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 <net/tcp.h> 23 24 int sysctl_tcp_syn_retries __read_mostly = TCP_SYN_RETRIES; 25 int sysctl_tcp_synack_retries __read_mostly = TCP_SYNACK_RETRIES; 26 int sysctl_tcp_keepalive_time __read_mostly = TCP_KEEPALIVE_TIME; 27 int sysctl_tcp_keepalive_probes __read_mostly = TCP_KEEPALIVE_PROBES; 28 int sysctl_tcp_keepalive_intvl __read_mostly = TCP_KEEPALIVE_INTVL; 29 int sysctl_tcp_retries1 __read_mostly = TCP_RETR1; 30 int sysctl_tcp_retries2 __read_mostly = TCP_RETR2; 31 int sysctl_tcp_orphan_retries __read_mostly; 32 33 static void tcp_write_timer(unsigned long); 34 static void tcp_delack_timer(unsigned long); 35 static void tcp_keepalive_timer (unsigned long data); 36 37 void tcp_init_xmit_timers(struct sock *sk) 38 { 39 inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer, 40 &tcp_keepalive_timer); 41 } 42 43 EXPORT_SYMBOL(tcp_init_xmit_timers); 44 45 static void tcp_write_err(struct sock *sk) 46 { 47 sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT; 48 sk->sk_error_report(sk); 49 50 tcp_done(sk); 51 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT); 52 } 53 54 /* Do not allow orphaned sockets to eat all our resources. 55 * This is direct violation of TCP specs, but it is required 56 * to prevent DoS attacks. It is called when a retransmission timeout 57 * or zero probe timeout occurs on orphaned socket. 58 * 59 * Criteria is still not confirmed experimentally and may change. 60 * We kill the socket, if: 61 * 1. If number of orphaned sockets exceeds an administratively configured 62 * limit. 63 * 2. If we have strong memory pressure. 64 */ 65 static int tcp_out_of_resources(struct sock *sk, int do_reset) 66 { 67 struct tcp_sock *tp = tcp_sk(sk); 68 int orphans = percpu_counter_read_positive(&tcp_orphan_count); 69 70 /* If peer does not open window for long time, or did not transmit 71 * anything for long time, penalize it. */ 72 if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset) 73 orphans <<= 1; 74 75 /* If some dubious ICMP arrived, penalize even more. */ 76 if (sk->sk_err_soft) 77 orphans <<= 1; 78 79 if (tcp_too_many_orphans(sk, orphans)) { 80 if (net_ratelimit()) 81 printk(KERN_INFO "Out of socket memory\n"); 82 83 /* Catch exceptional cases, when connection requires reset. 84 * 1. Last segment was sent recently. */ 85 if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN || 86 /* 2. Window is closed. */ 87 (!tp->snd_wnd && !tp->packets_out)) 88 do_reset = 1; 89 if (do_reset) 90 tcp_send_active_reset(sk, GFP_ATOMIC); 91 tcp_done(sk); 92 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY); 93 return 1; 94 } 95 return 0; 96 } 97 98 /* Calculate maximal number or retries on an orphaned socket. */ 99 static int tcp_orphan_retries(struct sock *sk, int alive) 100 { 101 int retries = sysctl_tcp_orphan_retries; /* May be zero. */ 102 103 /* We know from an ICMP that something is wrong. */ 104 if (sk->sk_err_soft && !alive) 105 retries = 0; 106 107 /* However, if socket sent something recently, select some safe 108 * number of retries. 8 corresponds to >100 seconds with minimal 109 * RTO of 200msec. */ 110 if (retries == 0 && alive) 111 retries = 8; 112 return retries; 113 } 114 115 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk) 116 { 117 /* Black hole detection */ 118 if (sysctl_tcp_mtu_probing) { 119 if (!icsk->icsk_mtup.enabled) { 120 icsk->icsk_mtup.enabled = 1; 121 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 122 } else { 123 struct tcp_sock *tp = tcp_sk(sk); 124 int mss; 125 126 mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1; 127 mss = min(sysctl_tcp_base_mss, mss); 128 mss = max(mss, 68 - tp->tcp_header_len); 129 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 130 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 131 } 132 } 133 } 134 135 /* This function calculates a "timeout" which is equivalent to the timeout of a 136 * TCP connection after "boundary" unsucessful, exponentially backed-off 137 * retransmissions with an initial RTO of TCP_RTO_MIN. 138 */ 139 static