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 /** 26 * tcp_write_err() - close socket and save error info 27 * @sk: The socket the error has appeared on. 28 * 29 * Returns: Nothing (void) 30 */ 31 32 static void tcp_write_err(struct sock *sk) 33 { 34 sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT; 35 sk->sk_error_report(sk); 36 37 tcp_done(sk); 38 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT); 39 } 40 41 /** 42 * tcp_out_of_resources() - Close socket if out of resources 43 * @sk: pointer to current socket 44 * @do_reset: send a last packet with reset flag 45 * 46 * Do not allow orphaned sockets to eat all our resources. 47 * This is direct violation of TCP specs, but it is required 48 * to prevent DoS attacks. It is called when a retransmission timeout 49 * or zero probe timeout occurs on orphaned socket. 50 * 51 * Criteria is still not confirmed experimentally and may change. 52 * We kill the socket, if: 53 * 1. If number of orphaned sockets exceeds an administratively configured 54 * limit. 55 * 2. If we have strong memory pressure. 56 */ 57 static int tcp_out_of_resources(struct sock *sk, bool do_reset) 58 { 59 struct tcp_sock *tp = tcp_sk(sk); 60 int shift = 0; 61 62 /* If peer does not open window for long time, or did not transmit 63 * anything for long time, penalize it. */ 64 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset) 65 shift++; 66 67 /* If some dubious ICMP arrived, penalize even more. */ 68 if (sk->sk_err_soft) 69 shift++; 70 71 if (tcp_check_oom(sk, shift)) { 72 /* Catch exceptional cases, when connection requires reset. 73 * 1. Last segment was sent recently. */ 74 if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN || 75 /* 2. Window is closed. */ 76 (!tp->snd_wnd && !tp->packets_out)) 77 do_reset = true; 78 if (do_reset) 79 tcp_send_active_reset(sk, GFP_ATOMIC); 80 tcp_done(sk); 81 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY); 82 return 1; 83 } 84 return 0; 85 } 86 87 /** 88 * tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket 89 * @sk: Pointer to the current socket. 90 * @alive: bool, socket alive state 91 */ 92 static int tcp_orphan_retries(struct sock *sk, bool alive) 93 { 94 int retries = sock_net(sk)->ipv4.sysctl_tcp_orphan_retries; /* May be zero. */ 95 96 /* We know from an ICMP that something is wrong. */ 97 if (sk->sk_err_soft && !alive) 98 retries = 0; 99 100 /* However, if socket sent something recently, select some safe 101 * number of retries. 8 corresponds to >100 seconds with minimal 102 * RTO of 200msec. */ 103 if (retries == 0 && alive) 104 retries = 8; 105 return retries; 106 } 107 108 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk) 109 { 110 const struct net *net = sock_net(sk); 111 int mss; 112 113 /* Black hole detection */ 114 if (!net->ipv4.sysctl_tcp_mtu_probing) 115 return; 116 117 if (!icsk->icsk_mtup.enabled) { 118 icsk->icsk_mtup.enabled = 1; 119 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 120 } else { 121 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1; 122 mss = min(net->ipv4.sysctl_tcp_base_mss, mss); 123 mss = max(mss, 68 - tcp_sk(sk)->tcp_header_len); 124 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 125 } 126 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 127 } 128 129 130 /** 131 * retransmits_timed_out() - returns true if this connection has timed out 132 * @sk: The current socket 133 * @boundary: max number of retransmissions 134 * @timeout: A custom timeout value. 135 * If set to 0 the default timeout is calculated and used. 136 * Using TCP_RTO_MIN and the number of unsuccessful retransmits. 