tcp_minisocks.c revision a928630a2ffeaf6aa9a6b78456935b6ab1be3066
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_minisocks.c,v 1.15 2002/02/01 22:01:04 davem 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/mm.h> 24#include <linux/module.h> 25#include <linux/sysctl.h> 26#include <linux/workqueue.h> 27#include <net/tcp.h> 28#include <net/inet_common.h> 29#include <net/xfrm.h> 30 31#ifdef CONFIG_SYSCTL 32#define SYNC_INIT 0 /* let the user enable it */ 33#else 34#define SYNC_INIT 1 35#endif 36 37int sysctl_tcp_syncookies __read_mostly = SYNC_INIT; 38int sysctl_tcp_abort_on_overflow __read_mostly; 39 40struct inet_timewait_death_row tcp_death_row = { 41 .sysctl_max_tw_buckets = NR_FILE * 2, 42 .period = TCP_TIMEWAIT_LEN / INET_TWDR_TWKILL_SLOTS, 43 .death_lock = __SPIN_LOCK_UNLOCKED(tcp_death_row.death_lock), 44 .hashinfo = &tcp_hashinfo, 45 .tw_timer = TIMER_INITIALIZER(inet_twdr_hangman, 0, 46 (unsigned long)&tcp_death_row), 47 .twkill_work = __WORK_INITIALIZER(tcp_death_row.twkill_work, 48 inet_twdr_twkill_work, 49 &tcp_death_row), 50/* Short-time timewait calendar */ 51 52 .twcal_hand = -1, 53 .twcal_timer = TIMER_INITIALIZER(inet_twdr_twcal_tick, 0, 54 (unsigned long)&tcp_death_row), 55}; 56 57EXPORT_SYMBOL_GPL(tcp_death_row); 58 59static __inline__ int tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win) 60{ 61 if (seq == s_win) 62 return 1; 63 if (after(end_seq, s_win) && before(seq, e_win)) 64 return 1; 65 return (seq == e_win && seq == end_seq); 66} 67 68/* 69 * * Main purpose of TIME-WAIT state is to close connection gracefully, 70 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN 71 * (and, probably, tail of data) and one or more our ACKs are lost. 72 * * What is TIME-WAIT timeout? It is associated with maximal packet 73 * lifetime in the internet, which results in wrong conclusion, that 74 * it is set to catch "old duplicate segments" wandering out of their path. 75 * It is not quite correct. This timeout is calculated so that it exceeds 76 * maximal retransmission timeout enough to allow to lose one (or more) 77 * segments sent by peer and our ACKs. This time may be calculated from RTO. 78 * * When TIME-WAIT socket receives RST, it means that another end 79 * finally closed and we are allowed to kill TIME-WAIT too. 80 * * Second purpose of TIME-WAIT is catching old duplicate segments. 81 * Well, certainly it is pure paranoia, but if we load TIME-WAIT 82 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs. 83 * * If we invented some more clever way to catch duplicates 84 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs. 85 * 86 * The algorithm below is based on FORMAL INTERPRETATION of RFCs. 87 * When you compare it to RFCs, please, read section SEGMENT ARRIVES 88 * from the very beginning. 89 * 90 * NOTE. With recycling (and later with fin-wait-2) TW bucket 91 * is _not_ stateless. It means, that strictly speaking we must 92 * spinlock it. I do not want! Well, probability of misbehaviour 93 * is ridiculously low and, seems, we could use some mb() tricks 94 * to avoid misread sequence numbers, states etc. --ANK 95 */ 96enum tcp_tw_status 97tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb, 98 const struct tcphdr *th) 99{ 100 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 101 struct tcp_options_received tmp_opt; 102 int paws_reject = 0; 103 104 tmp_opt.saw_tstamp = 0; 105 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) { 106 tcp_parse_options(skb, &tmp_opt, 0); 107 108 if (tmp_opt.saw_tstamp) { 109 tmp_opt.ts_recent = tcptw->tw_ts_recent; 110 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp; 111 paws_reject = tcp_paws_check(&tmp_opt, th->rst); 112 } 113 } 114 115 if (tw->tw_substate == TCP_FIN_WAIT2) { 116 /* Just repeat all the checks of tcp_rcv_state_process() */ 117 118 /* Out of window, send ACK */ 119 if (paws_reject || 120 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 121 tcptw->tw_rcv_nxt, 122 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd)) 123 return TCP_TW_ACK; 124 125 if (th->rst) 126 goto kill; 127 128 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt)) 129 goto kill_with_rst; 130 131 /* Dup ACK? */ 132 if (!after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) || 133 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) { 134 inet_twsk_put(tw); 135 return TCP_TW_SUCCESS; 136 } 137 138 /* New data or FIN. If new data arrive after half-duplex close, 139 * reset. 140 */ 141 if (!th->fin || 142 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) { 143kill_with_rst: 144 inet_twsk_deschedule(tw, &tcp_death_row); 145 inet_twsk_put(tw); 146 return TCP_TW_RST; 147 } 148 149 /* FIN arrived, enter true time-wait state. */ 150 tw->tw_substate = TCP_TIME_WAIT; 151 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq; 152 if (tmp_opt.saw_tstamp) { 153 tcptw->tw_ts_recent_stamp = xtime.tv_sec; 154 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 155 } 156 157 /* I am shamed, but failed to make it more elegant. 158 * Yes, it is direct reference to IP, which is impossible 159 * to generalize to IPv6. Taking into account that IPv6 160 * do not understand recycling in any case, it not 161 * a big problem in practice. --ANK */ 162 if (tw->tw_family == AF_INET && 163 tcp_death_row.sysctl_tw_recycle && tcptw->tw_ts_recent_stamp && 164 tcp_v4_tw_remember_stamp(tw)) 165 inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout, 166 TCP_TIMEWAIT_LEN); 167 else 168 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 169 TCP_TIMEWAIT_LEN); 170 return TCP_TW_ACK; 171 } 172 173 /* 174 * Now real TIME-WAIT state. 175 * 176 * RFC 1122: 177 * "When a connection is [...] on TIME-WAIT state [...] 178 * [a TCP] MAY accept a new SYN from the remote TCP to 179 * reopen the connection directly, if it: 180 * 181 * (1) assigns its initial sequence number for the new 182 * connection to be larger than the largest sequence 183 * number it used on the previous connection incarnation, 184 * and 185 * 186 * (2) returns to TIME-WAIT state if the SYN turns out 187 * to be an old duplicate". 188 */ 189 190 if (!paws_reject && 191 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt && 192 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 193 /* In window segment, it may be only reset or bare ack. */ 194 195 if (th->rst) { 196 /* This is TIME_WAIT assassination, in two flavors. 197 * Oh well... nobody has a sufficient solution to this 198 * protocol bug yet. 199 */ 200 if (sysctl_tcp_rfc1337 == 0) { 201kill: 202 inet_twsk_deschedule(tw, &tcp_death_row); 203 inet_twsk_put(tw); 204 return TCP_TW_SUCCESS; 205 } 206 } 207 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 208 TCP_TIMEWAIT_LEN); 209 210 if (tmp_opt.saw_tstamp) { 211 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 212 tcptw->tw_ts_recent_stamp = xtime.tv_sec; 213 } 214 215 inet_twsk_put(tw); 216 return TCP_TW_SUCCESS; 217 } 218 219 /* Out of window segment. 220 221 All the segments are ACKed immediately. 222 223 The only exception is new SYN. We accept it, if it is 224 not old duplicate and we are not in danger to be killed 225 by delayed old duplicates. RFC check is that it has 226 newer sequence number works at rates <40Mbit/sec. 227 However, if paws works, it is reliable AND even more, 228 we even may relax silly seq space cutoff. 