tcp_minisocks.c revision 4bc2f18ba4f22a90ab593c0a580fc9a19c4777b6
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 30int sysctl_tcp_syncookies __read_mostly = 1; 31EXPORT_SYMBOL(sysctl_tcp_syncookies); 32 33int sysctl_tcp_abort_on_overflow __read_mostly; 34 35struct 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}; 50EXPORT_SYMBOL_GPL(tcp_death_row); 51 52static __inline__ int tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win) 53{ 54 if (seq == s_win) 55 return 1; 56 if (after(end_seq, s_win) && before(seq, e_win)) 57 return 1; 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 */ 89enum tcp_tw_status 90tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb, 91 const struct tcphdr *th) 92{ 93 struct tcp_options_received tmp_opt; 94 u8 *hash_location; 95 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 96 int paws_reject = 0; 97 98 tmp_opt.saw_tstamp = 0; 99 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) { 100 tcp_parse_options(skb, &tmp_opt, &hash_location, 0); 101 102 if (tmp_opt.saw_tstamp) { 103 tmp_opt.ts_recent = tcptw->tw_ts_recent; 104 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp; 105 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 106 } 107 } 108 109 if (tw->tw_substate == TCP_FIN_WAIT2) { 110 /* Just repeat all the checks of tcp_rcv_state_process() */ 111 112 /* Out of window, send ACK */ 113 if (paws_reject || 114 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 115 tcptw->tw_rcv_nxt, 116 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd)) 117 return TCP_TW_ACK; 118 119 if (th->rst) 120 goto kill; 121 122 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt)) 123 goto kill_with_rst; 124 125 /* Dup ACK? */ 126 if (!th->ack || 127 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) || 128 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) { 129 inet_twsk_put(tw); 130 return TCP_TW_SUCCESS; 131 } 132 133 /* New data or FIN. If new data arrive after half-duplex close, 134 * reset. 135 */ 136 if (!th->fin || 137 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) { 138kill_with_rst: 139 inet_twsk_deschedule(tw, &tcp_death_row); 140 inet_twsk_put(tw); 141 return TCP_TW_RST; 142 } 143 144 /* FIN arrived, enter true time-wait state. */ 145 tw->tw_substate = TCP_TIME_WAIT; 146 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq; 147 if (tmp_opt.saw_tstamp) { 148 tcptw->tw_ts_recent_stamp = get_seconds(); 149 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 150 } 151 152 /* I am shamed, but failed to make it more elegant. 153 * Yes, it is direct reference to IP, which is impossible 154 * to generalize to IPv6. Taking into account that IPv6 155 * do not understand recycling in any case, it not 156 * a big problem in practice. --ANK */ 157 if (tw->tw_family == AF_INET && 158 tcp_death_row.sysctl_tw_recycle && tcptw->tw_ts_recent_stamp && 159 tcp_v4_tw_remember_stamp(tw)) 160 inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout, 161 TCP_TIMEWAIT_LEN); 162 else 163 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 164 TCP_TIMEWAIT_LEN); 165 return TCP_TW_ACK; 166 } 167 168 /* 169 * Now real TIME-WAIT state. 170 * 171 * RFC 1122: 172 * "When a connection is [...] on TIME-WAIT state [...] 173 * [a TCP] MAY accept a new SYN from the remote TCP to 174 * reopen the connection directly, if it: 175 * 176 * (1) assigns its initial sequence number for the new 177 * connection to be larger than the largest sequence 178 * number it used on the previous connection incarnation, 179 * and 180 * 181 * (2) returns to TIME-WAIT state if the SYN turns out 182 * to be an old duplicate". 183 */ 184 185 if (!paws_reject && 186 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt && 187 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 188 /* In window segment, it may be only reset or bare ack. */ 189 190 if (th->rst) { 191 /* This is TIME_WAIT assassination, in two flavors. 