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_syncookies __read_mostly = 1; 31 EXPORT_SYMBOL(sysctl_tcp_syncookies); 32 33 int sysctl_tcp_abort_on_overflow __read_mostly; 34 35 struct inet_timewait_death_row tcp_death_row = { 36 .sysctl_max_tw_buckets = NR_FILE * 2, 37 .hashinfo = &tcp_hashinfo, 38 }; 39 EXPORT_SYMBOL_GPL(tcp_death_row); 40 41 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win) 42 { 43 if (seq == s_win) 44 return true; 45 if (after(end_seq, s_win) && before(seq, e_win)) 46 return true; 47 return seq == e_win && seq == end_seq; 48 } 49 50 static enum tcp_tw_status 51 tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw, 52 const struct sk_buff *skb, int mib_idx) 53 { 54 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 55 56 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx, 57 &tcptw->tw_last_oow_ack_time)) { 58 /* Send ACK. Note, we do not put the bucket, 59 * it will be released by caller. 60 */ 61 return TCP_TW_ACK; 62 } 63 64 /* We are rate-limiting, so just release the tw sock and drop skb. */ 65 inet_twsk_put(tw); 66 return TCP_TW_SUCCESS; 67 } 68 69 /* 70 * * Main purpose of TIME-WAIT state is to close connection gracefully, 71 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN 72 * (and, probably, tail of data) and one or more our ACKs are lost. 73 * * What is TIME-WAIT timeout? It is associated with maximal packet 74 * lifetime in the internet, which results in wrong conclusion, that 75 * it is set to catch "old duplicate segments" wandering out of their path. 76 * It is not quite correct. This timeout is calculated so that it exceeds 77 * maximal retransmission timeout enough to allow to lose one (or more) 78 * segments sent by peer and our ACKs. This time may be calculated from RTO. 79 * * When TIME-WAIT socket receives RST, it means that another end 80 * finally closed and we are allowed to kill TIME-WAIT too. 81 * * Second purpose of TIME-WAIT is catching old duplicate segments. 82 * Well, certainly it is pure paranoia, but if we load TIME-WAIT 83 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs. 84 * * If we invented some more clever way to catch duplicates 85 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs. 86 * 87 * The algorithm below is based on FORMAL INTERPRETATION of RFCs. 88 * When you compare it to RFCs, please, read section SEGMENT ARRIVES 89 * from the very beginning. 90 * 91 * NOTE. With recycling (and later with fin-wait-2) TW bucket 92 * is _not_ stateless. It means, that strictly speaking we must 93 * spinlock it. I do not want! Well, probability of misbehaviour 94 * is ridiculously low and, seems, we could use some mb() tricks 95 * to avoid misread sequence numbers, states etc. --ANK 96 * 97 * We don't need to initialize tmp_out.sack_ok as we don't use the results 98 */ 99 enum tcp_tw_status 100 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb, 101 const struct tcphdr *th) 102 { 103 struct tcp_options_received tmp_opt; 104 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 105 bool paws_reject = false; 106 107 tmp_opt.saw_tstamp = 0; 108 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) { 109 tcp_parse_options(skb, &tmp_opt, 0, NULL); 110 111 if (tmp_opt.saw_tstamp) { 112 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset; 113 tmp_opt.ts_recent = tcptw->tw_ts_recent; 114 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp; 115 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 116 } 117 } 118 119 if (tw->tw_substate == TCP_FIN_WAIT2) { 120 /* Just repeat all the checks of tcp_rcv_state_process() */ 121 122 /* Out of window, send ACK */ 123 if (paws_reject || 124 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 125 tcptw->tw_rcv_nxt, 126 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd)) 127 return tcp_timewait_check_oow_rate_limit( 128 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2); 129 130 if (th->rst) 131 goto kill; 132 133 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt)) 134 goto kill_with_rst; 135 136 /* Dup ACK? */ 137 if (!th->ack || 138 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) || 139 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) { 140 inet_twsk_put(tw); 141 return TCP_TW_SUCCESS; 142 } 143 144 /* New data or FIN. If new data arrive after half-duplex close, 145 * reset. 