229 230 RED-PEN: we violate main RFC requirement, if this SYN will appear 231 old duplicate (i.e. we receive RST in reply to SYN-ACK), 232 we must return socket to time-wait state. It is not good, 233 but not fatal yet. 234 */ 235 236 if (th->syn && !th->rst && !th->ack && !paws_reject && 237 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) || 238 (tmp_opt.saw_tstamp && 239 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) { 240 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 241 if (isn == 0) 242 isn++; 243 TCP_SKB_CB(skb)->when = isn; 244 return TCP_TW_SYN; 245 } 246 247 if (paws_reject) 248 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 249 250 if(!th->rst) { 251 /* In this case we must reset the TIMEWAIT timer. 252 * 253 * If it is ACKless SYN it may be both old duplicate 254 * and new good SYN with random sequence number <rcv_nxt. 255 * Do not reschedule in the last case. 256 */ 257 if (paws_reject || th->ack) 258 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 259 TCP_TIMEWAIT_LEN); 260 261 /* Send ACK. Note, we do not put the bucket, 262 * it will be released by caller. 263 */ 264 return TCP_TW_ACK; 265 } 266 inet_twsk_put(tw); 267 return TCP_TW_SUCCESS; 268} 269 270/* 271 * Move a socket to time-wait or dead fin-wait-2 state. 272 */ 273void tcp_time_wait(struct sock *sk, int state, int timeo) 274{ 275 struct inet_timewait_sock *tw = NULL; 276 const struct inet_connection_sock *icsk = inet_csk(sk); 277 const struct tcp_sock *tp = tcp_sk(sk); 278 int recycle_ok = 0; 279 280 if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp) 281 recycle_ok = icsk->icsk_af_ops->remember_stamp(sk); 282 283 if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets) 284 tw = inet_twsk_alloc(sk, state); 285 286 if (tw != NULL) { 287 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 288 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 289 290 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 291 tcptw->tw_rcv_nxt = tp->rcv_nxt; 292 tcptw->tw_snd_nxt = tp->snd_nxt; 293 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 294 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 295 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 296 297#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) 298 if (tw->tw_family == PF_INET6) { 299 struct ipv6_pinfo *np = inet6_sk(sk); 300 struct inet6_timewait_sock *tw6; 301 302 tw->tw_ipv6_offset = inet6_tw_offset(sk->sk_prot); 303 tw6 = inet6_twsk((struct sock *)tw); 304 ipv6_addr_copy(&tw6->tw_v6_daddr, &np->daddr); 305 ipv6_addr_copy(&tw6->tw_v6_rcv_saddr, &np->rcv_saddr); 306 tw->tw_ipv6only = np->ipv6only; 307 } 308#endif 309 310#ifdef CONFIG_TCP_MD5SIG 311 /* 312 * The timewait bucket does not have the key DB from the 313 * sock structure. We just make a quick copy of the 314 * md5 key being used (if indeed we are using one) 315 * so the timewait ack generating code has the key. 316 */ 317 do { 318 struct tcp_md5sig_key *key; 319 memset(tcptw->tw_md5_key, 0, sizeof(tcptw->tw_md5_key)); 320 tcptw->tw_md5_keylen = 0; 321 key = tp->af_specific->md5_lookup(sk, sk); 322 if (key != NULL) { 323 memcpy(&tcptw->tw_md5_key, key->key, key->keylen); 324 tcptw->tw_md5_keylen = key->keylen; 325 if (tcp_alloc_md5sig_pool() == NULL) 326 BUG(); 327 } 328 } while(0); 329#endif 330 331 /* Linkage updates. */ 332 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo); 333 334 /* Get the TIME_WAIT timeout firing. */ 335 if (timeo < rto) 336 timeo = rto; 337 338 if (recycle_ok) { 339 tw->tw_timeout = rto; 340 } else { 341 tw->tw_timeout = TCP_TIMEWAIT_LEN; 342 if (state == TCP_TIME_WAIT) 343 timeo = TCP_TIMEWAIT_LEN; 344 } 345 346 inet_twsk_schedule(tw, &tcp_death_row, timeo, 347 TCP_TIMEWAIT_LEN); 348 inet_twsk_put(tw); 349 } else { 350 /* Sorry, if we're out of memory, just CLOSE this 351 * socket up. We've got bigger problems than 352 * non-graceful socket closings. 