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 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win) 53 { 54 if (seq == s_win) 55 return true; 56 if (after(end_seq, s_win) && before(seq, e_win)) 57 return true; 58 return seq == e_win && seq == end_seq; 59 } 60 61 /* 62 * * Main purpose of TIME-WAIT state is to close connection gracefully, 63 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN 64 * (and, probably, tail of data) and one or more our ACKs are lost. 65 * * What is TIME-WAIT timeout? It is associated with maximal packet 66 * lifetime in the internet, which results in wrong conclusion, that 67 * it is set to catch "old duplicate segments" wandering out of their path. 68 * It is not quite correct. This timeout is calculated so that it exceeds 69 * maximal retransmission timeout enough to allow to lose one (or more) 70 * segments sent by peer and our ACKs. This time may be calculated from RTO. 71 * * When TIME-WAIT socket receives RST, it means that another end 72 * finally closed and we are allowed to kill TIME-WAIT too. 73 * * Second purpose of TIME-WAIT is catching old duplicate segments. 74 * Well, certainly it is pure paranoia, but if we load TIME-WAIT 75 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs. 76 * * If we invented some more clever way to catch duplicates 77 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs. 78 * 79 * The algorithm below is based on FORMAL INTERPRETATION of RFCs. 80 * When you compare it to RFCs, please, read section SEGMENT ARRIVES 81 * from the very beginning. 82 * 83 * NOTE. With recycling (and later with fin-wait-2) TW bucket 84 * is _not_ stateless. It means, that strictly speaking we must 85 * spinlock it. I do not want! Well, probability of misbehaviour 86 * is ridiculously low and, seems, we could use some mb() tricks 87 * to avoid misread sequence numbers, states etc. --ANK 88 * 89 * We don't need to initialize tmp_out.sack_ok as we don't use the results 90 */ 91 enum tcp_tw_status 92 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb, 93 const struct tcphdr *th) 94 { 95 struct tcp_options_received tmp_opt; 96 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 97 bool paws_reject = false; 98 99 tmp_opt.saw_tstamp = 0; 100 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) { 101 tcp_parse_options(skb, &tmp_opt, 0, NULL); 102 103 if (tmp_opt.saw_tstamp) { 104 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset; 105 tmp_opt.ts_recent = tcptw->tw_ts_recent; 106 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp; 107 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 108 } 109 } 110 111 if (tw->tw_substate == TCP_FIN_WAIT2) { 112 /* Just repeat all the checks of tcp_rcv_state_process() */ 113 114 /* Out of window, send ACK */ 115 if (paws_reject || 116 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 117 tcptw->tw_rcv_nxt, 118 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd)) 119 return TCP_TW_ACK; 120 121 if (th->rst) 122 goto kill; 123 124 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt)) 125 goto kill_with_rst; 126 127 /* Dup ACK? */ 128 if (!th->ack || 129 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) || 130 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) { 131 inet_twsk_put(tw); 132 return TCP_TW_SUCCESS; 133 } 134 135 /* New data or FIN. If new data arrive after half-duplex close, 136 * reset. 137 */ 138 if (!th->fin || 139 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) { 140 kill_with_rst: 141 inet_twsk_deschedule(tw, &tcp_death_row); 142 inet_twsk_put(tw); 143 return TCP_TW_RST; 144 } 145 146 /* FIN arrived, enter true time-wait state. */ 147 tw->tw_substate = TCP_TIME_WAIT; 148 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq; 149 if (tmp_opt.saw_tstamp) { 150 tcptw->tw_ts_recent_stamp = get_seconds(); 151 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 152 } 153 154 if (tcp_death_row.sysctl_tw_recycle && 155 tcptw->tw_ts_recent_stamp && 156 tcp_tw_remember_stamp(tw)) 157 inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout, 158 TCP_TIMEWAIT_LEN); 159 else 160 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 161 TCP_TIMEWAIT_LEN); 162 return TCP_TW_ACK; 163 } 164 165 /* 166 * Now real TIME-WAIT state. 