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 <net/tcp.h> 27 #include <net/inet_common.h> 28 #include <net/xfrm.h> 29 30 int sysctl_tcp_abort_on_overflow __read_mostly; 31 32 struct inet_timewait_death_row tcp_death_row = { 33 .sysctl_max_tw_buckets = NR_FILE * 2, 34 .hashinfo = &tcp_hashinfo, 35 }; 36 EXPORT_SYMBOL_GPL(tcp_death_row); 37 38 static bool tcp_in_window(u32 seq, u32 end_seq, u32 s_win, u32 e_win) 39 { 40 if (seq == s_win) 41 return true; 42 if (after(end_seq, s_win) && before(seq, e_win)) 43 return true; 44 return seq == e_win && seq == end_seq; 45 } 46 47 static enum tcp_tw_status 48 tcp_timewait_check_oow_rate_limit(struct inet_timewait_sock *tw, 49 const struct sk_buff *skb, int mib_idx) 50 { 51 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 52 53 if (!tcp_oow_rate_limited(twsk_net(tw), skb, mib_idx, 54 &tcptw->tw_last_oow_ack_time)) { 55 /* Send ACK. Note, we do not put the bucket, 56 * it will be released by caller. 57 */ 58 return TCP_TW_ACK; 59 } 60 61 /* We are rate-limiting, so just release the tw sock and drop skb. */ 62 inet_twsk_put(tw); 63 return TCP_TW_SUCCESS; 64 } 65 66 /* 67 * * Main purpose of TIME-WAIT state is to close connection gracefully, 68 * when one of ends sits in LAST-ACK or CLOSING retransmitting FIN 69 * (and, probably, tail of data) and one or more our ACKs are lost. 70 * * What is TIME-WAIT timeout? It is associated with maximal packet 71 * lifetime in the internet, which results in wrong conclusion, that 72 * it is set to catch "old duplicate segments" wandering out of their path. 73 * It is not quite correct. This timeout is calculated so that it exceeds 74 * maximal retransmission timeout enough to allow to lose one (or more) 75 * segments sent by peer and our ACKs. This time may be calculated from RTO. 76 * * When TIME-WAIT socket receives RST, it means that another end 77 * finally closed and we are allowed to kill TIME-WAIT too. 78 * * Second purpose of TIME-WAIT is catching old duplicate segments. 79 * Well, certainly it is pure paranoia, but if we load TIME-WAIT 80 * with this semantics, we MUST NOT kill TIME-WAIT state with RSTs. 81 * * If we invented some more clever way to catch duplicates 82 * (f.e. based on PAWS), we could truncate TIME-WAIT to several RTOs. 83 * 84 * The algorithm below is based on FORMAL INTERPRETATION of RFCs. 85 * When you compare it to RFCs, please, read section SEGMENT ARRIVES 86 * from the very beginning. 87 * 88 * NOTE. With recycling (and later with fin-wait-2) TW bucket 89 * is _not_ stateless. It means, that strictly speaking we must 90 * spinlock it. I do not want! Well, probability of misbehaviour 91 * is ridiculously low and, seems, we could use some mb() tricks 92 * to avoid misread sequence numbers, states etc. --ANK 93 * 94 * We don't need to initialize tmp_out.sack_ok as we don't use the results 95 */ 96 enum tcp_tw_status 97 tcp_timewait_state_process(struct inet_timewait_sock *tw, struct sk_buff *skb, 98 const struct tcphdr *th) 99 { 100 struct tcp_options_received tmp_opt; 101 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 102 bool paws_reject = false; 103 104 tmp_opt.saw_tstamp = 0; 105 if (th->doff > (sizeof(*th) >> 2) && tcptw->tw_ts_recent_stamp) { 106 tcp_parse_options(skb, &tmp_opt, 0, NULL); 107 108 if (tmp_opt.saw_tstamp) { 109 tmp_opt.rcv_tsecr -= tcptw->tw_ts_offset; 110 tmp_opt.ts_recent = tcptw->tw_ts_recent; 111 tmp_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp; 112 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 113 } 114 } 115 116 if (tw->tw_substate == TCP_FIN_WAIT2) { 117 /* Just repeat all the checks of tcp_rcv_state_process() */ 118 119 /* Out of window, send ACK */ 120 if (paws_reject || 121 !