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 * Version: $Id: tcp_timer.c,v 1.1.1.1 2007/08/03 18:53:51 Exp $ 9 * 10 * Authors: Ross Biro 11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 12 * Mark Evans, <evansmp@uhura.aston.ac.uk> 13 * Corey Minyard <wf-rch!minyard@relay.EU.net> 14 * Florian La Roche, <flla@stud.uni-sb.de> 15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 16 * Linus Torvalds, <torvalds@cs.helsinki.fi> 17 * Alan Cox, <gw4pts@gw4pts.ampr.org> 18 * Matthew Dillon, <dillon@apollo.west.oic.com> 19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 20 * Jorge Cwik, <jorge@laser.satlink.net> 21 */ 22 23#include <linux/module.h> 24#include <net/tcp.h> 25 26int sysctl_tcp_syn_retries __read_mostly = TCP_SYN_RETRIES; 27int sysctl_tcp_synack_retries __read_mostly = TCP_SYNACK_RETRIES; 28int sysctl_tcp_keepalive_time __read_mostly = TCP_KEEPALIVE_TIME; 29int sysctl_tcp_keepalive_probes __read_mostly = TCP_KEEPALIVE_PROBES; 30int sysctl_tcp_keepalive_intvl __read_mostly = TCP_KEEPALIVE_INTVL; 31int sysctl_tcp_retries1 __read_mostly = TCP_RETR1; 32int sysctl_tcp_retries2 __read_mostly = TCP_RETR2; 33int sysctl_tcp_orphan_retries __read_mostly; 34 35static void tcp_write_timer(unsigned long); 36static void tcp_delack_timer(unsigned long); 37static void tcp_keepalive_timer (unsigned long data); 38 39void 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 45EXPORT_SYMBOL(tcp_init_xmit_timers); 46 47static void tcp_write_err(struct sock *sk) 48{ 49 sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT; 50 sk->sk_error_report(sk); 51 52 tcp_done(sk); 53 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONTIMEOUT); 54} 55 56/* Do not allow orphaned sockets to eat all our resources. 57 * This is direct violation of TCP specs, but it is required 58 * to prevent DoS attacks. It is called when a retransmission timeout 59 * or zero probe timeout occurs on orphaned socket. 60 * 61 * Criteria is still not confirmed experimentally and may change. 62 * We kill the socket, if: 63 * 1. If number of orphaned sockets exceeds an administratively configured 64 * limit. 65 * 2. If we have strong memory pressure. 66 */ 67static int tcp_out_of_resources(struct sock *sk, int do_reset) 68{ 69 struct tcp_sock *tp = tcp_sk(sk); 70 int orphans = atomic_read(&tcp_orphan_count); 71 72 /* If peer does not open window for long time, or did not transmit 73 * anything for long time, penalize it. */ 74 if ((s32)(tcp_time_stamp - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset) 75 orphans <<= 1; 76 77 /* If some dubious ICMP arrived, penalize even more. */ 78 if (sk->sk_err_soft) 79 orphans <<= 1; 80 81 if (tcp_too_many_orphans(sk, orphans)) { 82 if (net_ratelimit()) 83 printk(KERN_INFO "Out of socket memory\n"); 84 85 /* Catch exceptional cases, when connection requires reset. 86 * 1. Last segment was sent recently. */ 87 if ((s32)(tcp_time_stamp - tp->lsndtime) <= TCP_TIMEWAIT_LEN || 88 /* 2. Window is closed. */ 89 (!tp->snd_wnd && !tp->packets_out)) 90 do_reset = 1; 91 if (do_reset) 92 tcp_send_active_reset(sk, GFP_ATOMIC); 93 tcp_done(sk); 94 NET_INC_STATS_BH(LINUX_MIB_TCPABORTONMEMORY); 95 return 1; 96 } 97 return 0; 98} 99 100/* Calculate maximal number or retries on an orphaned socket. */ 101static int tcp_orphan_retries(struct sock *sk, int alive) 102{ 103 int retries = sysctl_tcp_orphan_retries; /* May be zero. */ 104 105 /* We know from an ICMP that something is wrong. */ 106 if (sk->sk_err_soft && !alive) 107 retries = 0; 108 109 /* However, if socket sent something recently, select some safe 110 * number of retries. 