tcp_minisocks.c revision 8336886f786fdacbc19b719c1f7ea91eb70706d4
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 */ 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 const u8 *hash_location; 95 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 96 bool paws_reject = false; 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, NULL); 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 if (tcp_death_row.sysctl_tw_recycle && 153 tcptw->tw_ts_recent_stamp && 154 tcp_tw_remember_stamp(tw)) 155 inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout, 156 TCP_TIMEWAIT_LEN); 157 else 158 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 159 TCP_TIMEWAIT_LEN); 160 return TCP_TW_ACK; 161 } 162 163 /* 164 * Now real TIME-WAIT state. 165 * 166 * RFC 1122: 167 * "When a connection is [...] on TIME-WAIT state [...] 168 * [a TCP] MAY accept a new SYN from the remote TCP to 169 * reopen the connection directly, if it: 170 * 171 * (1) assigns its initial sequence number for the new 172 * connection to be larger than the largest sequence 173 * number it used on the previous connection incarnation, 174 * and 175 * 176 * (2) returns to TIME-WAIT state if the SYN turns out 177 * to be an old duplicate". 178 */ 179 180 if (!paws_reject && 181 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt && 182 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 183 /* In window segment, it may be only reset or bare ack. */ 184 185 if (th->rst) { 186 /* This is TIME_WAIT assassination, in two flavors. 187 * Oh well... nobody has a sufficient solution to this 188 * protocol bug yet. 189 */ 190 if (sysctl_tcp_rfc1337 == 0) { 191kill: 192 inet_twsk_deschedule(tw, &tcp_death_row); 193 inet_twsk_put(tw); 194 return TCP_TW_SUCCESS; 195 } 196 } 197 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 198 TCP_TIMEWAIT_LEN); 199 200 if (tmp_opt.saw_tstamp) { 201 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 202 tcptw->tw_ts_recent_stamp = get_seconds(); 203 } 204 205 inet_twsk_put(tw); 206 return TCP_TW_SUCCESS; 207 } 208 209 /* Out of window segment. 210 211 All the segments are ACKed immediately. 212 213 The only exception is new SYN. We accept it, if it is 214 not old duplicate and we are not in danger to be killed 215 by delayed old duplicates. RFC check is that it has 216 newer sequence number works at rates <40Mbit/sec. 217 However, if paws works, it is reliable AND even more, 218 we even may relax silly seq space cutoff. 219 220 RED-PEN: we violate main RFC requirement, if this SYN will appear 221 old duplicate (i.e. we receive RST in reply to SYN-ACK), 222 we must return socket to time-wait state. It is not good, 223 but not fatal yet. 224 */ 225 226 if (th->syn && !th->rst && !th->ack && !paws_reject && 227 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) || 228 (tmp_opt.saw_tstamp && 229 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) { 230 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 231 if (isn == 0) 232 isn++; 233 TCP_SKB_CB(skb)->when = isn; 234 return TCP_TW_SYN; 235 } 236 237 if (paws_reject) 238 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED); 239 240 if (!th->rst) { 241 /* In this case we must reset the TIMEWAIT timer. 242 * 243 * If it is ACKless SYN it may be both old duplicate 244 * and new good SYN with random sequence number <rcv_nxt. 245 * Do not reschedule in the last case. 246 */ 247 if (paws_reject || th->ack) 248 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 249 TCP_TIMEWAIT_LEN); 250 251 /* Send ACK. Note, we do not put the bucket, 252 * it will be released by caller. 253 */ 254 return TCP_TW_ACK; 255 } 256 inet_twsk_put(tw); 257 return TCP_TW_SUCCESS; 258} 259EXPORT_SYMBOL(tcp_timewait_state_process); 260 261/* 262 * Move a socket to time-wait or dead fin-wait-2 state. 