net/ipv4/tcp_timer.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              Implementation of the Transmission Control Protocol(TCP).
   7  *
   8  * Authors:     Ross Biro
   9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  12  *              Florian La Roche, <flla@stud.uni-sb.de>
  13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18  *              Jorge Cwik, <jorge@laser.satlink.net>
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/gfp.h>
  23 #include <net/tcp.h>
  24
  25 int sysctl_tcp_syn_retries __read_mostly = TCP_SYN_RETRIES;
  26 int sysctl_tcp_synack_retries __read_mostly = TCP_SYNACK_RETRIES;
  27 int sysctl_tcp_keepalive_time __read_mostly = TCP_KEEPALIVE_TIME;
  28 int sysctl_tcp_keepalive_probes __read_mostly = TCP_KEEPALIVE_PROBES;
  29 int sysctl_tcp_keepalive_intvl __read_mostly = TCP_KEEPALIVE_INTVL;
  30 int sysctl_tcp_retries1 __read_mostly = TCP_RETR1;
  31 int sysctl_tcp_retries2 __read_mostly = TCP_RETR2;
  32 int sysctl_tcp_orphan_retries __read_mostly;
  33 int sysctl_tcp_thin_linear_timeouts __read_mostly;
  34
  35 static void tcp_write_err(struct sock *sk)
  36 {
  37         sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
  38         sk->sk_error_report(sk);
  39
  40         tcp_done(sk);
  41         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
  42 }
  43
  44 /* Do not allow orphaned sockets to eat all our resources.
  45  * This is direct violation of TCP specs, but it is required
  46  * to prevent DoS attacks. It is called when a retransmission timeout
  47  * or zero probe timeout occurs on orphaned socket.
  48  *
  49  * Criteria is still not confirmed experimentally and may change.
  50  * We kill the socket, if:
  51  * 1. If number of orphaned sockets exceeds an administratively configured
  52  *    limit.
  53  * 2. If we have strong memory pressure.
  54  */
  55 static int tcp_out_of_resources(struct sock *sk, int do_reset)
  56 {
  57         struct tcp_sock *tp = tcp_sk(sk);
  58         int shift = 0;
  59
  60         /* If peer does not open window for long time, or did not transmit
  61          * anything for long time, penalize it. */
  62         if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
  63                 shift++;
  64
  65         /* If some dubious ICMP arrived, penalize even more. */
  66         if (sk->sk_err_soft)
  67                 shift++;
  68
  69         if (tcp_check_oom(sk, shift)) {
  70                 /* Catch exceptional cases, when connection requires reset.
  71                  *      1. Last segment was sent recently. */
  72                 if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
  73                     /*  2. Window is closed. */
  74                     (!tp->snd_wnd && !tp->packets_out))
  75                         do_reset = 1;
  76                 if (do_reset)
  77                         tcp_send_active_reset(sk, GFP_ATOMIC);
  78                 tcp_done(sk);
  79                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
  80                 return 1;
  81         }
  82         return 0;
  83 }
  84
  85 /* Calculate maximal number or retries on an orphaned socket. */
  86 static int tcp_orphan_retries(struct sock *sk, int alive)
  87 {
  88         int retries = sysctl_tcp_orphan_retries; /* May be zero. */
  89
  90         /* We know from an ICMP that something is wrong. */
  91         if (sk->sk_err_soft && !alive)
  92                 retries = 0;
  93
  94         /* However, if socket sent something recently, select some safe
  95          * number of retries. 8 corresponds to >100 seconds with minimal
  96          * RTO of 200msec. */
  97         if (retries == 0 && alive)
  98                 retries = 8;
  99         return retries;
 100 }
 101
 102 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
 103 {
 104         /* Black hole detection */
 105         if (sysctl_tcp_mtu_probing) {
 106                 if (!icsk->icsk_mtup.enabled) {
 107                         icsk->icsk_mtup.enabled = 1;
 108                         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
 109                 } else {
 110                         struct tcp_sock *tp = tcp_sk(sk);
 111                         int mss;
 112
 113                         mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
 114                         mss = min(sysctl_tcp_base_mss, mss);
 115                         mss = max(mss, 68 - tp->tcp_header_len);
 116                         icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
 117                         tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
 118                 }
 119         }
 120 }
 121
 122 /* This function calculates a "timeout" which is equivalent to the timeout of a
 123  * TCP connection after "boundary" unsuccessful, exponentially backed-off
 124  * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
 125  * syn_set flag is set.
