1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * INET An implementation of the TCP/IP protocol suite for the LINUX 4 * operating system. INET is implemented using the BSD Socket 5 * interface as the means of communication with the user level. 6 * 7 * Implementation of the Transmission Control Protocol(TCP). 8 * 9 * Authors: Ross Biro 10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 11 * Mark Evans, <evansmp@uhura.aston.ac.uk> 12 * Corey Minyard <wf-rch!minyard@relay.EU.net> 13 * Florian La Roche, <flla@stud.uni-sb.de> 14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 15 * Linus Torvalds, <torvalds@cs.helsinki.fi> 16 * Alan Cox, <gw4pts@gw4pts.ampr.org> 17 * Matthew Dillon, <dillon@apollo.west.oic.com> 18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 19 * Jorge Cwik, <jorge@laser.satlink.net> 20 */ 21 22 #include <net/tcp.h> 23 #include <net/xfrm.h> 24 #include <net/busy_poll.h> 25 #include <net/rstreason.h> 26 27 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win) 28 { 29 if (seq == s_win) 30 return true; 31 if (after(end_seq, s_win) && before(seq, e_win)) 32 return true; 33 return seq == e_win && seq == end_seq; 34 } 35 36 static enum tcp_tw_status 37 tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw, 38 const struct sk_buff *skb, int mib_idx) 39 { 40 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 41 42 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx, 43 &tcptw->tw_last_oow_ack_time)) { 44 /* Send ACK. Note, we do not put the bucket, 45 * it will be released by caller. 46 */ 47 return TCP_TW_ACK; 48 } 49 50 /* We are rate-limiting, so just release the tw sock and drop skb. */ 51 inet_twsk_put(tw); 52 return TCP_TW_SUCCESS; 53 } 54 55 static void twsk_rcv_nxt_update(struct tcp_timewait_sock *tcptw, u32 seq, 56 u32 rcv_nxt) 57 { 58 #ifdef CONFIG_TCP_AO 59 struct tcp_ao_info *ao; 60 61 ao = rcu_dereference(tcptw->ao_info); 62 if (unlikely(ao && seq < rcv_nxt)) 63 WRITE_ONCE(ao->rcv_sne, ao->rcv_sne + 1); 64 #endif 65 WRITE_ONCE(tcptw->tw_rcv_nxt, seq); 66 } 67 68 /* 69 * * Main purpose of TIME-WAIT state is to close connection gracefully, 70 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN 71 * (and, probably, tail of data) and one or more our ACKs are lost. 72 * * What is TIME-WAIT timeout? It is associated with maximal packet 73 * lifetime in the internet, which results in wrong conclusion, that 74 * it is set to catch "old duplicate segments" wandering out of their path. 75 * It is not quite correct. This timeout is calculated so that it exceeds 76 * maximal retransmission timeout enough to allow to lose one (or more) 77 * segments sent by peer and our ACKs. This time may be calculated from RTO. 78 * * When TIME-WAIT socket receives RST, it means that another end 79 * finally closed and we are allowed to kill TIME-WAIT too. 80 * * Second purpose of TIME-WAIT is catching old duplicate segments. 81 * Well, certainly it is pure paranoia, but if we load TIME-WAIT 82 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs. 83 * * If we invented some more clever way to catch duplicates 84 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs. 85 * 86 * The algorithm below is based on FORMAL INTERPRETATION of RFCs. 87 * When you compare it to RFCs, please, read section SEGMENT ARRIVES 88 * from the very beginning. 89 * 90 * NOTE. With recycling (and later with fin-wait-2) TW bucket 91 * is _not_ stateless. It means, that strictly speaking we must 92 * spinlock it. I do not want! Well, probability of misbehaviour 93 * is ridiculously low and, seems, we could use some mb() tricks 94 * to avoid misread sequence numbers, states etc. --ANK 95 * 96 * We don't need to initialize tmp_out.sack_ok as we don't use the results 97 */ 98 enum tcp_tw_status 99 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb, 100 const struct tcphdr *th, u32 *tw_isn) 101 { 102 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 103 u32 rcv_nxt = READ_ONCE(tcptw->tw_rcv_nxt); 104 struct tcp_options_received tmp_opt; 105 bool paws_reject = false; 106 int ts_recent_stamp; 107 108 tmp_opt.saw_tstamp = 0; 109 ts_recent_stamp = READ_ONCE(tcptw->tw_ts_recent_stamp); 110 if (th->doff > (sizeof(*th) >> 2) && ts_recent_stamp) { 111 tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL); 112 113 if (tmp_opt.saw_tstamp) { 114 if (tmp_opt.rcv_tsecr) 115 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset; 116 tmp_opt.ts_recent = READ_ONCE(tcptw->tw_ts_recent); 117 tmp_opt.ts_recent_stamp = ts_recent_stamp; 118 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 119 } 120 } 121 122 if (READ_ONCE(tw->tw_substate) == TCP_FIN_WAIT2) { 123 /* Just repeat all the checks of tcp_rcv_state_process() */ 124 125 /* Out of window, send ACK */ 126 if (paws_reject || 127 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 128 rcv_nxt, 129 rcv_nxt + tcptw->tw_rcv_wnd)) 130 return tcp_timewait_check_oow_rate_limit( 131 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2); 132 133 if (th->rst) 134 goto kill; 135 136 if (th->syn && !before(TCP_SKB_CB(skb)->seq, rcv_nxt)) 137 return TCP_TW_RST; 138 139 /* Dup ACK? */ 140 if (!th->ack || 141 !after(TCP_SKB_CB(skb)->end_seq, rcv_nxt) || 142 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) { 143 inet_twsk_put(tw); 144 return TCP_TW_SUCCESS; 145 } 146 147 /* New data or FIN. If new data arrive after half-duplex close, 148 * reset. 149 */ 150 if (!th->fin || 151 TCP_SKB_CB(skb)->end_seq != rcv_nxt + 1) 152 return TCP_TW_RST; 153 154 /* FIN arrived, enter true time-wait state. */ 155 WRITE_ONCE(tw->tw_substate, TCP_TIME_WAIT); 156 twsk_rcv_nxt_update(tcptw, TCP_SKB_CB(skb)->end_seq, 157 rcv_nxt); 158 159 if (tmp_opt.saw_tstamp) { 160 WRITE_ONCE(tcptw->tw_ts_recent_stamp, 161 ktime_get_seconds()); 162 WRITE_ONCE(tcptw->tw_ts_recent, 163 tmp_opt.rcv_tsval); 164 } 165 166 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 167 return TCP_TW_ACK; 168 } 169 170 /* 171 * Now real TIME-WAIT state. 172 * 173 * RFC 1122: 174 * "When a connection is [...] on TIME-WAIT state [...] 175 * [a TCP] MAY accept a new SYN from the remote TCP to 176 * reopen the connection directly, if it: 177 * 178 * (1) assigns its initial sequence number for the new 179 * connection to be larger than the largest sequence 180 * number it used on the previous connection incarnation, 181 * and 182 * 183 * (2) returns to TIME-WAIT state if the SYN turns out 184 * to be an old duplicate". 185 */ 186 187 if (!paws_reject && 188 (TCP_SKB_CB(skb)->seq == rcv_nxt && 189 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 190 /* In window segment, it may be only reset or bare ack. */ 191 192 if (th->rst) { 193 /* This is TIME_WAIT assassination, in two flavors. 194 * Oh well... nobody has a sufficient solution to this 195 * protocol bug yet. 196 */ 197 if (!READ_ONCE(twsk_net(tw)->ipv4.sysctl_tcp_rfc1337)) { 198 kill: 199 inet_twsk_deschedule_put(tw); 200 return TCP_TW_SUCCESS; 201 } 202 } else { 203 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 204 } 205 206 if (tmp_opt.saw_tstamp) { 207 WRITE_ONCE(tcptw->tw_ts_recent, 208 tmp_opt.rcv_tsval); 209 WRITE_ONCE(tcptw->tw_ts_recent_stamp, 210 ktime_get_seconds()); 211 } 212 213 inet_twsk_put(tw); 214 return TCP_TW_SUCCESS; 215 } 216 217 /* Out of window segment. 218 219 All the segments are ACKed immediately. 220 221 The only exception is new SYN. We accept it, if it is 222 not old duplicate and we are not in danger to be killed 223 by delayed old duplicates. RFC check is that it has 224 newer sequence number works at rates <40Mbit/sec. 225 However, if paws works, it is reliable AND even more, 226 we even may relax silly seq space cutoff. 