263 */ 264void tcp_time_wait(struct sock *sk, int state, int timeo) 265{ 266 struct inet_timewait_sock *tw = NULL; 267 const struct inet_connection_sock *icsk = inet_csk(sk); 268 const struct tcp_sock *tp = tcp_sk(sk); 269 bool recycle_ok = false; 270 271 if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp) 272 recycle_ok = tcp_remember_stamp(sk); 273 274 if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets) 275 tw = inet_twsk_alloc(sk, state); 276 277 if (tw != NULL) { 278 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 279 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 280 struct inet_sock *inet = inet_sk(sk); 281 282 tw->tw_transparent = inet->transparent; 283 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 284 tcptw->tw_rcv_nxt = tp->rcv_nxt; 285 tcptw->tw_snd_nxt = tp->snd_nxt; 286 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 287 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 288 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 289 290#if IS_ENABLED(CONFIG_IPV6) 291 if (tw->tw_family == PF_INET6) { 292 struct ipv6_pinfo *np = inet6_sk(sk); 293 struct inet6_timewait_sock *tw6; 294 295 tw->tw_ipv6_offset = inet6_tw_offset(sk->sk_prot); 296 tw6 = inet6_twsk((struct sock *)tw); 297 tw6->tw_v6_daddr = np->daddr; 298 tw6->tw_v6_rcv_saddr = np->rcv_saddr; 299 tw->tw_tclass = np->tclass; 300 tw->tw_ipv6only = np->ipv6only; 301 } 302#endif 303 304#ifdef CONFIG_TCP_MD5SIG 305 /* 306 * The timewait bucket does not have the key DB from the 307 * sock structure. We just make a quick copy of the 308 * md5 key being used (if indeed we are using one) 309 * so the timewait ack generating code has the key. 310 */ 311 do { 312 struct tcp_md5sig_key *key; 313 tcptw->tw_md5_key = NULL; 314 key = tp->af_specific->md5_lookup(sk, sk); 315 if (key != NULL) { 316 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC); 317 if (tcptw->tw_md5_key && tcp_alloc_md5sig_pool(sk) == NULL) 318 BUG(); 319 } 320 } while (0); 321#endif 322 323 /* Linkage updates. */ 324 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo); 325 326 /* Get the TIME_WAIT timeout firing. */ 327 if (timeo < rto) 328 timeo = rto; 329 330 if (recycle_ok) { 331 tw->tw_timeout = rto; 332 } else { 333 tw->tw_timeout = TCP_TIMEWAIT_LEN; 334 if (state == TCP_TIME_WAIT) 335 timeo = TCP_TIMEWAIT_LEN; 336 } 337 338 inet_twsk_schedule(tw, &tcp_death_row, timeo, 339 TCP_TIMEWAIT_LEN); 340 inet_twsk_put(tw); 341 } else { 342 /* Sorry, if we're out of memory, just CLOSE this 343 * socket up. We've got bigger problems than 344 * non-graceful socket closings. 345 */ 346 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW); 347 } 348 349 tcp_update_metrics(sk); 350 tcp_done(sk); 351} 352 353void tcp_twsk_destructor(struct sock *sk) 354{ 355#ifdef CONFIG_TCP_MD5SIG 356 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 357 358 if (twsk->tw_md5_key) { 359 tcp_free_md5sig_pool(); 360 kfree_rcu(twsk->tw_md5_key, rcu); 361 } 362#endif 363} 364EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 365 366static inline void TCP_ECN_openreq_child(struct tcp_sock *tp, 367 struct request_sock *req) 368{ 369 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 370} 371 372/* This is not only more efficient than what we used to do, it eliminates 373 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 374 * 375 * Actually, we could lots of memory writes here. tp of listening 376 * socket contains all necessary default parameters. 377 */ 378struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb) 379{ 380 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 381 382 if (newsk != NULL) { 383 const struct inet_request_sock *ireq = inet_rsk(req); 384 struct tcp_request_sock *treq = tcp_rsk(req); 385 struct inet_connection_sock *newicsk = inet_csk(newsk); 386 struct tcp_sock *newtp = tcp_sk(newsk); 387 struct tcp_sock *oldtp = tcp_sk(sk); 388 struct tcp_cookie_values *oldcvp = oldtp->cookie_values; 389 390 /* TCP Cookie Transactions require space for the cookie pair, 391 * as it differs for each connection. There is no need to 392 * copy any s_data_payload stored at the original socket. 393 * Failure will prevent resuming the connection. 