xref: /freebsd/sys/netinet/tcp_timer.c (revision ba3c1f5972d7b90feb6e6da47905ff2757e0fe57)
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 __FBSDID("$FreeBSD$");
36 
37 #include "opt_inet.h"
38 #include "opt_inet6.h"
39 #include "opt_rss.h"
40 
41 #include <sys/param.h>
42 #include <sys/kernel.h>
43 #include <sys/lock.h>
44 #include <sys/mbuf.h>
45 #include <sys/mutex.h>
46 #include <sys/protosw.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
51 #include <sys/systm.h>
52 
53 #include <net/if.h>
54 #include <net/route.h>
55 #include <net/rss_config.h>
56 #include <net/vnet.h>
57 #include <net/netisr.h>
58 
59 #include <netinet/in.h>
60 #include <netinet/in_kdtrace.h>
61 #include <netinet/in_pcb.h>
62 #include <netinet/in_rss.h>
63 #include <netinet/in_systm.h>
64 #ifdef INET6
65 #include <netinet6/in6_pcb.h>
66 #endif
67 #include <netinet/ip_var.h>
68 #include <netinet/tcp.h>
69 #include <netinet/tcp_fsm.h>
70 #include <netinet/tcp_timer.h>
71 #include <netinet/tcp_var.h>
72 #include <netinet/tcp_log_buf.h>
73 #include <netinet/tcp_seq.h>
74 #include <netinet/cc/cc.h>
75 #ifdef INET6
76 #include <netinet6/tcp6_var.h>
77 #endif
78 #include <netinet/tcpip.h>
79 
80 int    tcp_persmin;
81 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, persmin,
82     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
83     &tcp_persmin, 0, sysctl_msec_to_ticks, "I",
84     "minimum persistence interval");
85 
86 int    tcp_persmax;
87 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, persmax,
88     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
89     &tcp_persmax, 0, sysctl_msec_to_ticks, "I",
90     "maximum persistence interval");
91 
92 int	tcp_keepinit;
93 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINIT, keepinit,
94     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
95     &tcp_keepinit, 0, sysctl_msec_to_ticks, "I",
96     "time to establish connection");
97 
98 int	tcp_keepidle;
99 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPIDLE, keepidle,
100     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
101     &tcp_keepidle, 0, sysctl_msec_to_ticks, "I",
102     "time before keepalive probes begin");
103 
104 int	tcp_keepintvl;
105 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINTVL, keepintvl,
106     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
107     &tcp_keepintvl, 0, sysctl_msec_to_ticks, "I",
108     "time between keepalive probes");
109 
110 int	tcp_delacktime;
111 SYSCTL_PROC(_net_inet_tcp, TCPCTL_DELACKTIME, delacktime,
112     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
113     &tcp_delacktime, 0, sysctl_msec_to_ticks, "I",
114     "Time before a delayed ACK is sent");
115 
116 VNET_DEFINE(int, tcp_msl);
117 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, msl,
118     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_VNET,
119     &VNET_NAME(tcp_msl), 0, sysctl_msec_to_ticks, "I",
120     "Maximum segment lifetime");
121 
122 int	tcp_rexmit_initial;
123 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_initial,
124    CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
125     &tcp_rexmit_initial, 0, sysctl_msec_to_ticks, "I",
126     "Initial Retransmission Timeout");
127 
128 int	tcp_rexmit_min;
129 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_min,
130     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
131     &tcp_rexmit_min, 0, sysctl_msec_to_ticks, "I",
132     "Minimum Retransmission Timeout");
133 
134 int	tcp_rexmit_slop;
135 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_slop,
136     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
137     &tcp_rexmit_slop, 0, sysctl_msec_to_ticks, "I",
138     "Retransmission Timer Slop");
139 
140 VNET_DEFINE(int, tcp_always_keepalive) = 1;
141 SYSCTL_INT(_net_inet_tcp, OID_AUTO, always_keepalive, CTLFLAG_VNET|CTLFLAG_RW,
142     &VNET_NAME(tcp_always_keepalive) , 0,
143     "Assume SO_KEEPALIVE on all TCP connections");
144 
145 int    tcp_fast_finwait2_recycle = 0;
