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 30 int sysctl_tcp_abort_on_overflow __read_mostly; 31 32 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win) 33 { 34 if (seq == s_win) 35 return true; 36 if (after(end_seq, s_win) && before(seq, e_win)) 37 return true; 38 return seq == e_win && seq == end_seq; 39 } 40 41 static enum tcp_tw_status 42 tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw, 43 const struct sk_buff *skb, int mib_idx) 44 { 45 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 46 47 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx, 48 &tcptw->tw_last_oow_ack_time)) { 49 /* Send ACK. Note, we do not put the bucket, 50 * it will be released by caller. 51 */ 52 return TCP_TW_ACK; 53 } 54 55 /* We are rate-limiting, so just release the tw sock and drop skb. */ 56 inet_twsk_put(tw); 57 return TCP_TW_SUCCESS; 58 } 59 60 /* 61 * * Main purpose of TIME-WAIT state is to close connection gracefully, 62 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN 63 * (and, probably, tail of data) and one or more our ACKs are lost. 64 * * What is TIME-WAIT timeout? It is associated with maximal packet 65 * lifetime in the internet, which results in wrong conclusion, that 66 * it is set to catch "old duplicate segments" wandering out of their path. 67 * It is not quite correct. This timeout is calculated so that it exceeds 68 * maximal retransmission timeout enough to allow to lose one (or more) 69 * segments sent by peer and our ACKs. This time may be calculated from RTO. 70 * * When TIME-WAIT socket receives RST, it means that another end 71 * finally closed and we are allowed to kill TIME-WAIT too. 72 * * Second purpose of TIME-WAIT is catching old duplicate segments. 73 * Well, certainly it is pure paranoia, but if we load TIME-WAIT 74 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs. 75 * * If we invented some more clever way to catch duplicates 76 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs. 77 * 78 * The algorithm below is based on FORMAL INTERPRETATION of RFCs. 79 * When you compare it to RFCs, please, read section SEGMENT ARRIVES 80 * from the very beginning. 81 * 82 * NOTE. With recycling (and later with fin-wait-2) TW bucket 83 * is _not_ stateless. It means, that strictly speaking we must 84 * spinlock it. I do not want! Well, probability of misbehaviour 85 * is ridiculously low and, seems, we could use some mb() tricks 86 * to avoid misread sequence numbers, states etc. --ANK 87 * 88 * We don't need to initialize tmp_out.sack_ok as we don't use the results 89 */ 90 enum tcp_tw_status 91 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb, 92 const struct tcphdr *th) 93 { 94 struct tcp_options_received tmp_opt; 95 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 96 bool paws_reject = false; 97 struct inet_timewait_death_row *tcp_death_row = &sock_net((struct sock*)tw)->ipv4.tcp_death_row; 98 99 tmp_opt.saw_tstamp = 0; 100 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) { 101 tcp_parse_options(skb, &tmp_opt, 0, NULL); 102 103 if (tmp_opt.saw_tstamp) { 104 if (tmp_opt.rcv_tsecr) 105 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset; 106 tmp_opt.ts_recent = tcptw->tw_ts_recent; 107 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp; 108 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 109 } 110 } 111 112 if (tw->tw_substate == TCP_FIN_WAIT2) { 113 /* Just repeat all the checks of tcp_rcv_state_process() */ 114 115 /* Out of window, send ACK */ 116 if (paws_reject || 117 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 118 tcptw->tw_rcv_nxt, 119 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd)) 120 return tcp_timewait_check_oow_rate_limit( 121 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2); 122 123 if (th->rst) 124 goto kill; 125 126 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt)) 127 return TCP_TW_RST; 128 129 /* Dup ACK? */ 130 if (!th->ack || 131 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) || 132 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) { 133 inet_twsk_put(tw); 134 return TCP_TW_SUCCESS; 135 } 136 137 /* New data or FIN. If new data arrive after half-duplex close, 138 * reset. 