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_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
  34 static 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  */
  59 static 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_jiffies32 - 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_jiffies32 - 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  */
  94 static 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
 110 static 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_jiffies32;
 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  *
 143  * The default "timeout" value this function can calculate and use
 144  * is equivalent to the timeout of a TCP Connection
 145  * after "boundary" unsuccessful, exponentially backed-off
 146  * retransmissions with an initial RTO of TCP_RTO_MIN.
 147  */
 148 static bool retransmits_timed_out(struct sock *sk,
 149                                   unsigned int boundary,
 150                                   unsigned int timeout)
 151 {
 152         const unsigned int rto_base = TCP_RTO_MIN;
 153         unsigned int linear_backoff_thresh, start_ts;
 154
 155         if (!inet_csk(sk)->icsk_retransmits)
 156                 return false;
 157
 158         start_ts = tcp_sk(sk)->retrans_stamp;
 159         if (unlikely(!start_ts))
 160                 start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
 161
 162         if (likely(timeout == 0)) {
 163                 linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
 164
 165                 if (boundary <= linear_backoff_thresh)
 166                         timeout = ((2 << boundary) - 1) * rto_base;
 167                 else
 168                         timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
 169                                 (boundary - linear_backoff_thresh) * TCP_RTO_MAX;
 170         }
 171         return (tcp_time_stamp(tcp_sk(sk)) - start_ts) >= jiffies_to_msecs(timeout);
 172 }
 173
 174 /* A write timeout has occurred. Process the after effects. */
 175 static int tcp_write_timeout(struct sock *sk)
 176 {
 177         struct inet_connection_sock *icsk = inet_csk(sk);
 178         struct tcp_sock *tp = tcp_sk(sk);
 179         struct net *net = sock_net(sk);
 180         bool expired, do_reset;
 181         int retry_until;
 182
 183         if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 184                 if (icsk->icsk_retransmits) {
 185                         dst_negative_advice(sk);
 186                         if (tp->syn_fastopen || tp->syn_data)
 187                                 tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
 188                         if (tp->syn_data && icsk->icsk_retransmits == 1)
 189                                 NET_INC_STATS(sock_net(sk),
 190                                               LINUX_MIB_TCPFASTOPENACTIVEFAIL);
 191                 } else if (!tp->syn_data && !tp->syn_fastopen) {
 192                         sk_rethink_txhash(sk);
 193                 }
 194                 retry_until = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
 195                 expired = icsk->icsk_retransmits >= retry_until;
 196         } else {
 197                 if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1, 0)) {
 198                         /* Some middle-boxes may black-hole Fast Open _after_
 199                          * the handshake. Therefore we conservatively disable
 200                          * Fast Open on this path on recurring timeouts after
 201                          * successful Fast Open.
 202                          */
 203                         if (tp->syn_data_acked) {
 204                                 tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
 205                                 if (icsk->icsk_retransmits == net->ipv4.sysctl_tcp_retries1)
 206                                         NET_INC_STATS(sock_net(sk),
 207                                                       LINUX_MIB_TCPFASTOPENACTIVEFAIL);
 208                         }
 209                         /* Black hole detection */
 210                         tcp_mtu_probing(icsk, sk);
 211
 212                         dst_negative_advice(sk);
 213                 } else {
 214                         sk_rethink_txhash(sk);
 215                 }
 216
 217                 retry_until = net->ipv4.sysctl_tcp_retries2;
 218                 if (sock_flag(sk, SOCK_DEAD)) {
 219                         const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