167 * 168 * RFC 1122: 169 * "When a connection is [...] on TIME-WAIT state [...] 170 * [a TCP] MAY accept a new SYN from the remote TCP to 171 * reopen the connection directly, if it: 172 * 173 * (1) assigns its initial sequence number for the new 174 * connection to be larger than the largest sequence 175 * number it used on the previous connection incarnation, 176 * and 177 * 178 * (2) returns to TIME-WAIT state if the SYN turns out 179 * to be an old duplicate". 180 */ 181 182 if (!paws_reject && 183 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt && 184 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 185 /* In window segment, it may be only reset or bare ack. */ 186 187 if (th->rst) { 188 /* This is TIME_WAIT assassination, in two flavors. 189 * Oh well... nobody has a sufficient solution to this 190 * protocol bug yet. 191 */ 192 if (sysctl_tcp_rfc1337 == 0) { 193 kill: 194 inet_twsk_deschedule(tw, &tcp_death_row); 195 inet_twsk_put(tw); 196 return TCP_TW_SUCCESS; 197 } 198 } 199 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 200 TCP_TIMEWAIT_LEN); 201 202 if (tmp_opt.saw_tstamp) { 203 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 204 tcptw->tw_ts_recent_stamp = get_seconds(); 205 } 206 207 inet_twsk_put(tw); 208 return TCP_TW_SUCCESS; 209 } 210 211 /* Out of window segment. 212 213 All the segments are ACKed immediately. 214 215 The only exception is new SYN. We accept it, if it is 216 not old duplicate and we are not in danger to be killed 217 by delayed old duplicates. RFC check is that it has 218 newer sequence number works at rates <40Mbit/sec. 219 However, if paws works, it is reliable AND even more, 220 we even may relax silly seq space cutoff. 221 222 RED-PEN: we violate main RFC requirement, if this SYN will appear 223 old duplicate (i.e. we receive RST in reply to SYN-ACK), 224 we must return socket to time-wait state. It is not good, 225 but not fatal yet. 226 */ 227 228 if (th->syn && !th->rst && !th->ack && !paws_reject && 229 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) || 230 (tmp_opt.saw_tstamp && 231 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) { 232 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 233 if (isn == 0) 234 isn++; 235 TCP_SKB_CB(skb)->when = isn; 236 return TCP_TW_SYN; 237 } 238 239 if (paws_reject) 240 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED); 241 242 if (!th->rst) { 243 /* In this case we must reset the TIMEWAIT timer. 244 * 245 * If it is ACKless SYN it may be both old duplicate 246 * and new good SYN with random sequence number <rcv_nxt. 247 * Do not reschedule in the last case. 248 */ 249 if (paws_reject || th->ack) 250 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 251 TCP_TIMEWAIT_LEN); 252 253 /* Send ACK. Note, we do not put the bucket, 254 * it will be released by caller. 255 */ 256 return TCP_TW_ACK; 257 } 258 inet_twsk_put(tw); 259 return TCP_TW_SUCCESS; 260 } 261 EXPORT_SYMBOL(tcp_timewait_state_process); 262 263 /* 264 * Move a socket to time-wait or dead fin-wait-2 state. 265 */ 266 void tcp_time_wait(struct sock *sk, int state, int timeo) 267 { 268 struct inet_timewait_sock *tw = NULL; 269 const struct inet_connection_sock *icsk = inet_csk(sk); 270 const struct tcp_sock *tp = tcp_sk(sk); 271 bool recycle_ok = false; 272 273 if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp) 274 recycle_ok = tcp_remember_stamp(sk); 275 276 if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets) 277 tw = inet_twsk_alloc(sk, state); 278 279 if (tw != NULL) { 280 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 281 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 282 struct inet_sock *inet = inet_sk(sk); 283 284 tw->tw_transparent = inet->transparent; 285 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 