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  * 4. 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 #ifndef INVARIANTS
  51 #include <sys/syslog.h>
  52 #endif
  53 #include <sys/protosw.h>
  54 #include <sys/random.h>
  55
  56 #include <vm/uma.h>
  57
  58 #include <net/route.h>
  59 #include <net/if.h>
  60 #include <net/if_var.h>
  61 #include <net/vnet.h>
  62
  63 #include <netinet/in.h>
  64 #include <netinet/in_kdtrace.h>
  65 #include <netinet/in_pcb.h>
  66 #include <netinet/in_systm.h>
  67 #include <netinet/in_var.h>
  68 #include <netinet/ip.h>
  69 #include <netinet/ip_icmp.h>
  70 #include <netinet/ip_var.h>
  71 #ifdef INET6
  72 #include <netinet/ip6.h>
  73 #include <netinet6/in6_pcb.h>
  74 #include <netinet6/ip6_var.h>
  75 #include <netinet6/scope6_var.h>
  76 #include <netinet6/nd6.h>
  77 #endif
  78 #include <netinet/tcp.h>
  79 #include <netinet/tcp_fsm.h>
  80 #include <netinet/tcp_seq.h>
  81 #include <netinet/tcp_timer.h>
  82 #include <netinet/tcp_var.h>
  83 #ifdef INET6
  84 #include <netinet6/tcp6_var.h>
  85 #endif
  86 #include <netinet/tcpip.h>
  87 #ifdef TCPDEBUG
  88 #include <netinet/tcp_debug.h>
  89 #endif
  90 #ifdef INET6
  91 #include <netinet6/ip6protosw.h>
  92 #endif
  93
  94 #include <machine/in_cksum.h>
  95
  96 #include <security/mac/mac_framework.h>
  97
  98 static VNET_DEFINE(uma_zone_t, tcptw_zone);
  99 #define V_tcptw_zone            VNET(tcptw_zone)
 100 static int      maxtcptw;
 101
 102 /*
 103  * The timed wait queue contains references to each of the TCP sessions
 104  * currently in the TIME_WAIT state.  The queue pointers, including the
 105  * queue pointers in each tcptw structure, are protected using the global
 106  * timewait lock, which must be held over queue iteration and modification.
 107  *
 108  * Rules on tcptw usage:
 109  *  - a inpcb is always freed _after_ its tcptw
 110  *  - a tcptw relies on its inpcb reference counting for memory stability
 111  *  - a tcptw is dereferenceable only while its inpcb is locked
 112  */
 113 static VNET_DEFINE(TAILQ_HEAD(, tcptw), twq_2msl);
 114 #define V_twq_2msl              VNET(twq_2msl)
 115
 116 /* Global timewait lock */
 117 static VNET_DEFINE(struct rwlock, tw_lock);
 118 #define V_tw_lock               VNET(tw_lock)
 119
 120 #define TW_LOCK_INIT(tw, d)     rw_init_flags(&(tw), (d), 0)
 121 #define TW_LOCK_DESTROY(tw)     rw_destroy(&(tw))
 122 #define TW_RLOCK(tw)            rw_rlock(&(tw))
 123 #define TW_WLOCK(tw)            rw_wlock(&(tw))
 124 #define TW_RUNLOCK(tw)          rw_runlock(&(tw))
 125 #define TW_WUNLOCK(tw)          rw_wunlock(&(tw))
 126 #define TW_LOCK_ASSERT(tw)      rw_assert(&(tw), RA_LOCKED)
 127 #define TW_RLOCK_ASSERT(tw)     rw_assert(&(tw), RA_RLOCKED)
 128 #define TW_WLOCK_ASSERT(tw)     rw_assert(&(tw), RA_WLOCKED)
 129 #define TW_UNLOCK_ASSERT(tw)    rw_assert(&(tw), RA_UNLOCKED)
 130
 131 static void     tcp_tw_2msl_reset(struct tcptw *, int);
 132 static void     tcp_tw_2msl_stop(struct tcptw *, int);
 133 static int      tcp_twrespond(struct tcptw *, int);
 134
 135 static int
 136 tcptw_auto_size(void)
 137 {
 138         int halfrange;
 139
 140         /*
 141          * Max out at half the ephemeral port range so that TIME_WAIT
 142          * sockets don't tie up too many ephemeral ports.
