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 struct sk_buff *head = tcp_rtx_queue_head(sk); 161 162 if (!head) 163 return false; 164 start_ts = tcp_skb_timestamp(head); 165 } 166 167 if (likely(timeout == 0)) { 168 linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base); 169 170 if (boundary <= linear_backoff_thresh) 171 timeout = ((2 << boundary) - 1) * rto_base; 172 else 173 timeout = ((2 << linear_backoff_thresh) - 1) * rto_base + 174 (boundary - linear_backoff_thresh) * TCP_RTO_MAX; 175 } 176 return (tcp_time_stamp(tcp_sk(sk)) - start_ts) >= jiffies_to_msecs(timeout); 177 } 178 179 /* A write timeout has occurred. Process the after effects. */ 180 static int tcp_write_timeout(struct sock *sk) 181 { 182 struct inet_connection_sock *icsk = inet_csk(sk); 183 struct tcp_sock *tp = tcp_sk(sk); 184 struct net *net = sock_net(sk); 185 bool expired, do_reset; 186 int retry_until; 187 188 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) { 189 if (icsk->icsk_retransmits) { 190 dst_negative_advice(sk); 191 if (tp->syn_fastopen || tp->syn_data) 192 tcp_fastopen_cache_set(sk, 0, NULL, true, 0); 193 if (tp->syn_data && icsk->icsk_retransmits == 1) 194 NET_INC_STATS(sock_net(sk), 195 LINUX_MIB_TCPFASTOPENACTIVEFAIL); 196 } else if (!tp->syn_data && !tp->syn_fastopen) { 197 sk_rethink_txhash(sk); 198 } 199 retry_until = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries; 200 expired = icsk->icsk_retransmits >= retry_until; 201 } else { 202 if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1, 0)) { 203 /* Some middle-boxes may black-hole Fast Open _after_ 204 * the handshake. Therefore we conservatively disable 205 * Fast Open on this path on recurring timeouts after 206 * successful Fast Open. 207 */ 208 if (tp->syn_data_acked) { 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); 229 230 if (tcp_out_of_resources(sk, do_reset)) 231 return 1; 232 } 233 expired = retransmits_timed_out(sk, retry_until, 234 icsk->icsk_user_timeout); 235 } 236 if (expired) { 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 */ 245 void tcp_delack_timer_handler(struct sock *sk) 246 { 247 struct inet_connection_sock *icsk = inet_csk(sk); 248 249 sk_mem_reclaim_partial(sk); 250 251 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 252 !(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 (inet_csk_ack_scheduled(sk)) { 262 if (!icsk->icsk_ack.pingpong) { 263 /* Delayed ACK missed: inflate ATO. */ 264 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto); 265 } else { 266 /* Delayed ACK missed: leave pingpong mode and 267 * deflate ATO. 268 */ 269 icsk->icsk_ack.pingpong = 0; 270 icsk->icsk_ack.ato = TCP_ATO_MIN; 271 } 272 tcp_send_ack(sk); 273 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS); 274 } 275 276 out: 277 if (tcp_under_memory_pressure(sk)) 278 sk_mem_reclaim(sk); 279 } 280 281 282 /** 283 * tcp_delack_timer() - The TCP delayed ACK timeout handler 284 * @data: Pointer to the current socket. (gets casted to struct sock *) 285 * 286 * This function gets (indirectly) called when the kernel timer for a TCP packet 287 * of this socket expires. Calls tcp_delack_timer_handler() to do the actual work. 288 * 289 * Returns: Nothing (void) 290 */ 291 static void tcp_delack_timer(unsigned long data) 292 { 293 struct sock *sk = (struct sock *)data; 294 295 bh_lock_sock(sk); 296 if (!sock_owned_by_user(sk)) { 297 tcp_delack_timer_handler(sk); 298 } else { 299 inet_csk(sk)->icsk_ack.blocked = 1; 300 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED); 301 /* deleguate our work to tcp_release_cb() */ 302 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags)) 303 sock_hold(sk); 304 } 305 bh_unlock_sock(sk); 306 sock_put(sk); 307 } 308 309 static void tcp_probe_timer(struct sock *sk) 310 { 311 struct inet_connection_sock *icsk = inet_csk(sk); 312 struct sk_buff *skb = tcp_send_head(sk); 313 struct tcp_sock *tp = tcp_sk(sk); 314 int max_probes; 315 u32 start_ts; 316 317 if (tp->packets_out || !skb) { 318 icsk->icsk_probes_out = 0; 319 return; 320 } 321 322 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as 323 * long as the receiver continues to respond probes. We support this by 324 * default and reset icsk_probes_out with incoming ACKs. But if the 325 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we 326 * kill the socket when the retry count and the time exceeds the 327 * corresponding system limit. We also implement similar policy when 328 * we use RTO to probe window in tcp_retransmit_timer(). 