xref: /freebsd/sys/netinet/tcp_timer.c (revision 7a7741af18d6c8a804cc643cb7ecda9d730c6aa6)
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 
32 #include <sys/cdefs.h>
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_rss.h"
36 
37 #include <sys/param.h>
38 #include <sys/kernel.h>
39 #include <sys/lock.h>
40 #include <sys/mbuf.h>
41 #include <sys/mutex.h>
42 #include <sys/protosw.h>
43 #include <sys/smp.h>
44 #include <sys/socket.h>
45 #include <sys/socketvar.h>
46 #include <sys/sysctl.h>
47 #include <sys/systm.h>
48 
49 #include <net/if.h>
50 #include <net/route.h>
51 #include <net/rss_config.h>
52 #include <net/vnet.h>
53 #include <net/netisr.h>
54 
55 #include <netinet/in.h>
56 #include <netinet/in_kdtrace.h>
57 #include <netinet/in_pcb.h>
58 #include <netinet/in_rss.h>
59 #include <netinet/in_systm.h>
60 #ifdef INET6
61 #include <netinet6/in6_pcb.h>
62 #endif
63 #include <netinet/ip_var.h>
64 #include <netinet/tcp.h>
65 #include <netinet/tcp_fsm.h>
66 #include <netinet/tcp_timer.h>
67 #include <netinet/tcp_var.h>
68 #include <netinet/tcp_log_buf.h>
69 #include <netinet/tcp_seq.h>
70 #include <netinet/cc/cc.h>
71 #ifdef INET6
72 #include <netinet6/tcp6_var.h>
73 #endif
74 #include <netinet/tcpip.h>
75 
76 int    tcp_persmin;
77 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, persmin,
78     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
79     &tcp_persmin, 0, sysctl_msec_to_ticks, "I",
80     "minimum persistence interval");
81 
82 int    tcp_persmax;
83 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, persmax,
84     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
85     &tcp_persmax, 0, sysctl_msec_to_ticks, "I",
86     "maximum persistence interval");
87 
88 int	tcp_keepinit;
89 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINIT, keepinit,
90     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
91     &tcp_keepinit, 0, sysctl_msec_to_ticks, "I",
92     "time to establish connection");
93 
94 int	tcp_keepidle;
95 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPIDLE, keepidle,
96     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
97     &tcp_keepidle, 0, sysctl_msec_to_ticks, "I",
98     "time before keepalive probes begin");
99 
100 int	tcp_keepintvl;
101 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINTVL, keepintvl,
102     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
103     &tcp_keepintvl, 0, sysctl_msec_to_ticks, "I",
104     "time between keepalive probes");
105 
106 int	tcp_delacktime;
107 SYSCTL_PROC(_net_inet_tcp, TCPCTL_DELACKTIME, delacktime,
108     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
109     &tcp_delacktime, 0, sysctl_msec_to_ticks, "I",
110     "Time before a delayed ACK is sent");
111 
112 VNET_DEFINE(int, tcp_msl);
113 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, msl,
114     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_VNET,
115     &VNET_NAME(tcp_msl), 0, sysctl_msec_to_ticks, "I",
116     "Maximum segment lifetime");
117 
118 int	tcp_rexmit_initial;
119 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_initial,
120    CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
121     &tcp_rexmit_initial, 0, sysctl_msec_to_ticks, "I",
122     "Initial Retransmission Timeout");
123 
124 int	tcp_rexmit_min;
125 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_min,
126     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
127     &tcp_rexmit_min, 0, sysctl_msec_to_ticks, "I",
128     "Minimum Retransmission Timeout");
129 
130 int	tcp_rexmit_slop;
