1 // SPDX-License-Identifier: GPL-2.0 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 /* 23 * Changes: 24 * Pedro Roque : Fast Retransmit/Recovery. 25 * Two receive queues. 26 * Retransmit queue handled by TCP. 27 * Better retransmit timer handling. 28 * New congestion avoidance. 29 * Header prediction. 30 * Variable renaming. 31 * 32 * Eric : Fast Retransmit. 33 * Randy Scott : MSS option defines. 34 * Eric Schenk : Fixes to slow start algorithm. 35 * Eric Schenk : Yet another double ACK bug. 36 * Eric Schenk : Delayed ACK bug fixes. 37 * Eric Schenk : Floyd style fast retrans war avoidance. 38 * David S. Miller : Don't allow zero congestion window. 39 * Eric Schenk : Fix retransmitter so that it sends 40 * next packet on ack of previous packet. 41 * Andi Kleen : Moved open_request checking here 42 * and process RSTs for open_requests. 43 * Andi Kleen : Better prune_queue, and other fixes. 44 * Andrey Savochkin: Fix RTT measurements in the presence of 45 * timestamps. 46 * Andrey Savochkin: Check sequence numbers correctly when 47 * removing SACKs due to in sequence incoming 48 * data segments. 49 * Andi Kleen: Make sure we never ack data there is not 50 * enough room for. Also make this condition 51 * a fatal error if it might still happen. 52 * Andi Kleen: Add tcp_measure_rcv_mss to make 53 * connections with MSS<min(MTU,ann. MSS) 54 * work without delayed acks. 55 * Andi Kleen: Process packets with PSH set in the 56 * fast path. 57 * J Hadi Salim: ECN support 58 * Andrei Gurtov, 59 * Pasi Sarolahti, 60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission 61 * engine. Lots of bugs are found. 62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs 63 */ 64 65 #define pr_fmt(fmt) "TCP: " fmt 66 67 #include <linux/mm.h> 68 #include <linux/slab.h> 69 #include <linux/module.h> 70 #include <linux/sysctl.h> 71 #include <linux/kernel.h> 72 #include <linux/prefetch.h> 73 #include <linux/bitops.h> 74 #include <net/dst.h> 75 #include <net/tcp.h> 76 #include <net/tcp_ecn.h> 77 #include <net/proto_memory.h> 78 #include <net/inet_common.h> 79 #include <linux/ipsec.h> 80 #include <linux/unaligned.h> 81 #include <linux/errqueue.h> 82 #include <trace/events/tcp.h> 83 #include <linux/jump_label_ratelimit.h> 84 #include <net/busy_poll.h> 85 #include <net/mptcp.h> 86 87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 88 89 #define FLAG_DATA 0x01 /* Incoming frame contained data. */ 90 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 91 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 92 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 93 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 94 #define FLAG_DATA_SACKED 0x20 /* New SACK. */ 95 #define FLAG_ECE 0x40 /* ECE in this ACK */ 96 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */ 97 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 98 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ 99 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 100 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 101 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */ 102 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 103 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ 104 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */ 105 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */ 106 #define FLAG_DSACK_TLP 0x20000 /* DSACK for tail loss probe */ 107 #define FLAG_TS_PROGRESS 0x40000 /* Positive timestamp delta */ 108 109 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 110 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) 111 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK) 112 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 113 114 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 115 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 116 117 #define REXMIT_NONE 0 /* no loss recovery to do */ 118 #define REXMIT_LOST 1 /* retransmit packets marked lost */ 119 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */ 120 121 #if IS_ENABLED(CONFIG_TLS_DEVICE) 122 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ); 123 124 void clean_acked_data_enable(struct tcp_sock *tp, 125 void (*cad)(struct sock *sk, u32 ack_seq)) 126 { 127 tp->tcp_clean_acked = cad; 128 static_branch_deferred_inc(&clean_acked_data_enabled); 129 } 130 EXPORT_SYMBOL_GPL(clean_acked_data_enable); 131 132 void clean_acked_data_disable(struct tcp_sock *tp) 133 { 134 static_branch_slow_dec_deferred(&clean_acked_data_enabled); 135 tp->tcp_clean_acked = NULL; 136 } 137 EXPORT_SYMBOL_GPL(clean_acked_data_disable); 138 139 void clean_acked_data_flush(void) 140 { 141 static_key_deferred_flush(&clean_acked_data_enabled); 142 } 143 EXPORT_SYMBOL_GPL(clean_acked_data_flush); 144 #endif 145 146 #ifdef CONFIG_CGROUP_BPF 147 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 148 { 149 bool unknown_opt = tcp_sk(sk)->rx_opt.saw_unknown && 150 BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 151 BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG); 152 bool parse_all_opt = BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 153 BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG); 154 struct bpf_sock_ops_kern sock_ops; 155 156 if (likely(!unknown_opt && !parse_all_opt)) 157 return; 158 159 /* The skb will be handled in the 160 * bpf_skops_established() or 161 * bpf_skops_write_hdr_opt(). 162 */ 163 switch (sk->sk_state) { 164 case TCP_SYN_RECV: 165 case TCP_SYN_SENT: 166 case TCP_LISTEN: 167 return; 168 } 169 170 sock_owned_by_me(sk); 171 172 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 173 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB; 174 sock_ops.is_fullsock = 1; 175 sock_ops.is_locked_tcp_sock = 1; 176 sock_ops.sk = sk; 177 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb)); 178 179 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); 180 } 181 182 static void bpf_skops_established(struct sock *sk, int bpf_op, 183 struct sk_buff *skb) 184 { 185 struct bpf_sock_ops_kern sock_ops; 186 187 sock_owned_by_me(sk); 188 189 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 190 sock_ops.op = bpf_op; 191 sock_ops.is_fullsock = 1; 192 sock_ops.is_locked_tcp_sock = 1; 193 sock_ops.sk = sk; 194 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */ 195 if (skb) 196 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb)); 197 198 BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); 199 } 200 #else 201 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 202 { 203 } 204 205 static void bpf_skops_established(struct sock *sk, int bpf_op, 206 struct sk_buff *skb) 207 { 208 } 209 #endif 210 211 static __cold void tcp_gro_dev_warn(const struct sock *sk, const struct sk_buff *skb, 212 unsigned int len) 213 { 214 struct net_device *dev; 215 216 rcu_read_lock(); 217 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif); 218 if (!dev || len >= READ_ONCE(dev->mtu)) 219 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n", 220 dev ? dev->name : "Unknown driver"); 221 rcu_read_unlock(); 222 } 223 224 /* Adapt the MSS value used to make delayed ack decision to the 225 * real world. 226 */ 227 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 228 { 229 struct inet_connection_sock *icsk = inet_csk(sk); 230 const unsigned int lss = icsk->icsk_ack.last_seg_size; 231 unsigned int len; 232 233 icsk->icsk_ack.last_seg_size = 0; 234 235 /* skb->len may jitter because of SACKs, even if peer 236 * sends good full-sized frames. 237 */ 238 len = skb_shinfo(skb)->gso_size ? : skb->len; 239 if (len >= icsk->icsk_ack.rcv_mss) { 240 /* Note: divides are still a bit expensive. 241 * For the moment, only adjust scaling_ratio 242 * when we update icsk_ack.rcv_mss. 243 */ 244 if (unlikely(len != icsk->icsk_ack.rcv_mss)) { 245 u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE; 246 u8 old_ratio = tcp_sk(sk)->scaling_ratio; 247 248 do_div(val, skb->truesize); 249 tcp_sk(sk)->scaling_ratio = val ? val : 1; 250 251 if (old_ratio != tcp_sk(sk)->scaling_ratio) { 252 struct tcp_sock *tp = tcp_sk(sk); 253 254 val = tcp_win_from_space(sk, sk->sk_rcvbuf); 255 tcp_set_window_clamp(sk, val); 256 257 if (tp->window_clamp < tp->rcvq_space.space) 258 tp->rcvq_space.space = tp->window_clamp; 259 } 260 } 261 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len, 262 tcp_sk(sk)->advmss); 263 /* Account for possibly-removed options */ 264 DO_ONCE_LITE_IF(len > icsk->icsk_ack.rcv_mss + MAX_TCP_OPTION_SPACE, 265 tcp_gro_dev_warn, sk, skb, len); 266 /* If the skb has a len of exactly 1*MSS and has the PSH bit 267 * set then it is likely the end of an application write. So 268 * more data may not be arriving soon, and yet the data sender 269 * may be waiting for an ACK if cwnd-bound or using TX zero 270 * copy. So we set ICSK_ACK_PUSHED here so that 271 * tcp_cleanup_rbuf() will send an ACK immediately if the app 272 * reads all of the data and is not ping-pong. If len > MSS 273 * then this logic does not matter (and does not hurt) because 274 * tcp_cleanup_rbuf() will always ACK immediately if the app 275 * reads data and there is more than an MSS of unACKed data. 276 */ 277 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH) 278 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 279 } else { 280 /* Otherwise, we make more careful check taking into account, 281 * that SACKs block is variable. 282 * 283 * "len" is invariant segment length, including TCP header. 284 */ 285 len += skb->data - skb_transport_header(skb); 286 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 287 /* If PSH is not set, packet should be 288 * full sized, provided peer TCP is not badly broken. 289 * This observation (if it is correct 8)) allows 290 * to handle super-low mtu links fairly. 291 */ 292 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 293 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 294 /* Subtract also invariant (if peer is RFC compliant), 295 * tcp header plus fixed timestamp option length. 296 * Resulting "len" is MSS free of SACK jitter. 297 */ 298 len -= tcp_sk(sk)->tcp_header_len; 299 icsk->icsk_ack.last_seg_size = len; 300 if (len == lss) { 301 icsk->icsk_ack.rcv_mss = len; 302 return; 303 } 304 } 305 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 306 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 307 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 308 } 309 } 310 311 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks) 312 { 313 struct inet_connection_sock *icsk = inet_csk(sk); 314 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 315 316 if (quickacks == 0) 317 quickacks = 2; 318 quickacks = min(quickacks, max_quickacks); 319 if (quickacks > icsk->icsk_ack.quick) 320 icsk->icsk_ack.quick = quickacks; 321 } 322 323 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks) 324 { 325 struct inet_connection_sock *icsk = inet_csk(sk); 326 327 tcp_incr_quickack(sk, max_quickacks); 328 inet_csk_exit_pingpong_mode(sk); 329 icsk->icsk_ack.ato = TCP_ATO_MIN; 330 } 331 332 /* Send ACKs quickly, if "quick" count is not exhausted 333 * and the session is not interactive. 334 */ 335 336 static bool tcp_in_quickack_mode(struct sock *sk) 337 { 338 const struct inet_connection_sock *icsk = inet_csk(sk); 339 340 return icsk->icsk_ack.dst_quick_ack || 341 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk)); 342 } 343 344 static void tcp_data_ecn_check(struct sock *sk, const struct sk_buff *skb) 345 { 346 struct tcp_sock *tp = tcp_sk(sk); 347 348 if (tcp_ecn_disabled(tp)) 349 return; 350 351 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { 352 case INET_ECN_NOT_ECT: 353 /* Funny extension: if ECT is not set on a segment, 354 * and we already seen ECT on a previous segment, 355 * it is probably a retransmit. 356 */ 357 if (tp->ecn_flags & TCP_ECN_SEEN) 358 tcp_enter_quickack_mode(sk, 2); 359 break; 360 case INET_ECN_CE: 361 if (tcp_ca_needs_ecn(sk)) 362 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE); 363 364 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR) && 365 tcp_ecn_mode_rfc3168(tp)) { 366 /* Better not delay acks, sender can have a very low cwnd */ 367 tcp_enter_quickack_mode(sk, 2); 368 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 369 } 370 /* As for RFC3168 ECN, the TCP_ECN_SEEN flag is set by 371 * tcp_data_ecn_check() when the ECN codepoint of 372 * received TCP data contains ECT(0), ECT(1), or CE. 373 */ 374 if (!tcp_ecn_mode_rfc3168(tp)) 375 break; 376 tp->ecn_flags |= TCP_ECN_SEEN; 377 break; 378 default: 379 if (tcp_ca_needs_ecn(sk)) 380 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE); 381 if (!tcp_ecn_mode_rfc3168(tp)) 382 break; 383 tp->ecn_flags |= TCP_ECN_SEEN; 384 break; 385 } 386 } 387 388 /* Returns true if the byte counters can be used */ 389 static bool tcp_accecn_process_option(struct tcp_sock *tp, 390 const struct sk_buff *skb, 391 u32 delivered_bytes, int flag) 392 { 393 u8 estimate_ecnfield = tp->est_ecnfield; 394 bool ambiguous_ecn_bytes_incr = false; 395 bool first_changed = false; 396 unsigned int optlen; 397 bool order1, res; 398 unsigned int i; 399 u8 *ptr; 400 401 if (tcp_accecn_opt_fail_recv(tp)) 402 return false; 403 404 if (!(flag & FLAG_SLOWPATH) || !tp->rx_opt.accecn) { 405 if (!tp->saw_accecn_opt) { 406 /* Too late to enable after this point due to 407 * potential counter wraps 408 */ 409 if (tp->bytes_sent >= (1 << 23) - 1) { 410 u8 saw_opt = TCP_ACCECN_OPT_FAIL_SEEN; 411 412 tcp_accecn_saw_opt_fail_recv(tp, saw_opt); 413 } 414 return false; 415 } 416 417 if (estimate_ecnfield) { 418 u8 ecnfield = estimate_ecnfield - 1; 419 420 tp->delivered_ecn_bytes[ecnfield] += delivered_bytes; 421 return true; 422 } 423 return false; 424 } 425 426 ptr = skb_transport_header(skb) + tp->rx_opt.accecn; 427 optlen = ptr[1] - 2; 428 if (WARN_ON_ONCE(ptr[0] != TCPOPT_ACCECN0 && ptr[0] != TCPOPT_ACCECN1)) 429 return false; 430 order1 = (ptr[0] == TCPOPT_ACCECN1); 431 ptr += 2; 432 433 if (tp->saw_accecn_opt < TCP_ACCECN_OPT_COUNTER_SEEN) { 434 tp->saw_accecn_opt = tcp_accecn_option_init(skb, 435 tp->rx_opt.accecn); 436 if (tp->saw_accecn_opt == TCP_ACCECN_OPT_FAIL_SEEN) 437 tcp_accecn_fail_mode_set(tp, TCP_ACCECN_OPT_FAIL_RECV); 438 } 439 440 res = !!estimate_ecnfield; 441 for (i = 0; i < 3; i++) { 442 u32 init_offset; 443 u8 ecnfield; 444 s32 delta; 445 u32 *cnt; 446 447 if (optlen < TCPOLEN_ACCECN_PERFIELD) 448 break; 449 450 ecnfield = tcp_accecn_optfield_to_ecnfield(i, order1); 451 init_offset = tcp_accecn_field_init_offset(ecnfield); 452 cnt = &tp->delivered_ecn_bytes[ecnfield - 1]; 453 delta = tcp_update_ecn_bytes(cnt, ptr, init_offset); 454 if (delta && delta < 0) { 455 res = false; 456 ambiguous_ecn_bytes_incr = true; 457 } 458 if (delta && ecnfield != estimate_ecnfield) { 459 if (!first_changed) { 460 tp->est_ecnfield = ecnfield; 461 first_changed = true; 462 } else { 463 res = false; 464 ambiguous_ecn_bytes_incr = true; 465 } 466 } 467 468 optlen -= TCPOLEN_ACCECN_PERFIELD; 469 ptr += TCPOLEN_ACCECN_PERFIELD; 470 } 471 if (ambiguous_ecn_bytes_incr) 472 tp->est_ecnfield = 0; 473 474 return res; 475 } 476 477 static void tcp_count_delivered_ce(struct tcp_sock *tp, u32 ecn_count) 478 { 479 tp->delivered_ce += ecn_count; 480 } 481 482 /* Updates the delivered and delivered_ce counts */ 483 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered, 484 bool ece_ack) 485 { 486 tp->delivered += delivered; 487 if (tcp_ecn_mode_rfc3168(tp) && ece_ack) 488 tcp_count_delivered_ce(tp, delivered); 489 } 490 491 /* Returns the ECN CE delta */ 492 static u32 __tcp_accecn_process(struct sock *sk, const struct sk_buff *skb, 493 u32 delivered_pkts, u32 delivered_bytes, 494 int flag) 495 { 496 u32 old_ceb = tcp_sk(sk)->delivered_ecn_bytes[INET_ECN_CE - 1]; 497 const struct tcphdr *th = tcp_hdr(skb); 498 struct tcp_sock *tp = tcp_sk(sk); 499 u32 delta, safe_delta, d_ceb; 500 bool opt_deltas_valid; 501 u32 corrected_ace; 502 503 /* Reordered ACK or uncertain due to lack of data to send and ts */ 504 if (!(flag & (FLAG_FORWARD_PROGRESS | FLAG_TS_PROGRESS))) 505 return 0; 506 507 opt_deltas_valid = tcp_accecn_process_option(tp, skb, 508 delivered_bytes, flag); 509 510 if (!(flag & FLAG_SLOWPATH)) { 511 /* AccECN counter might overflow on large ACKs */ 512 if (delivered_pkts <= TCP_ACCECN_CEP_ACE_MASK) 513 return 0; 514 } 515 516 /* ACE field is not available during handshake */ 517 if (flag & FLAG_SYN_ACKED) 518 return 0; 519 520 if (tp->received_ce_pending >= TCP_ACCECN_ACE_MAX_DELTA) 521 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 522 523 corrected_ace = tcp_accecn_ace(th) - TCP_ACCECN_CEP_INIT_OFFSET; 524 delta = (corrected_ace - tp->delivered_ce) & TCP_ACCECN_CEP_ACE_MASK; 525 if (delivered_pkts <= TCP_ACCECN_CEP_ACE_MASK) 526 return delta; 527 528 safe_delta = delivered_pkts - 529 ((delivered_pkts - delta) & TCP_ACCECN_CEP_ACE_MASK); 530 531 if (opt_deltas_valid) { 532 d_ceb = tp->delivered_ecn_bytes[INET_ECN_CE - 1] - old_ceb; 533 if (!d_ceb) 534 return delta; 535 536 if ((delivered_pkts >= (TCP_ACCECN_CEP_ACE_MASK + 1) * 2) && 537 (tcp_is_sack(tp) || 538 ((1 << inet_csk(sk)->icsk_ca_state) & 539 (TCPF_CA_Open | TCPF_CA_CWR)))) { 540 u32 est_d_cep; 541 542 if (delivered_bytes <= d_ceb) 543 return safe_delta; 544 545 est_d_cep = DIV_ROUND_UP_ULL((u64)d_ceb * 546 delivered_pkts, 547 delivered_bytes); 548 return min(safe_delta, 549 delta + 550 (est_d_cep & ~TCP_ACCECN_CEP_ACE_MASK)); 551 } 552 553 if (d_ceb > delta * tp->mss_cache) 554 return safe_delta; 555 if (d_ceb < 556 safe_delta * tp->mss_cache >> TCP_ACCECN_SAFETY_SHIFT) 557 return delta; 558 } 559 560 return safe_delta; 561 } 562 563 static u32 tcp_accecn_process(struct sock *sk, const struct sk_buff *skb, 564 u32 delivered_pkts, u32 delivered_bytes, 565 int *flag) 566 { 567 struct tcp_sock *tp = tcp_sk(sk); 568 u32 delta; 569 570 delta = __tcp_accecn_process(sk, skb, delivered_pkts, 571 delivered_bytes, *flag); 572 if (delta > 0) { 573 tcp_count_delivered_ce(tp, delta); 574 *flag |= FLAG_ECE; 575 /* Recalculate header predictor */ 576 if (tp->pred_flags) 577 tcp_fast_path_on(tp); 578 } 579 return delta; 580 } 581 582 /* Buffer size and advertised window tuning. 583 * 584 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 585 */ 586 587 static void tcp_sndbuf_expand(struct sock *sk) 588 { 589 const struct tcp_sock *tp = tcp_sk(sk); 590 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 591 int sndmem, per_mss; 592 u32 nr_segs; 593 594 /* Worst case is non GSO/TSO : each frame consumes one skb 595 * and skb->head is kmalloced using power of two area of memory 596 */ 597 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 598 MAX_TCP_HEADER + 599 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 600 601 per_mss = roundup_pow_of_two(per_mss) + 602 SKB_DATA_ALIGN(sizeof(struct sk_buff)); 603 604 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp)); 605 nr_segs = max_t(u32, nr_segs, tp->reordering + 1); 606 607 /* Fast Recovery (RFC 5681 3.2) : 608 * Cubic needs 1.7 factor, rounded to 2 to include 609 * extra cushion (application might react slowly to EPOLLOUT) 610 */ 611 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2; 612 sndmem *= nr_segs * per_mss; 613 614 if (sk->sk_sndbuf < sndmem) 615 WRITE_ONCE(sk->sk_sndbuf, 616 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2]))); 617 } 618 619 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 620 * 621 * All tcp_full_space() is split to two parts: "network" buffer, allocated 622 * forward and advertised in receiver window (tp->rcv_wnd) and 623 * "application buffer", required to isolate scheduling/application 624 * latencies from network. 625 * window_clamp is maximal advertised window. It can be less than 626 * tcp_full_space(), in this case tcp_full_space() - window_clamp 627 * is reserved for "application" buffer. The less window_clamp is 628 * the smoother our behaviour from viewpoint of network, but the lower 629 * throughput and the higher sensitivity of the connection to losses. 8) 630 * 631 * rcv_ssthresh is more strict window_clamp used at "slow start" 632 * phase to predict further behaviour of this connection. 633 * It is used for two goals: 634 * - to enforce header prediction at sender, even when application 635 * requires some significant "application buffer". It is check #1. 636 * - to prevent pruning of receive queue because of misprediction 637 * of receiver window. Check #2. 638 * 639 * The scheme does not work when sender sends good segments opening 640 * window and then starts to feed us spaghetti. But it should work 641 * in common situations. Otherwise, we have to rely on queue collapsing. 642 */ 643 644 /* Slow part of check#2. */ 645 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb, 646 unsigned int skbtruesize) 647 { 648 const struct tcp_sock *tp = tcp_sk(sk); 649 /* Optimize this! */ 650 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1; 651 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1; 652 653 while (tp->rcv_ssthresh <= window) { 654 if (truesize <= skb->len) 655 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 656 657 truesize >>= 1; 658 window >>= 1; 659 } 660 return 0; 661 } 662 663 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing 664 * can play nice with us, as sk_buff and skb->head might be either 665 * freed or shared with up to MAX_SKB_FRAGS segments. 666 * Only give a boost to drivers using page frag(s) to hold the frame(s), 667 * and if no payload was pulled in skb->head before reaching us. 668 */ 669 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb) 670 { 671 u32 truesize = skb->truesize; 672 673 if (adjust && !skb_headlen(skb)) { 674 truesize -= SKB_TRUESIZE(skb_end_offset(skb)); 675 /* paranoid check, some drivers might be buggy */ 676 if (unlikely((int)truesize < (int)skb->len)) 677 truesize = skb->truesize; 678 } 679 return truesize; 680 } 681 682 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb, 683 bool adjust) 684 { 685 struct tcp_sock *tp = tcp_sk(sk); 686 int room; 687 688 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh; 689 690 if (room <= 0) 691 return; 692 693 /* Check #1 */ 694 if (!tcp_under_memory_pressure(sk)) { 695 unsigned int truesize = truesize_adjust(adjust, skb); 696 int incr; 697 698 /* Check #2. Increase window, if skb with such overhead 699 * will fit to rcvbuf in future. 700 */ 701 if (tcp_win_from_space(sk, truesize) <= skb->len) 702 incr = 2 * tp->advmss; 703 else 704 incr = __tcp_grow_window(sk, skb, truesize); 705 706 if (incr) { 707 incr = max_t(int, incr, 2 * skb->len); 708 tp->rcv_ssthresh += min(room, incr); 709 inet_csk(sk)->icsk_ack.quick |= 1; 710 } 711 } else { 712 /* Under pressure: 713 * Adjust rcv_ssthresh according to reserved mem 714 */ 715 tcp_adjust_rcv_ssthresh(sk); 716 } 717 } 718 719 /* 3. Try to fixup all. It is made immediately after connection enters 720 * established state. 721 */ 722 static void tcp_init_buffer_space(struct sock *sk) 723 { 724 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win); 725 struct tcp_sock *tp = tcp_sk(sk); 726 int maxwin; 727 728 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 729 tcp_sndbuf_expand(sk); 730 731 tcp_mstamp_refresh(tp); 732 tp->rcvq_space.time = tp->tcp_mstamp; 733 tp->rcvq_space.seq = tp->copied_seq; 734 735 maxwin = tcp_full_space(sk); 736 737 if (tp->window_clamp >= maxwin) { 738 WRITE_ONCE(tp->window_clamp, maxwin); 739 740 if (tcp_app_win && maxwin > 4 * tp->advmss) 741 WRITE_ONCE(tp->window_clamp, 742 max(maxwin - (maxwin >> tcp_app_win), 743 4 * tp->advmss)); 744 } 745 746 /* Force reservation of one segment. */ 747 if (tcp_app_win && 748 tp->window_clamp > 2 * tp->advmss && 749 tp->window_clamp + tp->advmss > maxwin) 750 WRITE_ONCE(tp->window_clamp, 751 max(2 * tp->advmss, maxwin - tp->advmss)); 752 753 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 754 tp->snd_cwnd_stamp = tcp_jiffies32; 755 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd, 756 (u32)TCP_INIT_CWND * tp->advmss); 757 } 758 759 /* 4. Recalculate window clamp after socket hit its memory bounds. */ 760 static void tcp_clamp_window(struct sock *sk) 761 { 762 struct tcp_sock *tp = tcp_sk(sk); 763 struct inet_connection_sock *icsk = inet_csk(sk); 764 struct net *net = sock_net(sk); 765 int rmem2; 766 767 icsk->icsk_ack.quick = 0; 768 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]); 769 770 if (sk->sk_rcvbuf < rmem2 && 771 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 772 !tcp_under_memory_pressure(sk) && 773 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { 774 WRITE_ONCE(sk->sk_rcvbuf, 775 min(atomic_read(&sk->sk_rmem_alloc), rmem2)); 776 } 777 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 778 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 779 } 780 781 /* Initialize RCV_MSS value. 782 * RCV_MSS is an our guess about MSS used by the peer. 783 * We haven't any direct information about the MSS. 784 * It's better to underestimate the RCV_MSS rather than overestimate. 785 * Overestimations make us ACKing less frequently than needed. 786 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 787 */ 788 void tcp_initialize_rcv_mss(struct sock *sk) 789 { 790 const struct tcp_sock *tp = tcp_sk(sk); 791 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 792 793 hint = min(hint, tp->rcv_wnd / 2); 794 hint = min(hint, TCP_MSS_DEFAULT); 795 hint = max(hint, TCP_MIN_MSS); 796 797 inet_csk(sk)->icsk_ack.rcv_mss = hint; 798 } 799 EXPORT_IPV6_MOD(tcp_initialize_rcv_mss); 800 801 /* Receiver "autotuning" code. 802 * 803 * The algorithm for RTT estimation w/o timestamps is based on 804 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 805 * <https://public.lanl.gov/radiant/pubs.html#DRS> 806 * 807 * More detail on this code can be found at 808 * <http://staff.psc.edu/jheffner/>, 809 * though this reference is out of date. A new paper 810 * is pending. 811 */ 812 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 813 { 814 u32 new_sample, old_sample = tp->rcv_rtt_est.rtt_us; 815 long m = sample << 3; 816 817 if (old_sample == 0 || m < old_sample) { 818 new_sample = m; 819 } else { 820 /* If we sample in larger samples in the non-timestamp 821 * case, we could grossly overestimate the RTT especially 822 * with chatty applications or bulk transfer apps which 823 * are stalled on filesystem I/O. 824 * 825 * Also, since we are only going for a minimum in the 826 * non-timestamp case, we do not smooth things out 827 * else with timestamps disabled convergence takes too 828 * long. 829 */ 830 if (win_dep) 831 return; 832 /* Do not use this sample if receive queue is not empty. */ 833 if (tp->rcv_nxt != tp->copied_seq) 834 return; 835 new_sample = old_sample - (old_sample >> 3) + sample; 836 } 837 838 tp->rcv_rtt_est.rtt_us = new_sample; 839 } 840 841 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 842 { 843 u32 delta_us; 844 845 if (tp->rcv_rtt_est.time == 0) 846 goto new_measure; 847 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 848 return; 849 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time); 850 if (!delta_us) 851 delta_us = 1; 852 tcp_rcv_rtt_update(tp, delta_us, 1); 853 854 new_measure: 855 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 856 tp->rcv_rtt_est.time = tp->tcp_mstamp; 857 } 858 859 static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp, u32 min_delta) 860 { 861 u32 delta, delta_us; 862 863 delta = tcp_time_stamp_ts(tp) - tp->rx_opt.rcv_tsecr; 864 if (tp->tcp_usec_ts) 865 return delta; 866 867 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) { 868 if (!delta) 869 delta = min_delta; 870 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ); 871 return delta_us; 872 } 873 return -1; 874 } 875 876 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 877 const struct sk_buff *skb) 878 { 879 struct tcp_sock *tp = tcp_sk(sk); 880 881 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr) 882 return; 883 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 884 885 if (TCP_SKB_CB(skb)->end_seq - 886 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) { 887 s32 delta = tcp_rtt_tsopt_us(tp, 0); 888 889 if (delta > 0) 890 tcp_rcv_rtt_update(tp, delta, 0); 891 } 892 } 893 894 void tcp_rcvbuf_grow(struct sock *sk) 895 { 896 const struct net *net = sock_net(sk); 897 struct tcp_sock *tp = tcp_sk(sk); 898 int rcvwin, rcvbuf, cap; 899 900 if (!READ_ONCE(net->ipv4.sysctl_tcp_moderate_rcvbuf) || 901 (sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 902 return; 903 904 /* slow start: allow the sender to double its rate. */ 905 rcvwin = tp->rcvq_space.space << 1; 906 907 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) 908 rcvwin += TCP_SKB_CB(tp->ooo_last_skb)->end_seq - tp->rcv_nxt; 909 910 cap = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]); 911 912 rcvbuf = min_t(u32, tcp_space_from_win(sk, rcvwin), cap); 913 if (rcvbuf > sk->sk_rcvbuf) { 914 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 915 /* Make the window clamp follow along. */ 916 WRITE_ONCE(tp->window_clamp, 917 tcp_win_from_space(sk, rcvbuf)); 918 } 919 } 920 /* 921 * This function should be called every time data is copied to user space. 922 * It calculates the appropriate TCP receive buffer space. 