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 <linux/mm.h> 23 #include <linux/module.h> 24 #include <linux/slab.h> 25 #include <linux/sysctl.h> 26 #include <linux/workqueue.h> 27 #include <linux/static_key.h> 28 #include <net/tcp.h> 29 #include <net/inet_common.h> 30 #include <net/xfrm.h> 31 #include <net/busy_poll.h> 32 33 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win) 34 { 35 if (seq == s_win) 36 return true; 37 if (after(end_seq, s_win) && before(seq, e_win)) 38 return true; 39 return seq == e_win && seq == end_seq; 40 } 41 42 static enum tcp_tw_status 43 tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw, 44 const struct sk_buff *skb, int mib_idx) 45 { 46 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 47 48 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx, 49 &tcptw->tw_last_oow_ack_time)) { 50 /* Send ACK. Note, we do not put the bucket, 51 * it will be released by caller. 52 */ 53 return TCP_TW_ACK; 54 } 55 56 /* We are rate-limiting, so just release the tw sock and drop skb. */ 57 inet_twsk_put(tw); 58 return TCP_TW_SUCCESS; 59 } 60 61 /* 62 * * Main purpose of TIME-WAIT state is to close connection gracefully, 63 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN 64 * (and, probably, tail of data) and one or more our ACKs are lost. 65 * * What is TIME-WAIT timeout? It is associated with maximal packet 66 * lifetime in the internet, which results in wrong conclusion, that 67 * it is set to catch "old duplicate segments" wandering out of their path. 68 * It is not quite correct. This timeout is calculated so that it exceeds 69 * maximal retransmission timeout enough to allow to lose one (or more) 70 * segments sent by peer and our ACKs. This time may be calculated from RTO. 71 * * When TIME-WAIT socket receives RST, it means that another end 72 * finally closed and we are allowed to kill TIME-WAIT too. 73 * * Second purpose of TIME-WAIT is catching old duplicate segments. 74 * Well, certainly it is pure paranoia, but if we load TIME-WAIT 75 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs. 76 * * If we invented some more clever way to catch duplicates 77 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs. 78 * 79 * The algorithm below is based on FORMAL INTERPRETATION of RFCs. 80 * When you compare it to RFCs, please, read section SEGMENT ARRIVES 81 * from the very beginning. 82 * 83 * NOTE. With recycling (and later with fin-wait-2) TW bucket 84 * is _not_ stateless. It means, that strictly speaking we must 85 * spinlock it. I do not want! Well, probability of misbehaviour 86 * is ridiculously low and, seems, we could use some mb() tricks 87 * to avoid misread sequence numbers, states etc. --ANK 88 * 89 * We don't need to initialize tmp_out.sack_ok as we don't use the results 90 */ 91 enum tcp_tw_status 92 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb, 93 const struct tcphdr *th) 94 { 95 struct tcp_options_received tmp_opt; 96 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 97 bool paws_reject = false; 98 99 tmp_opt.saw_tstamp = 0; 100 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) { 101 tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL); 102 103 if (tmp_opt.saw_tstamp) { 104 if (tmp_opt.rcv_tsecr) 105 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset; 106 tmp_opt.ts_recent = tcptw->tw_ts_recent; 107 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp; 108 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 109 } 110 } 111 112 if (tw->tw_substate == TCP_FIN_WAIT2) { 113 /* Just repeat all the checks of tcp_rcv_state_process() */ 114 115 /* Out of window, send ACK */ 116 if (paws_reject || 117 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 118 tcptw->tw_rcv_nxt, 119 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd)) 120 return tcp_timewait_check_oow_rate_limit( 121 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2); 122 123 if (th->rst) 124 goto kill; 125 126 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt)) 127 return TCP_TW_RST; 128 129 /* Dup ACK? */ 130 if (!th->ack || 131 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) || 132 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) { 133 inet_twsk_put(tw); 134 return TCP_TW_SUCCESS; 135 } 136 137 /* New data or FIN. If new data arrive after half-duplex close, 138 * reset. 