137 * 138 * The default "timeout" value this function can calculate and use 139 * is equivalent to the timeout of a TCP Connection 140 * after "boundary" unsuccessful, exponentially backed-off 141 * retransmissions with an initial RTO of TCP_RTO_MIN. 142 */ 143 static bool retransmits_timed_out(struct sock *sk, 144 unsigned int boundary, 145 unsigned int timeout) 146 { 147 const unsigned int rto_base = TCP_RTO_MIN; 148 unsigned int linear_backoff_thresh, start_ts; 149 150 if (!inet_csk(sk)->icsk_retransmits) 151 return false; 152 153 start_ts = tcp_sk(sk)->retrans_stamp; 154 if (unlikely(!start_ts)) { 155 struct sk_buff *head = tcp_rtx_queue_head(sk); 156 157 if (!head) 158 return false; 159 start_ts = tcp_skb_timestamp(head); 160 } 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 inet_connection_sock *icsk = inet_csk(sk); 243 244 sk_mem_reclaim_partial(sk); 245 246 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 247 !(icsk->icsk_ack.pending & ICSK_ACK_TIMER)) 248 goto out; 249 250 if (time_after(icsk->icsk_ack.timeout, jiffies)) { 251 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout); 252 goto out; 253 } 254 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER; 255 256 if (inet_csk_ack_scheduled(sk)) { 257 if (!icsk->icsk_ack.pingpong) { 258 /* Delayed ACK missed: inflate ATO. */ 259 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto); 260 } else { 261 /* Delayed ACK missed: leave pingpong mode and 262 * deflate ATO. 263 */ 264 icsk->icsk_ack.pingpong = 0; 265 icsk->icsk_ack.ato = TCP_ATO_MIN; 266 } 267 tcp_send_ack(sk); 268 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS); 269 } 270 271 out: 272 if (tcp_under_memory_pressure(sk)) 273 sk_mem_reclaim(sk); 274 } 275 276 277 /** 278 * tcp_delack_timer() - The TCP delayed ACK timeout handler 279 * @data: Pointer to the current socket. (gets casted to struct sock *) 280 * 281 * This function gets (indirectly) called when the kernel timer for a TCP packet 282 * of this socket expires. Calls tcp_delack_timer_handler() to do the actual work. 283 * 284 * Returns: Nothing (void) 285 */ 286 static void tcp_delack_timer(struct timer_list *t) 287 { 288 struct inet_connection_sock *icsk = 289 from_timer(icsk, t, icsk_delack_timer); 290 struct sock *sk = &icsk->icsk_inet.sk; 291 292 bh_lock_sock(sk); 293 if (!sock_owned_by_user(sk)) { 294 tcp_delack_timer_handler(sk); 295 } else { 296 icsk->icsk_ack.blocked = 1; 297 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED); 298 /* deleguate our work to tcp_release_cb() */ 299 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags)) 300 sock_hold(sk); 301 } 302 bh_unlock_sock(sk); 303 sock_put(sk); 304 } 305 306 static void tcp_probe_timer(struct sock *sk) 307 { 308 struct inet_connection_sock *icsk = inet_csk(sk); 309 struct sk_buff *skb = tcp_send_head(sk); 310 struct tcp_sock *tp = tcp_sk(sk); 311 int max_probes; 312 u32 start_ts; 313 314 if (tp->packets_out || !skb) { 315 icsk->icsk_probes_out = 0; 316 return; 317 } 318 319 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as 320 * long as the receiver continues to respond probes. We support this by 321 * default and reset icsk_probes_out with incoming ACKs. But if the 322 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we 323 * kill the socket when the retry count and the time exceeds the 324 * corresponding system limit. We also implement similar policy when 325 * we use RTO to probe window in tcp_retransmit_timer(). 