1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * INET An implementation of the TCP/IP protocol suite for the LINUX 4 * operating system. INET is implemented using the BSD Socket 5 * interface as the means of communication with the user level. 6 * 7 * Implementation of the Transmission Control Protocol(TCP). 8 * 9 * Authors: Ross Biro 10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 11 * Mark Evans, <evansmp@uhura.aston.ac.uk> 12 * Corey Minyard <wf-rch!minyard@relay.EU.net> 13 * Florian La Roche, <flla@stud.uni-sb.de> 14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 15 * Linus Torvalds, <torvalds@cs.helsinki.fi> 16 * Alan Cox, <gw4pts@gw4pts.ampr.org> 17 * Matthew Dillon, <dillon@apollo.west.oic.com> 18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 19 * Jorge Cwik, <jorge@laser.satlink.net> 20 */ 21 22 #include <linux/module.h> 23 #include <linux/gfp.h> 24 #include <net/tcp.h> 25 26 static u32 tcp_clamp_rto_to_user_timeout(const struct sock *sk) 27 { 28 struct inet_connection_sock *icsk = inet_csk(sk); 29 u32 elapsed, start_ts; 30 s32 remaining; 31 32 start_ts = tcp_sk(sk)->retrans_stamp; 33 if (!icsk->icsk_user_timeout) 34 return icsk->icsk_rto; 35 elapsed = tcp_time_stamp(tcp_sk(sk)) - start_ts; 36 remaining = icsk->icsk_user_timeout - elapsed; 37 if (remaining <= 0) 38 return 1; /* user timeout has passed; fire ASAP */ 39 40 return min_t(u32, icsk->icsk_rto, msecs_to_jiffies(remaining)); 41 } 42 43 u32 tcp_clamp_probe0_to_user_timeout(const struct sock *sk, u32 when) 44 { 45 struct inet_connection_sock *icsk = inet_csk(sk); 46 u32 remaining; 47 s32 elapsed; 48 49 if (!icsk->icsk_user_timeout || !icsk->icsk_probes_tstamp) 50 return when; 51 52 elapsed = tcp_jiffies32 - icsk->icsk_probes_tstamp; 53 if (unlikely(elapsed < 0)) 54 elapsed = 0; 55 remaining = msecs_to_jiffies(icsk->icsk_user_timeout) - elapsed; 56 remaining = max_t(u32, remaining, TCP_TIMEOUT_MIN); 57 58 return min_t(u32, remaining, when); 59 } 60 61 /** 62 * tcp_write_err() - close socket and save error info 63 * @sk: The socket the error has appeared on. 64 * 65 * Returns: Nothing (void) 66 */ 67 68 static void tcp_write_err(struct sock *sk) 69 { 70 WRITE_ONCE(sk->sk_err, READ_ONCE(sk->sk_err_soft) ? : ETIMEDOUT); 71 sk_error_report(sk); 72 73 tcp_write_queue_purge(sk); 74 tcp_done(sk); 75 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT); 76 } 77 78 /** 79 * tcp_out_of_resources() - Close socket if out of resources 80 * @sk: pointer to current socket 81 * @do_reset: send a last packet with reset flag 82 * 83 * Do not allow orphaned sockets to eat all our resources. 84 * This is direct violation of TCP specs, but it is required 85 * to prevent DoS attacks. It is called when a retransmission timeout 86 * or zero probe timeout occurs on orphaned socket. 87 * 88 * Also close if our net namespace is exiting; in that case there is no 89 * hope of ever communicating again since all netns interfaces are already 90 * down (or about to be down), and we need to release our dst references, 91 * which have been moved to the netns loopback interface, so the namespace 92 * can finish exiting. This condition is only possible if we are a kernel 93 * socket, as those do not hold references to the namespace. 94 * 95 * Criteria is still not confirmed experimentally and may change. 96 * We kill the socket, if: 97 * 1. If number of orphaned sockets exceeds an administratively configured 98 * limit. 99 * 2. If we have strong memory pressure. 100 * 3. If our net namespace is exiting. 