xref: /freebsd/sys/netinet/tcp_timewait.c (revision 5bf5ca772c6de2d53344a78cf461447cc322ccea)
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 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 *tw;
229 	struct inpcb *inp = tp->t_inpcb;
230 	int acknow;
231 	struct socket *so;
232 #ifdef INET6
233 	int 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 		int error = 0;
245 #ifdef INET6
246 		if (isipv6)
247 			error = in6_localaddr(&inp->in6p_faddr);
248 #endif
249 #if defined(INET6) && defined(INET)
250 		else
251 #endif
252 #ifdef INET
253 			error = in_localip(inp->inp_faddr);
254 #endif
255 		if (error) {
256 			tp = tcp_close(tp);
257 			if (tp != NULL)
258 				INP_WUNLOCK(inp);
259 			return;
260 		}
261 	}
262 
263 
264 	/*
265 	 * For use only by DTrace.  We do not reference the state
266 	 * after this point so modifying it in place is not a problem.
267 	 */
268 	tcp_state_change(tp, TCPS_TIME_WAIT);
269 
270 	tw = uma_zalloc(V_tcptw_zone, M_NOWAIT);
271 	if (tw == NULL) {
272 		/*
273 		 * Reached limit on total number of TIMEWAIT connections
274 		 * allowed. Remove a connection from TIMEWAIT queue in LRU
275 		 * fashion to make room for this connection.
276 		 *
277 		 * XXX:  Check if it possible to always have enough room
278 		 * in advance based on guarantees provided by uma_zalloc().
279 		 */
280 		tw = tcp_tw_2msl_scan(1);
281 		if (tw == NULL) {
282 			tp = tcp_close(tp);
283 			if (tp != NULL)
284 				INP_WUNLOCK(inp);
285 			return;
286 		}
287 	}
288 	/*
289 	 * The tcptw will hold a reference on its inpcb until tcp_twclose
290 	 * is called
291 	 */
292 	tw->tw_inpcb = inp;
293 	in_pcbref(inp);	/* Reference from tw */
294 
295 	/*
296 	 * Recover last window size sent.
297 	 */
298 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt))
299 		tw->last_win = (tp->rcv_adv - tp->rcv_nxt) >> tp->rcv_scale;
300 	else
301 		tw->last_win = 0;
302 
303 	/*
304 	 * Set t_recent if timestamps are used on the connection.
305 	 */
306 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_RCVD_TSTMP|TF_NOOPT)) ==
307 	    (TF_REQ_TSTMP|TF_RCVD_TSTMP)) {
308 		tw->t_recent = tp->ts_recent;
309 		tw->ts_offset = tp->ts_offset;
310 	} else {
311 		tw->t_recent = 0;
312 		tw->ts_offset = 0;
313 	}
314 
315 	tw->snd_nxt = tp->snd_nxt;
316 	tw->rcv_nxt = tp->rcv_nxt;
317 	tw->iss     = tp->iss;
318 	tw->irs     = tp->irs;
319 	tw->t_starttime = tp->t_starttime;
320 	tw->tw_time = 0;
321 
322 /* XXX
323  * If this code will
324  * be used for fin-wait-2 state also, then we may need
325  * a ts_recent from the last segment.
326  */
327 	acknow = tp->t_flags & TF_ACKNOW;
328 
329 	/*
330 	 * First, discard tcpcb state, which includes stopping its timers and
331 	 * freeing it.  tcp_discardcb() used to also release the inpcb, but
332 	 * that work is now done in the caller.
333 	 *
334 	 * Note: soisdisconnected() call used to be made in tcp_discardcb(),
335 	 * and might not be needed here any longer.
