xref: /freebsd/sys/netinet/tcp_timewait.c (revision d0ba1baed3f6e4936a0c1b89c25f6c59168ef6de)
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_subr.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_tcpdebug.h"
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/callout.h>
44 #include <sys/kernel.h>
45 #include <sys/sysctl.h>
46 #include <sys/malloc.h>
47 #include <sys/mbuf.h>
48 #include <sys/priv.h>
49 #include <sys/proc.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #ifndef INVARIANTS
53 #include <sys/syslog.h>
54 #endif
55 #include <sys/protosw.h>
56 #include <sys/random.h>
57 
58 #include <vm/uma.h>
59 
60 #include <net/route.h>
61 #include <net/if.h>
62 #include <net/if_var.h>
63 #include <net/vnet.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/in_systm.h>
68 #include <netinet/in_var.h>
69 #include <netinet/ip.h>
70 #include <netinet/ip_icmp.h>
71 #include <netinet/ip_var.h>
72 #ifdef INET6
73 #include <netinet/ip6.h>
74 #include <netinet6/in6_pcb.h>
75 #include <netinet6/ip6_var.h>
76 #include <netinet6/scope6_var.h>
77 #include <netinet6/nd6.h>
78 #endif
79 #include <netinet/tcp.h>
80 #include <netinet/tcp_fsm.h>
81 #include <netinet/tcp_seq.h>
82 #include <netinet/tcp_timer.h>
83 #include <netinet/tcp_var.h>
84 #ifdef INET6
85 #include <netinet6/tcp6_var.h>
86 #endif
87 #include <netinet/tcpip.h>
88 #ifdef TCPDEBUG
89 #include <netinet/tcp_debug.h>
90 #endif
91 #ifdef INET6
92 #include <netinet6/ip6protosw.h>
93 #endif
94 
95 #include <machine/in_cksum.h>
96 
97 #include <security/mac/mac_framework.h>
98 
99 static VNET_DEFINE(uma_zone_t, tcptw_zone);
100 #define	V_tcptw_zone		VNET(tcptw_zone)
101 static int	maxtcptw;
102 
103 /*
104  * The timed wait queue contains references to each of the TCP sessions
105  * currently in the TIME_WAIT state.  The queue pointers, including the
106  * queue pointers in each tcptw structure, are protected using the global
107  * timewait lock, which must be held over queue iteration and modification.
108  *
109  * Rules on tcptw usage:
110  *  - a inpcb is always freed _after_ its tcptw
111  *  - a tcptw relies on its inpcb reference counting for memory stability
112  *  - a tcptw is dereferenceable only while its inpcb is locked
113  */
114 static VNET_DEFINE(TAILQ_HEAD(, tcptw), twq_2msl);
115 #define	V_twq_2msl		VNET(twq_2msl)
116 
117 /* Global timewait lock */
118 static VNET_DEFINE(struct rwlock, tw_lock);
119 #define	V_tw_lock		VNET(tw_lock)
120 
121 #define	TW_LOCK_INIT(tw, d)	rw_init_flags(&(tw), (d), 0)
122 #define	TW_LOCK_DESTROY(tw)	rw_destroy(&(tw))
123 #define	TW_RLOCK(tw)		rw_rlock(&(tw))
124 #define	TW_WLOCK(tw)		rw_wlock(&(tw))
125 #define	TW_RUNLOCK(tw)		rw_runlock(&(tw))
126 #define	TW_WUNLOCK(tw)		rw_wunlock(&(tw))
127 #define	TW_LOCK_ASSERT(tw)	rw_assert(&(tw), RA_LOCKED)
128 #define	TW_RLOCK_ASSERT(tw)	rw_assert(&(tw), RA_RLOCKED)
129 #define	TW_WLOCK_ASSERT(tw)	rw_assert(&(tw), RA_WLOCKED)
130 #define	TW_UNLOCK_ASSERT(tw)	rw_assert(&(tw), RA_UNLOCKED)
131 
132 static void	tcp_tw_2msl_reset(struct tcptw *, int);
133 static void	tcp_tw_2msl_stop(struct tcptw *, int);
134 static int	tcp_twrespond(struct tcptw *, int);
135 
136 static int
137 tcptw_auto_size(void)
138 {
139 	int halfrange;
140 
141 	/*
142 	 * Max out at half the ephemeral port range so that TIME_WAIT
143 	 * sockets don't tie up too many ephemeral ports.
