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, max_retransmits; 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); 244 245 max_retransmits = retry_until; 246 if (sk->sk_state == TCP_SYN_SENT) 247 max_retransmits += READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts); 248 249 expired = icsk->icsk_retransmits >= max_retransmits; 250 } else { 251 if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1), 0)) { 252 /* Black hole detection */ 253 tcp_mtu_probing(icsk, sk); 254 255 __dst_negative_advice(sk); 256 } 257 258 retry_until = READ_ONCE(net->ipv4.sysctl_tcp_retries2); 259 if (sock_flag(sk, SOCK_DEAD)) { 260 const bool alive = icsk->icsk_rto < TCP_RTO_MAX; 261 262 retry_until = tcp_orphan_retries(sk, alive); 263 do_reset = alive || 264 !retransmits_timed_out(sk, retry_until, 0); 265 266 if (tcp_out_of_resources(sk, do_reset)) 267 return 1; 268 } 269 } 270 if (!expired) 271 expired = retransmits_timed_out(sk, retry_until, 272 icsk->icsk_user_timeout); 273 tcp_fastopen_active_detect_blackhole(sk, expired); 274 275 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG)) 276 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB, 277 icsk->icsk_retransmits, 278 icsk->icsk_rto, (int)expired); 279 280 if (expired) { 281 /* Has it gone just too far? */ 282 tcp_write_err(sk); 283 return 1; 284 } 285 286 if (sk_rethink_txhash(sk)) { 287 tp->timeout_rehash++; 288 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH); 289 } 290 291 return 0; 292 } 293 294 /* Called with BH disabled */ 295 void tcp_delack_timer_handler(struct sock *sk) 296 { 297 struct inet_connection_sock *icsk = inet_csk(sk); 298 struct tcp_sock *tp = tcp_sk(sk); 299 300 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 301 return; 302 303 /* Handling the sack compression case */ 304 if (tp->compressed_ack) { 305 tcp_mstamp_refresh(tp); 306 tcp_sack_compress_send_ack(sk); 307 return; 308 } 309 310 if (!(icsk->icsk_ack.pending & ICSK_ACK_TIMER)) 311 return; 312 313 if (time_after(icsk->icsk_ack.timeout, jiffies)) { 314 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout); 315 return; 316 } 317 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER; 318 319 if (inet_csk_ack_scheduled(sk)) { 320 if (!inet_csk_in_pingpong_mode(sk)) { 321 /* Delayed ACK missed: inflate ATO. */ 322 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto); 323 } else { 324 /* Delayed ACK missed: leave pingpong mode and 325 * deflate ATO. 326 */ 327 inet_csk_exit_pingpong_mode(sk); 328 icsk->icsk_ack.ato = TCP_ATO_MIN; 329 } 330 tcp_mstamp_refresh(tp); 331 tcp_send_ack(sk); 332 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS); 333 } 334 } 335 336 337 /** 338 * tcp_delack_timer() - The TCP delayed ACK timeout handler 339 * @t: Pointer to the timer. (gets casted to struct sock *) 340 * 341 * This function gets (indirectly) called when the kernel timer for a TCP packet 342 * of this socket expires. Calls tcp_delack_timer_handler() to do the actual work. 343 * 344 * Returns: Nothing (void) 345 */ 346 static void tcp_delack_timer(struct timer_list *t) 347 { 348 struct inet_connection_sock *icsk = 349 from_timer(icsk, t, icsk_delack_timer); 350 struct sock *sk = &icsk->icsk_inet.sk; 351 352 bh_lock_sock(sk); 353 if (!sock_owned_by_user(sk)) { 354 tcp_delack_timer_handler(sk); 355 } else { 356 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED); 357 /* deleguate our work to tcp_release_cb() */ 358 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags)) 359 sock_hold(sk); 360 } 361 bh_unlock_sock(sk); 362 sock_put(sk); 363 } 364 365 static void tcp_probe_timer(struct sock *sk) 366 { 367 struct inet_connection_sock *icsk = inet_csk(sk); 368 struct sk_buff *skb = tcp_send_head(sk); 369 struct tcp_sock *tp = tcp_sk(sk); 370 int max_probes; 371 372 if (tp->packets_out || !skb) { 373 icsk->icsk_probes_out = 0; 374 icsk->icsk_probes_tstamp = 0; 375 return; 376 } 377 378 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as 379 * long as the receiver continues to respond probes. We support this by 380 * default and reset icsk_probes_out with incoming ACKs. But if the 381 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we 382 * kill the socket when the retry count and the time exceeds the 383 * corresponding system limit. We also implement similar policy when 384 * we use RTO to probe window in tcp_retransmit_timer(). 