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_too_many_orphans(sk, shift)) { 81 if (net_ratelimit()) 82 printk(KERN_INFO "Out of socket memory\n"); 83 84 /* Catch exceptional cases, when connection requires reset. 85 * 1. Last segment was sent recently. */ 86 if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN || 87 /* 2. Window is closed. */ 88 (!tp->snd_wnd && !tp->packets_out)) 89 do_reset = 1; 90 if (do_reset) 91 tcp_send_active_reset(sk, GFP_ATOMIC); 92 tcp_done(sk); 93 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY); 94 return 1; 95 } 96 return 0; 97 } 98 99 /* Calculate maximal number or retries on an orphaned socket. */ 100 static int tcp_orphan_retries(struct sock *sk, int alive) 101 { 102 int retries = sysctl_tcp_orphan_retries; /* May be zero. */ 103 104 /* We know from an ICMP that something is wrong. */ 105 if (sk->sk_err_soft && !alive) 106 retries = 0; 107 108 /* However, if socket sent something recently, select some safe 109 * number of retries. 8 corresponds to >100 seconds with minimal 110 * RTO of 200msec. */ 111 if (retries == 0 && alive) 112 retries = 8; 113 return retries; 114 } 115 116 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk) 117 { 118 /* Black hole detection */ 119 if (sysctl_tcp_mtu_probing) { 120 if (!icsk->icsk_mtup.enabled) { 121 icsk->icsk_mtup.enabled = 1; 122 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 123 } else { 124 struct tcp_sock *tp = tcp_sk(sk); 125 int mss; 126 127 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1; 128 mss = min(sysctl_tcp_base_mss, mss); 129 mss = max(mss, 68 - tp->tcp_header_len); 130 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 131 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 132 } 133 } 134 } 135 136 /* This function calculates a "timeout" which is equivalent to the timeout of a 137 * TCP connection after "boundary" unsuccessful, exponentially backed-off 138 * retransmissions with an initial RTO of TCP_RTO_MIN or TCP_TIMEOUT_INIT if 139 * syn_set flag is set. 140 */ 141 static bool retransmits_timed_out(struct sock *sk, 142 unsigned int boundary, 143 bool syn_set) 144 { 145 unsigned int timeout, linear_backoff_thresh; 146 unsigned int start_ts; 147 unsigned int rto_base = syn_set ? TCP_TIMEOUT_INIT : TCP_RTO_MIN; 148 149 if (!inet_csk(sk)->icsk_retransmits) 150 return false; 151 152 if (unlikely(!tcp_sk(sk)->retrans_stamp)) 153 start_ts = TCP_SKB_CB(tcp_write_queue_head(sk))->when; 154 else 155 start_ts = tcp_sk(sk)->retrans_stamp; 156 157 linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base); 158 159 if (boundary <= linear_backoff_thresh) 160 timeout = ((2 << boundary) - 1) * rto_base; 161 else 162 timeout = ((2 << linear_backoff_thresh) - 1) * rto_base + 163 (boundary - linear_backoff_thresh) * TCP_RTO_MAX; 164 165 return (tcp_time_stamp - start_ts) >= timeout; 166 } 167 168 /* A write timeout has occurred. Process the after effects. */ 169 static int tcp_write_timeout(struct sock *sk) 170 { 171 struct inet_connection_sock *icsk = inet_csk(sk); 172 int retry_until; 173 bool do_reset, syn_set = 0; 174 175 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) { 176 if (icsk->icsk_retransmits) 177 dst_negative_advice(sk); 178 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 179 syn_set = 1; 180 } else { 181 if (retransmits_timed_out(sk, sysctl_tcp_retries1, 0)) { 182 /* Black hole detection */ 183 tcp_mtu_probing(icsk, sk); 184 185 dst_negative_advice(sk); 186 } 187 188 retry_until = sysctl_tcp_retries2; 189 if (sock_flag(sk, SOCK_DEAD)) { 190 const int alive = (icsk->icsk_rto < TCP_RTO_MAX); 191 192 retry_until = tcp_orphan_retries(sk, alive); 193 do_reset = alive || 194 !retransmits_timed_out(sk, retry_until, 0); 195 196 if (tcp_out_of_resources(sk, do_reset)) 197 return 1; 198 } 199 } 200 201 if (retransmits_timed_out(sk, retry_until, syn_set)) { 