sys/netinet/tcp_timer.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_timer.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_rss.h"
  40
  41 #include <sys/param.h>
  42 #include <sys/kernel.h>
  43 #include <sys/lock.h>
  44 #include <sys/mbuf.h>
  45 #include <sys/mutex.h>
  46 #include <sys/protosw.h>
  47 #include <sys/smp.h>
  48 #include <sys/socket.h>
  49 #include <sys/socketvar.h>
  50 #include <sys/sysctl.h>
  51 #include <sys/systm.h>
  52
  53 #include <net/if.h>
  54 #include <net/route.h>
  55 #include <net/rss_config.h>
  56 #include <net/vnet.h>
  57 #include <net/netisr.h>
  58
  59 #include <netinet/in.h>
  60 #include <netinet/in_kdtrace.h>
  61 #include <netinet/in_pcb.h>
  62 #include <netinet/in_rss.h>
  63 #include <netinet/in_systm.h>
  64 #ifdef INET6
  65 #include <netinet6/in6_pcb.h>
  66 #endif
  67 #include <netinet/ip_var.h>
  68 #include <netinet/tcp.h>
  69 #include <netinet/tcp_fsm.h>
  70 #include <netinet/tcp_timer.h>
  71 #include <netinet/tcp_var.h>
  72 #include <netinet/tcp_log_buf.h>
  73 #include <netinet/tcp_seq.h>
  74 #include <netinet/cc/cc.h>
  75 #ifdef INET6
  76 #include <netinet6/tcp6_var.h>
  77 #endif
  78 #include <netinet/tcpip.h>
  79
  80 int    tcp_persmin;
  81 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, persmin,
  82     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
  83     &tcp_persmin, 0, sysctl_msec_to_ticks, "I",
  84     "minimum persistence interval");
  85
  86 int    tcp_persmax;
  87 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, persmax,
  88     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
  89     &tcp_persmax, 0, sysctl_msec_to_ticks, "I",
  90     "maximum persistence interval");
  91
  92 int     tcp_keepinit;
  93 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINIT, keepinit,
  94     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
  95     &tcp_keepinit, 0, sysctl_msec_to_ticks, "I",
  96     "time to establish connection");
  97
  98 int     tcp_keepidle;
  99 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPIDLE, keepidle,
 100     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
 101     &tcp_keepidle, 0, sysctl_msec_to_ticks, "I",
 102     "time before keepalive probes begin");
 103
 104 int     tcp_keepintvl;
 105 SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINTVL, keepintvl,
 106     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
 107     &tcp_keepintvl, 0, sysctl_msec_to_ticks, "I",
 108     "time between keepalive probes");
 109
 110 int     tcp_delacktime;
 111 SYSCTL_PROC(_net_inet_tcp, TCPCTL_DELACKTIME, delacktime,
 112     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
 113     &tcp_delacktime, 0, sysctl_msec_to_ticks, "I",
 114     "Time before a delayed ACK is sent");
 115
 116 VNET_DEFINE(int, tcp_msl);
 117 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, msl,
 118     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_VNET,
 119     &VNET_NAME(tcp_msl), 0, sysctl_msec_to_ticks, "I",
 120     "Maximum segment lifetime");
 121
 122 int     tcp_rexmit_initial;
 123 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_initial,
 124    CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
 125     &tcp_rexmit_initial, 0, sysctl_msec_to_ticks, "I",
 126     "Initial Retransmission Timeout");
 127
 128 int     tcp_rexmit_min;
 129 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_min,
 130     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
 131     &tcp_rexmit_min, 0, sysctl_msec_to_ticks, "I",
 132     "Minimum Retransmission Timeout");
 133
 134 int     tcp_rexmit_slop;
 135 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_slop,
 136     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
 137     &tcp_rexmit_slop, 0, sysctl_msec_to_ticks, "I",
 138     "Retransmission Timer Slop");
 139
 140 VNET_DEFINE(int, tcp_always_keepalive) = 1;
 141 SYSCTL_INT(_net_inet_tcp, OID_AUTO, always_keepalive, CTLFLAG_VNET|CTLFLAG_RW,
 142     &VNET_NAME(tcp_always_keepalive) , 0,
 143     "Assume SO_KEEPALIVE on all TCP connections");
 144
 145 int    tcp_fast_finwait2_recycle = 0;
