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, bool 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 = true;
  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         start_ts = tcp_sk(sk)->retrans_stamp;
 139         if (unlikely(!start_ts))
 140                 start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
 141
 142         if (likely(timeout == 0)) {
 143                 linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
 144
 145                 if (boundary <= linear_backoff_thresh)
 146                         timeout = ((2 << boundary) - 1) * rto_base;
 147                 else
 148                         timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
 149                                 (boundary - linear_backoff_thresh) * TCP_RTO_MAX;
 150         }
 151         return (tcp_time_stamp - start_ts) >= timeout;
 152 }
 153
 154 /* A write timeout has occurred. Process the after effects. */
 155 static int tcp_write_timeout(struct sock *sk)
 156 {
 157         struct inet_connection_sock *icsk = inet_csk(sk);
 158         struct tcp_sock *tp = tcp_sk(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                         if (tp->syn_fastopen || tp->syn_data)
 166                                 tcp_fastopen_cache_set(sk, 0, NULL, true);
 167                         if (tp->syn_data)
 168                                 NET_INC_STATS_BH(sock_net(sk),
 169                                                  LINUX_MIB_TCPFASTOPENACTIVEFAIL);
 170                 }
 171                 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
 172                 syn_set = true;
 173         } else {
 174                 if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0, 0)) {
 175                         /* Black hole detection */
 176                         tcp_mtu_probing(icsk, sk);
 177
 178                         dst_negative_advice(sk);
 179                 }
 180
 181                 retry_until = sysctl_tcp_retries2;
 182                 if (sock_flag(sk, SOCK_DEAD)) {
 183                         const int alive = icsk->icsk_rto < TCP_RTO_MAX;
 184
 185                         retry_until = tcp_orphan_retries(sk, alive);
 186                         do_reset = alive ||
 187                                 !retransmits_timed_out(sk, retry_until, 0, 0);
 188
 189                         if (tcp_out_of_resources(sk, do_reset))
 190                                 return 1;
 191                 }
 192         }
 193
 194         if (retransmits_timed_out(sk, retry_until,
 195                                   syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
 196                 /* Has it gone just too far? */
 197                 tcp_write_err(sk);
 198                 return 1;
 199         }
 200         return 0;
 201 }
 202
 203 void tcp_delack_timer_handler(struct sock *sk)
 204 {
 205         struct tcp_sock *tp = tcp_sk(sk);
 206         struct inet_connection_sock *icsk = inet_csk(sk);
 207
 208         sk_mem_reclaim_partial(sk);
 209
 210         if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
 211                 goto out;
 212
 213         if (time_after(icsk->icsk_ack.timeout, jiffies)) {
 214                 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
 215                 goto out;
 216         }
 217         icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
 218
 219         if (!skb_queue_empty(&tp->ucopy.prequeue)) {
 220                 struct sk_buff *skb;
 221
 222                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
 223
 224                 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
 225                         sk_backlog_rcv(sk, skb);
 226
 227                 tp->ucopy.memory = 0;
 228         }
 229
 230         if (inet_csk_ack_scheduled(sk)) {
 231                 if (!icsk->icsk_ack.pingpong) {
 232                         /* Delayed ACK missed: inflate ATO. */
 233                         icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
 234                 } else {
 235                         /* Delayed ACK missed: leave pingpong mode and
 236                          * deflate ATO.
 237                          */
 238                         icsk->icsk_ack.pingpong = 0;
 239                         icsk->icsk_ack.ato      = TCP_ATO_MIN;
 240                 }
 241                 tcp_send_ack(sk);
 242                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS);
 243         }
 244
 245 out:
 246         if (sk_under_memory_pressure(sk))
 247                 sk_mem_reclaim(sk);
 248 }
 249
 250 static void tcp_delack_timer(unsigned long data)
 251 {
 252         struct sock *sk = (struct sock *)data;
 253
 254         bh_lock_sock(sk);
 255         if (!sock_owned_by_user(sk)) {
 256                 tcp_delack_timer_handler(sk);
 257         } else {
 258                 inet_csk(sk)->icsk_ack.blocked = 1;
 259                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
 260                 /* deleguate our work to tcp_release_cb() */
 261                 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
 262                         sock_hold(sk);
 263         }
 264         bh_unlock_sock(sk);
 265         sock_put(sk);
 266 }
 267
 268 static void tcp_probe_timer(struct sock *sk)
 269 {
 270         struct inet_connection_sock *icsk = inet_csk(sk);
 271         struct tcp_sock *tp = tcp_sk(sk);
 272         int max_probes;
 273         u32 start_ts;
 274
 275         if (tp->packets_out || !tcp_send_head(sk)) {
 276                 icsk->icsk_probes_out = 0;
 277                 return;
 278         }
 279
 280         /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
 281          * long as the receiver continues to respond probes. We support this by
 282          * default and reset icsk_probes_out with incoming ACKs. But if the
 283          * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
 284          * kill the socket when the retry count and the time exceeds the
 285          * corresponding system limit. We also implement similar policy when
 286          * we use RTO to probe window in tcp_retransmit_timer().
