FILE
tcp_input.c
File preview with syntax highlighting.
linux-6.19.14/net/ipv4/tcp_input.c
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 87int 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) 122static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ); 123 124void 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} 130EXPORT_SYMBOL_GPL(clean_acked_data_enable); 131 132void 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} 137EXPORT_SYMBOL_GPL(clean_acked_data_disable); 138 139void clean_acked_data_flush(void) 140{ 141 static_key_deferred_flush(&clean_acked_data_enabled); 142} 143EXPORT_SYMBOL_GPL(clean_acked_data_flush); 144#endif 145 146#ifdef CONFIG_CGROUP_BPF 147static 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 182static 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 201static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 202{ 203} 204 205static void bpf_skops_established(struct sock *sk, int bpf_op, 206 struct sk_buff *skb) 207{ 208} 209#endif 210 211static __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 */ 227static 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 311static 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 323static 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 336static 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 344static 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 */ 389static 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 477static 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 */ 483static 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 */ 492static 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 563static 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 587static 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. */ 645static 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 */ 669static 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 682static 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 */ 722static 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. */ 760static 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 */ 788void 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} 799EXPORT_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 */ 812static 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 841static 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 854new_measure: 855 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 856 tp->rcv_rtt_est.time = tp->tcp_mstamp; 857} 858 859static 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 876static 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 894void tcp_rcvbuf_grow(struct sock *sk, u32 newval) 895{ 896 const struct net *net = sock_net(sk); 897 struct tcp_sock *tp = tcp_sk(sk); 898 u32 rcvwin, rcvbuf, cap, oldval; 899 u32 rtt_threshold, rtt_us; 900 u64 grow; 901 902 oldval = tp->rcvq_space.space; 903 tp->rcvq_space.space = newval; 904 905 if (!READ_ONCE(net->ipv4.sysctl_tcp_moderate_rcvbuf) || 906 (sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 907 return; 908 909 /* DRS is always one RTT late. */ 910 rcvwin = newval << 1; 911 912 rtt_us = tp->rcv_rtt_est.rtt_us >> 3; 913 rtt_threshold = READ_ONCE(net->ipv4.sysctl_tcp_rcvbuf_low_rtt); 914 if (rtt_us < rtt_threshold) { 915 /* For small RTT, we set @grow to rcvwin * rtt_us/rtt_threshold. 916 * It might take few additional ms to reach 'line rate', 917 * but will avoid sk_rcvbuf inflation and poor cache use. 918 */ 919 grow = div_u64((u64)rcvwin * rtt_us, rtt_threshold); 920 } else { 921 /* slow start: allow the sender to double its rate. */ 922 grow = div_u64(((u64)rcvwin << 1) * (newval - oldval), oldval); 923 } 924 rcvwin += grow; 925 926 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) 927 rcvwin += TCP_SKB_CB(tp->ooo_last_skb)->end_seq - tp->rcv_nxt; 928 929 cap = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]); 930 931 rcvbuf = min_t(u32, tcp_space_from_win(sk, rcvwin), cap); 932 if (rcvbuf > sk->sk_rcvbuf) { 933 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 934 /* Make the window clamp follow along. */ 935 WRITE_ONCE(tp->window_clamp, 936 tcp_win_from_space(sk, rcvbuf)); 937 } 938} 939/* 940 * This function should be called every time data is copied to user space. 941 * It calculates the appropriate TCP receive buffer space. 942 */ 943void tcp_rcv_space_adjust(struct sock *sk) 944{ 945 struct tcp_sock *tp = tcp_sk(sk); 946 int time, inq, copied; 947 948 trace_tcp_rcv_space_adjust(sk); 949 950 if (unlikely(!tp->rcv_rtt_est.rtt_us)) 951 return; 952 953 /* We do not refresh tp->tcp_mstamp here. 954 * Some platforms have expensive ktime_get() implementations. 955 * Using the last cached value is enough for DRS. 956 */ 957 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time); 958 if (time < (tp->rcv_rtt_est.rtt_us >> 3)) 959 return; 960 961 /* Number of bytes copied to user in last RTT */ 962 copied = tp->copied_seq - tp->rcvq_space.seq; 963 /* Number of bytes in receive queue. */ 964 inq = tp->rcv_nxt - tp->copied_seq; 965 copied -= inq; 966 if (copied <= tp->rcvq_space.space) 967 goto new_measure; 968 969 trace_tcp_rcvbuf_grow(sk, time); 970 971 tcp_rcvbuf_grow(sk, copied); 972 973new_measure: 974 tp->rcvq_space.seq = tp->copied_seq; 975 tp->rcvq_space.time = tp->tcp_mstamp; 976} 977 978static void tcp_save_lrcv_flowlabel(struct sock *sk, const struct sk_buff *skb) 979{ 980#if IS_ENABLED(CONFIG_IPV6) 981 struct inet_connection_sock *icsk = inet_csk(sk); 982 983 if (skb->protocol == htons(ETH_P_IPV6)) 984 icsk->icsk_ack.lrcv_flowlabel = ntohl(ip6_flowlabel(ipv6_hdr(skb))); 985#endif 986} 987 988/* There is something which you must keep in mind when you analyze the 989 * behavior of the tp->ato delayed ack timeout interval. When a 990 * connection starts up, we want to ack as quickly as possible. The 991 * problem is that "good" TCP's do slow start at the beginning of data 992 * transmission. The means that until we send the first few ACK's the 993 * sender will sit on his end and only queue most of his data, because 994 * he can only send snd_cwnd unacked packets at any given time. For 995 * each ACK we send, he increments snd_cwnd and transmits more of his 996 * queue. -DaveM 997 */ 998static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 999{ 1000 struct tcp_sock *tp = tcp_sk(sk); 1001 struct inet_connection_sock *icsk = inet_csk(sk); 1002 u32 now; 1003 1004 inet_csk_schedule_ack(sk); 1005 1006 tcp_measure_rcv_mss(sk, skb); 1007 1008 tcp_rcv_rtt_measure(tp); 1009 1010 now = tcp_jiffies32; 1011 1012 if (!icsk->icsk_ack.ato) { 1013 /* The _first_ data packet received, initialize 1014 * delayed ACK engine. 1015 */ 1016 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 1017 icsk->icsk_ack.ato = TCP_ATO_MIN; 1018 } else { 1019 int m = now - icsk->icsk_ack.lrcvtime; 1020 1021 if (m <= TCP_ATO_MIN / 2) { 1022 /* The fastest case is the first. */ 1023 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 1024 } else if (m < icsk->icsk_ack.ato) { 1025 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; 1026 if (icsk->icsk_ack.ato > icsk->icsk_rto) 1027 icsk->icsk_ack.ato = icsk->icsk_rto; 1028 } else if (m > icsk->icsk_rto) { 1029 /* Too long gap. Apparently sender failed to 1030 * restart window, so that we send ACKs quickly. 1031 */ 1032 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 1033 } 1034 } 1035 icsk->icsk_ack.lrcvtime = now; 1036 tcp_save_lrcv_flowlabel(sk, skb); 1037 1038 tcp_data_ecn_check(sk, skb); 1039 1040 if (skb->len >= 128) 1041 tcp_grow_window(sk, skb, true); 1042} 1043 1044/* Called to compute a smoothed rtt estimate. The data fed to this 1045 * routine either comes from timestamps, or from segments that were 1046 * known _not_ to have been retransmitted [see Karn/Partridge 1047 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 1048 * piece by Van Jacobson. 1049 * NOTE: the next three routines used to be one big routine. 1050 * To save cycles in the RFC 1323 implementation it was better to break 1051 * it up into three procedures. -- erics 1052 */ 1053static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) 1054{ 1055 struct tcp_sock *tp = tcp_sk(sk); 1056 long m = mrtt_us; /* RTT */ 1057 u32 srtt = tp->srtt_us; 1058 1059 /* The following amusing code comes from Jacobson's 1060 * article in SIGCOMM '88. Note that rtt and mdev 1061 * are scaled versions of rtt and mean deviation. 1062 * This is designed to be as fast as possible 1063 * m stands for "measurement". 1064 * 1065 * On a 1990 paper the rto value is changed to: 1066 * RTO = rtt + 4 * mdev 1067 * 1068 * Funny. This algorithm seems to be very broken. 1069 * These formulae increase RTO, when it should be decreased, increase 1070 * too slowly, when it should be increased quickly, decrease too quickly 1071 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 1072 * does not matter how to _calculate_ it. Seems, it was trap 1073 * that VJ failed to avoid. 8) 1074 */ 1075 if (srtt != 0) { 1076 m -= (srtt >> 3); /* m is now error in rtt est */ 1077 srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 1078 if (m < 0) { 1079 m = -m; /* m is now abs(error) */ 1080 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 1081 /* This is similar to one of Eifel findings. 1082 * Eifel blocks mdev updates when rtt decreases. 1083 * This solution is a bit different: we use finer gain 1084 * for mdev in this case (alpha*beta). 1085 * Like Eifel it also prevents growth of rto, 1086 * but also it limits too fast rto decreases, 1087 * happening in pure Eifel. 1088 */ 1089 if (m > 0) 1090 m >>= 3; 1091 } else { 1092 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 1093 } 1094 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ 1095 if (tp->mdev_us > tp->mdev_max_us) { 1096 tp->mdev_max_us = tp->mdev_us; 1097 if (tp->mdev_max_us > tp->rttvar_us) 1098 tp->rttvar_us = tp->mdev_max_us; 1099 } 1100 if (after(tp->snd_una, tp->rtt_seq)) { 1101 if (tp->mdev_max_us < tp->rttvar_us) 1102 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; 1103 tp->rtt_seq = tp->snd_nxt; 1104 tp->mdev_max_us = tcp_rto_min_us(sk); 1105 1106 tcp_bpf_rtt(sk, mrtt_us, srtt); 1107 } 1108 } else { 1109 /* no previous measure. */ 1110 srtt = m << 3; /* take the measured time to be rtt */ 1111 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ 1112 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); 1113 tp->mdev_max_us = tp->rttvar_us; 1114 tp->rtt_seq = tp->snd_nxt; 1115 1116 tcp_bpf_rtt(sk, mrtt_us, srtt); 1117 } 1118 tp->srtt_us = max(1U, srtt); 1119} 1120 1121void tcp_update_pacing_rate(struct sock *sk) 1122{ 1123 const struct tcp_sock *tp = tcp_sk(sk); 1124 u64 rate; 1125 1126 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ 1127 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); 1128 1129 /* current rate is (cwnd * mss) / srtt 1130 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. 1131 * In Congestion Avoidance phase, set it to 120 % the current rate. 1132 * 1133 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) 1134 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching 1135 * end of slow start and should slow down. 1136 */ 1137 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2) 1138 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio); 1139 else 1140 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio); 1141 1142 rate *= max(tcp_snd_cwnd(tp), tp->packets_out); 1143 1144 if (likely(tp->srtt_us)) 1145 do_div(rate, tp->srtt_us); 1146 1147 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate 1148 * without any lock. We want to make sure compiler wont store 1149 * intermediate values in this location. 1150 */ 1151 WRITE_ONCE(sk->sk_pacing_rate, 1152 min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate))); 1153} 1154 1155/* Calculate rto without backoff. This is the second half of Van Jacobson's 1156 * routine referred to above. 1157 */ 1158void tcp_set_rto(struct sock *sk) 1159{ 1160 const struct tcp_sock *tp = tcp_sk(sk); 1161 /* Old crap is replaced with new one. 8) 1162 * 1163 * More seriously: 1164 * 1. If rtt variance happened to be less 50msec, it is hallucination. 1165 * It cannot be less due to utterly erratic ACK generation made 1166 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 1167 * to do with delayed acks, because at cwnd>2 true delack timeout 1168 * is invisible. Actually, Linux-2.4 also generates erratic 1169 * ACKs in some circumstances. 1170 */ 1171 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 1172 1173 /* 2. Fixups made earlier cannot be right. 1174 * If we do not estimate RTO correctly without them, 1175 * all the algo is pure shit and should be replaced 1176 * with correct one. It is exactly, which we pretend to do. 1177 */ 1178 1179 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 1180 * guarantees that rto is higher. 1181 */ 1182 tcp_bound_rto(sk); 1183} 1184 1185__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 1186{ 1187 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 1188 1189 if (!cwnd) 1190 cwnd = TCP_INIT_CWND; 1191 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 1192} 1193 1194struct tcp_sacktag_state { 1195 /* Timestamps for earliest and latest never-retransmitted segment 1196 * that was SACKed. RTO needs the earliest RTT to stay conservative, 1197 * but congestion control should still get an accurate delay signal. 1198 */ 1199 u64 first_sackt; 1200 u64 last_sackt; 1201 u32 reord; 1202 u32 sack_delivered; 1203 u32 delivered_bytes; 1204 int flag; 1205 unsigned int mss_now; 1206 struct rate_sample *rate; 1207}; 1208 1209/* Take a notice that peer is sending D-SACKs. Skip update of data delivery 1210 * and spurious retransmission information if this DSACK is unlikely caused by 1211 * sender's action: 1212 * - DSACKed sequence range is larger than maximum receiver's window. 1213 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs. 1214 */ 1215static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq, 1216 u32 end_seq, struct tcp_sacktag_state *state) 1217{ 1218 u32 seq_len, dup_segs = 1; 1219 1220 if (!before(start_seq, end_seq)) 1221 return 0; 1222 1223 seq_len = end_seq - start_seq; 1224 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */ 1225 if (seq_len > tp->max_window) 1226 return 0; 1227 if (seq_len > tp->mss_cache) 1228 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache); 1229 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq) 1230 state->flag |= FLAG_DSACK_TLP; 1231 1232 tp->dsack_dups += dup_segs; 1233 /* Skip the DSACK if dup segs weren't retransmitted by sender */ 1234 if (tp->dsack_dups > tp->total_retrans) 1235 return 0; 1236 1237 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 1238 /* We increase the RACK ordering window in rounds where we receive 1239 * DSACKs that may have been due to reordering causing RACK to trigger 1240 * a spurious fast recovery. Thus RACK ignores DSACKs that happen 1241 * without having seen reordering, or that match TLP probes (TLP 1242 * is timer-driven, not triggered by RACK). 1243 */ 1244 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP)) 1245 tp->rack.dsack_seen = 1; 1246 1247 state->flag |= FLAG_DSACKING_ACK; 1248 /* A spurious retransmission is delivered */ 1249 state->sack_delivered += dup_segs; 1250 1251 return dup_segs; 1252} 1253 1254/* It's reordering when higher sequence was delivered (i.e. sacked) before 1255 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering 1256 * distance is approximated in full-mss packet distance ("reordering"). 1257 */ 1258static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq, 1259 const int ts) 1260{ 1261 struct tcp_sock *tp = tcp_sk(sk); 1262 const u32 mss = tp->mss_cache; 1263 u32 fack, metric; 1264 1265 fack = tcp_highest_sack_seq(tp); 1266 if (!before(low_seq, fack)) 1267 return; 1268 1269 metric = fack - low_seq; 1270 if ((metric > tp->reordering * mss) && mss) { 1271#if FASTRETRANS_DEBUG > 1 1272 pr_debug("Disorder%d %d %u f%u s%u rr%d\n", 1273 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 1274 tp->reordering, 1275 0, 1276 tp->sacked_out, 1277 tp->undo_marker ? tp->undo_retrans : 0); 1278#endif 1279 tp->reordering = min_t(u32, (metric + mss - 1) / mss, 1280 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)); 1281 } 1282 1283 /* This exciting event is worth to be remembered. 8) */ 1284 tp->reord_seen++; 1285 NET_INC_STATS(sock_net(sk), 1286 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER); 1287} 1288 1289 /* This must be called before lost_out or retrans_out are updated 1290 * on a new loss, because we want to know if all skbs previously 1291 * known to be lost have already been retransmitted, indicating 1292 * that this newly lost skb is our next skb to retransmit. 1293 */ 1294static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 1295{ 1296 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) || 1297 (tp->retransmit_skb_hint && 1298 before(TCP_SKB_CB(skb)->seq, 1299 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))) 1300 tp->retransmit_skb_hint = skb; 1301} 1302 1303/* Sum the number of packets on the wire we have marked as lost, and 1304 * notify the congestion control module that the given skb was marked lost. 1305 */ 1306static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb) 1307{ 1308 tp->lost += tcp_skb_pcount(skb); 1309} 1310 1311void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb) 1312{ 1313 __u8 sacked = TCP_SKB_CB(skb)->sacked; 1314 struct tcp_sock *tp = tcp_sk(sk); 1315 1316 if (sacked & TCPCB_SACKED_ACKED) 1317 return; 1318 1319 tcp_verify_retransmit_hint(tp, skb); 1320 if (sacked & TCPCB_LOST) { 1321 if (sacked & TCPCB_SACKED_RETRANS) { 1322 /* Account for retransmits that are lost again */ 1323 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1324 tp->retrans_out -= tcp_skb_pcount(skb); 1325 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT, 1326 tcp_skb_pcount(skb)); 1327 tcp_notify_skb_loss_event(tp, skb); 1328 } 1329 } else { 1330 tp->lost_out += tcp_skb_pcount(skb); 1331 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1332 tcp_notify_skb_loss_event(tp, skb); 1333 } 1334} 1335 1336/* This procedure tags the retransmission queue when SACKs arrive. 1337 * 1338 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1339 * Packets in queue with these bits set are counted in variables 1340 * sacked_out, retrans_out and lost_out, correspondingly. 1341 * 1342 * Valid combinations are: 1343 * Tag InFlight Description 1344 * 0 1 - orig segment is in flight. 1345 * S 0 - nothing flies, orig reached receiver. 1346 * L 0 - nothing flies, orig lost by net. 1347 * R 2 - both orig and retransmit are in flight. 1348 * L|R 1 - orig is lost, retransmit is in flight. 1349 * S|R 1 - orig reached receiver, retrans is still in flight. 1350 * (L|S|R is logically valid, it could occur when L|R is sacked, 1351 * but it is equivalent to plain S and code short-circuits it to S. 1352 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1353 * 1354 * These 6 states form finite state machine, controlled by the following events: 1355 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1356 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1357 * 3. Loss detection event of two flavors: 1358 * A. Scoreboard estimator decided the packet is lost. 1359 * A'. Reno "three dupacks" marks head of queue lost. 1360 * B. SACK arrives sacking SND.NXT at the moment, when the 1361 * segment was retransmitted. 1362 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1363 * 1364 * It is pleasant to note, that state diagram turns out to be commutative, 1365 * so that we are allowed not to be bothered by order of our actions, 1366 * when multiple events arrive simultaneously. (see the function below). 1367 * 1368 * Reordering detection. 1369 * -------------------- 1370 * Reordering metric is maximal distance, which a packet can be displaced 1371 * in packet stream. With SACKs we can estimate it: 1372 * 1373 * 1. SACK fills old hole and the corresponding segment was not 1374 * ever retransmitted -> reordering. Alas, we cannot use it 1375 * when segment was retransmitted. 1376 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1377 * for retransmitted and already SACKed segment -> reordering.. 1378 * Both of these heuristics are not used in Loss state, when we cannot 1379 * account for retransmits accurately. 1380 * 1381 * SACK block validation. 1382 * ---------------------- 1383 * 1384 * SACK block range validation checks that the received SACK block fits to 1385 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1386 * Note that SND.UNA is not included to the range though being valid because 1387 * it means that the receiver is rather inconsistent with itself reporting 1388 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1389 * perfectly valid, however, in light of RFC2018 which explicitly states 1390 * that "SACK block MUST reflect the newest segment. Even if the newest 1391 * segment is going to be discarded ...", not that it looks very clever 1392 * in case of head skb. Due to potentional receiver driven attacks, we 1393 * choose to avoid immediate execution of a walk in write queue due to 1394 * reneging and defer head skb's loss recovery to standard loss recovery 1395 * procedure that will eventually trigger (nothing forbids us doing this). 1396 * 1397 * Implements also blockage to start_seq wrap-around. Problem lies in the 1398 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1399 * there's no guarantee that it will be before snd_nxt (n). The problem 1400 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1401 * wrap (s_w): 1402 * 1403 * <- outs wnd -> <- wrapzone -> 1404 * u e n u_w e_w s n_w 1405 * | | | | | | | 1406 * |<------------+------+----- TCP seqno space --------------+---------->| 1407 * ...-- <2^31 ->| |<--------... 1408 * ...---- >2^31 ------>| |<--------... 1409 * 1410 * Current code wouldn't be vulnerable but it's better still to discard such 1411 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1412 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1413 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1414 * equal to the ideal case (infinite seqno space without wrap caused issues). 1415 * 1416 * With D-SACK the lower bound is extended to cover sequence space below 1417 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1418 * again, D-SACK block must not to go across snd_una (for the same reason as 1419 * for the normal SACK blocks, explained above). But there all simplicity 1420 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1421 * fully below undo_marker they do not affect behavior in anyway and can 1422 * therefore be safely ignored. In rare cases (which are more or less 1423 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1424 * fragmentation and packet reordering past skb's retransmission. To consider 1425 * them correctly, the acceptable range must be extended even more though 1426 * the exact amount is rather hard to quantify. However, tp->max_window can 1427 * be used as an exaggerated estimate. 1428 */ 1429static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, 1430 u32 start_seq, u32 end_seq) 1431{ 1432 /* Too far in future, or reversed (interpretation is ambiguous) */ 1433 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1434 return false; 1435 1436 /* Nasty start_seq wrap-around check (see comments above) */ 1437 if (!before(start_seq, tp->snd_nxt)) 1438 return false; 1439 1440 /* In outstanding window? ...This is valid exit for D-SACKs too. 1441 * start_seq == snd_una is non-sensical (see comments above) 1442 */ 1443 if (after(start_seq, tp->snd_una)) 1444 return true; 1445 1446 if (!is_dsack || !tp->undo_marker) 1447 return false; 1448 1449 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1450 if (after(end_seq, tp->snd_una)) 1451 return false; 1452 1453 if (!before(start_seq, tp->undo_marker)) 1454 return true; 1455 1456 /* Too old */ 1457 if (!after(end_seq, tp->undo_marker)) 1458 return false; 1459 1460 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1461 * start_seq < undo_marker and end_seq >= undo_marker. 1462 */ 1463 return !before(start_seq, end_seq - tp->max_window); 1464} 1465 1466static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1467 struct tcp_sack_block_wire *sp, int num_sacks, 1468 u32 prior_snd_una, struct tcp_sacktag_state *state) 1469{ 1470 struct tcp_sock *tp = tcp_sk(sk); 1471 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1472 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1473 u32 dup_segs; 1474 1475 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1476 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1477 } else if (num_sacks > 1) { 1478 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1479 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1480 1481 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1)) 1482 return false; 1483 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV); 1484 } else { 1485 return false; 1486 } 1487 1488 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state); 1489 if (!dup_segs) { /* Skip dubious DSACK */ 1490 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS); 1491 return false; 1492 } 1493 1494 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs); 1495 1496 /* D-SACK for already forgotten data... Do dumb counting. */ 1497 if (tp->undo_marker && tp->undo_retrans > 0 && 1498 !after(end_seq_0, prior_snd_una) && 1499 after(end_seq_0, tp->undo_marker)) 1500 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs); 1501 1502 return true; 1503} 1504 1505/* Check if skb is fully within the SACK block. In presence of GSO skbs, 1506 * the incoming SACK may not exactly match but we can find smaller MSS 1507 * aligned portion of it that matches. Therefore we might need to fragment 1508 * which may fail and creates some hassle (caller must handle error case 1509 * returns). 1510 * 1511 * FIXME: this could be merged to shift decision code 1512 */ 1513static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1514 u32 start_seq, u32 end_seq) 1515{ 1516 int err; 1517 bool in_sack; 1518 unsigned int pkt_len; 1519 unsigned int mss; 1520 1521 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1522 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1523 1524 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1525 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1526 mss = tcp_skb_mss(skb); 1527 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1528 1529 if (!in_sack) { 1530 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1531 if (pkt_len < mss) 1532 pkt_len = mss; 1533 } else { 1534 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1535 if (pkt_len < mss) 1536 return -EINVAL; 1537 } 1538 1539 /* Round if necessary so that SACKs cover only full MSSes 1540 * and/or the remaining small portion (if present) 1541 */ 1542 if (pkt_len > mss) { 1543 unsigned int new_len = (pkt_len / mss) * mss; 1544 if (!in_sack && new_len < pkt_len) 1545 new_len += mss; 1546 pkt_len = new_len; 1547 } 1548 1549 if (pkt_len >= skb->len && !in_sack) 1550 return 0; 1551 1552 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 1553 pkt_len, mss, GFP_ATOMIC); 1554 if (err < 0) 1555 return err; 1556 } 1557 1558 return in_sack; 1559} 1560 1561/* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1562static u8 tcp_sacktag_one(struct sock *sk, 1563 struct tcp_sacktag_state *state, u8 sacked, 1564 u32 start_seq, u32 end_seq, 1565 int dup_sack, int pcount, u32 plen, 1566 u64 xmit_time) 1567{ 1568 struct tcp_sock *tp = tcp_sk(sk); 1569 1570 /* Account D-SACK for retransmitted packet. */ 1571 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1572 if (tp->undo_marker && tp->undo_retrans > 0 && 1573 after(end_seq, tp->undo_marker)) 1574 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount); 1575 if ((sacked & TCPCB_SACKED_ACKED) && 1576 before(start_seq, state->reord)) 1577 state->reord = start_seq; 1578 } 1579 1580 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1581 if (!after(end_seq, tp->snd_una)) 1582 return sacked; 1583 1584 if (!