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 #include <net/rstreason.h> 26 27 static u32 tcp_clamp_rto_to_user_timeout(const struct sock *sk) 28 { 29 const struct inet_connection_sock *icsk = inet_csk(sk); 30 const struct tcp_sock *tp = tcp_sk(sk); 31 u32 elapsed, user_timeout; 32 s32 remaining; 33 34 user_timeout = READ_ONCE(icsk->icsk_user_timeout); 35 if (!user_timeout) 36 return icsk->icsk_rto; 37 38 elapsed = tcp_time_stamp_ts(tp) - tp->retrans_stamp; 39 if (tp->tcp_usec_ts) 40 elapsed /= USEC_PER_MSEC; 41 42 remaining = user_timeout - elapsed; 43 if (remaining <= 0) 44 return 1; /* user timeout has passed; fire ASAP */ 45 46 return min_t(u32, icsk->icsk_rto, msecs_to_jiffies(remaining)); 47 } 48 49 u32 tcp_clamp_probe0_to_user_timeout(const struct sock *sk, u32 when) 50 { 51 const struct inet_connection_sock *icsk = inet_csk(sk); 52 u32 remaining, user_timeout; 53 s32 elapsed; 54 55 user_timeout = READ_ONCE(icsk->icsk_user_timeout); 56 if (!user_timeout || !icsk->icsk_probes_tstamp) 57 return when; 58 59 elapsed = tcp_jiffies32 - icsk->icsk_probes_tstamp; 60 if (unlikely(elapsed < 0)) 61 elapsed = 0; 62 remaining = msecs_to_jiffies(user_timeout) - elapsed; 63 remaining = max_t(u32, remaining, TCP_TIMEOUT_MIN); 64 65 return min_t(u32, remaining, when); 66 } 67 68 /** 69 * tcp_write_err() - close socket and save error info 70 * @sk: The socket the error has appeared on. 71 * 72 * Returns: Nothing (void) 73 */ 74 75 static void tcp_write_err(struct sock *sk) 76 { 77 tcp_done_with_error(sk, READ_ONCE(sk->sk_err_soft) ? : ETIMEDOUT); 78 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT); 79 } 80 81 /** 82 * tcp_out_of_resources() - Close socket if out of resources 83 * @sk: pointer to current socket 84 * @do_reset: send a last packet with reset flag 85 * 86 * Do not allow orphaned sockets to eat all our resources. 87 * This is direct violation of TCP specs, but it is required 88 * to prevent DoS attacks. It is called when a retransmission timeout 89 * or zero probe timeout occurs on orphaned socket. 90 * 91 * Also close if our net namespace is exiting; in that case there is no 92 * hope of ever communicating again since all netns interfaces are already 93 * down (or about to be down), and we need to release our dst references, 94 * which have been moved to the netns loopback interface, so the namespace 95 * can finish exiting. This condition is only possible if we are a kernel 96 * socket, as those do not hold references to the namespace. 97 * 98 * Criteria is still not confirmed experimentally and may change. 99 * We kill the socket, if: 100 * 1. If number of orphaned sockets exceeds an administratively configured 101 * limit. 102 * 2. If we have strong memory pressure. 103 * 3. If our net namespace is exiting. 104 */ 105 static int tcp_out_of_resources(struct sock *sk, bool do_reset) 106 { 107 struct tcp_sock *tp = tcp_sk(sk); 108 int shift = 0; 109 110 /* If peer does not open window for long time, or did not transmit 111 * anything for long time, penalize it. */ 112 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset) 113 shift++; 114 115 /* If some dubious ICMP arrived, penalize even more. */ 116 if (READ_ONCE(sk->sk_err_soft)) 117 shift++; 118 119 if (tcp_check_oom(sk, shift)) { 120 /* Catch exceptional cases, when connection requires reset. 