336 	 */
337 	tcp_discardcb(tp);
338 	so = inp->inp_socket;
339 	soisdisconnected(so);
340 	tw->tw_cred = crhold(so->so_cred);
341 	SOCK_LOCK(so);
342 	tw->tw_so_options = so->so_options;
343 	SOCK_UNLOCK(so);
344 	if (acknow)
345 		tcp_twrespond(tw, TH_ACK);
346 	inp->inp_ppcb = tw;
347 	inp->inp_flags |= INP_TIMEWAIT;
348 	TCPSTATES_INC(TCPS_TIME_WAIT);
349 	tcp_tw_2msl_reset(tw, 0);
350 
351 	/*
352 	 * If the inpcb owns the sole reference to the socket, then we can
353 	 * detach and free the socket as it is not needed in time wait.
354 	 */
355 	if (inp->inp_flags & INP_SOCKREF) {
356 		KASSERT(so->so_state & SS_PROTOREF,
357 		    ("tcp_twstart: !SS_PROTOREF"));
358 		inp->inp_flags &= ~INP_SOCKREF;
359 		INP_WUNLOCK(inp);
360 		SOCK_LOCK(so);
361 		so->so_state &= ~SS_PROTOREF;
362 		sofree(so);
363 	} else
364 		INP_WUNLOCK(inp);
365 }
366 
367 /*
368  * Returns 1 if the TIME_WAIT state was killed and we should start over,
369  * looking for a pcb in the listen state.  Returns 0 otherwise.
370  */
371 int
372 tcp_twcheck(struct inpcb *inp, struct tcpopt *to __unused, struct tcphdr *th,
373     struct mbuf *m, int tlen)
374 {
375 	struct tcptw *tw;
376 	int thflags;
377 	tcp_seq seq;
378 
379 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
380 	INP_WLOCK_ASSERT(inp);
381 
382 	/*
383 	 * XXXRW: Time wait state for inpcb has been recycled, but inpcb is
384 	 * still present.  This is undesirable, but temporarily necessary
385 	 * until we work out how to handle inpcb's who's timewait state has
386 	 * been removed.
387 	 */
388 	tw = intotw(inp);
389 	if (tw == NULL)
390 		goto drop;
391 
392 	thflags = th->th_flags;
393 
394 	/*
395 	 * NOTE: for FIN_WAIT_2 (to be added later),
396 	 * must validate sequence number before accepting RST
397 	 */
398 
399 	/*
400 	 * If the segment contains RST:
401 	 *	Drop the segment - see Stevens, vol. 2, p. 964 and
402 	 *      RFC 1337.
403 	 */
404 	if (thflags & TH_RST)
405 		goto drop;
406 
407 #if 0
408 /* PAWS not needed at the moment */
409 	/*
410 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment
411 	 * and it's less than ts_recent, drop it.
412 	 */
413 	if ((to.to_flags & TOF_TS) != 0 && tp->ts_recent &&
414 	    TSTMP_LT(to.to_tsval, tp->ts_recent)) {
415 		if ((thflags & TH_ACK) == 0)
416 			goto drop;
417 		goto ack;
418 	}
419 	/*
420 	 * ts_recent is never updated because we never accept new segments.
421 	 */
422 #endif
423 
424 	/*
425 	 * If a new connection request is received
426 	 * while in TIME_WAIT, drop the old connection
427 	 * and start over if the sequence numbers
428 	 * are above the previous ones.
429 	 */
430 	if ((thflags & TH_SYN) && SEQ_GT(th->th_seq, tw->rcv_nxt)) {
431 		tcp_twclose(tw, 0);
432 		return (1);
433 	}
434 
435 	/*
436 	 * Drop the segment if it does not contain an ACK.
437 	 */
438 	if ((thflags & TH_ACK) == 0)
439 		goto drop;
440 
441 	/*
442 	 * Reset the 2MSL timer if this is a duplicate FIN.
443 	 */
444 	if (thflags & TH_FIN) {
445 		seq = th->th_seq + tlen + (thflags & TH_SYN ? 1 : 0);
446 		if (seq + 1 == tw->rcv_nxt)
447 			tcp_tw_2msl_reset(tw, 1);
448 	}
449 
450 	/*
451 	 * Acknowledge the segment if it has data or is not a duplicate ACK.
