sys/netinet/tcp_timewait.c

   1 /*-
   2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
   3  *      The Regents of the University of California.  All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  * 3. Neither the name of the University nor the names of its contributors
  14  *    may be used to endorse or promote products derived from this software
  15  *    without specific prior written permission.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  *
  29  *      @(#)tcp_subr.c  8.2 (Berkeley) 5/24/95
  30  */
  31
  32 #include <sys/cdefs.h>
  33 __FBSDID("$FreeBSD$");
  34
  35 #include "opt_inet.h"
  36 #include "opt_inet6.h"
  37 #include "opt_tcpdebug.h"
  38
  39 #include <sys/param.h>
  40 #include <sys/systm.h>
  41 #include <sys/callout.h>
  42 #include <sys/kernel.h>
  43 #include <sys/sysctl.h>
  44 #include <sys/malloc.h>
  45 #include <sys/mbuf.h>
  46 #include <sys/priv.h>
  47 #include <sys/proc.h>
  48 #include <sys/socket.h>
  49 #include <sys/socketvar.h>
  50 #include <sys/protosw.h>
  51 #include <sys/random.h>
  52
  53 #include <vm/uma.h>
  54
  55 #include <net/route.h>
  56 #include <net/if.h>
  57 #include <net/if_var.h>
  58 #include <net/vnet.h>
  59
  60 #include <netinet/in.h>
  61 #include <netinet/in_pcb.h>
  62 #include <netinet/in_systm.h>
  63 #include <netinet/in_var.h>
  64 #include <netinet/ip.h>
  65 #include <netinet/ip_icmp.h>
  66 #include <netinet/ip_var.h>
  67 #ifdef INET6
  68 #include <netinet/ip6.h>
  69 #include <netinet6/in6_pcb.h>
  70 #include <netinet6/ip6_var.h>
  71 #include <netinet6/scope6_var.h>
  72 #include <netinet6/nd6.h>
  73 #endif
  74 #include <netinet/tcp.h>
  75 #include <netinet/tcp_fsm.h>
  76 #include <netinet/tcp_seq.h>
  77 #include <netinet/tcp_timer.h>
  78 #include <netinet/tcp_var.h>
  79 #ifdef INET6
  80 #include <netinet6/tcp6_var.h>
  81 #endif
  82 #include <netinet/tcpip.h>
  83 #ifdef TCPDEBUG
  84 #include <netinet/tcp_debug.h>
  85 #endif
  86 #ifdef INET6
  87 #include <netinet6/ip6protosw.h>
  88 #endif
  89
  90 #include <machine/in_cksum.h>
  91
  92 #include <security/mac/mac_framework.h>
  93
  94 static VNET_DEFINE(uma_zone_t, tcptw_zone);
  95 #define V_tcptw_zone            VNET(tcptw_zone)
  96 static int      maxtcptw;
  97
  98 /*
  99  * The timed wait queue contains references to each of the TCP sessions
 100  * currently in the TIME_WAIT state.  The queue pointers, including the
 101  * queue pointers in each tcptw structure, are protected using the global
 102  * timewait lock, which must be held over queue iteration and modification.