121 * 1. Last segment was sent recently. */ 122 if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN || 123 /* 2. Window is closed. */ 124 (!tp->snd_wnd && !tp->packets_out)) 125 do_reset = true; 126 if (do_reset) 127 tcp_send_active_reset(sk, GFP_ATOMIC, 128 SK_RST_REASON_NOT_SPECIFIED); 129 tcp_done(sk); 130 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY); 131 return 1; 132 } 133 134 if (!check_net(sock_net(sk))) { 135 /* Not possible to send reset; just close */ 136 tcp_done(sk); 137 return 1; 138 } 139 140 return 0; 141 } 142 143 /** 144 * tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket 145 * @sk: Pointer to the current socket. 146 * @alive: bool, socket alive state 147 */ 148 static int tcp_orphan_retries(struct sock *sk, bool alive) 149 { 150 int retries = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_orphan_retries); /* May be zero. */ 151 152 /* We know from an ICMP that something is wrong. */ 153 if (READ_ONCE(sk->sk_err_soft) && !alive) 154 retries = 0; 155 156 /* However, if socket sent something recently, select some safe 157 * number of retries. 8 corresponds to >100 seconds with minimal 158 * RTO of 200msec. */ 159 if (retries == 0 && alive) 160 retries = 8; 161 return retries; 162 } 163 164 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk) 165 { 166 const struct net *net = sock_net(sk); 167 int mss; 168 169 /* Black hole detection */ 170 if (!READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing)) 171 return; 172 173 if (!icsk->icsk_mtup.enabled) { 174 icsk->icsk_mtup.enabled = 1; 175 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 176 } else { 177 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1; 178 mss = min(READ_ONCE(net->ipv4.sysctl_tcp_base_mss), mss); 179 mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_mtu_probe_floor)); 180 mss = max(mss, READ_ONCE(net->ipv4.sysctl_tcp_min_snd_mss)); 181 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 182 } 183 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 184 } 185 186 static unsigned int tcp_model_timeout(struct sock *sk, 187 unsigned int boundary, 188 unsigned int rto_base) 189 { 190 unsigned int linear_backoff_thresh, timeout; 191 192 linear_backoff_thresh = ilog2(TCP_RTO_MAX / rto_base); 193 if (boundary <= linear_backoff_thresh) 194 timeout = ((2 << boundary) - 1) * rto_base; 195 else 196 timeout = ((2 << linear_backoff_thresh) - 1) * rto_base + 197 (boundary - linear_backoff_thresh) * TCP_RTO_MAX; 198 return jiffies_to_msecs(timeout); 199 } 200 /** 201 * retransmits_timed_out() - returns true if this connection has timed out 202 * @sk: The current socket 203 * @boundary: max number of retransmissions 204 * @timeout: A custom timeout value. 205 * If set to 0 the default timeout is calculated and used. 206 * Using TCP_RTO_MIN and the number of unsuccessful retransmits. 207 * 208 * The default "timeout" value this function can calculate and use 209 * is equivalent to the timeout of a TCP Connection 210 * after "boundary" unsuccessful, exponentially backed-off 211 * retransmissions with an initial RTO of TCP_RTO_MIN. 212 */ 213 static bool retransmits_timed_out(struct sock *sk, 214 unsigned int boundary, 215 unsigned int timeout) 216 { 217 struct tcp_sock *tp = tcp_sk(sk); 218 unsigned int start_ts, delta; 219 220 if (!inet_csk(sk)->icsk_retransmits) 221 return false; 222 223 start_ts = tp->retrans_stamp; 224 if (likely(timeout == 0)) { 225 unsigned int rto_base = TCP_RTO_MIN; 226 227 