 143          */
 144         if (V_ipport_lastauto > V_ipport_firstauto)
 145                 halfrange = (V_ipport_lastauto - V_ipport_firstauto) / 2;
 146         else
 147                 halfrange = (V_ipport_firstauto - V_ipport_lastauto) / 2;
 148         /* Protect against goofy port ranges smaller than 32. */
 149         return (imin(imax(halfrange, 32), maxsockets / 5));
 150 }
 151
 152 static int
 153 sysctl_maxtcptw(SYSCTL_HANDLER_ARGS)
 154 {
 155         int error, new;
 156
 157         if (maxtcptw == 0)
 158                 new = tcptw_auto_size();
 159         else
 160                 new = maxtcptw;
 161         error = sysctl_handle_int(oidp, &new, 0, req);
 162         if (error == 0 && req->newptr)
 163                 if (new >= 32) {
 164                         maxtcptw = new;
 165                         uma_zone_set_max(V_tcptw_zone, maxtcptw);
 166                 }
 167         return (error);
 168 }
 169
 170 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, maxtcptw, CTLTYPE_INT|CTLFLAG_RW,
 171     &maxtcptw, 0, sysctl_maxtcptw, "IU",
 172     "Maximum number of compressed TCP TIME_WAIT entries");
 173
 174 VNET_DEFINE(int, nolocaltimewait) = 0;
 175 #define V_nolocaltimewait       VNET(nolocaltimewait)
 176 SYSCTL_INT(_net_inet_tcp, OID_AUTO, nolocaltimewait, CTLFLAG_VNET | CTLFLAG_RW,
 177     &VNET_NAME(nolocaltimewait), 0,
 178     "Do not create compressed TCP TIME_WAIT entries for local connections");
 179
 180 void
 181 tcp_tw_zone_change(void)
 182 {
 183
 184         if (maxtcptw == 0)
 185                 uma_zone_set_max(V_tcptw_zone, tcptw_auto_size());
 186 }
 187
 188 void
 189 tcp_tw_init(void)
 190 {
 191
 192         V_tcptw_zone = uma_zcreate("tcptw", sizeof(struct tcptw),
 193             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
 194         TUNABLE_INT_FETCH("net.inet.tcp.maxtcptw", &maxtcptw);
 195         if (maxtcptw == 0)
 196                 uma_zone_set_max(V_tcptw_zone, tcptw_auto_size());
 197         else
 198                 uma_zone_set_max(V_tcptw_zone, maxtcptw);
 199         TAILQ_INIT(&V_twq_2msl);
 200         TW_LOCK_INIT(V_tw_lock, "tcptw");
 201 }
 202
 203 #ifdef VIMAGE
 204 void
 205 tcp_tw_destroy(void)
 206 {
 207         struct tcptw *tw;
 208
 209         INP_INFO_RLOCK(&V_tcbinfo);
 210         while ((tw = TAILQ_FIRST(&V_twq_2msl)) != NULL)
 211                 tcp_twclose(tw, 0);
 212         INP_INFO_RUNLOCK(&V_tcbinfo);
 213
 214         TW_LOCK_DESTROY(V_tw_lock);
 215         uma_zdestroy(V_tcptw_zone);
 216 }
 217 #endif
 218
 219 /*
 220  * Move a TCP connection into TIME_WAIT state.
 221  *    tcbinfo is locked.
 222  *    inp is locked, and is unlocked before returning.
 223  */
 224 void
 225 tcp_twstart(struct tcpcb *tp)
 226 {
 227         struct tcptw *tw;
 228         struct inpcb *inp = tp->t_inpcb;
 229         int acknow;
 230         struct socket *so;
 231         uint32_t recwin;
 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         so = inp->inp_socket;
 299         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
 300             (long)TCP_MAXWIN << tp->rcv_scale);
 301         if (recwin < (so->so_rcv.sb_hiwat / 4) &&
 302             recwin < tp->t_maxseg)
 303                 recwin = 0;
 304         if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
 305             recwin < (tp->rcv_adv - tp->rcv_nxt))
 306                 recwin = (tp->rcv_adv - tp->rcv_nxt);
 307         tw->last_win = (u_short)(recwin >> tp->rcv_scale);
 308
 309         /*
 310          * Set t_recent if timestamps are used on the connection.
