rack is *not* possible and
6678 		 * we switch to the default code.
6679 		 */
6680 		if ((tp->t_flags & TF_SACK_PERMIT) == 0) {
6681 			tcp_switch_back_to_default(tp);
6682 			(*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen,
6683 			    tlen, iptos);
6684 			return;
6685 		}
6686 		/* Set the flag */
6687 		rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
6688 		tcp_set_hpts(tp->t_inpcb);
6689 		rack_stop_all_timers(tp);
6690 		sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack);
6691 	}
6692 	/*
6693 	 * This is the one exception case where we set the rack state
6694 	 * always. All other times (timers etc) we must have a rack-state
6695 	 * set (so we assure we have done the checks above for SACK).
6696 	 */
6697 	if (rack->r_state != tp->t_state)
6698 		rack_set_state(tp, rack);
6699 	if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&rack->r_ctl.rc_map)) != NULL)
6700 		kern_prefetch(rsm, &prev_state);
6701 	prev_state = rack->r_state;
6702 	rack->r_ctl.rc_tlp_send_cnt = 0;
6703 	rack_clear_rate_sample(rack);
6704 	retval = (*rack->r_substate) (m, th, so,
6705 	    tp, &to, drop_hdrlen,
6706 	    tlen, tiwin, thflags, nxt_pkt);
6707 #ifdef INVARIANTS
6708 	if ((retval == 0) &&
6709 	    (tp->t_inpcb == NULL)) {
6710 		panic("retval:%d tp:%p t_inpcb:NULL state:%d",
6711 		    retval, tp, prev_state);
6712 	}
6713 #endif
6714 	if (retval == 0) {
6715 		/*
6716 		 * If retval is 1 the tcb is unlocked and most likely the tp
6717 		 * is gone.
6718 		 */
6719 		INP_WLOCK_ASSERT(tp->t_inpcb);
6720 		tcp_rack_xmit_timer_commit(rack, tp);
6721 		if (((tp->snd_max - tp->snd_una) > tp->snd_wnd) &&
6722 		    (rack->rc_in_persist == 0)){
6723 			/*
6724 			 * The peer shrunk its window on us to the point
6725 			 * where we have sent too much. The only thing
6726 			 * we can do here is stop any timers and
6727 			 * enter persist. We most likely lost the last
6728 			 * bytes we sent but oh well, we will have to
6729 			 * retransmit them after the peer is caught up.
6730 			 */
6731 			if (rack->rc_inp->inp_in_hpts)
6732 				tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
6733 			rack_timer_cancel(tp, rack, cts, __LINE__);
6734 			rack_enter_persist(tp, rack, cts);
6735 			rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0);
6736 			way_out = 3;
6737 			goto done_with_input;
6738 		}
6739 		if (nxt_pkt == 0) {
6740 			if (rack->r_wanted_output != 0) {
6741 				did_out = 1;
6742 				(void)tp->t_fb->tfb_tcp_output(tp);
6743 			}
6744 			rack_start_hpts_timer(rack, tp, cts, __LINE__, 0, 0, 0);
6745 		}
6746 		if (((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) &&
6747 		    (SEQ_GT(tp->snd_max, tp->snd_una) ||
6748 		     (tp->t_flags & TF_DELACK) ||
6749 		     ((tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) &&
6750 		      (tp->t_state <= TCPS_CLOSING)))) {
6751 			/* We could not send (probably in the hpts but stopped the timer earlier)? */
6752 			if ((tp->snd_max == tp->snd_una) &&
6753 			    ((tp->t_flags & TF_DELACK) == 0) &&
6754 			    (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) {
6755 				/* keep alive not needed if we are hptsi output yet */
6756 				;
6757 			} else {
6758 				if (rack->rc_inp->inp_in_hpts)
6759 					tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT);
6760 				rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), __LINE__, 0, 0, 0);
6761 			}
6762 			way_out = 1;
6763 		} else {
6764 			/* Do we have the correct timer running? */
6765 			rack_timer_audit(tp, rack, &so->so_snd);
6766 			way_out = 2;
6767 		}
6768 	done_with_input:
6769 		rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out);
6770 		if (did_out)
6771 			rack->r_wanted_output = 0;
6772 #ifdef INVARIANTS
6773 		if (tp->t_inpcb == NULL) {
6774 			panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d",
6775 			      did_out,
6776 			      retval, tp, prev_state);
6777 		}
6778 #endif
6779 		INP_WUNLOCK(tp->t_inpcb);
6780 	}
6781 }
6782 
6783 void
6784 rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
6785     struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos)
6786 {
6787 	struct timeval tv;
6788 #ifdef RSS
6789 	struct tcp_function_block *tfb;
6790 	struct tcp_rack *rack;
6791 	struct epoch_tracker et;
6792 
6793 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6794 	if (rack->r_state == 0) {
6795 		/*
6796 		 * Initial input (ACK to SYN-ACK etc)lets go ahead and get
6797 		 * it processed
6798 		 */
6799 		INP_INFO_RLOCK_ET(&V_tcbinfo, et);
6800 		tcp_get_usecs(&tv);
6801 		rack_hpts_do_segment(m, th, so, tp, drop_hdrlen,
6802 		    tlen, iptos, 0, &tv);
6803 		INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
6804 		return;
6805 	}
6806 	tcp_queue_to_input(tp, m, th, tlen, drop_hdrlen, iptos);
6807 	INP_WUNLOCK(tp->t_inpcb);
6808 #else
6809 	tcp_get_usecs(&tv);
6810 	rack_hpts_do_segment(m, th, so, tp, drop_hdrlen,
6811 	    tlen, iptos, 0, &tv);
6812 #endif
6813 }
6814 
6815 struct rack_sendmap *
6816 tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused)
6817 {
6818 	struct rack_sendmap *rsm = NULL;
6819 	int32_t idx;
6820 	uint32_t srtt_cur, srtt = 0, thresh = 0, ts_low = 0;
6821 
6822 	/* Return the next guy to be re-transmitted */
6823 	if (TAILQ_EMPTY(&rack->r_ctl.rc_map)) {
6824 		return (NULL);
6825 	}
6826 	if (tp->t_flags & TF_SENTFIN) {
6827 		/* retran the end FIN? */
6828 		return (NULL);
6829 	}
6830 	/* ok lets look at this one */
6831 	rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap);
6832 	if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) {
6833 		goto check_it;
6834 	}
6835 	rsm = rack_find_lowest_rsm(rack);
6836 	if (rsm == NULL) {
6837 		return (NULL);
6838 	}
6839 check_it:
6840 	srtt_cur = tp->t_srtt >> TCP_RTT_SHIFT;
6841 	srtt = TICKS_2_MSEC(srtt_cur);
6842 	if (rack->rc_rack_rtt && (srtt > rack->rc_rack_rtt))
6843 		srtt = rack->rc_rack_rtt;
6844 	if (rsm->r_flags & RACK_ACKED) {
6845 		return (NULL);
6846 	}
6847 	if ((rsm->r_flags & RACK_SACK_PASSED) == 0) {
6848 		/* Its not yet ready */
6849 		return (NULL);
6850 	}
6851 	idx = rsm->r_rtr_cnt - 1;
6852 	ts_low = rsm->r_tim_lastsent[idx];
6853 	thresh = rack_calc_thresh_rack(rack, srtt, tsused);
6854 	if (tsused <= ts_low) {
6855 		return (NULL);
6856 	}
6857 	if ((tsused - ts_low) >= thresh) {
6858 		return (rsm);
6859 	}
6860 	return (NULL);
6861 }
6862 
6863 static int
6864 rack_output(struct tcpcb *tp)
6865 {
6866 	struct socket *so;
6867 	uint32_t recwin, sendwin;
6868 	uint32_t sb_offset;
6869 	int32_t len, flags, error = 0;
6870 	struct mbuf *m;
6871 	struct mbuf *mb;
6872 	uint32_t if_hw_tsomaxsegcount = 0;
6873 	uint32_t if_hw_tsomaxsegsize;
6874 	long tot_len_this_send = 0;
6875 	struct ip *ip = NULL;
6876 #ifdef TCPDEBUG
6877 	struct ipovly *ipov = NULL;
6878 #endif
6879 	struct udphdr *udp = NULL;
6880 	struct tcp_rack *rack;
6881 	struct tcphdr *th;
6882 	uint8_t pass = 0;
6883 	uint8_t wanted_cookie = 0;
6884 	u_char opt[TCP_MAXOLEN];
6885 	unsigned ipoptlen, optlen, hdrlen, ulen=0;
6886 	uint32_t rack_seq;
6887 
6888 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
6889 	unsigned ipsec_optlen = 0;
6890 
6891 #endif
6892 	int32_t idle, sendalot;
6893 	int32_t sub_from_prr = 0;
6894 	volatile int32_t sack_rxmit;
6895 	struct rack_sendmap *rsm = NULL;
6896 	int32_t tso, mtu, would_have_fin = 0;
6897 	struct tcpopt to;
6898 	int32_t slot = 0;
6899 	uint32_t cts;
6900 	uint8_t hpts_calling, doing_tlp = 0;
6901 	int32_t do_a_prefetch;
6902 	int32_t prefetch_rsm = 0;
6903 	int32_t prefetch_so_done = 0;
6904 	struct tcp_log_buffer *lgb = NULL;
6905 	struct inpcb *inp;
6906 	struct sockbuf *sb;
6907 #ifdef INET6
6908 	struct ip6_hdr *ip6 = NULL;
6909 	int32_t isipv6;
6910 #endif
6911 	/* setup and take the cache hits here */
6912 	rack = (struct tcp_rack *)tp->t_fb_ptr;
6913 	inp = rack->rc_inp;
6914 	so = inp->inp_socket;
6915 	sb = &so->so_snd;
6916 	kern_prefetch(sb, &do_a_prefetch);
6917 	do_a_prefetch = 1;
6918 
6919 	INP_WLOCK_ASSERT(inp);
6920 #ifdef TCP_OFFLOAD
6921 	if (tp->t_flags & TF_TOE)
6922 		return (tcp_offload_output(tp));
6923 #endif
6924 
6925 	/*
6926 	 * For TFO connections in SYN_RECEIVED, only allow the initial
6927 	 * SYN|ACK and those sent by the retransmit timer.
6928 	 */
6929 	if (IS_FASTOPEN(tp->t_flags) &&
6930 	    (tp->t_state == TCPS_SYN_RECEIVED) &&
6931 	    SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
6932 	    (rack->r_ctl.rc_resend == NULL))         /* not a retransmit */
6933 		return (0);
6934 #ifdef INET6
6935 	if (rack->r_state) {
6936 		/* Use the cache line loaded if possible */
6937 		isipv6 = rack->r_is_v6;
6938 	} else {
6939 		isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
6940 	}
6941 #endif
6942 	cts = tcp_ts_getticks();
6943 	if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) &&
6944 	    inp->inp_in_hpts) {
6945 		/*
6946 		 * We are on the hpts for some timer but not hptsi output.
6947 		 * Remove from the hpts unconditionally.
6948 		 */
6949 		rack_timer_cancel(tp, rack, cts, __LINE__);
6950 	}
6951 	/* Mark that we have called rack_output(). */
6952 	if ((rack->r_timer_override) ||
6953 	    (tp->t_flags & TF_FORCEDATA) ||
6954 	    (tp->t_state < TCPS_ESTABLISHED)) {
6955 		if (tp->t_inpcb->inp_in_hpts)
6956 			tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT);
6957 	} else if (tp->t_inpcb->inp_in_hpts) {
6958 		/*
6959 		 * On the hpts you can't pass even if ACKNOW is on, we will
6960 		 * when the hpts fires.
6961 		 */
6962 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1);
6963 		return (0);
6964 	}
6965 	hpts_calling = inp->inp_hpts_calls;
6966 	inp->inp_hpts_calls = 0;
6967 	if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) {
6968 		if (rack_process_timers(tp, rack, cts, hpts_calling)) {
6969 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1);
6970 			return (0);
6971 		}
6972 	}
6973 	rack->r_wanted_output = 0;
6974 	rack->r_timer_override = 0;
6975 	/*
6976 	 * Determine length of data that should be transmitted, and flags
6977 	 * that will be used. If there is some data or critical controls
6978 	 * (SYN, RST) to send, then transmit; otherwise, investigate
6979 	 * further.