 146 SYSCTL_INT(_net_inet_tcp, OID_AUTO, fast_finwait2_recycle, CTLFLAG_RW,
 147     &tcp_fast_finwait2_recycle, 0,
 148     "Recycle closed FIN_WAIT_2 connections faster");
 149
 150 int    tcp_finwait2_timeout;
 151 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, finwait2_timeout,
 152     CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
 153     &tcp_finwait2_timeout, 0, sysctl_msec_to_ticks, "I",
 154     "FIN-WAIT2 timeout");
 155
 156 int     tcp_keepcnt = TCPTV_KEEPCNT;
 157 SYSCTL_INT(_net_inet_tcp, OID_AUTO, keepcnt, CTLFLAG_RW, &tcp_keepcnt, 0,
 158     "Number of keepalive probes to send");
 159
 160         /* max idle probes */
 161 int     tcp_maxpersistidle;
 162
 163 int     tcp_rexmit_drop_options = 0;
 164 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rexmit_drop_options, CTLFLAG_RW,
 165     &tcp_rexmit_drop_options, 0,
 166     "Drop TCP options from 3rd and later retransmitted SYN");
 167
 168 int     tcp_maxunacktime = TCPTV_MAXUNACKTIME;
 169 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, maxunacktime,
 170     CTLTYPE_INT|CTLFLAG_RW | CTLFLAG_NEEDGIANT,
 171     &tcp_maxunacktime, 0, sysctl_msec_to_ticks, "I",
 172     "Maximum time (in ms) that a session can linger without making progress");
 173
 174 VNET_DEFINE(int, tcp_pmtud_blackhole_detect);
 175 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_detection,
 176     CTLFLAG_RW|CTLFLAG_VNET,
 177     &VNET_NAME(tcp_pmtud_blackhole_detect), 0,
 178     "Path MTU Discovery Black Hole Detection Enabled");
 179
 180 #ifdef INET
 181 VNET_DEFINE(int, tcp_pmtud_blackhole_mss) = 1200;
 182 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_mss,
 183     CTLFLAG_RW|CTLFLAG_VNET,
 184     &VNET_NAME(tcp_pmtud_blackhole_mss), 0,
 185     "Path MTU Discovery Black Hole Detection lowered MSS");
 186 #endif
 187
 188 #ifdef INET6
 189 VNET_DEFINE(int, tcp_v6pmtud_blackhole_mss) = 1220;
 190 SYSCTL_INT(_net_inet_tcp, OID_AUTO, v6pmtud_blackhole_mss,
 191     CTLFLAG_RW|CTLFLAG_VNET,
 192     &VNET_NAME(tcp_v6pmtud_blackhole_mss), 0,
 193     "Path MTU Discovery IPv6 Black Hole Detection lowered MSS");
 194 #endif
 195
 196 #ifdef  RSS
 197 static int      per_cpu_timers = 1;
 198 #else
 199 static int      per_cpu_timers = 0;
 200 #endif
 201 SYSCTL_INT(_net_inet_tcp, OID_AUTO, per_cpu_timers, CTLFLAG_RW,
 202     &per_cpu_timers , 0, "run tcp timers on all cpus");
 203
 204 /*
 205  * Map the given inp to a CPU id.
 206  *
 207  * This queries RSS if it's compiled in, else it defaults to the current
 208  * CPU ID.
 209  */
 210 inline int
 211 inp_to_cpuid(struct inpcb *inp)
 212 {
 213         u_int cpuid;
 214
 215         if (per_cpu_timers) {
 216 #ifdef  RSS
 217                 cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype);
 218                 if (cpuid == NETISR_CPUID_NONE)
 219                         return (curcpu);        /* XXX */
 220                 else
 221                         return (cpuid);
 222 #endif
 223                 /*
 224                  * We don't have a flowid -> cpuid mapping, so cheat and
 225                  * just map unknown cpuids to curcpu.  Not the best, but
 226                  * apparently better than defaulting to swi 0.
 227                  */
 228                 cpuid = inp->inp_flowid % (mp_maxid + 1);
 229                 if (! CPU_ABSENT(cpuid))
 230                         return (cpuid);
 231                 return (curcpu);
 232         } else {
 233                 return (0);
 234         }
 235 }
 236
 237 int     tcp_backoff[TCP_MAXRXTSHIFT + 1] =
 238     { 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 512, 512, 512 };
 239
 240 int tcp_totbackoff = 2559;      /* sum of tcp_backoff[] */
 241
 242 /*
 243  * TCP timer processing.
 244  *
 245  * Each connection has 5 timers associated with it, which can be scheduled
 246  * simultaneously.  They all are serviced by one callout tcp_timer_enter().
 247  * This function executes the next timer via tcp_timersw[] vector.  Each
 248  * timer is supposed to return 'true' unless the connection was destroyed.
 249  * In the former case tcp_timer_enter() will schedule callout for next timer.
