1 /* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Authors: Ross Biro 9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 10 * Mark Evans, <evansmp@uhura.aston.ac.uk> 11 * Corey Minyard <wf-rch!minyard@relay.EU.net> 12 * Florian La Roche, <flla@stud.uni-sb.de> 13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 14 * Linus Torvalds, <torvalds@cs.helsinki.fi> 15 * Alan Cox, <gw4pts@gw4pts.ampr.org> 16 * Matthew Dillon, <dillon@apollo.west.oic.com> 17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 18 * Jorge Cwik, <jorge@laser.satlink.net> 19 */ 20 21 #include <linux/mm.h> 22 #include <linux/module.h> 23 #include <linux/slab.h> 24 #include <linux/sysctl.h> 25 #include <linux/workqueue.h> 26 #include <linux/static_key.h> 27 #include <net/tcp.h> 28 #include <net/inet_common.h> 29 #include <net/xfrm.h> 30 #include <net/busy_poll.h> 31 32 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win) 33 { 34 if (seq == s_win) 35 return true; 36 if (after(end_seq, s_win) && before(seq, e_win)) 37 return true; 38 return seq == e_win && seq == end_seq; 39 } 40 41 static enum tcp_tw_status 42 tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw, 43 const struct sk_buff *skb, int mib_idx) 44 { 45 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 46 47 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx, 48 &tcptw->tw_last_oow_ack_time)) { 49 /* Send ACK. Note, we do not put the bucket, 50 * it will be released by caller. 51 */ 52 return TCP_TW_ACK; 53 } 54 55 /* We are rate-limiting, so just release the tw sock and drop skb. */ 56 inet_twsk_put(tw); 57 return TCP_TW_SUCCESS; 58 } 59 60 /* 61 * * Main purpose of TIME-WAIT state is to close connection gracefully, 62 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN 63 * (and, probably, tail of data) and one or more our ACKs are lost. 64 * * What is TIME-WAIT timeout? It is associated with maximal packet 65 * lifetime in the internet, which results in wrong conclusion, that 66 * it is set to catch "old duplicate segments" wandering out of their path. 67 * It is not quite correct. This timeout is calculated so that it exceeds 68 * maximal retransmission timeout enough to allow to lose one (or more) 69 * segments sent by peer and our ACKs. This time may be calculated from RTO. 70 * * When TIME-WAIT socket receives RST, it means that another end 71 * finally closed and we are allowed to kill TIME-WAIT too. 72 * * Second purpose of TIME-WAIT is catching old duplicate segments. 73 * Well, certainly it is pure paranoia, but if we load TIME-WAIT 74 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs. 75 * * If we invented some more clever way to catch duplicates 76 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs. 77 * 78 * The algorithm below is based on FORMAL INTERPRETATION of RFCs. 79 * When you compare it to RFCs, please, read section SEGMENT ARRIVES 80 * from the very beginning. 81 * 82 * NOTE. With recycling (and later with fin-wait-2) TW bucket 83 * is _not_ stateless. It means, that strictly speaking we must 84 * spinlock it. I do not want! Well, probability of misbehaviour 85 * is ridiculously low and, seems, we could use some mb() tricks 86 * to avoid misread sequence numbers, states etc. --ANK 87 * 88 * We don't need to initialize tmp_out.sack_ok as we don't use the results 89 */ 90 enum tcp_tw_status 91 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb, 92 const struct tcphdr *th) 93 { 94 struct tcp_options_received tmp_opt; 95 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 96 bool paws_reject = false; 97 98 tmp_opt.saw_tstamp = 0; 99 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) { 100 tcp_parse_options(twsk_net(tw), skb, &tmp_opt, 0, NULL); 101 102 if (tmp_opt.saw_tstamp) { 103 if (tmp_opt.rcv_tsecr) 104 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset; 105 tmp_opt.ts_recent = tcptw->tw_ts_recent; 106 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp; 107 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 108 } 109 } 110 111 if (tw->tw_substate == TCP_FIN_WAIT2) { 112 /* Just repeat all the checks of tcp_rcv_state_process() */ 113 114 /* Out of window, send ACK */ 115 if (paws_reject || 116 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 117 tcptw->tw_rcv_nxt, 118 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd)) 119 return tcp_timewait_check_oow_rate_limit( 120 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2); 121 122 if (th->rst) 123 goto kill; 124 125 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt)) 126 return TCP_TW_RST; 127 128 /* Dup ACK? */ 129 if (!th->ack || 130 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) || 131 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) { 132 inet_twsk_put(tw); 133 return TCP_TW_SUCCESS; 134 } 135 136 /* New data or FIN. If new data arrive after half-duplex close, 137 * reset. 138 */ 139 if (!th->fin || 140 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) 141 return TCP_TW_RST; 142 143 /* FIN arrived, enter true time-wait state. */ 144 tw->tw_substate = TCP_TIME_WAIT; 145 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq; 146 if (tmp_opt.saw_tstamp) { 147 tcptw->tw_ts_recent_stamp = get_seconds(); 148 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 149 } 150 151 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 152 return TCP_TW_ACK; 153 } 154 155 /* 156 * Now real TIME-WAIT state. 157 * 158 * RFC 1122: 159 * "When a connection is [...] on TIME-WAIT state [...] 160 * [a TCP] MAY accept a new SYN from the remote TCP to 161 * reopen the connection directly, if it: 162 * 163 * (1) assigns its initial sequence number for the new 164 * connection to be larger than the largest sequence 165 * number it used on the previous connection incarnation, 166 * and 167 * 168 * (2) returns to TIME-WAIT state if the SYN turns out 169 * to be an old duplicate". 170 */ 171 172 if (!paws_reject && 173 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt && 174 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 175 /* In window segment, it may be only reset or bare ack. */ 176 177 if (th->rst) { 178 /* This is TIME_WAIT assassination, in two flavors. 179 * Oh well... nobody has a sufficient solution to this 180 * protocol bug yet. 181 */ 182 if (twsk_net(tw)->ipv4.sysctl_tcp_rfc1337 == 0) { 183 kill: 184 inet_twsk_deschedule_put(tw); 185 return TCP_TW_SUCCESS; 186 } 187 } 188 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 189 190 if (tmp_opt.saw_tstamp) { 191 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 192 tcptw->tw_ts_recent_stamp = get_seconds(); 193 } 194 195 inet_twsk_put(tw); 196 return TCP_TW_SUCCESS; 197 } 198 199 /* Out of window segment. 200 201 All the segments are ACKed immediately. 202 203 The only exception is new SYN. We accept it, if it is 204 not old duplicate and we are not in danger to be killed 205 by delayed old duplicates. RFC check is that it has 206 newer sequence number works at rates <40Mbit/sec. 207 However, if paws works, it is reliable AND even more, 208 we even may relax silly seq space cutoff. 209 210 RED-PEN: we violate main RFC requirement, if this SYN will appear 211 old duplicate (i.e. we receive RST in reply to SYN-ACK), 212 we must return socket to time-wait state. It is not good, 213 but not fatal yet. 214 */ 215 216 if (th->syn && !th->rst && !th->ack && !paws_reject && 217 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) || 218 (tmp_opt.saw_tstamp && 219 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) { 220 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 221 if (isn == 0) 222 isn++; 223 TCP_SKB_CB(skb)->tcp_tw_isn = isn; 224 return TCP_TW_SYN; 225 } 226 227 if (paws_reject) 228 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED); 229 230 if (!th->rst) { 231 /* In this case we must reset the TIMEWAIT timer. 