202 /* Has it gone just too far? */ 203 tcp_write_err(sk); 204 return 1; 205 } 206 return 0; 207 } 208 209 static void tcp_delack_timer(unsigned long data) 210 { 211 struct sock *sk = (struct sock *)data; 212 struct tcp_sock *tp = tcp_sk(sk); 213 struct inet_connection_sock *icsk = inet_csk(sk); 214 215 bh_lock_sock(sk); 216 if (sock_owned_by_user(sk)) { 217 /* Try again later. */ 218 icsk->icsk_ack.blocked = 1; 219 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED); 220 sk_reset_timer(sk, &icsk->icsk_delack_timer, jiffies + TCP_DELACK_MIN); 221 goto out_unlock; 222 } 223 224 sk_mem_reclaim_partial(sk); 225 226 if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER)) 227 goto out; 228 229 if (time_after(icsk->icsk_ack.timeout, jiffies)) { 230 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout); 231 goto out; 232 } 233 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER; 234 235 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 236 struct sk_buff *skb; 237 238 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED); 239 240 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 241 sk_backlog_rcv(sk, skb); 242 243 tp->ucopy.memory = 0; 244 } 245 246 if (inet_csk_ack_scheduled(sk)) { 247 if (!icsk->icsk_ack.pingpong) { 248 /* Delayed ACK missed: inflate ATO. */ 249 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto); 250 } else { 251 /* Delayed ACK missed: leave pingpong mode and 252 * deflate ATO. 253 */ 254 icsk->icsk_ack.pingpong = 0; 255 icsk->icsk_ack.ato = TCP_ATO_MIN; 256 } 257 tcp_send_ack(sk); 258 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS); 259 } 260 TCP_CHECK_TIMER(sk); 261 262 out: 263 if (tcp_memory_pressure) 264 sk_mem_reclaim(sk); 265 out_unlock: 266 bh_unlock_sock(sk); 267 sock_put(sk); 268 } 269 270 static void tcp_probe_timer(struct sock *sk) 271 { 272 struct inet_connection_sock *icsk = inet_csk(sk); 273 struct tcp_sock *tp = tcp_sk(sk); 274 int max_probes; 275 276 if (tp->packets_out || !tcp_send_head(sk)) { 277 icsk->icsk_probes_out = 0; 278 return; 279 } 280 281 /* *WARNING* RFC 1122 forbids this 282 * 283 * It doesn't AFAIK, because we kill the retransmit timer -AK 284 * 285 * FIXME: We ought not to do it, Solaris 2.5 actually has fixing 286 * this behaviour in Solaris down as a bug fix. [AC] 287 * 288 * Let me to explain. icsk_probes_out is zeroed by incoming ACKs 289 * even if they advertise zero window. Hence, connection is killed only 290 * if we received no ACKs for normal connection timeout. It is not killed 291 * only because window stays zero for some time, window may be zero 292 * until armageddon and even later. We are in full accordance 293 * with RFCs, only probe timer combines both retransmission timeout 294 * and probe timeout in one bottle. --ANK 295 */ 296 max_probes = sysctl_tcp_retries2; 297 298 if (sock_flag(sk, SOCK_DEAD)) { 299 const int alive = ((icsk->icsk_rto << icsk->icsk_backoff) < TCP_RTO_MAX); 300 301 max_probes = tcp_orphan_retries(sk, alive); 302 303 if (tcp_out_of_resources(sk, alive || icsk->icsk_probes_out <= max_probes)) 304 return; 305 } 306 307 if (icsk->icsk_probes_out > max_probes) { 308 tcp_write_err(sk); 309 } else { 310 /* Only send another probe if we didn't close things up. */ 311 tcp_send_probe0(sk); 312 } 313 } 314 315 /* 316 * The TCP retransmit timer. 317 */ 318 319 void tcp_retransmit_timer(struct sock *sk) 320 { 321 struct tcp_sock *tp = tcp_sk(sk); 322 struct inet_connection_sock *icsk = inet_csk(sk); 323 324 if (!tp->packets_out) 325 goto out; 326 327 WARN_ON(tcp_write_queue_empty(sk)); 328 329 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) && 330 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) { 331 /* Receiver dastardly shrinks window. Our retransmits 332 * become zero probes, but we should not timeout this 333 * connection. If the socket is an orphan, time it out, 334 * we cannot allow such beasts to hang infinitely. 