192 * Oh well... nobody has a sufficient solution to this 193 * protocol bug yet. 194 */ 195 if (sysctl_tcp_rfc1337 == 0) { 196kill: 197 inet_twsk_deschedule(tw, &tcp_death_row); 198 inet_twsk_put(tw); 199 return TCP_TW_SUCCESS; 200 } 201 } 202 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 203 TCP_TIMEWAIT_LEN); 204 205 if (tmp_opt.saw_tstamp) { 206 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 207 tcptw->tw_ts_recent_stamp = get_seconds(); 208 } 209 210 inet_twsk_put(tw); 211 return TCP_TW_SUCCESS; 212 } 213 214 /* Out of window segment. 215 216 All the segments are ACKed immediately. 217 218 The only exception is new SYN. We accept it, if it is 219 not old duplicate and we are not in danger to be killed 220 by delayed old duplicates. RFC check is that it has 221 newer sequence number works at rates <40Mbit/sec. 222 However, if paws works, it is reliable AND even more, 223 we even may relax silly seq space cutoff. 224 225 RED-PEN: we violate main RFC requirement, if this SYN will appear 226 old duplicate (i.e. we receive RST in reply to SYN-ACK), 227 we must return socket to time-wait state. It is not good, 228 but not fatal yet. 229 */ 230 231 if (th->syn && !th->rst && !th->ack && !paws_reject && 232 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) || 233 (tmp_opt.saw_tstamp && 234 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) { 235 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 236 if (isn == 0) 237 isn++; 238 TCP_SKB_CB(skb)->when = isn; 239 return TCP_TW_SYN; 240 } 241 242 if (paws_reject) 243 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED); 244 245 if (!th->rst) { 246 /* In this case we must reset the TIMEWAIT timer. 247 * 248 * If it is ACKless SYN it may be both old duplicate 249 * and new good SYN with random sequence number <rcv_nxt. 250 * Do not reschedule in the last case. 251 */ 252 if (paws_reject || th->ack) 253 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 254 TCP_TIMEWAIT_LEN); 255 256 /* Send ACK. Note, we do not put the bucket, 257 * it will be released by caller. 258 */ 259 return TCP_TW_ACK; 260 } 261 inet_twsk_put(tw); 262 return TCP_TW_SUCCESS; 263} 264EXPORT_SYMBOL(tcp_timewait_state_process); 265 266/* 267 * Move a socket to time-wait or dead fin-wait-2 state. 268 */ 269void tcp_time_wait(struct sock *sk, int state, int timeo) 270{ 271 struct inet_timewait_sock *tw = NULL; 272 const struct inet_connection_sock *icsk = inet_csk(sk); 273 const struct tcp_sock *tp = tcp_sk(sk); 274 int recycle_ok = 0; 275 276 if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp) 277 recycle_ok = icsk->icsk_af_ops->remember_stamp(sk); 278 279 if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets) 280 tw = inet_twsk_alloc(sk, state); 281 282 if (tw != NULL) { 283 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 284 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 285 286 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 287 tcptw->tw_rcv_nxt = tp->rcv_nxt; 288 tcptw->tw_snd_nxt = tp->snd_nxt; 289 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 290 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 291 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 292 293#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) 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 ipv6_addr_copy(&tw6->tw_v6_daddr, &np->daddr); 301 ipv6_addr_copy(&tw6->tw_v6_rcv_saddr, &np->rcv_saddr); 302 tw->tw_ipv6only = np->ipv6only; 303 } 304#endif 305 306#ifdef CONFIG_TCP_MD5SIG 307 /* 308 * The timewait bucket does not have the key DB from the 309 * sock structure. We just make a quick copy of the 310 * md5 key being used (if indeed we are using one) 311 * so the timewait ack generating code has the key. 312 */ 313 do { 314 struct tcp_md5sig_key *key; 315 memset(tcptw->tw_md5_key, 0, sizeof(tcptw->tw_md5_key)); 316 tcptw->tw_md5_keylen = 0; 317 key = tp->af_specific->md5_lookup(sk, sk); 318 if (key != NULL) { 319 memcpy(&tcptw->tw_md5_key, key->key, key->keylen); 320 tcptw->tw_md5_keylen = key->keylen; 321 if (tcp_alloc_md5sig_pool(sk) == NULL) 322 BUG(); 323 } 324 } while (0); 325#endif 326 327 /* Linkage updates. */ 328 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo); 329 330 /* Get the TIME_WAIT timeout firing. */ 331 if (timeo < rto) 332 timeo = rto; 333 334 if (recycle_ok) { 335 tw->tw_timeout = rto; 336 } else { 337 tw->tw_timeout = TCP_TIMEWAIT_LEN; 338 if (state == TCP_TIME_WAIT) 339 timeo = TCP_TIMEWAIT_LEN; 340 } 341 342 inet_twsk_schedule(tw, &tcp_death_row, timeo, 343 TCP_TIMEWAIT_LEN); 344 inet_twsk_put(tw); 345 } else { 346 /* Sorry, if we're out of memory, just CLOSE this 347 * socket up. We've got bigger problems than 348 * non-graceful socket closings. 