tcp_minisocks.c revision 4670fd819e7f47392c7c6fc6168ea2857c66d163
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 52/* VJ's idea. Save last timestamp seen from this destination 53 * and hold it at least for normal timewait interval to use for duplicate 54 * segment detection in subsequent connections, before they enter synchronized 55 * state. 56 */ 57 58static bool tcp_remember_stamp(struct sock *sk) 59{ 60 const struct inet_connection_sock *icsk = inet_csk(sk); 61 struct tcp_sock *tp = tcp_sk(sk); 62 struct inet_peer *peer; 63 64 peer = icsk->icsk_af_ops->get_peer(sk); 65 if (peer) { 66 if ((s32)(peer->tcp_ts - tp->rx_opt.ts_recent) <= 0 || 67 ((u32)get_seconds() - peer->tcp_ts_stamp > TCP_PAWS_MSL && 68 peer->tcp_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) { 69 peer->tcp_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp; 70 peer->tcp_ts = tp->rx_opt.ts_recent; 71 } 72 return true; 73 } 74 75 return false; 76} 77 78static bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw) 79{ 80 struct sock *sk = (struct sock *) tw; 81 struct inet_peer *peer; 82 83 peer = twsk_getpeer(sk); 84 if (peer) { 85 const struct tcp_timewait_sock *tcptw = tcp_twsk(sk); 86 87 if ((s32)(peer->tcp_ts - tcptw->tw_ts_recent) <= 0 || 88 ((u32)get_seconds() - peer->tcp_ts_stamp > TCP_PAWS_MSL && 89 peer->tcp_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) { 90 peer->tcp_ts_stamp = (u32)tcptw->tw_ts_recent_stamp; 91 peer->tcp_ts = tcptw->tw_ts_recent; 92 } 93 inet_putpeer(peer); 94 return true; 95 } 96 return false; 97} 98 99static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win) 100{ 101 if (seq == s_win) 102 return true; 103 if (after(end_seq, s_win) && before(seq, e_win)) 104 return true; 105 return seq == e_win && seq == end_seq; 106} 107 108/* 109 * * Main purpose of TIME-WAIT state is to close connection gracefully, 110 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN 111 * (and, probably, tail of data) and one or more our ACKs are lost. 112 * * What is TIME-WAIT timeout? It is associated with maximal packet 113 * lifetime in the internet, which results in wrong conclusion, that 114 * it is set to catch "old duplicate segments" wandering out of their path. 115 * It is not quite correct. This timeout is calculated so that it exceeds 116 * maximal retransmission timeout enough to allow to lose one (or more) 117 * segments sent by peer and our ACKs. This time may be calculated from RTO. 118 * * When TIME-WAIT socket receives RST, it means that another end 119 * finally closed and we are allowed to kill TIME-WAIT too. 120 * * Second purpose of TIME-WAIT is catching old duplicate segments. 121 * Well, certainly it is pure paranoia, but if we load TIME-WAIT 122 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs. 123 * * If we invented some more clever way to catch duplicates 124 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs. 125 * 126 * The algorithm below is based on FORMAL INTERPRETATION of RFCs. 127 * When you compare it to RFCs, please, read section SEGMENT ARRIVES 128 * from the very beginning. 129 * 130 * NOTE. With recycling (and later with fin-wait-2) TW bucket 131 * is _not_ stateless. It means, that strictly speaking we must 132 * spinlock it. I do not want! Well, probability of misbehaviour 133 * is ridiculously low and, seems, we could use some mb() tricks 134 * to avoid misread sequence numbers, states etc. --ANK 135 */ 136enum tcp_tw_status 137tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb, 138 const struct tcphdr *th) 139{ 140 struct tcp_options_received tmp_opt; 141 const u8 *hash_location; 142 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 143 bool paws_reject = false; 144 145 tmp_opt.saw_tstamp = 0; 146 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) { 147 tcp_parse_options(skb, &tmp_opt, &hash_location, 0); 148 149 if (tmp_opt.saw_tstamp) { 150 tmp_opt.ts_recent = tcptw->tw_ts_recent; 151 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp; 152 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 153 } 154 } 155 156 if (tw->tw_substate == TCP_FIN_WAIT2) { 157 /* Just repeat all the checks of tcp_rcv_state_process() */ 158 159 /* Out of window, send ACK */ 160 if (paws_reject || 161 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 162 tcptw->tw_rcv_nxt, 163 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd)) 164 return TCP_TW_ACK; 165 166 if (th->rst) 167 goto kill; 168 169 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt)) 170 goto kill_with_rst; 171 172 /* Dup ACK? */ 173 if (!th->ack || 174 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) || 175 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) { 176 inet_twsk_put(tw); 177 return TCP_TW_SUCCESS; 178 } 179 180 /* New data or FIN. If new data arrive after half-duplex close, 181 * reset. 182 */ 183 if (!th->fin || 184 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) { 185kill_with_rst: 186 inet_twsk_deschedule(tw, &tcp_death_row); 187 inet_twsk_put(tw); 188 return TCP_TW_RST; 189 } 190 191 /* FIN arrived, enter true time-wait state. */ 192 tw->tw_substate = TCP_TIME_WAIT; 193 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq; 194 if (tmp_opt.saw_tstamp) { 195 tcptw->tw_ts_recent_stamp = get_seconds(); 196 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 197 } 198 199 if (tcp_death_row.sysctl_tw_recycle && 200 tcptw->tw_ts_recent_stamp && 201 tcp_tw_remember_stamp(tw)) 202 inet_twsk_schedule(tw, &tcp_death_row, tw->tw_timeout, 203 TCP_TIMEWAIT_LEN); 204 else 205 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 206 TCP_TIMEWAIT_LEN); 207 return TCP_TW_ACK; 208 } 209 210 /* 211 * Now real TIME-WAIT state. 212 * 213 * RFC 1122: 214 * "When a connection is [...] on TIME-WAIT state [...] 215 * [a TCP] MAY accept a new SYN from the remote TCP to 216 * reopen the connection directly, if it: 217 * 218 * (1) assigns its initial sequence number for the new 219 * connection to be larger than the largest sequence 220 * number it used on the previous connection incarnation, 221 * and 222 * 223 * (2) returns to TIME-WAIT state if the SYN turns out 224 * to be an old duplicate". 225 */ 226 227 if (!paws_reject && 228 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt && 229 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 230 /* In window segment, it may be only reset or bare ack. */ 231 232 if (th->rst) { 233 /* This is TIME_WAIT assassination, in two flavors. 234 * Oh well... nobody has a sufficient solution to this 235 * protocol bug yet. 236 */ 237 if (sysctl_tcp_rfc1337 == 0) { 238kill: 239 inet_twsk_deschedule(tw, &tcp_death_row); 240 inet_twsk_put(tw); 241 return TCP_TW_SUCCESS; 242 } 243 } 244 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 245 TCP_TIMEWAIT_LEN); 246 247 if (tmp_opt.saw_tstamp) { 248 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 249 tcptw->tw_ts_recent_stamp = get_seconds(); 250 } 251 252 inet_twsk_put(tw); 253 return TCP_TW_SUCCESS; 254 } 255 256 /* Out of window segment. 257 258 All the segments are ACKed immediately. 259 260 The only exception is new SYN. We accept it, if it is 261 not old duplicate and we are not in danger to be killed 262 by delayed old duplicates. RFC check is that it has 263 newer sequence number works at rates <40Mbit/sec. 264 However, if paws works, it is reliable AND even more, 265 we even may relax silly seq space cutoff. 266 267 RED-PEN: we violate main RFC requirement, if this SYN will appear 268 old duplicate (i.e. we receive RST in reply to SYN-ACK), 269 we must return socket to time-wait state. It is not good, 270 but not fatal yet. 271 */ 272 273 if (th->syn && !th->rst && !th->ack && !paws_reject && 274 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) || 275 (tmp_opt.saw_tstamp && 276 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) { 277 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 278 if (isn == 0) 279 isn++; 280 TCP_SKB_CB(skb)->when = isn; 281 return TCP_TW_SYN; 282 } 283 284 if (paws_reject) 285 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED); 286 287 if (!th->rst) { 288 /* In this case we must reset the TIMEWAIT timer. 