131 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_slop,
132     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
133     &tcp_rexmit_slop, 0, sysctl_msec_to_ticks, "I",
134     "Retransmission Timer Slop");
135 
136 VNET_DEFINE(int, tcp_always_keepalive) = 1;
137 SYSCTL_INT(_net_inet_tcp, OID_AUTO, always_keepalive, CTLFLAG_VNET|CTLFLAG_RW,
138     &VNET_NAME(tcp_always_keepalive) , 0,
139     "Assume SO_KEEPALIVE on all TCP connections");
140 
141 int    tcp_fast_finwait2_recycle = 0;
142 SYSCTL_INT(_net_inet_tcp, OID_AUTO, fast_finwait2_recycle, CTLFLAG_RW,
143     &tcp_fast_finwait2_recycle, 0,
144     "Recycle closed FIN_WAIT_2 connections faster");
145 
146 int    tcp_finwait2_timeout;
147 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, finwait2_timeout,
148     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
149     &tcp_finwait2_timeout, 0, sysctl_msec_to_ticks, "I",
150     "FIN-WAIT2 timeout");
151 
152 int	tcp_keepcnt = TCPTV_KEEPCNT;
153 SYSCTL_INT(_net_inet_tcp, OID_AUTO, keepcnt, CTLFLAG_RW, &tcp_keepcnt, 0,
154     "Number of keepalive probes to send");
155 
156 	/* max idle probes */
157 int	tcp_maxpersistidle;
158 
159 int	tcp_rexmit_drop_options = 0;
160 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rexmit_drop_options, CTLFLAG_RW,
161     &tcp_rexmit_drop_options, 0,
162     "Drop TCP options from 3rd and later retransmitted SYN");
163 
164 int	tcp_maxunacktime = TCPTV_MAXUNACKTIME;
165 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, maxunacktime,
166     CTLTYPE_INT|CTLFLAG_RW | CTLFLAG_NEEDGIANT,
167     &tcp_maxunacktime, 0, sysctl_msec_to_ticks, "I",
168     "Maximum time (in ms) that a session can linger without making progress");
169 
170 VNET_DEFINE(int, tcp_pmtud_blackhole_detect);
171 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_detection,
172     CTLFLAG_RW|CTLFLAG_VNET,
173     &VNET_NAME(tcp_pmtud_blackhole_detect), 0,
174     "Path MTU Discovery Black Hole Detection Enabled");
175 
176 #ifdef INET
177 VNET_DEFINE(int, tcp_pmtud_blackhole_mss) = 1200;
178 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_mss,
179     CTLFLAG_RW|CTLFLAG_VNET,
180     &VNET_NAME(tcp_pmtud_blackhole_mss), 0,
181     "Path MTU Discovery Black Hole Detection lowered MSS");
182 #endif
183 
184 #ifdef INET6
185 VNET_DEFINE(int, tcp_v6pmtud_blackhole_mss) = 1220;
186 SYSCTL_INT(_net_inet_tcp, OID_AUTO, v6pmtud_blackhole_mss,
187     CTLFLAG_RW|CTLFLAG_VNET,
188     &VNET_NAME(tcp_v6pmtud_blackhole_mss), 0,
189     "Path MTU Discovery IPv6 Black Hole Detection lowered MSS");
190 #endif
191 
192 #ifdef	RSS
193 static int	per_cpu_timers = 1;
194 #else
195 static int	per_cpu_timers = 0;
196 #endif
197 SYSCTL_INT(_net_inet_tcp, OID_AUTO, per_cpu_timers, CTLFLAG_RW,
198     &per_cpu_timers , 0, "run tcp timers on all cpus");
199 
200 static int
201 sysctl_net_inet_tcp_retries(SYSCTL_HANDLER_ARGS)
202 {
203 	int error, new;
204 
205 	new = V_tcp_retries;
206 	error = sysctl_handle_int(oidp, &new, 0, req);
207 	if (error == 0 && req->newptr) {
208 		if ((new < 1) || (new > TCP_MAXRXTSHIFT))
209 			error = EINVAL;
210 		else
211 			V_tcp_retries = new;
212 	}
213 	return (error);
214 }
215 
216 VNET_DEFINE(int, tcp_retries) = TCP_MAXRXTSHIFT;
217 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, retries,
218     CTLTYPE_INT | CTLFLAG_VNET | CTLFLAG_RW,
219     &VNET_NAME(tcp_retries), 0, sysctl_net_inet_tcp_retries, "I",
220     "maximum number of consecutive timer based retransmissions");
221 
222 /*
223  * Map the given inp to a CPU id.