146 */ 147 if (!th->fin || 148 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) { 149 kill_with_rst: 150 inet_twsk_deschedule_put(tw); 151 return TCP_TW_RST; 152 } 153 154 /* FIN arrived, enter true time-wait state. */ 155 tw->tw_substate = TCP_TIME_WAIT; 156 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq; 157 if (tmp_opt.saw_tstamp) { 158 tcptw->tw_ts_recent_stamp = get_seconds(); 159 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 160 } 161 162 if (tcp_death_row.sysctl_tw_recycle && 163 tcptw->tw_ts_recent_stamp && 164 tcp_tw_remember_stamp(tw)) 165 inet_twsk_reschedule(tw, tw->tw_timeout); 166 else 167 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 168 return TCP_TW_ACK; 169 } 170 171 /* 172 * Now real TIME-WAIT state. 173 * 174 * RFC 1122: 175 * "When a connection is [...] on TIME-WAIT state [...] 176 * [a TCP] MAY accept a new SYN from the remote TCP to 177 * reopen the connection directly, if it: 178 * 179 * (1) assigns its initial sequence number for the new 180 * connection to be larger than the largest sequence 181 * number it used on the previous connection incarnation, 182 * and 183 * 184 * (2) returns to TIME-WAIT state if the SYN turns out 185 * to be an old duplicate". 186 */ 187 188 if (!paws_reject && 189 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt && 190 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 191 /* In window segment, it may be only reset or bare ack. */ 192 193 if (th->rst) { 194 /* This is TIME_WAIT assassination, in two flavors. 195 * Oh well... nobody has a sufficient solution to this 196 * protocol bug yet. 197 */ 198 if (sysctl_tcp_rfc1337 == 0) { 199 kill: 200 inet_twsk_deschedule_put(tw); 201 return TCP_TW_SUCCESS; 202 } 203 } 204 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 205 206 if (tmp_opt.saw_tstamp) { 207 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 208 tcptw->tw_ts_recent_stamp = get_seconds(); 209 } 210 211 inet_twsk_put(tw); 212 return TCP_TW_SUCCESS; 213 } 214 215 /* Out of window segment. 216 217 All the segments are ACKed immediately. 218 219 The only exception is new SYN. We accept it, if it is 220 not old duplicate and we are not in danger to be killed 221 by delayed old duplicates. RFC check is that it has 222 newer sequence number works at rates <40Mbit/sec. 223 However, if paws works, it is reliable AND even more, 224 we even may relax silly seq space cutoff. 225 226 RED-PEN: we violate main RFC requirement, if this SYN will appear 227 old duplicate (i.e. we receive RST in reply to SYN-ACK), 228 we must return socket to time-wait state. It is not good, 229 but not fatal yet. 230 */ 231 232 if (th->syn && !th->rst && !th->ack && !paws_reject && 233 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) || 234 (tmp_opt.saw_tstamp && 235 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) { 236 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 237 if (isn == 0) 238 isn++; 239 TCP_SKB_CB(skb)->tcp_tw_isn = isn; 240 return TCP_TW_SYN; 241 } 242 243 if (paws_reject) 244 NET_INC_STATS_BH(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED); 245 246 if (!th->rst) { 247 /* In this case we must reset the TIMEWAIT timer. 248 * 249 * If it is ACKless SYN it may be both old duplicate 250 * and new good SYN with random sequence number <rcv_nxt. 251 * Do not reschedule in the last case. 252 */ 253 if (paws_reject || th->ack) 254 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 255 256 return tcp_timewait_check_oow_rate_limit( 257 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT); 258 } 259 inet_twsk_put(tw); 260 return TCP_TW_SUCCESS; 261 } 262 EXPORT_SYMBOL(tcp_timewait_state_process); 263 264 /* 265 * Move a socket to time-wait or dead fin-wait-2 state. 