]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/netinet/tcp_timewait.c
epoch(9): allow preemptible epochs to compose
[FreeBSD/FreeBSD.git] / sys / netinet / tcp_timewait.c
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         struct epoch_tracker et;
210
211         INP_INFO_RLOCK_ET(&V_tcbinfo, et);
212         while ((tw = TAILQ_FIRST(&V_twq_2msl)) != NULL)
213                 tcp_twclose(tw, 0);
214         INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
215
216         TW_LOCK_DESTROY(V_tw_lock);
217         uma_zdestroy(V_tcptw_zone);
218 }
219 #endif
220
221 /*
222  * Move a TCP connection into TIME_WAIT state.
223  *    tcbinfo is locked.
224  *    inp is locked, and is unlocked before returning.
225  */
226 void
227 tcp_twstart(struct tcpcb *tp)
228 {
229         struct tcptw twlocal, *tw;
230         struct inpcb *inp = tp->t_inpcb;
231         struct socket *so;
232         bool acknow, local;
233 #ifdef INET6
234         bool isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6;
235 #endif
236
237         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
238         INP_WLOCK_ASSERT(inp);
239
240         /* A dropped inp should never transition to TIME_WAIT state. */
241         KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("tcp_twstart: "
242             "(inp->inp_flags & INP_DROPPED) != 0"));
243
244         if (V_nolocaltimewait) {
245 #ifdef INET6
246                 if (isipv6)
247                         local = in6_localaddr(&inp->in6p_faddr);
248                 else
249 #endif
250 #ifdef INET
251                         local = in_localip(inp->inp_faddr);
252 #else
253                         local = false;
254 #endif
255         } else
256                 local = false;
257
258         /*
259          * For use only by DTrace.  We do not reference the state
260          * after this point so modifying it in place is not a problem.
261          */
262         tcp_state_change(tp, TCPS_TIME_WAIT);
263
264         if (local)
265                 tw = &twlocal;
266         else
267                 tw = uma_zalloc(V_tcptw_zone, M_NOWAIT);
268         if (tw == NULL) {
269                 /*
270                  * Reached limit on total number of TIMEWAIT connections
271                  * allowed. Remove a connection from TIMEWAIT queue in LRU
272                  * fashion to make room for this connection.
273                  *
274                  * XXX:  Check if it possible to always have enough room
275                  * in advance based on guarantees provided by uma_zalloc().
276                  */
277                 tw = tcp_tw_2msl_scan(1);
278                 if (tw == NULL) {
279                         tp = tcp_close(tp);
280                         if (tp != NULL)
281                                 INP_WUNLOCK(inp);
282                         return;
283                 }
284         }
285         /*
286          * For !local case the tcptw will hold a reference on its inpcb
287          * until tcp_twclose is called.
288          */
289         tw->tw_inpcb = inp;
290
291         /*
292          * Recover last window size sent.
293          */
294         if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt))
295                 tw->last_win = (tp->rcv_adv - tp->rcv_nxt) >> tp->rcv_scale;
296         else
297                 tw->last_win = 0;
298
299         /*
300          * Set t_recent if timestamps are used on the connection.
301          */
302         if ((tp->t_flags & (TF_REQ_TSTMP|TF_RCVD_TSTMP|TF_NOOPT)) ==
303             (TF_REQ_TSTMP|TF_RCVD_TSTMP)) {
304                 tw->t_recent = tp->ts_recent;
305                 tw->ts_offset = tp->ts_offset;
306         } else {
307                 tw->t_recent = 0;
308                 tw->ts_offset = 0;
309         }
310
311         tw->snd_nxt = tp->snd_nxt;
312         tw->rcv_nxt = tp->rcv_nxt;
313         tw->iss     = tp->iss;
314         tw->irs     = tp->irs;
315         tw->t_starttime = tp->t_starttime;
316         tw->tw_time = 0;
317
318 /* XXX
319  * If this code will
320  * be used for fin-wait-2 state also, then we may need
321  * a ts_recent from the last segment.
322  */
323         acknow = tp->t_flags & TF_ACKNOW;
324
325         /*
326          * First, discard tcpcb state, which includes stopping its timers and
327          * freeing it.  tcp_discardcb() used to also release the inpcb, but
328          * that work is now done in the caller.
329          *
330          * Note: soisdisconnected() call used to be made in tcp_discardcb(),
331          * and might not be needed here any longer.