227 228 RED-PEN: we violate main RFC requirement, if this SYN will appear 229 old duplicate (i.e. we receive RST in reply to SYN-ACK), 230 we must return socket to time-wait state. It is not good, 231 but not fatal yet. 232 */ 233 234 if (th->syn && !th->rst && !th->ack && !paws_reject && 235 (after(TCP_SKB_CB(skb)->seq, rcv_nxt) || 236 (tmp_opt.saw_tstamp && 237 (s32)(READ_ONCE(tcptw->tw_ts_recent) - tmp_opt.rcv_tsval) < 0))) { 238 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 239 if (isn == 0) 240 isn++; 241 *tw_isn = isn; 242 return TCP_TW_SYN; 243 } 244 245 if (paws_reject) 246 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED); 247 248 if (!th->rst) { 249 /* In this case we must reset the TIMEWAIT timer. 250 * 251 * If it is ACKless SYN it may be both old duplicate 252 * and new good SYN with random sequence number <rcv_nxt. 253 * Do not reschedule in the last case. 254 */ 255 if (paws_reject || th->ack) 256 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 257 258 return tcp_timewait_check_oow_rate_limit( 259 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT); 260 } 261 inet_twsk_put(tw); 262 return TCP_TW_SUCCESS; 263 } 264 EXPORT_SYMBOL(tcp_timewait_state_process); 265 266 static void tcp_time_wait_init(struct sock *sk, struct tcp_timewait_sock *tcptw) 267 { 268 #ifdef CONFIG_TCP_MD5SIG 269 const struct tcp_sock *tp = tcp_sk(sk); 270 struct tcp_md5sig_key *key; 271 272 /* 273 * The timewait bucket does not have the key DB from the 274 * sock structure. We just make a quick copy of the 275 * md5 key being used (if indeed we are using one) 276 * so the timewait ack generating code has the key. 277 */ 278 tcptw->tw_md5_key = NULL; 279 if (!static_branch_unlikely(&tcp_md5_needed.key)) 280 return; 281 282 key = tp->af_specific->md5_lookup(sk, sk); 283 if (key) { 284 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC); 285 if (!tcptw->tw_md5_key) 286 return; 287 if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key)) 288 goto out_free; 289 tcp_md5_add_sigpool(); 290 } 291 return; 292 out_free: 293 WARN_ON_ONCE(1); 294 kfree(tcptw->tw_md5_key); 295 tcptw->tw_md5_key = NULL; 296 #endif 297 } 298 299 /* 300 * Move a socket to time-wait or dead fin-wait-2 state. 301 */ 302 void tcp_time_wait(struct sock *sk, int state, int timeo) 303 { 304 const struct inet_connection_sock *icsk = inet_csk(sk); 305 struct tcp_sock *tp = tcp_sk(sk); 306 struct net *net = sock_net(sk); 307 struct inet_timewait_sock *tw; 308 309 tw = inet_twsk_alloc(sk, &net->ipv4.tcp_death_row, state); 310 311 if (tw) { 312 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 313 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 314 315 tw->tw_transparent = inet_test_bit(TRANSPARENT, sk); 316 tw->tw_mark = sk->sk_mark; 317 tw->tw_priority = READ_ONCE(sk->sk_priority); 318 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 319 tcptw->tw_rcv_nxt = tp->rcv_nxt; 320 tcptw->tw_snd_nxt = tp->snd_nxt; 321 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 322 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 323 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 324 tcptw->tw_ts_offset = tp->tsoffset; 325 tw->tw_usec_ts = tp->tcp_usec_ts; 326 tcptw->tw_last_oow_ack_time = 0; 327 tcptw->tw_tx_delay = tp->tcp_tx_delay; 328 tw->tw_txhash = sk->sk_txhash; 329 tw->tw_tx_queue_mapping = sk->sk_tx_queue_mapping; 330 #ifdef CONFIG_SOCK_RX_QUEUE_MAPPING 331 tw->tw_rx_queue_mapping = sk->sk_rx_queue_mapping; 332 #endif 333 #if IS_ENABLED(CONFIG_IPV6) 334 if (tw->tw_family == PF_INET6) { 335 struct ipv6_pinfo *np = inet6_sk(sk); 336 337 tw->tw_v6_daddr = sk->sk_v6_daddr; 338 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr; 339 tw->tw_tclass = np->tclass; 340 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK); 341 tw->tw_ipv6only = sk->sk_ipv6only; 342 } 343 #endif 344 345 tcp_time_wait_init(sk, tcptw); 346 tcp_ao_time_wait(tcptw, tp); 347 348 /* Get the TIME_WAIT timeout firing. */ 349 if (timeo < rto) 350 timeo = rto; 351 352 if (state == TCP_TIME_WAIT) 353 timeo = TCP_TIMEWAIT_LEN; 354 355 /* Linkage updates. 356 * Note that access to tw after this point is illegal. 