xref: /freebsd/sys/netinet/tcp_timer.c (revision 2008043f386721d58158e37e0d7e50df8095942d)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *	@(#)tcp_timer.c	8.2 (Berkeley) 5/24/95
32  */
33 
34 #include <sys/cdefs.h>
35 #include "opt_inet.h"
36 #include "opt_inet6.h"
37 #include "opt_rss.h"
38 
39 #include <sys/param.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mbuf.h>
43 #include <sys/mutex.h>
44 #include <sys/protosw.h>
45 #include <sys/smp.h>
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/sysctl.h>
49 #include <sys/systm.h>
50 
51 #include <net/if.h>
52 #include <net/route.h>
53 #include <net/rss_config.h>
54 #include <net/vnet.h>
55 #include <net/netisr.h>
56 
57 #include <netinet/in.h>
58 #include <netinet/in_kdtrace.h>
59 #include <netinet/in_pcb.h>
60 #include <netinet/in_rss.h>
61 #include <netinet/in_systm.h>
62 #ifdef INET6
63 #include <netinet6/in6_pcb.h>
64 #endif
65 #include <netinet/ip_var.h>
66 #include <netinet/tcp.h>
67 #include <netinet/tcp_fsm.h>
68 #include <netinet/tcp_timer.h>
69 #include <netinet/tcp_var.h>
70 #include <netinet/tcp_log_buf.h>
71 #include <netinet/tcp_seq.h>
72 #include <netinet/cc/cc.h>
73 #ifdef INET6
74 #include <netinet6/tcp6_var.h>
75 #endif
76 #include <netinet/tcpip.h>
77 
78 int    tcp_persmin;
79 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, persmin,
80     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
81     &tcp_persmin, 0, sysctl_msec_to_ticks, "I",
82     "minimum persistence interval");
83 
84 int    tcp_persmax;
85 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, persmax,
86     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
87     &tcp_persmax, 0, sysctl_msec_to_ticks, "I",
88     "maximum persistence interval");
89 
90 int	tcp_keepinit;
91 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINIT, keepinit,
92     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
93     &tcp_keepinit, 0, sysctl_msec_to_ticks, "I",
94     "time to establish connection");
95 
96 int	tcp_keepidle;
97 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPIDLE, keepidle,
98     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
99     &tcp_keepidle, 0, sysctl_msec_to_ticks, "I",
100     "time before keepalive probes begin");
101 
102 int	tcp_keepintvl;
103 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINTVL, keepintvl,
104     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
105     &tcp_keepintvl, 0, sysctl_msec_to_ticks, "I",
106     "time between keepalive probes");
107 
108 int	tcp_delacktime;
109 SYSCTL_PROC(_net_inet_tcp, TCPCTL_DELACKTIME, delacktime,
110     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
111     &tcp_delacktime, 0, sysctl_msec_to_ticks, "I",
112     "Time before a delayed ACK is sent");
113 
114 VNET_DEFINE(int, tcp_msl);
115 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, msl,
116     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_VNET,
117     &VNET_NAME(tcp_msl), 0, sysctl_msec_to_ticks, "I",
118     "Maximum segment lifetime");
119 
120 int	tcp_rexmit_initial;
121 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_initial,
122    CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
123     &tcp_rexmit_initial, 0, sysctl_msec_to_ticks, "I",
124     "Initial Retransmission Timeout");
125 
126 int	tcp_rexmit_min;
127 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_min,
128     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
129     &tcp_rexmit_min, 0, sysctl_msec_to_ticks, "I",
130     "Minimum Retransmission Timeout");
131 
132 int	tcp_rexmit_slop;
133 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_slop,
134     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
135     &tcp_rexmit_slop, 0, sysctl_msec_to_ticks, "I",
136     "Retransmission Timer Slop");
137 
138 VNET_DEFINE(int, tcp_always_keepalive) = 1;
139 SYSCTL_INT(_net_inet_tcp, OID_AUTO, always_keepalive, CTLFLAG_VNET|CTLFLAG_RW,
140     &VNET_NAME(tcp_always_keepalive) , 0,
141     "Assume SO_KEEPALIVE on all TCP connections");
142 
143 int    tcp_fast_finwait2_recycle = 0;