923 */ 924 void tcp_rcv_space_adjust(struct sock *sk) 925 { 926 struct tcp_sock *tp = tcp_sk(sk); 927 int time, inq, copied; 928 929 trace_tcp_rcv_space_adjust(sk); 930 931 tcp_mstamp_refresh(tp); 932 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time); 933 if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0) 934 return; 935 936 /* Number of bytes copied to user in last RTT */ 937 copied = tp->copied_seq - tp->rcvq_space.seq; 938 /* Number of bytes in receive queue. */ 939 inq = tp->rcv_nxt - tp->copied_seq; 940 copied -= inq; 941 if (copied <= tp->rcvq_space.space) 942 goto new_measure; 943 944 trace_tcp_rcvbuf_grow(sk, time); 945 946 tp->rcvq_space.space = copied; 947 948 tcp_rcvbuf_grow(sk); 949 950 new_measure: 951 tp->rcvq_space.seq = tp->copied_seq; 952 tp->rcvq_space.time = tp->tcp_mstamp; 953 } 954 955 static void tcp_save_lrcv_flowlabel(struct sock *sk, const struct sk_buff *skb) 956 { 957 #if IS_ENABLED(CONFIG_IPV6) 958 struct inet_connection_sock *icsk = inet_csk(sk); 959 960 if (skb->protocol == htons(ETH_P_IPV6)) 961 icsk->icsk_ack.lrcv_flowlabel = ntohl(ip6_flowlabel(ipv6_hdr(skb))); 962 #endif 963 } 964 965 /* There is something which you must keep in mind when you analyze the 966 * behavior of the tp->ato delayed ack timeout interval. When a 967 * connection starts up, we want to ack as quickly as possible. The 968 * problem is that "good" TCP's do slow start at the beginning of data 969 * transmission. The means that until we send the first few ACK's the 970 * sender will sit on his end and only queue most of his data, because 971 * he can only send snd_cwnd unacked packets at any given time. For 972 * each ACK we send, he increments snd_cwnd and transmits more of his 973 * queue. -DaveM 974 */ 975 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 976 { 977 struct tcp_sock *tp = tcp_sk(sk); 978 struct inet_connection_sock *icsk = inet_csk(sk); 979 u32 now; 980 981 inet_csk_schedule_ack(sk); 982 983 tcp_measure_rcv_mss(sk, skb); 984 985 tcp_rcv_rtt_measure(tp); 986 987 now = tcp_jiffies32; 988 989 if (!icsk->icsk_ack.ato) { 990 /* The _first_ data packet received, initialize 991 * delayed ACK engine. 992 */ 993 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 994 icsk->icsk_ack.ato = TCP_ATO_MIN; 995 } else { 996 int m = now - icsk->icsk_ack.lrcvtime; 997 998 if (m <= TCP_ATO_MIN / 2) { 999 /* The fastest case is the first. */ 1000 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 1001 } else if (m < icsk->icsk_ack.ato) { 1002 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 1003 if (icsk->icsk_ack.ato > icsk->icsk_rto) 1004 icsk->icsk_ack.ato = icsk->icsk_rto; 1005 } else if (m > icsk->icsk_rto) { 1006 /* Too long gap. Apparently sender failed to 1007 * restart window, so that we send ACKs quickly. 1008 */ 1009 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 1010 } 1011 } 1012 icsk->icsk_ack.lrcvtime = now; 1013 tcp_save_lrcv_flowlabel(sk, skb); 1014 1015 tcp_data_ecn_check(sk, skb); 1016 1017 if (skb->len >= 128) 1018 tcp_grow_window(sk, skb, true); 1019 } 1020 1021 /* Called to compute a smoothed rtt estimate. The data fed to this 1022 * routine either comes from timestamps, or from segments that were 1023 * known _not_ to have been retransmitted [see Karn/Partridge 1024 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 1025 * piece by Van Jacobson. 1026 * NOTE: the next three routines used to be one big routine. 1027 * To save cycles in the RFC 1323 implementation it was better to break 1028 * it up into three procedures. -- erics 1029 */ 1030 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) 1031 { 1032 struct tcp_sock *tp = tcp_sk(sk); 1033 long m = mrtt_us; /* RTT */ 1034 u32 srtt = tp->srtt_us; 1035 1036 /* The following amusing code comes from Jacobson's 1037 * article in SIGCOMM '88. Note that rtt and mdev 1038 * are scaled versions of rtt and mean deviation. 1039 * This is designed to be as fast as possible 1040 * m stands for "measurement". 1041 * 1042 * On a 1990 paper the rto value is changed to: 1043 * RTO = rtt + 4 * mdev 1044 * 1045 * Funny. This algorithm seems to be very broken. 1046 * These formulae increase RTO, when it should be decreased, increase 1047 * too slowly, when it should be increased quickly, decrease too quickly 1048 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 1049 * does not matter how to _calculate_ it. Seems, it was trap 1050 * that VJ failed to avoid. 8) 1051 */ 1052 if (srtt != 0) { 1053 m -= (srtt >> 3); /* m is now error in rtt est */ 1054 srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 1055 if (m < 0) { 1056 m = -m; /* m is now abs(error) */ 1057 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 1058 /* This is similar to one of Eifel findings. 1059 * Eifel blocks mdev updates when rtt decreases. 1060 * This solution is a bit different: we use finer gain 1061 * for mdev in this case (alpha*beta). 1062 * Like Eifel it also prevents growth of rto, 1063 * but also it limits too fast rto decreases, 1064 * happening in pure Eifel. 1065 */ 1066 if (m > 0) 1067 m >>= 3; 1068 } else { 1069 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 1070 } 1071 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ 1072 if (tp->mdev_us > tp->mdev_max_us) { 1073 tp->mdev_max_us = tp->mdev_us; 1074 if (tp->mdev_max_us > tp->rttvar_us) 1075 tp->rttvar_us = tp->mdev_max_us; 1076 } 1077 if (after(tp->snd_una, tp->rtt_seq)) { 1078 if (tp->mdev_max_us < tp->rttvar_us) 1079 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; 1080 tp->rtt_seq = tp->snd_nxt; 1081 tp->mdev_max_us = tcp_rto_min_us(sk); 1082 1083 tcp_bpf_rtt(sk, mrtt_us, srtt); 1084 } 1085 } else { 1086 /* no previous measure. */ 1087 srtt = m << 3; /* take the measured time to be rtt */ 1088 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ 1089 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); 1090 tp->mdev_max_us = tp->rttvar_us; 1091 tp->rtt_seq = tp->snd_nxt; 1092 1093 tcp_bpf_rtt(sk, mrtt_us, srtt); 1094 } 1095 tp->srtt_us = max(1U, srtt); 1096 } 1097 1098 static void tcp_update_pacing_rate(struct sock *sk) 1099 { 1100 const struct tcp_sock *tp = tcp_sk(sk); 1101 u64 rate; 1102 1103 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ 1104 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); 1105 1106 /* current rate is (cwnd * mss) / srtt 1107 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. 1108 * In Congestion Avoidance phase, set it to 120 % the current rate. 1109 * 1110 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) 1111 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching 1112 * end of slow start and should slow down. 1113 */ 1114 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2) 1115 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio); 1116 else 1117 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio); 1118 1119 rate *= max(tcp_snd_cwnd(tp), tp->packets_out); 1120 1121 if (likely(tp->srtt_us)) 1122 do_div(rate, tp->srtt_us); 1123 1124 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate 1125 * without any lock. We want to make sure compiler wont store 1126 * intermediate values in this location. 1127 */ 1128 WRITE_ONCE(sk->sk_pacing_rate, 1129 min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate))); 1130 } 1131 1132 /* Calculate rto without backoff. This is the second half of Van Jacobson's 1133 * routine referred to above. 1134 */ 1135 static void tcp_set_rto(struct sock *sk) 1136 { 1137 const struct tcp_sock *tp = tcp_sk(sk); 1138 /* Old crap is replaced with new one. 8) 1139 * 1140 * More seriously: 1141 * 1. If rtt variance happened to be less 50msec, it is hallucination. 1142 * It cannot be less due to utterly erratic ACK generation made 1143 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 1144 * to do with delayed acks, because at cwnd>2 true delack timeout 1145 * is invisible. Actually, Linux-2.4 also generates erratic 1146 * ACKs in some circumstances. 1147 */ 1148 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 1149 1150 /* 2. Fixups made earlier cannot be right. 1151 * If we do not estimate RTO correctly without them, 1152 * all the algo is pure shit and should be replaced 1153 * with correct one. It is exactly, which we pretend to do. 1154 */ 1155 1156 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 1157 * guarantees that rto is higher. 1158 */ 1159 tcp_bound_rto(sk); 1160 } 1161 1162 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 1163 { 1164 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 1165 1166 if (!cwnd) 1167 cwnd = TCP_INIT_CWND; 1168 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 1169 } 1170 1171 struct tcp_sacktag_state { 1172 /* Timestamps for earliest and latest never-retransmitted segment 1173 * that was SACKed. RTO needs the earliest RTT to stay conservative, 1174 * but congestion control should still get an accurate delay signal. 1175 */ 1176 u64 first_sackt; 1177 u64 last_sackt; 1178 u32 reord; 1179 u32 sack_delivered; 1180 u32 delivered_bytes; 1181 int flag; 1182 unsigned int mss_now; 1183 struct rate_sample *rate; 1184 }; 1185 1186 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery 1187 * and spurious retransmission information if this DSACK is unlikely caused by 1188 * sender's action: 1189 * - DSACKed sequence range is larger than maximum receiver's window. 1190 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs. 1191 */ 1192 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq, 1193 u32 end_seq, struct tcp_sacktag_state *state) 1194 { 1195 u32 seq_len, dup_segs = 1; 1196 1197 if (!before(start_seq, end_seq)) 1198 return 0; 1199 1200 seq_len = end_seq - start_seq; 1201 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */ 1202 if (seq_len > tp->max_window) 1203 return 0; 1204 if (seq_len > tp->mss_cache) 1205 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache); 1206 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq) 1207 state->flag |= FLAG_DSACK_TLP; 1208 1209 tp->dsack_dups += dup_segs; 1210 /* Skip the DSACK if dup segs weren't retransmitted by sender */ 1211 if (tp->dsack_dups > tp->total_retrans) 1212 return 0; 1213 1214 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 1215 /* We increase the RACK ordering window in rounds where we receive 1216 * DSACKs that may have been due to reordering causing RACK to trigger 1217 * a spurious fast recovery. Thus RACK ignores DSACKs that happen 1218 * without having seen reordering, or that match TLP probes (TLP 1219 * is timer-driven, not triggered by RACK). 1220 */ 1221 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP)) 1222 tp->rack.dsack_seen = 1; 1223 1224 state->flag |= FLAG_DSACKING_ACK; 1225 /* A spurious retransmission is delivered */ 1226 state->sack_delivered += dup_segs; 1227 1228 return dup_segs; 1229 } 1230 1231 /* It's reordering when higher sequence was delivered (i.e. sacked) before 1232 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering 1233 * distance is approximated in full-mss packet distance ("reordering"). 1234 */ 1235 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq, 1236 const int ts) 1237 { 1238 struct tcp_sock *tp = tcp_sk(sk); 1239 const u32 mss = tp->mss_cache; 1240 u32 fack, metric; 1241 1242 fack = tcp_highest_sack_seq(tp); 1243 if (!before(low_seq, fack)) 1244 return; 1245 1246 metric = fack - low_seq; 1247 if ((metric > tp->reordering * mss) && mss) { 1248 #if FASTRETRANS_DEBUG > 1 1249 pr_debug("Disorder%d %d %u f%u s%u rr%d\n", 1250 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 1251 tp->reordering, 1252 0, 1253 tp->sacked_out, 1254 tp->undo_marker ? tp->undo_retrans : 0); 1255 #endif 1256 tp->reordering = min_t(u32, (metric + mss - 1) / mss, 1257 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)); 1258 } 1259 1260 /* This exciting event is worth to be remembered. 8) */ 1261 tp->reord_seen++; 1262 NET_INC_STATS(sock_net(sk), 1263 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER); 1264 } 1265 1266 /* This must be called before lost_out or retrans_out are updated 1267 * on a new loss, because we want to know if all skbs previously 1268 * known to be lost have already been retransmitted, indicating 1269 * that this newly lost skb is our next skb to retransmit. 1270 */ 1271 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 1272 { 1273 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) || 1274 (tp->retransmit_skb_hint && 1275 before(TCP_SKB_CB(skb)->seq, 1276 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))) 1277 tp->retransmit_skb_hint = skb; 1278 } 1279 1280 /* Sum the number of packets on the wire we have marked as lost, and 1281 * notify the congestion control module that the given skb was marked lost. 1282 */ 1283 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb) 1284 { 1285 tp->lost += tcp_skb_pcount(skb); 1286 } 1287 1288 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb) 1289 { 1290 __u8 sacked = TCP_SKB_CB(skb)->sacked; 1291 struct tcp_sock *tp = tcp_sk(sk); 1292 1293 if (sacked & TCPCB_SACKED_ACKED) 1294 return; 1295 1296 tcp_verify_retransmit_hint(tp, skb); 1297 if (sacked & TCPCB_LOST) { 1298 if (sacked & TCPCB_SACKED_RETRANS) { 1299 /* Account for retransmits that are lost again */ 1300 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1301 tp->retrans_out -= tcp_skb_pcount(skb); 1302 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT, 1303 tcp_skb_pcount(skb)); 1304 tcp_notify_skb_loss_event(tp, skb); 1305 } 1306 } else { 1307 tp->lost_out += tcp_skb_pcount(skb); 1308 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1309 tcp_notify_skb_loss_event(tp, skb); 1310 } 1311 } 1312 1313 /* This procedure tags the retransmission queue when SACKs arrive. 1314 * 1315 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1316 * Packets in queue with these bits set are counted in variables 1317 * sacked_out, retrans_out and lost_out, correspondingly. 1318 * 1319 * Valid combinations are: 1320 * Tag InFlight Description 1321 * 0 1 - orig segment is in flight. 1322 * S 0 - nothing flies, orig reached receiver. 1323 * L 0 - nothing flies, orig lost by net. 1324 * R 2 - both orig and retransmit are in flight. 1325 * L|R 1 - orig is lost, retransmit is in flight. 1326 * S|R 1 - orig reached receiver, retrans is still in flight. 1327 * (L|S|R is logically valid, it could occur when L|R is sacked, 1328 * but it is equivalent to plain S and code short-circuits it to S. 1329 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1330 * 1331 * These 6 states form finite state machine, controlled by the following events: 1332 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1333 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1334 * 3. Loss detection event of two flavors: 1335 * A. Scoreboard estimator decided the packet is lost. 1336 * A'. Reno "three dupacks" marks head of queue lost. 1337 * B. SACK arrives sacking SND.NXT at the moment, when the 1338 * segment was retransmitted. 1339 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1340 * 1341 * It is pleasant to note, that state diagram turns out to be commutative, 1342 * so that we are allowed not to be bothered by order of our actions, 1343 * when multiple events arrive simultaneously. (see the function below). 1344 * 1345 * Reordering detection. 1346 * -------------------- 1347 * Reordering metric is maximal distance, which a packet can be displaced 1348 * in packet stream. With SACKs we can estimate it: 1349 * 1350 * 1. SACK fills old hole and the corresponding segment was not 1351 * ever retransmitted -> reordering. Alas, we cannot use it 1352 * when segment was retransmitted. 1353 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1354 * for retransmitted and already SACKed segment -> reordering.. 1355 * Both of these heuristics are not used in Loss state, when we cannot 1356 * account for retransmits accurately. 1357 * 1358 * SACK block validation. 1359 * ---------------------- 1360 * 1361 * SACK block range validation checks that the received SACK block fits to 1362 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1363 * Note that SND.UNA is not included to the range though being valid because 1364 * it means that the receiver is rather inconsistent with itself reporting 1365 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1366 * perfectly valid, however, in light of RFC2018 which explicitly states 1367 * that "SACK block MUST reflect the newest segment. Even if the newest 1368 * segment is going to be discarded ...", not that it looks very clever 1369 * in case of head skb. Due to potentional receiver driven attacks, we 1370 * choose to avoid immediate execution of a walk in write queue due to 1371 * reneging and defer head skb's loss recovery to standard loss recovery 1372 * procedure that will eventually trigger (nothing forbids us doing this). 1373 * 1374 * Implements also blockage to start_seq wrap-around. Problem lies in the 1375 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1376 * there's no guarantee that it will be before snd_nxt (n). The problem 1377 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1378 * wrap (s_w): 1379 * 1380 * <- outs wnd -> <- wrapzone -> 1381 * u e n u_w e_w s n_w 1382 * | | | | | | | 1383 * |<------------+------+----- TCP seqno space --------------+---------->| 1384 * ...-- <2^31 ->| |<--------... 1385 * ...---- >2^31 ------>| |<--------... 1386 * 1387 * Current code wouldn't be vulnerable but it's better still to discard such 1388 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1389 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1390 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1391 * equal to the ideal case (infinite seqno space without wrap caused issues). 1392 * 1393 * With D-SACK the lower bound is extended to cover sequence space below 1394 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1395 * again, D-SACK block must not to go across snd_una (for the same reason as 1396 * for the normal SACK blocks, explained above). But there all simplicity 1397 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1398 * fully below undo_marker they do not affect behavior in anyway and can 1399 * therefore be safely ignored. In rare cases (which are more or less 1400 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1401 * fragmentation and packet reordering past skb's retransmission. To consider 1402 * them correctly, the acceptable range must be extended even more though 1403 * the exact amount is rather hard to quantify. However, tp->max_window can 1404 * be used as an exaggerated estimate. 1405 */ 1406 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, 1407 u32 start_seq, u32 end_seq) 1408 { 1409 /* Too far in future, or reversed (interpretation is ambiguous) */ 1410 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1411 return false; 1412 1413 /* Nasty start_seq wrap-around check (see comments above) */ 1414 if (!before(start_seq, tp->snd_nxt)) 1415 return false; 1416 1417 /* In outstanding window? ...This is valid exit for D-SACKs too. 1418 * start_seq == snd_una is non-sensical (see comments above) 1419 */ 1420 if (after(start_seq, tp->snd_una)) 1421 return true; 1422 1423 if (!is_dsack || !tp->undo_marker) 1424 return false; 1425 1426 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1427 if (after(end_seq, tp->snd_una)) 1428 return false; 1429 1430 if (!before(start_seq, tp->undo_marker)) 1431 return true; 1432 1433 /* Too old */ 1434 if (!after(end_seq, tp->undo_marker)) 1435 return false; 1436 1437 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1438 * start_seq < undo_marker and end_seq >= undo_marker. 1439 */ 1440 return !before(start_seq, end_seq - tp->max_window); 1441 } 1442 1443 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1444 struct tcp_sack_block_wire *sp, int num_sacks, 1445 u32 prior_snd_una, struct tcp_sacktag_state *state) 1446 { 1447 struct tcp_sock *tp = tcp_sk(sk); 1448 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1449 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1450 u32 dup_segs; 1451 1452 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1453 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1454 } else if (num_sacks > 1) { 1455 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1456 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1457 1458 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1)) 1459 return false; 1460 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV); 1461 } else { 1462 return false; 1463 } 1464 1465 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state); 1466 if (!dup_segs) { /* Skip dubious DSACK */ 1467 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS); 1468 return false; 1469 } 1470 1471 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs); 1472 1473 /* D-SACK for already forgotten data... Do dumb counting. */ 1474 if (tp->undo_marker && tp->undo_retrans > 0 && 1475 !after(end_seq_0, prior_snd_una) && 1476 after(end_seq_0, tp->undo_marker)) 1477 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs); 1478 1479 return true; 1480 } 1481 1482 /* Check if skb is fully within the SACK block. In presence of GSO skbs, 1483 * the incoming SACK may not exactly match but we can find smaller MSS 1484 * aligned portion of it that matches. Therefore we might need to fragment 1485 * which may fail and creates some hassle (caller must handle error case 1486 * returns). 1487 * 1488 * FIXME: this could be merged to shift decision code 1489 */ 1490 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1491 u32 start_seq, u32 end_seq) 1492 { 1493 int err; 1494 bool in_sack; 1495 unsigned int pkt_len; 1496 unsigned int mss; 1497 1498 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1499 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1500 1501 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1502 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1503 mss = tcp_skb_mss(skb); 1504 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1505 1506 if (!in_sack) { 1507 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1508 if (pkt_len < mss) 1509 pkt_len = mss; 1510 } else { 1511 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1512 if (pkt_len < mss) 1513 return -EINVAL; 1514 } 1515 1516 /* Round if necessary so that SACKs cover only full MSSes 1517 * and/or the remaining small portion (if present) 1518 */ 1519 if (pkt_len > mss) { 1520 unsigned int new_len = (pkt_len / mss) * mss; 1521 if (!in_sack && new_len < pkt_len) 1522 new_len += mss; 1523 pkt_len = new_len; 1524 } 1525 1526 if (pkt_len >= skb->len && !in_sack) 1527 return 0; 1528 1529 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 1530 pkt_len, mss, GFP_ATOMIC); 1531 if (err < 0) 1532 return err; 1533 } 1534 1535 return in_sack; 1536 } 1537 1538 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1539 static u8 tcp_sacktag_one(struct sock *sk, 1540 struct tcp_sacktag_state *state, u8 sacked, 1541 u32 start_seq, u32 end_seq, 1542 int dup_sack, int pcount, u32 plen, 1543 u64 xmit_time) 1544 { 1545 struct tcp_sock *tp = tcp_sk(sk); 1546 1547 /* Account D-SACK for retransmitted packet. */ 1548 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1549 if (tp->undo_marker && tp->undo_retrans > 0 && 1550 after(end_seq, tp->undo_marker)) 1551 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount); 1552 if ((sacked & TCPCB_SACKED_ACKED) && 1553 before(start_seq, state->reord)) 1554 state->reord = start_seq; 1555 } 1556 1557 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1558 if (!after(end_seq, tp->snd_una)) 1559 return sacked; 1560 1561 if (!(sacked & TCPCB_SACKED_ACKED)) { 1562 tcp_rack_advance(tp, sacked, end_seq, xmit_time); 1563 1564 if (sacked & TCPCB_SACKED_RETRANS) { 1565 /* If the segment is not tagged as lost, 1566 * we do not clear RETRANS, believing 1567 * that retransmission is still in flight. 1568 */ 1569 if (sacked & TCPCB_LOST) { 1570 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1571 tp->lost_out -= pcount; 1572 tp->retrans_out -= pcount; 1573 } 1574 } else { 1575 if (!(sacked & TCPCB_RETRANS)) { 1576 /* New sack for not retransmitted frame, 1577 * which was in hole. It is reordering. 1578 */ 1579 if (before(start_seq, 1580 tcp_highest_sack_seq(tp)) && 1581 before(start_seq, state->reord)) 1582 state->reord = start_seq; 1583 1584 if (!after(end_seq, tp->high_seq)) 1585 state->flag |= FLAG_ORIG_SACK_ACKED; 1586 if (state->first_sackt == 0) 1587 state->first_sackt = xmit_time; 1588 state->last_sackt = xmit_time; 1589 } 1590 1591 if (sacked & TCPCB_LOST) { 1592 sacked &= ~TCPCB_LOST; 1593 tp->lost_out -= pcount; 1594 } 1595 } 1596 1597 sacked |= TCPCB_SACKED_ACKED; 1598 state->flag |= FLAG_DATA_SACKED; 1599 tp->sacked_out += pcount; 1600 /* Out-of-order packets delivered */ 1601 state->sack_delivered += pcount; 1602 state->delivered_bytes += plen; 1603 } 1604 1605 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1606 * frames and clear it. undo_retrans is decreased above, L|R frames 1607 * are accounted above as well. 1608 */ 1609 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1610 sacked &= ~TCPCB_SACKED_RETRANS; 1611 tp->retrans_out -= pcount; 1612 } 1613 1614 return sacked; 1615 } 1616 1617 /* Shift newly-SACKed bytes from this skb to the immediately previous 1618 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1619 */ 1620 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, 1621 struct sk_buff *skb, 1622 struct tcp_sacktag_state *state, 1623 unsigned int pcount, int shifted, int mss, 1624 bool dup_sack) 1625 { 1626 struct tcp_sock *tp = tcp_sk(sk); 1627 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1628 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1629 1630 BUG_ON(!pcount); 1631 1632 /* Adjust counters and hints for the newly sacked sequence 1633 * range but discard the return value since prev is already 1634 * marked. We must tag the range first because the seq 1635 * advancement below implicitly advances 1636 * tcp_highest_sack_seq() when skb is highest_sack. 1637 */ 1638 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1639 start_seq, end_seq, dup_sack, pcount, skb->len, 1640 tcp_skb_timestamp_us(skb)); 1641 tcp_rate_skb_delivered(sk, skb, state->rate); 1642 1643 TCP_SKB_CB(prev)->end_seq += shifted; 1644 TCP_SKB_CB(skb)->seq += shifted; 1645 1646 tcp_skb_pcount_add(prev, pcount); 1647 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount); 1648 tcp_skb_pcount_add(skb, -pcount); 1649 1650 /* When we're adding to gso_segs == 1, gso_size will be zero, 1651 * in theory this shouldn't be necessary but as long as DSACK 1652 * code can come after this skb later on it's better to keep 1653 * setting gso_size to something. 1654 */ 1655 if (!TCP_SKB_CB(prev)->tcp_gso_size) 1656 TCP_SKB_CB(prev)->tcp_gso_size = mss; 1657 1658 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1659 if (tcp_skb_pcount(skb) <= 1) 1660 TCP_SKB_CB(skb)->tcp_gso_size = 0; 1661 1662 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1663 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1664 1665 if (skb->len > 0) { 1666 BUG_ON(!tcp_skb_pcount(skb)); 1667 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1668 return false; 1669 } 1670 1671 /* Whole SKB was eaten :-) */ 1672 1673 if (skb == tp->retransmit_skb_hint) 1674 tp->retransmit_skb_hint = prev; 1675 1676 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1677 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor; 1678 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1679 TCP_SKB_CB(prev)->end_seq++; 1680 1681 if (skb == tcp_highest_sack(sk)) 1682 tcp_advance_highest_sack(sk, skb); 1683 1684 tcp_skb_collapse_tstamp(prev, skb); 1685 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp)) 1686 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0; 1687 1688 tcp_rtx_queue_unlink_and_free(skb, sk); 1689 1690 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED); 1691 1692 return true; 1693 } 1694 1695 /* I wish gso_size would have a bit more sane initialization than 1696 * something-or-zero which complicates things 1697 */ 1698 static int tcp_skb_seglen(const struct sk_buff *skb) 1699 { 1700 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1701 } 1702 1703 /* Shifting pages past head area doesn't work */ 1704 static int skb_can_shift(const struct sk_buff *skb) 1705 { 1706 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1707 } 1708 1709 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from, 1710 int pcount, int shiftlen) 1711 { 1712 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE) 1713 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need 1714 * to make sure not storing more than 65535 * 8 bytes per skb, 1715 * even if current MSS is bigger. 1716 */ 1717 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE)) 1718 return 0; 1719 if (unlikely(tcp_skb_pcount(to) + pcount > 65535)) 1720 return 0; 1721 return skb_shift(to, from, shiftlen); 1722 } 1723 1724 /* Try collapsing SACK blocks spanning across multiple skbs to a single 1725 * skb. 1726 */ 1727 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1728 struct tcp_sacktag_state *state, 1729 u32 start_seq, u32 end_seq, 1730 bool dup_sack) 1731 { 1732 struct tcp_sock *tp = tcp_sk(sk); 1733 struct sk_buff *prev; 1734 int mss; 1735 int pcount = 0; 1736 int len; 1737 int in_sack; 1738 1739 /* Normally R but no L won't result in plain S */ 1740 if (!dup_sack && 1741 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1742 goto fallback; 1743 if (!skb_can_shift(skb)) 1744 goto fallback; 1745 /* This frame is about to be dropped (was ACKed). */ 1746 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1747 goto fallback; 1748 1749 /* Can only happen with delayed DSACK + discard craziness */ 1750 prev = skb_rb_prev(skb); 1751 if (!prev) 1752 goto fallback; 1753 1754 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1755 goto fallback; 1756 1757 if (!tcp_skb_can_collapse(prev, skb)) 1758 goto fallback; 1759 1760 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1761 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1762 1763 if (in_sack) { 1764 len = skb->len; 1765 pcount = tcp_skb_pcount(skb); 1766 mss = tcp_skb_seglen(skb); 1767 1768 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1769 * drop this restriction as unnecessary 1770 */ 1771 if (mss != tcp_skb_seglen(prev)) 1772 goto fallback; 1773 } else { 1774 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1775 goto noop; 1776 /* CHECKME: This is non-MSS split case only?, this will 1777 * cause skipped skbs due to advancing loop btw, original 1778 * has that feature too 1779 */ 1780 if (tcp_skb_pcount(skb) <= 1) 1781 goto noop; 1782 1783 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1784 if (!in_sack) { 1785 /* TODO: head merge to next could be attempted here 1786 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1787 * though it might not be worth of the additional hassle 1788 * 1789 * ...we can probably just fallback to what was done 1790 * previously. We could try merging non-SACKed ones 1791 * as well but it probably isn't going to buy off 1792 * because later SACKs might again split them, and 1793 * it would make skb timestamp tracking considerably 1794 * harder problem. 