357 */ 358 inet_twsk_hashdance_schedule(tw, sk, net->ipv4.tcp_death_row.hashinfo, timeo); 359 } else { 360 /* Sorry, if we're out of memory, just CLOSE this 361 * socket up. We've got bigger problems than 362 * non-graceful socket closings. 363 */ 364 NET_INC_STATS(net, LINUX_MIB_TCPTIMEWAITOVERFLOW); 365 } 366 367 tcp_update_metrics(sk); 368 tcp_done(sk); 369 } 370 EXPORT_SYMBOL(tcp_time_wait); 371 372 #ifdef CONFIG_TCP_MD5SIG 373 static void tcp_md5_twsk_free_rcu(struct rcu_head *head) 374 { 375 struct tcp_md5sig_key *key; 376 377 key = container_of(head, struct tcp_md5sig_key, rcu); 378 kfree(key); 379 static_branch_slow_dec_deferred(&tcp_md5_needed); 380 tcp_md5_release_sigpool(); 381 } 382 #endif 383 384 void tcp_twsk_destructor(struct sock *sk) 385 { 386 #ifdef CONFIG_TCP_MD5SIG 387 if (static_branch_unlikely(&tcp_md5_needed.key)) { 388 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 389 390 if (twsk->tw_md5_key) 391 call_rcu(&twsk->tw_md5_key->rcu, tcp_md5_twsk_free_rcu); 392 } 393 #endif 394 tcp_ao_destroy_sock(sk, true); 395 } 396 EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 397 398 void tcp_twsk_purge(struct list_head *net_exit_list) 399 { 400 bool purged_once = false; 401 struct net *net; 402 403 list_for_each_entry(net, net_exit_list, exit_list) { 404 if (net->ipv4.tcp_death_row.hashinfo->pernet) { 405 /* Even if tw_refcount == 1, we must clean up kernel reqsk */ 406 inet_twsk_purge(net->ipv4.tcp_death_row.hashinfo); 407 } else if (!purged_once) { 408 inet_twsk_purge(&tcp_hashinfo); 409 purged_once = true; 410 } 411 } 412 } 413 414 /* Warning : This function is called without sk_listener being locked. 415 * Be sure to read socket fields once, as their value could change under us. 416 */ 417 void tcp_openreq_init_rwin(struct request_sock *req, 418 const struct sock *sk_listener, 419 const struct dst_entry *dst) 420 { 421 struct inet_request_sock *ireq = inet_rsk(req); 422 const struct tcp_sock *tp = tcp_sk(sk_listener); 423 int full_space = tcp_full_space(sk_listener); 424 u32 window_clamp; 425 __u8 rcv_wscale; 426 u32 rcv_wnd; 427 int mss; 428 429 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 430 window_clamp = READ_ONCE(tp->window_clamp); 431 /* Set this up on the first call only */ 432 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW); 433 434 /* limit the window selection if the user enforce a smaller rx buffer */ 435 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK && 436 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0)) 437 req->rsk_window_clamp = full_space; 438 439 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req); 440 if (rcv_wnd == 0) 441 rcv_wnd = dst_metric(dst, RTAX_INITRWND); 442 else if (full_space < rcv_wnd * mss) 443 full_space = rcv_wnd * mss; 444 445 /* tcp_full_space because it is guaranteed to be the first packet */ 446 tcp_select_initial_window(sk_listener, full_space, 447 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 448 &req->rsk_rcv_wnd, 449 &req->rsk_window_clamp, 450 ireq->wscale_ok, 451 &rcv_wscale, 452 rcv_wnd); 453 ireq->rcv_wscale = rcv_wscale; 454 } 455 EXPORT_SYMBOL(tcp_openreq_init_rwin); 456 457 static void tcp_ecn_openreq_child(struct tcp_sock *tp, 458 const struct request_sock *req) 459 { 460 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 461 } 462 463 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst) 464 { 465 struct inet_connection_sock *icsk = inet_csk(sk); 466 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 467 bool ca_got_dst = false; 468 469 if (ca_key != TCP_CA_UNSPEC) { 470 const struct tcp_congestion_ops *ca; 471 472 rcu_read_lock(); 473 ca = tcp_ca_find_key(ca_key); 474 if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 475 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 476 icsk->icsk_ca_ops = ca; 477 ca_got_dst = true; 478 } 479 rcu_read_unlock(); 480 } 481 482 /* If no valid choice made yet, assign current system default ca. */ 483 if (!ca_got_dst && 484 (!icsk->icsk_ca_setsockopt || 485 !bpf_try_module_get(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner))) 486 tcp_assign_congestion_control(sk); 487 488 tcp_set_ca_state(sk, TCP_CA_Open); 489 } 490 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child); 491 492 static void smc_check_reset_syn_req(const struct tcp_sock *oldtp, 493 struct request_sock *req, 494 struct tcp_sock *newtp) 495 { 496 #if IS_ENABLED(CONFIG_SMC) 497 struct inet_request_sock *ireq; 498 499 if (static_branch_unlikely(&tcp_have_smc)) { 500 ireq = inet_rsk(req); 501 if (oldtp->syn_smc && !ireq->smc_ok) 502 newtp->syn_smc = 0; 503 } 504 #endif 505 } 506 507 /* This is not only more efficient than what we used to do, it eliminates 508 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 509 * 510 * Actually, we could lots of memory writes here. tp of listening 511 * socket contains all necessary default parameters. 512 */ 513 struct sock *tcp_create_openreq_child(const struct sock *sk, 514 struct request_sock *req, 515 struct sk_buff *skb) 516 { 517 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 518 const struct inet_request_sock *ireq = inet_rsk(req); 519 struct tcp_request_sock *treq = tcp_rsk(req); 520 struct inet_connection_sock *newicsk; 521 const struct tcp_sock *oldtp; 522 struct tcp_sock *newtp; 523 u32 seq; 524 525 if (!newsk) 526 return NULL; 527 528 newicsk = inet_csk(newsk); 529 newtp = tcp_sk(newsk); 530 oldtp = tcp_sk(sk); 531 532 smc_check_reset_syn_req(oldtp, req, newtp); 533 534 /* Now setup tcp_sock */ 535 newtp->pred_flags = 0; 536 537 seq = treq->rcv_isn + 1; 538 newtp->rcv_wup = seq; 539 WRITE_ONCE(newtp->copied_seq, seq); 540 WRITE_ONCE(newtp->rcv_nxt, seq); 541 newtp->segs_in = 1; 542 543 seq = treq->snt_isn + 1; 544 newtp->snd_sml = newtp->snd_una = seq; 545 WRITE_ONCE(newtp->snd_nxt, seq); 546 newtp->snd_up = seq; 547 548 INIT_LIST_HEAD(&newtp->tsq_node); 549 INIT_LIST_HEAD(&newtp->tsorted_sent_queue); 550 551 tcp_init_wl(newtp, treq->rcv_isn); 552 553 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U); 554 newicsk->icsk_ack.lrcvtime = tcp_jiffies32; 555 556 newtp->lsndtime = tcp_jiffies32; 557 newsk->sk_txhash = READ_ONCE(treq->txhash); 558 newtp->total_retrans = req->num_retrans; 559 560 tcp_init_xmit_timers(newsk); 561 WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1); 562 563 if (sock_flag(newsk, SOCK_KEEPOPEN)) 564 inet_csk_reset_keepalive_timer(newsk, 565 keepalive_time_when(newtp)); 566 567 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 568 newtp->rx_opt.sack_ok = ireq->sack_ok; 569 newtp->window_clamp = req->rsk_window_clamp; 570 newtp->rcv_ssthresh = req->rsk_rcv_wnd; 571 newtp->rcv_wnd = req->rsk_rcv_wnd; 572 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 573 if (newtp->rx_opt.wscale_ok) { 574 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 575 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 576 } else { 577 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 578 newtp->window_clamp = min(newtp->window_clamp, 65535U); 579 } 580 newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale; 581 newtp->max_window = newtp->snd_wnd; 582 583 if (newtp->rx_opt.tstamp_ok) { 584 newtp->tcp_usec_ts = treq->req_usec_ts; 585 newtp->rx_opt.ts_recent = READ_ONCE(req->ts_recent); 586 newtp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 587 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 588 } else { 589 newtp->tcp_usec_ts = 0; 590 newtp->rx_opt.ts_recent_stamp = 0; 591 newtp->tcp_header_len = sizeof(struct tcphdr); 592 } 593 if (req->num_timeout) { 594 newtp->total_rto = req->num_timeout; 595 newtp->undo_marker = treq->snt_isn; 596 if (newtp->tcp_usec_ts) { 597 newtp->retrans_stamp = treq->snt_synack; 598 newtp->total_rto_time = (u32)(tcp_clock_us() - 599 newtp->retrans_stamp) / USEC_PER_MSEC; 600 } else { 601 newtp->retrans_stamp = div_u64(treq->snt_synack, 602 USEC_PER_SEC / TCP_TS_HZ); 603 