 220
 221                         retry_until = tcp_orphan_retries(sk, alive);
 222                         do_reset = alive ||
 223                                 !retransmits_timed_out(sk, retry_until, 0);
 224
 225                         if (tcp_out_of_resources(sk, do_reset))
 226                                 return 1;
 227                 }
 228                 expired = retransmits_timed_out(sk, retry_until,
 229                                                 icsk->icsk_user_timeout);
 230         }
 231         if (expired) {
 232                 /* Has it gone just too far? */
 233                 tcp_write_err(sk);
 234                 return 1;
 235         }
 236         return 0;
 237 }
 238
 239 /* Called with BH disabled */
 240 void tcp_delack_timer_handler(struct sock *sk)
 241 {
 242         struct tcp_sock *tp = tcp_sk(sk);
 243         struct inet_connection_sock *icsk = inet_csk(sk);
 244
 245         sk_mem_reclaim_partial(sk);
 246
 247         if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
 248             !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
 249                 goto out;
 250
 251         if (time_after(icsk->icsk_ack.timeout, jiffies)) {
 252                 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
 253                 goto out;
 254         }
 255         icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
 256
 257         if (!skb_queue_empty(&tp->ucopy.prequeue)) {
 258                 struct sk_buff *skb;
 259
 260                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED);
 261
 262                 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
 263                         sk_backlog_rcv(sk, skb);
 264
 265                 tp->ucopy.memory = 0;
 266         }
 267
 268         if (inet_csk_ack_scheduled(sk)) {
 269                 if (!icsk->icsk_ack.pingpong) {
 270                         /* Delayed ACK missed: inflate ATO. */
 271                         icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
 272                 } else {
 273                         /* Delayed ACK missed: leave pingpong mode and
 274                          * deflate ATO.
 275                          */
 276                         icsk->icsk_ack.pingpong = 0;
 277                         icsk->icsk_ack.ato      = TCP_ATO_MIN;
 278                 }
 279                 tcp_send_ack(sk);
 280                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
 281         }
 282
 283 out:
 284         if (tcp_under_memory_pressure(sk))
 285                 sk_mem_reclaim(sk);
 286 }
 287
 288
 289 /**
 290  *  tcp_delack_timer() - The TCP delayed ACK timeout handler
 291  *  @data:  Pointer to the current socket. (gets casted to struct sock *)
 292  *
 293  *  This function gets (indirectly) called when the kernel timer for a TCP packet
 294  *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
 295  *
 296  *  Returns: Nothing (void)
 297  */
 298 static void tcp_delack_timer(unsigned long data)
 299 {
 300         struct sock *sk = (struct sock *)data;
 301
 302         bh_lock_sock(sk);
 303         if (!sock_owned_by_user(sk)) {
 304                 tcp_delack_timer_handler(sk);
 305         } else {
 306                 inet_csk(sk)->icsk_ack.blocked = 1;
 307                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
 308                 /* deleguate our work to tcp_release_cb() */
 309                 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
 310                         sock_hold(sk);
 311         }
 312         bh_unlock_sock(sk);
 313         sock_put(sk);
 314 }
 315
 316 static void tcp_probe_timer(struct sock *sk)
 317 {
 318         struct inet_connection_sock *icsk = inet_csk(sk);
 319         struct tcp_sock *tp = tcp_sk(sk);
 320         int max_probes;
 321         u32 start_ts;
 322
 323         if (tp->packets_out || !tcp_send_head(sk)) {
 324                 icsk->icsk_probes_out = 0;
 325                 return;
 326         }
 327
 328         /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
 329          * long as the receiver continues to respond probes. We support this by
 330          * default and reset icsk_probes_out with incoming ACKs. But if the
 331          * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
 332          * kill the socket when the retry count and the time exceeds the
 333          * corresponding system limit. We also implement similar policy when
 334          * we use RTO to probe window in tcp_retransmit_timer().