101 */ 102 static int tcp_out_of_resources(struct sock *sk, bool do_reset) 103 { 104 struct tcp_sock *tp = tcp_sk(sk); 105 int shift = 0; 106 107 /* If peer does not open window for long time, or did not transmit 108 * anything for long time, penalize it. */ 109 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset) 110 shift++; 111 112 /* If some dubious ICMP arrived, penalize even more. */ 113 if (READ_ONCE(sk->sk_err_soft)) 114 shift++; 115 116 if (tcp_check_oom(sk, shift)) { 117 /* Catch exceptional cases, when connection requires reset. 118 * 1. Last segment was sent recently. */ 119 if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN || 120 /* 2. Window is closed. */ 121 (!tp->snd_wnd && !tp->packets_out)) 122 do_reset = true; 123 if (do_reset) 124 tcp_send_active_reset(sk, GFP_ATOMIC); 125 tcp_done(sk); 126 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY); 127 return 1; 128 } 129 130 if (!check_net(sock_net(sk))) { 131 /* Not possible to send reset; just close */ 132 tcp_done(sk); 133 return 1; 134 } 135 136 return 0; 137 } 138 139 /** 140 * tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket 141 * @sk: Pointer to the current socket. 142 * @alive: bool, socket alive state 143 */ 144 static int tcp_orphan_retries(struct sock *sk, bool alive) 145 { 146 int retries = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_orphan_retries); /* May be zero. */ 147 148 /* We know from an ICMP that something is wrong. */ 149 if (READ_ONCE(sk->sk_err_soft) && !alive) 150 retries = 0; 151 152 /* However, if socket sent something recently, select some safe 153 * number of retries. 8 corresponds to >100 seconds with minimal 154 * RTO of 200msec. */ 155 if (retries == 0 && alive) 156 retries = 8; 157 return retries; 158 } 159 160 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk) 161 { 162 const struct net *net = sock_net(sk); 163 int mss; 164 165 /* Black hole detection */ 166 if (!READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing)) 167 return; 168 169 if (!icsk->icsk_mtup.enabled) { 170 icsk->icsk_mtup.enabled = 1; 171 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 172 } else { 173 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1; 174 mss = min(READ_ONCE(net->ipv4.sysctl_tcp_base_mss), mss); 175 mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_mtu_probe_floor)); 176 mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_min_snd_mss)); 177 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 178 } 179 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 180 } 181 182 static unsigned int tcp_model_timeout(struct sock *sk, 183 unsigned int boundary, 184 unsigned int rto_base) 185 { 186 unsigned int linear_backoff_thresh, timeout; 187 188 linear_backoff_thresh = ilog2(TCP_RTO_MAX / rto_base); 189 if (boundary <= linear_backoff_thresh) 190 timeout = ((2 << boundary) - 1) * rto_base; 191 else 192 timeout = ((2 << linear_backoff_thresh) - 1) * rto_base + 193 (boundary - linear_backoff_thresh) * TCP_RTO_MAX; 194 return jiffies_to_msecs(timeout); 195 } 196 /** 197 * retransmits_timed_out() - returns true if this connection has timed out 198 * @sk: The current socket 199 * @boundary: max number of retransmissions 200 * @timeout: A custom timeout value. 201 * If set to 0 the default timeout is calculated and used. 202 * Using TCP_RTO_MIN and the number of unsuccessful retransmits. 