 311          */
 312         if ((tp->t_flags & (TF_REQ_TSTMP|TF_RCVD_TSTMP|TF_NOOPT)) ==
 313             (TF_REQ_TSTMP|TF_RCVD_TSTMP)) {
 314                 tw->t_recent = tp->ts_recent;
 315                 tw->ts_offset = tp->ts_offset;
 316         } else {
 317                 tw->t_recent = 0;
 318                 tw->ts_offset = 0;
 319         }
 320
 321         tw->snd_nxt = tp->snd_nxt;
 322         tw->rcv_nxt = tp->rcv_nxt;
 323         tw->iss     = tp->iss;
 324         tw->irs     = tp->irs;
 325         tw->t_starttime = tp->t_starttime;
 326         tw->tw_time = 0;
 327
 328 /* XXX
 329  * If this code will
 330  * be used for fin-wait-2 state also, then we may need
 331  * a ts_recent from the last segment.
 332  */
 333         acknow = tp->t_flags & TF_ACKNOW;
 334
 335         /*
 336          * First, discard tcpcb state, which includes stopping its timers and
 337          * freeing it.  tcp_discardcb() used to also release the inpcb, but
 338          * that work is now done in the caller.
 339          *
 340          * Note: soisdisconnected() call used to be made in tcp_discardcb(),
 341          * and might not be needed here any longer.
 342          */
 343         tcp_discardcb(tp);
 344         soisdisconnected(so);
 345         tw->tw_cred = crhold(so->so_cred);
 346         SOCK_LOCK(so);
 347         tw->tw_so_options = so->so_options;
 348         SOCK_UNLOCK(so);
 349         if (acknow)
 350                 tcp_twrespond(tw, TH_ACK);
 351         inp->inp_ppcb = tw;
 352         inp->inp_flags |= INP_TIMEWAIT;
 353         TCPSTATES_INC(TCPS_TIME_WAIT);
 354         tcp_tw_2msl_reset(tw, 0);
 355
 356         /*
 357          * If the inpcb owns the sole reference to the socket, then we can
 358          * detach and free the socket as it is not needed in time wait.
 359          */
 360         if (inp->inp_flags & INP_SOCKREF) {
 361                 KASSERT(so->so_state & SS_PROTOREF,
 362                     ("tcp_twstart: !SS_PROTOREF"));
 363                 inp->inp_flags &= ~INP_SOCKREF;
 364                 INP_WUNLOCK(inp);
 365                 ACCEPT_LOCK();
 366                 SOCK_LOCK(so);
 367                 so->so_state &= ~SS_PROTOREF;
 368                 sofree(so);
 369         } else
 370                 INP_WUNLOCK(inp);
 371 }
 372
 373 /*
 374  * Returns 1 if the TIME_WAIT state was killed and we should start over,
 375  * looking for a pcb in the listen state.  Returns 0 otherwise.
 376  */
 377 int
 378 tcp_twcheck(struct inpcb *inp, struct tcpopt *to __unused, struct tcphdr *th,
 379     struct mbuf *m, int tlen)
 380 {
 381         struct tcptw *tw;
 382         int thflags;
 383         tcp_seq seq;
 384
 385         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
 386         INP_WLOCK_ASSERT(inp);
 387
 388         /*
 389          * XXXRW: Time wait state for inpcb has been recycled, but inpcb is
 390          * still present.  This is undesirable, but temporarily necessary
 391          * until we work out how to handle inpcb's who's timewait state has
 392          * been removed.
 393          */
 394         tw = intotw(inp);
 395         if (tw == NULL)
 396                 goto drop;
 397
 398         thflags = th->th_flags;
 399
 400         /*
 401          * NOTE: for FIN_WAIT_2 (to be added later),
 402          * must validate sequence number before accepting RST
 403          */
 404
 405         /*
 406          * If the segment contains RST:
 407          *      Drop the segment - see Stevens, vol. 2, p. 964 and
 408          *      RFC 1337.
 409          */
 410         if (thflags & TH_RST)
 411                 goto drop;
 412
 413 #if 0
 414 /* PAWS not needed at the moment */
 415         /*
 416          * RFC 1323 PAWS: If we have a timestamp reply on this segment
 417          * and it's less than ts_recent, drop it.
 418          */
 419         if ((to.to_flags & TOF_TS) != 0 && tp->ts_recent &&
 420             TSTMP_LT(to.to_tsval, tp->ts_recent)) {
 421                 if ((thflags & TH_ACK) == 0)
 422                         goto drop;
 423                 goto ack;
 424         }
 425         /*
 426          * ts_recent is never updated because we never accept new segments.
 427          */
 428 #endif
 429
 430         /*
 431          * If a new connection request is received
 432          * while in TIME_WAIT, drop the old connection
 433          * and start over if the sequence numbers
 434          * are above the previous ones.