 287          */
 288         start_ts = tcp_skb_timestamp(tcp_send_head(sk));
 289         if (!start_ts)
 290                 skb_mstamp_get(&tcp_send_head(sk)->skb_mstamp);
 291         else if (icsk->icsk_user_timeout &&
 292                  (s32)(tcp_time_stamp - start_ts) > icsk->icsk_user_timeout)
 293                 goto abort;
 294
 295         max_probes = sysctl_tcp_retries2;
 296         if (sock_flag(sk, SOCK_DEAD)) {
 297                 const int alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
 298
 299                 max_probes = tcp_orphan_retries(sk, alive);
 300                 if (!alive && icsk->icsk_backoff >= max_probes)
 301                         goto abort;
 302                 if (tcp_out_of_resources(sk, true))
 303                         return;
 304         }
 305
 306         if (icsk->icsk_probes_out > max_probes) {
 307 abort:          tcp_write_err(sk);
 308         } else {
 309                 /* Only send another probe if we didn't close things up. */
 310                 tcp_send_probe0(sk);
 311         }
 312 }
 313
 314 /*
 315  *      Timer for Fast Open socket to retransmit SYNACK. Note that the
 316  *      sk here is the child socket, not the parent (listener) socket.
 317  */
 318 static void tcp_fastopen_synack_timer(struct sock *sk)
 319 {
 320         struct inet_connection_sock *icsk = inet_csk(sk);
 321         int max_retries = icsk->icsk_syn_retries ? :
 322             sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
 323         struct request_sock *req;
 324
 325         req = tcp_sk(sk)->fastopen_rsk;
 326         req->rsk_ops->syn_ack_timeout(sk, req);
 327
 328         if (req->num_timeout >= max_retries) {
 329                 tcp_write_err(sk);
 330                 return;
 331         }
 332         /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
 333          * returned from rtx_syn_ack() to make it more persistent like
 334          * regular retransmit because if the child socket has been accepted
 335          * it's not good to give up too easily.
 336          */
 337         inet_rtx_syn_ack(sk, req);
 338         req->num_timeout++;
 339         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
 340                           TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
 341 }
 342
 343 /*
 344  *      The TCP retransmit timer.
 345  */
 346
 347 void tcp_retransmit_timer(struct sock *sk)
 348 {
 349         struct tcp_sock *tp = tcp_sk(sk);
 350         struct inet_connection_sock *icsk = inet_csk(sk);
 351
 352         if (tp->fastopen_rsk) {
 353                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
 354                              sk->sk_state != TCP_FIN_WAIT1);
 355                 tcp_fastopen_synack_timer(sk);
 356                 /* Before we receive ACK to our SYN-ACK don't retransmit
 357                  * anything else (e.g., data or FIN segments).
 358                  */
 359                 return;
 360         }
 361         if (!tp->packets_out)
 362                 goto out;
 363
 364         WARN_ON(tcp_write_queue_empty(sk));
 365
 366         tp->tlp_high_seq = 0;
 367
 368         if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
 369             !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
 370                 /* Receiver dastardly shrinks window. Our retransmits
 371                  * become zero probes, but we should not timeout this
 372                  * connection. If the socket is an orphan, time it out,
 373                  * we cannot allow such beasts to hang infinitely.