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) 228 rto_base = tcp_timeout_init(sk); 229 timeout = tcp_model_timeout(sk, boundary, rto_base); 230 } 231 232 if (tp->tcp_usec_ts) { 233 /* delta maybe off up to a jiffy due to timer granularity. */ 234 delta = tp->tcp_mstamp - start_ts + jiffies_to_usecs(1); 235 return (s32)(delta - timeout * USEC_PER_MSEC) >= 0; 236 } 237 return (s32)(tcp_time_stamp_ts(tp) - start_ts - timeout) >= 0; 238 } 239 240 /* A write timeout has occurred. Process the after effects. */ 241 static int tcp_write_timeout(struct sock *sk) 242 { 243 struct inet_connection_sock *icsk = inet_csk(sk); 244 struct tcp_sock *tp = tcp_sk(sk); 245 struct net *net = sock_net(sk); 246 bool expired = false, do_reset; 247 int retry_until, max_retransmits; 248 249 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) { 250 if (icsk->icsk_retransmits) 251 __dst_negative_advice(sk); 252 /* Paired with WRITE_ONCE() in tcp_sock_set_syncnt() */ 253 retry_until = READ_ONCE(icsk->icsk_syn_retries) ? : 254 READ_ONCE(net->ipv4.sysctl_tcp_syn_retries); 255 256 max_retransmits = retry_until; 257 if (sk->sk_state == TCP_SYN_SENT) 258 max_retransmits += READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts); 259 260 expired = icsk->icsk_retransmits >= max_retransmits; 261 } else { 262 if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1), 0)) { 263 /* Black hole detection */ 264 tcp_mtu_probing(icsk, sk); 265 266 __dst_negative_advice(sk); 267 } 268 269 retry_until = READ_ONCE(net->ipv4.sysctl_tcp_retries2); 270 if (sock_flag(sk, SOCK_DEAD)) { 271 const bool alive = icsk->icsk_rto < TCP_RTO_MAX; 272 273 retry_until = tcp_orphan_retries(sk, alive); 274 do_reset = alive || 275 !retransmits_timed_out(sk, retry_until, 0); 276 277 if (tcp_out_of_resources(sk, do_reset)) 278 return 1; 279 } 280 } 281 if (!expired) 282 expired = retransmits_timed_out(sk, retry_until, 283 READ_ONCE(icsk->icsk_user_timeout)); 284 tcp_fastopen_active_detect_blackhole(sk, expired); 285 286 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RTO_CB_FLAG)) 287 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RTO_CB, 288 icsk->icsk_retransmits, 289 icsk->icsk_rto, (int)expired); 290 291 if (expired) { 292 /* Has it gone just too far? */ 293 tcp_write_err(sk); 294 return 1; 295 } 296 297 if (sk_rethink_txhash(sk)) { 298 tp->timeout_rehash++; 299 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTREHASH); 300 } 301 302 return 0; 303 } 304 305 /* Called with BH disabled */ 306 void tcp_delack_timer_handler(struct sock *sk) 307 { 308 struct inet_connection_sock *icsk = inet_csk(sk); 309 struct tcp_sock *tp = tcp_sk(sk); 310 311 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 312 return; 313 314 /* Handling the sack compression case */ 315 if (tp->compressed_ack) { 316 tcp_mstamp_refresh(tp); 317 tcp_sack_compress_send_ack(sk); 318 return; 319 } 320 321 if (!(icsk->icsk_ack.pending & ICSK_ACK_TIMER)) 322 return; 323 324 if (time_after(icsk->icsk_ack.timeout, jiffies)) { 325 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout); 326 return; 327 } 328 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER; 329 330 if (inet_csk_ack_scheduled(sk)) { 331 if (!inet_csk_in_pingpong_mode(sk)) { 332 /* Delayed ACK missed: inflate ATO. */ 333 icsk->icsk_ack.ato = min_t(u32, icsk->icsk_ack.ato << 1, icsk->icsk_rto); 334 } else { 335 /* Delayed ACK missed: leave pingpong mode