newtp->total_rto_time = tcp_clock_ms() - 604 newtp->retrans_stamp; 605 } 606 newtp->total_rto_recoveries = 1; 607 } 608 newtp->tsoffset = treq->ts_off; 609 #ifdef CONFIG_TCP_MD5SIG 610 newtp->md5sig_info = NULL; /*XXX*/ 611 #endif 612 #ifdef CONFIG_TCP_AO 613 newtp->ao_info = NULL; 614 615 if (tcp_rsk_used_ao(req)) { 616 struct tcp_ao_key *ao_key; 617 618 ao_key = treq->af_specific->ao_lookup(sk, req, tcp_rsk(req)->ao_keyid, -1); 619 if (ao_key) 620 newtp->tcp_header_len += tcp_ao_len_aligned(ao_key); 621 } 622 #endif 623 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 624 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 625 newtp->rx_opt.mss_clamp = req->mss; 626 tcp_ecn_openreq_child(newtp, req); 627 newtp->fastopen_req = NULL; 628 RCU_INIT_POINTER(newtp->fastopen_rsk, NULL); 629 630 newtp->bpf_chg_cc_inprogress = 0; 631 tcp_bpf_clone(sk, newsk); 632 633 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS); 634 635 xa_init_flags(&newsk->sk_user_frags, XA_FLAGS_ALLOC1); 636 637 return newsk; 638 } 639 EXPORT_SYMBOL(tcp_create_openreq_child); 640 641 /* 642 * Process an incoming packet for SYN_RECV sockets represented as a 643 * request_sock. Normally sk is the listener socket but for TFO it 644 * points to the child socket. 645 * 646 * XXX (TFO) - The current impl contains a special check for ack 647 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better? 648 * 649 * We don't need to initialize tmp_opt.sack_ok as we don't use the results 650 * 651 * Note: If @fastopen is true, this can be called from process context. 652 * Otherwise, this is from BH context. 653 */ 654 655 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 656 struct request_sock *req, 657 bool fastopen, bool *req_stolen) 658 { 659 struct tcp_options_received tmp_opt; 660 struct sock *child; 661 const struct tcphdr *th = tcp_hdr(skb); 662 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 663 bool paws_reject = false; 664 bool own_req; 665 666 tmp_opt.saw_tstamp = 0; 667 if (th->doff > (sizeof(struct tcphdr)>>2)) { 668 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL); 669 670 if (tmp_opt.saw_tstamp) { 671 tmp_opt.ts_recent = READ_ONCE(req->ts_recent); 672 if (tmp_opt.rcv_tsecr) 673 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off; 674 /* We do not store true stamp, but it is not required, 675 * it can be estimated (approximately) 676 * from another data. 677 */ 678 tmp_opt.ts_recent_stamp = ktime_get_seconds() - reqsk_timeout(req, TCP_RTO_MAX) / HZ; 679 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 680 } 681 } 682 683 /* Check for pure retransmitted SYN. */ 684 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 685 flg == TCP_FLAG_SYN && 686 !paws_reject) { 687 /* 688 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 689 * this case on figure 6 and figure 8, but formal 690 * protocol description says NOTHING. 691 * To be more exact, it says that we should send ACK, 692 * because this segment (at least, if it has no data) 693 * is out of window. 694 * 695 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 696 * describe SYN-RECV state. All the description 697 * is wrong, we cannot believe to it and should 698 * rely only on common sense and implementation 699 * experience. 700 * 701 * Enforce "SYN-ACK" according to figure 8, figure 6 702 * of RFC793, fixed by RFC1122. 703 * 704 * Note that even if there is new data in the SYN packet 705 * they will be thrown away too. 706 * 707 * Reset timer after retransmitting SYNACK, similar to 708 * the idea of fast retransmit in recovery. 709 */ 710 if (!tcp_oow_rate_limited(sock_net(sk), skb, 711 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 712 &tcp_rsk(req)->last_oow_ack_time) && 713 714 !inet_rtx_syn_ack(sk, req)) { 715 unsigned long expires = jiffies; 716 717 expires += reqsk_timeout(req, TCP_RTO_MAX); 718 if (!fastopen) 719 mod_timer_pending(&req->rsk_timer, expires); 720 else 721 req->rsk_timer.expires = expires; 722 } 723 return NULL; 724 } 725 726 /* Further reproduces section "SEGMENT ARRIVES" 727 for state SYN-RECEIVED of RFC793. 