8 corresponds to >100 seconds with minimal 111 * RTO of 200msec. */ 112 if (retries == 0 && alive) 113 retries = 8; 114 return retries; 115} 116 117/* A write timeout has occurred. Process the after effects. */ 118static int tcp_write_timeout(struct sock *sk) 119{ 120 struct inet_connection_sock *icsk = inet_csk(sk); 121 struct tcp_sock *tp = tcp_sk(sk); 122 int retry_until; 123 int mss; 124 125 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) { 126 if (icsk->icsk_retransmits) 127 dst_negative_advice(&sk->sk_dst_cache); 128 retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries; 129 } else { 130 if (icsk->icsk_retransmits >= sysctl_tcp_retries1) { 131 /* Black hole detection */ 132 if (sysctl_tcp_mtu_probing) { 133 if (!icsk->icsk_mtup.enabled) { 134 icsk->icsk_mtup.enabled = 1; 135 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 136 } else { 137 mss = min(sysctl_tcp_base_mss, 138 tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)/2); 139 mss = max(mss, 68 - tp->tcp_header_len); 140 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 141 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 142 } 143 } 144 145 dst_negative_advice(&sk->sk_dst_cache); 146 } 147 148 retry_until = sysctl_tcp_retries2; 149 if (sock_flag(sk, SOCK_DEAD)) { 150 const int alive = (icsk->icsk_rto < TCP_RTO_MAX); 151 152 retry_until = tcp_orphan_retries(sk, alive); 153 154 if (tcp_out_of_resources(sk, alive || icsk->icsk_retransmits < retry_until)) 155 return 1; 156 } 157 } 158 159 if (icsk->icsk_retransmits >= retry_until) { 160 /* Has it gone just too far? */ 161 tcp_write_err(sk); 162 return 1; 163 } 164 return 0; 165} 166 167static void tcp_delack_timer(unsigned long data) 168{ 169 struct sock *sk = (struct sock*)data; 170 struct tcp_sock *tp = tcp_sk(sk); 171 struct inet_connection_sock *icsk = inet_csk(sk); 172 173 bh_lock_sock(sk); 174 if (sock_owned_by_user(sk)) { 175 /* Try again later. */ 176 icsk->icsk_ack.blocked = 1; 177 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOCKED); 178 sk_reset_timer(sk, &icsk->icsk_delack_timer, jiffies + TCP_DELACK_MIN); 179 goto out_unlock; 180 } 181 182 sk_stream_mem_reclaim(sk); 183 184 if (sk->sk_state == TCP_CLOSE || !(icsk->icsk_ack.pending & ICSK_ACK_TIMER)) 185 goto out; 186 187 if (time_after(icsk->icsk_ack.timeout, jiffies)) { 188 sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout); 189 goto out; 190 } 191 icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER; 192 193 if (!skb_queue_empty(&tp->ucopy.prequeue)) { 194 struct sk_buff *skb; 195 196 NET_INC_STATS_BH(LINUX_MIB_TCPSCHEDULERFAILED); 197 198 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL) 199 sk->sk_backlog_rcv(sk, skb); 200 201 tp->ucopy.memory = 0; 202 } 203 204 if (inet_csk_ack_scheduled(sk)) { 205 if (!icsk->icsk_ack.pingpong) { 206 /* Delayed ACK missed: inflate ATO. */ 207 icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto); 208 } else { 209 /* Delayed ACK missed: leave pingpong mode and 210 * deflate ATO. 211 */ 212 icsk->icsk_ack.pingpong = 0; 213 icsk->icsk_ack.ato = TCP_ATO_MIN; 214 } 215 tcp_send_ack(sk); 216 NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKS); 217 } 218 TCP_CHECK_TIMER(sk); 219 220out: 221 if (tcp_memory_pressure) 222 sk_stream_mem_reclaim(sk); 223out_unlock: 224 bh_unlock_sock(sk); 225 sock_put(sk); 226} 227 228static void tcp_probe_timer(struct sock *sk) 229{ 230 struct inet_connection_sock *icsk = inet_csk(sk); 231 struct tcp_sock *tp = tcp_sk(sk); 232 int max_probes; 233 234 if (tp->packets_out || !tcp_send_head(sk)) { 235 icsk->icsk_probes_out = 0; 236 return; 237 } 238 239 max_probes = sysctl_tcp_retries2; 240 241 if (sock_flag(sk, SOCK_DEAD)) { 242 const int alive = ((icsk->icsk_rto << icsk->icsk_backoff) < TCP_RTO_MAX); 243 244 max_probes = tcp_orphan_retries(sk, alive); 245 246 if (tcp_out_of_resources(sk, alive || icsk->icsk_probes_out <= max_probes)) 247 return; 248 } 249 250 if (icsk->icsk_probes_out > max_probes) { 251 tcp_write_err(sk); 252 } else { 253 /* Only send another probe if we didn't close things up. */ 254 tcp_send_probe0(sk); 255 } 256} 257 258/* 259 * The TCP retransmit timer. 