144 SYSCTL_INT(_net_inet_tcp, OID_AUTO, fast_finwait2_recycle, CTLFLAG_RW,
145     &tcp_fast_finwait2_recycle, 0,
146     "Recycle closed FIN_WAIT_2 connections faster");
147 
148 int    tcp_finwait2_timeout;
149 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, finwait2_timeout,
150     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
151     &tcp_finwait2_timeout, 0, sysctl_msec_to_ticks, "I",
152     "FIN-WAIT2 timeout");
153 
154 int	tcp_keepcnt = TCPTV_KEEPCNT;
155 SYSCTL_INT(_net_inet_tcp, OID_AUTO, keepcnt, CTLFLAG_RW, &tcp_keepcnt, 0,
156     "Number of keepalive probes to send");
157 
158 	/* max idle probes */
159 int	tcp_maxpersistidle;
160 
161 int	tcp_rexmit_drop_options = 0;
162 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rexmit_drop_options, CTLFLAG_RW,
163     &tcp_rexmit_drop_options, 0,
164     "Drop TCP options from 3rd and later retransmitted SYN");
165 
166 int	tcp_maxunacktime = TCPTV_MAXUNACKTIME;
167 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, maxunacktime,
168     CTLTYPE_INT|CTLFLAG_RW | CTLFLAG_NEEDGIANT,
169     &tcp_maxunacktime, 0, sysctl_msec_to_ticks, "I",
170     "Maximum time (in ms) that a session can linger without making progress");
171 
172 VNET_DEFINE(int, tcp_pmtud_blackhole_detect);
173 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_detection,
174     CTLFLAG_RW|CTLFLAG_VNET,
175     &VNET_NAME(tcp_pmtud_blackhole_detect), 0,
176     "Path MTU Discovery Black Hole Detection Enabled");
177 
178 #ifdef INET
179 VNET_DEFINE(int, tcp_pmtud_blackhole_mss) = 1200;
180 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_mss,
181     CTLFLAG_RW|CTLFLAG_VNET,
182     &VNET_NAME(tcp_pmtud_blackhole_mss), 0,
183     "Path MTU Discovery Black Hole Detection lowered MSS");
184 #endif
185 
186 #ifdef INET6
187 VNET_DEFINE(int, tcp_v6pmtud_blackhole_mss) = 1220;
188 SYSCTL_INT(_net_inet_tcp, OID_AUTO, v6pmtud_blackhole_mss,
189     CTLFLAG_RW|CTLFLAG_VNET,
190     &VNET_NAME(tcp_v6pmtud_blackhole_mss), 0,
191     "Path MTU Discovery IPv6 Black Hole Detection lowered MSS");
192 #endif
193 
194 #ifdef	RSS
195 static int	per_cpu_timers = 1;
196 #else
197 static int	per_cpu_timers = 0;
198 #endif
199 SYSCTL_INT(_net_inet_tcp, OID_AUTO, per_cpu_timers, CTLFLAG_RW,
200     &per_cpu_timers , 0, "run tcp timers on all cpus");
201 
202 static int
203 sysctl_net_inet_tcp_retries(SYSCTL_HANDLER_ARGS)
204 {
205 	int error, new;
206 
207 	new = V_tcp_retries;
208 	error = sysctl_handle_int(oidp, &new, 0, req);
209 	if (error == 0 && req->newptr) {
210 		if ((new < 1) || (new > TCP_MAXRXTSHIFT))
211 			error = EINVAL;
212 		else
213 			V_tcp_retries = new;
214 	}
215 	return (error);
216 }
217 
218 VNET_DEFINE(int, tcp_retries) = TCP_MAXRXTSHIFT;
219 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, retries,
220     CTLTYPE_INT | CTLFLAG_VNET | CTLFLAG_RW,
221     &VNET_NAME(tcp_retries), 0, sysctl_net_inet_tcp_retries, "I",
222     "maximum number of consecutive timer based retransmissions");
223 
224 /*
225  * Map the given inp to a CPU id.
226  *
227  * This queries RSS if it's compiled in, else it defaults to the current
228  * CPU ID.
229  */
230 inline int
231 inp_to_cpuid(struct inpcb *inp)
232 {
233 	u_int cpuid;
234 
235 	if (per_cpu_timers) {
236 #ifdef	RSS
237 		cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
238 		if (cpuid == NETISR_CPUID_NONE)
239 			return (curcpu);	/* XXX */
240 		else
241 			return (cpuid);
242 #endif
243 		/*
244 		 * We don't have a flowid -> cpuid mapping, so cheat and
245 		 * just map unknown cpuids to curcpu.  Not the best, but
246 		 * apparently better than defaulting to swi 0.
247 		 */
248 		cpuid = inp->inp_flowid % (mp_maxid + 1);
249 		if (! CPU_ABSENT(cpuid))
250 			return (cpuid);
251 		return (curcpu);
252 	} else {
253 		return (0);
254 	}
255 }
256 
257 int	tcp_backoff[TCP_MAXRXTSHIFT + 1] =
258     { 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 512, 512, 512 };
259 
260 int tcp_totbackoff = 2559;	/* sum of tcp_backoff[] */
261 
262 /*
263  * TCP timer processing.