139 */ 140 if (!th->fin || 141 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) 142 return TCP_TW_RST; 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_reschedule(tw, tw->tw_timeout); 156 else 157 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 158 return TCP_TW_ACK; 159 } 160 161 /* 162 * Now real TIME-WAIT state. 163 * 164 * RFC 1122: 165 * "When a connection is [...] on TIME-WAIT state [...] 166 * [a TCP] MAY accept a new SYN from the remote TCP to 167 * reopen the connection directly, if it: 168 * 169 * (1) assigns its initial sequence number for the new 170 * connection to be larger than the largest sequence 171 * number it used on the previous connection incarnation, 172 * and 173 * 174 * (2) returns to TIME-WAIT state if the SYN turns out 175 * to be an old duplicate". 176 */ 177 178 if (!paws_reject && 179 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt && 180 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 181 /* In window segment, it may be only reset or bare ack. */ 182 183 if (th->rst) { 184 /* This is TIME_WAIT assassination, in two flavors. 185 * Oh well... nobody has a sufficient solution to this 186 * protocol bug yet. 187 */ 188 if (sysctl_tcp_rfc1337 == 0) { 189 kill: 190 inet_twsk_deschedule_put(tw); 191 return TCP_TW_SUCCESS; 192 } 193 } 194 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 195 196 if (tmp_opt.saw_tstamp) { 197 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 198 tcptw->tw_ts_recent_stamp = get_seconds(); 199 } 200 201 inet_twsk_put(tw); 202 return TCP_TW_SUCCESS; 203 } 204 205 /* Out of window segment. 206 207 All the segments are ACKed immediately. 208 209 The only exception is new SYN. We accept it, if it is 210 not old duplicate and we are not in danger to be killed 211 by delayed old duplicates. RFC check is that it has 212 newer sequence number works at rates <40Mbit/sec. 213 However, if paws works, it is reliable AND even more, 214 we even may relax silly seq space cutoff. 215 216 RED-PEN: we violate main RFC requirement, if this SYN will appear 217 old duplicate (i.e. we receive RST in reply to SYN-ACK), 218 we must return socket to time-wait state. It is not good, 219 but not fatal yet. 220 */ 221 222 if (th->syn && !th->rst && !th->ack && !paws_reject && 223 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) || 224 (tmp_opt.saw_tstamp && 225 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) { 226 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 227 if (isn == 0) 228 isn++; 229 TCP_SKB_CB(skb)->tcp_tw_isn = isn; 230 return TCP_TW_SYN; 231 } 232 233 if (paws_reject) 234 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED); 235 236 if (!th->rst) { 237 /* In this case we must reset the TIMEWAIT timer. 238 * 239 * If it is ACKless SYN it may be both old duplicate 240 * and new good SYN with random sequence number <rcv_nxt. 241 * Do not reschedule in the last case. 242 */ 243 if (paws_reject || th->ack) 244 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 245 246 return tcp_timewait_check_oow_rate_limit( 247 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT); 248 } 249 inet_twsk_put(tw); 250 return TCP_TW_SUCCESS; 251 } 252 EXPORT_SYMBOL(tcp_timewait_state_process); 253 254 /* 255 * Move a socket to time-wait or dead fin-wait-2 state. 