260 */ 261 262static void tcp_retransmit_timer(struct sock *sk) 263{ 264 struct tcp_sock *tp = tcp_sk(sk); 265 struct inet_connection_sock *icsk = inet_csk(sk); 266 267 if (!tp->packets_out) 268 goto out; 269 270 BUG_TRAP(!tcp_write_queue_empty(sk)); 271 272 if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) && 273 !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) { 274 /* Receiver dastardly shrinks window. Our retransmits 275 * become zero probes, but we should not timeout this 276 * connection. If the socket is an orphan, time it out, 277 * we cannot allow such beasts to hang infinitely. 278 */ 279#ifdef TCP_DEBUG 280 if (1) { 281 struct inet_sock *inet = inet_sk(sk); 282 LIMIT_NETDEBUG(KERN_DEBUG "TCP: Treason uncloaked! Peer %u.%u.%u.%u:%u/%u shrinks window %u:%u. Repaired.\n", 283 NIPQUAD(inet->daddr), ntohs(inet->dport), 284 inet->num, tp->snd_una, tp->snd_nxt); 285 } 286#endif 287 if (tcp_time_stamp - tp->rcv_tstamp > TCP_RTO_MAX) { 288 tcp_write_err(sk); 289 goto out; 290 } 291 tcp_enter_loss(sk, 0); 292 tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); 293 __sk_dst_reset(sk); 294 goto out_reset_timer; 295 } 296 297 if (tcp_write_timeout(sk)) 298 goto out; 299 300 if (icsk->icsk_retransmits == 0) { 301 if (icsk->icsk_ca_state == TCP_CA_Disorder || 302 icsk->icsk_ca_state == TCP_CA_Recovery) { 303 if (tp->rx_opt.sack_ok) { 304 if (icsk->icsk_ca_state == TCP_CA_Recovery) 305 NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERYFAIL); 306 else 307 NET_INC_STATS_BH(LINUX_MIB_TCPSACKFAILURES); 308 } else { 309 if (icsk->icsk_ca_state == TCP_CA_Recovery) 310 NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERYFAIL); 311 else 312 NET_INC_STATS_BH(LINUX_MIB_TCPRENOFAILURES); 313 } 314 } else if (icsk->icsk_ca_state == TCP_CA_Loss) { 315 NET_INC_STATS_BH(LINUX_MIB_TCPLOSSFAILURES); 316 } else { 317 NET_INC_STATS_BH(LINUX_MIB_TCPTIMEOUTS); 318 } 319 } 320 321 if (tcp_use_frto(sk)) { 322 tcp_enter_frto(sk); 323 } else { 324 tcp_enter_loss(sk, 0); 325 } 326 327 if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk)) > 0) { 328 /* Retransmission failed because of local congestion, 329 * do not backoff. 330 */ 331 if (!icsk->icsk_retransmits) 332 icsk->icsk_retransmits = 1; 333 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 334 min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL), 335 TCP_RTO_MAX); 336 goto out; 337 } 338 339 /* Increase the timeout each time we retransmit. Note that 340 * we do not increase the rtt estimate. rto is initialized 341 * from rtt, but increases here. Jacobson (SIGCOMM 88) suggests 342 * that doubling rto each time is the least we can get away with. 343 * In KA9Q, Karn uses this for the first few times, and then 344 * goes to quadratic. netBSD doubles, but only goes up to *64, 345 * and clamps at 1 to 64 sec afterwards. Note that 120 sec is 346 * defined in the protocol as the maximum possible RTT. I guess 347 * we'll have to use something other than TCP to talk to the 348 * University of Mars. 349 * 350 * PAWS allows us longer timeouts and large windows, so once 351 * implemented ftp to mars will work nicely. We will have to fix 352 * the 120 second clamps though! 