1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * INET An implementation of the TCP/IP protocol suite for the LINUX 4 * operating system. INET is implemented using the BSD Socket 5 * interface as the means of communication with the user level. 6 * 7 * Implementation of the Transmission Control Protocol(TCP). 8 * 9 * Authors: Ross Biro 10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 11 * Mark Evans, <evansmp@uhura.aston.ac.uk> 12 * Corey Minyard <wf-rch!minyard@relay.EU.net> 13 * Florian La Roche, <flla@stud.uni-sb.de> 14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 15 * Linus Torvalds, <torvalds@cs.helsinki.fi> 16 * Alan Cox, <gw4pts@gw4pts.ampr.org> 17 * Matthew Dillon, <dillon@apollo.west.oic.com> 18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 19 * Jorge Cwik, <jorge@laser.satlink.net> 20 */ 21 22 /* 23 * Changes: 24 * Pedro Roque : Fast Retransmit/Recovery. 25 * Two receive queues. 26 * Retransmit queue handled by TCP. 27 * Better retransmit timer handling. 28 * New congestion avoidance. 29 * Header prediction. 30 * Variable renaming. 31 * 32 * Eric : Fast Retransmit. 33 * Randy Scott : MSS option defines. 34 * Eric Schenk : Fixes to slow start algorithm. 35 * Eric Schenk : Yet another double ACK bug. 36 * Eric Schenk : Delayed ACK bug fixes. 37 * Eric Schenk : Floyd style fast retrans war avoidance. 38 * David S. Miller : Don't allow zero congestion window. 39 * Eric Schenk : Fix retransmitter so that it sends 40 * next packet on ack of previous packet. 41 * Andi Kleen : Moved open_request checking here 42 * and process RSTs for open_requests. 43 * Andi Kleen : Better prune_queue, and other fixes. 44 * Andrey Savochkin: Fix RTT measurements in the presence of 45 * timestamps. 46 * Andrey Savochkin: Check sequence numbers correctly when 47 * removing SACKs due to in sequence incoming 48 * data segments. 49 * Andi Kleen: Make sure we never ack data there is not 50 * enough room for. Also make this condition 51 * a fatal error if it might still happen. 52 * Andi Kleen: Add tcp_measure_rcv_mss to make 53 * connections with MSS<min(MTU,ann. MSS) 54 * work without delayed acks. 55 * Andi Kleen: Process packets with PSH set in the 56 * fast path. 57 * J Hadi Salim: ECN support 58 * Andrei Gurtov, 59 * Pasi Sarolahti, 60 * Panu Kuhlberg: Experimental audit of TCP (re)transmission 61 * engine. Lots of bugs are found. 62 * Pasi Sarolahti: F-RTO for dealing with spurious RTOs 63 */ 64 65 #define pr_fmt(fmt) "TCP: " fmt 66 67 #include <linux/mm.h> 68 #include <linux/slab.h> 69 #include <linux/module.h> 70 #include <linux/sysctl.h> 71 #include <linux/kernel.h> 72 #include <linux/prefetch.h> 73 #include <linux/bitops.h> 74 #include <net/dst.h> 75 #include <net/tcp.h> 76 #include <net/tcp_ecn.h> 77 #include <net/proto_memory.h> 78 #include <net/inet_common.h> 79 #include <linux/ipsec.h> 80 #include <linux/unaligned.h> 81 #include <linux/errqueue.h> 82 #include <trace/events/tcp.h> 83 #include <linux/jump_label_ratelimit.h> 84 #include <net/busy_poll.h> 85 #include <net/mptcp.h> 86 87 int sysctl_tcp_max_orphans __read_mostly = NR_FILE; 88 89 #define FLAG_DATA 0x01 /* Incoming frame contained data. */ 90 #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ 91 #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ 92 #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ 93 #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ 94 #define FLAG_DATA_SACKED 0x20 /* New SACK. */ 95 #define FLAG_ECE 0x40 /* ECE in this ACK */ 96 #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */ 97 #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ 98 #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ 99 #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ 100 #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ 101 #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */ 102 #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ 103 #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ 104 #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */ 105 #define FLAG_ACK_MAYBE_DELAYED 0x10000 /* Likely a delayed ACK */ 106 #define FLAG_DSACK_TLP 0x20000 /* DSACK for tail loss probe */ 107 #define FLAG_TS_PROGRESS 0x40000 /* Positive timestamp delta */ 108 109 #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) 110 #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED| \ 111 FLAG_SND_UNA_ADVANCED) 112 #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE|FLAG_DSACKING_ACK) 113 #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) 114 115 #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) 116 #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) 117 118 #define REXMIT_NONE 0 /* no loss recovery to do */ 119 #define REXMIT_LOST 1 /* retransmit packets marked lost */ 120 #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */ 121 122 #if IS_ENABLED(CONFIG_TLS_DEVICE) 123 static DEFINE_STATIC_KEY_DEFERRED_FALSE(clean_acked_data_enabled, HZ); 124 125 void clean_acked_data_enable(struct tcp_sock *tp, 126 void (*cad)(struct sock *sk, u32 ack_seq)) 127 { 128 tp->tcp_clean_acked = cad; 129 static_branch_deferred_inc(&clean_acked_data_enabled); 130 } 131 EXPORT_SYMBOL_GPL(clean_acked_data_enable); 132 133 void clean_acked_data_disable(struct tcp_sock *tp) 134 { 135 static_branch_slow_dec_deferred(&clean_acked_data_enabled); 136 tp->tcp_clean_acked = NULL; 137 } 138 EXPORT_SYMBOL_GPL(clean_acked_data_disable); 139 140 void clean_acked_data_flush(void) 141 { 142 static_key_deferred_flush(&clean_acked_data_enabled); 143 } 144 EXPORT_SYMBOL_GPL(clean_acked_data_flush); 145 #endif 146 147 #ifdef CONFIG_CGROUP_BPF 148 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 149 { 150 struct bpf_sock_ops_kern sock_ops; 151 const struct tcp_sock *tp; 152 153 if (!cgroup_bpf_enabled(CGROUP_SOCK_OPS)) 154 return; 155 156 tp = tcp_sk(sk); 157 if (!(tp->rx_opt.saw_unknown && 158 BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_PARSE_UNKNOWN_HDR_OPT_CB_FLAG)) && 159 !BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_PARSE_ALL_HDR_OPT_CB_FLAG)) 160 return; 161 162 /* The skb will be handled in the 163 * bpf_skops_established() or 164 * bpf_skops_write_hdr_opt(). 165 */ 166 switch (sk->sk_state) { 167 case TCP_SYN_RECV: 168 case TCP_SYN_SENT: 169 case TCP_LISTEN: 170 return; 171 } 172 173 sock_owned_by_me(sk); 174 175 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 176 sock_ops.op = BPF_SOCK_OPS_PARSE_HDR_OPT_CB; 177 sock_ops.is_fullsock = 1; 178 sock_ops.is_locked_tcp_sock = 1; 179 sock_ops.sk = sk; 180 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb)); 181 182 __BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); 183 } 184 185 static void bpf_skops_established(struct sock *sk, int bpf_op, 186 struct sk_buff *skb) 187 { 188 struct bpf_sock_ops_kern sock_ops; 189 190 if (!cgroup_bpf_enabled(CGROUP_SOCK_OPS)) 191 return; 192 193 sock_owned_by_me(sk); 194 195 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 196 sock_ops.op = bpf_op; 197 sock_ops.is_fullsock = 1; 198 sock_ops.is_locked_tcp_sock = 1; 199 sock_ops.sk = sk; 200 /* sk with TCP_REPAIR_ON does not have skb in tcp_finish_connect */ 201 if (skb) 202 bpf_skops_init_skb(&sock_ops, skb, tcp_hdrlen(skb)); 203 204 __BPF_CGROUP_RUN_PROG_SOCK_OPS(&sock_ops); 205 } 206 #else 207 static void bpf_skops_parse_hdr(struct sock *sk, struct sk_buff *skb) 208 { 209 } 210 211 static void bpf_skops_established(struct sock *sk, int bpf_op, 212 struct sk_buff *skb) 213 { 214 } 215 #endif 216 217 static void bpf_tcp_ops_parse_hdr(struct sock *sk, struct sk_buff *skb) 218 { 219 const struct tcp_sock *tp; 220 221 if (!cgroup_bpf_enabled(CGROUP_TCP_SOCK_OPS)) 222 return; 223 224 switch (sk->sk_state) { 225 case TCP_SYN_RECV: 226 case TCP_SYN_SENT: 227 case TCP_LISTEN: 228 return; 229 } 230 231 tp = tcp_sk(sk); 232 233 if ((tp->rx_opt.saw_unknown && 234 BPF_TCP_OPS_TEST_FLAG(tp, PARSE_HDR_OPT_UNKNOWN)) || 235 BPF_TCP_OPS_TEST_FLAG(tp, PARSE_HDR_OPT_ALL)) 236 __bpf_tcp_ops_call(parse_hdr, sk, skb); 237 } 238 239 static __cold void tcp_gro_dev_warn(const struct sock *sk, const struct sk_buff *skb, 240 unsigned int len) 241 { 242 struct net_device *dev; 243 244 rcu_read_lock(); 245 dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif); 246 if (!dev || len >= READ_ONCE(dev->mtu)) 247 pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n", 248 dev ? dev->name : "Unknown driver"); 249 rcu_read_unlock(); 250 } 251 252 /* Adapt the MSS value used to make delayed ack decision to the 253 * real world. 254 */ 255 static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) 256 { 257 struct inet_connection_sock *icsk = inet_csk(sk); 258 const unsigned int lss = icsk->icsk_ack.last_seg_size; 259 unsigned int len; 260 261 icsk->icsk_ack.last_seg_size = 0; 262 263 /* skb->len may jitter because of SACKs, even if peer 264 * sends good full-sized frames. 265 */ 266 len = skb_shinfo(skb)->gso_size ? : skb->len; 267 if (len >= icsk->icsk_ack.rcv_mss) { 268 /* Note: divides are still a bit expensive. 269 * For the moment, only adjust scaling_ratio 270 * when we update icsk_ack.rcv_mss. 271 */ 272 if (unlikely(len != icsk->icsk_ack.rcv_mss)) { 273 u64 val = (u64)skb->len << TCP_RMEM_TO_WIN_SCALE; 274 u8 old_ratio = tcp_sk(sk)->scaling_ratio; 275 276 do_div(val, skb->truesize); 277 tcp_sk(sk)->scaling_ratio = val ? val : 1; 278 279 if (old_ratio != tcp_sk(sk)->scaling_ratio) { 280 struct tcp_sock *tp = tcp_sk(sk); 281 282 val = tcp_win_from_space(sk, sk->sk_rcvbuf); 283 WRITE_ONCE(tp->window_clamp, val); 284 285 if (tp->window_clamp < tp->rcvq_space.space) 286 tp->rcvq_space.space = tp->window_clamp; 287 } 288 } 289 icsk->icsk_ack.rcv_mss = min_t(unsigned int, len, 290 tcp_sk(sk)->advmss); 291 /* Account for possibly-removed options */ 292 DO_ONCE_LITE_IF(len > icsk->icsk_ack.rcv_mss + MAX_TCP_OPTION_SPACE, 293 tcp_gro_dev_warn, sk, skb, len); 294 /* If the skb has a len of exactly 1*MSS and has the PSH bit 295 * set then it is likely the end of an application write. So 296 * more data may not be arriving soon, and yet the data sender 297 * may be waiting for an ACK if cwnd-bound or using TX zero 298 * copy. So we set ICSK_ACK_PUSHED here so that 299 * tcp_cleanup_rbuf() will send an ACK immediately if the app 300 * reads all of the data and is not ping-pong. If len > MSS 301 * then this logic does not matter (and does not hurt) because 302 * tcp_cleanup_rbuf() will always ACK immediately if the app 303 * reads data and there is more than an MSS of unACKed data. 304 */ 305 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_PSH) 306 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 307 } else { 308 /* Otherwise, we make more careful check taking into account, 309 * that SACKs block is variable. 310 * 311 * "len" is invariant segment length, including TCP header. 312 */ 313 len += skb->data - skb_transport_header(skb); 314 if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || 315 /* If PSH is not set, packet should be 316 * full sized, provided peer TCP is not badly broken. 317 * This observation (if it is correct 8)) allows 318 * to handle super-low mtu links fairly. 319 */ 320 (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && 321 !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { 322 /* Subtract also invariant (if peer is RFC compliant), 323 * tcp header plus fixed timestamp option length. 324 * Resulting "len" is MSS free of SACK jitter. 325 */ 326 len -= tcp_sk(sk)->tcp_header_len; 327 icsk->icsk_ack.last_seg_size = len; 328 if (len == lss) { 329 icsk->icsk_ack.rcv_mss = len; 330 return; 331 } 332 } 333 if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) 334 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; 335 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; 336 } 337 } 338 339 static void tcp_incr_quickack(struct sock *sk, unsigned int max_quickacks) 340 { 341 struct inet_connection_sock *icsk = inet_csk(sk); 342 unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); 343 344 if (quickacks == 0) 345 quickacks = 2; 346 quickacks = min(quickacks, max_quickacks); 347 if (quickacks > icsk->icsk_ack.quick) 348 icsk->icsk_ack.quick = quickacks; 349 } 350 351 static void tcp_enter_quickack_mode(struct sock *sk, unsigned int max_quickacks) 352 { 353 struct inet_connection_sock *icsk = inet_csk(sk); 354 355 tcp_incr_quickack(sk, max_quickacks); 356 inet_csk_exit_pingpong_mode(sk); 357 icsk->icsk_ack.ato = TCP_ATO_MIN; 358 } 359 360 /* Send ACKs quickly, if "quick" count is not exhausted 361 * and the session is not interactive. 362 */ 363 364 static bool tcp_in_quickack_mode(struct sock *sk) 365 { 366 const struct inet_connection_sock *icsk = inet_csk(sk); 367 368 return icsk->icsk_ack.dst_quick_ack || 369 (icsk->icsk_ack.quick && !inet_csk_in_pingpong_mode(sk)); 370 } 371 372 static void tcp_data_ecn_check(struct sock *sk, const struct sk_buff *skb) 373 { 374 struct tcp_sock *tp = tcp_sk(sk); 375 376 if (tcp_ecn_disabled(tp)) 377 return; 378 379 switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { 380 case INET_ECN_NOT_ECT: 381 /* Funny extension: if ECT is not set on a segment, 382 * and we already seen ECT on a previous segment, 383 * it is probably a retransmit. 384 */ 385 if (tp->ecn_flags & TCP_ECN_SEEN) 386 tcp_enter_quickack_mode(sk, 2); 387 break; 388 case INET_ECN_CE: 389 if (tcp_ca_needs_ecn(sk)) 390 tcp_ca_event(sk, CA_EVENT_ECN_IS_CE); 391 392 if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR) && 393 tcp_ecn_mode_rfc3168(tp)) { 394 /* Better not delay acks, sender can have a very low cwnd */ 395 tcp_enter_quickack_mode(sk, 2); 396 tp->ecn_flags |= TCP_ECN_DEMAND_CWR; 397 } 398 /* As for RFC3168 ECN, the TCP_ECN_SEEN flag is set by 399 * tcp_data_ecn_check() when the ECN codepoint of 400 * received TCP data contains ECT(0), ECT(1), or CE. 401 */ 402 if (!tcp_ecn_mode_rfc3168(tp)) 403 break; 404 tp->ecn_flags |= TCP_ECN_SEEN; 405 break; 406 default: 407 if (tcp_ca_needs_ecn(sk)) 408 tcp_ca_event(sk, CA_EVENT_ECN_NO_CE); 409 if (!tcp_ecn_mode_rfc3168(tp)) 410 break; 411 tp->ecn_flags |= TCP_ECN_SEEN; 412 break; 413 } 414 } 415 416 /* Returns true if the byte counters can be used */ 417 static bool tcp_accecn_process_option(struct tcp_sock *tp, 418 const struct sk_buff *skb, 419 u32 delivered_bytes, int flag) 420 { 421 u8 estimate_ecnfield = tp->est_ecnfield; 422 bool ambiguous_ecn_bytes_incr = false; 423 bool first_changed = false; 424 unsigned int optlen; 425 bool order1, res; 426 unsigned int i; 427 u8 *ptr; 428 429 if (tcp_accecn_opt_fail_recv(tp)) 430 return false; 431 432 if (!(flag & FLAG_SLOWPATH) || !tp->rx_opt.accecn) { 433 if (!tp->saw_accecn_opt) { 434 /* Too late to enable after this point due to 435 * potential counter wraps 436 */ 437 if (tp->bytes_sent >= (1 << 23) - 1) { 438 u8 saw_opt = TCP_ACCECN_OPT_FAIL_SEEN; 439 440 tcp_accecn_saw_opt_fail_recv(tp, saw_opt); 441 } 442 return false; 443 } 444 445 if (estimate_ecnfield) { 446 u8 ecnfield = estimate_ecnfield - 1; 447 448 tp->delivered_ecn_bytes[ecnfield] += delivered_bytes; 449 return true; 450 } 451 return false; 452 } 453 454 ptr = skb_transport_header(skb) + tp->rx_opt.accecn; 455 optlen = ptr[1] - 2; 456 if (WARN_ON_ONCE(ptr[0] != TCPOPT_ACCECN0 && ptr[0] != TCPOPT_ACCECN1)) 457 return false; 458 order1 = (ptr[0] == TCPOPT_ACCECN1); 459 ptr += 2; 460 461 if (tp->saw_accecn_opt < TCP_ACCECN_OPT_COUNTER_SEEN) { 462 tp->saw_accecn_opt = tcp_accecn_option_init(skb, 463 tp->rx_opt.accecn); 464 if (tp->saw_accecn_opt == TCP_ACCECN_OPT_FAIL_SEEN) 465 tcp_accecn_fail_mode_set(tp, TCP_ACCECN_OPT_FAIL_RECV); 466 } 467 468 res = !!estimate_ecnfield; 469 for (i = 0; i < 3; i++) { 470 u32 init_offset; 471 u8 ecnfield; 472 s32 delta; 473 u32 *cnt; 474 475 if (optlen < TCPOLEN_ACCECN_PERFIELD) 476 break; 477 478 ecnfield = tcp_accecn_optfield_to_ecnfield(i, order1); 479 init_offset = tcp_accecn_field_init_offset(ecnfield); 480 cnt = &tp->delivered_ecn_bytes[ecnfield - 1]; 481 delta = tcp_update_ecn_bytes(cnt, ptr, init_offset); 482 if (delta && delta < 0) { 483 res = false; 484 ambiguous_ecn_bytes_incr = true; 485 } 486 if (delta && ecnfield != estimate_ecnfield) { 487 if (!first_changed) { 488 tp->est_ecnfield = ecnfield; 489 first_changed = true; 490 } else { 491 res = false; 492 ambiguous_ecn_bytes_incr = true; 493 } 494 } 495 496 optlen -= TCPOLEN_ACCECN_PERFIELD; 497 ptr += TCPOLEN_ACCECN_PERFIELD; 498 } 499 if (ambiguous_ecn_bytes_incr) 500 tp->est_ecnfield = 0; 501 502 return res; 503 } 504 505 static void tcp_count_delivered_ce(struct tcp_sock *tp, u32 ecn_count) 506 { 507 WRITE_ONCE(tp->delivered_ce, tp->delivered_ce + ecn_count); 508 } 509 510 /* Updates the delivered and delivered_ce counts */ 511 static void tcp_count_delivered(struct tcp_sock *tp, u32 delivered, 512 bool ece_ack) 513 { 514 WRITE_ONCE(tp->delivered, tp->delivered + delivered); 515 if (tcp_ecn_mode_rfc3168(tp) && ece_ack) 516 tcp_count_delivered_ce(tp, delivered); 517 } 518 519 #define PKTS_ACKED_WEIGHT 6 520 #define PKTS_ACKED_PREC 6 521 #define ACK_COMP_THRESH 4 522 523 /* Returns the ECN CE delta */ 524 static u32 __tcp_accecn_process(struct sock *sk, const struct sk_buff *skb, 525 u32 delivered_pkts, u32 delivered_bytes, 526 int flag) 527 { 528 u32 old_ceb = tcp_sk(sk)->delivered_ecn_bytes[INET_ECN_CE - 1]; 529 const struct tcphdr *th = tcp_hdr(skb); 530 struct tcp_sock *tp = tcp_sk(sk); 531 u32 delta, safe_delta, d_ceb; 532 bool opt_deltas_valid; 533 u32 corrected_ace; 534 u32 ewma; 535 536 /* Reordered ACK or uncertain due to lack of data to send and ts */ 537 if (!(flag & (FLAG_FORWARD_PROGRESS | FLAG_TS_PROGRESS))) 538 return 0; 539 540 opt_deltas_valid = tcp_accecn_process_option(tp, skb, 541 delivered_bytes, flag); 542 543 if (delivered_pkts) { 544 if (!tp->pkts_acked_ewma) { 545 ewma = delivered_pkts << PKTS_ACKED_PREC; 546 } else { 547 ewma = tp->pkts_acked_ewma; 548 ewma = (((ewma << PKTS_ACKED_WEIGHT) - ewma) + 549 (delivered_pkts << PKTS_ACKED_PREC)) >> 550 PKTS_ACKED_WEIGHT; 551 } 552 tp->pkts_acked_ewma = min_t(u32, ewma, 0xFFFFU); 553 } 554 555 if (!(flag & FLAG_SLOWPATH)) { 556 /* AccECN counter might overflow on large ACKs */ 557 if (delivered_pkts <= TCP_ACCECN_CEP_ACE_MASK) 558 return 0; 559 } 560 561 /* ACE field is not available during handshake */ 562 if (flag & FLAG_SYN_ACKED) 563 return 0; 564 565 if (tp->received_ce_pending >= TCP_ACCECN_ACE_MAX_DELTA) 566 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 567 568 corrected_ace = tcp_accecn_ace(th) - TCP_ACCECN_CEP_INIT_OFFSET; 569 delta = (corrected_ace - tp->delivered_ce) & TCP_ACCECN_CEP_ACE_MASK; 570 if (delivered_pkts <= TCP_ACCECN_CEP_ACE_MASK) 571 return delta; 572 573 safe_delta = delivered_pkts - 574 ((delivered_pkts - delta) & TCP_ACCECN_CEP_ACE_MASK); 575 576 if (opt_deltas_valid) { 577 d_ceb = tp->delivered_ecn_bytes[INET_ECN_CE - 1] - old_ceb; 578 if (!d_ceb) 579 return delta; 580 581 if ((delivered_pkts >= (TCP_ACCECN_CEP_ACE_MASK + 1) * 2) && 582 (tcp_is_sack(tp) || 583 ((1 << inet_csk(sk)->icsk_ca_state) & 584 (TCPF_CA_Open | TCPF_CA_CWR)))) { 585 u32 est_d_cep; 586 587 if (delivered_bytes <= d_ceb) 588 return safe_delta; 589 590 est_d_cep = DIV_ROUND_UP_ULL((u64)d_ceb * 591 delivered_pkts, 592 delivered_bytes); 593 return min(safe_delta, 594 delta + 595 (est_d_cep & ~TCP_ACCECN_CEP_ACE_MASK)); 596 } 597 598 if (d_ceb > delta * tp->mss_cache) 599 return safe_delta; 600 if (d_ceb < 601 safe_delta * tp->mss_cache >> TCP_ACCECN_SAFETY_SHIFT) 602 return delta; 603 } else if (tp->pkts_acked_ewma > (ACK_COMP_THRESH << PKTS_ACKED_PREC)) 604 return delta; 605 606 return safe_delta; 607 } 608 609 static u32 tcp_accecn_process(struct sock *sk, const struct sk_buff *skb, 610 u32 delivered_pkts, u32 delivered_bytes, 611 int *flag) 612 { 613 struct tcp_sock *tp = tcp_sk(sk); 614 u32 delta; 615 616 delta = __tcp_accecn_process(sk, skb, delivered_pkts, 617 delivered_bytes, *flag); 618 if (delta > 0) { 619 tcp_count_delivered_ce(tp, delta); 620 *flag |= FLAG_ECE; 621 /* Recalculate header predictor */ 622 if (tp->pred_flags) 623 tcp_fast_path_on(tp); 624 } 625 return delta; 626 } 627 628 /* Buffer size and advertised window tuning. 629 * 630 * 1. Tuning sk->sk_sndbuf, when connection enters established state. 631 */ 632 633 static void tcp_sndbuf_expand(struct sock *sk) 634 { 635 const struct tcp_sock *tp = tcp_sk(sk); 636 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 637 int sndmem, per_mss; 638 u32 nr_segs; 639 640 /* Worst case is non GSO/TSO : each frame consumes one skb 641 * and skb->head is kmalloced using power of two area of memory 642 */ 643 per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + 644 MAX_TCP_HEADER + 645 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 646 647 per_mss = roundup_pow_of_two(per_mss) + 648 SKB_DATA_ALIGN(sizeof(struct sk_buff)); 649 650 nr_segs = max_t(u32, TCP_INIT_CWND, tcp_snd_cwnd(tp)); 651 nr_segs = max_t(u32, nr_segs, tp->reordering + 1); 652 653 /* Fast Recovery (RFC 5681 3.2) : 654 * Cubic needs 1.7 factor, rounded to 2 to include 655 * extra cushion (application might react slowly to EPOLLOUT) 656 */ 657 sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2; 658 sndmem *= nr_segs * per_mss; 659 660 if (sk->sk_sndbuf < sndmem) 661 WRITE_ONCE(sk->sk_sndbuf, 662 min(sndmem, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[2]))); 663 } 664 665 /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) 666 * 667 * All tcp_full_space() is split to two parts: "network" buffer, allocated 668 * forward and advertised in receiver window (tp->rcv_wnd) and 669 * "application buffer", required to isolate scheduling/application 670 * latencies from network. 671 * window_clamp is maximal advertised window. It can be less than 672 * tcp_full_space(), in this case tcp_full_space() - window_clamp 673 * is reserved for "application" buffer. The less window_clamp is 674 * the smoother our behaviour from viewpoint of network, but the lower 675 * throughput and the higher sensitivity of the connection to losses. 8) 676 * 677 * rcv_ssthresh is more strict window_clamp used at "slow start" 678 * phase to predict further behaviour of this connection. 679 * It is used for two goals: 680 * - to enforce header prediction at sender, even when application 681 * requires some significant "application buffer". It is check #1. 682 * - to prevent pruning of receive queue because of misprediction 683 * of receiver window. Check #2. 684 * 685 * The scheme does not work when sender sends good segments opening 686 * window and then starts to feed us spaghetti. But it should work 687 * in common situations. Otherwise, we have to rely on queue collapsing. 688 */ 689 690 /* Slow part of check#2. */ 691 static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb, 692 unsigned int skbtruesize) 693 { 694 const struct tcp_sock *tp = tcp_sk(sk); 695 /* Optimize this! */ 696 int truesize = tcp_win_from_space(sk, skbtruesize) >> 1; 697 int window = tcp_win_from_space(sk, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])) >> 1; 698 699 while (tp->rcv_ssthresh <= window) { 700 if (truesize <= skb->len) 701 return 2 * inet_csk(sk)->icsk_ack.rcv_mss; 702 703 truesize >>= 1; 704 window >>= 1; 705 } 706 return 0; 707 } 708 709 /* Even if skb appears to have a bad len/truesize ratio, TCP coalescing 710 * can play nice with us, as sk_buff and skb->head might be either 711 * freed or shared with up to MAX_SKB_FRAGS segments. 712 * Only give a boost to drivers using page frag(s) to hold the frame(s), 713 * and if no payload was pulled in skb->head before reaching us. 714 */ 715 static u32 truesize_adjust(bool adjust, const struct sk_buff *skb) 716 { 717 u32 truesize = skb->truesize; 718 719 if (adjust && !skb_headlen(skb)) { 720 truesize -= SKB_TRUESIZE(skb_end_offset(skb)); 721 /* paranoid check, some drivers might be buggy */ 722 if (unlikely((int)truesize < (int)skb->len)) 723 truesize = skb->truesize; 724 } 725 return truesize; 726 } 727 728 static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb, 729 bool adjust) 730 { 731 struct tcp_sock *tp = tcp_sk(sk); 732 int room; 733 734 room = min_t(int, tp->window_clamp, tcp_space(sk)) - tp->rcv_ssthresh; 735 736 if (room <= 0) 737 return; 738 739 /* Check #1 */ 740 if (!tcp_under_memory_pressure(sk)) { 741 unsigned int truesize = truesize_adjust(adjust, skb); 742 int incr; 743 744 /* Check #2. Increase window, if skb with such overhead 745 * will fit to rcvbuf in future. 746 */ 747 if (tcp_win_from_space(sk, truesize) <= skb->len) 748 incr = 2 * tp->advmss; 749 else 750 incr = __tcp_grow_window(sk, skb, truesize); 751 752 if (incr) { 753 incr = max_t(int, incr, 2 * skb->len); 754 tp->rcv_ssthresh += min(room, incr); 755 inet_csk(sk)->icsk_ack.quick |= 1; 756 } 757 } else { 758 /* Under pressure: 759 * Adjust rcv_ssthresh according to reserved mem 760 */ 761 tcp_adjust_rcv_ssthresh(sk); 762 } 763 } 764 765 /* 3. Try to fixup all. It is made immediately after connection enters 766 * established state. 767 */ 768 static void tcp_init_buffer_space(struct sock *sk) 769 { 770 int tcp_app_win = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_app_win); 771 struct tcp_sock *tp = tcp_sk(sk); 772 int maxwin; 773 774 if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) 775 tcp_sndbuf_expand(sk); 776 777 tcp_mstamp_refresh(tp); 778 tp->rcvq_space.time = tp->tcp_mstamp; 779 tp->rcvq_space.seq = tp->copied_seq; 780 781 maxwin = tcp_full_space(sk); 782 783 if (tp->window_clamp >= maxwin) { 784 WRITE_ONCE(tp->window_clamp, maxwin); 785 786 if (tcp_app_win && maxwin > 4 * tp->advmss) 787 WRITE_ONCE(tp->window_clamp, 788 max(maxwin - (maxwin >> tcp_app_win), 789 4 * tp->advmss)); 790 } 791 792 /* Force reservation of one segment. */ 793 if (tcp_app_win && 794 tp->window_clamp > 2 * tp->advmss && 795 tp->window_clamp + tp->advmss > maxwin) 796 WRITE_ONCE(tp->window_clamp, 797 max(2 * tp->advmss, maxwin - tp->advmss)); 798 799 tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); 800 tp->snd_cwnd_stamp = tcp_jiffies32; 801 tp->rcvq_space.space = min3(tp->rcv_ssthresh, tp->rcv_wnd, 802 (u32)TCP_INIT_CWND * tp->advmss); 803 } 804 805 /* 4. Recalculate window clamp after socket hit its memory bounds. */ 806 static void tcp_clamp_window(struct sock *sk) 807 { 808 struct tcp_sock *tp = tcp_sk(sk); 809 struct inet_connection_sock *icsk = inet_csk(sk); 810 struct net *net = sock_net(sk); 811 int rmem2; 812 813 icsk->icsk_ack.quick = 0; 814 rmem2 = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]); 815 816 if (sk->sk_rcvbuf < rmem2 && 817 !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && 818 !tcp_under_memory_pressure(sk) && 819 sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { 820 WRITE_ONCE(sk->sk_rcvbuf, 821 min(atomic_read(&sk->sk_rmem_alloc), rmem2)); 822 } 823 if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) 824 tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); 825 } 826 827 /* Initialize RCV_MSS value. 828 * RCV_MSS is an our guess about MSS used by the peer. 829 * We haven't any direct information about the MSS. 830 * It's better to underestimate the RCV_MSS rather than overestimate. 831 * Overestimations make us ACKing less frequently than needed. 832 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). 833 */ 834 void tcp_initialize_rcv_mss(struct sock *sk) 835 { 836 const struct tcp_sock *tp = tcp_sk(sk); 837 unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); 838 839 hint = min(hint, tp->rcv_wnd / 2); 840 hint = min(hint, TCP_MSS_DEFAULT); 841 hint = max(hint, TCP_MIN_MSS); 842 843 inet_csk(sk)->icsk_ack.rcv_mss = hint; 844 } 845 846 /* Receiver "autotuning" code. 847 * 848 * The algorithm for RTT estimation w/o timestamps is based on 849 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. 850 * <https://public.lanl.gov/radiant/pubs.html#DRS> 851 * 852 * More detail on this code can be found at 853 * <http://staff.psc.edu/jheffner/>, 854 * though this reference is out of date. A new paper 855 * is pending. 856 */ 857 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) 858 { 859 u32 new_sample, old_sample = tp->rcv_rtt_est.rtt_us; 860 long m = sample << 3; 861 862 if (old_sample == 0 || m < old_sample) { 863 new_sample = m; 864 } else { 865 /* If we sample in larger samples in the non-timestamp 866 * case, we could grossly overestimate the RTT especially 867 * with chatty applications or bulk transfer apps which 868 * are stalled on filesystem I/O. 869 * 870 * Also, since we are only going for a minimum in the 871 * non-timestamp case, we do not smooth things out 872 * else with timestamps disabled convergence takes too 873 * long. 874 */ 875 if (win_dep) 876 return; 877 /* Do not use this sample if receive queue is not empty. */ 878 if (tp->rcv_nxt != tp->copied_seq) 879 return; 880 new_sample = old_sample - (old_sample >> 3) + sample; 881 } 882 883 tp->rcv_rtt_est.rtt_us = new_sample; 884 } 885 886 static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) 887 { 888 u32 delta_us; 889 890 if (tp->rcv_rtt_est.time == 0) 891 goto new_measure; 892 if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) 893 return; 894 delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time); 895 if (!delta_us) 896 delta_us = 1; 897 tcp_rcv_rtt_update(tp, delta_us, 1); 898 899 new_measure: 900 tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; 901 tp->rcv_rtt_est.time = tp->tcp_mstamp; 902 } 903 904 static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp, u32 min_delta) 905 { 906 u32 delta, delta_us; 907 908 delta = tcp_time_stamp_ts(tp) - tp->rx_opt.rcv_tsecr; 909 if (tp->tcp_usec_ts) 910 return delta; 911 912 if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) { 913 if (!delta) 914 delta = min_delta; 915 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ); 916 return delta_us; 917 } 918 return -1; 919 } 920 921 static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, 922 const struct sk_buff *skb) 923 { 924 struct tcp_sock *tp = tcp_sk(sk); 925 926 if (tp->rx_opt.rcv_tsecr == tp->rcv_rtt_last_tsecr) 927 return; 928 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 929 930 if (TCP_SKB_CB(skb)->end_seq - 931 TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) { 932 s32 delta = tcp_rtt_tsopt_us(tp, 0); 933 934 if (delta > 0) 935 tcp_rcv_rtt_update(tp, delta, 0); 936 } 937 } 938 939 void tcp_rcvbuf_grow(struct sock *sk, u32 newval) 940 { 941 const struct net *net = sock_net(sk); 942 struct tcp_sock *tp = tcp_sk(sk); 943 u32 rcvwin, rcvbuf, cap, oldval; 944 u32 rtt_threshold, rtt_us; 945 u64 grow; 946 947 oldval = tp->rcvq_space.space; 948 tp->rcvq_space.space = newval; 949 950 if (!READ_ONCE(net->ipv4.sysctl_tcp_moderate_rcvbuf) || 951 (sk->sk_userlocks & SOCK_RCVBUF_LOCK)) 952 return; 953 954 /* DRS is always one RTT late. */ 955 rcvwin = newval << 1; 956 957 rtt_us = tp->rcv_rtt_est.rtt_us >> 3; 958 rtt_threshold = READ_ONCE(net->ipv4.sysctl_tcp_rcvbuf_low_rtt); 959 if (rtt_us < rtt_threshold) { 960 /* For small RTT, we set @grow to rcvwin * rtt_us/rtt_threshold. 961 * It might take few additional ms to reach 'line rate', 962 * but will avoid sk_rcvbuf inflation and poor cache use. 963 */ 964 grow = div_u64((u64)rcvwin * rtt_us, rtt_threshold); 965 } else { 966 /* slow start: allow the sender to double its rate. */ 967 grow = div_u64(((u64)rcvwin << 1) * (newval - oldval), oldval); 968 } 969 rcvwin += grow; 970 971 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) 972 rcvwin += TCP_SKB_CB(tp->ooo_last_skb)->end_seq - tp->rcv_nxt; 973 974 cap = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]); 975 976 rcvbuf = min_t(u32, tcp_space_from_win(sk, rcvwin), cap); 977 if (rcvbuf > sk->sk_rcvbuf) { 978 WRITE_ONCE(sk->sk_rcvbuf, rcvbuf); 979 /* Make the window clamp follow along. */ 980 WRITE_ONCE(tp->window_clamp, 981 tcp_win_from_space(sk, rcvbuf)); 982 } 983 } 984 /* 985 * This function should be called every time data is copied to user space. 986 * It calculates the appropriate TCP receive buffer space. 987 */ 988 void tcp_rcv_space_adjust(struct sock *sk) 989 { 990 struct tcp_sock *tp = tcp_sk(sk); 991 int time, inq, copied; 992 993 trace_tcp_rcv_space_adjust(sk); 994 995 if (unlikely(!tp->rcv_rtt_est.rtt_us)) 996 return; 997 998 /* We do not refresh tp->tcp_mstamp here. 999 * Some platforms have expensive ktime_get() implementations. 