146 SYSCTL_INT(_net_inet_tcp, OID_AUTO, fast_finwait2_recycle, CTLFLAG_RW,
147     &tcp_fast_finwait2_recycle, 0,
148     "Recycle closed FIN_WAIT_2 connections faster");
149 
150 int    tcp_finwait2_timeout;
151 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, finwait2_timeout,
152     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
153     &tcp_finwait2_timeout, 0, sysctl_msec_to_ticks, "I",
154     "FIN-WAIT2 timeout");
155 
156 int	tcp_keepcnt = TCPTV_KEEPCNT;
157 SYSCTL_INT(_net_inet_tcp, OID_AUTO, keepcnt, CTLFLAG_RW, &tcp_keepcnt, 0,
158     "Number of keepalive probes to send");
159 
160 	/* max idle probes */
161 int	tcp_maxpersistidle;
162 
163 int	tcp_rexmit_drop_options = 0;
164 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rexmit_drop_options, CTLFLAG_RW,
165     &tcp_rexmit_drop_options, 0,
166     "Drop TCP options from 3rd and later retransmitted SYN");
167 
168 int	tcp_maxunacktime = TCPTV_MAXUNACKTIME;
169 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, maxunacktime,
170     CTLTYPE_INT|CTLFLAG_RW | CTLFLAG_NEEDGIANT,
171     &tcp_maxunacktime, 0, sysctl_msec_to_ticks, "I",
172     "Maximum time (in ms) that a session can linger without making progress");
173 
174 VNET_DEFINE(int, tcp_pmtud_blackhole_detect);
175 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_detection,
176     CTLFLAG_RW|CTLFLAG_VNET,
177     &VNET_NAME(tcp_pmtud_blackhole_detect), 0,
178     "Path MTU Discovery Black Hole Detection Enabled");
179 
180 #ifdef INET
181 VNET_DEFINE(int, tcp_pmtud_blackhole_mss) = 1200;
182 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_mss,
183     CTLFLAG_RW|CTLFLAG_VNET,
184     &VNET_NAME(tcp_pmtud_blackhole_mss), 0,
185     "Path MTU Discovery Black Hole Detection lowered MSS");
186 #endif
187 
188 #ifdef INET6
189 VNET_DEFINE(int, tcp_v6pmtud_blackhole_mss) = 1220;
190 SYSCTL_INT(_net_inet_tcp, OID_AUTO, v6pmtud_blackhole_mss,
191     CTLFLAG_RW|CTLFLAG_VNET,
192     &VNET_NAME(tcp_v6pmtud_blackhole_mss), 0,
193     "Path MTU Discovery IPv6 Black Hole Detection lowered MSS");
194 #endif
195 
196 #ifdef	RSS
197 static int	per_cpu_timers = 1;
198 #else
199 static int	per_cpu_timers = 0;
200 #endif
201 SYSCTL_INT(_net_inet_tcp, OID_AUTO, per_cpu_timers, CTLFLAG_RW,
202     &per_cpu_timers , 0, "run tcp timers on all cpus");
203 
204 static int
205 sysctl_net_inet_tcp_retries(SYSCTL_HANDLER_ARGS)
206 {
207 	int error, new;
208 
209 	new = V_tcp_retries;
210 	error = sysctl_handle_int(oidp, &new, 0, req);
211 	if (error == 0 && req->newptr) {
212 		if ((new < 1) || (new > TCP_MAXRXTSHIFT))
213 			error = EINVAL;
214 		else
215 			V_tcp_retries = new;
216 	}
217 	return (error);
218 }
219 
220 VNET_DEFINE(int, tcp_retries) = TCP_MAXRXTSHIFT;
221 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, retries,
222     CTLTYPE_INT | CTLFLAG_VNET | CTLFLAG_RW,
223     &VNET_NAME(tcp_retries), 0, sysctl_net_inet_tcp_retries, "I",
224     "maximum number of consecutive timer based retransmissions");
225 
226 /*
227  * Map the given inp to a CPU id.
228  *
229  * This queries RSS if it's compiled in, else it defaults to the current
230  * CPU ID.
231  */
232 inline int
233 inp_to_cpuid(struct inpcb *inp)
234 {
235 	u_int cpuid;
236 
237 	if (per_cpu_timers) {
238 #ifdef	RSS
239 		cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
240 		if (cpuid == NETISR_CPUID_NONE)
241 			return (curcpu);	/* XXX */
242 		else
243 			return (cpuid);
244 #endif
245 		/*
246 		 * We don't have a flowid -> cpuid mapping, so cheat and
247 		 * just map unknown cpuids to curcpu.  Not the best, but
248 		 * apparently better than defaulting to swi 0.
249 		 */
250 		cpuid = inp->inp_flowid % (mp_maxid + 1);
251 		if (! CPU_ABSENT(cpuid))
252 			return (cpuid);
253 		return (curcpu);
254 	} else {
255 		return (0);
256 	}
257 }
258 
259 int	tcp_backoff[TCP_MAXRXTSHIFT + 1] =
260     { 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 512, 512, 512 };
261 
262 int tcp_totbackoff = 2559;	/* sum of tcp_backoff[] */
263 
264 /*
265  * TCP timer processing.