264  *
265  * Each connection has 5 timers associated with it, which can be scheduled
266  * simultaneously.  They all are serviced by one callout tcp_timer_enter().
267  * This function executes the next timer via tcp_timersw[] vector.  Each
268  * timer is supposed to return 'true' unless the connection was destroyed.
269  * In the former case tcp_timer_enter() will schedule callout for next timer.
270  */
271 
272 typedef bool tcp_timer_t(struct tcpcb *);
273 static tcp_timer_t tcp_timer_delack;
274 static tcp_timer_t tcp_timer_2msl;
275 static tcp_timer_t tcp_timer_keep;
276 static tcp_timer_t tcp_timer_persist;
277 static tcp_timer_t tcp_timer_rexmt;
278 
279 static tcp_timer_t * const tcp_timersw[TT_N] = {
280 	[TT_DELACK] = tcp_timer_delack,
281 	[TT_REXMT] = tcp_timer_rexmt,
282 	[TT_PERSIST] = tcp_timer_persist,
283 	[TT_KEEP] = tcp_timer_keep,
284 	[TT_2MSL] = tcp_timer_2msl,
285 };
286 
287 /*
288  * tcp_output_locked() s a timer specific variation of call to tcp_output(),
289  * see tcp_var.h for the rest.  It handles drop request from advanced stacks,
290  * but keeps tcpcb locked unless tcp_drop() destroyed it.
291  * Returns true if tcpcb is valid and locked.
292  */
293 static inline bool
294 tcp_output_locked(struct tcpcb *tp)
295 {
296 	int rv;
297 
298 	INP_WLOCK_ASSERT(tptoinpcb(tp));
299 
300 	if ((rv = tp->t_fb->tfb_tcp_output(tp)) < 0) {
301 		KASSERT(tp->t_fb->tfb_flags & TCP_FUNC_OUTPUT_CANDROP,
302 		    ("TCP stack %s requested tcp_drop(%p)",
303 		    tp->t_fb->tfb_tcp_block_name, tp));
304 		tp = tcp_drop(tp, rv);
305 	}
306 
307 	return (tp != NULL);
308 }
309 
310 static bool
311 tcp_timer_delack(struct tcpcb *tp)
312 {
313 	struct epoch_tracker et;
314 #if defined(INVARIANTS) || defined(VIMAGE)
315 	struct inpcb *inp = tptoinpcb(tp);
316 #endif
317 	bool rv;
318 
319 	INP_WLOCK_ASSERT(inp);
320 
321 	CURVNET_SET(inp->inp_vnet);
322 	tp->t_flags |= TF_ACKNOW;
323 	TCPSTAT_INC(tcps_delack);
324 	NET_EPOCH_ENTER(et);
325 	rv = tcp_output_locked(tp);
326 	NET_EPOCH_EXIT(et);
327 	CURVNET_RESTORE();
328 
329 	return (rv);
330 }
331 
332 static bool
333 tcp_timer_2msl(struct tcpcb *tp)
334 {
335 	struct inpcb *inp = tptoinpcb(tp);
336 	bool close = false;
337 
338 	INP_WLOCK_ASSERT(inp);
339 
340 	TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
341 	CURVNET_SET(inp->inp_vnet);
342 	tcp_log_end_status(tp, TCP_EI_STATUS_2MSL);
343 	tcp_free_sackholes(tp);
344 	/*
345 	 * 2 MSL timeout in shutdown went off.  If we're closed but
346 	 * still waiting for peer to close and connection has been idle
347 	 * too long delete connection control block.  Otherwise, check
348 	 * again in a bit.
349 	 *
350 	 * If fastrecycle of FIN_WAIT_2, in FIN_WAIT_2 and receiver has closed,
351 	 * there's no point in hanging onto FIN_WAIT_2 socket. Just close it.
352 	 * Ignore fact that there were recent incoming segments.
353 	 *
354 	 * XXXGL: check if inp_socket shall always be !NULL here?