 103  *
 104  * Rules on tcptw usage:
 105  *  - a inpcb is always freed _after_ its tcptw
 106  *  - a tcptw relies on its inpcb reference counting for memory stability
 107  *  - a tcptw is dereferenceable only while its inpcb is locked
 108  */
 109 static VNET_DEFINE(TAILQ_HEAD(, tcptw), twq_2msl);
 110 #define V_twq_2msl              VNET(twq_2msl)
 111
 112 /* Global timewait lock */
 113 static VNET_DEFINE(struct rwlock, tw_lock);
 114 #define V_tw_lock               VNET(tw_lock)
 115
 116 #define TW_LOCK_INIT(tw, d)     rw_init_flags(&(tw), (d), 0)
 117 #define TW_LOCK_DESTROY(tw)     rw_destroy(&(tw))
 118 #define TW_RLOCK(tw)            rw_rlock(&(tw))
 119 #define TW_WLOCK(tw)            rw_wlock(&(tw))
 120 #define TW_RUNLOCK(tw)          rw_runlock(&(tw))
 121 #define TW_WUNLOCK(tw)          rw_wunlock(&(tw))
 122 #define TW_LOCK_ASSERT(tw)      rw_assert(&(tw), RA_LOCKED)
 123 #define TW_RLOCK_ASSERT(tw)     rw_assert(&(tw), RA_RLOCKED)
 124 #define TW_WLOCK_ASSERT(tw)     rw_assert(&(tw), RA_WLOCKED)
 125 #define TW_UNLOCK_ASSERT(tw)    rw_assert(&(tw), RA_UNLOCKED)
 126
 127 static void     tcp_tw_2msl_reset(struct tcptw *, int);
 128 static void     tcp_tw_2msl_stop(struct tcptw *, int);
 129 static int      tcp_twrespond(struct tcptw *, int);
 130
 131 static int
 132 tcptw_auto_size(void)
 133 {
 134         int halfrange;
 135
 136         /*
 137          * Max out at half the ephemeral port range so that TIME_WAIT
 138          * sockets don't tie up too many ephemeral ports.
 139          */
 140         if (V_ipport_lastauto > V_ipport_firstauto)
 141                 halfrange = (V_ipport_lastauto - V_ipport_firstauto) / 2;
 142         else
 143                 halfrange = (V_ipport_firstauto - V_ipport_lastauto) / 2;
 144         /* Protect against goofy port ranges smaller than 32. */
 145         return (imin(imax(halfrange, 32), maxsockets / 5));
 146 }
 147
 148 static int
 149 sysctl_maxtcptw(SYSCTL_HANDLER_ARGS)
 150 {
 151         int error, new;
 152
 153         if (maxtcptw == 0)
 154                 new = tcptw_auto_size();
 155         else
 156                 new = maxtcptw;
 157         error = sysctl_handle_int(oidp, &new, 0, req);
 158         if (error == 0 && req->newptr)
 159                 if (new >= 32) {
 160                         maxtcptw = new;
 161                         uma_zone_set_max(V_tcptw_zone, maxtcptw);
 162                 }
 163         return (error);
 164 }
 165
 166 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, maxtcptw, CTLTYPE_INT|CTLFLAG_RW,
 167     &maxtcptw, 0, sysctl_maxtcptw, "IU",
 168     "Maximum number of compressed TCP TIME_WAIT entries");
 169
 170 VNET_DEFINE(int, nolocaltimewait) = 0;
 171 #define V_nolocaltimewait       VNET(nolocaltimewait)
 172 SYSCTL_INT(_net_inet_tcp, OID_AUTO, nolocaltimewait, CTLFLAG_VNET | CTLFLAG_RW,
 173     &VNET_NAME(nolocaltimewait), 0,
 174     "Do not create compressed TCP TIME_WAIT entries for local connections");
 175
 176 void
 177 tcp_tw_zone_change(void)
 178 {
 179
 180         if (maxtcptw == 0)
 181                 uma_zone_set_max(V_tcptw_zone, tcptw_auto_size());
 182 }
 183
 184 void
 185 tcp_tw_init(void)
 186 {
 187
 188         V_tcptw_zone = uma_zcreate("tcptw", sizeof(struct tcptw),
 189             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
 190         TUNABLE_INT_FETCH("net.inet.tcp.maxtcptw", &maxtcptw);
 191         if (maxtcptw == 0)
 192                 uma_zone_set_max(V_tcptw_zone, tcptw_auto_size());
 193         else
 194                 uma_zone_set_max(V_tcptw_zone, maxtcptw);
 195         TAILQ_INIT(&V_twq_2msl);
 196         TW_LOCK_INIT(V_tw_lock, "tcptw");
 197 }
 198
 199 #ifdef VIMAGE
 200 void
 201 tcp_tw_destroy(void)
 202 {
 203         struct tcptw *tw;
 204
 205         INP_INFO_RLOCK(&V_tcbinfo);
 206         while ((tw = TAILQ_FIRST(&V_twq_2msl)) != NULL)
 207                 tcp_twclose(tw, 0);
 208         INP_INFO_RUNLOCK(&V_tcbinfo);
 209
 210         TW_LOCK_DESTROY(V_tw_lock);
 211         uma_zdestroy(V_tcptw_zone);
 212 }
 213 #endif
 214
 215 /*
 216  * Move a TCP connection into TIME_WAIT state.