286 tcptw->tw_rcv_nxt = tp->rcv_nxt; 287 tcptw->tw_snd_nxt = tp->snd_nxt; 288 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 289 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 290 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 291 tcptw->tw_ts_offset = tp->tsoffset; 292 293 #if IS_ENABLED(CONFIG_IPV6) 294 if (tw->tw_family == PF_INET6) { 295 struct ipv6_pinfo *np = inet6_sk(sk); 296 struct inet6_timewait_sock *tw6; 297 298 tw->tw_ipv6_offset = inet6_tw_offset(sk->sk_prot); 299 tw6 = inet6_twsk((struct sock *)tw); 300 tw6->tw_v6_daddr = np->daddr; 301 tw6->tw_v6_rcv_saddr = np->rcv_saddr; 302 tw->tw_tclass = np->tclass; 303 tw->tw_ipv6only = np->ipv6only; 304 } 305 #endif 306 307 #ifdef CONFIG_TCP_MD5SIG 308 /* 309 * The timewait bucket does not have the key DB from the 310 * sock structure. We just make a quick copy of the 311 * md5 key being used (if indeed we are using one) 312 * so the timewait ack generating code has the key. 313 */ 314 do { 315 struct tcp_md5sig_key *key; 316 tcptw->tw_md5_key = NULL; 317 key = tp->af_specific->md5_lookup(sk, sk); 318 if (key != NULL) { 319 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC); 320 if (tcptw->tw_md5_key && !tcp_alloc_md5sig_pool()) 321 BUG(); 322 } 323 } while (0); 324 #endif 325 326 /* Linkage updates. */ 327 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo); 328 329 /* Get the TIME_WAIT timeout firing. */ 330 if (timeo < rto) 331 timeo = rto; 332 333 if (recycle_ok) { 334 tw->tw_timeout = rto; 335 } else { 336 tw->tw_timeout = TCP_TIMEWAIT_LEN; 337 if (state == TCP_TIME_WAIT) 338 timeo = TCP_TIMEWAIT_LEN; 339 } 340 341 inet_twsk_schedule(tw, &tcp_death_row, timeo, 342 TCP_TIMEWAIT_LEN); 343 inet_twsk_put(tw); 344 } else { 345 /* Sorry, if we're out of memory, just CLOSE this 346 * socket up. We've got bigger problems than 347 * non-graceful socket closings. 348 */ 349 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW); 350 } 351 352 tcp_update_metrics(sk); 353 tcp_done(sk); 354 } 355 356 void tcp_twsk_destructor(struct sock *sk) 357 { 358 #ifdef CONFIG_TCP_MD5SIG 359 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 360 361 if (twsk->tw_md5_key) 362 kfree_rcu(twsk->tw_md5_key, rcu); 363 #endif 364 } 365 EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 366 367 static inline void TCP_ECN_openreq_child(struct tcp_sock *tp, 368 struct request_sock *req) 369 { 370 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 371 } 372 373 /* This is not only more efficient than what we used to do, it eliminates 374 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 375 * 376 * Actually, we could lots of memory writes here. tp of listening 377 * socket contains all necessary default parameters. 378 */ 379 struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb) 380 { 381 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 382 383 if (newsk != NULL) { 384 const struct inet_request_sock *ireq = inet_rsk(req); 385 struct tcp_request_sock *treq = tcp_rsk(req); 386 struct inet_connection_sock *newicsk = inet_csk(newsk); 387 struct tcp_sock *newtp = tcp_sk(newsk); 388 389 /* Now setup tcp_sock */ 390 newtp->pred_flags = 0; 391 392 newtp->rcv_wup = newtp->copied_seq = 393 newtp->rcv_nxt = treq->rcv_isn + 1; 394 395 newtp->snd_sml = newtp->snd_una = 396 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1; 397 398 tcp_prequeue_init(newtp); 399 INIT_LIST_HEAD(&newtp->tsq_node); 400 401 tcp_init_wl(newtp, treq->rcv_isn); 402 403 newtp->srtt = 0; 404 newtp->mdev = TCP_TIMEOUT_INIT; 405 newicsk->icsk_rto = TCP_TIMEOUT_INIT; 406 407 newtp->packets_out = 0; 408 newtp->retrans_out = 0; 409 newtp->sacked_out = 0; 410 newtp->fackets_out = 0; 411 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 412 tcp_enable_early_retrans(newtp); 413 newtp->tlp_high_seq = 0; 414 415 /* So many TCP implementations out there (incorrectly) count the 416 * initial SYN frame in their delayed-ACK and congestion control 417 * algorithms that we must have the