 126  */
 127 static bool retransmits_timed_out(struct sock *sk,
 128                                   unsigned int boundary,
 129                                   unsigned int timeout,
 130                                   bool syn_set)
 131 {
 132         unsigned int linear_backoff_thresh, start_ts;
 133         unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
 134
 135         if (!inet_csk(sk)->icsk_retransmits)
 136                 return false;
 137
 138         if (unlikely(!tcp_sk(sk)->retrans_stamp))
 139                 start_ts = TCP_SKB_CB(tcp_write_queue_head(sk))->when;
 140         else
 141                 start_ts = tcp_sk(sk)->retrans_stamp;
 142
 143         if (likely(timeout == 0)) {
 144                 linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
 145
 146                 if (boundary <= linear_backoff_thresh)
 147                         timeout = ((2 << boundary) - 1) * rto_base;
 148                 else
 149                         timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
 150                                 (boundary - linear_backoff_thresh) * TCP_RTO_MAX;
 151         }
 152         return (tcp_time_stamp - start_ts) >= timeout;
 153 }
 154
 155 /* A write timeout has occurred. Process the after effects. */
 156 static int tcp_write_timeout(struct sock *sk)
 157 {
 158         struct inet_connection_sock *icsk = inet_csk(sk);
 159         int retry_until;
 160         bool do_reset, syn_set = false;
 161
 162         if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 163                 if (icsk->icsk_retransmits)
 164                         dst_negative_advice(sk);
 165                 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
 166                 syn_set = true;
 167         } else {
 168                 if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0, 0)) {
 169                         /* Black hole detection */
 170                         tcp_mtu_probing(icsk, sk);
 171
 172                         dst_negative_advice(sk);
 173                 }
 174
 175                 retry_until = sysctl_tcp_retries2;
 176                 if (sock_flag(sk, SOCK_DEAD)) {
 177                         const int alive = (icsk->icsk_rto < TCP_RTO_MAX);
 178
 179                         retry_until = tcp_orphan_retries(sk, alive);
 180                         do_reset = alive ||
 181                                 !retransmits_timed_out(sk, retry_until, 0, 0);
 182
 183                         if (tcp_out_of_resources(sk, do_reset))
 184                                 return 1;
 185                 }
 186         }
 187
 188         if (retransmits_timed_out(sk, retry_until,
 189                                   syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
 190                 /* Has it gone just too far? */
 191                 tcp_write_err(sk);
 192                 return 1;
 193         }
 194         return 0;
 195 }
 196
 197 void tcp_delack_timer_handler(struct sock *sk)
 198 {
 199         struct tcp_sock *tp = tcp_sk(sk);
 200         struct inet_connection_sock *icsk = inet_csk(sk);
 201
 202         sk_mem_reclaim_partial(sk);
 203
 204         if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
 205                 goto out;
 206
 207         if (time_after(icsk->icsk_ack.timeout, jiffies)) {
 208                 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
 209                 goto out;
 210         }
 211         icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
 212
 213         if (!skb_queue_empty(&tp->ucopy.prequeue)) {
 214                 struct sk_buff *skb;
 215
 216                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
 217
 218                 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
 219                         sk_backlog_rcv(sk, skb);
 220
 221                 tp->ucopy.memory = 0;
 222         }
 223
 224         if (inet_csk_ack_scheduled(sk)) {
 225                 if (!icsk->icsk_ack.pingpong) {
 226                         /* Delayed ACK missed: inflate ATO. */
 227                         icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
 228                 } else {
 229                         /* Delayed ACK missed: leave pingpong mode and
 230                          * deflate ATO.