355 	 */
356 	if (tp->t_state == TCPS_TIME_WAIT) {
357 		close = true;
358 	} else if (tp->t_state == TCPS_FIN_WAIT_2 &&
359 	    tcp_fast_finwait2_recycle && inp->inp_socket &&
360 	    (inp->inp_socket->so_rcv.sb_state & SBS_CANTRCVMORE)) {
361 		TCPSTAT_INC(tcps_finwait2_drops);
362 		close = true;
363 	} else {
364 		if (ticks - tp->t_rcvtime <= TP_MAXIDLE(tp))
365 			tcp_timer_activate(tp, TT_2MSL, TP_KEEPINTVL(tp));
366 		else
367 			close = true;
368 	}
369 	if (close) {
370 		struct epoch_tracker et;
371 
372 		NET_EPOCH_ENTER(et);
373 		tp = tcp_close(tp);
374 		NET_EPOCH_EXIT(et);
375 	}
376 	CURVNET_RESTORE();
377 
378 	return (tp != NULL);
379 }
380 
381 static bool
382 tcp_timer_keep(struct tcpcb *tp)
383 {
384 	struct epoch_tracker et;
385 	struct inpcb *inp = tptoinpcb(tp);
386 	struct tcptemp *t_template;
387 
388 	INP_WLOCK_ASSERT(inp);
389 
390 	TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
391 	CURVNET_SET(inp->inp_vnet);
392 	/*
393 	 * Because we don't regularly reset the keepalive callout in
394 	 * the ESTABLISHED state, it may be that we don't actually need
395 	 * to send a keepalive yet. If that occurs, schedule another
396 	 * call for the next time the keepalive timer might expire.
397 	 */
398 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
399 		u_int idletime;
400 
401 		idletime = ticks - tp->t_rcvtime;
402 		if (idletime < TP_KEEPIDLE(tp)) {
403 			tcp_timer_activate(tp, TT_KEEP,
404 			    TP_KEEPIDLE(tp) - idletime);
405 			CURVNET_RESTORE();
406 			return (true);
407 		}
408 	}
409 
410 	/*
411 	 * Keep-alive timer went off; send something
412 	 * or drop connection if idle for too long.
413 	 */
414 	TCPSTAT_INC(tcps_keeptimeo);
415 	if (tp->t_state < TCPS_ESTABLISHED)
416 		goto dropit;
417 	if ((V_tcp_always_keepalive ||
418 	    inp->inp_socket->so_options & SO_KEEPALIVE) &&
419 	    tp->t_state <= TCPS_CLOSING) {
420 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
421 			goto dropit;
422 		/*
423 		 * Send a packet designed to force a response
424 		 * if the peer is up and reachable:
425 		 * either an ACK if the connection is still alive,
426 		 * or an RST if the peer has closed the connection
427 		 * due to timeout or reboot.
428 		 * Using sequence number tp->snd_una-1
429 		 * causes the transmitted zero-length segment
430 		 * to lie outside the receive window;
431 		 * by the protocol spec, this requires the
432 		 * correspondent TCP to respond.
433 		 */
434 		TCPSTAT_INC(tcps_keepprobe);
435 		t_template = tcpip_maketemplate(inp);
436 		if (t_template) {
437 			NET_EPOCH_ENTER(et);
438 			tcp_respond(tp, t_template->tt_ipgen,
439 				    &t_template->tt_t, (struct mbuf *)NULL,
440 				    tp->rcv_nxt, tp->snd_una - 1, 0);
441 			NET_EPOCH_EXIT(et);
442 			free(t_template, M_TEMP);
443 		}
444 		tcp_timer_activate(tp, TT_KEEP, TP_KEEPINTVL(tp));
445 	} else
446 		tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp));
447 
448 	CURVNET_RESTORE();
449 	return (true);
450 
451 dropit:
452 	TCPSTAT_INC(tcps_keepdrops);
453 	NET_EPOCH_ENTER(et);
454 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
455 	tp = tcp_drop(tp, ETIMEDOUT);
456 	NET_EPOCH_EXIT(et);
457 	CURVNET_RESTORE();
458 
459 	return (tp != NULL);
460 }
461 
462 /*
463  * Has this session exceeded the maximum time without seeing a substantive
464  * acknowledgement? If so, return true; otherwise false.
465  */
466 static bool
467 tcp_maxunacktime_check(struct tcpcb *tp)
468 {
469 
470 	/* Are we tracking this timer for this session? */
471 	if (TP_MAXUNACKTIME(tp) == 0)
472 		return false;
473 
474 	/* Do we have a current measurement. */
475 	if (tp->t_acktime == 0)
476 		return false;
477 
478 	/* Are we within the acceptable range? */
479 	if (TSTMP_GT(TP_MAXUNACKTIME(tp) + tp->t_acktime, (u_int)ticks))
480 		return false;
481 
482 	/* We exceeded the timer. */
483 	TCPSTAT_INC(tcps_progdrops);
484 	return true;
485 }
486 
487 static bool
488 tcp_timer_persist(struct tcpcb *tp)
489 {
490 	struct epoch_tracker et;
491 #if defined(INVARIANTS) || defined(VIMAGE)
492 	struct inpcb *inp = tptoinpcb(tp);
493 #endif
494 	bool progdrop, rv;
495 
496 	INP_WLOCK_ASSERT(inp);
497 
498 	TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
499 	CURVNET_SET(inp->inp_vnet);
500 	/*
501 	 * Persistence timer into zero window.