 217  *    tcbinfo is locked.
 218  *    inp is locked, and is unlocked before returning.
 219  */
 220 void
 221 tcp_twstart(struct tcpcb *tp)
 222 {
 223         struct tcptw *tw;
 224         struct inpcb *inp = tp->t_inpcb;
 225         int acknow;
 226         struct socket *so;
 227 #ifdef INET6
 228         int isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6;
 229 #endif
 230
 231         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
 232         INP_WLOCK_ASSERT(inp);
 233
 234         /* A dropped inp should never transition to TIME_WAIT state. */
 235         KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("tcp_twstart: "
 236             "(inp->inp_flags & INP_DROPPED) != 0"));
 237
 238         if (V_nolocaltimewait) {
 239                 int error = 0;
 240 #ifdef INET6
 241                 if (isipv6)
 242                         error = in6_localaddr(&inp->in6p_faddr);
 243 #endif
 244 #if defined(INET6) && defined(INET)
 245                 else
 246 #endif
 247 #ifdef INET
 248                         error = in_localip(inp->inp_faddr);
 249 #endif
 250                 if (error) {
 251                         tp = tcp_close(tp);
 252                         if (tp != NULL)
 253                                 INP_WUNLOCK(inp);
 254                         return;
 255                 }
 256         }
 257
 258
 259         /*
 260          * For use only by DTrace.  We do not reference the state
 261          * after this point so modifying it in place is not a problem.
 262          */
 263         tcp_state_change(tp, TCPS_TIME_WAIT);
 264
 265         tw = uma_zalloc(V_tcptw_zone, M_NOWAIT);
 266         if (tw == NULL) {
 267                 /*
 268                  * Reached limit on total number of TIMEWAIT connections
 269                  * allowed. Remove a connection from TIMEWAIT queue in LRU
 270                  * fashion to make room for this connection.
 271                  *
 272                  * XXX:  Check if it possible to always have enough room
 273                  * in advance based on guarantees provided by uma_zalloc().
 274                  */
 275                 tw = tcp_tw_2msl_scan(1);
 276                 if (tw == NULL) {
 277                         tp = tcp_close(tp);
 278                         if (tp != NULL)
 279                                 INP_WUNLOCK(inp);
 280                         return;
 281                 }
 282         }
 283         /*
 284          * The tcptw will hold a reference on its inpcb until tcp_twclose
 285          * is called
 286          */
 287         tw->tw_inpcb = inp;
 288         in_pcbref(inp); /* Reference from tw */
 289
 290         /*
 291          * Recover last window size sent.
 292          */
 293         if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt))
 294                 tw->last_win = (tp->rcv_adv - tp->rcv_nxt) >> tp->rcv_scale;
 295         else
 296                 tw->last_win = 0;
 297
 298         /*
 299          * Set t_recent if timestamps are used on the connection.
 300          */
 301         if ((tp->t_flags & (TF_REQ_TSTMP|TF_RCVD_TSTMP|TF_NOOPT)) ==
 302             (TF_REQ_TSTMP|TF_RCVD_TSTMP)) {
 303                 tw->t_recent = tp->ts_recent;
 304                 tw->ts_offset = tp->ts_offset;
 305         } else {
 306                 tw->t_recent = 0;
 307                 tw->ts_offset = 0;
 308         }
 309
 310         tw->snd_nxt = tp->snd_nxt;
 311         tw->rcv_nxt = tp->rcv_nxt;
 312         tw->iss     = tp->iss;
 313         tw->irs     = tp->irs;
 314         tw->t_starttime = tp->t_starttime;
 315         tw->tw_time = 0;
 316
 317 /* XXX
 318  * If this code will
 319  * be used for fin-wait-2 state also, then we may need
 320  * a ts_recent from the last segment.