 374                  */
 375                 struct inet_sock *inet = inet_sk(sk);
 376                 if (sk->sk_family == AF_INET) {
 377                         net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
 378                                             &inet->inet_daddr,
 379                                             ntohs(inet->inet_dport),
 380                                             inet->inet_num,
 381                                             tp->snd_una, tp->snd_nxt);
 382                 }
 383 #if IS_ENABLED(CONFIG_IPV6)
 384                 else if (sk->sk_family == AF_INET6) {
 385                         net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
 386                                             &sk->sk_v6_daddr,
 387                                             ntohs(inet->inet_dport),
 388                                             inet->inet_num,
 389                                             tp->snd_una, tp->snd_nxt);
 390                 }
 391 #endif
 392                 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
 393                         tcp_write_err(sk);
 394                         goto out;
 395                 }
 396                 tcp_enter_loss(sk);
 397                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
 398                 __sk_dst_reset(sk);
 399                 goto out_reset_timer;
 400         }
 401
 402         if (tcp_write_timeout(sk))
 403                 goto out;
 404
 405         if (icsk->icsk_retransmits == 0) {
 406                 int mib_idx;
 407
 408                 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
 409                         if (tcp_is_sack(tp))
 410                                 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
 411                         else
 412                                 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
 413                 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
 414                         mib_idx = LINUX_MIB_TCPLOSSFAILURES;
 415                 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
 416                            tp->sacked_out) {
 417                         if (tcp_is_sack(tp))
 418                                 mib_idx = LINUX_MIB_TCPSACKFAILURES;
 419                         else
 420                                 mib_idx = LINUX_MIB_TCPRENOFAILURES;
 421                 } else {
 422                         mib_idx = LINUX_MIB_TCPTIMEOUTS;
 423                 }
 424                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
 425         }
 426
 427         tcp_enter_loss(sk);
 428
 429         if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) {
 430                 /* Retransmission failed because of local congestion,
 431                  * do not backoff.
 432                  */
 433                 if (!icsk->icsk_retransmits)
 434                         icsk->icsk_retransmits = 1;
 435                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
 436                                           min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
 437                                           TCP_RTO_MAX);
 438                 goto out;
 439         }
 440
 441         /* Increase the timeout each time we retransmit.  Note that
 442          * we do not increase the rtt estimate.  rto is initialized
 443          * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
 444          * that doubling rto each time is the least we can get away with.
 445          * In KA9Q, Karn uses this for the first few times, and then
 446          * goes to quadratic.  netBSD doubles, but only goes up to *64,
 447          * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
 448          * defined in the protocol as the maximum possible RTT.  I guess
 449          * we'll have to use something other than TCP to talk to the
 450          * University of Mars.
 451          *
 452          * PAWS allows us longer timeouts and large windows, so once
 453          * implemented ftp to mars will work nicely. We will have to fix
 454          * the 120 second clamps though!
 455          */
 456         icsk->icsk_backoff++;
 457         icsk->icsk_retransmits++;
 458
 459 out_reset_timer:
 460         /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
 461          * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
 462          * might be increased if the stream oscillates between thin and thick,
 463          * thus the old value might already be too high compared to the value
 464          * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
 465          * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
 466          * exponential backoff behaviour to avoid continue hammering
 467          * linear-timeout retransmissions into a black hole
 468          */
 469         if (sk->sk_state == TCP_ESTABLISHED &&
 470             (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
 471             tcp_stream_is_thin(tp) &&
 472             icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
 473                 icsk->icsk_backoff = 0;
 474                 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
 475         } else {
 476                 /* Use normal (exponential) backoff */
 477                 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
 478         }
 479         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
 480         if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0, 0))
 481                 __sk_dst_reset(sk);
 482
 483 out:;
 484 }
 485
 486 void tcp_write_timer_handler(struct sock *sk)
 487 {
 488         struct inet_connection_sock *icsk = inet_csk(sk);
 489         int event;
 490