326 */ 327 start_ts = tcp_skb_timestamp(skb); 328 if (!start_ts) 329 skb->skb_mstamp = tp->tcp_mstamp; 330 else if (icsk->icsk_user_timeout && 331 (s32)(tcp_time_stamp(tp) - start_ts) > 332 jiffies_to_msecs(icsk->icsk_user_timeout)) 333 goto abort; 334 335 max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2; 336 if (sock_flag(sk, SOCK_DEAD)) { 337 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX; 338 339 max_probes = tcp_orphan_retries(sk, alive); 340 if (!alive && icsk->icsk_backoff >= max_probes) 341 goto abort; 342 if (tcp_out_of_resources(sk, true)) 343 return; 344 } 345 346 if (icsk->icsk_probes_out > max_probes) { 347 abort: tcp_write_err(sk); 348 } else { 349 /* Only send another probe if we didn't close things up. */ 350 tcp_send_probe0(sk); 351 } 352 } 353 354 /* 355 * Timer for Fast Open socket to retransmit SYNACK. Note that the 356 * sk here is the child socket, not the parent (listener) socket. 357 */ 358 static void tcp_fastopen_synack_timer(struct sock *sk) 359 { 360 struct inet_connection_sock *icsk = inet_csk(sk); 361 int max_retries = icsk->icsk_syn_retries ? : 362 sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */ 363 struct request_sock *req; 364 365 req = tcp_sk(sk)->fastopen_rsk; 366 req->rsk_ops->syn_ack_timeout(req); 367 368 if (req->num_timeout >= max_retries) { 369 tcp_write_err(sk); 370 return; 371 } 372 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error 373 * returned from rtx_syn_ack() to make it more persistent like 374 * regular retransmit because if the child socket has been accepted 375 * it's not good to give up too easily. 376 */ 377 inet_rtx_syn_ack(sk, req); 378 req->num_timeout++; 379 icsk->icsk_retransmits++; 380 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 381 TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX); 382 } 383 384 385 /** 386 * tcp_retransmit_timer() - The TCP retransmit timeout handler 387 * @sk: Pointer to the current socket. 388 * 389 * This function gets called when the kernel timer for a TCP packet 390 * of this socket expires. 391 * 392 * It handles retransmission, timer adjustment and other necesarry measures. 393 * 394 * Returns: Nothing (void) 395 */ 396 void tcp_retransmit_timer(struct sock *sk) 397 { 398 struct tcp_sock *tp = tcp_sk(sk); 399 struct net *net = sock_net(sk); 400 struct inet_connection_sock *icsk = inet_csk(sk); 401 402 if (tp->fastopen_rsk) { 403 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 404 sk->sk_state != TCP_FIN_WAIT1); 405 tcp_fastopen_synack_timer(sk); 406 /* Before we receive ACK to our SYN-ACK don't retransmit 407 * anything else (e.g., data or FIN segments). 408 */ 409 return; 410 } 411 if (!tp->packets_out) 412 goto out; 413 414 WARN_ON(tcp_rtx_queue_empty(sk)); 415 416 tp->tlp_high_seq = 0; 417 418 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) && 419 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) { 420 /* Receiver dastardly shrinks window. Our retransmits 421 * become zero probes, but we should not timeout this 422 * connection. If the socket is an orphan, time it out, 423 * we cannot allow such beasts to hang infinitely. 424 */ 425 struct inet_sock *inet = inet_sk(sk); 426 if (sk->sk_family == AF_INET) { 427 net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 428 &inet->inet_daddr, 429 ntohs(inet->inet_dport), 430 inet->inet_num, 431 tp->snd_una, tp->snd_nxt); 432 } 433 #if IS_ENABLED(CONFIG_IPV6) 434 else if (sk->sk_family == AF_INET6) { 435 net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 436 &sk->sk_v6_daddr, 437 ntohs(inet->inet_dport), 438 inet->inet_num, 439 tp->snd_una, tp->snd_nxt); 440 } 441 #endif 442 if (tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX) { 443 tcp_write_err(sk); 444 goto out; 445 } 446 tcp_enter_loss(sk); 447 tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1); 448 __sk_dst_reset(sk); 449 goto out_reset_timer; 450 } 451 452 if (tcp_write_timeout(sk)) 453 goto out; 454 455 if (icsk->icsk_retransmits == 0) { 456 int mib_idx; 457 458 if (icsk->icsk_ca_state == TCP_CA_Recovery) { 459 if (tcp_is_sack(tp)) 460 