 250  */
 251
 252 typedef bool tcp_timer_t(struct tcpcb *);
 253 static tcp_timer_t tcp_timer_delack;
 254 static tcp_timer_t tcp_timer_2msl;
 255 static tcp_timer_t tcp_timer_keep;
 256 static tcp_timer_t tcp_timer_persist;
 257 static tcp_timer_t tcp_timer_rexmt;
 258
 259 static tcp_timer_t * const tcp_timersw[TT_N] = {
 260         [TT_DELACK] = tcp_timer_delack,
 261         [TT_REXMT] = tcp_timer_rexmt,
 262         [TT_PERSIST] = tcp_timer_persist,
 263         [TT_KEEP] = tcp_timer_keep,
 264         [TT_2MSL] = tcp_timer_2msl,
 265 };
 266
 267 /*
 268  * tcp_output_locked() s a timer specific variation of call to tcp_output(),
 269  * see tcp_var.h for the rest.  It handles drop request from advanced stacks,
 270  * but keeps tcpcb locked unless tcp_drop() destroyed it.
 271  * Returns true if tcpcb is valid and locked.
 272  */
 273 static inline bool
 274 tcp_output_locked(struct tcpcb *tp)
 275 {
 276         int rv;
 277
 278         INP_WLOCK_ASSERT(tptoinpcb(tp));
 279
 280         if ((rv = tp->t_fb->tfb_tcp_output(tp)) < 0) {
 281                 KASSERT(tp->t_fb->tfb_flags & TCP_FUNC_OUTPUT_CANDROP,
 282                     ("TCP stack %s requested tcp_drop(%p)",
 283                     tp->t_fb->tfb_tcp_block_name, tp));
 284                 tp = tcp_drop(tp, rv);
 285         }
 286
 287         return (tp != NULL);
 288 }
 289
 290 static bool
 291 tcp_timer_delack(struct tcpcb *tp)
 292 {
 293         struct epoch_tracker et;
 294 #if defined(INVARIANTS) || defined(VIMAGE)
 295         struct inpcb *inp = tptoinpcb(tp);
 296 #endif
 297         bool rv;
 298
 299         INP_WLOCK_ASSERT(inp);
 300
 301         CURVNET_SET(inp->inp_vnet);
 302         tp->t_flags |= TF_ACKNOW;
 303         TCPSTAT_INC(tcps_delack);
 304         NET_EPOCH_ENTER(et);
 305         rv = tcp_output_locked(tp);
 306         NET_EPOCH_EXIT(et);
 307         CURVNET_RESTORE();
 308
 309         return (rv);
 310 }
 311
 312 static bool
 313 tcp_timer_2msl(struct tcpcb *tp)
 314 {
 315         struct inpcb *inp = tptoinpcb(tp);
 316         bool close = false;
 317
 318         INP_WLOCK_ASSERT(inp);
 319
 320         TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
 321         CURVNET_SET(inp->inp_vnet);
 322         tcp_log_end_status(tp, TCP_EI_STATUS_2MSL);
 323         tcp_free_sackholes(tp);
 324         /*
 325          * 2 MSL timeout in shutdown went off.  If we're closed but
 326          * still waiting for peer to close and connection has been idle
 327          * too long delete connection control block.  Otherwise, check
 328          * again in a bit.
 329          *
 330          * If fastrecycle of FIN_WAIT_2, in FIN_WAIT_2 and receiver has closed,
 331          * there's no point in hanging onto FIN_WAIT_2 socket. Just close it.
 332          * Ignore fact that there were recent incoming segments.
 333          *
 334          * XXXGL: check if inp_socket shall always be !NULL here?
 335          */
 336         if (tp->t_state == TCPS_TIME_WAIT) {
 337                 close = true;
 338         } else if (tp->t_state == TCPS_FIN_WAIT_2 &&
 339             tcp_fast_finwait2_recycle && inp->inp_socket &&
 340             (inp->inp_socket->so_rcv.sb_state & SBS_CANTRCVMORE)) {
 341                 TCPSTAT_INC(tcps_finwait2_drops);
 342                 close = true;
 343         } else {
 344                 if (ticks - tp->t_rcvtime <= TP_MAXIDLE(tp))
 345                         tcp_timer_activate(tp, TT_2MSL, TP_KEEPINTVL(tp));
 346                 else
 347                         close = true;
 348         }
 349         if (close) {
 350                 struct epoch_tracker et;
 351
 352                 NET_EPOCH_ENTER(et);
 353                 tp = tcp_close(tp);
 354                 NET_EPOCH_EXIT(et);
 355         }
 356         CURVNET_RESTORE();
 357
 358         return (tp != NULL);
 359 }
 360
 361 static bool
 362 tcp_timer_keep(struct tcpcb *tp)
 363 {
 364         struct epoch_tracker et;
 365         struct inpcb *inp = tptoinpcb(tp);
 366         struct tcptemp *t_template;
 367
 368         INP_WLOCK_ASSERT(inp);
 369
 370         TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
 371         CURVNET_SET(inp->inp_vnet);
 372         /*
 373          * Because we don't regularly reset the keepalive callout in
 374          * the ESTABLISHED state, it may be that we don't actually need
 375          * to send a keepalive yet. If that occurs, schedule another
 376          * call for the next time the keepalive timer might expire.