139 */ 140 if (!th->fin || 141 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) 142 return TCP_TW_RST; 143 144 /* FIN arrived, enter true time-wait state. */ 145 tw->tw_substate = TCP_TIME_WAIT; 146 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq; 147 if (tmp_opt.saw_tstamp) { 148 tcptw->tw_ts_recent_stamp = ktime_get_seconds(); 149 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 150 } 151 152 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 153 return TCP_TW_ACK; 154 } 155 156 /* 157 * Now real TIME-WAIT state. 158 * 159 * RFC 1122: 160 * "When a connection is [...] on TIME-WAIT state [...] 161 * [a TCP] MAY accept a new SYN from the remote TCP to 162 * reopen the connection directly, if it: 163 * 164 * (1) assigns its initial sequence number for the new 165 * connection to be larger than the largest sequence 166 * number it used on the previous connection incarnation, 167 * and 168 * 169 * (2) returns to TIME-WAIT state if the SYN turns out 170 * to be an old duplicate". 171 */ 172 173 if (!paws_reject && 174 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt && 175 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 176 /* In window segment, it may be only reset or bare ack. */ 177 178 if (th->rst) { 179 /* This is TIME_WAIT assassination, in two flavors. 180 * Oh well... nobody has a sufficient solution to this 181 * protocol bug yet. 182 */ 183 if (twsk_net(tw)->ipv4.sysctl_tcp_rfc1337 == 0) { 184 kill: 185 inet_twsk_deschedule_put(tw); 186 return TCP_TW_SUCCESS; 187 } 188 } else { 189 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 190 } 191 192 if (tmp_opt.saw_tstamp) { 193 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 194 tcptw->tw_ts_recent_stamp = ktime_get_seconds(); 195 } 196 197 inet_twsk_put(tw); 198 return TCP_TW_SUCCESS; 199 } 200 201 /* Out of window segment. 202 203 All the segments are ACKed immediately. 204 205 The only exception is new SYN. We accept it, if it is 206 not old duplicate and we are not in danger to be killed 207 by delayed old duplicates. RFC check is that it has 208 newer sequence number works at rates <40Mbit/sec. 209 However, if paws works, it is reliable AND even more, 210 we even may relax silly seq space cutoff. 211 212 RED-PEN: we violate main RFC requirement, if this SYN will appear 213 old duplicate (i.e. we receive RST in reply to SYN-ACK), 214 we must return socket to time-wait state. It is not good, 215 but not fatal yet. 216 */ 217 218 if (th->syn && !th->rst && !th->ack && !paws_reject && 219 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) || 220 (tmp_opt.saw_tstamp && 221 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) { 222 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 223 if (isn == 0) 224 isn++; 225 TCP_SKB_CB(skb)->tcp_tw_isn = isn; 226 return TCP_TW_SYN; 227 } 228 229 if (paws_reject) 230 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED); 231 232 if (!th->rst) { 233 /* In this case we must reset the TIMEWAIT timer. 234 * 235 * If it is ACKless SYN it may be both old duplicate 236 * and new good SYN with random sequence number <rcv_nxt. 237 * Do not reschedule in the last case. 238 */ 239 if (paws_reject || th->ack) 240 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 241 242 return tcp_timewait_check_oow_rate_limit( 243 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT); 244 } 245 inet_twsk_put(tw); 246 return TCP_TW_SUCCESS; 247 } 248 EXPORT_SYMBOL(tcp_timewait_state_process); 249 250 /* 251 * Move a socket to time-wait or dead fin-wait-2 state. 