289 * 290 * If it is ACKless SYN it may be both old duplicate 291 * and new good SYN with random sequence number <rcv_nxt. 292 * Do not reschedule in the last case. 293 */ 294 if (paws_reject || th->ack) 295 inet_twsk_schedule(tw, &tcp_death_row, TCP_TIMEWAIT_LEN, 296 TCP_TIMEWAIT_LEN); 297 298 /* Send ACK. Note, we do not put the bucket, 299 * it will be released by caller. 300 */ 301 return TCP_TW_ACK; 302 } 303 inet_twsk_put(tw); 304 return TCP_TW_SUCCESS; 305} 306EXPORT_SYMBOL(tcp_timewait_state_process); 307 308/* 309 * Move a socket to time-wait or dead fin-wait-2 state. 310 */ 311void tcp_time_wait(struct sock *sk, int state, int timeo) 312{ 313 struct inet_timewait_sock *tw = NULL; 314 const struct inet_connection_sock *icsk = inet_csk(sk); 315 const struct tcp_sock *tp = tcp_sk(sk); 316 bool recycle_ok = false; 317 318 if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp) 319 recycle_ok = tcp_remember_stamp(sk); 320 321 if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets) 322 tw = inet_twsk_alloc(sk, state); 323 324 if (tw != NULL) { 325 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 326 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 327 328 tw->tw_transparent = inet_sk(sk)->transparent; 329 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 330 tcptw->tw_rcv_nxt = tp->rcv_nxt; 331 tcptw->tw_snd_nxt = tp->snd_nxt; 332 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 333 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 334 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 335 336#if IS_ENABLED(CONFIG_IPV6) 337 if (tw->tw_family == PF_INET6) { 338 struct ipv6_pinfo *np = inet6_sk(sk); 339 struct inet6_timewait_sock *tw6; 340 341 tw->tw_ipv6_offset = inet6_tw_offset(sk->sk_prot); 342 tw6 = inet6_twsk((struct sock *)tw); 343 tw6->tw_v6_daddr = np->daddr; 344 tw6->tw_v6_rcv_saddr = np->rcv_saddr; 345 tw->tw_tclass = np->tclass; 346 tw->tw_ipv6only = np->ipv6only; 347 } 348#endif 349 350#ifdef CONFIG_TCP_MD5SIG 351 /* 352 * The timewait bucket does not have the key DB from the 353 * sock structure. We just make a quick copy of the 354 * md5 key being used (if indeed we are using one) 355 * so the timewait ack generating code has the key. 356 */ 357 do { 358 struct tcp_md5sig_key *key; 359 tcptw->tw_md5_key = NULL; 360 key = tp->af_specific->md5_lookup(sk, sk); 361 if (key != NULL) { 362 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC); 363 if (tcptw->tw_md5_key && tcp_alloc_md5sig_pool(sk) == NULL) 364 BUG(); 365 } 366 } while (0); 367#endif 368 369 /* Linkage updates. */ 370 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo); 371 372 /* Get the TIME_WAIT timeout firing. */ 373 if (timeo < rto) 374 timeo = rto; 375 376 if (recycle_ok) { 377 tw->tw_timeout = rto; 378 } else { 379 tw->tw_timeout = TCP_TIMEWAIT_LEN; 380 if (state == TCP_TIME_WAIT) 381 timeo = TCP_TIMEWAIT_LEN; 382 } 383 384 inet_twsk_schedule(tw, &tcp_death_row, timeo, 385 TCP_TIMEWAIT_LEN); 386 inet_twsk_put(tw); 387 } else { 388 /* Sorry, if we're out of memory, just CLOSE this 389 * socket up. We've got bigger problems than 390 * non-graceful socket closings. 391 */ 392 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW); 393 } 394 395 tcp_update_metrics(sk); 396 tcp_done(sk); 397} 398 399void tcp_twsk_destructor(struct sock *sk) 400{ 401#ifdef CONFIG_TCP_MD5SIG 402 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 403 if (twsk->tw_md5_key) { 404 tcp_free_md5sig_pool(); 405 kfree_rcu(twsk->tw_md5_key, rcu); 406 } 407#endif 408} 409EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 410 411static inline void TCP_ECN_openreq_child(struct tcp_sock *tp, 412 struct request_sock *req) 413{ 414 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 415} 416 417/* This is not only more efficient than what we used to do, it eliminates 418 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 419 * 420 * Actually, we could lots of memory writes here. tp of listening 421 * socket contains all necessary default parameters. 