(sacked & TCPCB_SACKED_ACKED)) { 1585 tcp_rack_advance(tp, sacked, end_seq, xmit_time); 1586 1587 if (sacked & TCPCB_SACKED_RETRANS) { 1588 /* If the segment is not tagged as lost, 1589 * we do not clear RETRANS, believing 1590 * that retransmission is still in flight. 1591 */ 1592 if (sacked & TCPCB_LOST) { 1593 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1594 tp->lost_out -= pcount; 1595 tp->retrans_out -= pcount; 1596 } 1597 } else { 1598 if (!(sacked & TCPCB_RETRANS)) { 1599 /* New sack for not retransmitted frame, 1600 * which was in hole. It is reordering. 1601 */ 1602 if (before(start_seq, 1603 tcp_highest_sack_seq(tp)) && 1604 before(start_seq, state->reord)) 1605 state->reord = start_seq; 1606 1607 if (!after(end_seq, tp->high_seq)) 1608 state->flag |= FLAG_ORIG_SACK_ACKED; 1609 if (state->first_sackt == 0) 1610 state->first_sackt = xmit_time; 1611 state->last_sackt = xmit_time; 1612 } 1613 1614 if (sacked & TCPCB_LOST) { 1615 sacked &= ~TCPCB_LOST; 1616 tp->lost_out -= pcount; 1617 } 1618 } 1619 1620 sacked |= TCPCB_SACKED_ACKED; 1621 state->flag |= FLAG_DATA_SACKED; 1622 tp->sacked_out += pcount; 1623 /* Out-of-order packets delivered */ 1624 state->sack_delivered += pcount; 1625 state->delivered_bytes += plen; 1626 } 1627 1628 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1629 * frames and clear it. undo_retrans is decreased above, L|R frames 1630 * are accounted above as well. 1631 */ 1632 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1633 sacked &= ~TCPCB_SACKED_RETRANS; 1634 tp->retrans_out -= pcount; 1635 } 1636 1637 return sacked; 1638} 1639 1640/* Shift newly-SACKed bytes from this skb to the immediately previous 1641 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1642 */ 1643static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, 1644 struct sk_buff *skb, 1645 struct tcp_sacktag_state *state, 1646 unsigned int pcount, int shifted, int mss, 1647 bool dup_sack) 1648{ 1649 struct tcp_sock *tp = tcp_sk(sk); 1650 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1651 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1652 1653 BUG_ON(!pcount); 1654 1655 /* Adjust counters and hints for the newly sacked sequence 1656 * range but discard the return value since prev is already 1657 * marked. We must tag the range first because the seq 1658 * advancement below implicitly advances 1659 * tcp_highest_sack_seq() when skb is highest_sack. 1660 */ 1661 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1662 start_seq, end_seq, dup_sack, pcount, skb->len, 1663 tcp_skb_timestamp_us(skb)); 1664 tcp_rate_skb_delivered(sk, skb, state->rate); 1665 1666 TCP_SKB_CB(prev)->end_seq += shifted; 1667 TCP_SKB_CB(skb)->seq += shifted; 1668 1669 tcp_skb_pcount_add(prev, pcount); 1670 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount); 1671 tcp_skb_pcount_add(skb, -pcount); 1672 1673 /* When we're adding to gso_segs == 1, gso_size will be zero, 1674 * in theory this shouldn't be necessary but as long as DSACK 1675 * code can come after this skb later on it's better to keep 1676 * setting gso_size to something. 1677 */ 1678 if (!TCP_SKB_CB(prev)->tcp_gso_size) 1679 TCP_SKB_CB(prev)->tcp_gso_size = mss; 1680 1681 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1682 if (tcp_skb_pcount(skb) <= 1) 1683 TCP_SKB_CB(skb)->tcp_gso_size = 0; 1684 1685 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1686 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1687 1688 if (skb->len > 0) { 1689 BUG_ON(!tcp_skb_pcount(skb)); 1690 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1691 return false; 1692 } 1693 1694 /* Whole SKB was eaten :-) */ 1695 1696 if (skb == tp->retransmit_skb_hint) 1697 tp->retransmit_skb_hint = prev; 1698 1699 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1700 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor; 1701 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1702 TCP_SKB_CB(prev)->end_seq++; 1703 1704 if (skb == tcp_highest_sack(sk)) 1705 tcp_advance_highest_sack(sk, skb); 1706 1707 tcp_skb_collapse_tstamp(prev, skb); 1708 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp)) 1709 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0; 1710 1711 tcp_rtx_queue_unlink_and_free(skb, sk); 1712 1713 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED); 1714 1715 return true; 1716} 1717 1718/* I wish gso_size would have a bit more sane initialization than 1719 * something-or-zero which complicates things 1720 */ 1721static int tcp_skb_seglen(const struct sk_buff *skb) 1722{ 1723 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1724} 1725 1726/* Shifting pages past head area doesn't work */ 1727static int skb_can_shift(const struct sk_buff *skb) 1728{ 1729 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1730} 1731 1732int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from, 1733 int pcount, int shiftlen) 1734{ 1735 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE) 1736 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need 1737 * to make sure not storing more than 65535 * 8 bytes per skb, 1738 * even if current MSS is bigger. 1739 */ 1740 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE)) 1741 return 0; 1742 if (unlikely(tcp_skb_pcount(to) + pcount > 65535)) 1743 return 0; 1744 return skb_shift(to, from, shiftlen); 1745} 1746 1747/* Try collapsing SACK blocks spanning across multiple skbs to a single 1748 * skb. 1749 */ 1750static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1751 struct tcp_sacktag_state *state, 1752 u32 start_seq, u32 end_seq, 1753 bool dup_sack) 1754{ 1755 struct tcp_sock *tp = tcp_sk(sk); 1756 struct sk_buff *prev; 1757 int mss; 1758 int pcount = 0; 1759 int len; 1760 int in_sack; 1761 1762 /* Normally R but no L won't result in plain S */ 1763 if (!dup_sack && 1764 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1765 goto fallback; 1766 if (!skb_can_shift(skb)) 1767 goto fallback; 1768 /* This frame is about to be dropped (was ACKed). */ 1769 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1770 goto fallback; 1771 1772 /* Can only happen with delayed DSACK + discard craziness */ 1773 prev = skb_rb_prev(skb); 1774 if (!prev) 1775 goto fallback; 1776 1777 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 1778 goto fallback; 1779 1780 if (!tcp_skb_can_collapse(prev, skb)) 1781 goto fallback; 1782 1783 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1784 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1785 1786 if (in_sack) { 1787 len = skb->len; 1788 pcount = tcp_skb_pcount(skb); 1789 mss = tcp_skb_seglen(skb); 1790 1791 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1792 * drop this restriction as unnecessary 1793 */ 1794 if (mss != tcp_skb_seglen(prev)) 1795 goto fallback; 1796 } else { 1797 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 1798 goto noop; 1799 /* CHECKME: This is non-MSS split case only?, this will 1800 * cause skipped skbs due to advancing loop btw, original 1801 * has that feature too 1802 */ 1803 if (tcp_skb_pcount(skb) <= 1) 1804 goto noop; 1805 1806 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1807 if (!in_sack) { 1808 /* TODO: head merge to next could be attempted here 1809 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 1810 * though it might not be worth of the additional hassle 1811 * 1812 * ...we can probably just fallback to what was done 1813 * previously. We could try merging non-SACKed ones 1814 * as well but it probably isn't going to buy off 1815 * because later SACKs might again split them, and 1816 * it would make skb timestamp tracking considerably 1817 * harder problem. 1818 */ 1819 goto fallback; 1820 } 1821 1822 len = end_seq - TCP_SKB_CB(skb)->seq; 1823 BUG_ON(len < 0); 1824 BUG_ON(len > skb->len); 1825 1826 /* MSS boundaries should be honoured or else pcount will 1827 * severely break even though it makes things bit trickier. 1828 * Optimize common case to avoid most of the divides 1829 */ 1830 mss = tcp_skb_mss(skb); 1831 1832 /* TODO: Fix DSACKs to not fragment already SACKed and we can 1833 * drop this restriction as unnecessary 1834 */ 1835 if (mss != tcp_skb_seglen(prev)) 1836 goto fallback; 1837 1838 if (len == mss) { 1839 pcount = 1; 1840 } else if (len < mss) { 1841 goto noop; 1842 } else { 1843 pcount = len / mss; 1844 len = pcount * mss; 1845 } 1846 } 1847 1848 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ 1849 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) 1850 goto fallback; 1851 1852 if (!tcp_skb_shift(prev, skb, pcount, len)) 1853 goto fallback; 1854 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack)) 1855 goto out; 1856 1857 /* Hole filled allows collapsing with the next as well, this is very 1858 * useful when hole on every nth skb pattern happens 1859 */ 1860 skb = skb_rb_next(prev); 1861 if (!skb) 1862 goto out; 1863 1864 if (!skb_can_shift(skb) || 1865 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 1866 (mss != tcp_skb_seglen(skb))) 1867 goto out; 1868 1869 if (!tcp_skb_can_collapse(prev, skb)) 1870 goto out; 1871 len = skb->len; 1872 pcount = tcp_skb_pcount(skb); 1873 if (tcp_skb_shift(prev, skb, pcount, len)) 1874 tcp_shifted_skb(sk, prev, skb, state, pcount, 1875 len, mss, 0); 1876 1877out: 1878 return prev; 1879 1880noop: 1881 return skb; 1882 1883fallback: 1884 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 1885 return NULL; 1886} 1887 1888static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 1889 struct tcp_sack_block *next_dup, 1890 struct tcp_sacktag_state *state, 1891 u32 start_seq, u32 end_seq, 1892 bool dup_sack_in) 1893{ 1894 struct tcp_sock *tp = tcp_sk(sk); 1895 struct sk_buff *tmp; 1896 1897 skb_rbtree_walk_from(skb) { 1898 int in_sack = 0; 1899 bool dup_sack = dup_sack_in; 1900 1901 /* queue is in-order => we can short-circuit the walk early */ 1902 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 1903 break; 1904 1905 if (next_dup && 1906 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 1907 in_sack = tcp_match_skb_to_sack(sk, skb, 1908 next_dup->start_seq, 1909 next_dup->end_seq); 1910 if (in_sack > 0) 1911 dup_sack = true; 1912 } 1913 1914 /* skb reference here is a bit tricky to get right, since 1915 * shifting can eat and free both this skb and the next, 1916 * so not even _safe variant of the loop is enough. 1917 */ 1918 if (in_sack <= 0) { 1919 tmp = tcp_shift_skb_data(sk, skb, state, 1920 start_seq, end_seq, dup_sack); 1921 if (tmp) { 1922 if (tmp != skb) { 1923 skb = tmp; 1924 continue; 1925 } 1926 1927 in_sack = 0; 1928 } else { 1929 in_sack = tcp_match_skb_to_sack(sk, skb, 1930 start_seq, 1931 end_seq); 1932 } 1933 } 1934 1935 if (unlikely(in_sack < 0)) 1936 break; 1937 1938 if (in_sack) { 1939 TCP_SKB_CB(skb)->sacked = 1940 tcp_sacktag_one(sk, 1941 state, 1942 TCP_SKB_CB(skb)->sacked, 1943 TCP_SKB_CB(skb)->seq, 1944 TCP_SKB_CB(skb)->end_seq, 1945 dup_sack, 1946 tcp_skb_pcount(skb), 1947 skb->len, 1948 tcp_skb_timestamp_us(skb)); 1949 tcp_rate_skb_delivered(sk, skb, state->rate); 1950 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1951 list_del_init(&skb->tcp_tsorted_anchor); 1952 1953 if (!before(TCP_SKB_CB(skb)->seq, 1954 tcp_highest_sack_seq(tp))) 1955 tcp_advance_highest_sack(sk, skb); 1956 } 1957 } 1958 return skb; 1959} 1960 1961static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq) 1962{ 1963 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node; 1964 struct sk_buff *skb; 1965 1966 while (*p) { 1967 parent = *p; 1968 skb = rb_to_skb(parent); 1969 if (before(seq, TCP_SKB_CB(skb)->seq)) { 1970 p = &parent->rb_left; 1971 continue; 1972 } 1973 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) { 1974 p = &parent->rb_right; 1975 continue; 1976 } 1977 return skb; 1978 } 1979 return NULL; 1980} 1981 1982static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 1983 u32 skip_to_seq) 1984{ 1985 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq)) 1986 return skb; 1987 1988 return tcp_sacktag_bsearch(sk, skip_to_seq); 1989} 1990 1991static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 1992 struct sock *sk, 1993 struct tcp_sack_block *next_dup, 1994 struct tcp_sacktag_state *state, 1995 u32 skip_to_seq) 1996{ 1997 if (!next_dup) 1998 return skb; 1999 2000 if (before(next_dup->start_seq, skip_to_seq)) { 2001 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq); 2002 skb = tcp_sacktag_walk(skb, sk, NULL, state, 2003 next_dup->start_seq, next_dup->end_seq, 2004 1); 2005 } 2006 2007 return skb; 2008} 2009 2010static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 2011{ 2012 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 2013} 2014 2015static int 2016tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 2017 u32 prior_snd_una, struct tcp_sacktag_state *state) 2018{ 2019 struct tcp_sock *tp = tcp_sk(sk); 2020 const unsigned char *ptr = (skb_transport_header(ack_skb) + 2021 TCP_SKB_CB(ack_skb)->sacked); 2022 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 2023 struct tcp_sack_block sp[TCP_NUM_SACKS]; 2024 struct tcp_sack_block *cache; 2025 struct sk_buff *skb; 2026 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 2027 int used_sacks; 2028 bool found_dup_sack = false; 2029 int i, j; 2030 int first_sack_index; 2031 2032 state->flag = 0; 2033 state->reord = tp->snd_nxt; 2034 2035 if (!tp->sacked_out) 2036 tcp_highest_sack_reset(sk); 2037 2038 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 2039 num_sacks, prior_snd_una, state); 2040 2041 /* Eliminate too old ACKs, but take into 2042 * account more or less fresh ones, they can 2043 * contain valid SACK info. 2044 */ 2045 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 2046 return 0; 2047 2048 if (!tp->packets_out) 2049 goto out; 2050 2051 used_sacks = 0; 2052 first_sack_index = 0; 2053 for (i = 0; i < num_sacks; i++) { 2054 bool dup_sack = !i && found_dup_sack; 2055 2056 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 2057 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 2058 2059 if (!tcp_is_sackblock_valid(tp, dup_sack, 2060 sp[used_sacks].start_seq, 2061 sp[used_sacks].end_seq)) { 2062 int mib_idx; 2063 2064 if (dup_sack) { 2065 if (!tp->undo_marker) 2066 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 2067 else 2068 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 2069 } else { 2070 /* Don't count olds caused by ACK reordering */ 2071 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 2072 !after(sp[used_sacks].end_seq, tp->snd_una)) 2073 continue; 2074 mib_idx = LINUX_MIB_TCPSACKDISCARD; 2075 } 2076 2077 NET_INC_STATS(sock_net(sk), mib_idx); 2078 if (i == 0) 2079 first_sack_index = -1; 2080 continue; 2081 } 2082 2083 /* Ignore very old stuff early */ 2084 if (!after(sp[used_sacks].end_seq, prior_snd_una)) { 2085 if (i == 0) 2086 first_sack_index = -1; 2087 continue; 2088 } 2089 2090 used_sacks++; 2091 } 2092 2093 /* order SACK blocks to allow in order walk of the retrans queue */ 2094 for (i = used_sacks - 1; i > 0; i--) { 2095 for (j = 0; j < i; j++) { 2096 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 2097 swap(sp[j], sp[j + 1]); 2098 2099 /* Track where the first SACK block goes to */ 2100 if (j == first_sack_index) 2101 first_sack_index = j + 1; 2102 } 2103 } 2104 } 2105 2106 state->mss_now = tcp_current_mss(sk); 2107 skb = NULL; 2108 i = 0; 2109 2110 if (!tp->sacked_out) { 2111 /* It's already past, so skip checking against it */ 2112 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 2113 } else { 2114 cache = tp->recv_sack_cache; 2115 /* Skip empty blocks in at head of the cache */ 2116 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 2117 !cache->end_seq) 2118 cache++; 2119 } 2120 2121 while (i < used_sacks) { 2122 u32 start_seq = sp[i].start_seq; 2123 u32 end_seq = sp[i].end_seq; 2124 bool dup_sack = (found_dup_sack && (i == first_sack_index)); 2125 struct tcp_sack_block *next_dup = NULL; 2126 2127 if (found_dup_sack && ((i + 1) == first_sack_index)) 2128 next_dup = &sp[i + 1]; 2129 2130 /* Skip too early cached blocks */ 2131 while (tcp_sack_cache_ok(tp, cache) && 2132 !before(start_seq, cache->end_seq)) 2133 cache++; 2134 2135 /* Can skip some work by looking recv_sack_cache? */ 2136 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 2137 after(end_seq, cache->start_seq)) { 2138 2139 /* Head todo? */ 2140 if (before(start_seq, cache->start_seq)) { 2141 skb = tcp_sacktag_skip(skb, sk, start_seq); 2142 skb = tcp_sacktag_walk(skb, sk, next_dup, 2143 state, 2144 start_seq, 2145 cache->start_seq, 2146 dup_sack); 2147 } 2148 2149 /* Rest of the block already fully processed? */ 2150 if (!after(end_seq, cache->end_seq)) 2151 goto advance_sp; 2152 2153 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 2154 state, 2155 cache->end_seq); 2156 2157 /* ...tail remains todo... */ 2158 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 2159 /* ...but better entrypoint exists! */ 2160 skb = tcp_highest_sack(sk); 2161 if (!skb) 2162 break; 2163 cache++; 2164 goto walk; 2165 } 2166 2167 skb = tcp_sacktag_skip(skb, sk, cache->end_seq); 2168 /* Check overlap against next cached too (past this one already) */ 2169 cache++; 2170 continue; 2171 } 2172 2173 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 2174 skb = tcp_highest_sack(sk); 2175 if (!skb) 2176 break; 2177 } 2178 skb = tcp_sacktag_skip(skb, sk, start_seq); 2179 2180walk: 2181 skb = tcp_sacktag_walk(skb, sk, next_dup, state, 2182 start_seq, end_seq, dup_sack); 2183 2184advance_sp: 2185 i++; 2186 } 2187 2188 /* Clear the head of the cache sack blocks so we can skip it next time */ 2189 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 2190 tp->recv_sack_cache[i].start_seq = 0; 2191 tp->recv_sack_cache[i].end_seq = 0; 2192 } 2193 for (j = 0; j < used_sacks; j++) 2194 tp->recv_sack_cache[i++] = sp[j]; 2195 2196 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker) 2197 tcp_check_sack_reordering(sk, state->reord, 0); 2198 2199 tcp_verify_left_out(tp); 2200out: 2201 2202#if FASTRETRANS_DEBUG > 0 2203 WARN_ON((int)tp->sacked_out < 0); 2204 WARN_ON((int)tp->lost_out < 0); 2205 WARN_ON((int)tp->retrans_out < 0); 2206 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 2207#endif 2208 return state->flag; 2209} 2210 2211/* Limits sacked_out so that sum with lost_out isn't ever larger than 2212 * packets_out. Returns false if sacked_out adjustement wasn't necessary. 2213 */ 2214static bool tcp_limit_reno_sacked(struct tcp_sock *tp) 2215{ 2216 u32 holes; 2217 2218 holes = max(tp->lost_out, 1U); 2219 holes = min(holes, tp->packets_out); 2220 2221 if ((tp->sacked_out + holes) > tp->packets_out) { 2222 tp->sacked_out = tp->packets_out - holes; 2223 return true; 2224 } 2225 return false; 2226} 2227 2228/* If we receive more dupacks than we expected counting segments 2229 * in assumption of absent reordering, interpret this as reordering. 2230 * The only another reason could be bug in receiver TCP. 2231 */ 2232static void tcp_check_reno_reordering(struct sock *sk, const int addend) 2233{ 2234 struct tcp_sock *tp = tcp_sk(sk); 2235 2236 if (!tcp_limit_reno_sacked(tp)) 2237 return; 2238 2239 tp->reordering = min_t(u32, tp->packets_out + addend, 2240 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering)); 2241 tp->reord_seen++; 2242 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER); 2243} 2244 2245/* Emulate SACKs for SACKless connection: account for a new dupack. */ 2246 2247static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack) 2248{ 2249 if (num_dupack) { 2250 struct tcp_sock *tp = tcp_sk(sk); 2251 u32 prior_sacked = tp->sacked_out; 2252 s32 delivered; 2253 2254 tp->sacked_out += num_dupack; 2255 tcp_check_reno_reordering(sk, 0); 2256 delivered = tp->sacked_out - prior_sacked; 2257 if (delivered > 0) 2258 tcp_count_delivered(tp, delivered, ece_ack); 2259 tcp_verify_left_out(tp); 2260 } 2261} 2262 2263/* Account for ACK, ACKing some data in Reno Recovery phase. */ 2264 2265static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack) 2266{ 2267 struct tcp_sock *tp = tcp_sk(sk); 2268 2269 if (acked > 0) { 2270 /* One ACK acked hole. The rest eat duplicate ACKs. */ 2271 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1), 2272 ece_ack); 2273 if (acked - 1 >= tp->sacked_out) 2274 tp->sacked_out = 0; 2275 else 2276 tp->sacked_out -= acked - 1; 2277 } 2278 tcp_check_reno_reordering(sk, acked); 2279 tcp_verify_left_out(tp); 2280} 2281 2282static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 2283{ 2284 tp->sacked_out = 0; 2285} 2286 2287void tcp_clear_retrans(struct tcp_sock *tp) 2288{ 2289 tp->retrans_out = 0; 2290 tp->lost_out = 0; 2291 tp->undo_marker = 0; 2292 tp->undo_retrans = -1; 2293 tp->sacked_out = 0; 2294 tp->rto_stamp = 0; 2295 tp->total_rto = 0; 2296 tp->total_rto_recoveries = 0; 2297 tp->total_rto_time = 0; 2298} 2299 2300static inline void tcp_init_undo(struct tcp_sock *tp) 2301{ 2302 tp->undo_marker = tp->snd_una; 2303 2304 /* Retransmission still in flight may cause DSACKs later. */ 2305 /* First, account for regular retransmits in flight: */ 2306 tp->undo_retrans = tp->retrans_out; 2307 /* Next, account for TLP retransmits in flight: */ 2308 if (tp->tlp_high_seq && tp->tlp_retrans) 2309 tp->undo_retrans++; 2310 /* Finally, avoid 0, because undo_retrans==0 means "can undo now": */ 2311 if (!tp->undo_retrans) 2312 tp->undo_retrans = -1; 2313} 2314 2315/* If we detect SACK reneging, forget all SACK information 2316 * and reset tags completely, otherwise preserve SACKs. If receiver 2317 * dropped its ofo queue, we will know this due to reneging detection. 2318 */ 2319static void tcp_timeout_mark_lost(struct sock *sk) 2320{ 2321 struct tcp_sock *tp = tcp_sk(sk); 2322 struct sk_buff *skb, *head; 2323 bool is_reneg; /* is receiver reneging on SACKs? */ 2324 2325 head = tcp_rtx_queue_head(sk); 2326 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED); 2327 if (is_reneg) { 2328 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 2329 tp->sacked_out = 0; 2330 /* Mark SACK reneging until we recover from this loss event. */ 2331 tp->is_sack_reneg = 1; 2332 } else if (tcp_is_reno(tp)) { 2333 tcp_reset_reno_sack(tp); 2334 } 2335 2336 skb = head; 2337 skb_rbtree_walk_from(skb) { 2338 if (is_reneg) 2339 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2340 else if (skb != head && tcp_rack_skb_timeout(tp, skb, 0) > 0) 2341 continue; /* Don't mark recently sent ones lost yet */ 2342 tcp_mark_skb_lost(sk, skb); 2343 } 2344 tcp_verify_left_out(tp); 2345 tcp_clear_all_retrans_hints(tp); 2346} 2347 2348/* Enter Loss state. */ 2349void tcp_enter_loss(struct sock *sk) 2350{ 2351 const struct inet_connection_sock *icsk = inet_csk(sk); 2352 struct tcp_sock *tp = tcp_sk(sk); 2353 struct net *net = sock_net(sk); 2354 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery; 2355 u8 reordering; 2356 2357 tcp_timeout_mark_lost(sk); 2358 2359 /* Reduce ssthresh if it has not yet been made inside this window. */ 2360 if (icsk->icsk_ca_state <= TCP_CA_Disorder || 2361 !after(tp->high_seq, tp->snd_una) || 2362 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2363 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2364 tp->prior_cwnd = tcp_snd_cwnd(tp); 2365 tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); 2366 tcp_ca_event(sk, CA_EVENT_LOSS); 2367 tcp_init_undo(tp); 2368 } 2369 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1); 2370 tp->snd_cwnd_cnt = 0; 2371 tp->snd_cwnd_stamp = tcp_jiffies32; 2372 2373 /* Timeout in disordered state after receiving substantial DUPACKs 2374 * suggests that the degree of reordering is over-estimated. 2375 */ 2376 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering); 2377 if (icsk->icsk_ca_state <= TCP_CA_Disorder && 2378 tp->sacked_out >= reordering) 2379 tp->reordering = min_t(unsigned int, tp->reordering, 2380 reordering); 2381 2382 tcp_set_ca_state(sk, TCP_CA_Loss); 2383 tp->high_seq = tp->snd_nxt; 2384 tp->tlp_high_seq = 0; 2385 tcp_ecn_queue_cwr(tp); 2386 2387 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous 2388 * loss recovery is underway except recurring timeout(s) on 2389 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing 2390 */ 2391 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) && 2392 (new_recovery || icsk->icsk_retransmits) && 2393 !inet_csk(sk)->icsk_mtup.probe_size; 2394} 2395 2396/* If ACK arrived pointing to a remembered SACK, it means that our 2397 * remembered SACKs do not reflect real state of receiver i.e. 2398 * receiver _host_ is heavily congested (or buggy). 2399 * 2400 * To avoid big spurious retransmission bursts due to transient SACK 2401 * scoreboard oddities that look like reneging, we give the receiver a 2402 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will 2403 * restore sanity to the SACK scoreboard. If the apparent reneging 2404 * persists until this RTO then we'll clear the SACK scoreboard. 2405 */ 2406static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag) 2407{ 2408 if (*ack_flag & FLAG_SACK_RENEGING && 2409 *ack_flag & FLAG_SND_UNA_ADVANCED) { 2410 struct tcp_sock *tp = tcp_sk(sk); 2411 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4), 2412 msecs_to_jiffies(10)); 2413 2414 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, delay, false); 2415 *ack_flag &= ~FLAG_SET_XMIT_TIMER; 2416 return true; 2417 } 2418 return false; 2419} 2420 2421/* Linux NewReno/SACK/ECN state machine. 2422 * -------------------------------------- 2423 * 2424 * "Open" Normal state, no dubious events, fast path. 2425 * "Disorder" In all the respects it is "Open", 2426 * but requires a bit more attention. It is entered when 2427 * we see some SACKs or dupacks. It is split of "Open" 2428 * mainly to move some processing from fast path to slow one. 2429 * "CWR" CWND was reduced due to some Congestion Notification event. 2430 * It can be ECN, ICMP source quench, local device congestion. 2431 * "Recovery" CWND was reduced, we are fast-retransmitting. 