329 */ 330 start_ts = tcp_skb_timestamp(skb); 331 if (!start_ts) 332 skb->skb_mstamp = tp->tcp_mstamp; 333 else if (icsk->icsk_user_timeout && 334 (s32)(tcp_time_stamp(tp) - start_ts) > 335 jiffies_to_msecs(icsk->icsk_user_timeout)) 336 goto abort; 337 338 max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2; 339 if (sock_flag(sk, SOCK_DEAD)) { 340 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX; 341 342 max_probes = tcp_orphan_retries(sk, alive); 343 if (!alive && icsk->icsk_backoff >= max_probes) 344 goto abort; 345 if (tcp_out_of_resources(sk, true)) 346 return; 347 } 348 349 if (icsk->icsk_probes_out > max_probes) { 350 abort: tcp_write_err(sk); 351 } else { 352 /* Only send another probe if we didn't close things up. */ 353 tcp_send_probe0(sk); 354 } 355 } 356 357 /* 358 * Timer for Fast Open socket to retransmit SYNACK. Note that the 359 * sk here is the child socket, not the parent (listener) socket. 360 */ 361 static void tcp_fastopen_synack_timer(struct sock *sk) 362 { 363 struct inet_connection_sock *icsk = inet_csk(sk); 364 int max_retries = icsk->icsk_syn_retries ? : 365 sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */ 366 struct request_sock *req; 367 368 req = tcp_sk(sk)->fastopen_rsk; 369 req->rsk_ops->syn_ack_timeout(req); 370 371 if (req->num_timeout >= max_retries) { 372 tcp_write_err(sk); 373 return; 374 } 375 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error 376 * returned from rtx_syn_ack() to make it more persistent like 377 * regular retransmit because if the child socket has been accepted 378 * it's not good to give up too easily. 379 */ 380 inet_rtx_syn_ack(sk, req); 381 req->num_timeout++; 382 icsk->icsk_retransmits++; 383 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 384 TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX); 385 } 386 387 388 /** 389 * tcp_retransmit_timer() - The TCP retransmit timeout handler 390 * @sk: Pointer to the current socket. 391 * 392 * This function gets called when the kernel timer for a TCP packet 393 * of this socket expires. 394 * 395 * It handles retransmission, timer adjustment and other necesarry measures. 396 * 397 * Returns: Nothing (void) 398 */ 399 void tcp_retransmit_timer(struct sock *sk) 400 { 401 struct tcp_sock *tp = tcp_sk(sk); 402 struct net *net = sock_net(sk); 403 struct inet_connection_sock *icsk = inet_csk(sk); 404 405 if (tp->fastopen_rsk) { 406 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 407 sk->sk_state != TCP_FIN_WAIT1); 408 tcp_fastopen_synack_timer(sk); 409 /* Before we receive ACK to our SYN-ACK don't retransmit 410 * anything else (e.g., data or FIN segments). 411 */ 412 return; 413 } 414 if (!tp->packets_out) 415 goto out; 416 417 WARN_ON(tcp_rtx_queue_empty(sk)); 418 419 tp->tlp_high_seq = 0; 420 421 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) && 422 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) { 423 /* Receiver dastardly shrinks window. Our retransmits 424 * become zero probes, but we should not timeout this 425 * connection. If the socket is an orphan, time it out, 426 * we cannot allow such beasts to hang infinitely. 