203 * 204 * The default "timeout" value this function can calculate and use 205 * is equivalent to the timeout of a TCP Connection 206 * after "boundary" unsuccessful, exponentially backed-off 207 * retransmissions with an initial RTO of TCP_RTO_MIN. 208 */ 209 static bool retransmits_timed_out(struct sock *sk, 210 unsigned int boundary, 211 unsigned int timeout) 212 { 213 unsigned int start_ts; 214 215 if (!inet_csk(sk)->icsk_retransmits) 216 return false; 217 218 start_ts = tcp_sk(sk)->retrans_stamp; 219 if (likely(timeout == 0)) { 220 unsigned int rto_base = TCP_RTO_MIN; 221 222 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 223 rto_base = tcp_timeout_init(sk); 224 timeout = tcp_model_timeout(sk, boundary, rto_base); 225 } 226 227 return (s32)(tcp_time_stamp(tcp_sk(sk)) - start_ts - timeout) >= 0; 228 } 229 230 /* A write timeout has occurred. Process the after effects. */ 231 static int tcp_write_timeout(struct sock *sk) 232 { 233 struct inet_connection_sock *icsk = inet_csk(sk); 234 struct tcp_sock *tp = tcp_sk(sk); 235 struct net *net = sock_net(sk); 236 bool expired = false, do_reset; 237 int retry_until; 238 239 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) { 240 if (icsk->icsk_retransmits) 241 __dst_negative_advice(sk); 242 retry_until = icsk->icsk_syn_retries ? : 243 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries); 244 expired = icsk->icsk_retransmits >= retry_until; 245 } else { 246 if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1), 0)) { 247 /* Black hole detection */ 248 tcp_mtu_probing(icsk, sk); 249 250 __dst_negative_advice(sk); 251 } 252 253 retry_until = READ_ONCE(net->ipv4.sysctl_tcp_retries2); 254 if (sock_flag(sk, SOCK_DEAD)) { 255 const bool alive = icsk->icsk_rto < TCP_RTO_MAX; 256 257 retry_until = tcp_orphan_retries(sk, alive); 258 do_reset = alive || 259 !retransmits_timed_out(sk, retry_until, 0); 260 261 if (tcp_out_of_resources(sk, do_reset)) 262 return 1; 263 } 264 } 265 if (!expired) 266 expired = retransmits_timed_out(sk, retry_until, 267 icsk->icsk_user_timeout); 268 tcp_fastopen_active_detect_blackhole(sk, expired); 269 270 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG)) 271 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB, 272 icsk->icsk_retransmits, 273 icsk->icsk_rto, (int)expired); 274 275 if (expired) { 276 /* Has it gone just too far? */ 277 tcp_write_err(sk); 278 return 1; 279 } 280 281 if (sk_rethink_txhash(sk)) { 282 tp->timeout_rehash++; 283 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH); 284 } 285 286 return 0; 287 } 288 289 /* Called with BH disabled */ 290 void tcp_delack_timer_handler(struct sock *sk) 291 { 292 struct inet_connection_sock *icsk = inet_csk(sk); 293 struct tcp_sock *tp = tcp_sk(sk); 294 295 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 296 return; 297 298 /* Handling the sack compression case */ 299 if (tp->compressed_ack) { 300 tcp_mstamp_refresh(tp); 301 tcp_sack_compress_send_ack(sk); 302 return; 303 } 304 305 if (!(icsk->icsk_ack.pending & ICSK_ACK_TIMER)) 306 return; 307 308 if (time_after(icsk->icsk_ack.timeout, jiffies)) { 309 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout); 310 return; 311 } 312 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER; 313 314 if (inet_csk_ack_scheduled(sk)) { 315 if (!inet_csk_in_pingpong_mode(sk)) { 316 /* Delayed ACK missed: inflate ATO. */ 317 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto); 318 } else { 319 /* Delayed ACK missed: leave pingpong mode and 320 * deflate ATO. 321 */ 322 inet_csk_exit_pingpong_mode(sk); 323 icsk->icsk_ack.ato = TCP_ATO_MIN; 324 } 325 tcp_mstamp_refresh(tp); 326 tcp_send_ack(sk); 327 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS); 328 } 329 } 330 331 332 /** 333 * tcp_delack_timer() - The TCP delayed ACK timeout handler 334 * @t: Pointer to the timer. (gets casted to struct sock *) 335 * 336 * This