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 static void tcp_time_wait_init(struct sock *sk, struct tcp_timewait_sock *tcptw) 244 { 245 #ifdef CONFIG_TCP_MD5SIG 246 const struct tcp_sock *tp = tcp_sk(sk); 247 struct tcp_md5sig_key *key; 248 249 /* 250 * The timewait bucket does not have the key DB from the 251 * sock structure. We just make a quick copy of the 252 * md5 key being used (if indeed we are using one) 253 * so the timewait ack generating code has the key. 254 */ 255 tcptw->tw_md5_key = NULL; 256 if (!static_branch_unlikely(&tcp_md5_needed.key)) 257 return; 258 259 key = tp->af_specific->md5_lookup(sk, sk); 260 if (key) { 261 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC); 262 if (!tcptw->tw_md5_key) 263 return; 264 if (!tcp_alloc_md5sig_pool()) 265 goto out_free; 266 if (!static_key_fast_inc_not_disabled(&tcp_md5_needed.key.key)) 267 goto out_free; 268 } 269 return; 270 out_free: 271 WARN_ON_ONCE(1); 272 kfree(tcptw->tw_md5_key); 273 tcptw->tw_md5_key = NULL; 274 #endif 275 } 276 277 /* 278 * Move a socket to time-wait or dead fin-wait-2 state. 279 */ 280 void tcp_time_wait(struct sock *sk, int state, int timeo) 281 { 282 const struct inet_connection_sock *icsk = inet_csk(sk); 283 const struct tcp_sock *tp = tcp_sk(sk); 284 struct net *net = sock_net(sk); 285 struct inet_timewait_sock *tw; 286 287 tw = inet_twsk_alloc(sk, &net->ipv4.tcp_death_row, state); 288 289 if (tw) { 290 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 291 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 292 293 tw->tw_transparent = inet_test_bit(TRANSPARENT, sk); 294 tw->tw_mark = sk->sk_mark; 295 tw->tw_priority = sk->sk_priority; 296 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 297 tcptw->tw_rcv_nxt = tp->rcv_nxt; 298 tcptw->tw_snd_nxt = tp->snd_nxt; 299 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 300 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 301 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 302 tcptw->tw_ts_offset = tp->tsoffset; 303 tcptw->tw_last_oow_ack_time = 0; 304 tcptw->tw_tx_delay = tp->tcp_tx_delay; 305 tw->tw_txhash = sk->sk_txhash; 306 #if IS_ENABLED(CONFIG_IPV6) 307 if (tw->tw_family == PF_INET6) { 308 struct ipv6_pinfo *np = inet6_sk(sk); 309 310 tw->tw_v6_daddr = sk->sk_v6_daddr; 311 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr; 312 tw->tw_tclass = np->tclass; 313 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK); 314 tw->tw_ipv6only = sk->sk_ipv6only; 315 } 316 #endif 317 318 tcp_time_wait_init(sk, tcptw); 319 320 /* Get the TIME_WAIT timeout firing. */ 321 if (timeo < rto) 322 timeo = rto; 323 324 if (state == TCP_TIME_WAIT) 325 timeo = TCP_TIMEWAIT_LEN; 326 327 /* tw_timer is pinned, so we need to make sure BH are disabled 328 * in following section, otherwise timer handler could run before 329 * we complete the initialization. 330 */ 331 local_bh_disable(); 332 inet_twsk_schedule(tw, timeo); 333 /* Linkage updates. 334 * Note that access to tw after this point is illegal. 335 */ 336 inet_twsk_hashdance(tw, sk, net->ipv4.tcp_death_row.hashinfo); 337 local_bh_enable(); 338 } else { 339 /* Sorry, if we're out of memory, just CLOSE this 340 * socket up. We've got bigger problems than 341 * non-graceful socket closings. 342 */ 343 NET_INC_STATS(net, LINUX_MIB_TCPTIMEWAITOVERFLOW); 344 } 345 346 tcp_update_metrics(sk); 347 tcp_done(sk); 348 } 349 EXPORT_SYMBOL(tcp_time_wait); 350 351 void tcp_twsk_destructor(struct sock *sk) 352 { 353 #ifdef CONFIG_TCP_MD5SIG 354 if (static_branch_unlikely(&tcp_md5_needed.key)) { 355 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 356 357 if (twsk->tw_md5_key) { 358 kfree_rcu(twsk->tw_md5_key, rcu); 359 static_branch_slow_dec_deferred(&tcp_md5_needed); 360 } 361 } 362 #endif 363 } 364 EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 365 366 void tcp_twsk_purge(struct list_head *net_exit_list) 367 { 368 bool purged_once = false; 369 struct net *net; 370 371 list_for_each_entry(net, net_exit_list, exit_list) { 372 if (net->ipv4.tcp_death_row.hashinfo->pernet) { 373 /* Even if tw_refcount == 1, we must clean up kernel reqsk */ 374 inet_twsk_purge(net->ipv4.tcp_death_row.hashinfo); 375 } else if (!purged_once) { 376 inet_twsk_purge(&tcp_hashinfo); 377 purged_once = true; 378 } 379 } 380 } 381 382 /* Warning : This function is called without sk_listener being locked. 