452 	 */
453 	if (thflags != TH_ACK || tlen != 0 ||
454 	    th->th_seq != tw->rcv_nxt || th->th_ack != tw->snd_nxt)
455 		tcp_twrespond(tw, TH_ACK);
456 drop:
457 	INP_WUNLOCK(inp);
458 	m_freem(m);
459 	return (0);
460 }
461 
462 void
463 tcp_twclose(struct tcptw *tw, int reuse)
464 {
465 	struct socket *so;
466 	struct inpcb *inp;
467 
468 	/*
469 	 * At this point, we are in one of two situations:
470 	 *
471 	 * (1) We have no socket, just an inpcb<->twtcp pair.  We can free
472 	 *     all state.
473 	 *
474 	 * (2) We have a socket -- if we own a reference, release it and
475 	 *     notify the socket layer.
476 	 */
477 	inp = tw->tw_inpcb;
478 	KASSERT((inp->inp_flags & INP_TIMEWAIT), ("tcp_twclose: !timewait"));
479 	KASSERT(intotw(inp) == tw, ("tcp_twclose: inp_ppcb != tw"));
480 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);	/* in_pcbfree() */
481 	INP_WLOCK_ASSERT(inp);
482 
483 	tcp_tw_2msl_stop(tw, reuse);
484 	inp->inp_ppcb = NULL;
485 	in_pcbdrop(inp);
486 
487 	so = inp->inp_socket;
488 	if (so != NULL) {
489 		/*
490 		 * If there's a socket, handle two cases: first, we own a
491 		 * strong reference, which we will now release, or we don't
492 		 * in which case another reference exists (XXXRW: think
493 		 * about this more), and we don't need to take action.
494 		 */
495 		if (inp->inp_flags & INP_SOCKREF) {
496 			inp->inp_flags &= ~INP_SOCKREF;
497 			INP_WUNLOCK(inp);
498 			SOCK_LOCK(so);
499 			KASSERT(so->so_state & SS_PROTOREF,
500 			    ("tcp_twclose: INP_SOCKREF && !SS_PROTOREF"));
501 			so->so_state &= ~SS_PROTOREF;
502 			sofree(so);
503 		} else {
504 			/*
505 			 * If we don't own the only reference, the socket and
506 			 * inpcb need to be left around to be handled by
507 			 * tcp_usr_detach() later.
508 			 */
509 			INP_WUNLOCK(inp);
510 		}
511 	} else {
512 		/*
513 		 * The socket has been already cleaned-up for us, only free the
514 		 * inpcb.
515 		 */
516 		in_pcbfree(inp);
517 	}
518 	TCPSTAT_INC(tcps_closed);
519 }
520 
521 static int
522 tcp_twrespond(struct tcptw *tw, int flags)
523 {
524 	struct inpcb *inp = tw->tw_inpcb;
525 #if defined(INET6) || defined(INET)
526 	struct tcphdr *th = NULL;
527 #endif
528 	struct mbuf *m;
529 #ifdef INET
530 	struct ip *ip = NULL;
531 #endif
532 	u_int hdrlen, optlen;
533 	int error = 0;			/* Keep compiler happy */
534 	struct tcpopt to;
535 #ifdef INET6
536 	struct ip6_hdr *ip6 = NULL;
537 	int isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6;
538 #endif
539 	hdrlen = 0;                     /* Keep compiler happy */
540 
541 	INP_WLOCK_ASSERT(inp);
542 
543 	m = m_gethdr(M_NOWAIT, MT_DATA);
544 	if (m == NULL)
545 		return (ENOBUFS);
546 	m->m_data += max_linkhdr;
547 
548 #ifdef MAC
549 	mac_inpcb_create_mbuf(inp, m);
550 #endif
551 
552 #ifdef INET6
553 	if (isipv6) {
554 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
555 		ip6 = mtod(m, struct ip6_hdr *);
556 		th = (struct tcphdr *)(ip6 + 1);
557 		tcpip_fillheaders(inp, ip6, th);
558 	}
559 #endif
560 #if defined(INET6) && defined(INET)
561 	else
562 #endif
563 #ifdef INET
564 	{
565 		hdrlen = sizeof(struct tcpiphdr);
566 		ip = mtod(m, struct ip *);
567 		th = (struct tcphdr *)(ip + 1);
568 		tcpip_fillheaders(inp, ip, th);
569 	}
570 #endif
571 	to.to_flags = 0;
572 
573 	/*
574 	 * Send a timestamp and echo-reply if both our side and our peer
575 	 * have sent timestamps in our SYN's and this is not a RST.