6980 	 */
6981 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
6982 #ifdef NETFLIX_CWV
6983 	if (tp->cwv_enabled) {
6984 		if ((tp->cwv_cwnd_valid == 0) &&
6985 		    TCPS_HAVEESTABLISHED(tp->t_state) &&
6986 		    (tp->snd_cwnd > tp->snd_cwv.init_cwnd))
6987 			tcp_newcwv_nvp_closedown(tp);
6988 	} else
6989 #endif
6990 	if (tp->t_idle_reduce) {
6991 		if (idle && ((ticks - tp->t_rcvtime) >= tp->t_rxtcur))
6992 			rack_cc_after_idle(tp,
6993 		            (rack->r_idle_reduce_largest ? 1 :0));
6994 	}
6995 	tp->t_flags &= ~TF_LASTIDLE;
6996 	if (idle) {
6997 		if (tp->t_flags & TF_MORETOCOME) {
6998 			tp->t_flags |= TF_LASTIDLE;
6999 			idle = 0;
7000 		}
7001 	}
7002 again:
7003 	/*
7004 	 * If we've recently taken a timeout, snd_max will be greater than
7005 	 * snd_nxt.  There may be SACK information that allows us to avoid
7006 	 * resending already delivered data.  Adjust snd_nxt accordingly.
7007 	 */
7008 	sendalot = 0;
7009 	cts = tcp_ts_getticks();
7010 	tso = 0;
7011 	mtu = 0;
7012 	sb_offset = tp->snd_max - tp->snd_una;
7013 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
7014 
7015 	flags = tcp_outflags[tp->t_state];
7016 	/*
7017 	 * Send any SACK-generated retransmissions.  If we're explicitly
7018 	 * trying to send out new data (when sendalot is 1), bypass this
7019 	 * function. If we retransmit in fast recovery mode, decrement
7020 	 * snd_cwnd, since we're replacing a (future) new transmission with
7021 	 * a retransmission now, and we previously incremented snd_cwnd in
7022 	 * tcp_input().
7023 	 */
7024 	/*
7025 	 * Still in sack recovery , reset rxmit flag to zero.
7026 	 */
7027 	while (rack->rc_free_cnt < rack_free_cache) {
7028 		rsm = rack_alloc(rack);
7029 		if (rsm == NULL) {
7030 			if (inp->inp_hpts_calls)
7031 				/* Retry in a ms */
7032 				slot = 1;
7033 			goto just_return_nolock;
7034 		}
7035 		TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_next);
7036 		rack->rc_free_cnt++;
7037 		rsm = NULL;
7038 	}
7039 	if (inp->inp_hpts_calls)
7040 		inp->inp_hpts_calls = 0;
7041 	sack_rxmit = 0;
7042 	len = 0;
7043 	rsm = NULL;
7044 	if (flags & TH_RST) {
7045 		SOCKBUF_LOCK(sb);
7046 		goto send;
7047 	}
7048 	if (rack->r_ctl.rc_tlpsend) {
7049 		/* Tail loss probe */
7050 		long cwin;
7051 		long tlen;
7052 
7053 		doing_tlp = 1;
7054 		rsm = rack->r_ctl.rc_tlpsend;
7055 		rack->r_ctl.rc_tlpsend = NULL;
7056 		sack_rxmit = 1;
7057 		tlen = rsm->r_end - rsm->r_start;
7058 		if (tlen > tp->t_maxseg)
7059 			tlen = tp->t_maxseg;
7060 #ifdef INVARIANTS
7061 		if (SEQ_GT(tp->snd_una, rsm->r_start)) {
7062 			panic("tp:%p rack:%p snd_una:%u rsm:%p r_start:%u",
7063 			    tp, rack, tp->snd_una, rsm, rsm->r_start);
7064 		}
7065 #endif
7066 		sb_offset = rsm->r_start - tp->snd_una;
7067 		cwin = min(tp->snd_wnd, tlen);
7068 		len = cwin;
7069 	} else if (rack->r_ctl.rc_resend) {
7070 		/* Retransmit timer */
7071 		rsm = rack->r_ctl.rc_resend;
7072 		rack->r_ctl.rc_resend = NULL;
7073 		len = rsm->r_end - rsm->r_start;
7074 		sack_rxmit = 1;
7075 		sendalot = 0;
7076 		sb_offset = rsm->r_start - tp->snd_una;
7077 		if (len >= tp->t_maxseg) {
7078 			len = tp->t_maxseg;
7079 		}
7080 		KASSERT(sb_offset >= 0, ("%s: sack block to the left of una : %d",
7081 		    __func__, sb_offset));
7082 	} else if ((rack->rc_in_persist == 0) &&
7083 	    ((rsm = tcp_rack_output(tp, rack, cts)) != NULL)) {
7084 		long tlen;
7085 
7086 		if ((!IN_RECOVERY(tp->t_flags)) &&
7087 		    ((tp->t_flags & (TF_WASFRECOVERY | TF_WASCRECOVERY)) == 0)) {
7088 			/* Enter recovery if not induced by a time-out */
7089 			rack->r_ctl.rc_rsm_start = rsm->r_start;
7090 			rack->r_ctl.rc_cwnd_at = tp->snd_cwnd;
7091 			rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh;
7092 			rack_cong_signal(tp, NULL, CC_NDUPACK);
7093 			/*
7094 			 * When we enter recovery we need to assure we send
7095 			 * one packet.
7096 			 */
7097 			rack->r_ctl.rc_prr_sndcnt = tp->t_maxseg;
7098 		}
7099 #ifdef INVARIANTS
7100 		if (SEQ_LT(rsm->r_start, tp->snd_una)) {
7101 			panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n",
7102 			    tp, rack, rsm, rsm->r_start, tp->snd_una);
7103 		}
7104 #endif
7105 		tlen = rsm->r_end - rsm->r_start;
7106 		sb_offset = rsm->r_start - tp->snd_una;
7107 		if (tlen > rack->r_ctl.rc_prr_sndcnt) {
7108 			len = rack->r_ctl.rc_prr_sndcnt;
7109 		} else {
7110 			len = tlen;
7111 		}
7112 		if (len >= tp->t_maxseg) {
7113 			sendalot = 1;
7114 			len = tp->t_maxseg;
7115 		} else {
7116 			sendalot = 0;
7117 			if ((rack->rc_timer_up == 0) &&
7118 			    (len < tlen)) {
7119 				/*
7120 				 * If its not a timer don't send a partial
7121 				 * segment.
7122 				 */
7123 				len = 0;
7124 				goto just_return_nolock;
7125 			}
7126 		}
7127 		KASSERT(sb_offset >= 0, ("%s: sack block to the left of una : %d",
7128 		    __func__, sb_offset));
7129 		if (len > 0) {
7130 			sub_from_prr = 1;
7131 			sack_rxmit = 1;
7132 			TCPSTAT_INC(tcps_sack_rexmits);
7133 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
7134 			    min(len, tp->t_maxseg));
7135 			counter_u64_add(rack_rtm_prr_retran, 1);
7136 		}
7137 	}
7138 	if (rsm && (rsm->r_flags & RACK_HAS_FIN)) {
7139 		/* we are retransmitting the fin */
7140 		len--;
7141 		if (len) {
7142 			/*
7143 			 * When retransmitting data do *not* include the
7144 			 * FIN. This could happen from a TLP probe.
7145 			 */
7146 			flags &= ~TH_FIN;
7147 		}
7148 	}
7149 #ifdef INVARIANTS
7150 	/* For debugging */
7151 	rack->r_ctl.rc_rsm_at_retran = rsm;
7152 #endif
7153 	/*
7154 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
7155 	 * state flags.
7156 	 */
7157 	if (tp->t_flags & TF_NEEDFIN)
7158 		flags |= TH_FIN;
7159 	if (tp->t_flags & TF_NEEDSYN)
7160 		flags |= TH_SYN;
7161 	if ((sack_rxmit == 0) && (prefetch_rsm == 0)) {
7162 		void *end_rsm;
7163 		end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext);
7164 		if (end_rsm)
7165 			kern_prefetch(end_rsm, &prefetch_rsm);
7166 		prefetch_rsm = 1;
7167 	}
7168 	SOCKBUF_LOCK(sb);
7169 	/*
7170 	 * If in persist timeout with window of 0, send 1 byte. Otherwise,
7171 	 * if window is small but nonzero and time TF_SENTFIN expired, we
7172 	 * will send what we can and go to transmit state.
7173 	 */
7174 	if (tp->t_flags & TF_FORCEDATA) {
7175 		if (sendwin == 0) {
7176 			/*
7177 			 * If we still have some data to send, then clear
7178 			 * the FIN bit.  Usually this would happen below
7179 			 * when it realizes that we aren't sending all the
7180 			 * data.  However, if we have exactly 1 byte of
7181 			 * unsent data, then it won't clear the FIN bit
7182 			 * below, and if we are in persist state, we wind up
7183 			 * sending the packet without recording that we sent
7184 			 * the FIN bit.
7185 			 *
7186 			 * We can't just blindly clear the FIN bit, because
7187 			 * if we don't have any more data to send then the
7188 			 * probe will be the FIN itself.
7189 			 */
7190 			if (sb_offset < sbused(sb))
7191 				flags &= ~TH_FIN;
7192 			sendwin = 1;
7193 		} else {
7194 			if (rack->rc_in_persist)
7195 				rack_exit_persist(tp, rack);
7196 			/*
7197 			 * If we are dropping persist mode then we need to
7198 			 * correct snd_nxt/snd_max and off.
7199 			 */
7200 			tp->snd_nxt = tp->snd_max;
7201 			sb_offset = tp->snd_nxt - tp->snd_una;
7202 		}
7203 	}
7204 	/*
7205 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
7206 	 * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a
7207 	 * negative length.  This can also occur when TCP opens up its
7208 	 * congestion window while receiving additional duplicate acks after
7209 	 * fast-retransmit because TCP will reset snd_nxt to snd_max after
7210 	 * the fast-retransmit.
7211 	 *
7212 	 * In the normal retransmit-FIN-only case, however, snd_nxt will be
7213 	 * set to snd_una, the sb_offset will be 0, and the length may wind
7214 	 * up 0.
7215 	 *
7216 	 * If sack_rxmit is true we are retransmitting from the scoreboard
7217 	 * in which case len is already set.
7218 	 */
7219 	if (sack_rxmit == 0) {
7220 		uint32_t avail;
7221 
7222 		avail = sbavail(sb);
7223 		if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail)
7224 			sb_offset = tp->snd_nxt - tp->snd_una;
7225 		else
7226 			sb_offset = 0;
7227 		if (IN_RECOVERY(tp->t_flags) == 0) {
7228 			if (rack->r_ctl.rc_tlp_new_data) {
7229 				/* TLP is forcing out new data */
7230 				if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) {
7231 					rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset);
7232 				}
7233 				if (rack->r_ctl.rc_tlp_new_data > tp->snd_wnd)
7234 					len = tp->snd_wnd;
7235 				else
7236 					len = rack->r_ctl.rc_tlp_new_data;
7237 				rack->r_ctl.rc_tlp_new_data = 0;
7238 				doing_tlp = 1;
7239 			} else {
7240 				if (sendwin > avail) {
7241 					/* use the available */
7242 					if (avail > sb_offset) {
7243 						len = (int32_t)(avail - sb_offset);
7244 					} else {
7245 						len = 0;
7246 					}
7247 				} else {
7248 					if (sendwin > sb_offset) {
7249 						len = (int32_t)(sendwin - sb_offset);
7250 					} else {
7251 						len = 0;
7252 					}
7253 				}
7254 			}
7255 		} else {
7256 			uint32_t outstanding;
7257 
7258 			/*
7259 			 * We are inside of a SACK recovery episode and are
7260 			 * sending new data, having retransmitted all the
7261 			 * data possible so far in the scoreboard.