 335          */
 336         start_ts = tcp_skb_timestamp(tcp_send_head(sk));
 337         if (!start_ts)
 338                 tcp_send_head(sk)->skb_mstamp = tp->tcp_mstamp;
 339         else if (icsk->icsk_user_timeout &&
 340                  (s32)(tcp_time_stamp(tp) - start_ts) >
 341                  jiffies_to_msecs(icsk->icsk_user_timeout))
 342                 goto abort;
 343
 344         max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2;
 345         if (sock_flag(sk, SOCK_DEAD)) {
 346                 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
 347
 348                 max_probes = tcp_orphan_retries(sk, alive);
 349                 if (!alive && icsk->icsk_backoff >= max_probes)
 350                         goto abort;
 351                 if (tcp_out_of_resources(sk, true))
 352                         return;
 353         }
 354
 355         if (icsk->icsk_probes_out > max_probes) {
 356 abort:          tcp_write_err(sk);
 357         } else {
 358                 /* Only send another probe if we didn't close things up. */
 359                 tcp_send_probe0(sk);
 360         }
 361 }
 362
 363 /*
 364  *      Timer for Fast Open socket to retransmit SYNACK. Note that the
 365  *      sk here is the child socket, not the parent (listener) socket.
 366  */
 367 static void tcp_fastopen_synack_timer(struct sock *sk)
 368 {
 369         struct inet_connection_sock *icsk = inet_csk(sk);
 370         int max_retries = icsk->icsk_syn_retries ? :
 371             sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
 372         struct request_sock *req;
 373
 374         req = tcp_sk(sk)->fastopen_rsk;
 375         req->rsk_ops->syn_ack_timeout(req);
 376
 377         if (req->num_timeout >= max_retries) {
 378                 tcp_write_err(sk);
 379                 return;
 380         }
 381         /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
 382          * returned from rtx_syn_ack() to make it more persistent like
 383          * regular retransmit because if the child socket has been accepted
 384          * it's not good to give up too easily.
 385          */
 386         inet_rtx_syn_ack(sk, req);
 387         req->num_timeout++;
 388         icsk->icsk_retransmits++;
 389         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
 390                           TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
 391 }
 392
 393
 394 /**
 395  *  tcp_retransmit_timer() - The TCP retransmit timeout handler
 396  *  @sk:  Pointer to the current socket.
 397  *
 398  *  This function gets called when the kernel timer for a TCP packet
 399  *  of this socket expires.
 400  *
 401  *  It handles retransmission, timer adjustment and other necesarry measures.
 402  *
 403  *  Returns: Nothing (void)
 404  */
 405 void tcp_retransmit_timer(struct sock *sk)
 406 {
 407         struct tcp_sock *tp = tcp_sk(sk);
 408         struct net *net = sock_net(sk);
 409         struct inet_connection_sock *icsk = inet_csk(sk);
 410
 411         if (tp->fastopen_rsk) {
 412                 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
 413                              sk->sk_state != TCP_FIN_WAIT1);
 414                 tcp_fastopen_synack_timer(sk);
 415                 /* Before we receive ACK to our SYN-ACK don't retransmit
 416                  * anything else (e.g., data or FIN segments).
 417                  */
 418                 return;
 419         }
 420         if (!tp->packets_out)
 421                 goto out;
 422
 423         WARN_ON(tcp_write_queue_empty(sk));
 424
 425         tp->tlp_high_seq = 0;
 426
 427         if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
 428             !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
 429                 /* Receiver dastardly shrinks window. Our retransmits
 430                  * become zero probes, but we should not timeout this
 431                  * connection. If the socket is an orphan, time it out,
 432                  * we cannot allow such beasts to hang infinitely.