252 */ 253 void tcp_time_wait(struct sock *sk, int state, int timeo) 254 { 255 const struct inet_connection_sock *icsk = inet_csk(sk); 256 const struct tcp_sock *tp = tcp_sk(sk); 257 struct inet_timewait_sock *tw; 258 struct inet_timewait_death_row *tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row; 259 260 tw = inet_twsk_alloc(sk, tcp_death_row, state); 261 262 if (tw) { 263 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 264 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 265 struct inet_sock *inet = inet_sk(sk); 266 267 tw->tw_transparent = inet->transparent; 268 tw->tw_mark = sk->sk_mark; 269 tw->tw_priority = sk->sk_priority; 270 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 271 tcptw->tw_rcv_nxt = tp->rcv_nxt; 272 tcptw->tw_snd_nxt = tp->snd_nxt; 273 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 274 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 275 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 276 tcptw->tw_ts_offset = tp->tsoffset; 277 tcptw->tw_last_oow_ack_time = 0; 278 tcptw->tw_tx_delay = tp->tcp_tx_delay; 279 #if IS_ENABLED(CONFIG_IPV6) 280 if (tw->tw_family == PF_INET6) { 281 struct ipv6_pinfo *np = inet6_sk(sk); 282 283 tw->tw_v6_daddr = sk->sk_v6_daddr; 284 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr; 285 tw->tw_tclass = np->tclass; 286 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK); 287 tw->tw_txhash = sk->sk_txhash; 288 tw->tw_ipv6only = sk->sk_ipv6only; 289 } 290 #endif 291 292 #ifdef CONFIG_TCP_MD5SIG 293 /* 294 * The timewait bucket does not have the key DB from the 295 * sock structure. We just make a quick copy of the 296 * md5 key being used (if indeed we are using one) 297 * so the timewait ack generating code has the key. 298 */ 299 do { 300 tcptw->tw_md5_key = NULL; 301 if (static_branch_unlikely(&tcp_md5_needed)) { 302 struct tcp_md5sig_key *key; 303 304 key = tp->af_specific->md5_lookup(sk, sk); 305 if (key) { 306 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC); 307 BUG_ON(tcptw->tw_md5_key && !tcp_alloc_md5sig_pool()); 308 } 309 } 310 } while (0); 311 #endif 312 313 /* Get the TIME_WAIT timeout firing. */ 314 if (timeo < rto) 315 timeo = rto; 316 317 if (state == TCP_TIME_WAIT) 318 timeo = TCP_TIMEWAIT_LEN; 319 320 /* tw_timer is pinned, so we need to make sure BH are disabled 321 * in following section, otherwise timer handler could run before 322 * we complete the initialization. 323 */ 324 local_bh_disable(); 325 inet_twsk_schedule(tw, timeo); 326 /* Linkage updates. 327 * Note that access to tw after this point is illegal. 328 */ 329 inet_twsk_hashdance(tw, sk, &tcp_hashinfo); 330 local_bh_enable(); 331 } else { 332 /* Sorry, if we're out of memory, just CLOSE this 333 * socket up. We've got bigger problems than 334 * non-graceful socket closings. 335 */ 336 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW); 337 } 338 339 tcp_update_metrics(sk); 340 tcp_done(sk); 341 } 342 EXPORT_SYMBOL(tcp_time_wait); 343 344 void tcp_twsk_destructor(struct sock *sk) 345 { 346 #ifdef CONFIG_TCP_MD5SIG 347 if (static_branch_unlikely(&tcp_md5_needed)) { 348 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 349 350 if (twsk->tw_md5_key) 351 kfree_rcu(twsk->tw_md5_key, rcu); 352 } 353 #endif 354 } 355 EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 356 357 /* Warning : This function is called without sk_listener being locked. 358 * Be sure to read socket fields once, as their value could change under us. 359 */ 360 void tcp_openreq_init_rwin(struct request_sock *req, 361 const struct sock *sk_listener, 362 const struct dst_entry *dst) 363 { 364 struct inet_request_sock *ireq = inet_rsk(req); 365 const struct tcp_sock *tp = tcp_sk(sk_listener); 366 int full_space = tcp_full_space(sk_listener); 367 u32 window_clamp; 368 __u8 rcv_wscale; 369 u32 rcv_wnd; 370 int mss; 371 372 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 373 window_clamp = READ_ONCE(tp->window_clamp); 374 /* Set this up on the first call only */ 375 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW); 376 377 /* limit the window selection if the user enforce a smaller rx buffer */ 378 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK && 379 