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 122 tcptw->tw_rcv_nxt, 123 tcptw->tw_rcv_nxt + tcptw->tw_rcv_wnd)) 124 return tcp_timewait_check_oow_rate_limit( 125 tw, skb, LINUX_MIB_TCPACKSKIPPEDFINWAIT2); 126 127 if (th->rst) 128 goto kill; 129 130 if (th->syn && !before(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt)) 131 return TCP_TW_RST; 132 133 /* Dup ACK? */ 134 if (!th->ack || 135 !after(TCP_SKB_CB(skb)->end_seq, tcptw->tw_rcv_nxt) || 136 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) { 137 inet_twsk_put(tw); 138 return TCP_TW_SUCCESS; 139 } 140 141 /* New data or FIN. If new data arrive after half-duplex close, 142 * reset. 143 */ 144 if (!th->fin || 145 TCP_SKB_CB(skb)->end_seq != tcptw->tw_rcv_nxt + 1) 146 return TCP_TW_RST; 147 148 /* FIN arrived, enter true time-wait state. */ 149 tw->tw_substate = TCP_TIME_WAIT; 150 tcptw->tw_rcv_nxt = TCP_SKB_CB(skb)->end_seq; 151 if (tmp_opt.saw_tstamp) { 152 tcptw->tw_ts_recent_stamp = get_seconds(); 153 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 154 } 155 156 if (tcp_death_row.sysctl_tw_recycle && 157 tcptw->tw_ts_recent_stamp && 158 tcp_tw_remember_stamp(tw)) 159 inet_twsk_reschedule(tw, tw->tw_timeout); 160 else 161 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 162 return TCP_TW_ACK; 163 } 164 165 /* 166 * Now real TIME-WAIT state. 167 * 168 * RFC 1122: 169 * "When a connection is [...] on TIME-WAIT state [...] 170 * [a TCP] MAY accept a new SYN from the remote TCP to 171 * reopen the connection directly, if it: 172 * 173 * (1) assigns its initial sequence number for the new 174 * connection to be larger than the largest sequence 175 * number it used on the previous connection incarnation, 176 * and 177 * 178 * (2) returns to TIME-WAIT state if the SYN turns out 179 * to be an old duplicate". 180 */ 181 182 if (!paws_reject && 183 (TCP_SKB_CB(skb)->seq == tcptw->tw_rcv_nxt && 184 (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq || th->rst))) { 185 /* In window segment, it may be only reset or bare ack. */ 186 187 if (th->rst) { 188 /* This is TIME_WAIT assassination, in two flavors. 189 * Oh well... nobody has a sufficient solution to this 190 * protocol bug yet. 191 */ 192 if (sysctl_tcp_rfc1337 == 0) { 193 kill: 194 inet_twsk_deschedule_put(tw); 195 return TCP_TW_SUCCESS; 196 } 197 } 198 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 199 200 if (tmp_opt.saw_tstamp) { 201 tcptw->tw_ts_recent = tmp_opt.rcv_tsval; 202 tcptw->tw_ts_recent_stamp = get_seconds(); 203 } 204 205 inet_twsk_put(tw); 206 return TCP_TW_SUCCESS; 207 } 208 209 /* Out of window segment. 210 211 All the segments are ACKed immediately. 212 213 The only exception is new SYN. We accept it, if it is 214 not old duplicate and we are not in danger to be killed 215 by delayed old duplicates. RFC check is that it has 216 newer sequence number works at rates <40Mbit/sec. 217 However, if paws works, it is reliable AND even more, 218 we even may relax silly seq space cutoff. 