sys/netinet/tcp_timer.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_timer.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 #include "opt_rss.h"
  39
  40 #include <sys/param.h>
  41 #include <sys/kernel.h>
  42 #include <sys/lock.h>
  43 #include <sys/mbuf.h>
  44 #include <sys/mutex.h>
  45 #include <sys/protosw.h>
  46 #include <sys/smp.h>
  47 #include <sys/socket.h>
  48 #include <sys/socketvar.h>
  49 #include <sys/sysctl.h>
  50 #include <sys/systm.h>
  51
  52 #include <net/if.h>
  53 #include <net/route.h>
  54 #include <net/vnet.h>
  55 #include <net/netisr.h>
  56
  57 #include <netinet/cc.h>
  58 #include <netinet/in.h>
  59 #include <netinet/in_pcb.h>
  60 #include <netinet/in_rss.h>
  61 #include <netinet/in_systm.h>
  62 #ifdef INET6
  63 #include <netinet6/in6_pcb.h>
  64 #endif
  65 #include <netinet/ip_var.h>
  66 #include <netinet/tcp_fsm.h>
  67 #include <netinet/tcp_timer.h>
  68 #include <netinet/tcp_var.h>
  69 #include <netinet/tcpip.h>
  70 #ifdef TCPDEBUG
  71 #include <netinet/tcp_debug.h>
  72 #endif
  73
  74 int     tcp_keepinit;
  75 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINIT, keepinit, CTLTYPE_INT|CTLFLAG_RW,
  76     &tcp_keepinit, 0, sysctl_msec_to_ticks, "I", "time to establish connection");
  77
  78 int     tcp_keepidle;
  79 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPIDLE, keepidle, CTLTYPE_INT|CTLFLAG_RW,
  80     &tcp_keepidle, 0, sysctl_msec_to_ticks, "I", "time before keepalive probes begin");
  81
  82 int     tcp_keepintvl;
  83 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINTVL, keepintvl, CTLTYPE_INT|CTLFLAG_RW,
  84     &tcp_keepintvl, 0, sysctl_msec_to_ticks, "I", "time between keepalive probes");
  85
  86 int     tcp_delacktime;
  87 SYSCTL_PROC(_net_inet_tcp, TCPCTL_DELACKTIME, delacktime, CTLTYPE_INT|CTLFLAG_RW,
  88     &tcp_delacktime, 0, sysctl_msec_to_ticks, "I",
  89     "Time before a delayed ACK is sent");
  90
  91 int     tcp_msl;
  92 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, msl, CTLTYPE_INT|CTLFLAG_RW,
  93     &tcp_msl, 0, sysctl_msec_to_ticks, "I", "Maximum segment lifetime");
  94
  95 int     tcp_rexmit_min;
  96 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_min, CTLTYPE_INT|CTLFLAG_RW,
  97     &tcp_rexmit_min, 0, sysctl_msec_to_ticks, "I",
  98     "Minimum Retransmission Timeout");
  99
 100 int     tcp_rexmit_slop;
 101 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_slop, CTLTYPE_INT|CTLFLAG_RW,
 102     &tcp_rexmit_slop, 0, sysctl_msec_to_ticks, "I",
 103     "Retransmission Timer Slop");
 104
 105 static int      always_keepalive = 1;
 106 SYSCTL_INT(_net_inet_tcp, OID_AUTO, always_keepalive, CTLFLAG_RW,
 107     &always_keepalive , 0, "Assume SO_KEEPALIVE on all TCP connections");
 108
 109 int    tcp_fast_finwait2_recycle = 0;
 110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, fast_finwait2_recycle, CTLFLAG_RW,
 111     &tcp_fast_finwait2_recycle, 0,
 112     "Recycle closed FIN_WAIT_2 connections faster");
 113
 114 int    tcp_finwait2_timeout;
 115 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, finwait2_timeout, CTLTYPE_INT|CTLFLAG_RW,
 116     &tcp_finwait2_timeout, 0, sysctl_msec_to_ticks, "I", "FIN-WAIT2 timeout");
 117
 118 int     tcp_keepcnt = TCPTV_KEEPCNT;
