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_OOW; 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 enum skb_drop_reason *drop_reason) 102 { 103 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 104 u32 rcv_nxt = READ_ONCE(tcptw->tw_rcv_nxt); 105 struct tcp_options_received tmp_opt; 106 bool paws_reject = false; 107 int ts_recent_stamp; 108 109 tmp_opt.saw_tstamp = 0; 110 ts_recent_stamp = READ_ONCE(tcptw->tw_ts_recent_stamp); 111 if (th->doff > (sizeof(*th) >> 2) && ts_recent_stamp) { 112 tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL); 113 114 if (tmp_opt.saw_tstamp) { 115 if (tmp_opt.rcv_tsecr) 116 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset; 117 tmp_opt.ts_recent = READ_ONCE(tcptw->tw_ts_recent); 118 tmp_opt.ts_recent_stamp = ts_recent_stamp; 119 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 120 } 121 } 122 123 if (READ_ONCE(tw->tw_substate) == TCP_FIN_WAIT2) { 124 /* Just repeat all the checks of tcp_rcv_state_process() */ 125 126 /* Out of window, send ACK */ 127 if (paws_reject || 128 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 129 rcv_nxt, 130 rcv_nxt + tcptw->tw_rcv_wnd)) 131 return tcp_timewait_check_oow_rate_limit( 132 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2); 133 134 if (th->rst) 135 goto kill; 136 137 if (th->syn && !before(TCP_SKB_CB(skb)->seq, rcv_nxt)) 138 return TCP_TW_RST; 139 140 /* Dup ACK? */ 141 if (!th->ack || 142 !after(TCP_SKB_CB(skb)->end_seq, rcv_nxt) || 143 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) { 144 inet_twsk_put(tw); 145 return TCP_TW_SUCCESS; 146 } 147 148 /* New data or FIN. If new data arrive after half-duplex close, 149 * reset. 150 */ 151 if (!th->fin || 152 TCP_SKB_CB(skb)->end_seq != rcv_nxt + 1) 153 return TCP_TW_RST; 154 155 /* FIN arrived, enter true time-wait state. */ 156 WRITE_ONCE(tw->tw_substate, TCP_TIME_WAIT); 157 twsk_rcv_nxt_update(tcptw, TCP_SKB_CB(skb)->end_seq, 158 rcv_nxt); 159 160 if (tmp_opt.saw_tstamp) { 161 u64 ts = tcp_clock_ms(); 162 163 WRITE_ONCE(tw->tw_entry_stamp, ts); 164 WRITE_ONCE(tcptw->tw_ts_recent_stamp, 165 div_u64(ts, MSEC_PER_SEC)); 166 WRITE_ONCE(tcptw->tw_ts_recent, 167 tmp_opt.rcv_tsval); 168 } 169 170 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 171 return TCP_TW_ACK; 172 } 173 174 /* 175 * Now real TIME-WAIT state. 176 * 177 * RFC 1122: 178 * "When a connection is [...] on TIME-WAIT state [...] 179 * [a TCP] MAY accept a new SYN from the remote TCP to 180 * reopen the connection directly, if it: 181 * 182 * (1) assigns its initial sequence number for the new 183 * connection to be larger than the largest sequence 184 * number it used on the previous connection incarnation, 185 * and 186 * 187 * (2) returns to TIME-WAIT state if the SYN turns out 188 * to be an old duplicate". 189 */ 190 191 if (!paws_reject && 192 (TCP_SKB_CB(skb)->seq == rcv_nxt && 193 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 194 /* In window segment, it may be only reset or bare ack. */ 195 196 if (th->rst) { 197 /* This is TIME_WAIT assassination, in two flavors. 198 * Oh well... nobody has a sufficient solution to this 199 * protocol bug yet. 200 */ 201 if (!READ_ONCE(twsk_net(tw)->ipv4.sysctl_tcp_rfc1337)) { 202 kill: 203 inet_twsk_deschedule_put(tw); 204 return TCP_TW_SUCCESS; 205 } 206 } else { 207 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 208 } 209 210 if (tmp_opt.saw_tstamp) { 211 WRITE_ONCE(tcptw->tw_ts_recent, 212 tmp_opt.rcv_tsval); 213 WRITE_ONCE(tcptw->tw_ts_recent_stamp, 214 ktime_get_seconds()); 215 } 216 217 inet_twsk_put(tw); 218 return TCP_TW_SUCCESS; 219 } 220 221 /* Out of window segment. 