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 int sysctl_tcp_syn_retries __read_mostly = TCP_SYN_RETRIES; 26 int sysctl_tcp_synack_retries __read_mostly = TCP_SYNACK_RETRIES; 27 int sysctl_tcp_keepalive_time __read_mostly = TCP_KEEPALIVE_TIME; 28 int sysctl_tcp_keepalive_probes __read_mostly = TCP_KEEPALIVE_PROBES; 29 int sysctl_tcp_keepalive_intvl __read_mostly = TCP_KEEPALIVE_INTVL; 30 int sysctl_tcp_retries1 __read_mostly = TCP_RETR1; 31 int sysctl_tcp_retries2 __read_mostly = TCP_RETR2; 32 int sysctl_tcp_orphan_retries __read_mostly; 33 int sysctl_tcp_thin_linear_timeouts __read_mostly; 34 35 static void tcp_write_timer(unsigned long); 36 static void tcp_delack_timer(unsigned long); 37 static void tcp_keepalive_timer (unsigned long data); 38 39 void tcp_init_xmit_timers(struct sock *sk) 40 { 41 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer, 42 &tcp_keepalive_timer); 43 } 44 EXPORT_SYMBOL(tcp_init_xmit_timers); 45 46 static void tcp_write_err(struct sock *sk) 47 { 48 sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT; 49 sk->sk_error_report(sk); 50 51 tcp_done(sk); 52 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT); 53 } 54 55 /* Do not allow orphaned sockets to eat all our resources. 56 * This is direct violation of TCP specs, but it is required 57 * to prevent DoS attacks. It is called when a retransmission timeout 58 * or zero probe timeout occurs on orphaned socket. 59 * 60 * Criteria is still not confirmed experimentally and may change. 61 * We kill the socket, if: 62 * 1. If number of orphaned sockets exceeds an administratively configured 63 * limit. 64 * 2. If we have strong memory pressure. 65 */ 66 static int tcp_out_of_resources(struct sock *sk, int do_reset) 67 { 68 struct tcp_sock *tp = tcp_sk(sk); 69 int shift = 0; 70 71 /* If peer does not open window for long time, or did not transmit 72 * anything for long time, penalize it. */ 73 if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset) 74 shift++; 75 76 /* If some dubious ICMP arrived, penalize even more. */ 77 if (sk->sk_err_soft) 78 shift++; 79 80 if (tcp_check_oom(sk, shift)) { 81 /* Catch exceptional cases, when connection requires reset. 82 * 1. Last segment was sent recently. */ 83 if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN || 84 /* 2. Window is closed. */ 85 (!tp->snd_wnd && !tp->packets_out)) 86 do_reset = 1; 87 if (do_reset) 88 tcp_send_active_reset(sk, GFP_ATOMIC); 89 tcp_done(sk); 90 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY); 91 return 1; 92 } 93 return 0; 94 } 95 96 /* Calculate maximal number or retries on an orphaned socket. */ 97 static int tcp_orphan_retries(struct sock *sk, int alive) 98 { 99 int retries = sysctl_tcp_orphan_retries; /* May be zero. */ 100 101 /* We know from an ICMP that something is wrong. */ 102 if (sk->sk_err_soft && !alive) 103 retries = 0; 104 105 /* However, if socket sent something recently, select some safe 106 * number of retries. 8 corresponds to >100 seconds with minimal 107 * RTO of 200msec. */ 108 if (retries == 0 && alive) 109 retries = 8; 110 return retries; 111 } 112 113 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk) 114 { 115 /* Black hole detection */ 116 if (sysctl_tcp_mtu_probing) { 117 if (!icsk->icsk_mtup.enabled) { 118 icsk->icsk_mtup.enabled = 1; 119 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 120 } else { 121 struct tcp_sock *tp = tcp_sk(sk); 122 int mss; 123 124 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1; 125 mss = min(sysctl_tcp_base_mss, mss); 126 mss = max(mss, 68 - tp->tcp_header_len); 127 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 128 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 129 } 130 } 131 } 132 133 /* This function calculates a "timeout" which is equivalent to the timeout of a 134 * TCP connection after "boundary" unsuccessful, exponentially backed-off 135 * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if 136 * syn_set flag is set. 137 */ 138 static bool retransmits_timed_out(struct sock *sk, 139 unsigned int boundary, 140 unsigned int timeout, 141 bool syn_set) 142 { 143 unsigned int linear_backoff_thresh, start_ts; 144 unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN; 145 146 if (!inet_csk(sk)->icsk_retransmits) 147 return false; 148 149 if (unlikely(!tcp_sk(sk)->retrans_stamp)) 150 start_ts = TCP_SKB_CB(tcp_write_queue_head(sk))->when; 151 else 152 start_ts = tcp_sk(sk)->retrans_stamp; 153 154 if (likely(timeout == 0)) { 155 linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base); 156 157 if (boundary <= linear_backoff_thresh) 158 timeout = ((2 << boundary) - 1) * rto_base; 159 else 160 timeout = ((2 << linear_backoff_thresh) - 1) * rto_base + 161 (boundary - linear_backoff_thresh) * TCP_RTO_MAX; 162 } 163 return (tcp_time_stamp - start_ts) >= timeout; 164 } 165 166 /* A write timeout has occurred. Process the after effects. */ 167 static int tcp_write_timeout(struct sock *sk) 168 { 169 struct inet_connection_sock *icsk = inet_csk(sk); 170 int retry_until; 171 bool do_reset, syn_set = false; 172 173 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) { 174 if (icsk->icsk_retransmits) 175 dst_negative_advice(sk); 176 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 177 syn_set = true; 178 } else { 179 if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0, 0)) { 180 /* Black hole detection */ 181 tcp_mtu_probing(icsk, sk); 182 183 dst_negative_advice(sk); 184 } 185 186 retry_until = sysctl_tcp_retries2; 187 if (sock_flag(sk, SOCK_DEAD)) { 188 const int alive = (icsk->icsk_rto < TCP_RTO_MAX); 189 190 retry_until = tcp_orphan_retries(sk, alive); 191 do_reset = alive || 192 !retransmits_timed_out(sk, retry_until, 0, 0); 193 194 if (tcp_out_of_resources(sk, do_reset)) 195 return 1; 196 } 197 } 198 199 if (retransmits_timed_out(sk, retry_until, 200 syn_set ? 0 : icsk->icsk_user_timeout, syn_set)) { 201 /* Has it gone just too far? */ 202 tcp_write_err(sk); 203 return 1; 204 } 205 return 0; 206 } 207 208 static void tcp_delack_timer(unsigned long data) 209 { 210 struct sock *sk = (struct sock *)data; 211 struct tcp_sock *tp = tcp_sk(sk); 212 struct inet_connection_sock *icsk = inet_csk(sk); 213 214 bh_lock_sock(sk); 215 if (sock_owned_by_user(sk)) { 216 /* Try again later. */ 217 icsk->icsk_ack.blocked = 1; 218 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED); 219 sk_reset_timer(sk, &icsk->icsk_delack_timer, jiffies + TCP_DELACK_MIN); 220 goto out_unlock; 221 } 222 223 sk_mem_reclaim_partial(sk); 224 225 if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER)) 226 goto out; 227 228 if (time_after(icsk->icsk_ack.timeout, jiffies)) { 229 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout); 230 goto out; 231 } 232 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER; 233 234 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 235 struct sk_buff *skb; 236 237 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED); 238 239 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 240 sk_backlog_rcv(sk, skb); 241 242 tp->ucopy.memory = 0; 243 } 244 245 if (inet_csk_ack_scheduled(sk)) { 246 if (!icsk->icsk_ack.pingpong) { 247 /* Delayed ACK missed: inflate ATO. */ 248 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto); 249 } else { 250 /* Delayed ACK missed: leave pingpong mode and 251 * deflate ATO. 252 */ 253 icsk->icsk_ack.pingpong = 0; 254 icsk->icsk_ack.ato = TCP_ATO_MIN; 255 } 256 tcp_send_ack(sk); 257 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS); 258 } 259 260 out: 261 if (sk_under_memory_pressure(sk)) 262 sk_mem_reclaim(sk); 263 out_unlock: 264 bh_unlock_sock(sk); 265 sock_put(sk); 266 } 267 268 static void tcp_probe_timer(struct sock *sk) 269 { 270 struct inet_connection_sock *icsk = inet_csk(sk); 271 struct tcp_sock *tp = tcp_sk(sk); 272 int max_probes; 273 274 if (tp->packets_out || !tcp_send_head(sk)) { 275 icsk->icsk_probes_out = 0; 276 return; 277 } 278 279 /* *WARNING* RFC 1122 forbids this 280 * 281 * It doesn't AFAIK, because we kill the retransmit timer -AK 282 * 283 * FIXME: We ought not to do it, Solaris 2.5 actually has fixing 284 * this behaviour in Solaris down as a bug fix. [AC] 285 * 286 * Let me to explain. icsk_probes_out is zeroed by incoming ACKs 287 * even if they advertise zero window. Hence, connection is killed only 288 * if we received no ACKs for normal connection timeout. It is not killed 289 * only because window stays zero for some time, window may be zero 290 * until armageddon and even later. We are in full accordance 291 * with RFCs, only probe timer combines both retransmission timeout 292 * and probe timeout in one bottle. --ANK 293 */ 294 max_probes = sysctl_tcp_retries2; 295 296 if (sock_flag(sk, SOCK_DEAD)) { 297 const int alive = ((icsk->icsk_rto << icsk->icsk_backoff) < TCP_RTO_MAX); 298 299 max_probes = tcp_orphan_retries(sk, alive); 300 301 if (tcp_out_of_resources(sk, alive || icsk->icsk_probes_out <= max_probes)) 302 return; 303 } 304 305 if (icsk->icsk_probes_out > max_probes) { 306 tcp_write_err(sk); 307 } else { 308 /* Only send another probe if we didn't close things up. */ 309 tcp_send_probe0(sk); 310 } 311 } 312 313 /* 314 * The TCP retransmit timer. 