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 <linux/static_key.h> 27 #include <net/tcp.h> 28 #include <net/inet_common.h> 29 #include <net/xfrm.h> 30 #include <net/busy_poll.h> 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 98 tmp_opt.saw_tstamp = 0; 99 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) { 100 tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL); 101 102 if (tmp_opt.saw_tstamp) { 103 if (tmp_opt.rcv_tsecr) 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 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 152 return TCP_TW_ACK; 153 } 154 155 /* 156 * Now real TIME-WAIT state. 157 * 158 * RFC 1122: 159 * "When a connection is [...] on TIME-WAIT state [...] 160 * [a TCP] MAY accept a new SYN from the remote TCP to 161 * reopen the connection directly, if it: 162 * 163 * (1) assigns its initial sequence number for the new 164 * connection to be larger than the largest sequence 165 * number it used on the previous connection incarnation, 166 * and 167 * 168 * (2) returns to TIME-WAIT state if the SYN turns out 169 * to be an old duplicate". 170 */ 171 172 if (!paws_reject && 173 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt && 174 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 175 /* In window segment, it may be only reset or bare ack. */ 176 177 if (th->rst) { 178 /* This is TIME_WAIT assassination, in two flavors. 179 * Oh well... nobody has a sufficient solution to this 180 * protocol bug yet. 181 */ 182 if (twsk_net(tw)->ipv4.sysctl_tcp_rfc1337 == 0) { 183 kill: 184 inet_twsk_deschedule_put(tw); 185 return TCP_TW_SUCCESS; 186 } 187 } 188 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 189 190 if (tmp_opt.saw_tstamp) { 191 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 192 tcptw->tw_ts_recent_stamp = get_seconds(); 193 } 194 195 inet_twsk_put(tw); 196 return TCP_TW_SUCCESS; 197 } 198 199 /* Out of window segment. 200 201 All the segments are ACKed immediately. 202 203 The only exception is new SYN. We accept it, if it is 204 not old duplicate and we are not in danger to be killed 205 by delayed old duplicates. RFC check is that it has 206 newer sequence number works at rates <40Mbit/sec. 207 However, if paws works, it is reliable AND even more, 208 we even may relax silly seq space cutoff. 209 210 RED-PEN: we violate main RFC requirement, if this SYN will appear 211 old duplicate (i.e. we receive RST in reply to SYN-ACK), 212 we must return socket to time-wait state. It is not good, 213 but not fatal yet. 214 */ 215 216 if (th->syn && !th->rst && !th->ack && !paws_reject && 217 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) || 218 (tmp_opt.saw_tstamp && 219 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) { 220 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 221 if (isn == 0) 222 isn++; 223 TCP_SKB_CB(skb)->tcp_tw_isn = isn; 224 return TCP_TW_SYN; 225 } 226 227 if (paws_reject) 228 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED); 229 230 if (!th->rst) { 231 /* In this case we must reset the TIMEWAIT timer. 232 * 233 * If it is ACKless SYN it may be both old duplicate 234 * and new good SYN with random sequence number <rcv_nxt. 235 * Do not reschedule in the last case. 236 */ 237 if (paws_reject || th->ack) 238 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 239 240 return tcp_timewait_check_oow_rate_limit( 241 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT); 242 } 243 inet_twsk_put(tw); 244 return TCP_TW_SUCCESS; 245 } 246 EXPORT_SYMBOL(tcp_timewait_state_process); 247 248 /* 249 * Move a socket to time-wait or dead fin-wait-2 state. 