following bandaid to talk 418 * efficiently to them. -DaveM 419 */ 420 newtp->snd_cwnd = TCP_INIT_CWND; 421 newtp->snd_cwnd_cnt = 0; 422 423 if (newicsk->icsk_ca_ops != &tcp_init_congestion_ops && 424 !try_module_get(newicsk->icsk_ca_ops->owner)) 425 newicsk->icsk_ca_ops = &tcp_init_congestion_ops; 426 427 tcp_set_ca_state(newsk, TCP_CA_Open); 428 tcp_init_xmit_timers(newsk); 429 skb_queue_head_init(&newtp->out_of_order_queue); 430 newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1; 431 432 newtp->rx_opt.saw_tstamp = 0; 433 434 newtp->rx_opt.dsack = 0; 435 newtp->rx_opt.num_sacks = 0; 436 437 newtp->urg_data = 0; 438 439 if (sock_flag(newsk, SOCK_KEEPOPEN)) 440 inet_csk_reset_keepalive_timer(newsk, 441 keepalive_time_when(newtp)); 442 443 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 444 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) { 445 if (sysctl_tcp_fack) 446 tcp_enable_fack(newtp); 447 } 448 newtp->window_clamp = req->window_clamp; 449 newtp->rcv_ssthresh = req->rcv_wnd; 450 newtp->rcv_wnd = req->rcv_wnd; 451 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 452 if (newtp->rx_opt.wscale_ok) { 453 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 454 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 455 } else { 456 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 457 newtp->window_clamp = min(newtp->window_clamp, 65535U); 458 } 459 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) << 460 newtp->rx_opt.snd_wscale); 461 newtp->max_window = newtp->snd_wnd; 462 463 if (newtp->rx_opt.tstamp_ok) { 464 newtp->rx_opt.ts_recent = req->ts_recent; 465 newtp->rx_opt.ts_recent_stamp = get_seconds(); 466 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 467 } else { 468 newtp->rx_opt.ts_recent_stamp = 0; 469 newtp->tcp_header_len = sizeof(struct tcphdr); 470 } 471 newtp->tsoffset = 0; 472 #ifdef CONFIG_TCP_MD5SIG 473 newtp->md5sig_info = NULL; /*XXX*/ 474 if (newtp->af_specific->md5_lookup(sk, newsk)) 475 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 476 #endif 477 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 478 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 479 newtp->rx_opt.mss_clamp = req->mss; 480 TCP_ECN_openreq_child(newtp, req); 481 newtp->fastopen_rsk = NULL; 482 newtp->syn_data_acked = 0; 483 484 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS); 485 } 486 return newsk; 487 } 488 EXPORT_SYMBOL(tcp_create_openreq_child); 489 490 /* 491 * Process an incoming packet for SYN_RECV sockets represented as a 492 * request_sock. Normally sk is the listener socket but for TFO it 493 * points to the child socket. 494 * 495 * XXX (TFO) - The current impl contains a special check for ack 496 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better? 497 * 498 * We don't need to initialize tmp_opt.sack_ok as we don't use the results 499 */ 500 501 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 502 struct request_sock *req, 503 struct request_sock **prev, 504 bool fastopen) 505 { 506 struct tcp_options_received tmp_opt; 507 struct sock *child; 508 const struct tcphdr *th = tcp_hdr(skb); 509 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 510 bool paws_reject = false; 511 512 BUG_ON(fastopen == (sk->sk_state == TCP_LISTEN)); 513 514 tmp_opt.saw_tstamp = 0; 515 if (th->doff > (sizeof(struct tcphdr)>>2)) { 516 tcp_parse_options(skb, &tmp_opt, 0, NULL); 517 518 if (tmp_opt.saw_tstamp) { 519 tmp_opt.ts_recent = req->ts_recent; 520 /* We do not store true stamp, but it is not required, 521 * it can be estimated (approximately) 522 * from another data. 523 */ 524 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout); 525 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 526 } 527 } 528 529 /* Check for pure retransmitted SYN. */ 530 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 531 flg == TCP_FLAG_SYN && 532 !paws_reject) { 533 /* 534 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 535 * this case on figure 6 and figure 8, but formal 536 * protocol description says NOTHING. 