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL; 461 else 462 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL; 463 } else if (icsk->icsk_ca_state == TCP_CA_Loss) { 464 mib_idx = LINUX_MIB_TCPLOSSFAILURES; 465 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) || 466 tp->sacked_out) { 467 if (tcp_is_sack(tp)) 468 mib_idx = LINUX_MIB_TCPSACKFAILURES; 469 else 470 mib_idx = LINUX_MIB_TCPRENOFAILURES; 471 } else { 472 mib_idx = LINUX_MIB_TCPTIMEOUTS; 473 } 474 __NET_INC_STATS(sock_net(sk), mib_idx); 475 } 476 477 tcp_enter_loss(sk); 478 479 if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) { 480 /* Retransmission failed because of local congestion, 481 * do not backoff. 482 */ 483 if (!icsk->icsk_retransmits) 484 icsk->icsk_retransmits = 1; 485 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 486 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL), 487 TCP_RTO_MAX); 488 goto out; 489 } 490 491 /* Increase the timeout each time we retransmit. Note that 492 * we do not increase the rtt estimate. rto is initialized 493 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests 494 * that doubling rto each time is the least we can get away with. 495 * In KA9Q, Karn uses this for the first few times, and then 496 * goes to quadratic. netBSD doubles, but only goes up to *64, 497 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is 498 * defined in the protocol as the maximum possible RTT. I guess 499 * we'll have to use something other than TCP to talk to the 500 * University of Mars. 501 * 502 * PAWS allows us longer timeouts and large windows, so once 503 * implemented ftp to mars will work nicely. We will have to fix 504 * the 120 second clamps though! 505 */ 506 icsk->icsk_backoff++; 507 icsk->icsk_retransmits++; 508 509 out_reset_timer: 510 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is 511 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this 512 * might be increased if the stream oscillates between thin and thick, 513 * thus the old value might already be too high compared to the value 514 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without 515 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating 516 * exponential backoff behaviour to avoid continue hammering 517 * linear-timeout retransmissions into a black hole 518 */ 519 if (sk->sk_state == TCP_ESTABLISHED && 520 (tp->thin_lto || net->ipv4.sysctl_tcp_thin_linear_timeouts) && 521 tcp_stream_is_thin(tp) && 522 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) { 523 icsk->icsk_backoff = 0; 524 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX); 525 } else { 526 /* Use normal (exponential) backoff */ 527 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX); 528 } 529 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX); 530 if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0)) 531 __sk_dst_reset(sk); 532 533 out:; 534 } 535 536 /* Called with bottom-half processing disabled. 537 Called by tcp_write_timer() */ 538 void tcp_write_timer_handler(struct sock *sk) 539 { 540 struct inet_connection_sock *icsk = inet_csk(sk); 541 int event; 542 543 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 544 !icsk->icsk_pending) 545 goto out; 546 547 if (time_after(icsk->icsk_timeout, jiffies)) { 548 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout); 549 goto out; 550 } 551 552 tcp_mstamp_refresh(tcp_sk(sk)); 553 event = icsk->icsk_pending; 554 555 switch (event) { 556 case ICSK_TIME_REO_TIMEOUT: 557 tcp_rack_reo_timeout(sk); 558 break; 559 case ICSK_TIME_LOSS_PROBE: 560 tcp_send_loss_probe(sk); 561 break; 562 case ICSK_TIME_RETRANS: 563 icsk->icsk_pending = 0; 564 tcp_retransmit_timer(sk); 565 break; 566 case ICSK_TIME_PROBE0: 567 icsk->icsk_pending = 0; 