xref: /freebsd/sys/netinet/tcp_timewait.c (revision b9f654b163bce26de79705e77b872427c9f2afa1)
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_kdtrace.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/in_var.h>
70 #include <netinet/ip.h>
71 #include <netinet/ip_icmp.h>
72 #include <netinet/ip_var.h>
73 #ifdef INET6
74 #include <netinet/ip6.h>
75 #include <netinet6/in6_pcb.h>
76 #include <netinet6/ip6_var.h>
77 #include <netinet6/scope6_var.h>
78 #include <netinet6/nd6.h>
79 #endif
80 #include <netinet/tcp.h>
81 #include <netinet/tcp_fsm.h>
82 #include <netinet/tcp_seq.h>
83 #include <netinet/tcp_timer.h>
84 #include <netinet/tcp_var.h>
85 #ifdef INET6
86 #include <netinet6/tcp6_var.h>
87 #endif
88 #include <netinet/tcpip.h>
89 #ifdef TCPDEBUG
90 #include <netinet/tcp_debug.h>
91 #endif
92 #ifdef INET6
93 #include <netinet6/ip6protosw.h>
94 #endif
95 
96 #include <machine/in_cksum.h>
97 
98 #include <security/mac/mac_framework.h>
99 
100 VNET_DEFINE_STATIC(uma_zone_t, tcptw_zone);
101 #define	V_tcptw_zone		VNET(tcptw_zone)
102 static int	maxtcptw;
103 
104 /*
105  * The timed wait queue contains references to each of the TCP sessions
106  * currently in the TIME_WAIT state.  The queue pointers, including the
107  * queue pointers in each tcptw structure, are protected using the global
108  * timewait lock, which must be held over queue iteration and modification.
109  *
110  * Rules on tcptw usage:
111  *  - a inpcb is always freed _after_ its tcptw
112  *  - a tcptw relies on its inpcb reference counting for memory stability
113  *  - a tcptw is dereferenceable only while its inpcb is locked
114  */
115 VNET_DEFINE_STATIC(TAILQ_HEAD(, tcptw), twq_2msl);
116 #define	V_twq_2msl		VNET(twq_2msl)
117 
118 /* Global timewait lock */
119 VNET_DEFINE_STATIC(struct rwlock, tw_lock);
120 #define	V_tw_lock		VNET(tw_lock)
121 
122 #define	TW_LOCK_INIT(tw, d)	rw_init_flags(&(tw), (d), 0)
123 #define	TW_LOCK_DESTROY(tw)	rw_destroy(&(tw))
124 #define	TW_RLOCK(tw)		rw_rlock(&(tw))
125 #define	TW_WLOCK(tw)		rw_wlock(&(tw))
126 #define	TW_RUNLOCK(tw)		rw_runlock(&(tw))
127 #define	TW_WUNLOCK(tw)		rw_wunlock(&(tw))
128 #define	TW_LOCK_ASSERT(tw)	rw_assert(&(tw), RA_LOCKED)
129 #define	TW_RLOCK_ASSERT(tw)	rw_assert(&(tw), RA_RLOCKED)
130 #define	TW_WLOCK_ASSERT(tw)	rw_assert(&(tw), RA_WLOCKED)
131 #define	TW_UNLOCK_ASSERT(tw)	rw_assert(&(tw), RA_UNLOCKED)
132 
133 static void	tcp_tw_2msl_reset(struct tcptw *, int);
134 static void	tcp_tw_2msl_stop(struct tcptw *, int);
135 static int	tcp_twrespond(struct tcptw *, int);
136 
137 static int
138 tcptw_auto_size(void)
139 {
140 	int halfrange;
141 
142 	/*
143 	 * Max out at half the ephemeral port range so that TIME_WAIT
144 	 * sockets don't tie up too many ephemeral ports.
