1 // SPDX-License-Identifier: GPL-2.0-only 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: Pedro Roque : Retransmit queue handled by TCP. 24 * : Fragmentation on mtu decrease 25 * : Segment collapse on retransmit 26 * : AF independence 27 * 28 * Linus Torvalds : send_delayed_ack 29 * David S. Miller : Charge memory using the right skb 30 * during syn/ack processing. 31 * David S. Miller : Output engine completely rewritten. 32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 33 * Cacophonix Gaul : draft-minshall-nagle-01 34 * J Hadi Salim : ECN support 35 * 36 */ 37 38 #define pr_fmt(fmt) "TCP: " fmt 39 40 #include <net/tcp.h> 41 #include <net/mptcp.h> 42 43 #include <linux/compiler.h> 44 #include <linux/gfp.h> 45 #include <linux/module.h> 46 #include <linux/static_key.h> 47 48 #include <trace/events/tcp.h> 49 50 /* Refresh clocks of a TCP socket, 51 * ensuring monotically increasing values. 52 */ 53 void tcp_mstamp_refresh(struct tcp_sock *tp) 54 { 55 u64 val = tcp_clock_ns(); 56 57 tp->tcp_clock_cache = val; 58 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC); 59 } 60 61 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 62 int push_one, gfp_t gfp); 63 64 /* Account for new data that has been sent to the network. */ 65 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 66 { 67 struct inet_connection_sock *icsk = inet_csk(sk); 68 struct tcp_sock *tp = tcp_sk(sk); 69 unsigned int prior_packets = tp->packets_out; 70 71 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq); 72 73 __skb_unlink(skb, &sk->sk_write_queue); 74 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); 75 76 if (tp->highest_sack == NULL) 77 tp->highest_sack = skb; 78 79 tp->packets_out += tcp_skb_pcount(skb); 80 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 81 tcp_rearm_rto(sk); 82 83 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 84 tcp_skb_pcount(skb)); 85 } 86 87 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 88 * window scaling factor due to loss of precision. 89 * If window has been shrunk, what should we make? It is not clear at all. 90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 92 * invalid. OK, let's make this for now: 93 */ 94 static inline __u32 tcp_acceptable_seq(const struct sock *sk) 95 { 96 const struct tcp_sock *tp = tcp_sk(sk); 97 98 if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 99 (tp->rx_opt.wscale_ok && 100 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 101 return tp->snd_nxt; 102 else 103 return tcp_wnd_end(tp); 104 } 105 106 /* Calculate mss to advertise in SYN segment. 107 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 108 * 109 * 1. It is independent of path mtu. 110 * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 111 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 112 * attached devices, because some buggy hosts are confused by 113 * large MSS. 114 * 4. We do not make 3, we advertise MSS, calculated from first 115 * hop device mtu, but allow to raise it to ip_rt_min_advmss. 116 * This may be overridden via information stored in routing table. 117 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 118 * probably even Jumbo". 119 */ 120 static __u16 tcp_advertise_mss(struct sock *sk) 121 { 122 struct tcp_sock *tp = tcp_sk(sk); 123 const struct dst_entry *dst = __sk_dst_get(sk); 124 int mss = tp->advmss; 125 126 if (dst) { 127 unsigned int metric = dst_metric_advmss(dst); 128 129 if (metric < mss) { 130 mss = metric; 131 tp->advmss = mss; 132 } 133 } 134 135 return (__u16)mss; 136 } 137 138 /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 139 * This is the first part of cwnd validation mechanism. 