422 */ 423struct sock *tcp_create_openreq_child(struct sock *sk, struct request_sock *req, struct sk_buff *skb) 424{ 425 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 426 427 if (newsk != NULL) { 428 const struct inet_request_sock *ireq = inet_rsk(req); 429 struct tcp_request_sock *treq = tcp_rsk(req); 430 struct inet_connection_sock *newicsk = inet_csk(newsk); 431 struct tcp_sock *newtp = tcp_sk(newsk); 432 struct tcp_sock *oldtp = tcp_sk(sk); 433 struct tcp_cookie_values *oldcvp = oldtp->cookie_values; 434 435 /* TCP Cookie Transactions require space for the cookie pair, 436 * as it differs for each connection. There is no need to 437 * copy any s_data_payload stored at the original socket. 438 * Failure will prevent resuming the connection. 439 * 440 * Presumed copied, in order of appearance: 441 * cookie_in_always, cookie_out_never 442 */ 443 if (oldcvp != NULL) { 444 struct tcp_cookie_values *newcvp = 445 kzalloc(sizeof(*newtp->cookie_values), 446 GFP_ATOMIC); 447 448 if (newcvp != NULL) { 449 kref_init(&newcvp->kref); 450 newcvp->cookie_desired = 451 oldcvp->cookie_desired; 452 newtp->cookie_values = newcvp; 453 } else { 454 /* Not Yet Implemented */ 455 newtp->cookie_values = NULL; 456 } 457 } 458 459 /* Now setup tcp_sock */ 460 newtp->pred_flags = 0; 461 462 newtp->rcv_wup = newtp->copied_seq = 463 newtp->rcv_nxt = treq->rcv_isn + 1; 464 465 newtp->snd_sml = newtp->snd_una = 466 newtp->snd_nxt = newtp->snd_up = 467 treq->snt_isn + 1 + tcp_s_data_size(oldtp); 468 469 tcp_prequeue_init(newtp); 470 471 tcp_init_wl(newtp, treq->rcv_isn); 472 473 newtp->srtt = 0; 474 newtp->mdev = TCP_TIMEOUT_INIT; 475 newicsk->icsk_rto = TCP_TIMEOUT_INIT; 476 477 newtp->packets_out = 0; 478 newtp->retrans_out = 0; 479 newtp->sacked_out = 0; 480 newtp->fackets_out = 0; 481 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 482 tcp_enable_early_retrans(newtp); 483 484 /* So many TCP implementations out there (incorrectly) count the 485 * initial SYN frame in their delayed-ACK and congestion control 486 * algorithms that we must have the following bandaid to talk 487 * efficiently to them. -DaveM 488 */ 489 newtp->snd_cwnd = TCP_INIT_CWND; 490 newtp->snd_cwnd_cnt = 0; 491 newtp->bytes_acked = 0; 492 493 newtp->frto_counter = 0; 494 newtp->frto_highmark = 0; 495 496 if (newicsk->icsk_ca_ops != &tcp_init_congestion_ops && 497 !try_module_get(newicsk->icsk_ca_ops->owner)) 498 newicsk->icsk_ca_ops = &tcp_init_congestion_ops; 499 500 tcp_set_ca_state(newsk, TCP_CA_Open); 501 tcp_init_xmit_timers(newsk); 502 skb_queue_head_init(&newtp->out_of_order_queue); 503 newtp->write_seq = newtp->pushed_seq = 504 treq->snt_isn + 1 + tcp_s_data_size(oldtp); 505 506 newtp->rx_opt.saw_tstamp = 0; 507 508 newtp->rx_opt.dsack = 0; 509 newtp->rx_opt.num_sacks = 0; 510 511 newtp->urg_data = 0; 512 513 if (sock_flag(newsk, SOCK_KEEPOPEN)) 514 inet_csk_reset_keepalive_timer(newsk, 515 keepalive_time_when(newtp)); 516 517 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 518 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) { 519 if (sysctl_tcp_fack) 520 tcp_enable_fack(newtp); 521 } 522 newtp->window_clamp = req->window_clamp; 523 newtp->rcv_ssthresh = req->rcv_wnd; 524 newtp->rcv_wnd = req->rcv_wnd; 525 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 526 if (newtp->rx_opt.wscale_ok) { 527 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 528 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 529 } else { 530 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 531 newtp->window_clamp = min(newtp->window_clamp, 65535U); 532 } 533 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) << 534 newtp->rx_opt.snd_wscale); 535 newtp->max_window = newtp->snd_wnd; 536 537 if (newtp->rx_opt.tstamp_ok) { 538 newtp->rx_opt.ts_recent = req->ts_recent; 539 newtp->rx_opt.ts_recent_stamp = get_seconds(); 540 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 541 } else { 542 newtp->rx_opt.ts_recent_stamp = 0; 543 newtp->tcp_header_len = sizeof(struct tcphdr); 544 } 545#ifdef CONFIG_TCP_MD5SIG 546 newtp->md5sig_info = NULL; /*XXX*/ 547 if (newtp->af_specific->md5_lookup(sk, newsk)) 548 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 