502 	 * Force a byte to be output, if possible.
503 	 */
504 	TCPSTAT_INC(tcps_persisttimeo);
505 	/*
506 	 * Hack: if the peer is dead/unreachable, we do not
507 	 * time out if the window is closed.  After a full
508 	 * backoff, drop the connection if the idle time
509 	 * (no responses to probes) reaches the maximum
510 	 * backoff that we would use if retransmitting.
511 	 * Also, drop the connection if we haven't been making
512 	 * progress.
513 	 */
514 	progdrop = tcp_maxunacktime_check(tp);
515 	if (progdrop || (tp->t_rxtshift >= V_tcp_retries &&
516 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
517 	     ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff))) {
518 		if (!progdrop)
519 			TCPSTAT_INC(tcps_persistdrop);
520 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
521 		goto dropit;
522 	}
523 	/*
524 	 * If the user has closed the socket then drop a persisting
525 	 * connection after a much reduced timeout.
526 	 */
527 	if (tp->t_state > TCPS_CLOSE_WAIT &&
528 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
529 		TCPSTAT_INC(tcps_persistdrop);
530 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
531 		goto dropit;
532 	}
533 	tcp_setpersist(tp);
534 	tp->t_flags |= TF_FORCEDATA;
535 	NET_EPOCH_ENTER(et);
536 	if ((rv = tcp_output_locked(tp)))
537 		tp->t_flags &= ~TF_FORCEDATA;
538 	NET_EPOCH_EXIT(et);
539 	CURVNET_RESTORE();
540 
541 	return (rv);
542 
543 dropit:
544 	NET_EPOCH_ENTER(et);
545 	tp = tcp_drop(tp, ETIMEDOUT);
546 	NET_EPOCH_EXIT(et);
547 	CURVNET_RESTORE();
548 
549 	return (tp != NULL);
550 }
551 
552 static bool
553 tcp_timer_rexmt(struct tcpcb *tp)
554 {
555 	struct epoch_tracker et;
556 	struct inpcb *inp = tptoinpcb(tp);
557 	int rexmt;
558 	bool isipv6, rv;
559 
560 	INP_WLOCK_ASSERT(inp);
561 
562 	TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
563 	CURVNET_SET(inp->inp_vnet);
564 	tcp_free_sackholes(tp);
565 	if (tp->t_fb->tfb_tcp_rexmit_tmr) {
566 		/* The stack has a timer action too. */
567 		(*tp->t_fb->tfb_tcp_rexmit_tmr)(tp);
568 	}
569 	/*
570 	 * Retransmission timer went off.  Message has not
571 	 * been acked within retransmit interval.  Back off
572 	 * to a longer retransmit interval and retransmit one segment.
573 	 *
574 	 * If we've either exceeded the maximum number of retransmissions,
575 	 * or we've gone long enough without making progress, then drop
576 	 * the session.
577 	 */
578 	if (++tp->t_rxtshift > V_tcp_retries || tcp_maxunacktime_check(tp)) {
579 		if (tp->t_rxtshift > V_tcp_retries)
580 			TCPSTAT_INC(tcps_timeoutdrop);
581 		tp->t_rxtshift = V_tcp_retries;
582 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
583 		NET_EPOCH_ENTER(et);
584 		tp = tcp_drop(tp, ETIMEDOUT);
585 		NET_EPOCH_EXIT(et);
586 		CURVNET_RESTORE();
587 
588 		return (tp != NULL);
589 	}
590 	if (tp->t_state == TCPS_SYN_SENT) {
591 		/*
592 		 * If the SYN was retransmitted, indicate CWND to be
593 		 * limited to 1 segment in cc_conn_init().
594 		 */
595 		tp->snd_cwnd = 1;
596 	} else if (tp->t_rxtshift == 1) {
597 		/*
598 		 * first retransmit; record ssthresh and cwnd so they can
599 		 * be recovered if this turns out to be a "bad" retransmit.
600 		 * A retransmit is considered "bad" if an ACK for this
601 		 * segment is received within RTT/2 interval; the assumption
602 		 * here is that the ACK was already in flight.  See
603 		 * "On Estimating End-to-End Network Path Properties" by
604 		 * Allman and Paxson for more details.