222 223 All the segments are ACKed immediately. 224 225 The only exception is new SYN. We accept it, if it is 226 not old duplicate and we are not in danger to be killed 227 by delayed old duplicates. RFC check is that it has 228 newer sequence number works at rates <40Mbit/sec. 229 However, if paws works, it is reliable AND even more, 230 we even may relax silly seq space cutoff. 231 232 RED-PEN: we violate main RFC requirement, if this SYN will appear 233 old duplicate (i.e. we receive RST in reply to SYN-ACK), 234 we must return socket to time-wait state. It is not good, 235 but not fatal yet. 236 */ 237 238 if (th->syn && !th->rst && !th->ack && !paws_reject && 239 (after(TCP_SKB_CB(skb)->seq, rcv_nxt) || 240 (tmp_opt.saw_tstamp && 241 (s32)(READ_ONCE(tcptw->tw_ts_recent) - tmp_opt.rcv_tsval) < 0))) { 242 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 243 if (isn == 0) 244 isn++; 245 *tw_isn = isn; 246 return TCP_TW_SYN; 247 } 248 249 if (paws_reject) { 250 *drop_reason = SKB_DROP_REASON_TCP_RFC7323_TW_PAWS; 251 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWS_TW_REJECTED); 252 } 253 254 if (!th->rst) { 255 /* In this case we must reset the TIMEWAIT timer. 256 * 257 * If it is ACKless SYN it may be both old duplicate 258 * and new good SYN with random sequence number <rcv_nxt. 259 * Do not reschedule in the last case. 260 */ 261 if (paws_reject || th->ack) 262 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 263 264 return tcp_timewait_check_oow_rate_limit( 265 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT); 266 } 267 inet_twsk_put(tw); 268 return TCP_TW_SUCCESS; 269 } 270 EXPORT_IPV6_MOD(tcp_timewait_state_process); 271 272 static void tcp_time_wait_init(struct sock *sk, struct tcp_timewait_sock *tcptw) 273 { 274 #ifdef CONFIG_TCP_MD5SIG 275 const struct tcp_sock *tp = tcp_sk(sk); 276 struct tcp_md5sig_key *key; 277 278 /* 279 * The timewait bucket does not have the key DB from the 280 * sock structure. We just make a quick copy of the 281 * md5 key being used (if indeed we are using one) 282 * so the timewait ack generating code has the key. 283 */ 284 tcptw->tw_md5_key = NULL; 285 if (!static_branch_unlikely(&tcp_md5_needed.key)) 286 return; 287 288 key = tp->af_specific->md5_lookup(sk, sk); 289 if (key) { 290 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC); 291 if (!tcptw->tw_md5_key) 292 return; 293 if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key)) 294 goto out_free; 295 tcp_md5_add_sigpool(); 296 } 297 return; 298 out_free: 299 WARN_ON_ONCE(1); 300 kfree(tcptw->tw_md5_key); 301 tcptw->tw_md5_key = NULL; 302 #endif 303 } 304 305 /* 306 * Move a socket to time-wait or dead fin-wait-2 state. 307 */ 308 void tcp_time_wait(struct sock *sk, int state, int timeo) 309 { 310 const struct inet_connection_sock *icsk = inet_csk(sk); 311 struct tcp_sock *tp = tcp_sk(sk); 312 struct net *net = sock_net(sk); 313 struct inet_timewait_sock *tw; 314 315 tw = inet_twsk_alloc(sk, &net->ipv4.tcp_death_row, state); 316 317 if (tw) { 318 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 319 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 320 321 tw->tw_transparent = inet_test_bit(TRANSPARENT, sk); 322 tw->tw_mark = sk->sk_mark; 323 tw->tw_priority = READ_ONCE(sk->sk_priority); 324 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 