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