and 336 * deflate ATO. 337 */ 338 inet_csk_exit_pingpong_mode(sk); 339 icsk->icsk_ack.ato = TCP_ATO_MIN; 340 } 341 tcp_mstamp_refresh(tp); 342 tcp_send_ack(sk); 343 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS); 344 } 345 } 346 347 348 /** 349 * tcp_delack_timer() - The TCP delayed ACK timeout handler 350 * @t: Pointer to the timer. (gets casted to struct sock *) 351 * 352 * This function gets (indirectly) called when the kernel timer for a TCP packet 353 * of this socket expires. Calls tcp_delack_timer_handler() to do the actual work. 354 * 355 * Returns: Nothing (void) 356 */ 357 static void tcp_delack_timer(struct timer_list *t) 358 { 359 struct inet_connection_sock *icsk = 360 from_timer(icsk, t, icsk_delack_timer); 361 struct sock *sk = &icsk->icsk_inet.sk; 362 363 bh_lock_sock(sk); 364 if (!sock_owned_by_user(sk)) { 365 tcp_delack_timer_handler(sk); 366 } else { 367 __NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED); 368 /* deleguate our work to tcp_release_cb() */ 369 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags)) 370 sock_hold(sk); 371 } 372 bh_unlock_sock(sk); 373 sock_put(sk); 374 } 375 376 static void tcp_probe_timer(struct sock *sk) 377 { 378 struct inet_connection_sock *icsk = inet_csk(sk); 379 struct sk_buff *skb = tcp_send_head(sk); 380 struct tcp_sock *tp = tcp_sk(sk); 381 int max_probes; 382 383 if (tp->packets_out || !skb) { 384 icsk->icsk_probes_out = 0; 385 icsk->icsk_probes_tstamp = 0; 386 return; 387 } 388 389 /* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as 390 * long as the receiver continues to respond probes. We support this by 391 * default and reset icsk_probes_out with incoming ACKs. But if the 392 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we 393 * kill the socket when the retry count and the time exceeds the 394 * corresponding system limit. We also implement similar policy when 395 * we use RTO to probe window in tcp_retransmit_timer(). 396 */ 397 if (!icsk->icsk_probes_tstamp) { 398 icsk->icsk_probes_tstamp = tcp_jiffies32; 399 } else { 400 u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout); 401 402 if (user_timeout && 403 (s32)(tcp_jiffies32 - icsk->icsk_probes_tstamp) >= 404 msecs_to_jiffies(user_timeout)) 405 goto abort; 406 } 407 max_probes = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retries2); 408 if (sock_flag(sk, SOCK_DEAD)) { 409 const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX; 410 411 max_probes = tcp_orphan_retries(sk, alive); 412 if (!alive && icsk->icsk_backoff >= max_probes) 413 goto abort; 414 if (tcp_out_of_resources(sk, true)) 415 return; 416 } 417 418 if (icsk->icsk_probes_out >= max_probes) { 419 abort: tcp_write_err(sk); 420 } else { 421 /* Only send another probe if we didn't close things up. */ 422 tcp_send_probe0(sk); 423 } 424 } 425 426 static void tcp_update_rto_stats(struct sock *sk) 427 { 428 struct inet_connection_sock *icsk = inet_csk(sk); 429 struct tcp_sock *tp = tcp_sk(sk); 430 431 if (!icsk->icsk_retransmits) { 432 tp->total_rto_recoveries++; 433 tp->rto_stamp = tcp_time_stamp_ms(tp); 434 } 435 icsk->icsk_retransmits++; 436 tp->total_rto++; 437 } 438 439 /* 440 * Timer for Fast Open socket to retransmit SYNACK. Note that the 441 * sk here is the child socket, not the parent (listener) socket. 