140 */ 141 void tcp_cwnd_restart(struct sock *sk, s32 delta) 142 { 143 struct tcp_sock *tp = tcp_sk(sk); 144 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 145 u32 cwnd = tp->snd_cwnd; 146 147 tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 148 149 tp->snd_ssthresh = tcp_current_ssthresh(sk); 150 restart_cwnd = min(restart_cwnd, cwnd); 151 152 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 153 cwnd >>= 1; 154 tp->snd_cwnd = max(cwnd, restart_cwnd); 155 tp->snd_cwnd_stamp = tcp_jiffies32; 156 tp->snd_cwnd_used = 0; 157 } 158 159 /* Congestion state accounting after a packet has been sent. */ 160 static void tcp_event_data_sent(struct tcp_sock *tp, 161 struct sock *sk) 162 { 163 struct inet_connection_sock *icsk = inet_csk(sk); 164 const u32 now = tcp_jiffies32; 165 166 if (tcp_packets_in_flight(tp) == 0) 167 tcp_ca_event(sk, CA_EVENT_TX_START); 168 169 /* If this is the first data packet sent in response to the 170 * previous received data, 171 * and it is a reply for ato after last received packet, 172 * increase pingpong count. 173 */ 174 if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) && 175 (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 176 inet_csk_inc_pingpong_cnt(sk); 177 178 tp->lsndtime = now; 179 } 180 181 /* Account for an ACK we sent. */ 182 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts, 183 u32 rcv_nxt) 184 { 185 struct tcp_sock *tp = tcp_sk(sk); 186 187 if (unlikely(tp->compressed_ack)) { 188 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 189 tp->compressed_ack); 190 tp->compressed_ack = 0; 191 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 192 __sock_put(sk); 193 } 194 195 if (unlikely(rcv_nxt != tp->rcv_nxt)) 196 return; /* Special ACK sent by DCTCP to reflect ECN */ 197 tcp_dec_quickack_mode(sk, pkts); 198 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 199 } 200 201 /* Determine a window scaling and initial window to offer. 202 * Based on the assumption that the given amount of space 203 * will be offered. Store the results in the tp structure. 204 * NOTE: for smooth operation initial space offering should 205 * be a multiple of mss if possible. We assume here that mss >= 1. 206 * This MUST be enforced by all callers. 207 */ 208 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, 209 __u32 *rcv_wnd, __u32 *window_clamp, 210 int wscale_ok, __u8 *rcv_wscale, 211 __u32 init_rcv_wnd) 212 { 213 unsigned int space = (__space < 0 ? 0 : __space); 214 215 /* If no clamp set the clamp to the max possible scaled window */ 216 if (*window_clamp == 0) 217 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 218 space = min(*window_clamp, space); 219 220 /* Quantize space offering to a multiple of mss if possible. */ 221 if (space > mss) 222 space = rounddown(space, mss); 223 224 /* NOTE: offering an initial window larger than 32767 225 * will break some buggy TCP stacks. If the admin tells us 226 * it is likely we could be speaking with such a buggy stack 227 * we will truncate our initial window offering to 32K-1 228 * unless the remote has sent us a window scaling option, 229 * which we interpret as a sign the remote TCP is not 230 * misinterpreting the window field as a signed quantity. 231 */ 232 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 233 (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 234 else 235 (*rcv_wnd) = min_t(u32, space, U16_MAX); 236 237 if (init_rcv_wnd) 238 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 239 240 *rcv_wscale = 0; 241 if (wscale_ok) { 242 /* Set window scaling on max possible window */ 243 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); 244 space = max_t(u32, space, sysctl_rmem_max); 245 space = min_t(u32, space, *window_clamp); 246 *rcv_wscale = clamp_t(int, ilog2(space) - 15, 247 0, TCP_MAX_WSCALE); 248 } 249 /* Set the clamp no higher than max representable value */ 250 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); 251 } 252 EXPORT_SYMBOL(tcp_select_initial_window); 253 254 /* Chose a new window to advertise, update state in tcp_sock for the 255 * socket, and return result with RFC1323 scaling applied. The return 256 * value can be stuffed directly into th->window for an outgoing 257 * frame. 