224  *
225  * This queries RSS if it's compiled in, else it defaults to the current
226  * CPU ID.
227  */
228 inline int
229 inp_to_cpuid(struct inpcb *inp)
230 {
231 	u_int cpuid;
232 
233 	if (per_cpu_timers) {
234 #ifdef	RSS
235 		cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
236 		if (cpuid == NETISR_CPUID_NONE)
237 			return (curcpu);	/* XXX */
238 		else
239 			return (cpuid);
240 #endif
241 		/*
242 		 * We don't have a flowid -> cpuid mapping, so cheat and
243 		 * just map unknown cpuids to curcpu.  Not the best, but
244 		 * apparently better than defaulting to swi 0.
245 		 */
246 		cpuid = inp->inp_flowid % (mp_maxid + 1);
247 		if (! CPU_ABSENT(cpuid))
248 			return (cpuid);
249 		return (curcpu);
250 	} else {
251 		return (0);
252 	}
253 }
254 
255 int	tcp_backoff[TCP_MAXRXTSHIFT + 1] =
256     { 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 512, 512, 512 };
257 
258 int tcp_totbackoff = 2559;	/* sum of tcp_backoff[] */
259 
260 /*
261  * TCP timer processing.
262  *
263  * Each connection has 5 timers associated with it, which can be scheduled
264  * simultaneously.  They all are serviced by one callout tcp_timer_enter().
265  * This function executes the next timer via tcp_timersw[] vector.  Each
266  * timer is supposed to return 'true' unless the connection was destroyed.
267  * In the former case tcp_timer_enter() will schedule callout for next timer.
268  */
269 
270 typedef bool tcp_timer_t(struct tcpcb *);
271 static tcp_timer_t tcp_timer_delack;
272 static tcp_timer_t tcp_timer_2msl;
273 static tcp_timer_t tcp_timer_keep;
274 static tcp_timer_t tcp_timer_persist;
275 static tcp_timer_t tcp_timer_rexmt;
276 
277 static tcp_timer_t * const tcp_timersw[TT_N] = {
278 	[TT_DELACK] = tcp_timer_delack,
279 	[TT_REXMT] = tcp_timer_rexmt,
280 	[TT_PERSIST] = tcp_timer_persist,
281 	[TT_KEEP] = tcp_timer_keep,
282 	[TT_2MSL] = tcp_timer_2msl,
283 };
284 
285 /*
286  * tcp_output_locked() s a timer specific variation of call to tcp_output(),
287  * see tcp_var.h for the rest.  It handles drop request from advanced stacks,
288  * but keeps tcpcb locked unless tcp_drop() destroyed it.
289  * Returns true if tcpcb is valid and locked.
290  */
291 static inline bool
292 tcp_output_locked(struct tcpcb *tp)
293 {
294 	int rv;
295 
296 	INP_WLOCK_ASSERT(tptoinpcb(tp));
297 
298 	if ((rv = tp->t_fb->tfb_tcp_output(tp)) < 0) {
299 		KASSERT(tp->t_fb->tfb_flags & TCP_FUNC_OUTPUT_CANDROP,
300 		    ("TCP stack %s requested tcp_drop(%p)",
301 		    tp->t_fb->tfb_tcp_block_name, tp));
302 		tp = tcp_drop(tp, -rv);
303 	}
304 
305 	return (tp != NULL);
306 }
307 
308 static bool
309 tcp_timer_delack(struct tcpcb *tp)
310 {
311 	struct epoch_tracker et;
312 #if defined(INVARIANTS) || defined(VIMAGE)
313 	struct inpcb *inp = tptoinpcb(tp);
314 #endif
315 	bool rv;
316 
317 	INP_WLOCK_ASSERT(inp);
318 
319 	CURVNET_SET(inp->inp_vnet);
320 	tp->t_flags |= TF_ACKNOW;
321 	TCPSTAT_INC(tcps_delack);
322 	NET_EPOCH_ENTER(et);
323 	rv = tcp_output_locked(tp);
324 	NET_EPOCH_EXIT(et);
325 	CURVNET_RESTORE();
326 
327 	return (rv);
328 }
329 
330 static bool
331 tcp_timer_2msl(struct tcpcb *tp)
332 {
333 	struct inpcb *inp = tptoinpcb(tp);
334 	bool close = false;
335 
336 	INP_WLOCK_ASSERT(inp);
337 
338 	TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
339 	CURVNET_SET(inp->inp_vnet);
340 	tcp_log_end_status(tp, TCP_EI_STATUS_2MSL);
341 	tcp_free_sackholes(tp);
342 	/*
343 	 * 2 MSL timeout in shutdown went off.  If we're closed but
344 	 * still waiting for peer to close and connection has been idle
345 	 * too long delete connection control block.  Otherwise, check
346 	 * again in a bit.