232 * 233 * If it is ACKless SYN it may be both old duplicate 234 * and new good SYN with random sequence number <rcv_nxt. 235 * Do not reschedule in the last case. 236 */ 237 if (paws_reject || th->ack) 238 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 239 240 return tcp_timewait_check_oow_rate_limit( 241 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT); 242 } 243 inet_twsk_put(tw); 244 return TCP_TW_SUCCESS; 245 } 246 EXPORT_SYMBOL(tcp_timewait_state_process); 247 248 /* 249 * Move a socket to time-wait or dead fin-wait-2 state. 250 */ 251 void tcp_time_wait(struct sock *sk, int state, int timeo) 252 { 253 const struct inet_connection_sock *icsk = inet_csk(sk); 254 const struct tcp_sock *tp = tcp_sk(sk); 255 struct inet_timewait_sock *tw; 256 struct inet_timewait_death_row *tcp_death_row = &sock_net(sk)->ipv4.tcp_death_row; 257 258 tw = inet_twsk_alloc(sk, tcp_death_row, state); 259 260 if (tw) { 261 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 262 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 263 struct inet_sock *inet = inet_sk(sk); 264 265 tw->tw_transparent = inet->transparent; 266 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 267 tcptw->tw_rcv_nxt = tp->rcv_nxt; 268 tcptw->tw_snd_nxt = tp->snd_nxt; 269 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 270 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 271 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 272 tcptw->tw_ts_offset = tp->tsoffset; 273 tcptw->tw_last_oow_ack_time = 0; 274 275 #if IS_ENABLED(CONFIG_IPV6) 276 if (tw->tw_family == PF_INET6) { 277 struct ipv6_pinfo *np = inet6_sk(sk); 278 279 tw->tw_v6_daddr = sk->sk_v6_daddr; 280 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr; 281 tw->tw_tclass = np->tclass; 282 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK); 283 tw->tw_ipv6only = sk->sk_ipv6only; 284 } 285 #endif 286 287 #ifdef CONFIG_TCP_MD5SIG 288 /* 289 * The timewait bucket does not have the key DB from the 290 * sock structure. We just make a quick copy of the 291 * md5 key being used (if indeed we are using one) 292 * so the timewait ack generating code has the key. 293 */ 294 do { 295 struct tcp_md5sig_key *key; 296 tcptw->tw_md5_key = NULL; 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 } while (0); 303 #endif 304 305 /* Get the TIME_WAIT timeout firing. */ 306 if (timeo < rto) 307 timeo = rto; 308 309 tw->tw_timeout = TCP_TIMEWAIT_LEN; 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 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo); 321 inet_twsk_put(tw); 322 local_bh_enable(); 323 } else { 324 /* Sorry, if we're out of memory, just CLOSE this 325 * socket up. We've got bigger problems than 326 * non-graceful socket closings. 327 */ 328 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW); 329 } 330 331 tcp_update_metrics(sk); 332 tcp_done(sk); 333 } 334 335 void tcp_twsk_destructor(struct sock *sk) 336 { 337 #ifdef CONFIG_TCP_MD5SIG 338 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 339 340 if (twsk->tw_md5_key) 341 kfree_rcu(twsk->tw_md5_key, rcu); 342 #endif 343 } 344 EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 345 346 /* Warning : This function is called without sk_listener being locked. 347 * Be sure to read socket fields once, as their value could change under us. 348 */ 349 void tcp_openreq_init_rwin(struct request_sock *req, 350 const struct sock *sk_listener, 351 const struct dst_entry *dst) 352 { 353 struct inet_request_sock *ireq = inet_rsk(req); 354 const struct tcp_sock *tp = tcp_sk(sk_listener); 355 int full_space = tcp_full_space(sk_listener); 356 u32 window_clamp; 357 __u8 rcv_wscale; 358 u32 rcv_wnd; 359 int mss; 360 361 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 362 window_clamp = READ_ONCE(tp->window_clamp); 363 /* Set this up on the first call only */ 364 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW); 365 366 /* limit the window selection if the user enforce a smaller rx buffer */ 367 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK && 368 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0)) 369 req->rsk_window_clamp = full_space; 370 371 rcv_wnd = tcp_rwnd_init_bpf((struct sock *)req); 372 if (rcv_wnd == 0) 373 rcv_wnd = dst_metric(dst, RTAX_INITRWND); 