353 */ 354 if (net_ratelimit()) 355 printk(KERN_INFO "TCP: time wait bucket table overflow\n"); 356 } 357 358 tcp_update_metrics(sk); 359 tcp_done(sk); 360} 361 362void tcp_twsk_destructor(struct sock *sk) 363{ 364#ifdef CONFIG_TCP_MD5SIG 365 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 366 if (twsk->tw_md5_keylen) 367 tcp_put_md5sig_pool(); 368#endif 369} 370 371EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 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 */ 379struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb) 380{ 381 struct sock *newsk = inet_csk_clone(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(sk); 387 struct tcp_sock *newtp; 388 389 /* Now setup tcp_sock */ 390 newtp = tcp_sk(newsk); 391 newtp->pred_flags = 0; 392 newtp->rcv_nxt = treq->rcv_isn + 1; 393 newtp->snd_nxt = newtp->snd_una = newtp->snd_sml = treq->snt_isn + 1; 394 395 tcp_prequeue_init(newtp); 396 397 tcp_init_wl(newtp, treq->snt_isn, treq->rcv_isn); 398 399 newtp->srtt = 0; 400 newtp->mdev = TCP_TIMEOUT_INIT; 401 newicsk->icsk_rto = TCP_TIMEOUT_INIT; 402 403 newtp->packets_out = 0; 404 newtp->left_out = 0; 405 newtp->retrans_out = 0; 406 newtp->sacked_out = 0; 407 newtp->fackets_out = 0; 408 newtp->snd_ssthresh = 0x7fffffff; 409 410 /* So many TCP implementations out there (incorrectly) count the 411 * initial SYN frame in their delayed-ACK and congestion control 412 * algorithms that we must have the following bandaid to talk 413 * efficiently to them. -DaveM 414 */ 415 newtp->snd_cwnd = 2; 416 newtp->snd_cwnd_cnt = 0; 417 newtp->bytes_acked = 0; 418 419 newtp->frto_counter = 0; 420 newtp->frto_highmark = 0; 421 422 newicsk->icsk_ca_ops = &tcp_init_congestion_ops; 423 424 tcp_set_ca_state(newsk, TCP_CA_Open); 425 tcp_init_xmit_timers(newsk); 426 skb_queue_head_init(&newtp->out_of_order_queue); 427 newtp->rcv_wup = treq->rcv_isn + 1; 428 newtp->write_seq = treq->snt_isn + 1; 429 newtp->pushed_seq = newtp->write_seq; 430 newtp->copied_seq = treq->rcv_isn + 1; 431 432 newtp->rx_opt.saw_tstamp = 0; 433 434 newtp->rx_opt.dsack = 0; 435 newtp->rx_opt.eff_sacks = 0; 436 437 newtp->rx_opt.num_sacks = 0; 438 newtp->urg_data = 0; 439 440 if (sock_flag(newsk, SOCK_KEEPOPEN)) 441 inet_csk_reset_keepalive_timer(newsk, 442 keepalive_time_when(newtp)); 443 444 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 445 if((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) { 446 if (sysctl_tcp_fack) 447 newtp->rx_opt.sack_ok |= 2; 448 } 449 newtp->window_clamp = req->window_clamp; 450 newtp->rcv_ssthresh = req->rcv_wnd; 451 newtp->rcv_wnd = req->rcv_wnd; 452 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 453 if (newtp->rx_opt.wscale_ok) { 454 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 455 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 456 } else { 457 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 458 newtp->window_clamp = min(newtp->window_clamp, 65535U); 459 } 460 newtp->snd_wnd = ntohs(skb->h.th->window) << 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 = xtime.tv_sec; 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#ifdef CONFIG_TCP_MD5SIG 472 newtp->md5sig_info = NULL; /*XXX*/ 473 if (newtp->af_specific->md5_lookup(sk, newsk)) 474 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 475#endif 476 if (skb->len >= TCP_MIN_RCVMSS+newtp->tcp_header_len) 477 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 478 newtp->rx_opt.mss_clamp = req->mss; 479 TCP_ECN_openreq_child(newtp, req); 480 481 TCP_INC_STATS_BH(TCP_MIB_PASSIVEOPENS); 482 } 483 return newsk; 484} 485 486/* 487 * Process an incoming packet for SYN_RECV sockets represented 488 * as a request_sock. 489 */ 490 491struct sock *tcp_check_req(struct sock *sk,struct sk_buff *skb, 492 struct request_sock *req, 493 struct request_sock **prev) 494{ 495 struct tcphdr *th = skb->h.th; 496 u32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 497 int paws_reject = 0; 498 struct tcp_options_received tmp_opt; 499 struct sock *child; 500 501 tmp_opt.saw_tstamp = 0; 502 if (th->doff > (sizeof(struct tcphdr)>>2)) { 503 tcp_parse_options(skb, &tmp_opt, 0); 