145 	 */
146 	if (V_ipport_lastauto > V_ipport_firstauto)
147 		halfrange = (V_ipport_lastauto - V_ipport_firstauto) / 2;
148 	else
149 		halfrange = (V_ipport_firstauto - V_ipport_lastauto) / 2;
150 	/* Protect against goofy port ranges smaller than 32. */
151 	return (imin(imax(halfrange, 32), maxsockets / 5));
152 }
153 
154 static int
155 sysctl_maxtcptw(SYSCTL_HANDLER_ARGS)
156 {
157 	int error, new;
158 
159 	if (maxtcptw == 0)
160 		new = tcptw_auto_size();
161 	else
162 		new = maxtcptw;
163 	error = sysctl_handle_int(oidp, &new, 0, req);
164 	if (error == 0 && req->newptr)
165 		if (new >= 32) {
166 			maxtcptw = new;
167 			uma_zone_set_max(V_tcptw_zone, maxtcptw);
168 		}
169 	return (error);
170 }
171 
172 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, maxtcptw, CTLTYPE_INT|CTLFLAG_RW,
173     &maxtcptw, 0, sysctl_maxtcptw, "IU",
174     "Maximum number of compressed TCP TIME_WAIT entries");
175 
176 VNET_DEFINE_STATIC(int, nolocaltimewait) = 0;
177 #define	V_nolocaltimewait	VNET(nolocaltimewait)
178 SYSCTL_INT(_net_inet_tcp, OID_AUTO, nolocaltimewait, CTLFLAG_VNET | CTLFLAG_RW,
179     &VNET_NAME(nolocaltimewait), 0,
180     "Do not create compressed TCP TIME_WAIT entries for local connections");
181 
182 void
183 tcp_tw_zone_change(void)
184 {
185 
186 	if (maxtcptw == 0)
187 		uma_zone_set_max(V_tcptw_zone, tcptw_auto_size());
188 }
189 
190 void
191 tcp_tw_init(void)
192 {
193 
194 	V_tcptw_zone = uma_zcreate("tcptw", sizeof(struct tcptw),
195 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
196 	TUNABLE_INT_FETCH("net.inet.tcp.maxtcptw", &maxtcptw);
197 	if (maxtcptw == 0)
198 		uma_zone_set_max(V_tcptw_zone, tcptw_auto_size());
199 	else
200 		uma_zone_set_max(V_tcptw_zone, maxtcptw);
201 	TAILQ_INIT(&V_twq_2msl);
202 	TW_LOCK_INIT(V_tw_lock, "tcptw");
203 }
204 
205 #ifdef VIMAGE
206 void
207 tcp_tw_destroy(void)
208 {
209 	struct tcptw *tw;
210 	struct epoch_tracker et;
211 
212 	INP_INFO_RLOCK_ET(&V_tcbinfo, et);
213 	while ((tw = TAILQ_FIRST(&V_twq_2msl)) != NULL)
214 		tcp_twclose(tw, 0);
215 	INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
216 
217 	TW_LOCK_DESTROY(V_tw_lock);
218 	uma_zdestroy(V_tcptw_zone);
219 }
220 #endif
221 
222 /*
223  * Move a TCP connection into TIME_WAIT state.
224  *    tcbinfo is locked.
225  *    inp is locked, and is unlocked before returning.
226  */
227 void
228 tcp_twstart(struct tcpcb *tp)
229 {
230 	struct tcptw twlocal, *tw;
231 	struct inpcb *inp = tp->t_inpcb;
232 	struct socket *so;
233 	uint32_t recwin;
234 	bool acknow, local;
235 #ifdef INET6
236 	bool isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6;
237 #endif
238 
239 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
240 	INP_WLOCK_ASSERT(inp);
241 
242 	/* A dropped inp should never transition to TIME_WAIT state. */
243 	KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("tcp_twstart: "
244 	    "(inp->inp_flags & INP_DROPPED) != 0"));
245 
246 	if (V_nolocaltimewait) {
247 #ifdef INET6
248 		if (isipv6)
249 			local = in6_localaddr(&inp->in6p_faddr);
250 		else
251 #endif
252 #ifdef INET
253 			local = in_localip(inp->inp_faddr);
254 #else
255 			local = false;
256 #endif
257 	} else
258 		local = false;
259 
260 	/*
261 	 * For use only by DTrace.  We do not reference the state
262 	 * after this point so modifying it in place is not a problem.
263 	 */
264 	tcp_state_change(tp, TCPS_TIME_WAIT);
265 
266 	if (local)
267 		tw = &twlocal;
268 	else
269 		tw = uma_zalloc(V_tcptw_zone, M_NOWAIT);
270 	if (tw == NULL) {
271 		/*
272 		 * Reached limit on total number of TIMEWAIT connections
273 		 * allowed. Remove a connection from TIMEWAIT queue in LRU
274 		 * fashion to make room for this connection.
275 		 *
276 		 * XXX:  Check if it possible to always have enough room
277 		 * in advance based on guarantees provided by uma_zalloc().
