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 /* 55 * * Main purpose of TIME-WAIT state is to close connection gracefully, 56 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN 57 * (and, probably, tail of data) and one or more our ACKs are lost. 58 * * What is TIME-WAIT timeout? It is associated with maximal packet 59 * lifetime in the internet, which results in wrong conclusion, that 60 * it is set to catch "old duplicate segments" wandering out of their path. 61 * It is not quite correct. This timeout is calculated so that it exceeds 62 * maximal retransmission timeout enough to allow to lose one (or more) 63 * segments sent by peer and our ACKs. This time may be calculated from RTO. 64 * * When TIME-WAIT socket receives RST, it means that another end 65 * finally closed and we are allowed to kill TIME-WAIT too. 66 * * Second purpose of TIME-WAIT is catching old duplicate segments. 67 * Well, certainly it is pure paranoia, but if we load TIME-WAIT 68 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs. 69 * * If we invented some more clever way to catch duplicates 70 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs. 71 * 72 * The algorithm below is based on FORMAL INTERPRETATION of RFCs. 73 * When you compare it to RFCs, please, read section SEGMENT ARRIVES 74 * from the very beginning. 75 * 76 * NOTE. With recycling (and later with fin-wait-2) TW bucket 77 * is _not_ stateless. It means, that strictly speaking we must 78 * spinlock it. I do not want! Well, probability of misbehaviour 79 * is ridiculously low and, seems, we could use some mb() tricks 80 * to avoid misread sequence numbers, states etc. --ANK 81 * 82 * We don't need to initialize tmp_out.sack_ok as we don't use the results 83 */ 84 enum tcp_tw_status 85 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb, 86 const struct tcphdr *th) 87 { 88 struct tcp_options_received tmp_opt; 89 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 90 bool paws_reject = false; 91 92 tmp_opt.saw_tstamp = 0; 93 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) { 94 tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL); 95 96 if (tmp_opt.saw_tstamp) { 97 if (tmp_opt.rcv_tsecr) 98 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset; 99 tmp_opt.ts_recent = tcptw->tw_ts_recent; 100 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp; 101 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 102 } 103 } 104 105 if (tw->tw_substate == TCP_FIN_WAIT2) { 106 /* Just repeat all the checks of tcp_rcv_state_process() */ 107 108 /* Out of window, send ACK */ 109 if (paws_reject || 110 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 111 tcptw->tw_rcv_nxt, 112 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd)) 113 return tcp_timewait_check_oow_rate_limit( 114 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2); 115 116 if (th->rst) 117 goto kill; 118 119 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt)) 120 return TCP_TW_RST; 121 122 /* Dup ACK? */ 123 if (!th->ack || 124 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) || 125 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) { 126 inet_twsk_put(tw); 127 return TCP_TW_SUCCESS; 128 } 129 130 /* New data or FIN. If new data arrive after half-duplex close, 131 * reset. 132 */ 133 if (!th->fin || 134 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) 135 return TCP_TW_RST; 136 137 /* FIN arrived, enter true time-wait state. */ 138 tw->tw_substate = TCP_TIME_WAIT; 139 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq; 140 if (tmp_opt.saw_tstamp) { 141 tcptw->tw_ts_recent_stamp = ktime_get_seconds(); 142 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 143 } 144 145 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 146 return TCP_TW_ACK; 147 } 148 149 /* 150 * Now real TIME-WAIT state. 151 * 152 * RFC 1122: 153 * "When a connection is [...] on TIME-WAIT state [...] 