315 */ 316 317 void tcp_retransmit_timer(struct sock *sk) 318 { 319 struct tcp_sock *tp = tcp_sk(sk); 320 struct inet_connection_sock *icsk = inet_csk(sk); 321 322 if (tp->early_retrans_delayed) { 323 tcp_resume_early_retransmit(sk); 324 return; 325 } 326 327 if (!tp->packets_out) 328 goto out; 329 330 WARN_ON(tcp_write_queue_empty(sk)); 331 332 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) && 333 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) { 334 /* Receiver dastardly shrinks window. Our retransmits 335 * become zero probes, but we should not timeout this 336 * connection. If the socket is an orphan, time it out, 337 * we cannot allow such beasts to hang infinitely. 338 */ 339 struct inet_sock *inet = inet_sk(sk); 340 if (sk->sk_family == AF_INET) { 341 LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n"), 342 &inet->inet_daddr, 343 ntohs(inet->inet_dport), inet->inet_num, 344 tp->snd_una, tp->snd_nxt); 345 } 346 #if IS_ENABLED(CONFIG_IPV6) 347 else if (sk->sk_family == AF_INET6) { 348 struct ipv6_pinfo *np = inet6_sk(sk); 349 LIMIT_NETDEBUG(KERN_DEBUG pr_fmt("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n"), 350 &np->daddr, 351 ntohs(inet->inet_dport), inet->inet_num, 352 tp->snd_una, tp->snd_nxt); 353 } 354 #endif 355 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) { 356 tcp_write_err(sk); 357 goto out; 358 } 359 tcp_enter_loss(sk, 0); 360 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 361 __sk_dst_reset(sk); 362 goto out_reset_timer; 363 } 364 365 if (tcp_write_timeout(sk)) 366 goto out; 367 368 if (icsk->icsk_retransmits == 0) { 369 int mib_idx; 370 371 if (icsk->icsk_ca_state == TCP_CA_Recovery) { 372 if (tcp_is_sack(tp)) 373 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL; 374 else 375 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL; 376 } else if (icsk->icsk_ca_state == TCP_CA_Loss) { 377 mib_idx = LINUX_MIB_TCPLOSSFAILURES; 378 } else if ((icsk->icsk_ca_state == TCP_CA_Disorder) || 379 tp->sacked_out) { 380 if (tcp_is_sack(tp)) 381 mib_idx = LINUX_MIB_TCPSACKFAILURES; 382 else 383 mib_idx = LINUX_MIB_TCPRENOFAILURES; 384 } else { 385 mib_idx = LINUX_MIB_TCPTIMEOUTS; 386 } 387 NET_INC_STATS_BH(sock_net(sk), mib_idx); 388 } 389 390 if (tcp_use_frto(sk)) { 391 tcp_enter_frto(sk); 392 } else { 393 tcp_enter_loss(sk, 0); 394 } 395 396 if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) { 397 /* Retransmission failed because of local congestion, 398 * do not backoff. 399 */ 400 if (!icsk->icsk_retransmits) 401 icsk->icsk_retransmits = 1; 402 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 403 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL), 404 TCP_RTO_MAX); 405 goto out; 406 } 407 408 /* Increase the timeout each time we retransmit. Note that 409 * we do not increase the rtt estimate. rto is initialized 410 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests 411 * that doubling rto each time is the least we can get away with. 412 * In KA9Q, Karn uses this for the first few times, and then 413 * goes to quadratic. netBSD doubles, but only goes up to *64, 414 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is 415 * defined in the protocol as the maximum possible RTT. I guess 416 * we'll have to use something other than TCP to talk to the 417 * University of Mars. 418 * 419 * PAWS allows us longer timeouts and large windows, so once 420 * implemented ftp to mars will work nicely. We will have to fix 421 * the 120 second clamps though! 