 377          */
 378         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
 379                 u_int idletime;
 380
 381                 idletime = ticks - tp->t_rcvtime;
 382                 if (idletime < TP_KEEPIDLE(tp)) {
 383                         tcp_timer_activate(tp, TT_KEEP,
 384                             TP_KEEPIDLE(tp) - idletime);
 385                         CURVNET_RESTORE();
 386                         return (true);
 387                 }
 388         }
 389
 390         /*
 391          * Keep-alive timer went off; send something
 392          * or drop connection if idle for too long.
 393          */
 394         TCPSTAT_INC(tcps_keeptimeo);
 395         if (tp->t_state < TCPS_ESTABLISHED)
 396                 goto dropit;
 397         if ((V_tcp_always_keepalive ||
 398             inp->inp_socket->so_options & SO_KEEPALIVE) &&
 399             tp->t_state <= TCPS_CLOSING) {
 400                 if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp))
 401                         goto dropit;
 402                 /*
 403                  * Send a packet designed to force a response
 404                  * if the peer is up and reachable:
 405                  * either an ACK if the connection is still alive,
 406                  * or an RST if the peer has closed the connection
 407                  * due to timeout or reboot.
 408                  * Using sequence number tp->snd_una-1
 409                  * causes the transmitted zero-length segment
 410                  * to lie outside the receive window;
 411                  * by the protocol spec, this requires the
 412                  * correspondent TCP to respond.
 413                  */
 414                 TCPSTAT_INC(tcps_keepprobe);
 415                 t_template = tcpip_maketemplate(inp);
 416                 if (t_template) {
 417                         NET_EPOCH_ENTER(et);
 418                         tcp_respond(tp, t_template->tt_ipgen,
 419                                     &t_template->tt_t, (struct mbuf *)NULL,
 420                                     tp->rcv_nxt, tp->snd_una - 1, 0);
 421                         NET_EPOCH_EXIT(et);
 422                         free(t_template, M_TEMP);
 423                 }
 424                 tcp_timer_activate(tp, TT_KEEP, TP_KEEPINTVL(tp));
 425         } else
 426                 tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp));
 427
 428         CURVNET_RESTORE();
 429         return (true);
 430
 431 dropit:
 432         TCPSTAT_INC(tcps_keepdrops);
 433         NET_EPOCH_ENTER(et);
 434         tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX);
 435         tp = tcp_drop(tp, ETIMEDOUT);
 436         NET_EPOCH_EXIT(et);
 437         CURVNET_RESTORE();
 438
 439         return (tp != NULL);
 440 }
 441
 442 /*
 443  * Has this session exceeded the maximum time without seeing a substantive
 444  * acknowledgement? If so, return true; otherwise false.
 445  */
 446 static bool
 447 tcp_maxunacktime_check(struct tcpcb *tp)
 448 {
 449
 450         /* Are we tracking this timer for this session? */
 451         if (TP_MAXUNACKTIME(tp) == 0)
 452                 return false;
 453
 454         /* Do we have a current measurement. */
 455         if (tp->t_acktime == 0)
 456                 return false;
 457
 458         /* Are we within the acceptable range? */
 459         if (TSTMP_GT(TP_MAXUNACKTIME(tp) + tp->t_acktime, (u_int)ticks))
 460                 return false;
 461
 462         /* We exceeded the timer. */
 463         TCPSTAT_INC(tcps_progdrops);
 464         return true;
 465 }
 466
 467 static bool
 468 tcp_timer_persist(struct tcpcb *tp)
 469 {
 470         struct epoch_tracker et;
 471 #if defined(INVARIANTS) || defined(VIMAGE)
 472         struct inpcb *inp = tptoinpcb(tp);
 473 #endif
 474         bool progdrop, rv;
 475
 476         INP_WLOCK_ASSERT(inp);
 477
 478         TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
 479         CURVNET_SET(inp->inp_vnet);
 480         /*
 481          * Persistence timer into zero window.
 482          * Force a byte to be output, if possible.
 483          */
 484         TCPSTAT_INC(tcps_persisttimeo);
 485         /*
 486          * Hack: if the peer is dead/unreachable, we do not
 487          * time out if the window is closed.  After a full
 488          * backoff, drop the connection if the idle time
 489          * (no responses to probes) reaches the maximum
 490          * backoff that we would use if retransmitting.
 491          * Also, drop the connection if we haven't been making
 492          * progress.