1795 */ 1796 goto fallback; 1797 } 1798 1799 len = end_seq - TCP_SKB_CB(skb)->seq; 1800 BUG_ON(len < 0); 1801 BUG_ON(len > skb->len); 1802 1803 /* MSS boundaries should be honoured or else pcount will 1804 * severely break even though it makes things bit trickier. 1805 * Optimize common case to avoid most of the divides 1806 */ 1807 mss = tcp_skb_mss(skb); 1808 1809 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1810 * drop this restriction as unnecessary 1811 */ 1812 if (mss != tcp_skb_seglen(prev)) 1813 goto fallback; 1814 1815 if (len == mss) { 1816 pcount = 1; 1817 } else if (len < mss) { 1818 goto noop; 1819 } else { 1820 pcount = len / mss; 1821 len = pcount * mss; 1822 } 1823 } 1824 1825 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ 1826 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) 1827 goto fallback; 1828 1829 if (!tcp_skb_shift(prev, skb, pcount, len)) 1830 goto fallback; 1831 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack)) 1832 goto out; 1833 1834 /* Hole filled allows collapsing with the next as well, this is very 1835 * useful when hole on every nth skb pattern happens 1836 */ 1837 skb = skb_rb_next(prev); 1838 if (!skb) 1839 goto out; 1840 1841 if (!skb_can_shift(skb) || 1842 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1843 (mss != tcp_skb_seglen(skb))) 1844 goto out; 1845 1846 if (!tcp_skb_can_collapse(prev, skb)) 1847 goto out; 1848 len = skb->len; 1849 pcount = tcp_skb_pcount(skb); 1850 if (tcp_skb_shift(prev, skb, pcount, len)) 1851 tcp_shifted_skb(sk, prev, skb, state, pcount, 1852 len, mss, 0); 1853 1854 out: 1855 return prev; 1856 1857 noop: 1858 return skb; 1859 1860 fallback: 1861 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1862 return NULL; 1863 } 1864 1865 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1866 struct tcp_sack_block *next_dup, 1867 struct tcp_sacktag_state *state, 1868 u32 start_seq, u32 end_seq, 1869 bool dup_sack_in) 1870 { 1871 struct tcp_sock *tp = tcp_sk(sk); 1872 struct sk_buff *tmp; 1873 1874 skb_rbtree_walk_from(skb) { 1875 int in_sack = 0; 1876 bool dup_sack = dup_sack_in; 1877 1878 /* queue is in-order => we can short-circuit the walk early */ 1879 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1880 break; 1881 1882 if (next_dup && 1883 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1884 in_sack = tcp_match_skb_to_sack(sk, skb, 1885 next_dup->start_seq, 1886 next_dup->end_seq); 1887 if (in_sack > 0) 1888 dup_sack = true; 1889 } 1890 1891 /* skb reference here is a bit tricky to get right, since 1892 * shifting can eat and free both this skb and the next, 1893 * so not even _safe variant of the loop is enough. 1894 */ 1895 if (in_sack <= 0) { 1896 tmp = tcp_shift_skb_data(sk, skb, state, 1897 start_seq, end_seq, dup_sack); 1898 if (tmp) { 1899 if (tmp != skb) { 1900 skb = tmp; 1901 continue; 1902 } 1903 1904 in_sack = 0; 1905 } else { 1906 in_sack = tcp_match_skb_to_sack(sk, skb, 1907 start_seq, 1908 end_seq); 1909 } 1910 } 1911 1912 if (unlikely(in_sack < 0)) 1913 break; 1914 1915 if (in_sack) { 1916 TCP_SKB_CB(skb)->sacked = 1917 tcp_sacktag_one(sk, 1918 state, 1919 TCP_SKB_CB(skb)->sacked, 1920 TCP_SKB_CB(skb)->seq, 1921 TCP_SKB_CB(skb)->end_seq, 1922 dup_sack, 1923 tcp_skb_pcount(skb), 1924 skb->len, 1925 tcp_skb_timestamp_us(skb)); 1926 tcp_rate_skb_delivered(sk, skb, state->rate); 1927 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1928 list_del_init(&skb->tcp_tsorted_anchor); 1929 1930 if (!before(TCP_SKB_CB(skb)->seq, 1931 tcp_highest_sack_seq(tp))) 1932 tcp_advance_highest_sack(sk, skb); 1933 } 1934 } 1935 return skb; 1936 } 1937 1938 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq) 1939 { 1940 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node; 1941 struct sk_buff *skb; 1942 1943 while (*p) { 1944 parent = *p; 1945 skb = rb_to_skb(parent); 1946 if (before(seq, TCP_SKB_CB(skb)->seq)) { 1947 p = &parent->rb_left; 1948 continue; 1949 } 1950 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) { 1951 p = &parent->rb_right; 1952 continue; 1953 } 1954 return skb; 1955 } 1956 return NULL; 1957 } 1958 1959 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1960 u32 skip_to_seq) 1961 { 1962 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq)) 1963 return skb; 1964 1965 return tcp_sacktag_bsearch(sk, skip_to_seq); 1966 } 1967 1968 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1969 struct sock *sk, 1970 struct tcp_sack_block *next_dup, 1971 struct tcp_sacktag_state *state, 1972 u32 skip_to_seq) 1973 { 1974 if (!next_dup) 1975 return skb; 1976 1977 if (before(next_dup->start_seq, skip_to_seq)) { 1978 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq); 1979 skb = tcp_sacktag_walk(skb, sk, NULL, state, 1980 next_dup->start_seq, next_dup->end_seq, 1981 1); 1982 } 1983 1984 return skb; 1985 } 1986 1987 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 1988 { 1989 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 1990 } 1991 1992 static int 1993 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 1994 u32 prior_snd_una, struct tcp_sacktag_state *state) 1995 { 1996 struct tcp_sock *tp = tcp_sk(sk); 1997 const unsigned char *ptr = (skb_transport_header(ack_skb) + 1998 TCP_SKB_CB(ack_skb)->sacked); 1999 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 2000 struct tcp_sack_block sp[TCP_NUM_SACKS]; 2001 struct tcp_sack_block *cache; 2002 struct sk_buff *skb; 2003 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 2004 int used_sacks; 2005 bool found_dup_sack = false; 2006 int i, j; 2007 int first_sack_index; 2008 2009 state->flag = 0; 2010 state->reord = tp->snd_nxt; 2011 2012 if (!tp->sacked_out) 2013 tcp_highest_sack_reset(sk); 2014 2015 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 2016 num_sacks, prior_snd_una, state); 2017 2018 /* Eliminate too old ACKs, but take into 2019 * account more or less fresh ones, they can 2020 * contain valid SACK info. 2021 */ 2022 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 2023 return 0; 2024 2025 if (!tp->packets_out) 2026 goto out; 2027 2028 used_sacks = 0; 2029 first_sack_index = 0; 2030 for (i = 0; i < num_sacks; i++) { 2031 bool dup_sack = !i && found_dup_sack; 2032 2033 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 2034 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 2035 2036 if (!tcp_is_sackblock_valid(tp, dup_sack, 2037 sp[used_sacks].start_seq, 2038 sp[used_sacks].end_seq)) { 2039 int mib_idx; 2040 2041 if (dup_sack) { 2042 if (!tp->undo_marker) 2043 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 2044 else 2045 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 2046 } else { 2047 /* Don't count olds caused by ACK reordering */ 2048 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 2049 !after(sp[used_sacks].end_seq, tp->snd_una)) 2050 continue; 2051 mib_idx = LINUX_MIB_TCPSACKDISCARD; 2052 } 2053 2054 NET_INC_STATS(sock_net(sk), mib_idx); 2055 if (i == 0) 2056 first_sack_index = -1; 2057 continue; 2058 } 2059 2060 /* Ignore very old stuff early */ 2061 if (!after(sp[used_sacks].end_seq, prior_snd_una)) { 2062 if (i == 0) 2063 first_sack_index = -1; 2064 continue; 2065 } 2066 2067 used_sacks++; 2068 } 2069 2070 /* order SACK blocks to allow in order walk of the retrans queue */ 2071 for (i = used_sacks - 1; i > 0; i--) { 2072 for (j = 0; j < i; j++) { 2073 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 2074 swap(sp[j], sp[j + 1]); 2075 2076 /* Track where the first SACK block goes to */ 2077 if (j == first_sack_index) 2078 first_sack_index = j + 1; 2079 } 2080 } 2081 } 2082 2083 state->mss_now = tcp_current_mss(sk); 2084 skb = NULL; 2085 i = 0; 2086 2087 if (!tp->sacked_out) { 2088 /* It's already past, so skip checking against it */ 2089 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 2090 } else { 2091 cache = tp->recv_sack_cache; 2092 /* Skip empty blocks in at head of the cache */ 2093 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 2094 !cache->end_seq) 2095 cache++; 2096 } 2097 2098 while (i < used_sacks) { 2099 u32 start_seq = sp[i].start_seq; 2100 u32 end_seq = sp[i].end_seq; 2101 bool dup_sack = (found_dup_sack && (i == first_sack_index)); 2102 struct tcp_sack_block *next_dup = NULL; 2103 2104 if (found_dup_sack && ((i + 1) == first_sack_index)) 2105 next_dup = &sp[i + 1]; 2106 2107 /* Skip too early cached blocks */ 2108 while (tcp_sack_cache_ok(tp, cache) && 2109 !before(start_seq, cache->end_seq)) 2110 cache++; 2111 2112 /* Can skip some work by looking recv_sack_cache? */ 2113 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 2114 after(end_seq, cache->start_seq)) { 2115 2116 /* Head todo? */ 2117 if (before(start_seq, cache->start_seq)) { 2118 skb = tcp_sacktag_skip(skb, sk, start_seq); 2119 skb = tcp_sacktag_walk(skb, sk, next_dup, 2120 state, 2121 start_seq, 2122 cache->start_seq, 2123 dup_sack); 2124 } 2125 2126 /* Rest of the block already fully processed? */ 2127 if (!after(end_seq, cache->end_seq)) 2128 goto advance_sp; 2129 2130 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 2131 state, 2132 cache->end_seq); 2133 2134 /* ...tail remains todo... */ 2135 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 2136 /* ...but better entrypoint exists! */ 2137 skb = tcp_highest_sack(sk); 2138 if (!skb) 2139 break; 2140 cache++; 2141 goto walk; 2142 } 2143 2144 skb = tcp_sacktag_skip(skb, sk, cache->end_seq); 2145 /* Check overlap against next cached too (past this one already) */ 2146 cache++; 2147 continue; 2148 } 2149 2150 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 2151 skb = tcp_highest_sack(sk); 2152 if (!skb) 2153 break; 2154 } 2155 skb = tcp_sacktag_skip(skb, sk, start_seq); 2156 2157 walk: 2158 skb = tcp_sacktag_walk(skb, sk, next_dup, state, 2159 start_seq, end_seq, dup_sack); 2160 2161 advance_sp: 2162 i++; 2163 } 2164 2165 /* Clear the head of the cache sack blocks so we can skip it next time */ 2166 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 2167 tp->recv_sack_cache[i].start_seq = 0; 2168 tp->recv_sack_cache[i].end_seq = 0; 2169 } 2170 for (j = 0; j < used_sacks; j++) 2171 tp->recv_sack_cache[i++] = sp[j]; 2172 2173 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker) 2174 tcp_check_sack_reordering(sk, state->reord, 0); 2175 2176 tcp_verify_left_out(tp); 2177 out: 2178 2179 #if FASTRETRANS_DEBUG > 0 2180 WARN_ON((int)tp->sacked_out < 0); 2181 WARN_ON((int)tp->lost_out < 0); 2182 WARN_ON((int)tp->retrans_out < 0); 2183 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 2184 #endif 2185 return state->flag; 2186 } 2187 2188 /* Limits sacked_out so that sum with lost_out isn't ever larger than 2189 * packets_out. Returns false if sacked_out adjustement wasn't necessary. 2190 */ 2191 static bool tcp_limit_reno_sacked(struct tcp_sock *tp) 2192 { 2193 u32 holes; 2194 2195 holes = max(tp->lost_out, 1U); 2196 holes = min(holes, tp->packets_out); 2197 2198 if ((tp->sacked_out + holes) > tp->packets_out) { 2199 tp->sacked_out = tp->packets_out - holes; 2200 return true; 2201 } 2202 return false; 2203 } 2204 2205 /* If we receive more dupacks than we expected counting segments 2206 * in assumption of absent reordering, interpret this as reordering. 2207 * The only another reason could be bug in receiver TCP. 2208 */ 2209 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 2210 { 2211 struct tcp_sock *tp = tcp_sk(sk); 2212 2213 if (!tcp_limit_reno_sacked(tp)) 2214 return; 2215 2216 tp->reordering = min_t(u32, tp->packets_out + addend, 2217 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)); 2218 tp->reord_seen++; 2219 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER); 2220 } 2221 2222 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 2223 2224 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack) 2225 { 2226 if (num_dupack) { 2227 struct tcp_sock *tp = tcp_sk(sk); 2228 u32 prior_sacked = tp->sacked_out; 2229 s32 delivered; 2230 2231 tp->sacked_out += num_dupack; 2232 tcp_check_reno_reordering(sk, 0); 2233 delivered = tp->sacked_out - prior_sacked; 2234 if (delivered > 0) 2235 tcp_count_delivered(tp, delivered, ece_ack); 2236 tcp_verify_left_out(tp); 2237 } 2238 } 2239 2240 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 2241 2242 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack) 2243 { 2244 struct tcp_sock *tp = tcp_sk(sk); 2245 2246 if (acked > 0) { 2247 /* One ACK acked hole. The rest eat duplicate ACKs. */ 2248 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1), 2249 ece_ack); 2250 if (acked - 1 >= tp->sacked_out) 2251 tp->sacked_out = 0; 2252 else 2253 tp->sacked_out -= acked - 1; 2254 } 2255 tcp_check_reno_reordering(sk, acked); 2256 tcp_verify_left_out(tp); 2257 } 2258 2259 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 2260 { 2261 tp->sacked_out = 0; 2262 } 2263 2264 void tcp_clear_retrans(struct tcp_sock *tp) 2265 { 2266 tp->retrans_out = 0; 2267 tp->lost_out = 0; 2268 tp->undo_marker = 0; 2269 tp->undo_retrans = -1; 2270 tp->sacked_out = 0; 2271 tp->rto_stamp = 0; 2272 tp->total_rto = 0; 2273 tp->total_rto_recoveries = 0; 2274 tp->total_rto_time = 0; 2275 } 2276 2277 static inline void tcp_init_undo(struct tcp_sock *tp) 2278 { 2279 tp->undo_marker = tp->snd_una; 2280 2281 /* Retransmission still in flight may cause DSACKs later. */ 2282 /* First, account for regular retransmits in flight: */ 2283 tp->undo_retrans = tp->retrans_out; 2284 /* Next, account for TLP retransmits in flight: */ 2285 if (tp->tlp_high_seq && tp->tlp_retrans) 2286 tp->undo_retrans++; 2287 /* Finally, avoid 0, because undo_retrans==0 means "can undo now": */ 2288 if (!tp->undo_retrans) 2289 tp->undo_retrans = -1; 2290 } 2291 2292 /* If we detect SACK reneging, forget all SACK information 2293 * and reset tags completely, otherwise preserve SACKs. If receiver 2294 * dropped its ofo queue, we will know this due to reneging detection. 2295 */ 2296 static void tcp_timeout_mark_lost(struct sock *sk) 2297 { 2298 struct tcp_sock *tp = tcp_sk(sk); 2299 struct sk_buff *skb, *head; 2300 bool is_reneg; /* is receiver reneging on SACKs? */ 2301 2302 head = tcp_rtx_queue_head(sk); 2303 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED); 2304 if (is_reneg) { 2305 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 2306 tp->sacked_out = 0; 2307 /* Mark SACK reneging until we recover from this loss event. */ 2308 tp->is_sack_reneg = 1; 2309 } else if (tcp_is_reno(tp)) { 2310 tcp_reset_reno_sack(tp); 2311 } 2312 2313 skb = head; 2314 skb_rbtree_walk_from(skb) { 2315 if (is_reneg) 2316 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2317 else if (skb != head && tcp_rack_skb_timeout(tp, skb, 0) > 0) 2318 continue; /* Don't mark recently sent ones lost yet */ 2319 tcp_mark_skb_lost(sk, skb); 2320 } 2321 tcp_verify_left_out(tp); 2322 tcp_clear_all_retrans_hints(tp); 2323 } 2324 2325 /* Enter Loss state. */ 2326 void tcp_enter_loss(struct sock *sk) 2327 { 2328 const struct inet_connection_sock *icsk = inet_csk(sk); 2329 struct tcp_sock *tp = tcp_sk(sk); 2330 struct net *net = sock_net(sk); 2331 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery; 2332 u8 reordering; 2333 2334 tcp_timeout_mark_lost(sk); 2335 2336 /* Reduce ssthresh if it has not yet been made inside this window. */ 2337 if (icsk->icsk_ca_state <= TCP_CA_Disorder || 2338 !after(tp->high_seq, tp->snd_una) || 2339 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2340 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2341 tp->prior_cwnd = tcp_snd_cwnd(tp); 2342 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2343 tcp_ca_event(sk, CA_EVENT_LOSS); 2344 tcp_init_undo(tp); 2345 } 2346 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1); 2347 tp->snd_cwnd_cnt = 0; 2348 tp->snd_cwnd_stamp = tcp_jiffies32; 2349 2350 /* Timeout in disordered state after receiving substantial DUPACKs 2351 * suggests that the degree of reordering is over-estimated. 2352 */ 2353 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering); 2354 if (icsk->icsk_ca_state <= TCP_CA_Disorder && 2355 tp->sacked_out >= reordering) 2356 tp->reordering = min_t(unsigned int, tp->reordering, 2357 reordering); 2358 2359 tcp_set_ca_state(sk, TCP_CA_Loss); 2360 tp->high_seq = tp->snd_nxt; 2361 tp->tlp_high_seq = 0; 2362 tcp_ecn_queue_cwr(tp); 2363 2364 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous 2365 * loss recovery is underway except recurring timeout(s) on 2366 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing 2367 */ 2368 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) && 2369 (new_recovery || icsk->icsk_retransmits) && 2370 !inet_csk(sk)->icsk_mtup.probe_size; 2371 } 2372 2373 /* If ACK arrived pointing to a remembered SACK, it means that our 2374 * remembered SACKs do not reflect real state of receiver i.e. 2375 * receiver _host_ is heavily congested (or buggy). 2376 * 2377 * To avoid big spurious retransmission bursts due to transient SACK 2378 * scoreboard oddities that look like reneging, we give the receiver a 2379 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will 2380 * restore sanity to the SACK scoreboard. If the apparent reneging 2381 * persists until this RTO then we'll clear the SACK scoreboard. 2382 */ 2383 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag) 2384 { 2385 if (*ack_flag & FLAG_SACK_RENEGING && 2386 *ack_flag & FLAG_SND_UNA_ADVANCED) { 2387 struct tcp_sock *tp = tcp_sk(sk); 2388 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4), 2389 msecs_to_jiffies(10)); 2390 2391 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, delay, false); 2392 *ack_flag &= ~FLAG_SET_XMIT_TIMER; 2393 return true; 2394 } 2395 return false; 2396 } 2397 2398 /* Linux NewReno/SACK/ECN state machine. 2399 * -------------------------------------- 2400 * 2401 * "Open" Normal state, no dubious events, fast path. 2402 * "Disorder" In all the respects it is "Open", 2403 * but requires a bit more attention. It is entered when 2404 * we see some SACKs or dupacks. It is split of "Open" 2405 * mainly to move some processing from fast path to slow one. 2406 * "CWR" CWND was reduced due to some Congestion Notification event. 2407 * It can be ECN, ICMP source quench, local device congestion. 2408 * "Recovery" CWND was reduced, we are fast-retransmitting. 2409 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2410 * 2411 * tcp_fastretrans_alert() is entered: 2412 * - each incoming ACK, if state is not "Open" 2413 * - when arrived ACK is unusual, namely: 2414 * * SACK 2415 * * Duplicate ACK. 2416 * * ECN ECE. 2417 * 2418 * Counting packets in flight is pretty simple. 2419 * 2420 * in_flight = packets_out - left_out + retrans_out 2421 * 2422 * packets_out is SND.NXT-SND.UNA counted in packets. 2423 * 2424 * retrans_out is number of retransmitted segments. 2425 * 2426 * left_out is number of segments left network, but not ACKed yet. 2427 * 2428 * left_out = sacked_out + lost_out 2429 * 2430 * sacked_out: Packets, which arrived to receiver out of order 2431 * and hence not ACKed. With SACKs this number is simply 2432 * amount of SACKed data. Even without SACKs 2433 * it is easy to give pretty reliable estimate of this number, 2434 * counting duplicate ACKs. 2435 * 2436 * lost_out: Packets lost by network. TCP has no explicit 2437 * "loss notification" feedback from network (for now). 2438 * It means that this number can be only _guessed_. 2439 * Actually, it is the heuristics to predict lossage that 2440 * distinguishes different algorithms. 2441 * 2442 * F.e. after RTO, when all the queue is considered as lost, 2443 * lost_out = packets_out and in_flight = retrans_out. 2444 * 2445 * Essentially, we have now a few algorithms detecting 2446 * lost packets. 2447 * 2448 * If the receiver supports SACK: 2449 * 2450 * RACK (RFC8985): RACK is a newer loss detection algorithm 2451 * (2017-) that checks timing instead of counting DUPACKs. 2452 * Essentially a packet is considered lost if it's not S/ACKed 2453 * after RTT + reordering_window, where both metrics are 2454 * dynamically measured and adjusted. This is implemented in 2455 * tcp_rack_mark_lost. 2456 * 2457 * If the receiver does not support SACK: 2458 * 2459 * NewReno (RFC6582): in Recovery we assume that one segment 2460 * is lost (classic Reno). While we are in Recovery and 2461 * a partial ACK arrives, we assume that one more packet 2462 * is lost (NewReno). This heuristics are the same in NewReno 2463 * and SACK. 2464 * 2465 * The really tricky (and requiring careful tuning) part of the algorithm 2466 * is hidden in the RACK code in tcp_recovery.c and tcp_xmit_retransmit_queue(). 2467 * The first determines the moment _when_ we should reduce CWND and, 2468 * hence, slow down forward transmission. In fact, it determines the moment 2469 * when we decide that hole is caused by loss, rather than by a reorder. 2470 * 2471 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2472 * holes, caused by lost packets. 2473 * 2474 * And the most logically complicated part of algorithm is undo 2475 * heuristics. We detect false retransmits due to both too early 2476 * fast retransmit (reordering) and underestimated RTO, analyzing 2477 * timestamps and D-SACKs. When we detect that some segments were 2478 * retransmitted by mistake and CWND reduction was wrong, we undo 2479 * window reduction and abort recovery phase. This logic is hidden 2480 * inside several functions named tcp_try_undo_<something>. 2481 */ 2482 2483 /* This function decides, when we should leave Disordered state 2484 * and enter Recovery phase, reducing congestion window. 2485 * 2486 * Main question: may we further continue forward transmission 2487 * with the same cwnd? 2488 */ 2489 static bool tcp_time_to_recover(const struct tcp_sock *tp) 2490 { 2491 /* Has loss detection marked at least one packet lost? */ 2492 return tp->lost_out != 0; 2493 } 2494 2495 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when) 2496 { 2497 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2498 before(tp->rx_opt.rcv_tsecr, when); 2499 } 2500 2501 /* skb is spurious retransmitted if the returned timestamp echo 2502 * reply is prior to the skb transmission time 2503 */ 2504 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp, 2505 const struct sk_buff *skb) 2506 { 2507 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) && 2508 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb)); 2509 } 2510 2511 /* Nothing was retransmitted or returned timestamp is less 2512 * than timestamp of the first retransmission. 2513 */ 2514 static inline bool tcp_packet_delayed(const struct tcp_sock *tp) 2515 { 2516 const struct sock *sk = (const struct sock *)tp; 2517 2518 /* Received an echoed timestamp before the first retransmission? */ 2519 if (tp->retrans_stamp) 2520 return tcp_tsopt_ecr_before(tp, tp->retrans_stamp); 2521 2522 /* We set tp->retrans_stamp upon the first retransmission of a loss 2523 * recovery episode, so normally if tp->retrans_stamp is 0 then no 2524 * retransmission has happened yet (likely due to TSQ, which can cause 2525 * fast retransmits to be delayed). So if snd_una advanced while 2526 * (tp->retrans_stamp is 0 then apparently a packet was merely delayed, 2527 * not lost. But there are exceptions where we retransmit but then 2528 * clear tp->retrans_stamp, so we check for those exceptions. 2529 */ 2530 2531 /* (1) For non-SACK connections, tcp_is_non_sack_preventing_reopen() 2532 * clears tp->retrans_stamp when snd_una == high_seq. 2533 */ 2534 if (!tcp_is_sack(tp) && !before(tp->snd_una, tp->high_seq)) 2535 return false; 2536 2537 /* (2) In TCP_SYN_SENT tcp_clean_rtx_queue() clears tp->retrans_stamp 2538 * when setting FLAG_SYN_ACKED is set, even if the SYN was 2539 * retransmitted. 2540 */ 2541 if (sk->sk_state == TCP_SYN_SENT) 2542 return false; 2543 2544 return true; /* tp->retrans_stamp is zero; no retransmit yet */ 2545 } 2546 2547 /* Undo procedures. */ 2548 2549 /* We can clear retrans_stamp when there are no retransmissions in the 2550 * window. It would seem that it is trivially available for us in 2551 * tp->retrans_out, however, that kind of assumptions doesn't consider 2552 * what will happen if errors occur when sending retransmission for the 2553 * second time. ...It could the that such segment has only 2554 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2555 * the head skb is enough except for some reneging corner cases that 2556 * are not worth the effort. 2557 * 2558 * Main reason for all this complexity is the fact that connection dying 2559 * time now depends on the validity of the retrans_stamp, in particular, 2560 * that successive retransmissions of a segment must not advance 2561 * retrans_stamp under any conditions. 2562 */ 2563 static bool tcp_any_retrans_done(const struct sock *sk) 2564 { 2565 const struct tcp_sock *tp = tcp_sk(sk); 2566 struct sk_buff *skb; 2567 2568 if (tp->retrans_out) 2569 return true; 2570 2571 skb = tcp_rtx_queue_head(sk); 2572 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2573 return true; 2574 2575 return false; 2576 } 2577 2578 /* If loss recovery is finished and there are no retransmits out in the 2579 * network, then we clear retrans_stamp so that upon the next loss recovery 2580 * retransmits_timed_out() and timestamp-undo are using the correct value. 2581 */ 2582 static void tcp_retrans_stamp_cleanup(struct sock *sk) 2583 { 2584 if (!tcp_any_retrans_done(sk)) 2585 tcp_sk(sk)->retrans_stamp = 0; 2586 } 2587 2588 static void DBGUNDO(struct sock *sk, const char *msg) 2589 { 2590 #if FASTRETRANS_DEBUG > 1 2591 struct tcp_sock *tp = tcp_sk(sk); 2592 struct inet_sock *inet = inet_sk(sk); 2593 2594 if (sk->sk_family == AF_INET) { 2595 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2596 msg, 2597 &inet->inet_daddr, ntohs(inet->inet_dport), 2598 tcp_snd_cwnd(tp), tcp_left_out(tp), 2599 tp->snd_ssthresh, tp->prior_ssthresh, 2600 tp->packets_out); 2601 } 2602 #if IS_ENABLED(CONFIG_IPV6) 2603 else if (sk->sk_family == AF_INET6) { 2604 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2605 msg, 2606 &sk->sk_v6_daddr, ntohs(inet->inet_dport), 2607 tcp_snd_cwnd(tp), tcp_left_out(tp), 2608 tp->snd_ssthresh, tp->prior_ssthresh, 2609 tp->packets_out); 2610 } 2611 #endif 2612 #endif 2613 } 2614 2615 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss) 2616 { 2617 struct tcp_sock *tp = tcp_sk(sk); 2618 2619 if (unmark_loss) { 2620 struct sk_buff *skb; 2621 2622 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2623 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2624 } 2625 tp->lost_out = 0; 2626 tcp_clear_all_retrans_hints(tp); 2627 } 2628 2629 if (tp->prior_ssthresh) { 2630 const struct inet_connection_sock *icsk = inet_csk(sk); 2631 2632 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk)); 2633 2634 if (tp->prior_ssthresh > tp->snd_ssthresh) { 2635 tp->snd_ssthresh = tp->prior_ssthresh; 2636 tcp_ecn_withdraw_cwr(tp); 2637 } 2638 } 2639 tp->snd_cwnd_stamp = tcp_jiffies32; 2640 tp->undo_marker = 0; 2641 tp->rack.advanced = 1; /* Force RACK to re-exam losses */ 2642 } 2643 2644 static inline bool tcp_may_undo(const struct tcp_sock *tp) 2645 { 2646 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2647 } 2648 2649 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk) 2650 { 2651 struct tcp_sock *tp = tcp_sk(sk); 2652 2653 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2654 /* Hold old state until something *above* high_seq 2655 * is ACKed. For Reno it is MUST to prevent false 2656 * fast retransmits (RFC2582). SACK TCP is safe. */ 2657 if (!tcp_any_retrans_done(sk)) 2658 tp->retrans_stamp = 0; 2659 return true; 2660 } 2661 return false; 2662 } 2663 2664 /* People celebrate: "We love our President!" */ 2665 static bool tcp_try_undo_recovery(struct sock *sk) 2666 { 2667 struct tcp_sock *tp = tcp_sk(sk); 2668 2669 if (tcp_may_undo(tp)) { 2670 int mib_idx; 2671 2672 /* Happy end! We did not retransmit anything 2673 * or our original transmission succeeded. 2674 */ 2675 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2676 tcp_undo_cwnd_reduction(sk, false); 2677 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2678 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2679 else 2680 mib_idx = LINUX_MIB_TCPFULLUNDO; 2681 2682 NET_INC_STATS(sock_net(sk), mib_idx); 2683 } else if (tp->rack.reo_wnd_persist) { 2684 tp->rack.reo_wnd_persist--; 2685 } 2686 if (tcp_is_non_sack_preventing_reopen(sk)) 2687 return true; 2688 tcp_set_ca_state(sk, TCP_CA_Open); 2689 tp->is_sack_reneg = 0; 2690 return false; 2691 } 2692 2693 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2694 static bool tcp_try_undo_dsack(struct sock *sk) 2695 { 2696 struct tcp_sock *tp = tcp_sk(sk); 2697 2698 if (tp->undo_marker && !tp->undo_retrans) { 2699 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH, 2700 tp->rack.reo_wnd_persist + 1); 2701 DBGUNDO(sk, "D-SACK"); 2702 tcp_undo_cwnd_reduction(sk, false); 2703 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2704 return true; 2705 } 2706 return false; 2707 } 2708 2709 /* Undo during loss recovery after partial ACK or using F-RTO. */ 2710 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo) 2711 { 2712 struct tcp_sock *tp = tcp_sk(sk); 2713 2714 if (frto_undo || tcp_may_undo(tp)) { 2715 tcp_undo_cwnd_reduction(sk, true); 2716 2717 DBGUNDO(sk, "partial loss"); 2718 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2719 if (frto_undo) 2720 NET_INC_STATS(sock_net(sk), 2721 LINUX_MIB_TCPSPURIOUSRTOS); 2722 WRITE_ONCE(inet_csk(sk)->icsk_retransmits, 0); 2723 if (tcp_is_non_sack_preventing_reopen(sk)) 2724 return true; 2725 if (frto_undo || tcp_is_sack(tp)) { 2726 tcp_set_ca_state(sk, TCP_CA_Open); 2727 tp->is_sack_reneg = 0; 2728 } 2729 return true; 2730 } 2731 return false; 2732 } 2733 2734 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937. 2735 * It computes the number of packets to send (sndcnt) based on packets newly 2736 * delivered: 2737 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 2738 * cwnd reductions across a full RTT. 2739 * 2) Otherwise PRR uses packet conservation to send as much as delivered. 2740 * But when SND_UNA is acked without further losses, 2741 * slow starts cwnd up to ssthresh to speed up the recovery. 