504 505 if (tmp_opt.saw_tstamp) { 506 tmp_opt.ts_recent = req->ts_recent; 507 /* We do not store true stamp, but it is not required, 508 * it can be estimated (approximately) 509 * from another data. 510 */ 511 tmp_opt.ts_recent_stamp = xtime.tv_sec - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans); 512 paws_reject = tcp_paws_check(&tmp_opt, th->rst); 513 } 514 } 515 516 /* Check for pure retransmitted SYN. */ 517 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 518 flg == TCP_FLAG_SYN && 519 !paws_reject) { 520 /* 521 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 522 * this case on figure 6 and figure 8, but formal 523 * protocol description says NOTHING. 524 * To be more exact, it says that we should send ACK, 525 * because this segment (at least, if it has no data) 526 * is out of window. 527 * 528 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 529 * describe SYN-RECV state. All the description 530 * is wrong, we cannot believe to it and should 531 * rely only on common sense and implementation 532 * experience. 533 * 534 * Enforce "SYN-ACK" according to figure 8, figure 6 535 * of RFC793, fixed by RFC1122. 536 */ 537 req->rsk_ops->rtx_syn_ack(sk, req, NULL); 538 return NULL; 539 } 540 541 /* Further reproduces section "SEGMENT ARRIVES" 542 for state SYN-RECEIVED of RFC793. 543 It is broken, however, it does not work only 544 when SYNs are crossed. 545 546 You would think that SYN crossing is impossible here, since 547 we should have a SYN_SENT socket (from connect()) on our end, 548 but this is not true if the crossed SYNs were sent to both 549 ends by a malicious third party. We must defend against this, 550 and to do that we first verify the ACK (as per RFC793, page 551 36) and reset if it is invalid. Is this a true full defense? 552 To convince ourselves, let us consider a way in which the ACK 553 test can still pass in this 'malicious crossed SYNs' case. 554 Malicious sender sends identical SYNs (and thus identical sequence 555 numbers) to both A and B: 556 557 A: gets SYN, seq=7 558 B: gets SYN, seq=7 559 560 By our good fortune, both A and B select the same initial 561 send sequence number of seven :-) 562 563 A: sends SYN|ACK, seq=7, ack_seq=8 564 B: sends SYN|ACK, seq=7, ack_seq=8 565 566 So we are now A eating this SYN|ACK, ACK test passes. So 567 does sequence test, SYN is truncated, and thus we consider 568 it a bare ACK. 569 570 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 571 bare ACK. Otherwise, we create an established connection. Both 572 ends (listening sockets) accept the new incoming connection and try 573 to talk to each other. 8-) 574 575 Note: This case is both harmless, and rare. Possibility is about the 576 same as us discovering intelligent life on another plant tomorrow. 577 578 But generally, we should (RFC lies!) to accept ACK 579 from SYNACK both here and in tcp_rcv_state_process(). 580 tcp_rcv_state_process() does not, hence, we do not too. 581 582 Note that the case is absolutely generic: 583 we cannot optimize anything here without 584 violating protocol. All the checks must be made 585 before attempt to create socket. 586 */ 587 588 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 589 * and the incoming segment acknowledges something not yet 590 * sent (the segment carries an unacceptable ACK) ... 591 * a reset is sent." 592 * 593 * Invalid ACK: reset will be sent by listening socket 594 */ 595 if ((flg & TCP_FLAG_ACK) && 596 (TCP_SKB_CB(skb)->ack_seq != tcp_rsk(req)->snt_isn + 1)) 597 return sk; 598 599 /* Also, it would be not so bad idea to check rcv_tsecr, which 600 * is essentially ACK extension and too early or too late values 601 * should cause reset in unsynchronized states. 