256 */ 257 void tcp_time_wait(struct sock *sk, int state, int timeo) 258 { 259 const struct inet_connection_sock *icsk = inet_csk(sk); 260 const struct tcp_sock *tp = tcp_sk(sk); 261 struct inet_timewait_sock *tw; 262 bool recycle_ok = false; 263 struct inet_timewait_death_row *tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row; 264 265 if (tcp_death_row->sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp) 266 recycle_ok = tcp_remember_stamp(sk); 267 268 tw = inet_twsk_alloc(sk, tcp_death_row, state); 269 270 if (tw) { 271 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 272 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 273 struct inet_sock *inet = inet_sk(sk); 274 275 tw->tw_transparent = inet->transparent; 276 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 277 tcptw->tw_rcv_nxt = tp->rcv_nxt; 278 tcptw->tw_snd_nxt = tp->snd_nxt; 279 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 280 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 281 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 282 tcptw->tw_ts_offset = tp->tsoffset; 283 tcptw->tw_last_oow_ack_time = 0; 284 285 #if IS_ENABLED(CONFIG_IPV6) 286 if (tw->tw_family == PF_INET6) { 287 struct ipv6_pinfo *np = inet6_sk(sk); 288 289 tw->tw_v6_daddr = sk->sk_v6_daddr; 290 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr; 291 tw->tw_tclass = np->tclass; 292 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK); 293 tw->tw_ipv6only = sk->sk_ipv6only; 294 } 295 #endif 296 297 #ifdef CONFIG_TCP_MD5SIG 298 /* 299 * The timewait bucket does not have the key DB from the 300 * sock structure. We just make a quick copy of the 301 * md5 key being used (if indeed we are using one) 302 * so the timewait ack generating code has the key. 303 */ 304 do { 305 struct tcp_md5sig_key *key; 306 tcptw->tw_md5_key = NULL; 307 key = tp->af_specific->md5_lookup(sk, sk); 308 if (key) { 309 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC); 310 if (tcptw->tw_md5_key && !tcp_alloc_md5sig_pool()) 311 BUG(); 312 } 313 } while (0); 314 #endif 315 316 /* Get the TIME_WAIT timeout firing. */ 317 if (timeo < rto) 318 timeo = rto; 319 320 if (recycle_ok) { 321 tw->tw_timeout = rto; 322 } else { 323 tw->tw_timeout = TCP_TIMEWAIT_LEN; 324 if (state == TCP_TIME_WAIT) 325 timeo = TCP_TIMEWAIT_LEN; 326 } 327 328 inet_twsk_schedule(tw, timeo); 329 /* Linkage updates. */ 330 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo); 331 inet_twsk_put(tw); 332 } else { 333 /* Sorry, if we're out of memory, just CLOSE this 334 * socket up. We've got bigger problems than 335 * non-graceful socket closings. 336 */ 337 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW); 338 } 339 340 tcp_update_metrics(sk); 341 tcp_done(sk); 342 } 343 344 void tcp_twsk_destructor(struct sock *sk) 345 { 346 #ifdef CONFIG_TCP_MD5SIG 347 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 348 349 if (twsk->tw_md5_key) 350 kfree_rcu(twsk->tw_md5_key, rcu); 351 #endif 352 } 353 EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 354 355 /* Warning : This function is called without sk_listener being locked. 356 * Be sure to read socket fields once, as their value could change under us. 357 */ 358 void tcp_openreq_init_rwin(struct request_sock *req, 359 const struct sock *sk_listener, 360 const struct dst_entry *dst) 361 { 362 struct inet_request_sock *ireq = inet_rsk(req); 363 const struct tcp_sock *tp = tcp_sk(sk_listener); 364 int full_space = tcp_full_space(sk_listener); 365 u32 window_clamp; 366 __u8 rcv_wscale; 367 int mss; 368 369 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 370 window_clamp = READ_ONCE(tp->window_clamp); 371 /* Set this up on the first call only */ 372 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW); 373 374 /* limit the window selection if the user enforce a smaller rx buffer */ 375 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK && 376 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0)) 377 req->rsk_window_clamp = full_space; 378 379 /* tcp_full_space because it is