347 	 *
348 	 * If fastrecycle of FIN_WAIT_2, in FIN_WAIT_2 and receiver has closed,
349 	 * there's no point in hanging onto FIN_WAIT_2 socket. Just close it.
350 	 * Ignore fact that there were recent incoming segments.
351 	 *
352 	 * XXXGL: check if inp_socket shall always be !NULL here?
353 	 */
354 	if (tp->t_state == TCPS_TIME_WAIT) {
355 		close = true;
356 	} else if (tp->t_state == TCPS_FIN_WAIT_2 &&
357 	    tcp_fast_finwait2_recycle && inp->inp_socket &&
358 	    (inp->inp_socket->so_rcv.sb_state & SBS_CANTRCVMORE)) {
359 		TCPSTAT_INC(tcps_finwait2_drops);
360 		close = true;
361 	} else {
362 		if (ticks - tp->t_rcvtime <= TP_MAXIDLE(tp))
363 			tcp_timer_activate(tp, TT_2MSL, TP_KEEPINTVL(tp));
364 		else
365 			close = true;
366 	}
367 	if (close) {
368 		struct epoch_tracker et;
369 
370 		NET_EPOCH_ENTER(et);
371 		tp = tcp_close(tp);
372 		NET_EPOCH_EXIT(et);
373 	}
374 	CURVNET_RESTORE();
375 
376 	return (tp != NULL);
377 }
378 
379 static bool
380 tcp_timer_keep(struct tcpcb *tp)
381 {
382 	struct epoch_tracker et;
383 	struct inpcb *inp = tptoinpcb(tp);
384 	struct tcptemp *t_template;
385 
386 	INP_WLOCK_ASSERT(inp);
387 
388 	TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
389 	CURVNET_SET(inp->inp_vnet);
390 	/*
391 	 * Because we don't regularly reset the keepalive callout in
392 	 * the ESTABLISHED state, it may be that we don't actually need
393 	 * to send a keepalive yet. If that occurs, schedule another
394 	 * call for the next time the keepalive timer might expire.
395 	 */
396 	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
397 		u_int idletime;
398 
399 		idletime = ticks - tp->t_rcvtime;
400 		if (idletime < TP_KEEPIDLE(tp)) {
401 			tcp_timer_activate(tp, TT_KEEP,
402 			    TP_KEEPIDLE(tp) - idletime);
403 			CURVNET_RESTORE();
404 			return (true);
405 		}
406 	}
407 
408 	/*
409 	 * Keep-alive timer went off; send something
410 	 * or drop connection if idle for too long.
411 	 */
412 	TCPSTAT_INC(tcps_keeptimeo);
413 	if (tp->t_state < TCPS_ESTABLISHED)
414 		goto dropit;
415 	if ((V_tcp_always_keepalive ||
416 	    inp->inp_socket->so_options & SO_KEEPALIVE) &&
417 	    tp->t_state <= TCPS_CLOSING) {
418 		if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
419 			goto dropit;
420 		/*
421 		 * Send a packet designed to force a response
422 		 * if the peer is up and reachable:
423 		 * either an ACK if the connection is still alive,
424 		 * or an RST if the peer has closed the connection
425 		 * due to timeout or reboot.
426 		 * Using sequence number tp->snd_una-1
427 		 * causes the transmitted zero-length segment
428 		 * to lie outside the receive window;
429 		 * by the protocol spec, this requires the
430 		 * correspondent TCP to respond.