278 		 */
279 		tw = tcp_tw_2msl_scan(1);
280 		if (tw == NULL) {
281 			tp = tcp_close(tp);
282 			if (tp != NULL)
283 				INP_WUNLOCK(inp);
284 			return;
285 		}
286 	}
287 	/*
288 	 * For !local case the tcptw will hold a reference on its inpcb
289 	 * until tcp_twclose is called.
290 	 */
291 	tw->tw_inpcb = inp;
292 
293 	/*
294 	 * Recover last window size sent.
295 	 */
296 	so = inp->inp_socket;
297 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
298 	    (long)TCP_MAXWIN << tp->rcv_scale);
299 	if (recwin < (so->so_rcv.sb_hiwat / 4) &&
300 	    recwin < tp->t_maxseg)
301 		recwin = 0;
302 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
303 	    recwin < (tp->rcv_adv - tp->rcv_nxt))
304 		recwin = (tp->rcv_adv - tp->rcv_nxt);
305 	tw->last_win = (u_short)(recwin >> tp->rcv_scale);
306 
307 	/*
308 	 * Set t_recent if timestamps are used on the connection.
309 	 */
310 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_RCVD_TSTMP|TF_NOOPT)) ==
311 	    (TF_REQ_TSTMP|TF_RCVD_TSTMP)) {
312 		tw->t_recent = tp->ts_recent;
313 		tw->ts_offset = tp->ts_offset;
314 	} else {
315 		tw->t_recent = 0;
316 		tw->ts_offset = 0;
317 	}
318 
319 	tw->snd_nxt = tp->snd_nxt;
320 	tw->rcv_nxt = tp->rcv_nxt;
321 	tw->iss     = tp->iss;
322 	tw->irs     = tp->irs;
323 	tw->t_starttime = tp->t_starttime;
324 	tw->tw_time = 0;
325 
326 /* XXX
327  * If this code will
328  * be used for fin-wait-2 state also, then we may need
329  * a ts_recent from the last segment.
330  */
331 	acknow = tp->t_flags & TF_ACKNOW;
332 
333 	/*
334 	 * First, discard tcpcb state, which includes stopping its timers and
335 	 * freeing it.  tcp_discardcb() used to also release the inpcb, but
336 	 * that work is now done in the caller.
337 	 *
338 	 * Note: soisdisconnected() call used to be made in tcp_discardcb(),
339 	 * and might not be needed here any longer.
340 	 */
341 	tcp_discardcb(tp);
342 	soisdisconnected(so);
343 	tw->tw_so_options = so->so_options;
344 	inp->inp_flags |= INP_TIMEWAIT;
345 	if (acknow)
346 		tcp_twrespond(tw, TH_ACK);
347 	if (local)
348 		in_pcbdrop(inp);
349 	else {
350 		in_pcbref(inp);	/* Reference from tw */
351 		tw->tw_cred = crhold(so->so_cred);
352 		inp->inp_ppcb = tw;
353 		TCPSTATES_INC(TCPS_TIME_WAIT);
354 		tcp_tw_2msl_reset(tw, 0);
355 	}
356 
357 	/*
358 	 * If the inpcb owns the sole reference to the socket, then we can
359 	 * detach and free the socket as it is not needed in time wait.
360 	 */
361 	if (inp->inp_flags & INP_SOCKREF) {
362 		KASSERT(so->so_state & SS_PROTOREF,
363 		    ("tcp_twstart: !SS_PROTOREF"));
364 		inp->inp_flags &= ~INP_SOCKREF;
365 		INP_WUNLOCK(inp);
366 		SOCK_LOCK(so);
367 		so->so_state &= ~SS_PROTOREF;
368 		sofree(so);
369 	} else
370 		INP_WUNLOCK(inp);
371 }
372 
373 /*
374  * Returns 1 if the TIME_WAIT state was killed and we should start over,
375  * looking for a pcb in the listen state.  Returns 0 otherwise.
376  */
377 int
378 tcp_twcheck(struct inpcb *inp, struct tcpopt *to __unused, struct tcphdr *th,
379     struct mbuf *m, int tlen)
380 {
381 	struct tcptw *tw;
382 	int thflags;
383 	tcp_seq seq;
384 
385 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
386 	INP_WLOCK_ASSERT(inp);
387 
388 	/*
389 	 * XXXRW: Time wait state for inpcb has been recycled, but inpcb is
390 	 * still present.  This is undesirable, but temporarily necessary
391 	 * until we work out how to handle inpcb's who's timewait state has
392 	 * been removed.
