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 } else if (!tp->syn_data && !tp->syn_fastopen) { 187 sk_rethink_txhash(sk); 188 } 189 retry_until = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries; 190 expired = icsk->icsk_retransmits >= retry_until; 191 } else { 192 if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1, 0)) { 193 /* Black hole detection */ 194 tcp_mtu_probing(icsk, sk); 195 196 dst_negative_advice(sk); 197 } else { 198 sk_rethink_txhash(sk); 199 } 200 201 retry_until = net->ipv4.sysctl_tcp_retries2; 202 if (sock_flag(sk, SOCK_DEAD)) { 203 const bool alive = icsk->icsk_rto < TCP_RTO_MAX; 204 205 retry_until = tcp_orphan_retries(sk, alive); 206 do_reset = alive || 207 !retransmits_timed_out(sk, retry_until, 0); 208 209 if (tcp_out_of_resources(sk, do_reset)) 210 return 1; 211 } 212 expired = retransmits_timed_out(sk, retry_until, 213 icsk->icsk_user_timeout); 214 } 215 tcp_fastopen_active_detect_blackhole(sk, expired); 216 if (expired) { 217 /* Has it gone just too far? */ 218 tcp_write_err(sk); 219 return 1; 220 } 221 return 0; 222 } 223 224 /* Called with BH disabled */ 225 void tcp_delack_timer_handler(struct sock *sk) 226 { 227 struct inet_connection_sock *icsk = inet_csk(sk); 228 229 sk_mem_reclaim_partial(sk); 230 231 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 232 !(icsk->icsk_ack.pending & ICSK_ACK_TIMER)) 233 goto out; 234 235 if (time_after(icsk->icsk_ack.timeout, jiffies)) { 236 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout); 237 goto out; 238 } 239 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER; 240 241 if (inet_csk_ack_scheduled(sk)) { 242 if (!icsk->icsk_ack.pingpong) { 243 /* Delayed ACK missed: inflate ATO. */ 244 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto); 245 } else { 246 /* Delayed ACK missed: leave pingpong mode and 247 * deflate ATO. 248 */ 249 icsk->icsk_ack.pingpong = 0; 250 icsk->icsk_ack.ato = TCP_ATO_MIN; 251 } 252 tcp_mstamp_refresh(tcp_sk(sk)); 253 tcp_send_ack(sk); 254 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS); 255 } 256 257 out: 258 if (tcp_under_memory_pressure(sk)) 259 sk_mem_reclaim(sk); 260 } 261 262 263 /** 264 * tcp_delack_timer() - The TCP delayed ACK timeout handler 265 * @data: Pointer to the current socket. (gets casted to struct sock *) 266 * 267 * This function gets (indirectly) called when the kernel timer for a TCP packet 268 * of this socket expires. Calls tcp_delack_timer_handler() to do the actual work. 269 * 270 * Returns: Nothing (void) 271 */ 272 static void tcp_delack_timer(struct timer_list *t) 273 { 274 struct inet_connection_sock *icsk = 275 from_timer(icsk, t, icsk_delack_timer); 276 struct sock *sk = &icsk->icsk_inet.sk; 277 278 bh_lock_sock(sk); 279 if (!sock_owned_by_user(sk)) { 280 tcp_delack_timer_handler(sk); 281 } else { 282 icsk->icsk_ack.blocked = 1; 283 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED); 284 /* deleguate our work to tcp_release_cb() */ 285 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags)) 286 sock_hold(sk); 287 } 288 bh_unlock_sock(sk); 289 sock_put(sk); 290 } 291 292 static void tcp_probe_timer(struct sock *sk) 293 { 294 struct inet_connection_sock *icsk = inet_csk(sk); 295 struct sk_buff *skb = tcp_send_head(sk); 296 struct tcp_sock *tp = tcp_sk(sk); 297 int max_probes; 298 u32 start_ts; 299 300 if (tp->packets_out || !skb) { 301 icsk->icsk_probes_out = 0; 302 return; 303 } 304 305 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as 306 * long as the receiver continues to respond probes. We support this by 307 * default and reset icsk_probes_out with incoming ACKs. But if the 308 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we 309 * kill the socket when the retry count and the time exceeds the 310 * corresponding system limit. We also implement similar policy when 311 * we use RTO to probe window in tcp_retransmit_timer(). 