linux/net/ipv4/tcp_timer.c
<<
>>
Prefs
   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
  25int sysctl_tcp_thin_linear_timeouts __read_mostly;
  26
  27/**
  28 *  tcp_write_err() - close socket and save error info
  29 *  @sk:  The socket the error has appeared on.
  30 *
  31 *  Returns: Nothing (void)
  32 */
  33
  34static void tcp_write_err(struct sock *sk)
  35{
  36        sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
  37        sk->sk_error_report(sk);
  38
  39        tcp_done(sk);
  40        __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
  41}
  42
  43/**
  44 *  tcp_out_of_resources() - Close socket if out of resources
  45 *  @sk:        pointer to current socket
  46 *  @do_reset:  send a last packet with reset flag
  47 *
  48 *  Do not allow orphaned sockets to eat all our resources.
  49 *  This is direct violation of TCP specs, but it is required
  50 *  to prevent DoS attacks. It is called when a retransmission timeout
  51 *  or zero probe timeout occurs on orphaned socket.
  52 *
  53 *  Criteria is still not confirmed experimentally and may change.
  54 *  We kill the socket, if:
  55 *  1. If number of orphaned sockets exceeds an administratively configured
  56 *     limit.
  57 *  2. If we have strong memory pressure.
  58 */
  59static int tcp_out_of_resources(struct sock *sk, bool do_reset)
  60{
  61        struct tcp_sock *tp = tcp_sk(sk);
  62        int shift = 0;
  63
  64        /* If peer does not open window for long time, or did not transmit
  65         * anything for long time, penalize it. */
  66        if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
  67                shift++;
  68
  69        /* If some dubious ICMP arrived, penalize even more. */
  70        if (sk->sk_err_soft)
  71                shift++;
  72
  73        if (tcp_check_oom(sk, shift)) {
  74                /* Catch exceptional cases, when connection requires reset.
  75                 *      1. Last segment was sent recently. */
  76                if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
  77                    /*  2. Window is closed. */
  78                    (!tp->snd_wnd && !tp->packets_out))
  79                        do_reset = true;
  80                if (do_reset)
  81                        tcp_send_active_reset(sk, GFP_ATOMIC);
  82                tcp_done(sk);
  83                __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
  84                return 1;
  85        }
  86        return 0;
  87}
  88
  89/**
  90 *  tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
  91 *  @sk:    Pointer to the current socket.
  92 *  @alive: bool, socket alive state
  93 */
  94static int tcp_orphan_retries(struct sock *sk, bool alive)
  95{
  96        int retries = sock_net(sk)->ipv4.sysctl_tcp_orphan_retries; /* May be zero. */
  97
  98        /* We know from an ICMP that something is wrong. */
  99        if (sk->sk_err_soft && !alive)
 100                retries = 0;
 101
 102        /* However, if socket sent something recently, select some safe
 103         * number of retries. 8 corresponds to >100 seconds with minimal
 104         * RTO of 200msec. */
 105        if (retries == 0 && alive)
 106                retries = 8;
 107        return retries;
 108}
 109
 110static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
 111{
 112        struct net *net = sock_net(sk);
 113
 114        /* Black hole detection */
 115        if (net->ipv4.sysctl_tcp_mtu_probing) {
 116                if (!icsk->icsk_mtup.enabled) {
 117                        icsk->icsk_mtup.enabled = 1;
 118                        icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
 119                        tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
 120                } else {
 121                        struct net *net = sock_net(sk);
 122                        struct tcp_sock *tp = tcp_sk(sk);
 123                        int mss;
 124
 125                        mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
 126                        mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
 127                        mss = max(mss, 68 - tp->tcp_header_len);
 128                        icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
 129                        tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
 130                }
 131        }
 132}
 133
 134
 135/**
 136 *  retransmits_timed_out() - returns true if this connection has timed out
 137 *  @sk:       The current socket
 138 *  @boundary: max number of retransmissions
 139 *  @timeout:  A custom timeout value.