349 */ 350 LIMIT_NETDEBUG(KERN_INFO "TCP: time wait bucket table overflow\n"); 351 } 352 353 tcp_update_metrics(sk); 354 tcp_done(sk); 355} 356 357void tcp_twsk_destructor(struct sock *sk) 358{ 359#ifdef CONFIG_TCP_MD5SIG 360 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 361 if (twsk->tw_md5_keylen) 362 tcp_free_md5sig_pool(); 363#endif 364} 365EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 366 367static 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 */ 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(newsk); 387 struct tcp_sock *newtp = tcp_sk(newsk); 388 struct tcp_sock *oldtp = tcp_sk(sk); 389 struct tcp_cookie_values *oldcvp = oldtp->cookie_values; 390 391 /* TCP Cookie Transactions require space for the cookie pair, 392 * as it differs for each connection. There is no need to 393 * copy any s_data_payload stored at the original socket. 394 * Failure will prevent resuming the connection. 395 * 396 * Presumed copied, in order of appearance: 397 * cookie_in_always, cookie_out_never 398 */ 399 if (oldcvp != NULL) { 400 struct tcp_cookie_values *newcvp = 401 kzalloc(sizeof(*newtp->cookie_values), 402 GFP_ATOMIC); 403 404 if (newcvp != NULL) { 405 kref_init(&newcvp->kref); 406 newcvp->cookie_desired = 407 oldcvp->cookie_desired; 408 newtp->cookie_values = newcvp; 409 } else { 410 /* Not Yet Implemented */ 411 newtp->cookie_values = NULL; 412 } 413 } 414 415 /* Now setup tcp_sock */ 416 newtp->pred_flags = 0; 417 418 newtp->rcv_wup = newtp->copied_seq = 419 newtp->rcv_nxt = treq->rcv_isn + 1; 420 421 newtp->snd_sml = newtp->snd_una = 422 newtp->snd_nxt = newtp->snd_up = 423 treq->snt_isn + 1 + tcp_s_data_size(oldtp); 424 425 tcp_prequeue_init(newtp); 426 427 tcp_init_wl(newtp, treq->rcv_isn); 428 429 newtp->srtt = 0; 430 newtp->mdev = TCP_TIMEOUT_INIT; 431 newicsk->icsk_rto = TCP_TIMEOUT_INIT; 432 433 newtp->packets_out = 0; 434 newtp->retrans_out = 0; 435 newtp->sacked_out = 0; 436 newtp->fackets_out = 0; 437 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 438 439 /* So many TCP implementations out there (incorrectly) count the 440 * initial SYN frame in their delayed-ACK and congestion control 441 * algorithms that we must have the following bandaid to talk 442 * efficiently to them. -DaveM 443 */ 444 newtp->snd_cwnd = 2; 445 newtp->snd_cwnd_cnt = 0; 446 newtp->bytes_acked = 0; 447 448 newtp->frto_counter = 0; 449 newtp->frto_highmark = 0; 450 451 newicsk->icsk_ca_ops = &tcp_init_congestion_ops; 452 453 tcp_set_ca_state(newsk, TCP_CA_Open); 454 tcp_init_xmit_timers(newsk); 455 skb_queue_head_init(&newtp->out_of_order_queue); 456 newtp->write_seq = newtp->pushed_seq = 457 treq->snt_isn + 1 + tcp_s_data_size(oldtp); 458 459 newtp->rx_opt.saw_tstamp = 0; 460 461 newtp->rx_opt.dsack = 0; 462 newtp->rx_opt.num_sacks = 0; 463 464 newtp->urg_data = 0; 465 466 if (sock_flag(newsk, SOCK_KEEPOPEN)) 467 inet_csk_reset_keepalive_timer(newsk, 468 keepalive_time_when(newtp)); 469 470 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 471 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) { 472 if (sysctl_tcp_fack) 473 tcp_enable_fack(newtp); 474 } 475 newtp->window_clamp = req->window_clamp; 476 newtp->rcv_ssthresh = req->rcv_wnd; 477 newtp->rcv_wnd = req->rcv_wnd; 478 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 479 if (newtp->rx_opt.wscale_ok) { 480 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 481 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 482 } else { 483 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 484 newtp->window_clamp = min(newtp->window_clamp, 65535U); 485 } 486 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) << 487 newtp->rx_opt.snd_wscale); 488 newtp->max_window = newtp->snd_wnd; 489 490 if (newtp->rx_opt.tstamp_ok) { 491 newtp->rx_opt.ts_recent = req->ts_recent; 492 