 231                          */
 232                         icsk->icsk_ack.pingpong = 0;
 233                         icsk->icsk_ack.ato      = TCP_ATO_MIN;
 234                 }
 235                 tcp_send_ack(sk);
 236                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
 237         }
 238
 239 out:
 240         if (sk_under_memory_pressure(sk))
 241                 sk_mem_reclaim(sk);
 242 }
 243
 244 static void tcp_delack_timer(unsigned long data)
 245 {
 246         struct sock *sk = (struct sock *)data;
 247
 248         bh_lock_sock(sk);
 249         if (!sock_owned_by_user(sk)) {
 250                 tcp_delack_timer_handler(sk);
 251         } else {
 252                 inet_csk(sk)->icsk_ack.blocked = 1;
 253                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
 254                 /* deleguate our work to tcp_release_cb() */
 255                 set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags);
 256         }
 257         bh_unlock_sock(sk);
 258         sock_put(sk);
 259 }
 260
 261 static void tcp_probe_timer(struct sock *sk)
 262 {
 263         struct inet_connection_sock *icsk = inet_csk(sk);
 264         struct tcp_sock *tp = tcp_sk(sk);
 265         int max_probes;
 266
 267         if (tp->packets_out || !tcp_send_head(sk)) {
 268                 icsk->icsk_probes_out = 0;
 269                 return;
 270         }
 271
 272         /* *WARNING* RFC 1122 forbids this
 273          *
 274          * It doesn't AFAIK, because we kill the retransmit timer -AK
 275          *
 276          * FIXME: We ought not to do it, Solaris 2.5 actually has fixing
 277          * this behaviour in Solaris down as a bug fix. [AC]
 278          *
 279          * Let me to explain. icsk_probes_out is zeroed by incoming ACKs
 280          * even if they advertise zero window. Hence, connection is killed only
 281          * if we received no ACKs for normal connection timeout. It is not killed
 282          * only because window stays zero for some time, window may be zero
 283          * until armageddon and even later. We are in full accordance
 284          * with RFCs, only probe timer combines both retransmission timeout
 285          * and probe timeout in one bottle.                             --ANK
 286          */
 287         max_probes = sysctl_tcp_retries2;
 288
 289         if (sock_flag(sk, SOCK_DEAD)) {
 290                 const int alive = ((icsk->icsk_rto << icsk->icsk_backoff) < TCP_RTO_MAX);
 291
 292                 max_probes = tcp_orphan_retries(sk, alive);
 293
 294                 if (tcp_out_of_resources(sk, alive || icsk->icsk_probes_out <= max_probes))
 295                         return;
 296         }
 297
 298         if (icsk->icsk_probes_out > max_probes) {
 299                 tcp_write_err(sk);
 300         } else {
 301                 /* Only send another probe if we didn't close things up. */
 302                 tcp_send_probe0(sk);
 303         }
 304 }
 305
 306 /*
 307  *      The TCP retransmit timer.
 308  */
 309
 310 void tcp_retransmit_timer(struct sock *sk)
 311 {
 312         struct tcp_sock *tp = tcp_sk(sk);
 313         struct inet_connection_sock *icsk = inet_csk(sk);
 314
 315         if (tp->early_retrans_delayed) {
 316                 tcp_resume_early_retransmit(sk);
 317                 return;
 318         }
 319
 320         if (!tp->packets_out)
 321                 goto out;
 322
 323         WARN_ON(tcp_write_queue_empty(sk));
 324
 325         if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
 326             !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
 327                 /* Receiver dastardly shrinks window. Our retransmits
 328                  * become zero probes, but we should not timeout this
 329                  * connection. If the socket is an orphan, time it out,
 330                  * we cannot allow such beasts to hang infinitely.