guaranteed to be the first packet */ 380 tcp_select_initial_window(full_space, 381 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 382 &req->rsk_rcv_wnd, 383 &req->rsk_window_clamp, 384 ireq->wscale_ok, 385 &rcv_wscale, 386 dst_metric(dst, RTAX_INITRWND)); 387 ireq->rcv_wscale = rcv_wscale; 388 } 389 EXPORT_SYMBOL(tcp_openreq_init_rwin); 390 391 static void tcp_ecn_openreq_child(struct tcp_sock *tp, 392 const struct request_sock *req) 393 { 394 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 395 } 396 397 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst) 398 { 399 struct inet_connection_sock *icsk = inet_csk(sk); 400 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 401 bool ca_got_dst = false; 402 403 if (ca_key != TCP_CA_UNSPEC) { 404 const struct tcp_congestion_ops *ca; 405 406 rcu_read_lock(); 407 ca = tcp_ca_find_key(ca_key); 408 if (likely(ca && try_module_get(ca->owner))) { 409 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 410 icsk->icsk_ca_ops = ca; 411 ca_got_dst = true; 412 } 413 rcu_read_unlock(); 414 } 415 416 /* If no valid choice made yet, assign current system default ca. */ 417 if (!ca_got_dst && 418 (!icsk->icsk_ca_setsockopt || 419 !try_module_get(icsk->icsk_ca_ops->owner))) 420 tcp_assign_congestion_control(sk); 421 422 tcp_set_ca_state(sk, TCP_CA_Open); 423 } 424 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child); 425 426 /* This is not only more efficient than what we used to do, it eliminates 427 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 428 * 429 * Actually, we could lots of memory writes here. tp of listening 430 * socket contains all necessary default parameters. 431 */ 432 struct sock *tcp_create_openreq_child(const struct sock *sk, 433 struct request_sock *req, 434 struct sk_buff *skb) 435 { 436 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 437 438 if (newsk) { 439 const struct inet_request_sock *ireq = inet_rsk(req); 440 struct tcp_request_sock *treq = tcp_rsk(req); 441 struct inet_connection_sock *newicsk = inet_csk(newsk); 442 struct tcp_sock *newtp = tcp_sk(newsk); 443 444 /* Now setup tcp_sock */ 445 newtp->pred_flags = 0; 446 447 newtp->rcv_wup = newtp->copied_seq = 448 newtp->rcv_nxt = treq->rcv_isn + 1; 449 newtp->segs_in = 1; 450 451 newtp->snd_sml = newtp->snd_una = 452 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1; 453 454 tcp_prequeue_init(newtp); 455 INIT_LIST_HEAD(&newtp->tsq_node); 456 457 tcp_init_wl(newtp, treq->rcv_isn); 458 459 newtp->srtt_us = 0; 460 newtp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 461 minmax_reset(&newtp->rtt_min, tcp_time_stamp, ~0U); 462 newicsk->icsk_rto = TCP_TIMEOUT_INIT; 463 newicsk->icsk_ack.lrcvtime = tcp_time_stamp; 464 465 newtp->packets_out = 0; 466 newtp->retrans_out = 0; 467 newtp->sacked_out = 0; 468 newtp->fackets_out = 0; 469 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 470 newtp->tlp_high_seq = 0; 471 newtp->lsndtime = treq->snt_synack.stamp_jiffies; 472 newsk->sk_txhash = treq->txhash; 473 newtp->last_oow_ack_time = 0; 474 newtp->total_retrans = req->num_retrans; 475 476 /* So many TCP implementations out there (incorrectly) count the 477 * initial SYN frame in their delayed-ACK and congestion control 478 * algorithms that we must have the following bandaid to talk 479 * efficiently to them. -DaveM 480 */ 481 newtp->snd_cwnd = TCP_INIT_CWND; 482 newtp->snd_cwnd_cnt = 0; 483 484 /* There's a bubble in the pipe until at least the first ACK. */ 485 newtp->app_limited = ~0U; 486 487 tcp_init_xmit_timers(newsk); 488 newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1; 489 490 newtp->rx_opt.saw_tstamp = 0; 491 492 newtp->rx_opt.dsack = 0; 493 newtp->rx_opt.num_sacks = 0; 494 495 newtp->urg_data = 0; 496 497 if (sock_flag(newsk, SOCK_KEEPOPEN)) 498 inet_csk_reset_keepalive_timer(newsk, 499 keepalive_time_when(newtp)); 