function gets (indirectly) called when the kernel timer for a TCP packet 337 * of this socket expires. Calls tcp_delack_timer_handler() to do the actual work. 338 * 339 * Returns: Nothing (void) 340 */ 341 static void tcp_delack_timer(struct timer_list *t) 342 { 343 struct inet_connection_sock *icsk = 344 from_timer(icsk, t, icsk_delack_timer); 345 struct sock *sk = &icsk->icsk_inet.sk; 346 347 bh_lock_sock(sk); 348 if (!sock_owned_by_user(sk)) { 349 tcp_delack_timer_handler(sk); 350 } else { 351 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED); 352 /* deleguate our work to tcp_release_cb() */ 353 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags)) 354 sock_hold(sk); 355 } 356 bh_unlock_sock(sk); 357 sock_put(sk); 358 } 359 360 static void tcp_probe_timer(struct sock *sk) 361 { 362 struct inet_connection_sock *icsk = inet_csk(sk); 363 struct sk_buff *skb = tcp_send_head(sk); 364 struct tcp_sock *tp = tcp_sk(sk); 365 int max_probes; 366 367 if (tp->packets_out || !skb) { 368 icsk->icsk_probes_out = 0; 369 icsk->icsk_probes_tstamp = 0; 370 return; 371 } 372 373 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as 374 * long as the receiver continues to respond probes. We support this by 375 * default and reset icsk_probes_out with incoming ACKs. But if the 376 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we 377 * kill the socket when the retry count and the time exceeds the 378 * corresponding system limit. We also implement similar policy when 379 * we use RTO to probe window in tcp_retransmit_timer(). 380 */ 381 if (!icsk->icsk_probes_tstamp) 382 icsk->icsk_probes_tstamp = tcp_jiffies32; 383 else if (icsk->icsk_user_timeout && 384 (s32)(tcp_jiffies32 - icsk->icsk_probes_tstamp) >= 385 msecs_to_jiffies(icsk->icsk_user_timeout)) 386 goto abort; 387 388 max_probes = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retries2); 389 if (sock_flag(sk, SOCK_DEAD)) { 390 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX; 391 392 max_probes = tcp_orphan_retries(sk, alive); 393 if (!alive && icsk->icsk_backoff >= max_probes) 394 goto abort; 395 if (tcp_out_of_resources(sk, true)) 396 return; 397 } 398 399 if (icsk->icsk_probes_out >= max_probes) { 400 abort: tcp_write_err(sk); 401 } else { 402 /* Only send another probe if we didn't close things up. */ 403 tcp_send_probe0(sk); 404 } 405 } 406 407 /* 408 * Timer for Fast Open socket to retransmit SYNACK. Note that the 409 * sk here is the child socket, not the parent (listener) socket. 410 */ 411 static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req) 412 { 413 struct inet_connection_sock *icsk = inet_csk(sk); 414 struct tcp_sock *tp = tcp_sk(sk); 415 int max_retries; 416 417 req->rsk_ops->syn_ack_timeout(req); 418 419 /* add one more retry for fastopen */ 420 max_retries = icsk->icsk_syn_retries ? : 421 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_synack_retries) + 1; 422 423 if (req->num_timeout >= max_retries) { 424 tcp_write_err(sk); 425 return; 426 } 427 /* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */ 428 if (icsk->icsk_retransmits == 1) 429 tcp_enter_loss(sk); 430 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error 431 * returned from rtx_syn_ack() to make it more persistent like 432 * regular retransmit because if the child socket has been accepted 433 * it's not good to give up too easily. 434 */ 435 inet_rtx_syn_ack(sk, req); 436 req->num_timeout++; 437 icsk->icsk_retransmits++; 438 if (!tp->retrans_stamp) 439 tp->retrans_stamp = tcp_time_stamp(tp); 440 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 441 req->timeout << req->num_timeout, TCP_RTO_MAX); 442 } 443 444 445 /** 446 * tcp_retransmit_timer() - The TCP retransmit timeout handler 447 * @sk: Pointer to the current socket. 