2742 */ 2743 static void tcp_init_cwnd_reduction(struct sock *sk) 2744 { 2745 struct tcp_sock *tp = tcp_sk(sk); 2746 2747 tp->high_seq = tp->snd_nxt; 2748 tp->tlp_high_seq = 0; 2749 tp->snd_cwnd_cnt = 0; 2750 tp->prior_cwnd = tcp_snd_cwnd(tp); 2751 tp->prr_delivered = 0; 2752 tp->prr_out = 0; 2753 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); 2754 tcp_ecn_queue_cwr(tp); 2755 } 2756 2757 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag) 2758 { 2759 struct tcp_sock *tp = tcp_sk(sk); 2760 int sndcnt = 0; 2761 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 2762 2763 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd)) 2764 return; 2765 2766 trace_tcp_cwnd_reduction_tp(sk, newly_acked_sacked, newly_lost, flag); 2767 2768 tp->prr_delivered += newly_acked_sacked; 2769 if (delta < 0) { 2770 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 2771 tp->prior_cwnd - 1; 2772 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 2773 } else { 2774 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out, 2775 newly_acked_sacked); 2776 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost) 2777 sndcnt++; 2778 sndcnt = min(delta, sndcnt); 2779 } 2780 /* Force a fast retransmit upon entering fast recovery */ 2781 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1)); 2782 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt); 2783 } 2784 2785 static inline void tcp_end_cwnd_reduction(struct sock *sk) 2786 { 2787 struct tcp_sock *tp = tcp_sk(sk); 2788 2789 if (inet_csk(sk)->icsk_ca_ops->cong_control) 2790 return; 2791 2792 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ 2793 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH && 2794 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) { 2795 tcp_snd_cwnd_set(tp, tp->snd_ssthresh); 2796 tp->snd_cwnd_stamp = tcp_jiffies32; 2797 } 2798 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2799 } 2800 2801 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */ 2802 void tcp_enter_cwr(struct sock *sk) 2803 { 2804 struct tcp_sock *tp = tcp_sk(sk); 2805 2806 tp->prior_ssthresh = 0; 2807 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 2808 tp->undo_marker = 0; 2809 tcp_init_cwnd_reduction(sk); 2810 tcp_set_ca_state(sk, TCP_CA_CWR); 2811 } 2812 } 2813 EXPORT_SYMBOL(tcp_enter_cwr); 2814 2815 static void tcp_try_keep_open(struct sock *sk) 2816 { 2817 struct tcp_sock *tp = tcp_sk(sk); 2818 int state = TCP_CA_Open; 2819 2820 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 2821 state = TCP_CA_Disorder; 2822 2823 if (inet_csk(sk)->icsk_ca_state != state) { 2824 tcp_set_ca_state(sk, state); 2825 tp->high_seq = tp->snd_nxt; 2826 } 2827 } 2828 2829 static void tcp_try_to_open(struct sock *sk, int flag) 2830 { 2831 struct tcp_sock *tp = tcp_sk(sk); 2832 2833 tcp_verify_left_out(tp); 2834 2835 if (!tcp_any_retrans_done(sk)) 2836 tp->retrans_stamp = 0; 2837 2838 if (flag & FLAG_ECE) 2839 tcp_enter_cwr(sk); 2840 2841 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2842 tcp_try_keep_open(sk); 2843 } 2844 } 2845 2846 static void tcp_mtup_probe_failed(struct sock *sk) 2847 { 2848 struct inet_connection_sock *icsk = inet_csk(sk); 2849 2850 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2851 icsk->icsk_mtup.probe_size = 0; 2852 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL); 2853 } 2854 2855 static void tcp_mtup_probe_success(struct sock *sk) 2856 { 2857 struct tcp_sock *tp = tcp_sk(sk); 2858 struct inet_connection_sock *icsk = inet_csk(sk); 2859 u64 val; 2860 2861 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2862 2863 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache); 2864 do_div(val, icsk->icsk_mtup.probe_size); 2865 DEBUG_NET_WARN_ON_ONCE((u32)val != val); 2866 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val)); 2867 2868 tp->snd_cwnd_cnt = 0; 2869 tp->snd_cwnd_stamp = tcp_jiffies32; 2870 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2871 2872 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2873 icsk->icsk_mtup.probe_size = 0; 2874 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2875 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS); 2876 } 2877 2878 /* Sometimes we deduce that packets have been dropped due to reasons other than 2879 * congestion, like path MTU reductions or failed client TFO attempts. In these 2880 * cases we call this function to retransmit as many packets as cwnd allows, 2881 * without reducing cwnd. Given that retransmits will set retrans_stamp to a 2882 * non-zero value (and may do so in a later calling context due to TSQ), we 2883 * also enter CA_Loss so that we track when all retransmitted packets are ACKed 2884 * and clear retrans_stamp when that happens (to ensure later recurring RTOs 2885 * are using the correct retrans_stamp and don't declare ETIMEDOUT 2886 * prematurely). 2887 */ 2888 static void tcp_non_congestion_loss_retransmit(struct sock *sk) 2889 { 2890 const struct inet_connection_sock *icsk = inet_csk(sk); 2891 struct tcp_sock *tp = tcp_sk(sk); 2892 2893 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2894 tp->high_seq = tp->snd_nxt; 2895 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2896 tp->prior_ssthresh = 0; 2897 tp->undo_marker = 0; 2898 tcp_set_ca_state(sk, TCP_CA_Loss); 2899 } 2900 tcp_xmit_retransmit_queue(sk); 2901 } 2902 2903 /* Do a simple retransmit without using the backoff mechanisms in 2904 * tcp_timer. This is used for path mtu discovery. 2905 * The socket is already locked here. 2906 */ 2907 void tcp_simple_retransmit(struct sock *sk) 2908 { 2909 struct tcp_sock *tp = tcp_sk(sk); 2910 struct sk_buff *skb; 2911 int mss; 2912 2913 /* A fastopen SYN request is stored as two separate packets within 2914 * the retransmit queue, this is done by tcp_send_syn_data(). 2915 * As a result simply checking the MSS of the frames in the queue 2916 * will not work for the SYN packet. 2917 * 2918 * Us being here is an indication of a path MTU issue so we can 2919 * assume that the fastopen SYN was lost and just mark all the 2920 * frames in the retransmit queue as lost. We will use an MSS of 2921 * -1 to mark all frames as lost, otherwise compute the current MSS. 2922 */ 2923 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT) 2924 mss = -1; 2925 else 2926 mss = tcp_current_mss(sk); 2927 2928 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2929 if (tcp_skb_seglen(skb) > mss) 2930 tcp_mark_skb_lost(sk, skb); 2931 } 2932 2933 if (!tp->lost_out) 2934 return; 2935 2936 if (tcp_is_reno(tp)) 2937 tcp_limit_reno_sacked(tp); 2938 2939 tcp_verify_left_out(tp); 2940 2941 /* Don't muck with the congestion window here. 2942 * Reason is that we do not increase amount of _data_ 2943 * in network, but units changed and effective 2944 * cwnd/ssthresh really reduced now. 2945 */ 2946 tcp_non_congestion_loss_retransmit(sk); 2947 } 2948 EXPORT_IPV6_MOD(tcp_simple_retransmit); 2949 2950 void tcp_enter_recovery(struct sock *sk, bool ece_ack) 2951 { 2952 struct tcp_sock *tp = tcp_sk(sk); 2953 int mib_idx; 2954 2955 /* Start the clock with our fast retransmit, for undo and ETIMEDOUT. */ 2956 tcp_retrans_stamp_cleanup(sk); 2957 2958 if (tcp_is_reno(tp)) 2959 mib_idx = LINUX_MIB_TCPRENORECOVERY; 2960 else 2961 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 2962 2963 NET_INC_STATS(sock_net(sk), mib_idx); 2964 2965 tp->prior_ssthresh = 0; 2966 tcp_init_undo(tp); 2967 2968 if (!tcp_in_cwnd_reduction(sk)) { 2969 if (!ece_ack) 2970 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2971 tcp_init_cwnd_reduction(sk); 2972 } 2973 tcp_set_ca_state(sk, TCP_CA_Recovery); 2974 } 2975 2976 static void tcp_update_rto_time(struct tcp_sock *tp) 2977 { 2978 if (tp->rto_stamp) { 2979 tp->total_rto_time += tcp_time_stamp_ms(tp) - tp->rto_stamp; 2980 tp->rto_stamp = 0; 2981 } 2982 } 2983 2984 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are 2985 * recovered or spurious. Otherwise retransmits more on partial ACKs. 2986 */ 2987 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack, 2988 int *rexmit) 2989 { 2990 struct tcp_sock *tp = tcp_sk(sk); 2991 bool recovered = !before(tp->snd_una, tp->high_seq); 2992 2993 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) && 2994 tcp_try_undo_loss(sk, false)) 2995 return; 2996 2997 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */ 2998 /* Step 3.b. A timeout is spurious if not all data are 2999 * lost, i.e., never-retransmitted data are (s)acked. 3000 */ 3001 if ((flag & FLAG_ORIG_SACK_ACKED) && 3002 tcp_try_undo_loss(sk, true)) 3003 return; 3004 3005 if (after(tp->snd_nxt, tp->high_seq)) { 3006 if (flag & FLAG_DATA_SACKED || num_dupack) 3007 tp->frto = 0; /* Step 3.a. loss was real */ 3008 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) { 3009 tp->high_seq = tp->snd_nxt; 3010 /* Step 2.b. Try send new data (but deferred until cwnd 3011 * is updated in tcp_ack()). Otherwise fall back to 3012 * the conventional recovery. 3013 */ 3014 if (!tcp_write_queue_empty(sk) && 3015 after(tcp_wnd_end(tp), tp->snd_nxt)) { 3016 *rexmit = REXMIT_NEW; 3017 return; 3018 } 3019 tp->frto = 0; 3020 } 3021 } 3022 3023 if (recovered) { 3024 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */ 3025 tcp_try_undo_recovery(sk); 3026 return; 3027 } 3028 if (tcp_is_reno(tp)) { 3029 /* A Reno DUPACK means new data in F-RTO step 2.b above are 3030 * delivered. Lower inflight to clock out (re)transmissions. 3031 */ 3032 if (after(tp->snd_nxt, tp->high_seq) && num_dupack) 3033 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE); 3034 else if (flag & FLAG_SND_UNA_ADVANCED) 3035 tcp_reset_reno_sack(tp); 3036 } 3037 *rexmit = REXMIT_LOST; 3038 } 3039 3040 /* Undo during fast recovery after partial ACK. */ 3041 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una) 3042 { 3043 struct tcp_sock *tp = tcp_sk(sk); 3044 3045 if (tp->undo_marker && tcp_packet_delayed(tp)) { 3046 /* Plain luck! Hole if filled with delayed 3047 * packet, rather than with a retransmit. Check reordering. 3048 */ 3049 tcp_check_sack_reordering(sk, prior_snd_una, 1); 3050 3051 /* We are getting evidence that the reordering degree is higher 3052 * than we realized. If there are no retransmits out then we 3053 * can undo. Otherwise we clock out new packets but do not 3054 * mark more packets lost or retransmit more. 3055 */ 3056 if (tp->retrans_out) 3057 return true; 3058 3059 if (!tcp_any_retrans_done(sk)) 3060 tp->retrans_stamp = 0; 3061 3062 DBGUNDO(sk, "partial recovery"); 3063 tcp_undo_cwnd_reduction(sk, true); 3064 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 3065 tcp_try_keep_open(sk); 3066 } 3067 return false; 3068 } 3069 3070 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag) 3071 { 3072 struct tcp_sock *tp = tcp_sk(sk); 3073 3074 if (tcp_rtx_queue_empty(sk)) 3075 return; 3076 3077 if (unlikely(tcp_is_reno(tp))) { 3078 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED); 3079 } else { 3080 u32 prior_retrans = tp->retrans_out; 3081 3082 if (tcp_rack_mark_lost(sk)) 3083 *ack_flag &= ~FLAG_SET_XMIT_TIMER; 3084 if (prior_retrans > tp->retrans_out) 3085 *ack_flag |= FLAG_LOST_RETRANS; 3086 } 3087 } 3088 3089 /* Process an event, which can update packets-in-flight not trivially. 3090 * Main goal of this function is to calculate new estimate for left_out, 3091 * taking into account both packets sitting in receiver's buffer and 3092 * packets lost by network. 3093 * 3094 * Besides that it updates the congestion state when packet loss or ECN 3095 * is detected. But it does not reduce the cwnd, it is done by the 3096 * congestion control later. 3097 * 3098 * It does _not_ decide what to send, it is made in function 3099 * tcp_xmit_retransmit_queue(). 3100 */ 3101 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una, 3102 int num_dupack, int *ack_flag, int *rexmit) 3103 { 3104 struct inet_connection_sock *icsk = inet_csk(sk); 3105 struct tcp_sock *tp = tcp_sk(sk); 3106 int flag = *ack_flag; 3107 bool ece_ack = flag & FLAG_ECE; 3108 3109 if (!tp->packets_out && tp->sacked_out) 3110 tp->sacked_out = 0; 3111 3112 /* Now state machine starts. 3113 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 3114 if (ece_ack) 3115 tp->prior_ssthresh = 0; 3116 3117 /* B. In all the states check for reneging SACKs. */ 3118 if (tcp_check_sack_reneging(sk, ack_flag)) 3119 return; 3120 3121 /* C. Check consistency of the current state. */ 3122 tcp_verify_left_out(tp); 3123 3124 /* D. Check state exit conditions. State can be terminated 3125 * when high_seq is ACKed. */ 3126 if (icsk->icsk_ca_state == TCP_CA_Open) { 3127 WARN_ON(tp->retrans_out != 0 && !tp->syn_data); 3128 tp->retrans_stamp = 0; 3129 } else if (!before(tp->snd_una, tp->high_seq)) { 3130 switch (icsk->icsk_ca_state) { 3131 case TCP_CA_CWR: 3132 /* CWR is to be held something *above* high_seq 3133 * is ACKed for CWR bit to reach receiver. */ 3134 if (tp->snd_una != tp->high_seq) { 3135 tcp_end_cwnd_reduction(sk); 3136 tcp_set_ca_state(sk, TCP_CA_Open); 3137 } 3138 break; 3139 3140 case TCP_CA_Recovery: 3141 if (tcp_is_reno(tp)) 3142 tcp_reset_reno_sack(tp); 3143 if (tcp_try_undo_recovery(sk)) 3144 return; 3145 tcp_end_cwnd_reduction(sk); 3146 break; 3147 } 3148 } 3149 3150 /* E. Process state. */ 3151 switch (icsk->icsk_ca_state) { 3152 case TCP_CA_Recovery: 3153 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 3154 if (tcp_is_reno(tp)) 3155 tcp_add_reno_sack(sk, num_dupack, ece_ack); 3156 } else if (tcp_try_undo_partial(sk, prior_snd_una)) 3157 return; 3158 3159 if (tcp_try_undo_dsack(sk)) 3160 tcp_try_to_open(sk, flag); 3161 3162 tcp_identify_packet_loss(sk, ack_flag); 3163 if (icsk->icsk_ca_state != TCP_CA_Recovery) { 3164 if (!tcp_time_to_recover(tp)) 3165 return; 3166 /* Undo reverts the recovery state. If loss is evident, 3167 * starts a new recovery (e.g. reordering then loss); 3168 */ 3169 tcp_enter_recovery(sk, ece_ack); 3170 } 3171 break; 3172 case TCP_CA_Loss: 3173 tcp_process_loss(sk, flag, num_dupack, rexmit); 3174 if (icsk->icsk_ca_state != TCP_CA_Loss) 3175 tcp_update_rto_time(tp); 3176 tcp_identify_packet_loss(sk, ack_flag); 3177 if (!(icsk->icsk_ca_state == TCP_CA_Open || 3178 (*ack_flag & FLAG_LOST_RETRANS))) 3179 return; 3180 /* Change state if cwnd is undone or retransmits are lost */ 3181 fallthrough; 3182 default: 3183 if (tcp_is_reno(tp)) { 3184 if (flag & FLAG_SND_UNA_ADVANCED) 3185 tcp_reset_reno_sack(tp); 3186 tcp_add_reno_sack(sk, num_dupack, ece_ack); 3187 } 3188 3189 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 3190 tcp_try_undo_dsack(sk); 3191 3192 tcp_identify_packet_loss(sk, ack_flag); 3193 if (!tcp_time_to_recover(tp)) { 3194 tcp_try_to_open(sk, flag); 3195 return; 3196 } 3197 3198 /* MTU probe failure: don't reduce cwnd */ 3199 if (icsk->icsk_ca_state < TCP_CA_CWR && 3200 icsk->icsk_mtup.probe_size && 3201 tp->snd_una == tp->mtu_probe.probe_seq_start) { 3202 tcp_mtup_probe_failed(sk); 3203 /* Restores the reduction we did in tcp_mtup_probe() */ 3204 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1); 3205 tcp_simple_retransmit(sk); 3206 return; 3207 } 3208 3209 /* Otherwise enter Recovery state */ 3210 tcp_enter_recovery(sk, ece_ack); 3211 } 3212 3213 *rexmit = REXMIT_LOST; 3214 } 3215 3216 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag) 3217 { 3218 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ; 3219 struct tcp_sock *tp = tcp_sk(sk); 3220 3221 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) { 3222 /* If the remote keeps returning delayed ACKs, eventually 3223 * the min filter would pick it up and overestimate the 3224 * prop. delay when it expires. Skip suspected delayed ACKs. 3225 */ 3226 return; 3227 } 3228 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32, 3229 rtt_us ? : jiffies_to_usecs(1)); 3230 } 3231 3232 static bool tcp_ack_update_rtt(struct sock *sk, const int flag, 3233 long seq_rtt_us, long sack_rtt_us, 3234 long ca_rtt_us, struct rate_sample *rs) 3235 { 3236 const struct tcp_sock *tp = tcp_sk(sk); 3237 3238 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because 3239 * broken middle-boxes or peers may corrupt TS-ECR fields. But 3240 * Karn's algorithm forbids taking RTT if some retransmitted data 3241 * is acked (RFC6298). 3242 */ 3243 if (seq_rtt_us < 0) 3244 seq_rtt_us = sack_rtt_us; 3245 3246 /* RTTM Rule: A TSecr value received in a segment is used to 3247 * update the averaged RTT measurement only if the segment 3248 * acknowledges some new data, i.e., only if it advances the 3249 * left edge of the send window. 3250 * See draft-ietf-tcplw-high-performance-00, section 3.3. 3251 */ 3252 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && 3253 tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED) 3254 seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp, 1); 3255 3256 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */ 3257 if (seq_rtt_us < 0) 3258 return false; 3259 3260 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is 3261 * always taken together with ACK, SACK, or TS-opts. Any negative 3262 * values will be skipped with the seq_rtt_us < 0 check above. 3263 */ 3264 tcp_update_rtt_min(sk, ca_rtt_us, flag); 3265 tcp_rtt_estimator(sk, seq_rtt_us); 3266 tcp_set_rto(sk); 3267 3268 /* RFC6298: only reset backoff on valid RTT measurement. */ 3269 inet_csk(sk)->icsk_backoff = 0; 3270 return true; 3271 } 3272 3273 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ 3274 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) 3275 { 3276 struct rate_sample rs; 3277 long rtt_us = -1L; 3278 3279 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack) 3280 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack); 3281 3282 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs); 3283 } 3284 3285 3286 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 3287 { 3288 const struct inet_connection_sock *icsk = inet_csk(sk); 3289 3290 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked); 3291 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32; 3292 } 3293 3294 /* Restart timer after forward progress on connection. 3295 * RFC2988 recommends to restart timer to now+rto. 3296 */ 3297 void tcp_rearm_rto(struct sock *sk) 3298 { 3299 const struct inet_connection_sock *icsk = inet_csk(sk); 3300 struct tcp_sock *tp = tcp_sk(sk); 3301 3302 /* If the retrans timer is currently being used by Fast Open 3303 * for SYN-ACK retrans purpose, stay put. 3304 */ 3305 if (rcu_access_pointer(tp->fastopen_rsk)) 3306 return; 3307 3308 if (!tp->packets_out) { 3309 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3310 } else { 3311 u32 rto = inet_csk(sk)->icsk_rto; 3312 /* Offset the time elapsed after installing regular RTO */ 3313 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT || 3314 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 3315 s64 delta_us = tcp_rto_delta_us(sk); 3316 /* delta_us may not be positive if the socket is locked 3317 * when the retrans timer fires and is rescheduled. 3318 */ 3319 rto = usecs_to_jiffies(max_t(int, delta_us, 1)); 3320 } 3321 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, true); 3322 } 3323 } 3324 3325 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */ 3326 static void tcp_set_xmit_timer(struct sock *sk) 3327 { 3328 if (!tcp_schedule_loss_probe(sk, true)) 3329 tcp_rearm_rto(sk); 3330 } 3331 3332 /* If we get here, the whole TSO packet has not been acked. */ 3333 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3334 { 3335 struct tcp_sock *tp = tcp_sk(sk); 3336 u32 packets_acked; 3337 3338 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3339 3340 packets_acked = tcp_skb_pcount(skb); 3341 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3342 return 0; 3343 packets_acked -= tcp_skb_pcount(skb); 3344 3345 if (packets_acked) { 3346 BUG_ON(tcp_skb_pcount(skb) == 0); 3347 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3348 } 3349 3350 return packets_acked; 3351 } 3352 3353 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb, 3354 const struct sk_buff *ack_skb, u32 prior_snd_una) 3355 { 3356 const struct skb_shared_info *shinfo; 3357 3358 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */ 3359 if (likely(!TCP_SKB_CB(skb)->txstamp_ack)) 3360 return; 3361 3362 shinfo = skb_shinfo(skb); 3363 if (!before(shinfo->tskey, prior_snd_una) && 3364 before(shinfo->tskey, tcp_sk(sk)->snd_una)) { 3365 tcp_skb_tsorted_save(skb) { 3366 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK); 3367 } tcp_skb_tsorted_restore(skb); 3368 } 3369 } 3370 3371 /* Remove acknowledged frames from the retransmission queue. If our packet 3372 * is before the ack sequence we can discard it as it's confirmed to have 3373 * arrived at the other end. 3374 */ 3375 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb, 3376 u32 prior_fack, u32 prior_snd_una, 3377 struct tcp_sacktag_state *sack, bool ece_ack) 3378 { 3379 const struct inet_connection_sock *icsk = inet_csk(sk); 3380 u64 first_ackt, last_ackt; 3381 struct tcp_sock *tp = tcp_sk(sk); 3382 u32 prior_sacked = tp->sacked_out; 3383 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */ 3384 struct sk_buff *skb, *next; 3385 bool fully_acked = true; 3386 long sack_rtt_us = -1L; 3387 long seq_rtt_us = -1L; 3388 long ca_rtt_us = -1L; 3389 u32 pkts_acked = 0; 3390 bool rtt_update; 3391 int flag = 0; 3392 3393 first_ackt = 0; 3394 3395 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) { 3396 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3397 const u32 start_seq = scb->seq; 3398 u8 sacked = scb->sacked; 3399 u32 acked_pcount; 3400 3401 /* Determine how many packets and what bytes were acked, tso and else */ 3402 if (after(scb->end_seq, tp->snd_una)) { 3403 if (tcp_skb_pcount(skb) == 1 || 3404 !after(tp->snd_una, scb->seq)) 3405 break; 3406 3407 acked_pcount = tcp_tso_acked(sk, skb); 3408 if (!acked_pcount) 3409 break; 3410 fully_acked = false; 3411 } else { 3412 acked_pcount = tcp_skb_pcount(skb); 3413 } 3414 3415 if (unlikely(sacked & TCPCB_RETRANS)) { 3416 if (sacked & TCPCB_SACKED_RETRANS) 3417 tp->retrans_out -= acked_pcount; 3418 flag |= FLAG_RETRANS_DATA_ACKED; 3419 } else if (!(sacked & TCPCB_SACKED_ACKED)) { 3420 last_ackt = tcp_skb_timestamp_us(skb); 3421 WARN_ON_ONCE(last_ackt == 0); 3422 if (!first_ackt) 3423 first_ackt = last_ackt; 3424 3425 if (before(start_seq, reord)) 3426 reord = start_seq; 3427 if (!after(scb->end_seq, tp->high_seq)) 3428 flag |= FLAG_ORIG_SACK_ACKED; 3429 } 3430 3431 if (sacked & TCPCB_SACKED_ACKED) { 3432 tp->sacked_out -= acked_pcount; 3433 /* snd_una delta covers these skbs */ 3434 sack->delivered_bytes -= skb->len; 3435 } else if (tcp_is_sack(tp)) { 3436 tcp_count_delivered(tp, acked_pcount, ece_ack); 3437 if (!tcp_skb_spurious_retrans(tp, skb)) 3438 tcp_rack_advance(tp, sacked, scb->end_seq, 3439 tcp_skb_timestamp_us(skb)); 3440 } 3441 if (sacked & TCPCB_LOST) 3442 tp->lost_out -= acked_pcount; 3443 3444 tp->packets_out -= acked_pcount; 3445 pkts_acked += acked_pcount; 3446 tcp_rate_skb_delivered(sk, skb, sack->rate); 3447 3448 /* Initial outgoing SYN's get put onto the write_queue 3449 * just like anything else we transmit. It is not 3450 * true data, and if we misinform our callers that 3451 * this ACK acks real data, we will erroneously exit 3452 * connection startup slow start one packet too 3453 * quickly. This is severely frowned upon behavior. 3454 */ 3455 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) { 3456 flag |= FLAG_DATA_ACKED; 3457 } else { 3458 flag |= FLAG_SYN_ACKED; 3459 tp->retrans_stamp = 0; 3460 } 3461 3462 if (!fully_acked) 3463 break; 3464 3465 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); 3466 3467 next = skb_rb_next(skb); 3468 if (unlikely(skb == tp->retransmit_skb_hint)) 3469 tp->retransmit_skb_hint = NULL; 3470 tcp_highest_sack_replace(sk, skb, next); 3471 tcp_rtx_queue_unlink_and_free(skb, sk); 3472 } 3473 3474 if (!skb) 3475 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 3476 3477 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3478 tp->snd_up = tp->snd_una; 3479 3480 if (skb) { 3481 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); 3482 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3483 flag |= FLAG_SACK_RENEGING; 3484 } 3485 3486 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) { 3487 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt); 3488 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt); 3489 3490 if (pkts_acked == 1 && fully_acked && !prior_sacked && 3491 (tp->snd_una - prior_snd_una) < tp->mss_cache && 3492 sack->rate->prior_delivered + 1 == tp->delivered && 3493 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) { 3494 /* Conservatively mark a delayed ACK. It's typically 3495 * from a lone runt packet over the round trip to 3496 * a receiver w/o out-of-order or CE events. 3497 */ 3498 flag |= FLAG_ACK_MAYBE_DELAYED; 3499 } 3500 } 3501 if (sack->first_sackt) { 3502 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt); 3503 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt); 3504 } 3505 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us, 3506 ca_rtt_us, sack->rate); 3507 3508 if (flag & FLAG_ACKED) { 3509 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3510 if (unlikely(icsk->icsk_mtup.probe_size && 3511 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3512 tcp_mtup_probe_success(sk); 3513 } 3514 3515 if (tcp_is_reno(tp)) { 3516 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack); 3517 3518 /* If any of the cumulatively ACKed segments was 3519 * retransmitted, non-SACK case cannot confirm that 3520 * progress was due to original transmission due to 3521 * lack of TCPCB_SACKED_ACKED bits even if some of 3522 * the packets may have been never retransmitted. 3523 */ 3524 if (flag & FLAG_RETRANS_DATA_ACKED) 3525 flag &= ~FLAG_ORIG_SACK_ACKED; 3526 } else { 3527 /* Non-retransmitted hole got filled? That's reordering */ 3528 if (before(reord, prior_fack)) 3529 tcp_check_sack_reordering(sk, reord, 0); 3530 } 3531 3532 sack->delivered_bytes = (skb ? 3533 TCP_SKB_CB(skb)->seq : tp->snd_una) - 3534 prior_snd_una; 3535 } else if (skb && rtt_update && sack_rtt_us >= 0 && 3536 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, 3537 tcp_skb_timestamp_us(skb))) { 3538 /* Do not re-arm RTO if the sack RTT is measured from data sent 3539 * after when the head was last (re)transmitted. Otherwise the 3540 * timeout may continue to extend in loss recovery. 3541 */ 3542 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3543 } 3544 3545 if (icsk->icsk_ca_ops->pkts_acked) { 3546 struct ack_sample sample = { .pkts_acked = pkts_acked, 3547 .rtt_us = sack->rate->rtt_us }; 3548 3549 sample.in_flight = tp->mss_cache * 3550 (tp->delivered - sack->rate->prior_delivered); 3551 icsk->icsk_ca_ops->pkts_acked(sk, &sample); 3552 } 3553 3554 #if FASTRETRANS_DEBUG > 0 3555 WARN_ON((int)tp->sacked_out < 0); 3556 WARN_ON((int)tp->lost_out < 0); 3557 WARN_ON((int)tp->retrans_out < 0); 3558 if (!tp->packets_out && tcp_is_sack(tp)) { 3559 icsk = inet_csk(sk); 3560 if (tp->lost_out) { 3561 pr_debug("Leak l=%u %d\n", 3562 tp->lost_out, icsk->icsk_ca_state); 3563 tp->lost_out = 0; 3564 } 3565 if (tp->sacked_out) { 3566 pr_debug("Leak s=%u %d\n", 3567 tp->sacked_out, icsk->icsk_ca_state); 3568 tp->sacked_out = 0; 3569 } 3570 if (tp->retrans_out) { 3571 pr_debug("Leak r=%u %d\n", 3572 tp->retrans_out, icsk->icsk_ca_state); 3573 tp->retrans_out = 0; 3574 } 3575 } 3576 #endif 3577 return flag; 3578 } 3579 3580 static void tcp_ack_probe(struct sock *sk) 3581 { 3582 struct inet_connection_sock *icsk = inet_csk(sk); 3583 struct sk_buff *head = tcp_send_head(sk); 3584 const struct tcp_sock *tp = tcp_sk(sk); 3585 3586 /* Was it a usable window open? */ 3587 if (!head) 3588 return; 3589 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) { 3590 icsk->icsk_backoff = 0; 3591 icsk->icsk_probes_tstamp = 0; 3592 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3593 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3594 * This function is not for random using! 3595 */ 3596 } else { 3597 unsigned long when = tcp_probe0_when(sk, tcp_rto_max(sk)); 3598 3599 when = tcp_clamp_probe0_to_user_timeout(sk, when); 3600 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, true); 3601 } 3602 } 3603 3604 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) 3605 { 3606 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3607 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3608 } 3609 3610 /* Decide wheather to run the increase function of congestion control. */ 3611 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3612 { 3613 /* If reordering is high then always grow cwnd whenever data is 3614 * delivered regardless of its ordering. Otherwise stay conservative 3615 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ 3616 * new SACK or ECE mark may first advance cwnd here and later reduce 3617 * cwnd in tcp_fastretrans_alert() based on more states. 3618 */ 3619 if (tcp_sk(sk)->reordering > 3620 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering)) 3621 return flag & FLAG_FORWARD_PROGRESS; 3622 3623 return flag & FLAG_DATA_ACKED; 3624 } 3625 3626 /* The "ultimate" congestion control function that aims to replace the rigid 3627 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction). 3628 * It's called toward the end of processing an ACK with precise rate 3629 * information. All transmission or retransmission are delayed afterwards. 3630 */ 3631 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked, 3632 int flag, const struct rate_sample *rs) 3633 { 3634 const struct inet_connection_sock *icsk = inet_csk(sk); 3635 3636 if (icsk->icsk_ca_ops->cong_control) { 3637 icsk->icsk_ca_ops->cong_control(sk, ack, flag, rs); 3638 return; 3639 } 3640 3641 if (tcp_in_cwnd_reduction(sk)) { 3642 /* Reduce cwnd if state mandates */ 3643 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag); 3644 } else if (tcp_may_raise_cwnd(sk, flag)) { 3645 /* Advance cwnd if state allows */ 3646 tcp_cong_avoid(sk, ack, acked_sacked); 3647 } 3648 tcp_update_pacing_rate(sk); 3649 } 3650 3651 /* Check that window update is acceptable. 