 435          */
 436         if ((thflags & TH_SYN) && SEQ_GT(th->th_seq, tw->rcv_nxt)) {
 437                 tcp_twclose(tw, 0);
 438                 return (1);
 439         }
 440
 441         /*
 442          * Drop the segment if it does not contain an ACK.
 443          */
 444         if ((thflags & TH_ACK) == 0)
 445                 goto drop;
 446
 447         /*
 448          * Reset the 2MSL timer if this is a duplicate FIN.
 449          */
 450         if (thflags & TH_FIN) {
 451                 seq = th->th_seq + tlen + (thflags & TH_SYN ? 1 : 0);
 452                 if (seq + 1 == tw->rcv_nxt)
 453                         tcp_tw_2msl_reset(tw, 1);
 454         }
 455
 456         /*
 457          * Acknowledge the segment if it has data or is not a duplicate ACK.
 458          */
 459         if (thflags != TH_ACK || tlen != 0 ||
 460             th->th_seq != tw->rcv_nxt || th->th_ack != tw->snd_nxt) {
 461                 TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
 462                 tcp_twrespond(tw, TH_ACK);
 463                 goto dropnoprobe;
 464         }
 465 drop:
 466         TCP_PROBE5(receive, NULL, NULL, m, NULL, th);
 467 dropnoprobe:
 468         INP_WUNLOCK(inp);
 469         m_freem(m);
 470         return (0);
 471 }
 472
 473 void
 474 tcp_twclose(struct tcptw *tw, int reuse)
 475 {
 476         struct socket *so;
 477         struct inpcb *inp;
 478
 479         /*
 480          * At this point, we are in one of two situations:
 481          *
 482          * (1) We have no socket, just an inpcb<->twtcp pair.  We can free
 483          *     all state.
 484          *
 485          * (2) We have a socket -- if we own a reference, release it and
 486          *     notify the socket layer.
 487          */
 488         inp = tw->tw_inpcb;
 489         KASSERT((inp->inp_flags & INP_TIMEWAIT), ("tcp_twclose: !timewait"));
 490         KASSERT(intotw(inp) == tw, ("tcp_twclose: inp_ppcb != tw"));
 491         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);      /* in_pcbfree() */
 492         INP_WLOCK_ASSERT(inp);
 493
 494         tcp_tw_2msl_stop(tw, reuse);
 495         inp->inp_ppcb = NULL;
 496         in_pcbdrop(inp);
 497
 498         so = inp->inp_socket;
 499         if (so != NULL) {
 500                 /*
 501                  * If there's a socket, handle two cases: first, we own a
 502                  * strong reference, which we will now release, or we don't
 503                  * in which case another reference exists (XXXRW: think
 504                  * about this more), and we don't need to take action.
 505                  */
 506                 if (inp->inp_flags & INP_SOCKREF) {
 507                         inp->inp_flags &= ~INP_SOCKREF;
 508                         INP_WUNLOCK(inp);
 509                         ACCEPT_LOCK();
 510                         SOCK_LOCK(so);
 511                         KASSERT(so->so_state & SS_PROTOREF,
 512                             ("tcp_twclose: INP_SOCKREF && !SS_PROTOREF"));
 513                         so->so_state &= ~SS_PROTOREF;
 514                         sofree(so);
 515                 } else {
 516                         /*
 517                          * If we don't own the only reference, the socket and
 518                          * inpcb need to be left around to be handled by
 519                          * tcp_usr_detach() later.
 520                          */
 521                         INP_WUNLOCK(inp);
 522                 }
 523         } else {
 524                 /*
 525                  * The socket has been already cleaned-up for us, only free the
 526                  * inpcb.