431 		 */
432 		TCPSTAT_INC(tcps_keepprobe);
433 		t_template = tcpip_maketemplate(inp);
434 		if (t_template) {
435 			NET_EPOCH_ENTER(et);
436 			tcp_respond(tp, t_template->tt_ipgen,
437 				    &t_template->tt_t, (struct mbuf *)NULL,
438 				    tp->rcv_nxt, tp->snd_una - 1, 0);
439 			NET_EPOCH_EXIT(et);
440 			free(t_template, M_TEMP);
441 		}
442 		tcp_timer_activate(tp, TT_KEEP, TP_KEEPINTVL(tp));
443 	} else
444 		tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp));
445 
446 	CURVNET_RESTORE();
447 	return (true);
448 
449 dropit:
450 	TCPSTAT_INC(tcps_keepdrops);
451 	NET_EPOCH_ENTER(et);
452 	tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
453 	tp = tcp_drop(tp, ETIMEDOUT);
454 	NET_EPOCH_EXIT(et);
455 	CURVNET_RESTORE();
456 
457 	return (tp != NULL);
458 }
459 
460 /*
461  * Has this session exceeded the maximum time without seeing a substantive
462  * acknowledgement? If so, return true; otherwise false.
463  */
464 static bool
465 tcp_maxunacktime_check(struct tcpcb *tp)
466 {
467 
468 	/* Are we tracking this timer for this session? */
469 	if (TP_MAXUNACKTIME(tp) == 0)
470 		return false;
471 
472 	/* Do we have a current measurement. */
473 	if (tp->t_acktime == 0)
474 		return false;
475 
476 	/* Are we within the acceptable range? */
477 	if (TSTMP_GT(TP_MAXUNACKTIME(tp) + tp->t_acktime, (u_int)ticks))
478 		return false;
479 
480 	/* We exceeded the timer. */
481 	TCPSTAT_INC(tcps_progdrops);
482 	return true;
483 }
484 
485 static bool
486 tcp_timer_persist(struct tcpcb *tp)
487 {
488 	struct epoch_tracker et;
489 #if defined(INVARIANTS) || defined(VIMAGE)
490 	struct inpcb *inp = tptoinpcb(tp);
491 #endif
492 	bool progdrop, rv;
493 
494 	INP_WLOCK_ASSERT(inp);
495 
496 	TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
497 	CURVNET_SET(inp->inp_vnet);
498 	/*
499 	 * Persistence timer into zero window.
500 	 * Force a byte to be output, if possible.
501 	 */
502 	TCPSTAT_INC(tcps_persisttimeo);
503 	/*
504 	 * Hack: if the peer is dead/unreachable, we do not
505 	 * time out if the window is closed.  After a full
506 	 * backoff, drop the connection if the idle time
507 	 * (no responses to probes) reaches the maximum
508 	 * backoff that we would use if retransmitting.
509 	 * Also, drop the connection if we haven't been making
510 	 * progress.
511 	 */
512 	progdrop = tcp_maxunacktime_check(tp);
513 	if (progdrop || (tp->t_rxtshift >= V_tcp_retries &&
514 	    (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
515 	     ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff))) {
516 		if (!progdrop)
517 			TCPSTAT_INC(tcps_persistdrop);
518 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
519 		goto dropit;
520 	}
521 	/*
522 	 * If the user has closed the socket then drop a persisting
523 	 * connection after a much reduced timeout.
