tcp_minisocks.c revision 4308fc58dceda40fac8a8d9c05b7cfba0a6bbed3
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 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 */ 91enum tcp_tw_status 92tcp_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 const u8 *hash_location; 97 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 98 bool paws_reject = false; 99 100 tmp_opt.saw_tstamp = 0; 101 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) { 102 tcp_parse_options(skb, &tmp_opt, &hash_location, 0, NULL); 103 104 if (tmp_opt.saw_tstamp) { 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) { 140kill_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) { 193kill: 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} 261EXPORT_SYMBOL(tcp_timewait_state_process); 262 263/* 264 * Move a socket to time-wait or dead fin-wait-2 state. 265 */ 266void 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; 291 292#if IS_ENABLED(CONFIG_IPV6) 293 if (tw->tw_family == PF_INET6) { 294 struct ipv6_pinfo *np = inet6_sk(sk); 295 struct inet6_timewait_sock *tw6; 296 297 tw->tw_ipv6_offset = inet6_tw_offset(sk->sk_prot); 298 tw6 = inet6_twsk((struct sock *)tw); 299 tw6->tw_v6_daddr = np->daddr; 300 tw6->tw_v6_rcv_saddr = np->rcv_saddr; 301 tw->tw_tclass = np->tclass; 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 tcptw->tw_md5_key = NULL; 316 key = tp->af_specific->md5_lookup(sk, sk); 317 if (key != NULL) { 318 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC); 319 if (tcptw->tw_md5_key && tcp_alloc_md5sig_pool(sk) == NULL) 320 BUG(); 321 } 322 } while (0); 323#endif 324 325 /* Linkage updates. */ 326 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo); 327 328 /* Get the TIME_WAIT timeout firing. */ 329 if (timeo < rto) 330 timeo = rto; 331 332 if (recycle_ok) { 333 tw->tw_timeout = rto; 334 } else { 335 tw->tw_timeout = TCP_TIMEWAIT_LEN; 336 if (state == TCP_TIME_WAIT) 337 timeo = TCP_TIMEWAIT_LEN; 338 } 339 340 inet_twsk_schedule(tw, &tcp_death_row, timeo, 341 TCP_TIMEWAIT_LEN); 342 inet_twsk_put(tw); 343 } else { 344 /* Sorry, if we're out of memory, just CLOSE this 345 * socket up. We've got bigger problems than 346 * non-graceful socket closings. 347 */ 348 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW); 349 } 350 351 tcp_update_metrics(sk); 352 tcp_done(sk); 353} 354 355void tcp_twsk_destructor(struct sock *sk) 356{ 357#ifdef CONFIG_TCP_MD5SIG 358 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 359 360 if (twsk->tw_md5_key) { 361 tcp_free_md5sig_pool(); 362 kfree_rcu(twsk->tw_md5_key, rcu); 363 } 364#endif 365} 366EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 367 368static inline void TCP_ECN_openreq_child(struct tcp_sock *tp, 369 struct request_sock *req) 370{ 371 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 372} 373 374/* This is not only more efficient than what we used to do, it eliminates 375 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 376 * 377 * Actually, we could lots of memory writes here. tp of listening 378 * socket contains all necessary default parameters. 379 */ 380struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb) 381{ 382 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 383 384 if (newsk != NULL) { 385 const struct inet_request_sock *ireq = inet_rsk(req); 386 struct tcp_request_sock *treq = tcp_rsk(req); 387 struct inet_connection_sock *newicsk = inet_csk(newsk); 388 struct tcp_sock *newtp = tcp_sk(newsk); 389 struct tcp_sock *oldtp = tcp_sk(sk); 390 struct tcp_cookie_values *oldcvp = oldtp->cookie_values; 391 392 /* TCP Cookie Transactions require space for the cookie pair, 393 * as it differs for each connection. There is no need to 394 * copy any s_data_payload stored at the original socket. 395 * Failure will prevent resuming the connection. 