2432 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2433 * 2434 * tcp_fastretrans_alert() is entered: 2435 * - each incoming ACK, if state is not "Open" 2436 * - when arrived ACK is unusual, namely: 2437 * * SACK 2438 * * Duplicate ACK. 2439 * * ECN ECE. 2440 * 2441 * Counting packets in flight is pretty simple. 2442 * 2443 * in_flight = packets_out - left_out + retrans_out 2444 * 2445 * packets_out is SND.NXT-SND.UNA counted in packets. 2446 * 2447 * retrans_out is number of retransmitted segments. 2448 * 2449 * left_out is number of segments left network, but not ACKed yet. 2450 * 2451 * left_out = sacked_out + lost_out 2452 * 2453 * sacked_out: Packets, which arrived to receiver out of order 2454 * and hence not ACKed. With SACKs this number is simply 2455 * amount of SACKed data. Even without SACKs 2456 * it is easy to give pretty reliable estimate of this number, 2457 * counting duplicate ACKs. 2458 * 2459 * lost_out: Packets lost by network. TCP has no explicit 2460 * "loss notification" feedback from network (for now). 2461 * It means that this number can be only _guessed_. 2462 * Actually, it is the heuristics to predict lossage that 2463 * distinguishes different algorithms. 2464 * 2465 * F.e. after RTO, when all the queue is considered as lost, 2466 * lost_out = packets_out and in_flight = retrans_out. 2467 * 2468 * Essentially, we have now a few algorithms detecting 2469 * lost packets. 2470 * 2471 * If the receiver supports SACK: 2472 * 2473 * RACK (RFC8985): RACK is a newer loss detection algorithm 2474 * (2017-) that checks timing instead of counting DUPACKs. 2475 * Essentially a packet is considered lost if it's not S/ACKed 2476 * after RTT + reordering_window, where both metrics are 2477 * dynamically measured and adjusted. This is implemented in 2478 * tcp_rack_mark_lost. 2479 * 2480 * If the receiver does not support SACK: 2481 * 2482 * NewReno (RFC6582): in Recovery we assume that one segment 2483 * is lost (classic Reno). While we are in Recovery and 2484 * a partial ACK arrives, we assume that one more packet 2485 * is lost (NewReno). This heuristics are the same in NewReno 2486 * and SACK. 2487 * 2488 * The really tricky (and requiring careful tuning) part of the algorithm 2489 * is hidden in the RACK code in tcp_recovery.c and tcp_xmit_retransmit_queue(). 2490 * The first determines the moment _when_ we should reduce CWND and, 2491 * hence, slow down forward transmission. In fact, it determines the moment 2492 * when we decide that hole is caused by loss, rather than by a reorder. 2493 * 2494 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2495 * holes, caused by lost packets. 2496 * 2497 * And the most logically complicated part of algorithm is undo 2498 * heuristics. We detect false retransmits due to both too early 2499 * fast retransmit (reordering) and underestimated RTO, analyzing 2500 * timestamps and D-SACKs. When we detect that some segments were 2501 * retransmitted by mistake and CWND reduction was wrong, we undo 2502 * window reduction and abort recovery phase. This logic is hidden 2503 * inside several functions named tcp_try_undo_<something>. 2504 */ 2505 2506/* This function decides, when we should leave Disordered state 2507 * and enter Recovery phase, reducing congestion window. 2508 * 2509 * Main question: may we further continue forward transmission 2510 * with the same cwnd? 2511 */ 2512static bool tcp_time_to_recover(const struct tcp_sock *tp) 2513{ 2514 /* Has loss detection marked at least one packet lost? */ 2515 return tp->lost_out != 0; 2516} 2517 2518static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when) 2519{ 2520 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2521 before(tp->rx_opt.rcv_tsecr, when); 2522} 2523 2524/* skb is spurious retransmitted if the returned timestamp echo 2525 * reply is prior to the skb transmission time 2526 */ 2527static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp, 2528 const struct sk_buff *skb) 2529{ 2530 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) && 2531 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb)); 2532} 2533 2534/* Nothing was retransmitted or returned timestamp is less 2535 * than timestamp of the first retransmission. 2536 */ 2537static inline bool tcp_packet_delayed(const struct tcp_sock *tp) 2538{ 2539 const struct sock *sk = (const struct sock *)tp; 2540 2541 /* Received an echoed timestamp before the first retransmission? */ 2542 if (tp->retrans_stamp) 2543 return tcp_tsopt_ecr_before(tp, tp->retrans_stamp); 2544 2545 /* We set tp->retrans_stamp upon the first retransmission of a loss 2546 * recovery episode, so normally if tp->retrans_stamp is 0 then no 2547 * retransmission has happened yet (likely due to TSQ, which can cause 2548 * fast retransmits to be delayed). So if snd_una advanced while 2549 * (tp->retrans_stamp is 0 then apparently a packet was merely delayed, 2550 * not lost. But there are exceptions where we retransmit but then 2551 * clear tp->retrans_stamp, so we check for those exceptions. 2552 */ 2553 2554 /* (1) For non-SACK connections, tcp_is_non_sack_preventing_reopen() 2555 * clears tp->retrans_stamp when snd_una == high_seq. 2556 */ 2557 if (!tcp_is_sack(tp) && !before(tp->snd_una, tp->high_seq)) 2558 return false; 2559 2560 /* (2) In TCP_SYN_SENT tcp_clean_rtx_queue() clears tp->retrans_stamp 2561 * when setting FLAG_SYN_ACKED is set, even if the SYN was 2562 * retransmitted. 2563 */ 2564 if (sk->sk_state == TCP_SYN_SENT) 2565 return false; 2566 2567 return true; /* tp->retrans_stamp is zero; no retransmit yet */ 2568} 2569 2570/* Undo procedures. */ 2571 2572/* We can clear retrans_stamp when there are no retransmissions in the 2573 * window. It would seem that it is trivially available for us in 2574 * tp->retrans_out, however, that kind of assumptions doesn't consider 2575 * what will happen if errors occur when sending retransmission for the 2576 * second time. ...It could the that such segment has only 2577 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2578 * the head skb is enough except for some reneging corner cases that 2579 * are not worth the effort. 2580 * 2581 * Main reason for all this complexity is the fact that connection dying 2582 * time now depends on the validity of the retrans_stamp, in particular, 2583 * that successive retransmissions of a segment must not advance 2584 * retrans_stamp under any conditions. 2585 */ 2586static bool tcp_any_retrans_done(const struct sock *sk) 2587{ 2588 const struct tcp_sock *tp = tcp_sk(sk); 2589 struct sk_buff *skb; 2590 2591 if (tp->retrans_out) 2592 return true; 2593 2594 skb = tcp_rtx_queue_head(sk); 2595 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2596 return true; 2597 2598 return false; 2599} 2600 2601/* If loss recovery is finished and there are no retransmits out in the 2602 * network, then we clear retrans_stamp so that upon the next loss recovery 2603 * retransmits_timed_out() and timestamp-undo are using the correct value. 2604 */ 2605static void tcp_retrans_stamp_cleanup(struct sock *sk) 2606{ 2607 if (!tcp_any_retrans_done(sk)) 2608 tcp_sk(sk)->retrans_stamp = 0; 2609} 2610 2611static void DBGUNDO(struct sock *sk, const char *msg) 2612{ 2613#if FASTRETRANS_DEBUG > 1 2614 struct tcp_sock *tp = tcp_sk(sk); 2615 struct inet_sock *inet = inet_sk(sk); 2616 2617 if (sk->sk_family == AF_INET) { 2618 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2619 msg, 2620 &inet->inet_daddr, ntohs(inet->inet_dport), 2621 tcp_snd_cwnd(tp), tcp_left_out(tp), 2622 tp->snd_ssthresh, tp->prior_ssthresh, 2623 tp->packets_out); 2624 } 2625#if IS_ENABLED(CONFIG_IPV6) 2626 else if (sk->sk_family == AF_INET6) { 2627 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2628 msg, 2629 &sk->sk_v6_daddr, ntohs(inet->inet_dport), 2630 tcp_snd_cwnd(tp), tcp_left_out(tp), 2631 tp->snd_ssthresh, tp->prior_ssthresh, 2632 tp->packets_out); 2633 } 2634#endif 2635#endif 2636} 2637 2638static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss) 2639{ 2640 struct tcp_sock *tp = tcp_sk(sk); 2641 2642 if (unmark_loss) { 2643 struct sk_buff *skb; 2644 2645 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2646 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2647 } 2648 tp->lost_out = 0; 2649 tcp_clear_all_retrans_hints(tp); 2650 } 2651 2652 if (tp->prior_ssthresh) { 2653 const struct inet_connection_sock *icsk = inet_csk(sk); 2654 2655 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk)); 2656 2657 if (tp->prior_ssthresh > tp->snd_ssthresh) { 2658 tp->snd_ssthresh = tp->prior_ssthresh; 2659 tcp_ecn_withdraw_cwr(tp); 2660 } 2661 } 2662 tp->snd_cwnd_stamp = tcp_jiffies32; 2663 tp->undo_marker = 0; 2664 tp->rack.advanced = 1; /* Force RACK to re-exam losses */ 2665} 2666 2667static inline bool tcp_may_undo(const struct tcp_sock *tp) 2668{ 2669 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2670} 2671 2672static bool tcp_is_non_sack_preventing_reopen(struct sock *sk) 2673{ 2674 struct tcp_sock *tp = tcp_sk(sk); 2675 2676 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2677 /* Hold old state until something *above* high_seq 2678 * is ACKed. For Reno it is MUST to prevent false 2679 * fast retransmits (RFC2582). SACK TCP is safe. */ 2680 if (!tcp_any_retrans_done(sk)) 2681 tp->retrans_stamp = 0; 2682 return true; 2683 } 2684 return false; 2685} 2686 2687/* People celebrate: "We love our President!" */ 2688static bool tcp_try_undo_recovery(struct sock *sk) 2689{ 2690 struct tcp_sock *tp = tcp_sk(sk); 2691 2692 if (tcp_may_undo(tp)) { 2693 int mib_idx; 2694 2695 /* Happy end! We did not retransmit anything 2696 * or our original transmission succeeded. 2697 */ 2698 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2699 tcp_undo_cwnd_reduction(sk, false); 2700 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2701 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2702 else 2703 mib_idx = LINUX_MIB_TCPFULLUNDO; 2704 2705 NET_INC_STATS(sock_net(sk), mib_idx); 2706 } else if (tp->rack.reo_wnd_persist) { 2707 tp->rack.reo_wnd_persist--; 2708 } 2709 if (tcp_is_non_sack_preventing_reopen(sk)) 2710 return true; 2711 tcp_set_ca_state(sk, TCP_CA_Open); 2712 tp->is_sack_reneg = 0; 2713 return false; 2714} 2715 2716/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2717static bool tcp_try_undo_dsack(struct sock *sk) 2718{ 2719 struct tcp_sock *tp = tcp_sk(sk); 2720 2721 if (tp->undo_marker && !tp->undo_retrans) { 2722 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH, 2723 tp->rack.reo_wnd_persist + 1); 2724 DBGUNDO(sk, "D-SACK"); 2725 tcp_undo_cwnd_reduction(sk, false); 2726 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2727 return true; 2728 } 2729 return false; 2730} 2731 2732/* Undo during loss recovery after partial ACK or using F-RTO. */ 2733static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo) 2734{ 2735 struct tcp_sock *tp = tcp_sk(sk); 2736 2737 if (frto_undo || tcp_may_undo(tp)) { 2738 tcp_undo_cwnd_reduction(sk, true); 2739 2740 DBGUNDO(sk, "partial loss"); 2741 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2742 if (frto_undo) 2743 NET_INC_STATS(sock_net(sk), 2744 LINUX_MIB_TCPSPURIOUSRTOS); 2745 WRITE_ONCE(inet_csk(sk)->icsk_retransmits, 0); 2746 if (tcp_is_non_sack_preventing_reopen(sk)) 2747 return true; 2748 if (frto_undo || tcp_is_sack(tp)) { 2749 tcp_set_ca_state(sk, TCP_CA_Open); 2750 tp->is_sack_reneg = 0; 2751 } 2752 return true; 2753 } 2754 return false; 2755} 2756 2757/* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937. 2758 * It computes the number of packets to send (sndcnt) based on packets newly 2759 * delivered: 2760 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 2761 * cwnd reductions across a full RTT. 2762 * 2) Otherwise PRR uses packet conservation to send as much as delivered. 2763 * But when SND_UNA is acked without further losses, 2764 * slow starts cwnd up to ssthresh to speed up the recovery. 2765 */ 2766static void tcp_init_cwnd_reduction(struct sock *sk) 2767{ 2768 struct tcp_sock *tp = tcp_sk(sk); 2769 2770 tp->high_seq = tp->snd_nxt; 2771 tp->tlp_high_seq = 0; 2772 tp->snd_cwnd_cnt = 0; 2773 tp->prior_cwnd = tcp_snd_cwnd(tp); 2774 tp->prr_delivered = 0; 2775 tp->prr_out = 0; 2776 tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); 2777 tcp_ecn_queue_cwr(tp); 2778} 2779 2780void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag) 2781{ 2782 struct tcp_sock *tp = tcp_sk(sk); 2783 int sndcnt = 0; 2784 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 2785 2786 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd)) 2787 return; 2788 2789 trace_tcp_cwnd_reduction_tp(sk, newly_acked_sacked, newly_lost, flag); 2790 2791 tp->prr_delivered += newly_acked_sacked; 2792 if (delta < 0) { 2793 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 2794 tp->prior_cwnd - 1; 2795 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 2796 } else { 2797 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out, 2798 newly_acked_sacked); 2799 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost) 2800 sndcnt++; 2801 sndcnt = min(delta, sndcnt); 2802 } 2803 /* Force a fast retransmit upon entering fast recovery */ 2804 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1)); 2805 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt); 2806} 2807 2808static inline void tcp_end_cwnd_reduction(struct sock *sk) 2809{ 2810 struct tcp_sock *tp = tcp_sk(sk); 2811 2812 if (inet_csk(sk)->icsk_ca_ops->cong_control) 2813 return; 2814 2815 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ 2816 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH && 2817 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) { 2818 tcp_snd_cwnd_set(tp, tp->snd_ssthresh); 2819 tp->snd_cwnd_stamp = tcp_jiffies32; 2820 } 2821 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 2822} 2823 2824/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */ 2825void tcp_enter_cwr(struct sock *sk) 2826{ 2827 struct tcp_sock *tp = tcp_sk(sk); 2828 2829 tp->prior_ssthresh = 0; 2830 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 2831 tp->undo_marker = 0; 2832 tcp_init_cwnd_reduction(sk); 2833 tcp_set_ca_state(sk, TCP_CA_CWR); 2834 } 2835} 2836EXPORT_SYMBOL(tcp_enter_cwr); 2837 2838static void tcp_try_keep_open(struct sock *sk) 2839{ 2840 struct tcp_sock *tp = tcp_sk(sk); 2841 int state = TCP_CA_Open; 2842 2843 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 2844 state = TCP_CA_Disorder; 2845 2846 if (inet_csk(sk)->icsk_ca_state != state) { 2847 tcp_set_ca_state(sk, state); 2848 tp->high_seq = tp->snd_nxt; 2849 } 2850} 2851 2852static void tcp_try_to_open(struct sock *sk, int flag) 2853{ 2854 struct tcp_sock *tp = tcp_sk(sk); 2855 2856 tcp_verify_left_out(tp); 2857 2858 if (!tcp_any_retrans_done(sk)) 2859 tp->retrans_stamp = 0; 2860 2861 if (flag & FLAG_ECE) 2862 tcp_enter_cwr(sk); 2863 2864 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 2865 tcp_try_keep_open(sk); 2866 } 2867} 2868 2869static void tcp_mtup_probe_failed(struct sock *sk) 2870{ 2871 struct inet_connection_sock *icsk = inet_csk(sk); 2872 2873 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 2874 icsk->icsk_mtup.probe_size = 0; 2875 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL); 2876} 2877 2878static void tcp_mtup_probe_success(struct sock *sk) 2879{ 2880 struct tcp_sock *tp = tcp_sk(sk); 2881 struct inet_connection_sock *icsk = inet_csk(sk); 2882 u64 val; 2883 2884 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2885 2886 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache); 2887 do_div(val, icsk->icsk_mtup.probe_size); 2888 DEBUG_NET_WARN_ON_ONCE((u32)val != val); 2889 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val)); 2890 2891 tp->snd_cwnd_cnt = 0; 2892 tp->snd_cwnd_stamp = tcp_jiffies32; 2893 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2894 2895 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 2896 icsk->icsk_mtup.probe_size = 0; 2897 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 2898 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS); 2899} 2900 2901/* Sometimes we deduce that packets have been dropped due to reasons other than 2902 * congestion, like path MTU reductions or failed client TFO attempts. In these 2903 * cases we call this function to retransmit as many packets as cwnd allows, 2904 * without reducing cwnd. Given that retransmits will set retrans_stamp to a 2905 * non-zero value (and may do so in a later calling context due to TSQ), we 2906 * also enter CA_Loss so that we track when all retransmitted packets are ACKed 2907 * and clear retrans_stamp when that happens (to ensure later recurring RTOs 2908 * are using the correct retrans_stamp and don't declare ETIMEDOUT 2909 * prematurely). 2910 */ 2911static void tcp_non_congestion_loss_retransmit(struct sock *sk) 2912{ 2913 const struct inet_connection_sock *icsk = inet_csk(sk); 2914 struct tcp_sock *tp = tcp_sk(sk); 2915 2916 if (icsk->icsk_ca_state != TCP_CA_Loss) { 2917 tp->high_seq = tp->snd_nxt; 2918 tp->snd_ssthresh = tcp_current_ssthresh(sk); 2919 tp->prior_ssthresh = 0; 2920 tp->undo_marker = 0; 2921 tcp_set_ca_state(sk, TCP_CA_Loss); 2922 } 2923 tcp_xmit_retransmit_queue(sk); 2924} 2925 2926/* Do a simple retransmit without using the backoff mechanisms in 2927 * tcp_timer. This is used for path mtu discovery. 2928 * The socket is already locked here. 2929 */ 2930void tcp_simple_retransmit(struct sock *sk) 2931{ 2932 struct tcp_sock *tp = tcp_sk(sk); 2933 struct sk_buff *skb; 2934 int mss; 2935 2936 /* A fastopen SYN request is stored as two separate packets within 2937 * the retransmit queue, this is done by tcp_send_syn_data(). 2938 * As a result simply checking the MSS of the frames in the queue 2939 * will not work for the SYN packet. 2940 * 2941 * Us being here is an indication of a path MTU issue so we can 2942 * assume that the fastopen SYN was lost and just mark all the 2943 * frames in the retransmit queue as lost. We will use an MSS of 2944 * -1 to mark all frames as lost, otherwise compute the current MSS. 2945 */ 2946 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT) 2947 mss = -1; 2948 else 2949 mss = tcp_current_mss(sk); 2950 2951 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2952 if (tcp_skb_seglen(skb) > mss) 2953 tcp_mark_skb_lost(sk, skb); 2954 } 2955 2956 if (!tp->lost_out) 2957 return; 2958 2959 if (tcp_is_reno(tp)) 2960 tcp_limit_reno_sacked(tp); 2961 2962 tcp_verify_left_out(tp); 2963 2964 /* Don't muck with the congestion window here. 2965 * Reason is that we do not increase amount of _data_ 2966 * in network, but units changed and effective 2967 * cwnd/ssthresh really reduced now. 2968 */ 2969 tcp_non_congestion_loss_retransmit(sk); 2970} 2971EXPORT_IPV6_MOD(tcp_simple_retransmit); 2972 2973void tcp_enter_recovery(struct sock *sk, bool ece_ack) 2974{ 2975 struct tcp_sock *tp = tcp_sk(sk); 2976 int mib_idx; 2977 2978 /* Start the clock with our fast retransmit, for undo and ETIMEDOUT. */ 2979 tcp_retrans_stamp_cleanup(sk); 2980 2981 if (tcp_is_reno(tp)) 2982 mib_idx = LINUX_MIB_TCPRENORECOVERY; 2983 else 2984 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 2985 2986 NET_INC_STATS(sock_net(sk), mib_idx); 2987 2988 tp->prior_ssthresh = 0; 2989 tcp_init_undo(tp); 2990 2991 if (!tcp_in_cwnd_reduction(sk)) { 2992 if (!ece_ack) 2993 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2994 tcp_init_cwnd_reduction(sk); 2995 } 2996 tcp_set_ca_state(sk, TCP_CA_Recovery); 2997} 2998 2999static void tcp_update_rto_time(struct tcp_sock *tp) 3000{ 3001 if (tp->rto_stamp) { 3002 tp->total_rto_time += tcp_time_stamp_ms(tp) - tp->rto_stamp; 3003 tp->rto_stamp = 0; 3004 } 3005} 3006 3007/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are 3008 * recovered or spurious. Otherwise retransmits more on partial ACKs. 3009 */ 3010static void tcp_process_loss(struct sock *sk, int flag, int num_dupack, 3011 int *rexmit) 3012{ 3013 struct tcp_sock *tp = tcp_sk(sk); 3014 bool recovered = !before(tp->snd_una, tp->high_seq); 3015 3016 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) && 3017 tcp_try_undo_loss(sk, false)) 3018 return; 3019 3020 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */ 3021 /* Step 3.b. A timeout is spurious if not all data are 3022 * lost, i.e., never-retransmitted data are (s)acked. 3023 */ 3024 if ((flag & FLAG_ORIG_SACK_ACKED) && 3025 tcp_try_undo_loss(sk, true)) 3026 return; 3027 3028 if (after(tp->snd_nxt, tp->high_seq)) { 3029 if (flag & FLAG_DATA_SACKED || num_dupack) 3030 tp->frto = 0; /* Step 3.a. loss was real */ 3031 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) { 3032 tp->high_seq = tp->snd_nxt; 3033 /* Step 2.b. Try send new data (but deferred until cwnd 3034 * is updated in tcp_ack()). Otherwise fall back to 3035 * the conventional recovery. 3036 */ 3037 if (!tcp_write_queue_empty(sk) && 3038 after(tcp_wnd_end(tp), tp->snd_nxt)) { 3039 *rexmit = REXMIT_NEW; 3040 return; 3041 } 3042 tp->frto = 0; 3043 } 3044 } 3045 3046 if (recovered) { 3047 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */ 3048 tcp_try_undo_recovery(sk); 3049 return; 3050 } 3051 if (tcp_is_reno(tp)) { 3052 /* A Reno DUPACK means new data in F-RTO step 2.b above are 3053 * delivered. Lower inflight to clock out (re)transmissions. 3054 */ 3055 if (after(tp->snd_nxt, tp->high_seq) && num_dupack) 3056 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE); 3057 else if (flag & FLAG_SND_UNA_ADVANCED) 3058 tcp_reset_reno_sack(tp); 3059 } 3060 *rexmit = REXMIT_LOST; 3061} 3062 3063/* Undo during fast recovery after partial ACK. */ 3064static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una) 3065{ 3066 struct tcp_sock *tp = tcp_sk(sk); 3067 3068 if (tp->undo_marker && tcp_packet_delayed(tp)) { 3069 /* Plain luck! Hole if filled with delayed 3070 * packet, rather than with a retransmit. Check reordering. 3071 */ 3072 tcp_check_sack_reordering(sk, prior_snd_una, 1); 3073 3074 /* We are getting evidence that the reordering degree is higher 3075 * than we realized. If there are no retransmits out then we 3076 * can undo. Otherwise we clock out new packets but do not 3077 * mark more packets lost or retransmit more. 3078 */ 3079 if (tp->retrans_out) 3080 return true; 3081 3082 if (!tcp_any_retrans_done(sk)) 3083 tp->retrans_stamp = 0; 3084 3085 DBGUNDO(sk, "partial recovery"); 3086 tcp_undo_cwnd_reduction(sk, true); 3087 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 3088 tcp_try_keep_open(sk); 3089 } 3090 return false; 3091} 3092 3093static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag) 3094{ 3095 struct tcp_sock *tp = tcp_sk(sk); 3096 3097 if (tcp_rtx_queue_empty(sk)) 3098 return; 3099 3100 if (unlikely(tcp_is_reno(tp))) { 3101 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED); 3102 } else { 3103 u32 prior_retrans = tp->retrans_out; 3104 3105 if (tcp_rack_mark_lost(sk)) 3106 *ack_flag &= ~FLAG_SET_XMIT_TIMER; 3107 if (prior_retrans > tp->retrans_out) 3108 *ack_flag |= FLAG_LOST_RETRANS; 3109 } 3110} 3111 3112/* Process an event, which can update packets-in-flight not trivially. 3113 * Main goal of this function is to calculate new estimate for left_out, 3114 * taking into account both packets sitting in receiver's buffer and 3115 * packets lost by network. 3116 * 3117 * Besides that it updates the congestion state when packet loss or ECN 3118 * is detected. But it does not reduce the cwnd, it is done by the 3119 * congestion control later. 3120 * 3121 * It does _not_ decide what to send, it is made in function 3122 * tcp_xmit_retransmit_queue(). 3123 */ 3124static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una, 3125 int num_dupack, int *ack_flag, int *rexmit) 3126{ 3127 struct inet_connection_sock *icsk = inet_csk(sk); 3128 struct tcp_sock *tp = tcp_sk(sk); 3129 int flag = *ack_flag; 3130 bool ece_ack = flag & FLAG_ECE; 3131 3132 if (!tp->packets_out && tp->sacked_out) 3133 tp->sacked_out = 0; 3134 3135 /* Now state machine starts. 3136 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 3137 if (ece_ack) 3138 tp->prior_ssthresh = 0; 3139 3140 /* B. In all the states check for reneging SACKs. */ 3141 if (tcp_check_sack_reneging(sk, ack_flag)) 3142 return; 3143 3144 /* C. Check consistency of the current state. */ 3145 tcp_verify_left_out(tp); 3146 3147 /* D. Check state exit conditions. State can be terminated 3148 * when high_seq is ACKed. */ 3149 if (icsk->icsk_ca_state == TCP_CA_Open) { 3150 WARN_ON(tp->retrans_out != 0 && !tp->syn_data); 3151 tp->retrans_stamp = 0; 3152 } else if (!before(tp->snd_una, tp->high_seq)) { 3153 switch (icsk->icsk_ca_state) { 3154 case TCP_CA_CWR: 3155 /* CWR is to be held something *above* high_seq 3156 * is ACKed for CWR bit to reach receiver. */ 3157 if (tp->snd_una != tp->high_seq) { 3158 tcp_end_cwnd_reduction(sk); 3159 tcp_set_ca_state(sk, TCP_CA_Open); 3160 } 3161 break; 3162 3163 case TCP_CA_Recovery: 3164 if (tcp_is_reno(tp)) 3165 tcp_reset_reno_sack(tp); 3166 if (tcp_try_undo_recovery(sk)) 3167 return; 3168 tcp_end_cwnd_reduction(sk); 3169 break; 3170 } 3171 } 3172 3173 /* E. Process state. */ 3174 switch (icsk->icsk_ca_state) { 3175 case TCP_CA_Recovery: 3176 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 3177 if (tcp_is_reno(tp)) 3178 tcp_add_reno_sack(sk, num_dupack, ece_ack); 3179 } else if (tcp_try_undo_partial(sk, prior_snd_una)) 3180 return; 3181 3182 if (tcp_try_undo_dsack(sk)) 3183 tcp_try_to_open(sk, flag); 3184 3185 tcp_identify_packet_loss(sk, ack_flag); 3186 if (icsk->icsk_ca_state != TCP_CA_Recovery) { 3187 if (!tcp_time_to_recover(tp)) 3188 return; 3189 /* Undo reverts the recovery state. If loss is evident, 3190 * starts a new recovery (e.g. reordering then loss); 3191 */ 3192 tcp_enter_recovery(sk, ece_ack); 3193 } 3194 break; 3195 case TCP_CA_Loss: 3196 tcp_process_loss(sk, flag, num_dupack, rexmit); 3197 if (icsk->icsk_ca_state != TCP_CA_Loss) 3198 tcp_update_rto_time(tp); 3199 tcp_identify_packet_loss(sk, ack_flag); 3200 if (!(icsk->icsk_ca_state == TCP_CA_Open || 3201 (*ack_flag & FLAG_LOST_RETRANS))) 3202 return; 3203 /* Change state if cwnd is undone or retransmits are lost */ 3204 fallthrough; 3205 default: 3206 if (tcp_is_reno(tp)) { 3207 if (flag & FLAG_SND_UNA_ADVANCED) 3208 tcp_reset_reno_sack(tp); 3209 tcp_add_reno_sack(sk, num_dupack, ece_ack); 3210 } 3211 3212 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 3213 tcp_try_undo_dsack(sk); 3214 3215 tcp_identify_packet_loss(sk, ack_flag); 3216 if (!tcp_time_to_recover(tp)) { 3217 tcp_try_to_open(sk, flag); 3218 return; 3219 } 3220 3221 /* MTU probe failure: don't reduce cwnd */ 3222 if (icsk->icsk_ca_state < TCP_CA_CWR && 3223 icsk->icsk_mtup.probe_size && 3224 tp->snd_una == tp->mtu_probe.probe_seq_start) { 3225 tcp_mtup_probe_failed(sk); 3226 /* Restores the reduction we did in tcp_mtup_probe() */ 3227 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1); 3228 tcp_simple_retransmit(sk); 3229 return; 3230 } 3231 3232 /* Otherwise enter Recovery state */ 3233 tcp_enter_recovery(sk, ece_ack); 3234 } 3235 3236 *rexmit = REXMIT_LOST; 3237} 3238 3239static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag) 3240{ 3241 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ; 3242 struct tcp_sock *tp = tcp_sk(sk); 3243 3244 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) { 3245 /* If the remote keeps returning delayed ACKs, eventually 3246 * the min filter would pick it up and overestimate the 3247 * prop. delay when it expires. Skip suspected delayed ACKs. 3248 */ 3249 return; 3250 } 3251 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32, 3252 rtt_us ? : jiffies_to_usecs(1)); 3253} 3254 3255static bool tcp_ack_update_rtt(struct sock *sk, const int flag, 3256 long seq_rtt_us, long sack_rtt_us, 3257 long ca_rtt_us, struct rate_sample *rs) 3258{ 3259 const struct tcp_sock *tp = tcp_sk(sk); 3260 3261 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because 3262 * broken middle-boxes or peers may corrupt TS-ECR fields. But 3263 * Karn's algorithm forbids taking RTT if some retransmitted data 3264 * is acked (RFC6298). 