385 */ 386 if (!icsk->icsk_probes_tstamp) 387 icsk->icsk_probes_tstamp = tcp_jiffies32; 388 else if (icsk->icsk_user_timeout && 389 (s32)(tcp_jiffies32 - icsk->icsk_probes_tstamp) >= 390 msecs_to_jiffies(icsk->icsk_user_timeout)) 391 goto abort; 392 393 max_probes = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retries2); 394 if (sock_flag(sk, SOCK_DEAD)) { 395 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX; 396 397 max_probes = tcp_orphan_retries(sk, alive); 398 if (!alive && icsk->icsk_backoff >= max_probes) 399 goto abort; 400 if (tcp_out_of_resources(sk, true)) 401 return; 402 } 403 404 if (icsk->icsk_probes_out >= max_probes) { 405 abort: tcp_write_err(sk); 406 } else { 407 /* Only send another probe if we didn't close things up. */ 408 tcp_send_probe0(sk); 409 } 410 } 411 412 /* 413 * Timer for Fast Open socket to retransmit SYNACK. Note that the 414 * sk here is the child socket, not the parent (listener) socket. 415 */ 416 static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req) 417 { 418 struct inet_connection_sock *icsk = inet_csk(sk); 419 struct tcp_sock *tp = tcp_sk(sk); 420 int max_retries; 421 422 req->rsk_ops->syn_ack_timeout(req); 423 424 /* add one more retry for fastopen */ 425 max_retries = icsk->icsk_syn_retries ? : 426 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_synack_retries) + 1; 427 428 if (req->num_timeout >= max_retries) { 429 tcp_write_err(sk); 430 return; 431 } 432 /* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */ 433 if (icsk->icsk_retransmits == 1) 434 tcp_enter_loss(sk); 435 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error 436 * returned from rtx_syn_ack() to make it more persistent like 437 * regular retransmit because if the child socket has been accepted 438 * it's not good to give up too easily. 439 */ 440 inet_rtx_syn_ack(sk, req); 441 req->num_timeout++; 442 icsk->icsk_retransmits++; 443 if (!tp->retrans_stamp) 444 tp->retrans_stamp = tcp_time_stamp(tp); 445 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 446 req->timeout << req->num_timeout, TCP_RTO_MAX); 447 } 448 449 450 /** 451 * tcp_retransmit_timer() - The TCP retransmit timeout handler 452 * @sk: Pointer to the current socket. 453 * 454 * This function gets called when the kernel timer for a TCP packet 455 * of this socket expires. 456 * 457 * It handles retransmission, timer adjustment and other necessary measures. 458 * 459 * Returns: Nothing (void) 460 */ 461 void tcp_retransmit_timer(struct sock *sk) 462 { 463 struct tcp_sock *tp = tcp_sk(sk); 464 struct net *net = sock_net(sk); 465 struct inet_connection_sock *icsk = inet_csk(sk); 466 struct request_sock *req; 467 struct sk_buff *skb; 468 469 req = rcu_dereference_protected(tp->fastopen_rsk, 470 lockdep_sock_is_held(sk)); 471 if (req) { 472 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 473 sk->sk_state != TCP_FIN_WAIT1); 474 tcp_fastopen_synack_timer(sk, req); 475 /* Before we receive ACK to our SYN-ACK don't retransmit 476 * anything else (e.g., data or FIN segments). 477 */ 478 return; 479 } 480 481 if (!tp->packets_out) 482 return; 483 484 skb = tcp_rtx_queue_head(sk); 485 if (WARN_ON_ONCE(!skb)) 486 return; 487 488 tp->tlp_high_seq = 0; 489 490 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) && 491 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) { 492 /* Receiver dastardly shrinks window. Our retransmits 493 * become zero probes, but we should not timeout this 494 * connection. If the socket is an orphan, time it out, 495 * we cannot allow such beasts to hang infinitely. 496 */ 497 struct inet_sock *inet = inet_sk(sk); 498 if (sk->sk_family == AF_INET) { 499 net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 500 &inet->inet_daddr, 501 ntohs(inet->inet_dport), 502 inet->inet_num, 503 tp->snd_una, tp->snd_nxt); 504 } 505 #if IS_ENABLED(CONFIG_IPV6) 506 else if (sk->sk_family == AF_INET6) { 507 net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 508 &sk->sk_v6_daddr, 509 ntohs(inet->inet_dport), 510 inet->inet_num, 511 tp->snd_una, tp->snd_nxt); 512 } 513 #endif 514 if (tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX) { 515 tcp_write_err(sk); 516 goto out; 517 } 518 tcp_enter_loss(sk); 519 tcp_retransmit_skb(sk, skb, 1); 520 __sk_dst_reset(sk); 521 goto out_reset_timer; 522 } 523 524 