332          */
333         tcp_discardcb(tp);
334         so = inp->inp_socket;
335         soisdisconnected(so);
336         tw->tw_so_options = so->so_options;
337         inp->inp_flags |= INP_TIMEWAIT;
338         if (acknow)
339                 tcp_twrespond(tw, TH_ACK);
340         if (local)
341                 in_pcbdrop(inp);
342         else {
343                 in_pcbref(inp); /* Reference from tw */
344                 tw->tw_cred = crhold(so->so_cred);
345                 inp->inp_ppcb = tw;
346                 TCPSTATES_INC(TCPS_TIME_WAIT);
347                 tcp_tw_2msl_reset(tw, 0);
348         }
349
350         /*
351          * If the inpcb owns the sole reference to the socket, then we can
352          * detach and free the socket as it is not needed in time wait.
353          */
354         if (inp->inp_flags & INP_SOCKREF) {
355                 KASSERT(so->so_state & SS_PROTOREF,
356                     ("tcp_twstart: !SS_PROTOREF"));
357                 inp->inp_flags &= ~INP_SOCKREF;
358                 INP_WUNLOCK(inp);
359                 SOCK_LOCK(so);
360                 so->so_state &= ~SS_PROTOREF;
361                 sofree(so);
362         } else
363                 INP_WUNLOCK(inp);
364 }
365
366 /*
367  * Returns 1 if the TIME_WAIT state was killed and we should start over,
368  * looking for a pcb in the listen state.  Returns 0 otherwise.
369  */
370 int
371 tcp_twcheck(struct inpcb *inp, struct tcpopt *to __unused, struct tcphdr *th,
372     struct mbuf *m, int tlen)
373 {
374         struct tcptw *tw;
375         int thflags;
376         tcp_seq seq;
377
378         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
379         INP_WLOCK_ASSERT(inp);
380
381         /*
382          * XXXRW: Time wait state for inpcb has been recycled, but inpcb is
383          * still present.  This is undesirable, but temporarily necessary
384          * until we work out how to handle inpcb's who's timewait state has
385          * been removed.
386          */
387         tw = intotw(inp);
388         if (tw == NULL)
389                 goto drop;
390
391         thflags = th->th_flags;
392
393         /*
394          * NOTE: for FIN_WAIT_2 (to be added later),
395          * must validate sequence number before accepting RST
396          */
397
398         /*
399          * If the segment contains RST:
400          *      Drop the segment - see Stevens, vol. 2, p. 964 and
401          *      RFC 1337.
402          */
403         if (thflags & TH_RST)
404                 goto drop;
405
406 #if 0
407 /* PAWS not needed at the moment */
408         /*
409          * RFC 1323 PAWS: If we have a timestamp reply on this segment
410          * and it's less than ts_recent, drop it.
411          */
412         if ((to.to_flags & TOF_TS) != 0 && tp->ts_recent &&
413             TSTMP_LT(to.to_tsval, tp->ts_recent)) {
414                 if ((thflags & TH_ACK) == 0)
415                         goto drop;
416                 goto ack;
417         }
418         /*
419          * ts_recent is never updated because we never accept new segments.
420          */
421 #endif
422
423         /*
424          * If a new connection request is received
425          * while in TIME_WAIT, drop the old connection
426          * and start over if the sequence numbers
427          * are above the previous ones.
428          */
429         if ((thflags & TH_SYN) && SEQ_GT(th->th_seq, tw->rcv_nxt)) {
430                 tcp_twclose(tw, 0);
431                 return (1);
432         }
433
434         /*
435          * Drop the segment if it does not contain an ACK.
436          */
437         if ((thflags & TH_ACK) == 0)
438                 goto drop;
439
440         /*
441          * Reset the 2MSL timer if this is a duplicate FIN.
442          */
443         if (thflags & TH_FIN) {
444                 seq = th->th_seq + tlen + (thflags & TH_SYN ? 1 : 0);
445                 if (seq + 1 == tw->rcv_nxt)
446                         tcp_tw_2msl_reset(tw, 1);
447         }
448
449         /*
450          * Acknowledge the segment if it has data or is not a duplicate ACK.
451          */
452         if (thflags != TH_ACK || tlen != 0 ||
453             th->th_seq != tw->rcv_nxt || th->th_ack != tw->snd_nxt)
454                 tcp_twrespond(tw, TH_ACK);
455 drop:
456         INP_WUNLOCK(inp);
457         m_freem(m);
458         return (0);
459 }
460
461 void
462 tcp_twclose(struct tcptw *tw, int reuse)
463 {
464         struct socket *so;
465         struct inpcb *inp;
466
467         /*
468          * At this point, we are in one of two situations:
469          *
470          * (1) We have no socket, just an inpcb<->twtcp pair.  We can free
471          *     all state.