524 	 */
525 	if (tp->t_state > TCPS_CLOSE_WAIT &&
526 	    (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
527 		TCPSTAT_INC(tcps_persistdrop);
528 		tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
529 		goto dropit;
530 	}
531 	tcp_setpersist(tp);
532 	tp->t_flags |= TF_FORCEDATA;
533 	NET_EPOCH_ENTER(et);
534 	if ((rv = tcp_output_locked(tp)))
535 		tp->t_flags &= ~TF_FORCEDATA;
536 	NET_EPOCH_EXIT(et);
537 	CURVNET_RESTORE();
538 
539 	return (rv);
540 
541 dropit:
542 	NET_EPOCH_ENTER(et);
543 	tp = tcp_drop(tp, ETIMEDOUT);
544 	NET_EPOCH_EXIT(et);
545 	CURVNET_RESTORE();
546 
547 	return (tp != NULL);
548 }
549 
550 static bool
551 tcp_timer_rexmt(struct tcpcb *tp)
552 {
553 	struct epoch_tracker et;
554 	struct inpcb *inp = tptoinpcb(tp);
555 	int rexmt;
556 	bool isipv6, rv;
557 
558 	INP_WLOCK_ASSERT(inp);
559 
560 	TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
561 	CURVNET_SET(inp->inp_vnet);
562 	if (tp->t_fb->tfb_tcp_rexmit_tmr) {
563 		/* The stack has a timer action too. */
564 		(*tp->t_fb->tfb_tcp_rexmit_tmr)(tp);
565 	}
566 	/*
567 	 * Retransmission timer went off.  Message has not
568 	 * been acked within retransmit interval.  Back off
569 	 * to a longer retransmit interval and retransmit one segment.
570 	 *
571 	 * If we've either exceeded the maximum number of retransmissions,
572 	 * or we've gone long enough without making progress, then drop
573 	 * the session.
574 	 */
575 	if (++tp->t_rxtshift > V_tcp_retries || tcp_maxunacktime_check(tp)) {
576 		if (tp->t_rxtshift > V_tcp_retries)
577 			TCPSTAT_INC(tcps_timeoutdrop);
578 		tp->t_rxtshift = V_tcp_retries;
579 		tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
580 		NET_EPOCH_ENTER(et);
581 		tp = tcp_drop(tp, ETIMEDOUT);
582 		NET_EPOCH_EXIT(et);
583 		CURVNET_RESTORE();
584 
585 		return (tp != NULL);
586 	}
587 	if (tp->t_state == TCPS_SYN_SENT) {
588 		/*
589 		 * If the SYN was retransmitted, indicate CWND to be
590 		 * limited to 1 segment in cc_conn_init().
591 		 */
592 		tp->snd_cwnd = 1;
593 	} else if (tp->t_rxtshift == 1) {
594 		/*
595 		 * first retransmit; record ssthresh and cwnd so they can
596 		 * be recovered if this turns out to be a "bad" retransmit.
597 		 * A retransmit is considered "bad" if an ACK for this
598 		 * segment is received within RTT/2 interval; the assumption
599 		 * here is that the ACK was already in flight.  See
600 		 * "On Estimating End-to-End Network Path Properties" by
601 		 * Allman and Paxson for more details.
602 		 */
603 		tp->snd_cwnd_prev = tp->snd_cwnd;
604 		tp->snd_ssthresh_prev = tp->snd_ssthresh;
605 		tp->snd_recover_prev = tp->snd_recover;
606 		if (IN_FASTRECOVERY(tp->t_flags))
607 			tp->t_flags |= TF_WASFRECOVERY;
608 		else
609 			tp->t_flags &= ~TF_WASFRECOVERY;
610 		if (IN_CONGRECOVERY(tp->t_flags))
611 			tp->t_flags |= TF_WASCRECOVERY;
612 		else
613 			tp->t_flags &= ~TF_WASCRECOVERY;
614 		if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
615 			tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
616 		/* In the event that we've negotiated timestamps
617 		 * badrxtwin will be set to the value that we set
618 		 * the retransmitted packet's to_tsval to by tcp_output
619 		 */
620 		tp->t_flags |= TF_PREVVALID;
621 		tcp_resend_sackholes(tp);
622 	} else {
623 		tp->t_flags &= ~TF_PREVVALID;
624 		tcp_free_sackholes(tp);
625 	}
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 				if (tp->t_maxseg < V_tcp_mssdflt) {
760 					/*
761 					 * The MSS is so small we should not
762 					 * process incoming SACK's since we are
763 					 * subject to attack in such a case.
764 					 */
765 					tp->t_flags2 |= TF2_PROC_SACK_PROHIBIT;
766 				} else {
767 					tp->t_flags2 &= ~TF2_PROC_SACK_PROHIBIT;
768 				}
769 				TCPSTAT_INC(tcps_pmtud_blackhole_failed);
770 				/*
771 				 * Reset the slow-start flight size as it
772 				 * may depend on the new MSS.