219 220 RED-PEN: we violate main RFC requirement, if this SYN will appear 221 old duplicate (i.e. we receive RST in reply to SYN-ACK), 222 we must return socket to time-wait state. It is not good, 223 but not fatal yet. 224 */ 225 226 if (th->syn && !th->rst && !th->ack && !paws_reject && 227 (after(TCP_SKB_CB(skb)->seq, tcptw->tw_rcv_nxt) || 228 (tmp_opt.saw_tstamp && 229 (s32)(tcptw->tw_ts_recent - tmp_opt.rcv_tsval) < 0))) { 230 u32 isn = tcptw->tw_snd_nxt + 65535 + 2; 231 if (isn == 0) 232 isn++; 233 TCP_SKB_CB(skb)->tcp_tw_isn = isn; 234 return TCP_TW_SYN; 235 } 236 237 if (paws_reject) 238 __NET_INC_STATS(twsk_net(tw), LINUX_MIB_PAWSESTABREJECTED); 239 240 if (!th->rst) { 241 /* In this case we must reset the TIMEWAIT timer. 242 * 243 * If it is ACKless SYN it may be both old duplicate 244 * and new good SYN with random sequence number <rcv_nxt. 245 * Do not reschedule in the last case. 246 */ 247 if (paws_reject || th->ack) 248 inet_twsk_reschedule(tw, TCP_TIMEWAIT_LEN); 249 250 return tcp_timewait_check_oow_rate_limit( 251 tw, skb, LINUX_MIB_TCPACKSKIPPEDTIMEWAIT); 252 } 253 inet_twsk_put(tw); 254 return TCP_TW_SUCCESS; 255 } 256 EXPORT_SYMBOL(tcp_timewait_state_process); 257 258 /* 259 * Move a socket to time-wait or dead fin-wait-2 state. 260 */ 261 void tcp_time_wait(struct sock *sk, int state, int timeo) 262 { 263 const struct inet_connection_sock *icsk = inet_csk(sk); 264 const struct tcp_sock *tp = tcp_sk(sk); 265 struct inet_timewait_sock *tw; 266 bool recycle_ok = false; 267 268 if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp) 269 recycle_ok = tcp_remember_stamp(sk); 270 271 tw = inet_twsk_alloc(sk, &tcp_death_row, state); 272 273 if (tw) { 274 struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw); 275 const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1); 276 struct inet_sock *inet = inet_sk(sk); 277 278 tw->tw_transparent = inet->transparent; 279 tw->tw_rcv_wscale = tp->rx_opt.rcv_wscale; 280 tcptw->tw_rcv_nxt = tp->rcv_nxt; 281 tcptw->tw_snd_nxt = tp->snd_nxt; 282 tcptw->tw_rcv_wnd = tcp_receive_window(tp); 283 tcptw->tw_ts_recent = tp->rx_opt.ts_recent; 284 tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp; 285 tcptw->tw_ts_offset = tp->tsoffset; 286 tcptw->tw_last_oow_ack_time = 0; 287 288 #if IS_ENABLED(CONFIG_IPV6) 289 if (tw->tw_family == PF_INET6) { 290 struct ipv6_pinfo *np = inet6_sk(sk); 291 292 tw->tw_v6_daddr = sk->sk_v6_daddr; 293 tw->tw_v6_rcv_saddr = sk->sk_v6_rcv_saddr; 294 tw->tw_tclass = np->tclass; 295 tw->tw_flowlabel = be32_to_cpu(np->flow_label & IPV6_FLOWLABEL_MASK); 296 tw->tw_ipv6only = sk->sk_ipv6only; 297 } 298 #endif 299 300 #ifdef CONFIG_TCP_MD5SIG 301 /* 302 * The timewait bucket does not have the key DB from the 303 * sock structure. We just make a quick copy of the 304 * md5 key being used (if indeed we are using one) 305 * so the timewait ack generating code has the key. 306 */ 307 do { 308 struct tcp_md5sig_key *key; 309 tcptw->tw_md5_key = NULL; 310 key = tp->af_specific->md5_lookup(sk, sk); 311 if (key) { 312 tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC); 313 if (tcptw->tw_md5_key && !tcp_alloc_md5sig_pool()) 314 BUG(); 315 } 316 } while (0); 317 #endif 318 319 /* Get the TIME_WAIT timeout firing. */ 320 if (timeo < rto) 321 timeo = rto; 322 323 if (recycle_ok) { 324 tw->tw_timeout = rto; 325 } else { 326 tw->tw_timeout = TCP_TIMEWAIT_LEN; 327 if (state == TCP_TIME_WAIT) 328 timeo = TCP_TIMEWAIT_LEN; 329 } 330 331 inet_twsk_schedule(tw, timeo); 332 /* Linkage updates. */ 333 __inet_twsk_hashdance(tw, sk, &tcp_hashinfo); 334 inet_twsk_put(tw); 335 } else { 336 /* Sorry, if we're out of memory, just CLOSE this 337 * socket up. We've got bigger problems than 338 * non-graceful socket closings. 