500 501 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 502 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) { 503 if (sysctl_tcp_fack) 504 tcp_enable_fack(newtp); 505 } 506 newtp->window_clamp = req->rsk_window_clamp; 507 newtp->rcv_ssthresh = req->rsk_rcv_wnd; 508 newtp->rcv_wnd = req->rsk_rcv_wnd; 509 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 510 if (newtp->rx_opt.wscale_ok) { 511 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 512 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 513 } else { 514 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 515 newtp->window_clamp = min(newtp->window_clamp, 65535U); 516 } 517 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) << 518 newtp->rx_opt.snd_wscale); 519 newtp->max_window = newtp->snd_wnd; 520 521 if (newtp->rx_opt.tstamp_ok) { 522 newtp->rx_opt.ts_recent = req->ts_recent; 523 newtp->rx_opt.ts_recent_stamp = get_seconds(); 524 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 525 } else { 526 newtp->rx_opt.ts_recent_stamp = 0; 527 newtp->tcp_header_len = sizeof(struct tcphdr); 528 } 529 newtp->tsoffset = treq->ts_off; 530 #ifdef CONFIG_TCP_MD5SIG 531 newtp->md5sig_info = NULL; /*XXX*/ 532 if (newtp->af_specific->md5_lookup(sk, newsk)) 533 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 534 #endif 535 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 536 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 537 newtp->rx_opt.mss_clamp = req->mss; 538 tcp_ecn_openreq_child(newtp, req); 539 newtp->fastopen_rsk = NULL; 540 newtp->syn_data_acked = 0; 541 newtp->rack.mstamp.v64 = 0; 542 newtp->rack.advanced = 0; 543 544 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS); 545 } 546 return newsk; 547 } 548 EXPORT_SYMBOL(tcp_create_openreq_child); 549 550 /* 551 * Process an incoming packet for SYN_RECV sockets represented as a 552 * request_sock. Normally sk is the listener socket but for TFO it 553 * points to the child socket. 554 * 555 * XXX (TFO) - The current impl contains a special check for ack 556 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better? 557 * 558 * We don't need to initialize tmp_opt.sack_ok as we don't use the results 559 */ 560 561 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 562 struct request_sock *req, 563 bool fastopen) 564 { 565 struct tcp_options_received tmp_opt; 566 struct sock *child; 567 const struct tcphdr *th = tcp_hdr(skb); 568 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 569 bool paws_reject = false; 570 bool own_req; 571 572 tmp_opt.saw_tstamp = 0; 573 if (th->doff > (sizeof(struct tcphdr)>>2)) { 574 tcp_parse_options(skb, &tmp_opt, 0, NULL); 575 576 if (tmp_opt.saw_tstamp) { 577 tmp_opt.ts_recent = req->ts_recent; 578 if (tmp_opt.rcv_tsecr) 579 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off; 580 /* We do not store true stamp, but it is not required, 581 * it can be estimated (approximately) 582 * from another data. 583 */ 584 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout); 585 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 586 } 587 } 588 589 /* Check for pure retransmitted SYN. */ 590 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 591 flg == TCP_FLAG_SYN && 592 !paws_reject) { 593 /* 594 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 595 * this case on figure 6 and figure 8, but formal 596 * protocol description says NOTHING. 597 * To be more exact, it says that we should send ACK, 598 * because this segment (at least, if it has no data) 599 * is out of window. 600 * 601 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 602 * describe SYN-RECV state. All the description 603 * is wrong, we cannot believe to it and should 604 * rely only on common sense and implementation 605 * experience. 606 * 607 * Enforce "SYN-ACK" according to figure 8, figure 6 608 * of RFC793, fixed by RFC1122. 