 140 *             If set to 0 the default timeout is calculated and used.
 141 *             Using TCP_RTO_MIN and the number of unsuccessful retransmits.
 142 *  @syn_set:  true if the SYN Bit was set.
 143 *
 144 * The default "timeout" value this function can calculate and use
 145 * is equivalent to the timeout of a TCP Connection
 146 * after "boundary" unsuccessful, exponentially backed-off
 147 * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if
 148 * syn_set flag is set.
 149 *
 150 */
 151static bool retransmits_timed_out(struct sock *sk,
 152                                  unsigned int boundary,
 153                                  unsigned int timeout,
 154                                  bool syn_set)
 155{
 156        unsigned int linear_backoff_thresh, start_ts;
 157        unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN;
 158
 159        if (!inet_csk(sk)->icsk_retransmits)
 160                return false;
 161
 162        start_ts = tcp_sk(sk)->retrans_stamp;
 163        if (unlikely(!start_ts))
 164                start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
 165
 166        if (likely(timeout == 0)) {
 167                linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
 168
 169                if (boundary <= linear_backoff_thresh)
 170                        timeout = ((2 << boundary) - 1) * rto_base;
 171                else
 172                        timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
 173                                (boundary - linear_backoff_thresh) * TCP_RTO_MAX;
 174        }
 175        return (tcp_time_stamp - start_ts) >= timeout;
 176}
 177
 178/* A write timeout has occurred. Process the after effects. */
 179static int tcp_write_timeout(struct sock *sk)
 180{
 181        struct inet_connection_sock *icsk = inet_csk(sk);
 182        struct tcp_sock *tp = tcp_sk(sk);
 183        struct net *net = sock_net(sk);
 184        int retry_until;
 185        bool do_reset, syn_set = false;
 186
 187        if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 188                if (icsk->icsk_retransmits) {
 189                        dst_negative_advice(sk);
 190                        if (tp->syn_fastopen || tp->syn_data)
 191                                tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
 192                        if (tp->syn_data && icsk->icsk_retransmits == 1)
 193                                NET_INC_STATS(sock_net(sk),
 194                                              LINUX_MIB_TCPFASTOPENACTIVEFAIL);
 195                } else if (!tp->syn_data && !tp->syn_fastopen) {
 196                        sk_rethink_txhash(sk);
 197                }
 198                retry_until = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
 199                syn_set = true;
 200        } else {
 201                if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1, 0, 0)) {
 202                        /* Some middle-boxes may black-hole Fast Open _after_
 203                         * the handshake. Therefore we conservatively disable
 204                         * Fast Open on this path on recurring timeouts with
 205                         * few or zero bytes acked after Fast Open.
 206                         */
 207                        if (tp->syn_data_acked &&
 208                            tp->bytes_acked <= tp->rx_opt.mss_clamp) {
 209                                tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
 210                                if (icsk->icsk_retransmits == net->ipv4.sysctl_tcp_retries1)
 211                                        NET_INC_STATS(sock_net(sk),
 212                                                      LINUX_MIB_TCPFASTOPENACTIVEFAIL);
 213                        }
 214                        /* Black hole detection */
 215                        tcp_mtu_probing(icsk, sk);
 216
 217                        dst_negative_advice(sk);
 218                } else {
 219                        sk_rethink_txhash(sk);
 220                }
 221
 222                retry_until = net->ipv4.sysctl_tcp_retries2;
 223                if (sock_flag(sk, SOCK_DEAD)) {
 224                        const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
 225
 226                        retry_until = tcp_orphan_retries(sk, alive);
 227                        do_reset = alive ||
 228                                !retransmits_timed_out(sk, retry_until, 0, 0);
 229
 230                        if (tcp_out_of_resources(sk, do_reset))
 231                                return 1;
 232                }
 233        }
 234
 235        if (retransmits_timed_out(sk, retry_until,
 236                                  syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) {
 237                /* Has it gone just too far? */
 238                tcp_write_err(sk);
 239                return 1;
 240        }
 241        return 0;
 242}
 243
 244/* Called with BH disabled */
 245void tcp_delack_timer_handler(struct sock *sk)
 246{
 247        struct tcp_sock *tp = tcp_sk(sk);
 248        struct inet_connection_sock *icsk = inet_csk(sk);
 249
 250        sk_mem_reclaim_partial(sk);
 251
 252        if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
 253                goto out;
 254
 255        if (time_after(icsk->icsk_ack.timeout, jiffies)) {
 256                sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
 257                goto out;
 258        }
 259        icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
 260
 261        if (!skb_queue_empty(&tp->ucopy.prequeue)) {
 262                struct sk_buff *skb;
 263
 264                __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
 265
 266                while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
 267                        sk_backlog_rcv(sk, skb);
 268
 269                tp->ucopy.memory = 0;
 270        }
 271
 272        if (inet_csk_ack_scheduled(sk)) {
 273                if (!icsk->icsk_ack.pingpong) {
 274                        /* Delayed ACK missed: inflate ATO. */
 275                        icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
 276                } else {
 277                        /* Delayed ACK missed: leave pingpong mode and
 278                         * deflate ATO.