266  *
267  * Each connection has 5 timers associated with it, which can be scheduled
268  * simultaneously.  They all are serviced by one callout tcp_timer_enter().
269  * This function executes the next timer via tcp_timersw[] vector.  Each
270  * timer is supposed to return 'true' unless the connection was destroyed.
271  * In the former case tcp_timer_enter() will schedule callout for next timer.
272  */
273 
274 typedef bool tcp_timer_t(struct tcpcb *);
275 static tcp_timer_t tcp_timer_delack;
276 static tcp_timer_t tcp_timer_2msl;
277 static tcp_timer_t tcp_timer_keep;
278 static tcp_timer_t tcp_timer_persist;
279 static tcp_timer_t tcp_timer_rexmt;
280 
281 static tcp_timer_t * const tcp_timersw[TT_N] = {
282 	[TT_DELACK] = tcp_timer_delack,
283 	[TT_REXMT] = tcp_timer_rexmt,
284 	[TT_PERSIST] = tcp_timer_persist,
285 	[TT_KEEP] = tcp_timer_keep,
286 	[TT_2MSL] = tcp_timer_2msl,
287 };
288 
289 /*
290  * tcp_output_locked() s a timer specific variation of call to tcp_output(),
291  * see tcp_var.h for the rest.  It handles drop request from advanced stacks,
292  * but keeps tcpcb locked unless tcp_drop() destroyed it.
293  * Returns true if tcpcb is valid and locked.
294  */
295 static inline bool
296 tcp_output_locked(struct tcpcb *tp)
297 {
298 	int rv;
299 
300 	INP_WLOCK_ASSERT(tptoinpcb(tp));
301 
302 	if ((rv = tp->t_fb->tfb_tcp_output(tp)) < 0) {
303 		KASSERT(tp->t_fb->tfb_flags & TCP_FUNC_OUTPUT_CANDROP,
304 		    ("TCP stack %s requested tcp_drop(%p)",
305 		    tp->t_fb->tfb_tcp_block_name, tp));
306 		tp = tcp_drop(tp, rv);
307 	}
308 
309 	return (tp != NULL);
310 }
311 
312 static bool
313 tcp_timer_delack(struct tcpcb *tp)
314 {
315 	struct epoch_tracker et;
316 #if defined(INVARIANTS) || defined(VIMAGE)
317 	struct inpcb *inp = tptoinpcb(tp);
318 #endif
319 	bool rv;
320 
321 	INP_WLOCK_ASSERT(inp);
322 
323 	CURVNET_SET(inp->inp_vnet);
324 	tp->t_flags |= TF_ACKNOW;
325 	TCPSTAT_INC(tcps_delack);
326 	NET_EPOCH_ENTER(et);
327 	rv = tcp_output_locked(tp);
328 	NET_EPOCH_EXIT(et);
329 	CURVNET_RESTORE();
330 
331 	return (rv);
332 }
333 
334 static bool
335 tcp_timer_2msl(struct tcpcb *tp)
336 {
337 	struct inpcb *inp = tptoinpcb(tp);
338 	bool close = false;
339 
340 	INP_WLOCK_ASSERT(inp);
341 
342 	TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
343 	CURVNET_SET(inp->inp_vnet);
344 	tcp_log_end_status(tp, TCP_EI_STATUS_2MSL);
345 	tcp_free_sackholes(tp);
346 	/*
347 	 * 2 MSL timeout in shutdown went off.  If we're closed but
348 	 * still waiting for peer to close and connection has been idle
349 	 * too long delete connection control block.  Otherwise, check
350 	 * again in a bit.
351 	 *
352 	 * If fastrecycle of FIN_WAIT_2, in FIN_WAIT_2 and receiver has closed,
353 	 * there's no point in hanging onto FIN_WAIT_2 socket. Just close it.
354 	 * Ignore fact that there were recent incoming segments.
355 	 *
356 	 * XXXGL: check if inp_socket shall always be !NULL here?
357 	 */
358 	if (tp->t_state == TCPS_TIME_WAIT) {
359 		close = true;
360 	} else if (tp->t_state == TCPS_FIN_WAIT_2 &&
361 	    tcp_fast_finwait2_recycle && inp->inp_socket &&
362 	    (inp->inp_socket->so_rcv.sb_state & SBS_CANTRCVMORE)) {
363 		TCPSTAT_INC(tcps_finwait2_drops);
364 		close = true;
365 	} else {
366 		if (ticks - tp->t_rcvtime <= TP_MAXIDLE(tp))
367 			tcp_timer_activate(tp, TT_2MSL, TP_KEEPINTVL(tp));
368 		else
369 			close = true;
370 	}
371 	if (close) {
372 		struct epoch_tracker et;
373 
374 		NET_EPOCH_ENTER(et);
375 		tp = tcp_close(tp);
376 		NET_EPOCH_EXIT(et);
377 	}
378 	CURVNET_RESTORE();
379 
380 	return (tp != NULL);
381 }
382 
383 static bool
384 tcp_timer_keep(struct tcpcb *tp)
385 {
386 	struct epoch_tracker et;
387 	struct inpcb *inp = tptoinpcb(tp);
388 	struct tcptemp *t_template;
389 
390 	INP_WLOCK_ASSERT(inp);
391 
392 	TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
393 	CURVNET_SET(inp->inp_vnet);
394 	/*
395 	 * Because we don't regularly reset the keepalive callout in
396 	 * the ESTABLISHED state, it may be that we don't actually need
397 	 * to send a keepalive yet. If that occurs, schedule another
398 	 * call for the next time the keepalive timer might expire.