394 * 395 * Presumed copied, in order of appearance: 396 * cookie_in_always, cookie_out_never 397 */ 398 if (oldcvp != NULL) { 399 struct tcp_cookie_values *newcvp = 400 kzalloc(sizeof(*newtp->cookie_values), 401 GFP_ATOMIC); 402 403 if (newcvp != NULL) { 404 kref_init(&newcvp->kref); 405 newcvp->cookie_desired = 406 oldcvp->cookie_desired; 407 newtp->cookie_values = newcvp; 408 } else { 409 /* Not Yet Implemented */ 410 newtp->cookie_values = NULL; 411 } 412 } 413 414 /* Now setup tcp_sock */ 415 newtp->pred_flags = 0; 416 417 newtp->rcv_wup = newtp->copied_seq = 418 newtp->rcv_nxt = treq->rcv_isn + 1; 419 420 newtp->snd_sml = newtp->snd_una = 421 newtp->snd_nxt = newtp->snd_up = 422 treq->snt_isn + 1 + tcp_s_data_size(oldtp); 423 424 tcp_prequeue_init(newtp); 425 INIT_LIST_HEAD(&newtp->tsq_node); 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 tcp_enable_early_retrans(newtp); 439 440 /* So many TCP implementations out there (incorrectly) count the 441 * initial SYN frame in their delayed-ACK and congestion control 442 * algorithms that we must have the following bandaid to talk 443 * efficiently to them. -DaveM 444 */ 445 newtp->snd_cwnd = TCP_INIT_CWND; 446 newtp->snd_cwnd_cnt = 0; 447 newtp->bytes_acked = 0; 448 449 newtp->frto_counter = 0; 450 newtp->frto_highmark = 0; 451 452 if (newicsk->icsk_ca_ops != &tcp_init_congestion_ops && 453 !try_module_get(newicsk->icsk_ca_ops->owner)) 454 newicsk->icsk_ca_ops = &tcp_init_congestion_ops; 455 456 tcp_set_ca_state(newsk, TCP_CA_Open); 457 tcp_init_xmit_timers(newsk); 458 skb_queue_head_init(&newtp->out_of_order_queue); 459 newtp->write_seq = newtp->pushed_seq = 460 treq->snt_isn + 1 + tcp_s_data_size(oldtp); 461 462 newtp->rx_opt.saw_tstamp = 0; 463 464 newtp->rx_opt.dsack = 0; 465 newtp->rx_opt.num_sacks = 0; 466 467 newtp->urg_data = 0; 468 469 if (sock_flag(newsk, SOCK_KEEPOPEN)) 470 inet_csk_reset_keepalive_timer(newsk, 471 keepalive_time_when(newtp)); 472 473 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 474 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) { 475 if (sysctl_tcp_fack) 476 tcp_enable_fack(newtp); 477 } 478 newtp->window_clamp = req->window_clamp; 479 newtp->rcv_ssthresh = req->rcv_wnd; 480 newtp->rcv_wnd = req->rcv_wnd; 481 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 482 if (newtp->rx_opt.wscale_ok) { 483 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 484 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 485 } else { 486 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 487 newtp->window_clamp = min(newtp->window_clamp, 65535U); 488 } 489 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) << 490 newtp->rx_opt.snd_wscale); 491 newtp->max_window = newtp->snd_wnd; 492 493 if (newtp->rx_opt.tstamp_ok) { 494 newtp->rx_opt.ts_recent = req->ts_recent; 495 newtp->rx_opt.ts_recent_stamp = get_seconds(); 496 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 497 } else { 498 newtp->rx_opt.ts_recent_stamp = 0; 499 newtp->tcp_header_len = sizeof(struct tcphdr); 500 } 501#ifdef CONFIG_TCP_MD5SIG 502 newtp->md5sig_info = NULL; /*XXX*/ 503 if (newtp->af_specific->md5_lookup(sk, newsk)) 504 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 505#endif 506 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 507 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 508 newtp->rx_opt.mss_clamp = req->mss; 509 TCP_ECN_openreq_child(newtp, req); 510 newtp->fastopen_rsk = NULL; 511 512 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS); 513 } 514 return newsk; 515} 516EXPORT_SYMBOL(tcp_create_openreq_child); 517 518/* 519 * Process an incoming packet for SYN_RECV sockets represented as a 520 * request_sock. Normally sk is the listener socket but for TFO it 521 * points to the child socket. 522 * 523 * XXX (TFO) - The current impl contains a special check for ack 524 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better? 525 */ 526 527struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 528 struct request_sock *req, 529 struct request_sock **prev, 530 bool fastopen) 531{ 532 struct tcp_options_received tmp_opt; 533 const u8 *hash_location; 534 struct sock *child; 535 const struct tcphdr *th = tcp_hdr(skb); 536 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 537 bool paws_reject = false; 538 539 BUG_ON(fastopen == (sk->sk_state == TCP_LISTEN)); 540 541 tmp_opt.saw_tstamp = 0; 542 if (th->doff > (sizeof(struct tcphdr)>>2)) { 543 tcp_parse_options(skb, &tmp_opt, &hash_location, 0, NULL); 544 545 if (tmp_opt.saw_tstamp) { 546 tmp_opt.ts_recent = req->ts_recent; 547 /* We do not store true stamp, but it is not required, 548 * it can be estimated (approximately) 549 * from another data. 550 */ 551 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans); 552 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 553 } 554 } 555 556 /* Check for pure retransmitted SYN. */ 557 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 558 flg == TCP_FLAG_SYN && 559 !paws_reject) { 560 /* 561 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 562 * this case on figure 6 and figure 8, but formal 563 * protocol description says NOTHING. 