266 */ 267 void tcp_time_wait(struct sock *sk, int state, int timeo) 268 { 269 const struct inet_connection_sock *icsk = inet_csk(sk); 270 const struct tcp_sock *tp = tcp_sk(sk); 271 struct inet_timewait_sock *tw; 272 bool recycle_ok = false; 273 274 if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp) 275 recycle_ok = tcp_remember_stamp(sk); 276 277 tw = inet_twsk_alloc(sk, &tcp_death_row, state); 278 279 if (tw) { 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 tcptw->tw_ts_offset = tp->tsoffset; 292 tcptw->tw_last_oow_ack_time = 0; 293 294 #if IS_ENABLED(CONFIG_IPV6) 295 if (tw->tw_family == PF_INET6) { 296 struct ipv6_pinfo *np = inet6_sk(sk); 297 298 tw->tw_v6_daddr = sk->sk_v6_daddr; 299 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr; 300 tw->tw_tclass = np->tclass; 301 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK); 302 tw->tw_ipv6only = sk->sk_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) { 318 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC); 319 if (tcptw->tw_md5_key && !tcp_alloc_md5sig_pool()) 320 BUG(); 321 } 322 } while (0); 323 #endif 324 325 /* Get the TIME_WAIT timeout firing. */ 326 if (timeo < rto) 327 timeo = rto; 328 329 if (recycle_ok) { 330 tw->tw_timeout = rto; 331 } else { 332 tw->tw_timeout = TCP_TIMEWAIT_LEN; 333 if (state == TCP_TIME_WAIT) 334 timeo = TCP_TIMEWAIT_LEN; 335 } 336 337 inet_twsk_schedule(tw, timeo); 338 /* Linkage updates. */ 339 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo); 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 } 352 353 void 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 kfree_rcu(twsk->tw_md5_key, rcu); 360 #endif 361 } 362 EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 363 364 /* Warning : This function is called without sk_listener being locked. 365 * Be sure to read socket fields once, as their value could change under us. 366 */ 367 void tcp_openreq_init_rwin(struct request_sock *req, 368 const struct sock *sk_listener, 369 const struct dst_entry *dst) 370 { 371 struct inet_request_sock *ireq = inet_rsk(req); 372 const struct tcp_sock *tp = tcp_sk(sk_listener); 373 u16 user_mss = READ_ONCE(tp->rx_opt.user_mss); 374 int full_space = tcp_full_space(sk_listener); 375 int mss = dst_metric_advmss(dst); 376 u32 window_clamp; 377 __u8 rcv_wscale; 378 379 if (user_mss && user_mss < mss) 380 mss = user_mss; 381 382 window_clamp = READ_ONCE(tp->window_clamp); 383 /* Set this up on the first call only */ 384 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW); 385 386 /* limit the window selection if the user enforce a smaller rx buffer */ 387 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK && 388 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0)) 389 req->rsk_window_clamp = full_space; 390 391 /* tcp_full_space because it is guaranteed to be the first packet */ 392 tcp_select_initial_window(full_space, 393 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 394 &req->rsk_rcv_wnd, 395 &req->rsk_window_clamp, 396 ireq->wscale_ok, 397 &rcv_wscale, 398 dst_metric(dst, RTAX_INITRWND)); 399 ireq->rcv_wscale = rcv_wscale; 400 } 401 EXPORT_SYMBOL(tcp_openreq_init_rwin); 402 403 static void tcp_ecn_openreq_child(struct tcp_sock *tp, 404 const struct request_sock *req) 405 { 406 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 407 } 408 409 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst) 410 { 411 struct inet_connection_sock *icsk = inet_csk(sk); 412 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 413 bool ca_got_dst = false; 414 415 if (ca_key != TCP_CA_UNSPEC) { 416 const struct tcp_congestion_ops *ca; 417 418 rcu_read_lock(); 419 ca = tcp_ca_find_key(ca_key); 420 if (likely(ca && try_module_get(ca->owner))) { 421 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 422 icsk->icsk_ca_ops = ca; 423 ca_got_dst = true; 424 } 425 rcu_read_unlock(); 426 } 427 428 /* If no valid choice made yet, assign current system default ca. */ 429 if (!ca_got_dst && 430 (!icsk->icsk_ca_setsockopt || 431 !try_module_get(icsk->icsk_ca_ops->owner))) 432 tcp_assign_congestion_control(sk); 433 434 tcp_set_ca_state(sk, TCP_CA_Open); 435 } 436 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child); 437 438 /* This is not only more efficient than what we used to do, it eliminates 439 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 440 * 441 * Actually, we could lots of memory writes here. tp of listening 442 * socket contains all necessary default parameters. 