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 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset; 105 tmp_opt.ts_recent = tcptw->tw_ts_recent; 106 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp; 107 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 108 } 109 } 110 111 if (tw->tw_substate == TCP_FIN_WAIT2) { 112 /* Just repeat all the checks of tcp_rcv_state_process() */ 113 114 /* Out of window, send ACK */ 115 if (paws_reject || 116 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 117 tcptw->tw_rcv_nxt, 118 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd)) 119 return tcp_timewait_check_oow_rate_limit( 120 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2); 121 122 if (th->rst) 123 goto kill; 124 125 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt)) 126 return TCP_TW_RST; 127 128 /* Dup ACK? */ 129 if (!th->ack || 130 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) || 131 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) { 132 inet_twsk_put(tw); 133 return TCP_TW_SUCCESS; 134 } 135 136 /* New data or FIN. If new data arrive after half-duplex close, 137 * reset. 138 */ 139 if (!th->fin || 140 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) 141 return TCP_TW_RST; 142 143 /* FIN arrived, enter true time-wait state. */ 144 tw->tw_substate = TCP_TIME_WAIT; 145 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq; 146 if (tmp_opt.saw_tstamp) { 147 tcptw->tw_ts_recent_stamp = get_seconds(); 148 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 149 } 150 151 if (tcp_death_row->sysctl_tw_recycle && 152 tcptw->tw_ts_recent_stamp && 153 tcp_tw_remember_stamp(tw)) 154 inet_twsk_reschedule(tw, tw->tw_timeout); 155 else 156 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 157 return TCP_TW_ACK; 158 } 159 160 /* 161 * Now real TIME-WAIT state. 162 * 163 * RFC 1122: 164 * "When a connection is [...] on TIME-WAIT state [...] 165 * [a TCP] MAY accept a new SYN from the remote TCP to 166 * reopen the connection directly, if it: 167 * 168 * (1) assigns its initial sequence number for the new 169 * connection to be larger than the largest sequence 170 * number it used on the previous connection incarnation, 171 * and 172 * 173 * (2) returns to TIME-WAIT state if the SYN turns out 174 * to be an old duplicate". 175 */ 176 177 if (!paws_reject && 178 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt && 179 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 180 /* In window segment, it may be only reset or bare ack. */ 181 182 if (th->rst) { 183 /* This is TIME_WAIT assassination, in two flavors. 184 * Oh well... nobody has a sufficient solution to this 185 * protocol bug yet. 186 */ 187 if (sysctl_tcp_rfc1337 == 0) { 188 kill: 189 inet_twsk_deschedule_put(tw); 190 return TCP_TW_SUCCESS; 191 } 192 } 193 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 194 195 if (tmp_opt.saw_tstamp) { 196 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 197 tcptw->tw_ts_recent_stamp = get_seconds(); 198 } 199 200 inet_twsk_put(tw); 201 return TCP_TW_SUCCESS; 202 } 203 204 /* Out of window segment. 205 206 All the segments are ACKed immediately. 207 208 The only exception is new SYN. We accept it, if it is 209 not old duplicate and we are not in danger to be killed 210 by delayed old duplicates. RFC check is that it has 211 newer sequence number works at rates <40Mbit/sec. 212 However, if paws works, it is reliable AND even more, 213 we even may relax silly seq space cutoff. 214 215 RED-PEN: we violate main RFC requirement, if this SYN will appear 216 old duplicate (i.e. we receive RST in reply to SYN-ACK), 217 we must return socket to time-wait state. It is not good, 218 but not fatal yet. 219 */ 220 221 if (th->syn && !th->rst && !th->ack && !paws_reject && 222 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) || 223 (tmp_opt.saw_tstamp && 224 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) { 225 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 226 if (isn == 0) 227 isn++; 228 TCP_SKB_CB(skb)->tcp_tw_isn = isn; 229 return TCP_TW_SYN; 230 } 231 232 if (paws_reject) 233 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED); 234 235 if (!th->rst) { 236 /* In this case we must reset the TIMEWAIT timer. 