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/mm.h> 22 #include <linux/module.h> 23 #include <linux/slab.h> 24 #include <linux/sysctl.h> 25 #include <linux/workqueue.h> 26 #include <net/tcp.h> 27 #include <net/inet_common.h> 28 #include <net/xfrm.h> 29 30 int sysctl_tcp_syncookies __read_mostly = 1; 31 EXPORT_SYMBOL(sysctl_tcp_syncookies); 32 33 int sysctl_tcp_abort_on_overflow __read_mostly; 34 35 struct inet_timewait_death_row tcp_death_row = { 36 .sysctl_max_tw_buckets = NR_FILE * 2, 37 .period = TCP_TIMEWAIT_LEN / INET_TWDR_TWKILL_SLOTS, 38 .death_lock = __SPIN_LOCK_UNLOCKED(tcp_death_row.death_lock), 39 .hashinfo = &tcp_hashinfo, 40 .tw_timer = TIMER_INITIALIZER(inet_twdr_hangman, 0, 41 (unsigned long)&tcp_death_row), 42 .twkill_work = __WORK_INITIALIZER(tcp_death_row.twkill_work, 43 inet_twdr_twkill_work), 44 /* Short-time timewait calendar */ 45 46 .twcal_hand = -1, 47 .twcal_timer = TIMER_INITIALIZER(inet_twdr_twcal_tick, 0, 48 (unsigned long)&tcp_death_row), 49 }; 50 EXPORT_SYMBOL_GPL(tcp_death_row); 51 52 /* VJ's idea. Save last timestamp seen from this destination 53 * and hold it at least for normal timewait interval to use for duplicate 54 * segment detection in subsequent connections, before they enter synchronized 55 * state. 56 */ 57 58 static int tcp_remember_stamp(struct sock *sk) 59 { 60 const struct inet_connection_sock *icsk = inet_csk(sk); 61 struct tcp_sock *tp = tcp_sk(sk); 62 struct inet_peer *peer; 63 bool release_it; 64 65 peer = icsk->icsk_af_ops->get_peer(sk, &release_it); 66 if (peer) { 67 if ((s32)(peer->tcp_ts - tp->rx_opt.ts_recent) <= 0 || 68 ((u32)get_seconds() - peer->tcp_ts_stamp > TCP_PAWS_MSL && 69 peer->tcp_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) { 70 peer->tcp_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp; 71 peer->tcp_ts = tp->rx_opt.ts_recent; 72 } 73 if (release_it) 74 inet_putpeer(peer); 75 return 1; 76 } 77 78 return 0; 79 } 80 81 static int tcp_tw_remember_stamp(struct inet_timewait_sock *tw) 82 { 83 struct sock *sk = (struct sock *) tw; 84 struct inet_peer *peer; 85 86 peer = twsk_getpeer(sk); 87 if (peer) { 88 const struct tcp_timewait_sock *tcptw = tcp_twsk(sk); 89 90 if ((s32)(peer->tcp_ts - tcptw->tw_ts_recent) <= 0 || 91 ((u32)get_seconds() - peer->tcp_ts_stamp > TCP_PAWS_MSL && 92 peer->tcp_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) { 93 peer->tcp_ts_stamp = (u32)tcptw->tw_ts_recent_stamp; 94 peer->tcp_ts = tcptw->tw_ts_recent; 95 } 96 inet_putpeer(peer); 97 return 1; 98 } 99 return 0; 100 } 101 102 static __inline__ int tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win) 103 { 104 if (seq == s_win) 105 return 1; 106 if (after(end_seq, s_win) && before(seq, e_win)) 107 return 1; 108 return seq == e_win && seq == end_seq; 109 } 110 111 /* 112 * * Main purpose of TIME-WAIT state is to close connection gracefully, 113 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN 114 * (and, probably, tail of data) and one or more our ACKs are lost. 115 * * What is TIME-WAIT timeout? It is associated with maximal packet 116 * lifetime in the internet, which results in wrong conclusion, that 117 * it is set to catch "old duplicate segments" wandering out of their path. 118 * It is not quite correct. This timeout is calculated so that it exceeds 119 * maximal retransmission timeout enough to allow to lose one (or more) 120 * segments sent by peer and our ACKs. This time may be calculated from RTO. 121 * * When TIME-WAIT socket receives RST, it means that another end 122 * finally closed and we are allowed to kill TIME-WAIT too. 123 * * Second purpose of TIME-WAIT is catching old duplicate segments. 124 * Well, certainly it is pure paranoia, but if we load TIME-WAIT 125 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs. 126 * * If we invented some more clever way to catch duplicates 127 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs. 128 * 129 * The algorithm below is based on FORMAL INTERPRETATION of RFCs. 130 * When you compare it to RFCs, please, read section SEGMENT ARRIVES 131 * from the very beginning. 