250 */ 251 void tcp_time_wait(struct sock *sk, int state, int timeo) 252 { 253 const struct inet_connection_sock *icsk = inet_csk(sk); 254 const struct tcp_sock *tp = tcp_sk(sk); 255 struct inet_timewait_sock *tw; 256 struct inet_timewait_death_row *tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row; 257 258 tw = inet_twsk_alloc(sk, tcp_death_row, state); 259 260 if (tw) { 261 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 262 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 263 struct inet_sock *inet = inet_sk(sk); 264 265 tw->tw_transparent = inet->transparent; 266 tw->tw_mark = sk->sk_mark; 267 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 268 tcptw->tw_rcv_nxt = tp->rcv_nxt; 269 tcptw->tw_snd_nxt = tp->snd_nxt; 270 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 271 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 272 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 273 tcptw->tw_ts_offset = tp->tsoffset; 274 tcptw->tw_last_oow_ack_time = 0; 275 276 #if IS_ENABLED(CONFIG_IPV6) 277 if (tw->tw_family == PF_INET6) { 278 struct ipv6_pinfo *np = inet6_sk(sk); 279 280 tw->tw_v6_daddr = sk->sk_v6_daddr; 281 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr; 282 tw->tw_tclass = np->tclass; 283 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK); 284 tw->tw_ipv6only = sk->sk_ipv6only; 285 } 286 #endif 287 288 #ifdef CONFIG_TCP_MD5SIG 289 /* 290 * The timewait bucket does not have the key DB from the 291 * sock structure. We just make a quick copy of the 292 * md5 key being used (if indeed we are using one) 293 * so the timewait ack generating code has the key. 294 */ 295 do { 296 struct tcp_md5sig_key *key; 297 tcptw->tw_md5_key = NULL; 298 key = tp->af_specific->md5_lookup(sk, sk); 299 if (key) { 300 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC); 301 BUG_ON(tcptw->tw_md5_key && !tcp_alloc_md5sig_pool()); 302 } 303 } while (0); 304 #endif 305 306 /* Get the TIME_WAIT timeout firing. */ 307 if (timeo < rto) 308 timeo = rto; 309 310 tw->tw_timeout = TCP_TIMEWAIT_LEN; 311 if (state == TCP_TIME_WAIT) 312 timeo = TCP_TIMEWAIT_LEN; 313 314 /* tw_timer is pinned, so we need to make sure BH are disabled 315 * in following section, otherwise timer handler could run before 316 * we complete the initialization. 317 */ 318 local_bh_disable(); 319 inet_twsk_schedule(tw, timeo); 320 /* Linkage updates. 321 * Note that access to tw after this point is illegal. 322 */ 323 inet_twsk_hashdance(tw, sk, &tcp_hashinfo); 324 local_bh_enable(); 325 } else { 326 /* Sorry, if we're out of memory, just CLOSE this 327 * socket up. We've got bigger problems than 328 * non-graceful socket closings. 329 */ 330 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW); 331 } 332 333 tcp_update_metrics(sk); 334 tcp_done(sk); 335 } 336 EXPORT_SYMBOL(tcp_time_wait); 337 338 void tcp_twsk_destructor(struct sock *sk) 339 { 340 #ifdef CONFIG_TCP_MD5SIG 341 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 342 343 if (twsk->tw_md5_key) 344 kfree_rcu(twsk->tw_md5_key, rcu); 345 #endif 346 } 347 EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 348 349 /* Warning : This function is called without sk_listener being locked. 350 * Be sure to read socket fields once, as their value could change under us. 351 */ 352 void tcp_openreq_init_rwin(struct request_sock *req, 353 const struct sock *sk_listener, 354 const struct dst_entry *dst) 355 { 356 struct inet_request_sock *ireq = inet_rsk(req); 357 const struct tcp_sock *tp = tcp_sk(sk_listener); 358 int full_space = tcp_full_space(sk_listener); 359 u32 window_clamp; 360 __u8 rcv_wscale; 361 u32 rcv_wnd; 362 int mss; 363 364 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 365 window_clamp = READ_ONCE(tp->window_clamp); 366 /* Set this up on the first call only */ 367 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW); 368 369 /* limit the window selection if the user enforce a smaller rx buffer */ 370 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK && 371 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0)) 372 req->rsk_window_clamp = full_space; 373 374 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req); 375 if (rcv_wnd == 0) 376 rcv_wnd = dst_metric(dst, RTAX_INITRWND); 377 else if (full_space < rcv_wnd * mss) 