 321  */
 322         acknow = tp->t_flags & TF_ACKNOW;
 323
 324         /*
 325          * First, discard tcpcb state, which includes stopping its timers and
 326          * freeing it.  tcp_discardcb() used to also release the inpcb, but
 327          * that work is now done in the caller.
 328          *
 329          * Note: soisdisconnected() call used to be made in tcp_discardcb(),
 330          * and might not be needed here any longer.
 331          */
 332         tcp_discardcb(tp);
 333         so = inp->inp_socket;
 334         soisdisconnected(so);
 335         tw->tw_cred = crhold(so->so_cred);
 336         SOCK_LOCK(so);
 337         tw->tw_so_options = so->so_options;
 338         SOCK_UNLOCK(so);
 339         if (acknow)
 340                 tcp_twrespond(tw, TH_ACK);
 341         inp->inp_ppcb = tw;
 342         inp->inp_flags |= INP_TIMEWAIT;
 343         TCPSTATES_INC(TCPS_TIME_WAIT);
 344         tcp_tw_2msl_reset(tw, 0);
 345
 346         /*
 347          * If the inpcb owns the sole reference to the socket, then we can
 348          * detach and free the socket as it is not needed in time wait.
 349          */
 350         if (inp->inp_flags & INP_SOCKREF) {
 351                 KASSERT(so->so_state & SS_PROTOREF,
 352                     ("tcp_twstart: !SS_PROTOREF"));
 353                 inp->inp_flags &= ~INP_SOCKREF;
 354                 INP_WUNLOCK(inp);
 355                 ACCEPT_LOCK();
 356                 SOCK_LOCK(so);
 357                 so->so_state &= ~SS_PROTOREF;
 358                 sofree(so);
 359         } else
 360                 INP_WUNLOCK(inp);
 361 }
 362
 363 /*
 364  * Returns 1 if the TIME_WAIT state was killed and we should start over,
 365  * looking for a pcb in the listen state.  Returns 0 otherwise.
 366  */
 367 int
 368 tcp_twcheck(struct inpcb *inp, struct tcpopt *to __unused, struct tcphdr *th,
 369     struct mbuf *m, int tlen)
 370 {
 371         struct tcptw *tw;
 372         int thflags;
 373         tcp_seq seq;
 374
 375         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
 376         INP_WLOCK_ASSERT(inp);
 377
 378         /*
 379          * XXXRW: Time wait state for inpcb has been recycled, but inpcb is
 380          * still present.  This is undesirable, but temporarily necessary
 381          * until we work out how to handle inpcb's who's timewait state has
 382          * been removed.
 383          */
 384         tw = intotw(inp);
 385         if (tw == NULL)
 386                 goto drop;
 387
 388         thflags = th->th_flags;
 389
 390         /*
 391          * NOTE: for FIN_WAIT_2 (to be added later),
 392          * must validate sequence number before accepting RST
 393          */
 394
 395         /*
 396          * If the segment contains RST:
 397          *      Drop the segment - see Stevens, vol. 2, p. 964 and
 398          *      RFC 1337.
 399          */
 400         if (thflags & TH_RST)
 401                 goto drop;
 402
 403 #if 0
 404 /* PAWS not needed at the moment */
 405         /*
 406          * RFC 1323 PAWS: If we have a timestamp reply on this segment
 407          * and it's less than ts_recent, drop it.
 408          */
 409         if ((to.to_flags & TOF_TS) != 0 && tp->ts_recent &&
 410             TSTMP_LT(to.to_tsval, tp->ts_recent)) {
 411                 if ((thflags & TH_ACK) == 0)
 412                         goto drop;
 413                 goto ack;
 414         }
 415         /*
 416          * ts_recent is never updated because we never accept new segments.
 417          */
 418 #endif
 419
 420         /*
 421          * If a new connection request is received
 422          * while in TIME_WAIT, drop the old connection
 423          * and start over if the sequence numbers
 424          * are above the previous ones.