312 */ 313 start_ts = tcp_skb_timestamp(skb); 314 if (!start_ts) 315 skb->skb_mstamp = tp->tcp_mstamp; 316 else if (icsk->icsk_user_timeout && 317 (s32)(tcp_time_stamp(tp) - start_ts) > 318 jiffies_to_msecs(icsk->icsk_user_timeout)) 319 goto abort; 320 321 max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2; 322 if (sock_flag(sk, SOCK_DEAD)) { 323 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX; 324 325 max_probes = tcp_orphan_retries(sk, alive); 326 if (!alive && icsk->icsk_backoff >= max_probes) 327 goto abort; 328 if (tcp_out_of_resources(sk, true)) 329 return; 330 } 331 332 if (icsk->icsk_probes_out > max_probes) { 333 abort: tcp_write_err(sk); 334 } else { 335 /* Only send another probe if we didn't close things up. */ 336 tcp_send_probe0(sk); 337 } 338 } 339 340 /* 341 * Timer for Fast Open socket to retransmit SYNACK. Note that the 342 * sk here is the child socket, not the parent (listener) socket. 343 */ 344 static void tcp_fastopen_synack_timer(struct sock *sk) 345 { 346 struct inet_connection_sock *icsk = inet_csk(sk); 347 int max_retries = icsk->icsk_syn_retries ? : 348 sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */ 349 struct request_sock *req; 350 351 req = tcp_sk(sk)->fastopen_rsk; 352 req->rsk_ops->syn_ack_timeout(req); 353 354 if (req->num_timeout >= max_retries) { 355 tcp_write_err(sk); 356 return; 357 } 358 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error 359 * returned from rtx_syn_ack() to make it more persistent like 360 * regular retransmit because if the child socket has been accepted 361 * it's not good to give up too easily. 362 */ 363 inet_rtx_syn_ack(sk, req); 364 req->num_timeout++; 365 icsk->icsk_retransmits++; 366 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 367 TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX); 368 } 369 370 371 /** 372 * tcp_retransmit_timer() - The TCP retransmit timeout handler 373 * @sk: Pointer to the current socket. 374 * 375 * This function gets called when the kernel timer for a TCP packet 376 * of this socket expires. 377 * 378 * It handles retransmission, timer adjustment and other necesarry measures. 379 * 380 * Returns: Nothing (void) 381 */ 382 void tcp_retransmit_timer(struct sock *sk) 383 { 384 struct tcp_sock *tp = tcp_sk(sk); 385 struct net *net = sock_net(sk); 386 struct inet_connection_sock *icsk = inet_csk(sk); 387 388 if (tp->fastopen_rsk) { 389 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 390 sk->sk_state != TCP_FIN_WAIT1); 391 tcp_fastopen_synack_timer(sk); 392 /* Before we receive ACK to our SYN-ACK don't retransmit 393 * anything else (e.g., data or FIN segments). 394 */ 395 return; 396 } 397 if (!tp->packets_out) 398 goto out; 399 400 WARN_ON(tcp_rtx_queue_empty(sk)); 401 402 tp->tlp_high_seq = 0; 403 404 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) && 405 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) { 406 /* Receiver dastardly shrinks window. Our retransmits 407 * become zero probes, but we should not timeout this 408 * connection. If the socket is an orphan, time it out, 409 * we cannot allow such beasts to hang infinitely. 410 */ 411 struct inet_sock *inet = inet_sk(sk); 412 if (sk->sk_family == AF_INET) { 413 net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 414 &inet->inet_daddr, 415 ntohs(inet->inet_dport), 416 inet->inet_num, 417 tp->snd_una, tp->snd_nxt); 418 } 419 #if IS_ENABLED(CONFIG_IPV6) 420 else if (sk->sk_family == AF_INET6) { 421 net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 422 &sk->sk_v6_daddr, 423 ntohs(inet->inet_dport), 424 inet->inet_num, 425 tp->snd_una, tp->snd_nxt); 426 } 427 #endif 428 if (tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX) { 429 tcp_write_err(sk); 430 goto out; 431 } 432 tcp_enter_loss(sk); 433 tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1); 434 __sk_dst_reset(sk); 435 goto out_reset_timer; 436 } 437 438 if (tcp_write_timeout(sk)) 439 goto out; 440 441 if (icsk->icsk_retransmits == 0) { 442 int mib_idx; 443 444 if (icsk->icsk_ca_state == TCP_CA_Recovery) { 445 if (tcp_is_sack(tp)) 446 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL; 447 else 448 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL; 449 } else if (icsk->icsk_ca_state == TCP_CA_Loss) { 450 mib_idx = LINUX_MIB_TCPLOSSFAILURES; 451 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) || 452 tp->sacked_out) { 453 if (tcp_is_sack(tp)) 454 mib_idx = LINUX_MIB_TCPSACKFAILURES; 455 else 456 mib_idx = LINUX_MIB_TCPRENOFAILURES; 457 } else { 458 mib_idx = LINUX_MIB_TCPTIMEOUTS; 459 } 460 __NET_INC_STATS(sock_net(sk), mib_idx); 461 } 462 463 tcp_enter_loss(sk); 464 465 if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) { 466 /* Retransmission failed because of local congestion, 467 * do not backoff. 468 */ 469 if (!icsk->icsk_retransmits) 470 icsk->icsk_retransmits = 1; 471 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 472 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL), 473 TCP_RTO_MAX); 474 goto out; 475 } 476 477 /* Increase the timeout each time we retransmit. Note that 478 * we do not increase the rtt estimate. rto is initialized 479 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests 480 * that doubling rto each time is the least we can get away with. 481 * In KA9Q, Karn uses this for the first few times, and then 482 * goes to quadratic. netBSD doubles, but only goes up to *64, 483 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is 484 * defined in the protocol as the maximum possible RTT. I guess 485 * we'll have to use something other than TCP to talk to the 486 * University of Mars. 487 * 488 * PAWS allows us longer timeouts and large windows, so once 489 * implemented ftp to mars will work nicely. We will have to fix 490 * the 120 second clamps though! 491 */ 492 icsk->icsk_backoff++; 493 icsk->icsk_retransmits++; 494 495 out_reset_timer: 496 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is 497 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this 498 * might be increased if the stream oscillates between thin and thick, 499 * thus the old value might already be too high compared to the value 500 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without 501 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating 502 * exponential backoff behaviour to avoid continue hammering 503 * linear-timeout retransmissions into a black hole 504 */ 505 if (sk->sk_state == TCP_ESTABLISHED && 506 (tp->thin_lto || net->ipv4.sysctl_tcp_thin_linear_timeouts) && 507 tcp_stream_is_thin(tp) && 508 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) { 509 icsk->icsk_backoff = 0; 510 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX); 511 } else { 512 /* Use normal (exponential) backoff */ 513 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX); 514 } 515 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX); 516 if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0)) 517 __sk_dst_reset(sk); 518 519 out:; 520 } 521 522 /* Called with bottom-half processing disabled. 523 Called by tcp_write_timer() */ 524 void tcp_write_timer_handler(struct sock *sk) 525 { 526 struct inet_connection_sock *icsk = inet_csk(sk); 527 int event; 528 529 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 530 !icsk->icsk_pending) 531 goto out; 532 533 if (time_after(icsk->icsk_timeout, jiffies)) { 534 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout); 535 goto out; 536 } 537 538 tcp_mstamp_refresh(tcp_sk(sk)); 539 event = icsk->icsk_pending; 540 541 switch (event) { 542 case ICSK_TIME_REO_TIMEOUT: 543 tcp_rack_reo_timeout(sk); 544 break; 545 case ICSK_TIME_LOSS_PROBE: 546 tcp_send_loss_probe(sk); 547 break; 548 case ICSK_TIME_RETRANS: 549 