3652 * The function assumes that snd_una<=ack<=snd_next. 3653 */ 3654 static inline bool tcp_may_update_window(const struct tcp_sock *tp, 3655 const u32 ack, const u32 ack_seq, 3656 const u32 nwin) 3657 { 3658 return after(ack, tp->snd_una) || 3659 after(ack_seq, tp->snd_wl1) || 3660 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin)); 3661 } 3662 3663 static void tcp_snd_sne_update(struct tcp_sock *tp, u32 ack) 3664 { 3665 #ifdef CONFIG_TCP_AO 3666 struct tcp_ao_info *ao; 3667 3668 if (!static_branch_unlikely(&tcp_ao_needed.key)) 3669 return; 3670 3671 ao = rcu_dereference_protected(tp->ao_info, 3672 lockdep_sock_is_held((struct sock *)tp)); 3673 if (ao && ack < tp->snd_una) { 3674 ao->snd_sne++; 3675 trace_tcp_ao_snd_sne_update((struct sock *)tp, ao->snd_sne); 3676 } 3677 #endif 3678 } 3679 3680 /* If we update tp->snd_una, also update tp->bytes_acked */ 3681 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack) 3682 { 3683 u32 delta = ack - tp->snd_una; 3684 3685 sock_owned_by_me((struct sock *)tp); 3686 tp->bytes_acked += delta; 3687 tcp_snd_sne_update(tp, ack); 3688 tp->snd_una = ack; 3689 } 3690 3691 static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq) 3692 { 3693 #ifdef CONFIG_TCP_AO 3694 struct tcp_ao_info *ao; 3695 3696 if (!static_branch_unlikely(&tcp_ao_needed.key)) 3697 return; 3698 3699 ao = rcu_dereference_protected(tp->ao_info, 3700 lockdep_sock_is_held((struct sock *)tp)); 3701 if (ao && seq < tp->rcv_nxt) { 3702 ao->rcv_sne++; 3703 trace_tcp_ao_rcv_sne_update((struct sock *)tp, ao->rcv_sne); 3704 } 3705 #endif 3706 } 3707 3708 /* If we update tp->rcv_nxt, also update tp->bytes_received */ 3709 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq) 3710 { 3711 u32 delta = seq - tp->rcv_nxt; 3712 3713 sock_owned_by_me((struct sock *)tp); 3714 tp->bytes_received += delta; 3715 tcp_rcv_sne_update(tp, seq); 3716 WRITE_ONCE(tp->rcv_nxt, seq); 3717 } 3718 3719 /* Update our send window. 3720 * 3721 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3722 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3723 */ 3724 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3725 u32 ack_seq) 3726 { 3727 struct tcp_sock *tp = tcp_sk(sk); 3728 int flag = 0; 3729 u32 nwin = ntohs(tcp_hdr(skb)->window); 3730 3731 if (likely(!tcp_hdr(skb)->syn)) 3732 nwin <<= tp->rx_opt.snd_wscale; 3733 3734 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3735 flag |= FLAG_WIN_UPDATE; 3736 tcp_update_wl(tp, ack_seq); 3737 3738 if (tp->snd_wnd != nwin) { 3739 tp->snd_wnd = nwin; 3740 3741 /* Note, it is the only place, where 3742 * fast path is recovered for sending TCP. 3743 */ 3744 tp->pred_flags = 0; 3745 tcp_fast_path_check(sk); 3746 3747 if (!tcp_write_queue_empty(sk)) 3748 tcp_slow_start_after_idle_check(sk); 3749 3750 if (nwin > tp->max_window) { 3751 tp->max_window = nwin; 3752 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3753 } 3754 } 3755 } 3756 3757 tcp_snd_una_update(tp, ack); 3758 3759 return flag; 3760 } 3761 3762 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx, 3763 u32 *last_oow_ack_time) 3764 { 3765 /* Paired with the WRITE_ONCE() in this function. */ 3766 u32 val = READ_ONCE(*last_oow_ack_time); 3767 3768 if (val) { 3769 s32 elapsed = (s32)(tcp_jiffies32 - val); 3770 3771 if (0 <= elapsed && 3772 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) { 3773 NET_INC_STATS(net, mib_idx); 3774 return true; /* rate-limited: don't send yet! */ 3775 } 3776 } 3777 3778 /* Paired with the prior READ_ONCE() and with itself, 3779 * as we might be lockless. 3780 */ 3781 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32); 3782 3783 return false; /* not rate-limited: go ahead, send dupack now! */ 3784 } 3785 3786 /* Return true if we're currently rate-limiting out-of-window ACKs and 3787 * thus shouldn't send a dupack right now. We rate-limit dupacks in 3788 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS 3789 * attacks that send repeated SYNs or ACKs for the same connection. To 3790 * do this, we do not send a duplicate SYNACK or ACK if the remote 3791 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate. 3792 */ 3793 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb, 3794 int mib_idx, u32 *last_oow_ack_time) 3795 { 3796 /* Data packets without SYNs are not likely part of an ACK loop. */ 3797 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) && 3798 !tcp_hdr(skb)->syn) 3799 return false; 3800 3801 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time); 3802 } 3803 3804 static void tcp_send_ack_reflect_ect(struct sock *sk, bool accecn_reflector) 3805 { 3806 struct tcp_sock *tp = tcp_sk(sk); 3807 u16 flags = 0; 3808 3809 if (accecn_reflector) 3810 flags = tcp_accecn_reflector_flags(tp->syn_ect_rcv); 3811 __tcp_send_ack(sk, tp->rcv_nxt, flags); 3812 } 3813 3814 /* RFC 5961 7 [ACK Throttling] */ 3815 static void tcp_send_challenge_ack(struct sock *sk, bool accecn_reflector) 3816 { 3817 struct tcp_sock *tp = tcp_sk(sk); 3818 struct net *net = sock_net(sk); 3819 u32 count, now, ack_limit; 3820 3821 /* First check our per-socket dupack rate limit. */ 3822 if (__tcp_oow_rate_limited(net, 3823 LINUX_MIB_TCPACKSKIPPEDCHALLENGE, 3824 &tp->last_oow_ack_time)) 3825 return; 3826 3827 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit); 3828 if (ack_limit == INT_MAX) 3829 goto send_ack; 3830 3831 /* Then check host-wide RFC 5961 rate limit. */ 3832 now = jiffies / HZ; 3833 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) { 3834 u32 half = (ack_limit + 1) >> 1; 3835 3836 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now); 3837 WRITE_ONCE(net->ipv4.tcp_challenge_count, 3838 get_random_u32_inclusive(half, ack_limit + half - 1)); 3839 } 3840 count = READ_ONCE(net->ipv4.tcp_challenge_count); 3841 if (count > 0) { 3842 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1); 3843 send_ack: 3844 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK); 3845 tcp_send_ack_reflect_ect(sk, accecn_reflector); 3846 } 3847 } 3848 3849 static void tcp_store_ts_recent(struct tcp_sock *tp) 3850 { 3851 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3852 tp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 3853 } 3854 3855 static int __tcp_replace_ts_recent(struct tcp_sock *tp, s32 tstamp_delta) 3856 { 3857 tcp_store_ts_recent(tp); 3858 return tstamp_delta > 0 ? FLAG_TS_PROGRESS : 0; 3859 } 3860 3861 static int tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3862 { 3863 s32 delta; 3864 3865 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3866 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3867 * extra check below makes sure this can only happen 3868 * for pure ACK frames. -DaveM 3869 * 3870 * Not only, also it occurs for expired timestamps. 3871 */ 3872 3873 if (tcp_paws_check(&tp->rx_opt, 0)) { 3874 delta = tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent; 3875 return __tcp_replace_ts_recent(tp, delta); 3876 } 3877 } 3878 3879 return 0; 3880 } 3881 3882 /* This routine deals with acks during a TLP episode and ends an episode by 3883 * resetting tlp_high_seq. Ref: TLP algorithm in RFC8985 3884 */ 3885 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) 3886 { 3887 struct tcp_sock *tp = tcp_sk(sk); 3888 3889 if (before(ack, tp->tlp_high_seq)) 3890 return; 3891 3892 if (!tp->tlp_retrans) { 3893 /* TLP of new data has been acknowledged */ 3894 tp->tlp_high_seq = 0; 3895 } else if (flag & FLAG_DSACK_TLP) { 3896 /* This DSACK means original and TLP probe arrived; no loss */ 3897 tp->tlp_high_seq = 0; 3898 } else if (after(ack, tp->tlp_high_seq)) { 3899 /* ACK advances: there was a loss, so reduce cwnd. Reset 3900 * tlp_high_seq in tcp_init_cwnd_reduction() 3901 */ 3902 tcp_init_cwnd_reduction(sk); 3903 tcp_set_ca_state(sk, TCP_CA_CWR); 3904 tcp_end_cwnd_reduction(sk); 3905 tcp_try_keep_open(sk); 3906 NET_INC_STATS(sock_net(sk), 3907 LINUX_MIB_TCPLOSSPROBERECOVERY); 3908 } else if (!(flag & (FLAG_SND_UNA_ADVANCED | 3909 FLAG_NOT_DUP | FLAG_DATA_SACKED))) { 3910 /* Pure dupack: original and TLP probe arrived; no loss */ 3911 tp->tlp_high_seq = 0; 3912 } 3913 } 3914 3915 static void tcp_in_ack_event(struct sock *sk, int flag) 3916 { 3917 const struct inet_connection_sock *icsk = inet_csk(sk); 3918 3919 if (icsk->icsk_ca_ops->in_ack_event) { 3920 u32 ack_ev_flags = 0; 3921 3922 if (flag & FLAG_WIN_UPDATE) 3923 ack_ev_flags |= CA_ACK_WIN_UPDATE; 3924 if (flag & FLAG_SLOWPATH) { 3925 ack_ev_flags |= CA_ACK_SLOWPATH; 3926 if (flag & FLAG_ECE) 3927 ack_ev_flags |= CA_ACK_ECE; 3928 } 3929 3930 icsk->icsk_ca_ops->in_ack_event(sk, ack_ev_flags); 3931 } 3932 } 3933 3934 /* Congestion control has updated the cwnd already. So if we're in 3935 * loss recovery then now we do any new sends (for FRTO) or 3936 * retransmits (for CA_Loss or CA_recovery) that make sense. 3937 */ 3938 static void tcp_xmit_recovery(struct sock *sk, int rexmit) 3939 { 3940 struct tcp_sock *tp = tcp_sk(sk); 3941 3942 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT) 3943 return; 3944 3945 if (unlikely(rexmit == REXMIT_NEW)) { 3946 __tcp_push_pending_frames(sk, tcp_current_mss(sk), 3947 TCP_NAGLE_OFF); 3948 if (after(tp->snd_nxt, tp->high_seq)) 3949 return; 3950 tp->frto = 0; 3951 } 3952 tcp_xmit_retransmit_queue(sk); 3953 } 3954 3955 /* Returns the number of packets newly acked or sacked by the current ACK */ 3956 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, 3957 u32 ecn_count, int flag) 3958 { 3959 const struct net *net = sock_net(sk); 3960 struct tcp_sock *tp = tcp_sk(sk); 3961 u32 delivered; 3962 3963 delivered = tp->delivered - prior_delivered; 3964 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered); 3965 3966 if (flag & FLAG_ECE) { 3967 if (tcp_ecn_mode_rfc3168(tp)) 3968 ecn_count = delivered; 3969 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, ecn_count); 3970 } 3971 3972 return delivered; 3973 } 3974 3975 /* This routine deals with incoming acks, but not outgoing ones. */ 3976 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 3977 { 3978 struct inet_connection_sock *icsk = inet_csk(sk); 3979 struct tcp_sock *tp = tcp_sk(sk); 3980 struct tcp_sacktag_state sack_state; 3981 struct rate_sample rs = { .prior_delivered = 0 }; 3982 u32 prior_snd_una = tp->snd_una; 3983 bool is_sack_reneg = tp->is_sack_reneg; 3984 u32 ack_seq = TCP_SKB_CB(skb)->seq; 3985 u32 ack = TCP_SKB_CB(skb)->ack_seq; 3986 int num_dupack = 0; 3987 int prior_packets = tp->packets_out; 3988 u32 delivered = tp->delivered; 3989 u32 lost = tp->lost; 3990 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */ 3991 u32 ecn_count = 0; /* Did we receive ECE/an AccECN ACE update? */ 3992 u32 prior_fack; 3993 3994 sack_state.first_sackt = 0; 3995 sack_state.rate = &rs; 3996 sack_state.sack_delivered = 0; 3997 sack_state.delivered_bytes = 0; 3998 3999 /* We very likely will need to access rtx queue. */ 4000 prefetch(sk->tcp_rtx_queue.rb_node); 4001 4002 /* If the ack is older than previous acks 4003 * then we can probably ignore it. 4004 */ 4005 if (before(ack, prior_snd_una)) { 4006 u32 max_window; 4007 4008 /* do not accept ACK for bytes we never sent. */ 4009 max_window = min_t(u64, tp->max_window, tp->bytes_acked); 4010 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */ 4011 if (before(ack, prior_snd_una - max_window)) { 4012 if (!(flag & FLAG_NO_CHALLENGE_ACK)) 4013 tcp_send_challenge_ack(sk, false); 4014 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK; 4015 } 4016 goto old_ack; 4017 } 4018 4019 /* If the ack includes data we haven't sent yet, discard 4020 * this segment (RFC793 Section 3.9). 4021 */ 4022 if (after(ack, tp->snd_nxt)) 4023 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA; 4024 4025 if (after(ack, prior_snd_una)) { 4026 flag |= FLAG_SND_UNA_ADVANCED; 4027 WRITE_ONCE(icsk->icsk_retransmits, 0); 4028 4029 #if IS_ENABLED(CONFIG_TLS_DEVICE) 4030 if (static_branch_unlikely(&clean_acked_data_enabled.key)) 4031 if (tp->tcp_clean_acked) 4032 tp->tcp_clean_acked(sk, ack); 4033 #endif 4034 } 4035 4036 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una; 4037 rs.prior_in_flight = tcp_packets_in_flight(tp); 4038 4039 /* ts_recent update must be made after we are sure that the packet 4040 * is in window. 4041 */ 4042 if (flag & FLAG_UPDATE_TS_RECENT) 4043 flag |= tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4044 4045 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) == 4046 FLAG_SND_UNA_ADVANCED) { 4047 /* Window is constant, pure forward advance. 4048 * No more checks are required. 4049 * Note, we use the fact that SND.UNA>=SND.WL2. 4050 */ 4051 tcp_update_wl(tp, ack_seq); 4052 tcp_snd_una_update(tp, ack); 4053 flag |= FLAG_WIN_UPDATE; 4054 4055 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS); 4056 } else { 4057 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 4058 flag |= FLAG_DATA; 4059 else 4060 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS); 4061 4062 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 4063 4064 if (TCP_SKB_CB(skb)->sacked) 4065 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 4066 &sack_state); 4067 4068 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) 4069 flag |= FLAG_ECE; 4070 4071 if (sack_state.sack_delivered) 4072 tcp_count_delivered(tp, sack_state.sack_delivered, 4073 flag & FLAG_ECE); 4074 } 4075 4076 /* This is a deviation from RFC3168 since it states that: 4077 * "When the TCP data sender is ready to set the CWR bit after reducing 4078 * the congestion window, it SHOULD set the CWR bit only on the first 4079 * new data packet that it transmits." 4080 * We accept CWR on pure ACKs to be more robust 4081 * with widely-deployed TCP implementations that do this. 4082 */ 4083 tcp_ecn_accept_cwr(sk, skb); 4084 4085 /* We passed data and got it acked, remove any soft error 4086 * log. Something worked... 4087 */ 4088 if (READ_ONCE(sk->sk_err_soft)) 4089 WRITE_ONCE(sk->sk_err_soft, 0); 4090 WRITE_ONCE(icsk->icsk_probes_out, 0); 4091 tp->rcv_tstamp = tcp_jiffies32; 4092 if (!prior_packets) 4093 goto no_queue; 4094 4095 /* See if we can take anything off of the retransmit queue. */ 4096 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una, 4097 &sack_state, flag & FLAG_ECE); 4098 4099 tcp_rack_update_reo_wnd(sk, &rs); 4100 4101 if (tcp_ecn_mode_accecn(tp)) 4102 ecn_count = tcp_accecn_process(sk, skb, 4103 tp->delivered - delivered, 4104 sack_state.delivered_bytes, 4105 &flag); 4106 4107 tcp_in_ack_event(sk, flag); 4108 4109 if (tp->tlp_high_seq) 4110 tcp_process_tlp_ack(sk, ack, flag); 4111 4112 if (tcp_ack_is_dubious(sk, flag)) { 4113 if (!(flag & (FLAG_SND_UNA_ADVANCED | 4114 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) { 4115 num_dupack = 1; 4116 /* Consider if pure acks were aggregated in tcp_add_backlog() */ 4117 if (!(flag & FLAG_DATA)) 4118 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs); 4119 } 4120 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4121 &rexmit); 4122 } 4123 4124 /* If needed, reset TLP/RTO timer when RACK doesn't set. */ 4125 if (flag & FLAG_SET_XMIT_TIMER) 4126 tcp_set_xmit_timer(sk); 4127 4128 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 4129 sk_dst_confirm(sk); 4130 4131 delivered = tcp_newly_delivered(sk, delivered, ecn_count, flag); 4132 4133 lost = tp->lost - lost; /* freshly marked lost */ 4134 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED); 4135 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate); 4136 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate); 4137 tcp_xmit_recovery(sk, rexmit); 4138 return 1; 4139 4140 no_queue: 4141 if (tcp_ecn_mode_accecn(tp)) 4142 ecn_count = tcp_accecn_process(sk, skb, 4143 tp->delivered - delivered, 4144 sack_state.delivered_bytes, 4145 &flag); 4146 tcp_in_ack_event(sk, flag); 4147 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 4148 if (flag & FLAG_DSACKING_ACK) { 4149 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4150 &rexmit); 4151 tcp_newly_delivered(sk, delivered, ecn_count, flag); 4152 } 4153 /* If this ack opens up a zero window, clear backoff. It was 4154 * being used to time the probes, and is probably far higher than 4155 * it needs to be for normal retransmission. 4156 */ 4157 tcp_ack_probe(sk); 4158 4159 if (tp->tlp_high_seq) 4160 tcp_process_tlp_ack(sk, ack, flag); 4161 return 1; 4162 4163 old_ack: 4164 /* If data was SACKed, tag it and see if we should send more data. 4165 * If data was DSACKed, see if we can undo a cwnd reduction. 4166 */ 4167 if (TCP_SKB_CB(skb)->sacked) { 4168 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 4169 &sack_state); 4170 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4171 &rexmit); 4172 tcp_newly_delivered(sk, delivered, ecn_count, flag); 4173 tcp_xmit_recovery(sk, rexmit); 4174 } 4175 4176 return 0; 4177 } 4178 4179 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie, 4180 bool syn, struct tcp_fastopen_cookie *foc, 4181 bool exp_opt) 4182 { 4183 /* Valid only in SYN or SYN-ACK with an even length. */ 4184 if (!foc || !syn || len < 0 || (len & 1)) 4185 return; 4186 4187 if (len >= TCP_FASTOPEN_COOKIE_MIN && 4188 len <= TCP_FASTOPEN_COOKIE_MAX) 4189 memcpy(foc->val, cookie, len); 4190 else if (len != 0) 4191 len = -1; 4192 foc->len = len; 4193 foc->exp = exp_opt; 4194 } 4195 4196 static bool smc_parse_options(const struct tcphdr *th, 4197 struct tcp_options_received *opt_rx, 4198 const unsigned char *ptr, 4199 int opsize) 4200 { 4201 #if IS_ENABLED(CONFIG_SMC) 4202 if (static_branch_unlikely(&tcp_have_smc)) { 4203 if (th->syn && !(opsize & 1) && 4204 opsize >= TCPOLEN_EXP_SMC_BASE && 4205 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) { 4206 opt_rx->smc_ok = 1; 4207 return true; 4208 } 4209 } 4210 #endif 4211 return false; 4212 } 4213 4214 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped 4215 * value on success. 4216 */ 4217 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss) 4218 { 4219 const unsigned char *ptr = (const unsigned char *)(th + 1); 4220 int length = (th->doff * 4) - sizeof(struct tcphdr); 4221 u16 mss = 0; 4222 4223 while (length > 0) { 4224 int opcode = *ptr++; 4225 int opsize; 4226 4227 switch (opcode) { 4228 case TCPOPT_EOL: 4229 return mss; 4230 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 4231 length--; 4232 continue; 4233 default: 4234 if (length < 2) 4235 return mss; 4236 opsize = *ptr++; 4237 if (opsize < 2) /* "silly options" */ 4238 return mss; 4239 if (opsize > length) 4240 return mss; /* fail on partial options */ 4241 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) { 4242 u16 in_mss = get_unaligned_be16(ptr); 4243 4244 if (in_mss) { 4245 if (user_mss && user_mss < in_mss) 4246 in_mss = user_mss; 4247 mss = in_mss; 4248 } 4249 } 4250 ptr += opsize - 2; 4251 length -= opsize; 4252 } 4253 } 4254 return mss; 4255 } 4256 4257 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 4258 * But, this can also be called on packets in the established flow when 4259 * the fast version below fails. 4260 */ 4261 void tcp_parse_options(const struct net *net, 4262 const struct sk_buff *skb, 4263 struct tcp_options_received *opt_rx, int estab, 4264 struct tcp_fastopen_cookie *foc) 4265 { 4266 const unsigned char *ptr; 4267 const struct tcphdr *th = tcp_hdr(skb); 4268 int length = (th->doff * 4) - sizeof(struct tcphdr); 4269 4270 ptr = (const unsigned char *)(th + 1); 4271 opt_rx->saw_tstamp = 0; 4272 opt_rx->accecn = 0; 4273 opt_rx->saw_unknown = 0; 4274 4275 while (length > 0) { 4276 int opcode = *ptr++; 4277 int opsize; 4278 4279 switch (opcode) { 4280 case TCPOPT_EOL: 4281 return; 4282 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 4283 length--; 4284 continue; 4285 default: 4286 if (length < 2) 4287 return; 4288 opsize = *ptr++; 4289 if (opsize < 2) /* "silly options" */ 4290 return; 4291 if (opsize > length) 4292 return; /* don't parse partial options */ 4293 switch (opcode) { 4294 case TCPOPT_MSS: 4295 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 4296 u16 in_mss = get_unaligned_be16(ptr); 4297 if (in_mss) { 4298 if (opt_rx->user_mss && 4299 opt_rx->user_mss < in_mss) 4300 in_mss = opt_rx->user_mss; 4301 opt_rx->mss_clamp = in_mss; 4302 } 4303 } 4304 break; 4305 case TCPOPT_WINDOW: 4306 if (opsize == TCPOLEN_WINDOW && th->syn && 4307 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) { 4308 __u8 snd_wscale = *(__u8 *)ptr; 4309 opt_rx->wscale_ok = 1; 4310 if (snd_wscale > TCP_MAX_WSCALE) { 4311 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n", 4312 __func__, 4313 snd_wscale, 4314 TCP_MAX_WSCALE); 4315 snd_wscale = TCP_MAX_WSCALE; 4316 } 4317 opt_rx->snd_wscale = snd_wscale; 4318 } 4319 break; 4320 case TCPOPT_TIMESTAMP: 4321 if ((opsize == TCPOLEN_TIMESTAMP) && 4322 ((estab && opt_rx->tstamp_ok) || 4323 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) { 4324 opt_rx->saw_tstamp = 1; 4325 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 4326 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 4327 } 4328 break; 4329 case TCPOPT_SACK_PERM: 4330 if (opsize == TCPOLEN_SACK_PERM && th->syn && 4331 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) { 4332 opt_rx->sack_ok = TCP_SACK_SEEN; 4333 tcp_sack_reset(opt_rx); 4334 } 4335 break; 4336 4337 case TCPOPT_SACK: 4338 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 4339 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 4340 opt_rx->sack_ok) { 4341 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 4342 } 4343 break; 4344 #ifdef CONFIG_TCP_MD5SIG 4345 case TCPOPT_MD5SIG: 4346 /* The MD5 Hash has already been 4347 * checked (see tcp_v{4,6}_rcv()). 4348 */ 4349 break; 4350 #endif 4351 #ifdef CONFIG_TCP_AO 4352 case TCPOPT_AO: 4353 /* TCP AO has already been checked 4354 * (see tcp_inbound_ao_hash()). 4355 */ 4356 break; 4357 #endif 4358 case TCPOPT_FASTOPEN: 4359 tcp_parse_fastopen_option( 4360 opsize - TCPOLEN_FASTOPEN_BASE, 4361 ptr, th->syn, foc, false); 4362 break; 4363 4364 case TCPOPT_ACCECN0: 4365 case TCPOPT_ACCECN1: 4366 /* Save offset of AccECN option in TCP header */ 4367 opt_rx->accecn = (ptr - 2) - (__u8 *)th; 4368 break; 4369 4370 case TCPOPT_EXP: 4371 /* Fast Open option shares code 254 using a 4372 * 16 bits magic number. 4373 */ 4374 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE && 4375 get_unaligned_be16(ptr) == 4376 TCPOPT_FASTOPEN_MAGIC) { 4377 tcp_parse_fastopen_option(opsize - 4378 TCPOLEN_EXP_FASTOPEN_BASE, 4379 ptr + 2, th->syn, foc, true); 4380 break; 4381 } 4382 4383 if (smc_parse_options(th, opt_rx, ptr, opsize)) 4384 break; 4385 4386 opt_rx->saw_unknown = 1; 4387 break; 4388 4389 default: 4390 opt_rx->saw_unknown = 1; 4391 } 4392 ptr += opsize-2; 4393 length -= opsize; 4394 } 4395 } 4396 } 4397 EXPORT_SYMBOL(tcp_parse_options); 4398 4399 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 4400 { 4401 const __be32 *ptr = (const __be32 *)(th + 1); 4402 4403 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4404 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 4405 tp->rx_opt.saw_tstamp = 1; 4406 ++ptr; 4407 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4408 ++ptr; 4409 if (*ptr) 4410 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset; 4411 else 4412 tp->rx_opt.rcv_tsecr = 0; 4413 return true; 4414 } 4415 return false; 4416 } 4417 4418 /* Fast parse options. This hopes to only see timestamps. 4419 * If it is wrong it falls back on tcp_parse_options(). 4420 */ 4421 static bool tcp_fast_parse_options(const struct net *net, 4422 const struct sk_buff *skb, 4423 const struct tcphdr *th, struct tcp_sock *tp) 4424 { 4425 /* In the spirit of fast parsing, compare doff directly to constant 4426 * values. Because equality is used, short doff can be ignored here. 4427 */ 4428 if (th->doff == (sizeof(*th) / 4)) { 4429 tp->rx_opt.saw_tstamp = 0; 4430 tp->rx_opt.accecn = 0; 4431 return false; 4432 } else if (tp->rx_opt.tstamp_ok && 4433 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 4434 if (tcp_parse_aligned_timestamp(tp, th)) { 4435 tp->rx_opt.accecn = 0; 4436 return true; 4437 } 4438 } 4439 4440 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL); 4441 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 4442 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 4443 4444 return true; 4445 } 4446 4447 #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) 4448 /* 4449 * Parse Signature options 4450 */ 4451 int tcp_do_parse_auth_options(const struct tcphdr *th, 4452 const u8 **md5_hash, const u8 **ao_hash) 4453 { 4454 int length = (th->doff << 2) - sizeof(*th); 4455 const u8 *ptr = (const u8 *)(th + 1); 4456 unsigned int minlen = TCPOLEN_MD5SIG; 4457 4458 if (IS_ENABLED(CONFIG_TCP_AO)) 4459 minlen = sizeof(struct tcp_ao_hdr) + 1; 4460 4461 *md5_hash = NULL; 4462 *ao_hash = NULL; 4463 4464 /* If not enough data remaining, we can short cut */ 4465 while (length >= minlen) { 4466 int opcode = *ptr++; 4467 int opsize; 4468 4469 switch (opcode) { 4470 case TCPOPT_EOL: 4471 return 0; 4472 case TCPOPT_NOP: 4473 length--; 4474 continue; 4475 default: 4476 opsize = *ptr++; 4477 if (opsize < 2 || opsize > length) 4478 return -EINVAL; 4479 if (opcode == TCPOPT_MD5SIG) { 4480 if (opsize != TCPOLEN_MD5SIG) 4481 return -EINVAL; 4482 if (unlikely(*md5_hash || *ao_hash)) 4483 return -EEXIST; 4484 *md5_hash = ptr; 4485 } else if (opcode == TCPOPT_AO) { 4486 if (opsize <= sizeof(struct tcp_ao_hdr)) 4487 return -EINVAL; 4488 if (unlikely(*md5_hash || *ao_hash)) 4489 return -EEXIST; 4490 *ao_hash = ptr; 4491 } 4492 } 4493 ptr += opsize - 2; 4494 length -= opsize; 4495 } 4496 return 0; 4497 } 4498 EXPORT_SYMBOL(tcp_do_parse_auth_options); 4499 #endif 4500 4501 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 4502 * 4503 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 4504 * it can pass through stack. So, the following predicate verifies that 4505 * this segment is not used for anything but congestion avoidance or 4506 * fast retransmit. Moreover, we even are able to eliminate most of such 4507 * second order effects, if we apply some small "replay" window (~RTO) 4508 * to timestamp space. 4509 * 4510 * All these measures still do not guarantee that we reject wrapped ACKs 4511 * on networks with high bandwidth, when sequence space is recycled fastly, 4512 * but it guarantees that such events will be very rare and do not affect 4513 * connection seriously. This doesn't look nice, but alas, PAWS is really 4514 * buggy extension. 4515 * 4516 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 4517 * states that events when retransmit arrives after original data are rare. 4518 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 4519 * the biggest problem on large power networks even with minor reordering. 4520 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 4521 * up to bandwidth of 18Gigabit/sec. 8) ] 4522 */ 4523 4524 /* Estimates max number of increments of remote peer TSval in 4525 * a replay window (based on our current RTO estimation). 4526 */ 4527 static u32 tcp_tsval_replay(const struct sock *sk) 4528 { 4529 /* If we use usec TS resolution, 4530 * then expect the remote peer to use the same resolution. 4531 */ 4532 if (tcp_sk(sk)->tcp_usec_ts) 4533 return inet_csk(sk)->icsk_rto * (USEC_PER_SEC / HZ); 4534 4535 /* RFC 7323 recommends a TSval clock between 1ms and 1sec. 4536 * We know that some OS (including old linux) can use 1200 Hz. 4537 */ 4538 return inet_csk(sk)->icsk_rto * 1200 / HZ; 4539 } 4540 4541 static enum skb_drop_reason tcp_disordered_ack_check(const struct sock *sk, 4542 const struct sk_buff *skb) 4543 { 4544 const struct tcp_sock *tp = tcp_sk(sk); 4545 const struct tcphdr *th = tcp_hdr(skb); 4546 SKB_DR_INIT(reason, TCP_RFC7323_PAWS); 4547 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4548 u32 seq = TCP_SKB_CB(skb)->seq; 4549 4550 /* 1. Is this not a pure ACK ? */ 4551 if (!th->ack || seq != TCP_SKB_CB(skb)->end_seq) 4552 return reason; 4553 4554 /* 2. Is its sequence not the expected one ? */ 4555 if (seq != tp->rcv_nxt) 4556 return before(seq, tp->rcv_nxt) ? 4557 SKB_DROP_REASON_TCP_RFC7323_PAWS_ACK : 4558 reason; 4559 4560 /* 3. Is this not a duplicate ACK ? */ 4561 if (ack != tp->snd_una) 4562 return reason; 4563 4564 /* 4. Is this updating the window ? */ 4565 if (tcp_may_update_window(tp, ack, seq, ntohs(th->window) << 4566 tp->rx_opt.snd_wscale)) 4567 return reason; 4568 4569 /* 5. Is this not in the replay window ? */ 4570 if ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > 4571 tcp_tsval_replay(sk)) 4572 return reason; 4573 4574 return 0; 4575 } 4576 4577 /* Check segment sequence number for validity. 4578 * 4579 * Segment controls are considered valid, if the segment 4580 * fits to the window after truncation to the window. Acceptability 4581 * of data (and SYN, FIN, of course) is checked separately. 4582 * See tcp_data_queue(), for example. 4583 * 4584 * Also, controls (RST is main one) are accepted using RCV.WUP instead 4585 * of RCV.NXT. Peer still did not advance his SND.UNA when we 4586 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4587 * (borrowed from freebsd) 4588 */ 4589 4590 static enum skb_drop_reason tcp_sequence(const struct sock *sk, 4591 u32 seq, u32 end_seq) 4592 { 4593 const struct tcp_sock *tp = tcp_sk(sk); 4594 4595 if (before(end_seq, tp->rcv_wup)) 4596 return SKB_DROP_REASON_TCP_OLD_SEQUENCE; 4597 4598 if (after(end_seq, tp->rcv_nxt + tcp_receive_window(tp))) { 4599 if (after(seq, tp->rcv_nxt + tcp_receive_window(tp))) 4600 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE; 4601 4602 /* Only accept this packet if receive queue is empty. */ 4603 if (skb_queue_len(&sk->sk_receive_queue)) 4604 return SKB_DROP_REASON_TCP_INVALID_END_SEQUENCE; 4605 } 4606 4607 return SKB_NOT_DROPPED_YET; 4608 } 4609 4610 4611 void tcp_done_with_error(struct sock *sk, int err) 4612 { 4613 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4614 WRITE_ONCE(sk->sk_err, err); 4615 smp_wmb(); 4616 4617 tcp_write_queue_purge(sk); 4618 tcp_done(sk); 4619 4620 if (!sock_flag(sk, SOCK_DEAD)) 4621 sk_error_report(sk); 4622 } 4623 EXPORT_IPV6_MOD(tcp_done_with_error); 4624 4625 /* When we get a reset we do this. */ 4626 void tcp_reset(struct sock *sk, struct sk_buff *skb) 4627 { 4628 int err; 4629 4630 trace_tcp_receive_reset(sk); 4631 4632 /* mptcp can't tell us to ignore reset pkts, 4633 * so just ignore the return value of mptcp_incoming_options(). 