154 * [a TCP] MAY accept a new SYN from the remote TCP to 155 * reopen the connection directly, if it: 156 * 157 * (1) assigns its initial sequence number for the new 158 * connection to be larger than the largest sequence 159 * number it used on the previous connection incarnation, 160 * and 161 * 162 * (2) returns to TIME-WAIT state if the SYN turns out 163 * to be an old duplicate". 164 */ 165 166 if (!paws_reject && 167 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt && 168 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 169 /* In window segment, it may be only reset or bare ack. */ 170 171 if (th->rst) { 172 /* This is TIME_WAIT assassination, in two flavors. 173 * Oh well... nobody has a sufficient solution to this 174 * protocol bug yet. 175 */ 176 if (!READ_ONCE(twsk_net(tw)->ipv4.sysctl_tcp_rfc1337)) { 177 kill: 178 inet_twsk_deschedule_put(tw); 179 return TCP_TW_SUCCESS; 180 } 181 } else { 182 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 183 } 184 185 if (tmp_opt.saw_tstamp) { 186 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 187 tcptw->tw_ts_recent_stamp = ktime_get_seconds(); 188 } 189 190 inet_twsk_put(tw); 191 return TCP_TW_SUCCESS; 192 } 193 194 /* Out of window segment. 195 196 All the segments are ACKed immediately. 197 198 The only exception is new SYN. We accept it, if it is 199 not old duplicate and we are not in danger to be killed 200 by delayed old duplicates. RFC check is that it has 201 newer sequence number works at rates <40Mbit/sec. 202 However, if paws works, it is reliable AND even more, 203 we even may relax silly seq space cutoff. 204 205 RED-PEN: we violate main RFC requirement, if this SYN will appear 206 old duplicate (i.e. we receive RST in reply to SYN-ACK), 207 we must return socket to time-wait state. It is not good, 208 but not fatal yet. 209 */ 210 211 if (th->syn && !th->rst && !th->ack && !paws_reject && 212 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) || 213 (tmp_opt.saw_tstamp && 214 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) { 215 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 216 if (isn == 0) 217 isn++; 218 TCP_SKB_CB(skb)->tcp_tw_isn = isn; 219 return TCP_TW_SYN; 220 } 221 222 if (paws_reject) 223 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED); 224 225 if (!th->rst) { 226 /* In this case we must reset the TIMEWAIT timer. 227 * 228 * If it is ACKless SYN it may be both old duplicate 229 * and new good SYN with random sequence number <rcv_nxt. 230 * Do not reschedule in the last case. 231 */ 232 if (paws_reject || th->ack) 233 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 234 235 return tcp_timewait_check_oow_rate_limit( 236 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT); 237 } 238 inet_twsk_put(tw); 239 return TCP_TW_SUCCESS; 240 } 241 EXPORT_SYMBOL(tcp_timewait_state_process); 242 243 /* 244 * Move a socket to time-wait or dead fin-wait-2 state. 245 */ 246 void tcp_time_wait(struct sock *sk, int state, int timeo) 247 { 248 const struct inet_connection_sock *icsk = inet_csk(sk); 249 const struct tcp_sock *tp = tcp_sk(sk); 250 struct net *net = sock_net(sk); 251 struct inet_timewait_sock *tw; 252 253 tw = inet_twsk_alloc(sk, &net->ipv4.tcp_death_row, state); 254 255 if (tw) { 256 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 257 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 258 struct inet_sock *inet = inet_sk(sk); 259 260 tw->tw_transparent = inet->transparent; 261 tw->tw_mark = sk->sk_mark; 262 tw->tw_priority = sk->sk_priority; 263 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 264 tcptw->tw_rcv_nxt = tp->rcv_nxt; 265 tcptw->tw_snd_nxt = tp->snd_nxt; 266 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 267 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 268 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 269 tcptw->tw_ts_offset = tp->tsoffset; 270 tcptw->tw_last_oow_ack_time = 0; 271 tcptw->tw_tx_delay = tp->tcp_tx_delay; 272 #if IS_ENABLED(CONFIG_IPV6) 273 if (tw->tw_family == PF_INET6) { 274 struct ipv6_pinfo *np = inet6_sk(sk); 275 276 tw->tw_v6_daddr = sk->sk_v6_daddr; 277 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr; 278 tw->tw_tclass = np->tclass; 279 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK); 280 tw->tw_txhash = sk->sk_txhash; 281 tw->tw_ipv6only = sk->sk_ipv6only; 282 } 283 #endif 284 285 #ifdef CONFIG_TCP_MD5SIG 286 /* 287 * The timewait bucket does not have the key DB from the 288 * sock structure. We just make a quick copy of the 289 * md5 key being used (if indeed we are using one) 290 * so the timewait ack generating code has the key. 291 */ 292 do { 293 tcptw->tw_md5_key = NULL; 294 if (static_branch_unlikely(&tcp_md5_needed)) { 295 struct tcp_md5sig_key *key; 296 297 key = tp->af_specific->md5_lookup(sk, sk); 298 if (key) { 299 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC); 300 BUG_ON(tcptw->tw_md5_key && !tcp_alloc_md5sig_pool()); 301 } 302 } 303 } while (0); 304 #endif 305 306 /* Get the TIME_WAIT timeout firing. */ 307 if (timeo < rto) 308 timeo = rto; 309 310 if (state == TCP_TIME_WAIT) 311 timeo = TCP_TIMEWAIT_LEN; 312 313 /* tw_timer is pinned, so we need to make sure BH are disabled 314 * in following section, otherwise timer handler could run before 315 * we complete the initialization. 316 */ 317 local_bh_disable(); 318 inet_twsk_schedule(tw, timeo); 319 /* Linkage updates. 320 * Note that access to tw after this point is illegal. 321 */ 322 inet_twsk_hashdance(tw, sk, net->ipv4.tcp_death_row.hashinfo); 323 local_bh_enable(); 324 } else { 325 /* Sorry, if we're out of memory, just CLOSE this 326 * socket up. We've got bigger problems than 327 * non-graceful socket closings. 328 */ 329 NET_INC_STATS(net, LINUX_MIB_TCPTIMEWAITOVERFLOW); 330 } 331 332 tcp_update_metrics(sk); 333 tcp_done(sk); 334 } 335 EXPORT_SYMBOL(tcp_time_wait); 336 337 void tcp_twsk_destructor(struct sock *sk) 338 { 339 #ifdef CONFIG_TCP_MD5SIG 340 if (static_branch_unlikely(&tcp_md5_needed)) { 341 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 342 343 if (twsk->tw_md5_key) 344 kfree_rcu(twsk->tw_md5_key, rcu); 345 } 346 #endif 347 } 348 EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 349 350 void tcp_twsk_purge(struct list_head *net_exit_list, int family) 351 { 352 bool purged_once = false; 353 struct net *net; 354 355 list_for_each_entry(net, net_exit_list, exit_list) { 356 /* The last refcount is decremented in tcp_sk_exit_batch() */ 357 if (refcount_read(&net->ipv4.tcp_death_row.tw_refcount) == 1) 358 continue; 359 360 if (net->ipv4.tcp_death_row.hashinfo->pernet) { 361 inet_twsk_purge(net->ipv4.tcp_death_row.hashinfo, family); 362 } else if (!purged_once) { 363 inet_twsk_purge(&tcp_hashinfo, family); 364 purged_once = true; 365 } 366 } 367 } 368 EXPORT_SYMBOL_GPL(tcp_twsk_purge); 369 370 /* Warning : This function is called without sk_listener being locked. 371 * Be sure to read socket fields once, as their value could change under us. 372 */ 373 void tcp_openreq_init_rwin(struct request_sock *req, 374 const struct sock *sk_listener, 375 const struct dst_entry *dst) 376 { 377 struct inet_request_sock *ireq = inet_rsk(req); 378 const struct tcp_sock *tp = tcp_sk(sk_listener); 379 int full_space = tcp_full_space(sk_listener); 380 u32 window_clamp; 381 __u8 rcv_wscale; 382 u32 rcv_wnd; 383 int mss; 384 385 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 386 window_clamp = READ_ONCE(tp->window_clamp); 387 /* Set this up on the first call only */ 388 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW); 389 390 /* limit the window selection if the user enforce a smaller rx buffer */ 391 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK && 392 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0)) 393 req->rsk_window_clamp = full_space; 394 395 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req); 396 if (rcv_wnd == 0) 397 rcv_wnd = dst_metric(dst, RTAX_INITRWND); 398 else if (full_space < rcv_wnd * mss) 399 full_space = rcv_wnd * mss; 400 401 /* tcp_full_space because it is guaranteed to be the first packet */ 402 tcp_select_initial_window(sk_listener, full_space, 403 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 404 &req->rsk_rcv_wnd, 405 &req->rsk_window_clamp, 406 ireq->wscale_ok, 407 &rcv_wscale, 408 rcv_wnd); 409 ireq->rcv_wscale = rcv_wscale; 410 } 411 EXPORT_SYMBOL(tcp_openreq_init_rwin); 412 413 static void tcp_ecn_openreq_child(struct tcp_sock *tp, 414 const struct request_sock *req) 415 { 416 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 417 } 418 419 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst) 420 { 421 struct inet_connection_sock *icsk = inet_csk(sk); 422 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 423 bool ca_got_dst = false; 424 425 if (ca_key != TCP_CA_UNSPEC) { 426 const struct tcp_congestion_ops *ca; 427 428 rcu_read_lock(); 429 ca = tcp_ca_find_key(ca_key); 430 if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 431 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 432 icsk->icsk_ca_ops = ca; 433 ca_got_dst = true; 434 } 435 rcu_read_unlock(); 436 } 437 438 /* If no valid choice made yet, assign current system default ca. */ 439 if (!ca_got_dst && 440 (!icsk->icsk_ca_setsockopt || 441 !bpf_try_module_get(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner))) 442 tcp_assign_congestion_control(sk); 443 444 tcp_set_ca_state(sk, TCP_CA_Open); 445 } 446 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child); 447 448 static void smc_check_reset_syn_req(struct tcp_sock *oldtp, 449 struct request_sock *req, 450 struct tcp_sock *newtp) 451 { 452 #if IS_ENABLED(CONFIG_SMC) 453 struct inet_request_sock *ireq; 454 455 if (static_branch_unlikely(&tcp_have_smc)) { 456 ireq = inet_rsk(req); 457 if (oldtp->syn_smc && !ireq->smc_ok) 458 newtp->syn_smc = 0; 459 } 460 #endif 461 } 462 463 /* This is not only more efficient than what we used to do, it eliminates 464 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 465 * 466 * Actually, we could lots of memory writes here. tp of listening 467 * socket contains all necessary default parameters. 468 */ 469 struct sock *tcp_create_openreq_child(const struct sock *sk, 470 struct request_sock *req, 471 struct sk_buff *skb) 472 { 473 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 474 const struct inet_request_sock *ireq = inet_rsk(req); 475 struct tcp_request_sock *treq = tcp_rsk(req); 476 struct inet_connection_sock *newicsk; 477 struct tcp_sock *oldtp, *newtp; 478 u32 seq; 479 480 if (!newsk) 481 return NULL; 482 483 newicsk = inet_csk(newsk); 484 newtp = tcp_sk(newsk); 485 oldtp = tcp_sk(sk); 486 487 smc_check_reset_syn_req(oldtp, req, newtp); 488 489 /* Now setup tcp_sock */ 490 newtp->pred_flags = 0; 491 492 seq = treq->rcv_isn + 1; 493 newtp->rcv_wup = seq; 494 WRITE_ONCE(newtp->copied_seq, seq); 495 WRITE_ONCE(newtp->rcv_nxt, seq); 496 newtp->segs_in = 1; 497 498 seq = treq->snt_isn + 1; 499 newtp->snd_sml = newtp->snd_una = seq; 500 WRITE_ONCE(newtp->snd_nxt, seq); 501 newtp->snd_up = seq; 502 503 INIT_LIST_HEAD(&newtp->tsq_node); 504 INIT_LIST_HEAD(&newtp->tsorted_sent_queue); 505 506 tcp_init_wl(newtp, treq->rcv_isn); 507 508 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U); 509 newicsk->icsk_ack.lrcvtime = tcp_jiffies32; 510 511 newtp->lsndtime = tcp_jiffies32; 512 newsk->sk_txhash = treq->txhash; 513 newtp->total_retrans = req->num_retrans; 514 515 tcp_init_xmit_timers(newsk); 516 WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1); 517 518 if (sock_flag(newsk, SOCK_KEEPOPEN)) 519 inet_csk_reset_keepalive_timer(newsk, 520 keepalive_time_when(newtp)); 521 522 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 523 newtp->rx_opt.sack_ok = ireq->sack_ok; 524 newtp->window_clamp = req->rsk_window_clamp; 525 newtp->rcv_ssthresh = req->rsk_rcv_wnd; 526 newtp->rcv_wnd = req->rsk_rcv_wnd; 527 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 528 if (newtp->rx_opt.wscale_ok) { 529 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 530 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 531 } else { 532 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 533 newtp->window_clamp = min(newtp->window_clamp, 65535U); 534 } 535 newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale; 536 newtp->max_window = newtp->snd_wnd; 537 538 if (newtp->rx_opt.tstamp_ok) { 539 newtp->rx_opt.ts_recent = req->ts_recent; 540 newtp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 541 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 542 } else { 543 newtp->rx_opt.ts_recent_stamp = 0; 544 newtp->tcp_header_len = sizeof(struct tcphdr); 545 } 546 if (req->num_timeout) { 547 newtp->undo_marker = treq->snt_isn; 548 newtp->retrans_stamp = div_u64(treq->snt_synack, 549 USEC_PER_SEC / TCP_TS_HZ); 550 } 551 newtp->tsoffset = treq->ts_off; 552 #ifdef CONFIG_TCP_MD5SIG 553 newtp->md5sig_info = NULL; /*XXX*/ 554 if (treq->af_specific->req_md5_lookup(sk, req_to_sk(req))) 555 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 556 #endif 557 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 558 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 559 newtp->rx_opt.mss_clamp = req->mss; 560 tcp_ecn_openreq_child(newtp, req); 561 newtp->fastopen_req = NULL; 562 RCU_INIT_POINTER(newtp->fastopen_rsk, NULL); 563 564 tcp_bpf_clone(sk, newsk); 565 566 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS); 567 568 return newsk; 569 } 570 EXPORT_SYMBOL(tcp_create_openreq_child); 571 572 /* 573 * Process an incoming packet for SYN_RECV sockets represented as a 574 * request_sock. Normally sk is the listener socket but for TFO it 575 * points to the child socket. 576 * 577 * XXX (TFO) - The current impl contains a special check for ack 578 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better? 579 * 580 * We don't need to initialize tmp_opt.sack_ok as we don't use the results 581 */ 582 583 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 584 struct request_sock *req, 585 bool fastopen, bool *req_stolen) 586 { 587 struct tcp_options_received tmp_opt; 588 struct sock *child; 589 const struct tcphdr *th = tcp_hdr(skb); 590 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 591 bool paws_reject = false; 592 bool own_req; 593 594 tmp_opt.saw_tstamp = 0; 595 if (th->doff > (sizeof(struct tcphdr)>>2)) { 596 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL); 597 598 if (tmp_opt.saw_tstamp) { 599 tmp_opt.ts_recent = req->ts_recent; 600 if (tmp_opt.rcv_tsecr) 601 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off; 602 /* We do not store true stamp, but it is not required, 603 * it can be estimated (approximately) 604 * from another data. 605 */ 606 tmp_opt.ts_recent_stamp = ktime_get_seconds() - reqsk_timeout(req, TCP_RTO_MAX) / HZ; 607 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 608 } 609 } 610 611 /* Check for pure retransmitted SYN. */ 612 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 613 flg == TCP_FLAG_SYN && 614 !paws_reject) { 615 /* 616 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 617 * this case on figure 6 and figure 8, but formal 618 * protocol description says NOTHING. 619 * To be more exact, it says that we should send ACK, 620 * because this segment (at least, if it has no data) 621 * is out of window. 622 * 623 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 624 * describe SYN-RECV state. All the description 625 * is wrong, we cannot believe to it and should 626 * rely only on common sense and implementation 627 * experience. 628 * 629 * Enforce "SYN-ACK" according to figure 8, figure 6 630 * of RFC793, fixed by RFC1122. 631 * 632 * Note that even if there is new data in the SYN packet 633 * they will be thrown away too. 634 * 635 * Reset timer after retransmitting SYNACK, similar to 636 * the idea of fast retransmit in recovery. 