 119 SYSCTL_INT(_net_inet_tcp, OID_AUTO, keepcnt, CTLFLAG_RW, &tcp_keepcnt, 0,
 120     "Number of keepalive probes to send");
 121
 122         /* max idle probes */
 123 int     tcp_maxpersistidle;
 124
 125 static int      tcp_rexmit_drop_options = 0;
 126 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rexmit_drop_options, CTLFLAG_RW,
 127     &tcp_rexmit_drop_options, 0,
 128     "Drop TCP options from 3rd and later retransmitted SYN");
 129
 130 #ifdef  RSS
 131 static int      per_cpu_timers = 1;
 132 #else
 133 static int      per_cpu_timers = 0;
 134 #endif
 135 SYSCTL_INT(_net_inet_tcp, OID_AUTO, per_cpu_timers, CTLFLAG_RW,
 136     &per_cpu_timers , 0, "run tcp timers on all cpus");
 137
 138 #if 0
 139 #define INP_CPU(inp)    (per_cpu_timers ? (!CPU_ABSENT(((inp)->inp_flowid % (mp_maxid+1))) ? \
 140                 ((inp)->inp_flowid % (mp_maxid+1)) : curcpu) : 0)
 141 #endif
 142
 143 /*
 144  * Map the given inp to a CPU id.
 145  *
 146  * This queries RSS if it's compiled in, else it defaults to the current
 147  * CPU ID.
 148  */
 149 static inline int
 150 inp_to_cpuid(struct inpcb *inp)
 151 {
 152         u_int cpuid;
 153
 154 #ifdef  RSS
 155         if (per_cpu_timers) {
 156                 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
 157                 if (cpuid == NETISR_CPUID_NONE)
 158                         return (curcpu);        /* XXX */
 159                 else
 160                         return (cpuid);
 161         }
 162 #else
 163         /* Legacy, pre-RSS behaviour */
 164         if (per_cpu_timers) {
 165                 /*
 166                  * We don't have a flowid -> cpuid mapping, so cheat and
 167                  * just map unknown cpuids to curcpu.  Not the best, but
 168                  * apparently better than defaulting to swi 0.
 169                  */
 170                 cpuid = inp->inp_flowid % (mp_maxid + 1);
 171                 if (! CPU_ABSENT(cpuid))
 172                         return (cpuid);
 173                 return (curcpu);
 174         }
 175 #endif
 176         /* Default for RSS and non-RSS - cpuid 0 */
 177         else {
 178                 return (0);
 179         }
 180 }
 181
 182 /*
 183  * Tcp protocol timeout routine called every 500 ms.
 184  * Updates timestamps used for TCP
 185  * causes finite state machine actions if timers expire.
 186  */
 187 void
 188 tcp_slowtimo(void)
 189 {
 190         VNET_ITERATOR_DECL(vnet_iter);
 191
 192         VNET_LIST_RLOCK_NOSLEEP();
 193         VNET_FOREACH(vnet_iter) {
 194                 CURVNET_SET(vnet_iter);
 195                 tcp_tw_2msl_scan();
 196                 CURVNET_RESTORE();
 197         }
 198         VNET_LIST_RUNLOCK_NOSLEEP();
 199 }
 200
 201 int     tcp_syn_backoff[TCP_MAXRXTSHIFT + 1] =
 202     { 1, 1, 1, 1, 1, 2, 4, 8, 16, 32, 64, 64, 64 };
 203
 204 int     tcp_backoff[TCP_MAXRXTSHIFT + 1] =
 205     { 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 512, 512, 512 };
 206
 207 static int tcp_totbackoff = 2559;       /* sum of tcp_backoff[] */
 208
 209 static int tcp_timer_race;
 210 SYSCTL_INT(_net_inet_tcp, OID_AUTO, timer_race, CTLFLAG_RD, &tcp_timer_race,
 211     0, "Count of t_inpcb races on tcp_discardcb");
 212
 213 /*
 214  * TCP timer processing.