602 */ 603 604 /* RFC793: "first check sequence number". */ 605 606 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 607 tcp_rsk(req)->rcv_isn + 1, tcp_rsk(req)->rcv_isn + 1 + req->rcv_wnd)) { 608 /* Out of window: send ACK and drop. */ 609 if (!(flg & TCP_FLAG_RST)) 610 req->rsk_ops->send_ack(skb, req); 611 if (paws_reject) 612 NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); 613 return NULL; 614 } 615 616 /* In sequence, PAWS is OK. */ 617 618 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_isn + 1)) 619 req->ts_recent = tmp_opt.rcv_tsval; 620 621 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 622 /* Truncate SYN, it is out of window starting 623 at tcp_rsk(req)->rcv_isn + 1. */ 624 flg &= ~TCP_FLAG_SYN; 625 } 626 627 /* RFC793: "second check the RST bit" and 628 * "fourth, check the SYN bit" 629 */ 630 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 631 TCP_INC_STATS_BH(TCP_MIB_ATTEMPTFAILS); 632 goto embryonic_reset; 633 } 634 635 /* ACK sequence verified above, just make sure ACK is 636 * set. If ACK not set, just silently drop the packet. 637 */ 638 if (!(flg & TCP_FLAG_ACK)) 639 return NULL; 640 641 /* If TCP_DEFER_ACCEPT is set, drop bare ACK. */ 642 if (inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 643 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 644 inet_rsk(req)->acked = 1; 645 return NULL; 646 } 647 648 /* OK, ACK is valid, create big socket and 649 * feed this segment to it. It will repeat all 650 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 651 * ESTABLISHED STATE. If it will be dropped after 652 * socket is created, wait for troubles. 653 */ 654 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, 655 req, NULL); 656 if (child == NULL) 657 goto listen_overflow; 658#ifdef CONFIG_TCP_MD5SIG 659 else { 660 /* Copy over the MD5 key from the original socket */ 661 struct tcp_md5sig_key *key; 662 struct tcp_sock *tp = tcp_sk(sk); 663 key = tp->af_specific->md5_lookup(sk, child); 664 if (key != NULL) { 665 /* 666 * We're using one, so create a matching key on the 667 * newsk structure. If we fail to get memory then we 668 * end up not copying the key across. Shucks. 669 */ 670 char *newkey = kmalloc(key->keylen, GFP_ATOMIC); 671 if (newkey) { 672 if (!tcp_alloc_md5sig_pool()) 673 BUG(); 674 memcpy(newkey, key->key, key->keylen); 675 tp->af_specific->md5_add(child, child, 676 newkey, 677 key->keylen); 678 } 679 } 680 } 681#endif 682 683 inet_csk_reqsk_queue_unlink(sk, req, prev); 684 inet_csk_reqsk_queue_removed(sk, req); 685 686 inet_csk_reqsk_queue_add(sk, req, child); 687 return child; 688 689 listen_overflow: 690 if (!sysctl_tcp_abort_on_overflow) { 691 inet_rsk(req)->acked = 1; 692 return NULL; 693 } 694 695 embryonic_reset: 696 NET_INC_STATS_BH(LINUX_MIB_EMBRYONICRSTS); 697 if (!(flg & TCP_FLAG_RST)) 698 req->rsk_ops->send_reset(sk, skb); 699 700 inet_csk_reqsk_queue_drop(sk, req, prev); 701 return NULL; 702} 703 704/* 705 * Queue segment on the new socket if the new socket is active, 706 * otherwise we just shortcircuit this and continue with 707 * the new socket. 708 */ 709 710int tcp_child_process(struct sock *parent, struct sock *child, 711 struct sk_buff *skb) 712{ 713 int ret = 0; 714 int state = child->sk_state; 715 716 if (!sock_owned_by_user(child)) { 717 ret = tcp_rcv_state_process(child, skb, skb->h.th, skb->len); 718 719 /* Wakeup parent, send SIGIO */ 720 if (state == TCP_SYN_RECV && child->sk_state != state) 721 parent->sk_data_ready(parent, 0); 722 } else { 723 /* Alas, it is possible again, because we do lookup 724 * in main socket hash table and lock on listening 725 * socket does not protect us more. 726 */ 727 sk_add_backlog(child, skb); 728 } 729 730 bh_unlock_sock(child); 731 sock_put(child); 732 return ret; 733} 734 735EXPORT_SYMBOL(tcp_check_req); 736EXPORT_SYMBOL(tcp_child_process); 737EXPORT_SYMBOL(tcp_create_openreq_child); 738EXPORT_SYMBOL(tcp_timewait_state_process); 739