 279                         */
 280                        icsk->icsk_ack.pingpong = 0;
 281                        icsk->icsk_ack.ato      = TCP_ATO_MIN;
 282                }
 283                tcp_send_ack(sk);
 284                __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
 285        }
 286
 287out:
 288        if (tcp_under_memory_pressure(sk))
 289                sk_mem_reclaim(sk);
 290}
 291
 292
 293/**
 294 *  tcp_delack_timer() - The TCP delayed ACK timeout handler
 295 *  @data:  Pointer to the current socket. (gets casted to struct sock *)
 296 *
 297 *  This function gets (indirectly) called when the kernel timer for a TCP packet
 298 *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
 299 *
 300 *  Returns: Nothing (void)
 301 */
 302static void tcp_delack_timer(unsigned long data)
 303{
 304        struct sock *sk = (struct sock *)data;
 305
 306        bh_lock_sock(sk);
 307        if (!sock_owned_by_user(sk)) {
 308                tcp_delack_timer_handler(sk);
 309        } else {
 310                inet_csk(sk)->icsk_ack.blocked = 1;
 311                __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
 312                /* deleguate our work to tcp_release_cb() */
 313                if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
 314                        sock_hold(sk);
 315        }
 316        bh_unlock_sock(sk);
 317        sock_put(sk);
 318}
 319
 320static void tcp_probe_timer(struct sock *sk)
 321{
 322        struct inet_connection_sock *icsk = inet_csk(sk);
 323        struct tcp_sock *tp = tcp_sk(sk);
 324        int max_probes;
 325        u32 start_ts;
 326
 327        if (tp->packets_out || !tcp_send_head(sk)) {
 328                icsk->icsk_probes_out = 0;
 329                return;
 330        }
 331
 332        /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
 333         * long as the receiver continues to respond probes. We support this by
 334         * default and reset icsk_probes_out with incoming ACKs. But if the
 335         * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
 336         * kill the socket when the retry count and the time exceeds the
 337         * corresponding system limit. We also implement similar policy when
 338         * we use RTO to probe window in tcp_retransmit_timer().
 339         */
 340        start_ts = tcp_skb_timestamp(tcp_send_head(sk));
 341        if (!start_ts)
 342                skb_mstamp_get(&tcp_send_head(sk)->skb_mstamp);
 343        else if (icsk->icsk_user_timeout &&
 344                 (s32)(tcp_time_stamp - start_ts) > icsk->icsk_user_timeout)
 345                goto abort;
 346
 347        max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2;
 348        if (sock_flag(sk, SOCK_DEAD)) {
 349                const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
 350
 351                max_probes = tcp_orphan_retries(sk, alive);
 352                if (!alive && icsk->icsk_backoff >= max_probes)
 353                        goto abort;
 354                if (tcp_out_of_resources(sk, true))
 355                        return;
 356        }
 357
 358        if (icsk->icsk_probes_out > max_probes) {
 359abort:          tcp_write_err(sk);
 360        } else {
 361                /* Only send another probe if we didn't close things up. */
 362                tcp_send_probe0(sk);
 363        }
 364}
 365
 366/*
 367 *      Timer for Fast Open socket to retransmit SYNACK. Note that the
 368 *      sk here is the child socket, not the parent (listener) socket.