144 	 */
145 	if (V_ipport_lastauto > V_ipport_firstauto)
146 		halfrange = (V_ipport_lastauto - V_ipport_firstauto) / 2;
147 	else
148 		halfrange = (V_ipport_firstauto - V_ipport_lastauto) / 2;
149 	/* Protect against goofy port ranges smaller than 32. */
150 	return (imin(imax(halfrange, 32), maxsockets / 5));
151 }
152 
153 static int
154 sysctl_maxtcptw(SYSCTL_HANDLER_ARGS)
155 {
156 	int error, new;
157 
158 	if (maxtcptw == 0)
159 		new = tcptw_auto_size();
160 	else
161 		new = maxtcptw;
162 	error = sysctl_handle_int(oidp, &new, 0, req);
163 	if (error == 0 && req->newptr)
164 		if (new >= 32) {
165 			maxtcptw = new;
166 			uma_zone_set_max(V_tcptw_zone, maxtcptw);
167 		}
168 	return (error);
169 }
170 
171 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, maxtcptw, CTLTYPE_INT|CTLFLAG_RW,
172     &maxtcptw, 0, sysctl_maxtcptw, "IU",
173     "Maximum number of compressed TCP TIME_WAIT entries");
174 
175 static VNET_DEFINE(int, nolocaltimewait) = 0;
176 #define	V_nolocaltimewait	VNET(nolocaltimewait)
177 SYSCTL_INT(_net_inet_tcp, OID_AUTO, nolocaltimewait, CTLFLAG_VNET | CTLFLAG_RW,
178     &VNET_NAME(nolocaltimewait), 0,
179     "Do not create compressed TCP TIME_WAIT entries for local connections");
180 
181 void
182 tcp_tw_zone_change(void)
183 {
184 
185 	if (maxtcptw == 0)
186 		uma_zone_set_max(V_tcptw_zone, tcptw_auto_size());
187 }
188 
189 void
190 tcp_tw_init(void)
191 {
192 
193 	V_tcptw_zone = uma_zcreate("tcptw", sizeof(struct tcptw),
194 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
195 	TUNABLE_INT_FETCH("net.inet.tcp.maxtcptw", &maxtcptw);
196 	if (maxtcptw == 0)
197 		uma_zone_set_max(V_tcptw_zone, tcptw_auto_size());
198 	else
199 		uma_zone_set_max(V_tcptw_zone, maxtcptw);
200 	TAILQ_INIT(&V_twq_2msl);
201 	TW_LOCK_INIT(V_tw_lock, "tcptw");
202 }
203 
204 #ifdef VIMAGE
205 void
206 tcp_tw_destroy(void)
207 {
208 	struct tcptw *tw;
209 
210 	INP_INFO_RLOCK(&V_tcbinfo);
211 	while ((tw = TAILQ_FIRST(&V_twq_2msl)) != NULL)
212 		tcp_twclose(tw, 0);
213 	INP_INFO_RUNLOCK(&V_tcbinfo);
214 
215 	TW_LOCK_DESTROY(V_tw_lock);
216 	uma_zdestroy(V_tcptw_zone);
217 }
218 #endif
219 
220 /*
221  * Move a TCP connection into TIME_WAIT state.
222  *    tcbinfo is locked.
223  *    inp is locked, and is unlocked before returning.