258 */ 259 static u16 tcp_select_window(struct sock *sk) 260 { 261 struct tcp_sock *tp = tcp_sk(sk); 262 u32 old_win = tp->rcv_wnd; 263 u32 cur_win = tcp_receive_window(tp); 264 u32 new_win = __tcp_select_window(sk); 265 266 /* Never shrink the offered window */ 267 if (new_win < cur_win) { 268 /* Danger Will Robinson! 269 * Don't update rcv_wup/rcv_wnd here or else 270 * we will not be able to advertise a zero 271 * window in time. --DaveM 272 * 273 * Relax Will Robinson. 274 */ 275 if (new_win == 0) 276 NET_INC_STATS(sock_net(sk), 277 LINUX_MIB_TCPWANTZEROWINDOWADV); 278 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 279 } 280 tp->rcv_wnd = new_win; 281 tp->rcv_wup = tp->rcv_nxt; 282 283 /* Make sure we do not exceed the maximum possible 284 * scaled window. 285 */ 286 if (!tp->rx_opt.rcv_wscale && 287 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 288 new_win = min(new_win, MAX_TCP_WINDOW); 289 else 290 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 291 292 /* RFC1323 scaling applied */ 293 new_win >>= tp->rx_opt.rcv_wscale; 294 295 /* If we advertise zero window, disable fast path. */ 296 if (new_win == 0) { 297 tp->pred_flags = 0; 298 if (old_win) 299 NET_INC_STATS(sock_net(sk), 300 LINUX_MIB_TCPTOZEROWINDOWADV); 301 } else if (old_win == 0) { 302 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 303 } 304 305 return new_win; 306 } 307 308 /* Packet ECN state for a SYN-ACK */ 309 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 310 { 311 const struct tcp_sock *tp = tcp_sk(sk); 312 313 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 314 if (!(tp->ecn_flags & TCP_ECN_OK)) 315 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 316 else if (tcp_ca_needs_ecn(sk) || 317 tcp_bpf_ca_needs_ecn(sk)) 318 INET_ECN_xmit(sk); 319 } 320 321 /* Packet ECN state for a SYN. */ 322 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 323 { 324 struct tcp_sock *tp = tcp_sk(sk); 325 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 326 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 327 tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 328 329 if (!use_ecn) { 330 const struct dst_entry *dst = __sk_dst_get(sk); 331 332 if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 333 use_ecn = true; 334 } 335 336 tp->ecn_flags = 0; 337 338 if (use_ecn) { 339 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 340 tp->ecn_flags = TCP_ECN_OK; 341 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 342 INET_ECN_xmit(sk); 343 } 344 } 345 346 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 347 { 348 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 349 /* tp->ecn_flags are cleared at a later point in time when 350 * SYN ACK is ultimatively being received. 351 */ 352 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 353 } 354 355 static void 356 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 357 { 358 if (inet_rsk(req)->ecn_ok) 359 th->ece = 1; 360 } 361 362 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 363 * be sent. 364 */ 365 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 366 struct tcphdr *th, int tcp_header_len) 367 { 368 struct tcp_sock *tp = tcp_sk(sk); 369 370 if (tp->ecn_flags & TCP_ECN_OK) { 371 /* Not-retransmitted data segment: set ECT and inject CWR. */ 372 if (skb->len != tcp_header_len && 373 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 374 INET_ECN_xmit(sk); 375 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 376 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 377 th->cwr = 1; 378 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 379 } 380 } else if (!tcp_ca_needs_ecn(sk)) { 381 /* ACK or retransmitted segment: clear ECT|CE */ 382 INET_ECN_dontxmit(sk); 383 } 384 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 385 th->ece = 1; 386 } 387 } 388 389 /* Constructs common control bits of non-data skb. If SYN/FIN is present, 390 * auto increment end seqno. 