 527                  */
 528                 in_pcbfree(inp);
 529         }
 530         TCPSTAT_INC(tcps_closed);
 531 }
 532
 533 static int
 534 tcp_twrespond(struct tcptw *tw, int flags)
 535 {
 536         struct inpcb *inp = tw->tw_inpcb;
 537 #if defined(INET6) || defined(INET)
 538         struct tcphdr *th = NULL;
 539 #endif
 540         struct mbuf *m;
 541 #ifdef INET
 542         struct ip *ip = NULL;
 543 #endif
 544         u_int hdrlen, optlen;
 545         int error = 0;                  /* Keep compiler happy */
 546         struct tcpopt to;
 547 #ifdef INET6
 548         struct ip6_hdr *ip6 = NULL;
 549         int isipv6 = inp->inp_inc.inc_flags & INC_ISIPV6;
 550 #endif
 551         hdrlen = 0;                     /* Keep compiler happy */
 552
 553         INP_WLOCK_ASSERT(inp);
 554
 555         m = m_gethdr(M_NOWAIT, MT_DATA);
 556         if (m == NULL)
 557                 return (ENOBUFS);
 558         m->m_data += max_linkhdr;
 559
 560 #ifdef MAC
 561         mac_inpcb_create_mbuf(inp, m);
 562 #endif
 563
 564 #ifdef INET6
 565         if (isipv6) {
 566                 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
 567                 ip6 = mtod(m, struct ip6_hdr *);
 568                 th = (struct tcphdr *)(ip6 + 1);
 569                 tcpip_fillheaders(inp, ip6, th);
 570         }
 571 #endif
 572 #if defined(INET6) && defined(INET)
 573         else
 574 #endif
 575 #ifdef INET
 576         {
 577                 hdrlen = sizeof(struct tcpiphdr);
 578                 ip = mtod(m, struct ip *);
 579                 th = (struct tcphdr *)(ip + 1);
 580                 tcpip_fillheaders(inp, ip, th);
 581         }
 582 #endif
 583         to.to_flags = 0;
 584
 585         /*
 586          * Send a timestamp and echo-reply if both our side and our peer
 587          * have sent timestamps in our SYN's and this is not a RST.
 588          */
 589         if (tw->t_recent && flags == TH_ACK) {
 590                 to.to_flags |= TOF_TS;
 591                 to.to_tsval = tcp_ts_getticks() + tw->ts_offset;
 592                 to.to_tsecr = tw->t_recent;
 593         }
 594         optlen = tcp_addoptions(&to, (u_char *)(th + 1));
 595
 596         m->m_len = hdrlen + optlen;
 597         m->m_pkthdr.len = m->m_len;
 598
 599         KASSERT(max_linkhdr + m->m_len <= MHLEN, ("tcptw: mbuf too small"));
 600
 601         th->th_seq = htonl(tw->snd_nxt);
 602         th->th_ack = htonl(tw->rcv_nxt);
 603         th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
 604         th->th_flags = flags;
 605         th->th_win = htons(tw->last_win);
 606
 607         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
 608 #ifdef INET6
 609         if (isipv6) {
 610                 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
 611                 th->th_sum = in6_cksum_pseudo(ip6,
 612                     sizeof(struct tcphdr) + optlen, IPPROTO_TCP, 0);
 613                 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
 614                 TCP_PROBE5(send, NULL, NULL, ip6, NULL, th);
 615                 error = ip6_output(m, inp->in6p_outputopts, NULL,
 616                     (tw->tw_so_options & SO_DONTROUTE), NULL, NULL, inp);
 617         }
 618 #endif
 619 #if defined(INET6) && defined(INET)
 620         else
 621 #endif
 622 #ifdef INET
 623         {
 624                 m->m_pkthdr.csum_flags = CSUM_TCP;
 625                 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
 626                     htons(sizeof(struct tcphdr) + optlen + IPPROTO_TCP));
 627                 ip->ip_len = htons(m->m_pkthdr.len);
 628                 if (V_path_mtu_discovery)
 629                         ip->ip_off |= htons(IP_DF);
 630                 TCP_PROBE5(send, NULL, NULL, ip, NULL, th);
 631                 error = ip_output(m, inp->inp_options, NULL,
 632                     ((tw->tw_so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
 633                     NULL, inp);
 634         }
 635 #endif
 636         if (flags & TH_ACK)
 637                 TCPSTAT_INC(tcps_sndacks);
 638         else
 639                 TCPSTAT_INC(tcps_sndctrl);
 640         TCPSTAT_INC(tcps_sndtotal);
 641         return (error);
 642 }
 643
 644 static void
 645 tcp_tw_2msl_reset(struct tcptw *tw, int rearm)
 646 {
 647
 648         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
 649         INP_WLOCK_ASSERT(tw->tw_inpcb);
 650
 651         TW_WLOCK(V_tw_lock);
 652         if (rearm)
 653                 TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl);
 654         tw->tw_time = ticks + 2 * tcp_msl;
 655         TAILQ_INSERT_TAIL(&V_twq_2msl, tw, tw_2msl);
 656         TW_WUNLOCK(V_tw_lock);
 657 }
 658
 659 static void
 660 tcp_tw_2msl_stop(struct tcptw *tw, int reuse)
 661 {
 662         struct ucred *cred;