3265 */ 3266 if (seq_rtt_us < 0) 3267 seq_rtt_us = sack_rtt_us; 3268 3269 /* RTTM Rule: A TSecr value received in a segment is used to 3270 * update the averaged RTT measurement only if the segment 3271 * acknowledges some new data, i.e., only if it advances the 3272 * left edge of the send window. 3273 * See draft-ietf-tcplw-high-performance-00, section 3.3. 3274 */ 3275 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && 3276 tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED) 3277 seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp, 1); 3278 3279 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */ 3280 if (seq_rtt_us < 0) 3281 return false; 3282 3283 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is 3284 * always taken together with ACK, SACK, or TS-opts. Any negative 3285 * values will be skipped with the seq_rtt_us < 0 check above. 3286 */ 3287 tcp_update_rtt_min(sk, ca_rtt_us, flag); 3288 tcp_rtt_estimator(sk, seq_rtt_us); 3289 tcp_set_rto(sk); 3290 3291 /* RFC6298: only reset backoff on valid RTT measurement. */ 3292 inet_csk(sk)->icsk_backoff = 0; 3293 return true; 3294} 3295 3296/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ 3297void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) 3298{ 3299 struct rate_sample rs; 3300 long rtt_us = -1L; 3301 3302 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack) 3303 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack); 3304 3305 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs); 3306} 3307 3308 3309static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 3310{ 3311 const struct inet_connection_sock *icsk = inet_csk(sk); 3312 3313 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked); 3314 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32; 3315} 3316 3317/* Restart timer after forward progress on connection. 3318 * RFC2988 recommends to restart timer to now+rto. 3319 */ 3320void tcp_rearm_rto(struct sock *sk) 3321{ 3322 const struct inet_connection_sock *icsk = inet_csk(sk); 3323 struct tcp_sock *tp = tcp_sk(sk); 3324 3325 /* If the retrans timer is currently being used by Fast Open 3326 * for SYN-ACK retrans purpose, stay put. 3327 */ 3328 if (rcu_access_pointer(tp->fastopen_rsk)) 3329 return; 3330 3331 if (!tp->packets_out) { 3332 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3333 } else { 3334 u32 rto = inet_csk(sk)->icsk_rto; 3335 /* Offset the time elapsed after installing regular RTO */ 3336 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT || 3337 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 3338 s64 delta_us = tcp_rto_delta_us(sk); 3339 /* delta_us may not be positive if the socket is locked 3340 * when the retrans timer fires and is rescheduled. 3341 */ 3342 rto = usecs_to_jiffies(max_t(int, delta_us, 1)); 3343 } 3344 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, true); 3345 } 3346} 3347 3348/* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */ 3349static void tcp_set_xmit_timer(struct sock *sk) 3350{ 3351 if (!tcp_schedule_loss_probe(sk, true)) 3352 tcp_rearm_rto(sk); 3353} 3354 3355/* If we get here, the whole TSO packet has not been acked. */ 3356static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3357{ 3358 struct tcp_sock *tp = tcp_sk(sk); 3359 u32 packets_acked; 3360 3361 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3362 3363 packets_acked = tcp_skb_pcount(skb); 3364 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3365 return 0; 3366 packets_acked -= tcp_skb_pcount(skb); 3367 3368 if (packets_acked) { 3369 BUG_ON(tcp_skb_pcount(skb) == 0); 3370 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3371 } 3372 3373 return packets_acked; 3374} 3375 3376static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb, 3377 const struct sk_buff *ack_skb, u32 prior_snd_una) 3378{ 3379 const struct skb_shared_info *shinfo; 3380 3381 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */ 3382 if (likely(!TCP_SKB_CB(skb)->txstamp_ack)) 3383 return; 3384 3385 shinfo = skb_shinfo(skb); 3386 if (!before(shinfo->tskey, prior_snd_una) && 3387 before(shinfo->tskey, tcp_sk(sk)->snd_una)) { 3388 tcp_skb_tsorted_save(skb) { 3389 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK); 3390 } tcp_skb_tsorted_restore(skb); 3391 } 3392} 3393 3394/* Remove acknowledged frames from the retransmission queue. If our packet 3395 * is before the ack sequence we can discard it as it's confirmed to have 3396 * arrived at the other end. 3397 */ 3398static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb, 3399 u32 prior_fack, u32 prior_snd_una, 3400 struct tcp_sacktag_state *sack, bool ece_ack) 3401{ 3402 const struct inet_connection_sock *icsk = inet_csk(sk); 3403 u64 first_ackt, last_ackt; 3404 struct tcp_sock *tp = tcp_sk(sk); 3405 u32 prior_sacked = tp->sacked_out; 3406 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */ 3407 struct sk_buff *skb, *next; 3408 bool fully_acked = true; 3409 long sack_rtt_us = -1L; 3410 long seq_rtt_us = -1L; 3411 long ca_rtt_us = -1L; 3412 u32 pkts_acked = 0; 3413 bool rtt_update; 3414 int flag = 0; 3415 3416 first_ackt = 0; 3417 3418 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) { 3419 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3420 const u32 start_seq = scb->seq; 3421 u8 sacked = scb->sacked; 3422 u32 acked_pcount; 3423 3424 /* Determine how many packets and what bytes were acked, tso and else */ 3425 if (after(scb->end_seq, tp->snd_una)) { 3426 if (tcp_skb_pcount(skb) == 1 || 3427 !after(tp->snd_una, scb->seq)) 3428 break; 3429 3430 acked_pcount = tcp_tso_acked(sk, skb); 3431 if (!acked_pcount) 3432 break; 3433 fully_acked = false; 3434 } else { 3435 acked_pcount = tcp_skb_pcount(skb); 3436 } 3437 3438 if (unlikely(sacked & TCPCB_RETRANS)) { 3439 if (sacked & TCPCB_SACKED_RETRANS) 3440 tp->retrans_out -= acked_pcount; 3441 flag |= FLAG_RETRANS_DATA_ACKED; 3442 } else if (!(sacked & TCPCB_SACKED_ACKED)) { 3443 last_ackt = tcp_skb_timestamp_us(skb); 3444 WARN_ON_ONCE(last_ackt == 0); 3445 if (!first_ackt) 3446 first_ackt = last_ackt; 3447 3448 if (before(start_seq, reord)) 3449 reord = start_seq; 3450 if (!after(scb->end_seq, tp->high_seq)) 3451 flag |= FLAG_ORIG_SACK_ACKED; 3452 } 3453 3454 if (sacked & TCPCB_SACKED_ACKED) { 3455 tp->sacked_out -= acked_pcount; 3456 /* snd_una delta covers these skbs */ 3457 sack->delivered_bytes -= skb->len; 3458 } else if (tcp_is_sack(tp)) { 3459 tcp_count_delivered(tp, acked_pcount, ece_ack); 3460 if (!tcp_skb_spurious_retrans(tp, skb)) 3461 tcp_rack_advance(tp, sacked, scb->end_seq, 3462 tcp_skb_timestamp_us(skb)); 3463 } 3464 if (sacked & TCPCB_LOST) 3465 tp->lost_out -= acked_pcount; 3466 3467 tp->packets_out -= acked_pcount; 3468 pkts_acked += acked_pcount; 3469 tcp_rate_skb_delivered(sk, skb, sack->rate); 3470 3471 /* Initial outgoing SYN's get put onto the write_queue 3472 * just like anything else we transmit. It is not 3473 * true data, and if we misinform our callers that 3474 * this ACK acks real data, we will erroneously exit 3475 * connection startup slow start one packet too 3476 * quickly. This is severely frowned upon behavior. 3477 */ 3478 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) { 3479 flag |= FLAG_DATA_ACKED; 3480 } else { 3481 flag |= FLAG_SYN_ACKED; 3482 tp->retrans_stamp = 0; 3483 } 3484 3485 if (!fully_acked) 3486 break; 3487 3488 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); 3489 3490 next = skb_rb_next(skb); 3491 if (unlikely(skb == tp->retransmit_skb_hint)) 3492 tp->retransmit_skb_hint = NULL; 3493 tcp_highest_sack_replace(sk, skb, next); 3494 tcp_rtx_queue_unlink_and_free(skb, sk); 3495 } 3496 3497 if (!skb) 3498 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 3499 3500 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3501 tp->snd_up = tp->snd_una; 3502 3503 if (skb) { 3504 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); 3505 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3506 flag |= FLAG_SACK_RENEGING; 3507 } 3508 3509 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) { 3510 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt); 3511 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt); 3512 3513 if (pkts_acked == 1 && fully_acked && !prior_sacked && 3514 (tp->snd_una - prior_snd_una) < tp->mss_cache && 3515 sack->rate->prior_delivered + 1 == tp->delivered && 3516 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) { 3517 /* Conservatively mark a delayed ACK. It's typically 3518 * from a lone runt packet over the round trip to 3519 * a receiver w/o out-of-order or CE events. 3520 */ 3521 flag |= FLAG_ACK_MAYBE_DELAYED; 3522 } 3523 } 3524 if (sack->first_sackt) { 3525 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt); 3526 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt); 3527 } 3528 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us, 3529 ca_rtt_us, sack->rate); 3530 3531 if (flag & FLAG_ACKED) { 3532 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3533 if (unlikely(icsk->icsk_mtup.probe_size && 3534 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3535 tcp_mtup_probe_success(sk); 3536 } 3537 3538 if (tcp_is_reno(tp)) { 3539 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack); 3540 3541 /* If any of the cumulatively ACKed segments was 3542 * retransmitted, non-SACK case cannot confirm that 3543 * progress was due to original transmission due to 3544 * lack of TCPCB_SACKED_ACKED bits even if some of 3545 * the packets may have been never retransmitted. 3546 */ 3547 if (flag & FLAG_RETRANS_DATA_ACKED) 3548 flag &= ~FLAG_ORIG_SACK_ACKED; 3549 } else { 3550 /* Non-retransmitted hole got filled? That's reordering */ 3551 if (before(reord, prior_fack)) 3552 tcp_check_sack_reordering(sk, reord, 0); 3553 } 3554 3555 sack->delivered_bytes = (skb ? 3556 TCP_SKB_CB(skb)->seq : tp->snd_una) - 3557 prior_snd_una; 3558 } else if (skb && rtt_update && sack_rtt_us >= 0 && 3559 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, 3560 tcp_skb_timestamp_us(skb))) { 3561 /* Do not re-arm RTO if the sack RTT is measured from data sent 3562 * after when the head was last (re)transmitted. Otherwise the 3563 * timeout may continue to extend in loss recovery. 3564 */ 3565 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3566 } 3567 3568 if (icsk->icsk_ca_ops->pkts_acked) { 3569 struct ack_sample sample = { .pkts_acked = pkts_acked, 3570 .rtt_us = sack->rate->rtt_us }; 3571 3572 sample.in_flight = tp->mss_cache * 3573 (tp->delivered - sack->rate->prior_delivered); 3574 icsk->icsk_ca_ops->pkts_acked(sk, &sample); 3575 } 3576 3577#if FASTRETRANS_DEBUG > 0 3578 WARN_ON((int)tp->sacked_out < 0); 3579 WARN_ON((int)tp->lost_out < 0); 3580 WARN_ON((int)tp->retrans_out < 0); 3581 if (!tp->packets_out && tcp_is_sack(tp)) { 3582 icsk = inet_csk(sk); 3583 if (tp->lost_out) { 3584 pr_debug("Leak l=%u %d\n", 3585 tp->lost_out, icsk->icsk_ca_state); 3586 tp->lost_out = 0; 3587 } 3588 if (tp->sacked_out) { 3589 pr_debug("Leak s=%u %d\n", 3590 tp->sacked_out, icsk->icsk_ca_state); 3591 tp->sacked_out = 0; 3592 } 3593 if (tp->retrans_out) { 3594 pr_debug("Leak r=%u %d\n", 3595 tp->retrans_out, icsk->icsk_ca_state); 3596 tp->retrans_out = 0; 3597 } 3598 } 3599#endif 3600 return flag; 3601} 3602 3603static void tcp_ack_probe(struct sock *sk) 3604{ 3605 struct inet_connection_sock *icsk = inet_csk(sk); 3606 struct sk_buff *head = tcp_send_head(sk); 3607 const struct tcp_sock *tp = tcp_sk(sk); 3608 3609 /* Was it a usable window open? */ 3610 if (!head) 3611 return; 3612 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) { 3613 icsk->icsk_backoff = 0; 3614 icsk->icsk_probes_tstamp = 0; 3615 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3616 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3617 * This function is not for random using! 3618 */ 3619 } else { 3620 unsigned long when = tcp_probe0_when(sk, tcp_rto_max(sk)); 3621 3622 when = tcp_clamp_probe0_to_user_timeout(sk, when); 3623 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, true); 3624 } 3625} 3626 3627static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) 3628{ 3629 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3630 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3631} 3632 3633/* Decide wheather to run the increase function of congestion control. */ 3634static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3635{ 3636 /* If reordering is high then always grow cwnd whenever data is 3637 * delivered regardless of its ordering. Otherwise stay conservative 3638 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ 3639 * new SACK or ECE mark may first advance cwnd here and later reduce 3640 * cwnd in tcp_fastretrans_alert() based on more states. 3641 */ 3642 if (tcp_sk(sk)->reordering > 3643 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering)) 3644 return flag & FLAG_FORWARD_PROGRESS; 3645 3646 return flag & FLAG_DATA_ACKED; 3647} 3648 3649/* The "ultimate" congestion control function that aims to replace the rigid 3650 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction). 3651 * It's called toward the end of processing an ACK with precise rate 3652 * information. All transmission or retransmission are delayed afterwards. 3653 */ 3654static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked, 3655 int flag, const struct rate_sample *rs) 3656{ 3657 const struct inet_connection_sock *icsk = inet_csk(sk); 3658 3659 if (icsk->icsk_ca_ops->cong_control) { 3660 icsk->icsk_ca_ops->cong_control(sk, ack, flag, rs); 3661 return; 3662 } 3663 3664 if (tcp_in_cwnd_reduction(sk)) { 3665 /* Reduce cwnd if state mandates */ 3666 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag); 3667 } else if (tcp_may_raise_cwnd(sk, flag)) { 3668 /* Advance cwnd if state allows */ 3669 tcp_cong_avoid(sk, ack, acked_sacked); 3670 } 3671 tcp_update_pacing_rate(sk); 3672} 3673 3674/* Check that window update is acceptable. 3675 * The function assumes that snd_una<=ack<=snd_next. 3676 */ 3677static inline bool tcp_may_update_window(const struct tcp_sock *tp, 3678 const u32 ack, const u32 ack_seq, 3679 const u32 nwin) 3680{ 3681 return after(ack, tp->snd_una) || 3682 after(ack_seq, tp->snd_wl1) || 3683 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin)); 3684} 3685 3686static void tcp_snd_sne_update(struct tcp_sock *tp, u32 ack) 3687{ 3688#ifdef CONFIG_TCP_AO 3689 struct tcp_ao_info *ao; 3690 3691 if (!static_branch_unlikely(&tcp_ao_needed.key)) 3692 return; 3693 3694 ao = rcu_dereference_protected(tp->ao_info, 3695 lockdep_sock_is_held((struct sock *)tp)); 3696 if (ao && ack < tp->snd_una) { 3697 ao->snd_sne++; 3698 trace_tcp_ao_snd_sne_update((struct sock *)tp, ao->snd_sne); 3699 } 3700#endif 3701} 3702 3703/* If we update tp->snd_una, also update tp->bytes_acked */ 3704static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack) 3705{ 3706 u32 delta = ack - tp->snd_una; 3707 3708 sock_owned_by_me((struct sock *)tp); 3709 tp->bytes_acked += delta; 3710 tcp_snd_sne_update(tp, ack); 3711 tp->snd_una = ack; 3712} 3713 3714static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq) 3715{ 3716#ifdef CONFIG_TCP_AO 3717 struct tcp_ao_info *ao; 3718 3719 if (!static_branch_unlikely(&tcp_ao_needed.key)) 3720 return; 3721 3722 ao = rcu_dereference_protected(tp->ao_info, 3723 lockdep_sock_is_held((struct sock *)tp)); 3724 if (ao && seq < tp->rcv_nxt) { 3725 ao->rcv_sne++; 3726 trace_tcp_ao_rcv_sne_update((struct sock *)tp, ao->rcv_sne); 3727 } 3728#endif 3729} 3730 3731/* If we update tp->rcv_nxt, also update tp->bytes_received */ 3732static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq) 3733{ 3734 u32 delta = seq - tp->rcv_nxt; 3735 3736 sock_owned_by_me((struct sock *)tp); 3737 tp->bytes_received += delta; 3738 tcp_rcv_sne_update(tp, seq); 3739 WRITE_ONCE(tp->rcv_nxt, seq); 3740} 3741 3742/* Update our send window. 3743 * 3744 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3745 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3746 */ 3747static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3748 u32 ack_seq) 3749{ 3750 struct tcp_sock *tp = tcp_sk(sk); 3751 int flag = 0; 3752 u32 nwin = ntohs(tcp_hdr(skb)->window); 3753 3754 if (likely(!tcp_hdr(skb)->syn)) 3755 nwin <<= tp->rx_opt.snd_wscale; 3756 3757 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3758 flag |= FLAG_WIN_UPDATE; 3759 tcp_update_wl(tp, ack_seq); 3760 3761 if (tp->snd_wnd != nwin) { 3762 tp->snd_wnd = nwin; 3763 3764 /* Note, it is the only place, where 3765 * fast path is recovered for sending TCP. 3766 */ 3767 tp->pred_flags = 0; 3768 tcp_fast_path_check(sk); 3769 3770 if (!tcp_write_queue_empty(sk)) 3771 tcp_slow_start_after_idle_check(sk); 3772 3773 if (nwin > tp->max_window) { 3774 tp->max_window = nwin; 3775 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 3776 } 3777 } 3778 } 3779 3780 tcp_snd_una_update(tp, ack); 3781 3782 return flag; 3783} 3784 3785static bool __tcp_oow_rate_limited(struct net *net, int mib_idx, 3786 u32 *last_oow_ack_time) 3787{ 3788 /* Paired with the WRITE_ONCE() in this function. */ 3789 u32 val = READ_ONCE(*last_oow_ack_time); 3790 3791 if (val) { 3792 s32 elapsed = (s32)(tcp_jiffies32 - val); 3793 3794 if (0 <= elapsed && 3795 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) { 3796 NET_INC_STATS(net, mib_idx); 3797 return true; /* rate-limited: don't send yet! */ 3798 } 3799 } 3800 3801 /* Paired with the prior READ_ONCE() and with itself, 3802 * as we might be lockless. 3803 */ 3804 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32); 3805 3806 return false; /* not rate-limited: go ahead, send dupack now! */ 3807} 3808 3809/* Return true if we're currently rate-limiting out-of-window ACKs and 3810 * thus shouldn't send a dupack right now. We rate-limit dupacks in 3811 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS 3812 * attacks that send repeated SYNs or ACKs for the same connection. To 3813 * do this, we do not send a duplicate SYNACK or ACK if the remote 3814 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate. 3815 */ 3816bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb, 3817 int mib_idx, u32 *last_oow_ack_time) 3818{ 3819 /* Data packets without SYNs are not likely part of an ACK loop. */ 3820 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) && 3821 !tcp_hdr(skb)->syn) 3822 return false; 3823 3824 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time); 3825} 3826 3827static void tcp_send_ack_reflect_ect(struct sock *sk, bool accecn_reflector) 3828{ 3829 struct tcp_sock *tp = tcp_sk(sk); 3830 u16 flags = 0; 3831 3832 if (accecn_reflector) 3833 flags = tcp_accecn_reflector_flags(tp->syn_ect_rcv); 3834 __tcp_send_ack(sk, tp->rcv_nxt, flags); 3835} 3836 3837/* RFC 5961 7 [ACK Throttling] */ 3838static void tcp_send_challenge_ack(struct sock *sk, bool accecn_reflector) 3839{ 3840 struct tcp_sock *tp = tcp_sk(sk); 3841 struct net *net = sock_net(sk); 3842 u32 count, now, ack_limit; 3843 3844 /* First check our per-socket dupack rate limit. */ 3845 if (__tcp_oow_rate_limited(net, 3846 LINUX_MIB_TCPACKSKIPPEDCHALLENGE, 3847 &tp->last_oow_ack_time)) 3848 return; 3849 3850 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit); 3851 if (ack_limit == INT_MAX) 3852 goto send_ack; 3853 3854 /* Then check host-wide RFC 5961 rate limit. */ 3855 now = jiffies / HZ; 3856 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) { 3857 u32 half = (ack_limit + 1) >> 1; 3858 3859 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now); 3860 WRITE_ONCE(net->ipv4.tcp_challenge_count, 3861 get_random_u32_inclusive(half, ack_limit + half - 1)); 3862 } 3863 count = READ_ONCE(net->ipv4.tcp_challenge_count); 3864 if (count > 0) { 3865 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1); 3866send_ack: 3867 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK); 3868 tcp_send_ack_reflect_ect(sk, accecn_reflector); 3869 } 3870} 3871 3872static void tcp_store_ts_recent(struct tcp_sock *tp) 3873{ 3874 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 3875 tp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 3876} 3877 3878static int __tcp_replace_ts_recent(struct tcp_sock *tp, s32 tstamp_delta) 3879{ 3880 tcp_store_ts_recent(tp); 3881 return tstamp_delta > 0 ? FLAG_TS_PROGRESS : 0; 3882} 3883 3884static int tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 3885{ 3886 s32 delta; 3887 3888 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 3889 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 3890 * extra check below makes sure this can only happen 3891 * for pure ACK frames. -DaveM 3892 * 3893 * Not only, also it occurs for expired timestamps. 3894 */ 3895 3896 if (tcp_paws_check(&tp->rx_opt, 0)) { 3897 delta = tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent; 3898 return __tcp_replace_ts_recent(tp, delta); 3899 } 3900 } 3901 3902 return 0; 3903} 3904 3905/* This routine deals with acks during a TLP episode and ends an episode by 3906 * resetting tlp_high_seq. Ref: TLP algorithm in RFC8985 3907 */ 3908static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) 3909{ 3910 struct tcp_sock *tp = tcp_sk(sk); 3911 3912 if (before(ack, tp->tlp_high_seq)) 3913 return; 3914 3915 if (!tp->tlp_retrans) { 3916 /* TLP of new data has been acknowledged */ 3917 tp->tlp_high_seq = 0; 3918 } else if (flag & FLAG_DSACK_TLP) { 3919 /* This DSACK means original and TLP probe arrived; no loss */ 3920 tp->tlp_high_seq = 0; 3921 } else if (after(ack, tp->tlp_high_seq)) { 3922 /* ACK advances: there was a loss, so reduce cwnd. Reset 3923 * tlp_high_seq in tcp_init_cwnd_reduction() 3924 */ 3925 tcp_init_cwnd_reduction(sk); 3926 tcp_set_ca_state(sk, TCP_CA_CWR); 3927 tcp_end_cwnd_reduction(sk); 3928 tcp_try_keep_open(sk); 3929 NET_INC_STATS(sock_net(sk), 3930 LINUX_MIB_TCPLOSSPROBERECOVERY); 3931 } else if (!(flag & (FLAG_SND_UNA_ADVANCED | 3932 FLAG_NOT_DUP | FLAG_DATA_SACKED))) { 3933 /* Pure dupack: original and TLP probe arrived; no loss */ 3934 tp->tlp_high_seq = 0; 3935 } 3936} 3937 3938static void tcp_in_ack_event(struct sock *sk, int flag) 3939{ 3940 const struct inet_connection_sock *icsk = inet_csk(sk); 3941 3942 if (icsk->icsk_ca_ops->in_ack_event) { 3943 u32 ack_ev_flags = 0; 3944 3945 if (flag & FLAG_WIN_UPDATE) 3946 ack_ev_flags |= CA_ACK_WIN_UPDATE; 3947 if (flag & FLAG_SLOWPATH) { 3948 ack_ev_flags |= CA_ACK_SLOWPATH; 3949 if (flag & FLAG_ECE) 3950 ack_ev_flags |= CA_ACK_ECE; 3951 } 3952 3953 icsk->icsk_ca_ops->in_ack_event(sk, ack_ev_flags); 3954 } 3955} 3956 3957/* Congestion control has updated the cwnd already. So if we're in 3958 * loss recovery then now we do any new sends (for FRTO) or 3959 * retransmits (for CA_Loss or CA_recovery) that make sense. 3960 */ 3961static void tcp_xmit_recovery(struct sock *sk, int rexmit) 3962{ 3963 struct tcp_sock *tp = tcp_sk(sk); 3964 3965 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT) 3966 return; 3967 3968 if (unlikely(rexmit == REXMIT_NEW)) { 3969 __tcp_push_pending_frames(sk, tcp_current_mss(sk), 3970 TCP_NAGLE_OFF); 3971 if (after(tp->snd_nxt, tp->high_seq)) 3972 return; 3973 tp->frto = 0; 3974 } 3975 tcp_xmit_retransmit_queue(sk); 3976} 3977 3978/* Returns the number of packets newly acked or sacked by the current ACK */ 3979static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, 3980 u32 ecn_count, int flag) 3981{ 3982 const struct net *net = sock_net(sk); 3983 struct tcp_sock *tp = tcp_sk(sk); 3984 u32 delivered; 3985 3986 delivered = tp->delivered - prior_delivered; 3987 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered); 3988 3989 if (flag & FLAG_ECE) { 3990 if (tcp_ecn_mode_rfc3168(tp)) 3991 ecn_count = delivered; 3992 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, ecn_count); 3993 } 3994 3995 return delivered; 3996} 3997 3998/* This routine deals with incoming acks, but not outgoing ones. */ 3999static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 4000{ 4001 struct inet_connection_sock *icsk = inet_csk(sk); 4002 struct tcp_sock *tp = tcp_sk(sk); 4003 struct tcp_sacktag_state sack_state; 4004 struct rate_sample rs = { .prior_delivered = 0 }; 4005 u32 prior_snd_una = tp->snd_una; 4006 bool is_sack_reneg = tp->is_sack_reneg; 4007 u32 ack_seq = TCP_SKB_CB(skb)->seq; 4008 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4009 int num_dupack = 0; 4010 int prior_packets = tp->packets_out; 4011 u32 delivered = tp->delivered; 4012 u32 lost = tp->lost; 4013 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */ 4014 u32 ecn_count = 0; /* Did we receive ECE/an AccECN ACE update? */ 4015 u32 prior_fack; 4016 4017 sack_state.first_sackt = 0; 4018 sack_state.rate = &rs; 4019 sack_state.sack_delivered = 0; 4020 sack_state.delivered_bytes = 0; 4021 4022 /* We very likely will need to access rtx queue. */ 4023 prefetch(sk->tcp_rtx_queue.rb_node); 4024 4025 /* If the ack is older than previous acks 4026 * then we can probably ignore it. 4027 */ 4028 if (before(ack, prior_snd_una)) { 4029 u32 max_window; 4030 4031 /* do not accept ACK for bytes we never sent. */ 4032 max_window = min_t(u64, tp->max_window, tp->bytes_acked); 4033 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */ 4034 if (before(ack, prior_snd_una - max_window)) { 4035 if (!(flag & FLAG_NO_CHALLENGE_ACK)) 4036 tcp_send_challenge_ack(sk, false); 4037 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK; 4038 } 4039 goto old_ack; 4040 } 4041 4042 /* If the ack includes data we haven't sent yet, discard 4043 * this segment (RFC793 Section 3.9). 4044 */ 4045 if (after(ack, tp->snd_nxt)) 4046 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA; 4047 4048 if (after(ack, prior_snd_una)) { 4049 flag |= FLAG_SND_UNA_ADVANCED; 4050 WRITE_ONCE(icsk->icsk_retransmits, 0); 4051 4052#if IS_ENABLED(CONFIG_TLS_DEVICE) 4053 if (static_branch_unlikely(&clean_acked_data_enabled.key)) 4054 if (tp->tcp_clean_acked) 4055 tp->tcp_clean_acked(sk, ack); 4056#endif 4057 } 4058 4059 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una; 4060 rs.prior_in_flight = tcp_packets_in_flight(tp); 4061 4062 /* ts_recent update must be made after we are sure that the packet 4063 * is in window. 4064 */ 4065 if (flag & FLAG_UPDATE_TS_RECENT) 4066 flag |= tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4067 4068 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) == 4069 FLAG_SND_UNA_ADVANCED) { 4070 /* Window is constant, pure forward advance. 