537 * To be more exact, it says that we should send ACK, 538 * because this segment (at least, if it has no data) 539 * is out of window. 540 * 541 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 542 * describe SYN-RECV state. All the description 543 * is wrong, we cannot believe to it and should 544 * rely only on common sense and implementation 545 * experience. 546 * 547 * Enforce "SYN-ACK" according to figure 8, figure 6 548 * of RFC793, fixed by RFC1122. 549 * 550 * Note that even if there is new data in the SYN packet 551 * they will be thrown away too. 552 * 553 * Reset timer after retransmitting SYNACK, similar to 554 * the idea of fast retransmit in recovery. 555 */ 556 if (!inet_rtx_syn_ack(sk, req)) 557 req->expires = min(TCP_TIMEOUT_INIT << req->num_timeout, 558 TCP_RTO_MAX) + jiffies; 559 return NULL; 560 } 561 562 /* Further reproduces section "SEGMENT ARRIVES" 563 for state SYN-RECEIVED of RFC793. 564 It is broken, however, it does not work only 565 when SYNs are crossed. 566 567 You would think that SYN crossing is impossible here, since 568 we should have a SYN_SENT socket (from connect()) on our end, 569 but this is not true if the crossed SYNs were sent to both 570 ends by a malicious third party. We must defend against this, 571 and to do that we first verify the ACK (as per RFC793, page 572 36) and reset if it is invalid. Is this a true full defense? 573 To convince ourselves, let us consider a way in which the ACK 574 test can still pass in this 'malicious crossed SYNs' case. 575 Malicious sender sends identical SYNs (and thus identical sequence 576 numbers) to both A and B: 577 578 A: gets SYN, seq=7 579 B: gets SYN, seq=7 580 581 By our good fortune, both A and B select the same initial 582 send sequence number of seven :-) 583 584 A: sends SYN|ACK, seq=7, ack_seq=8 585 B: sends SYN|ACK, seq=7, ack_seq=8 586 587 So we are now A eating this SYN|ACK, ACK test passes. So 588 does sequence test, SYN is truncated, and thus we consider 589 it a bare ACK. 590 591 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 592 bare ACK. Otherwise, we create an established connection. Both 593 ends (listening sockets) accept the new incoming connection and try 594 to talk to each other. 8-) 595 596 Note: This case is both harmless, and rare. Possibility is about the 597 same as us discovering intelligent life on another plant tomorrow. 598 599 But generally, we should (RFC lies!) to accept ACK 600 from SYNACK both here and in tcp_rcv_state_process(). 601 tcp_rcv_state_process() does not, hence, we do not too. 602 603 Note that the case is absolutely generic: 604 we cannot optimize anything here without 605 violating protocol. All the checks must be made 606 before attempt to create socket. 607 */ 608 609 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 610 * and the incoming segment acknowledges something not yet 611 * sent (the segment carries an unacceptable ACK) ... 612 * a reset is sent." 613 * 614 * Invalid ACK: reset will be sent by listening socket. 615 * Note that the ACK validity check for a Fast Open socket is done 616 * elsewhere and is checked directly against the child socket rather 617 * than req because user data may have been sent out. 618 */ 619 if ((flg & TCP_FLAG_ACK) && !fastopen && 620 (TCP_SKB_CB(skb)->ack_seq != 621 tcp_rsk(req)->snt_isn + 1)) 622 return sk; 623 624 /* Also, it would be not so bad idea to check rcv_tsecr, which 625 * is essentially ACK extension and too early or too late values 626 * should cause reset in unsynchronized states. 627 */ 628 629 /* RFC793: "first check sequence number". */ 630 631 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 632 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rcv_wnd)) { 633 /* Out of window: send ACK and drop. */ 634 if (!