383 * Be sure to read socket fields once, as their value could change under us. 384 */ 385 void tcp_openreq_init_rwin(struct request_sock *req, 386 const struct sock *sk_listener, 387 const struct dst_entry *dst) 388 { 389 struct inet_request_sock *ireq = inet_rsk(req); 390 const struct tcp_sock *tp = tcp_sk(sk_listener); 391 int full_space = tcp_full_space(sk_listener); 392 u32 window_clamp; 393 __u8 rcv_wscale; 394 u32 rcv_wnd; 395 int mss; 396 397 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 398 window_clamp = READ_ONCE(tp->window_clamp); 399 /* Set this up on the first call only */ 400 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW); 401 402 /* limit the window selection if the user enforce a smaller rx buffer */ 403 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK && 404 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0)) 405 req->rsk_window_clamp = full_space; 406 407 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req); 408 if (rcv_wnd == 0) 409 rcv_wnd = dst_metric(dst, RTAX_INITRWND); 410 else if (full_space < rcv_wnd * mss) 411 full_space = rcv_wnd * mss; 412 413 /* tcp_full_space because it is guaranteed to be the first packet */ 414 tcp_select_initial_window(sk_listener, full_space, 415 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 416 &req->rsk_rcv_wnd, 417 &req->rsk_window_clamp, 418 ireq->wscale_ok, 419 &rcv_wscale, 420 rcv_wnd); 421 ireq->rcv_wscale = rcv_wscale; 422 } 423 EXPORT_SYMBOL(tcp_openreq_init_rwin); 424 425 static void tcp_ecn_openreq_child(struct tcp_sock *tp, 426 const struct request_sock *req) 427 { 428 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 429 } 430 431 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst) 432 { 433 struct inet_connection_sock *icsk = inet_csk(sk); 434 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 435 bool ca_got_dst = false; 436 437 if (ca_key != TCP_CA_UNSPEC) { 438 const struct tcp_congestion_ops *ca; 439 440 rcu_read_lock(); 441 ca = tcp_ca_find_key(ca_key); 442 if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 443 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 444 icsk->icsk_ca_ops = ca; 445 ca_got_dst = true; 446 } 447 rcu_read_unlock(); 448 } 449 450 /* If no valid choice made yet, assign current system default ca. */ 451 if (!ca_got_dst && 452 (!icsk->icsk_ca_setsockopt || 453 !bpf_try_module_get(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner))) 454 tcp_assign_congestion_control(sk); 455 456 tcp_set_ca_state(sk, TCP_CA_Open); 457 } 458 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child); 459 460 static void smc_check_reset_syn_req(const struct tcp_sock *oldtp, 461 struct request_sock *req, 462 struct tcp_sock *newtp) 463 { 464 #if IS_ENABLED(CONFIG_SMC) 465 struct inet_request_sock *ireq; 466 467 if (static_branch_unlikely(&tcp_have_smc)) { 468 ireq = inet_rsk(req); 469 if (oldtp->syn_smc && !ireq->smc_ok) 470 newtp->syn_smc = 0; 471 } 472 #endif 473 } 474 475 /* This is not only more efficient than what we used to do, it eliminates 476 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 477 * 478 * Actually, we could lots of memory writes here. tp of listening 479 * socket contains all necessary default parameters. 480 */ 481 struct sock *tcp_create_openreq_child(const struct sock *sk, 482 struct request_sock *req, 483 struct sk_buff *skb) 484 { 485 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 486 const struct inet_request_sock *ireq = inet_rsk(req); 487 struct tcp_request_sock *treq = tcp_rsk(req); 488 struct inet_connection_sock *newicsk; 489 const struct tcp_sock *oldtp; 490 struct tcp_sock *newtp; 491 u32 seq; 492 493 if (!newsk) 494 return NULL; 495 496 newicsk = inet_csk(newsk); 497 newtp = tcp_sk(newsk); 498 oldtp = tcp_sk(sk); 499 500 smc_check_reset_syn_req(oldtp, req, newtp); 501 502 /* Now setup tcp_sock */ 503 newtp->pred_flags = 0; 504 505 seq = treq->rcv_isn + 1; 506 newtp->rcv_wup = seq; 507 WRITE_ONCE(newtp->copied_seq, seq); 508 WRITE_ONCE(newtp->rcv_nxt, seq); 509 newtp->segs_in = 1; 510 511 seq = treq->snt_isn + 1; 512 newtp->snd_sml = newtp->snd_una = seq; 513 WRITE_ONCE(newtp->snd_nxt, seq); 514 newtp->snd_up = seq; 515 516 INIT_LIST_HEAD(&newtp->tsq_node); 517 INIT_LIST_HEAD(&newtp->tsorted_sent_queue); 