442 */ 443 static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req) 444 { 445 struct inet_connection_sock *icsk = inet_csk(sk); 446 struct tcp_sock *tp = tcp_sk(sk); 447 int max_retries; 448 449 req->rsk_ops->syn_ack_timeout(req); 450 451 /* Add one more retry for fastopen. 452 * Paired with WRITE_ONCE() in tcp_sock_set_syncnt() 453 */ 454 max_retries = READ_ONCE(icsk->icsk_syn_retries) ? : 455 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_synack_retries) + 1; 456 457 if (req->num_timeout >= max_retries) { 458 tcp_write_err(sk); 459 return; 460 } 461 /* Lower cwnd after certain SYNACK timeout like tcp_init_transfer() */ 462 if (icsk->icsk_retransmits == 1) 463 tcp_enter_loss(sk); 464 /* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error 465 * returned from rtx_syn_ack() to make it more persistent like 466 * regular retransmit because if the child socket has been accepted 467 * it's not good to give up too easily. 468 */ 469 inet_rtx_syn_ack(sk, req); 470 req->num_timeout++; 471 tcp_update_rto_stats(sk); 472 if (!tp->retrans_stamp) 473 tp->retrans_stamp = tcp_time_stamp_ts(tp); 474 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 475 req->timeout << req->num_timeout, TCP_RTO_MAX); 476 } 477 478 static bool tcp_rtx_probe0_timed_out(const struct sock *sk, 479 const struct sk_buff *skb, 480 u32 rtx_delta) 481 { 482 const struct inet_connection_sock *icsk = inet_csk(sk); 483 u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout); 484 const struct tcp_sock *tp = tcp_sk(sk); 485 int timeout = TCP_RTO_MAX * 2; 486 s32 rcv_delta; 487 488 if (user_timeout) { 489 /* If user application specified a TCP_USER_TIMEOUT, 490 * it does not want win 0 packets to 'reset the timer' 491 * while retransmits are not making progress. 492 */ 493 if (rtx_delta > user_timeout) 494 return true; 495 timeout = min_t(u32, timeout, msecs_to_jiffies(user_timeout)); 496 } 497 /* Note: timer interrupt might have been delayed by at least one jiffy, 498 * and tp->rcv_tstamp might very well have been written recently. 499 * rcv_delta can thus be negative. 500 */ 501 rcv_delta = icsk->icsk_timeout - tp->rcv_tstamp; 502 if (rcv_delta <= timeout) 503 return false; 504 505 return msecs_to_jiffies(rtx_delta) > timeout; 506 } 507 508 /** 509 * tcp_retransmit_timer() - The TCP retransmit timeout handler 510 * @sk: Pointer to the current socket. 511 * 512 * This function gets called when the kernel timer for a TCP packet 513 * of this socket expires. 514 * 515 * It handles retransmission, timer adjustment and other necessary measures. 516 * 517 * Returns: Nothing (void) 518 */ 519 void tcp_retransmit_timer(struct sock *sk) 520 { 521 struct tcp_sock *tp = tcp_sk(sk); 522 struct net *net = sock_net(sk); 523 struct inet_connection_sock *icsk = inet_csk(sk); 524 struct request_sock *req; 525 struct sk_buff *skb; 526 527 req = rcu_dereference_protected(tp->fastopen_rsk, 528 lockdep_sock_is_held(sk)); 529 if (req) { 530 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 531 sk->sk_state != TCP_FIN_WAIT1); 532 tcp_fastopen_synack_timer(sk, req); 533 /* Before we receive ACK to our SYN-ACK don't retransmit 534 * anything else (e.g., data or FIN segments). 535 */ 536 return; 537 } 538 539 if (!tp->packets_out) 540 return; 541 542 skb = tcp_rtx_queue_head(sk); 543 if (WARN_ON_ONCE(!skb)) 544 return; 545 546 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) && 547 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) { 548 /* Receiver dastardly shrinks window. Our retransmits 549 * become zero probes, but we should not timeout this 550 * connection. If the socket is an orphan, time it out, 551 * we cannot allow such beasts to hang infinitely. 