448 * 449 * This function gets called when the kernel timer for a TCP packet 450 * of this socket expires. 451 * 452 * It handles retransmission, timer adjustment and other necessary measures. 453 * 454 * Returns: Nothing (void) 455 */ 456 void tcp_retransmit_timer(struct sock *sk) 457 { 458 struct tcp_sock *tp = tcp_sk(sk); 459 struct net *net = sock_net(sk); 460 struct inet_connection_sock *icsk = inet_csk(sk); 461 struct request_sock *req; 462 struct sk_buff *skb; 463 464 req = rcu_dereference_protected(tp->fastopen_rsk, 465 lockdep_sock_is_held(sk)); 466 if (req) { 467 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 468 sk->sk_state != TCP_FIN_WAIT1); 469 tcp_fastopen_synack_timer(sk, req); 470 /* Before we receive ACK to our SYN-ACK don't retransmit 471 * anything else (e.g., data or FIN segments). 472 */ 473 return; 474 } 475 476 if (!tp->packets_out) 477 return; 478 479 skb = tcp_rtx_queue_head(sk); 480 if (WARN_ON_ONCE(!skb)) 481 return; 482 483 tp->tlp_high_seq = 0; 484 485 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) && 486 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) { 487 /* Receiver dastardly shrinks window. Our retransmits 488 * become zero probes, but we should not timeout this 489 * connection. If the socket is an orphan, time it out, 490 * we cannot allow such beasts to hang infinitely. 491 */ 492 struct inet_sock *inet = inet_sk(sk); 493 if (sk->sk_family == AF_INET) { 494 net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 495 &inet->inet_daddr, 496 ntohs(inet->inet_dport), 497 inet->inet_num, 498 tp->snd_una, tp->snd_nxt); 499 } 500 #if IS_ENABLED(CONFIG_IPV6) 501 else if (sk->sk_family == AF_INET6) { 502 net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 503 &sk->sk_v6_daddr, 504 ntohs(inet->inet_dport), 505 inet->inet_num, 506 tp->snd_una, tp->snd_nxt); 507 } 508 #endif 509 if (tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX) { 510 tcp_write_err(sk); 511 goto out; 512 } 513 tcp_enter_loss(sk); 514 tcp_retransmit_skb(sk, skb, 1); 515 __sk_dst_reset(sk); 516 goto out_reset_timer; 517 } 518 519 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS); 520 if (tcp_write_timeout(sk)) 521 goto out; 522 523 if (icsk->icsk_retransmits == 0) { 524 int mib_idx = 0; 525 526 if (icsk->icsk_ca_state == TCP_CA_Recovery) { 527 if (tcp_is_sack(tp)) 528 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL; 529 else 530 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL; 531 } else if (icsk->icsk_ca_state == TCP_CA_Loss) { 532 mib_idx = LINUX_MIB_TCPLOSSFAILURES; 533 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) || 534 tp->sacked_out) { 535 if (tcp_is_sack(tp)) 536 mib_idx = LINUX_MIB_TCPSACKFAILURES; 537 else 538 mib_idx = LINUX_MIB_TCPRENOFAILURES; 539 } 540 if (mib_idx) 541 __NET_INC_STATS(sock_net(sk), mib_idx); 542 } 543 544 tcp_enter_loss(sk); 545 546 icsk->icsk_retransmits++; 547 if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) { 548 /* Retransmission failed because of local congestion, 549 * Let senders fight for local resources conservatively. 