132 * 133 * NOTE. With recycling (and later with fin-wait-2) TW bucket 134 * is _not_ stateless. It means, that strictly speaking we must 135 * spinlock it. I do not want! Well, probability of misbehaviour 136 * is ridiculously low and, seems, we could use some mb() tricks 137 * to avoid misread sequence numbers, states etc. --ANK 138 */ 139 enum tcp_tw_status 140 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb, 141 const struct tcphdr *th) 142 { 143 struct tcp_options_received tmp_opt; 144 u8 *hash_location; 145 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 146 int paws_reject = 0; 147 148 tmp_opt.saw_tstamp = 0; 149 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) { 150 tcp_parse_options(skb, &tmp_opt, &hash_location, 0); 151 152 if (tmp_opt.saw_tstamp) { 153 tmp_opt.ts_recent = tcptw->tw_ts_recent; 154 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp; 155 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 156 } 157 } 158 159 if (tw->tw_substate == TCP_FIN_WAIT2) { 160 /* Just repeat all the checks of tcp_rcv_state_process() */ 161 162 /* Out of window, send ACK */ 163 if (paws_reject || 164 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 165 tcptw->tw_rcv_nxt, 166 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd)) 167 return TCP_TW_ACK; 168 169 if (th->rst) 170 goto kill; 171 172 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt)) 173 goto kill_with_rst; 174 175 /* Dup ACK? */ 176 if (!th->ack || 177 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) || 178 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) { 179 inet_twsk_put(tw); 180 return TCP_TW_SUCCESS; 181 } 182 183 /* New data or FIN. If new data arrive after half-duplex close, 184 * reset. 185 */ 186 if (!th->fin || 187 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) { 188 kill_with_rst: 189 inet_twsk_deschedule(tw, &tcp_death_row); 190 inet_twsk_put(tw); 191 return TCP_TW_RST; 192 } 193 194 /* FIN arrived, enter true time-wait state. */ 195 tw->tw_substate = TCP_TIME_WAIT; 196 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq; 197 if (tmp_opt.saw_tstamp) { 198 tcptw->tw_ts_recent_stamp = get_seconds(); 199 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 200 } 201 202 if (tcp_death_row.sysctl_tw_recycle && 203 tcptw->tw_ts_recent_stamp && 204 tcp_tw_remember_stamp(tw)) 205 inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout, 206 TCP_TIMEWAIT_LEN); 207 else 208 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 209 TCP_TIMEWAIT_LEN); 210 return TCP_TW_ACK; 211 } 212 213 /* 214 * Now real TIME-WAIT state. 215 * 216 * RFC 1122: 217 * "When a connection is [...] on TIME-WAIT state [...] 218 * [a TCP] MAY accept a new SYN from the remote TCP to 219 * reopen the connection directly, if it: 220 * 221 * (1) assigns its initial sequence number for the new 222 * connection to be larger than the largest sequence 223 * number it used on the previous connection incarnation, 224 * and 225 * 226 * (2) returns to TIME-WAIT state if the SYN turns out 227 * to be an old duplicate". 228 */ 229 230 if (!paws_reject && 231 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt && 232 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 233 /* In window segment, it may be only reset or bare ack. */ 234 235 if (th->rst) { 236 /* This is TIME_WAIT assassination, in two flavors. 237 * Oh well... nobody has a sufficient solution to this 238 * protocol bug yet. 239 */ 240 if (sysctl_tcp_rfc1337 == 0) { 241 kill: 242 inet_twsk_deschedule(tw, &tcp_death_row); 243 inet_twsk_put(tw); 244 return TCP_TW_SUCCESS; 245 } 246 } 247 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 248 TCP_TIMEWAIT_LEN); 249 250 if (tmp_opt.saw_tstamp) { 251 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 252 tcptw->tw_ts_recent_stamp = get_seconds(); 253 } 254 255 inet_twsk_put(tw); 256 return TCP_TW_SUCCESS; 257 } 258 259 /* Out of window segment. 260 261 All the segments are ACKed immediately. 262 263 The only exception is new SYN. We accept it, if it is 264 not old duplicate and we are not in danger to be killed 265 by delayed old duplicates. RFC check is that it has 266 newer sequence number works at rates <40Mbit/sec. 267 However, if paws works, it is reliable AND even more, 268 we even may relax silly seq space cutoff. 