icsk->icsk_pending = 0; 550 tcp_retransmit_timer(sk); 551 break; 552 case ICSK_TIME_PROBE0: 553 icsk->icsk_pending = 0; 554 tcp_probe_timer(sk); 555 break; 556 } 557 558 out: 559 sk_mem_reclaim(sk); 560 } 561 562 static void tcp_write_timer(struct timer_list *t) 563 { 564 struct inet_connection_sock *icsk = 565 from_timer(icsk, t, icsk_retransmit_timer); 566 struct sock *sk = &icsk->icsk_inet.sk; 567 568 bh_lock_sock(sk); 569 if (!sock_owned_by_user(sk)) { 570 tcp_write_timer_handler(sk); 571 } else { 572 /* delegate our work to tcp_release_cb() */ 573 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags)) 574 sock_hold(sk); 575 } 576 bh_unlock_sock(sk); 577 sock_put(sk); 578 } 579 580 void tcp_syn_ack_timeout(const struct request_sock *req) 581 { 582 struct net *net = read_pnet(&inet_rsk(req)->ireq_net); 583 584 __NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS); 585 } 586 EXPORT_SYMBOL(tcp_syn_ack_timeout); 587 588 void tcp_set_keepalive(struct sock *sk, int val) 589 { 590 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 591 return; 592 593 if (val && !sock_flag(sk, SOCK_KEEPOPEN)) 594 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk))); 595 else if (!val) 596 inet_csk_delete_keepalive_timer(sk); 597 } 598 EXPORT_SYMBOL_GPL(tcp_set_keepalive); 599 600 601 static void tcp_keepalive_timer (struct timer_list *t) 602 { 603 struct sock *sk = from_timer(sk, t, sk_timer); 604 struct inet_connection_sock *icsk = inet_csk(sk); 605 struct tcp_sock *tp = tcp_sk(sk); 606 u32 elapsed; 607 608 /* Only process if socket is not in use. */ 609 bh_lock_sock(sk); 610 if (sock_owned_by_user(sk)) { 611 /* Try again later. */ 612 inet_csk_reset_keepalive_timer (sk, HZ/20); 613 goto out; 614 } 615 616 if (sk->sk_state == TCP_LISTEN) { 617 pr_err("Hmm... keepalive on a LISTEN ???\n"); 618 goto out; 619 } 620 621 tcp_mstamp_refresh(tp); 622 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) { 623 if (tp->linger2 >= 0) { 624 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN; 625 626 if (tmo > 0) { 627 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 628 goto out; 629 } 630 } 631 tcp_send_active_reset(sk, GFP_ATOMIC); 632 goto death; 633 } 634 635 if (!sock_flag(sk, SOCK_KEEPOPEN) || 636 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT))) 637 goto out; 638 639 elapsed = keepalive_time_when(tp); 640 641 /* It is alive without keepalive 8) */ 642 if (tp->packets_out || !tcp_write_queue_empty(sk)) 643 goto resched; 644 645 elapsed = keepalive_time_elapsed(tp); 646 647 if (elapsed >= keepalive_time_when(tp)) { 648 /* If the TCP_USER_TIMEOUT option is enabled, use that 649 * to determine when to timeout instead. 650 */ 651 if ((icsk->icsk_user_timeout != 0 && 652 elapsed >= icsk->icsk_user_timeout && 653 icsk->icsk_probes_out > 0) || 654 (icsk->icsk_user_timeout == 0 && 655 icsk->icsk_probes_out >= keepalive_probes(tp))) { 656 tcp_send_active_reset(sk, GFP_ATOMIC); 657 tcp_write_err(sk); 658 goto out; 659 } 660 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) { 661 icsk->icsk_probes_out++; 662 elapsed = keepalive_intvl_when(tp); 663 } else { 664 /* If keepalive was lost due to local congestion, 665 * try harder. 666 */ 667 elapsed = TCP_RESOURCE_PROBE_INTERVAL; 668 } 669 } else { 670 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */ 671 elapsed = keepalive_time_when(tp) - elapsed; 672 } 673 674 sk_mem_reclaim(sk); 675 676 resched: 677 inet_csk_reset_keepalive_timer (sk, elapsed); 678 goto out; 679 680 death: 681 tcp_done(sk); 682 683 out: 684 bh_unlock_sock(sk); 685 sock_put(sk); 686 } 687 688 void tcp_init_xmit_timers(struct sock *sk) 689 { 690 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer, 691 &tcp_keepalive_timer); 692 hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC, 693 HRTIMER_MODE_ABS_PINNED); 694 tcp_sk(sk)->pacing_timer.function = tcp_pace_kick; 695 } 696