325 /* refreshed when we enter true TIME-WAIT state */ 326 tw->tw_entry_stamp = tcp_time_stamp_ms(tp); 327 tcptw->tw_rcv_nxt = tp->rcv_nxt; 328 tcptw->tw_snd_nxt = tp->snd_nxt; 329 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 330 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 331 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 332 tcptw->tw_ts_offset = tp->tsoffset; 333 tw->tw_usec_ts = tp->tcp_usec_ts; 334 tcptw->tw_last_oow_ack_time = 0; 335 tcptw->tw_tx_delay = tp->tcp_tx_delay; 336 tw->tw_txhash = sk->sk_txhash; 337 tw->tw_tx_queue_mapping = sk->sk_tx_queue_mapping; 338 #ifdef CONFIG_SOCK_RX_QUEUE_MAPPING 339 tw->tw_rx_queue_mapping = sk->sk_rx_queue_mapping; 340 #endif 341 #if IS_ENABLED(CONFIG_IPV6) 342 if (tw->tw_family == PF_INET6) { 343 struct ipv6_pinfo *np = inet6_sk(sk); 344 345 tw->tw_v6_daddr = sk->sk_v6_daddr; 346 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr; 347 tw->tw_tclass = np->tclass; 348 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK); 349 tw->tw_ipv6only = sk->sk_ipv6only; 350 } 351 #endif 352 353 tcp_time_wait_init(sk, tcptw); 354 tcp_ao_time_wait(tcptw, tp); 355 356 /* Get the TIME_WAIT timeout firing. */ 357 if (timeo < rto) 358 timeo = rto; 359 360 if (state == TCP_TIME_WAIT) 361 timeo = TCP_TIMEWAIT_LEN; 362 363 /* Linkage updates. 364 * Note that access to tw after this point is illegal. 365 */ 366 inet_twsk_hashdance_schedule(tw, sk, net->ipv4.tcp_death_row.hashinfo, timeo); 367 } else { 368 /* Sorry, if we're out of memory, just CLOSE this 369 * socket up. We've got bigger problems than 370 * non-graceful socket closings. 371 */ 372 NET_INC_STATS(net, LINUX_MIB_TCPTIMEWAITOVERFLOW); 373 } 374 375 tcp_update_metrics(sk); 376 tcp_done(sk); 377 } 378 EXPORT_SYMBOL(tcp_time_wait); 379 380 void tcp_twsk_destructor(struct sock *sk) 381 { 382 #ifdef CONFIG_TCP_MD5SIG 383 if (static_branch_unlikely(&tcp_md5_needed.key)) { 384 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 385 386 if (twsk->tw_md5_key) { 387 kfree(twsk->tw_md5_key); 388 static_branch_slow_dec_deferred(&tcp_md5_needed); 389 tcp_md5_release_sigpool(); 390 } 391 } 392 #endif 393 tcp_ao_destroy_sock(sk, true); 394 } 395 396 void tcp_twsk_purge(struct list_head *net_exit_list) 397 { 398 bool purged_once = false; 399 struct net *net; 400 401 list_for_each_entry(net, net_exit_list, exit_list) { 402 if (net->ipv4.tcp_death_row.hashinfo->pernet) { 403 /* Even if tw_refcount == 1, we must clean up kernel reqsk */ 404 inet_twsk_purge(net->ipv4.tcp_death_row.hashinfo); 405 } else if (!purged_once) { 406 inet_twsk_purge(&tcp_hashinfo); 407 purged_once = true; 408 } 409 } 410 } 411 412 /* Warning : This function is called without sk_listener being locked. 413 * Be sure to read socket fields once, as their value could change under us. 414 */ 415 void tcp_openreq_init_rwin(struct request_sock *req, 416 const struct sock *sk_listener, 417 const struct dst_entry *dst) 418 { 419 struct inet_request_sock *ireq = inet_rsk(req); 420 const struct tcp_sock *tp = tcp_sk(sk_listener); 421 int full_space = tcp_full_space(sk_listener); 422 u32 window_clamp; 423 __u8 rcv_wscale; 424 u32 rcv_wnd; 425 int mss; 426 427 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 428 window_clamp = READ_ONCE(tp->window_clamp); 429 /* Set this up on the first call only */ 430 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW); 431 432 /* limit the window selection if the user enforce a smaller rx buffer */ 433 