237 * 238 * If it is ACKless SYN it may be both old duplicate 239 * and new good SYN with random sequence number <rcv_nxt. 240 * Do not reschedule in the last case. 241 */ 242 if (paws_reject || th->ack) 243 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 244 245 return tcp_timewait_check_oow_rate_limit( 246 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT); 247 } 248 inet_twsk_put(tw); 249 return TCP_TW_SUCCESS; 250 } 251 EXPORT_SYMBOL(tcp_timewait_state_process); 252 253 /* 254 * Move a socket to time-wait or dead fin-wait-2 state. 255 */ 256 void tcp_time_wait(struct sock *sk, int state, int timeo) 257 { 258 const struct inet_connection_sock *icsk = inet_csk(sk); 259 const struct tcp_sock *tp = tcp_sk(sk); 260 struct inet_timewait_sock *tw; 261 bool recycle_ok = false; 262 struct inet_timewait_death_row *tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row; 263 264 if (tcp_death_row->sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp) 265 recycle_ok = tcp_remember_stamp(sk); 266 267 tw = inet_twsk_alloc(sk, tcp_death_row, state); 268 269 if (tw) { 270 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 271 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 272 struct inet_sock *inet = inet_sk(sk); 273 274 tw->tw_transparent = inet->transparent; 275 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 276 tcptw->tw_rcv_nxt = tp->rcv_nxt; 277 tcptw->tw_snd_nxt = tp->snd_nxt; 278 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 279 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 280 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 281 tcptw->tw_ts_offset = tp->tsoffset; 282 tcptw->tw_last_oow_ack_time = 0; 283 284 #if IS_ENABLED(CONFIG_IPV6) 285 if (tw->tw_family == PF_INET6) { 286 struct ipv6_pinfo *np = inet6_sk(sk); 287 288 tw->tw_v6_daddr = sk->sk_v6_daddr; 289 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr; 290 tw->tw_tclass = np->tclass; 291 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK); 292 tw->tw_ipv6only = sk->sk_ipv6only; 293 } 294 #endif 295 296 #ifdef CONFIG_TCP_MD5SIG 297 /* 298 * The timewait bucket does not have the key DB from the 299 * sock structure. We just make a quick copy of the 300 * md5 key being used (if indeed we are using one) 301 * so the timewait ack generating code has the key. 302 */ 303 do { 304 struct tcp_md5sig_key *key; 305 tcptw->tw_md5_key = NULL; 306 key = tp->af_specific->md5_lookup(sk, sk); 307 if (key) { 308 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC); 309 if (tcptw->tw_md5_key && !tcp_alloc_md5sig_pool()) 310 BUG(); 311 } 312 } while (0); 313 #endif 314 315 /* Get the TIME_WAIT timeout firing. */ 316 if (timeo < rto) 317 timeo = rto; 318 319 if (recycle_ok) { 320 tw->tw_timeout = rto; 321 } else { 322 tw->tw_timeout = TCP_TIMEWAIT_LEN; 323 if (state == TCP_TIME_WAIT) 324 timeo = TCP_TIMEWAIT_LEN; 325 } 326 327 inet_twsk_schedule(tw, timeo); 328 /* Linkage updates. */ 329 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo); 330 inet_twsk_put(tw); 331 } else { 332 /* Sorry, if we're out of memory, just CLOSE this 333 * socket up. We've got bigger problems than 334 * non-graceful socket closings. 335 */ 336 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW); 337 } 338 339 tcp_update_metrics(sk); 340 tcp_done(sk); 341 } 342 343 void tcp_twsk_destructor(struct sock *sk) 344 { 345 #ifdef CONFIG_TCP_MD5SIG 346 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 347 348 if (twsk->tw_md5_key) 349 kfree_rcu(twsk->tw_md5_key, rcu); 350 #endif 351 } 352 EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 353 354 /* Warning : This function is called without sk_listener being locked. 