(flg & TCP_FLAG_RST)) 635 req->rsk_ops->send_ack(sk, skb, req); 636 if (paws_reject) 637 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 638 return NULL; 639 } 640 641 /* In sequence, PAWS is OK. */ 642 643 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt)) 644 req->ts_recent = tmp_opt.rcv_tsval; 645 646 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 647 /* Truncate SYN, it is out of window starting 648 at tcp_rsk(req)->rcv_isn + 1. */ 649 flg &= ~TCP_FLAG_SYN; 650 } 651 652 /* RFC793: "second check the RST bit" and 653 * "fourth, check the SYN bit" 654 */ 655 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 656 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 657 goto embryonic_reset; 658 } 659 660 /* ACK sequence verified above, just make sure ACK is 661 * set. If ACK not set, just silently drop the packet. 662 * 663 * XXX (TFO) - if we ever allow "data after SYN", the 664 * following check needs to be removed. 665 */ 666 if (!(flg & TCP_FLAG_ACK)) 667 return NULL; 668 669 /* Got ACK for our SYNACK, so update baseline for SYNACK RTT sample. */ 670 if (tmp_opt.saw_tstamp && tmp_opt.rcv_tsecr) 671 tcp_rsk(req)->snt_synack = tmp_opt.rcv_tsecr; 672 else if (req->num_retrans) /* don't take RTT sample if retrans && ~TS */ 673 tcp_rsk(req)->snt_synack = 0; 674 675 /* For Fast Open no more processing is needed (sk is the 676 * child socket). 677 */ 678 if (fastopen) 679 return sk; 680 681 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 682 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 683 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 684 inet_rsk(req)->acked = 1; 685 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 686 return NULL; 687 } 688 689 /* OK, ACK is valid, create big socket and 690 * feed this segment to it. It will repeat all 691 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 692 * ESTABLISHED STATE. If it will be dropped after 693 * socket is created, wait for troubles. 694 */ 695 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL); 696 if (child == NULL) 697 goto listen_overflow; 698 699 inet_csk_reqsk_queue_unlink(sk, req, prev); 700 inet_csk_reqsk_queue_removed(sk, req); 701 702 inet_csk_reqsk_queue_add(sk, req, child); 703 return child; 704 705 listen_overflow: 706 if (!sysctl_tcp_abort_on_overflow) { 707 inet_rsk(req)->acked = 1; 708 return NULL; 709 } 710 711 embryonic_reset: 712 if (!(flg & TCP_FLAG_RST)) { 713 /* Received a bad SYN pkt - for TFO We try not to reset 714 * the local connection unless it's really necessary to 715 * avoid becoming vulnerable to outside attack aiming at 716 * resetting legit local connections. 717 */ 718 req->rsk_ops->send_reset(sk, skb); 719 } else if (fastopen) { /* received a valid RST pkt */ 720 reqsk_fastopen_remove(sk, req, true); 721 tcp_reset(sk); 722 } 723 if (!fastopen) { 724 inet_csk_reqsk_queue_drop(sk, req, prev); 725 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 726 } 727 return NULL; 728 } 729 EXPORT_SYMBOL(tcp_check_req); 730 731 /* 732 * Queue segment on the new socket if the new socket is active, 733 * otherwise we just shortcircuit this and continue with 734 * the new socket. 735 * 736 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV 737 * when entering. But other states are possible due to a race condition 738 * where after __inet_lookup_established() fails but before the listener 739 * locked is obtained, other packets cause the same connection to 740 * be created. 741 */ 742 743 int tcp_child_process(struct sock *parent, struct sock *child, 744 struct sk_buff *skb) 745 { 746 int ret = 0; 747 int state = child->sk_state; 748 749 if (!sock_owned_by_user(child)) { 750 ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb), 751 skb->len); 752 /* Wakeup parent, send SIGIO */ 753 if (state == TCP_SYN_RECV && child->sk_state != state) 754 parent->sk_data_ready(parent, 0); 755 } else { 756 /* Alas, it is possible again, because we do lookup 757 * in main socket hash table and lock on listening 758 * socket does not protect us more. 759 */ 760 __sk_add_backlog(child, skb); 761 } 762 763 bh_unlock_sock(child); 764 sock_put(child); 765 return ret; 766 } 767 EXPORT_SYMBOL(tcp_child_process); 768