 215  */
 216
 217 void
 218 tcp_timer_delack(void *xtp)
 219 {
 220         struct tcpcb *tp = xtp;
 221         struct inpcb *inp;
 222         CURVNET_SET(tp->t_vnet);
 223
 224         inp = tp->t_inpcb;
 225         /*
 226          * XXXRW: While this assert is in fact correct, bugs in the tcpcb
 227          * tear-down mean we need it as a work-around for races between
 228          * timers and tcp_discardcb().
 229          *
 230          * KASSERT(inp != NULL, ("tcp_timer_delack: inp == NULL"));
 231          */
 232         if (inp == NULL) {
 233                 tcp_timer_race++;
 234                 CURVNET_RESTORE();
 235                 return;
 236         }
 237         INP_WLOCK(inp);
 238         if (callout_pending(&tp->t_timers->tt_delack) ||
 239             !callout_active(&tp->t_timers->tt_delack)) {
 240                 INP_WUNLOCK(inp);
 241                 CURVNET_RESTORE();
 242                 return;
 243         }
 244         callout_deactivate(&tp->t_timers->tt_delack);
 245         if ((inp->inp_flags & INP_DROPPED) != 0) {
 246                 INP_WUNLOCK(inp);
 247                 CURVNET_RESTORE();
 248                 return;
 249         }
 250
 251         tp->t_flags |= TF_ACKNOW;
 252         TCPSTAT_INC(tcps_delack);
 253         (void) tcp_output(tp);
 254         INP_WUNLOCK(inp);
 255         CURVNET_RESTORE();
 256 }
 257
 258 void
 259 tcp_timer_2msl(void *xtp)
 260 {
 261         struct tcpcb *tp = xtp;
 262         struct inpcb *inp;
 263         CURVNET_SET(tp->t_vnet);
 264 #ifdef TCPDEBUG
 265         int ostate;
 266
 267         ostate = tp->t_state;
 268 #endif
 269         /*
 270          * XXXRW: Does this actually happen?
 271          */
 272         INP_INFO_WLOCK(&V_tcbinfo);
 273         inp = tp->t_inpcb;
 274         /*
 275          * XXXRW: While this assert is in fact correct, bugs in the tcpcb
 276          * tear-down mean we need it as a work-around for races between
 277          * timers and tcp_discardcb().
 278          *
 279          * KASSERT(inp != NULL, ("tcp_timer_2msl: inp == NULL"));
 280          */
 281         if (inp == NULL) {
 282                 tcp_timer_race++;
 283                 INP_INFO_WUNLOCK(&V_tcbinfo);
 284                 CURVNET_RESTORE();
 285                 return;
 286         }
 287         INP_WLOCK(inp);
 288         tcp_free_sackholes(tp);
 289         if (callout_pending(&tp->t_timers->tt_2msl) ||
 290             !callout_active(&tp->t_timers->tt_2msl)) {
 291                 INP_WUNLOCK(tp->t_inpcb);
 292                 INP_INFO_WUNLOCK(&V_tcbinfo);
 293                 CURVNET_RESTORE();
 294                 return;
 295         }
 296         callout_deactivate(&tp->t_timers->tt_2msl);
 297         if ((inp->inp_flags & INP_DROPPED) != 0) {
 298                 INP_WUNLOCK(inp);
 299                 INP_INFO_WUNLOCK(&V_tcbinfo);
 300                 CURVNET_RESTORE();
 301                 return;
 302         }
 303         /*
 304          * 2 MSL timeout in shutdown went off.  If we're closed but
 305          * still waiting for peer to close and connection has been idle
 306          * too long, or if 2MSL time is up from TIME_WAIT, delete connection
 307          * control block.  Otherwise, check again in a bit.
 308          *
 309          * If fastrecycle of FIN_WAIT_2, in FIN_WAIT_2 and receiver has closed,
 310          * there's no point in hanging onto FIN_WAIT_2 socket. Just close it.
 311          * Ignore fact that there were recent incoming segments.