605 		 */
606 		tp->snd_cwnd_prev = tp->snd_cwnd;
607 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
608 		tp->snd_recover_prev = tp->snd_recover;
609 		if (IN_FASTRECOVERY(tp->t_flags))
610 			tp->t_flags |= TF_WASFRECOVERY;
611 		else
612 			tp->t_flags &= ~TF_WASFRECOVERY;
613 		if (IN_CONGRECOVERY(tp->t_flags))
614 			tp->t_flags |= TF_WASCRECOVERY;
615 		else
616 			tp->t_flags &= ~TF_WASCRECOVERY;
617 		if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
618 			tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
619 		/* In the event that we've negotiated timestamps
620 		 * badrxtwin will be set to the value that we set
621 		 * the retransmitted packet's to_tsval to by tcp_output
622 		 */
623 		tp->t_flags |= TF_PREVVALID;
624 	} else
625 		tp->t_flags &= ~TF_PREVVALID;
626 	TCPSTAT_INC(tcps_rexmttimeo);
627 	if ((tp->t_state == TCPS_SYN_SENT) ||
628 	    (tp->t_state == TCPS_SYN_RECEIVED))
629 		rexmt = tcp_rexmit_initial * tcp_backoff[tp->t_rxtshift];
630 	else
631 		rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
632 	TCPT_RANGESET(tp->t_rxtcur, rexmt,
633 		      tp->t_rttmin, TCPTV_REXMTMAX);
634 
635 	/*
636 	 * We enter the path for PLMTUD if connection is established or, if
637 	 * connection is FIN_WAIT_1 status, reason for the last is that if
638 	 * amount of data we send is very small, we could send it in couple of
639 	 * packets and process straight to FIN. In that case we won't catch
640 	 * ESTABLISHED state.
641 	 */
642 #ifdef INET6
643 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
644 #else
645 	isipv6 = false;
646 #endif
647 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
648 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
649 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
650 	    ((tp->t_state == TCPS_ESTABLISHED) ||
651 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
652 		if (tp->t_rxtshift == 1) {
653 			/*
654 			 * We enter blackhole detection after the first
655 			 * unsuccessful timer based retransmission.
656 			 * Then we reduce up to two times the MSS, each
657 			 * candidate giving two tries of retransmissions.
658 			 * But we give a candidate only two tries, if it
659 			 * actually reduces the MSS.
660 			 */
661 			tp->t_blackhole_enter = 2;
662 			tp->t_blackhole_exit = tp->t_blackhole_enter;
663 			if (isipv6) {
664 #ifdef INET6
665 				if (tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss)
666 					tp->t_blackhole_exit += 2;
667 				if (tp->t_maxseg > V_tcp_v6mssdflt &&
668 				    V_tcp_v6pmtud_blackhole_mss > V_tcp_v6mssdflt)
669 					tp->t_blackhole_exit += 2;
670 #endif
671 			} else {
672 #ifdef INET
673 				if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss)
674 					tp->t_blackhole_exit += 2;
675 				if (tp->t_maxseg > V_tcp_mssdflt &&
676 				    V_tcp_pmtud_blackhole_mss > V_tcp_mssdflt)
677 					tp->t_blackhole_exit += 2;
678 #endif
679 			}
680 		}
681 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD|TF2_PLPMTU_MAXSEGSNT)) ==
682 		    (TF2_PLPMTU_PMTUD|TF2_PLPMTU_MAXSEGSNT)) &&
683 		    (tp->t_rxtshift >= tp->t_blackhole_enter &&
684 		    tp->t_rxtshift < tp->t_blackhole_exit &&
685 		    (tp->t_rxtshift - tp->t_blackhole_enter) % 2 == 0)) {
686 			/*
687 			 * Enter Path MTU Black-hole Detection mechanism:
688 			 * - Disable Path MTU Discovery (IP "DF" bit).
689 			 * - Reduce MTU to lower value than what we
690 			 *   negotiated with peer.
691 			 */
692 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
693 				/* Record that we may have found a black hole. */
694 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
695 				/* Keep track of previous MSS. */
696 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
697 			}
698 
699 			/*
700 			 * Reduce the MSS to blackhole value or to the default
701 			 * in an attempt to retransmit.
702 			 */
703 #ifdef INET6
704 			if (isipv6 &&
705 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss &&
706 			    V_tcp_v6pmtud_blackhole_mss > V_tcp_v6mssdflt) {
707 				/* Use the sysctl tuneable blackhole MSS. */
708 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
709 				TCPSTAT_INC(tcps_pmtud_blackhole_activated);
710 			} else if (isipv6) {
711 				/* Use the default MSS. */
712 				tp->t_maxseg = V_tcp_v6mssdflt;
713 				/*
714 				 * Disable Path MTU Discovery when we switch to
715 				 * minmss.