 493          */
 494         progdrop = tcp_maxunacktime_check(tp);
 495         if (progdrop || (tp->t_rxtshift == TCP_MAXRXTSHIFT &&
 496             (ticks - tp->t_rcvtime >= tcp_maxpersistidle ||
 497              ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff))) {
 498                 if (!progdrop)
 499                         TCPSTAT_INC(tcps_persistdrop);
 500                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
 501                 goto dropit;
 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                 tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX);
 511                 goto dropit;
 512         }
 513         tcp_setpersist(tp);
 514         tp->t_flags |= TF_FORCEDATA;
 515         NET_EPOCH_ENTER(et);
 516         if ((rv = tcp_output_locked(tp)))
 517                 tp->t_flags &= ~TF_FORCEDATA;
 518         NET_EPOCH_EXIT(et);
 519         CURVNET_RESTORE();
 520
 521         return (rv);
 522
 523 dropit:
 524         NET_EPOCH_ENTER(et);
 525         tp = tcp_drop(tp, ETIMEDOUT);
 526         NET_EPOCH_EXIT(et);
 527         CURVNET_RESTORE();
 528
 529         return (tp != NULL);
 530 }
 531
 532 static bool
 533 tcp_timer_rexmt(struct tcpcb *tp)
 534 {
 535         struct epoch_tracker et;
 536         struct inpcb *inp = tptoinpcb(tp);
 537         int rexmt;
 538         bool isipv6, rv;
 539
 540         INP_WLOCK_ASSERT(inp);
 541
 542         TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO);
 543         CURVNET_SET(inp->inp_vnet);
 544         tcp_free_sackholes(tp);
 545         if (tp->t_fb->tfb_tcp_rexmit_tmr) {
 546                 /* The stack has a timer action too. */
 547                 (*tp->t_fb->tfb_tcp_rexmit_tmr)(tp);
 548         }
 549         /*
 550          * Retransmission timer went off.  Message has not
 551          * been acked within retransmit interval.  Back off
 552          * to a longer retransmit interval and retransmit one segment.
 553          *
 554          * If we've either exceeded the maximum number of retransmissions,
 555          * or we've gone long enough without making progress, then drop
 556          * the session.
 557          */
 558         if (++tp->t_rxtshift > TCP_MAXRXTSHIFT || tcp_maxunacktime_check(tp)) {
 559                 if (tp->t_rxtshift > TCP_MAXRXTSHIFT)
 560                         TCPSTAT_INC(tcps_timeoutdrop);
 561                 tp->t_rxtshift = TCP_MAXRXTSHIFT;
 562                 tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN);
 563                 NET_EPOCH_ENTER(et);
 564                 tp = tcp_drop(tp, ETIMEDOUT);
 565                 NET_EPOCH_EXIT(et);
 566                 CURVNET_RESTORE();
 567
 568                 return (tp != NULL);
 569         }
 570         if (tp->t_state == TCPS_SYN_SENT) {
 571                 /*
 572                  * If the SYN was retransmitted, indicate CWND to be
 573                  * limited to 1 segment in cc_conn_init().
 574                  */
 575                 tp->snd_cwnd = 1;
 576         } else if (tp->t_rxtshift == 1) {
 577                 /*
 578                  * first retransmit; record ssthresh and cwnd so they can
 579                  * be recovered if this turns out to be a "bad" retransmit.
 580                  * A retransmit is considered "bad" if an ACK for this
 581                  * segment is received within RTT/2 interval; the assumption
 582                  * here is that the ACK was already in flight.  See
 583                  * "On Estimating End-to-End Network Path Properties" by
 584                  * Allman and Paxson for more details.
 585                  */
 586                 tp->snd_cwnd_prev = tp->snd_cwnd;
 587                 tp->snd_ssthresh_prev = tp->snd_ssthresh;
 588                 tp->snd_recover_prev = tp->snd_recover;
 589                 if (IN_FASTRECOVERY(tp->t_flags))
 590                         tp->t_flags |= TF_WASFRECOVERY;
 591                 else
 592                         tp->t_flags &= ~TF_WASFRECOVERY;
 593                 if (IN_CONGRECOVERY(tp->t_flags))
 594                         tp->t_flags |= TF_WASCRECOVERY;
 595                 else
 596                         tp->t_flags &= ~TF_WASCRECOVERY;
 597                 if ((tp->t_flags & TF_RCVD_TSTMP) == 0)
 598                         tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1));
 599                 /* In the event that we've negotiated timestamps
 600                  * badrxtwin will be set to the value that we set
 601                  * the retransmitted packet's to_tsval to by tcp_output
 602                  */
 603                 tp->t_flags |= TF_PREVVALID;
 604         } else
 605                 tp->t_flags &= ~TF_PREVVALID;
 606         TCPSTAT_INC(tcps_rexmttimeo);
 607         if ((tp->t_state == TCPS_SYN_SENT) ||
 608             (tp->t_state == TCPS_SYN_RECEIVED))
 609                 rexmt = tcp_rexmit_initial * tcp_backoff[tp->t_rxtshift];
 610         else
 611                 rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift];
 612         TCPT_RANGESET(tp->t_rxtcur, rexmt,
 613                       tp->t_rttmin, TCPTV_REXMTMAX);
 614
 615         /*
 616          * We enter the path for PLMTUD if connection is established or, if
 617          * connection is FIN_WAIT_1 status, reason for the last is that if
 618          * amount of data we send is very small, we could send it in couple of
 619          * packets and process straight to FIN. In that case we won't catch
 620          * ESTABLISHED state.