443 */ 444 struct sock *tcp_create_openreq_child(const struct sock *sk, 445 struct request_sock *req, 446 struct sk_buff *skb) 447 { 448 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 449 450 if (newsk) { 451 const struct inet_request_sock *ireq = inet_rsk(req); 452 struct tcp_request_sock *treq = tcp_rsk(req); 453 struct inet_connection_sock *newicsk = inet_csk(newsk); 454 struct tcp_sock *newtp = tcp_sk(newsk); 455 456 /* Now setup tcp_sock */ 457 newtp->pred_flags = 0; 458 459 newtp->rcv_wup = newtp->copied_seq = 460 newtp->rcv_nxt = treq->rcv_isn + 1; 461 newtp->segs_in = 0; 462 463 newtp->snd_sml = newtp->snd_una = 464 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1; 465 466 tcp_prequeue_init(newtp); 467 INIT_LIST_HEAD(&newtp->tsq_node); 468 469 tcp_init_wl(newtp, treq->rcv_isn); 470 471 newtp->srtt_us = 0; 472 newtp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 473 newtp->rtt_min[0].rtt = ~0U; 474 newicsk->icsk_rto = TCP_TIMEOUT_INIT; 475 476 newtp->packets_out = 0; 477 newtp->retrans_out = 0; 478 newtp->sacked_out = 0; 479 newtp->fackets_out = 0; 480 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 481 tcp_enable_early_retrans(newtp); 482 newtp->tlp_high_seq = 0; 483 newtp->lsndtime = treq->snt_synack.stamp_jiffies; 484 newsk->sk_txhash = treq->txhash; 485 newtp->last_oow_ack_time = 0; 486 newtp->total_retrans = req->num_retrans; 487 488 /* So many TCP implementations out there (incorrectly) count the 489 * initial SYN frame in their delayed-ACK and congestion control 490 * algorithms that we must have the following bandaid to talk 491 * efficiently to them. -DaveM 492 */ 493 newtp->snd_cwnd = TCP_INIT_CWND; 494 newtp->snd_cwnd_cnt = 0; 495 496 tcp_init_xmit_timers(newsk); 497 __skb_queue_head_init(&newtp->out_of_order_queue); 498 newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1; 499 500 newtp->rx_opt.saw_tstamp = 0; 501 502 newtp->rx_opt.dsack = 0; 503 newtp->rx_opt.num_sacks = 0; 504 505 newtp->urg_data = 0; 506 507 if (sock_flag(newsk, SOCK_KEEPOPEN)) 508 inet_csk_reset_keepalive_timer(newsk, 509 keepalive_time_when(newtp)); 510 511 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 512 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) { 513 if (sysctl_tcp_fack) 514 tcp_enable_fack(newtp); 515 } 516 newtp->window_clamp = req->rsk_window_clamp; 517 newtp->rcv_ssthresh = req->rsk_rcv_wnd; 518 newtp->rcv_wnd = req->rsk_rcv_wnd; 519 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 520 if (newtp->rx_opt.wscale_ok) { 521 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 522 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 523 } else { 524 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 525 newtp->window_clamp = min(newtp->window_clamp, 65535U); 526 } 527 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) << 528 newtp->rx_opt.snd_wscale); 529 newtp->max_window = newtp->snd_wnd; 530 531 if (newtp->rx_opt.tstamp_ok) { 532 newtp->rx_opt.ts_recent = req->ts_recent; 533 newtp->rx_opt.ts_recent_stamp = get_seconds(); 534 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 535 } else { 536 newtp->rx_opt.ts_recent_stamp = 0; 537 newtp->tcp_header_len = sizeof(struct tcphdr); 538 } 539 newtp->tsoffset = 0; 540 #ifdef CONFIG_TCP_MD5SIG 541 newtp->md5sig_info = NULL; /*XXX*/ 542 if (newtp->af_specific->md5_lookup(sk, newsk)) 543 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 544 #endif 545 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 546 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 547 newtp->rx_opt.mss_clamp = req->mss; 548 tcp_ecn_openreq_child(newtp, req); 549 newtp->fastopen_rsk = NULL; 550 newtp->syn_data_acked = 0; 551 newtp->rack.mstamp.v64 = 0; 552 newtp->rack.advanced = 0; 553 554 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_PASSIVEOPENS); 555 } 556 return newsk; 557 } 558 EXPORT_SYMBOL(tcp_create_openreq_child); 559 560 /* 561 * Process an incoming packet for SYN_RECV sockets represented as a 562 * request_sock. Normally sk is the listener socket but for TFO it 563 * points to the child socket. 564 * 565 * XXX (TFO) - The current impl contains a special check for ack 566 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better? 