xref: /freebsd/sys/netinet/tcp_timewait.c (revision 3332f1b444d4a73238e9f59cca27bfc95fe936bd)
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 <netinet/udp.h>
97 #include <netinet/udp_var.h>
98 #include <machine/in_cksum.h>
99 
100 #include <security/mac/mac_framework.h>
101 
102 VNET_DEFINE_STATIC(uma_zone_t, tcptw_zone);
103 #define	V_tcptw_zone		VNET(tcptw_zone)
104 static int	maxtcptw;
105 
106 /*
107  * The timed wait queue contains references to each of the TCP sessions
108  * currently in the TIME_WAIT state.  The queue pointers, including the
109  * queue pointers in each tcptw structure, are protected using the global
110  * timewait lock, which must be held over queue iteration and modification.
111  *
112  * Rules on tcptw usage:
113  *  - a inpcb is always freed _after_ its tcptw
114  *  - a tcptw relies on its inpcb reference counting for memory stability
115  *  - a tcptw is dereferenceable only while its inpcb is locked
116  */
117 VNET_DEFINE_STATIC(TAILQ_HEAD(, tcptw), twq_2msl);
118 #define	V_twq_2msl		VNET(twq_2msl)
119 
120 /* Global timewait lock */
121 VNET_DEFINE_STATIC(struct rwlock, tw_lock);
122 #define	V_tw_lock		VNET(tw_lock)
123 
124 #define	TW_LOCK_INIT(tw, d)	rw_init_flags(&(tw), (d), 0)
125 #define	TW_LOCK_DESTROY(tw)	rw_destroy(&(tw))
126 #define	TW_RLOCK(tw)		rw_rlock(&(tw))
127 #define	TW_WLOCK(tw)		rw_wlock(&(tw))
128 #define	TW_RUNLOCK(tw)		rw_runlock(&(tw))
129 #define	TW_WUNLOCK(tw)		rw_wunlock(&(tw))
130 #define	TW_LOCK_ASSERT(tw)	rw_assert(&(tw), RA_LOCKED)
131 #define	TW_RLOCK_ASSERT(tw)	rw_assert(&(tw), RA_RLOCKED)
132 #define	TW_WLOCK_ASSERT(tw)	rw_assert(&(tw), RA_WLOCKED)
133 #define	TW_UNLOCK_ASSERT(tw)	rw_assert(&(tw), RA_UNLOCKED)
134 
135 static void	tcp_tw_2msl_reset(struct tcptw *, int);
136 static void	tcp_tw_2msl_stop(struct tcptw *, int);
137 static int	tcp_twrespond(struct tcptw *, int);
138 
139 static int
140 tcptw_auto_size(void)
141 {
142 	int halfrange;
143 
144 	/*
145 	 * Max out at half the ephemeral port range so that TIME_WAIT
146 	 * sockets don't tie up too many ephemeral ports.
147 	 */
148 	if (V_ipport_lastauto > V_ipport_firstauto)
149 		halfrange = (V_ipport_lastauto - V_ipport_firstauto) / 2;
150 	else
151 		halfrange = (V_ipport_firstauto - V_ipport_lastauto) / 2;
152 	/* Protect against goofy port ranges smaller than 32. */
153 	return (imin(imax(halfrange, 32), maxsockets / 5));
154 }
155 
156 static int
157 sysctl_maxtcptw(SYSCTL_HANDLER_ARGS)
158 {
159 	int error, new;
160 
161 	if (maxtcptw == 0)
162 		new = tcptw_auto_size();
163 	else
164 		new = maxtcptw;
165 	error = sysctl_handle_int(oidp, &new, 0, req);
166 	if (error == 0 && req->newptr)
167 		if (new >= 32) {
168 			maxtcptw = new;
169 			uma_zone_set_max(V_tcptw_zone, maxtcptw);
170 		}
171 	return (error);
172 }
173 
174 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, maxtcptw,
175     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
176     &maxtcptw, 0, sysctl_maxtcptw, "IU",
177     "Maximum number of compressed TCP TIME_WAIT entries");
178 
179 VNET_DEFINE_STATIC(bool, nolocaltimewait) = true;