 369 */
 370static void tcp_fastopen_synack_timer(struct sock *sk)
 371{
 372        struct inet_connection_sock *icsk = inet_csk(sk);
 373        int max_retries = icsk->icsk_syn_retries ? :
 374            sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
 375        struct request_sock *req;
 376
 377        req = tcp_sk(sk)->fastopen_rsk;
 378        req->rsk_ops->syn_ack_timeout(req);
 379
 380        if (req->num_timeout >= max_retries) {
 381                tcp_write_err(sk);
 382                return;
 383        }
 384        /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
 385         * returned from rtx_syn_ack() to make it more persistent like
 386         * regular retransmit because if the child socket has been accepted
 387         * it's not good to give up too easily.
 388         */
 389        inet_rtx_syn_ack(sk, req);
 390        req->num_timeout++;
 391        icsk->icsk_retransmits++;
 392        inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
 393                          TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
 394}
 395
 396
 397/**
 398 *  tcp_retransmit_timer() - The TCP retransmit timeout handler
 399 *  @sk:  Pointer to the current socket.
 400 *
 401 *  This function gets called when the kernel timer for a TCP packet
 402 *  of this socket expires.
 403 *
 404 *  It handles retransmission, timer adjustment and other necesarry measures.
 405 *
 406 *  Returns: Nothing (void)
 407 */
 408void tcp_retransmit_timer(struct sock *sk)
 409{
 410        struct tcp_sock *tp = tcp_sk(sk);
 411        struct net *net = sock_net(sk);
 412        struct inet_connection_sock *icsk = inet_csk(sk);
 413
 414        if (tp->fastopen_rsk) {
 415                WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
 416                             sk->sk_state != TCP_FIN_WAIT1);
 417                tcp_fastopen_synack_timer(sk);
 418                /* Before we receive ACK to our SYN-ACK don't retransmit
 419                 * anything else (e.g., data or FIN segments).
 420                 */
 421                return;
 422        }
 423        if (!tp->packets_out)
 424                goto out;
 425
 426        WARN_ON(tcp_write_queue_empty(sk));
 427
 428        tp->tlp_high_seq = 0;
 429
 430        if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
 431            !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
 432                /* Receiver dastardly shrinks window. Our retransmits
 433                 * become zero probes, but we should not timeout this
 434                 * connection. If the socket is an orphan, time it out,
 435                 * we cannot allow such beasts to hang infinitely.
 436                 */
 437                struct inet_sock *inet = inet_sk(sk);
 438                if (sk->sk_family == AF_INET) {
 439                        net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
 440                                            &inet->inet_daddr,
 441                                            ntohs(inet->inet_dport),
 442                                            inet->inet_num,
 443                                            tp->snd_una, tp->snd_nxt);
 444                }
 445#if IS_ENABLED(CONFIG_IPV6)
 446                else if (sk->sk_family == AF_INET6) {
 447                        net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
 448                                            &sk->sk_v6_daddr,
 449                                            ntohs(inet->inet_dport),
 450                                            inet->inet_num,
 451                                            tp->snd_una, tp->snd_nxt);
 452                }
 453#endif
 454                if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) {
 455                        tcp_write_err(sk);
 456                        goto out;
 457                }
 458                tcp_enter_loss(sk);
 459                tcp_retransmit_skb(sk, tcp_write_queue_head(sk), 1);
 460                __sk_dst_reset(sk);
 461                goto out_reset_timer;
 462        }
 463
 464        if (tcp_write_timeout(sk))
 465                goto out;
 466
 467        if (icsk->icsk_retransmits == 0) {
 468                int mib_idx;
 469
 470                if (icsk->icsk_ca_state == TCP_CA_Recovery) {
 471                        if (tcp_is_sack(tp))
 472                                mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
 473                        else
 474                                mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
 475                } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
 476                        mib_idx = LINUX_MIB_TCPLOSSFAILURES;
 477                } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
 478                           tp->sacked_out) {
 479                        if (tcp_is_sack(tp))
 480                                mib_idx = LINUX_MIB_TCPSACKFAILURES;
 481                        else
 482                                mib_idx = LINUX_MIB_TCPRENOFAILURES;
 483                } else {
 484                        mib_idx = LINUX_MIB_TCPTIMEOUTS;
 485                }
 486                __NET_INC_STATS(sock_net(sk), mib_idx);
 487        }
 488
 489        tcp_enter_loss(sk);
 490
 491        if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk), 1) > 0) {
 492                /* Retransmission failed because of local congestion,
 493                 * do not backoff.