576 	 */
577 	if (tw->t_recent && flags == TH_ACK) {
578 		to.to_flags |= TOF_TS;
579 		to.to_tsval = tcp_ts_getticks() + tw->ts_offset;
580 		to.to_tsecr = tw->t_recent;
581 	}
582 	optlen = tcp_addoptions(&to, (u_char *)(th + 1));
583 
584 	m->m_len = hdrlen + optlen;
585 	m->m_pkthdr.len = m->m_len;
586 
587 	KASSERT(max_linkhdr + m->m_len <= MHLEN, ("tcptw: mbuf too small"));
588 
589 	th->th_seq = htonl(tw->snd_nxt);
590 	th->th_ack = htonl(tw->rcv_nxt);
591 	th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
592 	th->th_flags = flags;
593 	th->th_win = htons(tw->last_win);
594 
595 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
596 #ifdef INET6
597 	if (isipv6) {
598 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
599 		th->th_sum = in6_cksum_pseudo(ip6,
600 		    sizeof(struct tcphdr) + optlen, IPPROTO_TCP, 0);
601 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
602 		error = ip6_output(m, inp->in6p_outputopts, NULL,
603 		    (tw->tw_so_options & SO_DONTROUTE), NULL, NULL, inp);
604 	}
605 #endif
606 #if defined(INET6) && defined(INET)
607 	else
608 #endif
609 #ifdef INET
610 	{
611 		m->m_pkthdr.csum_flags = CSUM_TCP;
612 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
613 		    htons(sizeof(struct tcphdr) + optlen + IPPROTO_TCP));
614 		ip->ip_len = htons(m->m_pkthdr.len);
615 		if (V_path_mtu_discovery)
616 			ip->ip_off |= htons(IP_DF);
617 		error = ip_output(m, inp->inp_options, NULL,
618 		    ((tw->tw_so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
619 		    NULL, inp);
620 	}
621 #endif
622 	if (flags & TH_ACK)
623 		TCPSTAT_INC(tcps_sndacks);
624 	else
625 		TCPSTAT_INC(tcps_sndctrl);
626 	TCPSTAT_INC(tcps_sndtotal);
627 	return (error);
628 }
629 
630 static void
631 tcp_tw_2msl_reset(struct tcptw *tw, int rearm)
632 {
633 
634 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
635 	INP_WLOCK_ASSERT(tw->tw_inpcb);
636 
637 	TW_WLOCK(V_tw_lock);
638 	if (rearm)
639 		TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl);
640 	tw->tw_time = ticks + 2 * tcp_msl;
641 	TAILQ_INSERT_TAIL(&V_twq_2msl, tw, tw_2msl);
642 	TW_WUNLOCK(V_tw_lock);
643 }
644 
645 static void
646 tcp_tw_2msl_stop(struct tcptw *tw, int reuse)
647 {
648 	struct ucred *cred;
649 	struct inpcb *inp;
650 	int released;
651 
652 	INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
653 
654 	TW_WLOCK(V_tw_lock);
655 	inp = tw->tw_inpcb;
656 	tw->tw_inpcb = NULL;
657 
658 	TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl);
659 	cred = tw->tw_cred;
660 	tw->tw_cred = NULL;
661 	TW_WUNLOCK(V_tw_lock);
662 
663 	if (cred != NULL)
664 		crfree(cred);
665 
666 	released = in_pcbrele_wlocked(inp);
667 	KASSERT(!released, ("%s: inp should not be released here", __func__));
668 
669 	if (!reuse)
670 		uma_zfree(V_tcptw_zone, tw);
671 	TCPSTATES_DEC(TCPS_TIME_WAIT);
672 }
673 
674 struct tcptw *
675 tcp_tw_2msl_scan(int reuse)
676 {
677 	struct tcptw *tw;
678 	struct inpcb *inp;
679 
680 #ifdef INVARIANTS
681 	if (reuse) {
682 		/*
683 		 * Exclusive pcbinfo lock is not required in reuse case even if
684 		 * two inpcb locks can be acquired simultaneously:
685 		 *  - the inpcb transitioning to TIME_WAIT state in
686 		 *    tcp_tw_start(),
687 		 *  - the inpcb closed by tcp_twclose().