716 				 */
717 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
718 				TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
719 			}
720 #endif
721 #if defined(INET6) && defined(INET)
722 			else
723 #endif
724 #ifdef INET
725 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss &&
726 			    V_tcp_pmtud_blackhole_mss > V_tcp_mssdflt) {
727 				/* Use the sysctl tuneable blackhole MSS. */
728 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
729 				TCPSTAT_INC(tcps_pmtud_blackhole_activated);
730 			} else {
731 				/* Use the default MSS. */
732 				tp->t_maxseg = V_tcp_mssdflt;
733 				/*
734 				 * Disable Path MTU Discovery when we switch to
735 				 * minmss.
736 				 */
737 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
738 				TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
739 			}
740 #endif
741 			/*
742 			 * Reset the slow-start flight size
743 			 * as it may depend on the new MSS.
744 			 */
745 			if (CC_ALGO(tp)->conn_init != NULL)
746 				CC_ALGO(tp)->conn_init(&tp->t_ccv);
747 		} else {
748 			/*
749 			 * If further retransmissions are still unsuccessful
750 			 * with a lowered MTU, maybe this isn't a blackhole and
751 			 * we restore the previous MSS and blackhole detection
752 			 * flags.
753 			 */
754 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
755 			    (tp->t_rxtshift >= tp->t_blackhole_exit)) {
756 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
757 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
758 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
759 				TCPSTAT_INC(tcps_pmtud_blackhole_failed);
760 				/*
761 				 * Reset the slow-start flight size as it
762 				 * may depend on the new MSS.
763 				 */
764 				if (CC_ALGO(tp)->conn_init != NULL)
765 					CC_ALGO(tp)->conn_init(&tp->t_ccv);
766 			}
767 		}
768 	}
769 
770 	/*
771 	 * Disable RFC1323 and SACK if we haven't got any response to
772 	 * our third SYN to work-around some broken terminal servers
773 	 * (most of which have hopefully been retired) that have bad VJ
774 	 * header compression code which trashes TCP segments containing
775 	 * unknown-to-them TCP options.
776 	 */
777 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
778 	    (tp->t_rxtshift == 3))
779 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
780 	/*
781 	 * If we backed off this far, notify the L3 protocol that we're having
782 	 * connection problems.
783 	 */
784 	if (tp->t_rxtshift > TCP_RTT_INVALIDATE) {
785 #ifdef INET6
786 		if ((inp->inp_vflag & INP_IPV6) != 0)
787 			in6_losing(inp);
788 		else
789 #endif
790 			in_losing(inp);
791 	}
792 	tp->snd_nxt = tp->snd_una;
793 	tp->snd_recover = tp->snd_max;
794 	/*
795 	 * Force a segment to be sent.
796 	 */
797 	tp->t_flags |= TF_ACKNOW;
798 	/*
799 	 * If timing a segment in this window, stop the timer.
800 	 */
801 	tp->t_rtttime = 0;
802 
803 	cc_cong_signal(tp, NULL, CC_RTO);
804 	NET_EPOCH_ENTER(et);
805 	rv = tcp_output_locked(tp);
806 	NET_EPOCH_EXIT(et);
807 	CURVNET_RESTORE();
808 
809 	return (rv);
810 }
811 
812 static void
813 tcp_bblog_timer(struct tcpcb *tp, tt_which which, tt_what what, uint32_t ticks)
814 {
815 	struct tcp_log_buffer *lgb;
816 	uint64_t ms;
817 
818 	INP_WLOCK_ASSERT(tptoinpcb(tp));
819 	if (tcp_bblogging_on(tp))
820 		lgb = tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_RTO, 0, 0,
821 		    NULL, false, NULL, NULL, 0, NULL);
822 	else
823 		lgb = NULL;
824 	if (lgb != NULL) {
825 		lgb->tlb_flex1 = (what << 8) | which;
826 		if (what == TT_STARTING) {
827 			/* Convert ticks to ms and store it in tlb_flex2. */
828 			if (hz == 1000)
829 				lgb->tlb_flex2 = ticks;
830 			else {
831 				ms = (((uint64_t)ticks * 1000) + (hz - 1)) / hz;
832 				if (ms > UINT32_MAX)
833 					lgb->tlb_flex2 = UINT32_MAX;
834 				else
835 					lgb->tlb_flex2 = (uint32_t)ms;
836 			}
837 		}
838 	}
839 }