399 	 */
400 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
401 		u_int idletime;
402 
403 		idletime = ticks - tp->t_rcvtime;
404 		if (idletime < TP_KEEPIDLE(tp)) {
405 			tcp_timer_activate(tp, TT_KEEP,
406 			    TP_KEEPIDLE(tp) - idletime);
407 			CURVNET_RESTORE();
408 			return (true);
409 		}
410 	}
411 
412 	/*
413 	 * Keep-alive timer went off; send something
414 	 * or drop connection if idle for too long.
415 	 */
416 	TCPSTAT_INC(tcps_keeptimeo);
417 	if (tp->t_state < TCPS_ESTABLISHED)
418 		goto dropit;
419 	if ((V_tcp_always_keepalive ||
420 	    inp->inp_socket->so_options & SO_KEEPALIVE) &&
421 	    tp->t_state <= TCPS_CLOSING) {
422 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
423 			goto dropit;
424 		/*
425 		 * Send a packet designed to force a response
426 		 * if the peer is up and reachable:
427 		 * either an ACK if the connection is still alive,
428 		 * or an RST if the peer has closed the connection
429 		 * due to timeout or reboot.
430 		 * Using sequence number tp->snd_una-1
431 		 * causes the transmitted zero-length segment
432 		 * to lie outside the receive window;
433 		 * by the protocol spec, this requires the
434 		 * correspondent TCP to respond.
435 		 */
436 		TCPSTAT_INC(tcps_keepprobe);
437 		t_template = tcpip_maketemplate(inp);
438 		if (t_template) {
439 			NET_EPOCH_ENTER(et);
440 			tcp_respond(tp, t_template->tt_ipgen,
441 				    &t_template->tt_t, (struct mbuf *)NULL,
442 				    tp->rcv_nxt, tp->snd_una - 1, 0);
443 			NET_EPOCH_EXIT(et);
444 			free(t_template, M_TEMP);
445 		}
446 		tcp_timer_activate(tp, TT_KEEP, TP_KEEPINTVL(tp));
447 	} else
448 		tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp));
449 
450 	CURVNET_RESTORE();
451 	return (true);
452 
453 dropit:
454 	TCPSTAT_INC(tcps_keepdrops);
455 	NET_EPOCH_ENTER(et);
456 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
457 	tp = tcp_drop(tp, ETIMEDOUT);
458 	NET_EPOCH_EXIT(et);
459 	CURVNET_RESTORE();
460 
461 	return (tp != NULL);
462 }
463 
464 /*
465  * Has this session exceeded the maximum time without seeing a substantive
466  * acknowledgement? If so, return true; otherwise false.
467  */
468 static bool
469 tcp_maxunacktime_check(struct tcpcb *tp)
470 {
471 
472 	/* Are we tracking this timer for this session? */
473 	if (TP_MAXUNACKTIME(tp) == 0)
474 		return false;
475 
476 	/* Do we have a current measurement. */
477 	if (tp->t_acktime == 0)
478 		return false;
479 
480 	/* Are we within the acceptable range? */
481 	if (TSTMP_GT(TP_MAXUNACKTIME(tp) + tp->t_acktime, (u_int)ticks))
482 		return false;
483 
484 	/* We exceeded the timer. */
485 	TCPSTAT_INC(tcps_progdrops);
486 	return true;
487 }
488 
489 static bool
490 tcp_timer_persist(struct tcpcb *tp)
491 {
492 	struct epoch_tracker et;
493 #if defined(INVARIANTS) || defined(VIMAGE)
494 	struct inpcb *inp = tptoinpcb(tp);
495 #endif
496 	bool progdrop, rv;
497 
498 	INP_WLOCK_ASSERT(inp);
499 
500 	TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
501 	CURVNET_SET(inp->inp_vnet);
502 	/*
503 	 * Persistence timer into zero window.
504 	 * Force a byte to be output, if possible.