567 * 568 * We don't need to initialize tmp_opt.sack_ok as we don't use the results 569 */ 570 571 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 572 struct request_sock *req, 573 bool fastopen) 574 { 575 struct tcp_options_received tmp_opt; 576 struct sock *child; 577 const struct tcphdr *th = tcp_hdr(skb); 578 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 579 bool paws_reject = false; 580 bool own_req; 581 582 tmp_opt.saw_tstamp = 0; 583 if (th->doff > (sizeof(struct tcphdr)>>2)) { 584 tcp_parse_options(skb, &tmp_opt, 0, NULL); 585 586 if (tmp_opt.saw_tstamp) { 587 tmp_opt.ts_recent = req->ts_recent; 588 /* We do not store true stamp, but it is not required, 589 * it can be estimated (approximately) 590 * from another data. 591 */ 592 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout); 593 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 594 } 595 } 596 597 /* Check for pure retransmitted SYN. */ 598 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 599 flg == TCP_FLAG_SYN && 600 !paws_reject) { 601 /* 602 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 603 * this case on figure 6 and figure 8, but formal 604 * protocol description says NOTHING. 605 * To be more exact, it says that we should send ACK, 606 * because this segment (at least, if it has no data) 607 * is out of window. 608 * 609 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 610 * describe SYN-RECV state. All the description 611 * is wrong, we cannot believe to it and should 612 * rely only on common sense and implementation 613 * experience. 614 * 615 * Enforce "SYN-ACK" according to figure 8, figure 6 616 * of RFC793, fixed by RFC1122. 617 * 618 * Note that even if there is new data in the SYN packet 619 * they will be thrown away too. 620 * 621 * Reset timer after retransmitting SYNACK, similar to 622 * the idea of fast retransmit in recovery. 623 */ 624 if (!tcp_oow_rate_limited(sock_net(sk), skb, 625 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 626 &tcp_rsk(req)->last_oow_ack_time) && 627 628 !inet_rtx_syn_ack(sk, req)) { 629 unsigned long expires = jiffies; 630 631 expires += min(TCP_TIMEOUT_INIT << req->num_timeout, 632 TCP_RTO_MAX); 633 if (!fastopen) 634 mod_timer_pending(&req->rsk_timer, expires); 635 else 636 req->rsk_timer.expires = expires; 637 } 638 return NULL; 639 } 640 641 /* Further reproduces section "SEGMENT ARRIVES" 642 for state SYN-RECEIVED of RFC793. 643 It is broken, however, it does not work only 644 when SYNs are crossed. 645 646 You would think that SYN crossing is impossible here, since 647 we should have a SYN_SENT socket (from connect()) on our end, 648 but this is not true if the crossed SYNs were sent to both 649 ends by a malicious third party. We must defend against this, 650 and to do that we first verify the ACK (as per RFC793, page 651 36) and reset if it is invalid. Is this a true full defense? 652 To convince ourselves, let us consider a way in which the ACK 653 test can still pass in this 'malicious crossed SYNs' case. 654 Malicious sender sends identical SYNs (and thus identical sequence 655 numbers) to both A and B: 656 657 A: gets SYN, seq=7 658 B: gets SYN, seq=7 659 660 By our good fortune, both A and B select the same initial 661 send sequence number of seven :-) 662 663 A: sends SYN|ACK, seq=7, ack_seq=8 664 B: sends SYN|ACK, seq=7, ack_seq=8 665 666 So we are now A eating this SYN|ACK, ACK test passes. So 667 does sequence test, SYN is truncated, and thus we consider 668 it a bare ACK. 669 670 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 671 bare ACK. Otherwise, we create an established connection. Both 672 ends (listening sockets) accept the new incoming connection and try 673 to talk to each other. 8-) 674 675 Note: This case is both harmless, and rare. Possibility is about the 676 same as us discovering intelligent life on another plant tomorrow. 677 678 But generally, we should (RFC lies!) to accept ACK 679 from SYNACK both here and in tcp_rcv_state_process(). 680 tcp_rcv_state_process() does not, hence, we do not too. 681 682 Note that the case is absolutely generic: 683 we cannot optimize anything here without 684 violating protocol. All the checks must be made 685 before attempt to create socket. 686 */ 687 688 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 689 * and the incoming segment acknowledges something not yet 690 * sent (the segment carries an unacceptable ACK) ... 691 * a reset is sent." 692 * 693 * Invalid ACK: reset will be sent by listening socket. 694 * Note that the ACK validity check for a Fast Open socket is done 695 * elsewhere and is checked directly against the child socket rather 696 * than req because user data may have been sent out. 697 */ 698 if ((flg & TCP_FLAG_ACK) && !fastopen && 699 (TCP_SKB_CB(skb)->ack_seq != 700 tcp_rsk(req)->snt_isn + 1)) 701 return sk; 702 703 /* Also, it would be not so bad idea to check rcv_tsecr, which 704 * is essentially ACK extension and too early or too late values 705 * should cause reset in unsynchronized states. 