773 				 */
774 				if (CC_ALGO(tp)->conn_init != NULL)
775 					CC_ALGO(tp)->conn_init(&tp->t_ccv);
776 			}
777 		}
778 	}
779 
780 	/*
781 	 * Disable RFC1323 and SACK if we haven't got any response to
782 	 * our third SYN to work-around some broken terminal servers
783 	 * (most of which have hopefully been retired) that have bad VJ
784 	 * header compression code which trashes TCP segments containing
785 	 * unknown-to-them TCP options.
786 	 */
787 	if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
788 	    (tp->t_rxtshift == 3))
789 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
790 	/*
791 	 * If we backed off this far, notify the L3 protocol that we're having
792 	 * connection problems.
793 	 */
794 	if (tp->t_rxtshift > TCP_RTT_INVALIDATE) {
795 #ifdef INET6
796 		if ((inp->inp_vflag & INP_IPV6) != 0)
797 			in6_losing(inp);
798 		else
799 #endif
800 			in_losing(inp);
801 	}
802 	tp->snd_nxt = tp->snd_una;
803 	tp->snd_recover = tp->snd_max;
804 	/*
805 	 * Force a segment to be sent.
806 	 */
807 	tp->t_flags |= TF_ACKNOW;
808 	/*
809 	 * If timing a segment in this window, stop the timer.
810 	 */
811 	tp->t_rtttime = 0;
812 
813 	cc_cong_signal(tp, NULL, CC_RTO);
814 	NET_EPOCH_ENTER(et);
815 	rv = tcp_output_locked(tp);
816 	NET_EPOCH_EXIT(et);
817 	CURVNET_RESTORE();
818 
819 	return (rv);
820 }
821 
822 static void
823 tcp_bblog_timer(struct tcpcb *tp, tt_which which, tt_what what, uint32_t ticks)
824 {
825 	struct tcp_log_buffer *lgb;
826 	uint64_t ms;
827 
828 	INP_WLOCK_ASSERT(tptoinpcb(tp));
829 	if (tcp_bblogging_on(tp))
830 		lgb = tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_RTO, 0, 0,
831 		    NULL, false, NULL, NULL, 0, NULL);
832 	else
833 		lgb = NULL;
834 	if (lgb != NULL) {
835 		lgb->tlb_flex1 = (what << 8) | which;
836 		if (what == TT_STARTING) {
837 			/* Convert ticks to ms and store it in tlb_flex2. */
838 			if (hz == 1000)
839 				lgb->tlb_flex2 = ticks;
840 			else {
841 				ms = (((uint64_t)ticks * 1000) + (hz - 1)) / hz;
842 				if (ms > UINT32_MAX)
843 					lgb->tlb_flex2 = UINT32_MAX;
844 				else
845 					lgb->tlb_flex2 = (uint32_t)ms;
846 			}
847 		}
848 	}
849 }
850 
851 static inline tt_which
852 tcp_timer_next(struct tcpcb *tp, sbintime_t *precision)
853 {
854 	tt_which i, rv;
855 	sbintime_t after, before;
856 
857 	for (i = 0, rv = TT_N, after = before = SBT_MAX; i < TT_N; i++) {
858 		if (tp->t_timers[i] < after) {
859 			after = tp->t_timers[i];
860 			rv = i;
861 		}
862 		before = MIN(before, tp->t_timers[i] + tp->t_precisions[i]);
863 	}
864 	if (precision != NULL)
865 		*precision = before - after;
866 
867 	return (rv);
868 }
869 
870 static void
871 tcp_timer_enter(void *xtp)
872 {
873 	struct tcpcb *tp = xtp;
874 	struct inpcb *inp = tptoinpcb(tp);
875 	sbintime_t precision;
876 	tt_which which;
877 	bool tp_valid;
878 
879 	INP_WLOCK_ASSERT(inp);
880 	MPASS((curthread->td_pflags & TDP_INTCPCALLOUT) == 0);
881 
882 	curthread->td_pflags |= TDP_INTCPCALLOUT;
883 
884 	which = tcp_timer_next(tp, NULL);
885 	MPASS(which < TT_N);
886 	tp->t_timers[which] = SBT_MAX;
887 	tp->t_precisions[which] = 0;
888 
889 	tcp_bblog_timer(tp, which, TT_PROCESSING, 0);
890 	tp_valid = tcp_timersw[which](tp);
891 	if (tp_valid) {
892 		tcp_bblog_timer(tp, which, TT_PROCESSED, 0);
893 		if ((which = tcp_timer_next(tp, &precision)) != TT_N) {
894 			MPASS(tp->t_state > TCPS_CLOSED);
895 			callout_reset_sbt_on(&tp->t_callout,
896 			    tp->t_timers[which], precision, tcp_timer_enter,
897 			    tp, inp_to_cpuid(inp), C_ABSOLUTE);
898 		}
899 		INP_WUNLOCK(inp);
900 	}
901 
902 	curthread->td_pflags &= ~TDP_INTCPCALLOUT;
903 }
904 
905 /*
906  * Activate or stop (delta == 0) a TCP timer.