 663         struct inpcb *inp;
 664         int released;
 665
 666         INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
 667
 668         TW_WLOCK(V_tw_lock);
 669         inp = tw->tw_inpcb;
 670         tw->tw_inpcb = NULL;
 671
 672         TAILQ_REMOVE(&V_twq_2msl, tw, tw_2msl);
 673         cred = tw->tw_cred;
 674         tw->tw_cred = NULL;
 675         TW_WUNLOCK(V_tw_lock);
 676
 677         if (cred != NULL)
 678                 crfree(cred);
 679
 680         released = in_pcbrele_wlocked(inp);
 681         KASSERT(!released, ("%s: inp should not be released here", __func__));
 682
 683         if (!reuse)
 684                 uma_zfree(V_tcptw_zone, tw);
 685         TCPSTATES_DEC(TCPS_TIME_WAIT);
 686 }
 687
 688 struct tcptw *
 689 tcp_tw_2msl_scan(int reuse)
 690 {
 691         struct tcptw *tw;
 692         struct inpcb *inp;
 693
 694 #ifdef INVARIANTS
 695         if (reuse) {
 696                 /*
 697                  * Exclusive pcbinfo lock is not required in reuse case even if
 698                  * two inpcb locks can be acquired simultaneously:
 699                  *  - the inpcb transitioning to TIME_WAIT state in
 700                  *    tcp_tw_start(),
 701                  *  - the inpcb closed by tcp_twclose().
 702                  *
 703                  * It is because only inpcbs in FIN_WAIT2 or CLOSING states can
 704                  * transition in TIME_WAIT state.  Then a pcbcb cannot be in
 705                  * TIME_WAIT list and transitioning to TIME_WAIT state at same
 706                  * time.
 707                  */
 708                 INP_INFO_RLOCK_ASSERT(&V_tcbinfo);
 709         }
 710 #endif
 711
 712         for (;;) {
 713                 TW_RLOCK(V_tw_lock);
 714                 tw = TAILQ_FIRST(&V_twq_2msl);
 715                 if (tw == NULL || (!reuse && (tw->tw_time - ticks) > 0)) {
 716                         TW_RUNLOCK(V_tw_lock);
 717                         break;
 718                 }
 719                 KASSERT(tw->tw_inpcb != NULL, ("%s: tw->tw_inpcb == NULL",
 720                     __func__));
 721
 722                 inp = tw->tw_inpcb;
 723                 in_pcbref(inp);
 724                 TW_RUNLOCK(V_tw_lock);
 725
 726                 if (INP_INFO_TRY_RLOCK(&V_tcbinfo)) {
 727
 728                         INP_WLOCK(inp);
 729                         tw = intotw(inp);
 730                         if (in_pcbrele_wlocked(inp)) {
 731                                 if (__predict_true(tw == NULL)) {
 732                                         INP_INFO_RUNLOCK(&V_tcbinfo);
 733                                         continue;
 734                                 } else {
 735                                         /* This should not happen as in TIMEWAIT
 736                                          * state the inp should not be destroyed
 737                                          * before its tcptw. If INVARIANTS is
 738                                          * defined panic.
 739                                          */
 740 #ifdef INVARIANTS
 741                                         panic("%s: Panic before an infinite "
 742                                             "loop: INP_TIMEWAIT && (INP_FREED "
 743                                             "|| inp last reference) && tw != "
 744                                             "NULL", __func__);
 745 #else
 746                                         log(LOG_ERR, "%s: Avoid an infinite "
 747                                             "loop: INP_TIMEWAIT && (INP_FREED "
 748                                             "|| inp last reference) && tw != "
 749                                             "NULL", __func__);
 750 #endif
 751                                         INP_INFO_RUNLOCK(&V_tcbinfo);
 752                                         break;
 753                                 }
 754                         }
 755
 756                         if (tw == NULL) {
 757                                 /* tcp_twclose() has already been called */
 758                                 INP_WUNLOCK(inp);
 759                                 INP_INFO_RUNLOCK(&V_tcbinfo);
 760                                 continue;
 761                         }
 762
 763                         tcp_twclose(tw, reuse);
 764                         INP_INFO_RUNLOCK(&V_tcbinfo);
 765                         if (reuse)
 766                             return tw;
 767                 } else {
 768                         /* INP_INFO lock is busy, continue later. */
 769                         INP_WLOCK(inp);
 770                         if (!in_pcbrele_wlocked(inp))
 771                                 INP_WUNLOCK(inp);
 772                         break;
 773                 }
 774         }
 775
 776         return NULL;
 777 }