4071 * No more checks are required. 4072 * Note, we use the fact that SND.UNA>=SND.WL2. 4073 */ 4074 tcp_update_wl(tp, ack_seq); 4075 tcp_snd_una_update(tp, ack); 4076 flag |= FLAG_WIN_UPDATE; 4077 4078 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS); 4079 } else { 4080 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 4081 flag |= FLAG_DATA; 4082 else 4083 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS); 4084 4085 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 4086 4087 if (TCP_SKB_CB(skb)->sacked) 4088 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 4089 &sack_state); 4090 4091 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) 4092 flag |= FLAG_ECE; 4093 4094 if (sack_state.sack_delivered) 4095 tcp_count_delivered(tp, sack_state.sack_delivered, 4096 flag & FLAG_ECE); 4097 } 4098 4099 /* This is a deviation from RFC3168 since it states that: 4100 * "When the TCP data sender is ready to set the CWR bit after reducing 4101 * the congestion window, it SHOULD set the CWR bit only on the first 4102 * new data packet that it transmits." 4103 * We accept CWR on pure ACKs to be more robust 4104 * with widely-deployed TCP implementations that do this. 4105 */ 4106 tcp_ecn_accept_cwr(sk, skb); 4107 4108 /* We passed data and got it acked, remove any soft error 4109 * log. Something worked... 4110 */ 4111 if (READ_ONCE(sk->sk_err_soft)) 4112 WRITE_ONCE(sk->sk_err_soft, 0); 4113 WRITE_ONCE(icsk->icsk_probes_out, 0); 4114 tp->rcv_tstamp = tcp_jiffies32; 4115 if (!prior_packets) 4116 goto no_queue; 4117 4118 /* See if we can take anything off of the retransmit queue. */ 4119 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una, 4120 &sack_state, flag & FLAG_ECE); 4121 4122 tcp_rack_update_reo_wnd(sk, &rs); 4123 4124 if (tcp_ecn_mode_accecn(tp)) 4125 ecn_count = tcp_accecn_process(sk, skb, 4126 tp->delivered - delivered, 4127 sack_state.delivered_bytes, 4128 &flag); 4129 4130 tcp_in_ack_event(sk, flag); 4131 4132 if (tp->tlp_high_seq) 4133 tcp_process_tlp_ack(sk, ack, flag); 4134 4135 if (tcp_ack_is_dubious(sk, flag)) { 4136 if (!(flag & (FLAG_SND_UNA_ADVANCED | 4137 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) { 4138 num_dupack = 1; 4139 /* Consider if pure acks were aggregated in tcp_add_backlog() */ 4140 if (!(flag & FLAG_DATA)) 4141 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs); 4142 } 4143 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4144 &rexmit); 4145 } 4146 4147 /* If needed, reset TLP/RTO timer when RACK doesn't set. */ 4148 if (flag & FLAG_SET_XMIT_TIMER) 4149 tcp_set_xmit_timer(sk); 4150 4151 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 4152 sk_dst_confirm(sk); 4153 4154 delivered = tcp_newly_delivered(sk, delivered, ecn_count, flag); 4155 4156 lost = tp->lost - lost; /* freshly marked lost */ 4157 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED); 4158 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate); 4159 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate); 4160 tcp_xmit_recovery(sk, rexmit); 4161 return 1; 4162 4163no_queue: 4164 if (tcp_ecn_mode_accecn(tp)) 4165 ecn_count = tcp_accecn_process(sk, skb, 4166 tp->delivered - delivered, 4167 sack_state.delivered_bytes, 4168 &flag); 4169 tcp_in_ack_event(sk, flag); 4170 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 4171 if (flag & FLAG_DSACKING_ACK) { 4172 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4173 &rexmit); 4174 tcp_newly_delivered(sk, delivered, ecn_count, flag); 4175 } 4176 /* If this ack opens up a zero window, clear backoff. It was 4177 * being used to time the probes, and is probably far higher than 4178 * it needs to be for normal retransmission. 4179 */ 4180 tcp_ack_probe(sk); 4181 4182 if (tp->tlp_high_seq) 4183 tcp_process_tlp_ack(sk, ack, flag); 4184 return 1; 4185 4186old_ack: 4187 /* If data was SACKed, tag it and see if we should send more data. 4188 * If data was DSACKed, see if we can undo a cwnd reduction. 4189 */ 4190 if (TCP_SKB_CB(skb)->sacked) { 4191 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 4192 &sack_state); 4193 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4194 &rexmit); 4195 tcp_newly_delivered(sk, delivered, ecn_count, flag); 4196 tcp_xmit_recovery(sk, rexmit); 4197 } 4198 4199 return 0; 4200} 4201 4202static void tcp_parse_fastopen_option(int len, const unsigned char *cookie, 4203 bool syn, struct tcp_fastopen_cookie *foc, 4204 bool exp_opt) 4205{ 4206 /* Valid only in SYN or SYN-ACK with an even length. */ 4207 if (!foc || !syn || len < 0 || (len & 1)) 4208 return; 4209 4210 if (len >= TCP_FASTOPEN_COOKIE_MIN && 4211 len <= TCP_FASTOPEN_COOKIE_MAX) 4212 memcpy(foc->val, cookie, len); 4213 else if (len != 0) 4214 len = -1; 4215 foc->len = len; 4216 foc->exp = exp_opt; 4217} 4218 4219static bool smc_parse_options(const struct tcphdr *th, 4220 struct tcp_options_received *opt_rx, 4221 const unsigned char *ptr, 4222 int opsize) 4223{ 4224#if IS_ENABLED(CONFIG_SMC) 4225 if (static_branch_unlikely(&tcp_have_smc)) { 4226 if (th->syn && !(opsize & 1) && 4227 opsize >= TCPOLEN_EXP_SMC_BASE && 4228 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) { 4229 opt_rx->smc_ok = 1; 4230 return true; 4231 } 4232 } 4233#endif 4234 return false; 4235} 4236 4237/* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped 4238 * value on success. 4239 */ 4240u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss) 4241{ 4242 const unsigned char *ptr = (const unsigned char *)(th + 1); 4243 int length = (th->doff * 4) - sizeof(struct tcphdr); 4244 u16 mss = 0; 4245 4246 while (length > 0) { 4247 int opcode = *ptr++; 4248 int opsize; 4249 4250 switch (opcode) { 4251 case TCPOPT_EOL: 4252 return mss; 4253 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 4254 length--; 4255 continue; 4256 default: 4257 if (length < 2) 4258 return mss; 4259 opsize = *ptr++; 4260 if (opsize < 2) /* "silly options" */ 4261 return mss; 4262 if (opsize > length) 4263 return mss; /* fail on partial options */ 4264 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) { 4265 u16 in_mss = get_unaligned_be16(ptr); 4266 4267 if (in_mss) { 4268 if (user_mss && user_mss < in_mss) 4269 in_mss = user_mss; 4270 mss = in_mss; 4271 } 4272 } 4273 ptr += opsize - 2; 4274 length -= opsize; 4275 } 4276 } 4277 return mss; 4278} 4279 4280/* Look for tcp options. Normally only called on SYN and SYNACK packets. 4281 * But, this can also be called on packets in the established flow when 4282 * the fast version below fails. 4283 */ 4284void tcp_parse_options(const struct net *net, 4285 const struct sk_buff *skb, 4286 struct tcp_options_received *opt_rx, int estab, 4287 struct tcp_fastopen_cookie *foc) 4288{ 4289 const unsigned char *ptr; 4290 const struct tcphdr *th = tcp_hdr(skb); 4291 int length = (th->doff * 4) - sizeof(struct tcphdr); 4292 4293 ptr = (const unsigned char *)(th + 1); 4294 opt_rx->saw_tstamp = 0; 4295 opt_rx->accecn = 0; 4296 opt_rx->saw_unknown = 0; 4297 4298 while (length > 0) { 4299 int opcode = *ptr++; 4300 int opsize; 4301 4302 switch (opcode) { 4303 case TCPOPT_EOL: 4304 return; 4305 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 4306 length--; 4307 continue; 4308 default: 4309 if (length < 2) 4310 return; 4311 opsize = *ptr++; 4312 if (opsize < 2) /* "silly options" */ 4313 return; 4314 if (opsize > length) 4315 return; /* don't parse partial options */ 4316 switch (opcode) { 4317 case TCPOPT_MSS: 4318 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 4319 u16 in_mss = get_unaligned_be16(ptr); 4320 if (in_mss) { 4321 if (opt_rx->user_mss && 4322 opt_rx->user_mss < in_mss) 4323 in_mss = opt_rx->user_mss; 4324 opt_rx->mss_clamp = in_mss; 4325 } 4326 } 4327 break; 4328 case TCPOPT_WINDOW: 4329 if (opsize == TCPOLEN_WINDOW && th->syn && 4330 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) { 4331 __u8 snd_wscale = *(__u8 *)ptr; 4332 opt_rx->wscale_ok = 1; 4333 if (snd_wscale > TCP_MAX_WSCALE) { 4334 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n", 4335 __func__, 4336 snd_wscale, 4337 TCP_MAX_WSCALE); 4338 snd_wscale = TCP_MAX_WSCALE; 4339 } 4340 opt_rx->snd_wscale = snd_wscale; 4341 } 4342 break; 4343 case TCPOPT_TIMESTAMP: 4344 if ((opsize == TCPOLEN_TIMESTAMP) && 4345 ((estab && opt_rx->tstamp_ok) || 4346 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) { 4347 opt_rx->saw_tstamp = 1; 4348 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 4349 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 4350 } 4351 break; 4352 case TCPOPT_SACK_PERM: 4353 if (opsize == TCPOLEN_SACK_PERM && th->syn && 4354 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) { 4355 opt_rx->sack_ok = TCP_SACK_SEEN; 4356 tcp_sack_reset(opt_rx); 4357 } 4358 break; 4359 4360 case TCPOPT_SACK: 4361 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 4362 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 4363 opt_rx->sack_ok) { 4364 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 4365 } 4366 break; 4367#ifdef CONFIG_TCP_MD5SIG 4368 case TCPOPT_MD5SIG: 4369 /* The MD5 Hash has already been 4370 * checked (see tcp_v{4,6}_rcv()). 4371 */ 4372 break; 4373#endif 4374#ifdef CONFIG_TCP_AO 4375 case TCPOPT_AO: 4376 /* TCP AO has already been checked 4377 * (see tcp_inbound_ao_hash()). 4378 */ 4379 break; 4380#endif 4381 case TCPOPT_FASTOPEN: 4382 tcp_parse_fastopen_option( 4383 opsize - TCPOLEN_FASTOPEN_BASE, 4384 ptr, th->syn, foc, false); 4385 break; 4386 4387 case TCPOPT_ACCECN0: 4388 case TCPOPT_ACCECN1: 4389 /* Save offset of AccECN option in TCP header */ 4390 opt_rx->accecn = (ptr - 2) - (__u8 *)th; 4391 break; 4392 4393 case TCPOPT_EXP: 4394 /* Fast Open option shares code 254 using a 4395 * 16 bits magic number. 4396 */ 4397 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE && 4398 get_unaligned_be16(ptr) == 4399 TCPOPT_FASTOPEN_MAGIC) { 4400 tcp_parse_fastopen_option(opsize - 4401 TCPOLEN_EXP_FASTOPEN_BASE, 4402 ptr + 2, th->syn, foc, true); 4403 break; 4404 } 4405 4406 if (smc_parse_options(th, opt_rx, ptr, opsize)) 4407 break; 4408 4409 opt_rx->saw_unknown = 1; 4410 break; 4411 4412 default: 4413 opt_rx->saw_unknown = 1; 4414 } 4415 ptr += opsize-2; 4416 length -= opsize; 4417 } 4418 } 4419} 4420EXPORT_SYMBOL(tcp_parse_options); 4421 4422static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 4423{ 4424 const __be32 *ptr = (const __be32 *)(th + 1); 4425 4426 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4427 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 4428 tp->rx_opt.saw_tstamp = 1; 4429 ++ptr; 4430 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4431 ++ptr; 4432 if (*ptr) 4433 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset; 4434 else 4435 tp->rx_opt.rcv_tsecr = 0; 4436 return true; 4437 } 4438 return false; 4439} 4440 4441/* Fast parse options. This hopes to only see timestamps. 4442 * If it is wrong it falls back on tcp_parse_options(). 4443 */ 4444static bool tcp_fast_parse_options(const struct net *net, 4445 const struct sk_buff *skb, 4446 const struct tcphdr *th, struct tcp_sock *tp) 4447{ 4448 /* In the spirit of fast parsing, compare doff directly to constant 4449 * values. Because equality is used, short doff can be ignored here. 4450 */ 4451 if (th->doff == (sizeof(*th) / 4)) { 4452 tp->rx_opt.saw_tstamp = 0; 4453 tp->rx_opt.accecn = 0; 4454 return false; 4455 } else if (tp->rx_opt.tstamp_ok && 4456 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 4457 if (tcp_parse_aligned_timestamp(tp, th)) { 4458 tp->rx_opt.accecn = 0; 4459 return true; 4460 } 4461 } 4462 4463 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL); 4464 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 4465 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 4466 4467 return true; 4468} 4469 4470#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) 4471/* 4472 * Parse Signature options 4473 */ 4474int tcp_do_parse_auth_options(const struct tcphdr *th, 4475 const u8 **md5_hash, const u8 **ao_hash) 4476{ 4477 int length = (th->doff << 2) - sizeof(*th); 4478 const u8 *ptr = (const u8 *)(th + 1); 4479 unsigned int minlen = TCPOLEN_MD5SIG; 4480 4481 if (IS_ENABLED(CONFIG_TCP_AO)) 4482 minlen = sizeof(struct tcp_ao_hdr) + 1; 4483 4484 *md5_hash = NULL; 4485 *ao_hash = NULL; 4486 4487 /* If not enough data remaining, we can short cut */ 4488 while (length >= minlen) { 4489 int opcode = *ptr++; 4490 int opsize; 4491 4492 switch (opcode) { 4493 case TCPOPT_EOL: 4494 return 0; 4495 case TCPOPT_NOP: 4496 length--; 4497 continue; 4498 default: 4499 opsize = *ptr++; 4500 if (opsize < 2 || opsize > length) 4501 return -EINVAL; 4502 if (opcode == TCPOPT_MD5SIG) { 4503 if (opsize != TCPOLEN_MD5SIG) 4504 return -EINVAL; 4505 if (unlikely(*md5_hash || *ao_hash)) 4506 return -EEXIST; 4507 *md5_hash = ptr; 4508 } else if (opcode == TCPOPT_AO) { 4509 if (opsize <= sizeof(struct tcp_ao_hdr)) 4510 return -EINVAL; 4511 if (unlikely(*md5_hash || *ao_hash)) 4512 return -EEXIST; 4513 *ao_hash = ptr; 4514 } 4515 } 4516 ptr += opsize - 2; 4517 length -= opsize; 4518 } 4519 return 0; 4520} 4521EXPORT_SYMBOL(tcp_do_parse_auth_options); 4522#endif 4523 4524/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 4525 * 4526 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 4527 * it can pass through stack. So, the following predicate verifies that 4528 * this segment is not used for anything but congestion avoidance or 4529 * fast retransmit. Moreover, we even are able to eliminate most of such 4530 * second order effects, if we apply some small "replay" window (~RTO) 4531 * to timestamp space. 4532 * 4533 * All these measures still do not guarantee that we reject wrapped ACKs 4534 * on networks with high bandwidth, when sequence space is recycled fastly, 4535 * but it guarantees that such events will be very rare and do not affect 4536 * connection seriously. This doesn't look nice, but alas, PAWS is really 4537 * buggy extension. 4538 * 4539 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 4540 * states that events when retransmit arrives after original data are rare. 4541 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 4542 * the biggest problem on large power networks even with minor reordering. 4543 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 4544 * up to bandwidth of 18Gigabit/sec. 8) ] 4545 */ 4546 4547/* Estimates max number of increments of remote peer TSval in 4548 * a replay window (based on our current RTO estimation). 4549 */ 4550static u32 tcp_tsval_replay(const struct sock *sk) 4551{ 4552 /* If we use usec TS resolution, 4553 * then expect the remote peer to use the same resolution. 4554 */ 4555 if (tcp_sk(sk)->tcp_usec_ts) 4556 return inet_csk(sk)->icsk_rto * (USEC_PER_SEC / HZ); 4557 4558 /* RFC 7323 recommends a TSval clock between 1ms and 1sec. 4559 * We know that some OS (including old linux) can use 1200 Hz. 4560 */ 4561 return inet_csk(sk)->icsk_rto * 1200 / HZ; 4562} 4563 4564static enum skb_drop_reason tcp_disordered_ack_check(const struct sock *sk, 4565 const struct sk_buff *skb) 4566{ 4567 const struct tcp_sock *tp = tcp_sk(sk); 4568 const struct tcphdr *th = tcp_hdr(skb); 4569 SKB_DR_INIT(reason, TCP_RFC7323_PAWS); 4570 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4571 u32 seq = TCP_SKB_CB(skb)->seq; 4572 4573 /* 1. Is this not a pure ACK ? */ 4574 if (!th->ack || seq != TCP_SKB_CB(skb)->end_seq) 4575 return reason; 4576 4577 /* 2. Is its sequence not the expected one ? */ 4578 if (seq != tp->rcv_nxt) 4579 return before(seq, tp->rcv_nxt) ? 4580 SKB_DROP_REASON_TCP_RFC7323_PAWS_ACK : 4581 reason; 4582 4583 /* 3. Is this not a duplicate ACK ? */ 4584 if (ack != tp->snd_una) 4585 return reason; 4586 4587 /* 4. Is this updating the window ? */ 4588 if (tcp_may_update_window(tp, ack, seq, ntohs(th->window) << 4589 tp->rx_opt.snd_wscale)) 4590 return reason; 4591 4592 /* 5. Is this not in the replay window ? */ 4593 if ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > 4594 tcp_tsval_replay(sk)) 4595 return reason; 4596 4597 return 0; 4598} 4599 4600/* Check segment sequence number for validity. 4601 * 4602 * Segment controls are considered valid, if the segment 4603 * fits to the window after truncation to the window. Acceptability 4604 * of data (and SYN, FIN, of course) is checked separately. 4605 * See tcp_data_queue(), for example. 4606 * 4607 * Also, controls (RST is main one) are accepted using RCV.WUP instead 4608 * of RCV.NXT. Peer still did not advance his SND.UNA when we 4609 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4610 * (borrowed from freebsd) 4611 */ 4612 4613static enum skb_drop_reason tcp_sequence(const struct sock *sk, 4614 u32 seq, u32 end_seq, 4615 const struct tcphdr *th) 4616{ 4617 const struct tcp_sock *tp = tcp_sk(sk); 4618 u32 seq_limit; 4619 4620 if (before(end_seq, tp->rcv_wup)) 4621 return SKB_DROP_REASON_TCP_OLD_SEQUENCE; 4622 4623 seq_limit = tp->rcv_nxt + tcp_receive_window(tp); 4624 if (unlikely(after(end_seq, seq_limit))) { 4625 /* Some stacks are known to handle FIN incorrectly; allow the 4626 * FIN to extend beyond the window and check it in detail later. 4627 */ 4628 if (!after(end_seq - th->fin, seq_limit)) 4629 return SKB_NOT_DROPPED_YET; 4630 4631 if (after(seq, seq_limit)) 4632 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE; 4633 4634 /* Only accept this packet if receive queue is empty. */ 4635 if (skb_queue_len(&sk->sk_receive_queue)) 4636 return SKB_DROP_REASON_TCP_INVALID_END_SEQUENCE; 4637 } 4638 4639 return SKB_NOT_DROPPED_YET; 4640} 4641 4642 4643void tcp_done_with_error(struct sock *sk, int err) 4644{ 4645 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4646 WRITE_ONCE(sk->sk_err, err); 4647 smp_wmb(); 4648 4649 tcp_write_queue_purge(sk); 4650 tcp_done(sk); 4651 4652 if (!sock_flag(sk, SOCK_DEAD)) 4653 sk_error_report(sk); 4654} 4655EXPORT_IPV6_MOD(tcp_done_with_error); 4656 4657/* When we get a reset we do this. */ 4658void tcp_reset(struct sock *sk, struct sk_buff *skb) 4659{ 4660 int err; 4661 4662 trace_tcp_receive_reset(sk); 4663 4664 /* mptcp can't tell us to ignore reset pkts, 4665 * so just ignore the return value of mptcp_incoming_options(). 4666 */ 4667 if (sk_is_mptcp(sk)) 4668 mptcp_incoming_options(sk, skb); 4669 4670 /* We want the right error as BSD sees it (and indeed as we do). */ 4671 switch (sk->sk_state) { 4672 case TCP_SYN_SENT: 4673 err = ECONNREFUSED; 4674 break; 4675 case TCP_CLOSE_WAIT: 4676 err = EPIPE; 4677 break; 4678 case TCP_CLOSE: 4679 return; 4680 default: 4681 err = ECONNRESET; 4682 } 4683 tcp_done_with_error(sk, err); 4684} 4685 4686/* 4687 * Process the FIN bit. This now behaves as it is supposed to work 4688 * and the FIN takes effect when it is validly part of sequence 4689 * space. Not before when we get holes. 4690 * 4691 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4692 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4693 * TIME-WAIT) 4694 * 4695 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4696 * close and we go into CLOSING (and later onto TIME-WAIT) 4697 * 4698 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4699 */ 4700void tcp_fin(struct sock *sk) 4701{ 4702 struct tcp_sock *tp = tcp_sk(sk); 4703 4704 inet_csk_schedule_ack(sk); 4705 4706 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN); 4707 sock_set_flag(sk, SOCK_DONE); 4708 4709 switch (sk->sk_state) { 4710 case TCP_SYN_RECV: 4711 case TCP_ESTABLISHED: 4712 /* Move to CLOSE_WAIT */ 4713 tcp_set_state(sk, TCP_CLOSE_WAIT); 4714 inet_csk_enter_pingpong_mode(sk); 4715 break; 4716 4717 case TCP_CLOSE_WAIT: 4718 case TCP_CLOSING: 4719 /* Received a retransmission of the FIN, do 4720 * nothing. 4721 */ 4722 break; 4723 case TCP_LAST_ACK: 4724 /* RFC793: Remain in the LAST-ACK state. */ 4725 break; 4726 4727 case TCP_FIN_WAIT1: 4728 /* This case occurs when a simultaneous close 4729 * happens, we must ack the received FIN and 4730 * enter the CLOSING state. 4731 */ 4732 tcp_send_ack(sk); 4733 tcp_set_state(sk, TCP_CLOSING); 4734 break; 4735 case TCP_FIN_WAIT2: 4736 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4737 tcp_send_ack(sk); 4738 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4739 break; 4740 default: 4741 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 4742 * cases we should never reach this piece of code. 4743 */ 4744 pr_err("%s: Impossible, sk->sk_state=%d\n", 4745 __func__, sk->sk_state); 4746 break; 4747 } 4748 4749 /* It _is_ possible, that we have something out-of-order _after_ FIN. 4750 * Probably, we should reset in this case. For now drop them. 4751 */ 4752 skb_rbtree_purge(&tp->out_of_order_queue); 4753 if (tcp_is_sack(tp)) 4754 tcp_sack_reset(&tp->rx_opt); 4755 4756 if (!sock_flag(sk, SOCK_DEAD)) { 4757 sk->sk_state_change(sk); 4758 4759 /* Do not send POLL_HUP for half duplex close. */ 4760 if (sk->sk_shutdown == SHUTDOWN_MASK || 4761 sk->sk_state == TCP_CLOSE) 4762 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 4763 else 4764 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 4765 } 4766} 4767 4768static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 4769 u32 end_seq) 4770{ 4771 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 4772 if (before(seq, sp->start_seq)) 4773 sp->start_seq = seq; 4774 if (after(end_seq, sp->end_seq)) 4775 sp->end_seq = end_seq; 4776 return true; 4777 } 4778 return false; 4779} 4780 4781static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 4782{ 4783 struct tcp_sock *tp = tcp_sk(sk); 4784 4785 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { 4786 int mib_idx; 4787 4788 if (before(seq, tp->rcv_nxt)) 4789 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 4790 else 4791 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 4792 4793 NET_INC_STATS(sock_net(sk), mib_idx); 4794 4795 tp->rx_opt.dsack = 1; 4796 tp->duplicate_sack[0].start_seq = seq; 4797 tp->duplicate_sack[0].end_seq = end_seq; 4798 } 4799} 4800 4801static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 4802{ 4803 struct tcp_sock *tp = tcp_sk(sk); 4804 4805 if (!tp->rx_opt.dsack) 4806 tcp_dsack_set(sk, seq, end_seq); 4807 else 4808 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 4809} 4810 4811static void tcp_rcv_spurious_retrans(struct sock *sk, const struct sk_buff *skb) 4812{ 4813 /* When the ACK path fails or drops most ACKs, the sender would 4814 * timeout and spuriously retransmit the same segment repeatedly. 4815 * If it seems our ACKs are not reaching the other side, 4816 * based on receiving a duplicate data segment with new flowlabel 4817 * (suggesting the sender suffered an RTO), and we are not already 4818 * repathing due to our own RTO, then rehash the socket to repath our 4819 * packets. 4820 */ 4821#if IS_ENABLED(CONFIG_IPV6) 4822 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss && 4823 skb->protocol == htons(ETH_P_IPV6) && 4824 (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel != 4825 ntohl(ip6_flowlabel(ipv6_hdr(skb)))) && 4826 sk_rethink_txhash(sk)) 4827 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH); 4828 4829 /* Save last flowlabel after a spurious retrans. */ 4830 tcp_save_lrcv_flowlabel(sk, skb); 4831#endif 4832} 4833 4834static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 4835{ 4836 struct tcp_sock *tp = tcp_sk(sk); 4837 4838 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 4839 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 4840 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 4841 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 4842 4843 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { 4844 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 4845 4846 tcp_rcv_spurious_retrans(sk, skb); 4847 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 4848 end_seq = tp->rcv_nxt; 4849 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 4850 } 4851 } 4852 4853 tcp_send_ack(sk); 4854} 4855 4856/* These routines update the SACK block as out-of-order packets arrive or 4857 * in-order packets close up the sequence space. 4858 */ 4859static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 4860{ 4861 int this_sack; 4862 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4863 struct tcp_sack_block *swalk = sp + 1; 4864 4865 /* See if the recent change to the first SACK eats into 4866 * or hits the sequence space of other SACK blocks, if so coalesce. 4867 */ 4868 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 4869 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 4870 int i; 4871 4872 /* Zap SWALK, by moving every further SACK up by one slot. 4873 * Decrease num_sacks. 4874 */ 4875 tp->rx_opt.num_sacks--; 4876 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 4877 sp[i] = sp[i + 1]; 4878 continue; 4879 } 4880 this_sack++; 4881 swalk++; 4882 } 4883} 4884 4885void tcp_sack_compress_send_ack(struct sock *sk) 4886{ 4887 struct tcp_sock *tp = tcp_sk(sk); 4888 4889 if (!tp->compressed_ack) 4890 return; 4891 4892 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 4893 __sock_put(sk); 4894 4895 /* Since we have to send one ack finally, 4896 * substract one from tp->compressed_ack to keep 4897 * LINUX_MIB_TCPACKCOMPRESSED accurate. 4898 */ 4899 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 4900 tp->compressed_ack - 1); 4901 4902 tp->compressed_ack = 0; 4903 tcp_send_ack(sk); 4904} 4905 4906/* Reasonable amount of sack blocks included in TCP SACK option 4907 * The max is 4, but this becomes 3 if TCP timestamps are there. 4908 * Given that SACK packets might be lost, be conservative and use 2. 4909 */ 4910#define TCP_SACK_BLOCKS_EXPECTED 2 4911 4912static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 4913{ 4914 struct tcp_sock *tp = tcp_sk(sk); 4915 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4916 int cur_sacks = tp->rx_opt.num_sacks; 4917 int this_sack; 4918 4919 if (!cur_sacks) 4920 goto new_sack; 4921 4922 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 4923 if (tcp_sack_extend(sp, seq, end_seq)) { 4924 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4925 tcp_sack_compress_send_ack(sk); 4926 /* Rotate this_sack to the first one. */ 4927 for (; this_sack > 0; this_sack--, sp--) 4928 swap(*sp, *(sp - 1)); 4929 if (cur_sacks > 1) 4930 tcp_sack_maybe_coalesce(tp); 4931 return; 4932 } 4933 } 4934 4935 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 4936 tcp_sack_compress_send_ack(sk); 4937 4938 /* Could not find an adjacent existing SACK, build a new one, 4939 * put it at the front, and shift everyone else down. We 4940 * always know there is at least one SACK present already here. 4941 * 4942 * If the sack array is full, forget about the last one. 4943 */ 4944 if (this_sack >= TCP_NUM_SACKS) { 4945 this_sack--; 4946 tp->rx_opt.num_sacks--; 4947 sp--; 4948 } 4949 for (; this_sack > 0; this_sack--, sp--) 4950 *sp = *(sp - 1); 4951 4952new_sack: 4953 /* Build the new head SACK, and we're done. */ 4954 sp->start_seq = seq; 4955 sp->end_seq = end_seq; 4956 tp->rx_opt.num_sacks++; 4957} 4958 4959/* RCV.NXT advances, some SACKs should be eaten. */ 4960 4961static void tcp_sack_remove(struct tcp_sock *tp) 4962{ 4963 struct tcp_sack_block *sp = &tp->selective_acks[0]; 4964 int num_sacks = tp->rx_opt.num_sacks; 4965 int this_sack; 4966 4967 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 4968 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 4969 tp->rx_opt.num_sacks = 0; 4970 return; 4971 } 4972 4973 for (this_sack = 0; this_sack < num_sacks;) { 4974 /* Check if the start of the sack is covered by RCV.NXT. */ 4975 if (!before(tp->rcv_nxt, sp->start_seq)) { 4976 int i; 4977 4978 /* RCV.NXT must cover all the block! */ 4979 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 4980 4981 /* Zap this SACK, by moving forward any other SACKS. */ 4982 for (i = this_sack+1; i < num_sacks; i++) 4983 tp->selective_acks[i-1] = tp->selective_acks[i]; 4984 num_sacks--; 4985 continue; 4986 } 4987 this_sack++; 4988 sp++; 4989 } 4990 tp->rx_opt.num_sacks = num_sacks; 4991} 4992 4993/** 4994 * tcp_try_coalesce - try to merge skb to prior one 4995 * @sk: socket 4996 * @to: prior buffer 4997 * @from: buffer to add in queue 4998 * @fragstolen: pointer to boolean 4999 * 5000 * Before queueing skb @from after @to, try to merge them 5001 * to reduce overall memory use and queue lengths, if cost is small. 