7262 			 */
7263 			outstanding = tp->snd_max - tp->snd_una;
7264 			if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd)
7265 				len = 0;
7266 			else if (avail > sb_offset)
7267 				len = avail - sb_offset;
7268 			else
7269 				len = 0;
7270 			if (len > 0) {
7271 				if (len > rack->r_ctl.rc_prr_sndcnt)
7272 					len = rack->r_ctl.rc_prr_sndcnt;
7273 
7274 				if (len > 0) {
7275 					sub_from_prr = 1;
7276 					counter_u64_add(rack_rtm_prr_newdata, 1);
7277 				}
7278 			}
7279 			if (len > tp->t_maxseg) {
7280 				/*
7281 				 * We should never send more than a MSS when
7282 				 * retransmitting or sending new data in prr
7283 				 * mode unless the override flag is on. Most
7284 				 * likely the PRR algorithm is not going to
7285 				 * let us send a lot as well :-)
7286 				 */
7287 				if (rack->r_ctl.rc_prr_sendalot == 0)
7288 					len = tp->t_maxseg;
7289 			} else if (len < tp->t_maxseg) {
7290 				/*
7291 				 * Do we send any? The idea here is if the
7292 				 * send empty's the socket buffer we want to
7293 				 * do it. However if not then lets just wait
7294 				 * for our prr_sndcnt to get bigger.
7295 				 */
7296 				long leftinsb;
7297 
7298 				leftinsb = sbavail(sb) - sb_offset;
7299 				if (leftinsb > len) {
7300 					/* This send does not empty the sb */
7301 					len = 0;
7302 				}
7303 			}
7304 		}
7305 	}
7306 	if (prefetch_so_done == 0) {
7307 		kern_prefetch(so, &prefetch_so_done);
7308 		prefetch_so_done = 1;
7309 	}
7310 	/*
7311 	 * Lop off SYN bit if it has already been sent.  However, if this is
7312 	 * SYN-SENT state and if segment contains data and if we don't know
7313 	 * that foreign host supports TAO, suppress sending segment.
7314 	 */
7315 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) &&
7316 	    ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) {
7317 		if (tp->t_state != TCPS_SYN_RECEIVED)
7318 			flags &= ~TH_SYN;
7319 		/*
7320 		 * When sending additional segments following a TFO SYN|ACK,
7321 		 * do not include the SYN bit.
7322 		 */
7323 		if (IS_FASTOPEN(tp->t_flags) &&
7324 		    (tp->t_state == TCPS_SYN_RECEIVED))
7325 			flags &= ~TH_SYN;
7326 		sb_offset--, len++;
7327 	}
7328 	/*
7329 	 * Be careful not to send data and/or FIN on SYN segments. This
7330 	 * measure is needed to prevent interoperability problems with not
7331 	 * fully conformant TCP implementations.
7332 	 */
7333 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
7334 		len = 0;
7335 		flags &= ~TH_FIN;
7336 	}
7337 	/*
7338 	 * On TFO sockets, ensure no data is sent in the following cases:
7339 	 *
7340 	 *  - When retransmitting SYN|ACK on a passively-created socket
7341 	 *
7342 	 *  - When retransmitting SYN on an actively created socket
7343 	 *
7344 	 *  - When sending a zero-length cookie (cookie request) on an
7345 	 *    actively created socket
7346 	 *
7347 	 *  - When the socket is in the CLOSED state (RST is being sent)
7348 	 */
7349 	if (IS_FASTOPEN(tp->t_flags) &&
7350 	    (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
7351 	     ((tp->t_state == TCPS_SYN_SENT) &&
7352 	      (tp->t_tfo_client_cookie_len == 0)) ||
7353 	     (flags & TH_RST)))
7354 		len = 0;
7355 	/* Without fast-open there should never be data sent on a SYN */
7356 	if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags)))
7357 		len = 0;
7358 	if (len <= 0) {
7359 		/*
7360 		 * If FIN has been sent but not acked, but we haven't been
7361 		 * called to retransmit, len will be < 0.  Otherwise, window
7362 		 * shrank after we sent into it.  If window shrank to 0,
7363 		 * cancel pending retransmit, pull snd_nxt back to (closed)
7364 		 * window, and set the persist timer if it isn't already
7365 		 * going.  If the window didn't close completely, just wait
7366 		 * for an ACK.
7367 		 *
7368 		 * We also do a general check here to ensure that we will
7369 		 * set the persist timer when we have data to send, but a
7370 		 * 0-byte window. This makes sure the persist timer is set
7371 		 * even if the packet hits one of the "goto send" lines
7372 		 * below.
7373 		 */
7374 		len = 0;
7375 		if ((tp->snd_wnd == 0) &&
7376 		    (TCPS_HAVEESTABLISHED(tp->t_state)) &&
7377 		    (sb_offset < (int)sbavail(sb))) {
7378 			tp->snd_nxt = tp->snd_una;
7379 			rack_enter_persist(tp, rack, cts);
7380 		}
7381 	}
7382 	/* len will be >= 0 after this point. */
7383 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
7384 	tcp_sndbuf_autoscale(tp, so, sendwin);
7385 	/*
7386 	 * Decide if we can use TCP Segmentation Offloading (if supported by
7387 	 * hardware).
7388 	 *
7389 	 * TSO may only be used if we are in a pure bulk sending state.  The
7390 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP
7391 	 * options prevent using TSO.  With TSO the TCP header is the same
7392 	 * (except for the sequence number) for all generated packets.  This
7393 	 * makes it impossible to transmit any options which vary per
7394 	 * generated segment or packet.
7395 	 *
7396 	 * IPv4 handling has a clear separation of ip options and ip header
7397 	 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
7398 	 * the right thing below to provide length of just ip options and thus
7399 	 * checking for ipoptlen is enough to decide if ip options are present.
7400 	 */
7401 
7402 #ifdef INET6
7403 	if (isipv6)
7404 		ipoptlen = ip6_optlen(tp->t_inpcb);
7405 	else
7406 #endif
7407 		if (tp->t_inpcb->inp_options)
7408 			ipoptlen = tp->t_inpcb->inp_options->m_len -
7409 			    offsetof(struct ipoption, ipopt_list);
7410 		else
7411 			ipoptlen = 0;
7412 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
7413 	/*
7414 	 * Pre-calculate here as we save another lookup into the darknesses
7415 	 * of IPsec that way and can actually decide if TSO is ok.
7416 	 */
7417 #ifdef INET6
7418 	if (isipv6 && IPSEC_ENABLED(ipv6))
7419 		ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
7420 #ifdef INET
7421 	else
7422 #endif
7423 #endif				/* INET6 */
7424 #ifdef INET
7425 	if (IPSEC_ENABLED(ipv4))
7426 		ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
7427 #endif				/* INET */
7428 #endif
7429 
7430 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
7431 	ipoptlen += ipsec_optlen;
7432 #endif
7433 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
7434 	    (tp->t_port == 0) &&
7435 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
7436 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
7437 	    ipoptlen == 0)
7438 		tso = 1;
7439 	{
7440 		uint32_t outstanding;
7441 
7442 		outstanding = tp->snd_max - tp->snd_una;
7443 		if (tp->t_flags & TF_SENTFIN) {
7444 			/*
7445 			 * If we sent a fin, snd_max is 1 higher than
7446 			 * snd_una
7447 			 */
7448 			outstanding--;
7449 		}
7450 		if (outstanding > 0) {
7451 			/*
7452 			 * This is sub-optimal. We only send a stand alone
7453 			 * FIN on its own segment.
7454 			 */
7455 			if (flags & TH_FIN) {
7456 				flags &= ~TH_FIN;
7457 				would_have_fin = 1;
7458 			}
7459 		} else if (sack_rxmit) {
7460 			if ((rsm->r_flags & RACK_HAS_FIN) == 0)
7461 				flags &= ~TH_FIN;
7462 		} else {
7463 			if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
7464 			    sbused(sb)))
7465 				flags &= ~TH_FIN;
7466 		}
7467 	}
7468 	recwin = sbspace(&so->so_rcv);
7469 
7470 	/*
7471 	 * Sender silly window avoidance.   We transmit under the following
7472 	 * conditions when len is non-zero:
7473 	 *
7474 	 * - We have a full segment (or more with TSO) - This is the last
7475 	 * buffer in a write()/send() and we are either idle or running
7476 	 * NODELAY - we've timed out (e.g. persist timer) - we have more
7477 	 * then 1/2 the maximum send window's worth of data (receiver may be
7478 	 * limited the window size) - we need to retransmit
7479 	 */
7480 	if (len) {
7481 		if (len >= tp->t_maxseg) {
7482 			pass = 1;
7483 			goto send;
7484 		}
7485 		/*
7486 		 * NOTE! on localhost connections an 'ack' from the remote
7487 		 * end may occur synchronously with the output and cause us
7488 		 * to flush a buffer queued with moretocome.  XXX
7489 		 *
7490 		 */
7491 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
7492 		    (idle || (tp->t_flags & TF_NODELAY)) &&
7493 		    ((uint32_t)len + (uint32_t)sb_offset >= sbavail(&so->so_snd)) &&
7494 		    (tp->t_flags & TF_NOPUSH) == 0) {
7495 			pass = 2;
7496 			goto send;
7497 		}
7498 		if (tp->t_flags & TF_FORCEDATA) {	/* typ. timeout case */
7499 			pass = 3;
7500 			goto send;
7501 		}
7502 		if ((tp->snd_una == tp->snd_max) && len) {	/* Nothing outstanding */
7503 			goto send;
7504 		}
7505 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) {
7506 			pass = 4;
7507 			goto send;
7508 		}
7509 		if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {	/* retransmit case */
7510 			pass = 5;
7511 			goto send;
7512 		}
7513 		if (sack_rxmit) {
7514 			pass = 6;
7515 			goto send;
7516 		}
7517 	}
7518 	/*
7519 	 * Sending of standalone window updates.
7520 	 *
7521 	 * Window updates are important when we close our window due to a
7522 	 * full socket buffer and are opening it again after the application
7523 	 * reads data from it.  Once the window has opened again and the
7524 	 * remote end starts to send again the ACK clock takes over and
7525 	 * provides the most current window information.
7526 	 *
7527 	 * We must avoid the silly window syndrome whereas every read from
7528 	 * the receive buffer, no matter how small, causes a window update
7529 	 * to be sent.  We also should avoid sending a flurry of window
7530 	 * updates when the socket buffer had queued a lot of data and the
7531 	 * application is doing small reads.
7532 	 *
7533 	 * Prevent a flurry of pointless window updates by only sending an
7534 	 * update when we can increase the advertized window by more than
7535 	 * 1/4th of the socket buffer capacity.  When the buffer is getting
7536 	 * full or is very small be more aggressive and send an update
7537 	 * whenever we can increase by two mss sized segments. In all other
7538 	 * situations the ACK's to new incoming data will carry further
7539 	 * window increases.
7540 	 *
7541 	 * Don't send an independent window update if a delayed ACK is
7542 	 * pending (it will get piggy-backed on it) or the remote side
7543 	 * already has done a half-close and won't send more data.  Skip
7544 	 * this if the connection is in T/TCP half-open state.
7545 	 */
7546 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
7547 	    !(tp->t_flags & TF_DELACK) &&
7548 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
7549 		/*
7550 		 * "adv" is the amount we could increase the window, taking
7551 		 * into account that we are limited by TCP_MAXWIN <<
7552 		 * tp->rcv_scale.
7553 		 */
7554 		int32_t adv;
7555 		int oldwin;
7556 
7557 		adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
7558 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
7559 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
7560 			adv -= oldwin;
7561 		} else
7562 			oldwin = 0;
7563 
7564 		/*
7565 		 * If the new window size ends up being the same as the old
7566 		 * size when it is scaled, then don't force a window update.
7567 		 */
7568 		if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
7569 			goto dontupdate;
7570 
7571 		if (adv >= (int32_t)(2 * tp->t_maxseg) &&
7572 		    (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
7573 		    recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) ||
7574 		    so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg)) {
7575 			pass = 7;
7576 			goto send;
7577 		}
7578 		if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat)
7579 			goto send;
7580 	}
7581 dontupdate:
7582 
7583 	/*
7584 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
7585 	 * is also a catch-all for the retransmit timer timeout case.