269 270 RED-PEN: we violate main RFC requirement, if this SYN will appear 271 old duplicate (i.e. we receive RST in reply to SYN-ACK), 272 we must return socket to time-wait state. It is not good, 273 but not fatal yet. 274 */ 275 276 if (th->syn && !th->rst && !th->ack && !paws_reject && 277 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) || 278 (tmp_opt.saw_tstamp && 279 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) { 280 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 281 if (isn == 0) 282 isn++; 283 TCP_SKB_CB(skb)->when = isn; 284 return TCP_TW_SYN; 285 } 286 287 if (paws_reject) 288 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED); 289 290 if (!th->rst) { 291 /* In this case we must reset the TIMEWAIT timer. 292 * 293 * If it is ACKless SYN it may be both old duplicate 294 * and new good SYN with random sequence number <rcv_nxt. 295 * Do not reschedule in the last case. 296 */ 297 if (paws_reject || th->ack) 298 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 299 TCP_TIMEWAIT_LEN); 300 301 /* Send ACK. Note, we do not put the bucket, 302 * it will be released by caller. 303 */ 304 return TCP_TW_ACK; 305 } 306 inet_twsk_put(tw); 307 return TCP_TW_SUCCESS; 308 } 309 EXPORT_SYMBOL(tcp_timewait_state_process); 310 311 /* 312 * Move a socket to time-wait or dead fin-wait-2 state. 313 */ 314 void tcp_time_wait(struct sock *sk, int state, int timeo) 315 { 316 struct inet_timewait_sock *tw = NULL; 317 const struct inet_connection_sock *icsk = inet_csk(sk); 318 const struct tcp_sock *tp = tcp_sk(sk); 319 int recycle_ok = 0; 320 321 if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp) 322 recycle_ok = tcp_remember_stamp(sk); 323 324 if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets) 325 tw = inet_twsk_alloc(sk, state); 326 327 if (tw != NULL) { 328 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 329 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 330 331 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 332 tcptw->tw_rcv_nxt = tp->rcv_nxt; 333 tcptw->tw_snd_nxt = tp->snd_nxt; 334 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 335 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 336 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 337 338 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) 339 if (tw->tw_family == PF_INET6) { 340 struct ipv6_pinfo *np = inet6_sk(sk); 341 struct inet6_timewait_sock *tw6; 342 343 tw->tw_ipv6_offset = inet6_tw_offset(sk->sk_prot); 344 tw6 = inet6_twsk((struct sock *)tw); 345 ipv6_addr_copy(&tw6->tw_v6_daddr, &np->daddr); 346 ipv6_addr_copy(&tw6->tw_v6_rcv_saddr, &np->rcv_saddr); 347 tw->tw_ipv6only = np->ipv6only; 348 } 349 #endif 350 351 #ifdef CONFIG_TCP_MD5SIG 352 /* 353 * The timewait bucket does not have the key DB from the 354 * sock structure. We just make a quick copy of the 355 * md5 key being used (if indeed we are using one) 356 * so the timewait ack generating code has the key. 357 */ 358 do { 359 struct tcp_md5sig_key *key; 360 memset(tcptw->tw_md5_key, 0, sizeof(tcptw->tw_md5_key)); 361 tcptw->tw_md5_keylen = 0; 362 key = tp->af_specific->md5_lookup(sk, sk); 363 if (key != NULL) { 364 memcpy(&tcptw->tw_md5_key, key->key, key->keylen); 365 tcptw->tw_md5_keylen = key->keylen; 366 if (tcp_alloc_md5sig_pool(sk) == NULL) 367 BUG(); 368 } 369 } while (0); 370 #endif 371 372 /* Linkage updates. */ 373 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo); 374 375 /* Get the TIME_WAIT timeout firing. */ 376 if (timeo < rto) 377 timeo = rto; 378 379 if (recycle_ok) { 380 tw->tw_timeout = rto; 381 } else { 382 tw->tw_timeout = TCP_TIMEWAIT_LEN; 383 if (state == TCP_TIME_WAIT) 384 timeo = TCP_TIMEWAIT_LEN; 385 } 386 387 inet_twsk_schedule(tw, &tcp_death_row, timeo, 388 TCP_TIMEWAIT_LEN); 389 inet_twsk_put(tw); 390 } else { 391 /* Sorry, if we're out of memory, just CLOSE this 392 * socket up. We've got bigger problems than 393 * non-graceful socket closings. 