5002 * Packets in ofo or receive queues can stay a long time. 5003 * Better try to coalesce them right now to avoid future collapses. 5004 * Returns true if caller should free @from instead of queueing it 5005 */ 5006static bool tcp_try_coalesce(struct sock *sk, 5007 struct sk_buff *to, 5008 struct sk_buff *from, 5009 bool *fragstolen) 5010{ 5011 int delta; 5012 5013 *fragstolen = false; 5014 5015 /* Its possible this segment overlaps with prior segment in queue */ 5016 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq) 5017 return false; 5018 5019 if (!tcp_skb_can_collapse_rx(to, from)) 5020 return false; 5021 5022 if (!skb_try_coalesce(to, from, fragstolen, &delta)) 5023 return false; 5024 5025 atomic_add(delta, &sk->sk_rmem_alloc); 5026 sk_mem_charge(sk, delta); 5027 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); 5028 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; 5029 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; 5030 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags; 5031 5032 if (TCP_SKB_CB(from)->has_rxtstamp) { 5033 TCP_SKB_CB(to)->has_rxtstamp = true; 5034 to->tstamp = from->tstamp; 5035 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp; 5036 } 5037 5038 return true; 5039} 5040 5041static bool tcp_ooo_try_coalesce(struct sock *sk, 5042 struct sk_buff *to, 5043 struct sk_buff *from, 5044 bool *fragstolen) 5045{ 5046 bool res = tcp_try_coalesce(sk, to, from, fragstolen); 5047 5048 /* In case tcp_drop_reason() is called later, update to->gso_segs */ 5049 if (res) { 5050 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) + 5051 max_t(u16, 1, skb_shinfo(from)->gso_segs); 5052 5053 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF); 5054 } 5055 return res; 5056} 5057 5058noinline_for_tracing static void 5059tcp_drop_reason(struct sock *sk, struct sk_buff *skb, enum skb_drop_reason reason) 5060{ 5061 sk_drops_skbadd(sk, skb); 5062 sk_skb_reason_drop(sk, skb, reason); 5063} 5064 5065/* This one checks to see if we can put data from the 5066 * out_of_order queue into the receive_queue. 5067 */ 5068static void tcp_ofo_queue(struct sock *sk) 5069{ 5070 struct tcp_sock *tp = tcp_sk(sk); 5071 __u32 dsack_high = tp->rcv_nxt; 5072 bool fin, fragstolen, eaten; 5073 struct sk_buff *skb, *tail; 5074 struct rb_node *p; 5075 5076 p = rb_first(&tp->out_of_order_queue); 5077 while (p) { 5078 skb = rb_to_skb(p); 5079 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 5080 break; 5081 5082 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 5083 __u32 dsack = dsack_high; 5084 5085 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 5086 dsack = TCP_SKB_CB(skb)->end_seq; 5087 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 5088 } 5089 p = rb_next(p); 5090 rb_erase(&skb->rbnode, &tp->out_of_order_queue); 5091 5092 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) { 5093 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP); 5094 continue; 5095 } 5096 5097 tail = skb_peek_tail(&sk->sk_receive_queue); 5098 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen); 5099 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); 5100 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 5101 if (!eaten) 5102 tcp_add_receive_queue(sk, skb); 5103 else 5104 kfree_skb_partial(skb, fragstolen); 5105 5106 if (unlikely(fin)) { 5107 tcp_fin(sk); 5108 /* tcp_fin() purges tp->out_of_order_queue, 5109 * so we must end this loop right now. 5110 */ 5111 break; 5112 } 5113 } 5114} 5115 5116static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb); 5117static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb); 5118 5119static bool tcp_can_ingest(const struct sock *sk, const struct sk_buff *skb) 5120{ 5121 unsigned int rmem = atomic_read(&sk->sk_rmem_alloc); 5122 5123 return rmem <= sk->sk_rcvbuf; 5124} 5125 5126static int tcp_try_rmem_schedule(struct sock *sk, const struct sk_buff *skb, 5127 unsigned int size) 5128{ 5129 if (!tcp_can_ingest(sk, skb) || 5130 !sk_rmem_schedule(sk, skb, size)) { 5131 5132 if (tcp_prune_queue(sk, skb) < 0) 5133 return -1; 5134 5135 while (!sk_rmem_schedule(sk, skb, size)) { 5136 if (!tcp_prune_ofo_queue(sk, skb)) 5137 return -1; 5138 } 5139 } 5140 return 0; 5141} 5142 5143static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) 5144{ 5145 struct tcp_sock *tp = tcp_sk(sk); 5146 struct rb_node **p, *parent; 5147 struct sk_buff *skb1; 5148 u32 seq, end_seq; 5149 bool fragstolen; 5150 5151 tcp_save_lrcv_flowlabel(sk, skb); 5152 tcp_data_ecn_check(sk, skb); 5153 5154 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) { 5155 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP); 5156 READ_ONCE(sk->sk_data_ready)(sk); 5157 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM); 5158 return; 5159 } 5160 5161 tcp_measure_rcv_mss(sk, skb); 5162 /* Disable header prediction. */ 5163 tp->pred_flags = 0; 5164 inet_csk_schedule_ack(sk); 5165 5166 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs); 5167 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE); 5168 seq = TCP_SKB_CB(skb)->seq; 5169 end_seq = TCP_SKB_CB(skb)->end_seq; 5170 5171 p = &tp->out_of_order_queue.rb_node; 5172 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5173 /* Initial out of order segment, build 1 SACK. */ 5174 if (tcp_is_sack(tp)) { 5175 tp->rx_opt.num_sacks = 1; 5176 tp->selective_acks[0].start_seq = seq; 5177 tp->selective_acks[0].end_seq = end_seq; 5178 } 5179 rb_link_node(&skb->rbnode, NULL, p); 5180 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 5181 tp->ooo_last_skb = skb; 5182 goto end; 5183 } 5184 5185 /* In the typical case, we are adding an skb to the end of the list. 5186 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 5187 */ 5188 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb, 5189 skb, &fragstolen)) { 5190coalesce_done: 5191 /* For non sack flows, do not grow window to force DUPACK 5192 * and trigger fast retransmit. 5193 */ 5194 if (tcp_is_sack(tp)) 5195 tcp_grow_window(sk, skb, true); 5196 kfree_skb_partial(skb, fragstolen); 5197 skb = NULL; 5198 goto add_sack; 5199 } 5200 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 5201 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) { 5202 parent = &tp->ooo_last_skb->rbnode; 5203 p = &parent->rb_right; 5204 goto insert; 5205 } 5206 5207 /* Find place to insert this segment. Handle overlaps on the way. */ 5208 parent = NULL; 5209 while (*p) { 5210 parent = *p; 5211 skb1 = rb_to_skb(parent); 5212 if (before(seq, TCP_SKB_CB(skb1)->seq)) { 5213 p = &parent->rb_left; 5214 continue; 5215 } 5216 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) { 5217 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 5218 /* All the bits are present. Drop. */ 5219 NET_INC_STATS(sock_net(sk), 5220 LINUX_MIB_TCPOFOMERGE); 5221 tcp_drop_reason(sk, skb, 5222 SKB_DROP_REASON_TCP_OFOMERGE); 5223 skb = NULL; 5224 tcp_dsack_set(sk, seq, end_seq); 5225 goto add_sack; 5226 } 5227 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 5228 /* Partial overlap. */ 5229 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq); 5230 } else { 5231 /* skb's seq == skb1's seq and skb covers skb1. 5232 * Replace skb1 with skb. 5233 */ 5234 rb_replace_node(&skb1->rbnode, &skb->rbnode, 5235 &tp->out_of_order_queue); 5236 tcp_dsack_extend(sk, 5237 TCP_SKB_CB(skb1)->seq, 5238 TCP_SKB_CB(skb1)->end_seq); 5239 NET_INC_STATS(sock_net(sk), 5240 LINUX_MIB_TCPOFOMERGE); 5241 tcp_drop_reason(sk, skb1, 5242 SKB_DROP_REASON_TCP_OFOMERGE); 5243 goto merge_right; 5244 } 5245 } else if (tcp_ooo_try_coalesce(sk, skb1, 5246 skb, &fragstolen)) { 5247 goto coalesce_done; 5248 } 5249 p = &parent->rb_right; 5250 } 5251insert: 5252 /* Insert segment into RB tree. */ 5253 rb_link_node(&skb->rbnode, parent, p); 5254 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 5255 5256merge_right: 5257 /* Remove other segments covered by skb. */ 5258 while ((skb1 = skb_rb_next(skb)) != NULL) { 5259 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 5260 break; 5261 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 5262 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 5263 end_seq); 5264 break; 5265 } 5266 rb_erase(&skb1->rbnode, &tp->out_of_order_queue); 5267 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 5268 TCP_SKB_CB(skb1)->end_seq); 5269 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE); 5270 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE); 5271 } 5272 /* If there is no skb after us, we are the last_skb ! */ 5273 if (!skb1) 5274 tp->ooo_last_skb = skb; 5275 5276add_sack: 5277 if (tcp_is_sack(tp)) 5278 tcp_sack_new_ofo_skb(sk, seq, end_seq); 5279end: 5280 if (skb) { 5281 /* For non sack flows, do not grow window to force DUPACK 5282 * and trigger fast retransmit. 5283 */ 5284 if (tcp_is_sack(tp)) 5285 tcp_grow_window(sk, skb, false); 5286 skb_condense(skb); 5287 skb_set_owner_r(skb, sk); 5288 } 5289 /* do not grow rcvbuf for not-yet-accepted or orphaned sockets. */ 5290 if (sk->sk_socket) 5291 tcp_rcvbuf_grow(sk, tp->rcvq_space.space); 5292} 5293 5294static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, 5295 bool *fragstolen) 5296{ 5297 int eaten; 5298 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue); 5299 5300 eaten = (tail && 5301 tcp_try_coalesce(sk, tail, 5302 skb, fragstolen)) ? 1 : 0; 5303 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq); 5304 if (!eaten) { 5305 tcp_add_receive_queue(sk, skb); 5306 skb_set_owner_r(skb, sk); 5307 } 5308 return eaten; 5309} 5310 5311int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) 5312{ 5313 struct sk_buff *skb; 5314 int err = -ENOMEM; 5315 int data_len = 0; 5316 bool fragstolen; 5317 5318 if (size == 0) 5319 return 0; 5320 5321 if (size > PAGE_SIZE) { 5322 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS); 5323 5324 data_len = npages << PAGE_SHIFT; 5325 size = data_len + (size & ~PAGE_MASK); 5326 } 5327 skb = alloc_skb_with_frags(size - data_len, data_len, 5328 PAGE_ALLOC_COSTLY_ORDER, 5329 &err, sk->sk_allocation); 5330 if (!skb) 5331 goto err; 5332 5333 skb_put(skb, size - data_len); 5334 skb->data_len = data_len; 5335 skb->len = size; 5336 5337 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 5338 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 5339 goto err_free; 5340 } 5341 5342 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size); 5343 if (err) 5344 goto err_free; 5345 5346 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; 5347 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; 5348 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; 5349 5350 if (tcp_queue_rcv(sk, skb, &fragstolen)) { 5351 WARN_ON_ONCE(fragstolen); /* should not happen */ 5352 __kfree_skb(skb); 5353 } 5354 return size; 5355 5356err_free: 5357 kfree_skb(skb); 5358err: 5359 return err; 5360 5361} 5362 5363void tcp_data_ready(struct sock *sk) 5364{ 5365 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE)) 5366 READ_ONCE(sk->sk_data_ready)(sk); 5367} 5368 5369static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 5370{ 5371 struct tcp_sock *tp = tcp_sk(sk); 5372 enum skb_drop_reason reason; 5373 bool fragstolen; 5374 int eaten; 5375 5376 /* If a subflow has been reset, the packet should not continue 5377 * to be processed, drop the packet. 5378 */ 5379 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) { 5380 __kfree_skb(skb); 5381 return; 5382 } 5383 5384 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { 5385 __kfree_skb(skb); 5386 return; 5387 } 5388 tcp_cleanup_skb(skb); 5389 __skb_pull(skb, tcp_hdr(skb)->doff * 4); 5390 5391 reason = SKB_DROP_REASON_NOT_SPECIFIED; 5392 tp->rx_opt.dsack = 0; 5393 5394 /* Queue data for delivery to the user. 5395 * Packets in sequence go to the receive queue. 5396 * Out of sequence packets to the out_of_order_queue. 5397 */ 5398 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 5399 if (tcp_receive_window(tp) == 0) { 5400 /* Some stacks are known to send bare FIN packets 5401 * in a loop even if we send RWIN 0 in our ACK. 5402 * Accepting this FIN does not hurt memory pressure 5403 * because the FIN flag will simply be merged to the 5404 * receive queue tail skb in most cases. 5405 */ 5406 if (!skb->len && 5407 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 5408 goto queue_and_out; 5409 5410 reason = SKB_DROP_REASON_TCP_ZEROWINDOW; 5411 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5412 goto out_of_window; 5413 } 5414 5415 /* Ok. In sequence. In window. */ 5416queue_and_out: 5417 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 5418 /* TODO: maybe ratelimit these WIN 0 ACK ? */ 5419 inet_csk(sk)->icsk_ack.pending |= 5420 (ICSK_ACK_NOMEM | ICSK_ACK_NOW); 5421 inet_csk_schedule_ack(sk); 5422 READ_ONCE(sk->sk_data_ready)(sk); 5423 5424 if (skb_queue_len(&sk->sk_receive_queue) && skb->len) { 5425 reason = SKB_DROP_REASON_PROTO_MEM; 5426 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 5427 goto drop; 5428 } 5429 sk_forced_mem_schedule(sk, skb->truesize); 5430 } 5431 5432 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 5433 if (skb->len) 5434 tcp_event_data_recv(sk, skb); 5435 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 5436 tcp_fin(sk); 5437 5438 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5439 tcp_ofo_queue(sk); 5440 5441 /* RFC5681. 4.2. SHOULD send immediate ACK, when 5442 * gap in queue is filled. 5443 */ 5444 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5445 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 5446 } 5447 5448 if (tp->rx_opt.num_sacks) 5449 tcp_sack_remove(tp); 5450 5451 tcp_fast_path_check(sk); 5452 5453 if (eaten > 0) 5454 kfree_skb_partial(skb, fragstolen); 5455 if (!sock_flag(sk, SOCK_DEAD)) 5456 tcp_data_ready(sk); 5457 return; 5458 } 5459 5460 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 5461 tcp_rcv_spurious_retrans(sk, skb); 5462 /* A retransmit, 2nd most common case. Force an immediate ack. */ 5463 reason = SKB_DROP_REASON_TCP_OLD_DATA; 5464 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 5465 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 5466 5467out_of_window: 5468 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 5469 inet_csk_schedule_ack(sk); 5470drop: 5471 tcp_drop_reason(sk, skb, reason); 5472 return; 5473 } 5474 5475 /* Out of window. F.e. zero window probe. */ 5476 if (!before(TCP_SKB_CB(skb)->seq, 5477 tp->rcv_nxt + tcp_receive_window(tp))) { 5478 reason = SKB_DROP_REASON_TCP_OVERWINDOW; 5479 goto out_of_window; 5480 } 5481 5482 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5483 /* Partial packet, seq < rcv_next < end_seq */ 5484 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 5485 5486 /* If window is closed, drop tail of packet. But after 5487 * remembering D-SACK for its head made in previous line. 5488 */ 5489 if (!tcp_receive_window(tp)) { 5490 reason = SKB_DROP_REASON_TCP_ZEROWINDOW; 5491 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5492 goto out_of_window; 5493 } 5494 goto queue_and_out; 5495 } 5496 5497 tcp_data_queue_ofo(sk, skb); 5498} 5499 5500static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list) 5501{ 5502 if (list) 5503 return !skb_queue_is_last(list, skb) ? skb->next : NULL; 5504 5505 return skb_rb_next(skb); 5506} 5507 5508static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 5509 struct sk_buff_head *list, 5510 struct rb_root *root) 5511{ 5512 struct sk_buff *next = tcp_skb_next(skb, list); 5513 5514 if (list) 5515 __skb_unlink(skb, list); 5516 else 5517 rb_erase(&skb->rbnode, root); 5518 5519 __kfree_skb(skb); 5520 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 5521 5522 return next; 5523} 5524 5525/* Insert skb into rb tree, ordered by TCP_SKB_CB(skb)->seq */ 5526void tcp_rbtree_insert(struct rb_root *root, struct sk_buff *skb) 5527{ 5528 struct rb_node **p = &root->rb_node; 5529 struct rb_node *parent = NULL; 5530 struct sk_buff *skb1; 5531 5532 while (*p) { 5533 parent = *p; 5534 skb1 = rb_to_skb(parent); 5535 if (before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb1)->seq)) 5536 p = &parent->rb_left; 5537 else 5538 p = &parent->rb_right; 5539 } 5540 rb_link_node(&skb->rbnode, parent, p); 5541 rb_insert_color(&skb->rbnode, root); 5542} 5543 5544/* Collapse contiguous sequence of skbs head..tail with 5545 * sequence numbers start..end. 5546 * 5547 * If tail is NULL, this means until the end of the queue. 5548 * 5549 * Segments with FIN/SYN are not collapsed (only because this 5550 * simplifies code) 5551 */ 5552static void 5553tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root, 5554 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end) 5555{ 5556 struct sk_buff *skb = head, *n; 5557 struct sk_buff_head tmp; 5558 bool end_of_skbs; 5559 5560 /* First, check that queue is collapsible and find 5561 * the point where collapsing can be useful. 5562 */ 5563restart: 5564 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) { 5565 n = tcp_skb_next(skb, list); 5566 5567 if (!skb_frags_readable(skb)) 5568 goto skip_this; 5569 5570 /* No new bits? It is possible on ofo queue. */ 5571 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5572 skb = tcp_collapse_one(sk, skb, list, root); 5573 if (!skb) 5574 break; 5575 goto restart; 5576 } 5577 5578 /* The first skb to collapse is: 5579 * - not SYN/FIN and 5580 * - bloated or contains data before "start" or 5581 * overlaps to the next one and mptcp allow collapsing. 5582 */ 5583 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) && 5584 (tcp_win_from_space(sk, skb->truesize) > skb->len || 5585 before(TCP_SKB_CB(skb)->seq, start))) { 5586 end_of_skbs = false; 5587 break; 5588 } 5589 5590 if (n && n != tail && skb_frags_readable(n) && 5591 tcp_skb_can_collapse_rx(skb, n) && 5592 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) { 5593 end_of_skbs = false; 5594 break; 5595 } 5596 5597skip_this: 5598 /* Decided to skip this, advance start seq. */ 5599 start = TCP_SKB_CB(skb)->end_seq; 5600 } 5601 if (end_of_skbs || 5602 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) || 5603 !skb_frags_readable(skb)) 5604 return; 5605 5606 __skb_queue_head_init(&tmp); 5607 5608 while (before(start, end)) { 5609 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start); 5610 struct sk_buff *nskb; 5611 5612 nskb = alloc_skb(copy, GFP_ATOMIC); 5613 if (!nskb) 5614 break; 5615 5616 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 5617 skb_copy_decrypted(nskb, skb); 5618 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 5619 if (list) 5620 __skb_queue_before(list, skb, nskb); 5621 else 5622 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */ 5623 skb_set_owner_r(nskb, sk); 5624 mptcp_skb_ext_move(nskb, skb); 5625 5626 /* Copy data, releasing collapsed skbs. */ 5627 while (copy > 0) { 5628 int offset = start - TCP_SKB_CB(skb)->seq; 5629 int size = TCP_SKB_CB(skb)->end_seq - start; 5630 5631 BUG_ON(offset < 0); 5632 if (size > 0) { 5633 size = min(copy, size); 5634 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 5635 BUG(); 5636 TCP_SKB_CB(nskb)->end_seq += size; 5637 copy -= size; 5638 start += size; 5639 } 5640 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5641 skb = tcp_collapse_one(sk, skb, list, root); 5642 if (!skb || 5643 skb == tail || 5644 !tcp_skb_can_collapse_rx(nskb, skb) || 5645 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) || 5646 !skb_frags_readable(skb)) 5647 goto end; 5648 } 5649 } 5650 } 5651end: 5652 skb_queue_walk_safe(&tmp, skb, n) 5653 tcp_rbtree_insert(root, skb); 5654} 5655 5656/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 5657 * and tcp_collapse() them until all the queue is collapsed. 5658 */ 5659static void tcp_collapse_ofo_queue(struct sock *sk) 5660{ 5661 struct tcp_sock *tp = tcp_sk(sk); 5662 u32 range_truesize, sum_tiny = 0; 5663 struct sk_buff *skb, *head; 5664 u32 start, end; 5665 5666 skb = skb_rb_first(&tp->out_of_order_queue); 5667new_range: 5668 if (!skb) { 5669 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue); 5670 return; 5671 } 5672 start = TCP_SKB_CB(skb)->seq; 5673 end = TCP_SKB_CB(skb)->end_seq; 5674 range_truesize = skb->truesize; 5675 5676 for (head = skb;;) { 5677 skb = skb_rb_next(skb); 5678 5679 /* Range is terminated when we see a gap or when 5680 * we are at the queue end. 5681 */ 5682 if (!skb || 5683 after(TCP_SKB_CB(skb)->seq, end) || 5684 before(TCP_SKB_CB(skb)->end_seq, start)) { 5685 /* Do not attempt collapsing tiny skbs */ 5686 if (range_truesize != head->truesize || 5687 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) { 5688 tcp_collapse(sk, NULL, &tp->out_of_order_queue, 5689 head, skb, start, end); 5690 } else { 5691 sum_tiny += range_truesize; 5692 if (sum_tiny > sk->sk_rcvbuf >> 3) 5693 return; 5694 } 5695 goto new_range; 5696 } 5697 5698 range_truesize += skb->truesize; 5699 if (unlikely(before(TCP_SKB_CB(skb)->seq, start))) 5700 start = TCP_SKB_CB(skb)->seq; 5701 if (after(TCP_SKB_CB(skb)->end_seq, end)) 5702 end = TCP_SKB_CB(skb)->end_seq; 5703 } 5704} 5705 5706/* 5707 * Clean the out-of-order queue to make room. 5708 * We drop high sequences packets to : 5709 * 1) Let a chance for holes to be filled. 5710 * This means we do not drop packets from ooo queue if their sequence 5711 * is before incoming packet sequence. 5712 * 2) not add too big latencies if thousands of packets sit there. 5713 * (But if application shrinks SO_RCVBUF, we could still end up 5714 * freeing whole queue here) 5715 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks. 5716 * 5717 * Return true if queue has shrunk. 5718 */ 5719static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb) 5720{ 5721 struct tcp_sock *tp = tcp_sk(sk); 5722 struct rb_node *node, *prev; 5723 bool pruned = false; 5724 int goal; 5725 5726 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5727 return false; 5728 5729 goal = sk->sk_rcvbuf >> 3; 5730 node = &tp->ooo_last_skb->rbnode; 5731 5732 do { 5733 struct sk_buff *skb = rb_to_skb(node); 5734 5735 /* If incoming skb would land last in ofo queue, stop pruning. */ 5736 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq)) 5737 break; 5738 pruned = true; 5739 prev = rb_prev(node); 5740 rb_erase(node, &tp->out_of_order_queue); 5741 goal -= skb->truesize; 5742 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE); 5743 tp->ooo_last_skb = rb_to_skb(prev); 5744 if (!prev || goal <= 0) { 5745 if (tcp_can_ingest(sk, in_skb) && 5746 !tcp_under_memory_pressure(sk)) 5747 break; 5748 goal = sk->sk_rcvbuf >> 3; 5749 } 5750 node = prev; 5751 } while (node); 5752 5753 if (pruned) { 5754 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED); 5755 /* Reset SACK state. A conforming SACK implementation will 5756 * do the same at a timeout based retransmit. When a connection 5757 * is in a sad state like this, we care only about integrity 5758 * of the connection not performance. 5759 */ 5760 if (tp->rx_opt.sack_ok) 5761 tcp_sack_reset(&tp->rx_opt); 5762 } 5763 return pruned; 5764} 5765 5766/* Reduce allocated memory if we can, trying to get 5767 * the socket within its memory limits again. 5768 * 5769 * Return less than zero if we should start dropping frames 5770 * until the socket owning process reads some of the data 5771 * to stabilize the situation. 5772 */ 5773static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb) 5774{ 5775 struct tcp_sock *tp = tcp_sk(sk); 5776 5777 /* Do nothing if our queues are empty. */ 5778 if (!atomic_read(&sk->sk_rmem_alloc)) 5779 return -1; 5780 5781 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED); 5782 5783 if (!tcp_can_ingest(sk, in_skb)) 5784 tcp_clamp_window(sk); 5785 else if (tcp_under_memory_pressure(sk)) 5786 tcp_adjust_rcv_ssthresh(sk); 5787 5788 if (tcp_can_ingest(sk, in_skb)) 5789 return 0; 5790 5791 tcp_collapse_ofo_queue(sk); 5792 if (!skb_queue_empty(&sk->sk_receive_queue)) 5793 tcp_collapse(sk, &sk->sk_receive_queue, NULL, 5794 skb_peek(&sk->sk_receive_queue), 5795 NULL, 5796 tp->copied_seq, tp->rcv_nxt); 5797 5798 if (tcp_can_ingest(sk, in_skb)) 5799 return 0; 5800 5801 /* Collapsing did not help, destructive actions follow. 5802 * This must not ever occur. */ 5803 5804 tcp_prune_ofo_queue(sk, in_skb); 5805 5806 if (tcp_can_ingest(sk, in_skb)) 5807 return 0; 5808 5809 /* If we are really being abused, tell the caller to silently 5810 * drop receive data on the floor. It will get retransmitted 5811 * and hopefully then we'll have sufficient space. 5812 */ 5813 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED); 5814 5815 /* Massive buffer overcommit. */ 5816 tp->pred_flags = 0; 5817 return -1; 5818} 5819 5820static bool tcp_should_expand_sndbuf(struct sock *sk) 5821{ 5822 const struct tcp_sock *tp = tcp_sk(sk); 5823 5824 /* If the user specified a specific send buffer setting, do 5825 * not modify it. 5826 */ 5827 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 5828 return false; 5829 5830 /* If we are under global TCP memory pressure, do not expand. */ 5831 if (tcp_under_memory_pressure(sk)) { 5832 int unused_mem = sk_unused_reserved_mem(sk); 5833 5834 /* Adjust sndbuf according to reserved mem. But make sure 5835 * it never goes below SOCK_MIN_SNDBUF. 5836 * See sk_stream_moderate_sndbuf() for more details. 5837 */ 5838 if (unused_mem > SOCK_MIN_SNDBUF) 5839 WRITE_ONCE(sk->sk_sndbuf, unused_mem); 5840 5841 return false; 5842 } 5843 5844 /* If we are under soft global TCP memory pressure, do not expand. */ 5845 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 5846 return false; 5847 5848 /* If we filled the congestion window, do not expand. */ 5849 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)) 5850 return false; 5851 5852 return true; 5853} 5854 5855static void tcp_new_space(struct sock *sk) 5856{ 5857 struct tcp_sock *tp = tcp_sk(sk); 5858 5859 if (tcp_should_expand_sndbuf(sk)) { 5860 tcp_sndbuf_expand(sk); 5861 tp->snd_cwnd_stamp = tcp_jiffies32; 5862 } 5863 5864 INDIRECT_CALL_1(READ_ONCE(sk->sk_write_space), 5865 sk_stream_write_space, 5866 sk); 5867} 5868 5869/* Caller made space either from: 5870 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced) 5871 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt) 5872 * 5873 * We might be able to generate EPOLLOUT to the application if: 5874 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2 5875 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became 5876 * small enough that tcp_stream_memory_free() decides it 5877 * is time to generate EPOLLOUT. 