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS); 525 if (tcp_write_timeout(sk)) 526 goto out; 527 528 if (icsk->icsk_retransmits == 0) { 529 int mib_idx = 0; 530 531 if (icsk->icsk_ca_state == TCP_CA_Recovery) { 532 if (tcp_is_sack(tp)) 533 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL; 534 else 535 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL; 536 } else if (icsk->icsk_ca_state == TCP_CA_Loss) { 537 mib_idx = LINUX_MIB_TCPLOSSFAILURES; 538 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) || 539 tp->sacked_out) { 540 if (tcp_is_sack(tp)) 541 mib_idx = LINUX_MIB_TCPSACKFAILURES; 542 else 543 mib_idx = LINUX_MIB_TCPRENOFAILURES; 544 } 545 if (mib_idx) 546 __NET_INC_STATS(sock_net(sk), mib_idx); 547 } 548 549 tcp_enter_loss(sk); 550 551 icsk->icsk_retransmits++; 552 if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) { 553 /* Retransmission failed because of local congestion, 554 * Let senders fight for local resources conservatively. 555 */ 556 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 557 TCP_RESOURCE_PROBE_INTERVAL, 558 TCP_RTO_MAX); 559 goto out; 560 } 561 562 /* Increase the timeout each time we retransmit. Note that 563 * we do not increase the rtt estimate. rto is initialized 564 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests 565 * that doubling rto each time is the least we can get away with. 566 * In KA9Q, Karn uses this for the first few times, and then 567 * goes to quadratic. netBSD doubles, but only goes up to *64, 568 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is 569 * defined in the protocol as the maximum possible RTT. I guess 570 * we'll have to use something other than TCP to talk to the 571 * University of Mars. 572 * 573 * PAWS allows us longer timeouts and large windows, so once 574 * implemented ftp to mars will work nicely. We will have to fix 575 * the 120 second clamps though! 576 */ 577 icsk->icsk_backoff++; 578 579 out_reset_timer: 580 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is 581 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this 582 * might be increased if the stream oscillates between thin and thick, 583 * thus the old value might already be too high compared to the value 584 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without 585 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating 586 * exponential backoff behaviour to avoid continue hammering 587 * linear-timeout retransmissions into a black hole 588 */ 589 if (sk->sk_state == TCP_ESTABLISHED && 590 (tp->thin_lto || READ_ONCE(net->ipv4.sysctl_tcp_thin_linear_timeouts)) && 591 tcp_stream_is_thin(tp) && 592 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) { 593 icsk->icsk_backoff = 0; 594 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX); 595 } else if (sk->sk_state != TCP_SYN_SENT || 596 icsk->icsk_backoff > 597 READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts)) { 598 /* Use normal (exponential) backoff unless linear timeouts are 599 * activated. 600 */ 601 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX); 602 } 603 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 604 tcp_clamp_rto_to_user_timeout(sk), TCP_RTO_MAX); 605 if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1) + 1, 0)) 606 __sk_dst_reset(sk); 607 608 out:; 609 } 610 611 /* Called with bottom-half processing disabled. 612 Called by tcp_write_timer() */ 613 void tcp_write_timer_handler(struct sock *sk) 614 { 615 struct inet_connection_sock *icsk = inet_csk(sk); 616 int event; 617 618 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 619 !icsk->icsk_pending) 620 return; 621 622 if (time_after(icsk->icsk_timeout, jiffies)) { 623 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout); 624 return; 625 } 626 627 tcp_mstamp_refresh(tcp_sk(sk)); 628 event = icsk->icsk_pending; 629 630 switch (event) { 631 case ICSK_TIME_REO_TIMEOUT: 632 tcp_rack_reo_timeout(sk); 633 break; 634 case ICSK_TIME_LOSS_PROBE: 635 tcp_send_loss_probe(sk); 636 break; 637 case ICSK_TIME_RETRANS: 638 icsk->icsk_pending = 0; 639 tcp_retransmit_timer(sk); 640 break; 641 case ICSK_TIME_PROBE0: 642 icsk->icsk_pending = 0; 643 tcp_probe_timer(sk); 644 break; 645 } 646 } 647 648 static void tcp_write_timer(struct timer_list *t) 649 { 650 struct inet_connection_sock *icsk = 651 from_timer(icsk, t, icsk_retransmit_timer); 652 struct sock *sk = &icsk->icsk_inet.sk; 653 654 bh_lock_sock(sk); 655 if (!sock_owned_by_user(sk)) { 656 tcp_write_timer_handler(sk); 657 } else { 658 /* delegate our work to tcp_release_cb() */ 659 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags)) 660 sock_hold(sk); 661 } 662 bh_unlock_sock(sk); 663 sock_put(sk); 664 } 665 666 void tcp_syn_ack_timeout(const struct request_sock *req) 667 { 668 struct net *net = read_pnet(&inet_rsk(req)->ireq_net); 669 670 __NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS); 671 } 672 EXPORT_SYMBOL(tcp_syn_ack_timeout); 673 674 void tcp_set_keepalive(struct sock *sk, int val) 675 { 676 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 677 return; 678 679 if (val && !sock_flag(sk, SOCK_KEEPOPEN)) 680 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk))); 681 else if (!val) 682 inet_csk_delete_keepalive_timer(sk); 683 } 684 EXPORT_SYMBOL_GPL(tcp_set_keepalive); 685 686 687 static void tcp_keepalive_timer (struct timer_list *t) 688 { 689 struct sock *sk = from_timer(sk, t, sk_timer); 690 struct inet_connection_sock *icsk = inet_csk(sk); 691 struct tcp_sock *tp = tcp_sk(sk); 692 u32 elapsed; 693 694 /* Only process if socket is not in use. */ 695 bh_lock_sock(sk); 696 if (sock_owned_by_user(sk)) { 697 /* Try again later. */ 698 inet_csk_reset_keepalive_timer (sk, HZ/20); 699 goto out; 700 } 701 702 if (sk->sk_state == TCP_LISTEN) { 703 pr_err("Hmm... keepalive on a LISTEN ???\n"); 704 goto out; 705 } 706 707 tcp_mstamp_refresh(tp); 708 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) { 709 if (tp->linger2 >= 0) { 710 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN; 711 712 if (tmo > 0) { 713 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 714 goto out; 715 } 716 } 717 tcp_send_active_reset(sk, GFP_ATOMIC); 718 goto death; 719 } 720 721 if (!sock_flag(sk, SOCK_KEEPOPEN) || 722 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT))) 723 goto out; 724 725 elapsed = keepalive_time_when(tp); 726 727 /* It is alive without keepalive 8) */ 728 if (tp->packets_out || !tcp_write_queue_empty(sk)) 729 goto resched; 730 731 elapsed = keepalive_time_elapsed(tp); 732 733 if (elapsed >= keepalive_time_when(tp)) { 734 /* If the TCP_USER_TIMEOUT option is enabled, use that 735 * to determine when to timeout instead. 736 */ 737 if ((icsk->icsk_user_timeout != 0 && 738 elapsed >= msecs_to_jiffies(icsk->icsk_user_timeout) && 739 icsk->icsk_probes_out > 0) || 740 (icsk->icsk_user_timeout == 0 && 741 icsk->icsk_probes_out >= keepalive_probes(tp))) { 742 tcp_send_active_reset(sk, GFP_ATOMIC); 743 tcp_write_err(sk); 744 goto out; 745 } 746 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) { 747 icsk->icsk_probes_out++; 748 elapsed = keepalive_intvl_when(tp); 749 } else { 750 /* If keepalive was lost due to local congestion, 751 * try harder. 752 */ 753 elapsed = TCP_RESOURCE_PROBE_INTERVAL; 754 } 755 } else { 756 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */ 757 elapsed = keepalive_time_when(tp) - elapsed; 758 } 759 760 resched: 761 inet_csk_reset_keepalive_timer (sk, elapsed); 762 goto out; 763 764 death: 765 tcp_done(sk); 766 767 out: 768 bh_unlock_sock(sk); 769 sock_put(sk); 770 } 771 772 static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer) 773 { 774 struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer); 775 struct sock *sk = (struct sock *)tp; 776 777 bh_lock_sock(sk); 778 if (!sock_owned_by_user(sk)) { 779 if (tp->compressed_ack) { 780 /* Since we have to send one ack finally, 781 * subtract one from tp->compressed_ack to keep 782 * LINUX_MIB_TCPACKCOMPRESSED accurate. 783 */ 784 tp->compressed_ack--; 785 tcp_send_ack(sk); 786 } 787 } else { 788 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, 789 &sk->sk_tsq_flags)) 790 sock_hold(sk); 791 } 792 bh_unlock_sock(sk); 793 794 sock_put(sk); 795 796 return HRTIMER_NORESTART; 797 } 798 799 void tcp_init_xmit_timers(struct sock *sk) 800 { 801 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer, 802 &tcp_keepalive_timer); 803 hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC, 804 HRTIMER_MODE_ABS_PINNED_SOFT); 805 tcp_sk(sk)->pacing_timer.function = tcp_pace_kick; 806 807 hrtimer_init(&tcp_sk(sk)->compressed_ack_timer, CLOCK_MONOTONIC, 808 HRTIMER_MODE_REL_PINNED_SOFT); 809 tcp_sk(sk)->compressed_ack_timer.function = tcp_compressed_ack_kick; 810 } 811