505 	 */
506 	TCPSTAT_INC(tcps_persisttimeo);
507 	/*
508 	 * Hack: if the peer is dead/unreachable, we do not
509 	 * time out if the window is closed.  After a full
510 	 * backoff, drop the connection if the idle time
511 	 * (no responses to probes) reaches the maximum
512 	 * backoff that we would use if retransmitting.
513 	 * Also, drop the connection if we haven't been making
514 	 * progress.
515 	 */
516 	progdrop = tcp_maxunacktime_check(tp);
517 	if (progdrop || (tp->t_rxtshift >= V_tcp_retries &&
518 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
519 	     ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff))) {
520 		if (!progdrop)
521 			TCPSTAT_INC(tcps_persistdrop);
522 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
523 		goto dropit;
524 	}
525 	/*
526 	 * If the user has closed the socket then drop a persisting
527 	 * connection after a much reduced timeout.
528 	 */
529 	if (tp->t_state > TCPS_CLOSE_WAIT &&
530 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
531 		TCPSTAT_INC(tcps_persistdrop);
532 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
533 		goto dropit;
534 	}
535 	tcp_setpersist(tp);
536 	tp->t_flags |= TF_FORCEDATA;
537 	NET_EPOCH_ENTER(et);
538 	if ((rv = tcp_output_locked(tp)))
539 		tp->t_flags &= ~TF_FORCEDATA;
540 	NET_EPOCH_EXIT(et);
541 	CURVNET_RESTORE();
542 
543 	return (rv);
544 
545 dropit:
546 	NET_EPOCH_ENTER(et);
547 	tp = tcp_drop(tp, ETIMEDOUT);
548 	NET_EPOCH_EXIT(et);
549 	CURVNET_RESTORE();
550 
551 	return (tp != NULL);
552 }
553 
554 static bool
555 tcp_timer_rexmt(struct tcpcb *tp)
556 {
557 	struct epoch_tracker et;
558 	struct inpcb *inp = tptoinpcb(tp);
559 	int rexmt;
560 	bool isipv6, rv;
561 
562 	INP_WLOCK_ASSERT(inp);
563 
564 	TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
565 	CURVNET_SET(inp->inp_vnet);
566 	tcp_free_sackholes(tp);
567 	if (tp->t_fb->tfb_tcp_rexmit_tmr) {
568 		/* The stack has a timer action too. */
569 		(*tp->t_fb->tfb_tcp_rexmit_tmr)(tp);
570 	}
571 	/*
572 	 * Retransmission timer went off.  Message has not
573 	 * been acked within retransmit interval.  Back off
574 	 * to a longer retransmit interval and retransmit one segment.
575 	 *
576 	 * If we've either exceeded the maximum number of retransmissions,
577 	 * or we've gone long enough without making progress, then drop
578 	 * the session.
579 	 */
580 	if (++tp->t_rxtshift > V_tcp_retries || tcp_maxunacktime_check(tp)) {
581 		if (tp->t_rxtshift > V_tcp_retries)
582 			TCPSTAT_INC(tcps_timeoutdrop);
583 		tp->t_rxtshift = V_tcp_retries;
584 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
585 		NET_EPOCH_ENTER(et);
586 		tp = tcp_drop(tp, ETIMEDOUT);
587 		NET_EPOCH_EXIT(et);
588 		CURVNET_RESTORE();
589 
590 		return (tp != NULL);
591 	}
592 	if (tp->t_state == TCPS_SYN_SENT) {
593 		/*
594 		 * If the SYN was retransmitted, indicate CWND to be
595 		 * limited to 1 segment in cc_conn_init().
596 		 */
597 		tp->snd_cwnd = 1;
598 	} else if (tp->t_rxtshift == 1) {
599 		/*
600 		 * first retransmit; record ssthresh and cwnd so they can
601 		 * be recovered if this turns out to be a "bad" retransmit.
602 		 * A retransmit is considered "bad" if an ACK for this
603 		 * segment is received within RTT/2 interval; the assumption
604 		 * here is that the ACK was already in flight.  See
605 		 * "On Estimating End-to-End Network Path Properties" by
606 		 * Allman and Paxson for more details.