335 */ 336 #ifdef TCP_DEBUG 337 struct inet_sock *inet = inet_sk(sk); 338 if (sk->sk_family == AF_INET) { 339 LIMIT_NETDEBUG(KERN_DEBUG "TCP: Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 340 &inet->inet_daddr, ntohs(inet->inet_dport), 341 inet->inet_num, tp->snd_una, tp->snd_nxt); 342 } 343 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) 344 else if (sk->sk_family == AF_INET6) { 345 struct ipv6_pinfo *np = inet6_sk(sk); 346 LIMIT_NETDEBUG(KERN_DEBUG "TCP: Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 347 &np->daddr, ntohs(inet->inet_dport), 348 inet->inet_num, tp->snd_una, tp->snd_nxt); 349 } 350 #endif 351 #endif 352 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) { 353 tcp_write_err(sk); 354 goto out; 355 } 356 tcp_enter_loss(sk, 0); 357 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 358 __sk_dst_reset(sk); 359 goto out_reset_timer; 360 } 361 362 if (tcp_write_timeout(sk)) 363 goto out; 364 365 if (icsk->icsk_retransmits == 0) { 366 int mib_idx; 367 368 if (icsk->icsk_ca_state == TCP_CA_Disorder) { 369 if (tcp_is_sack(tp)) 370 mib_idx = LINUX_MIB_TCPSACKFAILURES; 371 else 372 mib_idx = LINUX_MIB_TCPRENOFAILURES; 373 } else if (icsk->icsk_ca_state == TCP_CA_Recovery) { 374 if (tcp_is_sack(tp)) 375 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL; 376 else 377 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL; 378 } else if (icsk->icsk_ca_state == TCP_CA_Loss) { 379 mib_idx = LINUX_MIB_TCPLOSSFAILURES; 380 } else { 381 mib_idx = LINUX_MIB_TCPTIMEOUTS; 382 } 383 NET_INC_STATS_BH(sock_net(sk), mib_idx); 384 } 385 386 if (tcp_use_frto(sk)) { 387 tcp_enter_frto(sk); 388 } else { 389 tcp_enter_loss(sk, 0); 390 } 391 392 if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) { 393 /* Retransmission failed because of local congestion, 394 * do not backoff. 395 */ 396 if (!icsk->icsk_retransmits) 397 icsk->icsk_retransmits = 1; 398 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 399 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL), 400 TCP_RTO_MAX); 401 goto out; 402 } 403 404 /* Increase the timeout each time we retransmit. Note that 405 * we do not increase the rtt estimate. rto is initialized 406 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests 407 * that doubling rto each time is the least we can get away with. 408 * In KA9Q, Karn uses this for the first few times, and then 409 * goes to quadratic. netBSD doubles, but only goes up to *64, 410 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is 411 * defined in the protocol as the maximum possible RTT. I guess 412 * we'll have to use something other than TCP to talk to the 413 * University of Mars. 414 * 415 * PAWS allows us longer timeouts and large windows, so once 416 * implemented ftp to mars will work nicely. We will have to fix 417 * the 120 second clamps though! 418 */ 419 icsk->icsk_backoff++; 420 icsk->icsk_retransmits++; 421 422 out_reset_timer: 423 /* If stream is thin, use linear timeouts. Since 'icsk_backoff' is 424 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this 425 * might be increased if the stream oscillates between thin and thick, 426 * thus the old value might already be too high compared to the value 427 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without 428 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating 429 * exponential backoff behaviour to avoid continue hammering 430 * linear-timeout retransmissions into a black hole 431 */ 432 if (sk->sk_state == TCP_ESTABLISHED && 433 (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) && 434 tcp_stream_is_thin(tp) && 435 icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) { 436 icsk->icsk_backoff = 0; 437 icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX); 438 } else { 439 /* Use normal (exponential) backoff */ 440 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX); 441 } 442 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX); 443 if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1, 0)) 444 __sk_dst_reset(sk); 445 446 out:; 447 } 448 449 static void tcp_write_timer(unsigned long data) 450 { 451 struct sock *sk = (struct sock *)data; 452 struct inet_connection_sock *icsk = inet_csk(sk); 453 int event; 454 455 bh_lock_sock(sk); 456 if (sock_owned_by_user(sk)) { 457 /* Try again later */ 458 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + (HZ / 20)); 459 goto out_unlock; 460 } 461 462 if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending) 463 goto out; 464 465 if (time_after(icsk->icsk_timeout, jiffies)) { 466 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout); 467 goto out; 468 } 469 470 event = icsk->icsk_pending; 471 icsk->icsk_pending = 0; 472 473 switch (event) { 474 case ICSK_TIME_RETRANS: 475 tcp_retransmit_timer(sk); 476 break; 477 case ICSK_TIME_PROBE0: 478 tcp_probe_timer(sk); 479 break; 480 } 481 TCP_CHECK_TIMER(sk); 482 483 out: 484 sk_mem_reclaim(sk); 485 out_unlock: 486 bh_unlock_sock(sk); 487 sock_put(sk); 488 } 489 490 /* 491 * Timer for listening sockets 492 */ 493 494 static void tcp_synack_timer(struct sock *sk) 495 { 496 inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL, 497 TCP_TIMEOUT_INIT, TCP_RTO_MAX); 498 } 499 500 void tcp_syn_ack_timeout(struct sock *sk, struct request_sock *req) 501 { 502 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEOUTS); 503 } 504 EXPORT_SYMBOL(tcp_syn_ack_timeout); 505 506 void tcp_set_keepalive(struct sock *sk, int val) 507 { 508 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 509 return; 510 511 if (val && !sock_flag(sk, SOCK_KEEPOPEN)) 512 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk))); 513 else if (!val) 514 inet_csk_delete_keepalive_timer(sk); 515 } 516 517 518 static void tcp_keepalive_timer (unsigned long data) 519 { 520 struct sock *sk = (struct sock *) data; 521 struct inet_connection_sock *icsk = inet_csk(sk); 522 struct tcp_sock *tp = tcp_sk(sk); 523 u32 elapsed; 524 525 /* Only process if socket is not in use. */ 526 bh_lock_sock(sk); 527 if (sock_owned_by_user(sk)) { 528 /* Try again later. */ 529 inet_csk_reset_keepalive_timer (sk, HZ/20); 530 goto out; 531 } 532 533 if (sk->sk_state == TCP_LISTEN) { 534 tcp_synack_timer(sk); 535 goto out; 536 } 537 538 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) { 539 if (tp->linger2 >= 0) { 540 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN; 541 542 if (tmo > 0) { 543 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 544 goto out; 545 } 546 } 547 tcp_send_active_reset(sk, GFP_ATOMIC); 548 goto death; 549 } 550 551 if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE) 552 goto out; 553 554 elapsed = keepalive_time_when(tp); 555 556 /* It is alive without keepalive 8) */ 557 if (tp->packets_out || tcp_send_head(sk)) 558 goto resched; 559 560 elapsed = keepalive_time_elapsed(tp); 561 562 if (elapsed >= keepalive_time_when(tp)) { 563 if (icsk->icsk_probes_out >= keepalive_probes(tp)) { 564 tcp_send_active_reset(sk, GFP_ATOMIC); 565 tcp_write_err(sk); 566 goto out; 567 } 568 if (tcp_write_wakeup(sk) <= 0) { 569 icsk->icsk_probes_out++; 570 elapsed = keepalive_intvl_when(tp); 571 } else { 572 /* If keepalive was lost due to local congestion, 573 * try harder. 574 */ 575 elapsed = TCP_RESOURCE_PROBE_INTERVAL; 576 } 577 } else { 578 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */ 579 elapsed = keepalive_time_when(tp) - elapsed; 580 } 581 582 TCP_CHECK_TIMER(sk); 583 sk_mem_reclaim(sk); 584 585 resched: 586 inet_csk_reset_keepalive_timer (sk, elapsed); 587 goto out; 588 589 death: 590 tcp_done(sk); 591 592 out: 593 bh_unlock_sock(sk); 594 sock_put(sk); 595 } 596