393 	 */
394 	tw = intotw(inp);
395 	if (tw == NULL)
396 		goto drop;
397 
398 	thflags = th->th_flags;
399 
400 	/*
401 	 * NOTE: for FIN_WAIT_2 (to be added later),
402 	 * must validate sequence number before accepting RST
403 	 */
404 
405 	/*
406 	 * If the segment contains RST:
407 	 *	Drop the segment - see Stevens, vol. 2, p. 964 and
408 	 *      RFC 1337.
409 	 */
410 	if (thflags & TH_RST)
411 		goto drop;
412 
413 #if 0
414 /* PAWS not needed at the moment */
415 	/*
416 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment
417 	 * and it's less than ts_recent, drop it.
418 	 */
419 	if ((to.to_flags & TOF_TS) != 0 && tp->ts_recent &&
420 	    TSTMP_LT(to.to_tsval, tp->ts_recent)) {
421 		if ((thflags & TH_ACK) == 0)
422 			goto drop;
423 		goto ack;
424 	}
425 	/*
426 	 * ts_recent is never updated because we never accept new segments.
427 	 */
428 #endif
429 
430 	/*
431 	 * If a new connection request is received
432 	 * while in TIME_WAIT, drop the old connection
433 	 * and start over if the sequence numbers
434 	 * are above the previous ones.
435 	 */
436 	if ((thflags & TH_SYN) && SEQ_GT(th->th_seq, tw->rcv_nxt)) {
437 		tcp_twclose(tw, 0);
438 		return (1);
439 	}
440 
441 	/*
442 	 * Drop the segment if it does not contain an ACK.
443 	 */
444 	if ((thflags & TH_ACK) == 0)
445 		goto drop;
446 
447 	/*
448 	 * Reset the 2MSL timer if this is a duplicate FIN.
449 	 */
450 	if (thflags & TH_FIN) {
451 		seq = th->th_seq + tlen + (thflags & TH_SYN ? 1 : 0);
452 		if (seq + 1 == tw->rcv_nxt)
453 			tcp_tw_2msl_reset(tw, 1);
454 	}
455 
456 	/*
457 	 * Acknowledge the segment if it has data or is not a duplicate ACK.
458 	 */
459 	if (thflags != TH_ACK || tlen != 0 ||
460 	    th->th_seq != tw->rcv_nxt || th->th_ack != tw->snd_nxt) {
461 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
462 		tcp_twrespond(tw, TH_ACK);
463 		goto dropnoprobe;
464 	}
465 drop:
466 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
467 dropnoprobe:
468 	INP_WUNLOCK(inp);
469 	m_freem(m);
470 	return (0);
471 }
472 
473 void
474 tcp_twclose(struct tcptw *tw, int reuse)
475 {
476 	struct socket *so;
477 	struct inpcb *inp;
478 
479 	/*
480 	 * At this point, we are in one of two situations:
481 	 *
482 	 * (1) We have no socket, just an inpcb<->twtcp pair.  We can free
483 	 *     all state.
484 	 *
485 	 * (2) We have a socket -- if we own a reference, release it and
486 	 *     notify the socket layer.
487 	 */
488 	inp = tw->tw_inpcb;
489 	KASSERT((inp->inp_flags & INP_TIMEWAIT), ("tcp_twclose: !timewait"));
490 	KASSERT(intotw(inp) == tw, ("tcp_twclose: inp_ppcb != tw"));
491 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);	/* in_pcbfree() */
492 	INP_WLOCK_ASSERT(inp);
493 
494 	tcp_tw_2msl_stop(tw, reuse);
495 	inp->inp_ppcb = NULL;
496 	in_pcbdrop(inp);
497 
498 	so = inp->inp_socket;
499 	if (so != NULL) {
500 		/*
501 		 * If there's a socket, handle two cases: first, we own a
502 		 * strong reference, which we will now release, or we don't
503 		 * in which case another reference exists (XXXRW: think
504 		 * about this more), and we don't need to take action.
505 		 */
506 		if (inp->inp_flags & INP_SOCKREF) {
507 			inp->inp_flags &= ~INP_SOCKREF;
508 			INP_WUNLOCK(inp);
509 			SOCK_LOCK(so);
510 			KASSERT(so->so_state & SS_PROTOREF,
511 			    ("tcp_twclose: INP_SOCKREF && !SS_PROTOREF"));
512 			so->so_state &= ~SS_PROTOREF;
513 			sofree(so);
514 		} else {
515 			/*
516 			 * If we don't own the only reference, the socket and
517 			 * inpcb need to be left around to be handled by
518 			 * tcp_usr_detach() later.