394 */ 395 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW); 396 } 397 398 tcp_update_metrics(sk); 399 tcp_done(sk); 400 } 401 402 void tcp_twsk_destructor(struct sock *sk) 403 { 404 #ifdef CONFIG_TCP_MD5SIG 405 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 406 if (twsk->tw_md5_keylen) 407 tcp_free_md5sig_pool(); 408 #endif 409 } 410 EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 411 412 static inline void TCP_ECN_openreq_child(struct tcp_sock *tp, 413 struct request_sock *req) 414 { 415 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 416 } 417 418 /* This is not only more efficient than what we used to do, it eliminates 419 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 420 * 421 * Actually, we could lots of memory writes here. tp of listening 422 * socket contains all necessary default parameters. 423 */ 424 struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb) 425 { 426 struct sock *newsk = inet_csk_clone(sk, req, GFP_ATOMIC); 427 428 if (newsk != NULL) { 429 const struct inet_request_sock *ireq = inet_rsk(req); 430 struct tcp_request_sock *treq = tcp_rsk(req); 431 struct inet_connection_sock *newicsk = inet_csk(newsk); 432 struct tcp_sock *newtp = tcp_sk(newsk); 433 struct tcp_sock *oldtp = tcp_sk(sk); 434 struct tcp_cookie_values *oldcvp = oldtp->cookie_values; 435 436 /* TCP Cookie Transactions require space for the cookie pair, 437 * as it differs for each connection. There is no need to 438 * copy any s_data_payload stored at the original socket. 439 * Failure will prevent resuming the connection. 440 * 441 * Presumed copied, in order of appearance: 442 * cookie_in_always, cookie_out_never 443 */ 444 if (oldcvp != NULL) { 445 struct tcp_cookie_values *newcvp = 446 kzalloc(sizeof(*newtp->cookie_values), 447 GFP_ATOMIC); 448 449 if (newcvp != NULL) { 450 kref_init(&newcvp->kref); 451 newcvp->cookie_desired = 452 oldcvp->cookie_desired; 453 newtp->cookie_values = newcvp; 454 } else { 455 /* Not Yet Implemented */ 456 newtp->cookie_values = NULL; 457 } 458 } 459 460 /* Now setup tcp_sock */ 461 newtp->pred_flags = 0; 462 463 newtp->rcv_wup = newtp->copied_seq = 464 newtp->rcv_nxt = treq->rcv_isn + 1; 465 466 newtp->snd_sml = newtp->snd_una = 467 newtp->snd_nxt = newtp->snd_up = 468 treq->snt_isn + 1 + tcp_s_data_size(oldtp); 469 470 tcp_prequeue_init(newtp); 471 472 tcp_init_wl(newtp, treq->rcv_isn); 473 474 newtp->srtt = 0; 475 newtp->mdev = TCP_TIMEOUT_INIT; 476 newicsk->icsk_rto = TCP_TIMEOUT_INIT; 477 478 newtp->packets_out = 0; 479 newtp->retrans_out = 0; 480 newtp->sacked_out = 0; 481 newtp->fackets_out = 0; 482 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 483 484 /* So many TCP implementations out there (incorrectly) count the 485 * initial SYN frame in their delayed-ACK and congestion control 486 * algorithms that we must have the following bandaid to talk 487 * efficiently to them. -DaveM 488 */ 489 newtp->snd_cwnd = 2; 490 newtp->snd_cwnd_cnt = 0; 491 newtp->bytes_acked = 0; 492 493 newtp->frto_counter = 0; 494 newtp->frto_highmark = 0; 495 496 newicsk->icsk_ca_ops = &tcp_init_congestion_ops; 497 498 tcp_set_ca_state(newsk, TCP_CA_Open); 499 tcp_init_xmit_timers(newsk); 500 skb_queue_head_init(&newtp->out_of_order_queue); 501 newtp->write_seq = newtp->pushed_seq = 502 treq->snt_isn + 1 + tcp_s_data_size(oldtp); 503 504 newtp->rx_opt.saw_tstamp = 0; 505 506 newtp->rx_opt.dsack = 0; 507 newtp->rx_opt.num_sacks = 0; 508 509 newtp->urg_data = 0; 510 511 if (sock_flag(newsk, SOCK_KEEPOPEN)) 512 inet_csk_reset_keepalive_timer(newsk, 513 keepalive_time_when(newtp)); 514 515 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 516 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) { 517 if (sysctl_tcp_fack) 518 tcp_enable_fack(newtp); 519 } 520 newtp->window_clamp = req->window_clamp; 521 newtp->rcv_ssthresh = req->rcv_wnd; 522 newtp->rcv_wnd = req->rcv_wnd; 523 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 524 if (newtp->rx_opt.wscale_ok) { 525 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 526 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 527 } else { 528 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 529 newtp->window_clamp = min(newtp->window_clamp, 65535U); 530 } 531 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) << 532 newtp->rx_opt.snd_wscale); 533 newtp->max_window = newtp->snd_wnd; 534 535 if (newtp->rx_opt.tstamp_ok) { 536 