378 full_space = rcv_wnd * mss; 379 380 /* tcp_full_space because it is guaranteed to be the first packet */ 381 tcp_select_initial_window(sk_listener, full_space, 382 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 383 &req->rsk_rcv_wnd, 384 &req->rsk_window_clamp, 385 ireq->wscale_ok, 386 &rcv_wscale, 387 rcv_wnd); 388 ireq->rcv_wscale = rcv_wscale; 389 } 390 EXPORT_SYMBOL(tcp_openreq_init_rwin); 391 392 static void tcp_ecn_openreq_child(struct tcp_sock *tp, 393 const struct request_sock *req) 394 { 395 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 396 } 397 398 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst) 399 { 400 struct inet_connection_sock *icsk = inet_csk(sk); 401 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 402 bool ca_got_dst = false; 403 404 if (ca_key != TCP_CA_UNSPEC) { 405 const struct tcp_congestion_ops *ca; 406 407 rcu_read_lock(); 408 ca = tcp_ca_find_key(ca_key); 409 if (likely(ca && try_module_get(ca->owner))) { 410 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 411 icsk->icsk_ca_ops = ca; 412 ca_got_dst = true; 413 } 414 rcu_read_unlock(); 415 } 416 417 /* If no valid choice made yet, assign current system default ca. */ 418 if (!ca_got_dst && 419 (!icsk->icsk_ca_setsockopt || 420 !try_module_get(icsk->icsk_ca_ops->owner))) 421 tcp_assign_congestion_control(sk); 422 423 tcp_set_ca_state(sk, TCP_CA_Open); 424 } 425 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child); 426 427 static void smc_check_reset_syn_req(struct tcp_sock *oldtp, 428 struct request_sock *req, 429 struct tcp_sock *newtp) 430 { 431 #if IS_ENABLED(CONFIG_SMC) 432 struct inet_request_sock *ireq; 433 434 if (static_branch_unlikely(&tcp_have_smc)) { 435 ireq = inet_rsk(req); 436 if (oldtp->syn_smc && !ireq->smc_ok) 437 newtp->syn_smc = 0; 438 } 439 #endif 440 } 441 442 /* This is not only more efficient than what we used to do, it eliminates 443 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 444 * 445 * Actually, we could lots of memory writes here. tp of listening 446 * socket contains all necessary default parameters. 447 */ 448 struct sock *tcp_create_openreq_child(const struct sock *sk, 449 struct request_sock *req, 450 struct sk_buff *skb) 451 { 452 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 453 454 if (newsk) { 455 const struct inet_request_sock *ireq = inet_rsk(req); 456 struct tcp_request_sock *treq = tcp_rsk(req); 457 struct inet_connection_sock *newicsk = inet_csk(newsk); 458 struct tcp_sock *newtp = tcp_sk(newsk); 459 struct tcp_sock *oldtp = tcp_sk(sk); 460 461 smc_check_reset_syn_req(oldtp, req, newtp); 462 463 /* Now setup tcp_sock */ 464 newtp->pred_flags = 0; 465 466 newtp->rcv_wup = newtp->copied_seq = 467 newtp->rcv_nxt = treq->rcv_isn + 1; 468 newtp->segs_in = 1; 469 470 newtp->snd_sml = newtp->snd_una = 471 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1; 472 473 INIT_LIST_HEAD(&newtp->tsq_node); 474 INIT_LIST_HEAD(&newtp->tsorted_sent_queue); 475 476 tcp_init_wl(newtp, treq->rcv_isn); 477 478 newtp->srtt_us = 0; 479 newtp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 480 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U); 481 newicsk->icsk_rto = TCP_TIMEOUT_INIT; 482 newicsk->icsk_ack.lrcvtime = tcp_jiffies32; 483 484 newtp->packets_out = 0; 485 newtp->retrans_out = 0; 486 newtp->sacked_out = 0; 487 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 488 newtp->tlp_high_seq = 0; 489 newtp->lsndtime = tcp_jiffies32; 490 newsk->sk_txhash = treq->txhash; 491 newtp->last_oow_ack_time = 0; 492 newtp->total_retrans = req->num_retrans; 493 494 /* So many TCP implementations out there (incorrectly) count the 495 * initial SYN frame in their delayed-ACK and congestion control 496 * algorithms that we must have the following bandaid to talk 497 * efficiently to them. -DaveM 498 */ 499 newtp->snd_cwnd = TCP_INIT_CWND; 500 newtp->snd_cwnd_cnt = 0; 501 502 /* There's a bubble in the pipe until at least