550 */ 551 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 552 TCP_RESOURCE_PROBE_INTERVAL, 553 TCP_RTO_MAX); 554 goto out; 555 } 556 557 /* Increase the timeout each time we retransmit. Note that 558 * we do not increase the rtt estimate. rto is initialized 559 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests 560 * that doubling rto each time is the least we can get away with. 561 * In KA9Q, Karn uses this for the first few times, and then 562 * goes to quadratic. netBSD doubles, but only goes up to *64, 563 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is 564 * defined in the protocol as the maximum possible RTT. I guess 565 * we'll have to use something other than TCP to talk to the 566 * University of Mars. 567 * 568 * PAWS allows us longer timeouts and large windows, so once 569 * implemented ftp to mars will work nicely. We will have to fix 570 * the 120 second clamps though! 571 */ 572 icsk->icsk_backoff++; 573 574 out_reset_timer: 575 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is 576 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this 577 * might be increased if the stream oscillates between thin and thick, 578 * thus the old value might already be too high compared to the value 579 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without 580 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating 581 * exponential backoff behaviour to avoid continue hammering 582 * linear-timeout retransmissions into a black hole 583 */ 584 if (sk->sk_state == TCP_ESTABLISHED && 585 (tp->thin_lto || READ_ONCE(net->ipv4.sysctl_tcp_thin_linear_timeouts)) && 586 tcp_stream_is_thin(tp) && 587 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) { 588 icsk->icsk_backoff = 0; 589 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX); 590 } else { 591 /* Use normal (exponential) backoff */ 592 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX); 593 } 594 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 595 tcp_clamp_rto_to_user_timeout(sk), TCP_RTO_MAX); 596 if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1) + 1, 0)) 597 __sk_dst_reset(sk); 598 599 out:; 600 } 601 602 /* Called with bottom-half processing disabled. 603 Called by tcp_write_timer() */ 604 void tcp_write_timer_handler(struct sock *sk) 605 { 606 struct inet_connection_sock *icsk = inet_csk(sk); 607 int event; 608 609 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 610 !icsk->icsk_pending) 611 return; 612 613 if (time_after(icsk->icsk_timeout, jiffies)) { 614 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout); 615 return; 616 } 617 618 tcp_mstamp_refresh(tcp_sk(sk)); 619 event = icsk->icsk_pending; 620 621 switch (event) { 622 case ICSK_TIME_REO_TIMEOUT: 623 tcp_rack_reo_timeout(sk); 624 break; 625 case ICSK_TIME_LOSS_PROBE: 626 tcp_send_loss_probe(sk); 627 break; 628 case ICSK_TIME_RETRANS: 629 icsk->icsk_pending = 0; 630 tcp_retransmit_timer(sk); 631 break; 632 case ICSK_TIME_PROBE0: 633 icsk->icsk_pending = 0; 634 tcp_probe_timer(sk); 635 break; 636 } 637 } 638 639 static void tcp_write_timer(struct timer_list *t) 640 { 641 struct inet_connection_sock *icsk = 642 from_timer(icsk, t, icsk_retransmit_timer); 643 struct sock *sk = &icsk->icsk_inet.sk; 644 645 bh_lock_sock(sk); 646 if (!sock_owned_by_user(sk)) { 647 tcp_write_timer_handler(sk); 648 } else { 649 /* delegate our work to tcp_release_cb() */ 650 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags)) 651 sock_hold(sk); 652 } 653 bh_unlock_sock(sk); 654 sock_put(sk); 655 } 656 657 void tcp_syn_ack_timeout(const struct request_sock *req) 658 { 659 struct net *net = read_pnet(&inet_rsk(req)->ireq_net); 660 661 __NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS); 662 } 663 EXPORT_SYMBOL(tcp_syn_ack_timeout); 664 665 void tcp_set_keepalive(struct sock *sk, int val) 666 { 667 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 668 return; 669 670 if (val && !sock_flag(sk, SOCK_KEEPOPEN)) 671 