1000 * Using the last cached value is enough for DRS. 1001 */ 1002 time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time); 1003 if (time < (tp->rcv_rtt_est.rtt_us >> 3)) 1004 return; 1005 1006 /* Number of bytes copied to user in last RTT */ 1007 copied = tp->copied_seq - tp->rcvq_space.seq; 1008 /* Number of bytes in receive queue. */ 1009 inq = tp->rcv_nxt - tp->copied_seq; 1010 copied -= inq; 1011 if (copied <= tp->rcvq_space.space) 1012 goto new_measure; 1013 1014 trace_tcp_rcvbuf_grow(sk, time); 1015 1016 tcp_rcvbuf_grow(sk, copied); 1017 1018 new_measure: 1019 tp->rcvq_space.seq = tp->copied_seq; 1020 tp->rcvq_space.time = tp->tcp_mstamp; 1021 } 1022 1023 static void tcp_save_lrcv_flowlabel(struct sock *sk, const struct sk_buff *skb) 1024 { 1025 #if IS_ENABLED(CONFIG_IPV6) 1026 struct inet_connection_sock *icsk = inet_csk(sk); 1027 1028 if (skb->protocol == htons(ETH_P_IPV6)) 1029 icsk->icsk_ack.lrcv_flowlabel = ntohl(ip6_flowlabel(ipv6_hdr(skb))); 1030 #endif 1031 } 1032 1033 /* There is something which you must keep in mind when you analyze the 1034 * behavior of the tp->ato delayed ack timeout interval. When a 1035 * connection starts up, we want to ack as quickly as possible. The 1036 * problem is that "good" TCP's do slow start at the beginning of data 1037 * transmission. The means that until we send the first few ACK's the 1038 * sender will sit on his end and only queue most of his data, because 1039 * he can only send snd_cwnd unacked packets at any given time. For 1040 * each ACK we send, he increments snd_cwnd and transmits more of his 1041 * queue. -DaveM 1042 */ 1043 static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) 1044 { 1045 struct tcp_sock *tp = tcp_sk(sk); 1046 struct inet_connection_sock *icsk = inet_csk(sk); 1047 u32 now; 1048 1049 inet_csk_schedule_ack(sk); 1050 1051 tcp_measure_rcv_mss(sk, skb); 1052 1053 tcp_rcv_rtt_measure(tp); 1054 1055 now = tcp_jiffies32; 1056 1057 if (!icsk->icsk_ack.ato) { 1058 /* The _first_ data packet received, initialize 1059 * delayed ACK engine. 1060 */ 1061 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 1062 icsk->icsk_ack.ato = TCP_ATO_MIN; 1063 } else { 1064 int m = now - icsk->icsk_ack.lrcvtime; 1065 1066 if (m <= TCP_ATO_MIN / 2) { 1067 /* The fastest case is the first. */ 1068 icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; 1069 } else if (m < icsk->icsk_ack.ato) { 1070 icsk->icsk_ack.ato = min3((icsk->icsk_ack.ato >> 1) + (u32)m, 1071 icsk->icsk_rto, 1072 (u32)TCP_DELACK_MAX); 1073 } else if (m > icsk->icsk_rto) { 1074 /* Too long gap. Apparently sender failed to 1075 * restart window, so that we send ACKs quickly. 1076 */ 1077 tcp_incr_quickack(sk, TCP_MAX_QUICKACKS); 1078 } 1079 } 1080 icsk->icsk_ack.lrcvtime = now; 1081 tcp_save_lrcv_flowlabel(sk, skb); 1082 1083 tcp_data_ecn_check(sk, skb); 1084 1085 if (skb->len >= 128) 1086 tcp_grow_window(sk, skb, true); 1087 } 1088 1089 /* Called to compute a smoothed rtt estimate. The data fed to this 1090 * routine either comes from timestamps, or from segments that were 1091 * known _not_ to have been retransmitted [see Karn/Partridge 1092 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 1093 * piece by Van Jacobson. 1094 * NOTE: the next three routines used to be one big routine. 1095 * To save cycles in the RFC 1323 implementation it was better to break 1096 * it up into three procedures. -- erics 1097 */ 1098 static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) 1099 { 1100 struct tcp_sock *tp = tcp_sk(sk); 1101 long m = mrtt_us; /* RTT */ 1102 u32 srtt = tp->srtt_us; 1103 1104 /* The following amusing code comes from Jacobson's 1105 * article in SIGCOMM '88. Note that rtt and mdev 1106 * are scaled versions of rtt and mean deviation. 1107 * This is designed to be as fast as possible 1108 * m stands for "measurement". 1109 * 1110 * On a 1990 paper the rto value is changed to: 1111 * RTO = rtt + 4 * mdev 1112 * 1113 * Funny. This algorithm seems to be very broken. 1114 * These formulae increase RTO, when it should be decreased, increase 1115 * too slowly, when it should be increased quickly, decrease too quickly 1116 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely 1117 * does not matter how to _calculate_ it. Seems, it was trap 1118 * that VJ failed to avoid. 8) 1119 */ 1120 if (srtt != 0) { 1121 m -= (srtt >> 3); /* m is now error in rtt est */ 1122 srtt += m; /* rtt = 7/8 rtt + 1/8 new */ 1123 if (m < 0) { 1124 m = -m; /* m is now abs(error) */ 1125 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 1126 /* This is similar to one of Eifel findings. 1127 * Eifel blocks mdev updates when rtt decreases. 1128 * This solution is a bit different: we use finer gain 1129 * for mdev in this case (alpha*beta). 1130 * Like Eifel it also prevents growth of rto, 1131 * but also it limits too fast rto decreases, 1132 * happening in pure Eifel. 1133 */ 1134 if (m > 0) 1135 m >>= 3; 1136 } else { 1137 m -= (tp->mdev_us >> 2); /* similar update on mdev */ 1138 } 1139 tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ 1140 if (tp->mdev_us > tp->mdev_max_us) { 1141 tp->mdev_max_us = tp->mdev_us; 1142 if (tp->mdev_max_us > tp->rttvar_us) 1143 tp->rttvar_us = tp->mdev_max_us; 1144 } 1145 if (after(tp->snd_una, tp->rtt_seq)) { 1146 if (tp->mdev_max_us < tp->rttvar_us) 1147 tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; 1148 tp->rtt_seq = tp->snd_nxt; 1149 tp->mdev_max_us = tcp_rto_min_us(sk); 1150 1151 tcp_bpf_rtt(sk, mrtt_us, srtt); 1152 } 1153 } else { 1154 /* no previous measure. */ 1155 srtt = m << 3; /* take the measured time to be rtt */ 1156 tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ 1157 tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); 1158 tp->mdev_max_us = tp->rttvar_us; 1159 tp->rtt_seq = tp->snd_nxt; 1160 1161 tcp_bpf_rtt(sk, mrtt_us, srtt); 1162 } 1163 WRITE_ONCE(tp->srtt_us, max(1U, srtt)); 1164 } 1165 1166 void tcp_update_pacing_rate(struct sock *sk) 1167 { 1168 const struct tcp_sock *tp = tcp_sk(sk); 1169 u64 rate; 1170 1171 /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ 1172 rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); 1173 1174 /* current rate is (cwnd * mss) / srtt 1175 * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. 1176 * In Congestion Avoidance phase, set it to 120 % the current rate. 1177 * 1178 * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) 1179 * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching 1180 * end of slow start and should slow down. 1181 */ 1182 if (tcp_snd_cwnd(tp) < tp->snd_ssthresh / 2) 1183 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ss_ratio); 1184 else 1185 rate *= READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_pacing_ca_ratio); 1186 1187 rate *= max(tcp_snd_cwnd(tp), tp->packets_out); 1188 1189 if (likely(tp->srtt_us)) 1190 do_div(rate, tp->srtt_us); 1191 1192 /* WRITE_ONCE() is needed because sch_fq fetches sk_pacing_rate 1193 * without any lock. We want to make sure compiler wont store 1194 * intermediate values in this location. 1195 */ 1196 WRITE_ONCE(sk->sk_pacing_rate, 1197 min_t(u64, rate, READ_ONCE(sk->sk_max_pacing_rate))); 1198 } 1199 1200 /* Calculate rto without backoff. This is the second half of Van Jacobson's 1201 * routine referred to above. 1202 */ 1203 void tcp_set_rto(struct sock *sk) 1204 { 1205 const struct tcp_sock *tp = tcp_sk(sk); 1206 /* Old crap is replaced with new one. 8) 1207 * 1208 * More seriously: 1209 * 1. If rtt variance happened to be less 50msec, it is hallucination. 1210 * It cannot be less due to utterly erratic ACK generation made 1211 * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ 1212 * to do with delayed acks, because at cwnd>2 true delack timeout 1213 * is invisible. Actually, Linux-2.4 also generates erratic 1214 * ACKs in some circumstances. 1215 */ 1216 inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); 1217 1218 /* 2. Fixups made earlier cannot be right. 1219 * If we do not estimate RTO correctly without them, 1220 * all the algo is pure shit and should be replaced 1221 * with correct one. It is exactly, which we pretend to do. 1222 */ 1223 1224 /* NOTE: clamping at TCP_RTO_MIN is not required, current algo 1225 * guarantees that rto is higher. 1226 */ 1227 tcp_bound_rto(sk); 1228 } 1229 1230 __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) 1231 { 1232 __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); 1233 1234 if (!cwnd) 1235 cwnd = TCP_INIT_CWND; 1236 return min_t(__u32, cwnd, tp->snd_cwnd_clamp); 1237 } 1238 1239 struct tcp_sacktag_state { 1240 /* Timestamps for earliest and latest never-retransmitted segment 1241 * that was SACKed. RTO needs the earliest RTT to stay conservative, 1242 * but congestion control should still get an accurate delay signal. 1243 */ 1244 u64 first_sackt; 1245 u64 last_sackt; 1246 u32 reord; 1247 u32 sack_delivered; 1248 u32 delivered_bytes; 1249 int flag; 1250 unsigned int mss_now; 1251 struct rate_sample *rate; 1252 }; 1253 1254 /* Take a notice that peer is sending D-SACKs. Skip update of data delivery 1255 * and spurious retransmission information if this DSACK is unlikely caused by 1256 * sender's action: 1257 * - DSACKed sequence range is larger than maximum receiver's window. 1258 * - Total no. of DSACKed segments exceed the total no. of retransmitted segs. 1259 */ 1260 static u32 tcp_dsack_seen(struct tcp_sock *tp, u32 start_seq, 1261 u32 end_seq, struct tcp_sacktag_state *state) 1262 { 1263 u32 seq_len, dup_segs = 1; 1264 1265 if (!before(start_seq, end_seq)) 1266 return 0; 1267 1268 seq_len = end_seq - start_seq; 1269 /* Dubious DSACK: DSACKed range greater than maximum advertised rwnd */ 1270 if (seq_len > tp->max_window) 1271 return 0; 1272 if (seq_len > tp->mss_cache) 1273 dup_segs = DIV_ROUND_UP(seq_len, tp->mss_cache); 1274 else if (tp->tlp_high_seq && tp->tlp_high_seq == end_seq) 1275 state->flag |= FLAG_DSACK_TLP; 1276 1277 WRITE_ONCE(tp->dsack_dups, tp->dsack_dups + dup_segs); 1278 /* Skip the DSACK if dup segs weren't retransmitted by sender */ 1279 if (tp->dsack_dups > tp->total_retrans) 1280 return 0; 1281 1282 tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; 1283 /* We increase the RACK ordering window in rounds where we receive 1284 * DSACKs that may have been due to reordering causing RACK to trigger 1285 * a spurious fast recovery. Thus RACK ignores DSACKs that happen 1286 * without having seen reordering, or that match TLP probes (TLP 1287 * is timer-driven, not triggered by RACK). 1288 */ 1289 if (tp->reord_seen && !(state->flag & FLAG_DSACK_TLP)) 1290 tp->rack.dsack_seen = 1; 1291 1292 state->flag |= FLAG_DSACKING_ACK; 1293 /* A spurious retransmission is delivered */ 1294 state->sack_delivered += dup_segs; 1295 1296 return dup_segs; 1297 } 1298 1299 /* It's reordering when higher sequence was delivered (i.e. sacked) before 1300 * some lower never-retransmitted sequence ("low_seq"). The maximum reordering 1301 * distance is approximated in full-mss packet distance ("reordering"). 1302 */ 1303 static void tcp_check_sack_reordering(struct sock *sk, const u32 low_seq, 1304 const int ts) 1305 { 1306 struct tcp_sock *tp = tcp_sk(sk); 1307 const u32 mss = tp->mss_cache; 1308 u32 fack, metric; 1309 1310 fack = tcp_highest_sack_seq(tp); 1311 if (!before(low_seq, fack)) 1312 return; 1313 1314 metric = fack - low_seq; 1315 if ((metric > tp->reordering * mss) && mss) { 1316 #if FASTRETRANS_DEBUG > 1 1317 pr_debug("Disorder%d %d %u f%u s%u rr%d\n", 1318 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, 1319 tp->reordering, 1320 0, 1321 tp->sacked_out, 1322 tp->undo_marker ? tp->undo_retrans : 0); 1323 #endif 1324 WRITE_ONCE(tp->reordering, 1325 min_t(u32, (metric + mss - 1) / mss, 1326 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering))); 1327 } 1328 1329 /* This exciting event is worth to be remembered. 8) */ 1330 WRITE_ONCE(tp->reord_seen, tp->reord_seen + 1); 1331 NET_INC_STATS(sock_net(sk), 1332 ts ? LINUX_MIB_TCPTSREORDER : LINUX_MIB_TCPSACKREORDER); 1333 } 1334 1335 /* This must be called before lost_out or retrans_out are updated 1336 * on a new loss, because we want to know if all skbs previously 1337 * known to be lost have already been retransmitted, indicating 1338 * that this newly lost skb is our next skb to retransmit. 1339 */ 1340 static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) 1341 { 1342 if ((!tp->retransmit_skb_hint && tp->retrans_out >= tp->lost_out) || 1343 (tp->retransmit_skb_hint && 1344 before(TCP_SKB_CB(skb)->seq, 1345 TCP_SKB_CB(tp->retransmit_skb_hint)->seq))) 1346 tp->retransmit_skb_hint = skb; 1347 } 1348 1349 /* Sum the number of packets on the wire we have marked as lost, and 1350 * notify the congestion control module that the given skb was marked lost. 1351 */ 1352 static void tcp_notify_skb_loss_event(struct tcp_sock *tp, const struct sk_buff *skb) 1353 { 1354 tp->lost += tcp_skb_pcount(skb); 1355 } 1356 1357 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb) 1358 { 1359 __u8 sacked = TCP_SKB_CB(skb)->sacked; 1360 struct tcp_sock *tp = tcp_sk(sk); 1361 1362 if (sacked & TCPCB_SACKED_ACKED) 1363 return; 1364 1365 tcp_verify_retransmit_hint(tp, skb); 1366 if (sacked & TCPCB_LOST) { 1367 if (sacked & TCPCB_SACKED_RETRANS) { 1368 /* Account for retransmits that are lost again */ 1369 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1370 tp->retrans_out -= tcp_skb_pcount(skb); 1371 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT, 1372 tcp_skb_pcount(skb)); 1373 tcp_notify_skb_loss_event(tp, skb); 1374 } 1375 } else { 1376 tp->lost_out += tcp_skb_pcount(skb); 1377 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 1378 tcp_notify_skb_loss_event(tp, skb); 1379 } 1380 } 1381 1382 /* This procedure tags the retransmission queue when SACKs arrive. 1383 * 1384 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). 1385 * Packets in queue with these bits set are counted in variables 1386 * sacked_out, retrans_out and lost_out, correspondingly. 1387 * 1388 * Valid combinations are: 1389 * Tag InFlight Description 1390 * 0 1 - orig segment is in flight. 1391 * S 0 - nothing flies, orig reached receiver. 1392 * L 0 - nothing flies, orig lost by net. 1393 * R 2 - both orig and retransmit are in flight. 1394 * L|R 1 - orig is lost, retransmit is in flight. 1395 * S|R 1 - orig reached receiver, retrans is still in flight. 1396 * (L|S|R is logically valid, it could occur when L|R is sacked, 1397 * but it is equivalent to plain S and code short-circuits it to S. 1398 * L|S is logically invalid, it would mean -1 packet in flight 8)) 1399 * 1400 * These 6 states form finite state machine, controlled by the following events: 1401 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) 1402 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) 1403 * 3. Loss detection event of two flavors: 1404 * A. Scoreboard estimator decided the packet is lost. 1405 * A'. Reno "three dupacks" marks head of queue lost. 1406 * B. SACK arrives sacking SND.NXT at the moment, when the 1407 * segment was retransmitted. 1408 * 4. D-SACK added new rule: D-SACK changes any tag to S. 1409 * 1410 * It is pleasant to note, that state diagram turns out to be commutative, 1411 * so that we are allowed not to be bothered by order of our actions, 1412 * when multiple events arrive simultaneously. (see the function below). 1413 * 1414 * Reordering detection. 1415 * -------------------- 1416 * Reordering metric is maximal distance, which a packet can be displaced 1417 * in packet stream. With SACKs we can estimate it: 1418 * 1419 * 1. SACK fills old hole and the corresponding segment was not 1420 * ever retransmitted -> reordering. Alas, we cannot use it 1421 * when segment was retransmitted. 1422 * 2. The last flaw is solved with D-SACK. D-SACK arrives 1423 * for retransmitted and already SACKed segment -> reordering.. 1424 * Both of these heuristics are not used in Loss state, when we cannot 1425 * account for retransmits accurately. 1426 * 1427 * SACK block validation. 1428 * ---------------------- 1429 * 1430 * SACK block range validation checks that the received SACK block fits to 1431 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. 1432 * Note that SND.UNA is not included to the range though being valid because 1433 * it means that the receiver is rather inconsistent with itself reporting 1434 * SACK reneging when it should advance SND.UNA. Such SACK block this is 1435 * perfectly valid, however, in light of RFC2018 which explicitly states 1436 * that "SACK block MUST reflect the newest segment. Even if the newest 1437 * segment is going to be discarded ...", not that it looks very clever 1438 * in case of head skb. Due to potentional receiver driven attacks, we 1439 * choose to avoid immediate execution of a walk in write queue due to 1440 * reneging and defer head skb's loss recovery to standard loss recovery 1441 * procedure that will eventually trigger (nothing forbids us doing this). 1442 * 1443 * Implements also blockage to start_seq wrap-around. Problem lies in the 1444 * fact that though start_seq (s) is before end_seq (i.e., not reversed), 1445 * there's no guarantee that it will be before snd_nxt (n). The problem 1446 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt 1447 * wrap (s_w): 1448 * 1449 * <- outs wnd -> <- wrapzone -> 1450 * u e n u_w e_w s n_w 1451 * | | | | | | | 1452 * |<------------+------+----- TCP seqno space --------------+---------->| 1453 * ...-- <2^31 ->| |<--------... 1454 * ...---- >2^31 ------>| |<--------... 1455 * 1456 * Current code wouldn't be vulnerable but it's better still to discard such 1457 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat 1458 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in 1459 * snd_nxt wrap -> snd_una region will then become "well defined", i.e., 1460 * equal to the ideal case (infinite seqno space without wrap caused issues). 1461 * 1462 * With D-SACK the lower bound is extended to cover sequence space below 1463 * SND.UNA down to undo_marker, which is the last point of interest. Yet 1464 * again, D-SACK block must not to go across snd_una (for the same reason as 1465 * for the normal SACK blocks, explained above). But there all simplicity 1466 * ends, TCP might receive valid D-SACKs below that. As long as they reside 1467 * fully below undo_marker they do not affect behavior in anyway and can 1468 * therefore be safely ignored. In rare cases (which are more or less 1469 * theoretical ones), the D-SACK will nicely cross that boundary due to skb 1470 * fragmentation and packet reordering past skb's retransmission. To consider 1471 * them correctly, the acceptable range must be extended even more though 1472 * the exact amount is rather hard to quantify. However, tp->max_window can 1473 * be used as an exaggerated estimate. 1474 */ 1475 static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, 1476 u32 start_seq, u32 end_seq) 1477 { 1478 /* Too far in future, or reversed (interpretation is ambiguous) */ 1479 if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) 1480 return false; 1481 1482 /* Nasty start_seq wrap-around check (see comments above) */ 1483 if (!before(start_seq, tp->snd_nxt)) 1484 return false; 1485 1486 /* In outstanding window? ...This is valid exit for D-SACKs too. 1487 * start_seq == snd_una is non-sensical (see comments above) 1488 */ 1489 if (after(start_seq, tp->snd_una)) 1490 return true; 1491 1492 if (!is_dsack || !tp->undo_marker) 1493 return false; 1494 1495 /* ...Then it's D-SACK, and must reside below snd_una completely */ 1496 if (after(end_seq, tp->snd_una)) 1497 return false; 1498 1499 if (!before(start_seq, tp->undo_marker)) 1500 return true; 1501 1502 /* Too old */ 1503 if (!after(end_seq, tp->undo_marker)) 1504 return false; 1505 1506 /* Undo_marker boundary crossing (overestimates a lot). Known already: 1507 * start_seq < undo_marker and end_seq >= undo_marker. 1508 */ 1509 return !before(start_seq, end_seq - tp->max_window); 1510 } 1511 1512 static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, 1513 struct tcp_sack_block_wire *sp, int num_sacks, 1514 u32 prior_snd_una, struct tcp_sacktag_state *state) 1515 { 1516 struct tcp_sock *tp = tcp_sk(sk); 1517 u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); 1518 u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); 1519 u32 dup_segs; 1520 1521 if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { 1522 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV); 1523 } else if (num_sacks > 1) { 1524 u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); 1525 u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); 1526 1527 if (after(end_seq_0, end_seq_1) || before(start_seq_0, start_seq_1)) 1528 return false; 1529 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKOFORECV); 1530 } else { 1531 return false; 1532 } 1533 1534 dup_segs = tcp_dsack_seen(tp, start_seq_0, end_seq_0, state); 1535 if (!dup_segs) { /* Skip dubious DSACK */ 1536 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKIGNOREDDUBIOUS); 1537 return false; 1538 } 1539 1540 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECVSEGS, dup_segs); 1541 1542 /* D-SACK for already forgotten data... Do dumb counting. */ 1543 if (tp->undo_marker && tp->undo_retrans > 0 && 1544 !after(end_seq_0, prior_snd_una) && 1545 after(end_seq_0, tp->undo_marker)) 1546 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - dup_segs); 1547 1548 return true; 1549 } 1550 1551 /* Check if skb is fully within the SACK block. In presence of GSO skbs, 1552 * the incoming SACK may not exactly match but we can find smaller MSS 1553 * aligned portion of it that matches. Therefore we might need to fragment 1554 * which may fail and creates some hassle (caller must handle error case 1555 * returns). 1556 * 1557 * FIXME: this could be merged to shift decision code 1558 */ 1559 static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, 1560 u32 start_seq, u32 end_seq) 1561 { 1562 int err; 1563 bool in_sack; 1564 unsigned int pkt_len; 1565 unsigned int mss; 1566 1567 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 1568 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 1569 1570 if (tcp_skb_pcount(skb) > 1 && !in_sack && 1571 after(TCP_SKB_CB(skb)->end_seq, start_seq)) { 1572 mss = tcp_skb_mss(skb); 1573 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 1574 1575 if (!in_sack) { 1576 pkt_len = start_seq - TCP_SKB_CB(skb)->seq; 1577 if (pkt_len < mss) 1578 pkt_len = mss; 1579 } else { 1580 pkt_len = end_seq - TCP_SKB_CB(skb)->seq; 1581 if (pkt_len < mss) 1582 return -EINVAL; 1583 } 1584 1585 /* Round if necessary so that SACKs cover only full MSSes 1586 * and/or the remaining small portion (if present) 1587 */ 1588 if (pkt_len > mss) { 1589 unsigned int new_len = (pkt_len / mss) * mss; 1590 if (!in_sack && new_len < pkt_len) 1591 new_len += mss; 1592 pkt_len = new_len; 1593 } 1594 1595 if (pkt_len >= skb->len && !in_sack) 1596 return 0; 1597 1598 err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 1599 pkt_len, mss, GFP_ATOMIC); 1600 if (err < 0) 1601 return err; 1602 } 1603 1604 return in_sack; 1605 } 1606 1607 /* Record the most recently (re)sent time among the (s)acked packets 1608 * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from 1609 * draft-cheng-tcpm-rack-00.txt 1610 */ 1611 static void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, 1612 u32 end_seq, u64 xmit_time) 1613 { 1614 u32 rtt_us; 1615 1616 rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, xmit_time); 1617 if (rtt_us < tcp_min_rtt(tp) && (sacked & TCPCB_RETRANS)) { 1618 /* If the sacked packet was retransmitted, it's ambiguous 1619 * whether the retransmission or the original (or the prior 1620 * retransmission) was sacked. 1621 * 1622 * If the original is lost, there is no ambiguity. Otherwise 1623 * we assume the original can be delayed up to aRTT + min_rtt. 1624 * the aRTT term is bounded by the fast recovery or timeout, 1625 * so it's at least one RTT (i.e., retransmission is at least 1626 * an RTT later). 1627 */ 1628 return; 1629 } 1630 tp->rack.advanced = 1; 1631 tp->rack.rtt_us = rtt_us; 1632 if (tcp_skb_sent_after(xmit_time, tp->rack.mstamp, 1633 end_seq, tp->rack.end_seq)) { 1634 tp->rack.mstamp = xmit_time; 1635 tp->rack.end_seq = end_seq; 1636 } 1637 } 1638 1639 /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ 1640 static u8 tcp_sacktag_one(struct sock *sk, 1641 struct tcp_sacktag_state *state, u8 sacked, 1642 u32 start_seq, u32 end_seq, 1643 int dup_sack, int pcount, u32 plen, 1644 u64 xmit_time) 1645 { 1646 struct tcp_sock *tp = tcp_sk(sk); 1647 1648 /* Account D-SACK for retransmitted packet. */ 1649 if (dup_sack && (sacked & TCPCB_RETRANS)) { 1650 if (tp->undo_marker && tp->undo_retrans > 0 && 1651 after(end_seq, tp->undo_marker)) 1652 tp->undo_retrans = max_t(int, 0, tp->undo_retrans - pcount); 1653 if ((sacked & TCPCB_SACKED_ACKED) && 1654 before(start_seq, state->reord)) 1655 state->reord = start_seq; 1656 } 1657 1658 /* Nothing to do; acked frame is about to be dropped (was ACKed). */ 1659 if (!after(end_seq, tp->snd_una)) 1660 return sacked; 1661 1662 if (!(sacked & TCPCB_SACKED_ACKED)) { 1663 tcp_rack_advance(tp, sacked, end_seq, xmit_time); 1664 1665 if (sacked & TCPCB_SACKED_RETRANS) { 1666 /* If the segment is not tagged as lost, 1667 * we do not clear RETRANS, believing 1668 * that retransmission is still in flight. 1669 */ 1670 if (sacked & TCPCB_LOST) { 1671 sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); 1672 tp->lost_out -= pcount; 1673 tp->retrans_out -= pcount; 1674 } 1675 } else { 1676 if (!(sacked & TCPCB_RETRANS)) { 1677 /* New sack for not retransmitted frame, 1678 * which was in hole. It is reordering. 1679 */ 1680 if (before(start_seq, 1681 tcp_highest_sack_seq(tp)) && 1682 before(start_seq, state->reord)) 1683 state->reord = start_seq; 1684 1685 if (!after(end_seq, tp->high_seq)) 1686 state->flag |= FLAG_ORIG_SACK_ACKED; 1687 if (state->first_sackt == 0) 1688 state->first_sackt = xmit_time; 1689 state->last_sackt = xmit_time; 1690 } 1691 1692 if (sacked & TCPCB_LOST) { 1693 sacked &= ~TCPCB_LOST; 1694 tp->lost_out -= pcount; 1695 } 1696 } 1697 1698 sacked |= TCPCB_SACKED_ACKED; 1699 state->flag |= FLAG_DATA_SACKED; 1700 tp->sacked_out += pcount; 1701 /* Out-of-order packets delivered */ 1702 state->sack_delivered += pcount; 1703 state->delivered_bytes += plen; 1704 } 1705 1706 /* D-SACK. We can detect redundant retransmission in S|R and plain R 1707 * frames and clear it. undo_retrans is decreased above, L|R frames 1708 * are accounted above as well. 1709 */ 1710 if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { 1711 sacked &= ~TCPCB_SACKED_RETRANS; 1712 tp->retrans_out -= pcount; 1713 } 1714 1715 return sacked; 1716 } 1717 1718 /* The bandwidth estimator estimates the rate at which the network 1719 * can currently deliver outbound data packets for this flow. At a high 1720 * level, it operates by taking a delivery rate sample for each ACK. 1721 * 1722 * A rate sample records the rate at which the network delivered packets 1723 * for this flow, calculated over the time interval between the transmission 1724 * of a data packet and the acknowledgment of that packet. 1725 * 1726 * Specifically, over the interval between each transmit and corresponding ACK, 1727 * the estimator generates a delivery rate sample. Typically it uses the rate 1728 * at which packets were acknowledged. However, the approach of using only the 1729 * acknowledgment rate faces a challenge under the prevalent ACK decimation or 1730 * compression: packets can temporarily appear to be delivered much quicker 1731 * than the bottleneck rate. Since it is physically impossible to do that in a 1732 * sustained fashion, when the estimator notices that the ACK rate is faster 1733 * than the transmit rate, it uses the latter: 1734 * 1735 * send_rate = #pkts_delivered/(last_snd_time - first_snd_time) 1736 * ack_rate = #pkts_delivered/(last_ack_time - first_ack_time) 1737 * bw = min(send_rate, ack_rate) 1738 * 1739 * Notice the estimator essentially estimates the goodput, not always the 1740 * network bottleneck link rate when the sending or receiving is limited by 1741 * other factors like applications or receiver window limits. The estimator 1742 * deliberately avoids using the inter-packet spacing approach because that 1743 * approach requires a large number of samples and sophisticated filtering. 1744 * 1745 * TCP flows can often be application-limited in request/response workloads. 1746 * The estimator marks a bandwidth sample as application-limited if there 1747 * was some moment during the sampled window of packets when there was no data 1748 * ready to send in the write queue. 1749 */ 1750 1751 /* Update the connection delivery information and generate a rate sample. */ 1752 static void tcp_rate_gen(struct sock *sk, u32 delivered, u32 lost, 1753 bool is_sack_reneg, struct rate_sample *rs) 1754 { 1755 struct tcp_sock *tp = tcp_sk(sk); 1756 u32 snd_us, ack_us; 1757 1758 /* Clear app limited if bubble is acked and gone. */ 1759 if (tp->app_limited && after(tp->delivered, tp->app_limited)) 1760 tp->app_limited = 0; 1761 1762 /* TODO: there are multiple places throughout tcp_ack() to get 1763 * current time. Refactor the code using a new "tcp_acktag_state" 1764 * to carry current time, flags, stats like "tcp_sacktag_state". 1765 */ 1766 if (delivered) 1767 tp->delivered_mstamp = tp->tcp_mstamp; 1768 1769 rs->acked_sacked = delivered; /* freshly ACKed or SACKed */ 1770 rs->losses = lost; /* freshly marked lost */ 1771 /* Return an invalid sample if no timing information is available or 1772 * in recovery from loss with SACK reneging. Rate samples taken during 1773 * a SACK reneging event may overestimate bw by including packets that 1774 * were SACKed before the reneg. 1775 */ 1776 if (!rs->prior_mstamp || is_sack_reneg) { 1777 rs->delivered = -1; 1778 rs->interval_us = -1; 1779 return; 1780 } 1781 rs->delivered = tp->delivered - rs->prior_delivered; 1782 1783 rs->delivered_ce = tp->delivered_ce - rs->prior_delivered_ce; 1784 /* delivered_ce occupies less than 32 bits in the skb control block */ 1785 rs->delivered_ce &= TCPCB_DELIVERED_CE_MASK; 1786 1787 /* Model sending data and receiving ACKs as separate pipeline phases 1788 * for a window. Usually the ACK phase is longer, but with ACK 1789 * compression the send phase can be longer. To be safe we use the 1790 * longer phase. 1791 */ 1792 snd_us = rs->interval_us; /* send phase */ 1793 ack_us = tcp_stamp_us_delta(tp->tcp_mstamp, 1794 rs->prior_mstamp); /* ack phase */ 1795 rs->interval_us = max(snd_us, ack_us); 1796 1797 /* Record both segment send and ack receive intervals */ 1798 rs->snd_interval_us = snd_us; 1799 rs->rcv_interval_us = ack_us; 1800 1801 /* Normally we expect interval_us >= min-rtt. 1802 * Note that rate may still be over-estimated when a spuriously 1803 * retransmistted skb was first (s)acked because "interval_us" 1804 * is under-estimated (up to an RTT). However continuously 1805 * measuring the delivery rate during loss recovery is crucial 1806 * for connections suffer heavy or prolonged losses. 1807 */ 1808 if (unlikely(rs->interval_us < tcp_min_rtt(tp))) { 1809 if (!rs->is_retrans) 1810 pr_debug("tcp rate: %ld %d %u %u %u\n", 1811 rs->interval_us, rs->delivered, 1812 inet_csk(sk)->icsk_ca_state, 1813 tp->rx_opt.sack_ok, tcp_min_rtt(tp)); 1814 rs->interval_us = -1; 1815 return; 1816 } 1817 1818 /* Record the last non-app-limited or the highest app-limited bw */ 1819 if (!rs->is_app_limited || 1820 ((u64)rs->delivered * tp->rate_interval_us >= 1821 (u64)tp->rate_delivered * rs->interval_us)) { 1822 tp->rate_delivered = rs->delivered; 1823 tp->rate_interval_us = rs->interval_us; 1824 tp->rate_app_limited = rs->is_app_limited; 1825 } 1826 } 1827 1828 /* When an skb is sacked or acked, we fill in the rate sample with the (prior) 1829 * delivery information when the skb was last transmitted. 1830 * 1831 * If an ACK (s)acks multiple skbs (e.g., stretched-acks), this function is 1832 * called multiple times. We favor the information from the most recently 1833 * sent skb, i.e., the skb with the most recently sent time and the highest 1834 * sequence. 