 491         if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
 492                 goto out;
 493
 494         if (time_after(icsk->icsk_timeout, jiffies)) {
 495                 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
 496                 goto out;
 497         }
 498
 499         event = icsk->icsk_pending;
 500
 501         switch (event) {
 502         case ICSK_TIME_EARLY_RETRANS:
 503                 tcp_resume_early_retransmit(sk);
 504                 break;
 505         case ICSK_TIME_LOSS_PROBE:
 506                 tcp_send_loss_probe(sk);
 507                 break;
 508         case ICSK_TIME_RETRANS:
 509                 icsk->icsk_pending = 0;
 510                 tcp_retransmit_timer(sk);
 511                 break;
 512         case ICSK_TIME_PROBE0:
 513                 icsk->icsk_pending = 0;
 514                 tcp_probe_timer(sk);
 515                 break;
 516         }
 517
 518 out:
 519         sk_mem_reclaim(sk);
 520 }
 521
 522 static void tcp_write_timer(unsigned long data)
 523 {
 524         struct sock *sk = (struct sock *)data;
 525
 526         bh_lock_sock(sk);
 527         if (!sock_owned_by_user(sk)) {
 528                 tcp_write_timer_handler(sk);
 529         } else {
 530                 /* deleguate our work to tcp_release_cb() */
 531                 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &tcp_sk(sk)->tsq_flags))
 532                         sock_hold(sk);
 533         }
 534         bh_unlock_sock(sk);
 535         sock_put(sk);
 536 }
 537
 538 /*
 539  *      Timer for listening sockets
 540  */
 541
 542 static void tcp_synack_timer(struct sock *sk)
 543 {
 544         inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL,
 545                                    TCP_TIMEOUT_INIT, TCP_RTO_MAX);
 546 }
 547
 548 void tcp_syn_ack_timeout(struct sock *sk, struct request_sock *req)
 549 {
 550         NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEOUTS);
 551 }
 552 EXPORT_SYMBOL(tcp_syn_ack_timeout);
 553
 554 void tcp_set_keepalive(struct sock *sk, int val)
 555 {
 556         if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
 557                 return;
 558
 559         if (val && !sock_flag(sk, SOCK_KEEPOPEN))
 560                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
 561         else if (!val)
 562                 inet_csk_delete_keepalive_timer(sk);
 563 }
 564
 565
 566 static void tcp_keepalive_timer (unsigned long data)
 567 {
 568         struct sock *sk = (struct sock *) data;
 569         struct inet_connection_sock *icsk = inet_csk(sk);
 570         struct tcp_sock *tp = tcp_sk(sk);
 571         u32 elapsed;
 572
 573         /* Only process if socket is not in use. */
 574         bh_lock_sock(sk);
 575         if (sock_owned_by_user(sk)) {
 576                 /* Try again later. */
 577                 inet_csk_reset_keepalive_timer (sk, HZ/20);
 578                 goto out;
 579         }
 580
 581         if (sk->sk_state == TCP_LISTEN) {
 582                 tcp_synack_timer(sk);
 583                 goto out;
 584         }
 585
 586         if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
 587                 if (tp->linger2 >= 0) {
 588                         const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
 589
 590                         if (tmo > 0) {
 591                                 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
 592                                 goto out;
 593                         }
 594                 }
 595                 tcp_send_active_reset(sk, GFP_ATOMIC);
 596                 goto death;
 597         }
 598
 599         if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
 600                 goto out;
 601
 602         elapsed = keepalive_time_when(tp);
 603
 604         /* It is alive without keepalive 8) */
 605         if (tp->packets_out || tcp_send_head(sk))
 606                 goto resched;
 607
 608         elapsed = keepalive_time_elapsed(tp);
 609
 610         if (elapsed >= keepalive_time_when(tp)) {
 611                 /* If the TCP_USER_TIMEOUT option is enabled, use that
 612                  * to determine when to timeout instead.
 613                  */
 614                 if ((icsk->icsk_user_timeout != 0 &&
 615                     elapsed >= icsk->icsk_user_timeout &&
 616                     icsk->icsk_probes_out > 0) ||
 617                     (icsk->icsk_user_timeout == 0 &&
 618                     icsk->icsk_probes_out >= keepalive_probes(tp))) {
 619                         tcp_send_active_reset(sk, GFP_ATOMIC);
 620                         tcp_write_err(sk);
 621                         goto out;
 622                 }
 623                 if (tcp_write_wakeup(sk) <= 0) {
 624                         icsk->icsk_probes_out++;
 625                         elapsed = keepalive_intvl_when(tp);
 626                 } else {
 627                         /* If keepalive was lost due to local congestion,
 628                          * try harder.
 629                          */
 630                         elapsed = TCP_RESOURCE_PROBE_INTERVAL;
 631                 }
 632         } else {
 633                 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
 634                 elapsed = keepalive_time_when(tp) - elapsed;
 635         }
 636
 637         sk_mem_reclaim(sk);
 638
 639 resched:
 640         inet_csk_reset_keepalive_timer (sk, elapsed);
 641         goto out;
 642
 643 death:
 644         tcp_done(sk);
 645
 646 out:
 647         bh_unlock_sock(sk);
 648         sock_put(sk);
 649 }
 650
 651 void tcp_init_xmit_timers(struct sock *sk)
 652 {
 653         inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
 654                                   &tcp_keepalive_timer);
 655 }
 656 EXPORT_SYMBOL(tcp_init_xmit_timers);