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK && 434 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0)) 435 req->rsk_window_clamp = full_space; 436 437 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req); 438 if (rcv_wnd == 0) 439 rcv_wnd = dst_metric(dst, RTAX_INITRWND); 440 else if (full_space < rcv_wnd * mss) 441 full_space = rcv_wnd * mss; 442 443 /* tcp_full_space because it is guaranteed to be the first packet */ 444 tcp_select_initial_window(sk_listener, full_space, 445 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 446 &req->rsk_rcv_wnd, 447 &req->rsk_window_clamp, 448 ireq->wscale_ok, 449 &rcv_wscale, 450 rcv_wnd); 451 ireq->rcv_wscale = rcv_wscale; 452 } 453 454 static void tcp_ecn_openreq_child(struct tcp_sock *tp, 455 const struct request_sock *req) 456 { 457 tcp_ecn_mode_set(tp, inet_rsk(req)->ecn_ok ? 458 TCP_ECN_MODE_RFC3168 : 459 TCP_ECN_DISABLED); 460 } 461 462 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst) 463 { 464 struct inet_connection_sock *icsk = inet_csk(sk); 465 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 466 bool ca_got_dst = false; 467 468 if (ca_key != TCP_CA_UNSPEC) { 469 const struct tcp_congestion_ops *ca; 470 471 rcu_read_lock(); 472 ca = tcp_ca_find_key(ca_key); 473 if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 474 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 475 icsk->icsk_ca_ops = ca; 476 ca_got_dst = true; 477 } 478 rcu_read_unlock(); 479 } 480 481 /* If no valid choice made yet, assign current system default ca. */ 482 if (!ca_got_dst && 483 (!icsk->icsk_ca_setsockopt || 484 !bpf_try_module_get(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner))) 485 tcp_assign_congestion_control(sk); 486 487 tcp_set_ca_state(sk, TCP_CA_Open); 488 } 489 EXPORT_IPV6_MOD_GPL(tcp_ca_openreq_child); 490 491 static void smc_check_reset_syn_req(const struct tcp_sock *oldtp, 492 struct request_sock *req, 493 struct tcp_sock *newtp) 494 { 495 #if IS_ENABLED(CONFIG_SMC) 496 struct inet_request_sock *ireq; 497 498 if (static_branch_unlikely(&tcp_have_smc)) { 499 ireq = inet_rsk(req); 500 if (oldtp->syn_smc && !ireq->smc_ok) 501 newtp->syn_smc = 0; 502 } 503 #endif 504 } 505 506 /* This is not only more efficient than what we used to do, it eliminates 507 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 508 * 509 * Actually, we could lots of memory writes here. tp of listening 510 * socket contains all necessary default parameters. 511 */ 512 struct sock *tcp_create_openreq_child(const struct sock *sk, 513 struct request_sock *req, 514 struct sk_buff *skb) 515 { 516 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 517 const struct inet_request_sock *ireq = inet_rsk(req); 518 struct tcp_request_sock *treq = tcp_rsk(req); 519 struct inet_connection_sock *newicsk; 520 const struct tcp_sock *oldtp; 521 struct tcp_sock *newtp; 522 u32 seq; 523 524 if (!newsk) 525 return NULL; 526 527 newicsk = inet_csk(newsk); 528 newtp = tcp_sk(newsk); 529 oldtp = tcp_sk(sk); 530 531 smc_check_reset_syn_req(oldtp, req, newtp); 532 533 /* Now setup tcp_sock */ 534 newtp->pred_flags = 0; 535 536 seq = treq->rcv_isn + 1; 537 newtp->rcv_wup = seq; 538 WRITE_ONCE(newtp->copied_seq, seq); 539 WRITE_ONCE(newtp->rcv_nxt, seq); 540 newtp->segs_in = 1; 541 542 seq = treq->snt_isn + 1; 543 newtp->snd_sml = newtp->snd_una = seq; 544 WRITE_ONCE(newtp->snd_nxt, seq); 545 newtp->snd_up = seq; 546 547 INIT_LIST_HEAD(&newtp->tsq_node); 548 INIT_LIST_HEAD(&newtp->tsorted_sent_queue); 