1835 */ 1836 static void tcp_rate_skb_delivered(struct sock *sk, struct sk_buff *skb, 1837 struct rate_sample *rs) 1838 { 1839 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 1840 struct tcp_sock *tp = tcp_sk(sk); 1841 u64 tx_tstamp; 1842 1843 if (!scb->tx.delivered_mstamp) 1844 return; 1845 1846 tx_tstamp = tcp_skb_timestamp_us(skb); 1847 if (!rs->prior_delivered || 1848 tcp_skb_sent_after(tx_tstamp, tp->first_tx_mstamp, 1849 scb->end_seq, rs->last_end_seq)) { 1850 rs->prior_delivered_ce = scb->tx.delivered_ce; 1851 rs->prior_delivered = scb->tx.delivered; 1852 rs->prior_mstamp = scb->tx.delivered_mstamp; 1853 rs->is_app_limited = scb->tx.is_app_limited; 1854 rs->is_retrans = scb->sacked & TCPCB_RETRANS; 1855 rs->last_end_seq = scb->end_seq; 1856 1857 /* Record send time of most recently ACKed packet: */ 1858 tp->first_tx_mstamp = tx_tstamp; 1859 /* Find the duration of the "send phase" of this window: */ 1860 rs->interval_us = tcp_stamp_us_delta(tp->first_tx_mstamp, 1861 scb->tx.first_tx_mstamp); 1862 1863 } 1864 /* Mark off the skb delivered once it's sacked to avoid being 1865 * used again when it's cumulatively acked. For acked packets 1866 * we don't need to reset since it'll be freed soon. 1867 */ 1868 if (scb->sacked & TCPCB_SACKED_ACKED) 1869 scb->tx.delivered_mstamp = 0; 1870 } 1871 1872 /* Shift newly-SACKed bytes from this skb to the immediately previous 1873 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. 1874 */ 1875 static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, 1876 struct sk_buff *skb, 1877 struct tcp_sacktag_state *state, 1878 unsigned int pcount, int shifted, int mss, 1879 bool dup_sack) 1880 { 1881 struct tcp_sock *tp = tcp_sk(sk); 1882 u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ 1883 u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ 1884 1885 BUG_ON(!pcount); 1886 1887 /* Adjust counters and hints for the newly sacked sequence 1888 * range but discard the return value since prev is already 1889 * marked. We must tag the range first because the seq 1890 * advancement below implicitly advances 1891 * tcp_highest_sack_seq() when skb is highest_sack. 1892 */ 1893 tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, 1894 start_seq, end_seq, dup_sack, pcount, skb->len, 1895 tcp_skb_timestamp_us(skb)); 1896 tcp_rate_skb_delivered(sk, skb, state->rate); 1897 1898 TCP_SKB_CB(prev)->end_seq += shifted; 1899 TCP_SKB_CB(skb)->seq += shifted; 1900 1901 tcp_skb_pcount_add(prev, pcount); 1902 WARN_ON_ONCE(tcp_skb_pcount(skb) < pcount); 1903 tcp_skb_pcount_add(skb, -pcount); 1904 1905 /* When we're adding to gso_segs == 1, gso_size will be zero, 1906 * in theory this shouldn't be necessary but as long as DSACK 1907 * code can come after this skb later on it's better to keep 1908 * setting gso_size to something. 1909 */ 1910 if (!TCP_SKB_CB(prev)->tcp_gso_size) 1911 TCP_SKB_CB(prev)->tcp_gso_size = mss; 1912 1913 /* CHECKME: To clear or not to clear? Mimics normal skb currently */ 1914 if (tcp_skb_pcount(skb) <= 1) 1915 TCP_SKB_CB(skb)->tcp_gso_size = 0; 1916 1917 /* Difference in this won't matter, both ACKed by the same cumul. ACK */ 1918 TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); 1919 1920 if (skb->len > 0) { 1921 BUG_ON(!tcp_skb_pcount(skb)); 1922 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED); 1923 return false; 1924 } 1925 1926 /* Whole SKB was eaten :-) */ 1927 1928 if (skb == tp->retransmit_skb_hint) 1929 tp->retransmit_skb_hint = prev; 1930 1931 TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1932 TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor; 1933 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1934 TCP_SKB_CB(prev)->end_seq++; 1935 1936 if (skb == tcp_highest_sack(sk)) 1937 tcp_advance_highest_sack(sk, skb); 1938 1939 tcp_skb_collapse_tstamp(prev, skb); 1940 if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp)) 1941 TCP_SKB_CB(prev)->tx.delivered_mstamp = 0; 1942 1943 tcp_rtx_queue_unlink_and_free(skb, sk); 1944 1945 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED); 1946 1947 return true; 1948 } 1949 1950 /* I wish gso_size would have a bit more sane initialization than 1951 * something-or-zero which complicates things 1952 */ 1953 static int tcp_skb_seglen(const struct sk_buff *skb) 1954 { 1955 return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); 1956 } 1957 1958 /* Shifting pages past head area doesn't work */ 1959 static int skb_can_shift(const struct sk_buff *skb) 1960 { 1961 return !skb_headlen(skb) && skb_is_nonlinear(skb); 1962 } 1963 1964 int tcp_skb_shift(struct sk_buff *to, struct sk_buff *from, 1965 int pcount, int shiftlen) 1966 { 1967 /* TCP min gso_size is 8 bytes (TCP_MIN_GSO_SIZE) 1968 * Since TCP_SKB_CB(skb)->tcp_gso_segs is 16 bits, we need 1969 * to make sure not storing more than 65535 * 8 bytes per skb, 1970 * even if current MSS is bigger. 1971 */ 1972 if (unlikely(to->len + shiftlen >= 65535 * TCP_MIN_GSO_SIZE)) 1973 return 0; 1974 if (unlikely(tcp_skb_pcount(to) + pcount > 65535)) 1975 return 0; 1976 return skb_shift(to, from, shiftlen); 1977 } 1978 1979 /* Try collapsing SACK blocks spanning across multiple skbs to a single 1980 * skb. 1981 */ 1982 static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, 1983 struct tcp_sacktag_state *state, 1984 u32 start_seq, u32 end_seq, 1985 bool dup_sack) 1986 { 1987 struct tcp_sock *tp = tcp_sk(sk); 1988 struct sk_buff *prev; 1989 int mss; 1990 int pcount = 0; 1991 int len; 1992 int in_sack; 1993 1994 /* Normally R but no L won't result in plain S */ 1995 if (!dup_sack && 1996 (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) 1997 goto fallback; 1998 if (!skb_can_shift(skb)) 1999 goto fallback; 2000 /* This frame is about to be dropped (was ACKed). */ 2001 if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 2002 goto fallback; 2003 2004 /* Can only happen with delayed DSACK + discard craziness */ 2005 prev = skb_rb_prev(skb); 2006 if (!prev) 2007 goto fallback; 2008 2009 if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) 2010 goto fallback; 2011 2012 if (!tcp_skb_can_collapse(prev, skb)) 2013 goto fallback; 2014 2015 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && 2016 !before(end_seq, TCP_SKB_CB(skb)->end_seq); 2017 2018 if (in_sack) { 2019 len = skb->len; 2020 pcount = tcp_skb_pcount(skb); 2021 mss = tcp_skb_seglen(skb); 2022 2023 /* TODO: Fix DSACKs to not fragment already SACKed and we can 2024 * drop this restriction as unnecessary 2025 */ 2026 if (mss != tcp_skb_seglen(prev)) 2027 goto fallback; 2028 } else { 2029 if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) 2030 goto noop; 2031 /* CHECKME: This is non-MSS split case only?, this will 2032 * cause skipped skbs due to advancing loop btw, original 2033 * has that feature too 2034 */ 2035 if (tcp_skb_pcount(skb) <= 1) 2036 goto noop; 2037 2038 in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); 2039 if (!in_sack) { 2040 /* TODO: head merge to next could be attempted here 2041 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), 2042 * though it might not be worth of the additional hassle 2043 * 2044 * ...we can probably just fallback to what was done 2045 * previously. We could try merging non-SACKed ones 2046 * as well but it probably isn't going to buy off 2047 * because later SACKs might again split them, and 2048 * it would make skb timestamp tracking considerably 2049 * harder problem. 2050 */ 2051 goto fallback; 2052 } 2053 2054 len = end_seq - TCP_SKB_CB(skb)->seq; 2055 BUG_ON(len < 0); 2056 BUG_ON(len > skb->len); 2057 2058 /* MSS boundaries should be honoured or else pcount will 2059 * severely break even though it makes things bit trickier. 2060 * Optimize common case to avoid most of the divides 2061 */ 2062 mss = tcp_skb_mss(skb); 2063 2064 /* TODO: Fix DSACKs to not fragment already SACKed and we can 2065 * drop this restriction as unnecessary 2066 */ 2067 if (mss != tcp_skb_seglen(prev)) 2068 goto fallback; 2069 2070 if (len == mss) { 2071 pcount = 1; 2072 } else if (len < mss) { 2073 goto noop; 2074 } else { 2075 pcount = len / mss; 2076 len = pcount * mss; 2077 } 2078 } 2079 2080 /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ 2081 if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) 2082 goto fallback; 2083 2084 if (!tcp_skb_shift(prev, skb, pcount, len)) 2085 goto fallback; 2086 if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack)) 2087 goto out; 2088 2089 /* Hole filled allows collapsing with the next as well, this is very 2090 * useful when hole on every nth skb pattern happens 2091 */ 2092 skb = skb_rb_next(prev); 2093 if (!skb) 2094 goto out; 2095 2096 if (!skb_can_shift(skb) || 2097 ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || 2098 (mss != tcp_skb_seglen(skb))) 2099 goto out; 2100 2101 if (!tcp_skb_can_collapse(prev, skb)) 2102 goto out; 2103 len = skb->len; 2104 pcount = tcp_skb_pcount(skb); 2105 if (tcp_skb_shift(prev, skb, pcount, len)) 2106 tcp_shifted_skb(sk, prev, skb, state, pcount, 2107 len, mss, 0); 2108 2109 out: 2110 return prev; 2111 2112 noop: 2113 return skb; 2114 2115 fallback: 2116 NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); 2117 return NULL; 2118 } 2119 2120 static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, 2121 struct tcp_sack_block *next_dup, 2122 struct tcp_sacktag_state *state, 2123 u32 start_seq, u32 end_seq, 2124 bool dup_sack_in) 2125 { 2126 struct tcp_sock *tp = tcp_sk(sk); 2127 struct sk_buff *tmp; 2128 2129 skb_rbtree_walk_from(skb) { 2130 int in_sack = 0; 2131 bool dup_sack = dup_sack_in; 2132 2133 /* queue is in-order => we can short-circuit the walk early */ 2134 if (!before(TCP_SKB_CB(skb)->seq, end_seq)) 2135 break; 2136 2137 if (next_dup && 2138 before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { 2139 in_sack = tcp_match_skb_to_sack(sk, skb, 2140 next_dup->start_seq, 2141 next_dup->end_seq); 2142 if (in_sack > 0) 2143 dup_sack = true; 2144 } 2145 2146 /* skb reference here is a bit tricky to get right, since 2147 * shifting can eat and free both this skb and the next, 2148 * so not even _safe variant of the loop is enough. 2149 */ 2150 if (in_sack <= 0) { 2151 tmp = tcp_shift_skb_data(sk, skb, state, 2152 start_seq, end_seq, dup_sack); 2153 if (tmp) { 2154 if (tmp != skb) { 2155 skb = tmp; 2156 continue; 2157 } 2158 2159 in_sack = 0; 2160 } else { 2161 in_sack = tcp_match_skb_to_sack(sk, skb, 2162 start_seq, 2163 end_seq); 2164 } 2165 } 2166 2167 if (unlikely(in_sack < 0)) 2168 break; 2169 2170 if (in_sack) { 2171 TCP_SKB_CB(skb)->sacked = 2172 tcp_sacktag_one(sk, 2173 state, 2174 TCP_SKB_CB(skb)->sacked, 2175 TCP_SKB_CB(skb)->seq, 2176 TCP_SKB_CB(skb)->end_seq, 2177 dup_sack, 2178 tcp_skb_pcount(skb), 2179 skb->len, 2180 tcp_skb_timestamp_us(skb)); 2181 tcp_rate_skb_delivered(sk, skb, state->rate); 2182 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2183 list_del_init(&skb->tcp_tsorted_anchor); 2184 2185 if (!before(TCP_SKB_CB(skb)->seq, 2186 tcp_highest_sack_seq(tp))) 2187 tcp_advance_highest_sack(sk, skb); 2188 } 2189 } 2190 return skb; 2191 } 2192 2193 static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, u32 seq) 2194 { 2195 struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node; 2196 struct sk_buff *skb; 2197 2198 while (*p) { 2199 parent = *p; 2200 skb = rb_to_skb(parent); 2201 if (before(seq, TCP_SKB_CB(skb)->seq)) { 2202 p = &parent->rb_left; 2203 continue; 2204 } 2205 if (!before(seq, TCP_SKB_CB(skb)->end_seq)) { 2206 p = &parent->rb_right; 2207 continue; 2208 } 2209 return skb; 2210 } 2211 return NULL; 2212 } 2213 2214 static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, 2215 u32 skip_to_seq) 2216 { 2217 if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq)) 2218 return skb; 2219 2220 return tcp_sacktag_bsearch(sk, skip_to_seq); 2221 } 2222 2223 static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, 2224 struct sock *sk, 2225 struct tcp_sack_block *next_dup, 2226 struct tcp_sacktag_state *state, 2227 u32 skip_to_seq) 2228 { 2229 if (!next_dup) 2230 return skb; 2231 2232 if (before(next_dup->start_seq, skip_to_seq)) { 2233 skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq); 2234 skb = tcp_sacktag_walk(skb, sk, NULL, state, 2235 next_dup->start_seq, next_dup->end_seq, 2236 1); 2237 } 2238 2239 return skb; 2240 } 2241 2242 static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) 2243 { 2244 return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 2245 } 2246 2247 static int 2248 tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, 2249 u32 prior_snd_una, struct tcp_sacktag_state *state) 2250 { 2251 struct tcp_sock *tp = tcp_sk(sk); 2252 const unsigned char *ptr = (skb_transport_header(ack_skb) + 2253 TCP_SKB_CB(ack_skb)->sacked); 2254 struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); 2255 struct tcp_sack_block sp[TCP_NUM_SACKS]; 2256 struct tcp_sack_block *cache; 2257 struct sk_buff *skb; 2258 int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); 2259 int used_sacks; 2260 bool found_dup_sack = false; 2261 int i, j; 2262 int first_sack_index; 2263 2264 state->flag = 0; 2265 state->reord = tp->snd_nxt; 2266 2267 if (!tp->sacked_out) 2268 tcp_highest_sack_reset(sk); 2269 2270 found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, 2271 num_sacks, prior_snd_una, state); 2272 2273 /* Eliminate too old ACKs, but take into 2274 * account more or less fresh ones, they can 2275 * contain valid SACK info. 2276 */ 2277 if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) 2278 return 0; 2279 2280 if (!tp->packets_out) 2281 goto out; 2282 2283 used_sacks = 0; 2284 first_sack_index = 0; 2285 for (i = 0; i < num_sacks; i++) { 2286 bool dup_sack = !i && found_dup_sack; 2287 2288 sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); 2289 sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); 2290 2291 if (!tcp_is_sackblock_valid(tp, dup_sack, 2292 sp[used_sacks].start_seq, 2293 sp[used_sacks].end_seq)) { 2294 int mib_idx; 2295 2296 if (dup_sack) { 2297 if (!tp->undo_marker) 2298 mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; 2299 else 2300 mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; 2301 } else { 2302 /* Don't count olds caused by ACK reordering */ 2303 if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && 2304 !after(sp[used_sacks].end_seq, tp->snd_una)) 2305 continue; 2306 mib_idx = LINUX_MIB_TCPSACKDISCARD; 2307 } 2308 2309 NET_INC_STATS(sock_net(sk), mib_idx); 2310 if (i == 0) 2311 first_sack_index = -1; 2312 continue; 2313 } 2314 2315 /* Ignore very old stuff early */ 2316 if (!after(sp[used_sacks].end_seq, prior_snd_una)) { 2317 if (i == 0) 2318 first_sack_index = -1; 2319 continue; 2320 } 2321 2322 used_sacks++; 2323 } 2324 2325 /* order SACK blocks to allow in order walk of the retrans queue */ 2326 for (i = used_sacks - 1; i > 0; i--) { 2327 for (j = 0; j < i; j++) { 2328 if (after(sp[j].start_seq, sp[j + 1].start_seq)) { 2329 swap(sp[j], sp[j + 1]); 2330 2331 /* Track where the first SACK block goes to */ 2332 if (j == first_sack_index) 2333 first_sack_index = j + 1; 2334 } 2335 } 2336 } 2337 2338 state->mss_now = tcp_current_mss(sk); 2339 skb = NULL; 2340 i = 0; 2341 2342 if (!tp->sacked_out) { 2343 /* It's already past, so skip checking against it */ 2344 cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); 2345 } else { 2346 cache = tp->recv_sack_cache; 2347 /* Skip empty blocks in at head of the cache */ 2348 while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && 2349 !cache->end_seq) 2350 cache++; 2351 } 2352 2353 while (i < used_sacks) { 2354 u32 start_seq = sp[i].start_seq; 2355 u32 end_seq = sp[i].end_seq; 2356 bool dup_sack = (found_dup_sack && (i == first_sack_index)); 2357 struct tcp_sack_block *next_dup = NULL; 2358 2359 if (found_dup_sack && ((i + 1) == first_sack_index)) 2360 next_dup = &sp[i + 1]; 2361 2362 /* Skip too early cached blocks */ 2363 while (tcp_sack_cache_ok(tp, cache) && 2364 !before(start_seq, cache->end_seq)) 2365 cache++; 2366 2367 /* Can skip some work by looking recv_sack_cache? */ 2368 if (tcp_sack_cache_ok(tp, cache) && !dup_sack && 2369 after(end_seq, cache->start_seq)) { 2370 2371 /* Head todo? */ 2372 if (before(start_seq, cache->start_seq)) { 2373 skb = tcp_sacktag_skip(skb, sk, start_seq); 2374 skb = tcp_sacktag_walk(skb, sk, next_dup, 2375 state, 2376 start_seq, 2377 cache->start_seq, 2378 dup_sack); 2379 } 2380 2381 /* Rest of the block already fully processed? */ 2382 if (!after(end_seq, cache->end_seq)) 2383 goto advance_sp; 2384 2385 skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, 2386 state, 2387 cache->end_seq); 2388 2389 /* ...tail remains todo... */ 2390 if (tcp_highest_sack_seq(tp) == cache->end_seq) { 2391 /* ...but better entrypoint exists! */ 2392 skb = tcp_highest_sack(sk); 2393 if (!skb) 2394 break; 2395 cache++; 2396 goto walk; 2397 } 2398 2399 skb = tcp_sacktag_skip(skb, sk, cache->end_seq); 2400 /* Check overlap against next cached too (past this one already) */ 2401 cache++; 2402 continue; 2403 } 2404 2405 if (!before(start_seq, tcp_highest_sack_seq(tp))) { 2406 skb = tcp_highest_sack(sk); 2407 if (!skb) 2408 break; 2409 } 2410 skb = tcp_sacktag_skip(skb, sk, start_seq); 2411 2412 walk: 2413 skb = tcp_sacktag_walk(skb, sk, next_dup, state, 2414 start_seq, end_seq, dup_sack); 2415 2416 advance_sp: 2417 i++; 2418 } 2419 2420 /* Clear the head of the cache sack blocks so we can skip it next time */ 2421 for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { 2422 tp->recv_sack_cache[i].start_seq = 0; 2423 tp->recv_sack_cache[i].end_seq = 0; 2424 } 2425 for (j = 0; j < used_sacks; j++) 2426 tp->recv_sack_cache[i++] = sp[j]; 2427 2428 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss || tp->undo_marker) 2429 tcp_check_sack_reordering(sk, state->reord, 0); 2430 2431 tcp_verify_left_out(tp); 2432 out: 2433 2434 #if FASTRETRANS_DEBUG > 0 2435 WARN_ON((int)tp->sacked_out < 0); 2436 WARN_ON((int)tp->lost_out < 0); 2437 WARN_ON((int)tp->retrans_out < 0); 2438 WARN_ON((int)tcp_packets_in_flight(tp) < 0); 2439 #endif 2440 return state->flag; 2441 } 2442 2443 /* Limits sacked_out so that sum with lost_out isn't ever larger than 2444 * packets_out. Returns false if sacked_out adjustement wasn't necessary. 2445 */ 2446 static bool tcp_limit_reno_sacked(struct tcp_sock *tp) 2447 { 2448 u32 holes; 2449 2450 holes = max(tp->lost_out, 1U); 2451 holes = min(holes, tp->packets_out); 2452 2453 if ((tp->sacked_out + holes) > tp->packets_out) { 2454 tp->sacked_out = tp->packets_out - holes; 2455 return true; 2456 } 2457 return false; 2458 } 2459 2460 /* If we receive more dupacks than we expected counting segments 2461 * in assumption of absent reordering, interpret this as reordering. 2462 * The only another reason could be bug in receiver TCP. 2463 */ 2464 static void tcp_check_reno_reordering(struct sock *sk, const int addend) 2465 { 2466 struct tcp_sock *tp = tcp_sk(sk); 2467 2468 if (!tcp_limit_reno_sacked(tp)) 2469 return; 2470 2471 WRITE_ONCE(tp->reordering, 2472 min_t(u32, tp->packets_out + addend, 2473 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_max_reordering))); 2474 WRITE_ONCE(tp->reord_seen, tp->reord_seen + 1); 2475 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRENOREORDER); 2476 } 2477 2478 /* Emulate SACKs for SACKless connection: account for a new dupack. */ 2479 2480 static void tcp_add_reno_sack(struct sock *sk, int num_dupack, bool ece_ack) 2481 { 2482 if (num_dupack) { 2483 struct tcp_sock *tp = tcp_sk(sk); 2484 u32 prior_sacked = tp->sacked_out; 2485 s32 delivered; 2486 2487 tp->sacked_out += num_dupack; 2488 tcp_check_reno_reordering(sk, 0); 2489 delivered = tp->sacked_out - prior_sacked; 2490 if (delivered > 0) 2491 tcp_count_delivered(tp, delivered, ece_ack); 2492 tcp_verify_left_out(tp); 2493 } 2494 } 2495 2496 /* Account for ACK, ACKing some data in Reno Recovery phase. */ 2497 2498 static void tcp_remove_reno_sacks(struct sock *sk, int acked, bool ece_ack) 2499 { 2500 struct tcp_sock *tp = tcp_sk(sk); 2501 2502 if (acked > 0) { 2503 /* One ACK acked hole. The rest eat duplicate ACKs. */ 2504 tcp_count_delivered(tp, max_t(int, acked - tp->sacked_out, 1), 2505 ece_ack); 2506 if (acked - 1 >= tp->sacked_out) 2507 tp->sacked_out = 0; 2508 else 2509 tp->sacked_out -= acked - 1; 2510 } 2511 tcp_check_reno_reordering(sk, acked); 2512 tcp_verify_left_out(tp); 2513 } 2514 2515 static inline void tcp_reset_reno_sack(struct tcp_sock *tp) 2516 { 2517 tp->sacked_out = 0; 2518 } 2519 2520 void tcp_clear_retrans(struct tcp_sock *tp) 2521 { 2522 tp->retrans_out = 0; 2523 tp->lost_out = 0; 2524 tp->undo_marker = 0; 2525 tp->undo_retrans = -1; 2526 tp->sacked_out = 0; 2527 tp->rto_stamp = 0; 2528 tp->total_rto = 0; 2529 tp->total_rto_recoveries = 0; 2530 tp->total_rto_time = 0; 2531 } 2532 2533 static inline void tcp_init_undo(struct tcp_sock *tp) 2534 { 2535 tp->undo_marker = tp->snd_una; 2536 2537 /* Retransmission still in flight may cause DSACKs later. */ 2538 /* First, account for regular retransmits in flight: */ 2539 tp->undo_retrans = tp->retrans_out; 2540 /* Next, account for TLP retransmits in flight: */ 2541 if (tp->tlp_high_seq && tp->tlp_retrans) 2542 tp->undo_retrans++; 2543 /* Finally, avoid 0, because undo_retrans==0 means "can undo now": */ 2544 if (!tp->undo_retrans) 2545 tp->undo_retrans = -1; 2546 } 2547 2548 /* If we detect SACK reneging, forget all SACK information 2549 * and reset tags completely, otherwise preserve SACKs. If receiver 2550 * dropped its ofo queue, we will know this due to reneging detection. 2551 */ 2552 static void tcp_timeout_mark_lost(struct sock *sk) 2553 { 2554 struct tcp_sock *tp = tcp_sk(sk); 2555 struct sk_buff *skb, *head; 2556 bool is_reneg; /* is receiver reneging on SACKs? */ 2557 2558 head = tcp_rtx_queue_head(sk); 2559 is_reneg = head && (TCP_SKB_CB(head)->sacked & TCPCB_SACKED_ACKED); 2560 if (is_reneg) { 2561 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); 2562 tp->sacked_out = 0; 2563 /* Mark SACK reneging until we recover from this loss event. */ 2564 tp->is_sack_reneg = 1; 2565 } else if (tcp_is_reno(tp)) { 2566 tcp_reset_reno_sack(tp); 2567 } 2568 2569 skb = head; 2570 skb_rbtree_walk_from(skb) { 2571 if (is_reneg) 2572 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; 2573 else if (skb != head && tcp_rack_skb_timeout(tp, skb, 0) > 0) 2574 continue; /* Don't mark recently sent ones lost yet */ 2575 tcp_mark_skb_lost(sk, skb); 2576 } 2577 tcp_verify_left_out(tp); 2578 tcp_clear_all_retrans_hints(tp); 2579 } 2580 2581 /* Enter Loss state. */ 2582 void tcp_enter_loss(struct sock *sk) 2583 { 2584 const struct inet_connection_sock *icsk = inet_csk(sk); 2585 struct tcp_sock *tp = tcp_sk(sk); 2586 struct net *net = sock_net(sk); 2587 bool new_recovery = icsk->icsk_ca_state < TCP_CA_Recovery; 2588 u8 reordering; 2589 2590 tcp_timeout_mark_lost(sk); 2591 2592 /* Reduce ssthresh if it has not yet been made inside this window. */ 2593 if (icsk->icsk_ca_state <= TCP_CA_Disorder || 2594 !after(tp->high_seq, tp->snd_una) || 2595 (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { 2596 tp->prior_ssthresh = tcp_current_ssthresh(sk); 2597 tp->prior_cwnd = tcp_snd_cwnd(tp); 2598 WRITE_ONCE(tp->snd_ssthresh, icsk->icsk_ca_ops->ssthresh(sk)); 2599 tcp_ca_event(sk, CA_EVENT_LOSS); 2600 tcp_init_undo(tp); 2601 } 2602 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + 1); 2603 tp->snd_cwnd_cnt = 0; 2604 tp->snd_cwnd_stamp = tcp_jiffies32; 2605 2606 /* Timeout in disordered state after receiving substantial DUPACKs 2607 * suggests that the degree of reordering is over-estimated. 2608 */ 2609 reordering = READ_ONCE(net->ipv4.sysctl_tcp_reordering); 2610 if (icsk->icsk_ca_state <= TCP_CA_Disorder && 2611 tp->sacked_out >= reordering) 2612 WRITE_ONCE(tp->reordering, 2613 min_t(unsigned int, tp->reordering, reordering)); 2614 2615 tcp_set_ca_state(sk, TCP_CA_Loss); 2616 tp->high_seq = tp->snd_nxt; 2617 tp->tlp_high_seq = 0; 2618 tcp_ecn_queue_cwr(tp); 2619 2620 /* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous 2621 * loss recovery is underway except recurring timeout(s) on 2622 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing 2623 */ 2624 tp->frto = READ_ONCE(net->ipv4.sysctl_tcp_frto) && 2625 (new_recovery || icsk->icsk_retransmits) && 2626 !inet_csk(sk)->icsk_mtup.probe_size; 2627 } 2628 2629 /* If ACK arrived pointing to a remembered SACK, it means that our 2630 * remembered SACKs do not reflect real state of receiver i.e. 2631 * receiver _host_ is heavily congested (or buggy). 2632 * 2633 * To avoid big spurious retransmission bursts due to transient SACK 2634 * scoreboard oddities that look like reneging, we give the receiver a 2635 * little time (max(RTT/2, 10ms)) to send us some more ACKs that will 2636 * restore sanity to the SACK scoreboard. If the apparent reneging 2637 * persists until this RTO then we'll clear the SACK scoreboard. 2638 */ 2639 static bool tcp_check_sack_reneging(struct sock *sk, int *ack_flag) 2640 { 2641 if (*ack_flag & FLAG_SACK_RENEGING && 2642 *ack_flag & FLAG_SND_UNA_ADVANCED) { 2643 struct tcp_sock *tp = tcp_sk(sk); 2644 unsigned long delay = max(usecs_to_jiffies(tp->srtt_us >> 4), 2645 msecs_to_jiffies(10)); 2646 2647 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, delay, false); 2648 *ack_flag &= ~FLAG_SET_XMIT_TIMER; 2649 return true; 2650 } 2651 return false; 2652 } 2653 2654 /* Linux NewReno/SACK/ECN state machine. 2655 * -------------------------------------- 2656 * 2657 * "Open" Normal state, no dubious events, fast path. 2658 * "Disorder" In all the respects it is "Open", 2659 * but requires a bit more attention. It is entered when 2660 * we see some SACKs or dupacks. It is split of "Open" 2661 * mainly to move some processing from fast path to slow one. 2662 * "CWR" CWND was reduced due to some Congestion Notification event. 2663 * It can be ECN, ICMP source quench, local device congestion. 2664 * "Recovery" CWND was reduced, we are fast-retransmitting. 2665 * "Loss" CWND was reduced due to RTO timeout or SACK reneging. 2666 * 2667 * tcp_fastretrans_alert() is entered: 2668 * - each incoming ACK, if state is not "Open" 2669 * - when arrived ACK is unusual, namely: 2670 * * SACK 2671 * * Duplicate ACK. 2672 * * ECN ECE. 2673 * 2674 * Counting packets in flight is pretty simple. 2675 * 2676 * in_flight = packets_out - left_out + retrans_out 2677 * 2678 * packets_out is SND.NXT-SND.UNA counted in packets. 2679 * 2680 * retrans_out is number of retransmitted segments. 2681 * 2682 * left_out is number of segments left network, but not ACKed yet. 2683 * 2684 * left_out = sacked_out + lost_out 2685 * 2686 * sacked_out: Packets, which arrived to receiver out of order 2687 * and hence not ACKed. With SACKs this number is simply 2688 * amount of SACKed data. Even without SACKs 2689 * it is easy to give pretty reliable estimate of this number, 2690 * counting duplicate ACKs. 2691 * 2692 * lost_out: Packets lost by network. TCP has no explicit 2693 * "loss notification" feedback from network (for now). 2694 * It means that this number can be only _guessed_. 2695 * Actually, it is the heuristics to predict lossage that 2696 * distinguishes different algorithms. 2697 * 2698 * F.e. after RTO, when all the queue is considered as lost, 2699 * lost_out = packets_out and in_flight = retrans_out. 2700 * 2701 * Essentially, we have now a few algorithms detecting 2702 * lost packets. 2703 * 2704 * If the receiver supports SACK: 2705 * 2706 * RACK (RFC8985): RACK is a newer loss detection algorithm 2707 * (2017-) that checks timing instead of counting DUPACKs. 2708 * Essentially a packet is considered lost if it's not S/ACKed 2709 * after RTT + reordering_window, where both metrics are 2710 * dynamically measured and adjusted. This is implemented in 2711 * tcp_rack_mark_lost. 2712 * 2713 * If the receiver does not support SACK: 2714 * 2715 * NewReno (RFC6582): in Recovery we assume that one segment 2716 * is lost (classic Reno). While we are in Recovery and 2717 * a partial ACK arrives, we assume that one more packet 2718 * is lost (NewReno). This heuristics are the same in NewReno 2719 * and SACK. 2720 * 2721 * The really tricky (and requiring careful tuning) part of the algorithm 2722 * is hidden in the RACK code in tcp_recovery.c and tcp_xmit_retransmit_queue(). 2723 * The first determines the moment _when_ we should reduce CWND and, 2724 * hence, slow down forward transmission. In fact, it determines the moment 2725 * when we decide that hole is caused by loss, rather than by a reorder. 2726 * 2727 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill 2728 * holes, caused by lost packets. 2729 * 2730 * And the most logically complicated part of algorithm is undo 2731 * heuristics. We detect false retransmits due to both too early 2732 * fast retransmit (reordering) and underestimated RTO, analyzing 2733 * timestamps and D-SACKs. When we detect that some segments were 2734 * retransmitted by mistake and CWND reduction was wrong, we undo 2735 * window reduction and abort recovery phase. This logic is hidden 2736 * inside several functions named tcp_try_undo_<something>. 2737 */ 2738 2739 /* This function decides, when we should leave Disordered state 2740 * and enter Recovery phase, reducing congestion window. 2741 * 2742 * Main question: may we further continue forward transmission 2743 * with the same cwnd? 2744 */ 2745 static bool tcp_time_to_recover(const struct tcp_sock *tp) 2746 { 2747 /* Has loss detection marked at least one packet lost? */ 2748 return tp->lost_out != 0; 2749 } 2750 2751 static bool tcp_tsopt_ecr_before(const struct tcp_sock *tp, u32 when) 2752 { 2753 return tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 2754 before(tp->rx_opt.rcv_tsecr, when); 2755 } 2756 2757 /* skb is spurious retransmitted if the returned timestamp echo 2758 * reply is prior to the skb transmission time 2759 */ 2760 static bool tcp_skb_spurious_retrans(const struct tcp_sock *tp, 2761 const struct sk_buff *skb) 2762 { 2763 return (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) && 2764 tcp_tsopt_ecr_before(tp, tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb)); 2765 } 2766 2767 /* Nothing was retransmitted or returned timestamp is less 2768 * than timestamp of the first retransmission. 2769 */ 2770 static inline bool tcp_packet_delayed(const struct tcp_sock *tp) 2771 { 2772 const struct sock *sk = (const struct sock *)tp; 2773 2774 /* Received an echoed timestamp before the first retransmission? */ 2775 if (tp->retrans_stamp) 2776 return tcp_tsopt_ecr_before(tp, tp->retrans_stamp); 2777 2778 /* We set tp->retrans_stamp upon the first retransmission of a loss 2779 * recovery episode, so normally if tp->retrans_stamp is 0 then no 2780 * retransmission has happened yet (likely due to TSQ, which can cause 2781 * fast retransmits to be delayed). So if snd_una advanced while 2782 * (tp->retrans_stamp is 0 then apparently a packet was merely delayed, 2783 * not lost. But there are exceptions where we retransmit but then 2784 * clear tp->retrans_stamp, so we check for those exceptions. 2785 */ 2786 2787 /* (1) For non-SACK connections, tcp_is_non_sack_preventing_reopen() 2788 * clears tp->retrans_stamp when snd_una == high_seq. 2789 */ 2790 if (!tcp_is_sack(tp) && !before(tp->snd_una, tp->high_seq)) 2791 return false; 2792 2793 /* (2) In TCP_SYN_SENT tcp_clean_rtx_queue() clears tp->retrans_stamp 2794 * when setting FLAG_SYN_ACKED is set, even if the SYN was 2795 * retransmitted. 2796 */ 2797 if (sk->sk_state == TCP_SYN_SENT) 2798 return false; 2799 2800 return true; /* tp->retrans_stamp is zero; no retransmit yet */ 2801 } 2802 2803 /* Undo procedures. */ 2804 2805 /* We can clear retrans_stamp when there are no retransmissions in the 2806 * window. It would seem that it is trivially available for us in 2807 * tp->retrans_out, however, that kind of assumptions doesn't consider 2808 * what will happen if errors occur when sending retransmission for the 2809 * second time. ...It could the that such segment has only 2810 * TCPCB_EVER_RETRANS set at the present time. It seems that checking 2811 * the head skb is enough except for some reneging corner cases that 2812 * are not worth the effort. 2813 * 2814 * Main reason for all this complexity is the fact that connection dying 2815 * time now depends on the validity of the retrans_stamp, in particular, 2816 * that successive retransmissions of a segment must not advance 2817 * retrans_stamp under any conditions. 