607 		 */
608 		tp->snd_cwnd_prev = tp->snd_cwnd;
609 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
610 		tp->snd_recover_prev = tp->snd_recover;
611 		if (IN_FASTRECOVERY(tp->t_flags))
612 			tp->t_flags |= TF_WASFRECOVERY;
613 		else
614 			tp->t_flags &= ~TF_WASFRECOVERY;
615 		if (IN_CONGRECOVERY(tp->t_flags))
616 			tp->t_flags |= TF_WASCRECOVERY;
617 		else
618 			tp->t_flags &= ~TF_WASCRECOVERY;
619 		if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
620 			tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
621 		/* In the event that we've negotiated timestamps
622 		 * badrxtwin will be set to the value that we set
623 		 * the retransmitted packet's to_tsval to by tcp_output
624 		 */
625 		tp->t_flags |= TF_PREVVALID;
626 	} else
627 		tp->t_flags &= ~TF_PREVVALID;
628 	TCPSTAT_INC(tcps_rexmttimeo);
629 	if ((tp->t_state == TCPS_SYN_SENT) ||
630 	    (tp->t_state == TCPS_SYN_RECEIVED))
631 		rexmt = tcp_rexmit_initial * tcp_backoff[tp->t_rxtshift];
632 	else
633 		rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
634 	TCPT_RANGESET(tp->t_rxtcur, rexmt,
635 		      tp->t_rttmin, TCPTV_REXMTMAX);
636 
637 	/*
638 	 * We enter the path for PLMTUD if connection is established or, if
639 	 * connection is FIN_WAIT_1 status, reason for the last is that if
640 	 * amount of data we send is very small, we could send it in couple of
641 	 * packets and process straight to FIN. In that case we won't catch
642 	 * ESTABLISHED state.
643 	 */
644 #ifdef INET6
645 	isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
646 #else
647 	isipv6 = false;
648 #endif
649 	if (((V_tcp_pmtud_blackhole_detect == 1) ||
650 	    (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
651 	    (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
652 	    ((tp->t_state == TCPS_ESTABLISHED) ||
653 	    (tp->t_state == TCPS_FIN_WAIT_1))) {
654 		if (tp->t_rxtshift == 1) {
655 			/*
656 			 * We enter blackhole detection after the first
657 			 * unsuccessful timer based retransmission.
658 			 * Then we reduce up to two times the MSS, each
659 			 * candidate giving two tries of retransmissions.
660 			 * But we give a candidate only two tries, if it
661 			 * actually reduces the MSS.
662 			 */
663 			tp->t_blackhole_enter = 2;
664 			tp->t_blackhole_exit = tp->t_blackhole_enter;
665 			if (isipv6) {
666 #ifdef INET6
667 				if (tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss)
668 					tp->t_blackhole_exit += 2;
669 				if (tp->t_maxseg > V_tcp_v6mssdflt &&
670 				    V_tcp_v6pmtud_blackhole_mss > V_tcp_v6mssdflt)
671 					tp->t_blackhole_exit += 2;
672 #endif
673 			} else {
674 #ifdef INET
675 				if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss)
676 					tp->t_blackhole_exit += 2;
677 				if (tp->t_maxseg > V_tcp_mssdflt &&
678 				    V_tcp_pmtud_blackhole_mss > V_tcp_mssdflt)
679 					tp->t_blackhole_exit += 2;
680 #endif
681 			}
682 		}
683 		if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD|TF2_PLPMTU_MAXSEGSNT)) ==
684 		    (TF2_PLPMTU_PMTUD|TF2_PLPMTU_MAXSEGSNT)) &&
685 		    (tp->t_rxtshift >= tp->t_blackhole_enter &&
686 		    tp->t_rxtshift < tp->t_blackhole_exit &&
687 		    (tp->t_rxtshift - tp->t_blackhole_enter) % 2 == 0)) {
688 			/*
689 			 * Enter Path MTU Black-hole Detection mechanism:
690 			 * - Disable Path MTU Discovery (IP "DF" bit).
691 			 * - Reduce MTU to lower value than what we
692 			 *   negotiated with peer.
693 			 */
694 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
695 				/* Record that we may have found a black hole. */
696 				tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
697 				/* Keep track of previous MSS. */
698 				tp->t_pmtud_saved_maxseg = tp->t_maxseg;
699 			}
700 
701 			/*
702 			 * Reduce the MSS to blackhole value or to the default
703 			 * in an attempt to retransmit.
704 			 */
705 #ifdef INET6
706 			if (isipv6 &&
707 			    tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss &&
708 			    V_tcp_v6pmtud_blackhole_mss > V_tcp_v6mssdflt) {
709 				/* Use the sysctl tuneable blackhole MSS. */
710 				tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
711 				TCPSTAT_INC(tcps_pmtud_blackhole_activated);
712 			} else if (isipv6) {
713 				/* Use the default MSS. */
714 				tp->t_maxseg = V_tcp_v6mssdflt;
715 				/*
716 				 * Disable Path MTU Discovery when we switch to
717 				 * minmss.