728 It is broken, however, it does not work only 729 when SYNs are crossed. 730 731 You would think that SYN crossing is impossible here, since 732 we should have a SYN_SENT socket (from connect()) on our end, 733 but this is not true if the crossed SYNs were sent to both 734 ends by a malicious third party. We must defend against this, 735 and to do that we first verify the ACK (as per RFC793, page 736 36) and reset if it is invalid. Is this a true full defense? 737 To convince ourselves, let us consider a way in which the ACK 738 test can still pass in this 'malicious crossed SYNs' case. 739 Malicious sender sends identical SYNs (and thus identical sequence 740 numbers) to both A and B: 741 742 A: gets SYN, seq=7 743 B: gets SYN, seq=7 744 745 By our good fortune, both A and B select the same initial 746 send sequence number of seven :-) 747 748 A: sends SYN|ACK, seq=7, ack_seq=8 749 B: sends SYN|ACK, seq=7, ack_seq=8 750 751 So we are now A eating this SYN|ACK, ACK test passes. So 752 does sequence test, SYN is truncated, and thus we consider 753 it a bare ACK. 754 755 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 756 bare ACK. Otherwise, we create an established connection. Both 757 ends (listening sockets) accept the new incoming connection and try 758 to talk to each other. 8-) 759 760 Note: This case is both harmless, and rare. Possibility is about the 761 same as us discovering intelligent life on another plant tomorrow. 762 763 But generally, we should (RFC lies!) to accept ACK 764 from SYNACK both here and in tcp_rcv_state_process(). 765 tcp_rcv_state_process() does not, hence, we do not too. 766 767 Note that the case is absolutely generic: 768 we cannot optimize anything here without 769 violating protocol. All the checks must be made 770 before attempt to create socket. 771 */ 772 773 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 774 * and the incoming segment acknowledges something not yet 775 * sent (the segment carries an unacceptable ACK) ... 776 * a reset is sent." 777 * 778 * Invalid ACK: reset will be sent by listening socket. 779 * Note that the ACK validity check for a Fast Open socket is done 780 * elsewhere and is checked directly against the child socket rather 781 * than req because user data may have been sent out. 782 */ 783 if ((flg & TCP_FLAG_ACK) && !fastopen && 784 (TCP_SKB_CB(skb)->ack_seq != 785 tcp_rsk(req)->snt_isn + 1)) 786 return sk; 787 788 /* Also, it would be not so bad idea to check rcv_tsecr, which 789 * is essentially ACK extension and too early or too late values 790 * should cause reset in unsynchronized states. 791 */ 792 793 /* RFC793: "first check sequence number". */ 794 795 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, 796 TCP_SKB_CB(skb)->end_seq, 797 tcp_rsk(req)->rcv_nxt, 798 tcp_rsk(req)->rcv_nxt + 799 tcp_synack_window(req))) { 800 /* Out of window: send ACK and drop. */ 801 if (!(flg & TCP_FLAG_RST) && 802 !tcp_oow_rate_limited(sock_net(sk), skb, 803 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 804 &tcp_rsk(req)->last_oow_ack_time)) 805 req->rsk_ops->send_ack(sk, skb, req); 806 if (paws_reject) 807 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 808 return NULL; 809 } 810 811 /* In sequence, PAWS is OK. */ 812 813 /* TODO: We probably should defer ts_recent change once 814 * we take ownership of @req. 815 */ 816 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt)) 817 WRITE_ONCE(req->ts_recent, tmp_opt.rcv_tsval); 818 819 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 820 /* Truncate SYN, it is out of window starting 821 at tcp_rsk(req)->rcv_isn + 1. */ 822 flg &= ~TCP_FLAG_SYN; 823 } 824 825 /* RFC793: "second check the RST bit" and 826 * "fourth, check the SYN bit" 827 */ 828 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 829 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 830 goto embryonic_reset; 831 } 832 833 /* ACK sequence verified above, just make sure ACK is 834 * set. If ACK not set, just silently drop the packet. 