2818 */ 2819 static bool tcp_any_retrans_done(const struct sock *sk) 2820 { 2821 const struct tcp_sock *tp = tcp_sk(sk); 2822 struct sk_buff *skb; 2823 2824 if (tp->retrans_out) 2825 return true; 2826 2827 skb = tcp_rtx_queue_head(sk); 2828 if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) 2829 return true; 2830 2831 return false; 2832 } 2833 2834 /* If loss recovery is finished and there are no retransmits out in the 2835 * network, then we clear retrans_stamp so that upon the next loss recovery 2836 * retransmits_timed_out() and timestamp-undo are using the correct value. 2837 */ 2838 static void tcp_retrans_stamp_cleanup(struct sock *sk) 2839 { 2840 if (!tcp_any_retrans_done(sk)) 2841 tcp_sk(sk)->retrans_stamp = 0; 2842 } 2843 2844 static void DBGUNDO(struct sock *sk, const char *msg) 2845 { 2846 #if FASTRETRANS_DEBUG > 1 2847 struct tcp_sock *tp = tcp_sk(sk); 2848 struct inet_sock *inet = inet_sk(sk); 2849 2850 if (sk->sk_family == AF_INET) { 2851 pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", 2852 msg, 2853 &inet->inet_daddr, ntohs(inet->inet_dport), 2854 tcp_snd_cwnd(tp), tcp_left_out(tp), 2855 tp->snd_ssthresh, tp->prior_ssthresh, 2856 tp->packets_out); 2857 } 2858 #if IS_ENABLED(CONFIG_IPV6) 2859 else if (sk->sk_family == AF_INET6) { 2860 pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", 2861 msg, 2862 &sk->sk_v6_daddr, ntohs(inet->inet_dport), 2863 tcp_snd_cwnd(tp), tcp_left_out(tp), 2864 tp->snd_ssthresh, tp->prior_ssthresh, 2865 tp->packets_out); 2866 } 2867 #endif 2868 #endif 2869 } 2870 2871 static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss) 2872 { 2873 struct tcp_sock *tp = tcp_sk(sk); 2874 2875 if (unmark_loss) { 2876 struct sk_buff *skb; 2877 2878 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 2879 TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; 2880 } 2881 tp->lost_out = 0; 2882 tcp_clear_all_retrans_hints(tp); 2883 } 2884 2885 if (tp->prior_ssthresh) { 2886 const struct inet_connection_sock *icsk = inet_csk(sk); 2887 2888 tcp_snd_cwnd_set(tp, icsk->icsk_ca_ops->undo_cwnd(sk)); 2889 2890 if (tp->prior_ssthresh > tp->snd_ssthresh) { 2891 WRITE_ONCE(tp->snd_ssthresh, tp->prior_ssthresh); 2892 tcp_ecn_withdraw_cwr(tp); 2893 } 2894 } 2895 tp->snd_cwnd_stamp = tcp_jiffies32; 2896 tp->undo_marker = 0; 2897 tp->rack.advanced = 1; /* Force RACK to re-exam losses */ 2898 } 2899 2900 static inline bool tcp_may_undo(const struct tcp_sock *tp) 2901 { 2902 return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); 2903 } 2904 2905 static bool tcp_is_non_sack_preventing_reopen(struct sock *sk) 2906 { 2907 struct tcp_sock *tp = tcp_sk(sk); 2908 2909 if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { 2910 /* Hold old state until something *above* high_seq 2911 * is ACKed. For Reno it is MUST to prevent false 2912 * fast retransmits (RFC2582). SACK TCP is safe. */ 2913 if (!tcp_any_retrans_done(sk)) 2914 tp->retrans_stamp = 0; 2915 return true; 2916 } 2917 return false; 2918 } 2919 2920 /* People celebrate: "We love our President!" */ 2921 static bool tcp_try_undo_recovery(struct sock *sk) 2922 { 2923 struct tcp_sock *tp = tcp_sk(sk); 2924 2925 if (tcp_may_undo(tp)) { 2926 int mib_idx; 2927 2928 /* Happy end! We did not retransmit anything 2929 * or our original transmission succeeded. 2930 */ 2931 DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); 2932 tcp_undo_cwnd_reduction(sk, false); 2933 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) 2934 mib_idx = LINUX_MIB_TCPLOSSUNDO; 2935 else 2936 mib_idx = LINUX_MIB_TCPFULLUNDO; 2937 2938 NET_INC_STATS(sock_net(sk), mib_idx); 2939 } else if (tp->rack.reo_wnd_persist) { 2940 tp->rack.reo_wnd_persist--; 2941 } 2942 if (tcp_is_non_sack_preventing_reopen(sk)) 2943 return true; 2944 tcp_set_ca_state(sk, TCP_CA_Open); 2945 tp->is_sack_reneg = 0; 2946 return false; 2947 } 2948 2949 /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ 2950 static bool tcp_try_undo_dsack(struct sock *sk) 2951 { 2952 struct tcp_sock *tp = tcp_sk(sk); 2953 2954 if (tp->undo_marker && !tp->undo_retrans) { 2955 tp->rack.reo_wnd_persist = min(TCP_RACK_RECOVERY_THRESH, 2956 tp->rack.reo_wnd_persist + 1); 2957 DBGUNDO(sk, "D-SACK"); 2958 tcp_undo_cwnd_reduction(sk, false); 2959 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); 2960 return true; 2961 } 2962 return false; 2963 } 2964 2965 /* Undo during loss recovery after partial ACK or using F-RTO. */ 2966 static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo) 2967 { 2968 struct tcp_sock *tp = tcp_sk(sk); 2969 2970 if (frto_undo || tcp_may_undo(tp)) { 2971 tcp_undo_cwnd_reduction(sk, true); 2972 2973 DBGUNDO(sk, "partial loss"); 2974 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); 2975 if (frto_undo) 2976 NET_INC_STATS(sock_net(sk), 2977 LINUX_MIB_TCPSPURIOUSRTOS); 2978 WRITE_ONCE(inet_csk(sk)->icsk_retransmits, 0); 2979 if (tcp_is_non_sack_preventing_reopen(sk)) 2980 return true; 2981 if (frto_undo || tcp_is_sack(tp)) { 2982 tcp_set_ca_state(sk, TCP_CA_Open); 2983 tp->is_sack_reneg = 0; 2984 } 2985 return true; 2986 } 2987 return false; 2988 } 2989 2990 /* The cwnd reduction in CWR and Recovery uses the PRR algorithm in RFC 6937. 2991 * It computes the number of packets to send (sndcnt) based on packets newly 2992 * delivered: 2993 * 1) If the packets in flight is larger than ssthresh, PRR spreads the 2994 * cwnd reductions across a full RTT. 2995 * 2) Otherwise PRR uses packet conservation to send as much as delivered. 2996 * But when SND_UNA is acked without further losses, 2997 * slow starts cwnd up to ssthresh to speed up the recovery. 2998 */ 2999 static void tcp_init_cwnd_reduction(struct sock *sk) 3000 { 3001 struct tcp_sock *tp = tcp_sk(sk); 3002 3003 tp->high_seq = tp->snd_nxt; 3004 tp->tlp_high_seq = 0; 3005 tp->snd_cwnd_cnt = 0; 3006 tp->prior_cwnd = tcp_snd_cwnd(tp); 3007 tp->prr_delivered = 0; 3008 tp->prr_out = 0; 3009 WRITE_ONCE(tp->snd_ssthresh, inet_csk(sk)->icsk_ca_ops->ssthresh(sk)); 3010 tcp_ecn_queue_cwr(tp); 3011 } 3012 3013 void tcp_cwnd_reduction(struct sock *sk, int newly_acked_sacked, int newly_lost, int flag) 3014 { 3015 struct tcp_sock *tp = tcp_sk(sk); 3016 int sndcnt = 0; 3017 int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp); 3018 3019 if (newly_acked_sacked <= 0 || WARN_ON_ONCE(!tp->prior_cwnd)) 3020 return; 3021 3022 trace_tcp_cwnd_reduction_tp(sk, newly_acked_sacked, newly_lost, flag); 3023 3024 tp->prr_delivered += newly_acked_sacked; 3025 if (delta < 0) { 3026 u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered + 3027 tp->prior_cwnd - 1; 3028 sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out; 3029 } else { 3030 sndcnt = max_t(int, tp->prr_delivered - tp->prr_out, 3031 newly_acked_sacked); 3032 if (flag & FLAG_SND_UNA_ADVANCED && !newly_lost) 3033 sndcnt++; 3034 sndcnt = min(delta, sndcnt); 3035 } 3036 /* Force a fast retransmit upon entering fast recovery */ 3037 sndcnt = max(sndcnt, (tp->prr_out ? 0 : 1)); 3038 tcp_snd_cwnd_set(tp, tcp_packets_in_flight(tp) + sndcnt); 3039 } 3040 3041 static inline void tcp_end_cwnd_reduction(struct sock *sk) 3042 { 3043 struct tcp_sock *tp = tcp_sk(sk); 3044 3045 if (inet_csk(sk)->icsk_ca_ops->cong_control) 3046 return; 3047 3048 /* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */ 3049 if (tp->snd_ssthresh < TCP_INFINITE_SSTHRESH && 3050 (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR || tp->undo_marker)) { 3051 tcp_snd_cwnd_set(tp, tp->snd_ssthresh); 3052 tp->snd_cwnd_stamp = tcp_jiffies32; 3053 } 3054 tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); 3055 } 3056 3057 /* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */ 3058 void tcp_enter_cwr(struct sock *sk) 3059 { 3060 struct tcp_sock *tp = tcp_sk(sk); 3061 3062 tp->prior_ssthresh = 0; 3063 if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { 3064 tp->undo_marker = 0; 3065 tcp_init_cwnd_reduction(sk); 3066 tcp_set_ca_state(sk, TCP_CA_CWR); 3067 } 3068 } 3069 EXPORT_SYMBOL(tcp_enter_cwr); 3070 3071 static void tcp_try_keep_open(struct sock *sk) 3072 { 3073 struct tcp_sock *tp = tcp_sk(sk); 3074 int state = TCP_CA_Open; 3075 3076 if (tcp_left_out(tp) || tcp_any_retrans_done(sk)) 3077 state = TCP_CA_Disorder; 3078 3079 if (inet_csk(sk)->icsk_ca_state != state) { 3080 tcp_set_ca_state(sk, state); 3081 tp->high_seq = tp->snd_nxt; 3082 } 3083 } 3084 3085 static void tcp_try_to_open(struct sock *sk, int flag) 3086 { 3087 struct tcp_sock *tp = tcp_sk(sk); 3088 3089 tcp_verify_left_out(tp); 3090 3091 if (!tcp_any_retrans_done(sk)) 3092 tp->retrans_stamp = 0; 3093 3094 if (flag & FLAG_ECE) 3095 tcp_enter_cwr(sk); 3096 3097 if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { 3098 tcp_try_keep_open(sk); 3099 } 3100 } 3101 3102 static void tcp_mtup_probe_failed(struct sock *sk) 3103 { 3104 struct inet_connection_sock *icsk = inet_csk(sk); 3105 3106 icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; 3107 icsk->icsk_mtup.probe_size = 0; 3108 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPFAIL); 3109 } 3110 3111 static void tcp_mtup_probe_success(struct sock *sk) 3112 { 3113 struct tcp_sock *tp = tcp_sk(sk); 3114 struct inet_connection_sock *icsk = inet_csk(sk); 3115 u64 val; 3116 3117 tp->prior_ssthresh = tcp_current_ssthresh(sk); 3118 3119 val = (u64)tcp_snd_cwnd(tp) * tcp_mss_to_mtu(sk, tp->mss_cache); 3120 do_div(val, icsk->icsk_mtup.probe_size); 3121 DEBUG_NET_WARN_ON_ONCE((u32)val != val); 3122 tcp_snd_cwnd_set(tp, max_t(u32, 1U, val)); 3123 3124 tp->snd_cwnd_cnt = 0; 3125 tp->snd_cwnd_stamp = tcp_jiffies32; 3126 tp->snd_ssthresh = tcp_current_ssthresh(sk); 3127 3128 icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; 3129 icsk->icsk_mtup.probe_size = 0; 3130 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 3131 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMTUPSUCCESS); 3132 } 3133 3134 /* Sometimes we deduce that packets have been dropped due to reasons other than 3135 * congestion, like path MTU reductions or failed client TFO attempts. In these 3136 * cases we call this function to retransmit as many packets as cwnd allows, 3137 * without reducing cwnd. Given that retransmits will set retrans_stamp to a 3138 * non-zero value (and may do so in a later calling context due to TSQ), we 3139 * also enter CA_Loss so that we track when all retransmitted packets are ACKed 3140 * and clear retrans_stamp when that happens (to ensure later recurring RTOs 3141 * are using the correct retrans_stamp and don't declare ETIMEDOUT 3142 * prematurely). 3143 */ 3144 static void tcp_non_congestion_loss_retransmit(struct sock *sk) 3145 { 3146 const struct inet_connection_sock *icsk = inet_csk(sk); 3147 struct tcp_sock *tp = tcp_sk(sk); 3148 3149 if (icsk->icsk_ca_state != TCP_CA_Loss) { 3150 tp->high_seq = tp->snd_nxt; 3151 WRITE_ONCE(tp->snd_ssthresh, tcp_current_ssthresh(sk)); 3152 tp->prior_ssthresh = 0; 3153 tp->undo_marker = 0; 3154 tcp_set_ca_state(sk, TCP_CA_Loss); 3155 } 3156 tcp_xmit_retransmit_queue(sk); 3157 } 3158 3159 /* Do a simple retransmit without using the backoff mechanisms in 3160 * tcp_timer. This is used for path mtu discovery. 3161 * The socket is already locked here. 3162 */ 3163 void tcp_simple_retransmit(struct sock *sk) 3164 { 3165 struct tcp_sock *tp = tcp_sk(sk); 3166 struct sk_buff *skb; 3167 int mss; 3168 3169 /* A fastopen SYN request is stored as two separate packets within 3170 * the retransmit queue, this is done by tcp_send_syn_data(). 3171 * As a result simply checking the MSS of the frames in the queue 3172 * will not work for the SYN packet. 3173 * 3174 * Us being here is an indication of a path MTU issue so we can 3175 * assume that the fastopen SYN was lost and just mark all the 3176 * frames in the retransmit queue as lost. We will use an MSS of 3177 * -1 to mark all frames as lost, otherwise compute the current MSS. 3178 */ 3179 if (tp->syn_data && sk->sk_state == TCP_SYN_SENT) 3180 mss = -1; 3181 else 3182 mss = tcp_current_mss(sk); 3183 3184 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) { 3185 if (tcp_skb_seglen(skb) > mss) 3186 tcp_mark_skb_lost(sk, skb); 3187 } 3188 3189 if (!tp->lost_out) 3190 return; 3191 3192 if (tcp_is_reno(tp)) 3193 tcp_limit_reno_sacked(tp); 3194 3195 tcp_verify_left_out(tp); 3196 3197 /* Don't muck with the congestion window here. 3198 * Reason is that we do not increase amount of _data_ 3199 * in network, but units changed and effective 3200 * cwnd/ssthresh really reduced now. 3201 */ 3202 tcp_non_congestion_loss_retransmit(sk); 3203 } 3204 3205 void tcp_enter_recovery(struct sock *sk, bool ece_ack) 3206 { 3207 struct tcp_sock *tp = tcp_sk(sk); 3208 int mib_idx; 3209 3210 /* Start the clock with our fast retransmit, for undo and ETIMEDOUT. */ 3211 tcp_retrans_stamp_cleanup(sk); 3212 3213 if (tcp_is_reno(tp)) 3214 mib_idx = LINUX_MIB_TCPRENORECOVERY; 3215 else 3216 mib_idx = LINUX_MIB_TCPSACKRECOVERY; 3217 3218 NET_INC_STATS(sock_net(sk), mib_idx); 3219 3220 tp->prior_ssthresh = 0; 3221 tcp_init_undo(tp); 3222 3223 if (!tcp_in_cwnd_reduction(sk)) { 3224 if (!ece_ack) 3225 tp->prior_ssthresh = tcp_current_ssthresh(sk); 3226 tcp_init_cwnd_reduction(sk); 3227 } 3228 tcp_set_ca_state(sk, TCP_CA_Recovery); 3229 } 3230 3231 static void tcp_update_rto_time(struct tcp_sock *tp) 3232 { 3233 if (tp->rto_stamp) { 3234 tp->total_rto_time += tcp_time_stamp_ms(tp) - tp->rto_stamp; 3235 tp->rto_stamp = 0; 3236 } 3237 } 3238 3239 /* Process an ACK in CA_Loss state. Move to CA_Open if lost data are 3240 * recovered or spurious. Otherwise retransmits more on partial ACKs. 3241 */ 3242 static void tcp_process_loss(struct sock *sk, int flag, int num_dupack, 3243 int *rexmit) 3244 { 3245 struct tcp_sock *tp = tcp_sk(sk); 3246 bool recovered = !before(tp->snd_una, tp->high_seq); 3247 3248 if ((flag & FLAG_SND_UNA_ADVANCED || rcu_access_pointer(tp->fastopen_rsk)) && 3249 tcp_try_undo_loss(sk, false)) 3250 return; 3251 3252 if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */ 3253 /* Step 3.b. A timeout is spurious if not all data are 3254 * lost, i.e., never-retransmitted data are (s)acked. 3255 */ 3256 if ((flag & FLAG_ORIG_SACK_ACKED) && 3257 tcp_try_undo_loss(sk, true)) 3258 return; 3259 3260 if (after(tp->snd_nxt, tp->high_seq)) { 3261 if (flag & FLAG_DATA_SACKED || num_dupack) 3262 tp->frto = 0; /* Step 3.a. loss was real */ 3263 } else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) { 3264 tp->high_seq = tp->snd_nxt; 3265 /* Step 2.b. Try send new data (but deferred until cwnd 3266 * is updated in tcp_ack()). Otherwise fall back to 3267 * the conventional recovery. 3268 */ 3269 if (!tcp_write_queue_empty(sk) && 3270 after(tcp_wnd_end(tp), tp->snd_nxt)) { 3271 *rexmit = REXMIT_NEW; 3272 return; 3273 } 3274 tp->frto = 0; 3275 } 3276 } 3277 3278 if (recovered) { 3279 /* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */ 3280 tcp_try_undo_recovery(sk); 3281 return; 3282 } 3283 if (tcp_is_reno(tp)) { 3284 /* A Reno DUPACK means new data in F-RTO step 2.b above are 3285 * delivered. Lower inflight to clock out (re)transmissions. 3286 */ 3287 if (after(tp->snd_nxt, tp->high_seq) && num_dupack) 3288 tcp_add_reno_sack(sk, num_dupack, flag & FLAG_ECE); 3289 else if (flag & FLAG_SND_UNA_ADVANCED) 3290 tcp_reset_reno_sack(tp); 3291 } 3292 *rexmit = REXMIT_LOST; 3293 } 3294 3295 /* Undo during fast recovery after partial ACK. */ 3296 static bool tcp_try_undo_partial(struct sock *sk, u32 prior_snd_una) 3297 { 3298 struct tcp_sock *tp = tcp_sk(sk); 3299 3300 if (tp->undo_marker && tcp_packet_delayed(tp)) { 3301 /* Plain luck! Hole if filled with delayed 3302 * packet, rather than with a retransmit. Check reordering. 3303 */ 3304 tcp_check_sack_reordering(sk, prior_snd_una, 1); 3305 3306 /* We are getting evidence that the reordering degree is higher 3307 * than we realized. If there are no retransmits out then we 3308 * can undo. Otherwise we clock out new packets but do not 3309 * mark more packets lost or retransmit more. 3310 */ 3311 if (tp->retrans_out) 3312 return true; 3313 3314 if (!tcp_any_retrans_done(sk)) 3315 tp->retrans_stamp = 0; 3316 3317 DBGUNDO(sk, "partial recovery"); 3318 tcp_undo_cwnd_reduction(sk, true); 3319 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); 3320 tcp_try_keep_open(sk); 3321 } 3322 return false; 3323 } 3324 3325 static void tcp_identify_packet_loss(struct sock *sk, int *ack_flag) 3326 { 3327 struct tcp_sock *tp = tcp_sk(sk); 3328 3329 if (tcp_rtx_queue_empty(sk)) 3330 return; 3331 3332 if (unlikely(tcp_is_reno(tp))) { 3333 tcp_newreno_mark_lost(sk, *ack_flag & FLAG_SND_UNA_ADVANCED); 3334 } else { 3335 u32 prior_retrans = tp->retrans_out; 3336 3337 if (tcp_rack_mark_lost(sk)) 3338 *ack_flag &= ~FLAG_SET_XMIT_TIMER; 3339 if (prior_retrans > tp->retrans_out) 3340 *ack_flag |= FLAG_LOST_RETRANS; 3341 } 3342 } 3343 3344 /* Process an event, which can update packets-in-flight not trivially. 3345 * Main goal of this function is to calculate new estimate for left_out, 3346 * taking into account both packets sitting in receiver's buffer and 3347 * packets lost by network. 3348 * 3349 * Besides that it updates the congestion state when packet loss or ECN 3350 * is detected. But it does not reduce the cwnd, it is done by the 3351 * congestion control later. 3352 * 3353 * It does _not_ decide what to send, it is made in function 3354 * tcp_xmit_retransmit_queue(). 3355 */ 3356 static void tcp_fastretrans_alert(struct sock *sk, const u32 prior_snd_una, 3357 int num_dupack, int *ack_flag, int *rexmit) 3358 { 3359 struct inet_connection_sock *icsk = inet_csk(sk); 3360 struct tcp_sock *tp = tcp_sk(sk); 3361 int flag = *ack_flag; 3362 bool ece_ack = flag & FLAG_ECE; 3363 3364 if (!tp->packets_out && tp->sacked_out) 3365 tp->sacked_out = 0; 3366 3367 /* Now state machine starts. 3368 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ 3369 if (ece_ack) 3370 tp->prior_ssthresh = 0; 3371 3372 /* B. In all the states check for reneging SACKs. */ 3373 if (tcp_check_sack_reneging(sk, ack_flag)) 3374 return; 3375 3376 /* C. Check consistency of the current state. */ 3377 tcp_verify_left_out(tp); 3378 3379 /* D. Check state exit conditions. State can be terminated 3380 * when high_seq is ACKed. */ 3381 if (icsk->icsk_ca_state == TCP_CA_Open) { 3382 WARN_ON(tp->retrans_out != 0 && !tp->syn_data); 3383 tp->retrans_stamp = 0; 3384 } else if (!before(tp->snd_una, tp->high_seq)) { 3385 switch (icsk->icsk_ca_state) { 3386 case TCP_CA_CWR: 3387 /* CWR is to be held something *above* high_seq 3388 * is ACKed for CWR bit to reach receiver. */ 3389 if (tp->snd_una != tp->high_seq) { 3390 tcp_end_cwnd_reduction(sk); 3391 tcp_set_ca_state(sk, TCP_CA_Open); 3392 } 3393 break; 3394 3395 case TCP_CA_Recovery: 3396 if (tcp_is_reno(tp)) 3397 tcp_reset_reno_sack(tp); 3398 if (tcp_try_undo_recovery(sk)) 3399 return; 3400 tcp_end_cwnd_reduction(sk); 3401 break; 3402 } 3403 } 3404 3405 /* E. Process state. */ 3406 switch (icsk->icsk_ca_state) { 3407 case TCP_CA_Recovery: 3408 if (!(flag & FLAG_SND_UNA_ADVANCED)) { 3409 if (tcp_is_reno(tp)) 3410 tcp_add_reno_sack(sk, num_dupack, ece_ack); 3411 } else if (tcp_try_undo_partial(sk, prior_snd_una)) 3412 return; 3413 3414 if (tcp_try_undo_dsack(sk)) 3415 tcp_try_to_open(sk, flag); 3416 3417 tcp_identify_packet_loss(sk, ack_flag); 3418 if (icsk->icsk_ca_state != TCP_CA_Recovery) { 3419 if (!tcp_time_to_recover(tp)) 3420 return; 3421 /* Undo reverts the recovery state. If loss is evident, 3422 * starts a new recovery (e.g. reordering then loss); 3423 */ 3424 tcp_enter_recovery(sk, ece_ack); 3425 } 3426 break; 3427 case TCP_CA_Loss: 3428 tcp_process_loss(sk, flag, num_dupack, rexmit); 3429 if (icsk->icsk_ca_state != TCP_CA_Loss) 3430 tcp_update_rto_time(tp); 3431 tcp_identify_packet_loss(sk, ack_flag); 3432 if (!(icsk->icsk_ca_state == TCP_CA_Open || 3433 (*ack_flag & FLAG_LOST_RETRANS))) 3434 return; 3435 /* Change state if cwnd is undone or retransmits are lost */ 3436 fallthrough; 3437 default: 3438 if (tcp_is_reno(tp)) { 3439 if (flag & FLAG_SND_UNA_ADVANCED) 3440 tcp_reset_reno_sack(tp); 3441 tcp_add_reno_sack(sk, num_dupack, ece_ack); 3442 } 3443 3444 if (icsk->icsk_ca_state <= TCP_CA_Disorder) 3445 tcp_try_undo_dsack(sk); 3446 3447 tcp_identify_packet_loss(sk, ack_flag); 3448 if (!tcp_time_to_recover(tp)) { 3449 tcp_try_to_open(sk, flag); 3450 return; 3451 } 3452 3453 /* MTU probe failure: don't reduce cwnd */ 3454 if (icsk->icsk_ca_state < TCP_CA_CWR && 3455 icsk->icsk_mtup.probe_size && 3456 tp->snd_una == tp->mtu_probe.probe_seq_start) { 3457 tcp_mtup_probe_failed(sk); 3458 /* Restores the reduction we did in tcp_mtup_probe() */ 3459 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1); 3460 tcp_simple_retransmit(sk); 3461 return; 3462 } 3463 3464 /* Otherwise enter Recovery state */ 3465 tcp_enter_recovery(sk, ece_ack); 3466 } 3467 3468 *rexmit = REXMIT_LOST; 3469 } 3470 3471 static void tcp_update_rtt_min(struct sock *sk, u32 rtt_us, const int flag) 3472 { 3473 u32 wlen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_rtt_wlen) * HZ; 3474 struct tcp_sock *tp = tcp_sk(sk); 3475 3476 if ((flag & FLAG_ACK_MAYBE_DELAYED) && rtt_us > tcp_min_rtt(tp)) { 3477 /* If the remote keeps returning delayed ACKs, eventually 3478 * the min filter would pick it up and overestimate the 3479 * prop. delay when it expires. Skip suspected delayed ACKs. 3480 */ 3481 return; 3482 } 3483 minmax_running_min(&tp->rtt_min, wlen, tcp_jiffies32, 3484 rtt_us ? : jiffies_to_usecs(1)); 3485 } 3486 3487 static bool tcp_ack_update_rtt(struct sock *sk, const int flag, 3488 long seq_rtt_us, long sack_rtt_us, 3489 long ca_rtt_us, struct rate_sample *rs) 3490 { 3491 const struct tcp_sock *tp = tcp_sk(sk); 3492 3493 /* Prefer RTT measured from ACK's timing to TS-ECR. This is because 3494 * broken middle-boxes or peers may corrupt TS-ECR fields. But 3495 * Karn's algorithm forbids taking RTT if some retransmitted data 3496 * is acked (RFC6298). 3497 */ 3498 if (seq_rtt_us < 0) 3499 seq_rtt_us = sack_rtt_us; 3500 3501 /* RTTM Rule: A TSecr value received in a segment is used to 3502 * update the averaged RTT measurement only if the segment 3503 * acknowledges some new data, i.e., only if it advances the 3504 * left edge of the send window. 3505 * See draft-ietf-tcplw-high-performance-00, section 3.3. 3506 */ 3507 if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && 3508 tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED) 3509 seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp, 1); 3510 3511 rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */ 3512 if (seq_rtt_us < 0) 3513 return false; 3514 3515 /* ca_rtt_us >= 0 is counting on the invariant that ca_rtt_us is 3516 * always taken together with ACK, SACK, or TS-opts. Any negative 3517 * values will be skipped with the seq_rtt_us < 0 check above. 3518 */ 3519 tcp_update_rtt_min(sk, ca_rtt_us, flag); 3520 tcp_rtt_estimator(sk, seq_rtt_us); 3521 tcp_set_rto(sk); 3522 3523 /* RFC6298: only reset backoff on valid RTT measurement. */ 3524 inet_csk(sk)->icsk_backoff = 0; 3525 return true; 3526 } 3527 3528 /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ 3529 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) 3530 { 3531 struct rate_sample rs; 3532 long rtt_us = -1L; 3533 3534 if (req && !req->num_retrans && tcp_rsk(req)->snt_synack) 3535 rtt_us = tcp_stamp_us_delta(tcp_clock_us(), tcp_rsk(req)->snt_synack); 3536 3537 tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, rtt_us, -1L, rtt_us, &rs); 3538 } 3539 3540 3541 static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 3542 { 3543 const struct inet_connection_sock *icsk = inet_csk(sk); 3544 3545 icsk->icsk_ca_ops->cong_avoid(sk, ack, acked); 3546 tcp_sk(sk)->snd_cwnd_stamp = tcp_jiffies32; 3547 } 3548 3549 /* Restart timer after forward progress on connection. 3550 * RFC2988 recommends to restart timer to now+rto. 3551 */ 3552 void tcp_rearm_rto(struct sock *sk) 3553 { 3554 const struct inet_connection_sock *icsk = inet_csk(sk); 3555 struct tcp_sock *tp = tcp_sk(sk); 3556 3557 /* If the retrans timer is currently being used by Fast Open 3558 * for SYN-ACK retrans purpose, stay put. 3559 */ 3560 if (rcu_access_pointer(tp->fastopen_rsk)) 3561 return; 3562 3563 if (!tp->packets_out) { 3564 inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); 3565 } else { 3566 u32 rto = inet_csk(sk)->icsk_rto; 3567 /* Offset the time elapsed after installing regular RTO */ 3568 if (icsk->icsk_pending == ICSK_TIME_REO_TIMEOUT || 3569 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 3570 s64 delta_us = tcp_rto_delta_us(sk); 3571 /* delta_us may not be positive if the socket is locked 3572 * when the retrans timer fires and is rescheduled. 3573 */ 3574 rto = usecs_to_jiffies(max_t(int, delta_us, 1)); 3575 } 3576 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto, true); 3577 } 3578 } 3579 3580 /* Try to schedule a loss probe; if that doesn't work, then schedule an RTO. */ 3581 static void tcp_set_xmit_timer(struct sock *sk) 3582 { 3583 if (!tcp_sk(sk)->packets_out || !tcp_schedule_loss_probe(sk, true)) 3584 tcp_rearm_rto(sk); 3585 } 3586 3587 /* If we get here, the whole TSO packet has not been acked. */ 3588 static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) 3589 { 3590 struct tcp_sock *tp = tcp_sk(sk); 3591 u32 packets_acked; 3592 3593 BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); 3594 3595 packets_acked = tcp_skb_pcount(skb); 3596 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3597 return 0; 3598 packets_acked -= tcp_skb_pcount(skb); 3599 3600 if (packets_acked) { 3601 BUG_ON(tcp_skb_pcount(skb) == 0); 3602 BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); 3603 } 3604 3605 return packets_acked; 3606 } 3607 3608 static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb, 3609 const struct sk_buff *ack_skb, u32 prior_snd_una) 3610 { 3611 const struct skb_shared_info *shinfo; 3612 3613 /* Avoid cache line misses to get skb_shinfo() and shinfo->tx_flags */ 3614 if (likely(!TCP_SKB_CB(skb)->txstamp_ack)) 3615 return; 3616 3617 shinfo = skb_shinfo(skb); 3618 if (!before(shinfo->tskey, prior_snd_una) && 3619 before(shinfo->tskey, tcp_sk(sk)->snd_una)) { 3620 tcp_skb_tsorted_save(skb) { 3621 __skb_tstamp_tx(skb, ack_skb, NULL, sk, SCM_TSTAMP_ACK); 3622 } tcp_skb_tsorted_restore(skb); 3623 } 3624 } 3625 3626 /* Remove acknowledged frames from the retransmission queue. If our packet 3627 * is before the ack sequence we can discard it as it's confirmed to have 3628 * arrived at the other end. 3629 */ 3630 static int tcp_clean_rtx_queue(struct sock *sk, const struct sk_buff *ack_skb, 3631 u32 prior_fack, u32 prior_snd_una, 3632 struct tcp_sacktag_state *sack, bool ece_ack) 3633 { 3634 const struct inet_connection_sock *icsk = inet_csk(sk); 3635 u64 first_ackt, last_ackt; 3636 struct tcp_sock *tp = tcp_sk(sk); 3637 u32 prior_sacked = tp->sacked_out; 3638 u32 reord = tp->snd_nxt; /* lowest acked un-retx un-sacked seq */ 3639 struct sk_buff *skb, *next; 3640 bool fully_acked = true; 3641 long sack_rtt_us = -1L; 3642 long seq_rtt_us = -1L; 3643 long ca_rtt_us = -1L; 3644 u32 pkts_acked = 0; 3645 bool rtt_update; 3646 int flag = 0; 3647 3648 first_ackt = 0; 3649 3650 for (skb = skb_rb_first(&sk->tcp_rtx_queue); skb; skb = next) { 3651 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 3652 const u32 start_seq = scb->seq; 3653 u8 sacked = scb->sacked; 3654 u32 acked_pcount; 3655 3656 /* Determine how many packets and what bytes were acked, tso and else */ 3657 if (after(scb->end_seq, tp->snd_una)) { 3658 if (tcp_skb_pcount(skb) == 1 || 3659 !after(tp->snd_una, scb->seq)) 3660 break; 3661 3662 acked_pcount = tcp_tso_acked(sk, skb); 3663 if (!acked_pcount) 3664 break; 3665 fully_acked = false; 3666 } else { 3667 acked_pcount = tcp_skb_pcount(skb); 3668 } 3669 3670 if (unlikely(sacked & TCPCB_RETRANS)) { 3671 if (sacked & TCPCB_SACKED_RETRANS) 3672 tp->retrans_out -= acked_pcount; 3673 flag |= FLAG_RETRANS_DATA_ACKED; 3674 } else if (!(sacked & TCPCB_SACKED_ACKED)) { 3675 last_ackt = tcp_skb_timestamp_us(skb); 3676 WARN_ON_ONCE(last_ackt == 0); 3677 if (!first_ackt) 3678 first_ackt = last_ackt; 3679 3680 if (before(start_seq, reord)) 3681 reord = start_seq; 3682 if (!after(scb->end_seq, tp->high_seq)) 3683 flag |= FLAG_ORIG_SACK_ACKED; 3684 } 3685 3686 if (sacked & TCPCB_SACKED_ACKED) { 3687 tp->sacked_out -= acked_pcount; 3688 /* snd_una delta covers these skbs */ 3689 sack->delivered_bytes -= skb->len; 3690 } else if (tcp_is_sack(tp)) { 3691 tcp_count_delivered(tp, acked_pcount, ece_ack); 3692 if (!tcp_skb_spurious_retrans(tp, skb)) 3693 tcp_rack_advance(tp, sacked, scb->end_seq, 3694 tcp_skb_timestamp_us(skb)); 3695 } 3696 if (sacked & TCPCB_LOST) 3697 tp->lost_out -= acked_pcount; 3698 3699 tp->packets_out -= acked_pcount; 3700 pkts_acked += acked_pcount; 3701 tcp_rate_skb_delivered(sk, skb, sack->rate); 3702 3703 /* Initial outgoing SYN's get put onto the write_queue 3704 * just like anything else we transmit. It is not 3705 * true data, and if we misinform our callers that 3706 * this ACK acks real data, we will erroneously exit 3707 * connection startup slow start one packet too 3708 * quickly. This is severely frowned upon behavior. 3709 */ 3710 if (likely(!(scb->tcp_flags & TCPHDR_SYN))) { 3711 flag |= FLAG_DATA_ACKED; 3712 } else { 3713 flag |= FLAG_SYN_ACKED; 3714 tp->retrans_stamp = 0; 3715 } 3716 3717 if (!fully_acked) 3718 break; 3719 3720 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); 3721 3722 next = skb_rb_next(skb); 3723 if (unlikely(skb == tp->retransmit_skb_hint)) 3724 tp->retransmit_skb_hint = NULL; 3725 tcp_highest_sack_replace(sk, skb, next); 3726 tcp_rtx_queue_unlink_and_free(skb, sk); 3727 } 3728 3729 if (!skb) 3730 tcp_chrono_stop(sk, TCP_CHRONO_BUSY); 3731 3732 if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) 3733 tp->snd_up = tp->snd_una; 3734 3735 if (skb) { 3736 tcp_ack_tstamp(sk, skb, ack_skb, prior_snd_una); 3737 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3738 flag |= FLAG_SACK_RENEGING; 3739 } 3740 3741 if (likely(first_ackt) && !(flag & FLAG_RETRANS_DATA_ACKED)) { 3742 seq_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, first_ackt); 3743 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, last_ackt); 3744 3745 if (pkts_acked == 1 && fully_acked && !prior_sacked && 3746 (tp->snd_una - prior_snd_una) < tp->mss_cache && 3747 sack->rate->prior_delivered + 1 == tp->delivered && 3748 !(flag & (FLAG_CA_ALERT | FLAG_SYN_ACKED))) { 3749 /* Conservatively mark a delayed ACK. It's typically 3750 * from a lone runt packet over the round trip to 3751 * a receiver w/o out-of-order or CE events. 3752 */ 3753 flag |= FLAG_ACK_MAYBE_DELAYED; 3754 } 3755 } 3756 if (sack->first_sackt) { 3757 sack_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->first_sackt); 3758 ca_rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, sack->last_sackt); 3759 } 3760 rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us, 3761 ca_rtt_us, sack->rate); 3762 3763 if (flag & FLAG_ACKED) { 3764 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3765 if (unlikely(icsk->icsk_mtup.probe_size && 3766 !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { 3767 tcp_mtup_probe_success(sk); 3768 } 3769 3770 if (tcp_is_reno(tp)) { 3771 tcp_remove_reno_sacks(sk, pkts_acked, ece_ack); 3772 3773 /* If any of the cumulatively ACKed segments was 3774 * retransmitted, non-SACK case cannot confirm that 3775 * progress was due to original transmission due to 3776 * lack of TCPCB_SACKED_ACKED bits even if some of 3777 * the packets may have been never retransmitted. 3778 */ 3779 if (flag & FLAG_RETRANS_DATA_ACKED) 3780 flag &= ~FLAG_ORIG_SACK_ACKED; 3781 } else { 3782 /* Non-retransmitted hole got filled? That's reordering */ 3783 if (before(reord, prior_fack)) 3784 tcp_check_sack_reordering(sk, reord, 0); 3785 } 3786 3787 sack->delivered_bytes = (skb ? 3788 TCP_SKB_CB(skb)->seq : tp->snd_una) - 3789 prior_snd_una; 3790 } else if (skb && rtt_update && sack_rtt_us >= 0 && 3791 sack_rtt_us > tcp_stamp_us_delta(tp->tcp_mstamp, 3792 tcp_skb_timestamp_us(skb))) { 3793 /* Do not re-arm RTO if the sack RTT is measured from data sent 3794 * after when the head was last (re)transmitted. Otherwise the 3795 * timeout may continue to extend in loss recovery. 