 433                  */
 434                 struct inet_sock *inet = inet_sk(sk);
 435                 if (sk->sk_family == AF_INET) {
 436                         net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
 437                                             &inet->inet_daddr,
 438                                             ntohs(inet->inet_dport),
 439                                             inet->inet_num,
 440                                             tp->snd_una, tp->snd_nxt);
 441                 }
 442 #if IS_ENABLED(CONFIG_IPV6)
 443                 else if (sk->sk_family == AF_INET6) {
 444                         net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
 445                                             &sk->sk_v6_daddr,
 446                                             ntohs(inet->inet_dport),
 447                                             inet->inet_num,
 448                                             tp->snd_una, tp->snd_nxt);
 449                 }
 450 #endif
 451                 if (tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX) {
 452                         tcp_write_err(sk);
 453                         goto out;
 454                 }
 455                 tcp_enter_loss(sk);
 456                 tcp_retransmit_skb(sk, tcp_write_queue_head(sk), 1);
 457                 __sk_dst_reset(sk);
 458                 goto out_reset_timer;
 459         }
 460
 461         if (tcp_write_timeout(sk))
 462                 goto out;
 463
 464         if (icsk->icsk_retransmits == 0) {
 465                 int mib_idx;
 466
 467                 if (icsk->icsk_ca_state == TCP_CA_Recovery) {
 468                         if (tcp_is_sack(tp))
 469                                 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
 470                         else
 471                                 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
 472                 } else if (icsk->icsk_ca_state == TCP_CA_Loss) {
 473                         mib_idx = LINUX_MIB_TCPLOSSFAILURES;
 474                 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
 475                            tp->sacked_out) {
 476                         if (tcp_is_sack(tp))
 477                                 mib_idx = LINUX_MIB_TCPSACKFAILURES;
 478                         else
 479                                 mib_idx = LINUX_MIB_TCPRENOFAILURES;
 480                 } else {
 481                         mib_idx = LINUX_MIB_TCPTIMEOUTS;
 482                 }
 483                 __NET_INC_STATS(sock_net(sk), mib_idx);
 484         }
 485
 486         tcp_enter_loss(sk);
 487
 488         if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk), 1) > 0) {
 489                 /* Retransmission failed because of local congestion,
 490                  * do not backoff.
 491                  */
 492                 if (!icsk->icsk_retransmits)
 493                         icsk->icsk_retransmits = 1;
 494                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
 495                                           min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
 496                                           TCP_RTO_MAX);
 497                 goto out;
 498         }
 499
 500         /* Increase the timeout each time we retransmit.  Note that
 501          * we do not increase the rtt estimate.  rto is initialized
 502          * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
 503          * that doubling rto each time is the least we can get away with.
 504          * In KA9Q, Karn uses this for the first few times, and then
 505          * goes to quadratic.  netBSD doubles, but only goes up to *64,
 506          * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
 507          * defined in the protocol as the maximum possible RTT.  I guess
 508          * we'll have to use something other than TCP to talk to the
 509          * University of Mars.
 510          *
 511          * PAWS allows us longer timeouts and large windows, so once
 512          * implemented ftp to mars will work nicely. We will have to fix
 513          * the 120 second clamps though!
 514          */
 515         icsk->icsk_backoff++;
 516         icsk->icsk_retransmits++;
 517
 518 out_reset_timer:
 519         /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
 520          * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
 521          * might be increased if the stream oscillates between thin and thick,
 522          * thus the old value might already be too high compared to the value
 523          * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
 524          * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
 525          * exponential backoff behaviour to avoid continue hammering
 526          * linear-timeout retransmissions into a black hole
 527          */
 528         if (sk->sk_state == TCP_ESTABLISHED &&
 529             (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
 530             tcp_stream_is_thin(tp) &&
 531             icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
 532                 icsk->icsk_backoff = 0;
 533                 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
 534         } else {
 535                 /* Use normal (exponential) backoff */
 536                 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
 537         }
 538         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
 539         if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0))
 540                 __sk_dst_reset(sk);
 541
 542 out:;
 543 }
 544
 545 /* Called with bottom-half processing disabled.