224  */
225 void
226 tcp_twstart(struct tcpcb *tp)
227 {
228 	struct tcptw twlocal, *tw;
229 	struct inpcb *inp = tp->t_inpcb;
230 	struct socket *so;
231 	bool acknow, local;
232 #ifdef INET6
233 	bool isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6;
234 #endif
235 
236 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
237 	INP_WLOCK_ASSERT(inp);
238 
239 	/* A dropped inp should never transition to TIME_WAIT state. */
240 	KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("tcp_twstart: "
241 	    "(inp->inp_flags & INP_DROPPED) != 0"));
242 
243 	if (V_nolocaltimewait) {
244 #ifdef INET6
245 		if (isipv6)
246 			local = in6_localaddr(&inp->in6p_faddr);
247 		else
248 #endif
249 #ifdef INET
250 			local = in_localip(inp->inp_faddr);
251 #else
252 			local = false;
253 #endif
254 	} else
255 		local = false;
256 
257 	/*
258 	 * For use only by DTrace.  We do not reference the state
259 	 * after this point so modifying it in place is not a problem.
260 	 */
261 	tcp_state_change(tp, TCPS_TIME_WAIT);
262 
263 	if (local)
264 		tw = &twlocal;
265 	else
266 		tw = uma_zalloc(V_tcptw_zone, M_NOWAIT);
267 	if (tw == NULL) {
268 		/*
269 		 * Reached limit on total number of TIMEWAIT connections
270 		 * allowed. Remove a connection from TIMEWAIT queue in LRU
271 		 * fashion to make room for this connection.
272 		 *
273 		 * XXX:  Check if it possible to always have enough room
274 		 * in advance based on guarantees provided by uma_zalloc().
275 		 */
276 		tw = tcp_tw_2msl_scan(1);
277 		if (tw == NULL) {
278 			tp = tcp_close(tp);
279 			if (tp != NULL)
280 				INP_WUNLOCK(inp);
281 			return;
282 		}
283 	}
284 	/*
285 	 * For !local case the tcptw will hold a reference on its inpcb
286 	 * until tcp_twclose is called.
287 	 */
288 	tw->tw_inpcb = inp;
289 
290 	/*
291 	 * Recover last window size sent.
292 	 */
293 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt))
294 		tw->last_win = (tp->rcv_adv - tp->rcv_nxt) >> tp->rcv_scale;
295 	else
296 		tw->last_win = 0;
297 
298 	/*
299 	 * Set t_recent if timestamps are used on the connection.
300 	 */
301 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_RCVD_TSTMP|TF_NOOPT)) ==
302 	    (TF_REQ_TSTMP|TF_RCVD_TSTMP)) {
303 		tw->t_recent = tp->ts_recent;
304 		tw->ts_offset = tp->ts_offset;
305 	} else {
306 		tw->t_recent = 0;
307 		tw->ts_offset = 0;
308 	}
309 
310 	tw->snd_nxt = tp->snd_nxt;
311 	tw->rcv_nxt = tp->rcv_nxt;
312 	tw->iss     = tp->iss;
313 	tw->irs     = tp->irs;
314 	tw->t_starttime = tp->t_starttime;
315 	tw->tw_time = 0;
316 
317 /* XXX
318  * If this code will
319  * be used for fin-wait-2 state also, then we may need
320  * a ts_recent from the last segment.
321  */
322 	acknow = tp->t_flags & TF_ACKNOW;
323 
324 	/*
325 	 * First, discard tcpcb state, which includes stopping its timers and
326 	 * freeing it.  tcp_discardcb() used to also release the inpcb, but
327 	 * that work is now done in the caller.
328 	 *
329 	 * Note: soisdisconnected() call used to be made in tcp_discardcb(),
330 	 * and might not be needed here any longer.