3796 */ 3797 flag |= FLAG_SET_XMIT_TIMER; /* set TLP or RTO timer */ 3798 } 3799 3800 if (icsk->icsk_ca_ops->pkts_acked) { 3801 struct ack_sample sample = { .pkts_acked = pkts_acked, 3802 .rtt_us = sack->rate->rtt_us }; 3803 3804 sample.in_flight = tp->mss_cache * 3805 (tp->delivered - sack->rate->prior_delivered); 3806 icsk->icsk_ca_ops->pkts_acked(sk, &sample); 3807 } 3808 3809 #if FASTRETRANS_DEBUG > 0 3810 WARN_ON((int)tp->sacked_out < 0); 3811 WARN_ON((int)tp->lost_out < 0); 3812 WARN_ON((int)tp->retrans_out < 0); 3813 if (!tp->packets_out && tcp_is_sack(tp)) { 3814 icsk = inet_csk(sk); 3815 if (tp->lost_out) { 3816 pr_debug("Leak l=%u %d\n", 3817 tp->lost_out, icsk->icsk_ca_state); 3818 tp->lost_out = 0; 3819 } 3820 if (tp->sacked_out) { 3821 pr_debug("Leak s=%u %d\n", 3822 tp->sacked_out, icsk->icsk_ca_state); 3823 tp->sacked_out = 0; 3824 } 3825 if (tp->retrans_out) { 3826 pr_debug("Leak r=%u %d\n", 3827 tp->retrans_out, icsk->icsk_ca_state); 3828 tp->retrans_out = 0; 3829 } 3830 } 3831 #endif 3832 return flag; 3833 } 3834 3835 static void tcp_ack_probe(struct sock *sk) 3836 { 3837 struct inet_connection_sock *icsk = inet_csk(sk); 3838 struct sk_buff *head = tcp_send_head(sk); 3839 const struct tcp_sock *tp = tcp_sk(sk); 3840 3841 /* Was it a usable window open? */ 3842 if (!head) 3843 return; 3844 if (!after(TCP_SKB_CB(head)->end_seq, tcp_wnd_end(tp))) { 3845 icsk->icsk_backoff = 0; 3846 icsk->icsk_probes_tstamp = 0; 3847 inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); 3848 /* Socket must be waked up by subsequent tcp_data_snd_check(). 3849 * This function is not for random using! 3850 */ 3851 } else { 3852 unsigned long when = tcp_probe0_when(sk, tcp_rto_max(sk)); 3853 3854 when = tcp_clamp_probe0_to_user_timeout(sk, when); 3855 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when, true); 3856 } 3857 } 3858 3859 static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) 3860 { 3861 return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || 3862 inet_csk(sk)->icsk_ca_state != TCP_CA_Open; 3863 } 3864 3865 /* Decide wheather to run the increase function of congestion control. */ 3866 static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) 3867 { 3868 /* If reordering is high then always grow cwnd whenever data is 3869 * delivered regardless of its ordering. Otherwise stay conservative 3870 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ 3871 * new SACK or ECE mark may first advance cwnd here and later reduce 3872 * cwnd in tcp_fastretrans_alert() based on more states. 3873 */ 3874 if (tcp_sk(sk)->reordering > 3875 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering)) 3876 return flag & FLAG_FORWARD_PROGRESS; 3877 3878 return flag & FLAG_DATA_ACKED; 3879 } 3880 3881 /* The "ultimate" congestion control function that aims to replace the rigid 3882 * cwnd increase and decrease control (tcp_cong_avoid,tcp_*cwnd_reduction). 3883 * It's called toward the end of processing an ACK with precise rate 3884 * information. All transmission or retransmission are delayed afterwards. 3885 */ 3886 static void tcp_cong_control(struct sock *sk, u32 ack, u32 acked_sacked, 3887 int flag, const struct rate_sample *rs) 3888 { 3889 const struct inet_connection_sock *icsk = inet_csk(sk); 3890 3891 if (icsk->icsk_ca_ops->cong_control) { 3892 icsk->icsk_ca_ops->cong_control(sk, ack, flag, rs); 3893 return; 3894 } 3895 3896 if (tcp_in_cwnd_reduction(sk)) { 3897 /* Reduce cwnd if state mandates */ 3898 tcp_cwnd_reduction(sk, acked_sacked, rs->losses, flag); 3899 } else if (tcp_may_raise_cwnd(sk, flag)) { 3900 /* Advance cwnd if state allows */ 3901 tcp_cong_avoid(sk, ack, acked_sacked); 3902 } 3903 tcp_update_pacing_rate(sk); 3904 } 3905 3906 /* Check that window update is acceptable. 3907 * The function assumes that snd_una<=ack<=snd_next. 3908 */ 3909 static inline bool tcp_may_update_window(const struct tcp_sock *tp, 3910 const u32 ack, const u32 ack_seq, 3911 const u32 nwin) 3912 { 3913 return after(ack, tp->snd_una) || 3914 after(ack_seq, tp->snd_wl1) || 3915 (ack_seq == tp->snd_wl1 && (nwin > tp->snd_wnd || !nwin)); 3916 } 3917 3918 static void tcp_snd_sne_update(struct tcp_sock *tp, u32 ack) 3919 { 3920 #ifdef CONFIG_TCP_AO 3921 struct tcp_ao_info *ao; 3922 3923 if (!static_branch_unlikely(&tcp_ao_needed.key)) 3924 return; 3925 3926 ao = rcu_dereference_protected(tp->ao_info, 3927 lockdep_sock_is_held((struct sock *)tp)); 3928 if (ao && ack < tp->snd_una) { 3929 ao->snd_sne++; 3930 trace_tcp_ao_snd_sne_update((struct sock *)tp, ao->snd_sne); 3931 } 3932 #endif 3933 } 3934 3935 /* If we update tp->snd_una, also update tp->bytes_acked */ 3936 static void tcp_snd_una_update(struct tcp_sock *tp, u32 ack) 3937 { 3938 u32 delta = ack - tp->snd_una; 3939 3940 sock_owned_by_me((struct sock *)tp); 3941 tp->bytes_acked += delta; 3942 tcp_snd_sne_update(tp, ack); 3943 WRITE_ONCE(tp->snd_una, ack); 3944 } 3945 3946 static void tcp_rcv_sne_update(struct tcp_sock *tp, u32 seq) 3947 { 3948 #ifdef CONFIG_TCP_AO 3949 struct tcp_ao_info *ao; 3950 3951 if (!static_branch_unlikely(&tcp_ao_needed.key)) 3952 return; 3953 3954 ao = rcu_dereference_protected(tp->ao_info, 3955 lockdep_sock_is_held((struct sock *)tp)); 3956 if (ao && seq < tp->rcv_nxt) { 3957 ao->rcv_sne++; 3958 trace_tcp_ao_rcv_sne_update((struct sock *)tp, ao->rcv_sne); 3959 } 3960 #endif 3961 } 3962 3963 /* If we update tp->rcv_nxt, also update tp->bytes_received */ 3964 static void tcp_rcv_nxt_update(struct tcp_sock *tp, u32 seq) 3965 { 3966 u32 delta = seq - tp->rcv_nxt; 3967 3968 sock_owned_by_me((struct sock *)tp); 3969 tp->bytes_received += delta; 3970 tcp_rcv_sne_update(tp, seq); 3971 WRITE_ONCE(tp->rcv_nxt, seq); 3972 } 3973 3974 /* Update our send window. 3975 * 3976 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 3977 * and in FreeBSD. NetBSD's one is even worse.) is wrong. 3978 */ 3979 static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack, 3980 u32 ack_seq) 3981 { 3982 struct tcp_sock *tp = tcp_sk(sk); 3983 int flag = 0; 3984 u32 nwin = ntohs(tcp_hdr(skb)->window); 3985 3986 if (likely(!tcp_hdr(skb)->syn)) 3987 nwin <<= tp->rx_opt.snd_wscale; 3988 3989 if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { 3990 tcp_update_wl(tp, ack_seq); 3991 3992 if (tp->snd_wnd != nwin) { 3993 tp->snd_wnd = nwin; 3994 flag |= FLAG_WIN_UPDATE; 3995 3996 /* Note, it is the only place, where 3997 * fast path is recovered for sending TCP. 3998 */ 3999 tp->pred_flags = 0; 4000 tcp_fast_path_check(sk); 4001 4002 if (!tcp_write_queue_empty(sk)) 4003 tcp_slow_start_after_idle_check(sk); 4004 4005 if (nwin > tp->max_window) { 4006 tp->max_window = nwin; 4007 tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); 4008 } 4009 } 4010 } 4011 4012 tcp_snd_una_update(tp, ack); 4013 4014 return flag; 4015 } 4016 4017 static bool __tcp_oow_rate_limited(struct net *net, int mib_idx, 4018 u32 *last_oow_ack_time) 4019 { 4020 /* Paired with the WRITE_ONCE() in this function. */ 4021 u32 val = READ_ONCE(*last_oow_ack_time); 4022 4023 if (val) { 4024 s32 elapsed = (s32)(tcp_jiffies32 - val); 4025 4026 if (0 <= elapsed && 4027 elapsed < READ_ONCE(net->ipv4.sysctl_tcp_invalid_ratelimit)) { 4028 NET_INC_STATS(net, mib_idx); 4029 return true; /* rate-limited: don't send yet! */ 4030 } 4031 } 4032 4033 /* Paired with the prior READ_ONCE() and with itself, 4034 * as we might be lockless. 4035 */ 4036 WRITE_ONCE(*last_oow_ack_time, tcp_jiffies32); 4037 4038 return false; /* not rate-limited: go ahead, send dupack now! */ 4039 } 4040 4041 /* Return true if we're currently rate-limiting out-of-window ACKs and 4042 * thus shouldn't send a dupack right now. We rate-limit dupacks in 4043 * response to out-of-window SYNs or ACKs to mitigate ACK loops or DoS 4044 * attacks that send repeated SYNs or ACKs for the same connection. To 4045 * do this, we do not send a duplicate SYNACK or ACK if the remote 4046 * endpoint is sending out-of-window SYNs or pure ACKs at a high rate. 4047 */ 4048 bool tcp_oow_rate_limited(struct net *net, const struct sk_buff *skb, 4049 int mib_idx, u32 *last_oow_ack_time) 4050 { 4051 /* Data packets without SYNs are not likely part of an ACK loop. */ 4052 if ((TCP_SKB_CB(skb)->seq != TCP_SKB_CB(skb)->end_seq) && 4053 !tcp_hdr(skb)->syn) 4054 return false; 4055 4056 return __tcp_oow_rate_limited(net, mib_idx, last_oow_ack_time); 4057 } 4058 4059 static void tcp_send_ack_reflect_ect(struct sock *sk, bool accecn_reflector) 4060 { 4061 struct tcp_sock *tp = tcp_sk(sk); 4062 u16 flags = 0; 4063 4064 if (accecn_reflector) 4065 flags = tcp_accecn_reflector_flags(tp->syn_ect_rcv); 4066 __tcp_send_ack(sk, tp->rcv_nxt, flags); 4067 } 4068 4069 /* Consume one slot from the per-netns RFC 5961 challenge ACK quota. 4070 * Returns true if a challenge ACK may be sent. 4071 */ 4072 static bool tcp_challenge_ack_allowed(struct net *net) 4073 { 4074 u32 count, now, ack_limit; 4075 4076 ack_limit = READ_ONCE(net->ipv4.sysctl_tcp_challenge_ack_limit); 4077 if (ack_limit == INT_MAX) 4078 return true; 4079 4080 now = jiffies / HZ; 4081 if (now != READ_ONCE(net->ipv4.tcp_challenge_timestamp)) { 4082 u32 half = (ack_limit + 1) >> 1; 4083 4084 WRITE_ONCE(net->ipv4.tcp_challenge_timestamp, now); 4085 WRITE_ONCE(net->ipv4.tcp_challenge_count, 4086 get_random_u32_inclusive(half, ack_limit + half - 1)); 4087 } 4088 count = READ_ONCE(net->ipv4.tcp_challenge_count); 4089 if (count > 0) { 4090 WRITE_ONCE(net->ipv4.tcp_challenge_count, count - 1); 4091 return true; 4092 } 4093 return false; 4094 } 4095 4096 /* RFC 5961 7 [ACK Throttling] */ 4097 static void tcp_send_challenge_ack(struct sock *sk, bool accecn_reflector) 4098 { 4099 struct tcp_sock *tp = tcp_sk(sk); 4100 struct net *net = sock_net(sk); 4101 4102 /* First check our per-socket dupack rate limit. */ 4103 if (__tcp_oow_rate_limited(net, 4104 LINUX_MIB_TCPACKSKIPPEDCHALLENGE, 4105 &tp->last_oow_ack_time)) 4106 return; 4107 4108 /* Then check the per-netns RFC 5961 rate limit. */ 4109 if (tcp_challenge_ack_allowed(net)) { 4110 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK); 4111 tcp_send_ack_reflect_ect(sk, accecn_reflector); 4112 } 4113 } 4114 4115 /* Send a challenge ACK from a SYN-RECEIVED request socket. Uses 4116 * __tcp_oow_rate_limited() directly so that an RST carrying payload 4117 * cannot bypass the per-request rate limit. 4118 */ 4119 void tcp_reqsk_send_challenge_ack(struct sock *sk, struct sk_buff *skb, 4120 struct request_sock *req) 4121 { 4122 struct net *net = sock_net(sk); 4123 4124 if (__tcp_oow_rate_limited(net, LINUX_MIB_TCPACKSKIPPEDCHALLENGE, 4125 &tcp_rsk(req)->last_oow_ack_time)) 4126 return; 4127 4128 if (tcp_challenge_ack_allowed(net)) { 4129 NET_INC_STATS(net, LINUX_MIB_TCPCHALLENGEACK); 4130 req->rsk_ops->send_ack(sk, skb, req); 4131 } 4132 } 4133 4134 static void tcp_store_ts_recent(struct tcp_sock *tp) 4135 { 4136 tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; 4137 tp->rx_opt.ts_recent_stamp = ktime_get_seconds(); 4138 } 4139 4140 static int __tcp_replace_ts_recent(struct tcp_sock *tp, s32 tstamp_delta) 4141 { 4142 tcp_store_ts_recent(tp); 4143 return tstamp_delta > 0 ? FLAG_TS_PROGRESS : 0; 4144 } 4145 4146 static int tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) 4147 { 4148 s32 delta; 4149 4150 if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { 4151 /* PAWS bug workaround wrt. ACK frames, the PAWS discard 4152 * extra check below makes sure this can only happen 4153 * for pure ACK frames. -DaveM 4154 * 4155 * Not only, also it occurs for expired timestamps. 4156 */ 4157 4158 if (tcp_paws_check(&tp->rx_opt, 0)) { 4159 delta = tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent; 4160 return __tcp_replace_ts_recent(tp, delta); 4161 } 4162 } 4163 4164 return 0; 4165 } 4166 4167 /* This routine deals with acks during a TLP episode and ends an episode by 4168 * resetting tlp_high_seq. Ref: TLP algorithm in RFC8985 4169 */ 4170 static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag) 4171 { 4172 struct tcp_sock *tp = tcp_sk(sk); 4173 4174 if (before(ack, tp->tlp_high_seq)) 4175 return; 4176 4177 if (!tp->tlp_retrans) { 4178 /* TLP of new data has been acknowledged */ 4179 tp->tlp_high_seq = 0; 4180 } else if (flag & FLAG_DSACK_TLP) { 4181 /* This DSACK means original and TLP probe arrived; no loss */ 4182 tp->tlp_high_seq = 0; 4183 } else if (after(ack, tp->tlp_high_seq)) { 4184 /* ACK advances: there was a loss, so reduce cwnd. Reset 4185 * tlp_high_seq in tcp_init_cwnd_reduction() 4186 */ 4187 tcp_init_cwnd_reduction(sk); 4188 tcp_set_ca_state(sk, TCP_CA_CWR); 4189 tcp_end_cwnd_reduction(sk); 4190 tcp_try_keep_open(sk); 4191 NET_INC_STATS(sock_net(sk), 4192 LINUX_MIB_TCPLOSSPROBERECOVERY); 4193 } else if (!(flag & (FLAG_SND_UNA_ADVANCED | 4194 FLAG_NOT_DUP | FLAG_DATA_SACKED))) { 4195 /* Pure dupack: original and TLP probe arrived; no loss */ 4196 tp->tlp_high_seq = 0; 4197 } 4198 } 4199 4200 static void tcp_in_ack_event(struct sock *sk, int flag) 4201 { 4202 const struct inet_connection_sock *icsk = inet_csk(sk); 4203 4204 if (icsk->icsk_ca_ops->in_ack_event) { 4205 u32 ack_ev_flags = 0; 4206 4207 if (flag & FLAG_WIN_UPDATE) 4208 ack_ev_flags |= CA_ACK_WIN_UPDATE; 4209 if (flag & FLAG_SLOWPATH) { 4210 ack_ev_flags |= CA_ACK_SLOWPATH; 4211 if (flag & FLAG_ECE) 4212 ack_ev_flags |= CA_ACK_ECE; 4213 } 4214 4215 icsk->icsk_ca_ops->in_ack_event(sk, ack_ev_flags); 4216 } 4217 } 4218 4219 /* Congestion control has updated the cwnd already. So if we're in 4220 * loss recovery then now we do any new sends (for FRTO) or 4221 * retransmits (for CA_Loss or CA_recovery) that make sense. 4222 */ 4223 static void tcp_xmit_recovery(struct sock *sk, int rexmit) 4224 { 4225 struct tcp_sock *tp = tcp_sk(sk); 4226 4227 if (rexmit == REXMIT_NONE || sk->sk_state == TCP_SYN_SENT) 4228 return; 4229 4230 if (unlikely(rexmit == REXMIT_NEW)) { 4231 __tcp_push_pending_frames(sk, tcp_current_mss(sk), 4232 TCP_NAGLE_OFF); 4233 if (after(tp->snd_nxt, tp->high_seq)) 4234 return; 4235 tp->frto = 0; 4236 } 4237 tcp_xmit_retransmit_queue(sk); 4238 } 4239 4240 /* Returns the number of packets newly acked or sacked by the current ACK */ 4241 static u32 tcp_newly_delivered(struct sock *sk, u32 prior_delivered, 4242 u32 ecn_count, int flag) 4243 { 4244 const struct net *net = sock_net(sk); 4245 struct tcp_sock *tp = tcp_sk(sk); 4246 u32 delivered; 4247 4248 delivered = tp->delivered - prior_delivered; 4249 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVERED, delivered); 4250 4251 if (flag & FLAG_ECE) { 4252 if (tcp_ecn_mode_rfc3168(tp)) 4253 ecn_count = delivered; 4254 NET_ADD_STATS(net, LINUX_MIB_TCPDELIVEREDCE, ecn_count); 4255 } 4256 4257 return delivered; 4258 } 4259 4260 /* Updates the RACK's reo_wnd based on DSACK and no. of recoveries. 4261 * 4262 * If a DSACK is received that seems like it may have been due to reordering 4263 * triggering fast recovery, increment reo_wnd by min_rtt/4 (upper bounded 4264 * by srtt), since there is possibility that spurious retransmission was 4265 * due to reordering delay longer than reo_wnd. 4266 * 4267 * Persist the current reo_wnd value for TCP_RACK_RECOVERY_THRESH (16) 4268 * no. of successful recoveries (accounts for full DSACK-based loss 4269 * recovery undo). After that, reset it to default (min_rtt/4). 4270 * 4271 * At max, reo_wnd is incremented only once per rtt. So that the new 4272 * DSACK on which we are reacting, is due to the spurious retx (approx) 4273 * after the reo_wnd has been updated last time. 4274 * 4275 * reo_wnd is tracked in terms of steps (of min_rtt/4), rather than 4276 * absolute value to account for change in rtt. 4277 */ 4278 static void tcp_rack_update_reo_wnd(struct sock *sk, struct rate_sample *rs) 4279 { 4280 struct tcp_sock *tp = tcp_sk(sk); 4281 4282 if ((READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) & 4283 TCP_RACK_STATIC_REO_WND) || 4284 !rs->prior_delivered) 4285 return; 4286 4287 /* Disregard DSACK if a rtt has not passed since we adjusted reo_wnd */ 4288 if (before(rs->prior_delivered, tp->rack.last_delivered)) 4289 tp->rack.dsack_seen = 0; 4290 4291 /* Adjust the reo_wnd if update is pending */ 4292 if (tp->rack.dsack_seen) { 4293 tp->rack.reo_wnd_steps = min_t(u32, 0xFF, 4294 tp->rack.reo_wnd_steps + 1); 4295 tp->rack.dsack_seen = 0; 4296 tp->rack.last_delivered = tp->delivered; 4297 tp->rack.reo_wnd_persist = TCP_RACK_RECOVERY_THRESH; 4298 } else if (!tp->rack.reo_wnd_persist) { 4299 tp->rack.reo_wnd_steps = 1; 4300 } 4301 } 4302 4303 /* This routine deals with incoming acks, but not outgoing ones. */ 4304 static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) 4305 { 4306 struct inet_connection_sock *icsk = inet_csk(sk); 4307 struct tcp_sock *tp = tcp_sk(sk); 4308 struct tcp_sacktag_state sack_state; 4309 struct rate_sample rs = { .prior_delivered = 0 }; 4310 u32 prior_snd_una = tp->snd_una; 4311 bool is_sack_reneg = tp->is_sack_reneg; 4312 u32 ack_seq = TCP_SKB_CB(skb)->seq; 4313 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4314 int num_dupack = 0; 4315 int prior_packets = tp->packets_out; 4316 u32 delivered = tp->delivered; 4317 u32 lost = tp->lost; 4318 int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */ 4319 u32 ecn_count = 0; /* Did we receive ECE/an AccECN ACE update? */ 4320 u32 prior_fack; 4321 4322 sack_state.first_sackt = 0; 4323 sack_state.rate = &rs; 4324 sack_state.sack_delivered = 0; 4325 sack_state.delivered_bytes = 0; 4326 4327 /* We very likely will need to access rtx queue. */ 4328 prefetch(sk->tcp_rtx_queue.rb_node); 4329 4330 /* If the ack is older than previous acks 4331 * then we can probably ignore it. 4332 */ 4333 if (before(ack, prior_snd_una)) { 4334 u32 max_window; 4335 4336 /* do not accept ACK for bytes we never sent. */ 4337 max_window = min_t(u64, tp->max_window, tp->bytes_acked); 4338 /* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */ 4339 if (before(ack, prior_snd_una - max_window)) { 4340 if (!(flag & FLAG_NO_CHALLENGE_ACK)) 4341 tcp_send_challenge_ack(sk, false); 4342 return -SKB_DROP_REASON_TCP_TOO_OLD_ACK; 4343 } 4344 goto old_ack; 4345 } 4346 4347 /* If the ack includes data we haven't sent yet, drop the 4348 * segment. RFC 793 Section 3.9 and RFC 5961 Section 5.2 4349 * require us to send an ACK back in that case. 4350 */ 4351 if (after(ack, tp->snd_nxt)) { 4352 if (!(flag & FLAG_NO_CHALLENGE_ACK)) 4353 tcp_send_challenge_ack(sk, false); 4354 return -SKB_DROP_REASON_TCP_ACK_UNSENT_DATA; 4355 } 4356 4357 if (after(ack, prior_snd_una)) { 4358 flag |= FLAG_SND_UNA_ADVANCED; 4359 WRITE_ONCE(icsk->icsk_retransmits, 0); 4360 4361 #if IS_ENABLED(CONFIG_TLS_DEVICE) 4362 if (static_branch_unlikely(&clean_acked_data_enabled.key)) 4363 if (tp->tcp_clean_acked) 4364 tp->tcp_clean_acked(sk, ack); 4365 #endif 4366 } 4367 4368 prior_fack = tcp_is_sack(tp) ? tcp_highest_sack_seq(tp) : tp->snd_una; 4369 rs.prior_in_flight = tcp_packets_in_flight(tp); 4370 4371 /* ts_recent update must be made after we are sure that the packet 4372 * is in window. 4373 */ 4374 if (flag & FLAG_UPDATE_TS_RECENT) 4375 flag |= tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4376 4377 if ((flag & (FLAG_SLOWPATH | FLAG_SND_UNA_ADVANCED)) == 4378 FLAG_SND_UNA_ADVANCED) { 4379 /* Window is constant, pure forward advance. 4380 * No more checks are required. 4381 * Note, we use the fact that SND.UNA>=SND.WL2. 4382 */ 4383 tcp_update_wl(tp, ack_seq); 4384 tcp_snd_una_update(tp, ack); 4385 4386 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPACKS); 4387 } else { 4388 if (ack_seq != TCP_SKB_CB(skb)->end_seq) 4389 flag |= FLAG_DATA; 4390 else 4391 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPUREACKS); 4392 4393 flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); 4394 4395 if (TCP_SKB_CB(skb)->sacked) 4396 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 4397 &sack_state); 4398 4399 if (tcp_ecn_rcv_ecn_echo(tp, tcp_hdr(skb))) 4400 flag |= FLAG_ECE; 4401 4402 if (sack_state.sack_delivered) 4403 tcp_count_delivered(tp, sack_state.sack_delivered, 4404 flag & FLAG_ECE); 4405 } 4406 4407 /* This is a deviation from RFC3168 since it states that: 4408 * "When the TCP data sender is ready to set the CWR bit after reducing 4409 * the congestion window, it SHOULD set the CWR bit only on the first 4410 * new data packet that it transmits." 4411 * We accept CWR on pure ACKs to be more robust 4412 * with widely-deployed TCP implementations that do this. 4413 */ 4414 tcp_ecn_accept_cwr(sk, skb); 4415 4416 /* We passed data and got it acked, remove any soft error 4417 * log. Something worked... 4418 */ 4419 if (READ_ONCE(sk->sk_err_soft)) 4420 WRITE_ONCE(sk->sk_err_soft, 0); 4421 WRITE_ONCE(icsk->icsk_probes_out, 0); 4422 tp->rcv_tstamp = tcp_jiffies32; 4423 if (!prior_packets) 4424 goto no_queue; 4425 4426 /* See if we can take anything off of the retransmit queue. */ 4427 flag |= tcp_clean_rtx_queue(sk, skb, prior_fack, prior_snd_una, 4428 &sack_state, flag & FLAG_ECE); 4429 4430 tcp_rack_update_reo_wnd(sk, &rs); 4431 4432 if (tcp_ecn_mode_accecn(tp)) 4433 ecn_count = tcp_accecn_process(sk, skb, 4434 tp->delivered - delivered, 4435 sack_state.delivered_bytes, 4436 &flag); 4437 4438 tcp_in_ack_event(sk, flag); 4439 4440 if (unlikely(tp->tlp_high_seq)) 4441 tcp_process_tlp_ack(sk, ack, flag); 4442 4443 if (tcp_ack_is_dubious(sk, flag)) { 4444 if (!(flag & (FLAG_SND_UNA_ADVANCED | 4445 FLAG_NOT_DUP | FLAG_DSACKING_ACK))) { 4446 num_dupack = 1; 4447 /* Consider if pure acks were aggregated in tcp_add_backlog() */ 4448 if (!(flag & FLAG_DATA)) 4449 num_dupack = max_t(u16, 1, skb_shinfo(skb)->gso_segs); 4450 } 4451 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4452 &rexmit); 4453 } 4454 4455 /* If needed, reset TLP/RTO timer when RACK doesn't set. */ 4456 if (flag & FLAG_SET_XMIT_TIMER) 4457 tcp_set_xmit_timer(sk); 4458 4459 if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) 4460 sk_dst_confirm(sk); 4461 4462 delivered = tcp_newly_delivered(sk, delivered, ecn_count, flag); 4463 4464 lost = tp->lost - lost; /* freshly marked lost */ 4465 rs.is_ack_delayed = !!(flag & FLAG_ACK_MAYBE_DELAYED); 4466 tcp_rate_gen(sk, delivered, lost, is_sack_reneg, sack_state.rate); 4467 tcp_cong_control(sk, ack, delivered, flag, sack_state.rate); 4468 tcp_xmit_recovery(sk, rexmit); 4469 return 1; 4470 4471 no_queue: 4472 if (tcp_ecn_mode_accecn(tp)) 4473 ecn_count = tcp_accecn_process(sk, skb, 4474 tp->delivered - delivered, 4475 sack_state.delivered_bytes, 4476 &flag); 4477 tcp_in_ack_event(sk, flag); 4478 /* If data was DSACKed, see if we can undo a cwnd reduction. */ 4479 if (flag & FLAG_DSACKING_ACK) { 4480 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4481 &rexmit); 4482 tcp_newly_delivered(sk, delivered, ecn_count, flag); 4483 } 4484 /* If this ack opens up a zero window, clear backoff. It was 4485 * being used to time the probes, and is probably far higher than 4486 * it needs to be for normal retransmission. 4487 */ 4488 tcp_ack_probe(sk); 4489 4490 if (unlikely(tp->tlp_high_seq)) 4491 tcp_process_tlp_ack(sk, ack, flag); 4492 return 1; 4493 4494 old_ack: 4495 /* An old ACK can carry new data. Update TS.Recent before SACK 4496 * processing can trigger a retransmission. 4497 */ 4498 if (flag & FLAG_UPDATE_TS_RECENT) 4499 tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); 4500 4501 /* If data was SACKed, tag it and see if we should send more data. 4502 * If data was DSACKed, see if we can undo a cwnd reduction. 4503 */ 4504 if (TCP_SKB_CB(skb)->sacked) { 4505 flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, 4506 &sack_state); 4507 tcp_fastretrans_alert(sk, prior_snd_una, num_dupack, &flag, 4508 &rexmit); 4509 tcp_newly_delivered(sk, delivered, ecn_count, flag); 4510 tcp_xmit_recovery(sk, rexmit); 4511 } 4512 4513 return 0; 4514 } 4515 4516 static void tcp_parse_fastopen_option(int len, const unsigned char *cookie, 4517 bool syn, struct tcp_fastopen_cookie *foc, 4518 bool exp_opt) 4519 { 4520 /* Valid only in SYN or SYN-ACK with an even length. */ 4521 if (!foc || !syn || len < 0 || (len & 1)) 4522 return; 4523 4524 if (len >= TCP_FASTOPEN_COOKIE_MIN && 4525 len <= TCP_FASTOPEN_COOKIE_MAX) 4526 memcpy(foc->val, cookie, len); 4527 else if (len != 0) 4528 len = -1; 4529 foc->len = len; 4530 foc->exp = exp_opt; 4531 } 4532 4533 static bool smc_parse_options(const struct tcphdr *th, 4534 struct tcp_options_received *opt_rx, 4535 const unsigned char *ptr, 4536 int opsize) 4537 { 4538 #if IS_ENABLED(CONFIG_SMC) 4539 if (static_branch_unlikely(&tcp_have_smc)) { 4540 if (th->syn && !(opsize & 1) && 4541 opsize >= TCPOLEN_EXP_SMC_BASE && 4542 get_unaligned_be32(ptr) == TCPOPT_SMC_MAGIC) { 4543 opt_rx->smc_ok = 1; 4544 return true; 4545 } 4546 } 4547 #endif 4548 return false; 4549 } 4550 4551 /* Try to parse the MSS option from the TCP header. Return 0 on failure, clamped 4552 * value on success. 4553 */ 4554 u16 tcp_parse_mss_option(const struct tcphdr *th, u16 user_mss) 4555 { 4556 const unsigned char *ptr = (const unsigned char *)(th + 1); 4557 int length = (th->doff * 4) - sizeof(struct tcphdr); 4558 u16 mss = 0; 4559 4560 while (length > 0) { 4561 int opcode = *ptr++; 4562 int opsize; 4563 4564 switch (opcode) { 4565 case TCPOPT_EOL: 4566 return mss; 4567 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 4568 length--; 4569 continue; 4570 default: 4571 if (length < 2) 4572 return mss; 4573 opsize = *ptr++; 4574 if (opsize < 2) /* "silly options" */ 4575 return mss; 4576 if (opsize > length) 4577 return mss; /* fail on partial options */ 4578 if (opcode == TCPOPT_MSS && opsize == TCPOLEN_MSS) { 4579 u16 in_mss = get_unaligned_be16(ptr); 4580 4581 if (in_mss) { 4582 if (user_mss && user_mss < in_mss) 4583 in_mss = user_mss; 4584 mss = in_mss; 4585 } 4586 } 4587 ptr += opsize - 2; 4588 length -= opsize; 4589 } 4590 } 4591 return mss; 4592 } 4593 4594 /* Look for tcp options. Normally only called on SYN and SYNACK packets. 4595 * But, this can also be called on packets in the established flow when 4596 * the fast version below fails. 4597 */ 4598 void tcp_parse_options(const struct net *net, 4599 const struct sk_buff *skb, 4600 struct tcp_options_received *opt_rx, int estab, 4601 struct tcp_fastopen_cookie *foc) 4602 { 4603 const unsigned char *ptr; 4604 const struct tcphdr *th = tcp_hdr(skb); 4605 int length = (th->doff * 4) - sizeof(struct tcphdr); 4606 4607 ptr = (const unsigned char *)(th + 1); 4608 opt_rx->saw_tstamp = 0; 4609 opt_rx->accecn = 0; 4610 opt_rx->saw_unknown = 0; 4611 4612 while (length > 0) { 4613 int opcode = *ptr++; 4614 int opsize; 4615 4616 switch (opcode) { 4617 case TCPOPT_EOL: 4618 return; 4619 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ 4620 length--; 4621 continue; 4622 default: 4623 if (length < 2) 4624 return; 4625 opsize = *ptr++; 4626 if (opsize < 2) /* "silly options" */ 4627 return; 4628 if (opsize > length) 4629 return; /* don't parse partial options */ 4630 switch (opcode) { 4631 case TCPOPT_MSS: 4632 if (opsize == TCPOLEN_MSS && th->syn && !estab) { 4633 u16 in_mss = get_unaligned_be16(ptr); 4634 if (in_mss) { 4635 if (opt_rx->user_mss && 4636 opt_rx->user_mss < in_mss) 4637 in_mss = opt_rx->user_mss; 4638 opt_rx->mss_clamp = in_mss; 4639 } 4640 } 4641 break; 4642 case TCPOPT_WINDOW: 4643 if (opsize == TCPOLEN_WINDOW && th->syn && 4644 !estab && READ_ONCE(net->ipv4.sysctl_tcp_window_scaling)) { 4645 __u8 snd_wscale = *(__u8 *)ptr; 4646 opt_rx->wscale_ok = 1; 4647 if (snd_wscale > TCP_MAX_WSCALE) { 4648 net_info_ratelimited("%s: Illegal window scaling value %d > %u received\n", 4649 __func__, 4650 snd_wscale, 4651 TCP_MAX_WSCALE); 4652 snd_wscale = TCP_MAX_WSCALE; 4653 } 4654 opt_rx->snd_wscale = snd_wscale; 4655 } 4656 break; 4657 case TCPOPT_TIMESTAMP: 4658 if ((opsize == TCPOLEN_TIMESTAMP) && 4659 ((estab && opt_rx->tstamp_ok) || 4660 (!estab && READ_ONCE(net->ipv4.sysctl_tcp_timestamps)))) { 4661 opt_rx->saw_tstamp = 1; 4662 opt_rx->rcv_tsval = get_unaligned_be32(ptr); 4663 opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); 4664 } 4665 break; 4666 case TCPOPT_SACK_PERM: 4667 if (opsize == TCPOLEN_SACK_PERM && th->syn && 4668 !estab && READ_ONCE(net->ipv4.sysctl_tcp_sack)) { 4669 opt_rx->sack_ok = TCP_SACK_SEEN; 4670 tcp_sack_reset(opt_rx); 4671 } 4672 break; 4673 4674 case TCPOPT_SACK: 4675 if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && 4676 !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && 4677 opt_rx->sack_ok) { 4678 TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; 4679 } 4680 break; 4681 #ifdef CONFIG_TCP_MD5SIG 4682 case TCPOPT_MD5SIG: 4683 /* The MD5 Hash has already been 4684 * checked (see tcp_v{4,6}_rcv()). 4685 */ 4686 break; 4687 #endif 4688 #ifdef CONFIG_TCP_AO 4689 case TCPOPT_AO: 4690 /* TCP AO has already been checked 4691 * (see tcp_inbound_ao_hash()). 4692 */ 4693 break; 4694 #endif 4695 case TCPOPT_FASTOPEN: 4696 tcp_parse_fastopen_option( 4697 opsize - TCPOLEN_FASTOPEN_BASE, 4698 ptr, th->syn, foc, false); 4699 break; 4700 4701 case TCPOPT_ACCECN0: 4702 case TCPOPT_ACCECN1: 4703 /* Save offset of AccECN option in TCP header */ 4704 opt_rx->accecn = (ptr - 2) - (__u8 *)th; 4705 break; 4706 4707 case TCPOPT_EXP: 4708 /* Fast Open option shares code 254 using a 4709 * 16 bits magic number. 4710 */ 4711 if (opsize >= TCPOLEN_EXP_FASTOPEN_BASE && 4712 get_unaligned_be16(ptr) == 4713 TCPOPT_FASTOPEN_MAGIC) { 4714 tcp_parse_fastopen_option(opsize - 4715 TCPOLEN_EXP_FASTOPEN_BASE, 4716 ptr + 2, th->syn, foc, true); 4717 break; 4718 } 4719 4720 if (smc_parse_options(th, opt_rx, ptr, opsize)) 4721 break; 4722 4723 opt_rx->saw_unknown = 1; 4724 break; 4725 4726 default: 4727 opt_rx->saw_unknown = 1; 4728 } 4729 ptr += opsize-2; 4730 length -= opsize; 4731 } 4732 } 4733 } 4734 EXPORT_SYMBOL(tcp_parse_options); 4735 4736 static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th) 4737 { 4738 const __be32 *ptr = (const __be32 *)(th + 1); 4739 4740 if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) 4741 | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { 4742 tp->rx_opt.saw_tstamp = 1; 4743 ++ptr; 4744 tp->rx_opt.rcv_tsval = ntohl(*ptr); 4745 ++ptr; 4746 if (*ptr) 4747 tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset; 4748 else 4749 tp->rx_opt.rcv_tsecr = 0; 4750 return true; 4751 } 4752 return false; 4753 } 4754 4755 /* Fast parse options. This hopes to only see timestamps. 4756 * If it is wrong it falls back on tcp_parse_options(). 4757 */ 4758 static bool tcp_fast_parse_options(const struct net *net, 4759 const struct sk_buff *skb, 4760 const struct tcphdr *th, struct tcp_sock *tp) 4761 { 4762 /* In the spirit of fast parsing, compare doff directly to constant 4763 * values. Because equality is used, short doff can be ignored here. 4764 */ 4765 if (th->doff == (sizeof(*th) / 4)) { 4766 tp->rx_opt.saw_tstamp = 0; 4767 tp->rx_opt.accecn = 0; 4768 return false; 4769 } else if (tp->rx_opt.tstamp_ok && 4770 th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { 4771 if (tcp_parse_aligned_timestamp(tp, th)) { 4772 tp->rx_opt.accecn = 0; 4773 return true; 4774 } 4775 } 4776 4777 tcp_parse_options(net, skb, &tp->rx_opt, 1, NULL); 4778 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 4779 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 4780 4781 return true; 4782 } 4783 4784 /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM 4785 * 4786 * It is not fatal. If this ACK does _not_ change critical state (seqs, window) 4787 * it can pass through stack. So, the following predicate verifies that 4788 * this segment is not used for anything but congestion avoidance or 4789 * fast retransmit. Moreover, we even are able to eliminate most of such 4790 * second order effects, if we apply some small "replay" window (~RTO) 4791 * to timestamp space. 4792 * 4793 * All these measures still do not guarantee that we reject wrapped ACKs 4794 * on networks with high bandwidth, when sequence space is recycled fastly, 4795 * but it guarantees that such events will be very rare and do not affect 4796 * connection seriously. This doesn't look nice, but alas, PAWS is really 4797 * buggy extension. 4798 * 4799 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC 4800 * states that events when retransmit arrives after original data are rare. 4801 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is 4802 * the biggest problem on large power networks even with minor reordering. 4803 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe 4804 * up to bandwidth of 18Gigabit/sec. 8) ] 4805 */ 4806 4807 /* Estimates max number of increments of remote peer TSval in 4808 * a replay window (based on our current RTO estimation). 4809 */ 4810 static u32 tcp_tsval_replay(const struct sock *sk) 4811 { 4812 /* If we use usec TS resolution, 4813 * then expect the remote peer to use the same resolution. 