339 */ 340 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPTIMEWAITOVERFLOW); 341 } 342 343 tcp_update_metrics(sk); 344 tcp_done(sk); 345 } 346 347 void tcp_twsk_destructor(struct sock *sk) 348 { 349 #ifdef CONFIG_TCP_MD5SIG 350 struct tcp_timewait_sock *twsk = tcp_twsk(sk); 351 352 if (twsk->tw_md5_key) 353 kfree_rcu(twsk->tw_md5_key, rcu); 354 #endif 355 } 356 EXPORT_SYMBOL_GPL(tcp_twsk_destructor); 357 358 /* Warning : This function is called without sk_listener being locked. 359 * Be sure to read socket fields once, as their value could change under us. 360 */ 361 void tcp_openreq_init_rwin(struct request_sock *req, 362 const struct sock *sk_listener, 363 const struct dst_entry *dst) 364 { 365 struct inet_request_sock *ireq = inet_rsk(req); 366 const struct tcp_sock *tp = tcp_sk(sk_listener); 367 u16 user_mss = READ_ONCE(tp->rx_opt.user_mss); 368 int full_space = tcp_full_space(sk_listener); 369 int mss = dst_metric_advmss(dst); 370 u32 window_clamp; 371 __u8 rcv_wscale; 372 373 if (user_mss && user_mss < mss) 374 mss = user_mss; 375 376 window_clamp = READ_ONCE(tp->window_clamp); 377 /* Set this up on the first call only */ 378 req->rsk_window_clamp = window_clamp ? : dst_metric(dst, RTAX_WINDOW); 379 380 /* limit the window selection if the user enforce a smaller rx buffer */ 381 if (sk_listener->sk_userlocks & SOCK_RCVBUF_LOCK && 382 (req->rsk_window_clamp > full_space || req->rsk_window_clamp == 0)) 383 req->rsk_window_clamp = full_space; 384 385 /* tcp_full_space because it is guaranteed to be the first packet */ 386 tcp_select_initial_window(full_space, 387 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 388 &req->rsk_rcv_wnd, 389 &req->rsk_window_clamp, 390 ireq->wscale_ok, 391 &rcv_wscale, 392 dst_metric(dst, RTAX_INITRWND)); 393 ireq->rcv_wscale = rcv_wscale; 394 } 395 EXPORT_SYMBOL(tcp_openreq_init_rwin); 396 397 static void tcp_ecn_openreq_child(struct tcp_sock *tp, 398 const struct request_sock *req) 399 { 400 tp->ecn_flags = inet_rsk(req)->ecn_ok ? TCP_ECN_OK : 0; 401 } 402 403 void tcp_ca_openreq_child(struct sock *sk, const struct dst_entry *dst) 404 { 405 struct inet_connection_sock *icsk = inet_csk(sk); 406 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 407 bool ca_got_dst = false; 408 409 if (ca_key != TCP_CA_UNSPEC) { 410 const struct tcp_congestion_ops *ca; 411 412 rcu_read_lock(); 413 ca = tcp_ca_find_key(ca_key); 414 if (likely(ca && try_module_get(ca->owner))) { 415 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 416 icsk->icsk_ca_ops = ca; 417 ca_got_dst = true; 418 } 419 rcu_read_unlock(); 420 } 421 422 /* If no valid choice made yet, assign current system default ca. */ 423 if (!ca_got_dst && 424 (!icsk->icsk_ca_setsockopt || 425 !try_module_get(icsk->icsk_ca_ops->owner))) 426 tcp_assign_congestion_control(sk); 427 428 tcp_set_ca_state(sk, TCP_CA_Open); 429 } 430 EXPORT_SYMBOL_GPL(tcp_ca_openreq_child); 431 432 /* This is not only more efficient than what we used to do, it eliminates 433 * a lot of code duplication between IPv4/IPv6 SYN recv processing. -DaveM 434 * 435 * Actually, we could lots of memory writes here. tp of listening 436 * socket contains all necessary default parameters. 