391 */ 392 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 393 { 394 skb->ip_summed = CHECKSUM_PARTIAL; 395 396 TCP_SKB_CB(skb)->tcp_flags = flags; 397 398 tcp_skb_pcount_set(skb, 1); 399 400 TCP_SKB_CB(skb)->seq = seq; 401 if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 402 seq++; 403 TCP_SKB_CB(skb)->end_seq = seq; 404 } 405 406 static inline bool tcp_urg_mode(const struct tcp_sock *tp) 407 { 408 return tp->snd_una != tp->snd_up; 409 } 410 411 #define OPTION_SACK_ADVERTISE (1 << 0) 412 #define OPTION_TS (1 << 1) 413 #define OPTION_MD5 (1 << 2) 414 #define OPTION_WSCALE (1 << 3) 415 #define OPTION_FAST_OPEN_COOKIE (1 << 8) 416 #define OPTION_SMC (1 << 9) 417 #define OPTION_MPTCP (1 << 10) 418 419 static void smc_options_write(__be32 *ptr, u16 *options) 420 { 421 #if IS_ENABLED(CONFIG_SMC) 422 if (static_branch_unlikely(&tcp_have_smc)) { 423 if (unlikely(OPTION_SMC & *options)) { 424 *ptr++ = htonl((TCPOPT_NOP << 24) | 425 (TCPOPT_NOP << 16) | 426 (TCPOPT_EXP << 8) | 427 (TCPOLEN_EXP_SMC_BASE)); 428 *ptr++ = htonl(TCPOPT_SMC_MAGIC); 429 } 430 } 431 #endif 432 } 433 434 struct tcp_out_options { 435 u16 options; /* bit field of OPTION_* */ 436 u16 mss; /* 0 to disable */ 437 u8 ws; /* window scale, 0 to disable */ 438 u8 num_sack_blocks; /* number of SACK blocks to include */ 439 u8 hash_size; /* bytes in hash_location */ 440 u8 bpf_opt_len; /* length of BPF hdr option */ 441 __u8 *hash_location; /* temporary pointer, overloaded */ 442 __u32 tsval, tsecr; /* need to include OPTION_TS */ 443 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 444 struct mptcp_out_options mptcp; 445 }; 446 447 static void mptcp_options_write(__be32 *ptr, const struct tcp_sock *tp, 448 struct tcp_out_options *opts) 449 { 450 #if IS_ENABLED(CONFIG_MPTCP) 451 if (unlikely(OPTION_MPTCP & opts->options)) 452 mptcp_write_options(ptr, tp, &opts->mptcp); 453 #endif 454 } 455 456 #ifdef CONFIG_CGROUP_BPF 457 static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb, 458 enum tcp_synack_type synack_type) 459 { 460 if (unlikely(!skb)) 461 return BPF_WRITE_HDR_TCP_CURRENT_MSS; 462 463 if (unlikely(synack_type == TCP_SYNACK_COOKIE)) 464 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE; 465 466 return 0; 467 } 468 469 /* req, syn_skb and synack_type are used when writing synack */ 470 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 471 struct request_sock *req, 472 struct sk_buff *syn_skb, 473 enum tcp_synack_type synack_type, 474 struct tcp_out_options *opts, 475 unsigned int *remaining) 476 { 477 struct bpf_sock_ops_kern sock_ops; 478 int err; 479 480 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), 481 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) || 482 !*remaining) 483 return; 484 485 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */ 486 487 /* init sock_ops */ 488 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 489 490 sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB; 491 492 if (req) { 493 /* The listen "sk" cannot be passed here because 494 * it is not locked. It would not make too much 495 * sense to do bpf_setsockopt(listen_sk) based 496 * on individual connection request also. 497 * 498 * Thus, "req" is passed here and the cgroup-bpf-progs 499 * of the listen "sk" will be run. 500 * 501 * "req" is also used here for fastopen even the "sk" here is 502 * a fullsock "child" sk. It is to keep the behavior 503 * consistent between fastopen and non-fastopen on 504 * the bpf programming side. 505 */ 506 sock_ops.sk = (struct sock *)req; 507 sock_ops.syn_skb = syn_skb; 508 } else { 509 sock_owned_by_me(sk); 510 511 sock_ops.is_fullsock = 1; 512 sock_ops.sk = sk; 513 } 514 515 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 516 