568 tcp_probe_timer(sk); 569 break; 570 } 571 572 out: 573 sk_mem_reclaim(sk); 574 } 575 576 static void tcp_write_timer(struct timer_list *t) 577 { 578 struct inet_connection_sock *icsk = 579 from_timer(icsk, t, icsk_retransmit_timer); 580 struct sock *sk = &icsk->icsk_inet.sk; 581 582 bh_lock_sock(sk); 583 if (!sock_owned_by_user(sk)) { 584 tcp_write_timer_handler(sk); 585 } else { 586 /* delegate our work to tcp_release_cb() */ 587 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags)) 588 sock_hold(sk); 589 } 590 bh_unlock_sock(sk); 591 sock_put(sk); 592 } 593 594 void tcp_syn_ack_timeout(const struct request_sock *req) 595 { 596 struct net *net = read_pnet(&inet_rsk(req)->ireq_net); 597 598 __NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS); 599 } 600 EXPORT_SYMBOL(tcp_syn_ack_timeout); 601 602 void tcp_set_keepalive(struct sock *sk, int val) 603 { 604 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 605 return; 606 607 if (val && !sock_flag(sk, SOCK_KEEPOPEN)) 608 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk))); 609 else if (!val) 610 inet_csk_delete_keepalive_timer(sk); 611 } 612 EXPORT_SYMBOL_GPL(tcp_set_keepalive); 613 614 615 static void tcp_keepalive_timer (struct timer_list *t) 616 { 617 struct sock *sk = from_timer(sk, t, sk_timer); 618 struct inet_connection_sock *icsk = inet_csk(sk); 619 struct tcp_sock *tp = tcp_sk(sk); 620 u32 elapsed; 621 622 /* Only process if socket is not in use. */ 623 bh_lock_sock(sk); 624 if (sock_owned_by_user(sk)) { 625 /* Try again later. */ 626 inet_csk_reset_keepalive_timer (sk, HZ/20); 627 goto out; 628 } 629 630 if (sk->sk_state == TCP_LISTEN) { 631 pr_err("Hmm... keepalive on a LISTEN ???\n"); 632 goto out; 633 } 634 635 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) { 636 if (tp->linger2 >= 0) { 637 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN; 638 639 if (tmo > 0) { 640 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 641 goto out; 642 } 643 } 644 tcp_send_active_reset(sk, GFP_ATOMIC); 645 goto death; 646 } 647 648 if (!sock_flag(sk, SOCK_KEEPOPEN) || 649 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT))) 650 goto out; 651 652 elapsed = keepalive_time_when(tp); 653 654 /* It is alive without keepalive 8) */ 655 if (tp->packets_out || !tcp_write_queue_empty(sk)) 656 goto resched; 657 658 elapsed = keepalive_time_elapsed(tp); 659 660 if (elapsed >= keepalive_time_when(tp)) { 661 /* If the TCP_USER_TIMEOUT option is enabled, use that 662 * to determine when to timeout instead. 663 */ 664 if ((icsk->icsk_user_timeout != 0 && 665 elapsed >= icsk->icsk_user_timeout && 666 icsk->icsk_probes_out > 0) || 667 (icsk->icsk_user_timeout == 0 && 668 icsk->icsk_probes_out >= keepalive_probes(tp))) { 669 tcp_send_active_reset(sk, GFP_ATOMIC); 670 tcp_write_err(sk); 671 goto out; 672 } 673 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) { 674 icsk->icsk_probes_out++; 675 elapsed = keepalive_intvl_when(tp); 676 } else { 677 /* If keepalive was lost due to local congestion, 678 * try harder. 679 */ 680 elapsed = TCP_RESOURCE_PROBE_INTERVAL; 681 } 682 } else { 683 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */ 684 elapsed = keepalive_time_when(tp) - elapsed; 685 } 686 687 sk_mem_reclaim(sk); 688 689 resched: 690 inet_csk_reset_keepalive_timer (sk, elapsed); 691 goto out; 692 693 death: 694 tcp_done(sk); 695 696 out: 697 bh_unlock_sock(sk); 698 sock_put(sk); 699 } 700 701 void tcp_init_xmit_timers(struct sock *sk) 702 { 703 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer, 704 &tcp_keepalive_timer); 705 hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC, 706 HRTIMER_MODE_ABS_PINNED); 707 tcp_sk(sk)->pacing_timer.function = tcp_pace_kick; 708 } 709