newtp->rx_opt.ts_recent_stamp = get_seconds(); 493 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 494 } else { 495 newtp->rx_opt.ts_recent_stamp = 0; 496 newtp->tcp_header_len = sizeof(struct tcphdr); 497 } 498#ifdef CONFIG_TCP_MD5SIG 499 newtp->md5sig_info = NULL; /*XXX*/ 500 if (newtp->af_specific->md5_lookup(sk, newsk)) 501 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 502#endif 503 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 504 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 505 newtp->rx_opt.mss_clamp = req->mss; 506 TCP_ECN_openreq_child(newtp, req); 507 508 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS); 509 } 510 return newsk; 511} 512EXPORT_SYMBOL(tcp_create_openreq_child); 513 514/* 515 * Process an incoming packet for SYN_RECV sockets represented 516 * as a request_sock. 517 */ 518 519struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 520 struct request_sock *req, 521 struct request_sock **prev) 522{ 523 struct tcp_options_received tmp_opt; 524 u8 *hash_location; 525 struct sock *child; 526 const struct tcphdr *th = tcp_hdr(skb); 527 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 528 int paws_reject = 0; 529 530 tmp_opt.saw_tstamp = 0; 531 if (th->doff > (sizeof(struct tcphdr)>>2)) { 532 tcp_parse_options(skb, &tmp_opt, &hash_location, 0); 533 534 if (tmp_opt.saw_tstamp) { 535 tmp_opt.ts_recent = req->ts_recent; 536 /* We do not store true stamp, but it is not required, 537 * it can be estimated (approximately) 538 * from another data. 539 */ 540 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans); 541 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 542 } 543 } 544 545 /* Check for pure retransmitted SYN. */ 546 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 547 flg == TCP_FLAG_SYN && 548 !paws_reject) { 549 /* 550 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 551 * this case on figure 6 and figure 8, but formal 552 * protocol description says NOTHING. 553 * To be more exact, it says that we should send ACK, 554 * because this segment (at least, if it has no data) 555 * is out of window. 556 * 557 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 558 * describe SYN-RECV state. All the description 559 * is wrong, we cannot believe to it and should 560 * rely only on common sense and implementation 561 * experience. 562 * 563 * Enforce "SYN-ACK" according to figure 8, figure 6 564 * of RFC793, fixed by RFC1122. 565 */ 566 req->rsk_ops->rtx_syn_ack(sk, req, NULL); 567 return NULL; 568 } 569 570 /* Further reproduces section "SEGMENT ARRIVES" 571 for state SYN-RECEIVED of RFC793. 572 It is broken, however, it does not work only 573 when SYNs are crossed. 574 575 You would think that SYN crossing is impossible here, since 576 we should have a SYN_SENT socket (from connect()) on our end, 577 but this is not true if the crossed SYNs were sent to both 578 ends by a malicious third party. We must defend against this, 579 and to do that we first verify the ACK (as per RFC793, page 580 36) and reset if it is invalid. Is this a true full defense? 581 To convince ourselves, let us consider a way in which the ACK 582 test can still pass in this 'malicious crossed SYNs' case. 583 Malicious sender sends identical SYNs (and thus identical sequence 584 numbers) to both A and B: 585 586 A: gets SYN, seq=7 587 B: gets SYN, seq=7 588 589 By our good fortune, both A and B select the same initial 590 send sequence number of seven :-) 591 592 A: sends SYN|ACK, seq=7, ack_seq=8 593 B: sends SYN|ACK, seq=7, ack_seq=8 594 595 So we are now A eating this SYN|ACK, ACK test passes. So 596 does sequence test, SYN is truncated, and thus we consider 597 it a bare ACK. 598 599 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 600 bare ACK. Otherwise, we create an established connection. Both 601 ends (listening sockets) accept the new incoming connection and try 602 to talk to each other. 8-) 603 604 Note: This case is both harmless, and rare. Possibility is about the 605 same as us discovering intelligent life on another plant tomorrow. 606 607 But generally, we should (RFC lies!) to accept ACK 608 from SYNACK both here and in tcp_rcv_state_process(). 