5878 */ 5879void tcp_check_space(struct sock *sk) 5880{ 5881 /* pairs with tcp_poll() */ 5882 smp_mb(); 5883 if (sk->sk_socket && 5884 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 5885 tcp_new_space(sk); 5886 if (!test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) 5887 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 5888 } 5889} 5890 5891static inline void tcp_data_snd_check(struct sock *sk) 5892{ 5893 tcp_push_pending_frames(sk); 5894 tcp_check_space(sk); 5895} 5896 5897/* 5898 * Check if sending an ack is needed. 5899 */ 5900static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 5901{ 5902 struct tcp_sock *tp = tcp_sk(sk); 5903 struct net *net = sock_net(sk); 5904 unsigned long rtt; 5905 u64 delay; 5906 5907 /* More than one full frame received... */ 5908 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 5909 /* ... and right edge of window advances far enough. 5910 * (tcp_recvmsg() will send ACK otherwise). 5911 * If application uses SO_RCVLOWAT, we want send ack now if 5912 * we have not received enough bytes to satisfy the condition. 5913 */ 5914 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat || 5915 __tcp_select_window(sk) >= tp->rcv_wnd)) || 5916 /* We ACK each frame or... */ 5917 tcp_in_quickack_mode(sk) || 5918 /* Protocol state mandates a one-time immediate ACK */ 5919 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) { 5920 /* If we are running from __release_sock() in user context, 5921 * Defer the ack until tcp_release_cb(). 5922 */ 5923 if (sock_owned_by_user_nocheck(sk) && 5924 READ_ONCE(net->ipv4.sysctl_tcp_backlog_ack_defer)) { 5925 set_bit(TCP_ACK_DEFERRED, &sk->sk_tsq_flags); 5926 return; 5927 } 5928send_now: 5929 tcp_send_ack(sk); 5930 return; 5931 } 5932 5933 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5934 tcp_send_delayed_ack(sk); 5935 return; 5936 } 5937 5938 if (!tcp_is_sack(tp) || 5939 tp->compressed_ack >= READ_ONCE(net->ipv4.sysctl_tcp_comp_sack_nr)) 5940 goto send_now; 5941 5942 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) { 5943 tp->compressed_ack_rcv_nxt = tp->rcv_nxt; 5944 tp->dup_ack_counter = 0; 5945 } 5946 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) { 5947 tp->dup_ack_counter++; 5948 goto send_now; 5949 } 5950 tp->compressed_ack++; 5951 if (hrtimer_is_queued(&tp->compressed_ack_timer)) 5952 return; 5953 5954 /* compress ack timer : comp_sack_rtt_percent of rtt, 5955 * but no more than tcp_comp_sack_delay_ns. 5956 */ 5957 5958 rtt = tp->rcv_rtt_est.rtt_us; 5959 if (tp->srtt_us && tp->srtt_us < rtt) 5960 rtt = tp->srtt_us; 5961 5962 /* delay = (rtt >> 3) * NSEC_PER_USEC * comp_sack_rtt_percent / 100 5963 * -> 5964 * delay = rtt * 1.25 * comp_sack_rtt_percent 5965 */ 5966 delay = (u64)(rtt + (rtt >> 2)) * 5967 READ_ONCE(net->ipv4.sysctl_tcp_comp_sack_rtt_percent); 5968 5969 delay = min(delay, READ_ONCE(net->ipv4.sysctl_tcp_comp_sack_delay_ns)); 5970 5971 sock_hold(sk); 5972 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay), 5973 READ_ONCE(net->ipv4.sysctl_tcp_comp_sack_slack_ns), 5974 HRTIMER_MODE_REL_PINNED_SOFT); 5975} 5976 5977static inline void tcp_ack_snd_check(struct sock *sk) 5978{ 5979 if (!inet_csk_ack_scheduled(sk)) { 5980 /* We sent a data segment already. */ 5981 return; 5982 } 5983 __tcp_ack_snd_check(sk, 1); 5984} 5985 5986/* 5987 * This routine is only called when we have urgent data 5988 * signaled. Its the 'slow' part of tcp_urg. It could be 5989 * moved inline now as tcp_urg is only called from one 5990 * place. We handle URGent data wrong. We have to - as 5991 * BSD still doesn't use the correction from RFC961. 5992 * For 1003.1g we should support a new option TCP_STDURG to permit 5993 * either form (or just set the sysctl tcp_stdurg). 5994 */ 5995 5996static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 5997{ 5998 struct tcp_sock *tp = tcp_sk(sk); 5999 u32 ptr = ntohs(th->urg_ptr); 6000 6001 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg)) 6002 ptr--; 6003 ptr += ntohl(th->seq); 6004 6005 /* Ignore urgent data that we've already seen and read. */ 6006 if (after(tp->copied_seq, ptr)) 6007 return; 6008 6009 /* Do not replay urg ptr. 6010 * 6011 * NOTE: interesting situation not covered by specs. 6012 * Misbehaving sender may send urg ptr, pointing to segment, 6013 * which we already have in ofo queue. We are not able to fetch 6014 * such data and will stay in TCP_URG_NOTYET until will be eaten 6015 * by recvmsg(). Seems, we are not obliged to handle such wicked 6016 * situations. But it is worth to think about possibility of some 6017 * DoSes using some hypothetical application level deadlock. 6018 */ 6019 if (before(ptr, tp->rcv_nxt)) 6020 return; 6021 6022 /* Do we already have a newer (or duplicate) urgent pointer? */ 6023 if (tp->urg_data && !after(ptr, tp->urg_seq)) 6024 return; 6025 6026 /* Tell the world about our new urgent pointer. */ 6027 sk_send_sigurg(sk); 6028 6029 /* We may be adding urgent data when the last byte read was 6030 * urgent. To do this requires some care. We cannot just ignore 6031 * tp->copied_seq since we would read the last urgent byte again 6032 * as data, nor can we alter copied_seq until this data arrives 6033 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 6034 * 6035 * NOTE. Double Dutch. Rendering to plain English: author of comment 6036 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 6037 * and expect that both A and B disappear from stream. This is _wrong_. 6038 * Though this happens in BSD with high probability, this is occasional. 6039 * Any application relying on this is buggy. Note also, that fix "works" 6040 * only in this artificial test. Insert some normal data between A and B and we will 6041 * decline of BSD again. Verdict: it is better to remove to trap 6042 * buggy users. 6043 */ 6044 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 6045 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 6046 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 6047 tp->copied_seq++; 6048 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 6049 __skb_unlink(skb, &sk->sk_receive_queue); 6050 __kfree_skb(skb); 6051 } 6052 } 6053 6054 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET); 6055 WRITE_ONCE(tp->urg_seq, ptr); 6056 6057 /* Disable header prediction. */ 6058 tp->pred_flags = 0; 6059} 6060 6061/* This is the 'fast' part of urgent handling. */ 6062static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 6063{ 6064 struct tcp_sock *tp = tcp_sk(sk); 6065 6066 /* Check if we get a new urgent pointer - normally not. */ 6067 if (unlikely(th->urg)) 6068 tcp_check_urg(sk, th); 6069 6070 /* Do we wait for any urgent data? - normally not... */ 6071 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) { 6072 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 6073 th->syn; 6074 6075 /* Is the urgent pointer pointing into this packet? */ 6076 if (ptr < skb->len) { 6077 u8 tmp; 6078 if (skb_copy_bits(skb, ptr, &tmp, 1)) 6079 BUG(); 6080 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp); 6081 if (!sock_flag(sk, SOCK_DEAD)) 6082 READ_ONCE(sk->sk_data_ready)(sk); 6083 } 6084 } 6085} 6086 6087/* Accept RST for rcv_nxt - 1 after a FIN. 6088 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a 6089 * FIN is sent followed by a RST packet. The RST is sent with the same 6090 * sequence number as the FIN, and thus according to RFC 5961 a challenge 6091 * ACK should be sent. However, Mac OSX rate limits replies to challenge 6092 * ACKs on the closed socket. In addition middleboxes can drop either the 6093 * challenge ACK or a subsequent RST. 6094 */ 6095static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb) 6096{ 6097 const struct tcp_sock *tp = tcp_sk(sk); 6098 6099 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) && 6100 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK | 6101 TCPF_CLOSING)); 6102} 6103 6104/* Does PAWS and seqno based validation of an incoming segment, flags will 6105 * play significant role here. 6106 */ 6107static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 6108 const struct tcphdr *th, int syn_inerr) 6109{ 6110 struct tcp_sock *tp = tcp_sk(sk); 6111 bool accecn_reflector = false; 6112 SKB_DR(reason); 6113 6114 /* RFC1323: H1. Apply PAWS check first. */ 6115 if (!tcp_fast_parse_options(sock_net(sk), skb, th, tp) || 6116 !tp->rx_opt.saw_tstamp || 6117 tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW)) 6118 goto step1; 6119 6120 reason = tcp_disordered_ack_check(sk, skb); 6121 if (!reason) 6122 goto step1; 6123 /* Reset is accepted even if it did not pass PAWS. */ 6124 if (th->rst) 6125 goto step1; 6126 if (unlikely(th->syn)) 6127 goto syn_challenge; 6128 6129 /* Old ACK are common, increment PAWS_OLD_ACK 6130 * and do not send a dupack. 6131 */ 6132 if (reason == SKB_DROP_REASON_TCP_RFC7323_PAWS_ACK) { 6133 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWS_OLD_ACK); 6134 goto discard; 6135 } 6136 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 6137 if (!tcp_oow_rate_limited(sock_net(sk), skb, 6138 LINUX_MIB_TCPACKSKIPPEDPAWS, 6139 &tp->last_oow_ack_time)) 6140 tcp_send_dupack(sk, skb); 6141 goto discard; 6142 6143step1: 6144 /* Step 1: check sequence number */ 6145 reason = tcp_sequence(sk, TCP_SKB_CB(skb)->seq, 6146 TCP_SKB_CB(skb)->end_seq, th); 6147 if (reason) { 6148 /* RFC793, page 37: "In all states except SYN-SENT, all reset 6149 * (RST) segments are validated by checking their SEQ-fields." 6150 * And page 69: "If an incoming segment is not acceptable, 6151 * an acknowledgment should be sent in reply (unless the RST 6152 * bit is set, if so drop the segment and return)". 6153 */ 6154 if (!th->rst) { 6155 if (th->syn) 6156 goto syn_challenge; 6157 6158 if (reason == SKB_DROP_REASON_TCP_INVALID_SEQUENCE || 6159 reason == SKB_DROP_REASON_TCP_INVALID_END_SEQUENCE) 6160 NET_INC_STATS(sock_net(sk), 6161 LINUX_MIB_BEYOND_WINDOW); 6162 if (!tcp_oow_rate_limited(sock_net(sk), skb, 6163 LINUX_MIB_TCPACKSKIPPEDSEQ, 6164 &tp->last_oow_ack_time)) 6165 tcp_send_dupack(sk, skb); 6166 } else if (tcp_reset_check(sk, skb)) { 6167 goto reset; 6168 } 6169 goto discard; 6170 } 6171 6172 /* Step 2: check RST bit */ 6173 if (th->rst) { 6174 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a 6175 * FIN and SACK too if available): 6176 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or 6177 * the right-most SACK block, 6178 * then 6179 * RESET the connection 6180 * else 6181 * Send a challenge ACK 6182 */ 6183 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt || 6184 tcp_reset_check(sk, skb)) 6185 goto reset; 6186 6187 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) { 6188 struct tcp_sack_block *sp = &tp->selective_acks[0]; 6189 int max_sack = sp[0].end_seq; 6190 int this_sack; 6191 6192 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; 6193 ++this_sack) { 6194 max_sack = after(sp[this_sack].end_seq, 6195 max_sack) ? 6196 sp[this_sack].end_seq : max_sack; 6197 } 6198 6199 if (TCP_SKB_CB(skb)->seq == max_sack) 6200 goto reset; 6201 } 6202 6203 /* Disable TFO if RST is out-of-order 6204 * and no data has been received 6205 * for current active TFO socket 6206 */ 6207 if (tp->syn_fastopen && !tp->data_segs_in && 6208 sk->sk_state == TCP_ESTABLISHED) 6209 tcp_fastopen_active_disable(sk); 6210 tcp_send_challenge_ack(sk, false); 6211 SKB_DR_SET(reason, TCP_RESET); 6212 goto discard; 6213 } 6214 6215 /* step 3: check security and precedence [ignored] */ 6216 6217 /* step 4: Check for a SYN 6218 * RFC 5961 4.2 : Send a challenge ack 6219 */ 6220 if (th->syn) { 6221 if (tcp_ecn_mode_accecn(tp)) { 6222 accecn_reflector = true; 6223 if (tp->rx_opt.accecn && 6224 tp->saw_accecn_opt < TCP_ACCECN_OPT_COUNTER_SEEN) { 6225 u8 saw_opt = tcp_accecn_option_init(skb, tp->rx_opt.accecn); 6226 6227 tcp_accecn_saw_opt_fail_recv(tp, saw_opt); 6228 tcp_accecn_opt_demand_min(sk, 1); 6229 } 6230 } 6231 if (sk->sk_state == TCP_SYN_RECV && sk->sk_socket && th->ack && 6232 TCP_SKB_CB(skb)->seq + 1 == TCP_SKB_CB(skb)->end_seq && 6233 TCP_SKB_CB(skb)->seq + 1 == tp->rcv_nxt && 6234 TCP_SKB_CB(skb)->ack_seq == tp->snd_nxt) 6235 goto pass; 6236syn_challenge: 6237 if (syn_inerr) 6238 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6239 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE); 6240 tcp_send_challenge_ack(sk, accecn_reflector); 6241 SKB_DR_SET(reason, TCP_INVALID_SYN); 6242 goto discard; 6243 } 6244 6245pass: 6246 bpf_skops_parse_hdr(sk, skb); 6247 6248 return true; 6249 6250discard: 6251 tcp_drop_reason(sk, skb, reason); 6252 return false; 6253 6254reset: 6255 tcp_reset(sk, skb); 6256 __kfree_skb(skb); 6257 return false; 6258} 6259 6260/* 6261 * TCP receive function for the ESTABLISHED state. 6262 * 6263 * It is split into a fast path and a slow path. The fast path is 6264 * disabled when: 6265 * - A zero window was announced from us - zero window probing 6266 * is only handled properly in the slow path. 6267 * - Out of order segments arrived. 6268 * - Urgent data is expected. 6269 * - There is no buffer space left 6270 * - Unexpected TCP flags/window values/header lengths are received 6271 * (detected by checking the TCP header against pred_flags) 6272 * - Data is sent in both directions. Fast path only supports pure senders 6273 * or pure receivers (this means either the sequence number or the ack 6274 * value must stay constant) 6275 * - Unexpected TCP option. 6276 * 6277 * When these conditions are not satisfied it drops into a standard 6278 * receive procedure patterned after RFC793 to handle all cases. 6279 * The first three cases are guaranteed by proper pred_flags setting, 6280 * the rest is checked inline. Fast processing is turned on in 6281 * tcp_data_queue when everything is OK. 6282 */ 6283void tcp_rcv_established(struct sock *sk, struct sk_buff *skb) 6284{ 6285 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED; 6286 const struct tcphdr *th = (const struct tcphdr *)skb->data; 6287 struct tcp_sock *tp = tcp_sk(sk); 6288 unsigned int len = skb->len; 6289 6290 /* TCP congestion window tracking */ 6291 trace_tcp_probe(sk, skb); 6292 6293 tcp_mstamp_refresh(tp); 6294 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst))) 6295 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb); 6296 /* 6297 * Header prediction. 6298 * The code loosely follows the one in the famous 6299 * "30 instruction TCP receive" Van Jacobson mail. 6300 * 6301 * Van's trick is to deposit buffers into socket queue 6302 * on a device interrupt, to call tcp_recv function 6303 * on the receive process context and checksum and copy 6304 * the buffer to user space. smart... 6305 * 6306 * Our current scheme is not silly either but we take the 6307 * extra cost of the net_bh soft interrupt processing... 6308 * We do checksum and copy also but from device to kernel. 6309 */ 6310 6311 tp->rx_opt.saw_tstamp = 0; 6312 tp->rx_opt.accecn = 0; 6313 6314 /* pred_flags is 0xS?10 << 16 + snd_wnd 6315 * if header_prediction is to be made 6316 * 'S' will always be tp->tcp_header_len >> 2 6317 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 6318 * turn it off (when there are holes in the receive 6319 * space for instance) 6320 * PSH flag is ignored. 6321 */ 6322 6323 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 6324 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 6325 !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 6326 int tcp_header_len = tp->tcp_header_len; 6327 s32 delta = 0; 6328 int flag = 0; 6329 6330 /* Timestamp header prediction: tcp_header_len 6331 * is automatically equal to th->doff*4 due to pred_flags 6332 * match. 6333 */ 6334 6335 /* Check timestamp */ 6336 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 6337 /* No? Slow path! */ 6338 if (!tcp_parse_aligned_timestamp(tp, th)) 6339 goto slow_path; 6340 6341 delta = tp->rx_opt.rcv_tsval - 6342 tp->rx_opt.ts_recent; 6343 /* If PAWS failed, check it more carefully in slow path */ 6344 if (delta < 0) 6345 goto slow_path; 6346 6347 /* DO NOT update ts_recent here, if checksum fails 6348 * and timestamp was corrupted part, it will result 6349 * in a hung connection since we will drop all 6350 * future packets due to the PAWS test. 6351 */ 6352 } 6353 6354 if (len <= tcp_header_len) { 6355 /* Bulk data transfer: sender */ 6356 if (len == tcp_header_len) { 6357 /* Predicted packet is in window by definition. 6358 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 6359 * Hence, check seq<=rcv_wup reduces to: 6360 */ 6361 if (tcp_header_len == 6362 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 6363 tp->rcv_nxt == tp->rcv_wup) 6364 flag |= __tcp_replace_ts_recent(tp, 6365 delta); 6366 6367 tcp_ecn_received_counters(sk, skb, 0); 6368 6369 /* We know that such packets are checksummed 6370 * on entry. 6371 */ 6372 tcp_ack(sk, skb, flag); 6373 __kfree_skb(skb); 6374 tcp_data_snd_check(sk); 6375 /* When receiving pure ack in fast path, update 6376 * last ts ecr directly instead of calling 6377 * tcp_rcv_rtt_measure_ts() 6378 */ 6379 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 6380 return; 6381 } else { /* Header too small */ 6382 reason = SKB_DROP_REASON_PKT_TOO_SMALL; 6383 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6384 goto discard; 6385 } 6386 } else { 6387 int eaten = 0; 6388 bool fragstolen = false; 6389 6390 if (tcp_checksum_complete(skb)) 6391 goto csum_error; 6392 6393 if (after(TCP_SKB_CB(skb)->end_seq, 6394 tp->rcv_nxt + tcp_receive_window(tp))) 6395 goto validate; 6396 6397 if ((int)skb->truesize > sk->sk_forward_alloc) 6398 goto step5; 6399 6400 /* Predicted packet is in window by definition. 6401 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 6402 * Hence, check seq<=rcv_wup reduces to: 6403 */ 6404 if (tcp_header_len == 6405 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 6406 tp->rcv_nxt == tp->rcv_wup) 6407 flag |= __tcp_replace_ts_recent(tp, 6408 delta); 6409 6410 tcp_rcv_rtt_measure_ts(sk, skb); 6411 6412 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS); 6413 6414 /* Bulk data transfer: receiver */ 6415 tcp_cleanup_skb(skb); 6416 __skb_pull(skb, tcp_header_len); 6417 tcp_ecn_received_counters(sk, skb, 6418 len - tcp_header_len); 6419 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 6420 6421 tcp_event_data_recv(sk, skb); 6422 6423 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 6424 /* Well, only one small jumplet in fast path... */ 6425 tcp_ack(sk, skb, flag | FLAG_DATA); 6426 tcp_data_snd_check(sk); 6427 if (!inet_csk_ack_scheduled(sk)) 6428 goto no_ack; 6429 } else { 6430 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq); 6431 } 6432 6433 __tcp_ack_snd_check(sk, 0); 6434no_ack: 6435 if (eaten) 6436 kfree_skb_partial(skb, fragstolen); 6437 tcp_data_ready(sk); 6438 return; 6439 } 6440 } 6441 6442slow_path: 6443 if (len < (th->doff << 2) || tcp_checksum_complete(skb)) 6444 goto csum_error; 6445 6446 if (!th->ack && !th->rst && !th->syn) { 6447 reason = SKB_DROP_REASON_TCP_FLAGS; 6448 goto discard; 6449 } 6450 6451 /* 6452 * Standard slow path. 6453 */ 6454validate: 6455 if (!tcp_validate_incoming(sk, skb, th, 1)) 6456 return; 6457 6458step5: 6459 tcp_ecn_received_counters_payload(sk, skb); 6460 6461 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT); 6462 if ((int)reason < 0) { 6463 reason = -reason; 6464 goto discard; 6465 } 6466 tcp_rcv_rtt_measure_ts(sk, skb); 6467 6468 /* Process urgent data. */ 6469 tcp_urg(sk, skb, th); 6470 6471 /* step 7: process the segment text */ 6472 tcp_data_queue(sk, skb); 6473 6474 tcp_data_snd_check(sk); 6475 tcp_ack_snd_check(sk); 6476 return; 6477 6478csum_error: 6479 reason = SKB_DROP_REASON_TCP_CSUM; 6480 trace_tcp_bad_csum(skb); 6481 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS); 6482 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6483 6484discard: 6485 tcp_drop_reason(sk, skb, reason); 6486} 6487EXPORT_IPV6_MOD(tcp_rcv_established); 6488 6489void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb) 6490{ 6491 struct inet_connection_sock *icsk = inet_csk(sk); 6492 struct tcp_sock *tp = tcp_sk(sk); 6493 6494 tcp_mtup_init(sk); 6495 icsk->icsk_af_ops->rebuild_header(sk); 6496 tcp_init_metrics(sk); 6497 6498 /* Initialize the congestion window to start the transfer. 6499 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 6500 * retransmitted. In light of RFC6298 more aggressive 1sec 6501 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 6502 * retransmission has occurred. 6503 */ 6504 if (tp->total_retrans > 1 && tp->undo_marker) 6505 tcp_snd_cwnd_set(tp, 1); 6506 else 6507 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk))); 6508 tp->snd_cwnd_stamp = tcp_jiffies32; 6509 6510 bpf_skops_established(sk, bpf_op, skb); 6511 /* Initialize congestion control unless BPF initialized it already: */ 6512 if (!icsk->icsk_ca_initialized) 6513 tcp_init_congestion_control(sk); 6514 tcp_init_buffer_space(sk); 6515} 6516 6517void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) 6518{ 6519 struct tcp_sock *tp = tcp_sk(sk); 6520 struct inet_connection_sock *icsk = inet_csk(sk); 6521 6522 tcp_ao_finish_connect(sk, skb); 6523 tcp_set_state(sk, TCP_ESTABLISHED); 6524 icsk->icsk_ack.lrcvtime = tcp_jiffies32; 6525 6526 if (skb) { 6527 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); 6528 security_inet_conn_established(sk, skb); 6529 sk_mark_napi_id(sk, skb); 6530 } 6531 6532 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb); 6533 6534 /* Prevent spurious tcp_cwnd_restart() on first data 6535 * packet. 6536 */ 6537 tp->lsndtime = tcp_jiffies32; 6538 6539 if (sock_flag(sk, SOCK_KEEPOPEN)) 6540 tcp_reset_keepalive_timer(sk, keepalive_time_when(tp)); 6541 6542 if (!tp->rx_opt.snd_wscale) 6543 __tcp_fast_path_on(tp, tp->snd_wnd); 6544 else 6545 tp->pred_flags = 0; 6546} 6547 6548static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack, 6549 struct tcp_fastopen_cookie *cookie) 6550{ 6551 struct tcp_sock *tp = tcp_sk(sk); 6552 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL; 6553 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0; 6554 bool syn_drop = false; 6555 6556 if (mss == READ_ONCE(tp->rx_opt.user_mss)) { 6557 struct tcp_options_received opt; 6558 6559 /* Get original SYNACK MSS value if user MSS sets mss_clamp */ 6560 tcp_clear_options(&opt); 6561 opt.user_mss = opt.mss_clamp = 0; 6562 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL); 6563 mss = opt.mss_clamp; 6564 } 6565 6566 if (!tp->syn_fastopen) { 6567 /* Ignore an unsolicited cookie */ 6568 cookie->len = -1; 6569 } else if (tp->total_retrans) { 6570 /* SYN timed out and the SYN-ACK neither has a cookie nor 6571 * acknowledges data. Presumably the remote received only 6572 * the retransmitted (regular) SYNs: either the original 6573 * SYN-data or the corresponding SYN-ACK was dropped. 6574 */ 6575 syn_drop = (cookie->len < 0 && data); 6576 } else if (cookie->len < 0 && !tp->syn_data) { 6577 /* We requested a cookie but didn't get it. If we did not use 6578 * the (old) exp opt format then try so next time (try_exp=1). 6579 * Otherwise we go back to use the RFC7413 opt (try_exp=2). 6580 */ 6581 try_exp = tp->syn_fastopen_exp ? 2 : 1; 6582 } 6583 6584 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp); 6585 6586 if (data) { /* Retransmit unacked data in SYN */ 6587 if (tp->total_retrans) 6588 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED; 6589 else 6590 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED; 6591 skb_rbtree_walk_from(data) 6592 tcp_mark_skb_lost(sk, data); 6593 tcp_non_congestion_loss_retransmit(sk); 6594 NET_INC_STATS(sock_net(sk), 6595 LINUX_MIB_TCPFASTOPENACTIVEFAIL); 6596 return true; 6597 } 6598 tp->syn_data_acked = tp->syn_data; 6599 if (tp->syn_data_acked) { 6600 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 6601 /* SYN-data is counted as two separate packets in tcp_ack() */ 6602 if (tp->delivered > 1) 6603 --tp->delivered; 6604 } 6605 6606 tcp_fastopen_add_skb(sk, synack); 6607 6608 return false; 6609} 6610 6611static void smc_check_reset_syn(struct tcp_sock *tp) 6612{ 6613#if IS_ENABLED(CONFIG_SMC) 6614 if (static_branch_unlikely(&tcp_have_smc)) { 6615 if (tp->syn_smc && !tp->rx_opt.smc_ok) 6616 tp->syn_smc = 0; 6617 } 6618#endif 6619} 6620 6621static void tcp_try_undo_spurious_syn(struct sock *sk) 6622{ 6623 struct tcp_sock *tp = tcp_sk(sk); 6624 u32 syn_stamp; 6625 6626 /* undo_marker is set when SYN or SYNACK times out. The timeout is 6627 * spurious if the ACK's timestamp option echo value matches the 6628 * original SYN timestamp. 6629 */ 6630 syn_stamp = tp->retrans_stamp; 6631 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp && 6632 syn_stamp == tp->rx_opt.rcv_tsecr) 6633 tp->undo_marker = 0; 6634} 6635 6636static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 6637 const struct tcphdr *th) 6638{ 6639 struct inet_connection_sock *icsk = inet_csk(sk); 6640 struct tcp_sock *tp = tcp_sk(sk); 6641 struct tcp_fastopen_cookie foc = { .len = -1 }; 6642 int saved_clamp = tp->rx_opt.mss_clamp; 6643 bool fastopen_fail; 6644 SKB_DR(reason); 6645 6646 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc); 6647 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 6648 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 6649 6650 if (th->ack) { 6651 /* rfc793: 6652 * "If the state is SYN-SENT then 6653 * first check the ACK bit 6654 * If the ACK bit is set 6655 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 6656 * a reset (unless the RST bit is set, if so drop 6657 * the segment and return)" 6658 */ 6659 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) || 6660 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 6661 /* Previous FIN/ACK or RST/ACK might be ignored. */ 6662 if (icsk->icsk_retransmits == 0) 6663 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 6664 TCP_TIMEOUT_MIN, false); 6665 SKB_DR_SET(reason, TCP_INVALID_ACK_SEQUENCE); 6666 goto reset_and_undo; 6667 } 6668 6669 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 6670 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 6671 tcp_time_stamp_ts(tp))) { 6672 NET_INC_STATS(sock_net(sk), 6673 LINUX_MIB_PAWSACTIVEREJECTED); 6674 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 6675 goto reset_and_undo; 6676 } 6677 6678 /* Now ACK is acceptable. 6679 * 6680 * "If the RST bit is set 6681 * If the ACK was acceptable then signal the user "error: 6682 * connection reset", drop the segment, enter CLOSED state, 6683 * delete TCB, and return." 