180 #define	V_nolocaltimewait	VNET(nolocaltimewait)
181 SYSCTL_BOOL(_net_inet_tcp, OID_AUTO, nolocaltimewait, CTLFLAG_VNET | CTLFLAG_RW,
182     &VNET_NAME(nolocaltimewait), true,
183     "Do not create compressed TCP TIME_WAIT entries for local connections");
184 
185 void
186 tcp_tw_zone_change(void)
187 {
188 
189 	if (maxtcptw == 0)
190 		uma_zone_set_max(V_tcptw_zone, tcptw_auto_size());
191 }
192 
193 void
194 tcp_tw_init(void)
195 {
196 
197 	V_tcptw_zone = uma_zcreate("tcptw", sizeof(struct tcptw),
198 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
199 	TUNABLE_INT_FETCH("net.inet.tcp.maxtcptw", &maxtcptw);
200 	if (maxtcptw == 0)
201 		uma_zone_set_max(V_tcptw_zone, tcptw_auto_size());
202 	else
203 		uma_zone_set_max(V_tcptw_zone, maxtcptw);
204 	TAILQ_INIT(&V_twq_2msl);
205 	TW_LOCK_INIT(V_tw_lock, "tcptw");
206 }
207 
208 #ifdef VIMAGE
209 void
210 tcp_tw_destroy(void)
211 {
212 	struct tcptw *tw;
213 	struct epoch_tracker et;
214 
215 	NET_EPOCH_ENTER(et);
216 	while ((tw = TAILQ_FIRST(&V_twq_2msl)) != NULL)
217 		tcp_twclose(tw, 0);
218 	NET_EPOCH_EXIT(et);
219 
220 	TW_LOCK_DESTROY(V_tw_lock);
221 	uma_zdestroy(V_tcptw_zone);
222 }
223 #endif
224 
225 /*
226  * Move a TCP connection into TIME_WAIT state.
227  *    tcbinfo is locked.
228  *    inp is locked, and is unlocked before returning.
229  */
230 void
231 tcp_twstart(struct tcpcb *tp)
232 {
233 	struct tcptw twlocal, *tw;
234 	struct inpcb *inp = tp->t_inpcb;
235 	struct socket *so;
236 	uint32_t recwin;
237 	bool acknow, local;
238 #ifdef INET6
239 	bool isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6;
240 #endif
241 
242 	NET_EPOCH_ASSERT();
243 	INP_WLOCK_ASSERT(inp);
244 
245 	/* A dropped inp should never transition to TIME_WAIT state. */
246 	KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("tcp_twstart: "
247 	    "(inp->inp_flags & INP_DROPPED) != 0"));
248 
249 	if (V_nolocaltimewait) {
250 #ifdef INET6
251 		if (isipv6)
252 			local = in6_localaddr(&inp->in6p_faddr);
253 		else
254 #endif
255 #ifdef INET
256 			local = in_localip(inp->inp_faddr);
257 #else
258 			local = false;
259 #endif
260 	} else
261 		local = false;
262 
263 	/*
264 	 * For use only by DTrace.  We do not reference the state
265 	 * after this point so modifying it in place is not a problem.
266 	 */
267 	tcp_state_change(tp, TCPS_TIME_WAIT);
268 
269 	if (local)
270 		tw = &twlocal;
271 	else
272 		tw = uma_zalloc(V_tcptw_zone, M_NOWAIT);
273 	if (tw == NULL) {
274 		/*
275 		 * Reached limit on total number of TIMEWAIT connections
276 		 * allowed. Remove a connection from TIMEWAIT queue in LRU
277 		 * fashion to make room for this connection.
278 		 *
279 		 * XXX:  Check if it possible to always have enough room
280 		 * in advance based on guarantees provided by uma_zalloc().
281 		 */
282 		tw = tcp_tw_2msl_scan(1);
283 		if (tw == NULL) {
284 			tp = tcp_close(tp);
285 			if (tp != NULL)
286 				INP_WUNLOCK(inp);
287 			return;
288 		}
289 	}
290 	/*
291 	 * For !local case the tcptw will hold a reference on its inpcb
292 	 * until tcp_twclose is called.