518 519 tcp_init_wl(newtp, treq->rcv_isn); 520 521 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U); 522 newicsk->icsk_ack.lrcvtime = tcp_jiffies32; 523 524 newtp->lsndtime = tcp_jiffies32; 525 newsk->sk_txhash = READ_ONCE(treq->txhash); 526 newtp->total_retrans = req->num_retrans; 527 528 tcp_init_xmit_timers(newsk); 529 WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1); 530 531 if (sock_flag(newsk, SOCK_KEEPOPEN)) 532 inet_csk_reset_keepalive_timer(newsk, 533 keepalive_time_when(newtp)); 534 535 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 536 newtp->rx_opt.sack_ok = ireq->sack_ok; 537 newtp->window_clamp = req->rsk_window_clamp; 538 newtp->rcv_ssthresh = req->rsk_rcv_wnd; 539 newtp->rcv_wnd = req->rsk_rcv_wnd; 540 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 541 if (newtp->rx_opt.wscale_ok) { 542 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 543 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 544 } else { 545 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 546 newtp->window_clamp = min(newtp->window_clamp, 65535U); 547 } 548 newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale; 549 newtp->max_window = newtp->snd_wnd; 550 551 if (newtp->rx_opt.tstamp_ok) { 552 newtp->rx_opt.ts_recent = READ_ONCE(req->ts_recent); 553 newtp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 554 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 555 } else { 556 newtp->rx_opt.ts_recent_stamp = 0; 557 newtp->tcp_header_len = sizeof(struct tcphdr); 558 } 559 if (req->num_timeout) { 560 newtp->undo_marker = treq->snt_isn; 561 newtp->retrans_stamp = div_u64(treq->snt_synack, 562 USEC_PER_SEC / TCP_TS_HZ); 563 } 564 newtp->tsoffset = treq->ts_off; 565 #ifdef CONFIG_TCP_MD5SIG 566 newtp->md5sig_info = NULL; /*XXX*/ 567 #endif 568 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 569 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 570 newtp->rx_opt.mss_clamp = req->mss; 571 tcp_ecn_openreq_child(newtp, req); 572 newtp->fastopen_req = NULL; 573 RCU_INIT_POINTER(newtp->fastopen_rsk, NULL); 574 575 newtp->bpf_chg_cc_inprogress = 0; 576 tcp_bpf_clone(sk, newsk); 577 578 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS); 579 580 return newsk; 581 } 582 EXPORT_SYMBOL(tcp_create_openreq_child); 583 584 /* 585 * Process an incoming packet for SYN_RECV sockets represented as a 586 * request_sock. Normally sk is the listener socket but for TFO it 587 * points to the child socket. 588 * 589 * XXX (TFO) - The current impl contains a special check for ack 590 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better? 591 * 592 * We don't need to initialize tmp_opt.sack_ok as we don't use the results 593 * 594 * Note: If @fastopen is true, this can be called from process context. 595 * Otherwise, this is from BH context. 596 */ 597 598 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 599 struct request_sock *req, 600 bool fastopen, bool *req_stolen) 601 { 602 struct tcp_options_received tmp_opt; 603 struct sock *child; 604 const struct tcphdr *th = tcp_hdr(skb); 605 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 606 bool paws_reject = false; 607 bool own_req; 608 609 tmp_opt.saw_tstamp = 0; 610 if (th->doff > (sizeof(struct tcphdr)>>2)) { 611 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL); 612 613 if (tmp_opt.saw_tstamp) { 614 tmp_opt.ts_recent = READ_ONCE(req->ts_recent); 615 if (tmp_opt.rcv_tsecr) 616 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off; 617 /* We do not store true stamp, but it is not required, 618 * it can be estimated (approximately) 619 * from another data. 620 */ 621 tmp_opt.ts_recent_stamp = ktime_get_seconds() - reqsk_timeout(req, TCP_RTO_MAX) / HZ; 622 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 623 } 624 } 625 626 /* Check for pure retransmitted SYN. */ 627 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 628 flg == TCP_FLAG_SYN && 629 !paws_reject) { 630 /* 631 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 632 * this case on figure 6 and figure 8, but formal 633 * protocol description says NOTHING. 634 * To be more exact, it says that we should send ACK, 635 * because this segment (at least, if it has no data) 636 * is out of window. 