718 				 */
719 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
720 				TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
721 			}
722 #endif
723 #if defined(INET6) && defined(INET)
724 			else
725 #endif
726 #ifdef INET
727 			if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss &&
728 			    V_tcp_pmtud_blackhole_mss > V_tcp_mssdflt) {
729 				/* Use the sysctl tuneable blackhole MSS. */
730 				tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
731 				TCPSTAT_INC(tcps_pmtud_blackhole_activated);
732 			} else {
733 				/* Use the default MSS. */
734 				tp->t_maxseg = V_tcp_mssdflt;
735 				/*
736 				 * Disable Path MTU Discovery when we switch to
737 				 * minmss.
738 				 */
739 				tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
740 				TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
741 			}
742 #endif
743 			/*
744 			 * Reset the slow-start flight size
745 			 * as it may depend on the new MSS.
746 			 */
747 			if (CC_ALGO(tp)->conn_init != NULL)
748 				CC_ALGO(tp)->conn_init(&tp->t_ccv);
749 		} else {
750 			/*
751 			 * If further retransmissions are still unsuccessful
752 			 * with a lowered MTU, maybe this isn't a blackhole and
753 			 * we restore the previous MSS and blackhole detection
754 			 * flags.
755 			 */
756 			if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
757 			    (tp->t_rxtshift >= tp->t_blackhole_exit)) {
758 				tp->t_flags2 |= TF2_PLPMTU_PMTUD;
759 				tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
760 				tp->t_maxseg = tp->t_pmtud_saved_maxseg;
761 				TCPSTAT_INC(tcps_pmtud_blackhole_failed);
762 				/*
763 				 * Reset the slow-start flight size as it
764 				 * may depend on the new MSS.
765 				 */
766 				if (CC_ALGO(tp)->conn_init != NULL)
767 					CC_ALGO(tp)->conn_init(&tp->t_ccv);
768 			}
769 		}
770 	}
771 
772 	/*
773 	 * Disable RFC1323 and SACK if we haven't got any response to
774 	 * our third SYN to work-around some broken terminal servers
775 	 * (most of which have hopefully been retired) that have bad VJ
776 	 * header compression code which trashes TCP segments containing
777 	 * unknown-to-them TCP options.
778 	 */
779 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
780 	    (tp->t_rxtshift == 3))
781 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
782 	/*
783 	 * If we backed off this far, notify the L3 protocol that we're having
784 	 * connection problems.
785 	 */
786 	if (tp->t_rxtshift > TCP_RTT_INVALIDATE) {
787 #ifdef INET6
788 		if ((inp->inp_vflag & INP_IPV6) != 0)
789 			in6_losing(inp);
790 		else
791 #endif
792 			in_losing(inp);
793 	}
794 	tp->snd_nxt = tp->snd_una;
795 	tp->snd_recover = tp->snd_max;
796 	/*
797 	 * Force a segment to be sent.
798 	 */
799 	tp->t_flags |= TF_ACKNOW;
800 	/*
801 	 * If timing a segment in this window, stop the timer.
802 	 */
803 	tp->t_rtttime = 0;
804 
805 	cc_cong_signal(tp, NULL, CC_RTO);
806 	NET_EPOCH_ENTER(et);
807 	rv = tcp_output_locked(tp);
808 	NET_EPOCH_EXIT(et);
809 	CURVNET_RESTORE();
810 
811 	return (rv);
812 }
813 
814 static void
815 tcp_bblog_timer(struct tcpcb *tp, tt_which which, tt_what what, uint32_t ticks)
816 {
817 	struct tcp_log_buffer *lgb;
818 	uint64_t ms;
819 
820 	INP_WLOCK_ASSERT(tptoinpcb(tp));
821 	if (tcp_bblogging_on(tp))
822 		lgb = tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_RTO, 0, 0,
823 		    NULL, false, NULL, NULL, 0, NULL);
824 	else
825 		lgb = NULL;
826 	if (lgb != NULL) {
827 		lgb->tlb_flex1 = (what << 8) | which;
828 		if (what == TT_STARTING) {
829 			/* Convert ticks to ms and store it in tlb_flex2. */
830 			if (hz == 1000)
831 				lgb->tlb_flex2 = ticks;
832 			else {
833 				ms = (((uint64_t)ticks * 1000) + (hz - 1)) / hz;
834 				if (ms > UINT32_MAX)
835 					lgb->tlb_flex2 = UINT32_MAX;
836 				else
837 					lgb->tlb_flex2 = (uint32_t)ms;
838 			}
839 		}
840 	}
841 }