549 550 tcp_init_wl(newtp, treq->rcv_isn); 551 552 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U); 553 newicsk->icsk_ack.lrcvtime = tcp_jiffies32; 554 555 newtp->lsndtime = tcp_jiffies32; 556 newsk->sk_txhash = READ_ONCE(treq->txhash); 557 newtp->total_retrans = req->num_retrans; 558 559 tcp_init_xmit_timers(newsk); 560 WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1); 561 562 if (sock_flag(newsk, SOCK_KEEPOPEN)) 563 tcp_reset_keepalive_timer(newsk, keepalive_time_when(newtp)); 564 565 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 566 newtp->rx_opt.sack_ok = ireq->sack_ok; 567 newtp->window_clamp = req->rsk_window_clamp; 568 newtp->rcv_ssthresh = req->rsk_rcv_wnd; 569 newtp->rcv_wnd = req->rsk_rcv_wnd; 570 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 571 if (newtp->rx_opt.wscale_ok) { 572 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 573 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 574 } else { 575 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 576 newtp->window_clamp = min(newtp->window_clamp, 65535U); 577 } 578 newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale; 579 newtp->max_window = newtp->snd_wnd; 580 581 if (newtp->rx_opt.tstamp_ok) { 582 newtp->tcp_usec_ts = treq->req_usec_ts; 583 newtp->rx_opt.ts_recent = req->ts_recent; 584 newtp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 585 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 586 } else { 587 newtp->tcp_usec_ts = 0; 588 newtp->rx_opt.ts_recent_stamp = 0; 589 newtp->tcp_header_len = sizeof(struct tcphdr); 590 } 591 if (req->num_timeout) { 592 newtp->total_rto = req->num_timeout; 593 newtp->undo_marker = treq->snt_isn; 594 if (newtp->tcp_usec_ts) { 595 newtp->retrans_stamp = treq->snt_synack; 596 newtp->total_rto_time = (u32)(tcp_clock_us() - 597 newtp->retrans_stamp) / USEC_PER_MSEC; 598 } else { 599 newtp->retrans_stamp = div_u64(treq->snt_synack, 600 USEC_PER_SEC / TCP_TS_HZ); 601 newtp->total_rto_time = tcp_clock_ms() - 602 newtp->retrans_stamp; 603 } 604 newtp->total_rto_recoveries = 1; 605 } 606 newtp->tsoffset = treq->ts_off; 607 #ifdef CONFIG_TCP_MD5SIG 608 newtp->md5sig_info = NULL; /*XXX*/ 609 #endif 610 #ifdef CONFIG_TCP_AO 611 newtp->ao_info = NULL; 612 613 if (tcp_rsk_used_ao(req)) { 614 struct tcp_ao_key *ao_key; 615 616 ao_key = treq->af_specific->ao_lookup(sk, req, tcp_rsk(req)->ao_keyid, -1); 617 if (ao_key) 618 newtp->tcp_header_len += tcp_ao_len_aligned(ao_key); 619 } 620 #endif 621 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 622 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 623 newtp->rx_opt.mss_clamp = req->mss; 624 tcp_ecn_openreq_child(newtp, req); 625 newtp->fastopen_req = NULL; 626 RCU_INIT_POINTER(newtp->fastopen_rsk, NULL); 627 628 newtp->bpf_chg_cc_inprogress = 0; 629 tcp_bpf_clone(sk, newsk); 630 631 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS); 632 633 xa_init_flags(&newsk->sk_user_frags, XA_FLAGS_ALLOC1); 634 635 return newsk; 636 } 637 EXPORT_SYMBOL(tcp_create_openreq_child); 638 639 /* 640 * Process an incoming packet for SYN_RECV sockets represented as a 641 * request_sock. Normally sk is the listener socket but for TFO it 642 * points to the child socket. 643 * 644 * XXX (TFO) - The current impl contains a special check for ack 645 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better? 646 * 647 * We don't need to initialize tmp_opt.sack_ok as we don't use the results 648 * 649 * Note: If @fastopen is true, this can be called from process context. 