637 * 638 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 639 * describe SYN-RECV state. All the description 640 * is wrong, we cannot believe to it and should 641 * rely only on common sense and implementation 642 * experience. 643 * 644 * Enforce "SYN-ACK" according to figure 8, figure 6 645 * of RFC793, fixed by RFC1122. 646 * 647 * Note that even if there is new data in the SYN packet 648 * they will be thrown away too. 649 * 650 * Reset timer after retransmitting SYNACK, similar to 651 * the idea of fast retransmit in recovery. 652 */ 653 if (!tcp_oow_rate_limited(sock_net(sk), skb, 654 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 655 &tcp_rsk(req)->last_oow_ack_time) && 656 657 !inet_rtx_syn_ack(sk, req)) { 658 unsigned long expires = jiffies; 659 660 expires += reqsk_timeout(req, TCP_RTO_MAX); 661 if (!fastopen) 662 mod_timer_pending(&req->rsk_timer, expires); 663 else 664 req->rsk_timer.expires = expires; 665 } 666 return NULL; 667 } 668 669 /* Further reproduces section "SEGMENT ARRIVES" 670 for state SYN-RECEIVED of RFC793. 671 It is broken, however, it does not work only 672 when SYNs are crossed. 673 674 You would think that SYN crossing is impossible here, since 675 we should have a SYN_SENT socket (from connect()) on our end, 676 but this is not true if the crossed SYNs were sent to both 677 ends by a malicious third party. We must defend against this, 678 and to do that we first verify the ACK (as per RFC793, page 679 36) and reset if it is invalid. Is this a true full defense? 680 To convince ourselves, let us consider a way in which the ACK 681 test can still pass in this 'malicious crossed SYNs' case. 682 Malicious sender sends identical SYNs (and thus identical sequence 683 numbers) to both A and B: 684 685 A: gets SYN, seq=7 686 B: gets SYN, seq=7 687 688 By our good fortune, both A and B select the same initial 689 send sequence number of seven :-) 690 691 A: sends SYN|ACK, seq=7, ack_seq=8 692 B: sends SYN|ACK, seq=7, ack_seq=8 693 694 So we are now A eating this SYN|ACK, ACK test passes. So 695 does sequence test, SYN is truncated, and thus we consider 696 it a bare ACK. 697 698 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 699 bare ACK. Otherwise, we create an established connection. Both 700 ends (listening sockets) accept the new incoming connection and try 701 to talk to each other. 8-) 702 703 Note: This case is both harmless, and rare. Possibility is about the 704 same as us discovering intelligent life on another plant tomorrow. 705 706 But generally, we should (RFC lies!) to accept ACK 707 from SYNACK both here and in tcp_rcv_state_process(). 708 tcp_rcv_state_process() does not, hence, we do not too. 709 710 Note that the case is absolutely generic: 711 we cannot optimize anything here without 712 violating protocol. All the checks must be made 713 before attempt to create socket. 714 */ 715 716 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 717 * and the incoming segment acknowledges something not yet 718 * sent (the segment carries an unacceptable ACK) ... 719 * a reset is sent." 720 * 721 * Invalid ACK: reset will be sent by listening socket. 722 * Note that the ACK validity check for a Fast Open socket is done 723 * elsewhere and is checked directly against the child socket rather 724 * than req because user data may have been sent out. 725 */ 726 if ((flg & TCP_FLAG_ACK) && !fastopen && 727 (TCP_SKB_CB(skb)->ack_seq != 728 tcp_rsk(req)->snt_isn + 1)) 729 return sk; 730 731 /* Also, it would be not so bad idea to check rcv_tsecr, which 732 * is essentially ACK extension and too early or too late values 733 * should cause reset in unsynchronized states. 734 */ 735 736 /* RFC793: "first check sequence number". */ 737 738 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 739 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) { 740 /* Out of window: send ACK and drop. */ 741 if (!