549#endif 550 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 551 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 552 newtp->rx_opt.mss_clamp = req->mss; 553 TCP_ECN_openreq_child(newtp, req); 554 555 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS); 556 } 557 return newsk; 558} 559EXPORT_SYMBOL(tcp_create_openreq_child); 560 561/* 562 * Process an incoming packet for SYN_RECV sockets represented 563 * as a request_sock. 564 */ 565 566struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 567 struct request_sock *req, 568 struct request_sock **prev) 569{ 570 struct tcp_options_received tmp_opt; 571 const u8 *hash_location; 572 struct sock *child; 573 const struct tcphdr *th = tcp_hdr(skb); 574 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 575 bool paws_reject = false; 576 577 tmp_opt.saw_tstamp = 0; 578 if (th->doff > (sizeof(struct tcphdr)>>2)) { 579 tcp_parse_options(skb, &tmp_opt, &hash_location, 0); 580 581 if (tmp_opt.saw_tstamp) { 582 tmp_opt.ts_recent = req->ts_recent; 583 /* We do not store true stamp, but it is not required, 584 * it can be estimated (approximately) 585 * from another data. 586 */ 587 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->retrans); 588 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 589 } 590 } 591 592 /* Check for pure retransmitted SYN. */ 593 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 594 flg == TCP_FLAG_SYN && 595 !paws_reject) { 596 /* 597 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 598 * this case on figure 6 and figure 8, but formal 599 * protocol description says NOTHING. 600 * To be more exact, it says that we should send ACK, 601 * because this segment (at least, if it has no data) 602 * is out of window. 603 * 604 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 605 * describe SYN-RECV state. All the description 606 * is wrong, we cannot believe to it and should 607 * rely only on common sense and implementation 608 * experience. 609 * 610 * Enforce "SYN-ACK" according to figure 8, figure 6 611 * of RFC793, fixed by RFC1122. 612 */ 613 req->rsk_ops->rtx_syn_ack(sk, req, NULL); 614 return NULL; 615 } 616 617 /* Further reproduces section "SEGMENT ARRIVES" 618 for state SYN-RECEIVED of RFC793. 619 It is broken, however, it does not work only 620 when SYNs are crossed. 621 622 You would think that SYN crossing is impossible here, since 623 we should have a SYN_SENT socket (from connect()) on our end, 624 but this is not true if the crossed SYNs were sent to both 625 ends by a malicious third party. We must defend against this, 626 and to do that we first verify the ACK (as per RFC793, page 627 36) and reset if it is invalid. Is this a true full defense? 628 To convince ourselves, let us consider a way in which the ACK 629 test can still pass in this 'malicious crossed SYNs' case. 630 Malicious sender sends identical SYNs (and thus identical sequence 631 numbers) to both A and B: 632 633 A: gets SYN, seq=7 634 B: gets SYN, seq=7 635 636 By our good fortune, both A and B select the same initial 637 send sequence number of seven :-) 638 639 A: sends SYN|ACK, seq=7, ack_seq=8 640 B: sends SYN|ACK, seq=7, ack_seq=8 641 642 So we are now A eating this SYN|ACK, ACK test passes. So 643 does sequence test, SYN is truncated, and thus we consider 644 it a bare ACK. 645 646 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 647 bare ACK. Otherwise, we create an established connection. Both 648 ends (listening sockets) accept the new incoming connection and try 649 to talk to each other. 8-) 650 651 Note: This case is both harmless, and rare. Possibility is about the 652 same as us discovering intelligent life on another plant tomorrow. 653 654 But generally, we should (RFC lies!) to accept ACK 655 from SYNACK both here and in tcp_rcv_state_process(). 656 tcp_rcv_state_process() does not, hence, we do not too. 657 658 Note that the case is absolutely generic: 659 we cannot optimize anything here without 660 violating protocol. All the checks must be made 661 before attempt to create socket. 