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 /* A write timeout has occurred. Process the after effects. */ 136 static int tcp_write_timeout(struct sock *sk) 137 { 138 struct inet_connection_sock *icsk = inet_csk(sk); 139 int retry_until; 140 bool do_reset; 141 142 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) { 143 if (icsk->icsk_retransmits) 144 dst_negative_advice(&sk->sk_dst_cache); 145 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 146 } else { 147 if (retransmits_timed_out(sk, sysctl_tcp_retries1)) { 148 /* Black hole detection */ 149 tcp_mtu_probing(icsk, sk); 150 151 dst_negative_advice(&sk->sk_dst_cache); 152 } 153 154 retry_until = sysctl_tcp_retries2; 155 if (sock_flag(sk, SOCK_DEAD)) { 156 const int alive = (icsk->icsk_rto < TCP_RTO_MAX); 157 158 retry_until = tcp_orphan_retries(sk, alive); 159 do_reset = alive || 160 !retransmits_timed_out(sk, retry_until); 161 162 if (tcp_out_of_resources(sk, do_reset)) 163 return 1; 164 } 165 } 166 167 if (retransmits_timed_out(sk, retry_until)) { 168 /* Has it gone just too far? */ 169 tcp_write_err(sk); 170 return 1; 171 } 172 return 0; 173 } 174 175 static void tcp_delack_timer(unsigned long data) 176 { 177 struct sock *sk = (struct sock *)data; 178 struct tcp_sock *tp = tcp_sk(sk); 179 struct inet_connection_sock *icsk = inet_csk(sk); 180 181 bh_lock_sock(sk); 182 if (sock_owned_by_user(sk)) { 183 /* Try again later. */ 184 icsk->icsk_ack.blocked = 1; 185 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED); 186 sk_reset_timer(sk, &icsk->icsk_delack_timer, jiffies + TCP_DELACK_MIN); 187 goto out_unlock; 188 } 189 190 sk_mem_reclaim_partial(sk); 191 192 if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER)) 193 goto out; 194 195 if (time_after(icsk->icsk_ack.timeout, jiffies)) { 196 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout); 197 goto out; 198 } 199 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER; 200 201 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 202 struct sk_buff *skb; 203 204 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSCHEDULERFAILED); 205 206 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 207 sk_backlog_rcv(sk, skb); 208 209 tp->ucopy.memory = 0; 210 } 211 212 if (inet_csk_ack_scheduled(sk)) { 213 if (!icsk->icsk_ack.pingpong) { 214 /* Delayed ACK missed: inflate ATO. */ 215 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto); 216 } else { 217 /* Delayed ACK missed: leave pingpong mode and 218 * deflate ATO. 219 */ 220 icsk->icsk_ack.pingpong = 0; 221 icsk->icsk_ack.ato = TCP_ATO_MIN; 222 } 223 tcp_send_ack(sk); 224 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKS); 225 } 226 TCP_CHECK_TIMER(sk); 227 228 out: 229 if (tcp_memory_pressure) 230 sk_mem_reclaim(sk); 231 out_unlock: 232 bh_unlock_sock(sk); 233 sock_put(sk); 234 } 235 236 static void tcp_probe_timer(struct sock *sk) 237 { 238 struct inet_connection_sock *icsk = inet_csk(sk); 239 struct tcp_sock *tp = tcp_sk(sk); 240 int max_probes; 241 242 if (tp->packets_out || !tcp_send_head(sk)) { 243 icsk->icsk_probes_out = 0; 244 return; 245 } 246 247 /* *WARNING* RFC 1122 forbids this 248 * 249 * It doesn't AFAIK, because we kill the retransmit timer -AK 250 * 251 * FIXME: We ought not to do it, Solaris 2.5 actually has fixing 252 * this behaviour in Solaris down as a bug fix. [AC] 253 * 254 * Let me to explain. icsk_probes_out is zeroed by incoming ACKs 255 * even if they advertise zero window. Hence, connection is killed only 256 * if we received no ACKs for normal connection timeout. It is not killed 257 * only because window stays zero for some time, window may be zero 258 * until armageddon and even later. We are in full accordance 259 * with RFCs, only probe timer combines both retransmission timeout 260 * and probe timeout in one bottle. --ANK 261 */ 262 max_probes = sysctl_tcp_retries2; 263 264 if (sock_flag(sk, SOCK_DEAD)) { 265 const int alive = ((icsk->icsk_rto << icsk->icsk_backoff) < TCP_RTO_MAX); 266 267 max_probes = tcp_orphan_retries(sk, alive); 268 269 if (tcp_out_of_resources(sk, alive || icsk->icsk_probes_out <= max_probes)) 270 return; 271 } 272 273 if (icsk->icsk_probes_out > max_probes) { 274 tcp_write_err(sk); 275 } else { 276 /* Only send another probe if we didn't close things up. */ 277 tcp_send_probe0(sk); 278 } 279 } 280 281 /* 282 * The TCP retransmit timer. 