422 */ 423 icsk->icsk_backoff++; 424 icsk->icsk_retransmits++; 425 426 out_reset_timer: 427 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is 428 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this 429 * might be increased if the stream oscillates between thin and thick, 430 * thus the old value might already be too high compared to the value 431 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without 432 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating 433 * exponential backoff behaviour to avoid continue hammering 434 * linear-timeout retransmissions into a black hole 435 */ 436 if (sk->sk_state == TCP_ESTABLISHED && 437 (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) && 438 tcp_stream_is_thin(tp) && 439 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) { 440 icsk->icsk_backoff = 0; 441 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX); 442 } else { 443 /* Use normal (exponential) backoff */ 444 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX); 445 } 446 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX); 447 if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0, 0)) 448 __sk_dst_reset(sk); 449 450 out:; 451 } 452 453 static void tcp_write_timer(unsigned long data) 454 { 455 struct sock *sk = (struct sock *)data; 456 struct inet_connection_sock *icsk = inet_csk(sk); 457 int event; 458 459 bh_lock_sock(sk); 460 if (sock_owned_by_user(sk)) { 461 /* Try again later */ 462 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + (HZ / 20)); 463 goto out_unlock; 464 } 465 466 if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending) 467 goto out; 468 469 if (time_after(icsk->icsk_timeout, jiffies)) { 470 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout); 471 goto out; 472 } 473 474 event = icsk->icsk_pending; 475 icsk->icsk_pending = 0; 476 477 switch (event) { 478 case ICSK_TIME_RETRANS: 479 tcp_retransmit_timer(sk); 480 break; 481 case ICSK_TIME_PROBE0: 482 tcp_probe_timer(sk); 483 break; 484 } 485 486 out: 487 sk_mem_reclaim(sk); 488 out_unlock: 489 bh_unlock_sock(sk); 490 sock_put(sk); 491 } 492 493 /* 494 * Timer for listening sockets 495 */ 496 497 static void tcp_synack_timer(struct sock *sk) 498 { 499 inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL, 500 TCP_TIMEOUT_INIT, TCP_RTO_MAX); 501 } 502 503 void tcp_syn_ack_timeout(struct sock *sk, struct request_sock *req) 504 { 505 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEOUTS); 506 } 507 EXPORT_SYMBOL(tcp_syn_ack_timeout); 508 509 void tcp_set_keepalive(struct sock *sk, int val) 510 { 511 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 512 return; 513 514 if (val && !sock_flag(sk, SOCK_KEEPOPEN)) 515 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk))); 516 else if (!val) 517 inet_csk_delete_keepalive_timer(sk); 518 } 519 520 521 static void tcp_keepalive_timer (unsigned long data) 522 { 523 struct sock *sk = (struct sock *) data; 524 struct inet_connection_sock *icsk = inet_csk(sk); 525 struct tcp_sock *tp = tcp_sk(sk); 526 u32 elapsed; 527 528 /* Only process if socket is not in use. */ 529 bh_lock_sock(sk); 530 if (sock_owned_by_user(sk)) { 531 /* Try again later. */ 532 inet_csk_reset_keepalive_timer (sk, HZ/20); 533 goto out; 534 } 535 536 if (sk->sk_state == TCP_LISTEN) { 537 tcp_synack_timer(sk); 538 goto out; 539 } 540 541 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) { 542 if (tp->linger2 >= 0) { 543 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN; 544 545 if (tmo > 0) { 546 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 547 goto out; 548 } 549 } 550 tcp_send_active_reset(sk, GFP_ATOMIC); 551 goto death; 552 } 553 554 if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE) 555 goto out; 556 557 elapsed = keepalive_time_when(tp); 558 559 /* It is alive without keepalive 8) */ 560 if (tp->packets_out || tcp_send_head(sk)) 561 goto resched; 562 563 elapsed = keepalive_time_elapsed(tp); 564 565 if (elapsed >= keepalive_time_when(tp)) { 566 /* If the TCP_USER_TIMEOUT option is enabled, use that 567 * to determine when to timeout instead. 568 */ 569 if ((icsk->icsk_user_timeout != 0 && 570 elapsed >= icsk->icsk_user_timeout && 571 icsk->icsk_probes_out > 0) || 572 (icsk->icsk_user_timeout == 0 && 573 icsk->icsk_probes_out >= keepalive_probes(tp))) { 574 tcp_send_active_reset(sk, GFP_ATOMIC); 575 tcp_write_err(sk); 576 goto out; 577 } 578 if (tcp_write_wakeup(sk) <= 0) { 579 icsk->icsk_probes_out++; 580 elapsed = keepalive_intvl_when(tp); 581 } else { 582 /* If keepalive was lost due to local congestion, 583 * try harder. 584 */ 585 elapsed = TCP_RESOURCE_PROBE_INTERVAL; 586 } 587 } else { 588 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */ 589 elapsed = keepalive_time_when(tp) - elapsed; 590 } 591 592 sk_mem_reclaim(sk); 593 594 resched: 595 inet_csk_reset_keepalive_timer (sk, elapsed); 596 goto out; 597 598 death: 599 tcp_done(sk); 600 601 out: 602 bh_unlock_sock(sk); 603 sock_put(sk); 604 } 605