newtp->rx_opt.ts_recent = req->ts_recent; 537 newtp->rx_opt.ts_recent_stamp = get_seconds(); 538 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 539 } else { 540 newtp->rx_opt.ts_recent_stamp = 0; 541 newtp->tcp_header_len = sizeof(struct tcphdr); 542 } 543 #ifdef CONFIG_TCP_MD5SIG 544 newtp->md5sig_info = NULL; /*XXX*/ 545 if (newtp->af_specific->md5_lookup(sk, newsk)) 546 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 547 #endif 548 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 549 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 550 newtp->rx_opt.mss_clamp = req->mss; 551 TCP_ECN_openreq_child(newtp, req); 552 553 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS); 554 } 555 return newsk; 556 } 557 EXPORT_SYMBOL(tcp_create_openreq_child); 558 559 /* 560 * Process an incoming packet for SYN_RECV sockets represented 561 * as a request_sock. 562 */ 563 564 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 565 struct request_sock *req, 566 struct request_sock **prev) 567 { 568 struct tcp_options_received tmp_opt; 569 u8 *hash_location; 570 struct sock *child; 571 const struct tcphdr *th = tcp_hdr(skb); 572 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 573 int paws_reject = 0; 574 575 tmp_opt.saw_tstamp = 0; 576 if (th->doff > (sizeof(struct tcphdr)>>2)) { 577 tcp_parse_options(skb, &tmp_opt, &hash_location, 0); 578 579 if (tmp_opt.saw_tstamp) { 580 tmp_opt.ts_recent = req->ts_recent; 581 /* We do not store true stamp, but it is not required, 582 * it can be estimated (approximately) 583 * from another data. 584 */ 585 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans); 586 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 587 } 588 } 589 590 /* Check for pure retransmitted SYN. */ 591 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 592 flg == TCP_FLAG_SYN && 593 !paws_reject) { 594 /* 595 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 596 * this case on figure 6 and figure 8, but formal 597 * protocol description says NOTHING. 598 * To be more exact, it says that we should send ACK, 599 * because this segment (at least, if it has no data) 600 * is out of window. 601 * 602 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 603 * describe SYN-RECV state. All the description 604 * is wrong, we cannot believe to it and should 605 * rely only on common sense and implementation 606 * experience. 607 * 608 * Enforce "SYN-ACK" according to figure 8, figure 6 609 * of RFC793, fixed by RFC1122. 610 */ 611 req->rsk_ops->rtx_syn_ack(sk, req, NULL); 612 return NULL; 613 } 614 615 /* Further reproduces section "SEGMENT ARRIVES" 616 for state SYN-RECEIVED of RFC793. 617 It is broken, however, it does not work only 618 when SYNs are crossed. 619 620 You would think that SYN crossing is impossible here, since 621 we should have a SYN_SENT socket (from connect()) on our end, 622 but this is not true if the crossed SYNs were sent to both 623 ends by a malicious third party. We must defend against this, 624 and to do that we first verify the ACK (as per RFC793, page 625 36) and reset if it is invalid. Is this a true full defense? 626 To convince ourselves, let us consider a way in which the ACK 627 test can still pass in this 'malicious crossed SYNs' case. 628 Malicious sender sends identical SYNs (and thus identical sequence 629 numbers) to both A and B: 630 631 A: gets SYN, seq=7 632 B: gets SYN, seq=7 633 634 By our good fortune, both A and B select the same initial 635 send sequence number of seven :-) 636 637 A: sends SYN|ACK, seq=7, ack_seq=8 638 B: sends SYN|ACK, seq=7, ack_seq=8 639 640 So we are now A eating this SYN|ACK, ACK test passes. So 641 does sequence test, SYN is truncated, and thus we consider 642 it a bare ACK. 643 644 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 645 bare ACK. Otherwise, we create an established connection. Both 646 ends (listening sockets) accept the new incoming connection and try 647 to talk to each other. 8-) 648 649 Note: This case is both harmless, and rare. Possibility is about the 650 same as us discovering intelligent life on another plant tomorrow. 651 652 But generally, we should (RFC lies!) to accept ACK 653 from SYNACK both here and in tcp_rcv_state_process(). 