907  */
908 void
909 tcp_timer_activate(struct tcpcb *tp, tt_which which, u_int delta)
910 {
911 	struct inpcb *inp = tptoinpcb(tp);
912 	sbintime_t precision;
913 	tt_what what;
914 
915 #ifdef TCP_OFFLOAD
916 	if (tp->t_flags & TF_TOE)
917 		return;
918 #endif
919 
920 	INP_WLOCK_ASSERT(inp);
921 	MPASS(tp->t_state > TCPS_CLOSED);
922 
923 	if (delta > 0) {
924 		what = TT_STARTING;
925 		callout_when(tick_sbt * delta, 0, C_HARDCLOCK,
926 		    &tp->t_timers[which], &tp->t_precisions[which]);
927 	} else {
928 		what = TT_STOPPING;
929 		tp->t_timers[which] = SBT_MAX;
930 	}
931 	tcp_bblog_timer(tp, which, what, delta);
932 
933 	if ((which = tcp_timer_next(tp, &precision)) != TT_N)
934 		callout_reset_sbt_on(&tp->t_callout, tp->t_timers[which],
935 		    precision, tcp_timer_enter, tp, inp_to_cpuid(inp),
936 		    C_ABSOLUTE);
937 	else
938 		callout_stop(&tp->t_callout);
939 }
940 
941 bool
942 tcp_timer_active(struct tcpcb *tp, tt_which which)
943 {
944 
945 	INP_WLOCK_ASSERT(tptoinpcb(tp));
946 
947 	return (tp->t_timers[which] != SBT_MAX);
948 }
949 
950 /*
951  * Stop all timers associated with tcpcb.
952  *
953  * Called when tcpcb moves to TCPS_CLOSED.
954  *
955  * XXXGL: unfortunately our callout(9) is not able to fully stop a locked
956  * callout even when only two threads are involved: the callout itself and the
957  * thread that does callout_stop().  See where softclock_call_cc() swaps the
958  * callwheel lock to callout lock and then checks cc_exec_cancel().  This is
959  * the race window.  If it happens, the tcp_timer_enter() won't be executed,
960  * however pcb lock will be locked and released, hence we can't free memory.
961  * Until callout(9) is improved, just keep retrying.  In my profiling I've seen
962  * such event happening less than 1 time per hour with 20-30 Gbit/s of traffic.
963  */
964 void
965 tcp_timer_stop(struct tcpcb *tp)
966 {
967 	struct inpcb *inp = tptoinpcb(tp);
968 
969 	INP_WLOCK_ASSERT(inp);
970 
971 	if (curthread->td_pflags & TDP_INTCPCALLOUT) {
972 		int stopped __diagused;
973 
974 		stopped = callout_stop(&tp->t_callout);
975 		MPASS(stopped == 0);
976 		for (tt_which i = 0; i < TT_N; i++)
977 			tp->t_timers[i] = SBT_MAX;
978 	} else while(__predict_false(callout_stop(&tp->t_callout) == 0)) {
979 		INP_WUNLOCK(inp);
980 		kern_yield(PRI_UNCHANGED);
981 		INP_WLOCK(inp);
982 	}
983 }
984