sock_ops.remaining_opt_len = *remaining; 517 /* tcp_current_mss() does not pass a skb */ 518 if (skb) 519 bpf_skops_init_skb(&sock_ops, skb, 0); 520 521 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 522 523 if (err || sock_ops.remaining_opt_len == *remaining) 524 return; 525 526 opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len; 527 /* round up to 4 bytes */ 528 opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3; 529 530 *remaining -= opts->bpf_opt_len; 531 } 532 533 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 534 struct request_sock *req, 535 struct sk_buff *syn_skb, 536 enum tcp_synack_type synack_type, 537 struct tcp_out_options *opts) 538 { 539 u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len; 540 struct bpf_sock_ops_kern sock_ops; 541 int err; 542 543 if (likely(!max_opt_len)) 544 return; 545 546 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); 547 548 sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB; 549 550 if (req) { 551 sock_ops.sk = (struct sock *)req; 552 sock_ops.syn_skb = syn_skb; 553 } else { 554 sock_owned_by_me(sk); 555 556 sock_ops.is_fullsock = 1; 557 sock_ops.sk = sk; 558 } 559 560 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); 561 sock_ops.remaining_opt_len = max_opt_len; 562 first_opt_off = tcp_hdrlen(skb) - max_opt_len; 563 bpf_skops_init_skb(&sock_ops, skb, first_opt_off); 564 565 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); 566 567 if (err) 568 nr_written = 0; 569 else 570 nr_written = max_opt_len - sock_ops.remaining_opt_len; 571 572 if (nr_written < max_opt_len) 573 memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP, 574 max_opt_len - nr_written); 575 } 576 #else 577 static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, 578 struct request_sock *req, 579 struct sk_buff *syn_skb, 580 enum tcp_synack_type synack_type, 581 struct tcp_out_options *opts, 582 unsigned int *remaining) 583 { 584 } 585 586 static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, 587 struct request_sock *req, 588 struct sk_buff *syn_skb, 589 enum tcp_synack_type synack_type, 590 struct tcp_out_options *opts) 591 { 592 } 593 #endif 594 595 /* Write previously computed TCP options to the packet. 596 * 597 * Beware: Something in the Internet is very sensitive to the ordering of 598 * TCP options, we learned this through the hard way, so be careful here. 599 * Luckily we can at least blame others for their non-compliance but from 600 * inter-operability perspective it seems that we're somewhat stuck with 601 * the ordering which we have been using if we want to keep working with 602 * those broken things (not that it currently hurts anybody as there isn't 603 * particular reason why the ordering would need to be changed). 604 * 605 * At least SACK_PERM as the first option is known to lead to a disaster 606 * (but it may well be that other scenarios fail similarly). 607 */ 608 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 609 struct tcp_out_options *opts) 610 { 611 u16 options = opts->options; /* mungable copy */ 612 613 if (unlikely(OPTION_MD5 & options)) { 614 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 615 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 616 /* overload cookie hash location */ 617 opts->hash_location = (__u8 *)ptr; 618 ptr += 4; 619 } 620 621 if (unlikely(opts->mss)) { 622 *ptr++ = htonl((TCPOPT_MSS << 24) | 623 (TCPOLEN_MSS << 16) | 624 opts->mss); 625 } 626 627 if (likely(OPTION_TS & options)) { 628 if (unlikely(OPTION_SACK_ADVERTISE & options)) { 629 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 630 (TCPOLEN_SACK_PERM << 16) | 631 (TCPOPT_TIMESTAMP << 8) | 632 TCPOLEN_TIMESTAMP); 633 options &= ~OPTION_SACK_ADVERTISE; 634 } else { 635 *ptr++ = htonl((TCPOPT_NOP << 24) | 636 (TCPOPT_NOP << 16) | 637 (TCPOPT_TIMESTAMP << 8) | 638 TCPOLEN_TIMESTAMP); 639 } 640 *ptr++ = htonl(opts->tsval); 641 *ptr++ = htonl(opts->tsecr); 642 } 643 644 if (unlikely(OPTION_SACK_ADVERTISE & options)) { 645 *ptr++ = htonl((TCPOPT_NOP << 24) | 646 (TCPOPT_NOP << 16) | 647 (TCPOPT_SACK_PERM << 8) | 648 TCPOLEN_SACK_PERM); 649 } 650 651 if (unlikely(OPTION_WSCALE & options)) { 652 *ptr++ = htonl((TCPOPT_NOP << 24) | 653 (TCPOPT_WINDOW << 16) | 654 (TCPOLEN_WINDOW << 8) | 655 opts->ws); 656 } 657 658 if (unlikely(opts->num_sack_blocks)) { 659 struct tcp_sack_block *sp = tp->rx_opt.dsack ? 