 494                 */
 495                if (!icsk->icsk_retransmits)
 496                        icsk->icsk_retransmits = 1;
 497                inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
 498                                          min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
 499                                          TCP_RTO_MAX);
 500                goto out;
 501        }
 502
 503        /* Increase the timeout each time we retransmit.  Note that
 504         * we do not increase the rtt estimate.  rto is initialized
 505         * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
 506         * that doubling rto each time is the least we can get away with.
 507         * In KA9Q, Karn uses this for the first few times, and then
 508         * goes to quadratic.  netBSD doubles, but only goes up to *64,
 509         * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
 510         * defined in the protocol as the maximum possible RTT.  I guess
 511         * we'll have to use something other than TCP to talk to the
 512         * University of Mars.
 513         *
 514         * PAWS allows us longer timeouts and large windows, so once
 515         * implemented ftp to mars will work nicely. We will have to fix
 516         * the 120 second clamps though!
 517         */
 518        icsk->icsk_backoff++;
 519        icsk->icsk_retransmits++;
 520
 521out_reset_timer:
 522        /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
 523         * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
 524         * might be increased if the stream oscillates between thin and thick,
 525         * thus the old value might already be too high compared to the value
 526         * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
 527         * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
 528         * exponential backoff behaviour to avoid continue hammering
 529         * linear-timeout retransmissions into a black hole
 530         */
 531        if (sk->sk_state == TCP_ESTABLISHED &&
 532            (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
 533            tcp_stream_is_thin(tp) &&
 534            icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
 535                icsk->icsk_backoff = 0;
 536                icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
 537        } else {
 538                /* Use normal (exponential) backoff */
 539                icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
 540        }
 541        inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
 542        if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0, 0))
 543                __sk_dst_reset(sk);
 544
 545out:;
 546}
 547
 548/* Called with bottom-half processing disabled.