 312          */
 313         if (tcp_fast_finwait2_recycle && tp->t_state == TCPS_FIN_WAIT_2 &&
 314             tp->t_inpcb && tp->t_inpcb->inp_socket &&
 315             (tp->t_inpcb->inp_socket->so_rcv.sb_state & SBS_CANTRCVMORE)) {
 316                 TCPSTAT_INC(tcps_finwait2_drops);
 317                 tp = tcp_close(tp);
 318         } else {
 319                 if (tp->t_state != TCPS_TIME_WAIT &&
 320                    ticks - tp->t_rcvtime <= TP_MAXIDLE(tp))
 321                        callout_reset_on(&tp->t_timers->tt_2msl,
 322                            TP_KEEPINTVL(tp), tcp_timer_2msl, tp,
 323                            inp_to_cpuid(inp));
 324                else
 325                        tp = tcp_close(tp);
 326        }
 327
 328 #ifdef TCPDEBUG
 329         if (tp != NULL && (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
 330                 tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0,
 331                           PRU_SLOWTIMO);
 332 #endif
 333         if (tp != NULL)
 334                 INP_WUNLOCK(inp);
 335         INP_INFO_WUNLOCK(&V_tcbinfo);
 336         CURVNET_RESTORE();
 337 }
 338
 339 void
 340 tcp_timer_keep(void *xtp)
 341 {
 342         struct tcpcb *tp = xtp;
 343         struct tcptemp *t_template;
 344         struct inpcb *inp;
 345         CURVNET_SET(tp->t_vnet);
 346 #ifdef TCPDEBUG
 347         int ostate;
 348
 349         ostate = tp->t_state;
 350 #endif
 351         INP_INFO_WLOCK(&V_tcbinfo);
 352         inp = tp->t_inpcb;
 353         /*
 354          * XXXRW: While this assert is in fact correct, bugs in the tcpcb
 355          * tear-down mean we need it as a work-around for races between
 356          * timers and tcp_discardcb().
 357          *
 358          * KASSERT(inp != NULL, ("tcp_timer_keep: inp == NULL"));
 359          */
 360         if (inp == NULL) {
 361                 tcp_timer_race++;
 362                 INP_INFO_WUNLOCK(&V_tcbinfo);
 363                 CURVNET_RESTORE();
 364                 return;
 365         }
 366         INP_WLOCK(inp);
 367         if (callout_pending(&tp->t_timers->tt_keep) ||
 368             !callout_active(&tp->t_timers->tt_keep)) {
 369                 INP_WUNLOCK(inp);
 370                 INP_INFO_WUNLOCK(&V_tcbinfo);
 371                 CURVNET_RESTORE();
 372                 return;
 373         }
 374         callout_deactivate(&tp->t_timers->tt_keep);
 375         if ((inp->inp_flags & INP_DROPPED) != 0) {
 376                 INP_WUNLOCK(inp);
 377                 INP_INFO_WUNLOCK(&V_tcbinfo);
 378                 CURVNET_RESTORE();
 379                 return;
 380         }
 381         /*
 382          * Keep-alive timer went off; send something
 383          * or drop connection if idle for too long.
 384          */
 385         TCPSTAT_INC(tcps_keeptimeo);
 386         if (tp->t_state < TCPS_ESTABLISHED)
 387                 goto dropit;
 388         if ((always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) &&
 389             tp->t_state <= TCPS_CLOSING) {
 390                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
 391                         goto dropit;
 392                 /*
 393                  * Send a packet designed to force a response
 394                  * if the peer is up and reachable:
 395                  * either an ACK if the connection is still alive,
 396                  * or an RST if the peer has closed the connection
 397                  * due to timeout or reboot.
 398                  * Using sequence number tp->snd_una-1
 399                  * causes the transmitted zero-length segment
 400                  * to lie outside the receive window;
 401                  * by the protocol spec, this requires the
 402                  * correspondent TCP to respond.