374 else if (full_space < rcv_wnd * mss) 375 full_space = rcv_wnd * mss; 376 377 /* tcp_full_space because it is guaranteed to be the first packet */ 378 tcp_select_initial_window(sk_listener, full_space, 379 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 380 &req->rsk_rcv_wnd, 381 &req->rsk_window_clamp, 382 ireq->wscale_ok, 383 &rcv_wscale, 384 rcv_wnd); 385 ireq->rcv_wscale = rcv_wscale; 386 } 387 EXPORT_SYMBOL(tcp_openreq_init_rwin); 388 389 static void tcp_ecn_openreq_child(struct tcp_sock *tp, 390 const struct request_sock *req) 391 { 392 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 393 } 394 395 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst) 396 { 397 struct inet_connection_sock *icsk = inet_csk(sk); 398 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 399 bool ca_got_dst = false; 400 401 if (ca_key != TCP_CA_UNSPEC) { 402 const struct tcp_congestion_ops *ca; 403 404 rcu_read_lock(); 405 ca = tcp_ca_find_key(ca_key); 406 if (likely(ca && try_module_get(ca->owner))) { 407 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 408 icsk->icsk_ca_ops = ca; 409 ca_got_dst = true; 410 } 411 rcu_read_unlock(); 412 } 413 414 /* If no valid choice made yet, assign current system default ca. */ 415 if (!ca_got_dst && 416 (!icsk->icsk_ca_setsockopt || 417 !try_module_get(icsk->icsk_ca_ops->owner))) 418 tcp_assign_congestion_control(sk); 419 420 tcp_set_ca_state(sk, TCP_CA_Open); 421 } 422 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child); 423 424 static void smc_check_reset_syn_req(struct tcp_sock *oldtp, 425 struct request_sock *req, 426 struct tcp_sock *newtp) 427 { 428 #if IS_ENABLED(CONFIG_SMC) 429 struct inet_request_sock *ireq; 430 431 if (static_branch_unlikely(&tcp_have_smc)) { 432 ireq = inet_rsk(req); 433 if (oldtp->syn_smc && !ireq->smc_ok) 434 newtp->syn_smc = 0; 435 } 436 #endif 437 } 438 439 /* This is not only more efficient than what we used to do, it eliminates 440 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 441 * 442 * Actually, we could lots of memory writes here. tp of listening 443 * socket contains all necessary default parameters. 444 */ 445 struct sock *tcp_create_openreq_child(const struct sock *sk, 446 struct request_sock *req, 447 struct sk_buff *skb) 448 { 449 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 450 451 if (newsk) { 452 const struct inet_request_sock *ireq = inet_rsk(req); 453 struct tcp_request_sock *treq = tcp_rsk(req); 454 struct inet_connection_sock *newicsk = inet_csk(newsk); 455 struct tcp_sock *newtp = tcp_sk(newsk); 456 struct tcp_sock *oldtp = tcp_sk(sk); 457 458 smc_check_reset_syn_req(oldtp, req, newtp); 459 460 /* Now setup tcp_sock */ 461 newtp->pred_flags = 0; 462 463 newtp->rcv_wup = newtp->copied_seq = 464 newtp->rcv_nxt = treq->rcv_isn + 1; 465 newtp->segs_in = 1; 466 467 newtp->snd_sml = newtp->snd_una = 468 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1; 469 470 INIT_LIST_HEAD(&newtp->tsq_node); 471 INIT_LIST_HEAD(&newtp->tsorted_sent_queue); 472 473 tcp_init_wl(newtp, treq->rcv_isn); 474 475 newtp->srtt_us = 0; 476 newtp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 477 minmax_reset(&newtp->rtt_min, tcp_jiffies32, ~0U); 478 newicsk->icsk_rto = TCP_TIMEOUT_INIT; 479 newicsk->icsk_ack.lrcvtime = tcp_jiffies32; 480 481 newtp->packets_out = 0; 482 newtp->retrans_out = 0; 483 newtp->sacked_out = 0; 484 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 485 newtp->tlp_high_seq = 0; 486 newtp->lsndtime = tcp_jiffies32; 487 newsk->sk_txhash = treq->txhash; 488 newtp->last_oow_ack_time = 0; 489 newtp->total_retrans = req->num_retrans; 490 491 /* So many TCP implementations out there (incorrectly) count the 492 * initial SYN frame in their delayed-ACK and congestion control 493 * algorithms that we must have the following bandaid to talk 494 * efficiently to them. -DaveM 495 */ 496 newtp->snd_cwnd = TCP_INIT_CWND; 497 newtp->snd_cwnd_cnt = 0; 498 499 /* There's a bubble in the pipe until at least the first ACK. */ 500 newtp->app_limited = ~0U; 501 502 tcp_init_xmit_timers(newsk); 503 newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1; 504 505 newtp->rx_opt.saw_tstamp = 0; 506 507 newtp->rx_opt.dsack = 0; 508 newtp->rx_opt.num_sacks = 0; 509 510 newtp->urg_data = 0; 511 512 if (sock_flag(newsk, SOCK_KEEPOPEN)) 513 inet_csk_reset_keepalive_timer(newsk, 514 keepalive_time_when(newtp)); 515 516 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 517 newtp->rx_opt.sack_ok = ireq->sack_ok; 518 newtp->window_clamp = req->rsk_window_clamp; 519 newtp->rcv_ssthresh = req->rsk_rcv_wnd; 520 newtp->rcv_wnd = req->rsk_rcv_wnd; 521 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 522 if (newtp->rx_opt.wscale_ok) { 523 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 524 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 525 } else { 526 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 527 newtp->window_clamp = min(newtp->window_clamp, 65535U); 528 } 529 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) << 530 newtp->rx_opt.snd_wscale); 531 newtp->max_window = newtp->snd_wnd; 532 533 if (newtp->rx_opt.tstamp_ok) { 534 newtp->rx_opt.ts_recent = req->ts_recent; 535 newtp->rx_opt.ts_recent_stamp = get_seconds(); 536 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 537 } else { 538 newtp->rx_opt.ts_recent_stamp = 0; 539 newtp->tcp_header_len = sizeof(struct tcphdr); 540 } 541 newtp->tsoffset = treq->ts_off; 542 #ifdef CONFIG_TCP_MD5SIG 543 newtp->md5sig_info = NULL; /*XXX*/ 544 if (newtp->af_specific->md5_lookup(sk, newsk)) 545 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 546 #endif 547 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 548 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 549 newtp->rx_opt.mss_clamp = req->mss; 550 tcp_ecn_openreq_child(newtp, req); 551 newtp->fastopen_req = NULL; 552 newtp->fastopen_rsk = NULL; 553 newtp->syn_data_acked = 0; 554 newtp->rack.mstamp = 0; 555 newtp->rack.advanced = 0; 556 newtp->rack.reo_wnd_steps = 1; 557 newtp->rack.last_delivered = 0; 558 newtp->rack.reo_wnd_persist = 0; 559 newtp->rack.dsack_seen = 0; 560 561 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS); 562 } 563 return newsk; 564 } 565 EXPORT_SYMBOL(tcp_create_openreq_child); 566 567 /* 568 * Process an incoming packet for SYN_RECV sockets represented as a 569 * request_sock. Normally sk is the listener socket but for TFO it 570 * points to the child socket. 571 * 572 * XXX (TFO) - The current impl contains a special check for ack 573 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better? 574 * 575 * We don't need to initialize tmp_opt.sack_ok as we don't use the results 576 */ 577 578 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 579 struct request_sock *req, 580 bool fastopen) 581 { 582 struct tcp_options_received tmp_opt; 583 struct sock *child; 584 const struct tcphdr *th = tcp_hdr(skb); 585 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 586 bool paws_reject = false; 587 bool own_req; 588 589 tmp_opt.saw_tstamp = 0; 590 if (th->doff > (sizeof(struct tcphdr)>>2)) { 591 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, NULL); 592 593 if (tmp_opt.saw_tstamp) { 594 tmp_opt.ts_recent = req->ts_recent; 595 if (tmp_opt.rcv_tsecr) 596 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off; 597 /* We do not store true stamp, but it is not required, 598 * it can be estimated (approximately) 599 * from another data. 600 */ 601 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout); 602 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 603 } 604 } 605 606 /* Check for pure retransmitted SYN. */ 607 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 608 flg == TCP_FLAG_SYN && 609 !paws_reject) { 610 /* 611 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 612 * this case on figure 6 and figure 8, but formal 613 * protocol description says NOTHING. 614 * To be more exact, it says that we should send ACK, 615 * because this segment (at least, if it has no data) 616 * is out of window. 