293 	 */
294 	tw->tw_inpcb = inp;
295 
296 	/*
297 	 * Recover last window size sent.
298 	 */
299 	so = inp->inp_socket;
300 	recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
301 	    (long)TCP_MAXWIN << tp->rcv_scale);
302 	if (recwin < (so->so_rcv.sb_hiwat / 4) &&
303 	    recwin < tp->t_maxseg)
304 		recwin = 0;
305 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
306 	    recwin < (tp->rcv_adv - tp->rcv_nxt))
307 		recwin = (tp->rcv_adv - tp->rcv_nxt);
308 	tw->last_win = (u_short)(recwin >> tp->rcv_scale);
309 
310 	/*
311 	 * Set t_recent if timestamps are used on the connection.
312 	 */
313 	if ((tp->t_flags & (TF_REQ_TSTMP|TF_RCVD_TSTMP|TF_NOOPT)) ==
314 	    (TF_REQ_TSTMP|TF_RCVD_TSTMP)) {
315 		tw->t_recent = tp->ts_recent;
316 		tw->ts_offset = tp->ts_offset;
317 	} else {
318 		tw->t_recent = 0;
319 		tw->ts_offset = 0;
320 	}
321 
322 	tw->snd_nxt = tp->snd_nxt;
323 	tw->t_port = tp->t_port;
324 	tw->rcv_nxt = tp->rcv_nxt;
325 	tw->iss     = tp->iss;
326 	tw->irs     = tp->irs;
327 	tw->t_starttime = tp->t_starttime;
328 	tw->tw_time = 0;
329 
330 /* XXX
331  * If this code will
332  * be used for fin-wait-2 state also, then we may need
333  * a ts_recent from the last segment.
334  */
335 	acknow = tp->t_flags & TF_ACKNOW;
336 
337 	/*
338 	 * First, discard tcpcb state, which includes stopping its timers and
339 	 * freeing it.  tcp_discardcb() used to also release the inpcb, but
340 	 * that work is now done in the caller.
341 	 *
342 	 * Note: soisdisconnected() call used to be made in tcp_discardcb(),
343 	 * and might not be needed here any longer.
344 	 */
345 	tcp_discardcb(tp);
346 	soisdisconnected(so);
347 	tw->tw_so_options = so->so_options;
348 	inp->inp_flags |= INP_TIMEWAIT;
349 	if (acknow)
350 		tcp_twrespond(tw, TH_ACK);
351 	if (local)
352 		in_pcbdrop(inp);
353 	else {
354 		in_pcbref(inp);	/* Reference from tw */
355 		tw->tw_cred = crhold(so->so_cred);
356 		inp->inp_ppcb = tw;
357 		TCPSTATES_INC(TCPS_TIME_WAIT);
358 		tcp_tw_2msl_reset(tw, 0);
359 	}
360 
361 	/*
362 	 * If the inpcb owns the sole reference to the socket, then we can
363 	 * detach and free the socket as it is not needed in time wait.
364 	 */
365 	if (inp->inp_flags & INP_SOCKREF) {
366 		KASSERT(so->so_state & SS_PROTOREF,
367 		    ("tcp_twstart: !SS_PROTOREF"));
368 		inp->inp_flags &= ~INP_SOCKREF;
369 		INP_WUNLOCK(inp);
370 		SOCK_LOCK(so);
371 		so->so_state &= ~SS_PROTOREF;
372 		sofree(so);
373 	} else
374 		INP_WUNLOCK(inp);
375 }
376 
377 /*
378  * Returns 1 if the TIME_WAIT state was killed and we should start over,
379  * looking for a pcb in the listen state.  Returns 0 otherwise.
380  *
381  * For pure SYN-segments the PCB shall be read-locked and the tcpopt pointer
382  * may be NULL.  For the rest write-lock and valid tcpopt.
383  */
384 int
385 tcp_twcheck(struct inpcb *inp, struct tcpopt *to, struct tcphdr *th,
386     struct mbuf *m, int tlen)
387 {
388 	struct tcptw *tw;
389 	int thflags;
390 	tcp_seq seq;
391 
392 	NET_EPOCH_ASSERT();
393 	INP_LOCK_ASSERT(inp);
394 
395 	/*
396 	 * XXXRW: Time wait state for inpcb has been recycled, but inpcb is
397 	 * still present.  This is undesirable, but temporarily necessary
398 	 * until we work out how to handle inpcb's who's timewait state has
399 	 * been removed.