355 * Be sure to read socket fields once, as their value could change under us. 356 */ 357 void tcp_openreq_init_rwin(struct request_sock *req, 358 const struct sock *sk_listener, 359 const struct dst_entry *dst) 360 { 361 struct inet_request_sock *ireq = inet_rsk(req); 362 const struct tcp_sock *tp = tcp_sk(sk_listener); 363 int full_space = tcp_full_space(sk_listener); 364 u32 window_clamp; 365 __u8 rcv_wscale; 366 int mss; 367 368 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 369 window_clamp = READ_ONCE(tp->window_clamp); 370 /* Set this up on the first call only */ 371 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW); 372 373 /* limit the window selection if the user enforce a smaller rx buffer */ 374 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK && 375 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0)) 376 req->rsk_window_clamp = full_space; 377 378 /* tcp_full_space because it is guaranteed to be the first packet */ 379 tcp_select_initial_window(full_space, 380 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 381 &req->rsk_rcv_wnd, 382 &req->rsk_window_clamp, 383 ireq->wscale_ok, 384 &rcv_wscale, 385 dst_metric(dst, RTAX_INITRWND)); 386 ireq->rcv_wscale = rcv_wscale; 387 } 388 EXPORT_SYMBOL(tcp_openreq_init_rwin); 389 390 static void tcp_ecn_openreq_child(struct tcp_sock *tp, 391 const struct request_sock *req) 392 { 393 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 394 } 395 396 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst) 397 { 398 struct inet_connection_sock *icsk = inet_csk(sk); 399 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 400 bool ca_got_dst = false; 401 402 if (ca_key != TCP_CA_UNSPEC) { 403 const struct tcp_congestion_ops *ca; 404 405 rcu_read_lock(); 406 ca = tcp_ca_find_key(ca_key); 407 if (likely(ca && try_module_get(ca->owner))) { 408 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 409 icsk->icsk_ca_ops = ca; 410 ca_got_dst = true; 411 } 412 rcu_read_unlock(); 413 } 414 415 /* If no valid choice made yet, assign current system default ca. */ 416 if (!ca_got_dst && 417 (!icsk->icsk_ca_setsockopt || 418 !try_module_get(icsk->icsk_ca_ops->owner))) 419 tcp_assign_congestion_control(sk); 420 421 tcp_set_ca_state(sk, TCP_CA_Open); 422 } 423 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child); 424 425 /* This is not only more efficient than what we used to do, it eliminates 426 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 427 * 428 * Actually, we could lots of memory writes here. tp of listening 429 * socket contains all necessary default parameters. 430 */ 431 struct sock *tcp_create_openreq_child(const struct sock *sk, 432 struct request_sock *req, 433 struct sk_buff *skb) 434 { 435 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 436 437 if (newsk) { 438 const struct inet_request_sock *ireq = inet_rsk(req); 439 struct tcp_request_sock *treq = tcp_rsk(req); 440 struct inet_connection_sock *newicsk = inet_csk(newsk); 441 struct tcp_sock *newtp = tcp_sk(newsk); 442 443 /* Now setup tcp_sock */ 444 newtp->pred_flags = 0; 445 446 newtp->rcv_wup = newtp->copied_seq = 447 newtp->rcv_nxt = treq->rcv_isn + 1; 448 newtp->segs_in = 1; 449 450 newtp->snd_sml = newtp->snd_una = 451 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1; 452 453 tcp_prequeue_init(newtp); 454 INIT_LIST_HEAD(&newtp->tsq_node); 455 456 tcp_init_wl(newtp, treq->rcv_isn); 457 458 newtp->srtt_us = 0; 459 newtp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 460 minmax_reset(&newtp->rtt_min, tcp_time_stamp, ~0U); 461 newicsk->icsk_rto = TCP_TIMEOUT_INIT; 462 463 newtp->packets_out = 0; 464 newtp->retrans_out = 0; 465 newtp->sacked_out = 0; 466 newtp->fackets_out = 0; 467 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 468 newtp->tlp_high_seq = 0; 469 newtp->lsndtime = treq->snt_synack.stamp_jiffies; 470 newsk->sk_txhash = treq->txhash; 471 newtp->last_oow_ack_time = 0; 472 newtp->total_retrans = req->num_retrans; 473 474 /* So many TCP implementations out there (incorrectly) count the 475 * initial SYN frame in their delayed-ACK and congestion control 476 * algorithms that we must have the following bandaid to talk 477 * efficiently to them. -DaveM 478 */ 479 newtp->snd_cwnd = TCP_INIT_CWND; 480 newtp->snd_cwnd_cnt = 0; 481 482 /* There's a bubble in the pipe until at least the first ACK. */ 483 newtp->app_limited = ~0U; 484 485 tcp_init_xmit_timers(newsk); 486 newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1; 487 488 newtp->rx_opt.saw_tstamp = 0; 489 490 newtp->rx_opt.dsack = 0; 491 newtp->rx_opt.num_sacks = 0; 492 493 newtp->urg_data = 0; 494 495 if (sock_flag(newsk, SOCK_KEEPOPEN)) 496 inet_csk_reset_keepalive_timer(newsk, 497 keepalive_time_when(newtp)); 498 499 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 500 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) { 501 if (sysctl_tcp_fack) 502 tcp_enable_fack(newtp); 503 } 504 newtp->window_clamp = req->rsk_window_clamp; 505 newtp->rcv_ssthresh = req->rsk_rcv_wnd; 506 newtp->rcv_wnd = req->rsk_rcv_wnd; 507 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 508 if (newtp->rx_opt.wscale_ok) { 509 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 510 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 511 } else { 512 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 513 newtp->window_clamp = min(newtp->window_clamp, 65535U); 514 } 515 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) << 516 newtp->rx_opt.snd_wscale); 517 newtp->max_window = newtp->snd_wnd; 518 519 if (newtp->rx_opt.tstamp_ok) { 520 newtp->rx_opt.ts_recent = req->ts_recent; 521 newtp->rx_opt.ts_recent_stamp = get_seconds(); 522 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 523 } else { 524 newtp->rx_opt.ts_recent_stamp = 0; 525 newtp->tcp_header_len = sizeof(struct tcphdr); 526 } 527 newtp->tsoffset = treq->ts_off; 528 #ifdef CONFIG_TCP_MD5SIG 529 newtp->md5sig_info = NULL; /*XXX*/ 530 if (newtp->af_specific->md5_lookup(sk, newsk)) 531 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 532 #endif 533 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 534 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 535 newtp->rx_opt.mss_clamp = req->mss; 536 tcp_ecn_openreq_child(newtp, req); 537 newtp->fastopen_rsk = NULL; 538 newtp->syn_data_acked = 0; 539 newtp->rack.mstamp.v64 = 0; 540 newtp->rack.advanced = 0; 541 542 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS); 543 } 544 return newsk; 545 } 546 EXPORT_SYMBOL(tcp_create_openreq_child); 547 548 /* 549 * Process an incoming packet for SYN_RECV sockets represented as a 550 * request_sock. Normally sk is the listener socket but for TFO it 551 * points to the child socket. 552 * 553 * XXX (TFO) - The current impl contains a special check for ack 554 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better? 555 * 556 * We don't need to initialize tmp_opt.sack_ok as we don't use the results 557 */ 558 559 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 560 struct request_sock *req, 561 bool fastopen) 562 { 563 struct tcp_options_received tmp_opt; 564 struct sock *child; 565 const struct tcphdr *th = tcp_hdr(skb); 566 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 567 bool paws_reject = false; 568 bool own_req; 569 570 tmp_opt.saw_tstamp = 0; 571 if (th->doff > (sizeof(struct tcphdr)>>2)) { 572 tcp_parse_options(skb, &tmp_opt, 0, NULL); 573 574 if (tmp_opt.saw_tstamp) { 575 tmp_opt.ts_recent = req->ts_recent; 576 if (tmp_opt.rcv_tsecr) 577 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off; 578 /* We do not store true stamp, but it is not required, 579 * it can be estimated (approximately) 580 * from another data. 581 */ 582 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout); 583 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 584 } 585 } 586 587 /* Check for pure retransmitted SYN. */ 588 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 589 flg == TCP_FLAG_SYN && 590 !paws_reject) { 591 /* 592 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 593 * this case on figure 6 and figure 8, but formal 594 * protocol description says NOTHING. 595 * To be more exact, it says that we should send ACK, 596 * because this segment (at least, if it has no data) 597 * is out of window. 