842 
843 static inline tt_which
844 tcp_timer_next(struct tcpcb *tp, sbintime_t *precision)
845 {
846 	tt_which i, rv;
847 	sbintime_t after, before;
848 
849 	for (i = 0, rv = TT_N, after = before = SBT_MAX; i < TT_N; i++) {
850 		if (tp->t_timers[i] < after) {
851 			after = tp->t_timers[i];
852 			rv = i;
853 		}
854 		before = MIN(before, tp->t_timers[i] + tp->t_precisions[i]);
855 	}
856 	if (precision != NULL)
857 		*precision = before - after;
858 
859 	return (rv);
860 }
861 
862 static void
863 tcp_timer_enter(void *xtp)
864 {
865 	struct tcpcb *tp = xtp;
866 	struct inpcb *inp = tptoinpcb(tp);
867 	sbintime_t precision;
868 	tt_which which;
869 	bool tp_valid;
870 
871 	INP_WLOCK_ASSERT(inp);
872 	MPASS((curthread->td_pflags & TDP_INTCPCALLOUT) == 0);
873 
874 	curthread->td_pflags |= TDP_INTCPCALLOUT;
875 
876 	which = tcp_timer_next(tp, NULL);
877 	MPASS(which < TT_N);
878 	tp->t_timers[which] = SBT_MAX;
879 	tp->t_precisions[which] = 0;
880 
881 	tcp_bblog_timer(tp, which, TT_PROCESSING, 0);
882 	tp_valid = tcp_timersw[which](tp);
883 	if (tp_valid) {
884 		tcp_bblog_timer(tp, which, TT_PROCESSED, 0);
885 		if ((which = tcp_timer_next(tp, &precision)) != TT_N) {
886 			callout_reset_sbt_on(&tp->t_callout,
887 			    tp->t_timers[which], precision, tcp_timer_enter,
888 			    tp, inp_to_cpuid(inp), C_ABSOLUTE);
889 		}
890 		INP_WUNLOCK(inp);
891 	}
892 
893 	curthread->td_pflags &= ~TDP_INTCPCALLOUT;
894 }
895 
896 /*
897  * Activate or stop (delta == 0) a TCP timer.
898  */
899 void
900 tcp_timer_activate(struct tcpcb *tp, tt_which which, u_int delta)
901 {
902 	struct inpcb *inp = tptoinpcb(tp);
903 	sbintime_t precision;
904 	tt_what what;
905 
906 #ifdef TCP_OFFLOAD
907 	if (tp->t_flags & TF_TOE)
908 		return;
909 #endif
910 
911 	INP_WLOCK_ASSERT(inp);
912 
913 	if (delta > 0) {
914 		what = TT_STARTING;
915 		callout_when(tick_sbt * delta, 0, C_HARDCLOCK,
916 		    &tp->t_timers[which], &tp->t_precisions[which]);
917 	} else {
918 		what = TT_STOPPING;
919 		tp->t_timers[which] = SBT_MAX;
920 	}
921 	tcp_bblog_timer(tp, which, what, delta);
922 
923 	if ((which = tcp_timer_next(tp, &precision)) != TT_N)
924 		callout_reset_sbt_on(&tp->t_callout, tp->t_timers[which],
925 		    precision, tcp_timer_enter, tp, inp_to_cpuid(inp),
926 		    C_ABSOLUTE);
927 	else
928 		callout_stop(&tp->t_callout);
929 }
930 
931 bool
932 tcp_timer_active(struct tcpcb *tp, tt_which which)
933 {
934 
935 	INP_WLOCK_ASSERT(tptoinpcb(tp));
936 
937 	return (tp->t_timers[which] != SBT_MAX);
938 }
939 
940 /*
941  * Stop all timers associated with tcpcb.
942  *
943  * Called only on tcpcb destruction.  The tcpcb shall already be dropped from
944  * the pcb lookup database and socket is not losing the last reference.
945  *
946  * XXXGL: unfortunately our callout(9) is not able to fully stop a locked
947  * callout even when only two threads are involved: the callout itself and the
948  * thread that does callout_stop().  See where softclock_call_cc() swaps the
949  * callwheel lock to callout lock and then checks cc_exec_cancel().  This is
950  * the race window.  If it happens, the tcp_timer_enter() won't be executed,
951  * however pcb lock will be locked and released, hence we can't free memory.
952  * Until callout(9) is improved, just keep retrying.  In my profiling I've seen
953  * such event happening less than 1 time per hour with 20-30 Gbit/s of traffic.
954  */
955 void
956 tcp_timer_stop(struct tcpcb *tp)
957 {
958 	struct inpcb *inp = tptoinpcb(tp);
959 
960 	INP_WLOCK_ASSERT(inp);
961 
962 	if (curthread->td_pflags & TDP_INTCPCALLOUT) {
963 		int stopped __diagused;
964 
965 		stopped = callout_stop(&tp->t_callout);
966 		MPASS(stopped == 0);
967 	} else while(__predict_false(callout_stop(&tp->t_callout) == 0)) {
968 		INP_WUNLOCK(inp);
969 		kern_yield(PRI_UNCHANGED);
970 		INP_WLOCK(inp);
971 	}
972 }
973