bool retransmits_timed_out(struct sock *sk, 140 unsigned int boundary) 141 { 142 unsigned int timeout, linear_backoff_thresh; 143 unsigned int start_ts; 144 145 if (!inet_csk(sk)->icsk_retransmits) 146 return false; 147 148 if (unlikely(!tcp_sk(sk)->retrans_stamp)) 149 start_ts = TCP_SKB_CB(tcp_write_queue_head(sk))->when; 150 else 151 start_ts = tcp_sk(sk)->retrans_stamp; 152 153 linear_backoff_thresh = ilog2(TCP_RTO_MAX/TCP_RTO_MIN); 154 155 if (boundary <= linear_backoff_thresh) 156 timeout = ((2 << boundary) - 1) * TCP_RTO_MIN; 157 else 158 timeout = ((2 << linear_backoff_thresh) - 1) * TCP_RTO_MIN + 159 (boundary - linear_backoff_thresh) * TCP_RTO_MAX; 160 161 return (tcp_time_stamp - start_ts) >= timeout; 162 } 163 164 /* A write timeout has occurred. Process the after effects. */ 165 static int tcp_write_timeout(struct sock *sk) 166 { 167 struct inet_connection_sock *icsk = inet_csk(sk); 168 int retry_until; 169 bool do_reset; 170 171 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) { 172 if (icsk->icsk_retransmits) 173 dst_negative_advice(&sk->sk_dst_cache, sk); 174 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 175 } else { 176 if (retransmits_timed_out(sk, sysctl_tcp_retries1)) { 177 /* Black hole detection */ 178 tcp_mtu_probing(icsk, sk); 179 180 dst_negative_advice(&sk->sk_dst_cache, sk); 181 } 182 183 retry_until = sysctl_tcp_retries2; 184 if (sock_flag(sk, SOCK_DEAD)) { 185 const int alive = (icsk->icsk_rto < TCP_RTO_MAX); 186 187 retry_until = tcp_orphan_retries(sk, alive); 188 do_reset = alive || 189 !retransmits_timed_out(sk, retry_until); 190 191 if (tcp_out_of_resources(sk, do_reset)) 192 return 1; 193 } 194 } 195 196 if (retransmits_timed_out(sk, retry_until)) { 197 /* Has it gone just too far? */ 198 tcp_write_err(sk); 199 return 1; 200 } 201 return 0; 202 } 203 204 static void tcp_delack_timer(unsigned long data) 205 { 206 struct sock *sk = (struct sock *)data; 207 struct tcp_sock *tp = tcp_sk(sk); 208 struct inet_connection_sock *icsk = inet_csk(sk); 209 210 bh_lock_sock(sk); 211 if (sock_owned_by_user(sk)) { 212 /* Try again later. */ 213 icsk->icsk_ack.blocked = 1; 214 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED); 215 sk_reset_timer(sk, &icsk->icsk_delack_timer, jiffies + TCP_DELACK_MIN); 216 goto out_unlock; 217 } 218 219 sk_mem_reclaim_partial(sk); 220 221 if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER)) 222 goto out; 223 224 if (time_after(icsk->icsk_ack.timeout, jiffies)) { 225 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout); 226 goto out; 227 } 228 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER; 229 230 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 231 struct sk_buff *skb; 232 233 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED); 234 235 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 236 sk_backlog_rcv(sk, skb); 237 238 tp->ucopy.memory = 0; 239 } 240 241 if (inet_csk_ack_scheduled(sk)) { 242 if (!icsk->icsk_ack.pingpong) { 243 /* Delayed ACK missed: inflate ATO. */ 244 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto); 245 } else { 246 /* Delayed ACK missed: leave pingpong mode and 247 * deflate ATO. 248 */ 249 icsk->icsk_ack.pingpong = 0; 250 icsk->icsk_ack.ato = TCP_ATO_MIN; 251 } 252 tcp_send_ack(sk); 253 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS); 254 } 255 TCP_CHECK_TIMER(sk); 256 257 out: 258 if (tcp_memory_pressure) 259 sk_mem_reclaim(sk); 260 out_unlock: 261 bh_unlock_sock(sk); 262 sock_put(sk); 263 } 264 265 static void tcp_probe_timer(struct sock *sk) 266 { 267 struct inet_connection_sock *icsk = inet_csk(sk); 268 struct tcp_sock *tp = tcp_sk(sk); 269 int max_probes; 270 271 if (tp->packets_out || !tcp_send_head(sk)) { 272 icsk->icsk_probes_out = 0; 273 return; 274 } 275 276 /* *WARNING* RFC 1122 forbids this 277 * 278 * It doesn't AFAIK, because we kill the retransmit timer -AK 279 * 280 * FIXME: We ought not to do it, Solaris 2.5 actually has fixing 281 * this behaviour in Solaris down as a bug fix. [AC] 282 * 283 * Let me to explain. icsk_probes_out