617 * 618 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 619 * describe SYN-RECV state. All the description 620 * is wrong, we cannot believe to it and should 621 * rely only on common sense and implementation 622 * experience. 623 * 624 * Enforce "SYN-ACK" according to figure 8, figure 6 625 * of RFC793, fixed by RFC1122. 626 * 627 * Note that even if there is new data in the SYN packet 628 * they will be thrown away too. 629 * 630 * Reset timer after retransmitting SYNACK, similar to 631 * the idea of fast retransmit in recovery. 632 */ 633 if (!tcp_oow_rate_limited(sock_net(sk), skb, 634 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 635 &tcp_rsk(req)->last_oow_ack_time) && 636 637 !inet_rtx_syn_ack(sk, req)) { 638 unsigned long expires = jiffies; 639 640 expires += min(TCP_TIMEOUT_INIT << req->num_timeout, 641 TCP_RTO_MAX); 642 if (!fastopen) 643 mod_timer_pending(&req->rsk_timer, expires); 644 else 645 req->rsk_timer.expires = expires; 646 } 647 return NULL; 648 } 649 650 /* Further reproduces section "SEGMENT ARRIVES" 651 for state SYN-RECEIVED of RFC793. 652 It is broken, however, it does not work only 653 when SYNs are crossed. 654 655 You would think that SYN crossing is impossible here, since 656 we should have a SYN_SENT socket (from connect()) on our end, 657 but this is not true if the crossed SYNs were sent to both 658 ends by a malicious third party. We must defend against this, 659 and to do that we first verify the ACK (as per RFC793, page 660 36) and reset if it is invalid. Is this a true full defense? 661 To convince ourselves, let us consider a way in which the ACK 662 test can still pass in this 'malicious crossed SYNs' case. 663 Malicious sender sends identical SYNs (and thus identical sequence 664 numbers) to both A and B: 665 666 A: gets SYN, seq=7 667 B: gets SYN, seq=7 668 669 By our good fortune, both A and B select the same initial 670 send sequence number of seven :-) 671 672 A: sends SYN|ACK, seq=7, ack_seq=8 673 B: sends SYN|ACK, seq=7, ack_seq=8 674 675 So we are now A eating this SYN|ACK, ACK test passes. So 676 does sequence test, SYN is truncated, and thus we consider 677 it a bare ACK. 678 679 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 680 bare ACK. Otherwise, we create an established connection. Both 681 ends (listening sockets) accept the new incoming connection and try 682 to talk to each other. 8-) 683 684 Note: This case is both harmless, and rare. Possibility is about the 685 same as us discovering intelligent life on another plant tomorrow. 686 687 But generally, we should (RFC lies!) to accept ACK 688 from SYNACK both here and in tcp_rcv_state_process(). 689 tcp_rcv_state_process() does not, hence, we do not too. 690 691 Note that the case is absolutely generic: 692 we cannot optimize anything here without 693 violating protocol. All the checks must be made 694 before attempt to create socket. 695 */ 696 697 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 698 * and the incoming segment acknowledges something not yet 699 * sent (the segment carries an unacceptable ACK) ... 700 * a reset is sent." 701 * 702 * Invalid ACK: reset will be sent by listening socket. 703 * Note that the ACK validity check for a Fast Open socket is done 704 * elsewhere and is checked directly against the child socket rather 705 * than req because user data may have been sent out. 706 */ 707 if ((flg & TCP_FLAG_ACK) && !fastopen && 708 (TCP_SKB_CB(skb)->ack_seq != 709 tcp_rsk(req)->snt_isn + 1)) 710 return sk; 711 712 /* Also, it would be not so bad idea to check rcv_tsecr, which 713 * is essentially ACK extension and too early or too late values 714 * should cause reset in unsynchronized states. 715 */ 716 717 /* RFC793: "first check sequence number". */ 718 719 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 720 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) { 721 /* Out of window: send ACK and drop. */ 722 if (!