400 	 */
401 	tw = intotw(inp);
402 	if (tw == NULL)
403 		goto drop;
404 
405 	thflags = th->th_flags;
406 #ifdef INVARIANTS
407 	if ((thflags & (TH_SYN | TH_ACK)) == TH_SYN)
408 		INP_RLOCK_ASSERT(inp);
409 	else {
410 		INP_WLOCK_ASSERT(inp);
411 		KASSERT(to != NULL,
412 		    ("%s: called without options on a non-SYN segment",
413 		    __func__));
414 	}
415 #endif
416 
417 	/*
418 	 * NOTE: for FIN_WAIT_2 (to be added later),
419 	 * must validate sequence number before accepting RST
420 	 */
421 
422 	/*
423 	 * If the segment contains RST:
424 	 *	Drop the segment - see Stevens, vol. 2, p. 964 and
425 	 *      RFC 1337.
426 	 */
427 	if (thflags & TH_RST)
428 		goto drop;
429 
430 #if 0
431 /* PAWS not needed at the moment */
432 	/*
433 	 * RFC 1323 PAWS: If we have a timestamp reply on this segment
434 	 * and it's less than ts_recent, drop it.
435 	 */
436 	if ((to.to_flags & TOF_TS) != 0 && tp->ts_recent &&
437 	    TSTMP_LT(to.to_tsval, tp->ts_recent)) {
438 		if ((thflags & TH_ACK) == 0)
439 			goto drop;
440 		goto ack;
441 	}
442 	/*
443 	 * ts_recent is never updated because we never accept new segments.
444 	 */
445 #endif
446 
447 	/*
448 	 * If a new connection request is received
449 	 * while in TIME_WAIT, drop the old connection
450 	 * and start over if the sequence numbers
451 	 * are above the previous ones.
452 	 * Allow UDP port number changes in this case.
453 	 */
454 	if ((thflags & TH_SYN) && SEQ_GT(th->th_seq, tw->rcv_nxt)) {
455 		/*
456 		 * In case we can't upgrade our lock just pretend we have
457 		 * lost this packet.
458 		 */
459 		if (((thflags & (TH_SYN | TH_ACK)) == TH_SYN) &&
460 		    INP_TRY_UPGRADE(inp) == 0)
461 			goto drop;
462 		tcp_twclose(tw, 0);
463 		return (1);
464 	}
465 
466 	/*
467 	 * Send RST if UDP port numbers don't match
468 	 */
469 	if (tw->t_port != m->m_pkthdr.tcp_tun_port) {
470 		if (th->th_flags & TH_ACK) {
471 			tcp_respond(NULL, mtod(m, void *), th, m,
472 			    (tcp_seq)0, th->th_ack, TH_RST);
473 		} else {
474 			if (th->th_flags & TH_SYN)
475 				tlen++;
476 			if (th->th_flags & TH_FIN)
477 				tlen++;
478 			tcp_respond(NULL, mtod(m, void *), th, m,
479 			    th->th_seq+tlen, (tcp_seq)0, TH_RST|TH_ACK);
480 		}
481 		INP_UNLOCK(inp);
482 		return (0);
483 	}
484 
485 	/*
486 	 * Drop the segment if it does not contain an ACK.
487 	 */
488 	if ((thflags & TH_ACK) == 0)
489 		goto drop;
490 
491 	INP_WLOCK_ASSERT(inp);
492 
493 	/*
494 	 * If timestamps were negotiated during SYN/ACK and a
495 	 * segment without a timestamp is received, silently drop
496 	 * the segment, unless the missing timestamps are tolerated.
497 	 * See section 3.2 of RFC 7323.
498 	 */
499 	if (((to->to_flags & TOF_TS) == 0) && (tw->t_recent != 0) &&
500 	    (V_tcp_tolerate_missing_ts == 0)) {
501 		goto drop;
502 	}
503 
504 	/*
505 	 * Reset the 2MSL timer if this is a duplicate FIN.