688 		 *
689 		 * It is because only inpcbs in FIN_WAIT2 or CLOSING states can
690 		 * transition in TIME_WAIT state.  Then a pcbcb cannot be in
691 		 * TIME_WAIT list and transitioning to TIME_WAIT state at same
692 		 * time.
693 		 */
694 		INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
695 	}
696 #endif
697 
698 	for (;;) {
699 		TW_RLOCK(V_tw_lock);
700 		tw = TAILQ_FIRST(&V_twq_2msl);
701 		if (tw == NULL || (!reuse && (tw->tw_time - ticks) > 0)) {
702 			TW_RUNLOCK(V_tw_lock);
703 			break;
704 		}
705 		KASSERT(tw->tw_inpcb != NULL, ("%s: tw->tw_inpcb == NULL",
706 		    __func__));
707 
708 		inp = tw->tw_inpcb;
709 		in_pcbref(inp);
710 		TW_RUNLOCK(V_tw_lock);
711 
712 		if (INP_INFO_TRY_RLOCK(&V_tcbinfo)) {
713 
714 			INP_WLOCK(inp);
715 			tw = intotw(inp);
716 			if (in_pcbrele_wlocked(inp)) {
717 				if (__predict_true(tw == NULL)) {
718 					INP_INFO_RUNLOCK(&V_tcbinfo);
719 					continue;
720 				} else {
721 					/* This should not happen as in TIMEWAIT
722 					 * state the inp should not be destroyed
723 					 * before its tcptw. If INVARIANTS is
724 					 * defined panic.
725 					 */
726 #ifdef INVARIANTS
727 					panic("%s: Panic before an infinite "
728 					    "loop: INP_TIMEWAIT && (INP_FREED "
729 					    "|| inp last reference) && tw != "
730 					    "NULL", __func__);
731 #else
732 					log(LOG_ERR, "%s: Avoid an infinite "
733 					    "loop: INP_TIMEWAIT && (INP_FREED "
734 					    "|| inp last reference) && tw != "
735 					    "NULL", __func__);
736 #endif
737 					INP_INFO_RUNLOCK(&V_tcbinfo);
738 					break;
739 				}
740 			}
741 
742 			if (tw == NULL) {
743 				/* tcp_twclose() has already been called */
744 				INP_WUNLOCK(inp);
745 				INP_INFO_RUNLOCK(&V_tcbinfo);
746 				continue;
747 			}
748 
749 			tcp_twclose(tw, reuse);
750 			INP_INFO_RUNLOCK(&V_tcbinfo);
751 			if (reuse)
752 			    return tw;
753 		} else {
754 			/* INP_INFO lock is busy, continue later. */
755 			INP_WLOCK(inp);
756 			if (!in_pcbrele_wlocked(inp))
757 				INP_WUNLOCK(inp);
758 			break;
759 		}
760 	}
761 
762 	return NULL;
763 }
764