519 			 */
520 			INP_WUNLOCK(inp);
521 		}
522 	} else {
523 		/*
524 		 * The socket has been already cleaned-up for us, only free the
525 		 * inpcb.
526 		 */
527 		in_pcbfree(inp);
528 	}
529 	TCPSTAT_INC(tcps_closed);
530 }
531 
532 static int
533 tcp_twrespond(struct tcptw *tw, int flags)
534 {
535 	struct inpcb *inp = tw->tw_inpcb;
536 #if defined(INET6) || defined(INET)
537 	struct tcphdr *th = NULL;
538 #endif
539 	struct mbuf *m;
540 #ifdef INET
541 	struct ip *ip = NULL;
542 #endif
543 	u_int hdrlen, optlen;
544 	int error = 0;			/* Keep compiler happy */
545 	struct tcpopt to;
546 #ifdef INET6
547 	struct ip6_hdr *ip6 = NULL;
548 	int isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6;
549 #endif
550 	hdrlen = 0;                     /* Keep compiler happy */
551 
552 	INP_WLOCK_ASSERT(inp);
553 
554 	m = m_gethdr(M_NOWAIT, MT_DATA);
555 	if (m == NULL)
556 		return (ENOBUFS);
557 	m->m_data += max_linkhdr;
558 
559 #ifdef MAC
560 	mac_inpcb_create_mbuf(inp, m);
561 #endif
562 
563 #ifdef INET6
564 	if (isipv6) {
565 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
566 		ip6 = mtod(m, struct ip6_hdr *);
567 		th = (struct tcphdr *)(ip6 + 1);
568 		tcpip_fillheaders(inp, ip6, th);
569 	}
570 #endif
571 #if defined(INET6) && defined(INET)
572 	else
573 #endif
574 #ifdef INET
575 	{
576 		hdrlen = sizeof(struct tcpiphdr);
577 		ip = mtod(m, struct ip *);
578 		th = (struct tcphdr *)(ip + 1);
579 		tcpip_fillheaders(inp, ip, th);
580 	}
581 #endif
582 	to.to_flags = 0;
583 
584 	/*
585 	 * Send a timestamp and echo-reply if both our side and our peer
586 	 * have sent timestamps in our SYN's and this is not a RST.
587 	 */
588 	if (tw->t_recent && flags == TH_ACK) {
589 		to.to_flags |= TOF_TS;
590 		to.to_tsval = tcp_ts_getticks() + tw->ts_offset;
591 		to.to_tsecr = tw->t_recent;
592 	}
593 	optlen = tcp_addoptions(&to, (u_char *)(th + 1));
594 
595 	m->m_len = hdrlen + optlen;
596 	m->m_pkthdr.len = m->m_len;
597 
598 	KASSERT(max_linkhdr + m->m_len <= MHLEN, ("tcptw: mbuf too small"));
599 
600 	th->th_seq = htonl(tw->snd_nxt);
601 	th->th_ack = htonl(tw->rcv_nxt);
602 	th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
603 	th->th_flags = flags;
604 	th->th_win = htons(tw->last_win);
605 
606 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
607 #ifdef INET6
608 	if (isipv6) {
609 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
610 		th->th_sum = in6_cksum_pseudo(ip6,
611 		    sizeof(struct tcphdr) + optlen, IPPROTO_TCP, 0);
612 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
613 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
614 		error = ip6_output(m, inp->in6p_outputopts, NULL,
615 		    (tw->tw_so_options & SO_DONTROUTE), NULL, NULL, inp);
616 	}
617 #endif
618 #if defined(INET6) && defined(INET)
619 	else
620 #endif
621 #ifdef INET
622 	{
623 		m->m_pkthdr.csum_flags = CSUM_TCP;
624 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
625 		    htons(sizeof(struct tcphdr) + optlen + IPPROTO_TCP));
626 		ip->ip_len = htons(m->m_pkthdr.len);
627 		if (V_path_mtu_discovery)
628 			ip->ip_off |= htons(IP_DF);
629 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
630 		error = ip_output(m, inp->inp_options, NULL,
631 		    ((tw->tw_so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
632 		    NULL, inp);
633 	}
634 #endif
635 	if (flags & TH_ACK)
636 		TCPSTAT_INC(tcps_sndacks);
637 	else
638 		TCPSTAT_INC(tcps_sndctrl);
639 	TCPSTAT_INC(tcps_sndtotal);
640 	return (error);
641 }
642 
643 static void