552 */ 553 struct inet_sock *inet = inet_sk(sk); 554 u32 rtx_delta; 555 556 rtx_delta = tcp_time_stamp_ts(tp) - (tp->retrans_stamp ?: 557 tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb)); 558 if (tp->tcp_usec_ts) 559 rtx_delta /= USEC_PER_MSEC; 560 561 if (sk->sk_family == AF_INET) { 562 net_dbg_ratelimited("Probing zero-window on %pI4:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n", 563 &inet->inet_daddr, ntohs(inet->inet_dport), 564 inet->inet_num, tp->snd_una, tp->snd_nxt, 565 jiffies_to_msecs(jiffies - tp->rcv_tstamp), 566 rtx_delta); 567 } 568 #if IS_ENABLED(CONFIG_IPV6) 569 else if (sk->sk_family == AF_INET6) { 570 net_dbg_ratelimited("Probing zero-window on %pI6:%u/%u, seq=%u:%u, recv %ums ago, lasting %ums\n", 571 &sk->sk_v6_daddr, ntohs(inet->inet_dport), 572 inet->inet_num, tp->snd_una, tp->snd_nxt, 573 jiffies_to_msecs(jiffies - tp->rcv_tstamp), 574 rtx_delta); 575 } 576 #endif 577 if (tcp_rtx_probe0_timed_out(sk, skb, rtx_delta)) { 578 tcp_write_err(sk); 579 goto out; 580 } 581 tcp_enter_loss(sk); 582 tcp_retransmit_skb(sk, skb, 1); 583 __sk_dst_reset(sk); 584 goto out_reset_timer; 585 } 586 587 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEOUTS); 588 if (tcp_write_timeout(sk)) 589 goto out; 590 591 if (icsk->icsk_retransmits == 0) { 592 int mib_idx = 0; 593 594 if (icsk->icsk_ca_state == TCP_CA_Recovery) { 595 if (tcp_is_sack(tp)) 596 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL; 597 else 598 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL; 599 } else if (icsk->icsk_ca_state == TCP_CA_Loss) { 600 mib_idx = LINUX_MIB_TCPLOSSFAILURES; 601 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) || 602 tp->sacked_out) { 603 if (tcp_is_sack(tp)) 604 mib_idx = LINUX_MIB_TCPSACKFAILURES; 605 else 606 mib_idx = LINUX_MIB_TCPRENOFAILURES; 607 } 608 if (mib_idx) 609 __NET_INC_STATS(sock_net(sk), mib_idx); 610 } 611 612 tcp_enter_loss(sk); 613 614 tcp_update_rto_stats(sk); 615 if (tcp_retransmit_skb(sk, tcp_rtx_queue_head(sk), 1) > 0) { 616 /* Retransmission failed because of local congestion, 617 * Let senders fight for local resources conservatively. 618 */ 619 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 620 TCP_RESOURCE_PROBE_INTERVAL, 621 TCP_RTO_MAX); 622 goto out; 623 } 624 625 /* Increase the timeout each time we retransmit. Note that 626 * we do not increase the rtt estimate. rto is initialized 627 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests 628 * that doubling rto each time is the least we can get away with. 629 * In KA9Q, Karn uses this for the first few times, and then 630 * goes to quadratic. netBSD doubles, but only goes up to *64, 631 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is 632 * defined in the protocol as the maximum possible RTT. I guess 633 * we'll have to use something other than TCP to talk to the 634 * University of Mars. 635 * 636 * PAWS allows us longer timeouts and large windows, so once 637 * implemented ftp to mars will work nicely. We will have to fix 638 * the 120 second clamps though! 