437 */ 438 struct sock *tcp_create_openreq_child(const struct sock *sk, 439 struct request_sock *req, 440 struct sk_buff *skb) 441 { 442 struct sock *newsk = inet_csk_clone_lock(sk, req, GFP_ATOMIC); 443 444 if (newsk) { 445 const struct inet_request_sock *ireq = inet_rsk(req); 446 struct tcp_request_sock *treq = tcp_rsk(req); 447 struct inet_connection_sock *newicsk = inet_csk(newsk); 448 struct tcp_sock *newtp = tcp_sk(newsk); 449 450 /* Now setup tcp_sock */ 451 newtp->pred_flags = 0; 452 453 newtp->rcv_wup = newtp->copied_seq = 454 newtp->rcv_nxt = treq->rcv_isn + 1; 455 newtp->segs_in = 1; 456 457 newtp->snd_sml = newtp->snd_una = 458 newtp->snd_nxt = newtp->snd_up = treq->snt_isn + 1; 459 460 tcp_prequeue_init(newtp); 461 INIT_LIST_HEAD(&newtp->tsq_node); 462 463 tcp_init_wl(newtp, treq->rcv_isn); 464 465 newtp->srtt_us = 0; 466 newtp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT); 467 minmax_reset(&newtp->rtt_min, tcp_time_stamp, ~0U); 468 newicsk->icsk_rto = TCP_TIMEOUT_INIT; 469 470 newtp->packets_out = 0; 471 newtp->retrans_out = 0; 472 newtp->sacked_out = 0; 473 newtp->fackets_out = 0; 474 newtp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 475 tcp_enable_early_retrans(newtp); 476 newtp->tlp_high_seq = 0; 477 newtp->lsndtime = treq->snt_synack.stamp_jiffies; 478 newsk->sk_txhash = treq->txhash; 479 newtp->last_oow_ack_time = 0; 480 newtp->total_retrans = req->num_retrans; 481 482 /* So many TCP implementations out there (incorrectly) count the 483 * initial SYN frame in their delayed-ACK and congestion control 484 * algorithms that we must have the following bandaid to talk 485 * efficiently to them. -DaveM 486 */ 487 newtp->snd_cwnd = TCP_INIT_CWND; 488 newtp->snd_cwnd_cnt = 0; 489 490 /* There's a bubble in the pipe until at least the first ACK. */ 491 newtp->app_limited = ~0U; 492 493 tcp_init_xmit_timers(newsk); 494 newtp->write_seq = newtp->pushed_seq = treq->snt_isn + 1; 495 496 newtp->rx_opt.saw_tstamp = 0; 497 498 newtp->rx_opt.dsack = 0; 499 newtp->rx_opt.num_sacks = 0; 500 501 newtp->urg_data = 0; 502 503 if (sock_flag(newsk, SOCK_KEEPOPEN)) 504 inet_csk_reset_keepalive_timer(newsk, 505 keepalive_time_when(newtp)); 506 507 newtp->rx_opt.tstamp_ok = ireq->tstamp_ok; 508 if ((newtp->rx_opt.sack_ok = ireq->sack_ok) != 0) { 509 if (sysctl_tcp_fack) 510 tcp_enable_fack(newtp); 511 } 512 newtp->window_clamp = req->rsk_window_clamp; 513 newtp->rcv_ssthresh = req->rsk_rcv_wnd; 514 newtp->rcv_wnd = req->rsk_rcv_wnd; 515 newtp->rx_opt.wscale_ok = ireq->wscale_ok; 516 if (newtp->rx_opt.wscale_ok) { 517 newtp->rx_opt.snd_wscale = ireq->snd_wscale; 518 newtp->rx_opt.rcv_wscale = ireq->rcv_wscale; 519 } else { 520 newtp->rx_opt.snd_wscale = newtp->rx_opt.rcv_wscale = 0; 521 newtp->window_clamp = min(newtp->window_clamp, 65535U); 522 } 523 newtp->snd_wnd = (ntohs(tcp_hdr(skb)->window) << 524 newtp->rx_opt.snd_wscale); 525 newtp->max_window = newtp->snd_wnd; 526 527 if (newtp->rx_opt.tstamp_ok) { 528 newtp->rx_opt.ts_recent = req->ts_recent; 529 newtp->rx_opt.ts_recent_stamp = get_seconds(); 530 newtp->tcp_header_len = sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 531 } else { 532 newtp->rx_opt.ts_recent_stamp = 0; 533 newtp->tcp_header_len = sizeof(struct tcphdr); 534 } 535 newtp->tsoffset = treq->ts_off; 536 #ifdef CONFIG_TCP_MD5SIG 537 newtp->md5sig_info = NULL; /*XXX*/ 538 if (newtp->af_specific->md5_lookup(sk, newsk)) 539 newtp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 540 #endif 541 if (skb->len >= TCP_MSS_DEFAULT + newtp->tcp_header_len) 542 