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0)) 380 req->rsk_window_clamp = full_space; 381 382 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req); 383 if (rcv_wnd == 0) 384 rcv_wnd = dst_metric(dst, RTAX_INITRWND); 385 else if (full_space < rcv_wnd * mss) 386 full_space = rcv_wnd * mss; 387 388 /* tcp_full_space because it is guaranteed to be the first packet */ 389 tcp_select_initial_window(sk_listener, full_space, 390 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 391 &req->rsk_rcv_wnd, 392 &req->rsk_window_clamp, 393 ireq->wscale_ok, 394 &rcv_wscale, 395 rcv_wnd); 396 ireq->rcv_wscale = rcv_wscale; 397 } 398 EXPORT_SYMBOL(tcp_openreq_init_rwin); 399 400 static void tcp_ecn_openreq_child(struct tcp_sock *tp, 401 const struct request_sock *req) 402 { 403 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 404 } 405 406 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst) 407 { 408 struct inet_connection_sock *icsk = inet_csk(sk); 409 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 410 bool ca_got_dst = false; 411 412 if (ca_key != TCP_CA_UNSPEC) { 413 const struct tcp_congestion_ops *ca; 414 415 rcu_read_lock(); 416 ca = tcp_ca_find_key(ca_key); 417 if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 418 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 419 icsk->icsk_ca_ops = ca; 420 ca_got_dst = true; 421 } 422 rcu_read_unlock(); 423 } 424 425 /* If no valid choice made yet, assign current system default ca. */ 426 if (!ca_got_dst && 427 (!icsk->icsk_ca_setsockopt || 428 !bpf_try_module_get(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner))) 429 tcp_assign_congestion_control(sk); 430 431 tcp_set_ca_state(sk, TCP_CA_Open); 432 } 433 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child); 434 435 static void smc_check_reset_syn_req(struct tcp_sock *oldtp, 436 struct request_sock *req, 437 struct tcp_sock *newtp) 438 { 439 #if IS_ENABLED(CONFIG_SMC) 440 struct inet_request_sock *ireq; 441 442 if (static_branch_unlikely(&tcp_have_smc)) { 443 ireq = inet_rsk(req); 444 if (oldtp->syn_smc && !ireq->smc_ok) 445 newtp->syn_smc = 0; 446 } 447 #endif 448 } 449 450 /* This is not only more efficient than what we used to do, it eliminates 451 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 452 * 453 * Actually, we could lots of memory writes here. tp of listening 454 * socket contains all necessary default parameters. 455 */ 456 struct sock *tcp_create_openreq_child(const struct sock *sk, 457 struct request_sock *req, 458 struct sk_buff *skb) 459 { 460 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 461 const struct inet_request_sock *ireq = inet_rsk(req); 462 struct tcp_request_sock *treq = tcp_rsk(req); 463 struct inet_connection_sock *newicsk; 464 struct tcp_sock *oldtp, *newtp; 465 u32 seq; 466 467 if (!newsk) 468 return NULL; 469 470 newicsk = inet_csk(newsk); 471 newtp = tcp_sk(newsk); 472 oldtp = tcp_sk(sk); 473 474 smc_check_reset_syn_req(oldtp, req, newtp); 475 476 /* Now setup tcp_sock */ 477 newtp->pred_flags = 0; 478 479 seq = treq->rcv_isn + 1; 480 newtp->rcv_wup = seq; 481 WRITE_ONCE(newtp->copied_seq, seq); 482 WRITE_ONCE(newtp->rcv_nxt, seq); 483 newtp->segs_in = 1; 484 485 seq = treq->snt_isn + 1; 486 newtp->snd_sml = newtp->snd_una = seq; 487 WRITE_ONCE(newtp->snd_nxt, seq); 488 newtp->snd_up = seq; 489 490 INIT_LIST_HEAD(&newtp->tsq_node); 491 INIT_LIST_HEAD(&newtp->tsorted_sent_queue); 492 493 tcp_init_wl(newtp, treq->rcv_isn); 494 495 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U); 496 newicsk->icsk_ack.lrcvtime = tcp_jiffies32; 497 498 newtp->lsndtime = tcp_jiffies32; 499 newsk->sk_txhash = treq->txhash; 500 newtp->total_retrans = req->num_retrans; 501 502 tcp_init_xmit_timers(newsk); 503 WRITE_ONCE(newtp->write_seq, newtp->pushed_seq = treq->snt_isn + 1); 504 505 if (sock_flag(newsk, SOCK_KEEPOPEN)) 506 inet_csk_reset_keepalive_timer(newsk, 507 keepalive_time_when(newtp)); 508 509 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 510 newtp->rx_opt.sack_ok = ireq->sack_ok; 511 newtp->window_clamp = req->rsk_window_clamp; 512 newtp->rcv_ssthresh = req->rsk_rcv_wnd; 513 newtp->rcv_wnd = req->rsk_rcv_wnd; 514 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 515 if (newtp->rx_opt.wscale_ok) { 516 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 517 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 518 } else { 519 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 520 newtp->window_clamp = min(newtp->window_clamp, 65535U); 521 } 522 newtp->snd_wnd = ntohs(tcp_hdr(skb)->window) << newtp->rx_opt.snd_wscale; 523 newtp->max_window = newtp->snd_wnd; 524 525 if (newtp->rx_opt.tstamp_ok) { 526 newtp->rx_opt.ts_recent = req->ts_recent; 527 newtp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 528 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 529 } else { 530 newtp->rx_opt.ts_recent_stamp = 0; 531 newtp->tcp_header_len = sizeof(struct tcphdr); 532 } 533 if (req->num_timeout) { 534 newtp->undo_marker = treq->snt_isn; 535 newtp->retrans_stamp = div_u64(treq->snt_synack, 536 USEC_PER_SEC / TCP_TS_HZ); 537 } 538 newtp->tsoffset = treq->ts_off; 539 #ifdef CONFIG_TCP_MD5SIG 540 newtp->md5sig_info = NULL; /*XXX*/ 541 if (newtp->af_specific->md5_lookup(sk, newsk)) 542 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 543 #endif 544 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 545 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 546 newtp->rx_opt.mss_clamp = req->mss; 547 tcp_ecn_openreq_child(newtp, req); 548 newtp->fastopen_req = NULL; 549 RCU_INIT_POINTER(newtp->fastopen_rsk, NULL); 550 551 tcp_bpf_clone(sk, newsk); 552 553 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS); 554 555 return newsk; 556 } 557 EXPORT_SYMBOL(tcp_create_openreq_child); 558 559 /* 560 * Process an incoming packet for SYN_RECV sockets represented as a 561 * request_sock. Normally sk is the listener socket but for TFO it 562 * points to the child socket. 563 * 564 * XXX (TFO) - The current impl contains a special check for ack 565 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better? 566 * 567 * We don't need to initialize tmp_opt.sack_ok as we don't use the results 568 */ 569 570 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 571 struct request_sock *req, 572 bool fastopen, bool *req_stolen) 573 { 574 struct tcp_options_received tmp_opt; 575 struct sock *child; 576 const struct tcphdr *th = tcp_hdr(skb); 577 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 578 bool paws_reject = false; 579 bool own_req; 580 581 tmp_opt.saw_tstamp = 0; 582 if (th->doff > (sizeof(struct tcphdr)>>2)) { 583 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL); 584 585 if (tmp_opt.saw_tstamp) { 586 tmp_opt.ts_recent = req->ts_recent; 587 if (tmp_opt.rcv_tsecr) 588 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off; 589 /* We do not store true stamp, but it is not required, 590 * it can be estimated (approximately) 591 * from another data. 592 */ 593 tmp_opt.ts_recent_stamp = ktime_get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout); 594 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 595 } 596 } 597 598 /* Check for pure retransmitted SYN. */ 599 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 600 flg == TCP_FLAG_SYN && 601 !paws_reject) { 602 /* 603 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 604 * this case on figure 6 and figure 8, but formal 605 * protocol description says NOTHING. 606 * To be more exact, it says that we should send ACK, 607 * because this segment (at least, if it has no data) 608 * is out of window. 609 * 610 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 611 * describe SYN-RECV state. All the description 612 * is wrong, we cannot believe to it and should 613 * rely only on common sense and implementation 614 * experience. 615 * 616 * Enforce "SYN-ACK" according to figure 8, figure 6 617 * of RFC793, fixed by RFC1122. 618 * 619 * Note that even if there is new data in the SYN packet 620 * they will be thrown away too. 621 * 622 * Reset timer after retransmitting SYNACK, similar to 623 * the idea of fast retransmit in recovery. 