644 tcp_tw_2msl_reset(struct tcptw *tw, int rearm)
645 {
646 
647 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
648 	INP_WLOCK_ASSERT(tw->tw_inpcb);
649 
650 	TW_WLOCK(V_tw_lock);
651 	if (rearm)
652 		TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl);
653 	tw->tw_time = ticks + 2 * tcp_msl;
654 	TAILQ_INSERT_TAIL(&V_twq_2msl, tw, tw_2msl);
655 	TW_WUNLOCK(V_tw_lock);
656 }
657 
658 static void
659 tcp_tw_2msl_stop(struct tcptw *tw, int reuse)
660 {
661 	struct ucred *cred;
662 	struct inpcb *inp;
663 	int released __unused;
664 
665 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
666 
667 	TW_WLOCK(V_tw_lock);
668 	inp = tw->tw_inpcb;
669 	tw->tw_inpcb = NULL;
670 
671 	TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl);
672 	cred = tw->tw_cred;
673 	tw->tw_cred = NULL;
674 	TW_WUNLOCK(V_tw_lock);
675 
676 	if (cred != NULL)
677 		crfree(cred);
678 
679 	released = in_pcbrele_wlocked(inp);
680 	KASSERT(!released, ("%s: inp should not be released here", __func__));
681 
682 	if (!reuse)
683 		uma_zfree(V_tcptw_zone, tw);
684 	TCPSTATES_DEC(TCPS_TIME_WAIT);
685 }
686 
687 struct tcptw *
688 tcp_tw_2msl_scan(int reuse)
689 {
690 	struct tcptw *tw;
691 	struct inpcb *inp;
692 	struct epoch_tracker et;
693 
694 #ifdef INVARIANTS
695 	if (reuse) {
696 		/*
697 		 * Exclusive pcbinfo lock is not required in reuse case even if
698 		 * two inpcb locks can be acquired simultaneously:
699 		 *  - the inpcb transitioning to TIME_WAIT state in
700 		 *    tcp_tw_start(),
701 		 *  - the inpcb closed by tcp_twclose().
702 		 *
703 		 * It is because only inpcbs in FIN_WAIT2 or CLOSING states can
704 		 * transition in TIME_WAIT state.  Then a pcbcb cannot be in
705 		 * TIME_WAIT list and transitioning to TIME_WAIT state at same
706 		 * time.
707 		 */
708 		INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
709 	}
710 #endif
711 
712 	for (;;) {
713 		TW_RLOCK(V_tw_lock);
714 		tw = TAILQ_FIRST(&V_twq_2msl);
715 		if (tw == NULL || (!reuse && (tw->tw_time - ticks) > 0)) {
716 			TW_RUNLOCK(V_tw_lock);
717 			break;
718 		}
719 		KASSERT(tw->tw_inpcb != NULL, ("%s: tw->tw_inpcb == NULL",
720 		    __func__));
721 
722 		inp = tw->tw_inpcb;
723 		in_pcbref(inp);
724 		TW_RUNLOCK(V_tw_lock);
725 
726 		INP_INFO_RLOCK_ET(&V_tcbinfo, et);
727 		INP_WLOCK(inp);
728 		tw = intotw(inp);
729 		if (in_pcbrele_wlocked(inp)) {
730 			if (__predict_true(tw == NULL)) {
731 				INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
732 				continue;
733 			} else {
734 				/* This should not happen as in TIMEWAIT
735 				 * state the inp should not be destroyed
736 				 * before its tcptw. If INVARIANTS is
737 				 * defined panic.
738 				 */
739 #ifdef INVARIANTS
740 				panic("%s: Panic before an infinite "
741 					  "loop: INP_TIMEWAIT && (INP_FREED "
742 					  "|| inp last reference) && tw != "
743 					  "NULL", __func__);
744 #else
745 				log(LOG_ERR, "%s: Avoid an infinite "
746 					"loop: INP_TIMEWAIT && (INP_FREED "
747 					"|| inp last reference) && tw != "
748 					"NULL", __func__);
749 #endif
750 				INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
751 				break;
752 			}
753 		}
754 
755 		if (tw == NULL) {
756 			/* tcp_twclose() has already been called */
757 			INP_WUNLOCK(inp);
758 			INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
759 			continue;
760 		}
761 
762 		tcp_twclose(tw, reuse);
763 		INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
764 		if (reuse)
765 			return tw;
766 	}
767 
768 	return NULL;
769 }
770