4634 */ 4635 if (sk_is_mptcp(sk)) 4636 mptcp_incoming_options(sk, skb); 4637 4638 /* We want the right error as BSD sees it (and indeed as we do). */ 4639 switch (sk->sk_state) { 4640 case TCP_SYN_SENT: 4641 err = ECONNREFUSED; 4642 break; 4643 case TCP_CLOSE_WAIT: 4644 err = EPIPE; 4645 break; 4646 case TCP_CLOSE: 4647 return; 4648 default: 4649 err = ECONNRESET; 4650 } 4651 tcp_done_with_error(sk, err); 4652 } 4653 4654 /* 4655 * Process the FIN bit. This now behaves as it is supposed to work 4656 * and the FIN takes effect when it is validly part of sequence 4657 * space. Not before when we get holes. 4658 * 4659 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4660 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4661 * TIME-WAIT) 4662 * 4663 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4664 * close and we go into CLOSING (and later onto TIME-WAIT) 4665 * 4666 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4667 */ 4668 void tcp_fin(struct sock *sk) 4669 { 4670 struct tcp_sock *tp = tcp_sk(sk); 4671 4672 inet_csk_schedule_ack(sk); 4673 4674 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN); 4675 sock_set_flag(sk, SOCK_DONE); 4676 4677 switch (sk->sk_state) { 4678 case TCP_SYN_RECV: 4679 case TCP_ESTABLISHED: 4680 /* Move to CLOSE_WAIT */ 4681 tcp_set_state(sk, TCP_CLOSE_WAIT); 4682 inet_csk_enter_pingpong_mode(sk); 4683 break; 4684 4685 case TCP_CLOSE_WAIT: 4686 case TCP_CLOSING: 4687 /* Received a retransmission of the FIN, do 4688 * nothing. 4689 */ 4690 break; 4691 case TCP_LAST_ACK: 4692 /* RFC793: Remain in the LAST-ACK state. */ 4693 break; 4694 4695 case TCP_FIN_WAIT1: 4696 /* This case occurs when a simultaneous close 4697 * happens, we must ack the received FIN and 4698 * enter the CLOSING state. 4699 */ 4700 tcp_send_ack(sk); 4701 tcp_set_state(sk, TCP_CLOSING); 4702 break; 4703 case TCP_FIN_WAIT2: 4704 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4705 tcp_send_ack(sk); 4706 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4707 break; 4708 default: 4709 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4710 * cases we should never reach this piece of code. 4711 */ 4712 pr_err("%s: Impossible, sk->sk_state=%d\n", 4713 __func__, sk->sk_state); 4714 break; 4715 } 4716 4717 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4718 * Probably, we should reset in this case. For now drop them. 4719 */ 4720 skb_rbtree_purge(&tp->out_of_order_queue); 4721 if (tcp_is_sack(tp)) 4722 tcp_sack_reset(&tp->rx_opt); 4723 4724 if (!sock_flag(sk, SOCK_DEAD)) { 4725 sk->sk_state_change(sk); 4726 4727 /* Do not send POLL_HUP for half duplex close. */ 4728 if (sk->sk_shutdown == SHUTDOWN_MASK || 4729 sk->sk_state == TCP_CLOSE) 4730 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4731 else 4732 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4733 } 4734 } 4735 4736 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4737 u32 end_seq) 4738 { 4739 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4740 if (before(seq, sp->start_seq)) 4741 sp->start_seq = seq; 4742 if (after(end_seq, sp->end_seq)) 4743 sp->end_seq = end_seq; 4744 return true; 4745 } 4746 return false; 4747 } 4748 4749 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4750 { 4751 struct tcp_sock *tp = tcp_sk(sk); 4752 4753 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { 4754 int mib_idx; 4755 4756 if (before(seq, tp->rcv_nxt)) 4757 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4758 else 4759 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4760 4761 NET_INC_STATS(sock_net(sk), mib_idx); 4762 4763 tp->rx_opt.dsack = 1; 4764 tp->duplicate_sack[0].start_seq = seq; 4765 tp->duplicate_sack[0].end_seq = end_seq; 4766 } 4767 } 4768 4769 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4770 { 4771 struct tcp_sock *tp = tcp_sk(sk); 4772 4773 if (!tp->rx_opt.dsack) 4774 tcp_dsack_set(sk, seq, end_seq); 4775 else 4776 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4777 } 4778 4779 static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb) 4780 { 4781 /* When the ACK path fails or drops most ACKs, the sender would 4782 * timeout and spuriously retransmit the same segment repeatedly. 4783 * If it seems our ACKs are not reaching the other side, 4784 * based on receiving a duplicate data segment with new flowlabel 4785 * (suggesting the sender suffered an RTO), and we are not already 4786 * repathing due to our own RTO, then rehash the socket to repath our 4787 * packets. 4788 */ 4789 #if IS_ENABLED(CONFIG_IPV6) 4790 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss && 4791 skb->protocol == htons(ETH_P_IPV6) && 4792 (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel != 4793 ntohl(ip6_flowlabel(ipv6_hdr(skb)))) && 4794 sk_rethink_txhash(sk)) 4795 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH); 4796 4797 /* Save last flowlabel after a spurious retrans. */ 4798 tcp_save_lrcv_flowlabel(sk, skb); 4799 #endif 4800 } 4801 4802 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 4803 { 4804 struct tcp_sock *tp = tcp_sk(sk); 4805 4806 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4807 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4808 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4809 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 4810 4811 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { 4812 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4813 4814 tcp_rcv_spurious_retrans(sk, skb); 4815 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4816 end_seq = tp->rcv_nxt; 4817 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4818 } 4819 } 4820 4821 tcp_send_ack(sk); 4822 } 4823 4824 /* These routines update the SACK block as out-of-order packets arrive or 4825 * in-order packets close up the sequence space. 4826 */ 4827 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4828 { 4829 int this_sack; 4830 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4831 struct tcp_sack_block *swalk = sp + 1; 4832 4833 /* See if the recent change to the first SACK eats into 4834 * or hits the sequence space of other SACK blocks, if so coalesce. 4835 */ 4836 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4837 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4838 int i; 4839 4840 /* Zap SWALK, by moving every further SACK up by one slot. 4841 * Decrease num_sacks. 4842 */ 4843 tp->rx_opt.num_sacks--; 4844 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4845 sp[i] = sp[i + 1]; 4846 continue; 4847 } 4848 this_sack++; 4849 swalk++; 4850 } 4851 } 4852 4853 void tcp_sack_compress_send_ack(struct sock *sk) 4854 { 4855 struct tcp_sock *tp = tcp_sk(sk); 4856 4857 if (!tp->compressed_ack) 4858 return; 4859 4860 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 4861 __sock_put(sk); 4862 4863 /* Since we have to send one ack finally, 4864 * substract one from tp->compressed_ack to keep 4865 * LINUX_MIB_TCPACKCOMPRESSED accurate. 4866 */ 4867 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 4868 tp->compressed_ack - 1); 4869 4870 tp->compressed_ack = 0; 4871 tcp_send_ack(sk); 4872 } 4873 4874 /* Reasonable amount of sack blocks included in TCP SACK option 4875 * The max is 4, but this becomes 3 if TCP timestamps are there. 4876 * Given that SACK packets might be lost, be conservative and use 2. 4877 */ 4878 #define TCP_SACK_BLOCKS_EXPECTED 2 4879 4880 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4881 { 4882 struct tcp_sock *tp = tcp_sk(sk); 4883 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4884 int cur_sacks = tp->rx_opt.num_sacks; 4885 int this_sack; 4886 4887 if (!cur_sacks) 4888 goto new_sack; 4889 4890 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4891 if (tcp_sack_extend(sp, seq, end_seq)) { 4892 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4893 tcp_sack_compress_send_ack(sk); 4894 /* Rotate this_sack to the first one. */ 4895 for (; this_sack > 0; this_sack--, sp--) 4896 swap(*sp, *(sp - 1)); 4897 if (cur_sacks > 1) 4898 tcp_sack_maybe_coalesce(tp); 4899 return; 4900 } 4901 } 4902 4903 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4904 tcp_sack_compress_send_ack(sk); 4905 4906 /* Could not find an adjacent existing SACK, build a new one, 4907 * put it at the front, and shift everyone else down. We 4908 * always know there is at least one SACK present already here. 4909 * 4910 * If the sack array is full, forget about the last one. 4911 */ 4912 if (this_sack >= TCP_NUM_SACKS) { 4913 this_sack--; 4914 tp->rx_opt.num_sacks--; 4915 sp--; 4916 } 4917 for (; this_sack > 0; this_sack--, sp--) 4918 *sp = *(sp - 1); 4919 4920 new_sack: 4921 /* Build the new head SACK, and we're done. */ 4922 sp->start_seq = seq; 4923 sp->end_seq = end_seq; 4924 tp->rx_opt.num_sacks++; 4925 } 4926 4927 /* RCV.NXT advances, some SACKs should be eaten. */ 4928 4929 static void tcp_sack_remove(struct tcp_sock *tp) 4930 { 4931 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4932 int num_sacks = tp->rx_opt.num_sacks; 4933 int this_sack; 4934 4935 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4936 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4937 tp->rx_opt.num_sacks = 0; 4938 return; 4939 } 4940 4941 for (this_sack = 0; this_sack < num_sacks;) { 4942 /* Check if the start of the sack is covered by RCV.NXT. */ 4943 if (!before(tp->rcv_nxt, sp->start_seq)) { 4944 int i; 4945 4946 /* RCV.NXT must cover all the block! */ 4947 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4948 4949 /* Zap this SACK, by moving forward any other SACKS. */ 4950 for (i = this_sack+1; i < num_sacks; i++) 4951 tp->selective_acks[i-1] = tp->selective_acks[i]; 4952 num_sacks--; 4953 continue; 4954 } 4955 this_sack++; 4956 sp++; 4957 } 4958 tp->rx_opt.num_sacks = num_sacks; 4959 } 4960 4961 /** 4962 * tcp_try_coalesce - try to merge skb to prior one 4963 * @sk: socket 4964 * @to: prior buffer 4965 * @from: buffer to add in queue 4966 * @fragstolen: pointer to boolean 4967 * 4968 * Before queueing skb @from after @to, try to merge them 4969 * to reduce overall memory use and queue lengths, if cost is small. 4970 * Packets in ofo or receive queues can stay a long time. 4971 * Better try to coalesce them right now to avoid future collapses. 4972 * Returns true if caller should free @from instead of queueing it 4973 */ 4974 static bool tcp_try_coalesce(struct sock *sk, 4975 struct sk_buff *to, 4976 struct sk_buff *from, 4977 bool *fragstolen) 4978 { 4979 int delta; 4980 4981 *fragstolen = false; 4982 4983 /* Its possible this segment overlaps with prior segment in queue */ 4984 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq) 4985 return false; 4986 4987 if (!tcp_skb_can_collapse_rx(to, from)) 4988 return false; 4989 4990 if (!skb_try_coalesce(to, from, fragstolen, &delta)) 4991 return false; 4992 4993 atomic_add(delta, &sk->sk_rmem_alloc); 4994 sk_mem_charge(sk, delta); 4995 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); 4996 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; 4997 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; 4998 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags; 4999 5000 if (TCP_SKB_CB(from)->has_rxtstamp) { 5001 TCP_SKB_CB(to)->has_rxtstamp = true; 5002 to->tstamp = from->tstamp; 5003 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp; 5004 } 5005 5006 return true; 5007 } 5008 5009 static bool tcp_ooo_try_coalesce(struct sock *sk, 5010 struct sk_buff *to, 5011 struct sk_buff *from, 5012 bool *fragstolen) 5013 { 5014 bool res = tcp_try_coalesce(sk, to, from, fragstolen); 5015 5016 /* In case tcp_drop_reason() is called later, update to->gso_segs */ 5017 if (res) { 5018 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) + 5019 max_t(u16, 1, skb_shinfo(from)->gso_segs); 5020 5021 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF); 5022 } 5023 return res; 5024 } 5025 5026 noinline_for_tracing static void 5027 tcp_drop_reason(struct sock *sk, struct sk_buff *skb, enum skb_drop_reason reason) 5028 { 5029 sk_drops_skbadd(sk, skb); 5030 sk_skb_reason_drop(sk, skb, reason); 5031 } 5032 5033 /* This one checks to see if we can put data from the 5034 * out_of_order queue into the receive_queue. 5035 */ 5036 static void tcp_ofo_queue(struct sock *sk) 5037 { 5038 struct tcp_sock *tp = tcp_sk(sk); 5039 __u32 dsack_high = tp->rcv_nxt; 5040 bool fin, fragstolen, eaten; 5041 struct sk_buff *skb, *tail; 5042 struct rb_node *p; 5043 5044 p = rb_first(&tp->out_of_order_queue); 5045 while (p) { 5046 skb = rb_to_skb(p); 5047 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 5048 break; 5049 5050 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 5051 __u32 dsack = dsack_high; 5052 5053 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 5054 dsack = TCP_SKB_CB(skb)->end_seq; 5055 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 5056 } 5057 p = rb_next(p); 5058 rb_erase(&skb->rbnode, &tp->out_of_order_queue); 5059 5060 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) { 5061 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP); 5062 continue; 5063 } 5064 5065 tail = skb_peek_tail(&sk->sk_receive_queue); 5066 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen); 5067 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); 5068 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 5069 if (!eaten) 5070 tcp_add_receive_queue(sk, skb); 5071 else 5072 kfree_skb_partial(skb, fragstolen); 5073 5074 if (unlikely(fin)) { 5075 tcp_fin(sk); 5076 /* tcp_fin() purges tp->out_of_order_queue, 5077 * so we must end this loop right now. 5078 */ 5079 break; 5080 } 5081 } 5082 } 5083 5084 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb); 5085 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb); 5086 5087 /* Check if this incoming skb can be added to socket receive queues 5088 * while satisfying sk->sk_rcvbuf limit. 5089 * 5090 * In theory we should use skb->truesize, but this can cause problems 5091 * when applications use too small SO_RCVBUF values. 5092 * When LRO / hw gro is used, the socket might have a high tp->scaling_ratio, 5093 * allowing RWIN to be close to available space. 5094 * Whenever the receive queue gets full, we can receive a small packet 5095 * filling RWIN, but with a high skb->truesize, because most NIC use 4K page 5096 * plus sk_buff metadata even when receiving less than 1500 bytes of payload. 5097 * 5098 * Note that we use skb->len to decide to accept or drop this packet, 5099 * but sk->sk_rmem_alloc is the sum of all skb->truesize. 5100 */ 5101 static bool tcp_can_ingest(const struct sock *sk, const struct sk_buff *skb) 5102 { 5103 unsigned int rmem = atomic_read(&sk->sk_rmem_alloc); 5104 5105 return rmem + skb->len <= sk->sk_rcvbuf; 5106 } 5107 5108 static int tcp_try_rmem_schedule(struct sock *sk, const struct sk_buff *skb, 5109 unsigned int size) 5110 { 5111 if (!tcp_can_ingest(sk, skb) || 5112 !sk_rmem_schedule(sk, skb, size)) { 5113 5114 if (tcp_prune_queue(sk, skb) < 0) 5115 return -1; 5116 5117 while (!sk_rmem_schedule(sk, skb, size)) { 5118 if (!tcp_prune_ofo_queue(sk, skb)) 5119 return -1; 5120 } 5121 } 5122 return 0; 5123 } 5124 5125 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) 5126 { 5127 struct tcp_sock *tp = tcp_sk(sk); 5128 struct rb_node **p, *parent; 5129 struct sk_buff *skb1; 5130 u32 seq, end_seq; 5131 bool fragstolen; 5132 5133 tcp_save_lrcv_flowlabel(sk, skb); 5134 tcp_data_ecn_check(sk, skb); 5135 5136 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) { 5137 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP); 5138 sk->sk_data_ready(sk); 5139 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM); 5140 return; 5141 } 5142 5143 tcp_measure_rcv_mss(sk, skb); 5144 /* Disable header prediction. */ 5145 tp->pred_flags = 0; 5146 inet_csk_schedule_ack(sk); 5147 5148 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs); 5149 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE); 5150 seq = TCP_SKB_CB(skb)->seq; 5151 end_seq = TCP_SKB_CB(skb)->end_seq; 5152 5153 p = &tp->out_of_order_queue.rb_node; 5154 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5155 /* Initial out of order segment, build 1 SACK. */ 5156 if (tcp_is_sack(tp)) { 5157 tp->rx_opt.num_sacks = 1; 5158 tp->selective_acks[0].start_seq = seq; 5159 tp->selective_acks[0].end_seq = end_seq; 5160 } 5161 rb_link_node(&skb->rbnode, NULL, p); 5162 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 5163 tp->ooo_last_skb = skb; 5164 goto end; 5165 } 5166 5167 /* In the typical case, we are adding an skb to the end of the list. 5168 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 5169 */ 5170 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb, 5171 skb, &fragstolen)) { 5172 coalesce_done: 5173 /* For non sack flows, do not grow window to force DUPACK 5174 * and trigger fast retransmit. 5175 */ 5176 if (tcp_is_sack(tp)) 5177 tcp_grow_window(sk, skb, true); 5178 kfree_skb_partial(skb, fragstolen); 5179 skb = NULL; 5180 goto add_sack; 5181 } 5182 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 5183 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) { 5184 parent = &tp->ooo_last_skb->rbnode; 5185 p = &parent->rb_right; 5186 goto insert; 5187 } 5188 5189 /* Find place to insert this segment. Handle overlaps on the way. */ 5190 parent = NULL; 5191 while (*p) { 5192 parent = *p; 5193 skb1 = rb_to_skb(parent); 5194 if (before(seq, TCP_SKB_CB(skb1)->seq)) { 5195 p = &parent->rb_left; 5196 continue; 5197 } 5198 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) { 5199 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 5200 /* All the bits are present. Drop. */ 5201 NET_INC_STATS(sock_net(sk), 5202 LINUX_MIB_TCPOFOMERGE); 5203 tcp_drop_reason(sk, skb, 5204 SKB_DROP_REASON_TCP_OFOMERGE); 5205 skb = NULL; 5206 tcp_dsack_set(sk, seq, end_seq); 5207 goto add_sack; 5208 } 5209 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 5210 /* Partial overlap. */ 5211 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq); 5212 } else { 5213 /* skb's seq == skb1's seq and skb covers skb1. 5214 * Replace skb1 with skb. 5215 */ 5216 rb_replace_node(&skb1->rbnode, &skb->rbnode, 5217 &tp->out_of_order_queue); 5218 tcp_dsack_extend(sk, 5219 TCP_SKB_CB(skb1)->seq, 5220 TCP_SKB_CB(skb1)->end_seq); 5221 NET_INC_STATS(sock_net(sk), 5222 LINUX_MIB_TCPOFOMERGE); 5223 tcp_drop_reason(sk, skb1, 5224 SKB_DROP_REASON_TCP_OFOMERGE); 5225 goto merge_right; 5226 } 5227 } else if (tcp_ooo_try_coalesce(sk, skb1, 5228 skb, &fragstolen)) { 5229 goto coalesce_done; 5230 } 5231 p = &parent->rb_right; 5232 } 5233 insert: 5234 /* Insert segment into RB tree. */ 5235 rb_link_node(&skb->rbnode, parent, p); 5236 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 5237 5238 merge_right: 5239 /* Remove other segments covered by skb. */ 5240 while ((skb1 = skb_rb_next(skb)) != NULL) { 5241 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 5242 break; 5243 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 5244 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 5245 end_seq); 5246 break; 5247 } 5248 rb_erase(&skb1->rbnode, &tp->out_of_order_queue); 5249 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 5250 TCP_SKB_CB(skb1)->end_seq); 5251 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE); 5252 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE); 5253 } 5254 /* If there is no skb after us, we are the last_skb ! */ 5255 if (!skb1) 5256 tp->ooo_last_skb = skb; 5257 5258 add_sack: 5259 if (tcp_is_sack(tp)) 5260 tcp_sack_new_ofo_skb(sk, seq, end_seq); 5261 end: 5262 if (skb) { 5263 /* For non sack flows, do not grow window to force DUPACK 5264 * and trigger fast retransmit. 5265 */ 5266 if (tcp_is_sack(tp)) 5267 tcp_grow_window(sk, skb, false); 5268 skb_condense(skb); 5269 skb_set_owner_r(skb, sk); 5270 } 5271 /* do not grow rcvbuf for not-yet-accepted or orphaned sockets. */ 5272 if (sk->sk_socket) 5273 tcp_rcvbuf_grow(sk); 5274 } 5275 5276 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, 5277 bool *fragstolen) 5278 { 5279 int eaten; 5280 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue); 5281 5282 eaten = (tail && 5283 tcp_try_coalesce(sk, tail, 5284 skb, fragstolen)) ? 1 : 0; 5285 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq); 5286 if (!eaten) { 5287 tcp_add_receive_queue(sk, skb); 5288 skb_set_owner_r(skb, sk); 5289 } 5290 return eaten; 5291 } 5292 5293 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) 5294 { 5295 struct sk_buff *skb; 5296 int err = -ENOMEM; 5297 int data_len = 0; 5298 bool fragstolen; 5299 5300 if (size == 0) 5301 return 0; 5302 5303 if (size > PAGE_SIZE) { 5304 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS); 5305 5306 data_len = npages << PAGE_SHIFT; 5307 size = data_len + (size & ~PAGE_MASK); 5308 } 5309 skb = alloc_skb_with_frags(size - data_len, data_len, 5310 PAGE_ALLOC_COSTLY_ORDER, 5311 &err, sk->sk_allocation); 5312 if (!skb) 5313 goto err; 5314 5315 skb_put(skb, size - data_len); 5316 skb->data_len = data_len; 5317 skb->len = size; 5318 5319 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 5320 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 5321 goto err_free; 5322 } 5323 5324 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size); 5325 if (err) 5326 goto err_free; 5327 5328 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; 5329 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; 5330 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; 5331 5332 if (tcp_queue_rcv(sk, skb, &fragstolen)) { 5333 WARN_ON_ONCE(fragstolen); /* should not happen */ 5334 __kfree_skb(skb); 5335 } 5336 return size; 5337 5338 err_free: 5339 kfree_skb(skb); 5340 err: 5341 return err; 5342 5343 } 5344 5345 void tcp_data_ready(struct sock *sk) 5346 { 5347 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE)) 5348 sk->sk_data_ready(sk); 5349 } 5350 5351 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 5352 { 5353 struct tcp_sock *tp = tcp_sk(sk); 5354 enum skb_drop_reason reason; 5355 bool fragstolen; 5356 int eaten; 5357 5358 /* If a subflow has been reset, the packet should not continue 5359 * to be processed, drop the packet. 5360 */ 5361 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) { 5362 __kfree_skb(skb); 5363 return; 5364 } 5365 5366 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { 5367 __kfree_skb(skb); 5368 return; 5369 } 5370 tcp_cleanup_skb(skb); 5371 __skb_pull(skb, tcp_hdr(skb)->doff * 4); 5372 5373 reason = SKB_DROP_REASON_NOT_SPECIFIED; 5374 tp->rx_opt.dsack = 0; 5375 5376 /* Queue data for delivery to the user. 5377 * Packets in sequence go to the receive queue. 5378 * Out of sequence packets to the out_of_order_queue. 5379 */ 5380 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 5381 if (tcp_receive_window(tp) == 0) { 5382 /* Some stacks are known to send bare FIN packets 5383 * in a loop even if we send RWIN 0 in our ACK. 5384 * Accepting this FIN does not hurt memory pressure 5385 * because the FIN flag will simply be merged to the 5386 * receive queue tail skb in most cases. 5387 */ 5388 if (!skb->len && 5389 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 5390 goto queue_and_out; 5391 5392 reason = SKB_DROP_REASON_TCP_ZEROWINDOW; 5393 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5394 goto out_of_window; 5395 } 5396 5397 /* Ok. In sequence. In window. */ 5398 queue_and_out: 5399 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 5400 /* TODO: maybe ratelimit these WIN 0 ACK ? */ 5401 inet_csk(sk)->icsk_ack.pending |= 5402 (ICSK_ACK_NOMEM | ICSK_ACK_NOW); 5403 inet_csk_schedule_ack(sk); 5404 sk->sk_data_ready(sk); 5405 5406 if (skb_queue_len(&sk->sk_receive_queue) && skb->len) { 5407 reason = SKB_DROP_REASON_PROTO_MEM; 5408 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 5409 goto drop; 5410 } 5411 sk_forced_mem_schedule(sk, skb->truesize); 5412 } 5413 5414 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 5415 if (skb->len) 5416 tcp_event_data_recv(sk, skb); 5417 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 5418 tcp_fin(sk); 5419 5420 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5421 tcp_ofo_queue(sk); 5422 5423 /* RFC5681. 4.2. SHOULD send immediate ACK, when 5424 * gap in queue is filled. 5425 */ 5426 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5427 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 5428 } 5429 5430 if (tp->rx_opt.num_sacks) 5431 tcp_sack_remove(tp); 5432 5433 tcp_fast_path_check(sk); 5434 5435 if (eaten > 0) 5436 kfree_skb_partial(skb, fragstolen); 5437 if (!sock_flag(sk, SOCK_DEAD)) 5438 tcp_data_ready(sk); 5439 return; 5440 } 5441 5442 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 5443 tcp_rcv_spurious_retrans(sk, skb); 5444 /* A retransmit, 2nd most common case. Force an immediate ack. */ 5445 reason = SKB_DROP_REASON_TCP_OLD_DATA; 5446 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 5447 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 5448 5449 out_of_window: 5450 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 5451 inet_csk_schedule_ack(sk); 5452 drop: 5453 tcp_drop_reason(sk, skb, reason); 5454 return; 5455 } 5456 5457 /* Out of window. F.e. zero window probe. */ 5458 if (!before(TCP_SKB_CB(skb)->seq, 5459 tp->rcv_nxt + tcp_receive_window(tp))) { 5460 reason = SKB_DROP_REASON_TCP_OVERWINDOW; 5461 goto out_of_window; 5462 } 5463 5464 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5465 /* Partial packet, seq < rcv_next < end_seq */ 5466 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 5467 5468 /* If window is closed, drop tail of packet. But after 5469 * remembering D-SACK for its head made in previous line. 5470 */ 5471 if (!tcp_receive_window(tp)) { 5472 reason = SKB_DROP_REASON_TCP_ZEROWINDOW; 5473 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5474 goto out_of_window; 5475 } 5476 goto queue_and_out; 5477 } 5478 5479 tcp_data_queue_ofo(sk, skb); 5480 } 5481 5482 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list) 5483 { 5484 if (list) 5485 return !skb_queue_is_last(list, skb) ? skb->next : NULL; 5486 5487 return skb_rb_next(skb); 5488 } 5489 5490 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 5491 struct sk_buff_head *list, 5492 struct rb_root *root) 5493 { 5494 struct sk_buff *next = tcp_skb_next(skb, list); 5495 5496 if (list) 5497 __skb_unlink(skb, list); 5498 else 5499 rb_erase(&skb->rbnode, root); 5500 5501 __kfree_skb(skb); 5502 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 5503 5504 return next; 5505 } 5506 5507 /* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */ 5508 void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb) 5509 { 5510 struct rb_node **p = &root->rb_node; 5511 struct rb_node *parent = NULL; 5512 struct sk_buff *skb1; 5513 5514 while (*p) { 5515 parent = *p; 5516 skb1 = rb_to_skb(parent); 5517 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq)) 5518 p = &parent->rb_left; 5519 else 5520 p = &parent->rb_right; 5521 } 5522 rb_link_node(&skb->rbnode, parent, p); 5523 rb_insert_color(&skb->rbnode, root); 5524 } 5525 5526 /* Collapse contiguous sequence of skbs head..tail with 5527 * sequence numbers start..end. 5528 * 5529 * If tail is NULL, this means until the end of the queue. 5530 * 5531 * Segments with FIN/SYN are not collapsed (only because this 5532 * simplifies code) 5533 */ 5534 static void 5535 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root, 5536 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end) 5537 { 5538 struct sk_buff *skb = head, *n; 5539 struct sk_buff_head tmp; 5540 bool end_of_skbs; 5541 5542 /* First, check that queue is collapsible and find 5543 * the point where collapsing can be useful. 5544 */ 5545 restart: 5546 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) { 5547 n = tcp_skb_next(skb, list); 5548 5549 if (!skb_frags_readable(skb)) 5550 goto skip_this; 5551 5552 /* No new bits? It is possible on ofo queue. */ 5553 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5554 skb = tcp_collapse_one(sk, skb, list, root); 5555 if (!skb) 5556 break; 5557 goto restart; 5558 } 5559 5560 /* The first skb to collapse is: 5561 * - not SYN/FIN and 5562 * - bloated or contains data before "start" or 5563 * overlaps to the next one and mptcp allow collapsing. 5564 */ 5565 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) && 5566 (tcp_win_from_space(sk, skb->truesize) > skb->len || 5567 before(TCP_SKB_CB(skb)->seq, start))) { 5568 end_of_skbs = false; 5569 break; 5570 } 5571 5572 if (n && n != tail && skb_frags_readable(n) && 5573 tcp_skb_can_collapse_rx(skb, n) && 5574 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) { 5575 end_of_skbs = false; 5576 break; 5577 } 5578 5579 skip_this: 5580 /* Decided to skip this, advance start seq. */ 5581 start = TCP_SKB_CB(skb)->end_seq; 5582 } 5583 if (end_of_skbs || 5584 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) || 5585 !skb_frags_readable(skb)) 5586 return; 5587 5588 __skb_queue_head_init(&tmp); 5589 5590 while (before(start, end)) { 5591 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start); 5592 struct sk_buff *nskb; 5593 5594 nskb = alloc_skb(copy, GFP_ATOMIC); 5595 if (!nskb) 5596 break; 5597 5598 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 5599 skb_copy_decrypted(nskb, skb); 5600 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 5601 if (list) 5602 __skb_queue_before(list, skb, nskb); 5603 else 5604 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */ 5605 skb_set_owner_r(nskb, sk); 5606 mptcp_skb_ext_move(nskb, skb); 5607 5608 /* Copy data, releasing collapsed skbs. */ 5609 while (copy > 0) { 5610 int offset = start - TCP_SKB_CB(skb)->seq; 5611 int size = TCP_SKB_CB(skb)->end_seq - start; 5612 5613 BUG_ON(offset < 0); 5614 if (size > 0) { 5615 size = min(copy, size); 5616 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 5617 BUG(); 5618 TCP_SKB_CB(nskb)->end_seq += size; 5619 copy -= size; 5620 start += size; 5621 } 5622 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5623 skb = tcp_collapse_one(sk, skb, list, root); 5624 if (!skb || 5625 skb == tail || 5626 !tcp_skb_can_collapse_rx(nskb, skb) || 5627 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) || 5628 !skb_frags_readable(skb)) 5629 goto end; 5630 } 5631 } 5632 } 5633 end: 5634 skb_queue_walk_safe(&tmp, skb, n) 5635 tcp_rbtree_insert(root, skb); 5636 } 5637 5638 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 5639 * and tcp_collapse() them until all the queue is collapsed. 