 403                  */
 404                 TCPSTAT_INC(tcps_keepprobe);
 405                 t_template = tcpip_maketemplate(inp);
 406                 if (t_template) {
 407                         tcp_respond(tp, t_template->tt_ipgen,
 408                                     &t_template->tt_t, (struct mbuf *)NULL,
 409                                     tp->rcv_nxt, tp->snd_una - 1, 0);
 410                         free(t_template, M_TEMP);
 411                 }
 412                 callout_reset_on(&tp->t_timers->tt_keep, TP_KEEPINTVL(tp),
 413                     tcp_timer_keep, tp, inp_to_cpuid(inp));
 414         } else
 415                 callout_reset_on(&tp->t_timers->tt_keep, TP_KEEPIDLE(tp),
 416                     tcp_timer_keep, tp, inp_to_cpuid(inp));
 417
 418 #ifdef TCPDEBUG
 419         if (inp->inp_socket->so_options & SO_DEBUG)
 420                 tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0,
 421                           PRU_SLOWTIMO);
 422 #endif
 423         INP_WUNLOCK(inp);
 424         INP_INFO_WUNLOCK(&V_tcbinfo);
 425         CURVNET_RESTORE();
 426         return;
 427
 428 dropit:
 429         TCPSTAT_INC(tcps_keepdrops);
 430         tp = tcp_drop(tp, ETIMEDOUT);
 431
 432 #ifdef TCPDEBUG
 433         if (tp != NULL && (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
 434                 tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0,
 435                           PRU_SLOWTIMO);
 436 #endif
 437         if (tp != NULL)
 438                 INP_WUNLOCK(tp->t_inpcb);
 439         INP_INFO_WUNLOCK(&V_tcbinfo);
 440         CURVNET_RESTORE();
 441 }
 442
 443 void
 444 tcp_timer_persist(void *xtp)
 445 {
 446         struct tcpcb *tp = xtp;
 447         struct inpcb *inp;
 448         CURVNET_SET(tp->t_vnet);
 449 #ifdef TCPDEBUG
 450         int ostate;
 451
 452         ostate = tp->t_state;
 453 #endif
 454         INP_INFO_WLOCK(&V_tcbinfo);
 455         inp = tp->t_inpcb;
 456         /*
 457          * XXXRW: While this assert is in fact correct, bugs in the tcpcb
 458          * tear-down mean we need it as a work-around for races between
 459          * timers and tcp_discardcb().
 460          *
 461          * KASSERT(inp != NULL, ("tcp_timer_persist: inp == NULL"));
 462          */
 463         if (inp == NULL) {
 464                 tcp_timer_race++;
 465                 INP_INFO_WUNLOCK(&V_tcbinfo);
 466                 CURVNET_RESTORE();
 467                 return;
 468         }
 469         INP_WLOCK(inp);
 470         if (callout_pending(&tp->t_timers->tt_persist) ||
 471             !callout_active(&tp->t_timers->tt_persist)) {
 472                 INP_WUNLOCK(inp);
 473                 INP_INFO_WUNLOCK(&V_tcbinfo);
 474                 CURVNET_RESTORE();
 475                 return;
 476         }
 477         callout_deactivate(&tp->t_timers->tt_persist);
 478         if ((inp->inp_flags & INP_DROPPED) != 0) {
 479                 INP_WUNLOCK(inp);
 480                 INP_INFO_WUNLOCK(&V_tcbinfo);
 481                 CURVNET_RESTORE();
 482                 return;
 483         }
 484         /*
 485          * Persistance timer into zero window.
 486          * Force a byte to be output, if possible.
 487          */
 488         TCPSTAT_INC(tcps_persisttimeo);
 489         /*
 490          * Hack: if the peer is dead/unreachable, we do not
 491          * time out if the window is closed.  After a full
 492          * backoff, drop the connection if the idle time
 493          * (no responses to probes) reaches the maximum
 494          * backoff that we would use if retransmitting.
 495          */
 496         if (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
 497             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
 498              ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) {
 499                 TCPSTAT_INC(tcps_persistdrop);
 500                 tp = tcp_drop(tp, ETIMEDOUT);
 501                 goto out;
 502         }
 503         /*
 504          * If the user has closed the socket then drop a persisting
 505          * connection after a much reduced timeout.