331 	 */
332 	tcp_discardcb(tp);
333 	so = inp->inp_socket;
334 	soisdisconnected(so);
335 	tw->tw_so_options = so->so_options;
336 	inp->inp_flags |= INP_TIMEWAIT;
337 	if (acknow)
338 		tcp_twrespond(tw, TH_ACK);
339 	if (local)
340 		in_pcbdrop(inp);
341 	else {
342 		in_pcbref(inp);	/* Reference from tw */
343 		tw->tw_cred = crhold(so->so_cred);
344 		inp->inp_ppcb = tw;
345 		TCPSTATES_INC(TCPS_TIME_WAIT);
346 		tcp_tw_2msl_reset(tw, 0);
347 	}
348 
349 	/*
350 	 * If the inpcb owns the sole reference to the socket, then we can
351 	 * detach and free the socket as it is not needed in time wait.
352 	 */
353 	if (inp->inp_flags & INP_SOCKREF) {
354 		KASSERT(so->so_state & SS_PROTOREF,
355 		    ("tcp_twstart: !SS_PROTOREF"));
356 		inp->inp_flags &= ~INP_SOCKREF;
357 		INP_WUNLOCK(inp);
358 		SOCK_LOCK(so);
359 		so->so_state &= ~SS_PROTOREF;
360 		sofree(so);
361 	} else
362 		INP_WUNLOCK(inp);
363 }
364 
365 /*
366  * Returns 1 if the TIME_WAIT state was killed and we should start over,
367  * looking for a pcb in the listen state.  Returns 0 otherwise.
368  */
369 int
370 tcp_twcheck(struct inpcb *inp, struct tcpopt *to __unused, struct tcphdr *th,
371     struct mbuf *m, int tlen)
372 {
373 	struct tcptw *tw;
374 	int thflags;
375 	tcp_seq seq;
376 
377 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
378 	INP_WLOCK_ASSERT(inp);
379 
380 	/*
381 	 * XXXRW: Time wait state for inpcb has been recycled, but inpcb is
382 	 * still present.  This is undesirable, but temporarily necessary
383 	 * until we work out how to handle inpcb's who's timewait state has
384 	 * been removed.
385 	 */
386 	tw = intotw(inp);
387 	if (tw == NULL)
388 		goto drop;
389 
390 	thflags = th->th_flags;
391 
392 	/*
393 	 * NOTE: for FIN_WAIT_2 (to be added later),
394 	 * must validate sequence number before accepting RST
395 	 */
396 
397 	/*
398 	 * If the segment contains RST:
399 	 *	Drop the segment - see Stevens, vol. 2, p. 964 and
400 	 *      RFC 1337.
401 	 */
402 	if (thflags & TH_RST)
403 		goto drop;
404 
405 #if 0
406 /* PAWS not needed at the moment */
407 	/*
408 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment
409 	 * and it's less than ts_recent, drop it.
410 	 */
411 	if ((to.to_flags & TOF_TS) != 0 && tp->ts_recent &&
412 	    TSTMP_LT(to.to_tsval, tp->ts_recent)) {
413 		if ((thflags & TH_ACK) == 0)
414 			goto drop;
415 		goto ack;
416 	}
417 	/*
418 	 * ts_recent is never updated because we never accept new segments.
419 	 */
420 #endif
421 
422 	/*
423 	 * If a new connection request is received
424 	 * while in TIME_WAIT, drop the old connection
425 	 * and start over if the sequence numbers
426 	 * are above the previous ones.
427 	 */
428 	if ((thflags & TH_SYN) && SEQ_GT(th->th_seq, tw->rcv_nxt)) {
429 		tcp_twclose(tw, 0);
430 		return (1);
431 	}
432 
433 	/*
434 	 * Drop the segment if it does not contain an ACK.
435 	 */
436 	if ((thflags & TH_ACK) == 0)
437 		goto drop;
438 
439 	/*
440 	 * Reset the 2MSL timer if this is a duplicate FIN.
441 	 */
442 	if (thflags & TH_FIN) {
443 		seq = th->th_seq + tlen + (thflags & TH_SYN ? 1 : 0);
444 		if (seq + 1 == tw->rcv_nxt)
445 			tcp_tw_2msl_reset(tw, 1);
446 	}
447 
448 	/*
449 	 * Acknowledge the segment if it has data or is not a duplicate ACK.