624 */ 625 if (!tcp_oow_rate_limited(sock_net(sk), skb, 626 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 627 &tcp_rsk(req)->last_oow_ack_time) && 628 629 !inet_rtx_syn_ack(sk, req)) { 630 unsigned long expires = jiffies; 631 632 expires += min(TCP_TIMEOUT_INIT << req->num_timeout, 633 TCP_RTO_MAX); 634 if (!fastopen) 635 mod_timer_pending(&req->rsk_timer, expires); 636 else 637 req->rsk_timer.expires = expires; 638 } 639 return NULL; 640 } 641 642 /* Further reproduces section "SEGMENT ARRIVES" 643 for state SYN-RECEIVED of RFC793. 644 It is broken, however, it does not work only 645 when SYNs are crossed. 646 647 You would think that SYN crossing is impossible here, since 648 we should have a SYN_SENT socket (from connect()) on our end, 649 but this is not true if the crossed SYNs were sent to both 650 ends by a malicious third party. We must defend against this, 651 and to do that we first verify the ACK (as per RFC793, page 652 36) and reset if it is invalid. Is this a true full defense? 653 To convince ourselves, let us consider a way in which the ACK 654 test can still pass in this 'malicious crossed SYNs' case. 655 Malicious sender sends identical SYNs (and thus identical sequence 656 numbers) to both A and B: 657 658 A: gets SYN, seq=7 659 B: gets SYN, seq=7 660 661 By our good fortune, both A and B select the same initial 662 send sequence number of seven :-) 663 664 A: sends SYN|ACK, seq=7, ack_seq=8 665 B: sends SYN|ACK, seq=7, ack_seq=8 666 667 So we are now A eating this SYN|ACK, ACK test passes. So 668 does sequence test, SYN is truncated, and thus we consider 669 it a bare ACK. 670 671 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 672 bare ACK. Otherwise, we create an established connection. Both 673 ends (listening sockets) accept the new incoming connection and try 674 to talk to each other. 8-) 675 676 Note: This case is both harmless, and rare. Possibility is about the 677 same as us discovering intelligent life on another plant tomorrow. 678 679 But generally, we should (RFC lies!) to accept ACK 680 from SYNACK both here and in tcp_rcv_state_process(). 681 tcp_rcv_state_process() does not, hence, we do not too. 682 683 Note that the case is absolutely generic: 684 we cannot optimize anything here without 685 violating protocol. All the checks must be made 686 before attempt to create socket. 687 */ 688 689 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 690 * and the incoming segment acknowledges something not yet 691 * sent (the segment carries an unacceptable ACK) ... 692 * a reset is sent." 693 * 694 * Invalid ACK: reset will be sent by listening socket. 695 * Note that the ACK validity check for a Fast Open socket is done 696 * elsewhere and is checked directly against the child socket rather 697 * than req because user data may have been sent out. 698 */ 699 if ((flg & TCP_FLAG_ACK) && !fastopen && 700 (TCP_SKB_CB(skb)->ack_seq != 701 tcp_rsk(req)->snt_isn + 1)) 702 return sk; 703 704 /* Also, it would be not so bad idea to check rcv_tsecr, which 705 * is essentially ACK extension and too early or too late values 706 * should cause reset in unsynchronized states. 707 */ 708 709 /* RFC793: "first check sequence number". */ 710 711 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 712 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) { 713 /* Out of window: send ACK and drop. */ 714 if (!(flg & TCP_FLAG_RST) && 715 !tcp_oow_rate_limited(sock_net(sk), skb, 716 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 717 &tcp_rsk(req)->last_oow_ack_time)) 718 req->rsk_ops->send_ack(sk, skb, req); 719 if (paws_reject) 720 __NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 721 return NULL; 722 } 723 724 /* In sequence, PAWS is OK. */ 725 726 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt)) 727 req->ts_recent = tmp_opt.rcv_tsval; 728 729 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 730 /* Truncate SYN, it is out of window starting 731 at tcp_rsk(req)->rcv_isn + 1. */ 732 flg &= ~TCP_FLAG_SYN; 733 } 734 735 /* RFC793: "second check the RST bit" and 736 * "fourth, check the SYN bit" 737 */ 738 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 739 __TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 740 goto embryonic_reset; 741 } 742 743 /* ACK sequence verified above, just make sure ACK is 744 * set. If ACK not set, just silently drop the packet. 