 425          */
 426         if ((thflags & TH_SYN) && SEQ_GT(th->th_seq, tw->rcv_nxt)) {
 427                 tcp_twclose(tw, 0);
 428                 return (1);
 429         }
 430
 431         /*
 432          * Drop the segment if it does not contain an ACK.
 433          */
 434         if ((thflags & TH_ACK) == 0)
 435                 goto drop;
 436
 437         /*
 438          * Reset the 2MSL timer if this is a duplicate FIN.
 439          */
 440         if (thflags & TH_FIN) {
 441                 seq = th->th_seq + tlen + (thflags & TH_SYN ? 1 : 0);
 442                 if (seq + 1 == tw->rcv_nxt)
 443                         tcp_tw_2msl_reset(tw, 1);
 444         }
 445
 446         /*
 447          * Acknowledge the segment if it has data or is not a duplicate ACK.
 448          */
 449         if (thflags != TH_ACK || tlen != 0 ||
 450             th->th_seq != tw->rcv_nxt || th->th_ack != tw->snd_nxt)
 451                 tcp_twrespond(tw, TH_ACK);
 452 drop:
 453         INP_WUNLOCK(inp);
 454         m_freem(m);
 455         return (0);
 456 }
 457
 458 void
 459 tcp_twclose(struct tcptw *tw, int reuse)
 460 {
 461         struct socket *so;
 462         struct inpcb *inp;
 463
 464         /*
 465          * At this point, we are in one of two situations:
 466          *
 467          * (1) We have no socket, just an inpcb<->twtcp pair.  We can free
 468          *     all state.
 469          *
 470          * (2) We have a socket -- if we own a reference, release it and
 471          *     notify the socket layer.
 472          */
 473         inp = tw->tw_inpcb;
 474         KASSERT((inp->inp_flags & INP_TIMEWAIT), ("tcp_twclose: !timewait"));
 475         KASSERT(intotw(inp) == tw, ("tcp_twclose: inp_ppcb != tw"));
 476         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);      /* in_pcbfree() */
 477         INP_WLOCK_ASSERT(inp);
 478
 479         tcp_tw_2msl_stop(tw, reuse);
 480         inp->inp_ppcb = NULL;
 481         in_pcbdrop(inp);
 482
 483         so = inp->inp_socket;
 484         if (so != NULL) {
 485                 /*
 486                  * If there's a socket, handle two cases: first, we own a
 487                  * strong reference, which we will now release, or we don't
 488                  * in which case another reference exists (XXXRW: think
 489                  * about this more), and we don't need to take action.
 490                  */
 491                 if (inp->inp_flags & INP_SOCKREF) {
 492                         inp->inp_flags &= ~INP_SOCKREF;
 493                         INP_WUNLOCK(inp);
 494                         ACCEPT_LOCK();
 495                         SOCK_LOCK(so);
 496                         KASSERT(so->so_state & SS_PROTOREF,
 497                             ("tcp_twclose: INP_SOCKREF && !SS_PROTOREF"));
 498                         so->so_state &= ~SS_PROTOREF;
 499                         sofree(so);
 500                 } else {
 501                         /*
 502                          * If we don't own the only reference, the socket and
 503                          * inpcb need to be left around to be handled by
 504                          * tcp_usr_detach() later.
 505                          */
 506                         INP_WUNLOCK(inp);
 507                 }
 508         } else {
 509                 /*
 510                  * The socket has been already cleaned-up for us, only free the
 511                  * inpcb.