450 	 */
451 	if (thflags != TH_ACK || tlen != 0 ||
452 	    th->th_seq != tw->rcv_nxt || th->th_ack != tw->snd_nxt)
453 		tcp_twrespond(tw, TH_ACK);
454 drop:
455 	INP_WUNLOCK(inp);
456 	m_freem(m);
457 	return (0);
458 }
459 
460 void
461 tcp_twclose(struct tcptw *tw, int reuse)
462 {
463 	struct socket *so;
464 	struct inpcb *inp;
465 
466 	/*
467 	 * At this point, we are in one of two situations:
468 	 *
469 	 * (1) We have no socket, just an inpcb<->twtcp pair.  We can free
470 	 *     all state.
471 	 *
472 	 * (2) We have a socket -- if we own a reference, release it and
473 	 *     notify the socket layer.
474 	 */
475 	inp = tw->tw_inpcb;
476 	KASSERT((inp->inp_flags & INP_TIMEWAIT), ("tcp_twclose: !timewait"));
477 	KASSERT(intotw(inp) == tw, ("tcp_twclose: inp_ppcb != tw"));
478 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);	/* in_pcbfree() */
479 	INP_WLOCK_ASSERT(inp);
480 
481 	tcp_tw_2msl_stop(tw, reuse);
482 	inp->inp_ppcb = NULL;
483 	in_pcbdrop(inp);
484 
485 	so = inp->inp_socket;
486 	if (so != NULL) {
487 		/*
488 		 * If there's a socket, handle two cases: first, we own a
489 		 * strong reference, which we will now release, or we don't
490 		 * in which case another reference exists (XXXRW: think
491 		 * about this more), and we don't need to take action.
492 		 */
493 		if (inp->inp_flags & INP_SOCKREF) {
494 			inp->inp_flags &= ~INP_SOCKREF;
495 			INP_WUNLOCK(inp);
496 			SOCK_LOCK(so);
497 			KASSERT(so->so_state & SS_PROTOREF,
498 			    ("tcp_twclose: INP_SOCKREF && !SS_PROTOREF"));
499 			so->so_state &= ~SS_PROTOREF;
500 			sofree(so);
501 		} else {
502 			/*
503 			 * If we don't own the only reference, the socket and
504 			 * inpcb need to be left around to be handled by
505 			 * tcp_usr_detach() later.
506 			 */
507 			INP_WUNLOCK(inp);
508 		}
509 	} else {
510 		/*
511 		 * The socket has been already cleaned-up for us, only free the
512 		 * inpcb.
513 		 */
514 		in_pcbfree(inp);
515 	}
516 	TCPSTAT_INC(tcps_closed);
517 }
518 
519 static int
520 tcp_twrespond(struct tcptw *tw, int flags)
521 {
522 	struct inpcb *inp = tw->tw_inpcb;
523 #if defined(INET6) || defined(INET)
524 	struct tcphdr *th = NULL;
525 #endif
526 	struct mbuf *m;
527 #ifdef INET
528 	struct ip *ip = NULL;
529 #endif
530 	u_int hdrlen, optlen;
531 	int error = 0;			/* Keep compiler happy */
532 	struct tcpopt to;
533 #ifdef INET6
534 	struct ip6_hdr *ip6 = NULL;
535 	int isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6;
536 #endif
537 	hdrlen = 0;                     /* Keep compiler happy */
538 
539 	INP_WLOCK_ASSERT(inp);
540 
541 	m = m_gethdr(M_NOWAIT, MT_DATA);
542 	if (m == NULL)
543 		return (ENOBUFS);
544 	m->m_data += max_linkhdr;
545 
546 #ifdef MAC
547 	mac_inpcb_create_mbuf(inp, m);
548 #endif
549 
550 #ifdef INET6
551 	if (isipv6) {
552 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
553 		ip6 = mtod(m, struct ip6_hdr *);
554 		th = (struct tcphdr *)(ip6 + 1);
555 		tcpip_fillheaders(inp, ip6, th);
556 	}
557 #endif
558 #if defined(INET6) && defined(INET)
559 	else
560 #endif
561 #ifdef INET
562 	{
563 		hdrlen = sizeof(struct tcpiphdr);
564 		ip = mtod(m, struct ip *);
565 		th = (struct tcphdr *)(ip + 1);
566 		tcpip_fillheaders(inp, ip, th);
567 	}
568 #endif
569 	to.to_flags = 0;
570 
571 	/*
572 	 * Send a timestamp and echo-reply if both our side and our peer
573 	 * have sent timestamps in our SYN's and this is not a RST.