the first ACK. */ 503 newtp->app_limited = ~0U; 504 505 tcp_init_xmit_timers(newsk); 506 newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1; 507 508 newtp->rx_opt.saw_tstamp = 0; 509 510 newtp->rx_opt.dsack = 0; 511 newtp->rx_opt.num_sacks = 0; 512 513 newtp->urg_data = 0; 514 515 if (sock_flag(newsk, SOCK_KEEPOPEN)) 516 inet_csk_reset_keepalive_timer(newsk, 517 keepalive_time_when(newtp)); 518 519 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 520 newtp->rx_opt.sack_ok = ireq->sack_ok; 521 newtp->window_clamp = req->rsk_window_clamp; 522 newtp->rcv_ssthresh = req->rsk_rcv_wnd; 523 newtp->rcv_wnd = req->rsk_rcv_wnd; 524 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 525 if (newtp->rx_opt.wscale_ok) { 526 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 527 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 528 } else { 529 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 530 newtp->window_clamp = min(newtp->window_clamp, 65535U); 531 } 532 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) << 533 newtp->rx_opt.snd_wscale); 534 newtp->max_window = newtp->snd_wnd; 535 536 if (newtp->rx_opt.tstamp_ok) { 537 newtp->rx_opt.ts_recent = req->ts_recent; 538 newtp->rx_opt.ts_recent_stamp = get_seconds(); 539 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 540 } else { 541 newtp->rx_opt.ts_recent_stamp = 0; 542 newtp->tcp_header_len = sizeof(struct tcphdr); 543 } 544 newtp->tsoffset = treq->ts_off; 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 newtp->fastopen_req = NULL; 555 newtp->fastopen_rsk = NULL; 556 newtp->syn_data_acked = 0; 557 newtp->rack.mstamp = 0; 558 newtp->rack.advanced = 0; 559 newtp->rack.reo_wnd_steps = 1; 560 newtp->rack.last_delivered = 0; 561 newtp->rack.reo_wnd_persist = 0; 562 newtp->rack.dsack_seen = 0; 563 564 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS); 565 } 566 return newsk; 567 } 568 EXPORT_SYMBOL(tcp_create_openreq_child); 569 570 /* 571 * Process an incoming packet for SYN_RECV sockets represented as a 572 * request_sock. Normally sk is the listener socket but for TFO it 573 * points to the child socket. 574 * 575 * XXX (TFO) - The current impl contains a special check for ack 576 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better? 577 * 578 * We don't need to initialize tmp_opt.sack_ok as we don't use the results 579 */ 580 581 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 582 struct request_sock *req, 583 bool fastopen, bool *req_stolen) 584 { 585 struct tcp_options_received tmp_opt; 586 struct sock *child; 587 const struct tcphdr *th = tcp_hdr(skb); 588 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 589 bool paws_reject = false; 590 bool own_req; 591 592 tmp_opt.saw_tstamp = 0; 593 if (th->doff > (sizeof(struct tcphdr)>>2)) { 594 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL); 595 596 if (tmp_opt.saw_tstamp) { 597 tmp_opt.ts_recent = req->ts_recent; 598 if (tmp_opt.rcv_tsecr) 599 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off; 600 /* We do not store true stamp, but it is not required, 601 * it can be estimated (approximately) 602 * from another data. 603 */ 604 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout); 605 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 606 } 607 } 608 609 /* Check for pure retransmitted SYN. */ 610 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 611 flg == TCP_FLAG_SYN && 612 !paws_reject) { 613 /* 614 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 615 * this case on figure 6 and figure 8, but formal 616 * protocol description says NOTHING. 617 * To be more exact, it says that we should send ACK, 618 * because this segment (at least, if it has no data) 619 * is out of window. 620 * 621 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 622 * describe SYN-RECV state. All the description 623 * is wrong, we cannot believe to it and should 624 * rely only on common sense and implementation 625 * experience. 