 621          */
 622 #ifdef INET6
 623         isipv6 = (inp->inp_vflag & INP_IPV6) ? true : false;
 624 #else
 625         isipv6 = false;
 626 #endif
 627         if (((V_tcp_pmtud_blackhole_detect == 1) ||
 628             (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) ||
 629             (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) &&
 630             ((tp->t_state == TCPS_ESTABLISHED) ||
 631             (tp->t_state == TCPS_FIN_WAIT_1))) {
 632                 if (tp->t_rxtshift == 1) {
 633                         /*
 634                          * We enter blackhole detection after the first
 635                          * unsuccessful timer based retransmission.
 636                          * Then we reduce up to two times the MSS, each
 637                          * candidate giving two tries of retransmissions.
 638                          * But we give a candidate only two tries, if it
 639                          * actually reduces the MSS.
 640                          */
 641                         tp->t_blackhole_enter = 2;
 642                         tp->t_blackhole_exit = tp->t_blackhole_enter;
 643                         if (isipv6) {
 644 #ifdef INET6
 645                                 if (tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss)
 646                                         tp->t_blackhole_exit += 2;
 647                                 if (tp->t_maxseg > V_tcp_v6mssdflt &&
 648                                     V_tcp_v6pmtud_blackhole_mss > V_tcp_v6mssdflt)
 649                                         tp->t_blackhole_exit += 2;
 650 #endif
 651                         } else {
 652 #ifdef INET
 653                                 if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss)
 654                                         tp->t_blackhole_exit += 2;
 655                                 if (tp->t_maxseg > V_tcp_mssdflt &&
 656                                     V_tcp_pmtud_blackhole_mss > V_tcp_mssdflt)
 657                                         tp->t_blackhole_exit += 2;
 658 #endif
 659                         }
 660                 }
 661                 if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD|TF2_PLPMTU_MAXSEGSNT)) ==
 662                     (TF2_PLPMTU_PMTUD|TF2_PLPMTU_MAXSEGSNT)) &&
 663                     (tp->t_rxtshift >= tp->t_blackhole_enter &&
 664                     tp->t_rxtshift < tp->t_blackhole_exit &&
 665                     (tp->t_rxtshift - tp->t_blackhole_enter) % 2 == 0)) {
 666                         /*
 667                          * Enter Path MTU Black-hole Detection mechanism:
 668                          * - Disable Path MTU Discovery (IP "DF" bit).
 669                          * - Reduce MTU to lower value than what we
 670                          *   negotiated with peer.
 671                          */
 672                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) {
 673                                 /* Record that we may have found a black hole. */
 674                                 tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE;
 675                                 /* Keep track of previous MSS. */
 676                                 tp->t_pmtud_saved_maxseg = tp->t_maxseg;
 677                         }
 678
 679                         /*
 680                          * Reduce the MSS to blackhole value or to the default
 681                          * in an attempt to retransmit.
 682                          */
 683 #ifdef INET6
 684                         if (isipv6 &&
 685                             tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss &&
 686                             V_tcp_v6pmtud_blackhole_mss > V_tcp_v6mssdflt) {
 687                                 /* Use the sysctl tuneable blackhole MSS. */
 688                                 tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss;
 689                                 TCPSTAT_INC(tcps_pmtud_blackhole_activated);
 690                         } else if (isipv6) {
 691                                 /* Use the default MSS. */
 692                                 tp->t_maxseg = V_tcp_v6mssdflt;
 693                                 /*
 694                                  * Disable Path MTU Discovery when we switch to
 695                                  * minmss.
 696                                  */
 697                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
 698                                 TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
 699                         }
 700 #endif
 701 #if defined(INET6) && defined(INET)
 702                         else
 703 #endif
 704 #ifdef INET
 705                         if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss &&
 706                             V_tcp_pmtud_blackhole_mss > V_tcp_mssdflt) {
 707                                 /* Use the sysctl tuneable blackhole MSS. */
 708                                 tp->t_maxseg = V_tcp_pmtud_blackhole_mss;
 709                                 TCPSTAT_INC(tcps_pmtud_blackhole_activated);
 710                         } else {
 711                                 /* Use the default MSS. */
 712                                 tp->t_maxseg = V_tcp_mssdflt;
 713                                 /*
 714                                  * Disable Path MTU Discovery when we switch to
 715                                  * minmss.