(flg & TCP_FLAG_RST) && 742 !tcp_oow_rate_limited(sock_net(sk), skb, 743 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 744 &tcp_rsk(req)->last_oow_ack_time)) 745 req->rsk_ops->send_ack(sk, skb, req); 746 if (paws_reject) 747 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 748 return NULL; 749 } 750 751 /* In sequence, PAWS is OK. */ 752 753 /* TODO: We probably should defer ts_recent change once 754 * we take ownership of @req. 755 */ 756 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt)) 757 WRITE_ONCE(req->ts_recent, tmp_opt.rcv_tsval); 758 759 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 760 /* Truncate SYN, it is out of window starting 761 at tcp_rsk(req)->rcv_isn + 1. */ 762 flg &= ~TCP_FLAG_SYN; 763 } 764 765 /* RFC793: "second check the RST bit" and 766 * "fourth, check the SYN bit" 767 */ 768 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 769 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 770 goto embryonic_reset; 771 } 772 773 /* ACK sequence verified above, just make sure ACK is 774 * set. If ACK not set, just silently drop the packet. 775 * 776 * XXX (TFO) - if we ever allow "data after SYN", the 777 * following check needs to be removed. 778 */ 779 if (!(flg & TCP_FLAG_ACK)) 780 return NULL; 781 782 /* For Fast Open no more processing is needed (sk is the 783 * child socket). 784 */ 785 if (fastopen) 786 return sk; 787 788 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 789 if (req->num_timeout < READ_ONCE(inet_csk(sk)->icsk_accept_queue.rskq_defer_accept) && 790 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 791 inet_rsk(req)->acked = 1; 792 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 793 return NULL; 794 } 795 796 /* OK, ACK is valid, create big socket and 797 * feed this segment to it. It will repeat all 798 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 799 * ESTABLISHED STATE. If it will be dropped after 800 * socket is created, wait for troubles. 801 */ 802 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL, 803 req, &own_req); 804 if (!child) 805 goto listen_overflow; 806 807 if (own_req && rsk_drop_req(req)) { 808 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req); 809 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, req); 810 return child; 811 } 812 813 sock_rps_save_rxhash(child, skb); 814 tcp_synack_rtt_meas(child, req); 815 *req_stolen = !own_req; 816 return inet_csk_complete_hashdance(sk, child, req, own_req); 817 818 listen_overflow: 819 if (sk != req->rsk_listener) 820 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE); 821 822 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow)) { 823 inet_rsk(req)->acked = 1; 824 return NULL; 825 } 826 827 embryonic_reset: 828 if (!(flg & TCP_FLAG_RST)) { 829 /* Received a bad SYN pkt - for TFO We try not to reset 830 * the local connection unless it's really necessary to 831 * avoid becoming vulnerable to outside attack aiming at 832 * resetting legit local connections. 833 */ 834 req->rsk_ops->send_reset(sk, skb); 835 } else if (fastopen) { /* received a valid RST pkt */ 836 reqsk_fastopen_remove(sk, req, true); 837 tcp_reset(sk, skb); 838 } 839 if (!fastopen) { 840 bool unlinked = inet_csk_reqsk_queue_drop(sk, req); 841 842 if (unlinked) 843 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 844 *req_stolen = !unlinked; 845 } 846 return NULL; 847 } 848 EXPORT_SYMBOL(tcp_check_req); 849 850 /* 851 * Queue segment on the new socket if the new socket is active, 852 * otherwise we just shortcircuit this and continue with 853 * the new socket. 854 * 855 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV 856 * when entering. But other states are possible due to a race condition 857 * where after __inet_lookup_established() fails but before the listener 858 * locked is obtained, other packets cause the same connection to 859 * be created. 860 */ 861 862 int tcp_child_process(struct sock *parent, struct sock *child, 863 struct sk_buff *skb) 864 __releases(&((child)->sk_lock.slock)) 865 { 866 int ret = 0; 867 int state = child->sk_state; 868 869 /* record sk_napi_id and sk_rx_queue_mapping of child. */ 870 sk_mark_napi_id_set(child, skb); 871 872 tcp_segs_in(tcp_sk(child), skb); 873 if (!sock_owned_by_user(child)) { 874 ret = tcp_rcv_state_process(child, skb); 875 /* Wakeup parent, send SIGIO */ 876 if (state == TCP_SYN_RECV && child->sk_state != state) 877 parent->sk_data_ready(parent); 878 } else { 879 /* Alas, it is possible again, because we do lookup 880 * in main socket hash table and lock on listening 881 * socket does not protect us more. 882 */ 883 __sk_add_backlog(child, skb); 884 } 885 886 bh_unlock_sock(child); 887 sock_put(child); 888 return ret; 889 } 890 EXPORT_SYMBOL(tcp_child_process); 891