574 	 */
575 	if (tw->t_recent && flags == TH_ACK) {
576 		to.to_flags |= TOF_TS;
577 		to.to_tsval = tcp_ts_getticks() + tw->ts_offset;
578 		to.to_tsecr = tw->t_recent;
579 	}
580 	optlen = tcp_addoptions(&to, (u_char *)(th + 1));
581 
582 	m->m_len = hdrlen + optlen;
583 	m->m_pkthdr.len = m->m_len;
584 
585 	KASSERT(max_linkhdr + m->m_len <= MHLEN, ("tcptw: mbuf too small"));
586 
587 	th->th_seq = htonl(tw->snd_nxt);
588 	th->th_ack = htonl(tw->rcv_nxt);
589 	th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
590 	th->th_flags = flags;
591 	th->th_win = htons(tw->last_win);
592 
593 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
594 #ifdef INET6
595 	if (isipv6) {
596 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
597 		th->th_sum = in6_cksum_pseudo(ip6,
598 		    sizeof(struct tcphdr) + optlen, IPPROTO_TCP, 0);
599 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
600 		error = ip6_output(m, inp->in6p_outputopts, NULL,
601 		    (tw->tw_so_options & SO_DONTROUTE), NULL, NULL, inp);
602 	}
603 #endif
604 #if defined(INET6) && defined(INET)
605 	else
606 #endif
607 #ifdef INET
608 	{
609 		m->m_pkthdr.csum_flags = CSUM_TCP;
610 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
611 		    htons(sizeof(struct tcphdr) + optlen + IPPROTO_TCP));
612 		ip->ip_len = htons(m->m_pkthdr.len);
613 		if (V_path_mtu_discovery)
614 			ip->ip_off |= htons(IP_DF);
615 		error = ip_output(m, inp->inp_options, NULL,
616 		    ((tw->tw_so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
617 		    NULL, inp);
618 	}
619 #endif
620 	if (flags & TH_ACK)
621 		TCPSTAT_INC(tcps_sndacks);
622 	else
623 		TCPSTAT_INC(tcps_sndctrl);
624 	TCPSTAT_INC(tcps_sndtotal);
625 	return (error);
626 }
627 
628 static void
629 tcp_tw_2msl_reset(struct tcptw *tw, int rearm)
630 {
631 
632 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
633 	INP_WLOCK_ASSERT(tw->tw_inpcb);
634 
635 	TW_WLOCK(V_tw_lock);
636 	if (rearm)
637 		TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl);
638 	tw->tw_time = ticks + 2 * tcp_msl;
639 	TAILQ_INSERT_TAIL(&V_twq_2msl, tw, tw_2msl);
640 	TW_WUNLOCK(V_tw_lock);
641 }
642 
643 static void
644 tcp_tw_2msl_stop(struct tcptw *tw, int reuse)
645 {
646 	struct ucred *cred;
647 	struct inpcb *inp;
648 	int released __unused;
649 
650 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
651 
652 	TW_WLOCK(V_tw_lock);
653 	inp = tw->tw_inpcb;
654 	tw->tw_inpcb = NULL;
655 
656 	TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl);
657 	cred = tw->tw_cred;
658 	tw->tw_cred = NULL;
659 	TW_WUNLOCK(V_tw_lock);
660 
661 	if (cred != NULL)
662 		crfree(cred);
663 
664 	released = in_pcbrele_wlocked(inp);
665 	KASSERT(!released, ("%s: inp should not be released here", __func__));
666 
667 	if (!reuse)
668 		uma_zfree(V_tcptw_zone, tw);
669 	TCPSTATES_DEC(TCPS_TIME_WAIT);
670 }
671 
672 struct tcptw *
673 tcp_tw_2msl_scan(int reuse)
674 {
675 	struct tcptw *tw;
676 	struct inpcb *inp;
677 
678 #ifdef INVARIANTS
679 	if (reuse) {
680 		/*
681 		 * Exclusive pcbinfo lock is not required in reuse case even if
682 		 * two inpcb locks can be acquired simultaneously:
683 		 *  - the inpcb transitioning to TIME_WAIT state in
684 		 *    tcp_tw_start(),
685 		 *  - the inpcb closed by tcp_twclose().