7586 	 */
7587 	if (tp->t_flags & TF_ACKNOW) {
7588 		pass = 8;
7589 		goto send;
7590 	}
7591 	if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) {
7592 		pass = 9;
7593 		goto send;
7594 	}
7595 	if (SEQ_GT(tp->snd_up, tp->snd_una)) {
7596 		pass = 10;
7597 		goto send;
7598 	}
7599 	/*
7600 	 * If our state indicates that FIN should be sent and we have not
7601 	 * yet done so, then we need to send.
7602 	 */
7603 	if (flags & TH_FIN) {
7604 		if ((tp->t_flags & TF_SENTFIN) ||
7605 		    (((tp->t_flags & TF_SENTFIN) == 0) &&
7606 		     (tp->snd_nxt == tp->snd_una))) {
7607 			pass = 11;
7608 			goto send;
7609 		}
7610 	}
7611 	/*
7612 	 * No reason to send a segment, just return.
7613 	 */
7614 just_return:
7615 	SOCKBUF_UNLOCK(sb);
7616 just_return_nolock:
7617 	if (tot_len_this_send == 0)
7618 		counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1);
7619 	rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, tot_len_this_send, 1);
7620 	rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling);
7621 	tp->t_flags &= ~TF_FORCEDATA;
7622 	return (0);
7623 
7624 send:
7625 	if (doing_tlp == 0) {
7626 		/*
7627 		 * Data not a TLP, and its not the rxt firing. If it is the
7628 		 * rxt firing, we want to leave the tlp_in_progress flag on
7629 		 * so we don't send another TLP. It has to be a rack timer
7630 		 * or normal send (response to acked data) to clear the tlp
7631 		 * in progress flag.
7632 		 */
7633 		rack->rc_tlp_in_progress = 0;
7634 	}
7635 	SOCKBUF_LOCK_ASSERT(sb);
7636 	if (len > 0) {
7637 		if (len >= tp->t_maxseg)
7638 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
7639 		else
7640 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
7641 	}
7642 	/*
7643 	 * Before ESTABLISHED, force sending of initial options unless TCP
7644 	 * set not to do any options. NOTE: we assume that the IP/TCP header
7645 	 * plus TCP options always fit in a single mbuf, leaving room for a
7646 	 * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr)
7647 	 * + optlen <= MCLBYTES
7648 	 */
7649 	optlen = 0;
7650 #ifdef INET6
7651 	if (isipv6)
7652 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
7653 	else
7654 #endif
7655 		hdrlen = sizeof(struct tcpiphdr);
7656 
7657 	/*
7658 	 * Compute options for segment. We only have to care about SYN and
7659 	 * established connection segments.  Options for SYN-ACK segments
7660 	 * are handled in TCP syncache.
7661 	 */
7662 	to.to_flags = 0;
7663 	if ((tp->t_flags & TF_NOOPT) == 0) {
7664 		/* Maximum segment size. */
7665 		if (flags & TH_SYN) {
7666 			tp->snd_nxt = tp->iss;
7667 			to.to_mss = tcp_mssopt(&inp->inp_inc);
7668 #ifdef NETFLIX_TCPOUDP
7669 			if (tp->t_port)
7670 				to.to_mss -= V_tcp_udp_tunneling_overhead;
7671 #endif
7672 			to.to_flags |= TOF_MSS;
7673 
7674 			/*
7675 			 * On SYN or SYN|ACK transmits on TFO connections,
7676 			 * only include the TFO option if it is not a
7677 			 * retransmit, as the presence of the TFO option may
7678 			 * have caused the original SYN or SYN|ACK to have
7679 			 * been dropped by a middlebox.
7680 			 */
7681 			if (IS_FASTOPEN(tp->t_flags) &&
7682 			    (tp->t_rxtshift == 0)) {
7683 				if (tp->t_state == TCPS_SYN_RECEIVED) {
7684 					to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
7685 					to.to_tfo_cookie =
7686 					    (u_int8_t *)&tp->t_tfo_cookie.server;
7687 					to.to_flags |= TOF_FASTOPEN;
7688 					wanted_cookie = 1;
7689 				} else if (tp->t_state == TCPS_SYN_SENT) {
7690 					to.to_tfo_len =
7691 					    tp->t_tfo_client_cookie_len;
7692 					to.to_tfo_cookie =
7693 					    tp->t_tfo_cookie.client;
7694 					to.to_flags |= TOF_FASTOPEN;
7695 					wanted_cookie = 1;
7696 					/*
7697 					 * If we wind up having more data to
7698 					 * send with the SYN than can fit in
7699 					 * one segment, don't send any more
7700 					 * until the SYN|ACK comes back from
7701 					 * the other end.
7702 					 */
7703 					sendalot = 0;
7704 				}
7705 			}
7706 		}
7707 		/* Window scaling. */
7708 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
7709 			to.to_wscale = tp->request_r_scale;
7710 			to.to_flags |= TOF_SCALE;
7711 		}
7712 		/* Timestamps. */
7713 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
7714 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
7715 			to.to_tsval = cts + tp->ts_offset;
7716 			to.to_tsecr = tp->ts_recent;
7717 			to.to_flags |= TOF_TS;
7718 		}
7719 		/* Set receive buffer autosizing timestamp. */
7720 		if (tp->rfbuf_ts == 0 &&
7721 		    (so->so_rcv.sb_flags & SB_AUTOSIZE))
7722 			tp->rfbuf_ts = tcp_ts_getticks();
7723 		/* Selective ACK's. */
7724 		if (flags & TH_SYN)
7725 			to.to_flags |= TOF_SACKPERM;
7726 		else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
7727 		    tp->rcv_numsacks > 0) {
7728 			to.to_flags |= TOF_SACK;
7729 			to.to_nsacks = tp->rcv_numsacks;
7730 			to.to_sacks = (u_char *)tp->sackblks;
7731 		}
7732 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
7733 		/* TCP-MD5 (RFC2385). */
7734 		if (tp->t_flags & TF_SIGNATURE)
7735 			to.to_flags |= TOF_SIGNATURE;
7736 #endif				/* TCP_SIGNATURE */
7737 
7738 		/* Processing the options. */
7739 		hdrlen += optlen = tcp_addoptions(&to, opt);
7740 		/*
7741 		 * If we wanted a TFO option to be added, but it was unable
7742 		 * to fit, ensure no data is sent.
7743 		 */
7744 		if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
7745 		    !(to.to_flags & TOF_FASTOPEN))
7746 			len = 0;
7747 	}
7748 #ifdef NETFLIX_TCPOUDP
7749 	if (tp->t_port) {
7750 		if (V_tcp_udp_tunneling_port == 0) {
7751 			/* The port was removed?? */
7752 			SOCKBUF_UNLOCK(&so->so_snd);
7753 			return (EHOSTUNREACH);
7754 		}
7755 		hdrlen += sizeof(struct udphdr);
7756 	}
7757 #endif
7758 	ipoptlen = 0;
7759 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
7760 	ipoptlen += ipsec_optlen;
7761 #endif
7762 
7763 	/*
7764 	 * Adjust data length if insertion of options will bump the packet
7765 	 * length beyond the t_maxseg length. Clear the FIN bit because we
7766 	 * cut off the tail of the segment.
7767 	 */
7768 	if (len + optlen + ipoptlen > tp->t_maxseg) {
7769 		if (flags & TH_FIN) {
7770 			would_have_fin = 1;
7771 			flags &= ~TH_FIN;
7772 		}
7773 		if (tso) {
7774 			uint32_t if_hw_tsomax;
7775 			uint32_t moff;
7776 			int32_t max_len;
7777 
7778 			/* extract TSO information */
7779 			if_hw_tsomax = tp->t_tsomax;
7780 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
7781 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
7782 			KASSERT(ipoptlen == 0,
7783 			    ("%s: TSO can't do IP options", __func__));
7784 
7785 			/*
7786 			 * Check if we should limit by maximum payload
7787 			 * length:
7788 			 */
7789 			if (if_hw_tsomax != 0) {
7790 				/* compute maximum TSO length */
7791 				max_len = (if_hw_tsomax - hdrlen -
7792 				    max_linkhdr);
7793 				if (max_len <= 0) {
7794 					len = 0;
7795 				} else if (len > max_len) {
7796 					sendalot = 1;
7797 					len = max_len;
7798 				}
7799 			}
7800 			/*
7801 			 * Prevent the last segment from being fractional
7802 			 * unless the send sockbuf can be emptied:
7803 			 */
7804 			max_len = (tp->t_maxseg - optlen);
7805 			if ((sb_offset + len) < sbavail(sb)) {
7806 				moff = len % (u_int)max_len;
7807 				if (moff != 0) {
7808 					len -= moff;
7809 					sendalot = 1;
7810 				}
7811 			}
7812 			/*
7813 			 * In case there are too many small fragments don't
7814 			 * use TSO:
7815 			 */
7816 			if (len <= max_len) {
7817 				len = max_len;
7818 				sendalot = 1;
7819 				tso = 0;
7820 			}
7821 			/*
7822 			 * Send the FIN in a separate segment after the bulk
7823 			 * sending is done. We don't trust the TSO
7824 			 * implementations to clear the FIN flag on all but
7825 			 * the last segment.
7826 			 */
7827 			if (tp->t_flags & TF_NEEDFIN)
7828 				sendalot = 1;
7829 
7830 		} else {
7831 			len = tp->t_maxseg - optlen - ipoptlen;
7832 			sendalot = 1;
7833 		}
7834 	} else
7835 		tso = 0;
7836 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
7837 	    ("%s: len > IP_MAXPACKET", __func__));
7838 #ifdef DIAGNOSTIC
7839 #ifdef INET6
7840 	if (max_linkhdr + hdrlen > MCLBYTES)
7841 #else
7842 	if (max_linkhdr + hdrlen > MHLEN)
7843 #endif
7844 		panic("tcphdr too big");
7845 #endif
7846 
7847 	/*
7848 	 * This KASSERT is here to catch edge cases at a well defined place.
7849 	 * Before, those had triggered (random) panic conditions further
7850 	 * down.
7851 	 */
7852 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
7853 	if ((len == 0) &&
7854 	    (flags & TH_FIN) &&
7855 	    (sbused(sb))) {
7856 		/*
7857 		 * We have outstanding data, don't send a fin by itself!.
7858 		 */
7859 		goto just_return;
7860 	}
7861 	/*
7862 	 * Grab a header mbuf, attaching a copy of data to be transmitted,
7863 	 * and initialize the header from the template for sends on this
7864 	 * connection.
7865 	 */
7866 	if (len) {
7867 		uint32_t max_val;
7868 		uint32_t moff;
7869 
7870 		if (rack->rc_pace_max_segs)
7871 			max_val = rack->rc_pace_max_segs * tp->t_maxseg;
7872 		else
7873 			max_val = len;
7874 		/*
7875 		 * We allow a limit on sending with hptsi.
7876 		 */
7877 		if (len > max_val) {
7878 			len = max_val;
7879 		}
7880 #ifdef INET6
7881 		if (MHLEN < hdrlen + max_linkhdr)
7882 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
7883 		else
7884 #endif
7885 			m = m_gethdr(M_NOWAIT, MT_DATA);
7886 
7887 		if (m == NULL) {
7888 			SOCKBUF_UNLOCK(sb);
7889 			error = ENOBUFS;
7890 			sack_rxmit = 0;
7891 			goto out;
7892 		}
7893 		m->m_data += max_linkhdr;
7894 		m->m_len = hdrlen;
7895 
7896 		/*
7897 		 * Start the m_copy functions from the closest mbuf to the
7898 		 * sb_offset in the socket buffer chain.