6684 */ 6685 6686 if (th->rst) { 6687 tcp_reset(sk, skb); 6688consume: 6689 __kfree_skb(skb); 6690 return 0; 6691 } 6692 6693 /* rfc793: 6694 * "fifth, if neither of the SYN or RST bits is set then 6695 * drop the segment and return." 6696 * 6697 * See note below! 6698 * --ANK(990513) 6699 */ 6700 if (!th->syn) { 6701 SKB_DR_SET(reason, TCP_FLAGS); 6702 goto discard_and_undo; 6703 } 6704 /* rfc793: 6705 * "If the SYN bit is on ... 6706 * are acceptable then ... 6707 * (our SYN has been ACKed), change the connection 6708 * state to ESTABLISHED..." 6709 */ 6710 6711 if (tcp_ecn_mode_any(tp)) 6712 tcp_ecn_rcv_synack(sk, skb, th, 6713 TCP_SKB_CB(skb)->ip_dsfield); 6714 6715 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 6716 tcp_try_undo_spurious_syn(sk); 6717 tcp_ack(sk, skb, FLAG_SLOWPATH); 6718 6719 /* Ok.. it's good. Set up sequence numbers and 6720 * move to established. 6721 */ 6722 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6723 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6724 6725 /* RFC1323: The window in SYN & SYN/ACK segments is 6726 * never scaled. 6727 */ 6728 tp->snd_wnd = ntohs(th->window); 6729 6730 if (!tp->rx_opt.wscale_ok) { 6731 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 6732 WRITE_ONCE(tp->window_clamp, 6733 min(tp->window_clamp, 65535U)); 6734 } 6735 6736 if (tp->rx_opt.saw_tstamp) { 6737 tp->rx_opt.tstamp_ok = 1; 6738 tp->tcp_header_len = 6739 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6740 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 6741 tcp_store_ts_recent(tp); 6742 } else { 6743 tp->tcp_header_len = sizeof(struct tcphdr); 6744 } 6745 6746 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6747 tcp_initialize_rcv_mss(sk); 6748 6749 /* Remember, tcp_poll() does not lock socket! 6750 * Change state from SYN-SENT only after copied_seq 6751 * is initialized. */ 6752 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6753 6754 smc_check_reset_syn(tp); 6755 6756 smp_mb(); 6757 6758 tcp_finish_connect(sk, skb); 6759 6760 fastopen_fail = (tp->syn_fastopen || tp->syn_data) && 6761 tcp_rcv_fastopen_synack(sk, skb, &foc); 6762 6763 if (!sock_flag(sk, SOCK_DEAD)) { 6764 sk->sk_state_change(sk); 6765 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 6766 } 6767 if (fastopen_fail) 6768 return -1; 6769 if (sk->sk_write_pending || 6770 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) || 6771 inet_csk_in_pingpong_mode(sk)) { 6772 /* Save one ACK. Data will be ready after 6773 * several ticks, if write_pending is set. 6774 * 6775 * It may be deleted, but with this feature tcpdumps 6776 * look so _wonderfully_ clever, that I was not able 6777 * to stand against the temptation 8) --ANK 6778 */ 6779 inet_csk_schedule_ack(sk); 6780 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 6781 tcp_reset_xmit_timer(sk, ICSK_TIME_DACK, 6782 TCP_DELACK_MAX, false); 6783 goto consume; 6784 } 6785 tcp_send_ack_reflect_ect(sk, tcp_ecn_mode_accecn(tp)); 6786 return -1; 6787 } 6788 6789 /* No ACK in the segment */ 6790 6791 if (th->rst) { 6792 /* rfc793: 6793 * "If the RST bit is set 6794 * 6795 * Otherwise (no ACK) drop the segment and return." 6796 */ 6797 SKB_DR_SET(reason, TCP_RESET); 6798 goto discard_and_undo; 6799 } 6800 6801 /* PAWS check. */ 6802 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 6803 tcp_paws_reject(&tp->rx_opt, 0)) { 6804 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 6805 goto discard_and_undo; 6806 } 6807 if (th->syn) { 6808 /* We see SYN without ACK. It is attempt of 6809 * simultaneous connect with crossed SYNs. 6810 * Particularly, it can be connect to self. 6811 */ 6812#ifdef CONFIG_TCP_AO 6813 struct tcp_ao_info *ao; 6814 6815 ao = rcu_dereference_protected(tp->ao_info, 6816 lockdep_sock_is_held(sk)); 6817 if (ao) { 6818 WRITE_ONCE(ao->risn, th->seq); 6819 ao->rcv_sne = 0; 6820 } 6821#endif 6822 tcp_set_state(sk, TCP_SYN_RECV); 6823 6824 if (tp->rx_opt.saw_tstamp) { 6825 tp->rx_opt.tstamp_ok = 1; 6826 tcp_store_ts_recent(tp); 6827 tp->tcp_header_len = 6828 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 6829 } else { 6830 tp->tcp_header_len = sizeof(struct tcphdr); 6831 } 6832 6833 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 6834 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 6835 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 6836 6837 /* RFC1323: The window in SYN & SYN/ACK segments is 6838 * never scaled. 6839 */ 6840 tp->snd_wnd = ntohs(th->window); 6841 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 6842 tp->max_window = tp->snd_wnd; 6843 6844 tcp_ecn_rcv_syn(sk, th, skb); 6845 6846 tcp_mtup_init(sk); 6847 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 6848 tcp_initialize_rcv_mss(sk); 6849 6850 tcp_send_synack(sk); 6851#if 0 6852 /* Note, we could accept data and URG from this segment. 6853 * There are no obstacles to make this (except that we must 6854 * either change tcp_recvmsg() to prevent it from returning data 6855 * before 3WHS completes per RFC793, or employ TCP Fast Open). 6856 * 6857 * However, if we ignore data in ACKless segments sometimes, 6858 * we have no reasons to accept it sometimes. 6859 * Also, seems the code doing it in step6 of tcp_rcv_state_process 6860 * is not flawless. So, discard packet for sanity. 6861 * Uncomment this return to process the data. 6862 */ 6863 return -1; 6864#else 6865 goto consume; 6866#endif 6867 } 6868 /* "fifth, if neither of the SYN or RST bits is set then 6869 * drop the segment and return." 6870 */ 6871 6872discard_and_undo: 6873 tcp_clear_options(&tp->rx_opt); 6874 tp->rx_opt.mss_clamp = saved_clamp; 6875 tcp_drop_reason(sk, skb, reason); 6876 return 0; 6877 6878reset_and_undo: 6879 tcp_clear_options(&tp->rx_opt); 6880 tp->rx_opt.mss_clamp = saved_clamp; 6881 /* we can reuse/return @reason to its caller to handle the exception */ 6882 return reason; 6883} 6884 6885static void tcp_rcv_synrecv_state_fastopen(struct sock *sk) 6886{ 6887 struct tcp_sock *tp = tcp_sk(sk); 6888 struct request_sock *req; 6889 6890 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows 6891 * undo. If peer SACKs triggered fast recovery, we can't undo here. 6892 */ 6893 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out) 6894 tcp_try_undo_recovery(sk); 6895 6896 tcp_update_rto_time(tp); 6897 WRITE_ONCE(inet_csk(sk)->icsk_retransmits, 0); 6898 /* In tcp_fastopen_synack_timer() on the first SYNACK RTO we set 6899 * retrans_stamp but don't enter CA_Loss, so in case that happened we 6900 * need to zero retrans_stamp here to prevent spurious 6901 * retransmits_timed_out(). However, if the ACK of our SYNACK caused us 6902 * to enter CA_Recovery then we need to leave retrans_stamp as it was 6903 * set entering CA_Recovery, for correct retransmits_timed_out() and 6904 * undo behavior. 6905 */ 6906 tcp_retrans_stamp_cleanup(sk); 6907 6908 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1, 6909 * we no longer need req so release it. 6910 */ 6911 req = rcu_dereference_protected(tp->fastopen_rsk, 6912 lockdep_sock_is_held(sk)); 6913 reqsk_fastopen_remove(sk, req, false); 6914 6915 /* Re-arm the timer because data may have been sent out. 6916 * This is similar to the regular data transmission case 6917 * when new data has just been ack'ed. 6918 * 6919 * (TFO) - we could try to be more aggressive and 6920 * retransmitting any data sooner based on when they 6921 * are sent out. 6922 */ 6923 tcp_rearm_rto(sk); 6924} 6925 6926/* 6927 * This function implements the receiving procedure of RFC 793 for 6928 * all states except ESTABLISHED and TIME_WAIT. 6929 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 6930 * address independent. 6931 */ 6932 6933enum skb_drop_reason 6934tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb) 6935{ 6936 struct tcp_sock *tp = tcp_sk(sk); 6937 struct inet_connection_sock *icsk = inet_csk(sk); 6938 const struct tcphdr *th = tcp_hdr(skb); 6939 struct request_sock *req; 6940 int queued = 0; 6941 SKB_DR(reason); 6942 6943 switch (sk->sk_state) { 6944 case TCP_CLOSE: 6945 SKB_DR_SET(reason, TCP_CLOSE); 6946 goto discard; 6947 6948 case TCP_LISTEN: 6949 if (th->ack) 6950 return SKB_DROP_REASON_TCP_FLAGS; 6951 6952 if (th->rst) { 6953 SKB_DR_SET(reason, TCP_RESET); 6954 goto discard; 6955 } 6956 if (th->syn) { 6957 if (th->fin) { 6958 SKB_DR_SET(reason, TCP_FLAGS); 6959 goto discard; 6960 } 6961 /* It is possible that we process SYN packets from backlog, 6962 * so we need to make sure to disable BH and RCU right there. 6963 */ 6964 rcu_read_lock(); 6965 local_bh_disable(); 6966 icsk->icsk_af_ops->conn_request(sk, skb); 6967 local_bh_enable(); 6968 rcu_read_unlock(); 6969 6970 consume_skb(skb); 6971 return 0; 6972 } 6973 SKB_DR_SET(reason, TCP_FLAGS); 6974 goto discard; 6975 6976 case TCP_SYN_SENT: 6977 tp->rx_opt.saw_tstamp = 0; 6978 tcp_mstamp_refresh(tp); 6979 queued = tcp_rcv_synsent_state_process(sk, skb, th); 6980 if (queued >= 0) 6981 return queued; 6982 6983 /* Do step6 onward by hand. */ 6984 tcp_urg(sk, skb, th); 6985 __kfree_skb(skb); 6986 tcp_data_snd_check(sk); 6987 return 0; 6988 } 6989 6990 tcp_mstamp_refresh(tp); 6991 tp->rx_opt.saw_tstamp = 0; 6992 req = rcu_dereference_protected(tp->fastopen_rsk, 6993 lockdep_sock_is_held(sk)); 6994 if (req) { 6995 bool req_stolen; 6996 6997 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 6998 sk->sk_state != TCP_FIN_WAIT1); 6999 7000 SKB_DR_SET(reason, TCP_FASTOPEN); 7001 if (!tcp_check_req(sk, skb, req, true, &req_stolen, &reason)) 7002 goto discard; 7003 } 7004 7005 if (!th->ack && !th->rst && !th->syn) { 7006 SKB_DR_SET(reason, TCP_FLAGS); 7007 goto discard; 7008 } 7009 if (!tcp_validate_incoming(sk, skb, th, 0)) 7010 return 0; 7011 7012 /* step 5: check the ACK field */ 7013 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | 7014 FLAG_UPDATE_TS_RECENT | 7015 FLAG_NO_CHALLENGE_ACK); 7016 7017 if ((int)reason <= 0) { 7018 if (sk->sk_state == TCP_SYN_RECV) { 7019 /* send one RST */ 7020 if (!reason) 7021 return SKB_DROP_REASON_TCP_OLD_ACK; 7022 return -reason; 7023 } 7024 /* accept old ack during closing */ 7025 if ((int)reason < 0) { 7026 tcp_send_challenge_ack(sk, false); 7027 reason = -reason; 7028 goto discard; 7029 } 7030 } 7031 SKB_DR_SET(reason, NOT_SPECIFIED); 7032 switch (sk->sk_state) { 7033 case TCP_SYN_RECV: 7034 tp->delivered++; /* SYN-ACK delivery isn't tracked in tcp_ack */ 7035 if (!tp->srtt_us) 7036 tcp_synack_rtt_meas(sk, req); 7037 7038 if (tp->rx_opt.tstamp_ok) 7039 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 7040 7041 if (req) { 7042 tcp_rcv_synrecv_state_fastopen(sk); 7043 } else { 7044 tcp_try_undo_spurious_syn(sk); 7045 tp->retrans_stamp = 0; 7046 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, 7047 skb); 7048 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 7049 } 7050 tcp_ao_established(sk); 7051 smp_mb(); 7052 tcp_set_state(sk, TCP_ESTABLISHED); 7053 sk->sk_state_change(sk); 7054 7055 /* Note, that this wakeup is only for marginal crossed SYN case. 7056 * Passively open sockets are not waked up, because 7057 * sk->sk_sleep == NULL and sk->sk_socket == NULL. 7058 */ 7059 if (sk->sk_socket) 7060 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 7061 7062 tp->snd_una = TCP_SKB_CB(skb)->ack_seq; 7063 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; 7064 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 7065 7066 if (!inet_csk(sk)->icsk_ca_ops->cong_control) 7067 tcp_update_pacing_rate(sk); 7068 7069 /* Prevent spurious tcp_cwnd_restart() on first data packet */ 7070 tp->lsndtime = tcp_jiffies32; 7071 7072 tcp_initialize_rcv_mss(sk); 7073 if (tcp_ecn_mode_accecn(tp)) 7074 tcp_accecn_third_ack(sk, skb, tp->syn_ect_snt); 7075 tcp_fast_path_on(tp); 7076 if (sk->sk_shutdown & SEND_SHUTDOWN) 7077 tcp_shutdown(sk, SEND_SHUTDOWN); 7078 7079 break; 7080 7081 case TCP_FIN_WAIT1: { 7082 int tmo; 7083 7084 if (req) 7085 tcp_rcv_synrecv_state_fastopen(sk); 7086 7087 if (tp->snd_una != tp->write_seq) 7088 break; 7089 7090 tcp_set_state(sk, TCP_FIN_WAIT2); 7091 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN); 7092 7093 sk_dst_confirm(sk); 7094 7095 if (!sock_flag(sk, SOCK_DEAD)) { 7096 /* Wake up lingering close() */ 7097 sk->sk_state_change(sk); 7098 break; 7099 } 7100 7101 if (READ_ONCE(tp->linger2) < 0) { 7102 tcp_done(sk); 7103 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 7104 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 7105 } 7106 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 7107 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 7108 /* Receive out of order FIN after close() */ 7109 if (tp->syn_fastopen && th->fin) 7110 tcp_fastopen_active_disable(sk); 7111 tcp_done(sk); 7112 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 7113 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 7114 } 7115 7116 tmo = tcp_fin_time(sk); 7117 if (tmo > TCP_TIMEWAIT_LEN) { 7118 tcp_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 7119 } else if (th->fin || sock_owned_by_user(sk)) { 7120 /* Bad case. We could lose such FIN otherwise. 7121 * It is not a big problem, but it looks confusing 7122 * and not so rare event. We still can lose it now, 7123 * if it spins in bh_lock_sock(), but it is really 7124 * marginal case. 7125 */ 7126 tcp_reset_keepalive_timer(sk, tmo); 7127 } else { 7128 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 7129 goto consume; 7130 } 7131 break; 7132 } 7133 7134 case TCP_CLOSING: 7135 if (tp->snd_una == tp->write_seq) { 7136 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 7137 goto consume; 7138 } 7139 break; 7140 7141 case TCP_LAST_ACK: 7142 if (tp->snd_una == tp->write_seq) { 7143 tcp_update_metrics(sk); 7144 tcp_done(sk); 7145 goto consume; 7146 } 7147 break; 7148 } 7149 7150 /* step 6: check the URG bit */ 7151 tcp_urg(sk, skb, th); 7152 7153 /* step 7: process the segment text */ 7154 switch (sk->sk_state) { 7155 case TCP_CLOSE_WAIT: 7156 case TCP_CLOSING: 7157 case TCP_LAST_ACK: 7158 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 7159 /* If a subflow has been reset, the packet should not 7160 * continue to be processed, drop the packet. 7161 */ 7162 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) 7163 goto discard; 7164 break; 7165 } 7166 fallthrough; 7167 case TCP_FIN_WAIT1: 7168 case TCP_FIN_WAIT2: 7169 /* RFC 793 says to queue data in these states, 7170 * RFC 1122 says we MUST send a reset. 7171 * BSD 4.4 also does reset. 7172 */ 7173 if (sk->sk_shutdown & RCV_SHUTDOWN) { 7174 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 7175 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 7176 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 7177 tcp_reset(sk, skb); 7178 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 7179 } 7180 } 7181 fallthrough; 7182 case TCP_ESTABLISHED: 7183 tcp_data_queue(sk, skb); 7184 queued = 1; 7185 break; 7186 } 7187 7188 /* tcp_data could move socket to TIME-WAIT */ 7189 if (sk->sk_state != TCP_CLOSE) { 7190 tcp_data_snd_check(sk); 7191 tcp_ack_snd_check(sk); 7192 } 7193 7194 if (!queued) { 7195discard: 7196 tcp_drop_reason(sk, skb, reason); 7197 } 7198 return 0; 7199 7200consume: 7201 __kfree_skb(skb); 7202 return 0; 7203} 7204EXPORT_IPV6_MOD(tcp_rcv_state_process); 7205 7206static inline void pr_drop_req(struct request_sock *req, __u16 port, int family) 7207{ 7208 struct inet_request_sock *ireq = inet_rsk(req); 7209 7210 if (family == AF_INET) 7211 net_dbg_ratelimited("drop open request from %pI4/%u\n", 7212 &ireq->ir_rmt_addr, port); 7213#if IS_ENABLED(CONFIG_IPV6) 7214 else if (family == AF_INET6) 7215 net_dbg_ratelimited("drop open request from %pI6/%u\n", 7216 &ireq->ir_v6_rmt_addr, port); 7217#endif 7218} 7219 7220/* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set 7221 * 7222 * If we receive a SYN packet with these bits set, it means a 7223 * network is playing bad games with TOS bits. In order to 7224 * avoid possible false congestion notifications, we disable 7225 * TCP ECN negotiation. 7226 * 7227 * Exception: tcp_ca wants ECN. This is required for DCTCP 7228 * congestion control: Linux DCTCP asserts ECT on all packets, 7229 * including SYN, which is most optimal solution; however, 7230 * others, such as FreeBSD do not. 7231 * 7232 * Exception: At least one of the reserved bits of the TCP header (th->res1) is 7233 * set, indicating the use of a future TCP extension (such as AccECN). See 7234 * RFC8311 §4.3 which updates RFC3168 to allow the development of such 7235 * extensions. 7236 */ 7237static void tcp_ecn_create_request(struct request_sock *req, 7238 const struct sk_buff *skb, 7239 const struct sock *listen_sk, 7240 const struct dst_entry *dst) 7241{ 7242 const struct tcphdr *th = tcp_hdr(skb); 7243 const struct net *net = sock_net(listen_sk); 7244 bool th_ecn = th->ece && th->cwr; 7245 bool ect, ecn_ok; 7246 u32 ecn_ok_dst; 7247 7248 if (tcp_accecn_syn_requested(th) && 7249 (READ_ONCE(net->ipv4.sysctl_tcp_ecn) >= 3 || 7250 tcp_ca_needs_accecn(listen_sk))) { 7251 inet_rsk(req)->ecn_ok = 1; 7252 tcp_rsk(req)->accecn_ok = 1; 7253 tcp_rsk(req)->syn_ect_rcv = TCP_SKB_CB(skb)->ip_dsfield & 7254 INET_ECN_MASK; 7255 return; 7256 } 7257 7258 if (!th_ecn) 7259 return; 7260 7261 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield); 7262 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK); 7263 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst; 7264 7265 if (((!ect || th->res1 || th->ae) && ecn_ok) || 7266 tcp_ca_needs_ecn(listen_sk) || 7267 (ecn_ok_dst & DST_FEATURE_ECN_CA) || 7268 tcp_bpf_ca_needs_ecn((struct sock *)req)) 7269 inet_rsk(req)->ecn_ok = 1; 7270} 7271 7272static void tcp_openreq_init(struct request_sock *req, 7273 const struct tcp_options_received *rx_opt, 7274 struct sk_buff *skb, const struct sock *sk) 7275{ 7276 struct inet_request_sock *ireq = inet_rsk(req); 7277 7278 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */ 7279 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq; 7280 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 7281 tcp_rsk(req)->snt_synack = 0; 7282 tcp_rsk(req)->snt_tsval_first = 0; 7283 tcp_rsk(req)->last_oow_ack_time = 0; 7284 tcp_rsk(req)->accecn_ok = 0; 7285 tcp_rsk(req)->saw_accecn_opt = TCP_ACCECN_OPT_NOT_SEEN; 7286 tcp_rsk(req)->accecn_fail_mode = 0; 7287 tcp_rsk(req)->syn_ect_rcv = 0; 7288 tcp_rsk(req)->syn_ect_snt = 0; 7289 req->mss = rx_opt->mss_clamp; 7290 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0; 7291 ireq->tstamp_ok = rx_opt->tstamp_ok; 7292 ireq->sack_ok = rx_opt->sack_ok; 7293 ireq->snd_wscale = rx_opt->snd_wscale; 7294 ireq->wscale_ok = rx_opt->wscale_ok; 7295 ireq->acked = 0; 7296 ireq->ecn_ok = 0; 7297 ireq->ir_rmt_port = tcp_hdr(skb)->source; 7298 ireq->ir_num = ntohs(tcp_hdr(skb)->dest); 7299 ireq->ir_mark = inet_request_mark(sk, skb); 7300#if IS_ENABLED(CONFIG_SMC) 7301 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested && 7302 tcp_sk(sk)->smc_hs_congested(sk)); 7303#endif 7304} 7305 7306/* 7307 * Return true if a syncookie should be sent 7308 */ 7309static bool tcp_syn_flood_action(struct sock *sk, const char *proto) 7310{ 7311 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; 7312 const char *msg = "Dropping request"; 7313 struct net *net = sock_net(sk); 7314 bool want_cookie = false; 7315 u8 syncookies; 7316 7317 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); 7318 7319#ifdef CONFIG_SYN_COOKIES 7320 if (syncookies) { 7321 msg = "Sending cookies"; 7322 want_cookie = true; 7323 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES); 7324 } else 7325#endif 7326 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP); 7327 7328 if (syncookies != 2 && !READ_ONCE(queue->synflood_warned)) { 7329 WRITE_ONCE(queue->synflood_warned, 1); 7330 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) { 7331 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n", 7332 proto, inet6_rcv_saddr(sk), 7333 sk->sk_num, msg); 7334 } else { 7335 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n", 7336 proto, &sk->sk_rcv_saddr, 7337 sk->sk_num, msg); 7338 } 7339 } 7340 7341 return want_cookie; 7342} 7343 7344static void tcp_reqsk_record_syn(const struct sock *sk, 7345 struct request_sock *req, 7346 const struct sk_buff *skb) 7347{ 7348 if (tcp_sk(sk)->save_syn) { 7349 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb); 7350 struct saved_syn *saved_syn; 7351 u32 mac_hdrlen; 7352 void *base; 7353 7354 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */ 7355 base = skb_mac_header(skb); 7356 mac_hdrlen = skb_mac_header_len(skb); 7357 len += mac_hdrlen; 7358 } else { 7359 base = skb_network_header(skb); 7360 mac_hdrlen = 0; 7361 } 7362 7363 saved_syn = kmalloc(struct_size(saved_syn, data, len), 7364 GFP_ATOMIC); 7365 if (saved_syn) { 7366 saved_syn->mac_hdrlen = mac_hdrlen; 7367 saved_syn->network_hdrlen = skb_network_header_len(skb); 7368 saved_syn->tcp_hdrlen = tcp_hdrlen(skb); 7369 memcpy(saved_syn->data, base, len); 7370 req->saved_syn = saved_syn; 7371 } 7372 } 7373} 7374 7375/* If a SYN cookie is required and supported, returns a clamped MSS value to be 7376 * used for SYN cookie generation. 7377 */ 7378u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops, 7379 const struct tcp_request_sock_ops *af_ops, 7380 struct sock *sk, struct tcphdr *th) 7381{ 7382 struct tcp_sock *tp = tcp_sk(sk); 7383 u16 mss; 7384 7385 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 && 7386 !inet_csk_reqsk_queue_is_full(sk)) 7387 return 0; 7388 7389 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name)) 7390 return 0; 7391 7392 if (sk_acceptq_is_full(sk)) { 7393 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 7394 return 0; 7395 } 7396 7397 mss = tcp_parse_mss_option(th, READ_ONCE(tp->rx_opt.user_mss)); 7398 if (!mss) 7399 mss = af_ops->mss_clamp; 7400 7401 return mss; 7402} 7403EXPORT_IPV6_MOD_GPL(tcp_get_syncookie_mss); 7404 7405int tcp_conn_request(struct request_sock_ops *rsk_ops, 7406 const struct tcp_request_sock_ops *af_ops, 7407 struct sock *sk, struct sk_buff *skb) 7408{ 7409 struct tcp_fastopen_cookie foc = { .len = -1 }; 7410 struct tcp_options_received tmp_opt; 7411 const struct tcp_sock *tp = tcp_sk(sk); 7412 struct net *net = sock_net(sk); 7413 struct sock *fastopen_sk = NULL; 7414 union tcp_seq_and_ts_off st; 7415 struct request_sock *req; 7416 bool want_cookie = false; 7417 struct dst_entry *dst; 7418 struct flowi fl; 7419 u8 syncookies; 7420 u32 isn; 7421 7422#ifdef CONFIG_TCP_AO 7423 const struct tcp_ao_hdr *aoh; 7424#endif 7425 7426 isn = __this_cpu_read(tcp_tw_isn); 7427 if (isn) { 7428 /* TW buckets are converted to open requests without 7429 * limitations, they conserve resources and peer is 7430 * evidently real one. 7431 */ 7432 __this_cpu_write(tcp_tw_isn, 0); 7433 } else { 7434 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); 7435 7436 if (syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) { 7437 want_cookie = tcp_syn_flood_action(sk, 7438 rsk_ops->slab_name); 7439 if (!want_cookie) 7440 goto drop; 7441 } 7442 } 7443 7444 if (sk_acceptq_is_full(sk)) { 7445 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 7446 goto drop; 7447 } 7448 7449 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie); 7450 if (!req) 7451 goto drop; 7452 7453 req->syncookie = want_cookie; 7454 tcp_rsk(req)->af_specific = af_ops; 7455 tcp_rsk(req)->ts_off = 0; 7456 tcp_rsk(req)->req_usec_ts = false; 7457#if IS_ENABLED(CONFIG_MPTCP) 7458 tcp_rsk(req)->is_mptcp = 0; 7459#endif 7460 7461 tcp_clear_options(&tmp_opt); 7462 tmp_opt.mss_clamp = af_ops->mss_clamp; 7463 tmp_opt.user_mss = READ_ONCE(tp->rx_opt.user_mss); 7464 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, 7465 want_cookie ? NULL : &foc); 7466 7467 if (want_cookie && !tmp_opt.saw_tstamp) 7468 tcp_clear_options(&tmp_opt); 7469 7470 if (IS_ENABLED(CONFIG_SMC) && want_cookie) 7471 tmp_opt.smc_ok = 0; 7472 7473 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp; 7474 tcp_openreq_init(req, &tmp_opt, skb, sk); 7475 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk); 7476 7477 /* Note: tcp_v6_init_req() might override ir_iif for link locals */ 7478 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb); 7479 7480 dst = af_ops->route_req(sk, skb, &fl, req, isn); 7481 if (!dst) 7482 goto drop_and_free; 7483 7484 if (tmp_opt.tstamp_ok || (!want_cookie && !isn)) 7485 st = af_ops->init_seq_and_ts_off(net, skb); 7486 7487 if (tmp_opt.tstamp_ok) { 7488 tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst); 7489 tcp_rsk(req)->ts_off = st.ts_off; 7490 } 7491 if (!want_cookie && !isn) { 7492 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog); 7493 7494 /* Kill the following clause, if you dislike this way. */ 7495 if (!syncookies && 7496 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) < 7497 (max_syn_backlog >> 2)) && 7498 !tcp_peer_is_proven(req, dst)) { 7499 /* Without syncookies last quarter of 7500 * backlog is filled with destinations, 7501 * proven to be alive. 7502 * It means that we continue to communicate 7503 * to destinations, already remembered 7504 * to the moment of synflood. 7505 */ 7506 pr_drop_req(req, ntohs(tcp_hdr(skb)->source), 7507 rsk_ops->family); 7508 goto drop_and_release; 7509 } 7510 7511 isn = st.seq; 7512 } 7513 7514 tcp_ecn_create_request(req, skb, sk, dst); 7515 7516 if (want_cookie) { 7517 isn = cookie_init_sequence(af_ops, sk, skb, &req->mss); 7518 if (!tmp_opt.tstamp_ok) 7519 inet_rsk(req)->ecn_ok = 0; 7520 } 7521 7522#ifdef CONFIG_TCP_AO 7523 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh)) 7524 goto drop_and_release; /* Invalid TCP options */ 7525 if (aoh) { 7526 tcp_rsk(req)->used_tcp_ao = true; 7527 tcp_rsk(req)->ao_rcv_next = aoh->keyid; 7528 tcp_rsk(req)->ao_keyid = aoh->rnext_keyid; 7529 7530 } else { 7531 tcp_rsk(req)->used_tcp_ao = false; 7532 } 7533#endif 7534 tcp_rsk(req)->snt_isn = isn; 7535 tcp_rsk(req)->txhash = net_tx_rndhash(); 7536 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield; 7537 tcp_openreq_init_rwin(req, sk, dst); 7538 sk_rx_queue_set(req_to_sk(req), skb); 7539 if (!want_cookie) { 7540 tcp_reqsk_record_syn(sk, req, skb); 7541 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst); 7542 } 7543 if (fastopen_sk) { 7544 af_ops->send_synack(fastopen_sk, dst, &fl, req, 7545 &foc, TCP_SYNACK_FASTOPEN, skb); 7546 /* Add the child socket directly into the accept queue */ 7547 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) { 7548 bh_unlock_sock(fastopen_sk); 7549 sock_put(fastopen_sk); 7550 goto drop_and_free; 7551 } 7552 READ_ONCE(sk->sk_data_ready)(sk); 7553 bh_unlock_sock(fastopen_sk); 7554 sock_put(fastopen_sk); 7555 } else { 7556 tcp_rsk(req)->tfo_listener = false; 7557 if (!want_cookie && 7558 unlikely(!inet_csk_reqsk_queue_hash_add(sk, req))) { 7559 reqsk_free(req); 7560 dst_release(dst); 7561 return 0; 7562 } 7563 af_ops->send_synack(sk, dst, &fl, req, &foc, 7564 !want_cookie ? TCP_SYNACK_NORMAL : 7565 TCP_SYNACK_COOKIE, 7566 skb); 7567 if (want_cookie) { 7568 reqsk_free(req); 7569 return 0; 7570 } 7571 } 7572 reqsk_put(req); 7573 return 0; 7574 7575drop_and_release: 7576 dst_release(dst); 7577drop_and_free: 7578 __reqsk_free(req); 7579drop: 7580 tcp_listendrop(sk); 7581 return 0; 7582} 7583EXPORT_IPV6_MOD(tcp_conn_request);
~/
Read-only file preview.