650 * Otherwise, this is from BH context. 651 */ 652 653 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 654 struct request_sock *req, 655 bool fastopen, bool *req_stolen, 656 enum skb_drop_reason *drop_reason) 657 { 658 struct tcp_options_received tmp_opt; 659 struct sock *child; 660 const struct tcphdr *th = tcp_hdr(skb); 661 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 662 bool tsecr_reject = false; 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 = req->ts_recent; 672 if (tmp_opt.rcv_tsecr) { 673 if (inet_rsk(req)->tstamp_ok && !fastopen) 674 tsecr_reject = !between(tmp_opt.rcv_tsecr, 675 tcp_rsk(req)->snt_tsval_first, 676 READ_ONCE(tcp_rsk(req)->snt_tsval_last)); 677 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off; 678 } 679 /* We do not store true stamp, but it is not required, 680 * it can be estimated (approximately) 681 * from another data. 682 */ 683 tmp_opt.ts_recent_stamp = ktime_get_seconds() - reqsk_timeout(req, TCP_RTO_MAX) / HZ; 684 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 685 } 686 } 687 688 /* Check for pure retransmitted SYN. */ 689 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 690 flg == TCP_FLAG_SYN && 691 !paws_reject) { 692 /* 693 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 694 * this case on figure 6 and figure 8, but formal 695 * protocol description says NOTHING. 696 * To be more exact, it says that we should send ACK, 697 * because this segment (at least, if it has no data) 698 * is out of window. 699 * 700 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 701 * describe SYN-RECV state. All the description 702 * is wrong, we cannot believe to it and should 703 * rely only on common sense and implementation 704 * experience. 705 * 706 * Enforce "SYN-ACK" according to figure 8, figure 6 707 * of RFC793, fixed by RFC1122. 708 * 709 * Note that even if there is new data in the SYN packet 710 * they will be thrown away too. 711 * 712 * Reset timer after retransmitting SYNACK, similar to 713 * the idea of fast retransmit in recovery. 714 */ 715 if (!tcp_oow_rate_limited(sock_net(sk), skb, 716 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 717 &tcp_rsk(req)->last_oow_ack_time) && 718 719 !tcp_rtx_synack(sk, req)) { 720 unsigned long expires = jiffies; 721 722 expires += reqsk_timeout(req, TCP_RTO_MAX); 723 if (!fastopen) 724 mod_timer_pending(&req->rsk_timer, expires); 725 else 726 req->rsk_timer.expires = expires; 727 } 728 return NULL; 729 } 730 731 /* Further reproduces section "SEGMENT ARRIVES" 732 for state SYN-RECEIVED of RFC793. 733 It is broken, however, it does not work only 734 when SYNs are crossed. 735 736 You would think that SYN crossing is impossible here, since 737 we should have a SYN_SENT socket (from connect()) on our end, 738 but this is not true if the crossed SYNs were sent to both 739 ends by a malicious third party. We must defend against this, 740 and to do that we first verify the ACK (as per RFC793, page 741 36) and reset if it is invalid. Is this a true full defense? 