newicsk->icsk_ack.last_seg_size = skb->len - newtp->tcp_header_len; 543 newtp->rx_opt.mss_clamp = req->mss; 544 tcp_ecn_openreq_child(newtp, req); 545 newtp->fastopen_rsk = NULL; 546 newtp->syn_data_acked = 0; 547 newtp->rack.mstamp.v64 = 0; 548 newtp->rack.advanced = 0; 549 550 __TCP_INC_STATS(sock_net(sk), TCP_MIB_PASSIVEOPENS); 551 } 552 return newsk; 553 } 554 EXPORT_SYMBOL(tcp_create_openreq_child); 555 556 /* 557 * Process an incoming packet for SYN_RECV sockets represented as a 558 * request_sock. Normally sk is the listener socket but for TFO it 559 * points to the child socket. 560 * 561 * XXX (TFO) - The current impl contains a special check for ack 562 * validation and inside tcp_v4_reqsk_send_ack(). Can we do better? 563 * 564 * We don't need to initialize tmp_opt.sack_ok as we don't use the results 565 */ 566 567 struct sock *tcp_check_req(struct sock *sk, struct sk_buff *skb, 568 struct request_sock *req, 569 bool fastopen) 570 { 571 struct tcp_options_received tmp_opt; 572 struct sock *child; 573 const struct tcphdr *th = tcp_hdr(skb); 574 __be32 flg = tcp_flag_word(th) & (TCP_FLAG_RST|TCP_FLAG_SYN|TCP_FLAG_ACK); 575 bool paws_reject = false; 576 bool own_req; 577 578 tmp_opt.saw_tstamp = 0; 579 if (th->doff > (sizeof(struct tcphdr)>>2)) { 580 tcp_parse_options(skb, &tmp_opt, 0, NULL); 581 582 if (tmp_opt.saw_tstamp) { 583 tmp_opt.ts_recent = req->ts_recent; 584 if (tmp_opt.rcv_tsecr) 585 tmp_opt.rcv_tsecr -= tcp_rsk(req)->ts_off; 586 /* We do not store true stamp, but it is not required, 587 * it can be estimated (approximately) 588 * from another data. 589 */ 590 tmp_opt.ts_recent_stamp = get_seconds() - ((TCP_TIMEOUT_INIT/HZ)<<req->num_timeout); 591 paws_reject = tcp_paws_reject(&tmp_opt, th->rst); 592 } 593 } 594 595 /* Check for pure retransmitted SYN. */ 596 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn && 597 flg == TCP_FLAG_SYN && 598 !paws_reject) { 599 /* 600 * RFC793 draws (Incorrectly! It was fixed in RFC1122) 601 * this case on figure 6 and figure 8, but formal 602 * protocol description says NOTHING. 603 * To be more exact, it says that we should send ACK, 604 * because this segment (at least, if it has no data) 605 * is out of window. 606 * 607 * CONCLUSION: RFC793 (even with RFC1122) DOES NOT 608 * describe SYN-RECV state. All the description 609 * is wrong, we cannot believe to it and should 610 * rely only on common sense and implementation 611 * experience. 612 * 613 * Enforce "SYN-ACK" according to figure 8, figure 6 614 * of RFC793, fixed by RFC1122. 615 * 616 * Note that even if there is new data in the SYN packet 617 * they will be thrown away too. 618 * 619 * Reset timer after retransmitting SYNACK, similar to 620 * the idea of fast retransmit in recovery. 621 */ 622 if (!tcp_oow_rate_limited(sock_net(sk), skb, 623 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 624 &tcp_rsk(req)->last_oow_ack_time) && 625 626 !inet_rtx_syn_ack(sk, req)) { 627 unsigned long expires = jiffies; 628 629 expires += min(TCP_TIMEOUT_INIT << req->num_timeout, 630 TCP_RTO_MAX); 631 if (!fastopen) 632 mod_timer_pending(&req->rsk_timer, expires); 633 else 634 req->rsk_timer.expires = expires; 635 } 636 return NULL; 637 } 638 639 /* Further reproduces section "SEGMENT ARRIVES" 640 for state SYN-RECEIVED of RFC793. 641 It is broken, however, it does not work only 642 when SYNs are crossed. 