396 * 397 * Presumed copied, in order of appearance: 398 * cookie_in_always, cookie_out_never 399 */ 400 if (oldcvp != NULL) { 401 struct tcp_cookie_values *newcvp = 402 kzalloc(sizeof(*newtp->cookie_values), 403 GFP_ATOMIC); 404 405 if (newcvp != NULL) { 406 kref_init(&newcvp->kref); 407 newcvp->cookie_desired = 408 oldcvp->cookie_desired; 409 newtp->cookie_values = newcvp; 410 } else { 411 /* Not Yet Implemented */ 412 newtp->cookie_values = NULL; 413 } 414 } 415 416 /* Now setup tcp_sock */ 417 newtp->pred_flags = 0; 418 419 newtp->rcv_wup = newtp->copied_seq = 420 newtp->rcv_nxt = treq->rcv_isn + 1; 421 422 newtp->snd_sml = newtp->snd_una = 423 newtp->snd_nxt = newtp->snd_up = 424 treq->snt_isn + 1 + tcp_s_data_size(oldtp); 425 426 tcp_prequeue_init(newtp); 427 INIT_LIST_HEAD(&newtp->tsq_node); 428 429 tcp_init_wl(newtp, treq->rcv_isn); 430 431 newtp->srtt = 0; 432 newtp->mdev = TCP_TIMEOUT_INIT; 433 newicsk->icsk_rto = TCP_TIMEOUT_INIT; 434 435 newtp->packets_out = 0; 436 newtp->retrans_out = 0; 437 newtp->sacked_out = 0; 438 newtp->fackets_out = 0; 439 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 440 tcp_enable_early_retrans(newtp); 441 442 /* So many TCP implementations out there (incorrectly) count the 443 * initial SYN frame in their delayed-ACK and congestion control 444 * algorithms that we must have the following bandaid to talk 445 * efficiently to them. -DaveM 446 */ 447 newtp->snd_cwnd = TCP_INIT_CWND; 448 newtp->snd_cwnd_cnt = 0; 449 newtp->bytes_acked = 0; 450 451 newtp->frto_counter = 0; 452 newtp->frto_highmark = 0; 453 454 if (newicsk->icsk_ca_ops != &tcp_init_congestion_ops && 455 !try_module_get(newicsk->icsk_ca_ops->owner)) 456 newicsk->icsk_ca_ops = &tcp_init_congestion_ops; 457 458 tcp_set_ca_state(newsk, TCP_CA_Open); 459 tcp_init_xmit_timers(newsk); 460 skb_queue_head_init(&newtp->out_of_order_queue); 461 newtp->write_seq = newtp->pushed_seq = 462 treq->snt_isn + 1 + tcp_s_data_size(oldtp); 463 464 newtp->rx_opt.saw_tstamp = 0; 465 466 newtp->rx_opt.dsack = 0; 467 newtp->rx_opt.num_sacks = 0; 468 469 newtp->urg_data = 0; 470 471 if (sock_flag(newsk, SOCK_KEEPOPEN)) 472 inet_csk_reset_keepalive_timer(newsk, 473 keepalive_time_when(newtp)); 474 475 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 476 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) { 477 if (sysctl_tcp_fack) 478 tcp_enable_fack(newtp); 479 } 480 newtp->window_clamp = req->window_clamp; 481 newtp->rcv_ssthresh = req->rcv_wnd; 482 newtp->rcv_wnd = req->rcv_wnd; 483 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 484 if (newtp->rx_opt.wscale_ok) { 485 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 486 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 487 } else { 488 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 489 newtp->window_clamp = min(newtp->window_clamp, 65535U); 490 } 491 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) << 492 newtp->rx_opt.snd_wscale); 493 newtp->max_window = newtp->snd_wnd; 494 495 if (newtp->rx_opt.tstamp_ok) { 496 newtp->rx_opt.ts_recent = req->ts_recent; 497 newtp->rx_opt.ts_recent_stamp = get_seconds(); 498 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 499 } else { 500 newtp->rx_opt.ts_recent_stamp = 0; 501 newtp->tcp_header_len = sizeof(struct tcphdr); 502 } 503#ifdef CONFIG_TCP_MD5SIG 504 newtp->md5sig_info = NULL; /*XXX*/ 505 if (newtp->af_specific->md5_lookup(sk, newsk)) 506 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 507#endif 508 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 509 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 510 newtp->rx_opt.mss_clamp = req->mss; 511 TCP_ECN_openreq_child(newtp, req); 512 newtp->fastopen_rsk = NULL; 513 514 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS); 515 } 516 return newsk; 517} 518EXPORT_SYMBOL(tcp_create_openreq_child); 519 520/* 521 * Process an incoming packet for SYN_RECV sockets represented as a 522 * request_sock. Normally sk is the listener socket but for TFO it 523 * points to the child socket. 524 * 525 * XXX (TFO) - The current impl contains a special check for ack 526 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better? 