7899 		 */
7900 		mb = sbsndptr_noadv(sb, sb_offset, &moff);
7901 		if (len <= MHLEN - hdrlen - max_linkhdr) {
7902 			m_copydata(mb, moff, (int)len,
7903 			    mtod(m, caddr_t)+hdrlen);
7904 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
7905 				sbsndptr_adv(sb, mb, len);
7906 			m->m_len += len;
7907 		} else {
7908 			struct sockbuf *msb;
7909 
7910 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
7911 				msb = NULL;
7912 			else
7913 				msb = sb;
7914 			m->m_next = tcp_m_copym(mb, moff, &len,
7915 			    if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb);
7916 			if (len <= (tp->t_maxseg - optlen)) {
7917 				/*
7918 				 * Must have ran out of mbufs for the copy
7919 				 * shorten it to no longer need tso. Lets
7920 				 * not put on sendalot since we are low on
7921 				 * mbufs.
7922 				 */
7923 				tso = 0;
7924 			}
7925 			if (m->m_next == NULL) {
7926 				SOCKBUF_UNLOCK(sb);
7927 				(void)m_free(m);
7928 				error = ENOBUFS;
7929 				sack_rxmit = 0;
7930 				goto out;
7931 			}
7932 		}
7933 		if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
7934 			TCPSTAT_INC(tcps_sndprobe);
7935 #ifdef NETFLIX_STATS
7936 			if (SEQ_LT(tp->snd_nxt, tp->snd_max))
7937 				stats_voi_update_abs_u32(tp->t_stats,
7938 				    VOI_TCP_RETXPB, len);
7939 			else
7940 				stats_voi_update_abs_u64(tp->t_stats,
7941 				    VOI_TCP_TXPB, len);
7942 #endif
7943 		} else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
7944 			if (rsm && (rsm->r_flags & RACK_TLP)) {
7945 				/*
7946 				 * TLP should not count in retran count, but
7947 				 * in its own bin
7948 				 */
7949 				counter_u64_add(rack_tlp_retran, 1);
7950 				counter_u64_add(rack_tlp_retran_bytes, len);
7951 			} else {
7952 				tp->t_sndrexmitpack++;
7953 				TCPSTAT_INC(tcps_sndrexmitpack);
7954 				TCPSTAT_ADD(tcps_sndrexmitbyte, len);
7955 			}
7956 #ifdef NETFLIX_STATS
7957 			stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
7958 			    len);
7959 #endif
7960 		} else {
7961 			TCPSTAT_INC(tcps_sndpack);
7962 			TCPSTAT_ADD(tcps_sndbyte, len);
7963 #ifdef NETFLIX_STATS
7964 			stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
7965 			    len);
7966 #endif
7967 		}
7968 		/*
7969 		 * If we're sending everything we've got, set PUSH. (This
7970 		 * will keep happy those implementations which only give
7971 		 * data to the user when a buffer fills or a PUSH comes in.)
7972 		 */
7973 		if (sb_offset + len == sbused(sb) &&
7974 		    sbused(sb) &&
7975 		    !(flags & TH_SYN))
7976 			flags |= TH_PUSH;
7977 
7978 		/*
7979 		 * Are we doing hptsi, if so we must calculate the slot. We
7980 		 * only do hptsi in ESTABLISHED and with no RESET being
7981 		 * sent where we have data to send.
7982 		 */
7983 		if (((tp->t_state == TCPS_ESTABLISHED) ||
7984 		    (tp->t_state == TCPS_CLOSE_WAIT) ||
7985 		    ((tp->t_state == TCPS_FIN_WAIT_1) &&
7986 		    ((tp->t_flags & TF_SENTFIN) == 0) &&
7987 		    ((flags & TH_FIN) == 0))) &&
7988 		    ((flags & TH_RST) == 0) &&
7989 		    (rack->rc_always_pace)) {
7990 			/*
7991 			 * We use the most optimistic possible cwnd/srtt for
7992 			 * sending calculations. This will make our
7993 			 * calculation anticipate getting more through
7994 			 * quicker then possible. But thats ok we don't want
7995 			 * the peer to have a gap in data sending.
7996 			 */
7997 			uint32_t srtt, cwnd, tr_perms = 0;
7998 
7999 			if (rack->r_ctl.rc_rack_min_rtt)
8000 				srtt = rack->r_ctl.rc_rack_min_rtt;
8001 			else
8002 				srtt = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT));
8003 			if (rack->r_ctl.rc_rack_largest_cwnd)
8004 				cwnd = rack->r_ctl.rc_rack_largest_cwnd;
8005 			else
8006 				cwnd = tp->snd_cwnd;
8007 			tr_perms = cwnd / srtt;
8008 			if (tr_perms == 0) {
8009 				tr_perms = tp->t_maxseg;
8010 			}
8011 			tot_len_this_send += len;
8012 			/*
8013 			 * Calculate how long this will take to drain, if
8014 			 * the calculation comes out to zero, thats ok we
8015 			 * will use send_a_lot to possibly spin around for
8016 			 * more increasing tot_len_this_send to the point
8017 			 * that its going to require a pace, or we hit the
8018 			 * cwnd. Which in that case we are just waiting for
8019 			 * a ACK.
8020 			 */
8021 			slot = tot_len_this_send / tr_perms;
8022 			/* Now do we reduce the time so we don't run dry? */
8023 			if (slot && rack->rc_pace_reduce) {
8024 				int32_t reduce;
8025 
8026 				reduce = (slot / rack->rc_pace_reduce);
8027 				if (reduce < slot) {
8028 					slot -= reduce;
8029 				} else
8030 					slot = 0;
8031 			}
8032 			if (rack->r_enforce_min_pace &&
8033 			    (slot == 0) &&
8034 			    (tot_len_this_send >= (rack->r_min_pace_seg_thresh * tp->t_maxseg))) {
8035 				/* We are enforcing a minimum pace time of 1ms */
8036 				slot = rack->r_enforce_min_pace;
8037 			}
8038 		}
8039 		SOCKBUF_UNLOCK(sb);
8040 	} else {
8041 		SOCKBUF_UNLOCK(sb);
8042 		if (tp->t_flags & TF_ACKNOW)
8043 			TCPSTAT_INC(tcps_sndacks);
8044 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
8045 			TCPSTAT_INC(tcps_sndctrl);
8046 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
8047 			TCPSTAT_INC(tcps_sndurg);
8048 		else
8049 			TCPSTAT_INC(tcps_sndwinup);
8050 
8051 		m = m_gethdr(M_NOWAIT, MT_DATA);
8052 		if (m == NULL) {
8053 			error = ENOBUFS;
8054 			sack_rxmit = 0;
8055 			goto out;
8056 		}
8057 #ifdef INET6
8058 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
8059 		    MHLEN >= hdrlen) {
8060 			M_ALIGN(m, hdrlen);
8061 		} else
8062 #endif
8063 			m->m_data += max_linkhdr;
8064 		m->m_len = hdrlen;
8065 	}
8066 	SOCKBUF_UNLOCK_ASSERT(sb);
8067 	m->m_pkthdr.rcvif = (struct ifnet *)0;
8068 #ifdef MAC
8069 	mac_inpcb_create_mbuf(inp, m);
8070 #endif
8071 #ifdef INET6
8072 	if (isipv6) {
8073 		ip6 = mtod(m, struct ip6_hdr *);
8074 #ifdef NETFLIX_TCPOUDP
8075 		if (tp->t_port) {
8076 			udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + sizeof(struct ip6_hdr));
8077 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
8078 			udp->uh_dport = tp->t_port;
8079 			ulen = hdrlen + len - sizeof(struct ip6_hdr);
8080 			udp->uh_ulen = htons(ulen);
8081 			th = (struct tcphdr *)(udp + 1);
8082 		} else
8083 #endif
8084 			th = (struct tcphdr *)(ip6 + 1);
8085 		tcpip_fillheaders(inp, ip6, th);
8086 	} else
8087 #endif				/* INET6 */
8088 	{
8089 		ip = mtod(m, struct ip *);
8090 #ifdef TCPDEBUG
8091 		ipov = (struct ipovly *)ip;
8092 #endif
8093 #ifdef NETFLIX_TCPOUDP
8094 		if (tp->t_port) {
8095 			udp = (struct udphdr *)((caddr_t)ip + ipoptlen + sizeof(struct ip));
8096 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
8097 			udp->uh_dport = tp->t_port;
8098 			ulen = hdrlen + len - sizeof(struct ip);
8099 			udp->uh_ulen = htons(ulen);
8100 			th = (struct tcphdr *)(udp + 1);
8101 		} else
8102 #endif
8103 			th = (struct tcphdr *)(ip + 1);
8104 		tcpip_fillheaders(inp, ip, th);
8105 	}
8106 	/*
8107 	 * Fill in fields, remembering maximum advertised window for use in
8108 	 * delaying messages about window sizes. If resending a FIN, be sure
8109 	 * not to use a new sequence number.
8110 	 */
8111 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
8112 	    tp->snd_nxt == tp->snd_max)
8113 		tp->snd_nxt--;
8114 	/*
8115 	 * If we are starting a connection, send ECN setup SYN packet. If we
8116 	 * are on a retransmit, we may resend those bits a number of times
8117 	 * as per RFC 3168.
8118 	 */
8119 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
8120 		if (tp->t_rxtshift >= 1) {
8121 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
8122 				flags |= TH_ECE | TH_CWR;
8123 		} else
8124 			flags |= TH_ECE | TH_CWR;
8125 	}
8126 	if (tp->t_state == TCPS_ESTABLISHED &&
8127 	    (tp->t_flags & TF_ECN_PERMIT)) {
8128 		/*
8129 		 * If the peer has ECN, mark data packets with ECN capable
8130 		 * transmission (ECT). Ignore pure ack packets,
8131 		 * retransmissions and window probes.
8132 		 */
8133 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
8134 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
8135 #ifdef INET6
8136 			if (isipv6)
8137 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
8138 			else
8139 #endif
8140 				ip->ip_tos |= IPTOS_ECN_ECT0;
8141 			TCPSTAT_INC(tcps_ecn_ect0);
8142 		}
8143 		/*
8144 		 * Reply with proper ECN notifications.
8145 		 */
8146 		if (tp->t_flags & TF_ECN_SND_CWR) {
8147 			flags |= TH_CWR;
8148 			tp->t_flags &= ~TF_ECN_SND_CWR;
8149 		}
8150 		if (tp->t_flags & TF_ECN_SND_ECE)
8151 			flags |= TH_ECE;
8152 	}
8153 	/*
8154 	 * If we are doing retransmissions, then snd_nxt will not reflect
8155 	 * the first unsent octet.  For ACK only packets, we do not want the
8156 	 * sequence number of the retransmitted packet, we want the sequence
8157 	 * number of the next unsent octet.  So, if there is no data (and no
8158 	 * SYN or FIN), use snd_max instead of snd_nxt when filling in
8159 	 * ti_seq.  But if we are in persist state, snd_max might reflect
8160 	 * one byte beyond the right edge of the window, so use snd_nxt in
8161 	 * that case, since we know we aren't doing a retransmission.
8162 	 * (retransmit and persist are mutually exclusive...)
8163 	 */
8164 	if (sack_rxmit == 0) {
8165 		if (len || (flags & (TH_SYN | TH_FIN)) ||
8166 		    rack->rc_in_persist) {
8167 			th->th_seq = htonl(tp->snd_nxt);
8168 			rack_seq = tp->snd_nxt;
8169 		} else if (flags & TH_RST) {
8170 			/*
8171 			 * For a Reset send the last cum ack in sequence
8172 			 * (this like any other choice may still generate a
8173 			 * challenge ack, if a ack-update packet is in
8174 			 * flight).
8175 			 */
8176 			th->th_seq = htonl(tp->snd_una);
8177 			rack_seq = tp->snd_una;
8178 		} else {
8179 			th->th_seq = htonl(tp->snd_max);
8180 			rack_seq = tp->snd_max;
8181 		}
8182 	} else {
8183 		th->th_seq = htonl(rsm->r_start);
8184 		rack_seq = rsm->r_start;
8185 	}
8186 	th->th_ack = htonl(tp->rcv_nxt);
8187 	if (optlen) {
8188 		bcopy(opt, th + 1, optlen);
8189 		th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
8190 	}
8191 	th->th_flags = flags;
8192 	/*
8193 	 * Calculate receive window.  Don't shrink window, but avoid silly
8194 	 * window syndrome.