598 * 599 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 600 * describe SYN-RECV state. All the description 601 * is wrong, we cannot believe to it and should 602 * rely only on common sense and implementation 603 * experience. 604 * 605 * Enforce "SYN-ACK" according to figure 8, figure 6 606 * of RFC793, fixed by RFC1122. 607 * 608 * Note that even if there is new data in the SYN packet 609 * they will be thrown away too. 610 * 611 * Reset timer after retransmitting SYNACK, similar to 612 * the idea of fast retransmit in recovery. 613 */ 614 if (!tcp_oow_rate_limited(sock_net(sk), skb, 615 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 616 &tcp_rsk(req)->last_oow_ack_time) && 617 618 !inet_rtx_syn_ack(sk, req)) { 619 unsigned long expires = jiffies; 620 621 expires += min(TCP_TIMEOUT_INIT << req->num_timeout, 622 TCP_RTO_MAX); 623 if (!fastopen) 624 mod_timer_pending(&req->rsk_timer, expires); 625 else 626 req->rsk_timer.expires = expires; 627 } 628 return NULL; 629 } 630 631 /* Further reproduces section "SEGMENT ARRIVES" 632 for state SYN-RECEIVED of RFC793. 633 It is broken, however, it does not work only 634 when SYNs are crossed. 635 636 You would think that SYN crossing is impossible here, since 637 we should have a SYN_SENT socket (from connect()) on our end, 638 but this is not true if the crossed SYNs were sent to both 639 ends by a malicious third party. We must defend against this, 640 and to do that we first verify the ACK (as per RFC793, page 641 36) and reset if it is invalid. Is this a true full defense? 642 To convince ourselves, let us consider a way in which the ACK 643 test can still pass in this 'malicious crossed SYNs' case. 644 Malicious sender sends identical SYNs (and thus identical sequence 645 numbers) to both A and B: 646 647 A: gets SYN, seq=7 648 B: gets SYN, seq=7 649 650 By our good fortune, both A and B select the same initial 651 send sequence number of seven :-) 652 653 A: sends SYN|ACK, seq=7, ack_seq=8 654 B: sends SYN|ACK, seq=7, ack_seq=8 655 656 So we are now A eating this SYN|ACK, ACK test passes. So 657 does sequence test, SYN is truncated, and thus we consider 658 it a bare ACK. 659 660 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 661 bare ACK. Otherwise, we create an established connection. Both 662 ends (listening sockets) accept the new incoming connection and try 663 to talk to each other. 8-) 664 665 Note: This case is both harmless, and rare. Possibility is about the 666 same as us discovering intelligent life on another plant tomorrow. 667 668 But generally, we should (RFC lies!) to accept ACK 669 from SYNACK both here and in tcp_rcv_state_process(). 670 tcp_rcv_state_process() does not, hence, we do not too. 671 672 Note that the case is absolutely generic: 673 we cannot optimize anything here without 674 violating protocol. All the checks must be made 675 before attempt to create socket. 676 */ 677 678 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 679 * and the incoming segment acknowledges something not yet 680 * sent (the segment carries an unacceptable ACK) ... 681 * a reset is sent." 682 * 683 * Invalid ACK: reset will be sent by listening socket. 684 * Note that the ACK validity check for a Fast Open socket is done 685 * elsewhere and is checked directly against the child socket rather 686 * than req because user data may have been sent out. 687 */ 688 if ((flg & TCP_FLAG_ACK) && !fastopen && 689 (TCP_SKB_CB(skb)->ack_seq != 690 tcp_rsk(req)->snt_isn + 1)) 691 return sk; 692 693 /* Also, it would be not so bad idea to check rcv_tsecr, which 694 * is essentially ACK extension and too early or too late values 695 * should cause reset in unsynchronized states. 696 */ 697 698 /* RFC793: "first check sequence number". */ 699 700 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 701 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) { 702 /* Out of window: send ACK and drop. */ 703 if (!