527 * 528 * We don't need to initialize tmp_opt.sack_ok as we don't use the results 529 */ 530 531struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 532 struct request_sock *req, 533 struct request_sock **prev, 534 bool fastopen) 535{ 536 struct tcp_options_received tmp_opt; 537 const u8 *hash_location; 538 struct sock *child; 539 const struct tcphdr *th = tcp_hdr(skb); 540 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 541 bool paws_reject = false; 542 543 BUG_ON(fastopen == (sk->sk_state == TCP_LISTEN)); 544 545 tmp_opt.saw_tstamp = 0; 546 if (th->doff > (sizeof(struct tcphdr)>>2)) { 547 tcp_parse_options(skb, &tmp_opt, &hash_location, 0, NULL); 548 549 if (tmp_opt.saw_tstamp) { 550 tmp_opt.ts_recent = req->ts_recent; 551 /* We do not store true stamp, but it is not required, 552 * it can be estimated (approximately) 553 * from another data. 554 */ 555 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans); 556 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 557 } 558 } 559 560 /* Check for pure retransmitted SYN. */ 561 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 562 flg == TCP_FLAG_SYN && 563 !paws_reject) { 564 /* 565 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 566 * this case on figure 6 and figure 8, but formal 567 * protocol description says NOTHING. 568 * To be more exact, it says that we should send ACK, 569 * because this segment (at least, if it has no data) 570 * is out of window. 571 * 572 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 573 * describe SYN-RECV state. All the description 574 * is wrong, we cannot believe to it and should 575 * rely only on common sense and implementation 576 * experience. 577 * 578 * Enforce "SYN-ACK" according to figure 8, figure 6 579 * of RFC793, fixed by RFC1122. 580 * 581 * Note that even if there is new data in the SYN packet 582 * they will be thrown away too. 583 */ 584 req->rsk_ops->rtx_syn_ack(sk, req, NULL); 585 return NULL; 586 } 587 588 /* Further reproduces section "SEGMENT ARRIVES" 589 for state SYN-RECEIVED of RFC793. 590 It is broken, however, it does not work only 591 when SYNs are crossed. 592 593 You would think that SYN crossing is impossible here, since 594 we should have a SYN_SENT socket (from connect()) on our end, 595 but this is not true if the crossed SYNs were sent to both 596 ends by a malicious third party. We must defend against this, 597 and to do that we first verify the ACK (as per RFC793, page 598 36) and reset if it is invalid. Is this a true full defense? 599 To convince ourselves, let us consider a way in which the ACK 600 test can still pass in this 'malicious crossed SYNs' case. 601 Malicious sender sends identical SYNs (and thus identical sequence 602 numbers) to both A and B: 603 604 A: gets SYN, seq=7 605 B: gets SYN, seq=7 606 607 By our good fortune, both A and B select the same initial 608 send sequence number of seven :-) 609 610 A: sends SYN|ACK, seq=7, ack_seq=8 611 B: sends SYN|ACK, seq=7, ack_seq=8 612 613 So we are now A eating this SYN|ACK, ACK test passes. So 614 does sequence test, SYN is truncated, and thus we consider 615 it a bare ACK. 616 617 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 618 bare ACK. Otherwise, we create an established connection. Both 619 ends (listening sockets) accept the new incoming connection and try 620 to talk to each other. 8-) 621 622 Note: This case is both harmless, and rare. Possibility is about the 623 same as us discovering intelligent life on another plant tomorrow. 624 625 But generally, we should (RFC lies!) to accept ACK 626 from SYNACK both here and in tcp_rcv_state_process(). 627 tcp_rcv_state_process() does not, hence, we do not too. 628 629 Note that the case is absolutely generic: 630 we cannot optimize anything here without 631 violating protocol. All the checks must be made 632 before attempt to create socket. 633 */ 634 635 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 636 * and the incoming segment acknowledges something not yet 637 * sent (the segment carries an unacceptable ACK) ... 638 * a reset is sent." 639 * 640 * Invalid ACK: reset will be sent by listening socket. 641 * Note that the ACK validity check for a Fast Open socket is done 642 * elsewhere and is checked directly against the child socket rather 643 * than req because user data may have been sent out. 644 */ 645 if ((flg & TCP_FLAG_ACK) && !fastopen && 646 (TCP_SKB_CB(skb)->ack_seq != 647 tcp_rsk(req)->snt_isn + 1 + tcp_s_data_size(tcp_sk(sk)))) 648 return sk; 649 650 /* Also, it would be not so bad idea to check rcv_tsecr, which 651 * is essentially ACK extension and too early or too late values 652 * should cause reset in unsynchronized states. 