745 * 746 * XXX (TFO) - if we ever allow "data after SYN", the 747 * following check needs to be removed. 748 */ 749 if (!(flg & TCP_FLAG_ACK)) 750 return NULL; 751 752 /* For Fast Open no more processing is needed (sk is the 753 * child socket). 754 */ 755 if (fastopen) 756 return sk; 757 758 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 759 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 760 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 761 inet_rsk(req)->acked = 1; 762 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 763 return NULL; 764 } 765 766 /* OK, ACK is valid, create big socket and 767 * feed this segment to it. It will repeat all 768 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 769 * ESTABLISHED STATE. If it will be dropped after 770 * socket is created, wait for troubles. 771 */ 772 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL, 773 req, &own_req); 774 if (!child) 775 goto listen_overflow; 776 777 if (own_req && rsk_drop_req(req)) { 778 reqsk_queue_removed(&inet_csk(req->rsk_listener)->icsk_accept_queue, req); 779 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener, req); 780 return child; 781 } 782 783 sock_rps_save_rxhash(child, skb); 784 tcp_synack_rtt_meas(child, req); 785 *req_stolen = !own_req; 786 return inet_csk_complete_hashdance(sk, child, req, own_req); 787 788 listen_overflow: 789 if (sk != req->rsk_listener) 790 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMIGRATEREQFAILURE); 791 792 if (!sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow) { 793 inet_rsk(req)->acked = 1; 794 return NULL; 795 } 796 797 embryonic_reset: 798 if (!(flg & TCP_FLAG_RST)) { 799 /* Received a bad SYN pkt - for TFO We try not to reset 800 * the local connection unless it's really necessary to 801 * avoid becoming vulnerable to outside attack aiming at 802 * resetting legit local connections. 803 */ 804 req->rsk_ops->send_reset(sk, skb); 805 } else if (fastopen) { /* received a valid RST pkt */ 806 reqsk_fastopen_remove(sk, req, true); 807 tcp_reset(sk, skb); 808 } 809 if (!fastopen) { 810 bool unlinked = inet_csk_reqsk_queue_drop(sk, req); 811 812 if (unlinked) 813 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 814 *req_stolen = !unlinked; 815 } 816 return NULL; 817 } 818 EXPORT_SYMBOL(tcp_check_req); 819 820 /* 821 * Queue segment on the new socket if the new socket is active, 822 * otherwise we just shortcircuit this and continue with 823 * the new socket. 824 * 825 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV 826 * when entering. But other states are possible due to a race condition 827 * where after __inet_lookup_established() fails but before the listener 828 * locked is obtained, other packets cause the same connection to 829 * be created. 830 */ 831 832 int tcp_child_process(struct sock *parent, struct sock *child, 833 struct sk_buff *skb) 834 __releases(&((child)->sk_lock.slock)) 835 { 836 int ret = 0; 837 int state = child->sk_state; 838 839 /* record NAPI ID of child */ 840 sk_mark_napi_id(child, skb); 841 842 tcp_segs_in(tcp_sk(child), skb); 843 if (!sock_owned_by_user(child)) { 844 ret = tcp_rcv_state_process(child, skb); 845 /* Wakeup parent, send SIGIO */ 846 if (state == TCP_SYN_RECV && child->sk_state != state) 847 parent->sk_data_ready(parent); 848 } else { 849 /* Alas, it is possible again, because we do lookup 850 * in main socket hash table and lock on listening 851 * socket does not protect us more. 852 */ 853 __sk_add_backlog(child, skb); 854 } 855 856 bh_unlock_sock(child); 857 sock_put(child); 858 return ret; 859 } 860 EXPORT_SYMBOL(tcp_child_process); 861