 506          */
 507         if (tp->t_state > TCPS_CLOSE_WAIT &&
 508             (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) {
 509                 TCPSTAT_INC(tcps_persistdrop);
 510                 tp = tcp_drop(tp, ETIMEDOUT);
 511                 goto out;
 512         }
 513         tcp_setpersist(tp);
 514         tp->t_flags |= TF_FORCEDATA;
 515         (void) tcp_output(tp);
 516         tp->t_flags &= ~TF_FORCEDATA;
 517
 518 out:
 519 #ifdef TCPDEBUG
 520         if (tp != NULL && tp->t_inpcb->inp_socket->so_options & SO_DEBUG)
 521                 tcp_trace(TA_USER, ostate, tp, NULL, NULL, PRU_SLOWTIMO);
 522 #endif
 523         if (tp != NULL)
 524                 INP_WUNLOCK(inp);
 525         INP_INFO_WUNLOCK(&V_tcbinfo);
 526         CURVNET_RESTORE();
 527 }
 528
 529 void
 530 tcp_timer_rexmt(void * xtp)
 531 {
 532         struct tcpcb *tp = xtp;
 533         CURVNET_SET(tp->t_vnet);
 534         int rexmt;
 535         int headlocked;
 536         struct inpcb *inp;
 537 #ifdef TCPDEBUG
 538         int ostate;
 539
 540         ostate = tp->t_state;
 541 #endif
 542         INP_INFO_RLOCK(&V_tcbinfo);
 543         inp = tp->t_inpcb;
 544         /*
 545          * XXXRW: While this assert is in fact correct, bugs in the tcpcb
 546          * tear-down mean we need it as a work-around for races between
 547          * timers and tcp_discardcb().
 548          *
 549          * KASSERT(inp != NULL, ("tcp_timer_rexmt: inp == NULL"));
 550          */
 551         if (inp == NULL) {
 552                 tcp_timer_race++;
 553                 INP_INFO_RUNLOCK(&V_tcbinfo);
 554                 CURVNET_RESTORE();
 555                 return;
 556         }
 557         INP_WLOCK(inp);
 558         if (callout_pending(&tp->t_timers->tt_rexmt) ||
 559             !callout_active(&tp->t_timers->tt_rexmt)) {
 560                 INP_WUNLOCK(inp);
 561                 INP_INFO_RUNLOCK(&V_tcbinfo);
 562                 CURVNET_RESTORE();
 563                 return;
 564         }
 565         callout_deactivate(&tp->t_timers->tt_rexmt);
 566         if ((inp->inp_flags & INP_DROPPED) != 0) {
 567                 INP_WUNLOCK(inp);
 568                 INP_INFO_RUNLOCK(&V_tcbinfo);
 569                 CURVNET_RESTORE();
 570                 return;
 571         }
 572         tcp_free_sackholes(tp);
 573         /*
 574          * Retransmission timer went off.  Message has not
 575          * been acked within retransmit interval.  Back off
 576          * to a longer retransmit interval and retransmit one segment.
 577          */
 578         if (++tp->t_rxtshift > TCP_MAXRXTSHIFT) {
 579                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
 580                 TCPSTAT_INC(tcps_timeoutdrop);
 581                 in_pcbref(inp);
 582                 INP_INFO_RUNLOCK(&V_tcbinfo);
 583                 INP_WUNLOCK(inp);
 584                 INP_INFO_WLOCK(&V_tcbinfo);
 585                 INP_WLOCK(inp);
 586                 if (in_pcbrele_wlocked(inp)) {
 587                         INP_INFO_WUNLOCK(&V_tcbinfo);
 588                         CURVNET_RESTORE();
 589                         return;
 590                 }
 591                 if (inp->inp_flags & INP_DROPPED) {
 592                         INP_WUNLOCK(inp);
 593                         INP_INFO_WUNLOCK(&V_tcbinfo);
 594                         CURVNET_RESTORE();
 595                         return;
 596                 }
 597
 598                 tp = tcp_drop(tp, tp->t_softerror ?