472          *
473          * (2) We have a socket -- if we own a reference, release it and
474          *     notify the socket layer.
475          */
476         inp = tw->tw_inpcb;
477         KASSERT((inp->inp_flags & INP_TIMEWAIT), ("tcp_twclose: !timewait"));
478         KASSERT(intotw(inp) == tw, ("tcp_twclose: inp_ppcb != tw"));
479         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);      /* in_pcbfree() */
480         INP_WLOCK_ASSERT(inp);
481
482         tcp_tw_2msl_stop(tw, reuse);
483         inp->inp_ppcb = NULL;
484         in_pcbdrop(inp);
485
486         so = inp->inp_socket;
487         if (so != NULL) {
488                 /*
489                  * If there's a socket, handle two cases: first, we own a
490                  * strong reference, which we will now release, or we don't
491                  * in which case another reference exists (XXXRW: think
492                  * about this more), and we don't need to take action.
493                  */
494                 if (inp->inp_flags & INP_SOCKREF) {
495                         inp->inp_flags &= ~INP_SOCKREF;
496                         INP_WUNLOCK(inp);
497                         SOCK_LOCK(so);
498                         KASSERT(so->so_state & SS_PROTOREF,
499                             ("tcp_twclose: INP_SOCKREF && !SS_PROTOREF"));
500                         so->so_state &= ~SS_PROTOREF;
501                         sofree(so);
502                 } else {
503                         /*
504                          * If we don't own the only reference, the socket and
505                          * inpcb need to be left around to be handled by
506                          * tcp_usr_detach() later.
507                          */
508                         INP_WUNLOCK(inp);
509                 }
510         } else {
511                 /*
512                  * The socket has been already cleaned-up for us, only free the
513                  * inpcb.
514                  */
515                 in_pcbfree(inp);
516         }
517         TCPSTAT_INC(tcps_closed);
518 }
519
520 static int
521 tcp_twrespond(struct tcptw *tw, int flags)
522 {
523         struct inpcb *inp = tw->tw_inpcb;
524 #if defined(INET6) || defined(INET)
525         struct tcphdr *th = NULL;
526 #endif
527         struct mbuf *m;
528 #ifdef INET
529         struct ip *ip = NULL;
530 #endif
531         u_int hdrlen, optlen;
532         int error = 0;                  /* Keep compiler happy */
533         struct tcpopt to;
534 #ifdef INET6
535         struct ip6_hdr *ip6 = NULL;
536         int isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6;
537 #endif
538         hdrlen = 0;                     /* Keep compiler happy */
539
540         INP_WLOCK_ASSERT(inp);
541
542         m = m_gethdr(M_NOWAIT, MT_DATA);
543         if (m == NULL)
544                 return (ENOBUFS);
545         m->m_data += max_linkhdr;
546
547 #ifdef MAC
548         mac_inpcb_create_mbuf(inp, m);
549 #endif
550
551 #ifdef INET6
552         if (isipv6) {
553                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
554                 ip6 = mtod(m, struct ip6_hdr *);
555                 th = (struct tcphdr *)(ip6 + 1);
556                 tcpip_fillheaders(inp, ip6, th);
557         }
558 #endif
559 #if defined(INET6) && defined(INET)
560         else
561 #endif
562 #ifdef INET
563         {
564                 hdrlen = sizeof(struct tcpiphdr);
565                 ip = mtod(m, struct ip *);
566                 th = (struct tcphdr *)(ip + 1);
567                 tcpip_fillheaders(inp, ip, th);
568         }
569 #endif
570         to.to_flags = 0;
571
572         /*
573          * Send a timestamp and echo-reply if both our side and our peer
574          * have sent timestamps in our SYN's and this is not a RST.