283 */ 284 285 void tcp_retransmit_timer(struct sock *sk) 286 { 287 struct tcp_sock *tp = tcp_sk(sk); 288 struct inet_connection_sock *icsk = inet_csk(sk); 289 290 if (!tp->packets_out) 291 goto out; 292 293 WARN_ON(tcp_write_queue_empty(sk)); 294 295 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) && 296 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) { 297 /* Receiver dastardly shrinks window. Our retransmits 298 * become zero probes, but we should not timeout this 299 * connection. If the socket is an orphan, time it out, 300 * we cannot allow such beasts to hang infinitely. 301 */ 302 #ifdef TCP_DEBUG 303 struct inet_sock *inet = inet_sk(sk); 304 if (sk->sk_family == AF_INET) { 305 LIMIT_NETDEBUG(KERN_DEBUG "TCP: Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 306 &inet->daddr, ntohs(inet->dport), 307 inet->num, tp->snd_una, tp->snd_nxt); 308 } 309 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) 310 else if (sk->sk_family == AF_INET6) { 311 struct ipv6_pinfo *np = inet6_sk(sk); 312 LIMIT_NETDEBUG(KERN_DEBUG "TCP: Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n", 313 &np->daddr, ntohs(inet->dport), 314 inet->num, tp->snd_una, tp->snd_nxt); 315 } 316 #endif 317 #endif 318 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) { 319 tcp_write_err(sk); 320 goto out; 321 } 322 tcp_enter_loss(sk, 0); 323 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 324 __sk_dst_reset(sk); 325 goto out_reset_timer; 326 } 327 328 if (tcp_write_timeout(sk)) 329 goto out; 330 331 if (icsk->icsk_retransmits == 0) { 332 int mib_idx; 333 334 if (icsk->icsk_ca_state == TCP_CA_Disorder) { 335 if (tcp_is_sack(tp)) 336 mib_idx = LINUX_MIB_TCPSACKFAILURES; 337 else 338 mib_idx = LINUX_MIB_TCPRENOFAILURES; 339 } else if (icsk->icsk_ca_state == TCP_CA_Recovery) { 340 if (tcp_is_sack(tp)) 341 mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL; 342 else 343 mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL; 344 } else if (icsk->icsk_ca_state == TCP_CA_Loss) { 345 mib_idx = LINUX_MIB_TCPLOSSFAILURES; 346 } else { 347 mib_idx = LINUX_MIB_TCPTIMEOUTS; 348 } 349 NET_INC_STATS_BH(sock_net(sk), mib_idx); 350 } 351 352 if (tcp_use_frto(sk)) { 353 tcp_enter_frto(sk); 354 } else { 355 tcp_enter_loss(sk, 0); 356 } 357 358 if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) { 359 /* Retransmission failed because of local congestion, 360 * do not backoff. 361 */ 362 if (!icsk->icsk_retransmits) 363 icsk->icsk_retransmits = 1; 364 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 365 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL), 366 TCP_RTO_MAX); 367 goto out; 368 } 369 370 /* Increase the timeout each time we retransmit. Note that 371 * we do not increase the rtt estimate. rto is initialized 372 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests 373 * that doubling rto each time is the least we can get away with. 374 * In KA9Q, Karn uses this for the first few times, and then 375 * goes to quadratic. netBSD doubles, but only goes up to *64, 376 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is 377 * defined in the protocol as the maximum possible RTT. I guess 378 * we'll have to use something other than TCP to talk to the 379 * University of Mars. 380 * 381 * PAWS allows us longer timeouts and large windows, so once 382 * implemented ftp to mars will work nicely. We will have to fix 383 * the 120 second clamps though! 