506 	 */
507 	if (thflags & TH_FIN) {
508 		seq = th->th_seq + tlen + (thflags & TH_SYN ? 1 : 0);
509 		if (seq + 1 == tw->rcv_nxt)
510 			tcp_tw_2msl_reset(tw, 1);
511 	}
512 
513 	/*
514 	 * Acknowledge the segment if it has data or is not a duplicate ACK.
515 	 */
516 	if (thflags != TH_ACK || tlen != 0 ||
517 	    th->th_seq != tw->rcv_nxt || th->th_ack != tw->snd_nxt) {
518 		TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
519 		tcp_twrespond(tw, TH_ACK);
520 		goto dropnoprobe;
521 	}
522 drop:
523 	TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
524 dropnoprobe:
525 	INP_UNLOCK(inp);
526 	m_freem(m);
527 	return (0);
528 }
529 
530 void
531 tcp_twclose(struct tcptw *tw, int reuse)
532 {
533 	struct socket *so;
534 	struct inpcb *inp;
535 
536 	/*
537 	 * At this point, we are in one of two situations:
538 	 *
539 	 * (1) We have no socket, just an inpcb<->twtcp pair.  We can free
540 	 *     all state.
541 	 *
542 	 * (2) We have a socket -- if we own a reference, release it and
543 	 *     notify the socket layer.
544 	 */
545 	inp = tw->tw_inpcb;
546 	KASSERT((inp->inp_flags & INP_TIMEWAIT), ("tcp_twclose: !timewait"));
547 	KASSERT(intotw(inp) == tw, ("tcp_twclose: inp_ppcb != tw"));
548 	NET_EPOCH_ASSERT();
549 	INP_WLOCK_ASSERT(inp);
550 
551 	tcp_tw_2msl_stop(tw, reuse);
552 	inp->inp_ppcb = NULL;
553 	in_pcbdrop(inp);
554 
555 	so = inp->inp_socket;
556 	if (so != NULL) {
557 		/*
558 		 * If there's a socket, handle two cases: first, we own a
559 		 * strong reference, which we will now release, or we don't
560 		 * in which case another reference exists (XXXRW: think
561 		 * about this more), and we don't need to take action.
562 		 */
563 		if (inp->inp_flags & INP_SOCKREF) {
564 			inp->inp_flags &= ~INP_SOCKREF;
565 			INP_WUNLOCK(inp);
566 			SOCK_LOCK(so);
567 			KASSERT(so->so_state & SS_PROTOREF,
568 			    ("tcp_twclose: INP_SOCKREF && !SS_PROTOREF"));
569 			so->so_state &= ~SS_PROTOREF;
570 			sofree(so);
571 		} else {
572 			/*
573 			 * If we don't own the only reference, the socket and
574 			 * inpcb need to be left around to be handled by
575 			 * tcp_usr_detach() later.
576 			 */
577 			INP_WUNLOCK(inp);
578 		}
579 	} else {
580 		/*
581 		 * The socket has been already cleaned-up for us, only free the
582 		 * inpcb.