706 */ 707 708 /* RFC793: "first check sequence number". */ 709 710 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 711 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) { 712 /* Out of window: send ACK and drop. */ 713 if (!(flg & TCP_FLAG_RST)) 714 req->rsk_ops->send_ack(sk, skb, req); 715 if (paws_reject) 716 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 717 return NULL; 718 } 719 720 /* In sequence, PAWS is OK. */ 721 722 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt)) 723 req->ts_recent = tmp_opt.rcv_tsval; 724 725 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 726 /* Truncate SYN, it is out of window starting 727 at tcp_rsk(req)->rcv_isn + 1. */ 728 flg &= ~TCP_FLAG_SYN; 729 } 730 731 /* RFC793: "second check the RST bit" and 732 * "fourth, check the SYN bit" 733 */ 734 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 735 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 736 goto embryonic_reset; 737 } 738 739 /* ACK sequence verified above, just make sure ACK is 740 * set. If ACK not set, just silently drop the packet. 741 * 742 * XXX (TFO) - if we ever allow "data after SYN", the 743 * following check needs to be removed. 744 */ 745 if (!(flg & TCP_FLAG_ACK)) 746 return NULL; 747 748 /* For Fast Open no more processing is needed (sk is the 749 * child socket). 750 */ 751 if (fastopen) 752 return sk; 753 754 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 755 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 756 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 757 inet_rsk(req)->acked = 1; 758 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 759 return NULL; 760 } 761 762 /* OK, ACK is valid, create big socket and 763 * feed this segment to it. It will repeat all 764 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 765 * ESTABLISHED STATE. If it will be dropped after 766 * socket is created, wait for troubles. 767 */ 768 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL, 769 req, &own_req); 770 if (!child) 771 goto listen_overflow; 772 773 sock_rps_save_rxhash(child, skb); 774 tcp_synack_rtt_meas(child, req); 775 return inet_csk_complete_hashdance(sk, child, req, own_req); 776 777 listen_overflow: 778 if (!sysctl_tcp_abort_on_overflow) { 779 inet_rsk(req)->acked = 1; 780 return NULL; 781 } 782 783 embryonic_reset: 784 if (!(flg & TCP_FLAG_RST)) { 785 /* Received a bad SYN pkt - for TFO We try not to reset 786 * the local connection unless it's really necessary to 787 * avoid becoming vulnerable to outside attack aiming at 788 * resetting legit local connections. 789 */ 790 req->rsk_ops->send_reset(sk, skb); 791 } else if (fastopen) { /* received a valid RST pkt */ 792 reqsk_fastopen_remove(sk, req, true); 793 tcp_reset(sk); 794 } 795 if (!fastopen) { 796 inet_csk_reqsk_queue_drop(sk, req); 797 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 798 } 799 return NULL; 800 } 801 EXPORT_SYMBOL(tcp_check_req); 802 803 /* 804 * Queue segment on the new socket if the new socket is active, 805 * otherwise we just shortcircuit this and continue with 806 * the new socket. 807 * 808 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV 809 * when entering. But other states are possible due to a race condition 810 * where after __inet_lookup_established() fails but before the listener 811 * locked is obtained, other packets cause the same connection to 812 * be created. 813 */ 814 815 int tcp_child_process(struct sock *parent, struct sock *child, 816 struct sk_buff *skb) 817 { 818 int ret = 0; 819 int state = child->sk_state; 820 821 if (!sock_owned_by_user(child)) { 822 ret = tcp_rcv_state_process(child, skb); 823 /* Wakeup parent, send SIGIO */ 824 if (state == TCP_SYN_RECV && child->sk_state != state) 825 parent->sk_data_ready(parent); 826 } else { 827 /* Alas, it is possible again, because we do lookup 828 * in main socket hash table and lock on listening 829 * socket does not protect us more. 830 */ 831 __sk_add_backlog(child, skb); 832 } 833 834 bh_unlock_sock(child); 835 sock_put(child); 836 return ret; 837 } 838 EXPORT_SYMBOL(tcp_child_process); 839