840 
841 static inline tt_which
842 tcp_timer_next(struct tcpcb *tp, sbintime_t *precision)
843 {
844 	tt_which i, rv;
845 	sbintime_t after, before;
846 
847 	for (i = 0, rv = TT_N, after = before = SBT_MAX; i < TT_N; i++) {
848 		if (tp->t_timers[i] < after) {
849 			after = tp->t_timers[i];
850 			rv = i;
851 		}
852 		before = MIN(before, tp->t_timers[i] + tp->t_precisions[i]);
853 	}
854 	if (precision != NULL)
855 		*precision = before - after;
856 
857 	return (rv);
858 }
859 
860 static void
861 tcp_timer_enter(void *xtp)
862 {
863 	struct tcpcb *tp = xtp;
864 	struct inpcb *inp = tptoinpcb(tp);
865 	sbintime_t precision;
866 	tt_which which;
867 	bool tp_valid;
868 
869 	INP_WLOCK_ASSERT(inp);
870 	MPASS((curthread->td_pflags & TDP_INTCPCALLOUT) == 0);
871 
872 	curthread->td_pflags |= TDP_INTCPCALLOUT;
873 
874 	which = tcp_timer_next(tp, NULL);
875 	MPASS(which < TT_N);
876 	tp->t_timers[which] = SBT_MAX;
877 	tp->t_precisions[which] = 0;
878 
879 	tcp_bblog_timer(tp, which, TT_PROCESSING, 0);
880 	tp_valid = tcp_timersw[which](tp);
881 	if (tp_valid) {
882 		tcp_bblog_timer(tp, which, TT_PROCESSED, 0);
883 		if ((which = tcp_timer_next(tp, &precision)) != TT_N) {
884 			callout_reset_sbt_on(&tp->t_callout,
885 			    tp->t_timers[which], precision, tcp_timer_enter,
886 			    tp, inp_to_cpuid(inp), C_ABSOLUTE);
887 		}
888 		INP_WUNLOCK(inp);
889 	}
890 
891 	curthread->td_pflags &= ~TDP_INTCPCALLOUT;
892 }
893 
894 /*
895  * Activate or stop (delta == 0) a TCP timer.
896  */
897 void
898 tcp_timer_activate(struct tcpcb *tp, tt_which which, u_int delta)
899 {
900 	struct inpcb *inp = tptoinpcb(tp);
901 	sbintime_t precision;
902 	tt_what what;
903 
904 #ifdef TCP_OFFLOAD
905 	if (tp->t_flags & TF_TOE)
906 		return;
907 #endif
908 
909 	INP_WLOCK_ASSERT(inp);
910 
911 	if (delta > 0) {
912 		what = TT_STARTING;
913 		callout_when(tick_sbt * delta, 0, C_HARDCLOCK,
914 		    &tp->t_timers[which], &tp->t_precisions[which]);
915 	} else {
916 		what = TT_STOPPING;
917 		tp->t_timers[which] = SBT_MAX;
918 	}
919 	tcp_bblog_timer(tp, which, what, delta);
920 
921 	if ((which = tcp_timer_next(tp, &precision)) != TT_N)
922 		callout_reset_sbt_on(&tp->t_callout, tp->t_timers[which],
923 		    precision, tcp_timer_enter, tp, inp_to_cpuid(inp),
924 		    C_ABSOLUTE);
925 	else
926 		callout_stop(&tp->t_callout);
927 }
928 
929 bool
930 tcp_timer_active(struct tcpcb *tp, tt_which which)
931 {
932 
933 	INP_WLOCK_ASSERT(tptoinpcb(tp));
934 
935 	return (tp->t_timers[which] != SBT_MAX);
936 }
937 
938 /*
939  * Stop all timers associated with tcpcb.
940  *
941  * Called only on tcpcb destruction.  The tcpcb shall already be dropped from
942  * the pcb lookup database and socket is not losing the last reference.
943  *
944  * XXXGL: unfortunately our callout(9) is not able to fully stop a locked
945  * callout even when only two threads are involved: the callout itself and the
946  * thread that does callout_stop().  See where softclock_call_cc() swaps the
947  * callwheel lock to callout lock and then checks cc_exec_cancel().  This is
948  * the race window.  If it happens, the tcp_timer_enter() won't be executed,
949  * however pcb lock will be locked and released, hence we can't free memory.
950  * Until callout(9) is improved, just keep retrying.  In my profiling I've seen
951  * such event happening less than 1 time per hour with 20-30 Gbit/s of traffic.
952  */
953 void
954 tcp_timer_stop(struct tcpcb *tp)
955 {
956 	struct inpcb *inp = tptoinpcb(tp);
957 
958 	INP_WLOCK_ASSERT(inp);
959 
960 	if (curthread->td_pflags & TDP_INTCPCALLOUT) {
961 		int stopped __diagused;
962 
963 		stopped = callout_stop(&tp->t_callout);
964 		MPASS(stopped == 0);
965 	} else while(__predict_false(callout_stop(&tp->t_callout) == 0)) {
966 		INP_WUNLOCK(inp);
967 		kern_yield(PRI_UNCHANGED);
968 		INP_WLOCK(inp);
969 	}
970 }
971