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk))); 672 else if (!val) 673 inet_csk_delete_keepalive_timer(sk); 674 } 675 EXPORT_SYMBOL_GPL(tcp_set_keepalive); 676 677 678 static void tcp_keepalive_timer (struct timer_list *t) 679 { 680 struct sock *sk = from_timer(sk, t, sk_timer); 681 struct inet_connection_sock *icsk = inet_csk(sk); 682 struct tcp_sock *tp = tcp_sk(sk); 683 u32 elapsed; 684 685 /* Only process if socket is not in use. */ 686 bh_lock_sock(sk); 687 if (sock_owned_by_user(sk)) { 688 /* Try again later. */ 689 inet_csk_reset_keepalive_timer (sk, HZ/20); 690 goto out; 691 } 692 693 if (sk->sk_state == TCP_LISTEN) { 694 pr_err("Hmm... keepalive on a LISTEN ???\n"); 695 goto out; 696 } 697 698 tcp_mstamp_refresh(tp); 699 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) { 700 if (tp->linger2 >= 0) { 701 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN; 702 703 if (tmo > 0) { 704 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 705 goto out; 706 } 707 } 708 tcp_send_active_reset(sk, GFP_ATOMIC); 709 goto death; 710 } 711 712 if (!sock_flag(sk, SOCK_KEEPOPEN) || 713 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT))) 714 goto out; 715 716 elapsed = keepalive_time_when(tp); 717 718 /* It is alive without keepalive 8) */ 719 if (tp->packets_out || !tcp_write_queue_empty(sk)) 720 goto resched; 721 722 elapsed = keepalive_time_elapsed(tp); 723 724 if (elapsed >= keepalive_time_when(tp)) { 725 /* If the TCP_USER_TIMEOUT option is enabled, use that 726 * to determine when to timeout instead. 727 */ 728 if ((icsk->icsk_user_timeout != 0 && 729 elapsed >= msecs_to_jiffies(icsk->icsk_user_timeout) && 730 icsk->icsk_probes_out > 0) || 731 (icsk->icsk_user_timeout == 0 && 732 icsk->icsk_probes_out >= keepalive_probes(tp))) { 733 tcp_send_active_reset(sk, GFP_ATOMIC); 734 tcp_write_err(sk); 735 goto out; 736 } 737 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) { 738 icsk->icsk_probes_out++; 739 elapsed = keepalive_intvl_when(tp); 740 } else { 741 /* If keepalive was lost due to local congestion, 742 * try harder. 743 */ 744 elapsed = TCP_RESOURCE_PROBE_INTERVAL; 745 } 746 } else { 747 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */ 748 elapsed = keepalive_time_when(tp) - elapsed; 749 } 750 751 resched: 752 inet_csk_reset_keepalive_timer (sk, elapsed); 753 goto out; 754 755 death: 756 tcp_done(sk); 757 758 out: 759 bh_unlock_sock(sk); 760 sock_put(sk); 761 } 762 763 static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer) 764 { 765 struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer); 766 struct sock *sk = (struct sock *)tp; 767 768 bh_lock_sock(sk); 769 if (!sock_owned_by_user(sk)) { 770 if (tp->compressed_ack) { 771 /* Since we have to send one ack finally, 772 * subtract one from tp->compressed_ack to keep 773 * LINUX_MIB_TCPACKCOMPRESSED accurate. 774 */ 775 tp->compressed_ack--; 776 tcp_send_ack(sk); 777 } 778 } else { 779 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, 780 &sk->sk_tsq_flags)) 781 sock_hold(sk); 782 } 783 bh_unlock_sock(sk); 784 785 sock_put(sk); 786 787 return HRTIMER_NORESTART; 788 } 789 790 void tcp_init_xmit_timers(struct sock *sk) 791 { 792 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer, 793 &tcp_keepalive_timer); 794 hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC, 795 HRTIMER_MODE_ABS_PINNED_SOFT); 796 tcp_sk(sk)->pacing_timer.function = tcp_pace_kick; 797 798 hrtimer_init(&tcp_sk(sk)->compressed_ack_timer, CLOCK_MONOTONIC, 799 HRTIMER_MODE_REL_PINNED_SOFT); 800 tcp_sk(sk)->compressed_ack_timer.function = tcp_compressed_ack_kick; 801 } 802