427 */ 428 struct inet_sock *inet = inet_sk(sk); 429 if (sk->sk_family == AF_INET) { 430 net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 431 &inet->inet_daddr, 432 ntohs(inet->inet_dport), 433 inet->inet_num, 434 tp->snd_una, tp->snd_nxt); 435 } 436 #if IS_ENABLED(CONFIG_IPV6) 437 else if (sk->sk_family == AF_INET6) { 438 net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 439 &sk->sk_v6_daddr, 440 ntohs(inet->inet_dport), 441 inet->inet_num, 442 tp->snd_una, tp->snd_nxt); 443 } 444 #endif 445 if (tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX) { 446 tcp_write_err(sk); 447 goto out; 448 } 449 tcp_enter_loss(sk); 450 tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1); 451 __sk_dst_reset(sk); 452 goto out_reset_timer; 453 } 454 455 if (tcp_write_timeout(sk)) 456 goto out; 457 458 if (icsk->icsk_retransmits == 0) { 459 int mib_idx; 460 461 if (icsk->icsk_ca_state == TCP_CA_Recovery) { 462 if (tcp_is_sack(tp)) 463 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL; 464 else 465 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL; 466 } else if (icsk->icsk_ca_state == TCP_CA_Loss) { 467 mib_idx = LINUX_MIB_TCPLOSSFAILURES; 468 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) || 469 tp->sacked_out) { 470 if (tcp_is_sack(tp)) 471 mib_idx = LINUX_MIB_TCPSACKFAILURES; 472 else 473 mib_idx = LINUX_MIB_TCPRENOFAILURES; 474 } else { 475 mib_idx = LINUX_MIB_TCPTIMEOUTS; 476 } 477 __NET_INC_STATS(sock_net(sk), mib_idx); 478 } 479 480 tcp_enter_loss(sk); 481 482 if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) { 483 /* Retransmission failed because of local congestion, 484 * do not backoff. 485 */ 486 if (!icsk->icsk_retransmits) 487 icsk->icsk_retransmits = 1; 488 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 489 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL), 490 TCP_RTO_MAX); 491 goto out; 492 } 493 494 /* Increase the timeout each time we retransmit. Note that 495 * we do not increase the rtt estimate. rto is initialized 496 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests 497 * that doubling rto each time is the least we can get away with. 498 * In KA9Q, Karn uses this for the first few times, and then 499 * goes to quadratic. netBSD doubles, but only goes up to *64, 500 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is 501 * defined in the protocol as the maximum possible RTT. I guess 502 * we'll have to use something other than TCP to talk to the 503 * University of Mars. 504 * 505 * PAWS allows us longer timeouts and large windows, so once 506 * implemented ftp to mars will work nicely. We will have to fix 507 * the 120 second clamps though! 508 */ 509 icsk->icsk_backoff++; 510 icsk->icsk_retransmits++; 511 512 out_reset_timer: 513 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is 514 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this 515 * might be increased if the stream oscillates between thin and thick, 516 * thus the old value might already be too high compared to the value 517 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without 518 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating 519 * exponential backoff behaviour to avoid continue hammering 520 * linear-timeout retransmissions into a black hole 521 */ 522 if (sk->sk_state == TCP_ESTABLISHED && 523 (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) && 524 tcp_stream_is_thin(tp) && 525 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) { 526 icsk->icsk_backoff = 0; 527 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX); 528 } else { 529 /* Use normal (exponential) backoff */ 530 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX); 531 } 532 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX); 533 if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0)) 534 __sk_dst_reset(sk); 535 536 out:; 537 } 538 539 /* Called with bottom-half processing disabled. 