 599                               tp->t_softerror : ETIMEDOUT);
 600                 headlocked = 1;
 601                 goto out;
 602         }
 603         INP_INFO_RUNLOCK(&V_tcbinfo);
 604         headlocked = 0;
 605         if (tp->t_state == TCPS_SYN_SENT) {
 606                 /*
 607                  * If the SYN was retransmitted, indicate CWND to be
 608                  * limited to 1 segment in cc_conn_init().
 609                  */
 610                 tp->snd_cwnd = 1;
 611         } else if (tp->t_rxtshift == 1) {
 612                 /*
 613                  * first retransmit; record ssthresh and cwnd so they can
 614                  * be recovered if this turns out to be a "bad" retransmit.
 615                  * A retransmit is considered "bad" if an ACK for this
 616                  * segment is received within RTT/2 interval; the assumption
 617                  * here is that the ACK was already in flight.  See
 618                  * "On Estimating End-to-End Network Path Properties" by
 619                  * Allman and Paxson for more details.
 620                  */
 621                 tp->snd_cwnd_prev = tp->snd_cwnd;
 622                 tp->snd_ssthresh_prev = tp->snd_ssthresh;
 623                 tp->snd_recover_prev = tp->snd_recover;
 624                 if (IN_FASTRECOVERY(tp->t_flags))
 625                         tp->t_flags |= TF_WASFRECOVERY;
 626                 else
 627                         tp->t_flags &= ~TF_WASFRECOVERY;
 628                 if (IN_CONGRECOVERY(tp->t_flags))
 629                         tp->t_flags |= TF_WASCRECOVERY;
 630                 else
 631                         tp->t_flags &= ~TF_WASCRECOVERY;
 632                 tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
 633                 tp->t_flags |= TF_PREVVALID;
 634         } else
 635                 tp->t_flags &= ~TF_PREVVALID;
 636         TCPSTAT_INC(tcps_rexmttimeo);
 637         if (tp->t_state == TCPS_SYN_SENT)
 638                 rexmt = TCPTV_RTOBASE * tcp_syn_backoff[tp->t_rxtshift];
 639         else
 640                 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
 641         TCPT_RANGESET(tp->t_rxtcur, rexmt,
 642                       tp->t_rttmin, TCPTV_REXMTMAX);
 643         /*
 644          * Disable RFC1323 and SACK if we haven't got any response to
 645          * our third SYN to work-around some broken terminal servers
 646          * (most of which have hopefully been retired) that have bad VJ
 647          * header compression code which trashes TCP segments containing
 648          * unknown-to-them TCP options.
 649          */
 650         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
 651             (tp->t_rxtshift == 3))
 652                 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
 653         /*
 654          * If we backed off this far, our srtt estimate is probably bogus.
 655          * Clobber it so we'll take the next rtt measurement as our srtt;
 656          * move the current srtt into rttvar to keep the current
 657          * retransmit times until then.
 658          */
 659         if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) {
 660 #ifdef INET6
 661                 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0)
 662                         in6_losing(tp->t_inpcb);
 663 #endif
 664                 tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT);
 665                 tp->t_srtt = 0;
 666         }
 667         tp->snd_nxt = tp->snd_una;
 668         tp->snd_recover = tp->snd_max;
 669         /*
 670          * Force a segment to be sent.
 671          */
 672         tp->t_flags |= TF_ACKNOW;
 673         /*
 674          * If timing a segment in this window, stop the timer.