654 tcp_rcv_state_process() does not, hence, we do not too. 655 656 Note that the case is absolutely generic: 657 we cannot optimize anything here without 658 violating protocol. All the checks must be made 659 before attempt to create socket. 660 */ 661 662 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 663 * and the incoming segment acknowledges something not yet 664 * sent (the segment carries an unacceptable ACK) ... 665 * a reset is sent." 666 * 667 * Invalid ACK: reset will be sent by listening socket 668 */ 669 if ((flg & TCP_FLAG_ACK) && 670 (TCP_SKB_CB(skb)->ack_seq != 671 tcp_rsk(req)->snt_isn + 1 + tcp_s_data_size(tcp_sk(sk)))) 672 return sk; 673 674 /* Also, it would be not so bad idea to check rcv_tsecr, which 675 * is essentially ACK extension and too early or too late values 676 * should cause reset in unsynchronized states. 677 */ 678 679 /* RFC793: "first check sequence number". */ 680 681 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 682 tcp_rsk(req)->rcv_isn + 1, tcp_rsk(req)->rcv_isn + 1 + req->rcv_wnd)) { 683 /* Out of window: send ACK and drop. */ 684 if (!(flg & TCP_FLAG_RST)) 685 req->rsk_ops->send_ack(sk, skb, req); 686 if (paws_reject) 687 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 688 return NULL; 689 } 690 691 /* In sequence, PAWS is OK. */ 692 693 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_isn + 1)) 694 req->ts_recent = tmp_opt.rcv_tsval; 695 696 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 697 /* Truncate SYN, it is out of window starting 698 at tcp_rsk(req)->rcv_isn + 1. */ 699 flg &= ~TCP_FLAG_SYN; 700 } 701 702 /* RFC793: "second check the RST bit" and 703 * "fourth, check the SYN bit" 704 */ 705 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 706 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 707 goto embryonic_reset; 708 } 709 710 /* ACK sequence verified above, just make sure ACK is 711 * set. If ACK not set, just silently drop the packet. 712 */ 713 if (!(flg & TCP_FLAG_ACK)) 714 return NULL; 715 716 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 717 if (req->retrans < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 718 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 719 inet_rsk(req)->acked = 1; 720 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 721 return NULL; 722 } 723 724 /* OK, ACK is valid, create big socket and 725 * feed this segment to it. It will repeat all 726 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 727 * ESTABLISHED STATE. If it will be dropped after 728 * socket is created, wait for troubles. 729 */ 730 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL); 731 if (child == NULL) 732 goto listen_overflow; 733 734 inet_csk_reqsk_queue_unlink(sk, req, prev); 735 inet_csk_reqsk_queue_removed(sk, req); 736 737 inet_csk_reqsk_queue_add(sk, req, child); 738 return child; 739 740 listen_overflow: 741 if (!sysctl_tcp_abort_on_overflow) { 742 inet_rsk(req)->acked = 1; 743 return NULL; 744 } 745 746 embryonic_reset: 747 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 748 if (!(flg & TCP_FLAG_RST)) 749 req->rsk_ops->send_reset(sk, skb); 750 751 inet_csk_reqsk_queue_drop(sk, req, prev); 752 return NULL; 753 } 754 EXPORT_SYMBOL(tcp_check_req); 755 756 /* 757 * Queue segment on the new socket if the new socket is active, 758 * otherwise we just shortcircuit this and continue with 759 * the new socket. 760 */ 761 762 int tcp_child_process(struct sock *parent, struct sock *child, 763 struct sk_buff *skb) 764 { 765 int ret = 0; 766 int state = child->sk_state; 767 768 if (!sock_owned_by_user(child)) { 769 ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb), 770 skb->len); 771 /* Wakeup parent, send SIGIO */ 772 if (state == TCP_SYN_RECV && child->sk_state != state) 773 parent->sk_data_ready(parent, 0); 774 } else { 775 /* Alas, it is possible again, because we do lookup 776 * in main socket hash table and lock on listening 777 * socket does not protect us more. 778 */ 779 __sk_add_backlog(child, skb); 780 } 781 782 bh_unlock_sock(child); 783 sock_put(child); 784 return ret; 785 } 786 EXPORT_SYMBOL(tcp_child_process); 787