575          */
576         if (tw->t_recent && flags == TH_ACK) {
577                 to.to_flags |= TOF_TS;
578                 to.to_tsval = tcp_ts_getticks() + tw->ts_offset;
579                 to.to_tsecr = tw->t_recent;
580         }
581         optlen = tcp_addoptions(&to, (u_char *)(th + 1));
582
583         m->m_len = hdrlen + optlen;
584         m->m_pkthdr.len = m->m_len;
585
586         KASSERT(max_linkhdr + m->m_len <= MHLEN, ("tcptw: mbuf too small"));
587
588         th->th_seq = htonl(tw->snd_nxt);
589         th->th_ack = htonl(tw->rcv_nxt);
590         th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
591         th->th_flags = flags;
592         th->th_win = htons(tw->last_win);
593
594         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
595 #ifdef INET6
596         if (isipv6) {
597                 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
598                 th->th_sum = in6_cksum_pseudo(ip6,
599                     sizeof(struct tcphdr) + optlen, IPPROTO_TCP, 0);
600                 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
601                 error = ip6_output(m, inp->in6p_outputopts, NULL,
602                     (tw->tw_so_options & SO_DONTROUTE), NULL, NULL, inp);
603         }
604 #endif
605 #if defined(INET6) && defined(INET)
606         else
607 #endif
608 #ifdef INET
609         {
610                 m->m_pkthdr.csum_flags = CSUM_TCP;
611                 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
612                     htons(sizeof(struct tcphdr) + optlen + IPPROTO_TCP));
613                 ip->ip_len = htons(m->m_pkthdr.len);
614                 if (V_path_mtu_discovery)
615                         ip->ip_off |= htons(IP_DF);
616                 error = ip_output(m, inp->inp_options, NULL,
617                     ((tw->tw_so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
618                     NULL, inp);
619         }
620 #endif
621         if (flags & TH_ACK)
622                 TCPSTAT_INC(tcps_sndacks);
623         else
624                 TCPSTAT_INC(tcps_sndctrl);
625         TCPSTAT_INC(tcps_sndtotal);
626         return (error);
627 }
628
629 static void
630 tcp_tw_2msl_reset(struct tcptw *tw, int rearm)
631 {
632
633         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
634         INP_WLOCK_ASSERT(tw->tw_inpcb);
635
636         TW_WLOCK(V_tw_lock);
637         if (rearm)
638                 TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl);
639         tw->tw_time = ticks + 2 * tcp_msl;
640         TAILQ_INSERT_TAIL(&V_twq_2msl, tw, tw_2msl);
641         TW_WUNLOCK(V_tw_lock);
642 }
643
644 static void
645 tcp_tw_2msl_stop(struct tcptw *tw, int reuse)
646 {
647         struct ucred *cred;
648         struct inpcb *inp;
649         int released __unused;
650
651         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
652
653         TW_WLOCK(V_tw_lock);
654         inp = tw->tw_inpcb;
655         tw->tw_inpcb = NULL;
656
657         TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl);
658         cred = tw->tw_cred;
659         tw->tw_cred = NULL;
660         TW_WUNLOCK(V_tw_lock);
661
662         if (cred != NULL)
663                 crfree(cred);
664
665         released = in_pcbrele_wlocked(inp);
666         KASSERT(!released, ("%s: inp should not be released here", __func__));
667
668         if (!reuse)
669                 uma_zfree(V_tcptw_zone, tw);
670         TCPSTATES_DEC(TCPS_TIME_WAIT);
671 }
672
673 struct tcptw *
674 tcp_tw_2msl_scan(int reuse)
675 {
676         struct tcptw *tw;
677         struct inpcb *inp;
678         struct epoch_tracker et;
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                 INP_INFO_RLOCK_ET(&V_tcbinfo, et);
713                 INP_WLOCK(inp);
714                 tw = intotw(inp);
715                 if (in_pcbrele_wlocked(inp)) {
716                         if (__predict_true(tw == NULL)) {
717                                 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
718                                 continue;
719                         } else {
720                                 /* This should not happen as in TIMEWAIT
721                                  * state the inp should not be destroyed
722                                  * before its tcptw. If INVARIANTS is
723                                  * defined panic.
724                                  */
725 #ifdef INVARIANTS
726                                 panic("%s: Panic before an infinite "
727                                           "loop: INP_TIMEWAIT && (INP_FREED "
728                                           "|| inp last reference) && tw != "
729                                           "NULL", __func__);
730 #else
731                                 log(LOG_ERR, "%s: Avoid an infinite "
732                                         "loop: INP_TIMEWAIT && (INP_FREED "
733                                         "|| inp last reference) && tw != "
734                                         "NULL", __func__);
735 #endif
736                                 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
737                                 break;
738                         }
739                 }
740
741                 if (tw == NULL) {
742                         /* tcp_twclose() has already been called */
743                         INP_WUNLOCK(inp);
744                         INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
745                         continue;
746                 }
747
748                 tcp_twclose(tw, reuse);
749                 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
750                 if (reuse)
751                         return tw;
752         }
753
754         return NULL;
755 }