626 * 627 * Enforce "SYN-ACK" according to figure 8, figure 6 628 * of RFC793, fixed by RFC1122. 629 * 630 * Note that even if there is new data in the SYN packet 631 * they will be thrown away too. 632 * 633 * Reset timer after retransmitting SYNACK, similar to 634 * the idea of fast retransmit in recovery. 635 */ 636 if (!tcp_oow_rate_limited(sock_net(sk), skb, 637 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 638 &tcp_rsk(req)->last_oow_ack_time) && 639 640 !inet_rtx_syn_ack(sk, req)) { 641 unsigned long expires = jiffies; 642 643 expires += min(TCP_TIMEOUT_INIT << req->num_timeout, 644 TCP_RTO_MAX); 645 if (!fastopen) 646 mod_timer_pending(&req->rsk_timer, expires); 647 else 648 req->rsk_timer.expires = expires; 649 } 650 return NULL; 651 } 652 653 /* Further reproduces section "SEGMENT ARRIVES" 654 for state SYN-RECEIVED of RFC793. 655 It is broken, however, it does not work only 656 when SYNs are crossed. 657 658 You would think that SYN crossing is impossible here, since 659 we should have a SYN_SENT socket (from connect()) on our end, 660 but this is not true if the crossed SYNs were sent to both 661 ends by a malicious third party. We must defend against this, 662 and to do that we first verify the ACK (as per RFC793, page 663 36) and reset if it is invalid. Is this a true full defense? 664 To convince ourselves, let us consider a way in which the ACK 665 test can still pass in this 'malicious crossed SYNs' case. 666 Malicious sender sends identical SYNs (and thus identical sequence 667 numbers) to both A and B: 668 669 A: gets SYN, seq=7 670 B: gets SYN, seq=7 671 672 By our good fortune, both A and B select the same initial 673 send sequence number of seven :-) 674 675 A: sends SYN|ACK, seq=7, ack_seq=8 676 B: sends SYN|ACK, seq=7, ack_seq=8 677 678 So we are now A eating this SYN|ACK, ACK test passes. So 679 does sequence test, SYN is truncated, and thus we consider 680 it a bare ACK. 681 682 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 683 bare ACK. Otherwise, we create an established connection. Both 684 ends (listening sockets) accept the new incoming connection and try 685 to talk to each other. 8-) 686 687 Note: This case is both harmless, and rare. Possibility is about the 688 same as us discovering intelligent life on another plant tomorrow. 689 690 But generally, we should (RFC lies!) to accept ACK 691 from SYNACK both here and in tcp_rcv_state_process(). 692 tcp_rcv_state_process() does not, hence, we do not too. 693 694 Note that the case is absolutely generic: 695 we cannot optimize anything here without 696 violating protocol. All the checks must be made 697 before attempt to create socket. 698 */ 699 700 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 701 * and the incoming segment acknowledges something not yet 702 * sent (the segment carries an unacceptable ACK) ... 703 * a reset is sent." 704 * 705 * Invalid ACK: reset will be sent by listening socket. 706 * Note that the ACK validity check for a Fast Open socket is done 707 * elsewhere and is checked directly against the child socket rather 708 * than req because user data may have been sent out. 709 */ 710 if ((flg & TCP_FLAG_ACK) && !fastopen && 711 (TCP_SKB_CB(skb)->ack_seq != 712 tcp_rsk(req)->snt_isn + 1)) 713 return sk; 714 715 /* Also, it would be not so bad idea to check rcv_tsecr, which 716 * is essentially ACK extension and too early or too late values 717 * should cause reset in unsynchronized states. 718 */ 719 720 /* RFC793: "first check sequence number". */ 721 722 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 723 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) { 724 /* Out of window: send ACK and drop. */ 725 if (!