 546    Called by tcp_write_timer() */
 547 void tcp_write_timer_handler(struct sock *sk)
 548 {
 549         struct inet_connection_sock *icsk = inet_csk(sk);
 550         int event;
 551
 552         if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
 553             !icsk->icsk_pending)
 554                 goto out;
 555
 556         if (time_after(icsk->icsk_timeout, jiffies)) {
 557                 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
 558                 goto out;
 559         }
 560
 561         tcp_mstamp_refresh(tcp_sk(sk));
 562         event = icsk->icsk_pending;
 563
 564         switch (event) {
 565         case ICSK_TIME_REO_TIMEOUT:
 566                 tcp_rack_reo_timeout(sk);
 567                 break;
 568         case ICSK_TIME_LOSS_PROBE:
 569                 tcp_send_loss_probe(sk);
 570                 break;
 571         case ICSK_TIME_RETRANS:
 572                 icsk->icsk_pending = 0;
 573                 tcp_retransmit_timer(sk);
 574                 break;
 575         case ICSK_TIME_PROBE0:
 576                 icsk->icsk_pending = 0;
 577                 tcp_probe_timer(sk);
 578                 break;
 579         }
 580
 581 out:
 582         sk_mem_reclaim(sk);
 583 }
 584
 585 static void tcp_write_timer(unsigned long data)
 586 {
 587         struct sock *sk = (struct sock *)data;
 588
 589         bh_lock_sock(sk);
 590         if (!sock_owned_by_user(sk)) {
 591                 tcp_write_timer_handler(sk);
 592         } else {
 593                 /* delegate our work to tcp_release_cb() */
 594                 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
 595                         sock_hold(sk);
 596         }
 597         bh_unlock_sock(sk);
 598         sock_put(sk);
 599 }
 600
 601 void tcp_syn_ack_timeout(const struct request_sock *req)
 602 {
 603         struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
 604
 605         __NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
 606 }
 607 EXPORT_SYMBOL(tcp_syn_ack_timeout);
 608
 609 void tcp_set_keepalive(struct sock *sk, int val)
 610 {
 611         if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
 612                 return;
 613
 614         if (val && !sock_flag(sk, SOCK_KEEPOPEN))
 615                 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
 616         else if (!val)
 617                 inet_csk_delete_keepalive_timer(sk);
 618 }
 619 EXPORT_SYMBOL_GPL(tcp_set_keepalive);
 620
 621
 622 static 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) ||
 656             ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
 657                 goto out;
 658
 659         elapsed = keepalive_time_when(tp);
 660
 661         /* It is alive without keepalive 8) */
 662         if (tp->packets_out || tcp_send_head(sk))
 663                 goto resched;
 664
 665         elapsed = keepalive_time_elapsed(tp);
 666
 667         if (elapsed >= keepalive_time_when(tp)) {
 668                 /* If the TCP_USER_TIMEOUT option is enabled, use that
 669                  * to determine when to timeout instead.
 670                  */
 671                 if ((icsk->icsk_user_timeout != 0 &&
 672                     elapsed >= icsk->icsk_user_timeout &&
 673                     icsk->icsk_probes_out > 0) ||
 674                     (icsk->icsk_user_timeout == 0 &&
 675                     icsk->icsk_probes_out >= keepalive_probes(tp))) {
 676                         tcp_send_active_reset(sk, GFP_ATOMIC);
 677                         tcp_write_err(sk);
 678                         goto out;
 679                 }
 680                 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
 681                         icsk->icsk_probes_out++;
 682                         elapsed = keepalive_intvl_when(tp);
 683                 } else {
 684                         /* If keepalive was lost due to local congestion,
 685                          * try harder.
 686                          */
 687                         elapsed = TCP_RESOURCE_PROBE_INTERVAL;
 688                 }
 689         } else {
 690                 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */
 691                 elapsed = keepalive_time_when(tp) - elapsed;
 692         }
 693
 694         sk_mem_reclaim(sk);
 695
 696 resched:
 697         inet_csk_reset_keepalive_timer (sk, elapsed);
 698         goto out;
 699
 700 death:
 701         tcp_done(sk);
 702
 703 out:
 704         bh_unlock_sock(sk);
 705         sock_put(sk);
 706 }
 707
 708 void tcp_init_xmit_timers(struct sock *sk)
 709 {
 710         inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
 711                                   &tcp_keepalive_timer);
 712         hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
 713                      HRTIMER_MODE_ABS_PINNED);
 714         tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;
 715 }