is zeroed by incoming ACKs 284 * even if they advertise zero window. Hence, connection is killed only 285 * if we received no ACKs for normal connection timeout. It is not killed 286 * only because window stays zero for some time, window may be zero 287 * until armageddon and even later. We are in full accordance 288 * with RFCs, only probe timer combines both retransmission timeout 289 * and probe timeout in one bottle. --ANK 290 */ 291 max_probes = sysctl_tcp_retries2; 292 293 if (sock_flag(sk, SOCK_DEAD)) { 294 const int alive = ((icsk->icsk_rto << icsk->icsk_backoff) < TCP_RTO_MAX); 295 296 max_probes = tcp_orphan_retries(sk, alive); 297 298 if (tcp_out_of_resources(sk, alive || icsk->icsk_probes_out <= max_probes)) 299 return; 300 } 301 302 if (icsk->icsk_probes_out > max_probes) { 303 tcp_write_err(sk); 304 } else { 305 /* Only send another probe if we didn't close things up. */ 306 tcp_send_probe0(sk); 307 } 308 } 309 310 /* 311 * The TCP retransmit timer. 312 */ 313 314 void tcp_retransmit_timer(struct sock *sk) 315 { 316 struct tcp_sock *tp = tcp_sk(sk); 317 struct inet_connection_sock *icsk = inet_csk(sk); 318 319 if (!tp->packets_out) 320 goto out; 321 322 WARN_ON(tcp_write_queue_empty(sk)); 323 324 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) && 325 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) { 326 /* Receiver dastardly shrinks window. Our retransmits 327 * become zero probes, but we should not timeout this 328 * connection. If the socket is an orphan, time it out, 329 * we cannot allow such beasts to hang infinitely. 330 */ 331 #ifdef TCP_DEBUG 332 struct inet_sock *inet = inet_sk(sk); 333 if (sk->sk_family == AF_INET) { 334 LIMIT_NETDEBUG(KERN_DEBUG "TCP: Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 335 &inet->inet_daddr, ntohs(inet->inet_dport), 336 inet->inet_num, tp->snd_una, tp->snd_nxt); 337 } 338 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) 339 else if (sk->sk_family == AF_INET6) { 340 struct ipv6_pinfo *np = inet6_sk(sk); 341 LIMIT_NETDEBUG(KERN_DEBUG "TCP: Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 342 &np->daddr, ntohs(inet->inet_dport), 343 inet->inet_num, tp->snd_una, tp->snd_nxt); 344 } 345 #endif 346 #endif 347 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) { 348 tcp_write_err(sk); 349 goto out; 350 } 351 tcp_enter_loss(sk, 0); 352 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 353 __sk_dst_reset(sk); 354 goto out_reset_timer; 355 } 356 357 if (tcp_write_timeout(sk)) 358 goto out; 359 360 if (icsk->icsk_retransmits == 0) { 361 int mib_idx; 362 363 if (icsk->icsk_ca_state == TCP_CA_Disorder) { 364 if (tcp_is_sack(tp)) 365 mib_idx = LINUX_MIB_TCPSACKFAILURES; 366 else 367 mib_idx = LINUX_MIB_TCPRENOFAILURES; 368 } else if (icsk->icsk_ca_state == TCP_CA_Recovery) { 369 if (tcp_is_sack(tp)) 370 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL; 371 else 372 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL; 373 } else if (icsk->icsk_ca_state == TCP_CA_Loss) { 374 mib_idx = LINUX_MIB_TCPLOSSFAILURES; 375 } else { 376 mib_idx = LINUX_MIB_TCPTIMEOUTS; 377 } 378 NET_INC_STATS_BH(sock_net(sk), mib_idx); 379 } 380 381 if (tcp_use_frto(sk)) { 382 tcp_enter_frto(sk); 383 } else { 384 tcp_enter_loss(sk, 0); 385 } 386 387 if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) { 388 /* Retransmission failed because of local congestion, 389 * do not backoff. 390 */ 391 if (!icsk->icsk_retransmits) 392 icsk->icsk_retransmits = 1; 393 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 394 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL), 395 TCP_RTO_MAX); 396 goto out; 397 } 398 399 /* Increase the timeout each time we retransmit. Note that 400 * we do not increase the rtt estimate. rto is initialized 401 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests 402 * that doubling rto each time is the least we can get away with. 403 * In KA9Q, Karn uses this for the first few times, and then 404 * goes to quadratic. netBSD doubles, but only goes up to *64, 405 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is 406 * defined in the protocol as the maximum possible RTT. I guess 407 * we'll have to use something other than TCP to talk to the 408 * University of Mars. 409 * 410 * PAWS allows us longer timeouts and large windows, so once 411 * implemented ftp to mars will work nicely. We will have to fix 412 * the 120 second clamps though! 