835 * 836 * XXX (TFO) - if we ever allow "data after SYN", the 837 * following check needs to be removed. 838 */ 839 if (!(flg & TCP_FLAG_ACK)) 840 return NULL; 841 842 /* For Fast Open no more processing is needed (sk is the 843 * child socket). 844 */ 845 if (fastopen) 846 return sk; 847 848 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 849 if (req->num_timeout < READ_ONCE(inet_csk(sk)->icsk_accept_queue.rskq_defer_accept) && 850 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 851 inet_rsk(req)->acked = 1; 852 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 853 return NULL; 854 } 855 856 /* OK, ACK is valid, create big socket and 857 * feed this segment to it. It will repeat all 858 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 859 * ESTABLISHED STATE. If it will be dropped after 860 * socket is created, wait for troubles. 861 */ 862 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL, 863 req, &own_req); 864 if (!child) 865 goto listen_overflow; 866 867 if (own_req && rsk_drop_req(req)) { 868 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req); 869 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, req); 870 return child; 871 } 872 873 sock_rps_save_rxhash(child, skb); 874 tcp_synack_rtt_meas(child, req); 875 *req_stolen = !own_req; 876 return inet_csk_complete_hashdance(sk, child, req, own_req); 877 878 listen_overflow: 879 if (sk != req->rsk_listener) 880 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE); 881 882 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow)) { 883 inet_rsk(req)->acked = 1; 884 return NULL; 885 } 886 887 embryonic_reset: 888 if (!(flg & TCP_FLAG_RST)) { 889 /* Received a bad SYN pkt - for TFO We try not to reset 890 * the local connection unless it's really necessary to 891 * avoid becoming vulnerable to outside attack aiming at 892 * resetting legit local connections. 893 */ 894 req->rsk_ops->send_reset(sk, skb, SK_RST_REASON_INVALID_SYN); 895 } else if (fastopen) { /* received a valid RST pkt */ 896 reqsk_fastopen_remove(sk, req, true); 897 tcp_reset(sk, skb); 898 } 899 if (!fastopen) { 900 bool unlinked = inet_csk_reqsk_queue_drop(sk, req); 901 902 if (unlinked) 903 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 904 *req_stolen = !unlinked; 905 } 906 return NULL; 907 } 908 EXPORT_SYMBOL(tcp_check_req); 909 910 /* 911 * Queue segment on the new socket if the new socket is active, 912 * otherwise we just shortcircuit this and continue with 913 * the new socket. 914 * 915 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV 916 * when entering. But other states are possible due to a race condition 917 * where after __inet_lookup_established() fails but before the listener 918 * locked is obtained, other packets cause the same connection to 919 * be created. 920 */ 921 922 enum skb_drop_reason tcp_child_process(struct sock *parent, struct sock *child, 923 struct sk_buff *skb) 924 __releases(&((child)->sk_lock.slock)) 925 { 926 enum skb_drop_reason reason = SKB_NOT_DROPPED_YET; 927 int state = child->sk_state; 928 929 /* record sk_napi_id and sk_rx_queue_mapping of child. */ 930 sk_mark_napi_id_set(child, skb); 931 932 tcp_segs_in(tcp_sk(child), skb); 933 if (!sock_owned_by_user(child)) { 934 reason = tcp_rcv_state_process(child, skb); 935 /* Wakeup parent, send SIGIO */ 936 if (state == TCP_SYN_RECV && child->sk_state != state) 937 parent->sk_data_ready(parent); 938 } else { 939 /* Alas, it is possible again, because we do lookup 940 * in main socket hash table and lock on listening 941 * socket does not protect us more. 942 */ 943 __sk_add_backlog(child, skb); 944 } 945 946 bh_unlock_sock(child); 947 sock_put(child); 948 return reason; 949 } 950 EXPORT_SYMBOL(tcp_child_process); 951