609 * 610 * Note that even if there is new data in the SYN packet 611 * they will be thrown away too. 612 * 613 * Reset timer after retransmitting SYNACK, similar to 614 * the idea of fast retransmit in recovery. 615 */ 616 if (!tcp_oow_rate_limited(sock_net(sk), skb, 617 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 618 &tcp_rsk(req)->last_oow_ack_time) && 619 620 !inet_rtx_syn_ack(sk, req)) { 621 unsigned long expires = jiffies; 622 623 expires += min(TCP_TIMEOUT_INIT << req->num_timeout, 624 TCP_RTO_MAX); 625 if (!fastopen) 626 mod_timer_pending(&req->rsk_timer, expires); 627 else 628 req->rsk_timer.expires = expires; 629 } 630 return NULL; 631 } 632 633 /* Further reproduces section "SEGMENT ARRIVES" 634 for state SYN-RECEIVED of RFC793. 635 It is broken, however, it does not work only 636 when SYNs are crossed. 637 638 You would think that SYN crossing is impossible here, since 639 we should have a SYN_SENT socket (from connect()) on our end, 640 but this is not true if the crossed SYNs were sent to both 641 ends by a malicious third party. We must defend against this, 642 and to do that we first verify the ACK (as per RFC793, page 643 36) and reset if it is invalid. Is this a true full defense? 644 To convince ourselves, let us consider a way in which the ACK 645 test can still pass in this 'malicious crossed SYNs' case. 646 Malicious sender sends identical SYNs (and thus identical sequence 647 numbers) to both A and B: 648 649 A: gets SYN, seq=7 650 B: gets SYN, seq=7 651 652 By our good fortune, both A and B select the same initial 653 send sequence number of seven :-) 654 655 A: sends SYN|ACK, seq=7, ack_seq=8 656 B: sends SYN|ACK, seq=7, ack_seq=8 657 658 So we are now A eating this SYN|ACK, ACK test passes. So 659 does sequence test, SYN is truncated, and thus we consider 660 it a bare ACK. 661 662 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 663 bare ACK. Otherwise, we create an established connection. Both 664 ends (listening sockets) accept the new incoming connection and try 665 to talk to each other. 8-) 666 667 Note: This case is both harmless, and rare. Possibility is about the 668 same as us discovering intelligent life on another plant tomorrow. 669 670 But generally, we should (RFC lies!) to accept ACK 671 from SYNACK both here and in tcp_rcv_state_process(). 672 tcp_rcv_state_process() does not, hence, we do not too. 673 674 Note that the case is absolutely generic: 675 we cannot optimize anything here without 676 violating protocol. All the checks must be made 677 before attempt to create socket. 678 */ 679 680 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 681 * and the incoming segment acknowledges something not yet 682 * sent (the segment carries an unacceptable ACK) ... 683 * a reset is sent." 684 * 685 * Invalid ACK: reset will be sent by listening socket. 686 * Note that the ACK validity check for a Fast Open socket is done 687 * elsewhere and is checked directly against the child socket rather 688 * than req because user data may have been sent out. 689 */ 690 if ((flg & TCP_FLAG_ACK) && !fastopen && 691 (TCP_SKB_CB(skb)->ack_seq != 692 tcp_rsk(req)->snt_isn + 1)) 693 return sk; 694 695 /* Also, it would be not so bad idea to check rcv_tsecr, which 696 * is essentially ACK extension and too early or too late values 697 * should cause reset in unsynchronized states. 698 */ 699 700 /* RFC793: "first check sequence number". */ 701 702 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 703 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) { 704 /* Out of window: send ACK and drop. */ 705 if (!