413 */ 414 icsk->icsk_backoff++; 415 icsk->icsk_retransmits++; 416 417 out_reset_timer: 418 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX); 419 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX); 420 if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1)) 421 __sk_dst_reset(sk); 422 423 out:; 424 } 425 426 static void tcp_write_timer(unsigned long data) 427 { 428 struct sock *sk = (struct sock *)data; 429 struct inet_connection_sock *icsk = inet_csk(sk); 430 int event; 431 432 bh_lock_sock(sk); 433 if (sock_owned_by_user(sk)) { 434 /* Try again later */ 435 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + (HZ / 20)); 436 goto out_unlock; 437 } 438 439 if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending) 440 goto out; 441 442 if (time_after(icsk->icsk_timeout, jiffies)) { 443 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout); 444 goto out; 445 } 446 447 event = icsk->icsk_pending; 448 icsk->icsk_pending = 0; 449 450 switch (event) { 451 case ICSK_TIME_RETRANS: 452 tcp_retransmit_timer(sk); 453 break; 454 case ICSK_TIME_PROBE0: 455 tcp_probe_timer(sk); 456 break; 457 } 458 TCP_CHECK_TIMER(sk); 459 460 out: 461 sk_mem_reclaim(sk); 462 out_unlock: 463 bh_unlock_sock(sk); 464 sock_put(sk); 465 } 466 467 /* 468 * Timer for listening sockets 469 */ 470 471 static void tcp_synack_timer(struct sock *sk) 472 { 473 inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL, 474 TCP_TIMEOUT_INIT, TCP_RTO_MAX); 475 } 476 477 void tcp_set_keepalive(struct sock *sk, int val) 478 { 479 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 480 return; 481 482 if (val && !sock_flag(sk, SOCK_KEEPOPEN)) 483 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk))); 484 else if (!val) 485 inet_csk_delete_keepalive_timer(sk); 486 } 487 488 489 static void tcp_keepalive_timer (unsigned long data) 490 { 491 struct sock *sk = (struct sock *) data; 492 struct inet_connection_sock *icsk = inet_csk(sk); 493 struct tcp_sock *tp = tcp_sk(sk); 494 __u32 elapsed; 495 496 /* Only process if socket is not in use. */ 497 bh_lock_sock(sk); 498 if (sock_owned_by_user(sk)) { 499 /* Try again later. */ 500 inet_csk_reset_keepalive_timer (sk, HZ/20); 501 goto out; 502 } 503 504 if (sk->sk_state == TCP_LISTEN) { 505 tcp_synack_timer(sk); 506 goto out; 507 } 508 509 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) { 510 if (tp->linger2 >= 0) { 511 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN; 512 513 if (tmo > 0) { 514 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 515 goto out; 516 } 517 } 518 tcp_send_active_reset(sk, GFP_ATOMIC); 519 goto death; 520 } 521 522 if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE) 523 goto out; 524 525 elapsed = keepalive_time_when(tp); 526 527 /* It is alive without keepalive 8) */ 528 if (tp->packets_out || tcp_send_head(sk)) 529 goto resched; 530 531 elapsed = tcp_time_stamp - tp->rcv_tstamp; 532 533 if (elapsed >= keepalive_time_when(tp)) { 534 if (icsk->icsk_probes_out >= keepalive_probes(tp)) { 535 tcp_send_active_reset(sk, GFP_ATOMIC); 536 tcp_write_err(sk); 537 goto out; 538 } 539 if (tcp_write_wakeup(sk) <= 0) { 540 icsk->icsk_probes_out++; 541 elapsed = keepalive_intvl_when(tp); 542 } else { 543 /* If keepalive was lost due to local congestion, 544 * try harder. 545 */ 546 elapsed = TCP_RESOURCE_PROBE_INTERVAL; 547 } 548 } else { 549 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */ 550 elapsed = keepalive_time_when(tp) - elapsed; 551 } 552 553 TCP_CHECK_TIMER(sk); 554 sk_mem_reclaim(sk); 555 556 resched: 557 inet_csk_reset_keepalive_timer (sk, elapsed); 558 goto out; 559 560 death: 561 tcp_done(sk); 562 563 out: 564 bh_unlock_sock(sk); 565 sock_put(sk); 566 } 567