4814 */ 4815 if (tcp_sk(sk)->tcp_usec_ts) 4816 return inet_csk(sk)->icsk_rto * (USEC_PER_SEC / HZ); 4817 4818 /* RFC 7323 recommends a TSval clock between 1ms and 1sec. 4819 * We know that some OS (including old linux) can use 1200 Hz. 4820 */ 4821 return inet_csk(sk)->icsk_rto * 1200 / HZ; 4822 } 4823 4824 static enum skb_drop_reason tcp_disordered_ack_check(const struct sock *sk, 4825 const struct sk_buff *skb) 4826 { 4827 const struct tcp_sock *tp = tcp_sk(sk); 4828 const struct tcphdr *th = tcp_hdr(skb); 4829 SKB_DR_INIT(reason, TCP_RFC7323_PAWS); 4830 u32 ack = TCP_SKB_CB(skb)->ack_seq; 4831 u32 seq = TCP_SKB_CB(skb)->seq; 4832 4833 /* 1. Is this not a pure ACK ? */ 4834 if (!th->ack || seq != TCP_SKB_CB(skb)->end_seq) 4835 return reason; 4836 4837 /* 2. Is its sequence not the expected one ? */ 4838 if (seq != tp->rcv_nxt) 4839 return before(seq, tp->rcv_nxt) ? 4840 SKB_DROP_REASON_TCP_RFC7323_PAWS_ACK : 4841 reason; 4842 4843 /* 3. Is this not a duplicate ACK ? */ 4844 if (ack != tp->snd_una) 4845 return reason; 4846 4847 /* 4. Is this updating the window ? */ 4848 if (tcp_may_update_window(tp, ack, seq, ntohs(th->window) << 4849 tp->rx_opt.snd_wscale)) 4850 return reason; 4851 4852 /* 5. Is this not in the replay window ? */ 4853 if ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > 4854 tcp_tsval_replay(sk)) 4855 return reason; 4856 4857 return 0; 4858 } 4859 4860 /* Check segment sequence number for validity. 4861 * 4862 * Segment controls are considered valid, if the segment 4863 * fits to the window after truncation to the window. Acceptability 4864 * of data (and SYN, FIN, of course) is checked separately. 4865 * See tcp_data_queue(), for example. 4866 * 4867 * Also, controls (RST is main one) are accepted using RCV.WUP instead 4868 * of RCV.NXT. Peer still did not advance his SND.UNA when we 4869 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. 4870 * (borrowed from freebsd) 4871 */ 4872 4873 static enum skb_drop_reason tcp_sequence(const struct sock *sk, 4874 u32 seq, u32 end_seq, 4875 const struct tcphdr *th) 4876 { 4877 const struct tcp_sock *tp = tcp_sk(sk); 4878 u32 seq_limit; 4879 4880 if (before(end_seq, tp->rcv_wup)) 4881 return SKB_DROP_REASON_TCP_OLD_SEQUENCE; 4882 4883 seq_limit = tp->rcv_nxt + tcp_max_receive_window(tp); 4884 if (unlikely(after(end_seq, seq_limit))) { 4885 /* Some stacks are known to handle FIN incorrectly; allow the 4886 * FIN to extend beyond the window and check it in detail later. 4887 */ 4888 if (!after(end_seq - th->fin, seq_limit)) 4889 return SKB_NOT_DROPPED_YET; 4890 4891 if (after(seq, seq_limit)) 4892 return SKB_DROP_REASON_TCP_INVALID_SEQUENCE; 4893 4894 /* Only accept this packet if receive queue is empty. */ 4895 if (skb_queue_len(&sk->sk_receive_queue)) 4896 return SKB_DROP_REASON_TCP_INVALID_END_SEQUENCE; 4897 } 4898 4899 return SKB_NOT_DROPPED_YET; 4900 } 4901 4902 4903 void tcp_done_with_error(struct sock *sk, int err) 4904 { 4905 /* This barrier is coupled with smp_rmb() in tcp_poll() */ 4906 WRITE_ONCE(sk->sk_err, err); 4907 smp_wmb(); 4908 4909 tcp_write_queue_purge(sk); 4910 tcp_done(sk); 4911 4912 if (!sock_flag(sk, SOCK_DEAD)) 4913 sk_error_report(sk); 4914 } 4915 4916 /* When we get a reset we do this. */ 4917 void tcp_reset(struct sock *sk, struct sk_buff *skb) 4918 { 4919 int err; 4920 4921 trace_tcp_receive_reset(sk); 4922 4923 /* mptcp can't tell us to ignore reset pkts, 4924 * so just ignore the return value of mptcp_incoming_options(). 4925 */ 4926 if (sk_is_mptcp(sk)) 4927 mptcp_incoming_options(sk, skb); 4928 4929 /* We want the right error as BSD sees it (and indeed as we do). */ 4930 switch (sk->sk_state) { 4931 case TCP_SYN_SENT: 4932 err = ECONNREFUSED; 4933 break; 4934 case TCP_CLOSE_WAIT: 4935 err = EPIPE; 4936 break; 4937 case TCP_CLOSE: 4938 return; 4939 default: 4940 err = ECONNRESET; 4941 } 4942 tcp_done_with_error(sk, err); 4943 } 4944 4945 /* 4946 * Process the FIN bit. This now behaves as it is supposed to work 4947 * and the FIN takes effect when it is validly part of sequence 4948 * space. Not before when we get holes. 4949 * 4950 * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT 4951 * (and thence onto LAST-ACK and finally, CLOSE, we never enter 4952 * TIME-WAIT) 4953 * 4954 * If we are in FINWAIT-1, a received FIN indicates simultaneous 4955 * close and we go into CLOSING (and later onto TIME-WAIT) 4956 * 4957 * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. 4958 */ 4959 void tcp_fin(struct sock *sk) 4960 { 4961 struct tcp_sock *tp = tcp_sk(sk); 4962 4963 inet_csk_schedule_ack(sk); 4964 4965 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN); 4966 sock_set_flag(sk, SOCK_DONE); 4967 4968 switch (sk->sk_state) { 4969 case TCP_SYN_RECV: 4970 case TCP_ESTABLISHED: 4971 /* Move to CLOSE_WAIT */ 4972 tcp_set_state(sk, TCP_CLOSE_WAIT); 4973 inet_csk_enter_pingpong_mode(sk); 4974 break; 4975 4976 case TCP_CLOSE_WAIT: 4977 case TCP_CLOSING: 4978 /* Received a retransmission of the FIN, do 4979 * nothing. 4980 */ 4981 break; 4982 case TCP_LAST_ACK: 4983 /* RFC793: Remain in the LAST-ACK state. */ 4984 break; 4985 4986 case TCP_FIN_WAIT1: 4987 /* This case occurs when a simultaneous close 4988 * happens, we must ack the received FIN and 4989 * enter the CLOSING state. 4990 */ 4991 tcp_send_ack(sk); 4992 tcp_set_state(sk, TCP_CLOSING); 4993 break; 4994 case TCP_FIN_WAIT2: 4995 /* Received a FIN -- send ACK and enter TIME_WAIT. */ 4996 tcp_send_ack(sk); 4997 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 4998 break; 4999 default: 5000 /* Only TCP_LISTEN and TCP_CLOSE are left, in these 5001 * cases we should never reach this piece of code. 5002 */ 5003 pr_err("%s: Impossible, sk->sk_state=%d\n", 5004 __func__, sk->sk_state); 5005 break; 5006 } 5007 5008 /* It _is_ possible, that we have something out-of-order _after_ FIN. 5009 * Probably, we should reset in this case. For now drop them. 5010 */ 5011 skb_rbtree_purge(&tp->out_of_order_queue); 5012 if (tcp_is_sack(tp)) 5013 tcp_sack_reset(&tp->rx_opt); 5014 5015 if (!sock_flag(sk, SOCK_DEAD)) { 5016 sk->sk_state_change(sk); 5017 5018 /* Do not send POLL_HUP for half duplex close. */ 5019 if (sk->sk_shutdown == SHUTDOWN_MASK || 5020 sk->sk_state == TCP_CLOSE) 5021 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); 5022 else 5023 sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); 5024 } 5025 } 5026 5027 static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, 5028 u32 end_seq) 5029 { 5030 if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { 5031 if (before(seq, sp->start_seq)) 5032 sp->start_seq = seq; 5033 if (after(end_seq, sp->end_seq)) 5034 sp->end_seq = end_seq; 5035 return true; 5036 } 5037 return false; 5038 } 5039 5040 static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) 5041 { 5042 struct tcp_sock *tp = tcp_sk(sk); 5043 5044 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { 5045 int mib_idx; 5046 5047 if (before(seq, tp->rcv_nxt)) 5048 mib_idx = LINUX_MIB_TCPDSACKOLDSENT; 5049 else 5050 mib_idx = LINUX_MIB_TCPDSACKOFOSENT; 5051 5052 NET_INC_STATS(sock_net(sk), mib_idx); 5053 5054 tp->rx_opt.dsack = 1; 5055 tp->duplicate_sack[0].start_seq = seq; 5056 tp->duplicate_sack[0].end_seq = end_seq; 5057 } 5058 } 5059 5060 static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) 5061 { 5062 struct tcp_sock *tp = tcp_sk(sk); 5063 5064 if (!tp->rx_opt.dsack) 5065 tcp_dsack_set(sk, seq, end_seq); 5066 else 5067 tcp_sack_extend(tp->duplicate_sack, seq, end_seq); 5068 } 5069 5070 static void tcp_rcv_spurious_retrans(struct sock *sk, 5071 const struct sk_buff *skb) 5072 { 5073 struct tcp_sock *tp = tcp_sk(sk); 5074 5075 /* When the ACK path fails or drops most ACKs, the sender would 5076 * timeout and spuriously retransmit the same segment repeatedly. 5077 * If it seems our ACKs are not reaching the other side, 5078 * based on receiving a duplicate data segment with new flowlabel 5079 * (suggesting the sender suffered an RTO), and we are not already 5080 * repathing due to our own RTO, then rehash the socket to repath our 5081 * packets. 5082 */ 5083 #if IS_ENABLED(CONFIG_IPV6) 5084 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Loss && 5085 skb->protocol == htons(ETH_P_IPV6) && 5086 (tcp_sk(sk)->inet_conn.icsk_ack.lrcv_flowlabel != 5087 ntohl(ip6_flowlabel(ipv6_hdr(skb)))) && 5088 __sk_rethink_txhash_reset_dst(sk)) { 5089 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDUPLICATEDATAREHASH); 5090 } 5091 5092 /* Save last flowlabel after a spurious retrans. */ 5093 tcp_save_lrcv_flowlabel(sk, skb); 5094 #endif 5095 /* Check DSACK info to detect that the previous ACK carrying the 5096 * AccECN option was lost after the second retransmision, and then 5097 * stop sending AccECN option in all subsequent ACKs. 5098 */ 5099 if (tcp_ecn_mode_accecn(tp) && 5100 tp->accecn_opt_sent_w_dsack && 5101 TCP_SKB_CB(skb)->seq == tp->duplicate_sack[0].start_seq) 5102 tcp_accecn_fail_mode_set(tp, TCP_ACCECN_OPT_FAIL_SEND); 5103 } 5104 5105 static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb) 5106 { 5107 struct tcp_sock *tp = tcp_sk(sk); 5108 5109 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 5110 before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5111 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 5112 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 5113 5114 if (tcp_is_sack(tp) && READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_dsack)) { 5115 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 5116 5117 tcp_rcv_spurious_retrans(sk, skb); 5118 if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) 5119 end_seq = tp->rcv_nxt; 5120 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); 5121 } 5122 } 5123 5124 tcp_send_ack(sk); 5125 } 5126 5127 /* These routines update the SACK block as out-of-order packets arrive or 5128 * in-order packets close up the sequence space. 5129 */ 5130 static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) 5131 { 5132 int this_sack; 5133 struct tcp_sack_block *sp = &tp->selective_acks[0]; 5134 struct tcp_sack_block *swalk = sp + 1; 5135 5136 /* See if the recent change to the first SACK eats into 5137 * or hits the sequence space of other SACK blocks, if so coalesce. 5138 */ 5139 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { 5140 if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { 5141 int i; 5142 5143 /* Zap SWALK, by moving every further SACK up by one slot. 5144 * Decrease num_sacks. 5145 */ 5146 tp->rx_opt.num_sacks--; 5147 for (i = this_sack; i < tp->rx_opt.num_sacks; i++) 5148 sp[i] = sp[i + 1]; 5149 continue; 5150 } 5151 this_sack++; 5152 swalk++; 5153 } 5154 } 5155 5156 void tcp_sack_compress_send_ack(struct sock *sk) 5157 { 5158 struct tcp_sock *tp = tcp_sk(sk); 5159 5160 if (!tp->compressed_ack) 5161 return; 5162 5163 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 5164 __sock_put(sk); 5165 5166 /* Since we have to send one ack finally, 5167 * substract one from tp->compressed_ack to keep 5168 * LINUX_MIB_TCPACKCOMPRESSED accurate. 5169 */ 5170 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 5171 tp->compressed_ack - 1); 5172 5173 tp->compressed_ack = 0; 5174 tcp_send_ack(sk); 5175 } 5176 5177 /* Reasonable amount of sack blocks included in TCP SACK option 5178 * The max is 4, but this becomes 3 if TCP timestamps are there. 5179 * Given that SACK packets might be lost, be conservative and use 2. 5180 */ 5181 #define TCP_SACK_BLOCKS_EXPECTED 2 5182 5183 static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) 5184 { 5185 struct tcp_sock *tp = tcp_sk(sk); 5186 struct tcp_sack_block *sp = &tp->selective_acks[0]; 5187 int cur_sacks = tp->rx_opt.num_sacks; 5188 int this_sack; 5189 5190 if (!cur_sacks) 5191 goto new_sack; 5192 5193 for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { 5194 if (tcp_sack_extend(sp, seq, end_seq)) { 5195 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 5196 tcp_sack_compress_send_ack(sk); 5197 /* Rotate this_sack to the first one. */ 5198 for (; this_sack > 0; this_sack--, sp--) 5199 swap(*sp, *(sp - 1)); 5200 if (cur_sacks > 1) 5201 tcp_sack_maybe_coalesce(tp); 5202 return; 5203 } 5204 } 5205 5206 if (this_sack >= TCP_SACK_BLOCKS_EXPECTED) 5207 tcp_sack_compress_send_ack(sk); 5208 5209 /* Could not find an adjacent existing SACK, build a new one, 5210 * put it at the front, and shift everyone else down. We 5211 * always know there is at least one SACK present already here. 5212 * 5213 * If the sack array is full, forget about the last one. 5214 */ 5215 if (this_sack >= TCP_NUM_SACKS) { 5216 this_sack--; 5217 tp->rx_opt.num_sacks--; 5218 sp--; 5219 } 5220 for (; this_sack > 0; this_sack--, sp--) 5221 *sp = *(sp - 1); 5222 5223 new_sack: 5224 /* Build the new head SACK, and we're done. */ 5225 sp->start_seq = seq; 5226 sp->end_seq = end_seq; 5227 tp->rx_opt.num_sacks++; 5228 } 5229 5230 /* RCV.NXT advances, some SACKs should be eaten. */ 5231 5232 static void tcp_sack_remove(struct tcp_sock *tp) 5233 { 5234 struct tcp_sack_block *sp = &tp->selective_acks[0]; 5235 int num_sacks = tp->rx_opt.num_sacks; 5236 int this_sack; 5237 5238 /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ 5239 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5240 tp->rx_opt.num_sacks = 0; 5241 return; 5242 } 5243 5244 for (this_sack = 0; this_sack < num_sacks;) { 5245 /* Check if the start of the sack is covered by RCV.NXT. */ 5246 if (!before(tp->rcv_nxt, sp->start_seq)) { 5247 int i; 5248 5249 /* RCV.NXT must cover all the block! */ 5250 WARN_ON(before(tp->rcv_nxt, sp->end_seq)); 5251 5252 /* Zap this SACK, by moving forward any other SACKS. */ 5253 for (i = this_sack+1; i < num_sacks; i++) 5254 tp->selective_acks[i-1] = tp->selective_acks[i]; 5255 num_sacks--; 5256 continue; 5257 } 5258 this_sack++; 5259 sp++; 5260 } 5261 tp->rx_opt.num_sacks = num_sacks; 5262 } 5263 5264 /** 5265 * tcp_try_coalesce - try to merge skb to prior one 5266 * @sk: socket 5267 * @to: prior buffer 5268 * @from: buffer to add in queue 5269 * @fragstolen: pointer to boolean 5270 * 5271 * Before queueing skb @from after @to, try to merge them 5272 * to reduce overall memory use and queue lengths, if cost is small. 5273 * Packets in ofo or receive queues can stay a long time. 5274 * Better try to coalesce them right now to avoid future collapses. 5275 * Returns true if caller should free @from instead of queueing it 5276 */ 5277 static bool tcp_try_coalesce(struct sock *sk, 5278 struct sk_buff *to, 5279 struct sk_buff *from, 5280 bool *fragstolen) 5281 { 5282 int delta; 5283 5284 *fragstolen = false; 5285 5286 /* Its possible this segment overlaps with prior segment in queue */ 5287 if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq) 5288 return false; 5289 5290 if (!tcp_skb_can_collapse_rx(to, from)) 5291 return false; 5292 5293 /* Nothing follows a FIN. */ 5294 if (unlikely(TCP_SKB_CB(to)->tcp_flags & TCPHDR_FIN)) 5295 return false; 5296 5297 if (!skb_try_coalesce(to, from, fragstolen, &delta)) 5298 return false; 5299 5300 atomic_add(delta, &sk->sk_rmem_alloc); 5301 sk_mem_charge(sk, delta); 5302 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE); 5303 TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq; 5304 TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq; 5305 TCP_SKB_CB(to)->tcp_flags |= TCP_SKB_CB(from)->tcp_flags; 5306 5307 if (TCP_SKB_CB(from)->has_rxtstamp) { 5308 TCP_SKB_CB(to)->has_rxtstamp = true; 5309 to->tstamp = from->tstamp; 5310 skb_hwtstamps(to)->hwtstamp = skb_hwtstamps(from)->hwtstamp; 5311 } 5312 5313 return true; 5314 } 5315 5316 static bool tcp_ooo_try_coalesce(struct sock *sk, 5317 struct sk_buff *to, 5318 struct sk_buff *from, 5319 bool *fragstolen) 5320 { 5321 bool res = tcp_try_coalesce(sk, to, from, fragstolen); 5322 5323 /* In case tcp_drop_reason() is called later, update to->gso_segs */ 5324 if (res) { 5325 u32 gso_segs = max_t(u16, 1, skb_shinfo(to)->gso_segs) + 5326 max_t(u16, 1, skb_shinfo(from)->gso_segs); 5327 5328 skb_shinfo(to)->gso_segs = min_t(u32, gso_segs, 0xFFFF); 5329 } 5330 return res; 5331 } 5332 5333 noinline_for_tracing static void 5334 tcp_drop_reason(struct sock *sk, struct sk_buff *skb, enum skb_drop_reason reason) 5335 { 5336 sk_drops_skbadd(sk, skb); 5337 sk_skb_reason_drop(sk, skb, reason); 5338 } 5339 5340 /* This one checks to see if we can put data from the 5341 * out_of_order queue into the receive_queue. 5342 */ 5343 static void tcp_ofo_queue(struct sock *sk) 5344 { 5345 struct tcp_sock *tp = tcp_sk(sk); 5346 __u32 dsack_high = tp->rcv_nxt; 5347 bool fin, fragstolen, eaten; 5348 struct sk_buff *skb, *tail; 5349 struct rb_node *p; 5350 5351 p = rb_first(&tp->out_of_order_queue); 5352 while (p) { 5353 skb = rb_to_skb(p); 5354 if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) 5355 break; 5356 5357 if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { 5358 __u32 dsack = dsack_high; 5359 5360 if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) 5361 dsack = TCP_SKB_CB(skb)->end_seq; 5362 tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); 5363 } 5364 p = rb_next(p); 5365 rb_erase(&skb->rbnode, &tp->out_of_order_queue); 5366 5367 if (unlikely(!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))) { 5368 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_DROP); 5369 continue; 5370 } 5371 5372 bpf_tcp_ops_call_flag(enqueue_rcvq, RCVQ, sk, skb); 5373 5374 tail = skb_peek_tail(&sk->sk_receive_queue); 5375 eaten = tail && tcp_try_coalesce(sk, tail, skb, &fragstolen); 5376 tcp_rcv_nxt_update(tp, TCP_SKB_CB(skb)->end_seq); 5377 fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN; 5378 if (!eaten) 5379 tcp_add_receive_queue(sk, skb); 5380 else 5381 kfree_skb_partial(skb, fragstolen); 5382 5383 if (unlikely(fin)) { 5384 tcp_fin(sk); 5385 /* tcp_fin() purges tp->out_of_order_queue, 5386 * so we must end this loop right now. 5387 */ 5388 break; 5389 } 5390 } 5391 } 5392 5393 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb); 5394 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb); 5395 5396 static bool tcp_can_ingest(const struct sock *sk, const struct sk_buff *skb) 5397 { 5398 unsigned int rmem = atomic_read(&sk->sk_rmem_alloc); 5399 5400 return rmem <= sk->sk_rcvbuf; 5401 } 5402 5403 static int tcp_try_rmem_schedule(struct sock *sk, const struct sk_buff *skb, 5404 unsigned int size) 5405 { 5406 if (!tcp_can_ingest(sk, skb) || 5407 !sk_rmem_schedule(sk, skb, size)) { 5408 5409 if (tcp_prune_queue(sk, skb) < 0) 5410 return -1; 5411 5412 while (!sk_rmem_schedule(sk, skb, size)) { 5413 if (!tcp_prune_ofo_queue(sk, skb)) 5414 return -1; 5415 } 5416 } 5417 return 0; 5418 } 5419 5420 /* Credit the memory added to the receive queues since @rmem was sampled 5421 * toward the next allowed collapse. Saturate at sk_rcvbuf, not at the goal 5422 * checked in tcp_prune_queue(): sk_rcvbuf (and the goal) can grow later, 5423 * for instance in tcp_clamp_window(). 5424 */ 5425 static void tcp_collapse_credit_add(struct sock *sk, int rmem) 5426 { 5427 struct tcp_sock *tp = tcp_sk(sk); 5428 int delta = atomic_read(&sk->sk_rmem_alloc) - rmem; 5429 5430 if (delta > 0 && tp->rcv_collapse_credit < (u32)sk->sk_rcvbuf) 5431 tp->rcv_collapse_credit += delta; 5432 } 5433 5434 static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) 5435 { 5436 struct tcp_sock *tp = tcp_sk(sk); 5437 struct rb_node **p, *parent; 5438 struct sk_buff *skb1; 5439 u32 seq, end_seq; 5440 bool fragstolen; 5441 int rmem; 5442 5443 tcp_save_lrcv_flowlabel(sk, skb); 5444 tcp_data_ecn_check(sk, skb); 5445 5446 if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) { 5447 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFODROP); 5448 READ_ONCE(sk->sk_data_ready)(sk); 5449 tcp_drop_reason(sk, skb, SKB_DROP_REASON_PROTO_MEM); 5450 return; 5451 } 5452 /* Sampled after tcp_try_rmem_schedule(), which can prune. */ 5453 rmem = atomic_read(&sk->sk_rmem_alloc); 5454 5455 tcp_measure_rcv_mss(sk, skb); 5456 /* Disable header prediction. */ 5457 tp->pred_flags = 0; 5458 inet_csk_schedule_ack(sk); 5459 5460 tp->rcv_ooopack += max_t(u16, 1, skb_shinfo(skb)->gso_segs); 5461 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOQUEUE); 5462 seq = TCP_SKB_CB(skb)->seq; 5463 end_seq = TCP_SKB_CB(skb)->end_seq; 5464 5465 p = &tp->out_of_order_queue.rb_node; 5466 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5467 /* Initial out of order segment, build 1 SACK. */ 5468 if (tcp_is_sack(tp)) { 5469 tp->rx_opt.num_sacks = 1; 5470 tp->selective_acks[0].start_seq = seq; 5471 tp->selective_acks[0].end_seq = end_seq; 5472 } 5473 rb_link_node(&skb->rbnode, NULL, p); 5474 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 5475 tp->ooo_last_skb = skb; 5476 goto end; 5477 } 5478 5479 /* In the typical case, we are adding an skb to the end of the list. 5480 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup. 5481 */ 5482 if (tcp_ooo_try_coalesce(sk, tp->ooo_last_skb, 5483 skb, &fragstolen)) { 5484 coalesce_done: 5485 /* For non sack flows, do not grow window to force DUPACK 5486 * and trigger fast retransmit. 5487 */ 5488 if (tcp_is_sack(tp)) 5489 tcp_grow_window(sk, skb, true); 5490 kfree_skb_partial(skb, fragstolen); 5491 skb = NULL; 5492 goto add_sack; 5493 } 5494 /* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */ 5495 if (!before(seq, TCP_SKB_CB(tp->ooo_last_skb)->end_seq)) { 5496 parent = &tp->ooo_last_skb->rbnode; 5497 p = &parent->rb_right; 5498 goto insert; 5499 } 5500 5501 /* Find place to insert this segment. Handle overlaps on the way. */ 5502 parent = NULL; 5503 while (*p) { 5504 parent = *p; 5505 skb1 = rb_to_skb(parent); 5506 if (before(seq, TCP_SKB_CB(skb1)->seq)) { 5507 p = &parent->rb_left; 5508 continue; 5509 } 5510 if (before(seq, TCP_SKB_CB(skb1)->end_seq)) { 5511 if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 5512 /* All the bits are present. Drop. */ 5513 NET_INC_STATS(sock_net(sk), 5514 LINUX_MIB_TCPOFOMERGE); 5515 tcp_drop_reason(sk, skb, 5516 SKB_DROP_REASON_TCP_OFOMERGE); 5517 skb = NULL; 5518 tcp_dsack_set(sk, seq, end_seq); 5519 goto add_sack; 5520 } 5521 if (after(seq, TCP_SKB_CB(skb1)->seq)) { 5522 /* Partial overlap. */ 5523 tcp_dsack_set(sk, seq, TCP_SKB_CB(skb1)->end_seq); 5524 } else { 5525 /* skb's seq == skb1's seq and skb covers skb1. 5526 * Replace skb1 with skb. 5527 */ 5528 rb_replace_node(&skb1->rbnode, &skb->rbnode, 5529 &tp->out_of_order_queue); 5530 tcp_dsack_extend(sk, 5531 TCP_SKB_CB(skb1)->seq, 5532 TCP_SKB_CB(skb1)->end_seq); 5533 NET_INC_STATS(sock_net(sk), 5534 LINUX_MIB_TCPOFOMERGE); 5535 tcp_drop_reason(sk, skb1, 5536 SKB_DROP_REASON_TCP_OFOMERGE); 5537 goto merge_right; 5538 } 5539 } else if (tcp_ooo_try_coalesce(sk, skb1, 5540 skb, &fragstolen)) { 5541 goto coalesce_done; 5542 } 5543 p = &parent->rb_right; 5544 } 5545 insert: 5546 /* Insert segment into RB tree. */ 5547 rb_link_node(&skb->rbnode, parent, p); 5548 rb_insert_color(&skb->rbnode, &tp->out_of_order_queue); 5549 5550 merge_right: 5551 /* Remove other segments covered by skb. */ 5552 while ((skb1 = skb_rb_next(skb)) != NULL) { 5553 if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) 5554 break; 5555 if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { 5556 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 5557 end_seq); 5558 break; 5559 } 5560 rb_erase(&skb1->rbnode, &tp->out_of_order_queue); 5561 tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, 5562 TCP_SKB_CB(skb1)->end_seq); 5563 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPOFOMERGE); 5564 tcp_drop_reason(sk, skb1, SKB_DROP_REASON_TCP_OFOMERGE); 5565 } 5566 /* If there is no skb after us, we are the last_skb ! */ 5567 if (!skb1) 5568 tp->ooo_last_skb = skb; 5569 5570 add_sack: 5571 if (tcp_is_sack(tp)) 5572 tcp_sack_new_ofo_skb(sk, seq, end_seq); 5573 end: 5574 if (skb) { 5575 /* For non sack flows, do not grow window to force DUPACK 5576 * and trigger fast retransmit. 5577 */ 5578 if (tcp_is_sack(tp)) 5579 tcp_grow_window(sk, skb, false); 5580 skb_condense(skb); 5581 skb_set_owner_r(skb, sk); 5582 } 5583 /* Net growth: coalescing can charge nothing, and covered skbs 5584 * were freed. 5585 */ 5586 tcp_collapse_credit_add(sk, rmem); 5587 /* do not grow rcvbuf for not-yet-accepted or orphaned sockets. */ 5588 if (sk->sk_socket) 5589 tcp_rcvbuf_grow(sk, tp->rcvq_space.space); 5590 } 5591 5592 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, 5593 bool *fragstolen) 5594 { 5595 int rmem = atomic_read(&sk->sk_rmem_alloc); 5596 struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue); 5597 int eaten; 5598 5599 bpf_tcp_ops_call_flag(enqueue_rcvq, RCVQ, sk, skb); 5600 5601 eaten = (tail && 5602 tcp_try_coalesce(sk, tail, 5603 skb, fragstolen)) ? 1 : 0; 5604 tcp_rcv_nxt_update(tcp_sk(sk), TCP_SKB_CB(skb)->end_seq); 5605 if (!eaten) { 5606 tcp_add_receive_queue(sk, skb); 5607 skb_set_owner_r(skb, sk); 5608 } 5609 tcp_collapse_credit_add(sk, rmem); 5610 return eaten; 5611 } 5612 5613 int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size) 5614 { 5615 struct sk_buff *skb; 5616 int err = -ENOMEM; 5617 int data_len = 0; 5618 bool fragstolen; 5619 5620 if (size == 0) 5621 return 0; 5622 5623 if (size > PAGE_SIZE) { 5624 int npages = min_t(size_t, size >> PAGE_SHIFT, MAX_SKB_FRAGS); 5625 5626 data_len = npages << PAGE_SHIFT; 5627 size = data_len + (size & ~PAGE_MASK); 5628 } 5629 skb = alloc_skb_with_frags(size - data_len, data_len, 5630 PAGE_ALLOC_COSTLY_ORDER, 5631 &err, sk->sk_allocation); 5632 if (!skb) 5633 goto err; 5634 5635 skb_put(skb, size - data_len); 5636 skb->data_len = data_len; 5637 skb->len = size; 5638 5639 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 5640 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 5641 goto err_free; 5642 } 5643 5644 err = skb_copy_datagram_from_iter(skb, 0, &msg->msg_iter, size); 5645 if (err) 5646 goto err_free; 5647 5648 TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt; 5649 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size; 5650 TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1; 5651 5652 if (tcp_queue_rcv(sk, skb, &fragstolen)) { 5653 WARN_ON_ONCE(fragstolen); /* should not happen */ 5654 __kfree_skb(skb); 5655 } 5656 return size; 5657 5658 err_free: 5659 kfree_skb(skb); 5660 err: 5661 return err; 5662 5663 } 5664 5665 void tcp_data_ready(struct sock *sk) 5666 { 5667 if (tcp_epollin_ready(sk, sk->sk_rcvlowat) || sock_flag(sk, SOCK_DONE)) 5668 READ_ONCE(sk->sk_data_ready)(sk); 5669 } 5670 5671 static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) 5672 { 5673 struct tcp_sock *tp = tcp_sk(sk); 5674 enum skb_drop_reason reason; 5675 bool fragstolen; 5676 int eaten; 5677 5678 /* If a subflow has been reset, the packet should not continue 5679 * to be processed, drop the packet. 5680 */ 5681 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) { 5682 __kfree_skb(skb); 5683 return; 5684 } 5685 5686 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { 5687 __kfree_skb(skb); 5688 return; 5689 } 5690 tcp_cleanup_skb(skb); 5691 __skb_pull(skb, tcp_hdr(skb)->doff * 4); 5692 5693 reason = SKB_DROP_REASON_NOT_SPECIFIED; 5694 tp->rx_opt.dsack = 0; 5695 5696 /* Queue data for delivery to the user. 5697 * Packets in sequence go to the receive queue. 5698 * Out of sequence packets to the out_of_order_queue. 5699 */ 5700 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { 5701 if (tcp_receive_window(tp) == 0) { 5702 /* Some stacks are known to send bare FIN packets 5703 * in a loop even if we send RWIN 0 in our ACK. 5704 * Accepting this FIN does not hurt memory pressure 5705 * because the FIN flag will simply be merged to the 5706 * receive queue tail skb in most cases. 5707 */ 5708 if (!skb->len && 5709 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 5710 goto queue_and_out; 5711 5712 reason = SKB_DROP_REASON_TCP_ZEROWINDOW; 5713 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5714 goto out_of_window; 5715 } 5716 5717 /* Ok. In sequence. In window. */ 5718 queue_and_out: 5719 if (tcp_try_rmem_schedule(sk, skb, skb->truesize)) { 5720 /* TODO: maybe ratelimit these WIN 0 ACK ? */ 5721 inet_csk(sk)->icsk_ack.pending |= 5722 (ICSK_ACK_NOMEM | ICSK_ACK_NOW); 5723 inet_csk_schedule_ack(sk); 5724 READ_ONCE(sk->sk_data_ready)(sk); 5725 5726 if (skb_queue_len(&sk->sk_receive_queue) && skb->len) { 5727 reason = SKB_DROP_REASON_PROTO_MEM; 5728 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP); 5729 goto drop; 5730 } 5731 sk_forced_mem_schedule(sk, skb->truesize); 5732 } 5733 5734 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 5735 if (skb->len) 5736 tcp_event_data_recv(sk, skb); 5737 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 5738 tcp_fin(sk); 5739 5740 if (!RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 5741 tcp_ofo_queue(sk); 5742 5743 /* RFC5681. 4.2. SHOULD send immediate ACK, when 5744 * gap in queue is filled. 5745 */ 5746 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 5747 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_NOW; 5748 } 5749 5750 if (tp->rx_opt.num_sacks) 5751 tcp_sack_remove(tp); 5752 5753 tcp_fast_path_check(sk); 5754 5755 if (eaten > 0) 5756 kfree_skb_partial(skb, fragstolen); 5757 if (!sock_flag(sk, SOCK_DEAD)) 5758 tcp_data_ready(sk); 5759 return; 5760 } 5761 5762 if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { 5763 tcp_rcv_spurious_retrans(sk, skb); 5764 /* A retransmit, 2nd most common case. Force an immediate ack. */ 5765 reason = SKB_DROP_REASON_TCP_OLD_DATA; 5766 NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); 5767 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); 5768 5769 out_of_window: 5770 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 5771 inet_csk_schedule_ack(sk); 5772 drop: 5773 tcp_drop_reason(sk, skb, reason); 5774 return; 5775 } 5776 5777 /* Out of window. F.e. zero window probe. */ 5778 if (!before(TCP_SKB_CB(skb)->seq, 5779 tp->rcv_nxt + tcp_receive_window(tp))) { 5780 reason = SKB_DROP_REASON_TCP_OVERWINDOW; 5781 NET_INC_STATS(sock_net(sk), LINUX_MIB_BEYOND_WINDOW); 5782 goto out_of_window; 5783 } 5784 5785 if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 5786 /* Partial packet, seq < rcv_next < end_seq */ 5787 tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); 5788 5789 /* If window is closed, drop tail of packet. But after 5790 * remembering D-SACK for its head made in previous line. 5791 */ 5792 if (!tcp_receive_window(tp)) { 5793 reason = SKB_DROP_REASON_TCP_ZEROWINDOW; 5794 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPZEROWINDOWDROP); 5795 goto out_of_window; 5796 } 5797 goto queue_and_out; 5798 } 5799 5800 tcp_data_queue_ofo(sk, skb); 5801 } 5802 5803 static struct sk_buff *tcp_skb_next(struct sk_buff *skb, struct sk_buff_head *list) 5804 { 5805 if (list) 5806 return !skb_queue_is_last(list, skb) ? skb->next : NULL; 5807 5808 return skb_rb_next(skb); 5809 } 5810 5811 static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, 5812 struct sk_buff_head *list, 5813 struct rb_root *root) 5814 { 5815 struct sk_buff *next = tcp_skb_next(skb, list); 5816 5817 if (list) 5818 __skb_unlink(skb, list); 5819 else 5820 rb_erase(&skb->rbnode, root); 5821 5822 __kfree_skb(skb); 5823 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); 5824 5825 return next; 5826 } 5827 5828 /* Collapse contiguous sequence of skbs head..tail with 5829 * sequence numbers start..end. 5830 * 5831 * If tail is NULL, this means until the end of the queue. 5832 * 5833 * Segments with FIN/SYN are not collapsed (only because this 5834 * simplifies code) 5835 */ 5836 static void 5837 tcp_collapse(struct sock *sk, struct sk_buff_head *list, struct rb_root *root, 5838 struct sk_buff *head, struct sk_buff *tail, u32 start, u32 end) 5839 { 5840 struct sk_buff *skb = head, *n; 5841 struct sk_buff_head tmp; 5842 bool end_of_skbs; 5843 5844 /* First, check that queue is collapsible and find 5845 * the point where collapsing can be useful. 5846 */ 5847 restart: 5848 for (end_of_skbs = true; skb != NULL && skb != tail; skb = n) { 5849 n = tcp_skb_next(skb, list); 5850 5851 if (!skb_frags_readable(skb)) 5852 goto skip_this; 5853 5854 /* No new bits? It is possible on ofo queue. */ 5855 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5856 skb = tcp_collapse_one(sk, skb, list, root); 5857 if (!skb) 5858 break; 5859 goto restart; 5860 } 5861 5862 /* The first skb to collapse is: 5863 * - not SYN/FIN and 5864 * - bloated or contains data before "start" or 5865 * overlaps to the next one and mptcp allow collapsing. 5866 */ 5867 if (!(TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) && 5868 (tcp_win_from_space(sk, skb->truesize) > skb->len || 5869 before(TCP_SKB_CB(skb)->seq, start))) { 5870 end_of_skbs = false; 5871 break; 5872 } 5873 5874 if (n && n != tail && skb_frags_readable(n) && 5875 tcp_skb_can_collapse_rx(skb, n) && 5876 TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(n)->seq) { 5877 end_of_skbs = false; 5878 break; 5879 } 5880 5881 skip_this: 5882 /* Decided to skip this, advance start seq. */ 5883 start = TCP_SKB_CB(skb)->end_seq; 5884 } 5885 if (end_of_skbs || 5886 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) || 5887 !skb_frags_readable(skb)) 5888 return; 5889 5890 __skb_queue_head_init(&tmp); 5891 5892 while (before(start, end)) { 5893 int copy = min_t(int, SKB_MAX_ORDER(0, 0), end - start); 5894 struct sk_buff *nskb; 5895 5896 nskb = alloc_skb(copy, GFP_ATOMIC); 5897 if (!nskb) 5898 break; 5899 5900 memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); 5901 TCP_SKB_CB(nskb)->has_rxtstamp = false; 5902 skb_copy_decrypted(nskb, skb); 5903 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; 5904 if (list) 5905 __skb_queue_before(list, skb, nskb); 5906 else 5907 __skb_queue_tail(&tmp, nskb); /* defer rbtree insertion */ 5908 /* nskb is charged before the skbs it replaces are freed. 5909 * Make sure sk->sk_forward_alloc can not become negative. 5910 */ 5911 sk_forced_mem_schedule(sk, nskb->truesize); 5912 skb_set_owner_r(nskb, sk); 5913 mptcp_skb_ext_move(nskb, skb); 5914 5915 /* Copy data, releasing collapsed skbs. */ 5916 while (copy > 0) { 5917 int offset = start - TCP_SKB_CB(skb)->seq; 5918 int size = TCP_SKB_CB(skb)->end_seq - start; 5919 5920 BUG_ON(offset < 0); 5921 if (size > 0) { 5922 size = min(copy, size); 5923 if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) 5924 BUG(); 5925 TCP_SKB_CB(nskb)->end_seq += size; 5926 if (TCP_SKB_CB(skb)->has_rxtstamp) { 5927 TCP_SKB_CB(nskb)->has_rxtstamp = true; 5928 nskb->tstamp = skb->tstamp; 5929 skb_hwtstamps(nskb)->hwtstamp = 5930 skb_hwtstamps(skb)->hwtstamp; 5931 } 5932 copy -= size; 5933 start += size; 5934 } 5935 if (!before(start, TCP_SKB_CB(skb)->end_seq)) { 5936 skb = tcp_collapse_one(sk, skb, list, root); 5937 if (!skb || 5938 skb == tail || 5939 !tcp_skb_can_collapse_rx(nskb, skb) || 5940 (TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_SYN | TCPHDR_FIN)) || 5941 !skb_frags_readable(skb)) 5942 goto end; 5943 } 5944 } 5945 } 5946 end: 5947 skb_queue_walk_safe(&tmp, skb, n) 5948 tcp_rbtree_insert(root, skb); 5949 } 5950 5951 /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs 5952 * and tcp_collapse() them until all the queue is collapsed. 5953 */ 5954 static void tcp_collapse_ofo_queue(struct sock *sk) 5955 { 5956 struct tcp_sock *tp = tcp_sk(sk); 5957 u32 range_truesize, sum_tiny = 0; 5958 struct sk_buff *skb, *head; 5959 u32 start, end; 5960 5961 skb = skb_rb_first(&tp->out_of_order_queue); 5962 new_range: 5963 if (!skb) { 5964 tp->ooo_last_skb = skb_rb_last(&tp->out_of_order_queue); 5965 return; 5966 } 5967 start = TCP_SKB_CB(skb)->seq; 5968 end = TCP_SKB_CB(skb)->end_seq; 5969 range_truesize = skb->truesize; 5970 5971 for (head = skb;;) { 5972 skb = skb_rb_next(skb); 5973 5974 /* Range is terminated when we see a gap or when 5975 * we are at the queue end. 5976 */ 5977 if (!skb || 5978 after(TCP_SKB_CB(skb)->seq, end) || 5979 before(TCP_SKB_CB(skb)->end_seq, start)) { 5980 /* Do not attempt collapsing tiny skbs */ 5981 if (range_truesize != head->truesize || 5982 end - start >= SKB_WITH_OVERHEAD(PAGE_SIZE)) { 5983 tcp_collapse(sk, NULL, &tp->out_of_order_queue, 5984 head, skb, start, end); 5985 } else { 5986 sum_tiny += range_truesize; 5987 if (sum_tiny > sk->sk_rcvbuf >> 3) 5988 return; 5989 } 5990 goto new_range; 5991 } 5992 5993 range_truesize += skb->truesize; 5994 if (unlikely(before(TCP_SKB_CB(skb)->seq, start))) 5995 start = TCP_SKB_CB(skb)->seq; 5996 if (after(TCP_SKB_CB(skb)->end_seq, end)) 5997 end = TCP_SKB_CB(skb)->end_seq; 5998 } 5999 } 6000 6001 /* 6002 * Clean the out-of-order queue to make room. 6003 * We drop high sequences packets to : 6004 * 1) Let a chance for holes to be filled. 6005 * This means we do not drop packets from ooo queue if their sequence 6006 * is before incoming packet sequence. 6007 * 2) not add too big latencies if thousands of packets sit there. 6008 * (But if application shrinks SO_RCVBUF, we could still end up 6009 * freeing whole queue here) 6010 * 3) Drop at least 12.5 % of sk_rcvbuf to avoid malicious attacks. 6011 * 6012 * Return true if queue has shrunk. 6013 */ 6014 static bool tcp_prune_ofo_queue(struct sock *sk, const struct sk_buff *in_skb) 6015 { 6016 struct tcp_sock *tp = tcp_sk(sk); 6017 struct rb_node *node, *prev; 6018 bool pruned = false; 6019 int goal; 6020 6021 if (RB_EMPTY_ROOT(&tp->out_of_order_queue)) 6022 return false; 6023 6024 goal = sk->sk_rcvbuf >> 3; 6025 node = &tp->ooo_last_skb->rbnode; 6026 6027 do { 6028 struct sk_buff *skb = rb_to_skb(node); 6029 6030 /* If incoming skb would land last in ofo queue, stop pruning. */ 6031 if (after(TCP_SKB_CB(in_skb)->seq, TCP_SKB_CB(skb)->seq)) 6032 break; 6033 pruned = true; 6034 prev = rb_prev(node); 6035 rb_erase(node, &tp->out_of_order_queue); 6036 goal -= skb->truesize; 6037 tcp_drop_reason(sk, skb, SKB_DROP_REASON_TCP_OFO_QUEUE_PRUNE); 6038 tp->ooo_last_skb = rb_to_skb(prev); 6039 if (!prev || goal <= 0) { 6040 if (tcp_can_ingest(sk, in_skb) && 6041 !tcp_under_memory_pressure(sk)) 6042 break; 6043 goal = sk->sk_rcvbuf >> 3; 6044 } 6045 node = prev; 6046 } while (node); 6047 6048 if (pruned) { 6049 NET_INC_STATS(sock_net(sk), LINUX_MIB_OFOPRUNED); 6050 /* Reset SACK state. A conforming SACK implementation will 6051 * do the same at a timeout based retransmit. When a connection 6052 * is in a sad state like this, we care only about integrity 6053 * of the connection not performance. 6054 */ 6055 if (tp->rx_opt.sack_ok) 6056 tcp_sack_reset(&tp->rx_opt); 6057 } 6058 return pruned; 6059 } 6060 6061 /* Reduce allocated memory if we can, trying to get 6062 * the socket within its memory limits again. 6063 * 6064 * Return less than zero if we should start dropping frames 6065 * until the socket owning process reads some of the data 6066 * to stabilize the situation. 6067 */ 6068 static int tcp_prune_queue(struct sock *sk, const struct sk_buff *in_skb) 6069 { 6070 struct tcp_sock *tp = tcp_sk(sk); 6071 6072 /* Do nothing if our queues are empty. */ 6073 if (!atomic_read(&sk->sk_rmem_alloc)) 6074 return -1; 6075 6076 NET_INC_STATS(sock_net(sk), LINUX_MIB_PRUNECALLED); 6077 6078 if (!tcp_can_ingest(sk, in_skb)) 6079 tcp_clamp_window(sk); 6080 else if (tcp_under_memory_pressure(sk)) 6081 tcp_adjust_rcv_ssthresh(sk); 6082 6083 if (tcp_can_ingest(sk, in_skb)) 6084 return 0; 6085 6086 /* Collapsing walks both queues and is expensive, and can free 6087 * almost nothing. Require at least 1/8 of sk_rcvbuf of new memory 6088 * in the receive queues since the previous collapse, otherwise 6089 * a stream of segments could trigger one collapse per packet. 6090 */ 6091 if (tp->rcv_collapse_credit >= (u32)sk->sk_rcvbuf >> 3) { 6092 tp->rcv_collapse_credit = 0; 6093 6094 tcp_collapse_ofo_queue(sk); 6095 if (!skb_queue_empty(&sk->sk_receive_queue)) 6096 tcp_collapse(sk, &sk->sk_receive_queue, NULL, 6097 skb_peek(&sk->sk_receive_queue), 6098 NULL, 6099 tp->copied_seq, tp->rcv_nxt); 6100 6101 if (tcp_can_ingest(sk, in_skb)) 6102 return 0; 6103 } 6104 6105 /* Collapsing did not help (or was not allowed), 6106 * destructive actions follow. 6107 */ 6108 6109 tcp_prune_ofo_queue(sk, in_skb); 6110 6111 if (tcp_can_ingest(sk, in_skb)) 6112 return 0; 6113 6114 /* If we are really being abused, tell the caller to silently 6115 * drop receive data on the floor. It will get retransmitted 6116 * and hopefully then we'll have sufficient space. 6117 */ 6118 NET_INC_STATS(sock_net(sk), LINUX_MIB_RCVPRUNED); 6119 6120 /* Massive buffer overcommit. */ 6121 tp->pred_flags = 0; 6122 return -1; 6123 } 6124 6125 static bool tcp_should_expand_sndbuf(struct sock *sk) 6126 { 6127 const struct tcp_sock *tp = tcp_sk(sk); 6128 6129 /* If the user specified a specific send buffer setting, do 6130 * not modify it. 6131 */ 6132 if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) 6133 return false; 6134 6135 /* If we are under global TCP memory pressure, do not expand. */ 6136 if (tcp_under_memory_pressure(sk)) { 6137 int unused_mem = sk_unused_reserved_mem(sk); 6138 6139 /* Adjust sndbuf according to reserved mem. But make sure 6140 * it never goes below SOCK_MIN_SNDBUF. 6141 * See sk_stream_moderate_sndbuf() for more details. 6142 */ 6143 if (unused_mem > SOCK_MIN_SNDBUF) 6144 WRITE_ONCE(sk->sk_sndbuf, unused_mem); 6145 6146 return false; 6147 } 6148 6149 /* If we are under soft global TCP memory pressure, do not expand. */ 6150 if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0)) 6151 return false; 6152 6153 /* If we filled the congestion window, do not expand. */ 6154 if (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)) 6155 return false; 6156 6157 return true; 6158 } 6159 6160 static void tcp_new_space(struct sock *sk) 6161 { 6162 struct tcp_sock *tp = tcp_sk(sk); 6163 6164 if (tcp_should_expand_sndbuf(sk)) { 6165 tcp_sndbuf_expand(sk); 6166 tp->snd_cwnd_stamp = tcp_jiffies32; 6167 } 6168 6169 INDIRECT_CALL_1(READ_ONCE(sk->sk_write_space), 6170 sk_stream_write_space, 6171 sk); 6172 } 6173 6174 /* Caller made space either from: 6175 * 1) Freeing skbs in rtx queues (after tp->snd_una has advanced) 6176 * 2) Sent skbs from output queue (and thus advancing tp->snd_nxt) 6177 * 6178 * We might be able to generate EPOLLOUT to the application if: 6179 * 1) Space consumed in output/rtx queues is below sk->sk_sndbuf/2 6180 * 2) notsent amount (tp->write_seq - tp->snd_nxt) became 6181 * small enough that tcp_stream_memory_free() decides it 6182 * is time to generate EPOLLOUT. 6183 */ 6184 void __tcp_check_space(struct sock *sk) 6185 { 6186 struct socket *sock = sk->sk_socket; 6187 6188 /* tp->tcp_nospace is only a hint, SOCK_NOSPACE is authoritative. */ 6189 if (!sock || !test_bit(SOCK_NOSPACE, &sock->flags)) 6190 return; 6191 6192 tcp_new_space(sk); 6193 if (!test_bit(SOCK_NOSPACE, &sock->flags)) 6194 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED); 6195 } 6196 6197 static inline void tcp_data_snd_check(struct sock *sk) 6198 { 6199 tcp_push_pending_frames(sk); 6200 tcp_check_space(sk); 6201 } 6202 6203 /* 6204 * Check if sending an ack is needed. 6205 */ 6206 static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) 6207 { 6208 struct tcp_sock *tp = tcp_sk(sk); 6209 struct net *net = sock_net(sk); 6210 unsigned long rtt; 6211 u64 delay; 6212 6213 /* More than one full frame received... */ 6214 if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && 6215 /* ... and right edge of window advances far enough. 6216 * (tcp_recvmsg() will send ACK otherwise). 6217 * If application uses SO_RCVLOWAT, we want send ack now if 6218 * we have not received enough bytes to satisfy the condition. 6219 */ 6220 (tp->rcv_nxt - tp->copied_seq < sk->sk_rcvlowat || 6221 __tcp_select_window(sk) >= tp->rcv_wnd)) || 6222 /* We ACK each frame or... */ 6223 tcp_in_quickack_mode(sk) || 6224 /* Protocol state mandates a one-time immediate ACK */ 6225 inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) { 6226 /* If we are running from __release_sock() in user context, 6227 * Defer the ack until tcp_release_cb(). 6228 */ 6229 if (sock_owned_by_user_nocheck(sk) && 6230 READ_ONCE(net->ipv4.sysctl_tcp_backlog_ack_defer)) { 6231 set_bit(TCP_ACK_DEFERRED, &sk->sk_tsq_flags); 6232 return; 6233 } 6234 send_now: 6235 tcp_send_ack(sk); 6236 return; 6237 } 6238 6239 if (!ofo_possible || RB_EMPTY_ROOT(&tp->out_of_order_queue)) { 6240 tcp_send_delayed_ack(sk); 6241 return; 6242 } 6243 6244 if (!tcp_is_sack(tp) || 6245 tp->compressed_ack >= READ_ONCE(net->ipv4.sysctl_tcp_comp_sack_nr)) 6246 goto send_now; 6247 6248 if (tp->compressed_ack_rcv_nxt != tp->rcv_nxt) { 6249 tp->compressed_ack_rcv_nxt = tp->rcv_nxt; 6250 tp->dup_ack_counter = 0; 6251 } 6252 if (tp->dup_ack_counter < TCP_FASTRETRANS_THRESH) { 6253 tp->dup_ack_counter++; 6254 goto send_now; 6255 } 6256 tp->compressed_ack++; 6257 if (hrtimer_is_queued(&tp->compressed_ack_timer)) 6258 return; 6259 6260 /* compress ack timer : comp_sack_rtt_percent of rtt, 6261 * but no more than tcp_comp_sack_delay_ns. 6262 */ 6263 6264 rtt = tp->rcv_rtt_est.rtt_us; 6265 if (tp->srtt_us && tp->srtt_us < rtt) 6266 rtt = tp->srtt_us; 6267 6268 /* delay = (rtt >> 3) * NSEC_PER_USEC * comp_sack_rtt_percent / 100 6269 * -> 6270 * delay = rtt * 1.25 * comp_sack_rtt_percent 6271 */ 6272 delay = (u64)(rtt + (rtt >> 2)) * 6273 READ_ONCE(net->ipv4.sysctl_tcp_comp_sack_rtt_percent); 6274 6275 delay = min(delay, READ_ONCE(net->ipv4.sysctl_tcp_comp_sack_delay_ns)); 6276 6277 sock_hold(sk); 6278 hrtimer_start_range_ns(&tp->compressed_ack_timer, ns_to_ktime(delay), 6279 READ_ONCE(net->ipv4.sysctl_tcp_comp_sack_slack_ns), 6280 HRTIMER_MODE_REL_PINNED_SOFT); 6281 } 6282 6283 static inline void tcp_ack_snd_check(struct sock *sk) 6284 { 6285 if (!inet_csk_ack_scheduled(sk)) { 6286 /* We sent a data segment already. */ 6287 return; 6288 } 6289 __tcp_ack_snd_check(sk, 1); 6290 } 6291 6292 /* 6293 * This routine is only called when we have urgent data 6294 * signaled. Its the 'slow' part of tcp_urg. It could be 6295 * moved inline now as tcp_urg is only called from one 6296 * place. We handle URGent data wrong. We have to - as 6297 * BSD still doesn't use the correction from RFC961. 6298 * For 1003.1g we should support a new option TCP_STDURG to permit 6299 * either form (or just set the sysctl tcp_stdurg). 6300 */ 6301 6302 static void tcp_check_urg(struct sock *sk, const struct tcphdr *th) 6303 { 6304 struct tcp_sock *tp = tcp_sk(sk); 6305 u32 ptr = ntohs(th->urg_ptr); 6306 6307 if (ptr && !READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_stdurg)) 6308 ptr--; 6309 ptr += ntohl(th->seq); 6310 6311 /* Ignore urgent data that we've already seen and read. */ 6312 if (after(tp->copied_seq, ptr)) 6313 return; 6314 6315 /* Do not replay urg ptr. 6316 * 6317 * NOTE: interesting situation not covered by specs. 6318 * Misbehaving sender may send urg ptr, pointing to segment, 6319 * which we already have in ofo queue. We are not able to fetch 6320 * such data and will stay in TCP_URG_NOTYET until will be eaten 6321 * by recvmsg(). Seems, we are not obliged to handle such wicked 6322 * situations. But it is worth to think about possibility of some 6323 * DoSes using some hypothetical application level deadlock. 6324 */ 6325 if (before(ptr, tp->rcv_nxt)) 6326 return; 6327 6328 /* Do we already have a newer (or duplicate) urgent pointer? */ 6329 if (tp->urg_data && !after(ptr, tp->urg_seq)) 6330 return; 6331 6332 /* Tell the world about our new urgent pointer. */ 6333 sk_send_sigurg(sk); 6334 6335 /* We may be adding urgent data when the last byte read was 6336 * urgent. To do this requires some care. We cannot just ignore 6337 * tp->copied_seq since we would read the last urgent byte again 6338 * as data, nor can we alter copied_seq until this data arrives 6339 * or we break the semantics of SIOCATMARK (and thus sockatmark()) 6340 * 6341 * NOTE. Double Dutch. Rendering to plain English: author of comment 6342 * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); 6343 * and expect that both A and B disappear from stream. This is _wrong_. 6344 * Though this happens in BSD with high probability, this is occasional. 6345 * Any application relying on this is buggy. Note also, that fix "works" 6346 * only in this artificial test. Insert some normal data between A and B and we will 6347 * decline of BSD again. Verdict: it is better to remove to trap 6348 * buggy users. 6349 */ 6350 if (tp->urg_seq == tp->copied_seq && tp->urg_data && 6351 !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { 6352 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); 6353 tp->copied_seq++; 6354 if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { 6355 __skb_unlink(skb, &sk->sk_receive_queue); 6356 __kfree_skb(skb); 6357 } 6358 } 6359 6360 WRITE_ONCE(tp->urg_data, TCP_URG_NOTYET); 6361 WRITE_ONCE(tp->urg_seq, ptr); 6362 6363 /* Disable header prediction. */ 6364 tp->pred_flags = 0; 6365 } 6366 6367 /* This is the 'fast' part of urgent handling. */ 6368 static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th) 6369 { 6370 struct tcp_sock *tp = tcp_sk(sk); 6371 6372 /* Check if we get a new urgent pointer - normally not. */ 6373 if (unlikely(th->urg)) 6374 tcp_check_urg(sk, th); 6375 6376 /* Do we wait for any urgent data? - normally not... */ 6377 if (unlikely(tp->urg_data == TCP_URG_NOTYET)) { 6378 u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - 6379 th->syn; 6380 6381 /* Is the urgent pointer pointing into this packet? */ 6382 if (ptr < skb->len) { 6383 u8 tmp; 6384 if (skb_copy_bits(skb, ptr, &tmp, 1)) 6385 BUG(); 6386 WRITE_ONCE(tp->urg_data, TCP_URG_VALID | tmp); 6387 if (!sock_flag(sk, SOCK_DEAD)) 6388 READ_ONCE(sk->sk_data_ready)(sk); 6389 } 6390 } 6391 } 6392 6393 /* Accept RST for rcv_nxt - 1 after a FIN. 6394 * When tcp connections are abruptly terminated from Mac OSX (via ^C), a 6395 * FIN is sent followed by a RST packet. The RST is sent with the same 6396 * sequence number as the FIN, and thus according to RFC 5961 a challenge 6397 * ACK should be sent. However, Mac OSX rate limits replies to challenge 6398 * ACKs on the closed socket. In addition middleboxes can drop either the 6399 * challenge ACK or a subsequent RST. 6400 */ 6401 static bool tcp_reset_check(const struct sock *sk, const struct sk_buff *skb) 6402 { 6403 const struct tcp_sock *tp = tcp_sk(sk); 6404 6405 return unlikely(TCP_SKB_CB(skb)->seq == (tp->rcv_nxt - 1) && 6406 (1 << sk->sk_state) & (TCPF_CLOSE_WAIT | TCPF_LAST_ACK | 6407 TCPF_CLOSING)); 6408 } 6409 6410 /* Does PAWS and seqno based validation of an incoming segment, flags will 6411 * play significant role here. 6412 */ 6413 static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, 6414 const struct tcphdr *th, int syn_inerr) 6415 { 6416 struct tcp_sock *tp = tcp_sk(sk); 6417 bool accecn_reflector = false; 6418 SKB_DR(reason); 6419 6420 /* RFC1323: H1. Apply PAWS check first. */ 6421 if (!tcp_fast_parse_options(sock_net(sk), skb, th, tp) || 6422 !tp->rx_opt.saw_tstamp || 6423 tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW)) 6424 goto step1; 6425 6426 reason = tcp_disordered_ack_check(sk, skb); 6427 if (!reason) 6428 goto step1; 6429 /* Reset is accepted even if it did not pass PAWS. */ 6430 if (th->rst) 6431 goto step1; 6432 if (unlikely(th->syn)) 6433 goto syn_challenge; 6434 6435 /* Old ACK are common, increment PAWS_OLD_ACK 6436 * and do not send a dupack. 6437 */ 6438 if (reason == SKB_DROP_REASON_TCP_RFC7323_PAWS_ACK) { 6439 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWS_OLD_ACK); 6440 goto discard; 6441 } 6442 NET_INC_STATS(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); 6443 if (!tcp_oow_rate_limited(sock_net(sk), skb, 6444 LINUX_MIB_TCPACKSKIPPEDPAWS, 6445 &tp->last_oow_ack_time)) 6446 tcp_send_dupack(sk, skb); 6447 goto discard; 6448 6449 step1: 6450 /* Step 1: check sequence number */ 6451 reason = tcp_sequence(sk, TCP_SKB_CB(skb)->seq, 6452 TCP_SKB_CB(skb)->end_seq, th); 6453 if (reason) { 6454 /* RFC793, page 37: "In all states except SYN-SENT, all reset 6455 * (RST) segments are validated by checking their SEQ-fields." 6456 * And page 69: "If an incoming segment is not acceptable, 6457 * an acknowledgment should be sent in reply (unless the RST 6458 * bit is set, if so drop the segment and return)". 6459 */ 6460 if (!th->rst) { 6461 if (th->syn) 6462 goto syn_challenge; 6463 6464 if (reason == SKB_DROP_REASON_TCP_INVALID_SEQUENCE || 6465 reason == SKB_DROP_REASON_TCP_INVALID_END_SEQUENCE) 6466 NET_INC_STATS(sock_net(sk), 6467 LINUX_MIB_BEYOND_WINDOW); 6468 if (!tcp_oow_rate_limited(sock_net(sk), skb, 6469 LINUX_MIB_TCPACKSKIPPEDSEQ, 6470 &tp->last_oow_ack_time)) 6471 tcp_send_dupack(sk, skb); 6472 } else if (tcp_reset_check(sk, skb)) { 6473 goto reset; 6474 } 6475 goto discard; 6476 } 6477 6478 /* Step 2: check RST bit */ 6479 if (th->rst) { 6480 /* RFC 5961 3.2 (extend to match against (RCV.NXT - 1) after a 6481 * FIN and SACK too if available): 6482 * If seq num matches RCV.NXT or (RCV.NXT - 1) after a FIN, or 6483 * the right-most SACK block, 6484 * then 6485 * RESET the connection 6486 * else 6487 * Send a challenge ACK 6488 */ 6489 if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt || 6490 tcp_reset_check(sk, skb)) 6491 goto reset; 6492 6493 if (tcp_is_sack(tp) && tp->rx_opt.num_sacks > 0) { 6494 struct tcp_sack_block *sp = &tp->selective_acks[0]; 6495 int max_sack = sp[0].end_seq; 6496 int this_sack; 6497 6498 for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; 6499 ++this_sack) { 6500 max_sack = after(sp[this_sack].end_seq, 6501 max_sack) ? 6502 sp[this_sack].end_seq : max_sack; 6503 } 6504 6505 if (TCP_SKB_CB(skb)->seq == max_sack) 6506 goto reset; 6507 } 6508 6509 /* Disable TFO if RST is out-of-order 6510 * and no data has been received 6511 * for current active TFO socket 6512 */ 6513 if (tp->syn_fastopen && !tp->data_segs_in && 6514 sk->sk_state == TCP_ESTABLISHED) 6515 tcp_fastopen_active_disable(sk); 6516 tcp_send_challenge_ack(sk, false); 6517 SKB_DR_SET(reason, TCP_RESET); 6518 goto discard; 6519 } 6520 6521 /* step 3: check security and precedence [ignored] */ 6522 6523 /* step 4: Check for a SYN 6524 * RFC 5961 4.2 : Send a challenge ack 6525 */ 6526 if (th->syn) { 6527 if (tcp_ecn_mode_accecn(tp)) { 6528 accecn_reflector = true; 6529 tp->syn_ect_rcv = TCP_SKB_CB(skb)->ip_dsfield & 6530 INET_ECN_MASK; 6531 if (tp->rx_opt.accecn && 6532 tp->saw_accecn_opt < TCP_ACCECN_OPT_COUNTER_SEEN) { 6533 u8 saw_opt = tcp_accecn_option_init(skb, tp->rx_opt.accecn); 6534 6535 tcp_accecn_saw_opt_fail_recv(tp, saw_opt); 6536 tcp_accecn_opt_demand_min(sk, 1); 6537 } 6538 } 6539 if (sk->sk_state == TCP_SYN_RECV && sk->sk_socket && th->ack && 6540 TCP_SKB_CB(skb)->seq + 1 == TCP_SKB_CB(skb)->end_seq && 6541 TCP_SKB_CB(skb)->seq + 1 == tp->rcv_nxt && 6542 TCP_SKB_CB(skb)->ack_seq == tp->snd_nxt) 6543 goto pass; 6544 syn_challenge: 6545 if (syn_inerr) 6546 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6547 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE); 6548 tcp_send_challenge_ack(sk, accecn_reflector); 6549 SKB_DR_SET(reason, TCP_INVALID_SYN); 6550 goto discard; 6551 } 6552 6553 pass: 6554 bpf_skops_parse_hdr(sk, skb); 6555 bpf_tcp_ops_parse_hdr(sk, skb); 6556 6557 return true; 6558 6559 discard: 6560 tcp_drop_reason(sk, skb, reason); 6561 return false; 6562 6563 reset: 6564 tcp_reset(sk, skb); 6565 __kfree_skb(skb); 6566 return false; 6567 } 6568 6569 /* 6570 * TCP receive function for the ESTABLISHED state. 6571 * 6572 * It is split into a fast path and a slow path. The fast path is 6573 * disabled when: 6574 * - A zero window was announced from us - zero window probing 6575 * is only handled properly in the slow path. 6576 * - Out of order segments arrived. 6577 * - Urgent data is expected. 6578 * - There is no buffer space left 6579 * - Unexpected TCP flags/window values/header lengths are received 6580 * (detected by checking the TCP header against pred_flags) 6581 * - Data is sent in both directions. Fast path only supports pure senders 6582 * or pure receivers (this means either the sequence number or the ack 6583 * value must stay constant) 6584 * - Unexpected TCP option. 6585 * - ACK sequence number is outside [SND.UNA, SND.NXT]. 6586 * 6587 * When these conditions are not satisfied it drops into a standard 6588 * receive procedure patterned after RFC793 to handle all cases. 6589 * The first three cases are guaranteed by proper pred_flags setting, 6590 * the rest is checked inline. Fast processing is turned on in 6591 * tcp_data_queue when everything is OK. 6592 */ 6593 void tcp_rcv_established(struct sock *sk, struct sk_buff *skb) 6594 { 6595 enum skb_drop_reason reason = SKB_DROP_REASON_NOT_SPECIFIED; 6596 const struct tcphdr *th = (const struct tcphdr *)skb->data; 6597 struct tcp_sock *tp = tcp_sk(sk); 6598 unsigned int len = skb->len; 6599 6600 /* TCP congestion window tracking */ 6601 trace_tcp_probe(sk, skb); 6602 6603 tcp_mstamp_refresh_inline(tp); 6604 if (unlikely(!rcu_access_pointer(sk->sk_rx_dst))) 6605 inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb); 6606 /* 6607 * Header prediction. 6608 * The code loosely follows the one in the famous 6609 * "30 instruction TCP receive" Van Jacobson mail. 6610 * 6611 * Van's trick is to deposit buffers into socket queue 6612 * on a device interrupt, to call tcp_recv function 6613 * on the receive process context and checksum and copy 6614 * the buffer to user space. smart... 6615 * 6616 * Our current scheme is not silly either but we take the 6617 * extra cost of the net_bh soft interrupt processing... 6618 * We do checksum and copy also but from device to kernel. 6619 */ 6620 6621 tp->rx_opt.saw_tstamp = 0; 6622 tp->rx_opt.accecn = 0; 6623 6624 /* pred_flags is 0xS?10 << 16 + snd_wnd 6625 * if header_prediction is to be made 6626 * 'S' will always be tp->tcp_header_len >> 2 6627 * '?' will be 0 for the fast path, otherwise pred_flags is 0 to 6628 * turn it off (when there are holes in the receive 6629 * space for instance) 6630 * PSH flag is ignored. 6631 */ 6632 6633 if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && 6634 TCP_SKB_CB(skb)->seq == tp->rcv_nxt && 6635 between(TCP_SKB_CB(skb)->ack_seq, tp->snd_una, tp->snd_nxt)) { 6636 int tcp_header_len = tp->tcp_header_len; 6637 s32 delta = 0; 6638 int flag = 0; 6639 6640 /* Timestamp header prediction: tcp_header_len 6641 * is automatically equal to th->doff*4 due to pred_flags 6642 * match. 6643 */ 6644 6645 /* Check timestamp */ 6646 if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { 6647 /* No? Slow path! */ 6648 if (!tcp_parse_aligned_timestamp(tp, th)) 6649 goto slow_path; 6650 6651 delta = tp->rx_opt.rcv_tsval - 6652 tp->rx_opt.ts_recent; 6653 /* If PAWS failed, check it more carefully in slow path */ 6654 if (delta < 0) 6655 goto slow_path; 6656 6657 /* DO NOT update ts_recent here, if checksum fails 6658 * and timestamp was corrupted part, it will result 6659 * in a hung connection since we will drop all 6660 * future packets due to the PAWS test. 6661 */ 6662 } 6663 6664 if (len <= tcp_header_len) { 6665 /* Bulk data transfer: sender */ 6666 if (len == tcp_header_len) { 6667 /* Predicted packet is in window by definition. 6668 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 6669 * Hence, check seq<=rcv_wup reduces to: 6670 */ 6671 if (tcp_header_len == 6672 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 6673 tp->rcv_nxt == tp->rcv_wup) 6674 flag |= __tcp_replace_ts_recent(tp, 6675 delta); 6676 6677 tcp_ecn_received_counters(sk, skb, 0); 6678 6679 /* We know that such packets are checksummed 6680 * on entry. 6681 */ 6682 tcp_ack(sk, skb, flag); 6683 __kfree_skb(skb); 6684 tcp_data_snd_check(sk); 6685 /* When receiving pure ack in fast path, update 6686 * last ts ecr directly instead of calling 6687 * tcp_rcv_rtt_measure_ts() 6688 */ 6689 tp->rcv_rtt_last_tsecr = tp->rx_opt.rcv_tsecr; 6690 return; 6691 } else { /* Header too small */ 6692 reason = SKB_DROP_REASON_PKT_TOO_SMALL; 6693 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6694 goto discard; 6695 } 6696 } else { 6697 int eaten = 0; 6698 bool fragstolen = false; 6699 6700 if (tcp_checksum_complete(skb)) 6701 goto csum_error; 6702 6703 if (after(TCP_SKB_CB(skb)->end_seq, 6704 tp->rcv_nxt + tcp_receive_window(tp))) 6705 goto validate; 6706 6707 if ((int)skb->truesize > sk->sk_forward_alloc) 6708 goto step5; 6709 6710 /* Predicted packet is in window by definition. 6711 * seq == rcv_nxt and rcv_wup <= rcv_nxt. 6712 * Hence, check seq<=rcv_wup reduces to: 6713 */ 6714 if (tcp_header_len == 6715 (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && 6716 tp->rcv_nxt == tp->rcv_wup) 6717 flag |= __tcp_replace_ts_recent(tp, 6718 delta); 6719 6720 tcp_rcv_rtt_measure_ts(sk, skb); 6721 6722 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPHPHITS); 6723 6724 /* Bulk data transfer: receiver */ 6725 tcp_cleanup_skb(skb); 6726 __skb_pull(skb, tcp_header_len); 6727 tcp_ecn_received_counters(sk, skb, 6728 len - tcp_header_len); 6729 eaten = tcp_queue_rcv(sk, skb, &fragstolen); 6730 6731 tcp_event_data_recv(sk, skb); 6732 6733 if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { 6734 /* Well, only one small jumplet in fast path... */ 6735 tcp_ack(sk, skb, flag | FLAG_DATA); 6736 tcp_data_snd_check(sk); 6737 if (!inet_csk_ack_scheduled(sk)) 6738 goto no_ack; 6739 } else { 6740 tcp_update_wl(tp, TCP_SKB_CB(skb)->seq); 6741 } 6742 6743 __tcp_ack_snd_check(sk, 0); 6744 no_ack: 6745 if (eaten) 6746 kfree_skb_partial(skb, fragstolen); 6747 tcp_data_ready(sk); 6748 return; 6749 } 6750 } 6751 6752 slow_path: 6753 if (len < (th->doff << 2) || tcp_checksum_complete(skb)) 6754 goto csum_error; 6755 6756 if (!th->ack && !th->rst && !th->syn) { 6757 reason = SKB_DROP_REASON_TCP_FLAGS; 6758 goto discard; 6759 } 6760 6761 /* 6762 * Standard slow path. 6763 */ 6764 validate: 6765 if (!tcp_validate_incoming(sk, skb, th, 1)) 6766 return; 6767 6768 step5: 6769 tcp_ecn_received_counters_payload(sk, skb); 6770 6771 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT); 6772 if ((int)reason < 0) { 6773 reason = -reason; 6774 goto discard; 6775 } 6776 tcp_rcv_rtt_measure_ts(sk, skb); 6777 6778 /* Process urgent data. */ 6779 tcp_urg(sk, skb, th); 6780 6781 /* step 7: process the segment text */ 6782 tcp_data_queue(sk, skb); 6783 6784 tcp_data_snd_check(sk); 6785 tcp_ack_snd_check(sk); 6786 return; 6787 6788 csum_error: 6789 reason = SKB_DROP_REASON_TCP_CSUM; 6790 trace_tcp_bad_csum(skb); 6791 TCP_INC_STATS(sock_net(sk), TCP_MIB_CSUMERRORS); 6792 TCP_INC_STATS(sock_net(sk), TCP_MIB_INERRS); 6793 6794 discard: 6795 tcp_drop_reason(sk, skb, reason); 6796 } 6797 6798 void tcp_init_transfer(struct sock *sk, int bpf_op, struct sk_buff *skb) 6799 { 6800 struct inet_connection_sock *icsk = inet_csk(sk); 6801 struct tcp_sock *tp = tcp_sk(sk); 6802 6803 tcp_mtup_init(sk); 6804 icsk->icsk_af_ops->rebuild_header(sk); 6805 tcp_init_metrics(sk); 6806 6807 /* Initialize the congestion window to start the transfer. 6808 * Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 6809 * retransmitted. In light of RFC6298 more aggressive 1sec 6810 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 6811 * retransmission has occurred. 6812 */ 6813 if (tp->total_retrans > 1 && tp->undo_marker) 6814 tcp_snd_cwnd_set(tp, 1); 6815 else 6816 tcp_snd_cwnd_set(tp, tcp_init_cwnd(tp, __sk_dst_get(sk))); 6817 tp->snd_cwnd_stamp = tcp_jiffies32; 6818 6819 bpf_skops_established(sk, bpf_op, skb); 6820 if (bpf_op == BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB) 6821 bpf_tcp_ops_call(active_established, sk, skb); 6822 else 6823 bpf_tcp_ops_call(passive_established, sk, skb); 6824 /* Initialize congestion control unless BPF initialized it already: */ 6825 if (!icsk->icsk_ca_initialized) 6826 tcp_init_congestion_control(sk); 6827 tcp_init_buffer_space(sk); 6828 } 6829 6830 void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) 6831 { 6832 struct tcp_sock *tp = tcp_sk(sk); 6833 struct inet_connection_sock *icsk = inet_csk(sk); 6834 6835 tcp_ao_finish_connect(sk, skb); 6836 tcp_set_state(sk, TCP_ESTABLISHED); 6837 icsk->icsk_ack.lrcvtime = tcp_jiffies32; 6838 6839 if (skb) { 6840 icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); 6841 security_inet_conn_established(sk, skb); 6842 sk_mark_napi_id(sk, skb); 6843 } 6844 6845 tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB, skb); 6846 6847 /* Prevent spurious tcp_cwnd_restart() on first data 6848 * packet. 