 331                  */
 332                 struct inet_sock *inet = inet_sk(sk);
 333                 if (sk->sk_family == AF_INET) {
 334                         LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n"),
 335                                        &inet->inet_daddr,
 336                                        ntohs(inet->inet_dport), inet->inet_num,
 337                                        tp->snd_una, tp->snd_nxt);
 338                 }
 339 #if IS_ENABLED(CONFIG_IPV6)
 340                 else if (sk->sk_family == AF_INET6) {
 341                         struct ipv6_pinfo *np = inet6_sk(sk);
 342                         LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n"),
 343                                        &np->daddr,
 344                                        ntohs(inet->inet_dport), inet->inet_num,
 345                                        tp->snd_una, tp->snd_nxt);
 346                 }
 347 #endif
 348                 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
 349                         tcp_write_err(sk);
 350                         goto out;
 351                 }
 352                 tcp_enter_loss(sk, 0);
 353                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
 354                 __sk_dst_reset(sk);
 355                 goto out_reset_timer;
 356         }
 357
 358         if (tcp_write_timeout(sk))
 359                 goto out;
 360
 361         if (icsk->icsk_retransmits == 0) {
 362                 int mib_idx;
 363
 364                 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
 365                         if (tcp_is_sack(tp))
 366                                 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
 367                         else
 368                                 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
 369                 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
 370                         mib_idx = LINUX_MIB_TCPLOSSFAILURES;
 371                 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
 372                            tp->sacked_out) {
 373                         if (tcp_is_sack(tp))
 374                                 mib_idx = LINUX_MIB_TCPSACKFAILURES;
 375                         else
 376                                 mib_idx = LINUX_MIB_TCPRENOFAILURES;
 377                 } else {
 378                         mib_idx = LINUX_MIB_TCPTIMEOUTS;
 379                 }
 380                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
 381         }
 382
 383         if (tcp_use_frto(sk)) {
 384                 tcp_enter_frto(sk);
 385         } else {
 386                 tcp_enter_loss(sk, 0);
 387         }
 388
 389         if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
 390                 /* Retransmission failed because of local congestion,
 391                  * do not backoff.
 392                  */
 393                 if (!icsk->icsk_retransmits)
 394                         icsk->icsk_retransmits = 1;
 395                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
 396                                           min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
 397                                           TCP_RTO_MAX);
 398                 goto out;
 399         }
 400
 401         /* Increase the timeout each time we retransmit.  Note that
 402          * we do not increase the rtt estimate.  rto is initialized
 403          * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
 404          * that doubling rto each time is the least we can get away with.
 405          * In KA9Q, Karn uses this for the first few times, and then
 406          * goes to quadratic.  netBSD doubles, but only goes up to *64,
 407          * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
 408          * defined in the protocol as the maximum possible RTT.  I guess
 409          * we'll have to use something other than TCP to talk to the
 410          * University of Mars.
 411          *
 412          * PAWS allows us longer timeouts and large windows, so once
 413          * implemented ftp to mars will work nicely. We will have to fix
 414          * the 120 second clamps though!
 415          */
 416         icsk->icsk_backoff++;
 417         icsk->icsk_retransmits++;
 418
 419 out_reset_timer:
 420         /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
 421          * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
 422          * might be increased if the stream oscillates between thin and thick,
 423          * thus the old value might already be too high compared to the value
 424          * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
 425          * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
 426          * exponential backoff behaviour to avoid continue hammering
 427          * linear-timeout retransmissions into a black hole
 428          */
 429         if (sk->sk_state == TCP_ESTABLISHED &&
 430             (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
 431             tcp_stream_is_thin(tp) &&
 432             icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
 433                 icsk->icsk_backoff = 0;
 434                 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
 435         } else {
 436                 /* Use normal (exponential) backoff */
 437                 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
 438         }
 439         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
 440         if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0, 0))
 441                 __sk_dst_reset(sk);