353 */ 354 icsk->icsk_backoff++; 355 icsk->icsk_retransmits++; 356 357out_reset_timer: 358 icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX); 359 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX); 360 if (icsk->icsk_retransmits > sysctl_tcp_retries1) 361 __sk_dst_reset(sk); 362 363out:; 364} 365 366static void tcp_write_timer(unsigned long data) 367{ 368 struct sock *sk = (struct sock*)data; 369 struct inet_connection_sock *icsk = inet_csk(sk); 370 int event; 371 372 bh_lock_sock(sk); 373 if (sock_owned_by_user(sk)) { 374 /* Try again later */ 375 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + (HZ / 20)); 376 goto out_unlock; 377 } 378 379 if (sk->sk_state == TCP_CLOSE || !icsk->icsk_pending) 380 goto out; 381 382 if (time_after(icsk->icsk_timeout, jiffies)) { 383 sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout); 384 goto out; 385 } 386 387 event = icsk->icsk_pending; 388 icsk->icsk_pending = 0; 389 390 switch (event) { 391 case ICSK_TIME_RETRANS: 392 tcp_retransmit_timer(sk); 393 break; 394 case ICSK_TIME_PROBE0: 395 tcp_probe_timer(sk); 396 break; 397 } 398 TCP_CHECK_TIMER(sk); 399 400out: 401 sk_stream_mem_reclaim(sk); 402out_unlock: 403 bh_unlock_sock(sk); 404 sock_put(sk); 405} 406 407/* 408 * Timer for listening sockets 409 */ 410 411static void tcp_synack_timer(struct sock *sk) 412{ 413 inet_csk_reqsk_queue_prune(sk, TCP_SYNQ_INTERVAL, 414 TCP_TIMEOUT_INIT, TCP_RTO_MAX); 415} 416 417void tcp_set_keepalive(struct sock *sk, int val) 418{ 419 if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) 420 return; 421 422 if (val && !sock_flag(sk, SOCK_KEEPOPEN)) 423 inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk))); 424 else if (!val) 425 inet_csk_delete_keepalive_timer(sk); 426} 427 428 429static void tcp_keepalive_timer (unsigned long data) 430{ 431 struct sock *sk = (struct sock *) data; 432 struct inet_connection_sock *icsk = inet_csk(sk); 433 struct tcp_sock *tp = tcp_sk(sk); 434 __u32 elapsed; 435 436 /* Only process if socket is not in use. */ 437 bh_lock_sock(sk); 438 if (sock_owned_by_user(sk)) { 439 /* Try again later. */ 440 inet_csk_reset_keepalive_timer (sk, HZ/20); 441 goto out; 442 } 443 444 if (sk->sk_state == TCP_LISTEN) { 445 tcp_synack_timer(sk); 446 goto out; 447 } 448 449 if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) { 450 if (tp->linger2 >= 0) { 451 const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN; 452 453 if (tmo > 0) { 454 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 455 goto out; 456 } 457 } 458 tcp_send_active_reset(sk, GFP_ATOMIC); 459 goto death; 460 } 461 462 if (!sock_flag(sk, SOCK_KEEPOPEN) || sk->sk_state == TCP_CLOSE) 463 goto out; 464 465 elapsed = keepalive_time_when(tp); 466 467 /* It is alive without keepalive 8) */ 468 if (tp->packets_out || tcp_send_head(sk)) 469 goto resched; 470 471 elapsed = tcp_time_stamp - tp->rcv_tstamp; 472 473 if (elapsed >= keepalive_time_when(tp)) { 474 if ((!tp->keepalive_probes && icsk->icsk_probes_out >= sysctl_tcp_keepalive_probes) || 475 (tp->keepalive_probes && icsk->icsk_probes_out >= tp->keepalive_probes)) { 476 tcp_send_active_reset(sk, GFP_ATOMIC); 477 tcp_write_err(sk); 478 goto out; 479 } 480 if (tcp_write_wakeup(sk) <= 0) { 481 icsk->icsk_probes_out++; 482 elapsed = keepalive_intvl_when(tp); 483 } else { 484 /* If keepalive was lost due to local congestion, 485 * try harder. 486 */ 487 elapsed = TCP_RESOURCE_PROBE_INTERVAL; 488 } 489 } else { 490 /* It is tp->rcv_tstamp + keepalive_time_when(tp) */ 491 elapsed = keepalive_time_when(tp) - elapsed; 492 } 493 494 TCP_CHECK_TIMER(sk); 495 sk_stream_mem_reclaim(sk); 496 497resched: 498 inet_csk_reset_keepalive_timer (sk, elapsed); 499 goto out; 500 501death: 502 tcp_done(sk); 503 504out: 505 bh_unlock_sock(sk); 506 sock_put(sk); 507} 508