(flg & TCP_FLAG_RST) && 706 !tcp_oow_rate_limited(sock_net(sk), skb, 707 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 708 &tcp_rsk(req)->last_oow_ack_time)) 709 req->rsk_ops->send_ack(sk, skb, req); 710 if (paws_reject) 711 __NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 712 return NULL; 713 } 714 715 /* In sequence, PAWS is OK. */ 716 717 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt)) 718 req->ts_recent = tmp_opt.rcv_tsval; 719 720 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 721 /* Truncate SYN, it is out of window starting 722 at tcp_rsk(req)->rcv_isn + 1. */ 723 flg &= ~TCP_FLAG_SYN; 724 } 725 726 /* RFC793: "second check the RST bit" and 727 * "fourth, check the SYN bit" 728 */ 729 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 730 __TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 731 goto embryonic_reset; 732 } 733 734 /* ACK sequence verified above, just make sure ACK is 735 * set. If ACK not set, just silently drop the packet. 736 * 737 * XXX (TFO) - if we ever allow "data after SYN", the 738 * following check needs to be removed. 739 */ 740 if (!(flg & TCP_FLAG_ACK)) 741 return NULL; 742 743 /* For Fast Open no more processing is needed (sk is the 744 * child socket). 745 */ 746 if (fastopen) 747 return sk; 748 749 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 750 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 751 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 752 inet_rsk(req)->acked = 1; 753 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 754 return NULL; 755 } 756 757 /* OK, ACK is valid, create big socket and 758 * feed this segment to it. It will repeat all 759 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 760 * ESTABLISHED STATE. If it will be dropped after 761 * socket is created, wait for troubles. 762 */ 763 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL, 764 req, &own_req); 765 if (!child) 766 goto listen_overflow; 767 768 sock_rps_save_rxhash(child, skb); 769 tcp_synack_rtt_meas(child, req); 770 return inet_csk_complete_hashdance(sk, child, req, own_req); 771 772 listen_overflow: 773 if (!sysctl_tcp_abort_on_overflow) { 774 inet_rsk(req)->acked = 1; 775 return NULL; 776 } 777 778 embryonic_reset: 779 if (!(flg & TCP_FLAG_RST)) { 780 /* Received a bad SYN pkt - for TFO We try not to reset 781 * the local connection unless it's really necessary to 782 * avoid becoming vulnerable to outside attack aiming at 783 * resetting legit local connections. 784 */ 785 req->rsk_ops->send_reset(sk, skb); 786 } else if (fastopen) { /* received a valid RST pkt */ 787 reqsk_fastopen_remove(sk, req, true); 788 tcp_reset(sk); 789 } 790 if (!fastopen) { 791 inet_csk_reqsk_queue_drop(sk, req); 792 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 793 } 794 return NULL; 795 } 796 EXPORT_SYMBOL(tcp_check_req); 797 798 /* 799 * Queue segment on the new socket if the new socket is active, 800 * otherwise we just shortcircuit this and continue with 801 * the new socket. 802 * 803 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV 804 * when entering. But other states are possible due to a race condition 805 * where after __inet_lookup_established() fails but before the listener 806 * locked is obtained, other packets cause the same connection to 807 * be created. 808 */ 809 810 int tcp_child_process(struct sock *parent, struct sock *child, 811 struct sk_buff *skb) 812 { 813 int ret = 0; 814 int state = child->sk_state; 815 816 tcp_segs_in(tcp_sk(child), skb); 817 if (!sock_owned_by_user(child)) { 818 ret = tcp_rcv_state_process(child, skb); 819 /* Wakeup parent, send SIGIO */ 820 if (state == TCP_SYN_RECV && child->sk_state != state) 821 parent->sk_data_ready(parent); 822 } else { 823 /* Alas, it is possible again, because we do lookup 824 * in main socket hash table and lock on listening 825 * socket does not protect us more. 826 */ 827 __sk_add_backlog(child, skb); 828 } 829 830 bh_unlock_sock(child); 831 sock_put(child); 832 return ret; 833 } 834 EXPORT_SYMBOL(tcp_child_process); 835