583 		 */
584 		in_pcbfree(inp);
585 	}
586 	TCPSTAT_INC(tcps_closed);
587 }
588 
589 static int
590 tcp_twrespond(struct tcptw *tw, int flags)
591 {
592 	struct inpcb *inp = tw->tw_inpcb;
593 #if defined(INET6) || defined(INET)
594 	struct tcphdr *th = NULL;
595 #endif
596 	struct mbuf *m;
597 #ifdef INET
598 	struct ip *ip = NULL;
599 #endif
600 	u_int hdrlen, optlen, ulen;
601 	int error = 0;			/* Keep compiler happy */
602 	struct tcpopt to;
603 #ifdef INET6
604 	struct ip6_hdr *ip6 = NULL;
605 	int isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6;
606 #endif
607 	struct udphdr *udp = NULL;
608 	hdrlen = 0;                     /* Keep compiler happy */
609 
610 	INP_WLOCK_ASSERT(inp);
611 
612 	m = m_gethdr(M_NOWAIT, MT_DATA);
613 	if (m == NULL)
614 		return (ENOBUFS);
615 	m->m_data += max_linkhdr;
616 
617 #ifdef MAC
618 	mac_inpcb_create_mbuf(inp, m);
619 #endif
620 
621 #ifdef INET6
622 	if (isipv6) {
623 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
624 		ip6 = mtod(m, struct ip6_hdr *);
625 		if (tw->t_port) {
626 			udp = (struct udphdr *)(ip6 + 1);
627 			hdrlen += sizeof(struct udphdr);
628 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
629 			udp->uh_dport = tw->t_port;
630 			ulen = (hdrlen - sizeof(struct ip6_hdr));
631 			th = (struct tcphdr *)(udp + 1);
632 		} else
633 			th = (struct tcphdr *)(ip6 + 1);
634 		tcpip_fillheaders(inp, tw->t_port, ip6, th);
635 	}
636 #endif
637 #if defined(INET6) && defined(INET)
638 	else
639 #endif
640 #ifdef INET
641 	{
642 		hdrlen = sizeof(struct tcpiphdr);
643 		ip = mtod(m, struct ip *);
644 		if (tw->t_port) {
645 			udp = (struct udphdr *)(ip + 1);
646 			hdrlen += sizeof(struct udphdr);
647 			udp->uh_sport = htons(V_tcp_udp_tunneling_port);
648 			udp->uh_dport = tw->t_port;
649 			ulen = (hdrlen - sizeof(struct ip));
650 			th = (struct tcphdr *)(udp + 1);
651 		} else
652 			th = (struct tcphdr *)(ip + 1);
653 		tcpip_fillheaders(inp, tw->t_port, ip, th);
654 	}
655 #endif
656 	to.to_flags = 0;
657 
658 	/*
659 	 * Send a timestamp and echo-reply if both our side and our peer
660 	 * have sent timestamps in our SYN's and this is not a RST.
661 	 */
662 	if (tw->t_recent && flags == TH_ACK) {
663 		to.to_flags |= TOF_TS;
664 		to.to_tsval = tcp_ts_getticks() + tw->ts_offset;
665 		to.to_tsecr = tw->t_recent;
666 	}
667 	optlen = tcp_addoptions(&to, (u_char *)(th + 1));
668 
669 	if (udp) {
670 		ulen += optlen;
671 		udp->uh_ulen = htons(ulen);
672 	}
673 	m->m_len = hdrlen + optlen;
674 	m->m_pkthdr.len = m->m_len;
675 
676 	KASSERT(max_linkhdr + m->m_len <= MHLEN, ("tcptw: mbuf too small"));
677 
678 	th->th_seq = htonl(tw->snd_nxt);
679 	th->th_ack = htonl(tw->rcv_nxt);
680 	th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
681 	th->th_flags = flags;
682 	th->th_win = htons(tw->last_win);
683 
684 #ifdef INET6
685 	if (isipv6) {
686 		if (tw->t_port) {
687 			m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
688 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
689 			udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
690 			th->th_sum = htons(0);
691 		} else {
692 			m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
693 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
694 			th->th_sum = in6_cksum_pseudo(ip6,
695 			    sizeof(struct tcphdr) + optlen, IPPROTO_TCP, 0);
696 		}
697 		ip6->ip6_hlim = in6_selecthlim(inp, NULL);
698 		TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
699 		error = ip6_output(m, inp->in6p_outputopts, NULL,
700 		    (tw->tw_so_options & SO_DONTROUTE), NULL, NULL, inp);
701 	}
702 #endif
703 #if defined(INET6) && defined(INET)