637 */ 638 if (!tcp_oow_rate_limited(sock_net(sk), skb, 639 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 640 &tcp_rsk(req)->last_oow_ack_time) && 641 642 !inet_rtx_syn_ack(sk, req)) { 643 unsigned long expires = jiffies; 644 645 expires += reqsk_timeout(req, TCP_RTO_MAX); 646 if (!fastopen) 647 mod_timer_pending(&req->rsk_timer, expires); 648 else 649 req->rsk_timer.expires = expires; 650 } 651 return NULL; 652 } 653 654 /* Further reproduces section "SEGMENT ARRIVES" 655 for state SYN-RECEIVED of RFC793. 656 It is broken, however, it does not work only 657 when SYNs are crossed. 658 659 You would think that SYN crossing is impossible here, since 660 we should have a SYN_SENT socket (from connect()) on our end, 661 but this is not true if the crossed SYNs were sent to both 662 ends by a malicious third party. We must defend against this, 663 and to do that we first verify the ACK (as per RFC793, page 664 36) and reset if it is invalid. Is this a true full defense? 665 To convince ourselves, let us consider a way in which the ACK 666 test can still pass in this 'malicious crossed SYNs' case. 667 Malicious sender sends identical SYNs (and thus identical sequence 668 numbers) to both A and B: 669 670 A: gets SYN, seq=7 671 B: gets SYN, seq=7 672 673 By our good fortune, both A and B select the same initial 674 send sequence number of seven :-) 675 676 A: sends SYN|ACK, seq=7, ack_seq=8 677 B: sends SYN|ACK, seq=7, ack_seq=8 678 679 So we are now A eating this SYN|ACK, ACK test passes. So 680 does sequence test, SYN is truncated, and thus we consider 681 it a bare ACK. 682 683 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 684 bare ACK. Otherwise, we create an established connection. Both 685 ends (listening sockets) accept the new incoming connection and try 686 to talk to each other. 8-) 687 688 Note: This case is both harmless, and rare. Possibility is about the 689 same as us discovering intelligent life on another plant tomorrow. 690 691 But generally, we should (RFC lies!) to accept ACK 692 from SYNACK both here and in tcp_rcv_state_process(). 693 tcp_rcv_state_process() does not, hence, we do not too. 694 695 Note that the case is absolutely generic: 696 we cannot optimize anything here without 697 violating protocol. All the checks must be made 698 before attempt to create socket. 699 */ 700 701 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 702 * and the incoming segment acknowledges something not yet 703 * sent (the segment carries an unacceptable ACK) ... 704 * a reset is sent." 705 * 706 * Invalid ACK: reset will be sent by listening socket. 707 * Note that the ACK validity check for a Fast Open socket is done 708 * elsewhere and is checked directly against the child socket rather 709 * than req because user data may have been sent out. 710 */ 711 if ((flg & TCP_FLAG_ACK) && !fastopen && 712 (TCP_SKB_CB(skb)->ack_seq != 713 tcp_rsk(req)->snt_isn + 1)) 714 return sk; 715 716 /* Also, it would be not so bad idea to check rcv_tsecr, which 717 * is essentially ACK extension and too early or too late values 718 * should cause reset in unsynchronized states. 719 */ 720 721 /* RFC793: "first check sequence number". */ 722 723 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 724 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) { 725 /* Out of window: send ACK and drop. */ 726 if (!(flg & TCP_FLAG_RST) && 727 !tcp_oow_rate_limited(sock_net(sk), skb, 728 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 729 &tcp_rsk(req)->last_oow_ack_time)) 730 req->rsk_ops->send_ack(sk, skb, req); 731 if (paws_reject) 732 __NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 733 return NULL; 734 } 735 736 /* In sequence, PAWS is OK. */ 737 738 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt)) 739 req->ts_recent = tmp_opt.rcv_tsval; 740 741 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 742 /* Truncate SYN, it is out of window starting 743 at tcp_rsk(req)->rcv_isn + 1. */ 744 flg &= ~TCP_FLAG_SYN; 745 } 746 747 /* RFC793: "second check the RST bit" and 748 * "fourth, check the SYN bit" 749 */ 750 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 751 __TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 752 goto embryonic_reset; 753 } 754 755 /* ACK sequence verified above, just make sure ACK is 756 * set. If ACK not set, just silently drop the packet. 