6849 */ 6850 tp->lsndtime = tcp_jiffies32; 6851 6852 if (sock_flag(sk, SOCK_KEEPOPEN)) 6853 tcp_reset_keepalive_timer(sk, keepalive_time_when(tp)); 6854 6855 if (!tp->rx_opt.snd_wscale) 6856 __tcp_fast_path_on(tp, tp->snd_wnd); 6857 else 6858 tp->pred_flags = 0; 6859 } 6860 6861 static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack, 6862 struct tcp_fastopen_cookie *cookie) 6863 { 6864 struct tcp_sock *tp = tcp_sk(sk); 6865 struct sk_buff *data = tp->syn_data ? tcp_rtx_queue_head(sk) : NULL; 6866 u16 mss = tp->rx_opt.mss_clamp, try_exp = 0; 6867 bool syn_drop = false; 6868 6869 if (mss == READ_ONCE(tp->rx_opt.user_mss)) { 6870 struct tcp_options_received opt; 6871 6872 /* Get original SYNACK MSS value if user MSS sets mss_clamp */ 6873 tcp_clear_options(&opt); 6874 opt.user_mss = opt.mss_clamp = 0; 6875 tcp_parse_options(sock_net(sk), synack, &opt, 0, NULL); 6876 mss = opt.mss_clamp; 6877 } 6878 6879 if (!tp->syn_fastopen) { 6880 /* Ignore an unsolicited cookie */ 6881 cookie->len = -1; 6882 } else if (tp->total_retrans) { 6883 /* SYN timed out and the SYN-ACK neither has a cookie nor 6884 * acknowledges data. Presumably the remote received only 6885 * the retransmitted (regular) SYNs: either the original 6886 * SYN-data or the corresponding SYN-ACK was dropped. 6887 */ 6888 syn_drop = (cookie->len < 0 && data); 6889 } else if (cookie->len < 0 && !tp->syn_data) { 6890 /* We requested a cookie but didn't get it. If we did not use 6891 * the (old) exp opt format then try so next time (try_exp=1). 6892 * Otherwise we go back to use the RFC7413 opt (try_exp=2). 6893 */ 6894 try_exp = tp->syn_fastopen_exp ? 2 : 1; 6895 } 6896 6897 tcp_fastopen_cache_set(sk, mss, cookie, syn_drop, try_exp); 6898 6899 if (data) { /* Retransmit unacked data in SYN */ 6900 if (tp->total_retrans) 6901 tp->fastopen_client_fail = TFO_SYN_RETRANSMITTED; 6902 else 6903 tp->fastopen_client_fail = TFO_DATA_NOT_ACKED; 6904 skb_rbtree_walk_from(data) 6905 tcp_mark_skb_lost(sk, data); 6906 tcp_non_congestion_loss_retransmit(sk); 6907 NET_INC_STATS(sock_net(sk), 6908 LINUX_MIB_TCPFASTOPENACTIVEFAIL); 6909 return true; 6910 } 6911 tp->syn_data_acked = tp->syn_data; 6912 if (tp->syn_data_acked) { 6913 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 6914 /* SYN-data is counted as two separate packets in tcp_ack() */ 6915 if (tp->delivered > 1) 6916 WRITE_ONCE(tp->delivered, tp->delivered - 1); 6917 } 6918 6919 tcp_fastopen_add_skb(sk, synack); 6920 6921 return false; 6922 } 6923 6924 static void smc_check_reset_syn(struct tcp_sock *tp) 6925 { 6926 #if IS_ENABLED(CONFIG_SMC) 6927 if (static_branch_unlikely(&tcp_have_smc)) { 6928 if (tp->syn_smc && !tp->rx_opt.smc_ok) 6929 tp->syn_smc = 0; 6930 } 6931 #endif 6932 } 6933 6934 static void tcp_try_undo_spurious_syn(struct sock *sk) 6935 { 6936 struct tcp_sock *tp = tcp_sk(sk); 6937 u32 syn_stamp; 6938 6939 /* undo_marker is set when SYN or SYNACK times out. The timeout is 6940 * spurious if the ACK's timestamp option echo value matches the 6941 * original SYN timestamp. 6942 */ 6943 syn_stamp = tp->retrans_stamp; 6944 if (tp->undo_marker && syn_stamp && tp->rx_opt.saw_tstamp && 6945 syn_stamp == tp->rx_opt.rcv_tsecr) 6946 tp->undo_marker = 0; 6947 } 6948 6949 static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, 6950 const struct tcphdr *th) 6951 { 6952 struct inet_connection_sock *icsk = inet_csk(sk); 6953 struct tcp_sock *tp = tcp_sk(sk); 6954 struct tcp_fastopen_cookie foc = { .len = -1 }; 6955 int saved_clamp = tp->rx_opt.mss_clamp; 6956 bool fastopen_fail; 6957 SKB_DR(reason); 6958 6959 tcp_parse_options(sock_net(sk), skb, &tp->rx_opt, 0, &foc); 6960 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) 6961 tp->rx_opt.rcv_tsecr -= tp->tsoffset; 6962 6963 if (th->ack) { 6964 /* rfc793: 6965 * "If the state is SYN-SENT then 6966 * first check the ACK bit 6967 * If the ACK bit is set 6968 * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send 6969 * a reset (unless the RST bit is set, if so drop 6970 * the segment and return)" 6971 */ 6972 if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) || 6973 after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { 6974 /* Previous FIN/ACK or RST/ACK might be ignored. */ 6975 if (icsk->icsk_retransmits == 0) 6976 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 6977 TCP_TIMEOUT_MIN, false); 6978 SKB_DR_SET(reason, TCP_INVALID_ACK_SEQUENCE); 6979 goto reset_and_undo; 6980 } 6981 6982 if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && 6983 !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, 6984 tcp_time_stamp_ts(tp))) { 6985 NET_INC_STATS(sock_net(sk), 6986 LINUX_MIB_PAWSACTIVEREJECTED); 6987 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 6988 goto reset_and_undo; 6989 } 6990 6991 /* Now ACK is acceptable. 6992 * 6993 * "If the RST bit is set 6994 * If the ACK was acceptable then signal the user "error: 6995 * connection reset", drop the segment, enter CLOSED state, 6996 * delete TCB, and return." 6997 */ 6998 6999 if (th->rst) { 7000 tcp_reset(sk, skb); 7001 consume: 7002 __kfree_skb(skb); 7003 return 0; 7004 } 7005 7006 /* rfc793: 7007 * "fifth, if neither of the SYN or RST bits is set then 7008 * drop the segment and return." 7009 * 7010 * See note below! 7011 * --ANK(990513) 7012 */ 7013 if (!th->syn) { 7014 SKB_DR_SET(reason, TCP_FLAGS); 7015 goto discard_and_undo; 7016 } 7017 /* rfc793: 7018 * "If the SYN bit is on ... 7019 * are acceptable then ... 7020 * (our SYN has been ACKed), change the connection 7021 * state to ESTABLISHED..." 7022 */ 7023 7024 if (tcp_ecn_mode_any(tp)) 7025 tcp_ecn_rcv_synack(sk, skb, th, 7026 TCP_SKB_CB(skb)->ip_dsfield); 7027 7028 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 7029 tcp_try_undo_spurious_syn(sk); 7030 tcp_ack(sk, skb, FLAG_SLOWPATH); 7031 7032 /* Ok.. it's good. Set up sequence numbers and 7033 * move to established. 7034 */ 7035 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 7036 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 7037 tp->rcv_mwnd_seq = tp->rcv_wup + tp->rcv_wnd; 7038 7039 /* RFC1323: The window in SYN & SYN/ACK segments is 7040 * never scaled. 7041 */ 7042 tp->snd_wnd = ntohs(th->window); 7043 7044 if (!tp->rx_opt.wscale_ok) { 7045 tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; 7046 WRITE_ONCE(tp->window_clamp, 7047 min(tp->window_clamp, 65535U)); 7048 } 7049 7050 if (tp->rx_opt.saw_tstamp) { 7051 tp->rx_opt.tstamp_ok = 1; 7052 tp->tcp_header_len = 7053 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 7054 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 7055 tcp_store_ts_recent(tp); 7056 } else { 7057 tp->tcp_header_len = sizeof(struct tcphdr); 7058 } 7059 7060 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 7061 tcp_initialize_rcv_mss(sk); 7062 7063 /* Remember, tcp_poll() does not lock socket! 7064 * Change state from SYN-SENT only after copied_seq 7065 * is initialized. */ 7066 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 7067 7068 smc_check_reset_syn(tp); 7069 7070 smp_mb(); 7071 7072 tcp_finish_connect(sk, skb); 7073 7074 fastopen_fail = (tp->syn_fastopen || tp->syn_data) && 7075 tcp_rcv_fastopen_synack(sk, skb, &foc); 7076 7077 if (!sock_flag(sk, SOCK_DEAD)) { 7078 sk->sk_state_change(sk); 7079 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 7080 } 7081 if (fastopen_fail) 7082 return -1; 7083 if (sk->sk_write_pending || 7084 READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept) || 7085 inet_csk_in_pingpong_mode(sk)) { 7086 /* Save one ACK. Data will be ready after 7087 * several ticks, if write_pending is set. 7088 * 7089 * It may be deleted, but with this feature tcpdumps 7090 * look so _wonderfully_ clever, that I was not able 7091 * to stand against the temptation 8) --ANK 7092 */ 7093 inet_csk_schedule_ack(sk); 7094 tcp_enter_quickack_mode(sk, TCP_MAX_QUICKACKS); 7095 tcp_reset_xmit_timer(sk, ICSK_TIME_DACK, 7096 TCP_DELACK_MAX, false); 7097 goto consume; 7098 } 7099 tcp_send_ack_reflect_ect(sk, tcp_ecn_mode_accecn(tp)); 7100 return -1; 7101 } 7102 7103 /* No ACK in the segment */ 7104 7105 if (th->rst) { 7106 /* rfc793: 7107 * "If the RST bit is set 7108 * 7109 * Otherwise (no ACK) drop the segment and return." 7110 */ 7111 SKB_DR_SET(reason, TCP_RESET); 7112 goto discard_and_undo; 7113 } 7114 7115 /* PAWS check. */ 7116 if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && 7117 tcp_paws_reject(&tp->rx_opt, 0)) { 7118 SKB_DR_SET(reason, TCP_RFC7323_PAWS); 7119 goto discard_and_undo; 7120 } 7121 if (th->syn) { 7122 /* We see SYN without ACK. It is attempt of 7123 * simultaneous connect with crossed SYNs. 7124 * Particularly, it can be connect to self. 7125 */ 7126 #ifdef CONFIG_TCP_AO 7127 struct tcp_ao_info *ao; 7128 7129 ao = rcu_dereference_protected(tp->ao_info, 7130 lockdep_sock_is_held(sk)); 7131 if (ao) { 7132 WRITE_ONCE(ao->risn, th->seq); 7133 ao->rcv_sne = 0; 7134 } 7135 #endif 7136 tcp_set_state(sk, TCP_SYN_RECV); 7137 7138 if (tp->rx_opt.saw_tstamp) { 7139 tp->rx_opt.tstamp_ok = 1; 7140 tcp_store_ts_recent(tp); 7141 tp->tcp_header_len = 7142 sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; 7143 } else { 7144 tp->tcp_header_len = sizeof(struct tcphdr); 7145 } 7146 7147 WRITE_ONCE(tp->rcv_nxt, TCP_SKB_CB(skb)->seq + 1); 7148 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 7149 tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; 7150 tp->rcv_mwnd_seq = tp->rcv_wup + tp->rcv_wnd; 7151 7152 /* RFC1323: The window in SYN & SYN/ACK segments is 7153 * never scaled. 7154 */ 7155 tp->snd_wnd = ntohs(th->window); 7156 tp->snd_wl1 = TCP_SKB_CB(skb)->seq; 7157 tp->max_window = tp->snd_wnd; 7158 7159 tcp_ecn_rcv_syn(sk, th, skb); 7160 7161 tcp_mtup_init(sk); 7162 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); 7163 tcp_initialize_rcv_mss(sk); 7164 7165 tcp_send_synack(sk); 7166 #if 0 7167 /* Note, we could accept data and URG from this segment. 7168 * There are no obstacles to make this (except that we must 7169 * either change tcp_recvmsg() to prevent it from returning data 7170 * before 3WHS completes per RFC793, or employ TCP Fast Open). 7171 * 7172 * However, if we ignore data in ACKless segments sometimes, 7173 * we have no reasons to accept it sometimes. 7174 * Also, seems the code doing it in step6 of tcp_rcv_state_process 7175 * is not flawless. So, discard packet for sanity. 7176 * Uncomment this return to process the data. 7177 */ 7178 return -1; 7179 #else 7180 goto consume; 7181 #endif 7182 } 7183 /* "fifth, if neither of the SYN or RST bits is set then 7184 * drop the segment and return." 7185 */ 7186 7187 discard_and_undo: 7188 tcp_clear_options(&tp->rx_opt); 7189 tp->rx_opt.mss_clamp = saved_clamp; 7190 tcp_drop_reason(sk, skb, reason); 7191 return 0; 7192 7193 reset_and_undo: 7194 tcp_clear_options(&tp->rx_opt); 7195 tp->rx_opt.mss_clamp = saved_clamp; 7196 /* we can reuse/return @reason to its caller to handle the exception */ 7197 return reason; 7198 } 7199 7200 static void tcp_rcv_synrecv_state_fastopen(struct sock *sk) 7201 { 7202 struct tcp_sock *tp = tcp_sk(sk); 7203 struct request_sock *req; 7204 7205 /* If we are still handling the SYNACK RTO, see if timestamp ECR allows 7206 * undo. If peer SACKs triggered fast recovery, we can't undo here. 7207 */ 7208 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss && !tp->packets_out) 7209 tcp_try_undo_recovery(sk); 7210 7211 tcp_update_rto_time(tp); 7212 WRITE_ONCE(inet_csk(sk)->icsk_retransmits, 0); 7213 /* In tcp_fastopen_synack_timer() on the first SYNACK RTO we set 7214 * retrans_stamp but don't enter CA_Loss, so in case that happened we 7215 * need to zero retrans_stamp here to prevent spurious 7216 * retransmits_timed_out(). However, if the ACK of our SYNACK caused us 7217 * to enter CA_Recovery then we need to leave retrans_stamp as it was 7218 * set entering CA_Recovery, for correct retransmits_timed_out() and 7219 * undo behavior. 7220 */ 7221 tcp_retrans_stamp_cleanup(sk); 7222 7223 /* Once we leave TCP_SYN_RECV or TCP_FIN_WAIT_1, 7224 * we no longer need req so release it. 7225 */ 7226 req = rcu_dereference_protected(tp->fastopen_rsk, 7227 lockdep_sock_is_held(sk)); 7228 reqsk_fastopen_remove(sk, req, false); 7229 7230 /* Re-arm the timer because data may have been sent out. 7231 * This is similar to the regular data transmission case 7232 * when new data has just been ack'ed. 7233 * 7234 * (TFO) - we could try to be more aggressive and 7235 * retransmitting any data sooner based on when they 7236 * are sent out. 7237 */ 7238 tcp_rearm_rto(sk); 7239 } 7240 7241 /* 7242 * This function implements the receiving procedure of RFC 793 for 7243 * all states except ESTABLISHED and TIME_WAIT. 7244 * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be 7245 * address independent. 7246 */ 7247 7248 enum skb_drop_reason 7249 tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb) 7250 { 7251 struct tcp_sock *tp = tcp_sk(sk); 7252 struct inet_connection_sock *icsk = inet_csk(sk); 7253 const struct tcphdr *th = tcp_hdr(skb); 7254 struct request_sock *req; 7255 int queued = 0; 7256 SKB_DR(reason); 7257 7258 switch (sk->sk_state) { 7259 case TCP_CLOSE: 7260 SKB_DR_SET(reason, TCP_CLOSE); 7261 goto discard; 7262 7263 case TCP_LISTEN: 7264 if (th->ack) 7265 return SKB_DROP_REASON_TCP_FLAGS; 7266 7267 if (th->rst) { 7268 SKB_DR_SET(reason, TCP_RESET); 7269 goto discard; 7270 } 7271 if (th->syn) { 7272 if (th->fin) { 7273 SKB_DR_SET(reason, TCP_FLAGS); 7274 goto discard; 7275 } 7276 /* It is possible that we process SYN packets from backlog, 7277 * so we need to make sure to disable BH and RCU right there. 7278 */ 7279 rcu_read_lock(); 7280 local_bh_disable(); 7281 icsk->icsk_af_ops->conn_request(sk, skb); 7282 local_bh_enable(); 7283 rcu_read_unlock(); 7284 7285 consume_skb(skb); 7286 return 0; 7287 } 7288 SKB_DR_SET(reason, TCP_FLAGS); 7289 goto discard; 7290 7291 case TCP_SYN_SENT: 7292 /* We sent nothing yet, so no segment belongs to this connection. 7293 * Send no RST: it would be built from this socket, whose TCP-AO 7294 * keys are not set up either. 7295 */ 7296 if (unlikely(inet_test_bit(DEFER_CONNECT, sk))) { 7297 SKB_DR_SET(reason, NOT_SPECIFIED); 7298 goto discard; 7299 } 7300 7301 tp->rx_opt.saw_tstamp = 0; 7302 tcp_mstamp_refresh(tp); 7303 queued = tcp_rcv_synsent_state_process(sk, skb, th); 7304 if (queued >= 0) 7305 return queued; 7306 7307 /* Do step6 onward by hand. */ 7308 tcp_urg(sk, skb, th); 7309 __kfree_skb(skb); 7310 tcp_data_snd_check(sk); 7311 return 0; 7312 } 7313 7314 tcp_mstamp_refresh(tp); 7315 tp->rx_opt.saw_tstamp = 0; 7316 req = rcu_dereference_protected(tp->fastopen_rsk, 7317 lockdep_sock_is_held(sk)); 7318 if (req) { 7319 bool req_stolen; 7320 7321 WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV && 7322 sk->sk_state != TCP_FIN_WAIT1); 7323 7324 SKB_DR_SET(reason, TCP_FASTOPEN); 7325 if (!tcp_check_req(sk, skb, req, true, &req_stolen, &reason)) 7326 goto discard; 7327 } 7328 7329 if (!th->ack && !th->rst && !th->syn) { 7330 SKB_DR_SET(reason, TCP_FLAGS); 7331 goto discard; 7332 } 7333 if (!tcp_validate_incoming(sk, skb, th, 0)) 7334 return 0; 7335 7336 /* step 5: check the ACK field */ 7337 reason = tcp_ack(sk, skb, FLAG_SLOWPATH | 7338 FLAG_UPDATE_TS_RECENT | 7339 FLAG_NO_CHALLENGE_ACK); 7340 7341 /* accept old ack (reason == 0) during closing */ 7342 if ((int)reason < 0) { 7343 reason = -reason; 7344 if (sk->sk_state == TCP_SYN_RECV) 7345 /* send one RST */ 7346 return reason; 7347 tcp_send_challenge_ack(sk, false); 7348 goto discard; 7349 } 7350 SKB_DR_SET(reason, NOT_SPECIFIED); 7351 switch (sk->sk_state) { 7352 case TCP_SYN_RECV: 7353 WRITE_ONCE(tp->delivered, tp->delivered + 1); /* SYN-ACK delivery isn't tracked in tcp_ack */ 7354 if (!tp->srtt_us) 7355 tcp_synack_rtt_meas(sk, req); 7356 7357 if (tp->rx_opt.tstamp_ok) 7358 tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; 7359 7360 if (req) { 7361 tcp_rcv_synrecv_state_fastopen(sk); 7362 } else { 7363 tcp_try_undo_spurious_syn(sk); 7364 tp->retrans_stamp = 0; 7365 tcp_init_transfer(sk, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, 7366 skb); 7367 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 7368 } 7369 tcp_ao_established(sk); 7370 smp_mb(); 7371 tcp_set_state(sk, TCP_ESTABLISHED); 7372 sk->sk_state_change(sk); 7373 7374 /* Note, that this wakeup is only for marginal crossed SYN case. 7375 * Passively open sockets are not waked up, because 7376 * sk->sk_sleep == NULL and sk->sk_socket == NULL. 7377 */ 7378 if (sk->sk_socket) 7379 sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); 7380 7381 WRITE_ONCE(tp->snd_una, TCP_SKB_CB(skb)->ack_seq); 7382 tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; 7383 tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); 7384 7385 if (!inet_csk(sk)->icsk_ca_ops->cong_control) 7386 tcp_update_pacing_rate(sk); 7387 7388 /* Prevent spurious tcp_cwnd_restart() on first data packet */ 7389 tp->lsndtime = tcp_jiffies32; 7390 7391 tcp_initialize_rcv_mss(sk); 7392 if (tcp_ecn_mode_accecn(tp)) 7393 tcp_accecn_third_ack(sk, skb, tp->syn_ect_snt); 7394 tcp_fast_path_on(tp); 7395 if (sk->sk_shutdown & SEND_SHUTDOWN) 7396 tcp_shutdown(sk, SEND_SHUTDOWN); 7397 7398 break; 7399 7400 case TCP_FIN_WAIT1: { 7401 int tmo; 7402 7403 if (req) 7404 tcp_rcv_synrecv_state_fastopen(sk); 7405 7406 if (tp->snd_una != tp->write_seq) 7407 break; 7408 7409 tcp_set_state(sk, TCP_FIN_WAIT2); 7410 WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | SEND_SHUTDOWN); 7411 7412 sk_dst_confirm(sk); 7413 7414 if (!sock_flag(sk, SOCK_DEAD)) { 7415 /* Wake up lingering close() */ 7416 sk->sk_state_change(sk); 7417 break; 7418 } 7419 7420 if (READ_ONCE(tp->linger2) < 0) { 7421 tcp_done(sk); 7422 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 7423 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 7424 } 7425 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 7426 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 7427 /* Receive out of order FIN after close() */ 7428 if (tp->syn_fastopen && th->fin) 7429 tcp_fastopen_active_disable(sk); 7430 tcp_done(sk); 7431 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 7432 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 7433 } 7434 7435 tmo = tcp_fin_time(sk); 7436 if (tmo > TCP_TIMEWAIT_LEN) { 7437 tcp_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); 7438 } else if (th->fin || sock_owned_by_user(sk)) { 7439 /* Bad case. We could lose such FIN otherwise. 7440 * It is not a big problem, but it looks confusing 7441 * and not so rare event. We still can lose it now, 7442 * if it spins in bh_lock_sock(), but it is really 7443 * marginal case. 7444 */ 7445 tcp_reset_keepalive_timer(sk, tmo); 7446 } else { 7447 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); 7448 goto consume; 7449 } 7450 break; 7451 } 7452 7453 case TCP_CLOSING: 7454 if (tp->snd_una == tp->write_seq) { 7455 tcp_time_wait(sk, TCP_TIME_WAIT, 0); 7456 goto consume; 7457 } 7458 break; 7459 7460 case TCP_LAST_ACK: 7461 if (tp->snd_una == tp->write_seq) { 7462 tcp_update_metrics(sk); 7463 tcp_done(sk); 7464 goto consume; 7465 } 7466 break; 7467 } 7468 7469 /* step 6: check the URG bit */ 7470 tcp_urg(sk, skb, th); 7471 7472 /* step 7: process the segment text */ 7473 switch (sk->sk_state) { 7474 case TCP_CLOSE_WAIT: 7475 case TCP_CLOSING: 7476 case TCP_LAST_ACK: 7477 if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { 7478 /* If a subflow has been reset, the packet should not 7479 * continue to be processed, drop the packet. 7480 */ 7481 if (sk_is_mptcp(sk) && !mptcp_incoming_options(sk, skb)) 7482 goto discard; 7483 break; 7484 } 7485 fallthrough; 7486 case TCP_FIN_WAIT1: 7487 case TCP_FIN_WAIT2: 7488 /* RFC 793 says to queue data in these states, 7489 * RFC 1122 says we MUST send a reset. 7490 * BSD 4.4 also does reset. 7491 */ 7492 if (sk->sk_shutdown & RCV_SHUTDOWN) { 7493 if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && 7494 after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { 7495 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA); 7496 tcp_reset(sk, skb); 7497 return SKB_DROP_REASON_TCP_ABORT_ON_DATA; 7498 } 7499 } 7500 fallthrough; 7501 case TCP_ESTABLISHED: 7502 tcp_data_queue(sk, skb); 7503 queued = 1; 7504 break; 7505 } 7506 7507 /* tcp_data could move socket to TIME-WAIT */ 7508 if (sk->sk_state != TCP_CLOSE) { 7509 tcp_data_snd_check(sk); 7510 tcp_ack_snd_check(sk); 7511 } 7512 7513 if (!queued) { 7514 discard: 7515 tcp_drop_reason(sk, skb, reason); 7516 } 7517 return 0; 7518 7519 consume: 7520 __kfree_skb(skb); 7521 return 0; 7522 } 7523 7524 static inline void pr_drop_req(struct request_sock *req, __u16 port, int family) 7525 { 7526 struct inet_request_sock *ireq = inet_rsk(req); 7527 7528 if (family == AF_INET) 7529 net_dbg_ratelimited("drop open request from %pI4/%u\n", 7530 &ireq->ir_rmt_addr, port); 7531 #if IS_ENABLED(CONFIG_IPV6) 7532 else if (family == AF_INET6) 7533 net_dbg_ratelimited("drop open request from %pI6/%u\n", 7534 &ireq->ir_v6_rmt_addr, port); 7535 #endif 7536 } 7537 7538 /* RFC3168 : 6.1.1 SYN packets must not have ECT/ECN bits set 7539 * 7540 * If we receive a SYN packet with these bits set, it means a 7541 * network is playing bad games with TOS bits. In order to 7542 * avoid possible false congestion notifications, we disable 7543 * TCP ECN negotiation. 7544 * 7545 * Exception: tcp_ca wants ECN. This is required for DCTCP 7546 * congestion control: Linux DCTCP asserts ECT on all packets, 7547 * including SYN, which is most optimal solution; however, 7548 * others, such as FreeBSD do not. 7549 * 7550 * Exception: At least one of the reserved bits of the TCP header (th->res1) is 7551 * set, indicating the use of a future TCP extension (such as AccECN). See 7552 * RFC8311 §4.3 which updates RFC3168 to allow the development of such 7553 * extensions. 7554 */ 7555 static void tcp_ecn_create_request(struct request_sock *req, 7556 const struct sk_buff *skb, 7557 const struct sock *listen_sk, 7558 const struct dst_entry *dst) 7559 { 7560 const struct tcphdr *th = tcp_hdr(skb); 7561 const struct net *net = sock_net(listen_sk); 7562 bool th_ecn = th->ece && th->cwr; 7563 bool ect, ecn_ok; 7564 u32 ecn_ok_dst; 7565 7566 if (tcp_accecn_syn_requested(th) && 7567 (READ_ONCE(net->ipv4.sysctl_tcp_ecn) >= 3 || 7568 tcp_ca_needs_accecn(listen_sk))) { 7569 inet_rsk(req)->ecn_ok = 1; 7570 tcp_rsk(req)->accecn_ok = 1; 7571 tcp_rsk(req)->syn_ect_rcv = TCP_SKB_CB(skb)->ip_dsfield & 7572 INET_ECN_MASK; 7573 return; 7574 } 7575 7576 if (!th_ecn) 7577 return; 7578 7579 ect = !INET_ECN_is_not_ect(TCP_SKB_CB(skb)->ip_dsfield); 7580 ecn_ok_dst = dst_feature(dst, DST_FEATURE_ECN_MASK); 7581 ecn_ok = READ_ONCE(net->ipv4.sysctl_tcp_ecn) || ecn_ok_dst; 7582 7583 if (((!ect || th->res1 || th->ae) && ecn_ok) || 7584 tcp_ca_needs_ecn(listen_sk) || 7585 (ecn_ok_dst & DST_FEATURE_ECN_CA) || 7586 tcp_bpf_ca_needs_ecn((struct sock *)req)) 7587 inet_rsk(req)->ecn_ok = 1; 7588 } 7589 7590 static void tcp_openreq_init(struct request_sock *req, 7591 const struct tcp_options_received *rx_opt, 7592 struct sk_buff *skb, const struct sock *sk) 7593 { 7594 struct inet_request_sock *ireq = inet_rsk(req); 7595 7596 req->rsk_rcv_wnd = 0; /* So that tcp_send_synack() knows! */ 7597 tcp_rsk(req)->rcv_isn = TCP_SKB_CB(skb)->seq; 7598 tcp_rsk(req)->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; 7599 tcp_rsk(req)->snt_synack = 0; 7600 tcp_rsk(req)->snt_tsval_first = 0; 7601 tcp_rsk(req)->last_oow_ack_time = 0; 7602 tcp_rsk(req)->accecn_ok = 0; 7603 tcp_rsk(req)->saw_accecn_opt = TCP_ACCECN_OPT_NOT_SEEN; 7604 tcp_rsk(req)->accecn_fail_mode = 0; 7605 tcp_rsk(req)->syn_ect_rcv = 0; 7606 tcp_rsk(req)->syn_ect_snt = 0; 7607 req->mss = rx_opt->mss_clamp; 7608 req->ts_recent = rx_opt->saw_tstamp ? rx_opt->rcv_tsval : 0; 7609 ireq->tstamp_ok = rx_opt->tstamp_ok; 7610 ireq->sack_ok = rx_opt->sack_ok; 7611 ireq->snd_wscale = rx_opt->snd_wscale; 7612 ireq->wscale_ok = rx_opt->wscale_ok; 7613 ireq->acked = 0; 7614 ireq->ecn_ok = 0; 7615 ireq->ir_rmt_port = tcp_hdr(skb)->source; 7616 ireq->ir_num = ntohs(tcp_hdr(skb)->dest); 7617 ireq->ir_mark = inet_request_mark(sk, skb); 7618 #if IS_ENABLED(CONFIG_SMC) 7619 ireq->smc_ok = rx_opt->smc_ok && !(tcp_sk(sk)->smc_hs_congested && 7620 tcp_sk(sk)->smc_hs_congested(sk)); 7621 #endif 7622 } 7623 7624 /* 7625 * Return true if a syncookie should be sent 7626 */ 7627 static bool tcp_syn_flood_action(struct sock *sk, const char *proto) 7628 { 7629 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; 7630 const char *msg = "Dropping request"; 7631 struct net *net = sock_net(sk); 7632 bool want_cookie = false; 7633 u8 syncookies; 7634 7635 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); 7636 7637 #ifdef CONFIG_SYN_COOKIES 7638 if (syncookies) { 7639 msg = "Sending cookies"; 7640 want_cookie = true; 7641 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDOCOOKIES); 7642 } else 7643 #endif 7644 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPREQQFULLDROP); 7645 7646 if (syncookies != 2 && !READ_ONCE(queue->synflood_warned)) { 7647 WRITE_ONCE(queue->synflood_warned, 1); 7648 if (IS_ENABLED(CONFIG_IPV6) && sk->sk_family == AF_INET6) { 7649 net_info_ratelimited("%s: Possible SYN flooding on port [%pI6c]:%u. %s.\n", 7650 proto, inet6_rcv_saddr(sk), 7651 sk->sk_num, msg); 7652 } else { 7653 net_info_ratelimited("%s: Possible SYN flooding on port %pI4:%u. %s.\n", 7654 proto, &sk->sk_rcv_saddr, 7655 sk->sk_num, msg); 7656 } 7657 } 7658 7659 return want_cookie; 7660 } 7661 7662 static void tcp_reqsk_record_syn(const struct sock *sk, 7663 struct request_sock *req, 7664 const struct sk_buff *skb) 7665 { 7666 if (tcp_sk(sk)->save_syn) { 7667 u32 len = skb_network_header_len(skb) + tcp_hdrlen(skb); 7668 struct saved_syn *saved_syn; 7669 u32 mac_hdrlen; 7670 void *base; 7671 7672 if (tcp_sk(sk)->save_syn == 2) { /* Save full header. */ 7673 base = skb_mac_header(skb); 7674 mac_hdrlen = skb_mac_header_len(skb); 7675 len += mac_hdrlen; 7676 } else { 7677 base = skb_network_header(skb); 7678 mac_hdrlen = 0; 7679 } 7680 7681 saved_syn = kmalloc_flex(*saved_syn, data, len, GFP_ATOMIC); 7682 if (saved_syn) { 7683 saved_syn->mac_hdrlen = mac_hdrlen; 7684 saved_syn->network_hdrlen = skb_network_header_len(skb); 7685 saved_syn->tcp_hdrlen = tcp_hdrlen(skb); 7686 memcpy(saved_syn->data, base, len); 7687 RCU_INIT_POINTER(req->saved_syn, saved_syn); 7688 } 7689 } 7690 } 7691 7692 /* If a SYN cookie is required and supported, returns a clamped MSS value to be 7693 * used for SYN cookie generation. 7694 */ 7695 u16 tcp_get_syncookie_mss(struct request_sock_ops *rsk_ops, 7696 const struct tcp_request_sock_ops *af_ops, 7697 struct sock *sk, struct tcphdr *th) 7698 { 7699 struct tcp_sock *tp = tcp_sk(sk); 7700 u16 mss; 7701 7702 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_syncookies) != 2 && 7703 !inet_csk_reqsk_queue_is_full(sk)) 7704 return 0; 7705 7706 if (!tcp_syn_flood_action(sk, rsk_ops->slab_name)) 7707 return 0; 7708 7709 if (sk_acceptq_is_full(sk)) { 7710 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 7711 return 0; 7712 } 7713 7714 mss = tcp_parse_mss_option(th, READ_ONCE(tp->rx_opt.user_mss)); 7715 if (!mss) 7716 mss = af_ops->mss_clamp; 7717 7718 return mss; 7719 } 7720 7721 int tcp_conn_request(struct request_sock_ops *rsk_ops, 7722 const struct tcp_request_sock_ops *af_ops, 7723 struct sock *sk, struct sk_buff *skb) 7724 { 7725 struct tcp_fastopen_cookie foc = { .len = -1 }; 7726 u32 isn = TCP_SKB_CB(skb)->tcp_tw_isn; 7727 struct tcp_options_received tmp_opt; 7728 const struct tcp_sock *tp = tcp_sk(sk); 7729 struct net *net = sock_net(sk); 7730 struct sock *fastopen_sk = NULL; 7731 union tcp_seq_and_ts_off st; 7732 struct request_sock *req; 7733 bool want_cookie = false; 7734 struct dst_entry *dst; 7735 struct flowi fl; 7736 u8 syncookies; 7737 7738 #ifdef CONFIG_TCP_AO 7739 const struct tcp_ao_hdr *aoh; 7740 #endif 7741 7742 /* If isn is non-zero, this SYN originally matched a TIME_WAIT socket. 7743 * TW sockets are converted to open requests without limitations, 7744 * we skip the queue limits and syncookie checks in the block below. 7745 */ 7746 if (!isn) { 7747 syncookies = READ_ONCE(net->ipv4.sysctl_tcp_syncookies); 7748 7749 if (syncookies == 2 || inet_csk_reqsk_queue_is_full(sk)) { 7750 want_cookie = tcp_syn_flood_action(sk, 7751 rsk_ops->slab_name); 7752 if (!want_cookie) 7753 goto drop; 7754 } 7755 } 7756 7757 if (sk_acceptq_is_full(sk)) { 7758 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS); 7759 goto drop; 7760 } 7761 7762 req = inet_reqsk_alloc(rsk_ops, sk, !want_cookie); 7763 if (!req) 7764 goto drop; 7765 7766 req->syncookie = want_cookie; 7767 tcp_rsk(req)->af_specific = af_ops; 7768 tcp_rsk(req)->ts_off = 0; 7769 tcp_rsk(req)->req_usec_ts = false; 7770 tcp_rsk(req)->txhash = net_tx_rndhash(); 7771 #if IS_ENABLED(CONFIG_MPTCP) 7772 tcp_rsk(req)->is_mptcp = 0; 7773 #endif 7774 7775 tcp_clear_options(&tmp_opt); 7776 tmp_opt.mss_clamp = af_ops->mss_clamp; 7777 tmp_opt.user_mss = READ_ONCE(tp->rx_opt.user_mss); 7778 tcp_parse_options(sock_net(sk), skb, &tmp_opt, 0, 7779 want_cookie ? NULL : &foc); 7780 7781 if (want_cookie && !tmp_opt.saw_tstamp) 7782 tcp_clear_options(&tmp_opt); 7783 7784 if (IS_ENABLED(CONFIG_SMC) && want_cookie) 7785 tmp_opt.smc_ok = 0; 7786 7787 tmp_opt.tstamp_ok = tmp_opt.saw_tstamp; 7788 tcp_openreq_init(req, &tmp_opt, skb, sk); 7789 inet_rsk(req)->no_srccheck = inet_test_bit(TRANSPARENT, sk); 7790 7791 /* Note: tcp_v6_init_req() might override ir_iif for link locals */ 7792 inet_rsk(req)->ir_iif = inet_request_bound_dev_if(sk, skb); 7793 7794 if (want_cookie) { 7795 isn = cookie_init_sequence(af_ops, skb, &req->mss); 7796 /* Use the cookie as txhash so the SYN-ACK and the later full 7797 * socket make the same egress choice (IPv6 ECMP path; IPv4 TX queue). 7798 */ 7799 tcp_rsk(req)->txhash = isn; 7800 } 7801 7802 dst = af_ops->route_req(sk, skb, &fl, req, want_cookie ? 0 : isn); 7803 if (!dst) 7804 goto drop_and_free; 7805 7806 if (tmp_opt.tstamp_ok || (!want_cookie && !isn)) 7807 st = INDIRECT_CALL_INET(af_ops->init_seq_and_ts_off, 7808 tcp_v6_init_seq_and_ts_off, 7809 tcp_v4_init_seq_and_ts_off, 7810 net, skb); 7811 7812 if (tmp_opt.tstamp_ok) { 7813 tcp_rsk(req)->req_usec_ts = dst_tcp_usec_ts(dst); 7814 tcp_rsk(req)->ts_off = st.ts_off; 7815 } 7816 if (!want_cookie && !isn) { 7817 int max_syn_backlog = READ_ONCE(net->ipv4.sysctl_max_syn_backlog); 7818 7819 /* Kill the following clause, if you dislike this way. */ 7820 if (!syncookies && 7821 (max_syn_backlog - inet_csk_reqsk_queue_len(sk) < 7822 (max_syn_backlog >> 2)) && 7823 !tcp_peer_is_proven(req, dst)) { 7824 /* Without syncookies last quarter of 7825 * backlog is filled with destinations, 7826 * proven to be alive. 7827 * It means that we continue to communicate 7828 * to destinations, already remembered 7829 * to the moment of synflood. 7830 */ 7831 pr_drop_req(req, ntohs(tcp_hdr(skb)->source), 7832 rsk_ops->family); 7833 goto drop_and_release; 7834 } 7835 7836 isn = st.seq; 7837 } 7838 7839 tcp_ecn_create_request(req, skb, sk, dst); 7840 7841 if (want_cookie) { 7842 cookie_record_sent(sk); 7843 if (!tmp_opt.tstamp_ok) 7844 inet_rsk(req)->ecn_ok = 0; 7845 } 7846 7847 #ifdef CONFIG_TCP_AO 7848 if (tcp_parse_auth_options(tcp_hdr(skb), NULL, &aoh)) 7849 goto drop_and_release; /* Invalid TCP options */ 7850 if (aoh) { 7851 tcp_rsk(req)->used_tcp_ao = true; 7852 tcp_rsk(req)->ao_rcv_next = aoh->keyid; 7853 tcp_rsk(req)->ao_keyid = aoh->rnext_keyid; 7854 7855 } else { 7856 tcp_rsk(req)->used_tcp_ao = false; 7857 } 7858 #endif 7859 tcp_rsk(req)->snt_isn = isn; 7860 tcp_rsk(req)->syn_tos = TCP_SKB_CB(skb)->ip_dsfield; 7861 tcp_openreq_init_rwin(req, sk, dst); 7862 sk_rx_queue_set(req_to_sk(req), skb); 7863 if (!want_cookie) { 7864 tcp_reqsk_record_syn(sk, req, skb); 7865 fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst); 7866 } 7867 if (fastopen_sk) { 7868 af_ops->send_synack(fastopen_sk, dst, &fl, req, 7869 &foc, TCP_SYNACK_FASTOPEN, skb); 7870 /* Add the child socket directly into the accept queue */ 7871 if (!inet_csk_reqsk_queue_add(sk, req, fastopen_sk)) { 7872 bh_unlock_sock(fastopen_sk); 7873 sock_put(fastopen_sk); 7874 /* tcp_disconnect() already dropped one reqsk ref. */ 7875 reqsk_put(req); 7876 goto drop; 7877 } 7878 READ_ONCE(sk->sk_data_ready)(sk); 7879 bh_unlock_sock(fastopen_sk); 7880 sock_put(fastopen_sk); 7881 } else { 7882 tcp_rsk(req)->tfo_listener = false; 7883 if (!want_cookie && 7884 unlikely(!inet_csk_reqsk_queue_hash_add(sk, req))) { 7885 reqsk_free(req); 7886 dst_release(dst); 7887 return 0; 7888 } 7889 af_ops->send_synack(sk, dst, &fl, req, &foc, 7890 !want_cookie ? TCP_SYNACK_NORMAL : 7891 TCP_SYNACK_COOKIE, 7892 skb); 7893 if (want_cookie) { 7894 reqsk_free(req); 7895 return 0; 7896 } 7897 } 7898 reqsk_put(req); 7899 return 0; 7900 7901 drop_and_release: 7902 dst_release(dst); 7903 drop_and_free: 7904 __reqsk_free(req); 7905 drop: 7906 tcp_listendrop(sk); 7907 return 0; 7908 } 7909