5640 */ 5641 static void tcp_collapse_ofo_queue(struct sock *sk) 5642 { 5643 struct tcp_sock *tp = tcp_sk(sk); 5644 u32 range_truesize, sum_tiny = 0; 5645 struct sk_buff *skb, *head; 5646 u32 start, end; 5647 5648 skb = skb_rb_first(&tp->out_of_order_queue); 5649 new_range: 5650 if (!skb) { 5651 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue); 5652 return; 5653 } 5654 start = TCP_SKB_CB(skb)->seq; 5655 end = TCP_SKB_CB(skb)->end_seq; 5656 range_truesize = skb->truesize; 5657 5658 for (head = skb;;) { 5659 skb = skb_rb_next(skb); 5660 5661 /* Range is terminated when we see a gap or when 5662 * we are at the queue end. 5663 */ 5664 if (!skb || 5665 after(TCP_SKB_CB(skb)->seq, end) || 5666 before(TCP_SKB_CB(skb)->end_seq, start)) { 5667 /* Do not attempt collapsing tiny skbs */ 5668 if (range_truesize != head->truesize || 5669 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) { 5670 tcp_collapse(sk, NULL, &tp->out_of_order_queue, 5671 head, skb, start, end); 5672 } else { 5673 sum_tiny += range_truesize; 5674 if (sum_tiny > sk->sk_rcvbuf >> 3) 5675 return; 5676 } 5677 goto new_range; 5678 } 5679 5680 range_truesize += skb->truesize; 5681 if (unlikely(before(TCP_SKB_CB(skb)->seq, start))) 5682 start = TCP_SKB_CB(skb)->seq; 5683 if (after(TCP_SKB_CB(skb)->end_seq, end)) 5684 end = TCP_SKB_CB(skb)->end_seq; 5685 } 5686 } 5687 5688 /* 5689 * Clean the out-of-order queue to make room. 5690 * We drop high sequences packets to : 5691 * 1) Let a chance for holes to be filled. 5692 * This means we do not drop packets from ooo queue if their sequence 5693 * is before incoming packet sequence. 5694 * 2) not add too big latencies if thousands of packets sit there. 5695 * (But if application shrinks SO_RCVBUF, we could still end up 5696 * freeing whole queue here) 5697 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks. 5698 * 5699 * Return true if queue has shrunk. 5700 */ 5701 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb) 5702 { 5703 struct tcp_sock *tp = tcp_sk(sk); 5704 struct rb_node *node, *prev; 5705 bool pruned = false; 5706 int goal; 5707 5708 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5709 return false; 5710 5711 goal = sk->sk_rcvbuf >> 3; 5712 node = &tp->ooo_last_skb->rbnode; 5713 5714 do { 5715 struct sk_buff *skb = rb_to_skb(node); 5716 5717 /* If incoming skb would land last in ofo queue, stop pruning. */ 5718 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq)) 5719 break; 5720 pruned = true; 5721 prev = rb_prev(node); 5722 rb_erase(node, &tp->out_of_order_queue); 5723 goal -= skb->truesize; 5724 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE); 5725 tp->ooo_last_skb = rb_to_skb(prev); 5726 if (!prev || goal <= 0) { 5727 if (tcp_can_ingest(sk, in_skb) && 5728 !tcp_under_memory_pressure(sk)) 5729 break; 5730 goal = sk->sk_rcvbuf >> 3; 5731 } 5732 node = prev; 5733 } while (node); 5734 5735 if (pruned) { 5736 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED); 5737 /* Reset SACK state. A conforming SACK implementation will 5738 * do the same at a timeout based retransmit. When a connection 5739 * is in a sad state like this, we care only about integrity 5740 * of the connection not performance. 5741 */ 5742 if (tp->rx_opt.sack_ok) 5743 tcp_sack_reset(&tp->rx_opt); 5744 } 5745 return pruned; 5746 } 5747 5748 /* Reduce allocated memory if we can, trying to get 5749 * the socket within its memory limits again. 5750 * 5751 * Return less than zero if we should start dropping frames 5752 * until the socket owning process reads some of the data 5753 * to stabilize the situation. 5754 */ 5755 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb) 5756 { 5757 struct tcp_sock *tp = tcp_sk(sk); 5758 5759 /* Do nothing if our queues are empty. */ 5760 if (!atomic_read(&sk->sk_rmem_alloc)) 5761 return -1; 5762 5763 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED); 5764 5765 if (!tcp_can_ingest(sk, in_skb)) 5766 tcp_clamp_window(sk); 5767 else if (tcp_under_memory_pressure(sk)) 5768 tcp_adjust_rcv_ssthresh(sk); 5769 5770 if (tcp_can_ingest(sk, in_skb)) 5771 return 0; 5772 5773 tcp_collapse_ofo_queue(sk); 5774 if (!skb_queue_empty(&sk->sk_receive_queue)) 5775 tcp_collapse(sk, &sk->sk_receive_queue, NULL, 5776 skb_peek(&sk->sk_receive_queue), 5777 NULL, 5778 tp->copied_seq, tp->rcv_nxt); 5779 5780 if (tcp_can_ingest(sk, in_skb)) 5781 return 0; 5782 5783 /* Collapsing did not help, destructive actions follow. 5784 * This must not ever occur. */ 5785 5786 tcp_prune_ofo_queue(sk, in_skb); 5787 5788 if (tcp_can_ingest(sk, in_skb)) 5789 return 0; 5790 5791 /* If we are really being abused, tell the caller to silently 5792 * drop receive data on the floor. It will get retransmitted 5793 * and hopefully then we'll have sufficient space. 5794 */ 5795 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED); 5796 5797 /* Massive buffer overcommit. */ 5798 tp->pred_flags = 0; 5799 return -1; 5800 } 5801 5802 static bool tcp_should_expand_sndbuf(struct sock *sk) 5803 { 5804 const struct tcp_sock *tp = tcp_sk(sk); 5805 5806 /* If the user specified a specific send buffer setting, do 5807 * not modify it. 5808 */ 5809 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 5810 return false; 5811 5812 /* If we are under global TCP memory pressure, do not expand. */ 5813 if (tcp_under_memory_pressure(sk)) { 5814 int unused_mem = sk_unused_reserved_mem(sk); 5815 5816 /* Adjust sndbuf according to reserved mem. But make sure 5817 * it never goes below SOCK_MIN_SNDBUF. 5818 * See sk_stream_moderate_sndbuf() for more details. 5819 */ 5820 if (unused_mem > SOCK_MIN_SNDBUF) 5821 WRITE_ONCE(sk->sk_sndbuf, unused_mem); 5822 5823 return false; 5824 } 5825 5826 /* If we are under soft global TCP memory pressure, do not expand. */ 5827 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 5828 return false; 5829 5830 /* If we filled the congestion window, do not expand. */ 5831 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)) 5832 return false; 5833 5834 return true; 5835 } 5836 5837 static void tcp_new_space(struct sock *sk) 5838 { 5839 struct tcp_sock *tp = tcp_sk(sk); 5840 5841 if (tcp_should_expand_sndbuf(sk)) { 5842 tcp_sndbuf_expand(sk); 5843 tp->snd_cwnd_stamp = tcp_jiffies32; 5844 } 5845 5846 INDIRECT_CALL_1(sk->sk_write_space, sk_stream_write_space, sk); 5847 } 5848 5849 /* Caller made space either from: 5850 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced) 5851 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt) 5852 * 5853 * We might be able to generate EPOLLOUT to the application if: 5854 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2 5855 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became 5856 * small enough that tcp_stream_memory_free() decides it 5857 * is time to generate EPOLLOUT. 5858 */ 5859 void tcp_check_space(struct sock *sk) 5860 { 5861 /* pairs with tcp_poll() */ 5862 smp_mb(); 5863 if (sk->sk_socket && 5864 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 5865 tcp_new_space(sk); 5866 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 5867 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 5868 } 5869 } 5870 5871 static inline void tcp_data_snd_check(struct sock *sk) 5872 { 5873 tcp_push_pending_frames(sk); 5874 tcp_check_space(sk); 5875 } 5876 5877 /* 5878 * Check if sending an ack is needed. 5879 */ 5880 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 5881 { 5882 struct tcp_sock *tp = tcp_sk(sk); 5883 unsigned long rtt, delay; 5884 5885 /* More than one full frame received... */ 5886 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 5887 /* ... and right edge of window advances far enough. 5888 * (tcp_recvmsg() will send ACK otherwise). 5889 * If application uses SO_RCVLOWAT, we want send ack now if 5890 * we have not received enough bytes to satisfy the condition. 5891 */ 5892 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat || 5893 __tcp_select_window(sk) >= tp->rcv_wnd)) || 5894 /* We ACK each frame or... */ 5895 tcp_in_quickack_mode(sk) || 5896 /* Protocol state mandates a one-time immediate ACK */ 5897 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) { 5898 /* If we are running from __release_sock() in user context, 5899 * Defer the ack until tcp_release_cb(). 5900 */ 5901 if (sock_owned_by_user_nocheck(sk) && 5902 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_backlog_ack_defer)) { 5903 set_bit(TCP_ACK_DEFERRED, &sk->sk_tsq_flags); 5904 return; 5905 } 5906 send_now: 5907 tcp_send_ack(sk); 5908 return; 5909 } 5910 5911 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5912 tcp_send_delayed_ack(sk); 5913 return; 5914 } 5915 5916 if (!tcp_is_sack(tp) || 5917 tp->compressed_ack >= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_nr)) 5918 goto send_now; 5919 5920 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) { 5921 tp->compressed_ack_rcv_nxt = tp->rcv_nxt; 5922 tp->dup_ack_counter = 0; 5923 } 5924 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) { 5925 tp->dup_ack_counter++; 5926 goto send_now; 5927 } 5928 tp->compressed_ack++; 5929 if (hrtimer_is_queued(&tp->compressed_ack_timer)) 5930 return; 5931 5932 /* compress ack timer : 5 % of rtt, but no more than tcp_comp_sack_delay_ns */ 5933 5934 rtt = tp->rcv_rtt_est.rtt_us; 5935 if (tp->srtt_us && tp->srtt_us < rtt) 5936 rtt = tp->srtt_us; 5937 5938 delay = min_t(unsigned long, 5939 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_delay_ns), 5940 rtt * (NSEC_PER_USEC >> 3)/20); 5941 sock_hold(sk); 5942 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay), 5943 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_comp_sack_slack_ns), 5944 HRTIMER_MODE_REL_PINNED_SOFT); 5945 } 5946 5947 static inline void tcp_ack_snd_check(struct sock *sk) 5948 { 5949 if (!inet_csk_ack_scheduled(sk)) { 5950 /* We sent a data segment already. */ 5951 return; 5952 } 5953 __tcp_ack_snd_check(sk, 1); 5954 } 5955 5956 /* 5957 * This routine is only called when we have urgent data 5958 * signaled. Its the 'slow' part of tcp_urg. It could be 5959 * moved inline now as tcp_urg is only called from one 5960 * place. We handle URGent data wrong. We have to - as 5961 * BSD still doesn't use the correction from RFC961. 5962 * For 1003.1g we should support a new option TCP_STDURG to permit 5963 * either form (or just set the sysctl tcp_stdurg). 5964 */ 5965 5966 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 5967 { 5968 struct tcp_sock *tp = tcp_sk(sk); 5969 u32 ptr = ntohs(th->urg_ptr); 5970 5971 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg)) 5972 ptr--; 5973 ptr += ntohl(th->seq); 5974 5975 /* Ignore urgent data that we've already seen and read. */ 5976 if (after(tp->copied_seq, ptr)) 5977 return; 5978 5979 /* Do not replay urg ptr. 5980 * 5981 * NOTE: interesting situation not covered by specs. 5982 * Misbehaving sender may send urg ptr, pointing to segment, 5983 * which we already have in ofo queue. We are not able to fetch 5984 * such data and will stay in TCP_URG_NOTYET until will be eaten 5985 * by recvmsg(). Seems, we are not obliged to handle such wicked 5986 * situations. But it is worth to think about possibility of some 5987 * DoSes using some hypothetical application level deadlock. 5988 */ 5989 if (before(ptr, tp->rcv_nxt)) 5990 return; 5991 5992 /* Do we already have a newer (or duplicate) urgent pointer? */ 5993 if (tp->urg_data && !after(ptr, tp->urg_seq)) 5994 return; 5995 5996 /* Tell the world about our new urgent pointer. */ 5997 sk_send_sigurg(sk); 5998 5999 /* We may be adding urgent data when the last byte read was 6000 * urgent. To do this requires some care. We cannot just ignore 6001 * tp->copied_seq since we would read the last urgent byte again 6002 * as data, nor can we alter copied_seq until this data arrives 6003 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 6004 * 6005 * NOTE. Double Dutch. Rendering to plain English: author of comment 6006 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 6007 * and expect that both A and B disappear from stream. This is _wrong_. 6008 * Though this happens in BSD with high probability, this is occasional. 6009 * Any application relying on this is buggy. Note also, that fix "works" 6010 * only in this artificial test. Insert some normal data between A and B and we will 6011 * decline of BSD again. Verdict: it is better to remove to trap 6012 * buggy users. 6013 */ 6014 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 6015 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 6016 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 6017 tp->copied_seq++; 6018 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 6019 __skb_unlink(skb, &sk->sk_receive_queue); 6020 __kfree_skb(skb); 6021 } 6022 } 6023 6024 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET); 6025 WRITE_ONCE(tp->urg_seq, ptr); 6026 6027 /* Disable header prediction. */ 6028 tp->pred_flags = 0; 6029 } 6030 6031 /* This is the 'fast' part of urgent handling. */ 6032 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 6033 { 6034 struct tcp_sock *tp = tcp_sk(sk); 6035 6036 /* Check if we get a new urgent pointer - normally not. */ 6037 if (unlikely(th->urg)) 6038 tcp_check_urg(sk, th); 6039 6040 /* Do we wait for any urgent data? - normally not... */ 6041 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) { 6042 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 6043 th->syn; 6044 6045 /* Is the urgent pointer pointing into this packet? */ 6046 if (ptr < skb->len) { 6047 u8 tmp; 6048 if (skb_copy_bits(skb, ptr, &tmp, 1)) 6049 BUG(); 6050 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp); 6051 if (!sock_flag(sk, SOCK_DEAD)) 6052 sk->sk_data_ready(sk); 6053 } 6054 } 6055 } 6056 6057 /* Accept RST for rcv_nxt - 1 after a FIN. 6058 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a 6059 * FIN is sent followed by a RST packet. The RST is sent with the same 6060 * sequence number as the FIN, and thus according to RFC 5961 a challenge 6061 * ACK should be sent. However, Mac OSX rate limits replies to challenge 6062 * ACKs on the closed socket. In addition middleboxes can drop either the 6063 * challenge ACK or a subsequent RST. 6064 */ 6065 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb) 6066 { 6067 const struct tcp_sock *tp = tcp_sk(sk); 6068 6069 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) && 6070 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK | 6071 TCPF_CLOSING)); 6072 } 6073 6074 /* Does PAWS and seqno based validation of an incoming segment, flags will 6075 * play significant role here. 6076 */ 6077 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 6078 const struct tcphdr *th, int syn_inerr) 6079 { 6080 struct tcp_sock *tp = tcp_sk(sk); 6081 bool accecn_reflector = false; 6082 SKB_DR(reason); 6083 6084 /* RFC1323: H1. Apply PAWS check first. */ 6085 if (!tcp_fast_parse_options(sock_net(sk), skb, th, tp) || 6086 !tp->rx_opt.saw_tstamp || 6087 tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW)) 6088 goto step1; 6089 6090 reason = tcp_disordered_ack_check(sk, skb); 6091 if (!reason) 6092 goto step1; 6093 /* Reset is accepted even if it did not pass PAWS. */ 6094 if (th->rst) 6095 goto step1; 6096 if (unlikely(th->syn)) 6097 goto syn_challenge; 6098 6099 /* Old ACK are common, increment PAWS_OLD_ACK 6100 * and do not send a dupack. 6101 */ 6102 if (reason == SKB_DROP_REASON_TCP_RFC7323_PAWS_ACK) { 6103 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWS_OLD_ACK); 6104 goto discard; 6105 } 6106 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 6107 if (!tcp_oow_rate_limited(sock_net(sk), skb, 6108 LINUX_MIB_TCPACKSKIPPEDPAWS, 6109 &tp->last_oow_ack_time)) 6110 tcp_send_dupack(sk, skb); 6111 goto discard; 6112 6113 step1: 6114 /* Step 1: check sequence number */ 6115 reason = tcp_sequence(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 6116 if (reason) { 6117 /* RFC793, page 37: "In all states except SYN-SENT, all reset 6118 * (RST) segments are validated by checking their SEQ-fields." 6119 * And page 69: "If an incoming segment is not acceptable, 6120 * an acknowledgment should be sent in reply (unless the RST 6121 * bit is set, if so drop the segment and return)". 6122 */ 6123 if (!th->rst) { 6124 if (th->syn) 6125 goto syn_challenge; 6126 6127 if (reason == SKB_DROP_REASON_TCP_INVALID_SEQUENCE || 6128 reason == SKB_DROP_REASON_TCP_INVALID_END_SEQUENCE) 6129 NET_INC_STATS(sock_net(sk), 6130 LINUX_MIB_BEYOND_WINDOW); 6131 if (!tcp_oow_rate_limited(sock_net(sk), skb, 6132 LINUX_MIB_TCPACKSKIPPEDSEQ, 6133 &tp->last_oow_ack_time)) 6134 tcp_send_dupack(sk, skb); 6135 } else if (tcp_reset_check(sk, skb)) { 6136 goto reset; 6137 } 6138 goto discard; 6139 } 6140 6141 /* Step 2: check RST bit */ 6142 if (th->rst) { 6143 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a 6144 * FIN and SACK too if available): 6145 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or 6146 * the right-most SACK block, 6147 * then 6148 * RESET the connection 6149 * else 6150 * Send a challenge ACK 6151 */ 6152 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt || 6153 tcp_reset_check(sk, skb)) 6154 goto reset; 6155 6156 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) { 6157 struct tcp_sack_block *sp = &tp->selective_acks[0]; 6158 int max_sack = sp[0].end_seq; 6159 int this_sack; 6160 6161 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; 6162 ++this_sack) { 6163 max_sack = after(sp[this_sack].end_seq, 6164 max_sack) ? 6165 sp[this_sack].end_seq : max_sack; 6166 } 6167 6168 if (TCP_SKB_CB(skb)->seq == max_sack) 6169 goto reset; 6170 } 6171 6172 /* Disable TFO if RST is out-of-order 6173 * and no data has been received 6174 * for current active TFO socket 6175 */ 6176 if (tp->syn_fastopen && !tp->data_segs_in && 6177 sk->sk_state == TCP_ESTABLISHED) 6178 tcp_fastopen_active_disable(sk); 6179 tcp_send_challenge_ack(sk, false); 6180 SKB_DR_SET(reason, TCP_RESET); 6181 goto discard; 6182 } 6183 6184 /* step 3: check security and precedence [ignored] */ 6185 6186 /* step 4: Check for a SYN 6187 * RFC 5961 4.2 : Send a challenge ack 6188 */ 6189 if (th->syn) { 6190 if (tcp_ecn_mode_accecn(tp)) { 6191 accecn_reflector = true; 6192 if (tp->rx_opt.accecn && 6193 tp->saw_accecn_opt < TCP_ACCECN_OPT_COUNTER_SEEN) { 6194 u8 saw_opt = tcp_accecn_option_init(skb, tp->rx_opt.accecn); 6195 6196 tcp_accecn_saw_opt_fail_recv(tp, saw_opt); 6197 tcp_accecn_opt_demand_min(sk, 1); 6198 } 6199 } 6200 if (sk->sk_state == TCP_SYN_RECV && sk->sk_socket && th->ack && 6201 TCP_SKB_CB(skb)->seq + 1 == TCP_SKB_CB(skb)->end_seq && 6202 TCP_SKB_CB(skb)->seq + 1 == tp->rcv_nxt && 6203 TCP_SKB_CB(skb)->ack_seq == tp->snd_nxt) 6204 goto pass; 6205 syn_challenge: 6206 if (syn_inerr) 6207 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6208 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE); 6209 tcp_send_challenge_ack(sk, accecn_reflector); 6210 SKB_DR_SET(reason, TCP_INVALID_SYN); 6211 goto discard; 6212 } 6213 6214 pass: 6215 bpf_skops_parse_hdr(sk, skb); 6216 6217 return true; 6218 6219 discard: 6220 tcp_drop_reason(sk, skb, reason); 6221 return false; 6222 6223 reset: 6224 tcp_reset(sk, skb); 6225 __kfree_skb(skb); 6226 return false; 6227 } 6228 6229 /* 6230 * TCP receive function for the ESTABLISHED state. 6231 * 6232 * It is split into a fast path and a slow path. The fast path is 6233 * disabled when: 6234 * - A zero window was announced from us - zero window probing 6235 * is only handled properly in the slow path. 6236 * - Out of order segments arrived. 6237 * - Urgent data is expected. 6238 * - There is no buffer space left 6239 * - Unexpected TCP flags/window values/header lengths are received 6240 * (detected by checking the TCP header against pred_flags) 6241 * - Data is sent in both directions. Fast path only supports pure senders 6242 * or pure receivers (this means either the sequence number or the ack 6243 * value must stay constant) 6244 * - Unexpected TCP option. 6245 * 6246 * When these conditions are not satisfied it drops into a standard 6247 * receive procedure patterned after RFC793 to handle all cases. 6248 * The first three cases are guaranteed by proper pred_flags setting, 6249 * the rest is checked inline. Fast processing is turned on in 6250 * tcp_data_queue when everything is OK. 6251 */ 6252 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb) 6253 { 6254 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED; 6255 const struct tcphdr *th = (const struct tcphdr *)skb->data; 6256 struct tcp_sock *tp = tcp_sk(sk); 6257 unsigned int len = skb->len; 6258 6259 /* TCP congestion window tracking */ 6260 trace_tcp_probe(sk, skb); 6261 6262 tcp_mstamp_refresh(tp); 6263 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst))) 6264 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb); 6265 /* 6266 * Header prediction. 6267 * The code loosely follows the one in the famous 6268 * "30 instruction TCP receive" Van Jacobson mail. 6269 * 6270 * Van's trick is to deposit buffers into socket queue 6271 * on a device interrupt, to call tcp_recv function 6272 * on the receive process context and checksum and copy 6273 * the buffer to user space. smart... 6274 * 6275 * Our current scheme is not silly either but we take the 6276 * extra cost of the net_bh soft interrupt processing... 6277 * We do checksum and copy also but from device to kernel. 6278 */ 6279 6280 tp->rx_opt.saw_tstamp = 0; 6281 tp->rx_opt.accecn = 0; 6282 6283 /* pred_flags is 0xS?10 << 16 + snd_wnd 6284 * if header_prediction is to be made 6285 * 'S' will always be tp->tcp_header_len >> 2 6286 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 6287 * turn it off (when there are holes in the receive 6288 * space for instance) 6289 * PSH flag is ignored. 6290 */ 6291 6292 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 6293 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 6294 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 6295 int tcp_header_len = tp->tcp_header_len; 6296 s32 delta = 0; 6297 int flag = 0; 6298 6299 /* Timestamp header prediction: tcp_header_len 6300 * is automatically equal to th->doff*4 due to pred_flags 6301 * match. 6302 */ 6303 6304 /* Check timestamp */ 6305 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 6306 /* No? Slow path! */ 6307 if (!tcp_parse_aligned_timestamp(tp, th)) 6308 goto slow_path; 6309 6310 delta = tp->rx_opt.rcv_tsval - 6311 tp->rx_opt.ts_recent; 6312 /* If PAWS failed, check it more carefully in slow path */ 6313 if (delta < 0) 6314 goto slow_path; 6315 6316 /* DO NOT update ts_recent here, if checksum fails 6317 * and timestamp was corrupted part, it will result 6318 * in a hung connection since we will drop all 6319 * future packets due to the PAWS test. 6320 */ 6321 } 6322 6323 if (len <= tcp_header_len) { 6324 /* Bulk data transfer: sender */ 6325 if (len == tcp_header_len) { 6326 /* Predicted packet is in window by definition. 6327 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 6328 * Hence, check seq<=rcv_wup reduces to: 6329 */ 6330 if (tcp_header_len == 6331 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 6332 tp->rcv_nxt == tp->rcv_wup) 6333 flag |= __tcp_replace_ts_recent(tp, 6334 delta); 6335 6336 tcp_ecn_received_counters(sk, skb, 0); 6337 6338 /* We know that such packets are checksummed 6339 * on entry. 6340 */ 6341 tcp_ack(sk, skb, flag); 6342 __kfree_skb(skb); 6343 tcp_data_snd_check(sk); 6344 /* When receiving pure ack in fast path, update 6345 * last ts ecr directly instead of calling 6346 * tcp_rcv_rtt_measure_ts() 6347 */ 6348 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 6349 return; 6350 } else { /* Header too small */ 6351 reason = SKB_DROP_REASON_PKT_TOO_SMALL; 6352 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6353 goto discard; 6354 } 6355 } else { 6356 int eaten = 0; 6357 bool fragstolen = false; 6358 6359 if (tcp_checksum_complete(skb)) 6360 goto csum_error; 6361 6362 if (after(TCP_SKB_CB(skb)->end_seq, 6363 tp->rcv_nxt + tcp_receive_window(tp))) 6364 goto validate; 6365 6366 if ((int)skb->truesize > sk->sk_forward_alloc) 6367 goto step5; 6368 6369 /* Predicted packet is in window by definition. 6370 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 6371 * Hence, check seq<=rcv_wup reduces to: 6372 */ 6373 if (tcp_header_len == 6374 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 6375 tp->rcv_nxt == tp->rcv_wup) 6376 flag |= __tcp_replace_ts_recent(tp, 6377 delta); 6378 6379 tcp_rcv_rtt_measure_ts(sk, skb); 6380 6381 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS); 6382 6383 /* Bulk data transfer: receiver */ 6384 tcp_cleanup_skb(skb); 6385 __skb_pull(skb, tcp_header_len); 6386 tcp_ecn_received_counters(sk, skb, 6387 len - tcp_header_len); 6388 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 6389 6390 tcp_event_data_recv(sk, skb); 6391 6392 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 6393 /* Well, only one small jumplet in fast path... */ 6394 tcp_ack(sk, skb, flag | FLAG_DATA); 6395 tcp_data_snd_check(sk); 6396 if (!inet_csk_ack_scheduled(sk)) 6397 goto no_ack; 6398 } else { 6399 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq); 6400 } 6401 6402 __tcp_ack_snd_check(sk, 0); 6403 no_ack: 6404 if (eaten) 6405 kfree_skb_partial(skb, fragstolen); 6406 tcp_data_ready(sk); 6407 return; 6408 } 6409 } 6410 6411 slow_path: 6412 if (len < (th->doff << 2) || tcp_checksum_complete(skb)) 6413 goto csum_error; 6414 6415 if (!th->ack && !th->rst && !th->syn) { 6416 reason = SKB_DROP_REASON_TCP_FLAGS; 6417 goto discard; 6418 } 6419 6420 /* 6421 * Standard slow path. 6422 */ 6423 validate: 6424 if (!tcp_validate_incoming(sk, skb, th, 1)) 6425 return; 6426 6427 step5: 6428 tcp_ecn_received_counters_payload(sk, skb); 6429 6430 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT); 6431 if ((int)reason < 0) { 6432 reason = -reason; 6433 goto discard; 6434 } 6435 tcp_rcv_rtt_measure_ts(sk, skb); 6436 6437 /* Process urgent data. */ 6438 tcp_urg(sk, skb, th); 6439 6440 /* step 7: process the segment text */ 6441 tcp_data_queue(sk, skb); 6442 6443 tcp_data_snd_check(sk); 6444 tcp_ack_snd_check(sk); 6445 return; 6446 6447 csum_error: 6448 reason = SKB_DROP_REASON_TCP_CSUM; 6449 trace_tcp_bad_csum(skb); 6450 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS); 6451 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6452 6453 discard: 6454 tcp_drop_reason(sk, skb, reason); 6455 } 6456 EXPORT_IPV6_MOD(tcp_rcv_established); 6457 6458 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb) 6459 { 6460 struct inet_connection_sock *icsk = inet_csk(sk); 6461 struct tcp_sock *tp = tcp_sk(sk); 6462 6463 tcp_mtup_init(sk); 6464 icsk->icsk_af_ops->rebuild_header(sk); 6465 tcp_init_metrics(sk); 6466 6467 /* Initialize the congestion window to start the transfer. 6468 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 6469 * retransmitted. In light of RFC6298 more aggressive 1sec 6470 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 6471 * retransmission has occurred. 6472 */ 6473 if (tp->total_retrans > 1 && tp->undo_marker) 6474 tcp_snd_cwnd_set(tp, 1); 6475 else 6476 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk))); 6477 tp->snd_cwnd_stamp = tcp_jiffies32; 6478 6479 bpf_skops_established(sk, bpf_op, skb); 6480 /* Initialize congestion control unless BPF initialized it already: */ 6481 if (!icsk->icsk_ca_initialized) 6482 tcp_init_congestion_control(sk); 6483 tcp_init_buffer_space(sk); 6484 } 6485 6486 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) 6487 { 6488 struct tcp_sock *tp = tcp_sk(sk); 6489 struct inet_connection_sock *icsk = inet_csk(sk); 6490 6491 tcp_ao_finish_connect(sk, skb); 6492 tcp_set_state(sk, TCP_ESTABLISHED); 6493 icsk->icsk_ack.lrcvtime = tcp_jiffies32; 6494 6495 if (skb) { 6496 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); 6497 security_inet_conn_established(sk, skb); 6498 sk_mark_napi_id(sk, skb); 6499 } 6500 6501 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb); 6502 6503 /* Prevent spurious tcp_cwnd_restart() on first data 6504 * packet. 6505 */ 6506 tp->lsndtime = tcp_jiffies32; 6507 6508 if (sock_flag(sk, SOCK_KEEPOPEN)) 6509 tcp_reset_keepalive_timer(sk, keepalive_time_when(tp)); 6510 6511 if (!tp->rx_opt.snd_wscale) 6512 __tcp_fast_path_on(tp, tp->snd_wnd); 6513 else 6514 tp->pred_flags = 0; 6515 } 6516 6517 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack, 6518 struct tcp_fastopen_cookie *cookie) 6519 { 6520 struct tcp_sock *tp = tcp_sk(sk); 6521 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL; 6522 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0; 6523 bool syn_drop = false; 6524 6525 if (mss == READ_ONCE(tp->rx_opt.user_mss)) { 6526 struct tcp_options_received opt; 6527 6528 /* Get original SYNACK MSS value if user MSS sets mss_clamp */ 6529 tcp_clear_options(&opt); 6530 opt.user_mss = opt.mss_clamp = 0; 6531 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL); 6532 mss = opt.mss_clamp; 6533 } 6534 6535 if (!tp->syn_fastopen) { 6536 /* Ignore an unsolicited cookie */ 6537 cookie->len = -1; 6538 } else if (tp->total_retrans) { 6539 /* SYN timed out and the SYN-ACK neither has a cookie nor 6540 * acknowledges data. Presumably the remote received only 6541 * the retransmitted (regular) SYNs: either the original 6542 * SYN-data or the corresponding SYN-ACK was dropped. 6543 */ 6544 syn_drop = (cookie->len < 0 && data); 6545 } else if (cookie->len < 0 && !tp->syn_data) { 6546 /* We requested a cookie but didn't get it. If we did not use 6547 * the (old) exp opt format then try so next time (try_exp=1). 6548 * Otherwise we go back to use the RFC7413 opt (try_exp=2). 6549 */ 6550 try_exp = tp->syn_fastopen_exp ? 