704 	else
705 #endif
706 #ifdef INET
707 	{
708 		if (tw->t_port) {
709 			m->m_pkthdr.csum_flags = CSUM_UDP;
710 			m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
711 			udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
712 			    ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
713 			th->th_sum = htons(0);
714 		} else {
715 			m->m_pkthdr.csum_flags = CSUM_TCP;
716 			m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
717 			th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
718 			    htons(sizeof(struct tcphdr) + optlen + IPPROTO_TCP));
719 		}
720 		ip->ip_len = htons(m->m_pkthdr.len);
721 		if (V_path_mtu_discovery)
722 			ip->ip_off |= htons(IP_DF);
723 		TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
724 		error = ip_output(m, inp->inp_options, NULL,
725 		    ((tw->tw_so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
726 		    NULL, inp);
727 	}
728 #endif
729 	if (flags & TH_ACK)
730 		TCPSTAT_INC(tcps_sndacks);
731 	else
732 		TCPSTAT_INC(tcps_sndctrl);
733 	TCPSTAT_INC(tcps_sndtotal);
734 	return (error);
735 }
736 
737 static void
738 tcp_tw_2msl_reset(struct tcptw *tw, int rearm)
739 {
740 
741 	NET_EPOCH_ASSERT();
742 	INP_WLOCK_ASSERT(tw->tw_inpcb);
743 
744 	TW_WLOCK(V_tw_lock);
745 	if (rearm)
746 		TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl);
747 	tw->tw_time = ticks + 2 * tcp_msl;
748 	TAILQ_INSERT_TAIL(&V_twq_2msl, tw, tw_2msl);
749 	TW_WUNLOCK(V_tw_lock);
750 }
751 
752 static void
753 tcp_tw_2msl_stop(struct tcptw *tw, int reuse)
754 {
755 	struct ucred *cred;
756 	struct inpcb *inp;
757 	int released __unused;
758 
759 	NET_EPOCH_ASSERT();
760 
761 	TW_WLOCK(V_tw_lock);
762 	inp = tw->tw_inpcb;
763 	tw->tw_inpcb = NULL;
764 
765 	TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl);
766 	cred = tw->tw_cred;
767 	tw->tw_cred = NULL;
768 	TW_WUNLOCK(V_tw_lock);
769 
770 	if (cred != NULL)
771 		crfree(cred);
772 
773 	released = in_pcbrele_wlocked(inp);
774 	KASSERT(!released, ("%s: inp should not be released here", __func__));
775 
776 	if (!reuse)
777 		uma_zfree(V_tcptw_zone, tw);
778 	TCPSTATES_DEC(TCPS_TIME_WAIT);
779 }
780 
781 struct tcptw *
782 tcp_tw_2msl_scan(int reuse)
783 {
784 	struct tcptw *tw;
785 	struct inpcb *inp;
786 
787 	NET_EPOCH_ASSERT();
788 
789 	for (;;) {
790 		TW_RLOCK(V_tw_lock);
791 		tw = TAILQ_FIRST(&V_twq_2msl);
792 		if (tw == NULL || (!reuse && (tw->tw_time - ticks) > 0)) {
793 			TW_RUNLOCK(V_tw_lock);
794 			break;
795 		}
796 		KASSERT(tw->tw_inpcb != NULL, ("%s: tw->tw_inpcb == NULL",
797 		    __func__));
798 
799 		inp = tw->tw_inpcb;
800 		in_pcbref(inp);
801 		TW_RUNLOCK(V_tw_lock);
802 
803 		INP_WLOCK(inp);
804 		tw = intotw(inp);
805 		if (in_pcbrele_wlocked(inp)) {
806 			if (__predict_true(tw == NULL)) {
807 				continue;
808 			} else {
809 				/* This should not happen as in TIMEWAIT
810 				 * state the inp should not be destroyed
811 				 * before its tcptw. If INVARIANTS is
812 				 * defined panic.
813 				 */
814 #ifdef INVARIANTS
815 				panic("%s: Panic before an infinite "
816 					  "loop: INP_TIMEWAIT && (INP_FREED "
817 					  "|| inp last reference) && tw != "
818 					  "NULL", __func__);
819 #else
820 				log(LOG_ERR, "%s: Avoid an infinite "
821 					"loop: INP_TIMEWAIT && (INP_FREED "
822 					"|| inp last reference) && tw != "
823 					"NULL", __func__);
824 #endif
825 				break;
826 			}
827 		}
828 
829 		if (tw == NULL) {
830 			/* tcp_twclose() has already been called */
831 			INP_WUNLOCK(inp);
832 			continue;
833 		}
834 
835 		tcp_twclose(tw, reuse);
836 		if (reuse)
837 			return tw;
838 	}
839 
840 	return NULL;
841 }
842