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