2 : 1; 6551 } 6552 6553 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp); 6554 6555 if (data) { /* Retransmit unacked data in SYN */ 6556 if (tp->total_retrans) 6557 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED; 6558 else 6559 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED; 6560 skb_rbtree_walk_from(data) 6561 tcp_mark_skb_lost(sk, data); 6562 tcp_non_congestion_loss_retransmit(sk); 6563 NET_INC_STATS(sock_net(sk), 6564 LINUX_MIB_TCPFASTOPENACTIVEFAIL); 6565 return true; 6566 } 6567 tp->syn_data_acked = tp->syn_data; 6568 if (tp->syn_data_acked) { 6569 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 6570 /* SYN-data is counted as two separate packets in tcp_ack() */ 6571 if (tp->delivered > 1) 6572 --tp->delivered; 6573 } 6574 6575 tcp_fastopen_add_skb(sk, synack); 6576 6577 return false; 6578 } 6579 6580 static void smc_check_reset_syn(struct tcp_sock *tp) 6581 { 6582 #if IS_ENABLED(CONFIG_SMC) 6583 if (static_branch_unlikely(&tcp_have_smc)) { 6584 if (tp->syn_smc && !tp->rx_opt.smc_ok) 6585 tp->syn_smc = 0; 6586 } 6587 #endif 6588 } 6589 6590 static void tcp_try_undo_spurious_syn(struct sock *sk) 6591 { 6592 struct tcp_sock *tp = tcp_sk(sk); 6593 u32 syn_stamp; 6594 6595 /* undo_marker is set when SYN or SYNACK times out. The timeout is 6596 * spurious if the ACK's timestamp option echo value matches the 6597 * original SYN timestamp. 6598 */ 6599 syn_stamp = tp->retrans_stamp; 6600 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp && 6601 syn_stamp == tp->rx_opt.rcv_tsecr) 6602 tp->undo_marker = 0; 6603 } 6604 6605 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 6606 const struct tcphdr *th) 6607 { 6608 struct inet_connection_sock *icsk = inet_csk(sk); 6609 struct tcp_sock *tp = tcp_sk(sk); 6610 struct tcp_fastopen_cookie foc = { .len = -1 }; 6611 int saved_clamp = tp->rx_opt.mss_clamp; 6612 bool fastopen_fail; 6613 SKB_DR(reason); 6614 6615 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc); 6616 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 6617 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 6618 6619 if (th->ack) { 6620 /* rfc793: 6621 * "If the state is SYN-SENT then 6622 * first check the ACK bit 6623 * If the ACK bit is set 6624 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 6625 * a reset (unless the RST bit is set, if so drop 6626 * the segment and return)" 6627 */ 6628 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) || 6629 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 6630 /* Previous FIN/ACK or RST/ACK might be ignored. */ 6631 if (icsk->icsk_retransmits == 0) 6632 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 6633 TCP_TIMEOUT_MIN, false); 6634 SKB_DR_SET(reason, TCP_INVALID_ACK_SEQUENCE); 6635 goto reset_and_undo; 6636 } 6637 6638 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 6639 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 6640 tcp_time_stamp_ts(tp))) { 6641 NET_INC_STATS(sock_net(sk), 6642 LINUX_MIB_PAWSACTIVEREJECTED); 6643 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 6644 goto reset_and_undo; 6645 } 6646 6647 /* Now ACK is acceptable. 6648 * 6649 * "If the RST bit is set 6650 * If the ACK was acceptable then signal the user "error: 6651 * connection reset", drop the segment, enter CLOSED state, 6652 * delete TCB, and return." 6653 */ 6654 6655 if (th->rst) { 6656 tcp_reset(sk, skb); 6657 consume: 6658 __kfree_skb(skb); 6659 return 0; 6660 } 6661 6662 /* rfc793: 6663 * "fifth, if neither of the SYN or RST bits is set then 6664 * drop the segment and return." 6665 * 6666 * See note below! 6667 * --ANK(990513) 6668 */ 6669 if (!th->syn) { 6670 SKB_DR_SET(reason, TCP_FLAGS); 6671 goto discard_and_undo; 6672 } 6673 /* rfc793: 6674 * "If the SYN bit is on ... 6675 * are acceptable then ... 6676 * (our SYN has been ACKed), change the connection 6677 * state to ESTABLISHED..." 6678 */ 6679 6680 if (tcp_ecn_mode_any(tp)) 6681 tcp_ecn_rcv_synack(sk, skb, th, 6682 TCP_SKB_CB(skb)->ip_dsfield); 6683 6684 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6685 tcp_try_undo_spurious_syn(sk); 6686 tcp_ack(sk, skb, FLAG_SLOWPATH); 6687 6688 /* Ok.. it's good. Set up sequence numbers and 6689 * move to established. 6690 */ 6691 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6692 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6693 6694 /* RFC1323: The window in SYN & SYN/ACK segments is 6695 * never scaled. 6696 */ 6697 tp->snd_wnd = ntohs(th->window); 6698 6699 if (!tp->rx_opt.wscale_ok) { 6700 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 6701 WRITE_ONCE(tp->window_clamp, 6702 min(tp->window_clamp, 65535U)); 6703 } 6704 6705 if (tp->rx_opt.saw_tstamp) { 6706 tp->rx_opt.tstamp_ok = 1; 6707 tp->tcp_header_len = 6708 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6709 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6710 tcp_store_ts_recent(tp); 6711 } else { 6712 tp->tcp_header_len = sizeof(struct tcphdr); 6713 } 6714 6715 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6716 tcp_initialize_rcv_mss(sk); 6717 6718 /* Remember, tcp_poll() does not lock socket! 6719 * Change state from SYN-SENT only after copied_seq 6720 * is initialized. */ 6721 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6722 6723 smc_check_reset_syn(tp); 6724 6725 smp_mb(); 6726 6727 tcp_finish_connect(sk, skb); 6728 6729 fastopen_fail = (tp->syn_fastopen || tp->syn_data) && 6730 tcp_rcv_fastopen_synack(sk, skb, &foc); 6731 6732 if (!sock_flag(sk, SOCK_DEAD)) { 6733 sk->sk_state_change(sk); 6734 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6735 } 6736 if (fastopen_fail) 6737 return -1; 6738 if (sk->sk_write_pending || 6739 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) || 6740 inet_csk_in_pingpong_mode(sk)) { 6741 /* Save one ACK. Data will be ready after 6742 * several ticks, if write_pending is set. 6743 * 6744 * It may be deleted, but with this feature tcpdumps 6745 * look so _wonderfully_ clever, that I was not able 6746 * to stand against the temptation 8) --ANK 6747 */ 6748 inet_csk_schedule_ack(sk); 6749 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 6750 tcp_reset_xmit_timer(sk, ICSK_TIME_DACK, 6751 TCP_DELACK_MAX, false); 6752 goto consume; 6753 } 6754 tcp_send_ack_reflect_ect(sk, tcp_ecn_mode_accecn(tp)); 6755 return -1; 6756 } 6757 6758 /* No ACK in the segment */ 6759 6760 if (th->rst) { 6761 /* rfc793: 6762 * "If the RST bit is set 6763 * 6764 * Otherwise (no ACK) drop the segment and return." 6765 */ 6766 SKB_DR_SET(reason, TCP_RESET); 6767 goto discard_and_undo; 6768 } 6769 6770 /* PAWS check. */ 6771 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 6772 tcp_paws_reject(&tp->rx_opt, 0)) { 6773 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 6774 goto discard_and_undo; 6775 } 6776 if (th->syn) { 6777 /* We see SYN without ACK. It is attempt of 6778 * simultaneous connect with crossed SYNs. 6779 * Particularly, it can be connect to self. 6780 */ 6781 #ifdef CONFIG_TCP_AO 6782 struct tcp_ao_info *ao; 6783 6784 ao = rcu_dereference_protected(tp->ao_info, 6785 lockdep_sock_is_held(sk)); 6786 if (ao) { 6787 WRITE_ONCE(ao->risn, th->seq); 6788 ao->rcv_sne = 0; 6789 } 6790 #endif 6791 tcp_set_state(sk, TCP_SYN_RECV); 6792 6793 if (tp->rx_opt.saw_tstamp) { 6794 tp->rx_opt.tstamp_ok = 1; 6795 tcp_store_ts_recent(tp); 6796 tp->tcp_header_len = 6797 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6798 } else { 6799 tp->tcp_header_len = sizeof(struct tcphdr); 6800 } 6801 6802 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6803 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6804 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6805 6806 /* RFC1323: The window in SYN & SYN/ACK segments is 6807 * never scaled. 6808 */ 6809 tp->snd_wnd = ntohs(th->window); 6810 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 6811 tp->max_window = tp->snd_wnd; 6812 6813 tcp_ecn_rcv_syn(tp, th, skb); 6814 6815 tcp_mtup_init(sk); 6816 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6817 tcp_initialize_rcv_mss(sk); 6818 6819 tcp_send_synack(sk); 6820 #if 0 6821 /* Note, we could accept data and URG from this segment. 6822 * There are no obstacles to make this (except that we must 6823 * either change tcp_recvmsg() to prevent it from returning data 6824 * before 3WHS completes per RFC793, or employ TCP Fast Open). 6825 * 6826 * However, if we ignore data in ACKless segments sometimes, 6827 * we have no reasons to accept it sometimes. 6828 * Also, seems the code doing it in step6 of tcp_rcv_state_process 6829 * is not flawless. So, discard packet for sanity. 6830 * Uncomment this return to process the data. 6831 */ 6832 return -1; 6833 #else 6834 goto consume; 6835 #endif 6836 } 6837 /* "fifth, if neither of the SYN or RST bits is set then 6838 * drop the segment and return." 6839 */ 6840 6841 discard_and_undo: 6842 tcp_clear_options(&tp->rx_opt); 6843 tp->rx_opt.mss_clamp = saved_clamp; 6844 tcp_drop_reason(sk, skb, reason); 6845 return 0; 6846 6847 reset_and_undo: 6848 tcp_clear_options(&tp->rx_opt); 6849 tp->rx_opt.mss_clamp = saved_clamp; 6850 /* we can reuse/return @reason to its caller to handle the exception */ 6851 return reason; 6852 } 6853 6854 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk) 6855 { 6856 struct tcp_sock *tp = tcp_sk(sk); 6857 struct request_sock *req; 6858 6859 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows 6860 * undo. If peer SACKs triggered fast recovery, we can't undo here. 6861 */ 6862 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out) 6863 tcp_try_undo_recovery(sk); 6864 6865 tcp_update_rto_time(tp); 6866 WRITE_ONCE(inet_csk(sk)->icsk_retransmits, 0); 6867 /* In tcp_fastopen_synack_timer() on the first SYNACK RTO we set 6868 * retrans_stamp but don't enter CA_Loss, so in case that happened we 6869 * need to zero retrans_stamp here to prevent spurious 6870 * retransmits_timed_out(). However, if the ACK of our SYNACK caused us 6871 * to enter CA_Recovery then we need to leave retrans_stamp as it was 6872 * set entering CA_Recovery, for correct retransmits_timed_out() and 6873 * undo behavior. 6874 */ 6875 tcp_retrans_stamp_cleanup(sk); 6876 6877 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1, 6878 * we no longer need req so release it. 6879 */ 6880 req = rcu_dereference_protected(tp->fastopen_rsk, 6881 lockdep_sock_is_held(sk)); 6882 reqsk_fastopen_remove(sk, req, false); 6883 6884 /* Re-arm the timer because data may have been sent out. 6885 * This is similar to the regular data transmission case 6886 * when new data has just been ack'ed. 6887 * 6888 * (TFO) - we could try to be more aggressive and 6889 * retransmitting any data sooner based on when they 6890 * are sent out. 6891 */ 6892 tcp_rearm_rto(sk); 6893 } 6894 6895 /* 6896 * This function implements the receiving procedure of RFC 793 for 6897 * all states except ESTABLISHED and TIME_WAIT. 6898 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 6899 * address independent. 6900 */ 6901 6902 enum skb_drop_reason 6903 tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb) 6904 { 6905 struct tcp_sock *tp = tcp_sk(sk); 6906 struct inet_connection_sock *icsk = inet_csk(sk); 6907 const struct tcphdr *th = tcp_hdr(skb); 6908 struct request_sock *req; 6909 int queued = 0; 6910 SKB_DR(reason); 6911 6912 switch (sk->sk_state) { 6913 case TCP_CLOSE: 6914 SKB_DR_SET(reason, TCP_CLOSE); 6915 goto discard; 6916 6917 case TCP_LISTEN: 6918 if (th->ack) 6919 return SKB_DROP_REASON_TCP_FLAGS; 6920 6921 if (th->rst) { 6922 SKB_DR_SET(reason, TCP_RESET); 6923 goto discard; 6924 } 6925 if (th->syn) { 6926 if (th->fin) { 6927 SKB_DR_SET(reason, TCP_FLAGS); 6928 goto discard; 6929 } 6930 /* It is possible that we process SYN packets from backlog, 6931 * so we need to make sure to disable BH and RCU right there. 6932 */ 6933 rcu_read_lock(); 6934 local_bh_disable(); 6935 icsk->icsk_af_ops->conn_request(sk, skb); 6936 local_bh_enable(); 6937 rcu_read_unlock(); 6938 6939 consume_skb(skb); 6940 return 0; 6941 } 6942 SKB_DR_SET(reason, TCP_FLAGS); 6943 goto discard; 6944 6945 case TCP_SYN_SENT: 6946 tp->rx_opt.saw_tstamp = 0; 6947 tcp_mstamp_refresh(tp); 6948 queued = tcp_rcv_synsent_state_process(sk, skb, th); 6949 if (queued >= 0) 6950 return queued; 6951 6952 /* Do step6 onward by hand. */ 6953 tcp_urg(sk, skb, th); 6954 __kfree_skb(skb); 6955 tcp_data_snd_check(sk); 6956 return 0; 6957 } 6958 6959 tcp_mstamp_refresh(tp); 6960 tp->rx_opt.saw_tstamp = 0; 6961 req = rcu_dereference_protected(tp->fastopen_rsk, 6962 lockdep_sock_is_held(sk)); 6963 if (req) { 6964 bool req_stolen; 6965 6966 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 6967 sk->sk_state != TCP_FIN_WAIT1); 6968 6969 SKB_DR_SET(reason, TCP_FASTOPEN); 6970 if (!tcp_check_req(sk, skb, req, true, &req_stolen, &reason)) 6971 goto discard; 6972 } 6973 6974 if (!th->ack && !th->rst && !th->syn) { 6975 SKB_DR_SET(reason, TCP_FLAGS); 6976 goto discard; 6977 } 6978 if (!tcp_validate_incoming(sk, skb, th, 0)) 6979 return 0; 6980 6981 /* step 5: check the ACK field */ 6982 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | 6983 FLAG_UPDATE_TS_RECENT | 6984 FLAG_NO_CHALLENGE_ACK); 6985 6986 if ((int)reason <= 0) { 6987 if (sk->sk_state == TCP_SYN_RECV) { 6988 /* send one RST */ 6989 if (!reason) 6990 return SKB_DROP_REASON_TCP_OLD_ACK; 6991 return -reason; 6992 } 6993 /* accept old ack during closing */ 6994 if ((int)reason < 0) { 6995 tcp_send_challenge_ack(sk, false); 6996 reason = -reason; 6997 goto discard; 6998 } 6999 } 7000 SKB_DR_SET(reason, NOT_SPECIFIED); 7001 switch (sk->sk_state) { 7002 case TCP_SYN_RECV: 7003 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */ 7004 if (!tp->srtt_us) 7005 tcp_synack_rtt_meas(sk, req); 7006 7007 if (tp->rx_opt.tstamp_ok) 7008 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 7009 7010 if (req) { 7011 tcp_rcv_synrecv_state_fastopen(sk); 7012 } else { 7013 tcp_try_undo_spurious_syn(sk); 7014 tp->retrans_stamp = 0; 7015 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, 7016 skb); 7017 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 7018 } 7019 tcp_ao_established(sk); 7020 smp_mb(); 7021 tcp_set_state(sk, TCP_ESTABLISHED); 7022 sk->sk_state_change(sk); 7023 7024 /* Note, that this wakeup is only for marginal crossed SYN case. 7025 * Passively open sockets are not waked up, because 7026 * sk->sk_sleep == NULL and sk->sk_socket == NULL. 7027 */ 7028 if (sk->sk_socket) 7029 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 7030 7031 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 7032 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; 7033 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 7034 7035 if (!inet_csk(sk)->icsk_ca_ops->cong_control) 7036 tcp_update_pacing_rate(sk); 7037 7038 /* Prevent spurious tcp_cwnd_restart() on first data packet */ 7039 tp->lsndtime = tcp_jiffies32; 7040 7041 tcp_initialize_rcv_mss(sk); 7042 if (tcp_ecn_mode_accecn(tp)) 7043 tcp_accecn_third_ack(sk, skb, tp->syn_ect_snt); 7044 tcp_fast_path_on(tp); 7045 if (sk->sk_shutdown & SEND_SHUTDOWN) 7046 tcp_shutdown(sk, SEND_SHUTDOWN); 7047 7048 break; 7049 7050 case TCP_FIN_WAIT1: { 7051 int tmo; 7052 7053 if (req) 7054 tcp_rcv_synrecv_state_fastopen(sk); 7055 7056 if (tp->snd_una != tp->write_seq) 7057 break; 7058 7059 tcp_set_state(sk, TCP_FIN_WAIT2); 7060 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN); 7061 7062 sk_dst_confirm(sk); 7063 7064 if (!sock_flag(sk, SOCK_DEAD)) { 7065 /* Wake up lingering close() */ 7066 sk->sk_state_change(sk); 7067 break; 7068 } 7069 7070 if (READ_ONCE(tp->linger2) < 0) { 7071 tcp_done(sk); 7072 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 7073 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 7074 } 7075 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 7076 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 7077 /* Receive out of order FIN after close() */ 7078 if (tp->syn_fastopen && th->fin) 7079 tcp_fastopen_active_disable(sk); 7080 tcp_done(sk); 7081 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 7082 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 7083 } 7084 7085 tmo = tcp_fin_time(sk); 7086 if (tmo > TCP_TIMEWAIT_LEN) { 7087 tcp_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 7088 } else if (th->fin || sock_owned_by_user(sk)) { 7089 /* Bad case. We could lose such FIN otherwise. 7090 * It is not a big problem, but it looks confusing 7091 * and not so rare event. We still can lose it now, 7092 * if it spins in bh_lock_sock(), but it is really 7093 * marginal case. 7094 */ 7095 tcp_reset_keepalive_timer(sk, tmo); 7096 } else { 7097 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 7098 goto consume; 7099 } 7100 break; 7101 } 7102 7103 case TCP_CLOSING: 7104 if (tp->snd_una == tp->write_seq) { 7105 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 7106 goto consume; 7107 } 7108 break; 7109 7110 case TCP_LAST_ACK: 7111 if (tp->snd_una == tp->write_seq) { 7112 tcp_update_metrics(sk); 7113 tcp_done(sk); 7114 goto consume; 7115 } 7116 break; 7117 } 7118 7119 /* step 6: check the URG bit */ 7120 tcp_urg(sk, skb, th); 7121 7122 /* step 7: process the segment text */ 7123 switch (sk->sk_state) { 7124 case TCP_CLOSE_WAIT: 7125 case TCP_CLOSING: 7126 case TCP_LAST_ACK: 7127 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 7128 /* If a subflow has been reset, the packet should not 7129 * continue to be processed, drop the packet. 7130 */ 7131 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) 7132 goto discard; 7133 break; 7134 } 7135 fallthrough; 7136 case TCP_FIN_WAIT1: 7137 case TCP_FIN_WAIT2: 7138 /* RFC 793 says to queue data in these states, 7139 * RFC 1122 says we MUST send a reset. 7140 * BSD 4.4 also does reset. 7141 */ 7142 if (sk->sk_shutdown & RCV_SHUTDOWN) { 7143 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 7144 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 7145 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 7146 tcp_reset(sk, skb); 7147 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 7148 } 7149 } 7150 fallthrough; 7151 case TCP_ESTABLISHED: 7152 tcp_data_queue(sk, skb); 7153 queued = 1; 7154 break; 7155 } 7156 7157 /* tcp_data could move socket to TIME-WAIT */ 7158 if (sk->sk_state != TCP_CLOSE) { 7159 tcp_data_snd_check(sk); 7160 tcp_ack_snd_check(sk); 7161 } 7162 7163 if (!queued) { 7164 discard: 7165 tcp_drop_reason(sk, skb, reason); 7166 } 7167 return 0; 7168 7169 consume: 7170 __kfree_skb(skb); 7171 return 0; 7172 } 7173 EXPORT_IPV6_MOD(tcp_rcv_state_process); 7174 7175 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family) 7176 { 7177 struct inet_request_sock *ireq = inet_rsk(req); 7178 7179 if (family == AF_INET) 7180 net_dbg_ratelimited("drop open request from %pI4/%u\n", 7181 &ireq->ir_rmt_addr, port); 7182 #if IS_ENABLED(CONFIG_IPV6) 7183 else if (family == AF_INET6) 7184 net_dbg_ratelimited("drop open request from %pI6/%u\n", 7185 &ireq->ir_v6_rmt_addr, port); 7186 #endif 7187 } 7188 7189 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set 7190 * 7191 * If we receive a SYN packet with these bits set, it means a 7192 * network is playing bad games with TOS bits. In order to 7193 * avoid possible false congestion notifications, we disable 7194 * TCP ECN negotiation. 7195 * 7196 * Exception: tcp_ca wants ECN. This is required for DCTCP 7197 * congestion control: Linux DCTCP asserts ECT on all packets, 7198 * including SYN, which is most optimal solution; however, 7199 * others, such as FreeBSD do not. 7200 * 7201 * Exception: At least one of the reserved bits of the TCP header (th->res1) is 7202 * set, indicating the use of a future TCP extension (such as AccECN). See 7203 * RFC8311 §4.3 which updates RFC3168 to allow the development of such 7204 * extensions. 7205 */ 7206 static void tcp_ecn_create_request(struct request_sock *req, 7207 const struct sk_buff *skb, 7208 const struct sock *listen_sk, 7209 const struct dst_entry *dst) 7210 { 7211 const struct tcphdr *th = tcp_hdr(skb); 7212 const struct net *net = sock_net(listen_sk); 7213 bool th_ecn = th->ece && th->cwr; 7214 bool ect, ecn_ok; 7215 u32 ecn_ok_dst; 7216 7217 if (tcp_accecn_syn_requested(th) && 7218 READ_ONCE(net->ipv4.sysctl_tcp_ecn) >= 3) { 7219 inet_rsk(req)->ecn_ok = 1; 7220 tcp_rsk(req)->accecn_ok = 1; 7221 tcp_rsk(req)->syn_ect_rcv = TCP_SKB_CB(skb)->ip_dsfield & 7222 INET_ECN_MASK; 7223 return; 7224 } 7225 7226 if (!th_ecn) 7227 return; 7228 7229 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield); 7230 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK); 7231 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst; 7232 7233 if (((!ect || th->res1 || th->ae) && ecn_ok) || 7234 tcp_ca_needs_ecn(listen_sk) || 7235 (ecn_ok_dst & DST_FEATURE_ECN_CA) || 7236 tcp_bpf_ca_needs_ecn((struct sock *)req)) 7237 inet_rsk(req)->ecn_ok = 1; 7238 } 7239 7240 static void tcp_openreq_init(struct request_sock *req, 7241 const struct tcp_options_received *rx_opt, 7242 struct sk_buff *skb, const struct sock *sk) 7243 { 7244 struct inet_request_sock *ireq = inet_rsk(req); 7245 7246 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */ 7247 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq; 7248 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 7249 tcp_rsk(req)->snt_synack = 0; 7250 tcp_rsk(req)->snt_tsval_first = 0; 7251 tcp_rsk(req)->last_oow_ack_time = 0; 7252 tcp_rsk(req)->accecn_ok = 0; 7253 tcp_rsk(req)->saw_accecn_opt = TCP_ACCECN_OPT_NOT_SEEN; 7254 tcp_rsk(req)->accecn_fail_mode = 0; 7255 tcp_rsk(req)->syn_ect_rcv = 0; 7256 tcp_rsk(req)->syn_ect_snt = 0; 7257 req->mss = rx_opt->mss_clamp; 7258 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0; 7259 ireq->tstamp_ok = rx_opt->tstamp_ok; 7260 ireq->sack_ok = rx_opt->sack_ok; 7261 ireq->snd_wscale = rx_opt->snd_wscale; 7262 ireq->wscale_ok = rx_opt->wscale_ok; 7263 ireq->acked = 0; 7264 ireq->ecn_ok = 0; 7265 ireq->ir_rmt_port = tcp_hdr(skb)->source; 7266 ireq->ir_num = ntohs(tcp_hdr(skb)->dest); 7267 ireq->ir_mark = inet_request_mark(sk, skb); 7268 #if IS_ENABLED(CONFIG_SMC) 7269 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested && 7270 tcp_sk(sk)->smc_hs_congested(sk)); 7271 #endif 7272 } 7273 7274 /* 7275 * Return true if a syncookie should be sent 7276 */ 7277 static bool tcp_syn_flood_action(struct sock *sk, const char *proto) 7278 { 7279 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; 7280 const char *msg = "Dropping request"; 7281 struct net *net = sock_net(sk); 7282 bool want_cookie = false; 7283 u8 syncookies; 7284 7285 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); 7286 7287 #ifdef CONFIG_SYN_COOKIES 7288 if (syncookies) { 7289 msg = "Sending cookies"; 7290 want_cookie = true; 7291 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES); 7292 } else 7293 #endif 7294 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP); 7295 7296 if (syncookies != 2 && !READ_ONCE(queue->synflood_warned)) { 7297 WRITE_ONCE(queue->synflood_warned, 1); 7298 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) { 7299 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n", 7300 proto, inet6_rcv_saddr(sk), 7301 sk->sk_num, msg); 7302 } else { 7303 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n", 7304 proto, &sk->sk_rcv_saddr, 7305 sk->sk_num, msg); 7306 } 7307 } 7308 7309 return want_cookie; 7310 } 7311 7312 static void tcp_reqsk_record_syn(const struct sock *sk, 7313 struct request_sock *req, 7314 const struct sk_buff *skb) 7315 { 7316 if (tcp_sk(sk)->save_syn) { 7317 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb); 7318 struct saved_syn *saved_syn; 7319 u32 mac_hdrlen; 7320 void *base; 7321 7322 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */ 7323 base = skb_mac_header(skb); 7324 mac_hdrlen = skb_mac_header_len(skb); 7325 len += mac_hdrlen; 7326 } else { 7327 base = skb_network_header(skb); 7328 mac_hdrlen = 0; 7329 } 7330 7331 saved_syn = kmalloc(struct_size(saved_syn, data, len), 7332 GFP_ATOMIC); 7333 if (saved_syn) { 7334 saved_syn->mac_hdrlen = mac_hdrlen; 7335 saved_syn->network_hdrlen = skb_network_header_len(skb); 7336 saved_syn->tcp_hdrlen = tcp_hdrlen(skb); 7337 memcpy(saved_syn->data, base, len); 7338 req->saved_syn = saved_syn; 7339 } 7340 } 7341 } 7342 7343 /* If a SYN cookie is required and supported, returns a clamped MSS value to be 7344 * used for SYN cookie generation. 7345 */ 7346 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops, 7347 const struct tcp_request_sock_ops *af_ops, 7348 struct sock *sk, struct tcphdr *th) 7349 { 7350 struct tcp_sock *tp = tcp_sk(sk); 7351 u16 mss; 7352 7353 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 && 7354 !inet_csk_reqsk_queue_is_full(sk)) 7355 return 0; 7356 7357 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name)) 7358 return 0; 7359 7360 if (sk_acceptq_is_full(sk)) { 7361 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 7362 return 0; 7363 } 7364 7365 mss = tcp_parse_mss_option(th, READ_ONCE(tp->rx_opt.user_mss)); 7366 if (!mss) 7367 mss = af_ops->mss_clamp; 7368 7369 return mss; 7370 } 7371 EXPORT_IPV6_MOD_GPL(tcp_get_syncookie_mss); 7372 7373 int tcp_conn_request(struct request_sock_ops *rsk_ops, 7374 const struct tcp_request_sock_ops *af_ops, 7375 struct sock *sk, struct sk_buff *skb) 7376 { 7377 struct tcp_fastopen_cookie foc = { .len = -1 }; 7378 struct tcp_options_received tmp_opt; 7379 const struct tcp_sock *tp = tcp_sk(sk); 7380 struct net *net = sock_net(sk); 7381 struct sock *fastopen_sk = NULL; 7382 struct request_sock *req; 7383 bool want_cookie = false; 7384 struct dst_entry *dst; 7385 struct flowi fl; 7386 u8 syncookies; 7387 u32 isn; 7388 7389 #ifdef CONFIG_TCP_AO 7390 const struct tcp_ao_hdr *aoh; 7391 #endif 7392 7393 isn = __this_cpu_read(tcp_tw_isn); 7394 if (isn) { 7395 /* TW buckets are converted to open requests without 7396 * limitations, they conserve resources and peer is 7397 * evidently real one. 7398 */ 7399 __this_cpu_write(tcp_tw_isn, 0); 7400 } else { 7401 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); 7402 7403 if (syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) { 7404 want_cookie = tcp_syn_flood_action(sk, 7405 rsk_ops->slab_name); 7406 if (!want_cookie) 7407 goto drop; 7408 } 7409 } 7410 7411 if (sk_acceptq_is_full(sk)) { 7412 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 7413 goto drop; 7414 } 7415 7416 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie); 7417 if (!req) 7418 goto drop; 7419 7420 req->syncookie = want_cookie; 7421 tcp_rsk(req)->af_specific = af_ops; 7422 tcp_rsk(req)->ts_off = 0; 7423 tcp_rsk(req)->req_usec_ts = false; 7424 #if IS_ENABLED(CONFIG_MPTCP) 7425 tcp_rsk(req)->is_mptcp = 0; 7426 #endif 7427 7428 tcp_clear_options(&tmp_opt); 7429 tmp_opt.mss_clamp = af_ops->mss_clamp; 7430 tmp_opt.user_mss = READ_ONCE(tp->rx_opt.user_mss); 7431 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, 7432 want_cookie ? NULL : &foc); 7433 7434 if (want_cookie && !tmp_opt.saw_tstamp) 7435 tcp_clear_options(&tmp_opt); 7436 7437 if (IS_ENABLED(CONFIG_SMC) && want_cookie) 7438 tmp_opt.smc_ok = 0; 7439 7440 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp; 7441 tcp_openreq_init(req, &tmp_opt, skb, sk); 7442 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk); 7443 7444 /* Note: tcp_v6_init_req() might override ir_iif for link locals */ 7445 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb); 7446 7447 dst = af_ops->route_req(sk, skb, &fl, req, isn); 7448 if (!dst) 7449 goto drop_and_free; 7450 7451 if (tmp_opt.tstamp_ok) { 7452 tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst); 7453 tcp_rsk(req)->ts_off = af_ops->init_ts_off(net, skb); 7454 } 7455 if (!want_cookie && !isn) { 7456 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog); 7457 7458 /* Kill the following clause, if you dislike this way. */ 7459 if (!syncookies && 7460 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) < 7461 (max_syn_backlog >> 2)) && 7462 !tcp_peer_is_proven(req, dst)) { 7463 /* Without syncookies last quarter of 7464 * backlog is filled with destinations, 7465 * proven to be alive. 7466 * It means that we continue to communicate 7467 * to destinations, already remembered 7468 * to the moment of synflood. 7469 */ 7470 pr_drop_req(req, ntohs(tcp_hdr(skb)->source), 7471 rsk_ops->family); 7472 goto drop_and_release; 7473 } 7474 7475 isn = af_ops->init_seq(skb); 7476 } 7477 7478 tcp_ecn_create_request(req, skb, sk, dst); 7479 7480 if (want_cookie) { 7481 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss); 7482 if (!tmp_opt.tstamp_ok) 7483 inet_rsk(req)->ecn_ok = 0; 7484 } 7485 7486 #ifdef CONFIG_TCP_AO 7487 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh)) 7488 goto drop_and_release; /* Invalid TCP options */ 7489 if (aoh) { 7490 tcp_rsk(req)->used_tcp_ao = true; 7491 tcp_rsk(req)->ao_rcv_next = aoh->keyid; 7492 tcp_rsk(req)->ao_keyid = aoh->rnext_keyid; 7493 7494 } else { 7495 tcp_rsk(req)->used_tcp_ao = false; 7496 } 7497 #endif 7498 tcp_rsk(req)->snt_isn = isn; 7499 tcp_rsk(req)->txhash = net_tx_rndhash(); 7500 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield; 7501 tcp_openreq_init_rwin(req, sk, dst); 7502 sk_rx_queue_set(req_to_sk(req), skb); 7503 if (!want_cookie) { 7504 tcp_reqsk_record_syn(sk, req, skb); 7505 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst); 7506 } 7507 if (fastopen_sk) { 7508 af_ops->send_synack(fastopen_sk, dst, &fl, req, 7509 &foc, TCP_SYNACK_FASTOPEN, skb); 7510 /* Add the child socket directly into the accept queue */ 7511 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) { 7512 reqsk_fastopen_remove(fastopen_sk, req, false); 7513 bh_unlock_sock(fastopen_sk); 7514 sock_put(fastopen_sk); 7515 goto drop_and_free; 7516 } 7517 sk->sk_data_ready(sk); 7518 bh_unlock_sock(fastopen_sk); 7519 sock_put(fastopen_sk); 7520 } else { 7521 tcp_rsk(req)->tfo_listener = false; 7522 if (!want_cookie) { 7523 req->timeout = tcp_timeout_init((struct sock *)req); 7524 if (unlikely(!inet_csk_reqsk_queue_hash_add(sk, req, 7525 req->timeout))) { 7526 reqsk_free(req); 7527 dst_release(dst); 7528 return 0; 7529 } 7530 7531 } 7532 af_ops->send_synack(sk, dst, &fl, req, &foc, 7533 !want_cookie ? TCP_SYNACK_NORMAL : 7534 TCP_SYNACK_COOKIE, 7535 skb); 7536 if (want_cookie) { 7537 reqsk_free(req); 7538 return 0; 7539 } 7540 } 7541 reqsk_put(req); 7542 return 0; 7543 7544 drop_and_release: 7545 dst_release(dst); 7546 drop_and_free: 7547 __reqsk_free(req); 7548 drop: 7549 tcp_listendrop(sk); 7550 return 0; 7551 } 7552 EXPORT_IPV6_MOD(tcp_conn_request); 7553