8195 	 */
8196 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
8197 	    recwin < (long)tp->t_maxseg)
8198 		recwin = 0;
8199 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
8200 	    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
8201 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
8202 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
8203 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
8204 
8205 	/*
8206 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN> or
8207 	 * <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK> case is
8208 	 * handled in syncache.
8209 	 */
8210 	if (flags & TH_SYN)
8211 		th->th_win = htons((u_short)
8212 		    (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
8213 	else
8214 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
8215 	/*
8216 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0
8217 	 * window.  This may cause the remote transmitter to stall.  This
8218 	 * flag tells soreceive() to disable delayed acknowledgements when
8219 	 * draining the buffer.  This can occur if the receiver is
8220 	 * attempting to read more data than can be buffered prior to
8221 	 * transmitting on the connection.
8222 	 */
8223 	if (th->th_win == 0) {
8224 		tp->t_sndzerowin++;
8225 		tp->t_flags |= TF_RXWIN0SENT;
8226 	} else
8227 		tp->t_flags &= ~TF_RXWIN0SENT;
8228 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
8229 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
8230 		th->th_flags |= TH_URG;
8231 	} else
8232 		/*
8233 		 * If no urgent pointer to send, then we pull the urgent
8234 		 * pointer to the left edge of the send window so that it
8235 		 * doesn't drift into the send window on sequence number
8236 		 * wraparound.
8237 		 */
8238 		tp->snd_up = tp->snd_una;	/* drag it along */
8239 
8240 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
8241 	if (to.to_flags & TOF_SIGNATURE) {
8242 		/*
8243 		 * Calculate MD5 signature and put it into the place
8244 		 * determined before.
8245 		 * NOTE: since TCP options buffer doesn't point into
8246 		 * mbuf's data, calculate offset and use it.
8247 		 */
8248 		if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
8249 		    (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
8250 			/*
8251 			 * Do not send segment if the calculation of MD5
8252 			 * digest has failed.
8253 			 */
8254 			goto out;
8255 		}
8256 	}
8257 #endif
8258 
8259 	/*
8260 	 * Put TCP length in extended header, and then checksum extended
8261 	 * header and data.
8262 	 */
8263 	m->m_pkthdr.len = hdrlen + len;	/* in6_cksum() need this */
8264 #ifdef INET6
8265 	if (isipv6) {
8266 		/*
8267 		 * ip6_plen is not need to be filled now, and will be filled
8268 		 * in ip6_output.
8269 		 */
8270 		if (tp->t_port) {
8271 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
8272 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
8273 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
8274 			th->th_sum = htons(0);
8275 		} else {
8276 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
8277 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
8278 			th->th_sum = in6_cksum_pseudo(ip6,
8279 			    sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
8280 			    0);
8281 		}
8282 	}
8283 #endif
8284 #if defined(INET6) && defined(INET)
8285 	else
8286 #endif
8287 #ifdef INET
8288 	{
8289 		if (tp->t_port) {
8290 			m->m_pkthdr.csum_flags = CSUM_UDP;
8291 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
8292 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
8293 			   ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
8294 			th->th_sum = htons(0);
8295 		} else {
8296 			m->m_pkthdr.csum_flags = CSUM_TCP;
8297 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
8298 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
8299 			    ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
8300 			    IPPROTO_TCP + len + optlen));
8301 		}
8302 		/* IP version must be set here for ipv4/ipv6 checking later */
8303 		KASSERT(ip->ip_v == IPVERSION,
8304 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
8305 	}
8306 #endif
8307 
8308 	/*
8309 	 * Enable TSO and specify the size of the segments. The TCP pseudo
8310 	 * header checksum is always provided. XXX: Fixme: This is currently
8311 	 * not the case for IPv6.
8312 	 */
8313 	if (tso) {
8314 		KASSERT(len > tp->t_maxseg - optlen,
8315 		    ("%s: len <= tso_segsz", __func__));
8316 		m->m_pkthdr.csum_flags |= CSUM_TSO;
8317 		m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
8318 	}
8319 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
8320 	KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
8321 	    ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u",
8322 	    __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
8323 #else
8324 	KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
8325 	    ("%s: mbuf chain shorter than expected: %d + %u + %u != %u",
8326 	    __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
8327 #endif
8328 
8329 #ifdef TCP_HHOOK
8330 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
8331 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
8332 #endif
8333 
8334 #ifdef TCPDEBUG
8335 	/*
8336 	 * Trace.
8337 	 */
8338 	if (so->so_options & SO_DEBUG) {
8339 		u_short save = 0;
8340 
8341 #ifdef INET6
8342 		if (!isipv6)
8343 #endif
8344 		{
8345 			save = ipov->ih_len;
8346 			ipov->ih_len = htons(m->m_pkthdr.len	/* - hdrlen +
8347 			      * (th->th_off << 2) */ );
8348 		}
8349 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
8350 #ifdef INET6
8351 		if (!isipv6)
8352 #endif
8353 			ipov->ih_len = save;
8354 	}
8355 #endif				/* TCPDEBUG */
8356 
8357 	/* We're getting ready to send; log now. */
8358 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
8359 		union tcp_log_stackspecific log;
8360 
8361 		memset(&log.u_bbr, 0, sizeof(log.u_bbr));
8362 		log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts;
8363 		log.u_bbr.ininput = rack->rc_inp->inp_in_input;
8364 		log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt;
8365 		if (rsm || sack_rxmit) {
8366 			log.u_bbr.flex8 = 1;
8367 		} else {
8368 			log.u_bbr.flex8 = 0;
8369 		}
8370 		lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
8371 		    len, &log, false, NULL, NULL, 0, NULL);
8372 	} else
8373 		lgb = NULL;
8374 
8375 	/*
8376 	 * Fill in IP length and desired time to live and send to IP level.
8377 	 * There should be a better way to handle ttl and tos; we could keep
8378 	 * them in the template, but need a way to checksum without them.
8379 	 */
8380 	/*
8381 	 * m->m_pkthdr.len should have been set before cksum calcuration,
8382 	 * because in6_cksum() need it.
8383 	 */
8384 #ifdef INET6
8385 	if (isipv6) {
8386 		/*
8387 		 * we separately set hoplimit for every segment, since the
8388 		 * user might want to change the value via setsockopt. Also,
8389 		 * desired default hop limit might be changed via Neighbor
8390 		 * Discovery.
8391 		 */
8392 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
8393 
8394 		/*
8395 		 * Set the packet size here for the benefit of DTrace
8396 		 * probes. ip6_output() will set it properly; it's supposed
8397 		 * to include the option header lengths as well.
8398 		 */
8399 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
8400 
8401 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
8402 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
8403 		else
8404 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
8405 
8406 		if (tp->t_state == TCPS_SYN_SENT)
8407 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
8408 
8409 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
8410 		/* TODO: IPv6 IP6TOS_ECT bit on */
8411 		error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
8412 		    &inp->inp_route6,
8413 		    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
8414 		    NULL, NULL, inp);
8415 
8416 		if (error == EMSGSIZE && inp->inp_route6.ro_rt != NULL)
8417 			mtu = inp->inp_route6.ro_rt->rt_mtu;
8418 	}
8419 #endif				/* INET6 */
8420 #if defined(INET) && defined(INET6)
8421 	else
8422 #endif
8423 #ifdef INET
8424 	{
8425 		ip->ip_len = htons(m->m_pkthdr.len);
8426 #ifdef INET6
8427 		if (inp->inp_vflag & INP_IPV6PROTO)
8428 			ip->ip_ttl = in6_selecthlim(inp, NULL);
8429 #endif				/* INET6 */
8430 		/*
8431 		 * If we do path MTU discovery, then we set DF on every
8432 		 * packet. This might not be the best thing to do according
8433 		 * to RFC3390 Section 2. However the tcp hostcache migitates
8434 		 * the problem so it affects only the first tcp connection
8435 		 * with a host.
8436 		 *
8437 		 * NB: Don't set DF on small MTU/MSS to have a safe
8438 		 * fallback.
8439 		 */
8440 		if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
8441 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
8442 			if (tp->t_port == 0 || len < V_tcp_minmss) {
8443 				ip->ip_off |= htons(IP_DF);
8444 			}
8445 		} else {
8446 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
8447 		}
8448 
8449 		if (tp->t_state == TCPS_SYN_SENT)
8450 			TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
8451 
8452 		TCP_PROBE5(send, NULL, tp, ip, tp, th);
8453 
8454 		error = ip_output(m, tp->t_inpcb->inp_options, &inp->inp_route,
8455 		    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
8456 		    inp);
8457 		if (error == EMSGSIZE && inp->inp_route.ro_rt != NULL)
8458 			mtu = inp->inp_route.ro_rt->rt_mtu;
8459 	}
8460 #endif				/* INET */
8461 
8462 out:
8463 	if (lgb) {
8464 		lgb->tlb_errno = error;
8465 		lgb = NULL;
8466 	}
8467 	/*
8468 	 * In transmit state, time the transmission and arrange for the
8469 	 * retransmit.  In persist state, just set snd_max.
8470 	 */
8471 	if (error == 0) {
8472 		if (len == 0)
8473 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1);
8474 		else if (len == 1) {
8475 			counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1);
8476 		} else if (len > 1) {
8477 			int idx;
8478 
8479 			idx = (len / tp->t_maxseg) + 3;
8480 			if (idx >= TCP_MSS_ACCT_ATIMER)
8481 				counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1);
8482 			else
8483 				counter_u64_add(rack_out_size[idx], 1);
8484 		}
8485 	}
8486 	if (sub_from_prr && (error == 0)) {
8487 		rack->r_ctl.rc_prr_sndcnt -= len;
8488 	}
8489 	sub_from_prr = 0;
8490 	rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error, cts,
8491 	    pass, rsm);
8492 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
8493 	    (rack->rc_in_persist == 0)) {
8494 #ifdef NETFLIX_STATS
8495 		tcp_seq startseq = tp->snd_nxt;
8496 #endif
8497 
8498 		/*
8499 		 * Advance snd_nxt over sequence space of this segment.
8500 		 */
8501 		if (error)
8502 			/* We don't log or do anything with errors */
8503 			goto timer;
8504 
8505 		if (flags & (TH_SYN | TH_FIN)) {
8506 			if (flags & TH_SYN)
8507 				tp->snd_nxt++;
8508 			if (flags & TH_FIN) {
8509 				tp->snd_nxt++;
8510 				tp->t_flags |= TF_SENTFIN;
8511 			}
8512 		}
8513 		/* In the ENOBUFS case we do *not* update snd_max */
8514 		if (sack_rxmit)
8515 			goto timer;
8516 
8517 		tp->snd_nxt += len;
8518 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
8519 			if (tp->snd_una == tp->snd_max) {
8520 				/*
8521 				 * Update the time we just added data since
8522 				 * none was outstanding.
8523 				 */
8524 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
8525 				tp->t_acktime = ticks;
8526 			}
8527 			tp->snd_max = tp->snd_nxt;
8528 #ifdef NETFLIX_STATS
8529 			if (!(tp->t_flags & TF_GPUTINPROG) && len) {
8530 				tp->t_flags |= TF_GPUTINPROG;
8531 				tp->gput_seq = startseq;
8532 				tp->gput_ack = startseq +
8533 				    ulmin(sbavail(sb) - sb_offset, sendwin);
8534 				tp->gput_ts = tcp_ts_getticks();
8535 			}
8536 #endif
8537 		}
8538 		/*
8539 		 * Set retransmit timer if not currently set, and not doing
8540 		 * a pure ack or a keep-alive probe. Initial value for
8541 		 * retransmit timer is smoothed round-trip time + 2 *
8542 		 * round-trip time variance. Initialize shift counter which
8543 		 * is used for backoff of retransmit time.
8544 		 */
8545 timer:
8546 		if ((tp->snd_wnd == 0) &&
8547 		    TCPS_HAVEESTABLISHED(tp->t_state)) {
8548 			/*
8549 			 * If the persists timer was set above (right before
8550 			 * the goto send), and still needs to be on. Lets
8551 			 * make sure all is canceled. If the persist timer
8552 			 * is not running, we want to get it up.