 512                  */
 513                 in_pcbfree(inp);
 514         }
 515         TCPSTAT_INC(tcps_closed);
 516 }
 517
 518 static int
 519 tcp_twrespond(struct tcptw *tw, int flags)
 520 {
 521         struct inpcb *inp = tw->tw_inpcb;
 522 #if defined(INET6) || defined(INET)
 523         struct tcphdr *th = NULL;
 524 #endif
 525         struct mbuf *m;
 526 #ifdef INET
 527         struct ip *ip = NULL;
 528 #endif
 529         u_int hdrlen, optlen;
 530         int error = 0;                  /* Keep compiler happy */
 531         struct tcpopt to;
 532 #ifdef INET6
 533         struct ip6_hdr *ip6 = NULL;
 534         int isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6;
 535 #endif
 536         hdrlen = 0;                     /* Keep compiler happy */
 537
 538         INP_WLOCK_ASSERT(inp);
 539
 540         m = m_gethdr(M_NOWAIT, MT_DATA);
 541         if (m == NULL)
 542                 return (ENOBUFS);
 543         m->m_data += max_linkhdr;
 544
 545 #ifdef MAC
 546         mac_inpcb_create_mbuf(inp, m);
 547 #endif
 548
 549 #ifdef INET6
 550         if (isipv6) {
 551                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
 552                 ip6 = mtod(m, struct ip6_hdr *);
 553                 th = (struct tcphdr *)(ip6 + 1);
 554                 tcpip_fillheaders(inp, ip6, th);
 555         }
 556 #endif
 557 #if defined(INET6) && defined(INET)
 558         else
 559 #endif
 560 #ifdef INET
 561         {
 562                 hdrlen = sizeof(struct tcpiphdr);
 563                 ip = mtod(m, struct ip *);
 564                 th = (struct tcphdr *)(ip + 1);
 565                 tcpip_fillheaders(inp, ip, th);
 566         }
 567 #endif
 568         to.to_flags = 0;
 569
 570         /*
 571          * Send a timestamp and echo-reply if both our side and our peer
 572          * have sent timestamps in our SYN's and this is not a RST.
 573          */
 574         if (tw->t_recent && flags == TH_ACK) {
 575                 to.to_flags |= TOF_TS;
 576                 to.to_tsval = tcp_ts_getticks() + tw->ts_offset;
 577                 to.to_tsecr = tw->t_recent;
 578         }
 579         optlen = tcp_addoptions(&to, (u_char *)(th + 1));
 580
 581         m->m_len = hdrlen + optlen;
 582         m->m_pkthdr.len = m->m_len;
 583
 584         KASSERT(max_linkhdr + m->m_len <= MHLEN, ("tcptw: mbuf too small"));
 585
 586         th->th_seq = htonl(tw->snd_nxt);
 587         th->th_ack = htonl(tw->rcv_nxt);
 588         th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
 589         th->th_flags = flags;
 590         th->th_win = htons(tw->last_win);
 591
 592         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
 593 #ifdef INET6
 594         if (isipv6) {
 595                 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
 596                 th->th_sum = in6_cksum_pseudo(ip6,
 597                     sizeof(struct tcphdr) + optlen, IPPROTO_TCP, 0);
 598                 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
 599                 error = ip6_output(m, inp->in6p_outputopts, NULL,
 600                     (tw->tw_so_options & SO_DONTROUTE), NULL, NULL, inp);
 601         }
 602 #endif
 603 #if defined(INET6) && defined(INET)
 604         else
 605 #endif
 606 #ifdef INET
 607         {
 608                 m->m_pkthdr.csum_flags = CSUM_TCP;
 609                 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
 610                     htons(sizeof(struct tcphdr) + optlen + IPPROTO_TCP));
 611                 ip->ip_len = htons(m->m_pkthdr.len);
 612                 if (V_path_mtu_discovery)
 613                         ip->ip_off |= htons(IP_DF);
 614                 error = ip_output(m, inp->inp_options, NULL,
 615                     ((tw->tw_so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
 616                     NULL, inp);
 617         }
 618 #endif
 619         if (flags & TH_ACK)
 620                 TCPSTAT_INC(tcps_sndacks);
 621         else
 622                 TCPSTAT_INC(tcps_sndctrl);
 623         TCPSTAT_INC(tcps_sndtotal);
 624         return (error);
 625 }
 626
 627 static void
 628 tcp_tw_2msl_reset(struct tcptw *tw, int rearm)
 629 {
 630