742 To convince ourselves, let us consider a way in which the ACK 743 test can still pass in this 'malicious crossed SYNs' case. 744 Malicious sender sends identical SYNs (and thus identical sequence 745 numbers) to both A and B: 746 747 A: gets SYN, seq=7 748 B: gets SYN, seq=7 749 750 By our good fortune, both A and B select the same initial 751 send sequence number of seven :-) 752 753 A: sends SYN|ACK, seq=7, ack_seq=8 754 B: sends SYN|ACK, seq=7, ack_seq=8 755 756 So we are now A eating this SYN|ACK, ACK test passes. So 757 does sequence test, SYN is truncated, and thus we consider 758 it a bare ACK. 759 760 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 761 bare ACK. Otherwise, we create an established connection. Both 762 ends (listening sockets) accept the new incoming connection and try 763 to talk to each other. 8-) 764 765 Note: This case is both harmless, and rare. Possibility is about the 766 same as us discovering intelligent life on another plant tomorrow. 767 768 But generally, we should (RFC lies!) to accept ACK 769 from SYNACK both here and in tcp_rcv_state_process(). 770 tcp_rcv_state_process() does not, hence, we do not too. 771 772 Note that the case is absolutely generic: 773 we cannot optimize anything here without 774 violating protocol. All the checks must be made 775 before attempt to create socket. 776 */ 777 778 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 779 * and the incoming segment acknowledges something not yet 780 * sent (the segment carries an unacceptable ACK) ... 781 * a reset is sent." 782 * 783 * Invalid ACK: reset will be sent by listening socket. 784 * Note that the ACK validity check for a Fast Open socket is done 785 * elsewhere and is checked directly against the child socket rather 786 * than req because user data may have been sent out. 787 */ 788 if ((flg & TCP_FLAG_ACK) && !fastopen && 789 (TCP_SKB_CB(skb)->ack_seq != 790 tcp_rsk(req)->snt_isn + 1)) 791 return sk; 792 793 /* RFC793: "first check sequence number". */ 794 795 if (paws_reject || tsecr_reject || 796 !tcp_in_window(TCP_SKB_CB(skb)->seq, 797 TCP_SKB_CB(skb)->end_seq, 798 tcp_rsk(req)->rcv_nxt, 799 tcp_rsk(req)->rcv_nxt + 800 tcp_synack_window(req))) { 801 /* Out of window: send ACK and drop. */ 802 if (!(flg & TCP_FLAG_RST) && 803 !tcp_oow_rate_limited(sock_net(sk), skb, 804 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 805 &tcp_rsk(req)->last_oow_ack_time)) 806 req->rsk_ops->send_ack(sk, skb, req); 807 if (paws_reject) { 808 SKB_DR_SET(*drop_reason, TCP_RFC7323_PAWS); 809 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 810 } else if (tsecr_reject) { 811 SKB_DR_SET(*drop_reason, TCP_RFC7323_TSECR); 812 NET_INC_STATS(sock_net(sk), LINUX_MIB_TSECRREJECTED); 813 } else { 814 SKB_DR_SET(*drop_reason, TCP_OVERWINDOW); 815 } 816 return NULL; 817 } 818 819 /* In sequence, PAWS is OK. */ 820 821 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 822 /* Truncate SYN, it is out of window starting 823 at tcp_rsk(req)->rcv_isn + 1. */ 824 flg &= ~TCP_FLAG_SYN; 825 } 826 827 /* RFC793: "second check the RST bit" and 828 * "fourth, check the SYN bit" 829 */ 830 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 831 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 832 goto embryonic_reset; 833 } 834 835 /* ACK sequence verified above, just make sure ACK is 836 * set. If ACK not set, just silently drop the packet. 837 * 838 * XXX (TFO) - if we ever allow "data after SYN", the 839 * following check needs to be removed. 840 */ 841 if (!(flg & TCP_FLAG_ACK)) 842 return NULL; 843 844 /* For Fast Open no more processing is needed (sk is the 845 * child socket). 