643 644 You would think that SYN crossing is impossible here, since 645 we should have a SYN_SENT socket (from connect()) on our end, 646 but this is not true if the crossed SYNs were sent to both 647 ends by a malicious third party. We must defend against this, 648 and to do that we first verify the ACK (as per RFC793, page 649 36) and reset if it is invalid. Is this a true full defense? 650 To convince ourselves, let us consider a way in which the ACK 651 test can still pass in this 'malicious crossed SYNs' case. 652 Malicious sender sends identical SYNs (and thus identical sequence 653 numbers) to both A and B: 654 655 A: gets SYN, seq=7 656 B: gets SYN, seq=7 657 658 By our good fortune, both A and B select the same initial 659 send sequence number of seven :-) 660 661 A: sends SYN|ACK, seq=7, ack_seq=8 662 B: sends SYN|ACK, seq=7, ack_seq=8 663 664 So we are now A eating this SYN|ACK, ACK test passes. So 665 does sequence test, SYN is truncated, and thus we consider 666 it a bare ACK. 667 668 If icsk->icsk_accept_queue.rskq_defer_accept, we silently drop this 669 bare ACK. Otherwise, we create an established connection. Both 670 ends (listening sockets) accept the new incoming connection and try 671 to talk to each other. 8-) 672 673 Note: This case is both harmless, and rare. Possibility is about the 674 same as us discovering intelligent life on another plant tomorrow. 675 676 But generally, we should (RFC lies!) to accept ACK 677 from SYNACK both here and in tcp_rcv_state_process(). 678 tcp_rcv_state_process() does not, hence, we do not too. 679 680 Note that the case is absolutely generic: 681 we cannot optimize anything here without 682 violating protocol. All the checks must be made 683 before attempt to create socket. 684 */ 685 686 /* RFC793 page 36: "If the connection is in any non-synchronized state ... 687 * and the incoming segment acknowledges something not yet 688 * sent (the segment carries an unacceptable ACK) ... 689 * a reset is sent." 690 * 691 * Invalid ACK: reset will be sent by listening socket. 692 * Note that the ACK validity check for a Fast Open socket is done 693 * elsewhere and is checked directly against the child socket rather 694 * than req because user data may have been sent out. 695 */ 696 if ((flg & TCP_FLAG_ACK) && !fastopen && 697 (TCP_SKB_CB(skb)->ack_seq != 698 tcp_rsk(req)->snt_isn + 1)) 699 return sk; 700 701 /* Also, it would be not so bad idea to check rcv_tsecr, which 702 * is essentially ACK extension and too early or too late values 703 * should cause reset in unsynchronized states. 704 */ 705 706 /* RFC793: "first check sequence number". */ 707 708 if (paws_reject || !tcp_in_window(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, 709 tcp_rsk(req)->rcv_nxt, tcp_rsk(req)->rcv_nxt + req->rsk_rcv_wnd)) { 710 /* Out of window: send ACK and drop. */ 711 if (!(flg & TCP_FLAG_RST) && 712 !tcp_oow_rate_limited(sock_net(sk), skb, 713 LINUX_MIB_TCPACKSKIPPEDSYNRECV, 714 &tcp_rsk(req)->last_oow_ack_time)) 715 req->rsk_ops->send_ack(sk, skb, req); 716 if (paws_reject) 717 __NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 718 return NULL; 719 } 720 721 /* In sequence, PAWS is OK. */ 722 723 if (tmp_opt.saw_tstamp && !after(TCP_SKB_CB(skb)->seq, tcp_rsk(req)->rcv_nxt)) 724 req->ts_recent = tmp_opt.rcv_tsval; 725 726 if (TCP_SKB_CB(skb)->seq == tcp_rsk(req)->rcv_isn) { 727 /* Truncate SYN, it is out of window starting 728 at tcp_rsk(req)->rcv_isn + 1. */ 729 flg &= ~TCP_FLAG_SYN; 730 } 731 732 /* RFC793: "second check the RST bit" and 733 * "fourth, check the SYN bit" 734 */ 735 if (flg & (TCP_FLAG_RST|TCP_FLAG_SYN)) { 736 __TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS); 737 goto embryonic_reset; 738 } 739 740 /* ACK sequence verified above, just make sure ACK is 741 * set. If ACK not set, just silently drop the packet. 