609 tcp_rcv_state_process() does not, hence, we do not too. 610 611 Note that the case is absolutely generic: 612 we cannot optimize anything here without 613 violating protocol. All the checks must be made 614 before attempt to create socket. 615 */ 616 617 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 618 * and the incoming segment acknowledges something not yet 619 * sent (the segment carries an unacceptable ACK) ... 620 * a reset is sent." 621 * 622 * Invalid ACK: reset will be sent by listening socket 623 */ 624 if ((flg & TCP_FLAG_ACK) && 625 (TCP_SKB_CB(skb)->ack_seq != 626 tcp_rsk(req)->snt_isn + 1 + tcp_s_data_size(tcp_sk(sk)))) 627 return sk; 628 629 /* Also, it would be not so bad idea to check rcv_tsecr, which 630 * is essentially ACK extension and too early or too late values 631 * should cause reset in unsynchronized states. 632 */ 633 634 /* RFC793: "first check sequence number". */ 635 636 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 637 tcp_rsk(req)->rcv_isn + 1, tcp_rsk(req)->rcv_isn + 1 + req->rcv_wnd)) { 638 /* Out of window: send ACK and drop. */ 639 if (!(flg & TCP_FLAG_RST)) 640 req->rsk_ops->send_ack(sk, skb, req); 641 if (paws_reject) 642 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 643 return NULL; 644 } 645 646 /* In sequence, PAWS is OK. */ 647 648 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_isn + 1)) 649 req->ts_recent = tmp_opt.rcv_tsval; 650 651 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 652 /* Truncate SYN, it is out of window starting 653 at tcp_rsk(req)->rcv_isn + 1. */ 654 flg &= ~TCP_FLAG_SYN; 655 } 656 657 /* RFC793: "second check the RST bit" and 658 * "fourth, check the SYN bit" 659 */ 660 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 661 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 662 goto embryonic_reset; 663 } 664 665 /* ACK sequence verified above, just make sure ACK is 666 * set. If ACK not set, just silently drop the packet. 667 */ 668 if (!(flg & TCP_FLAG_ACK)) 669 return NULL; 670 671 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 672 if (req->retrans < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 673 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 674 inet_rsk(req)->acked = 1; 675 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 676 return NULL; 677 } 678 679 /* OK, ACK is valid, create big socket and 680 * feed this segment to it. It will repeat all 681 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 682 * ESTABLISHED STATE. If it will be dropped after 683 * socket is created, wait for troubles. 684 */ 685 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL); 686 if (child == NULL) 687 goto listen_overflow; 688 689 inet_csk_reqsk_queue_unlink(sk, req, prev); 690 inet_csk_reqsk_queue_removed(sk, req); 691 692 inet_csk_reqsk_queue_add(sk, req, child); 693 return child; 694 695listen_overflow: 696 if (!sysctl_tcp_abort_on_overflow) { 697 inet_rsk(req)->acked = 1; 698 return NULL; 699 } 700 701embryonic_reset: 702 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 703 if (!(flg & TCP_FLAG_RST)) 704 req->rsk_ops->send_reset(sk, skb); 705 706 inet_csk_reqsk_queue_drop(sk, req, prev); 707 return NULL; 708} 709EXPORT_SYMBOL(tcp_check_req); 710 711/* 712 * Queue segment on the new socket if the new socket is active, 713 * otherwise we just shortcircuit this and continue with 714 * the new socket. 715 */ 716 717int tcp_child_process(struct sock *parent, struct sock *child, 718 struct sk_buff *skb) 719{ 720 int ret = 0; 721 int state = child->sk_state; 722 723 if (!sock_owned_by_user(child)) { 724 ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb), 725 skb->len); 726 /* Wakeup parent, send SIGIO */ 727 if (state == TCP_SYN_RECV && child->sk_state != state) 728 parent->sk_data_ready(parent, 0); 729 } else { 730 /* Alas, it is possible again, because we do lookup 731 * in main socket hash table and lock on listening 732 * socket does not protect us more. 733 */ 734 __sk_add_backlog(child, skb); 735 } 736 737 bh_unlock_sock(child); 738 sock_put(child); 739 return ret; 740} 741EXPORT_SYMBOL(tcp_child_process); 742