8553 			 */
8554 			if (rack->rc_in_persist == 0) {
8555 				rack_enter_persist(tp, rack, cts);
8556 			}
8557 		}
8558 	} else {
8559 		/*
8560 		 * Persist case, update snd_max but since we are in persist
8561 		 * mode (no window) we do not update snd_nxt.
8562 		 */
8563 		int32_t xlen = len;
8564 
8565 		if (error)
8566 			goto nomore;
8567 
8568 		if (flags & TH_SYN)
8569 			++xlen;
8570 		if (flags & TH_FIN) {
8571 			++xlen;
8572 			tp->t_flags |= TF_SENTFIN;
8573 		}
8574 		/* In the ENOBUFS case we do *not* update snd_max */
8575 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max)) {
8576 			if (tp->snd_una == tp->snd_max) {
8577 				/*
8578 				 * Update the time we just added data since
8579 				 * none was outstanding.
8580 				 */
8581 				rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__);
8582 				tp->t_acktime = ticks;
8583 			}
8584 			tp->snd_max = tp->snd_nxt + len;
8585 		}
8586 	}
8587 nomore:
8588 	if (error) {
8589 		SOCKBUF_UNLOCK_ASSERT(sb);	/* Check gotos. */
8590 		/*
8591 		 * Failures do not advance the seq counter above. For the
8592 		 * case of ENOBUFS we will fall out and retry in 1ms with
8593 		 * the hpts. Everything else will just have to retransmit
8594 		 * with the timer.
8595 		 *
8596 		 * In any case, we do not want to loop around for another
8597 		 * send without a good reason.
8598 		 */
8599 		sendalot = 0;
8600 		switch (error) {
8601 		case EPERM:
8602 			tp->t_flags &= ~TF_FORCEDATA;
8603 			tp->t_softerror = error;
8604 			return (error);
8605 		case ENOBUFS:
8606 			if (slot == 0) {
8607 				/*
8608 				 * Pace us right away to retry in a some
8609 				 * time
8610 				 */
8611 				slot = 1 + rack->rc_enobuf;
8612 				if (rack->rc_enobuf < 255)
8613 					rack->rc_enobuf++;
8614 				if (slot > (rack->rc_rack_rtt / 2)) {
8615 					slot = rack->rc_rack_rtt / 2;
8616 				}
8617 				if (slot < 10)
8618 					slot = 10;
8619 			}
8620 			counter_u64_add(rack_saw_enobuf, 1);
8621 			error = 0;
8622 			goto enobufs;
8623 		case EMSGSIZE:
8624 			/*
8625 			 * For some reason the interface we used initially
8626 			 * to send segments changed to another or lowered
8627 			 * its MTU. If TSO was active we either got an
8628 			 * interface without TSO capabilits or TSO was
8629 			 * turned off. If we obtained mtu from ip_output()
8630 			 * then update it and try again.
8631 			 */
8632 			if (tso)
8633 				tp->t_flags &= ~TF_TSO;
8634 			if (mtu != 0) {
8635 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
8636 				goto again;
8637 			}
8638 			slot = 10;
8639 			rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, 0, 1);
8640 			tp->t_flags &= ~TF_FORCEDATA;
8641 			return (error);
8642 		case ENETUNREACH:
8643 			counter_u64_add(rack_saw_enetunreach, 1);
8644 		case EHOSTDOWN:
8645 		case EHOSTUNREACH:
8646 		case ENETDOWN:
8647 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
8648 				tp->t_softerror = error;
8649 			}
8650 			/* FALLTHROUGH */
8651 		default:
8652 			slot = 10;
8653 			rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, 0, 1);
8654 			tp->t_flags &= ~TF_FORCEDATA;
8655 			return (error);
8656 		}
8657 	} else {
8658 		rack->rc_enobuf = 0;
8659 	}
8660 	TCPSTAT_INC(tcps_sndtotal);
8661 
8662 	/*
8663 	 * Data sent (as far as we can tell). If this advertises a larger
8664 	 * window than any other segment, then remember the size of the
8665 	 * advertised window. Any pending ACK has now been sent.
8666 	 */
8667 	if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
8668 		tp->rcv_adv = tp->rcv_nxt + recwin;
8669 	tp->last_ack_sent = tp->rcv_nxt;
8670 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
8671 enobufs:
8672 	rack->r_tlp_running = 0;
8673 	if ((flags & TH_RST) || (would_have_fin == 1)) {
8674 		/*
8675 		 * We don't send again after a RST. We also do *not* send
8676 		 * again if we would have had a find, but now have
8677 		 * outstanding data.
8678 		 */
8679 		slot = 0;
8680 		sendalot = 0;
8681 	}
8682 	if (slot) {
8683 		/* set the rack tcb into the slot N */
8684 		counter_u64_add(rack_paced_segments, 1);
8685 	} else if (sendalot) {
8686 		if (len)
8687 			counter_u64_add(rack_unpaced_segments, 1);
8688 		sack_rxmit = 0;
8689 		tp->t_flags &= ~TF_FORCEDATA;
8690 		goto again;
8691 	} else if (len) {
8692 		counter_u64_add(rack_unpaced_segments, 1);
8693 	}
8694 	tp->t_flags &= ~TF_FORCEDATA;
8695 	rack_start_hpts_timer(rack, tp, cts, __LINE__, slot, tot_len_this_send, 1);
8696 	return (error);
8697 }
8698 
8699 /*
8700  * rack_ctloutput() must drop the inpcb lock before performing copyin on
8701  * socket option arguments.  When it re-acquires the lock after the copy, it
8702  * has to revalidate that the connection is still valid for the socket
8703  * option.
8704  */
8705 static int
8706 rack_set_sockopt(struct socket *so, struct sockopt *sopt,
8707     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
8708 {
8709 	int32_t error = 0, optval;
8710 
8711 	switch (sopt->sopt_name) {
8712 	case TCP_RACK_PROP_RATE:
8713 	case TCP_RACK_PROP:
8714 	case TCP_RACK_TLP_REDUCE:
8715 	case TCP_RACK_EARLY_RECOV:
8716 	case TCP_RACK_PACE_ALWAYS:
8717 	case TCP_DELACK:
8718 	case TCP_RACK_PACE_REDUCE:
8719 	case TCP_RACK_PACE_MAX_SEG:
8720 	case TCP_RACK_PRR_SENDALOT:
8721 	case TCP_RACK_MIN_TO:
8722 	case TCP_RACK_EARLY_SEG:
8723 	case TCP_RACK_REORD_THRESH:
8724 	case TCP_RACK_REORD_FADE:
8725 	case TCP_RACK_TLP_THRESH:
8726 	case TCP_RACK_PKT_DELAY:
8727 	case TCP_RACK_TLP_USE:
8728 	case TCP_RACK_TLP_INC_VAR:
8729 	case TCP_RACK_IDLE_REDUCE_HIGH:
8730 	case TCP_RACK_MIN_PACE:
8731 	case TCP_RACK_MIN_PACE_SEG:
8732 	case TCP_BBR_RACK_RTT_USE:
8733 	case TCP_DATA_AFTER_CLOSE:
8734 		break;
8735 	default:
8736 		return (tcp_default_ctloutput(so, sopt, inp, tp));
8737 		break;
8738 	}
8739 	INP_WUNLOCK(inp);
8740 	error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval));
8741 	if (error)
8742 		return (error);
8743 	INP_WLOCK(inp);
8744 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
8745 		INP_WUNLOCK(inp);
8746 		return (ECONNRESET);
8747 	}
8748 	tp = intotcpcb(inp);
8749 	rack = (struct tcp_rack *)tp->t_fb_ptr;
8750 	switch (sopt->sopt_name) {
8751 	case TCP_RACK_PROP_RATE:
8752 		if ((optval <= 0) || (optval >= 100)) {
8753 			error = EINVAL;
8754 			break;
8755 		}
8756 		RACK_OPTS_INC(tcp_rack_prop_rate);
8757 		rack->r_ctl.rc_prop_rate = optval;
8758 		break;
8759 	case TCP_RACK_TLP_USE:
8760 		if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) {
8761 			error = EINVAL;
8762 			break;
8763 		}
8764 		RACK_OPTS_INC(tcp_tlp_use);
8765 		rack->rack_tlp_threshold_use = optval;
8766 		break;
8767 	case TCP_RACK_PROP:
8768 		/* RACK proportional rate reduction (bool) */
8769 		RACK_OPTS_INC(tcp_rack_prop);
8770 		rack->r_ctl.rc_prop_reduce = optval;
8771 		break;
8772 	case TCP_RACK_TLP_REDUCE:
8773 		/* RACK TLP cwnd reduction (bool) */
8774 		RACK_OPTS_INC(tcp_rack_tlp_reduce);
8775 		rack->r_ctl.rc_tlp_cwnd_reduce = optval;
8776 		break;
8777 	case TCP_RACK_EARLY_RECOV:
8778 		/* Should recovery happen early (bool) */
8779 		RACK_OPTS_INC(tcp_rack_early_recov);
8780 		rack->r_ctl.rc_early_recovery = optval;
8781 		break;
8782 	case TCP_RACK_PACE_ALWAYS:
8783 		/* Use the always pace method (bool)  */
8784 		RACK_OPTS_INC(tcp_rack_pace_always);
8785 		if (optval > 0)
8786 			rack->rc_always_pace = 1;
8787 		else
8788 			rack->rc_always_pace = 0;
8789 		break;
8790 	case TCP_RACK_PACE_REDUCE:
8791 		/* RACK Hptsi reduction factor (divisor) */
8792 		RACK_OPTS_INC(tcp_rack_pace_reduce);
8793 		if (optval)
8794 			/* Must be non-zero */
8795 			rack->rc_pace_reduce = optval;
8796 		else
8797 			error = EINVAL;
8798 		break;
8799 	case TCP_RACK_PACE_MAX_SEG:
8800 		/* Max segments in a pace */
8801 		RACK_OPTS_INC(tcp_rack_max_seg);
8802 		rack->rc_pace_max_segs = optval;
8803 		break;
8804 	case TCP_RACK_PRR_SENDALOT:
8805 		/* Allow PRR to send more than one seg */
8806 		RACK_OPTS_INC(tcp_rack_prr_sendalot);
8807 		rack->r_ctl.rc_prr_sendalot = optval;
8808 		break;
8809 	case TCP_RACK_MIN_TO:
8810 		/* Minimum time between rack t-o's in ms */
8811 		RACK_OPTS_INC(tcp_rack_min_to);
8812 		rack->r_ctl.rc_min_to = optval;
8813 		break;
8814 	case TCP_RACK_EARLY_SEG:
8815 		/* If early recovery max segments */
8816 		RACK_OPTS_INC(tcp_rack_early_seg);
8817 		rack->r_ctl.rc_early_recovery_segs = optval;
8818 		break;
8819 	case TCP_RACK_REORD_THRESH:
8820 		/* RACK reorder threshold (shift amount) */
8821 		RACK_OPTS_INC(tcp_rack_reord_thresh);
8822 		if ((optval > 0) && (optval < 31))
8823 			rack->r_ctl.rc_reorder_shift = optval;
8824 		else
8825 			error = EINVAL;
8826 		break;
8827 	case TCP_RACK_REORD_FADE:
8828 		/* Does reordering fade after ms time */
8829 		RACK_OPTS_INC(tcp_rack_reord_fade);
8830 		rack->r_ctl.rc_reorder_fade = optval;
8831 		break;
8832 	case TCP_RACK_TLP_THRESH:
8833 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
8834 		RACK_OPTS_INC(tcp_rack_tlp_thresh);
8835 		if (optval)
8836 			rack->r_ctl.rc_tlp_threshold = optval;
8837 		else
8838 			error = EINVAL;
8839 		break;
8840 	case TCP_RACK_PKT_DELAY:
8841 		/* RACK added ms i.e. rack-rtt + reord + N */
8842 		RACK_OPTS_INC(tcp_rack_pkt_delay);
8843 		rack->r_ctl.rc_pkt_delay = optval;
8844 		break;
8845 	case TCP_RACK_TLP_INC_VAR:
8846 		/* Does TLP include rtt variance in t-o */
8847 		RACK_OPTS_INC(tcp_rack_tlp_inc_var);
8848 		rack->r_ctl.rc_prr_inc_var = optval;
8849 		break;
8850 	case TCP_RACK_IDLE_REDUCE_HIGH:
8851 		RACK_OPTS_INC(tcp_rack_idle_reduce_high);
8852 		if (optval)
8853 			rack->r_idle_reduce_largest = 1;
8854 		else
8855 			rack->r_idle_reduce_largest = 0;
8856 		break;
8857 	case TCP_DELACK:
8858 		if (optval == 0)
8859 			tp->t_delayed_ack = 0;
8860 		else
8861 			tp->t_delayed_ack = 1;
8862 		if (tp->t_flags & TF_DELACK) {
8863 			tp->t_flags &= ~TF_DELACK;
8864 			tp->t_flags |= TF_ACKNOW;
8865 			rack_output(tp);
8866 		}
8867 		break;
8868 	case TCP_RACK_MIN_PACE:
8869 		RACK_OPTS_INC(tcp_rack_min_pace);
8870 		if (optval > 3)
8871 			rack->r_enforce_min_pace = 3;
8872 		else
8873 			rack->r_enforce_min_pace = optval;
8874 		break;
8875 	case TCP_RACK_MIN_PACE_SEG:
8876 		RACK_OPTS_INC(tcp_rack_min_pace_seg);
8877 		if (optval >= 16)
8878 			rack->r_min_pace_seg_thresh = 15;
8879 		else
8880 			rack->r_min_pace_seg_thresh = optval;
8881 		break;
8882 	case TCP_BBR_RACK_RTT_USE:
8883 		if ((optval != USE_RTT_HIGH) &&
8884 		    (optval != USE_RTT_LOW) &&
8885 		    (optval != USE_RTT_AVG))
8886 			error = EINVAL;
8887 		else
8888 			rack->r_ctl.rc_rate_sample_method = optval;
8889 		break;
8890 	case TCP_DATA_AFTER_CLOSE:
8891 		if (optval)
8892 			rack->rc_allow_data_af_clo = 1;
8893 		else
8894 			rack->rc_allow_data_af_clo = 0;
8895 		break;
8896 	default:
8897 		return (tcp_default_ctloutput(so, sopt, inp, tp));
8898 		break;
8899 	}
8900 #ifdef NETFLIX_STATS
8901 	tcp_log_socket_option(tp, sopt->sopt_name, optval, error);
8902 #endif
8903 	INP_WUNLOCK(inp);
8904 	return (error);
8905 }
8906 
8907 static int
8908 rack_get_sockopt(struct socket *so, struct sockopt *sopt,
8909     struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack)
8910 {
8911 	int32_t error, optval;
8912 
8913 	/*
8914 	 * Because all our options are either boolean or an int, we can just
8915 	 * pull everything into optval and then unlock and copy. If we ever
8916 	 * add a option that is not a int, then this will have quite an
8917 	 * impact to this routine.