757 * 758 * XXX (TFO) - if we ever allow "data after SYN", the 759 * following check needs to be removed. 760 */ 761 if (!(flg & TCP_FLAG_ACK)) 762 return NULL; 763 764 /* For Fast Open no more processing is needed (sk is the 765 * child socket). 766 */ 767 if (fastopen) 768 return sk; 769 770 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 771 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 772 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 773 inet_rsk(req)->acked = 1; 774 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 775 return NULL; 776 } 777 778 /* OK, ACK is valid, create big socket and 779 * feed this segment to it. It will repeat all 780 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 781 * ESTABLISHED STATE. If it will be dropped after 782 * socket is created, wait for troubles. 783 */ 784 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL, 785 req, &own_req); 786 if (!child) 787 goto listen_overflow; 788 789 if (own_req && rsk_drop_req(req)) { 790 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req); 791 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, req); 792 return child; 793 } 794 795 sock_rps_save_rxhash(child, skb); 796 tcp_synack_rtt_meas(child, req); 797 *req_stolen = !own_req; 798 return inet_csk_complete_hashdance(sk, child, req, own_req); 799 800 listen_overflow: 801 if (sk != req->rsk_listener) 802 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE); 803 804 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow)) { 805 inet_rsk(req)->acked = 1; 806 return NULL; 807 } 808 809 embryonic_reset: 810 if (!(flg & TCP_FLAG_RST)) { 811 /* Received a bad SYN pkt - for TFO We try not to reset 812 * the local connection unless it's really necessary to 813 * avoid becoming vulnerable to outside attack aiming at 814 * resetting legit local connections. 815 */ 816 req->rsk_ops->send_reset(sk, skb); 817 } else if (fastopen) { /* received a valid RST pkt */ 818 reqsk_fastopen_remove(sk, req, true); 819 tcp_reset(sk, skb); 820 } 821 if (!fastopen) { 822 bool unlinked = inet_csk_reqsk_queue_drop(sk, req); 823 824 if (unlinked) 825 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 826 *req_stolen = !unlinked; 827 } 828 return NULL; 829 } 830 EXPORT_SYMBOL(tcp_check_req); 831 832 /* 833 * Queue segment on the new socket if the new socket is active, 834 * otherwise we just shortcircuit this and continue with 835 * the new socket. 836 * 837 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV 838 * when entering. But other states are possible due to a race condition 839 * where after __inet_lookup_established() fails but before the listener 840 * locked is obtained, other packets cause the same connection to 841 * be created. 842 */ 843 844 int tcp_child_process(struct sock *parent, struct sock *child, 845 struct sk_buff *skb) 846 __releases(&((child)->sk_lock.slock)) 847 { 848 int ret = 0; 849 int state = child->sk_state; 850 851 /* record sk_napi_id and sk_rx_queue_mapping of child. */ 852 sk_mark_napi_id_set(child, skb); 853 854 tcp_segs_in(tcp_sk(child), skb); 855 if (!sock_owned_by_user(child)) { 856 ret = tcp_rcv_state_process(child, skb); 857 /* Wakeup parent, send SIGIO */ 858 if (state == TCP_SYN_RECV && child->sk_state != state) 859 parent->sk_data_ready(parent); 860 } else { 861 /* Alas, it is possible again, because we do lookup 862 * in main socket hash table and lock on listening 863 * socket does not protect us more. 864 */ 865 __sk_add_backlog(child, skb); 866 } 867 868 bh_unlock_sock(child); 869 sock_put(child); 870 return ret; 871 } 872 EXPORT_SYMBOL(tcp_child_process); 873