(flg & TCP_FLAG_RST) && 726 !tcp_oow_rate_limited(sock_net(sk), skb, 727 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 728 &tcp_rsk(req)->last_oow_ack_time)) 729 req->rsk_ops->send_ack(sk, skb, req); 730 if (paws_reject) 731 __NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 732 return NULL; 733 } 734 735 /* In sequence, PAWS is OK. */ 736 737 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt)) 738 req->ts_recent = tmp_opt.rcv_tsval; 739 740 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 741 /* Truncate SYN, it is out of window starting 742 at tcp_rsk(req)->rcv_isn + 1. */ 743 flg &= ~TCP_FLAG_SYN; 744 } 745 746 /* RFC793: "second check the RST bit" and 747 * "fourth, check the SYN bit" 748 */ 749 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 750 __TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 751 goto embryonic_reset; 752 } 753 754 /* ACK sequence verified above, just make sure ACK is 755 * set. If ACK not set, just silently drop the packet. 756 * 757 * XXX (TFO) - if we ever allow "data after SYN", the 758 * following check needs to be removed. 759 */ 760 if (!(flg & TCP_FLAG_ACK)) 761 return NULL; 762 763 /* For Fast Open no more processing is needed (sk is the 764 * child socket). 765 */ 766 if (fastopen) 767 return sk; 768 769 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 770 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 771 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 772 inet_rsk(req)->acked = 1; 773 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 774 return NULL; 775 } 776 777 /* OK, ACK is valid, create big socket and 778 * feed this segment to it. It will repeat all 779 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 780 * ESTABLISHED STATE. If it will be dropped after 781 * socket is created, wait for troubles. 782 */ 783 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL, 784 req, &own_req); 785 if (!child) 786 goto listen_overflow; 787 788 sock_rps_save_rxhash(child, skb); 789 tcp_synack_rtt_meas(child, req); 790 *req_stolen = !own_req; 791 return inet_csk_complete_hashdance(sk, child, req, own_req); 792 793 listen_overflow: 794 if (!sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow) { 795 inet_rsk(req)->acked = 1; 796 return NULL; 797 } 798 799 embryonic_reset: 800 if (!(flg & TCP_FLAG_RST)) { 801 /* Received a bad SYN pkt - for TFO We try not to reset 802 * the local connection unless it's really necessary to 803 * avoid becoming vulnerable to outside attack aiming at 804 * resetting legit local connections. 805 */ 806 req->rsk_ops->send_reset(sk, skb); 807 } else if (fastopen) { /* received a valid RST pkt */ 808 reqsk_fastopen_remove(sk, req, true); 809 tcp_reset(sk); 810 } 811 if (!fastopen) { 812 inet_csk_reqsk_queue_drop(sk, req); 813 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 814 } 815 return NULL; 816 } 817 EXPORT_SYMBOL(tcp_check_req); 818 819 /* 820 * Queue segment on the new socket if the new socket is active, 821 * otherwise we just shortcircuit this and continue with 822 * the new socket. 823 * 824 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV 825 * when entering. But other states are possible due to a race condition 826 * where after __inet_lookup_established() fails but before the listener 827 * locked is obtained, other packets cause the same connection to 828 * be created. 829 */ 830 831 int tcp_child_process(struct sock *parent, struct sock *child, 832 struct sk_buff *skb) 833 { 834 int ret = 0; 835 int state = child->sk_state; 836 837 /* record NAPI ID of child */ 838 sk_mark_napi_id(child, skb); 839 840 tcp_segs_in(tcp_sk(child), skb); 841 if (!sock_owned_by_user(child)) { 842 ret = tcp_rcv_state_process(child, skb); 843 /* Wakeup parent, send SIGIO */ 844 if (state == TCP_SYN_RECV && child->sk_state != state) 845 parent->sk_data_ready(parent); 846 } else { 847 /* Alas, it is possible again, because we do lookup 848 * in main socket hash table and lock on listening 849 * socket does not protect us more. 850 */ 851 __sk_add_backlog(child, skb); 852 } 853 854 bh_unlock_sock(child); 855 sock_put(child); 856 return ret; 857 } 858 EXPORT_SYMBOL(tcp_child_process); 859