686 		 *
687 		 * It is because only inpcbs in FIN_WAIT2 or CLOSING states can
688 		 * transition in TIME_WAIT state.  Then a pcbcb cannot be in
689 		 * TIME_WAIT list and transitioning to TIME_WAIT state at same
690 		 * time.
691 		 */
692 		INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
693 	}
694 #endif
695 
696 	for (;;) {
697 		TW_RLOCK(V_tw_lock);
698 		tw = TAILQ_FIRST(&V_twq_2msl);
699 		if (tw == NULL || (!reuse && (tw->tw_time - ticks) > 0)) {
700 			TW_RUNLOCK(V_tw_lock);
701 			break;
702 		}
703 		KASSERT(tw->tw_inpcb != NULL, ("%s: tw->tw_inpcb == NULL",
704 		    __func__));
705 
706 		inp = tw->tw_inpcb;
707 		in_pcbref(inp);
708 		TW_RUNLOCK(V_tw_lock);
709 
710 		if (INP_INFO_TRY_RLOCK(&V_tcbinfo)) {
711 
712 			INP_WLOCK(inp);
713 			tw = intotw(inp);
714 			if (in_pcbrele_wlocked(inp)) {
715 				if (__predict_true(tw == NULL)) {
716 					INP_INFO_RUNLOCK(&V_tcbinfo);
717 					continue;
718 				} else {
719 					/* This should not happen as in TIMEWAIT
720 					 * state the inp should not be destroyed
721 					 * before its tcptw. If INVARIANTS is
722 					 * defined panic.
723 					 */
724 #ifdef INVARIANTS
725 					panic("%s: Panic before an infinite "
726 					    "loop: INP_TIMEWAIT && (INP_FREED "
727 					    "|| inp last reference) && tw != "
728 					    "NULL", __func__);
729 #else
730 					log(LOG_ERR, "%s: Avoid an infinite "
731 					    "loop: INP_TIMEWAIT && (INP_FREED "
732 					    "|| inp last reference) && tw != "
733 					    "NULL", __func__);
734 #endif
735 					INP_INFO_RUNLOCK(&V_tcbinfo);
736 					break;
737 				}
738 			}
739 
740 			if (tw == NULL) {
741 				/* tcp_twclose() has already been called */
742 				INP_WUNLOCK(inp);
743 				INP_INFO_RUNLOCK(&V_tcbinfo);
744 				continue;
745 			}
746 
747 			tcp_twclose(tw, reuse);
748 			INP_INFO_RUNLOCK(&V_tcbinfo);
749 			if (reuse)
750 			    return tw;
751 		} else {
752 			/* INP_INFO lock is busy, continue later. */
753 			INP_WLOCK(inp);
754 			if (!in_pcbrele_wlocked(inp))
755 				INP_WUNLOCK(inp);
756 			break;
757 		}
758 	}
759 
760 	return NULL;
761 }
762