(flg & TCP_FLAG_RST) && 704 !tcp_oow_rate_limited(sock_net(sk), skb, 705 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 706 &tcp_rsk(req)->last_oow_ack_time)) 707 req->rsk_ops->send_ack(sk, skb, req); 708 if (paws_reject) 709 __NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 710 return NULL; 711 } 712 713 /* In sequence, PAWS is OK. */ 714 715 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt)) 716 req->ts_recent = tmp_opt.rcv_tsval; 717 718 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 719 /* Truncate SYN, it is out of window starting 720 at tcp_rsk(req)->rcv_isn + 1. */ 721 flg &= ~TCP_FLAG_SYN; 722 } 723 724 /* RFC793: "second check the RST bit" and 725 * "fourth, check the SYN bit" 726 */ 727 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 728 __TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 729 goto embryonic_reset; 730 } 731 732 /* ACK sequence verified above, just make sure ACK is 733 * set. If ACK not set, just silently drop the packet. 734 * 735 * XXX (TFO) - if we ever allow "data after SYN", the 736 * following check needs to be removed. 737 */ 738 if (!(flg & TCP_FLAG_ACK)) 739 return NULL; 740 741 /* For Fast Open no more processing is needed (sk is the 742 * child socket). 743 */ 744 if (fastopen) 745 return sk; 746 747 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 748 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 749 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 750 inet_rsk(req)->acked = 1; 751 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 752 return NULL; 753 } 754 755 /* OK, ACK is valid, create big socket and 756 * feed this segment to it. It will repeat all 757 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 758 * ESTABLISHED STATE. If it will be dropped after 759 * socket is created, wait for troubles. 760 */ 761 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL, 762 req, &own_req); 763 if (!child) 764 goto listen_overflow; 765 766 sock_rps_save_rxhash(child, skb); 767 tcp_synack_rtt_meas(child, req); 768 return inet_csk_complete_hashdance(sk, child, req, own_req); 769 770 listen_overflow: 771 if (!sysctl_tcp_abort_on_overflow) { 772 inet_rsk(req)->acked = 1; 773 return NULL; 774 } 775 776 embryonic_reset: 777 if (!(flg & TCP_FLAG_RST)) { 778 /* Received a bad SYN pkt - for TFO We try not to reset 779 * the local connection unless it's really necessary to 780 * avoid becoming vulnerable to outside attack aiming at 781 * resetting legit local connections. 782 */ 783 req->rsk_ops->send_reset(sk, skb); 784 } else if (fastopen) { /* received a valid RST pkt */ 785 reqsk_fastopen_remove(sk, req, true); 786 tcp_reset(sk); 787 } 788 if (!fastopen) { 789 inet_csk_reqsk_queue_drop(sk, req); 790 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 791 } 792 return NULL; 793 } 794 EXPORT_SYMBOL(tcp_check_req); 795 796 /* 797 * Queue segment on the new socket if the new socket is active, 798 * otherwise we just shortcircuit this and continue with 799 * the new socket. 800 * 801 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV 802 * when entering. But other states are possible due to a race condition 803 * where after __inet_lookup_established() fails but before the listener 804 * locked is obtained, other packets cause the same connection to 805 * be created. 806 */ 807 808 int tcp_child_process(struct sock *parent, struct sock *child, 809 struct sk_buff *skb) 810 { 811 int ret = 0; 812 int state = child->sk_state; 813 814 tcp_segs_in(tcp_sk(child), skb); 815 if (!sock_owned_by_user(child)) { 816 ret = tcp_rcv_state_process(child, skb); 817 /* Wakeup parent, send SIGIO */ 818 if (state == TCP_SYN_RECV && child->sk_state != state) 819 parent->sk_data_ready(parent); 820 } else { 821 /* Alas, it is possible again, because we do lookup 822 * in main socket hash table and lock on listening 823 * socket does not protect us more. 824 */ 825 __sk_add_backlog(child, skb); 826 } 827 828 bh_unlock_sock(child); 829 sock_put(child); 830 return ret; 831 } 832 EXPORT_SYMBOL(tcp_child_process); 833