 716                                  */
 717                                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
 718                                 TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss);
 719                         }
 720 #endif
 721                         /*
 722                          * Reset the slow-start flight size
 723                          * as it may depend on the new MSS.
 724                          */
 725                         if (CC_ALGO(tp)->conn_init != NULL)
 726                                 CC_ALGO(tp)->conn_init(&tp->t_ccv);
 727                 } else {
 728                         /*
 729                          * If further retransmissions are still unsuccessful
 730                          * with a lowered MTU, maybe this isn't a blackhole and
 731                          * we restore the previous MSS and blackhole detection
 732                          * flags.
 733                          */
 734                         if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) &&
 735                             (tp->t_rxtshift >= tp->t_blackhole_exit)) {
 736                                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
 737                                 tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE;
 738                                 tp->t_maxseg = tp->t_pmtud_saved_maxseg;
 739                                 TCPSTAT_INC(tcps_pmtud_blackhole_failed);
 740                                 /*
 741                                  * Reset the slow-start flight size as it
 742                                  * may depend on the new MSS.
 743                                  */
 744                                 if (CC_ALGO(tp)->conn_init != NULL)
 745                                         CC_ALGO(tp)->conn_init(&tp->t_ccv);
 746                         }
 747                 }
 748         }
 749
 750         /*
 751          * Disable RFC1323 and SACK if we haven't got any response to
 752          * our third SYN to work-around some broken terminal servers
 753          * (most of which have hopefully been retired) that have bad VJ
 754          * header compression code which trashes TCP segments containing
 755          * unknown-to-them TCP options.
 756          */
 757         if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) &&
 758             (tp->t_rxtshift == 3))
 759                 tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT);
 760         /*
 761          * If we backed off this far, notify the L3 protocol that we're having
 762          * connection problems.
 763          */
 764         if (tp->t_rxtshift > TCP_RTT_INVALIDATE) {
 765 #ifdef INET6
 766                 if ((inp->inp_vflag & INP_IPV6) != 0)
 767                         in6_losing(inp);
 768                 else
 769 #endif
 770                         in_losing(inp);
 771         }
 772         tp->snd_nxt = tp->snd_una;
 773         tp->snd_recover = tp->snd_max;
 774         /*
 775          * Force a segment to be sent.
 776          */
 777         tp->t_flags |= TF_ACKNOW;
 778         /*
 779          * If timing a segment in this window, stop the timer.
 780          */
 781         tp->t_rtttime = 0;
 782
 783         cc_cong_signal(tp, NULL, CC_RTO);
 784         NET_EPOCH_ENTER(et);
 785         rv = tcp_output_locked(tp);
 786         NET_EPOCH_EXIT(et);
 787         CURVNET_RESTORE();
 788
 789         return (rv);
 790 }
 791
 792 static void
 793 tcp_bblog_timer(struct tcpcb *tp, tt_which which, tt_what what, uint32_t ticks)
 794 {
 795         struct tcp_log_buffer *lgb;
 796         uint64_t ms;
 797
 798         INP_WLOCK_ASSERT(tptoinpcb(tp));
 799         if (tcp_bblogging_on(tp))
 800                 lgb = tcp_log_event(tp, NULL, NULL, NULL, TCP_LOG_RTO, 0, 0,
 801                     NULL, false, NULL, NULL, 0, NULL);
 802         else
 803                 lgb = NULL;
 804         if (lgb != NULL) {
 805                 lgb->tlb_flex1 = (what << 8) | which;
 806                 if (what == TT_STARTING) {
 807                         /* Convert ticks to ms and store it in tlb_flex2. */
 808                         if (hz == 1000)
 809                                 lgb->tlb_flex2 = ticks;
 810                         else {
 811                                 ms = (((uint64_t)ticks * 1000) + (hz - 1)) / hz;
 812                                 if (ms > UINT32_MAX)
 813                                         lgb->tlb_flex2 = UINT32_MAX;
 814                                 else
 815                                         lgb->tlb_flex2 = (uint32_t)ms;
 816                         }
 817                 }