 549   Called by tcp_write_timer() */
 550void tcp_write_timer_handler(struct sock *sk)
 551{
 552        struct inet_connection_sock *icsk = inet_csk(sk);
 553        int event;
 554
 555        if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending)
 556                goto out;
 557
 558        if (time_after(icsk->icsk_timeout, jiffies)) {
 559                sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
 560                goto out;
 561        }
 562
 563        event = icsk->icsk_pending;
 564
 565        switch (event) {
 566        case ICSK_TIME_EARLY_RETRANS:
 567                tcp_resume_early_retransmit(sk);
 568                break;
 569        case ICSK_TIME_LOSS_PROBE:
 570                tcp_send_loss_probe(sk);
 571                break;
 572        case ICSK_TIME_RETRANS:
 573                icsk->icsk_pending = 0;
 574                tcp_retransmit_timer(sk);
 575                break;
 576        case ICSK_TIME_PROBE0:
 577                icsk->icsk_pending = 0;
 578                tcp_probe_timer(sk);
 579                break;
 580        }
 581
 582out:
 583        sk_mem_reclaim(sk);
 584}
 585
 586static void tcp_write_timer(unsigned long data)
 587{
 588        struct sock *sk = (struct sock *)data;
 589
 590        bh_lock_sock(sk);
 591        if (!sock_owned_by_user(sk)) {
 592                tcp_write_timer_handler(sk);
 593        } else {
 594                /* delegate our work to tcp_release_cb() */
 595                if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
 596                        sock_hold(sk);
 597        }
 598        bh_unlock_sock(sk);
 599        sock_put(sk);
 600}
 601
 602void tcp_syn_ack_timeout(const struct request_sock *req)
 603{
 604        struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
 605
 606        __NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
 607}
 608EXPORT_SYMBOL(tcp_syn_ack_timeout);
 609
 610void tcp_set_keepalive(struct sock *sk, int val)
 611{
 612        if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
 613                return;
 614
 615        if (val && !sock_flag(sk, SOCK_KEEPOPEN))
 616                inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
 617        else if (!val)
 618                inet_csk_delete_keepalive_timer(sk);
 619}
 620
 621
 622static void tcp_keepalive_timer (unsigned long data)
 623{
 624        struct sock *sk = (struct sock *) data;
 625        struct inet_connection_sock *icsk = inet_csk(sk);
 626        struct tcp_sock *tp = tcp_sk(sk);
 627        u32 elapsed;
 628
 629        /* Only process if socket is not in use. */
 630        bh_lock_sock(sk);
 631        if (sock_owned_by_user(sk)) {
 632                /* Try again later. */
 633                inet_csk_reset_keepalive_timer (sk, HZ/20);
 634                goto out;
 635        }
 636
 637        if (sk->sk_state == TCP_LISTEN) {
 638                pr_err("Hmm... keepalive on a LISTEN ???\n");
 639                goto out;
 640        }
 641
 642        if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
 643                if (tp->linger2 >= 0) {
 644                        const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
 645
 646                        if (tmo > 0) {
 647                                tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
 648                                goto out;
 649                        }
 650                }
 651                tcp_send_active_reset(sk, GFP_ATOMIC);
 652                goto death;
 653        }
 654
 655        if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE)
 656                goto out;
 657
 658        elapsed = keepalive_time_when(tp);
 659
 660        /* It is alive without keepalive 8) */
 661        if (tp->packets_out || tcp_send_head(sk))
 662                goto resched;
 663
 664        elapsed = keepalive_time_elapsed(tp);
 665
 666        if (elapsed >= keepalive_time_when(tp)) {
 667                /* If the TCP_USER_TIMEOUT option is enabled, use that
 668                 * to determine when to timeout instead.
 669                 */
 670                if ((icsk->icsk_user_timeout != 0 &&
 671                    elapsed >= icsk->icsk_user_timeout &&
 672                    icsk->icsk_probes_out > 0) ||
 673                    (icsk->icsk_user_timeout == 0 &&
 674                    icsk->icsk_probes_out >= keepalive_probes(tp))) {
 675                        tcp_send_active_reset(sk, GFP_ATOMIC);
 676                        tcp_write_err(sk);
 677                        goto out;
 678                }
 679                if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
 680                        icsk->icsk_probes_out++;
 681                        elapsed = keepalive_intvl_when(tp);
 682                } else {
 683                        /* If keepalive was lost due to local congestion,
 684                         * try harder.
 685                         */
 686                        elapsed = TCP_RESOURCE_PROBE_INTERVAL;
 687                }
 688        } else {
 689                /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
 690                elapsed = keepalive_time_when(tp) - elapsed;
 691        }
 692
 693        sk_mem_reclaim(sk);
 694
 695resched:
 696        inet_csk_reset_keepalive_timer (sk, elapsed);
 697        goto out;
 698
 699death:
 700        tcp_done(sk);
 701
 702out:
 703        bh_unlock_sock(sk);
 704        sock_put(sk);
 705}
 706
 707void tcp_init_xmit_timers(struct sock *sk)
 708{
 709        inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
 710                                  &tcp_keepalive_timer);
 711}
 712
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.