653 */ 654 655 /* RFC793: "first check sequence number". */ 656 657 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 658 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rcv_wnd)) { 659 /* Out of window: send ACK and drop. */ 660 if (!(flg & TCP_FLAG_RST)) 661 req->rsk_ops->send_ack(sk, skb, req); 662 if (paws_reject) 663 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 664 return NULL; 665 } 666 667 /* In sequence, PAWS is OK. */ 668 669 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt)) 670 req->ts_recent = tmp_opt.rcv_tsval; 671 672 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 673 /* Truncate SYN, it is out of window starting 674 at tcp_rsk(req)->rcv_isn + 1. */ 675 flg &= ~TCP_FLAG_SYN; 676 } 677 678 /* RFC793: "second check the RST bit" and 679 * "fourth, check the SYN bit" 680 */ 681 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 682 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 683 goto embryonic_reset; 684 } 685 686 /* ACK sequence verified above, just make sure ACK is 687 * set. If ACK not set, just silently drop the packet. 688 * 689 * XXX (TFO) - if we ever allow "data after SYN", the 690 * following check needs to be removed. 691 */ 692 if (!(flg & TCP_FLAG_ACK)) 693 return NULL; 694 695 /* For Fast Open no more processing is needed (sk is the 696 * child socket). 697 */ 698 if (fastopen) 699 return sk; 700 701 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 702 if (req->retrans < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 703 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 704 inet_rsk(req)->acked = 1; 705 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 706 return NULL; 707 } 708 if (tmp_opt.saw_tstamp && tmp_opt.rcv_tsecr) 709 tcp_rsk(req)->snt_synack = tmp_opt.rcv_tsecr; 710 else if (req->retrans) /* don't take RTT sample if retrans && ~TS */ 711 tcp_rsk(req)->snt_synack = 0; 712 713 /* OK, ACK is valid, create big socket and 714 * feed this segment to it. It will repeat all 715 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 716 * ESTABLISHED STATE. If it will be dropped after 717 * socket is created, wait for troubles. 718 */ 719 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL); 720 if (child == NULL) 721 goto listen_overflow; 722 723 inet_csk_reqsk_queue_unlink(sk, req, prev); 724 inet_csk_reqsk_queue_removed(sk, req); 725 726 inet_csk_reqsk_queue_add(sk, req, child); 727 return child; 728 729listen_overflow: 730 if (!sysctl_tcp_abort_on_overflow) { 731 inet_rsk(req)->acked = 1; 732 return NULL; 733 } 734 735embryonic_reset: 736 if (!(flg & TCP_FLAG_RST)) { 737 /* Received a bad SYN pkt - for TFO We try not to reset 738 * the local connection unless it's really necessary to 739 * avoid becoming vulnerable to outside attack aiming at 740 * resetting legit local connections. 741 */ 742 req->rsk_ops->send_reset(sk, skb); 743 } else if (fastopen) { /* received a valid RST pkt */ 744 reqsk_fastopen_remove(sk, req, true); 745 tcp_reset(sk); 746 } 747 if (!fastopen) { 748 inet_csk_reqsk_queue_drop(sk, req, prev); 749 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 750 } 751 return NULL; 752} 753EXPORT_SYMBOL(tcp_check_req); 754 755/* 756 * Queue segment on the new socket if the new socket is active, 757 * otherwise we just shortcircuit this and continue with 758 * the new socket. 759 * 760 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV 761 * when entering. But other states are possible due to a race condition 762 * where after __inet_lookup_established() fails but before the listener 763 * locked is obtained, other packets cause the same connection to 764 * be created. 765 */ 766 767int tcp_child_process(struct sock *parent, struct sock *child, 768 struct sk_buff *skb) 769{ 770 int ret = 0; 771 int state = child->sk_state; 772 773 if (!sock_owned_by_user(child)) { 774 ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb), 775 skb->len); 776 /* Wakeup parent, send SIGIO */ 777 if (state == TCP_SYN_RECV && child->sk_state != state) 778 parent->sk_data_ready(parent, 0); 779 } else { 780 /* Alas, it is possible again, because we do lookup 781 * in main socket hash table and lock on listening 782 * socket does not protect us more. 783 */ 784 __sk_add_backlog(child, skb); 785 } 786 787 bh_unlock_sock(child); 788 sock_put(child); 789 return ret; 790} 791EXPORT_SYMBOL(tcp_child_process); 792