846 */ 847 if (fastopen) 848 return sk; 849 850 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 851 if (req->num_timeout < READ_ONCE(inet_csk(sk)->icsk_accept_queue.rskq_defer_accept) && 852 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 853 inet_rsk(req)->acked = 1; 854 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 855 return NULL; 856 } 857 858 /* OK, ACK is valid, create big socket and 859 * feed this segment to it. It will repeat all 860 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 861 * ESTABLISHED STATE. If it will be dropped after 862 * socket is created, wait for troubles. 863 */ 864 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL, 865 req, &own_req); 866 if (!child) 867 goto listen_overflow; 868 869 if (own_req && tmp_opt.saw_tstamp && 870 !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt)) 871 tcp_sk(child)->rx_opt.ts_recent = tmp_opt.rcv_tsval; 872 873 if (own_req && rsk_drop_req(req)) { 874 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req); 875 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, req); 876 return child; 877 } 878 879 sock_rps_save_rxhash(child, skb); 880 tcp_synack_rtt_meas(child, req); 881 *req_stolen = !own_req; 882 return inet_csk_complete_hashdance(sk, child, req, own_req); 883 884 listen_overflow: 885 SKB_DR_SET(*drop_reason, TCP_LISTEN_OVERFLOW); 886 if (sk != req->rsk_listener) 887 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE); 888 889 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow)) { 890 inet_rsk(req)->acked = 1; 891 return NULL; 892 } 893 894 embryonic_reset: 895 if (!(flg & TCP_FLAG_RST)) { 896 /* Received a bad SYN pkt - for TFO We try not to reset 897 * the local connection unless it's really necessary to 898 * avoid becoming vulnerable to outside attack aiming at 899 * resetting legit local connections. 900 */ 901 req->rsk_ops->send_reset(sk, skb, SK_RST_REASON_INVALID_SYN); 902 } else if (fastopen) { /* received a valid RST pkt */ 903 reqsk_fastopen_remove(sk, req, true); 904 tcp_reset(sk, skb); 905 } 906 if (!fastopen) { 907 bool unlinked = inet_csk_reqsk_queue_drop(sk, req); 908 909 if (unlinked) 910 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 911 *req_stolen = !unlinked; 912 } 913 return NULL; 914 } 915 EXPORT_IPV6_MOD(tcp_check_req); 916 917 /* 918 * Queue segment on the new socket if the new socket is active, 919 * otherwise we just shortcircuit this and continue with 920 * the new socket. 921 * 922 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV 923 * when entering. But other states are possible due to a race condition 924 * where after __inet_lookup_established() fails but before the listener 925 * locked is obtained, other packets cause the same connection to 926 * be created. 927 */ 928 929 enum skb_drop_reason tcp_child_process(struct sock *parent, struct sock *child, 930 struct sk_buff *skb) 931 __releases(&((child)->sk_lock.slock)) 932 { 933 enum skb_drop_reason reason = SKB_NOT_DROPPED_YET; 934 int state = child->sk_state; 935 936 /* record sk_napi_id and sk_rx_queue_mapping of child. */ 937 sk_mark_napi_id_set(child, skb); 938 939 tcp_segs_in(tcp_sk(child), skb); 940 if (!sock_owned_by_user(child)) { 941 reason = tcp_rcv_state_process(child, skb); 942 /* Wakeup parent, send SIGIO */ 943 if (state == TCP_SYN_RECV && child->sk_state != state) 944 parent->sk_data_ready(parent); 945 } else { 946 /* Alas, it is possible again, because we do lookup 947 * in main socket hash table and lock on listening 948 * socket does not protect us more. 949 */ 950 __sk_add_backlog(child, skb); 951 } 952 953 bh_unlock_sock(child); 954 sock_put(child); 955 return reason; 956 } 957 EXPORT_IPV6_MOD(tcp_child_process); 958