(flg & TCP_FLAG_RST) && 723 !tcp_oow_rate_limited(sock_net(sk), skb, 724 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 725 &tcp_rsk(req)->last_oow_ack_time)) 726 req->rsk_ops->send_ack(sk, skb, req); 727 if (paws_reject) 728 __NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 729 return NULL; 730 } 731 732 /* In sequence, PAWS is OK. */ 733 734 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt)) 735 req->ts_recent = tmp_opt.rcv_tsval; 736 737 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 738 /* Truncate SYN, it is out of window starting 739 at tcp_rsk(req)->rcv_isn + 1. */ 740 flg &= ~TCP_FLAG_SYN; 741 } 742 743 /* RFC793: "second check the RST bit" and 744 * "fourth, check the SYN bit" 745 */ 746 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 747 __TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 748 goto embryonic_reset; 749 } 750 751 /* ACK sequence verified above, just make sure ACK is 752 * set. If ACK not set, just silently drop the packet. 753 * 754 * XXX (TFO) - if we ever allow "data after SYN", the 755 * following check needs to be removed. 756 */ 757 if (!(flg & TCP_FLAG_ACK)) 758 return NULL; 759 760 /* For Fast Open no more processing is needed (sk is the 761 * child socket). 762 */ 763 if (fastopen) 764 return sk; 765 766 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 767 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 768 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 769 inet_rsk(req)->acked = 1; 770 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 771 return NULL; 772 } 773 774 /* OK, ACK is valid, create big socket and 775 * feed this segment to it. It will repeat all 776 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 777 * ESTABLISHED STATE. If it will be dropped after 778 * socket is created, wait for troubles. 779 */ 780 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL, 781 req, &own_req); 782 if (!child) 783 goto listen_overflow; 784 785 sock_rps_save_rxhash(child, skb); 786 tcp_synack_rtt_meas(child, req); 787 return inet_csk_complete_hashdance(sk, child, req, own_req); 788 789 listen_overflow: 790 if (!sock_net(sk)->ipv4.sysctl_tcp_abort_on_overflow) { 791 inet_rsk(req)->acked = 1; 792 return NULL; 793 } 794 795 embryonic_reset: 796 if (!(flg & TCP_FLAG_RST)) { 797 /* Received a bad SYN pkt - for TFO We try not to reset 798 * the local connection unless it's really necessary to 799 * avoid becoming vulnerable to outside attack aiming at 800 * resetting legit local connections. 801 */ 802 req->rsk_ops->send_reset(sk, skb); 803 } else if (fastopen) { /* received a valid RST pkt */ 804 reqsk_fastopen_remove(sk, req, true); 805 tcp_reset(sk); 806 } 807 if (!fastopen) { 808 inet_csk_reqsk_queue_drop(sk, req); 809 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 810 } 811 return NULL; 812 } 813 EXPORT_SYMBOL(tcp_check_req); 814 815 /* 816 * Queue segment on the new socket if the new socket is active, 817 * otherwise we just shortcircuit this and continue with 818 * the new socket. 819 * 820 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV 821 * when entering. But other states are possible due to a race condition 822 * where after __inet_lookup_established() fails but before the listener 823 * locked is obtained, other packets cause the same connection to 824 * be created. 825 */ 826 827 int tcp_child_process(struct sock *parent, struct sock *child, 828 struct sk_buff *skb) 829 { 830 int ret = 0; 831 int state = child->sk_state; 832 833 /* record NAPI ID of child */ 834 sk_mark_napi_id(child, skb); 835 836 tcp_segs_in(tcp_sk(child), skb); 837 if (!sock_owned_by_user(child)) { 838 ret = tcp_rcv_state_process(child, skb); 839 /* Wakeup parent, send SIGIO */ 840 if (state == TCP_SYN_RECV && child->sk_state != state) 841 parent->sk_data_ready(parent); 842 } else { 843 /* Alas, it is possible again, because we do lookup 844 * in main socket hash table and lock on listening 845 * socket does not protect us more. 846 */ 847 __sk_add_backlog(child, skb); 848 } 849 850 bh_unlock_sock(child); 851 sock_put(child); 852 return ret; 853 } 854 EXPORT_SYMBOL(tcp_child_process); 855