564 * To be more exact, it says that we should send ACK, 565 * because this segment (at least, if it has no data) 566 * is out of window. 567 * 568 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 569 * describe SYN-RECV state. All the description 570 * is wrong, we cannot believe to it and should 571 * rely only on common sense and implementation 572 * experience. 573 * 574 * Enforce "SYN-ACK" according to figure 8, figure 6 575 * of RFC793, fixed by RFC1122. 576 * 577 * Note that even if there is new data in the SYN packet 578 * they will be thrown away too. 579 */ 580 req->rsk_ops->rtx_syn_ack(sk, req, NULL); 581 return NULL; 582 } 583 584 /* Further reproduces section "SEGMENT ARRIVES" 585 for state SYN-RECEIVED of RFC793. 586 It is broken, however, it does not work only 587 when SYNs are crossed. 588 589 You would think that SYN crossing is impossible here, since 590 we should have a SYN_SENT socket (from connect()) on our end, 591 but this is not true if the crossed SYNs were sent to both 592 ends by a malicious third party. We must defend against this, 593 and to do that we first verify the ACK (as per RFC793, page 594 36) and reset if it is invalid. Is this a true full defense? 595 To convince ourselves, let us consider a way in which the ACK 596 test can still pass in this 'malicious crossed SYNs' case. 597 Malicious sender sends identical SYNs (and thus identical sequence 598 numbers) to both A and B: 599 600 A: gets SYN, seq=7 601 B: gets SYN, seq=7 602 603 By our good fortune, both A and B select the same initial 604 send sequence number of seven :-) 605 606 A: sends SYN|ACK, seq=7, ack_seq=8 607 B: sends SYN|ACK, seq=7, ack_seq=8 608 609 So we are now A eating this SYN|ACK, ACK test passes. So 610 does sequence test, SYN is truncated, and thus we consider 611 it a bare ACK. 612 613 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 614 bare ACK. Otherwise, we create an established connection. Both 615 ends (listening sockets) accept the new incoming connection and try 616 to talk to each other. 8-) 617 618 Note: This case is both harmless, and rare. Possibility is about the 619 same as us discovering intelligent life on another plant tomorrow. 620 621 But generally, we should (RFC lies!) to accept ACK 622 from SYNACK both here and in tcp_rcv_state_process(). 623 tcp_rcv_state_process() does not, hence, we do not too. 624 625 Note that the case is absolutely generic: 626 we cannot optimize anything here without 627 violating protocol. All the checks must be made 628 before attempt to create socket. 629 */ 630 631 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 632 * and the incoming segment acknowledges something not yet 633 * sent (the segment carries an unacceptable ACK) ... 634 * a reset is sent." 635 * 636 * Invalid ACK: reset will be sent by listening socket. 637 * Note that the ACK validity check for a Fast Open socket is done 638 * elsewhere and is checked directly against the child socket rather 639 * than req because user data may have been sent out. 640 */ 641 if ((flg & TCP_FLAG_ACK) && !fastopen && 642 (TCP_SKB_CB(skb)->ack_seq != 643 tcp_rsk(req)->snt_isn + 1 + tcp_s_data_size(tcp_sk(sk)))) 644 return sk; 645 646 /* Also, it would be not so bad idea to check rcv_tsecr, which 647 * is essentially ACK extension and too early or too late values 648 * should cause reset in unsynchronized states. 649 */ 650 651 /* RFC793: "first check sequence number". */ 652 653 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 654 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rcv_wnd)) { 655 /* Out of window: send ACK and drop. */ 656 if (!