742 * 743 * XXX (TFO) - if we ever allow "data after SYN", the 744 * following check needs to be removed. 745 */ 746 if (!(flg & TCP_FLAG_ACK)) 747 return NULL; 748 749 /* For Fast Open no more processing is needed (sk is the 750 * child socket). 751 */ 752 if (fastopen) 753 return sk; 754 755 /* While TCP_DEFER_ACCEPT is active, drop bare ACK. */ 756 if (req->num_timeout < inet_csk(sk)->icsk_accept_queue.rskq_defer_accept && 757 TCP_SKB_CB(skb)->end_seq == tcp_rsk(req)->rcv_isn + 1) { 758 inet_rsk(req)->acked = 1; 759 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDEFERACCEPTDROP); 760 return NULL; 761 } 762 763 /* OK, ACK is valid, create big socket and 764 * feed this segment to it. It will repeat all 765 * the tests. THIS SEGMENT MUST MOVE SOCKET TO 766 * ESTABLISHED STATE. If it will be dropped after 767 * socket is created, wait for troubles. 768 */ 769 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL, 770 req, &own_req); 771 if (!child) 772 goto listen_overflow; 773 774 sock_rps_save_rxhash(child, skb); 775 tcp_synack_rtt_meas(child, req); 776 return inet_csk_complete_hashdance(sk, child, req, own_req); 777 778 listen_overflow: 779 if (!sysctl_tcp_abort_on_overflow) { 780 inet_rsk(req)->acked = 1; 781 return NULL; 782 } 783 784 embryonic_reset: 785 if (!(flg & TCP_FLAG_RST)) { 786 /* Received a bad SYN pkt - for TFO We try not to reset 787 * the local connection unless it's really necessary to 788 * avoid becoming vulnerable to outside attack aiming at 789 * resetting legit local connections. 790 */ 791 req->rsk_ops->send_reset(sk, skb); 792 } else if (fastopen) { /* received a valid RST pkt */ 793 reqsk_fastopen_remove(sk, req, true); 794 tcp_reset(sk); 795 } 796 if (!fastopen) { 797 inet_csk_reqsk_queue_drop(sk, req); 798 __NET_INC_STATS(sock_net(sk), LINUX_MIB_EMBRYONICRSTS); 799 } 800 return NULL; 801 } 802 EXPORT_SYMBOL(tcp_check_req); 803 804 /* 805 * Queue segment on the new socket if the new socket is active, 806 * otherwise we just shortcircuit this and continue with 807 * the new socket. 808 * 809 * For the vast majority of cases child->sk_state will be TCP_SYN_RECV 810 * when entering. But other states are possible due to a race condition 811 * where after __inet_lookup_established() fails but before the listener 812 * locked is obtained, other packets cause the same connection to 813 * be created. 814 */ 815 816 int tcp_child_process(struct sock *parent, struct sock *child, 817 struct sk_buff *skb) 818 { 819 int ret = 0; 820 int state = child->sk_state; 821 822 tcp_segs_in(tcp_sk(child), skb); 823 if (!sock_owned_by_user(child)) { 824 ret = tcp_rcv_state_process(child, skb); 825 /* Wakeup parent, send SIGIO */ 826 if (state == TCP_SYN_RECV && child->sk_state != state) 827 parent->sk_data_ready(parent); 828 } else { 829 /* Alas, it is possible again, because we do lookup 830 * in main socket hash table and lock on listening 831 * socket does not protect us more. 832 */ 833 __sk_add_backlog(child, skb); 834 } 835 836 bh_unlock_sock(child); 837 sock_put(child); 838 return ret; 839 } 840 EXPORT_SYMBOL(tcp_child_process); 841