8918 	 */
8919 	switch (sopt->sopt_name) {
8920 	case TCP_RACK_PROP_RATE:
8921 		optval = rack->r_ctl.rc_prop_rate;
8922 		break;
8923 	case TCP_RACK_PROP:
8924 		/* RACK proportional rate reduction (bool) */
8925 		optval = rack->r_ctl.rc_prop_reduce;
8926 		break;
8927 	case TCP_RACK_TLP_REDUCE:
8928 		/* RACK TLP cwnd reduction (bool) */
8929 		optval = rack->r_ctl.rc_tlp_cwnd_reduce;
8930 		break;
8931 	case TCP_RACK_EARLY_RECOV:
8932 		/* Should recovery happen early (bool) */
8933 		optval = rack->r_ctl.rc_early_recovery;
8934 		break;
8935 	case TCP_RACK_PACE_REDUCE:
8936 		/* RACK Hptsi reduction factor (divisor) */
8937 		optval = rack->rc_pace_reduce;
8938 		break;
8939 	case TCP_RACK_PACE_MAX_SEG:
8940 		/* Max segments in a pace */
8941 		optval = rack->rc_pace_max_segs;
8942 		break;
8943 	case TCP_RACK_PACE_ALWAYS:
8944 		/* Use the always pace method */
8945 		optval = rack->rc_always_pace;
8946 		break;
8947 	case TCP_RACK_PRR_SENDALOT:
8948 		/* Allow PRR to send more than one seg */
8949 		optval = rack->r_ctl.rc_prr_sendalot;
8950 		break;
8951 	case TCP_RACK_MIN_TO:
8952 		/* Minimum time between rack t-o's in ms */
8953 		optval = rack->r_ctl.rc_min_to;
8954 		break;
8955 	case TCP_RACK_EARLY_SEG:
8956 		/* If early recovery max segments */
8957 		optval = rack->r_ctl.rc_early_recovery_segs;
8958 		break;
8959 	case TCP_RACK_REORD_THRESH:
8960 		/* RACK reorder threshold (shift amount) */
8961 		optval = rack->r_ctl.rc_reorder_shift;
8962 		break;
8963 	case TCP_RACK_REORD_FADE:
8964 		/* Does reordering fade after ms time */
8965 		optval = rack->r_ctl.rc_reorder_fade;
8966 		break;
8967 	case TCP_RACK_TLP_THRESH:
8968 		/* RACK TLP theshold i.e. srtt+(srtt/N) */
8969 		optval = rack->r_ctl.rc_tlp_threshold;
8970 		break;
8971 	case TCP_RACK_PKT_DELAY:
8972 		/* RACK added ms i.e. rack-rtt + reord + N */
8973 		optval = rack->r_ctl.rc_pkt_delay;
8974 		break;
8975 	case TCP_RACK_TLP_USE:
8976 		optval = rack->rack_tlp_threshold_use;
8977 		break;
8978 	case TCP_RACK_TLP_INC_VAR:
8979 		/* Does TLP include rtt variance in t-o */
8980 		optval = rack->r_ctl.rc_prr_inc_var;
8981 		break;
8982 	case TCP_RACK_IDLE_REDUCE_HIGH:
8983 		optval = rack->r_idle_reduce_largest;
8984 		break;
8985 	case TCP_RACK_MIN_PACE:
8986 		optval = rack->r_enforce_min_pace;
8987 		break;
8988 	case TCP_RACK_MIN_PACE_SEG:
8989 		optval = rack->r_min_pace_seg_thresh;
8990 		break;
8991 	case TCP_BBR_RACK_RTT_USE:
8992 		optval = rack->r_ctl.rc_rate_sample_method;
8993 		break;
8994 	case TCP_DELACK:
8995 		optval = tp->t_delayed_ack;
8996 		break;
8997 	case TCP_DATA_AFTER_CLOSE:
8998 		optval = rack->rc_allow_data_af_clo;
8999 		break;
9000 	default:
9001 		return (tcp_default_ctloutput(so, sopt, inp, tp));
9002 		break;
9003 	}
9004 	INP_WUNLOCK(inp);
9005 	error = sooptcopyout(sopt, &optval, sizeof optval);
9006 	return (error);
9007 }
9008 
9009 static int
9010 rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp)
9011 {
9012 	int32_t error = EINVAL;
9013 	struct tcp_rack *rack;
9014 
9015 	rack = (struct tcp_rack *)tp->t_fb_ptr;
9016 	if (rack == NULL) {
9017 		/* Huh? */
9018 		goto out;
9019 	}
9020 	if (sopt->sopt_dir == SOPT_SET) {
9021 		return (rack_set_sockopt(so, sopt, inp, tp, rack));
9022 	} else if (sopt->sopt_dir == SOPT_GET) {
9023 		return (rack_get_sockopt(so, sopt, inp, tp, rack));
9024 	}
9025 out:
9026 	INP_WUNLOCK(inp);
9027 	return (error);
9028 }
9029 
9030 
9031 struct tcp_function_block __tcp_rack = {
9032 	.tfb_tcp_block_name = __XSTRING(STACKNAME),
9033 	.tfb_tcp_output = rack_output,
9034 	.tfb_tcp_do_segment = rack_do_segment,
9035 	.tfb_tcp_hpts_do_segment = rack_hpts_do_segment,
9036 	.tfb_tcp_ctloutput = rack_ctloutput,
9037 	.tfb_tcp_fb_init = rack_init,
9038 	.tfb_tcp_fb_fini = rack_fini,
9039 	.tfb_tcp_timer_stop_all = rack_stopall,
9040 	.tfb_tcp_timer_activate = rack_timer_activate,
9041 	.tfb_tcp_timer_active = rack_timer_active,
9042 	.tfb_tcp_timer_stop = rack_timer_stop,
9043 	.tfb_tcp_rexmit_tmr = rack_remxt_tmr,
9044 	.tfb_tcp_handoff_ok = rack_handoff_ok
9045 };
9046 
9047 static const char *rack_stack_names[] = {
9048 	__XSTRING(STACKNAME),
9049 #ifdef STACKALIAS
9050 	__XSTRING(STACKALIAS),
9051 #endif
9052 };
9053 
9054 static int
9055 rack_ctor(void *mem, int32_t size, void *arg, int32_t how)
9056 {
9057 	memset(mem, 0, size);
9058 	return (0);
9059 }
9060 
9061 static void
9062 rack_dtor(void *mem, int32_t size, void *arg)
9063 {
9064 
9065 }
9066 
9067 static bool rack_mod_inited = false;
9068 
9069 static int
9070 tcp_addrack(module_t mod, int32_t type, void *data)
9071 {
9072 	int32_t err = 0;
9073 	int num_stacks;
9074 
9075 	switch (type) {
9076 	case MOD_LOAD:
9077 		rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map",
9078 		    sizeof(struct rack_sendmap),
9079 		    rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0);
9080 
9081 		rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb",
9082 		    sizeof(struct tcp_rack),
9083 		    rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0);
9084 
9085 		sysctl_ctx_init(&rack_sysctl_ctx);
9086 		rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx,
9087 		    SYSCTL_STATIC_CHILDREN(_net_inet_tcp),
9088 		    OID_AUTO,
9089 		    __XSTRING(STACKNAME),
9090 		    CTLFLAG_RW, 0,
9091 		    "");
9092 		if (rack_sysctl_root == NULL) {
9093 			printf("Failed to add sysctl node\n");
9094 			err = EFAULT;
9095 			goto free_uma;
9096 		}
9097 		rack_init_sysctls();
9098 		num_stacks = nitems(rack_stack_names);
9099 		err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK,
9100 		    rack_stack_names, &num_stacks);
9101 		if (err) {
9102 			printf("Failed to register %s stack name for "
9103 			    "%s module\n", rack_stack_names[num_stacks],
9104 			    __XSTRING(MODNAME));
9105 			sysctl_ctx_free(&rack_sysctl_ctx);
9106 free_uma:
9107 			uma_zdestroy(rack_zone);
9108 			uma_zdestroy(rack_pcb_zone);
9109 			rack_counter_destroy();
9110 			printf("Failed to register rack module -- err:%d\n", err);
9111 			return (err);
9112 		}
9113 		rack_mod_inited = true;
9114 		break;
9115 	case MOD_QUIESCE:
9116 		err = deregister_tcp_functions(&__tcp_rack, true, false);
9117 		break;
9118 	case MOD_UNLOAD:
9119 		err = deregister_tcp_functions(&__tcp_rack, false, true);
9120 		if (err == EBUSY)
9121 			break;
9122 		if (rack_mod_inited) {
9123 			uma_zdestroy(rack_zone);
9124 			uma_zdestroy(rack_pcb_zone);
9125 			sysctl_ctx_free(&rack_sysctl_ctx);
9126 			rack_counter_destroy();
9127 			rack_mod_inited = false;
9128 		}
9129 		err = 0;
9130 		break;
9131 	default:
9132 		return (EOPNOTSUPP);
9133 	}
9134 	return (err);
9135 }
9136 
9137 static moduledata_t tcp_rack = {
9138 	.name = __XSTRING(MODNAME),
9139 	.evhand = tcp_addrack,
9140 	.priv = 0
9141 };
9142 
9143 MODULE_VERSION(MODNAME, 1);
9144 DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY);
9145 MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1);
9146