 631         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
 632         INP_WLOCK_ASSERT(tw->tw_inpcb);
 633
 634         TW_WLOCK(V_tw_lock);
 635         if (rearm)
 636                 TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl);
 637         tw->tw_time = ticks + 2 * tcp_msl;
 638         TAILQ_INSERT_TAIL(&V_twq_2msl, tw, tw_2msl);
 639         TW_WUNLOCK(V_tw_lock);
 640 }
 641
 642 static void
 643 tcp_tw_2msl_stop(struct tcptw *tw, int reuse)
 644 {
 645         struct ucred *cred;
 646         struct inpcb *inp;
 647         int released;
 648
 649         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
 650
 651         TW_WLOCK(V_tw_lock);
 652         inp = tw->tw_inpcb;
 653         tw->tw_inpcb = NULL;
 654
 655         TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl);
 656         cred = tw->tw_cred;
 657         tw->tw_cred = NULL;
 658         TW_WUNLOCK(V_tw_lock);
 659
 660         if (cred != NULL)
 661                 crfree(cred);
 662
 663         released = in_pcbrele_wlocked(inp);
 664         KASSERT(!released, ("%s: inp should not be released here", __func__));
 665
 666         if (!reuse)
 667                 uma_zfree(V_tcptw_zone, tw);
 668         TCPSTATES_DEC(TCPS_TIME_WAIT);
 669 }
 670
 671 struct tcptw *
 672 tcp_tw_2msl_scan(int reuse)
 673 {
 674         struct tcptw *tw;
 675         struct inpcb *inp;
 676
 677 #ifdef INVARIANTS
 678         if (reuse) {
 679                 /*
 680                  * Exclusive pcbinfo lock is not required in reuse case even if
 681                  * two inpcb locks can be acquired simultaneously:
 682                  *  - the inpcb transitioning to TIME_WAIT state in
 683                  *    tcp_tw_start(),
 684                  *  - the inpcb closed by tcp_twclose().
 685                  *
 686                  * It is because only inpcbs in FIN_WAIT2 or CLOSING states can
 687                  * transition in TIME_WAIT state.  Then a pcbcb cannot be in
 688                  * TIME_WAIT list and transitioning to TIME_WAIT state at same
 689                  * time.
 690                  */
 691                 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
 692         }
 693 #endif
 694
 695         for (;;) {
 696                 TW_RLOCK(V_tw_lock);
 697                 tw = TAILQ_FIRST(&V_twq_2msl);
 698                 if (tw == NULL || (!reuse && (tw->tw_time - ticks) > 0)) {
 699                         TW_RUNLOCK(V_tw_lock);
 700                         break;
 701                 }
 702                 KASSERT(tw->tw_inpcb != NULL, ("%s: tw->tw_inpcb == NULL",
 703                     __func__));
 704
 705                 inp = tw->tw_inpcb;
 706                 in_pcbref(inp);
 707                 TW_RUNLOCK(V_tw_lock);
 708
 709                 if (INP_INFO_TRY_RLOCK(&V_tcbinfo)) {
 710
 711                         INP_WLOCK(inp);
 712                         tw = intotw(inp);
 713                         if (in_pcbrele_wlocked(inp)) {
 714                                 KASSERT(tw == NULL, ("%s: held last inp "
 715                                     "reference but tw not NULL", __func__));
 716                                 INP_INFO_RUNLOCK(&V_tcbinfo);
 717                                 continue;
 718                         }
 719
 720                         if (tw == NULL) {
 721                                 /* tcp_twclose() has already been called */
 722                                 INP_WUNLOCK(inp);
 723                                 INP_INFO_RUNLOCK(&V_tcbinfo);
 724                                 continue;
 725                         }
 726
 727                         tcp_twclose(tw, reuse);
 728                         INP_INFO_RUNLOCK(&V_tcbinfo);
 729                         if (reuse)
 730                             return tw;
 731                 } else {
 732                         /* INP_INFO lock is busy, continue later. */
 733                         INP_WLOCK(inp);
 734                         if (!in_pcbrele_wlocked(inp))
 735                                 INP_WUNLOCK(inp);
 736                         break;
 737                 }
 738         }
 739
 740         return NULL;
 741 }