 675          */
 676         tp->t_rtttime = 0;
 677
 678         cc_cong_signal(tp, NULL, CC_RTO);
 679
 680         (void) tcp_output(tp);
 681
 682 out:
 683 #ifdef TCPDEBUG
 684         if (tp != NULL && (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
 685                 tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0,
 686                           PRU_SLOWTIMO);
 687 #endif
 688         if (tp != NULL)
 689                 INP_WUNLOCK(inp);
 690         if (headlocked)
 691                 INP_INFO_WUNLOCK(&V_tcbinfo);
 692         CURVNET_RESTORE();
 693 }
 694
 695 void
 696 tcp_timer_activate(struct tcpcb *tp, int timer_type, u_int delta)
 697 {
 698         struct callout *t_callout;
 699         void *f_callout;
 700         struct inpcb *inp = tp->t_inpcb;
 701         int cpu = inp_to_cpuid(inp);
 702
 703 #ifdef TCP_OFFLOAD
 704         if (tp->t_flags & TF_TOE)
 705                 return;
 706 #endif
 707
 708         switch (timer_type) {
 709                 case TT_DELACK:
 710                         t_callout = &tp->t_timers->tt_delack;
 711                         f_callout = tcp_timer_delack;
 712                         break;
 713                 case TT_REXMT:
 714                         t_callout = &tp->t_timers->tt_rexmt;
 715                         f_callout = tcp_timer_rexmt;
 716                         break;
 717                 case TT_PERSIST:
 718                         t_callout = &tp->t_timers->tt_persist;
 719                         f_callout = tcp_timer_persist;
 720                         break;
 721                 case TT_KEEP:
 722                         t_callout = &tp->t_timers->tt_keep;
 723                         f_callout = tcp_timer_keep;
 724                         break;
 725                 case TT_2MSL:
 726                         t_callout = &tp->t_timers->tt_2msl;
 727                         f_callout = tcp_timer_2msl;
 728                         break;
 729                 default:
 730                         panic("bad timer_type");
 731                 }
 732         if (delta == 0) {
 733                 callout_stop(t_callout);
 734         } else {
 735                 callout_reset_on(t_callout, delta, f_callout, tp, cpu);
 736         }
 737 }
 738
 739 int
 740 tcp_timer_active(struct tcpcb *tp, int timer_type)
 741 {
 742         struct callout *t_callout;
 743
 744         switch (timer_type) {
 745                 case TT_DELACK:
 746                         t_callout = &tp->t_timers->tt_delack;
 747                         break;
 748                 case TT_REXMT:
 749                         t_callout = &tp->t_timers->tt_rexmt;
 750                         break;
 751                 case TT_PERSIST:
 752                         t_callout = &tp->t_timers->tt_persist;
 753                         break;
 754                 case TT_KEEP:
 755                         t_callout = &tp->t_timers->tt_keep;
 756                         break;
 757                 case TT_2MSL:
 758                         t_callout = &tp->t_timers->tt_2msl;
 759                         break;
 760                 default:
 761                         panic("bad timer_type");
 762                 }
 763         return callout_active(t_callout);
 764 }
 765
 766 #define ticks_to_msecs(t)       (1000*(t) / hz)
 767
 768 void
 769 tcp_timer_to_xtimer(struct tcpcb *tp, struct tcp_timer *timer,
 770     struct xtcp_timer *xtimer)
 771 {
 772         sbintime_t now;
 773
 774         bzero(xtimer, sizeof(*xtimer));
 775         if (timer == NULL)
 776                 return;
 777         now = getsbinuptime();
 778         if (callout_active(&timer->tt_delack))
 779                 xtimer->tt_delack = (timer->tt_delack.c_time - now) / SBT_1MS;
 780         if (callout_active(&timer->tt_rexmt))
 781                 xtimer->tt_rexmt = (timer->tt_rexmt.c_time - now) / SBT_1MS;
 782         if (callout_active(&timer->tt_persist))
 783                 xtimer->tt_persist = (timer->tt_persist.c_time - now) / SBT_1MS;
 784         if (callout_active(&timer->tt_keep))
 785                 xtimer->tt_keep = (timer->tt_keep.c_time - now) / SBT_1MS;
 786         if (callout_active(&timer->tt_2msl))
 787                 xtimer->tt_2msl = (timer->tt_2msl.c_time - now) / SBT_1MS;
 788         xtimer->t_rcvtime = ticks_to_msecs(ticks - tp->t_rcvtime);
 789 }