 442
 443 out:;
 444 }
 445
 446 void tcp_write_timer_handler(struct sock *sk)
 447 {
 448         struct inet_connection_sock *icsk = inet_csk(sk);
 449         int event;
 450
 451         if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
 452                 goto out;
 453
 454         if (time_after(icsk->icsk_timeout, jiffies)) {
 455                 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
 456                 goto out;
 457         }
 458
 459         event = icsk->icsk_pending;
 460         icsk->icsk_pending = 0;
 461
 462         switch (event) {
 463         case ICSK_TIME_RETRANS:
 464                 tcp_retransmit_timer(sk);
 465                 break;
 466         case ICSK_TIME_PROBE0:
 467                 tcp_probe_timer(sk);
 468                 break;
 469         }
 470
 471 out:
 472         sk_mem_reclaim(sk);
 473 }
 474
 475 static void tcp_write_timer(unsigned long data)
 476 {
 477         struct sock *sk = (struct sock *)data;
 478
 479         bh_lock_sock(sk);
 480         if (!sock_owned_by_user(sk)) {
 481                 tcp_write_timer_handler(sk);
 482         } else {
 483                 /* deleguate our work to tcp_release_cb() */
 484                 set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags);
 485         }
 486         bh_unlock_sock(sk);
 487         sock_put(sk);
 488 }
 489
 490 /*
 491  *      Timer for listening sockets
 492  */
 493
 494 static void tcp_synack_timer(struct sock *sk)
 495 {
 496         inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL,
 497                                    TCP_TIMEOUT_INIT, TCP_RTO_MAX);
 498 }
 499
 500 void tcp_syn_ack_timeout(struct sock *sk, struct request_sock *req)
 501 {
 502         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
 503 }
 504 EXPORT_SYMBOL(tcp_syn_ack_timeout);
 505
 506 void tcp_set_keepalive(struct sock *sk, int val)
 507 {
 508         if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
 509                 return;
 510
 511         if (val && !sock_flag(sk, SOCK_KEEPOPEN))
 512                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
 513         else if (!val)
 514                 inet_csk_delete_keepalive_timer(sk);
 515 }
 516
 517
 518 static void tcp_keepalive_timer (unsigned long data)
 519 {
 520         struct sock *sk = (struct sock *) data;
 521         struct inet_connection_sock *icsk = inet_csk(sk);
 522         struct tcp_sock *tp = tcp_sk(sk);
 523         u32 elapsed;
 524
 525         /* Only process if socket is not in use. */
 526         bh_lock_sock(sk);
 527         if (sock_owned_by_user(sk)) {
 528                 /* Try again later. */
 529                 inet_csk_reset_keepalive_timer (sk, HZ/20);
 530                 goto out;
 531         }
 532
 533         if (sk->sk_state == TCP_LISTEN) {
 534                 tcp_synack_timer(sk);
 535                 goto out;
 536         }
 537
 538         if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
 539                 if (tp->linger2 >= 0) {
 540                         const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
 541
 542                         if (tmo > 0) {
 543                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
 544                                 goto out;
 545                         }
 546                 }
 547                 tcp_send_active_reset(sk, GFP_ATOMIC);
 548                 goto death;
 549         }
 550
 551         if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
 552                 goto out;
 553
 554         elapsed = keepalive_time_when(tp);
 555
 556         /* It is alive without keepalive 8) */
 557         if (tp->packets_out || tcp_send_head(sk))
 558                 goto resched;
 559
 560         elapsed = keepalive_time_elapsed(tp);
 561
 562         if (elapsed >= keepalive_time_when(tp)) {
 563                 /* If the TCP_USER_TIMEOUT option is enabled, use that
 564                  * to determine when to timeout instead.
 565                  */
 566                 if ((icsk->icsk_user_timeout != 0 &&
 567                     elapsed >= icsk->icsk_user_timeout &&
 568                     icsk->icsk_probes_out > 0) ||
 569                     (icsk->icsk_user_timeout == 0 &&
 570                     icsk->icsk_probes_out >= keepalive_probes(tp))) {
 571                         tcp_send_active_reset(sk, GFP_ATOMIC);
 572                         tcp_write_err(sk);
 573                         goto out;
 574                 }
 575                 if (tcp_write_wakeup(sk) <= 0) {
 576                         icsk->icsk_probes_out++;
 577                         elapsed = keepalive_intvl_when(tp);
 578                 } else {
 579                         /* If keepalive was lost due to local congestion,
 580                          * try harder.
 581                          */
 582                         elapsed = TCP_RESOURCE_PROBE_INTERVAL;
 583                 }
 584         } else {
 585                 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
 586                 elapsed = keepalive_time_when(tp) - elapsed;
 587         }
 588
 589         sk_mem_reclaim(sk);
 590
 591 resched:
 592         inet_csk_reset_keepalive_timer (sk, elapsed);
 593         goto out;
 594
 595 death:
 596         tcp_done(sk);
 597
 598 out:
 599         bh_unlock_sock(sk);
 600         sock_put(sk);
 601 }
 602
 603 void tcp_init_xmit_timers(struct sock *sk)
 604 {
 605         inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
 606                                   &tcp_keepalive_timer);
 607 }
 608 EXPORT_SYMBOL(tcp_init_xmit_timers);