540 Called by tcp_write_timer() */ 541 void tcp_write_timer_handler(struct sock *sk) 542 { 543 struct inet_connection_sock *icsk = inet_csk(sk); 544 int event; 545 546 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 547 !icsk->icsk_pending) 548 goto out; 549 550 if (time_after(icsk->icsk_timeout, jiffies)) { 551 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout); 552 goto out; 553 } 554 555 tcp_mstamp_refresh(tcp_sk(sk)); 556 event = icsk->icsk_pending; 557 558 switch (event) { 559 case ICSK_TIME_REO_TIMEOUT: 560 tcp_rack_reo_timeout(sk); 561 break; 562 case ICSK_TIME_LOSS_PROBE: 563 tcp_send_loss_probe(sk); 564 break; 565 case ICSK_TIME_RETRANS: 566 icsk->icsk_pending = 0; 567 tcp_retransmit_timer(sk); 568 break; 569 case ICSK_TIME_PROBE0: 570 icsk->icsk_pending = 0; 571 tcp_probe_timer(sk); 572 break; 573 } 574 575 out: 576 sk_mem_reclaim(sk); 577 } 578 579 static void tcp_write_timer(unsigned long data) 580 { 581 struct sock *sk = (struct sock *)data; 582 583 bh_lock_sock(sk); 584 if (!sock_owned_by_user(sk)) { 585 tcp_write_timer_handler(sk); 586 } else { 587 /* delegate our work to tcp_release_cb() */ 588 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags)) 589 sock_hold(sk); 590 } 591 bh_unlock_sock(sk); 592 sock_put(sk); 593 } 594 595 void tcp_syn_ack_timeout(const struct request_sock *req) 596 { 597 struct net *net = read_pnet(&inet_rsk(req)->ireq_net); 598 599 __NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS); 600 } 601 EXPORT_SYMBOL(tcp_syn_ack_timeout); 602 603 void tcp_set_keepalive(struct sock *sk, int val) 604 { 605 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 606 return; 607 608 if (val && !sock_flag(sk, SOCK_KEEPOPEN)) 609 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk))); 610 else if (!val) 611 inet_csk_delete_keepalive_timer(sk); 612 } 613 EXPORT_SYMBOL_GPL(tcp_set_keepalive); 614 615 616 static void tcp_keepalive_timer (unsigned long data) 617 { 618 struct sock *sk = (struct sock *) data; 619 struct inet_connection_sock *icsk = inet_csk(sk); 620 struct tcp_sock *tp = tcp_sk(sk); 621 u32 elapsed; 622 623 /* Only process if socket is not in use. */ 624 bh_lock_sock(sk); 625 if (sock_owned_by_user(sk)) { 626 /* Try again later. */ 627 inet_csk_reset_keepalive_timer (sk, HZ/20); 628 goto out; 629 } 630 631 if (sk->sk_state == TCP_LISTEN) { 632 pr_err("Hmm... keepalive on a LISTEN ???\n"); 633 goto out; 634 } 635 636 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) { 637 if (tp->linger2 >= 0) { 638 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN; 639 640 if (tmo > 0) { 641 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 642 goto out; 643 } 644 } 645 tcp_send_active_reset(sk, GFP_ATOMIC); 646 goto death; 647 } 648 649 if (!sock_flag(sk, SOCK_KEEPOPEN) || 650 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT))) 651 goto out; 652 653 elapsed = keepalive_time_when(tp); 654 655 /* It is alive without keepalive 8) */ 656 if (tp->packets_out || !tcp_write_queue_empty(sk)) 657 goto resched; 658 659 elapsed = keepalive_time_elapsed(tp); 660 661 if (elapsed >= keepalive_time_when(tp)) { 662 /* If the TCP_USER_TIMEOUT option is enabled, use that 663 * to determine when to timeout instead. 664 */ 665 if ((icsk->icsk_user_timeout != 0 && 666 elapsed >= icsk->icsk_user_timeout && 667 icsk->icsk_probes_out > 0) || 668 (icsk->icsk_user_timeout == 0 && 669 icsk->icsk_probes_out >= keepalive_probes(tp))) { 670 tcp_send_active_reset(sk, GFP_ATOMIC); 671 tcp_write_err(sk); 672 goto out; 673 } 674 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) { 675 icsk->icsk_probes_out++; 676 elapsed = keepalive_intvl_when(tp); 677 } else { 678 /* If keepalive was lost due to local congestion, 679 * try harder. 680 */ 681 elapsed = TCP_RESOURCE_PROBE_INTERVAL; 682 } 683 } else { 684 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */ 685 elapsed = keepalive_time_when(tp) - elapsed; 686 } 687 688 sk_mem_reclaim(sk); 689 690 resched: 691 inet_csk_reset_keepalive_timer (sk, elapsed); 692 goto out; 693 694 death: 695 tcp_done(sk); 696 697 out: 698 bh_unlock_sock(sk); 699 sock_put(sk); 700 } 701 702 void tcp_init_xmit_timers(struct sock *sk) 703 { 704 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer, 705 &tcp_keepalive_timer); 706 hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC, 707 HRTIMER_MODE_ABS_PINNED); 708 tcp_sk(sk)->pacing_timer.function = tcp_pace_kick; 709 } 710