(flg & TCP_FLAG_RST)) 657 req->rsk_ops->send_ack(sk, skb, req); 658 if (paws_reject) 659 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 660 return NULL; 661 } 662 663 /* In sequence, PAWS is OK. */ 664 665 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt)) 666 req->ts_recent = tmp_opt.rcv_tsval; 667 668 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 669 /* Truncate SYN, it is out of window starting 670 at tcp_rsk(req)->rcv_isn + 1. */ 671 flg &= ~TCP_FLAG_SYN; 672 } 673 674 /* RFC793: "second check the RST bit" and 675 * "fourth, check the SYN bit" 676 */ 677 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 678 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 679 goto embryonic_reset; 680 } 681 682 /* ACK sequence verified above, just make sure ACK is 683 * set. If ACK not set, just silently drop the packet. 684 * 685 * XXX (TFO) - if we ever allow "data after SYN", the 686 * following check needs to be removed. 687 */ 688 if (!(flg & TCP_FLAG_ACK)) 689 return NULL; 690 691 /* For Fast Open no more processing is needed (sk is the 692 * child socket). 693 */ 694 if (fastopen) 695 return sk; 696 697 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 698 if (req->retrans < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 699 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 700 inet_rsk(req)->acked = 1; 701 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 702 return NULL; 703 } 704 if (tmp_opt.saw_tstamp && tmp_opt.rcv_tsecr) 705 tcp_rsk(req)->snt_synack = tmp_opt.rcv_tsecr; 706 else if (req->retrans) /* don't take RTT sample if retrans && ~TS */ 707 tcp_rsk(req)->snt_synack = 0; 708 709 /* OK, ACK is valid, create big socket and 710 * feed this segment to it. It will repeat all 711 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 712 * ESTABLISHED STATE. If it will be dropped after 713 * socket is created, wait for troubles. 714 */ 715 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL); 716 if (child == NULL) 717 goto listen_overflow; 718 719 inet_csk_reqsk_queue_unlink(sk, req, prev); 720 inet_csk_reqsk_queue_removed(sk, req); 721 722 inet_csk_reqsk_queue_add(sk, req, child); 723 return child; 724 725listen_overflow: 726 if (!sysctl_tcp_abort_on_overflow) { 727 inet_rsk(req)->acked = 1; 728 return NULL; 729 } 730 731embryonic_reset: 732 if (!(flg & TCP_FLAG_RST)) { 733 /* Received a bad SYN pkt - for TFO We try not to reset 734 * the local connection unless it's really necessary to 735 * avoid becoming vulnerable to outside attack aiming at 736 * resetting legit local connections. 737 */ 738 req->rsk_ops->send_reset(sk, skb); 739 } else if (fastopen) { /* received a valid RST pkt */ 740 reqsk_fastopen_remove(sk, req, true); 741 tcp_reset(sk); 742 } 743 if (!fastopen) { 744 inet_csk_reqsk_queue_drop(sk, req, prev); 745 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 746 } 747 return NULL; 748} 749EXPORT_SYMBOL(tcp_check_req); 750 751/* 752 * Queue segment on the new socket if the new socket is active, 753 * otherwise we just shortcircuit this and continue with 754 * the new socket. 755 * 756 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV 757 * when entering. But other states are possible due to a race condition 758 * where after __inet_lookup_established() fails but before the listener 759 * locked is obtained, other packets cause the same connection to 760 * be created. 761 */ 762 763int tcp_child_process(struct sock *parent, struct sock *child, 764 struct sk_buff *skb) 765{ 766 int ret = 0; 767 int state = child->sk_state; 768 769 if (!sock_owned_by_user(child)) { 770 ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb), 771 skb->len); 772 /* Wakeup parent, send SIGIO */ 773 if (state == TCP_SYN_RECV && child->sk_state != state) 774 parent->sk_data_ready(parent, 0); 775 } else { 776 /* Alas, it is possible again, because we do lookup 777 * in main socket hash table and lock on listening 778 * socket does not protect us more. 779 */ 780 __sk_add_backlog(child, skb); 781 } 782 783 bh_unlock_sock(child); 784 sock_put(child); 785 return ret; 786} 787EXPORT_SYMBOL(tcp_child_process); 788