639 */ 640 641 out_reset_timer: 642 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is 643 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this 644 * might be increased if the stream oscillates between thin and thick, 645 * thus the old value might already be too high compared to the value 646 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without 647 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating 648 * exponential backoff behaviour to avoid continue hammering 649 * linear-timeout retransmissions into a black hole 650 */ 651 if (sk->sk_state == TCP_ESTABLISHED && 652 (tp->thin_lto || READ_ONCE(net->ipv4.sysctl_tcp_thin_linear_timeouts)) && 653 tcp_stream_is_thin(tp) && 654 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) { 655 icsk->icsk_backoff = 0; 656 icsk->icsk_rto = clamp(__tcp_set_rto(tp), 657 tcp_rto_min(sk), 658 TCP_RTO_MAX); 659 } else if (sk->sk_state != TCP_SYN_SENT || 660 tp->total_rto > 661 READ_ONCE(net->ipv4.sysctl_tcp_syn_linear_timeouts)) { 662 /* Use normal (exponential) backoff unless linear timeouts are 663 * activated. 664 */ 665 icsk->icsk_backoff++; 666 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX); 667 } 668 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 669 tcp_clamp_rto_to_user_timeout(sk), TCP_RTO_MAX); 670 if (retransmits_timed_out(sk, READ_ONCE(net->ipv4.sysctl_tcp_retries1) + 1, 0)) 671 __sk_dst_reset(sk); 672 673 out:; 674 } 675 676 /* Called with bottom-half processing disabled. 677 Called by tcp_write_timer() */ 678 void tcp_write_timer_handler(struct sock *sk) 679 { 680 struct inet_connection_sock *icsk = inet_csk(sk); 681 int event; 682 683 if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) || 684 !icsk->icsk_pending) 685 return; 686 687 if (time_after(icsk->icsk_timeout, jiffies)) { 688 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout); 689 return; 690 } 691 692 tcp_mstamp_refresh(tcp_sk(sk)); 693 event = icsk->icsk_pending; 694 695 switch (event) { 696 case ICSK_TIME_REO_TIMEOUT: 697 tcp_rack_reo_timeout(sk); 698 break; 699 case ICSK_TIME_LOSS_PROBE: 700 tcp_send_loss_probe(sk); 701 break; 702 case ICSK_TIME_RETRANS: 703 icsk->icsk_pending = 0; 704 tcp_retransmit_timer(sk); 705 break; 706 case ICSK_TIME_PROBE0: 707 icsk->icsk_pending = 0; 708 tcp_probe_timer(sk); 709 break; 710 } 711 } 712 713 static void tcp_write_timer(struct timer_list *t) 714 { 715 struct inet_connection_sock *icsk = 716 from_timer(icsk, t, icsk_retransmit_timer); 717 struct sock *sk = &icsk->icsk_inet.sk; 718 719 bh_lock_sock(sk); 720 if (!sock_owned_by_user(sk)) { 721 tcp_write_timer_handler(sk); 722 } else { 723 /* delegate our work to tcp_release_cb() */ 724 if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags)) 725 sock_hold(sk); 726 } 727 bh_unlock_sock(sk); 728 sock_put(sk); 729 } 730 731 void tcp_syn_ack_timeout(const struct request_sock *req) 732 { 733 struct net *net = read_pnet(&inet_rsk(req)->ireq_net); 734 735 __NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS); 736 } 737 EXPORT_SYMBOL(tcp_syn_ack_timeout); 738 739 void tcp_set_keepalive(struct sock *sk, int val) 740 { 741 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 742 return; 743 744 if (val && !sock_flag(sk, SOCK_KEEPOPEN)) 745 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk))); 746 else if (!val) 747 inet_csk_delete_keepalive_timer(sk); 748 } 749 EXPORT_SYMBOL_GPL(tcp_set_keepalive); 750 751 752 static void tcp_keepalive_timer (struct timer_list *t) 753 { 754 struct sock *sk = from_timer(sk, t, sk_timer); 755 struct inet_connection_sock *icsk = inet_csk(sk); 756 struct tcp_sock *tp = tcp_sk(sk); 757 u32 elapsed; 758 759 /* Only process if socket is not in use. */ 760 bh_lock_sock(sk); 761 if (sock_owned_by_user(sk)) { 762 /* Try again later. */ 763 inet_csk_reset_keepalive_timer (sk, HZ/20); 764 goto out; 765 } 766 767 if (sk->sk_state == TCP_LISTEN) { 768 pr_err("Hmm... keepalive on a LISTEN ???