662 */ 663 664 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 665 * and the incoming segment acknowledges something not yet 666 * sent (the segment carries an unacceptable ACK) ... 667 * a reset is sent." 668 * 669 * Invalid ACK: reset will be sent by listening socket 670 */ 671 if ((flg & TCP_FLAG_ACK) && 672 (TCP_SKB_CB(skb)->ack_seq != 673 tcp_rsk(req)->snt_isn + 1 + tcp_s_data_size(tcp_sk(sk)))) 674 return sk; 675 676 /* Also, it would be not so bad idea to check rcv_tsecr, which 677 * is essentially ACK extension and too early or too late values 678 * should cause reset in unsynchronized states. 679 */ 680 681 /* RFC793: "first check sequence number". */ 682 683 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 684 tcp_rsk(req)->rcv_isn + 1, tcp_rsk(req)->rcv_isn + 1 + req->rcv_wnd)) { 685 /* Out of window: send ACK and drop. */ 686 if (!(flg & TCP_FLAG_RST)) 687 req->rsk_ops->send_ack(sk, skb, req); 688 if (paws_reject) 689 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 690 return NULL; 691 } 692 693 /* In sequence, PAWS is OK. */ 694 695 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_isn + 1)) 696 req->ts_recent = tmp_opt.rcv_tsval; 697 698 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 699 /* Truncate SYN, it is out of window starting 700 at tcp_rsk(req)->rcv_isn + 1. */ 701 flg &= ~TCP_FLAG_SYN; 702 } 703 704 /* RFC793: "second check the RST bit" and 705 * "fourth, check the SYN bit" 706 */ 707 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 708 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 709 goto embryonic_reset; 710 } 711 712 /* ACK sequence verified above, just make sure ACK is 713 * set. If ACK not set, just silently drop the packet. 714 */ 715 if (!(flg & TCP_FLAG_ACK)) 716 return NULL; 717 718 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 719 if (req->retrans < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 720 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 721 inet_rsk(req)->acked = 1; 722 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 723 return NULL; 724 } 725 if (tmp_opt.saw_tstamp && tmp_opt.rcv_tsecr) 726 tcp_rsk(req)->snt_synack = tmp_opt.rcv_tsecr; 727 else if (req->retrans) /* don't take RTT sample if retrans && ~TS */ 728 tcp_rsk(req)->snt_synack = 0; 729 730 /* OK, ACK is valid, create big socket and 731 * feed this segment to it. It will repeat all 732 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 733 * ESTABLISHED STATE. If it will be dropped after 734 * socket is created, wait for troubles. 735 */ 736 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL); 737 if (child == NULL) 738 goto listen_overflow; 739 740 inet_csk_reqsk_queue_unlink(sk, req, prev); 741 inet_csk_reqsk_queue_removed(sk, req); 742 743 inet_csk_reqsk_queue_add(sk, req, child); 744 return child; 745 746listen_overflow: 747 if (!sysctl_tcp_abort_on_overflow) { 748 inet_rsk(req)->acked = 1; 749 return NULL; 750 } 751 752embryonic_reset: 753 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 754 if (!(flg & TCP_FLAG_RST)) 755 req->rsk_ops->send_reset(sk, skb); 756 757 inet_csk_reqsk_queue_drop(sk, req, prev); 758 return NULL; 759} 760EXPORT_SYMBOL(tcp_check_req); 761 762/* 763 * Queue segment on the new socket if the new socket is active, 764 * otherwise we just shortcircuit this and continue with 765 * the new socket. 766 */ 767 768int tcp_child_process(struct sock *parent, struct sock *child, 769 struct sk_buff *skb) 770{ 771 int ret = 0; 772 int state = child->sk_state; 773 774 if (!sock_owned_by_user(child)) { 775 ret = tcp_rcv_state_process(child, skb, tcp_hdr(skb), 776 skb->len); 777 /* Wakeup parent, send SIGIO */ 778 if (state == TCP_SYN_RECV && child->sk_state != state) 779 parent->sk_data_ready(parent, 0); 780 } else { 781 /* Alas, it is possible again, because we do lookup 782 * in main socket hash table and lock on listening 783 * socket does not protect us more. 784 */ 785 __sk_add_backlog(child, skb); 786 } 787 788 bh_unlock_sock(child); 789 sock_put(child); 790 return ret; 791} 792EXPORT_SYMBOL(tcp_child_process); 793