 818         }
 819 }
 820
 821 static inline tt_which
 822 tcp_timer_next(struct tcpcb *tp, sbintime_t *precision)
 823 {
 824         tt_which i, rv;
 825         sbintime_t after, before;
 826
 827         for (i = 0, rv = TT_N, after = before = SBT_MAX; i < TT_N; i++) {
 828                 if (tp->t_timers[i] < after) {
 829                         after = tp->t_timers[i];
 830                         rv = i;
 831                 }
 832                 before = MIN(before, tp->t_timers[i] + tp->t_precisions[i]);
 833         }
 834         if (precision != NULL)
 835                 *precision = before - after;
 836
 837         return (rv);
 838 }
 839
 840 static void
 841 tcp_timer_enter(void *xtp)
 842 {
 843         struct tcpcb *tp = xtp;
 844         struct inpcb *inp = tptoinpcb(tp);
 845         sbintime_t precision;
 846         tt_which which;
 847         bool tp_valid;
 848
 849         INP_WLOCK_ASSERT(inp);
 850         MPASS((curthread->td_pflags & TDP_INTCPCALLOUT) == 0);
 851
 852         curthread->td_pflags |= TDP_INTCPCALLOUT;
 853
 854         which = tcp_timer_next(tp, NULL);
 855         MPASS(which < TT_N);
 856         tp->t_timers[which] = SBT_MAX;
 857         tp->t_precisions[which] = 0;
 858
 859         tcp_bblog_timer(tp, which, TT_PROCESSING, 0);
 860         tp_valid = tcp_timersw[which](tp);
 861         if (tp_valid) {
 862                 tcp_bblog_timer(tp, which, TT_PROCESSED, 0);
 863                 if ((which = tcp_timer_next(tp, &precision)) != TT_N) {
 864                         callout_reset_sbt_on(&tp->t_callout,
 865                             tp->t_timers[which], precision, tcp_timer_enter,
 866                             tp, inp_to_cpuid(inp), C_ABSOLUTE);
 867                 }
 868                 INP_WUNLOCK(inp);
 869         }
 870
 871         curthread->td_pflags &= ~TDP_INTCPCALLOUT;
 872 }
 873
 874 /*
 875  * Activate or stop (delta == 0) a TCP timer.
 876  */
 877 void
 878 tcp_timer_activate(struct tcpcb *tp, tt_which which, u_int delta)
 879 {
 880         struct inpcb *inp = tptoinpcb(tp);
 881         sbintime_t precision;
 882         tt_what what;
 883
 884 #ifdef TCP_OFFLOAD
 885         if (tp->t_flags & TF_TOE)
 886                 return;
 887 #endif
 888
 889         INP_WLOCK_ASSERT(inp);
 890
 891         if (delta > 0) {
 892                 what = TT_STARTING;
 893                 callout_when(tick_sbt * delta, 0, C_HARDCLOCK,
 894                     &tp->t_timers[which], &tp->t_precisions[which]);
 895         } else {
 896                 what = TT_STOPPING;
 897                 tp->t_timers[which] = SBT_MAX;
 898         }
 899         tcp_bblog_timer(tp, which, what, delta);
 900
 901         if ((which = tcp_timer_next(tp, &precision)) != TT_N)
 902                 callout_reset_sbt_on(&tp->t_callout, tp->t_timers[which],
 903                     precision, tcp_timer_enter, tp, inp_to_cpuid(inp),
 904                     C_ABSOLUTE);
 905         else
 906                 callout_stop(&tp->t_callout);
 907 }
 908
 909 bool
 910 tcp_timer_active(struct tcpcb *tp, tt_which which)
 911 {
 912
 913         INP_WLOCK_ASSERT(tptoinpcb(tp));
 914
 915         return (tp->t_timers[which] != SBT_MAX);
 916 }
 917
 918 /*
 919  * Stop all timers associated with tcpcb.
 920  *
 921  * Called only on tcpcb destruction.  The tcpcb shall already be dropped from
 922  * the pcb lookup database and socket is not losing the last reference.
 923  *
 924  * XXXGL: unfortunately our callout(9) is not able to fully stop a locked
 925  * callout even when only two threads are involved: the callout itself and the
 926  * thread that does callout_stop().  See where softclock_call_cc() swaps the
 927  * callwheel lock to callout lock and then checks cc_exec_cancel().  This is
 928  * the race window.  If it happens, the tcp_timer_enter() won't be executed,
 929  * however pcb lock will be locked and released, hence we can't free memory.
 930  * Until callout(9) is improved, just keep retrying.  In my profiling I've seen
 931  * such event happening less than 1 time per hour with 20-30 Gbit/s of traffic.
 932  */
 933 void
 934 tcp_timer_stop(struct tcpcb *tp)
 935 {
 936         struct inpcb *inp = tptoinpcb(tp);
 937
 938         INP_WLOCK_ASSERT(inp);
 939
 940         if (curthread->td_pflags & TDP_INTCPCALLOUT) {
 941                 int stopped __diagused;
 942
 943                 stopped = callout_stop(&tp->t_callout);
 944                 MPASS(stopped == 0);
 945         } else while(__predict_false(callout_stop(&tp->t_callout) == 0)) {
 946                 INP_WUNLOCK(inp);
 947                 kern_yield(PRI_UNCHANGED);
 948                 INP_WLOCK(inp);
 949         }
 950 }