\n"); 769 goto out; 770 } 771 772 tcp_mstamp_refresh(tp); 773 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) { 774 if (READ_ONCE(tp->linger2) >= 0) { 775 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN; 776 777 if (tmo > 0) { 778 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 779 goto out; 780 } 781 } 782 tcp_send_active_reset(sk, GFP_ATOMIC, SK_RST_REASON_NOT_SPECIFIED); 783 goto death; 784 } 785 786 if (!sock_flag(sk, SOCK_KEEPOPEN) || 787 ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT))) 788 goto out; 789 790 elapsed = keepalive_time_when(tp); 791 792 /* It is alive without keepalive 8) */ 793 if (tp->packets_out || !tcp_write_queue_empty(sk)) 794 goto resched; 795 796 elapsed = keepalive_time_elapsed(tp); 797 798 if (elapsed >= keepalive_time_when(tp)) { 799 u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout); 800 801 /* If the TCP_USER_TIMEOUT option is enabled, use that 802 * to determine when to timeout instead. 803 */ 804 if ((user_timeout != 0 && 805 elapsed >= msecs_to_jiffies(user_timeout) && 806 icsk->icsk_probes_out > 0) || 807 (user_timeout == 0 && 808 icsk->icsk_probes_out >= keepalive_probes(tp))) { 809 tcp_send_active_reset(sk, GFP_ATOMIC, 810 SK_RST_REASON_NOT_SPECIFIED); 811 tcp_write_err(sk); 812 goto out; 813 } 814 if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) { 815 icsk->icsk_probes_out++; 816 elapsed = keepalive_intvl_when(tp); 817 } else { 818 /* If keepalive was lost due to local congestion, 819 * try harder. 820 */ 821 elapsed = TCP_RESOURCE_PROBE_INTERVAL; 822 } 823 } else { 824 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */ 825 elapsed = keepalive_time_when(tp) - elapsed; 826 } 827 828 resched: 829 inet_csk_reset_keepalive_timer (sk, elapsed); 830 goto out; 831 832 death: 833 tcp_done(sk); 834 835 out: 836 bh_unlock_sock(sk); 837 sock_put(sk); 838 } 839 840 static enum hrtimer_restart tcp_compressed_ack_kick(struct hrtimer *timer) 841 { 842 struct tcp_sock *tp = container_of(timer, struct tcp_sock, compressed_ack_timer); 843 struct sock *sk = (struct sock *)tp; 844 845 bh_lock_sock(sk); 846 if (!sock_owned_by_user(sk)) { 847 if (tp->compressed_ack) { 848 /* Since we have to send one ack finally, 849 * subtract one from tp->compressed_ack to keep 850 * LINUX_MIB_TCPACKCOMPRESSED accurate. 851 */ 852 tp->compressed_ack--; 853 tcp_send_ack(sk); 854 } 855 } else { 856 if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, 857 &sk->sk_tsq_flags)) 858 sock_hold(sk); 859 } 860 bh_unlock_sock(sk); 861 862 sock_put(sk); 863 864 return HRTIMER_NORESTART; 865 } 866 867 void tcp_init_xmit_timers(struct sock *sk) 868 { 869 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer, 870 &tcp_keepalive_timer); 871 hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC, 872 HRTIMER_MODE_ABS_PINNED_SOFT); 873 tcp_sk(sk)->pacing_timer.function = tcp_pace_kick; 874 875 hrtimer_init(&tcp_sk(sk)->compressed_ack_timer, CLOCK_MONOTONIC, 876 HRTIMER_MODE_REL_PINNED_SOFT); 877 tcp_sk(sk)->compressed_ack_timer.function = tcp_compressed_ack_kick; 878 } 879