384 */ 385 icsk->icsk_backoff++; 386 icsk->icsk_retransmits++; 387 388 out_reset_timer: 389 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX); 390 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX); 391 if (retransmits_timed_out(sk, sysctl_tcp_retries1 + 1)) 392 __sk_dst_reset(sk); 393 394 out:; 395 } 396 397 static void tcp_write_timer(unsigned long data) 398 { 399 struct sock *sk = (struct sock *)data; 400 struct inet_connection_sock *icsk = inet_csk(sk); 401 int event; 402 403 bh_lock_sock(sk); 404 if (sock_owned_by_user(sk)) { 405 /* Try again later */ 406 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + (HZ / 20)); 407 goto out_unlock; 408 } 409 410 if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending) 411 goto out; 412 413 if (time_after(icsk->icsk_timeout, jiffies)) { 414 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout); 415 goto out; 416 } 417 418 event = icsk->icsk_pending; 419 icsk->icsk_pending = 0; 420 421 switch (event) { 422 case ICSK_TIME_RETRANS: 423 tcp_retransmit_timer(sk); 424 break; 425 case ICSK_TIME_PROBE0: 426 tcp_probe_timer(sk); 427 break; 428 } 429 TCP_CHECK_TIMER(sk); 430 431 out: 432 sk_mem_reclaim(sk); 433 out_unlock: 434 bh_unlock_sock(sk); 435 sock_put(sk); 436 } 437 438 /* 439 * Timer for listening sockets 440 */ 441 442 static void tcp_synack_timer(struct sock *sk) 443 { 444 inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL, 445 TCP_TIMEOUT_INIT, TCP_RTO_MAX); 446 } 447 448 void tcp_set_keepalive(struct sock *sk, int val) 449 { 450 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 451 return; 452 453 if (val && !sock_flag(sk, SOCK_KEEPOPEN)) 454 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk))); 455 else if (!val) 456 inet_csk_delete_keepalive_timer(sk); 457 } 458 459 460 static void tcp_keepalive_timer (unsigned long data) 461 { 462 struct sock *sk = (struct sock *) data; 463 struct inet_connection_sock *icsk = inet_csk(sk); 464 struct tcp_sock *tp = tcp_sk(sk); 465 __u32 elapsed; 466 467 /* Only process if socket is not in use. */ 468 bh_lock_sock(sk); 469 if (sock_owned_by_user(sk)) { 470 /* Try again later. */ 471 inet_csk_reset_keepalive_timer (sk, HZ/20); 472 goto out; 473 } 474 475 if (sk->sk_state == TCP_LISTEN) { 476 tcp_synack_timer(sk); 477 goto out; 478 } 479 480 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) { 481 if (tp->linger2 >= 0) { 482 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN; 483 484 if (tmo > 0) { 485 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 486 goto out; 487 } 488 } 489 tcp_send_active_reset(sk, GFP_ATOMIC); 490 goto death; 491 } 492 493 if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE) 494 goto out; 495 496 elapsed = keepalive_time_when(tp); 497 498 /* It is alive without keepalive 8) */ 499 if (tp->packets_out || tcp_send_head(sk)) 500 goto resched; 501 502 elapsed = tcp_time_stamp - tp->rcv_tstamp; 503 504 if (elapsed >= keepalive_time_when(tp)) { 505 if (icsk->icsk_probes_out >= keepalive_probes(tp)) { 506 tcp_send_active_reset(sk, GFP_ATOMIC); 507 tcp_write_err(sk); 508 goto out; 509 } 510 if (tcp_write_wakeup(sk) <= 0) { 511 icsk->icsk_probes_out++; 512 elapsed = keepalive_intvl_when(tp); 513 } else { 514 /* If keepalive was lost due to local congestion, 515 * try harder. 516 */ 517 elapsed = TCP_RESOURCE_PROBE_INTERVAL; 518 } 519 } else { 520 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */ 521 elapsed = keepalive_time_when(tp) - elapsed; 522 } 523 524 TCP_CHECK_TIMER(sk); 525 sk_mem_reclaim(sk); 526 527 resched: 528 inet_csk_reset_keepalive_timer (sk, elapsed); 529 goto out; 530 531 death: 532 tcp_done(sk); 533 534 out: 535 bh_unlock_sock(sk); 536 sock_put(sk); 537 } 538