660 tp->duplicate_sack : tp->selective_acks; 661 int this_sack; 662 663 *ptr++ = htonl((TCPOPT_NOP << 24) | 664 (TCPOPT_NOP << 16) | 665 (TCPOPT_SACK << 8) | 666 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 667 TCPOLEN_SACK_PERBLOCK))); 668 669 for (this_sack = 0; this_sack < opts->num_sack_blocks; 670 ++this_sack) { 671 *ptr++ = htonl(sp[this_sack].start_seq); 672 *ptr++ = htonl(sp[this_sack].end_seq); 673 } 674 675 tp->rx_opt.dsack = 0; 676 } 677 678 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 679 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 680 u8 *p = (u8 *)ptr; 681 u32 len; /* Fast Open option length */ 682 683 if (foc->exp) { 684 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 685 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 686 TCPOPT_FASTOPEN_MAGIC); 687 p += TCPOLEN_EXP_FASTOPEN_BASE; 688 } else { 689 len = TCPOLEN_FASTOPEN_BASE + foc->len; 690 *p++ = TCPOPT_FASTOPEN; 691 *p++ = len; 692 } 693 694 memcpy(p, foc->val, foc->len); 695 if ((len & 3) == 2) { 696 p[foc->len] = TCPOPT_NOP; 697 p[foc->len + 1] = TCPOPT_NOP; 698 } 699 ptr += (len + 3) >> 2; 700 } 701 702 smc_options_write(ptr, &options); 703 704 mptcp_options_write(ptr, tp, opts); 705 } 706 707 static void smc_set_option(const struct tcp_sock *tp, 708 struct tcp_out_options *opts, 709 unsigned int *remaining) 710 { 711 #if IS_ENABLED(CONFIG_SMC) 712 if (static_branch_unlikely(&tcp_have_smc)) { 713 if (tp->syn_smc) { 714 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 715 opts->options |= OPTION_SMC; 716 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 717 } 718 } 719 } 720 #endif 721 } 722 723 static void smc_set_option_cond(const struct tcp_sock *tp, 724 const struct inet_request_sock *ireq, 725 struct tcp_out_options *opts, 726 unsigned int *remaining) 727 { 728 #if IS_ENABLED(CONFIG_SMC) 729 if (static_branch_unlikely(&tcp_have_smc)) { 730 if (tp->syn_smc && ireq->smc_ok) { 731 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 732 opts->options |= OPTION_SMC; 733 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 734 } 735 } 736 } 737 #endif 738 } 739 740 static void mptcp_set_option_cond(const struct request_sock *req, 741 struct tcp_out_options *opts, 742 unsigned int *remaining) 743 { 744 if (rsk_is_mptcp(req)) { 745 unsigned int size; 746 747 if (mptcp_synack_options(req, &size, &opts->mptcp)) { 748 if (*remaining >= size) { 749 opts->options |= OPTION_MPTCP; 750 *remaining -= size; 751 } 752 } 753 } 754 } 755 756 /* Compute TCP options for SYN packets. This is not the final 757 * network wire format yet. 758 */ 759 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 760 struct tcp_out_options *opts, 761 struct tcp_md5sig_key **md5) 762 { 763 struct tcp_sock *tp = tcp_sk(sk); 764 unsigned int remaining = MAX_TCP_OPTION_SPACE; 765 struct tcp_fastopen_request *fastopen = tp->fastopen_req; 766 767 *md5 = NULL; 768 #ifdef CONFIG_TCP_MD5SIG 769 if (static_branch_unlikely(&tcp_md5_needed) && 770 rcu_access_pointer(tp->md5sig_info)) { 771 *md5 = tp->af_specific->md5_lookup(sk, sk); 772 if (*md5) { 773 opts->options |= OPTION_MD5; 774 remaining -= TCPOLEN_MD5SIG_ALIGNED; 775 } 776 } 777 #endif 778 779 /* We always get an MSS option. The option bytes which will be seen in 780 * normal data packets should timestamps be used, must be in the MSS 781 * advertised. But we subtract them from tp->mss_cache so that 782 * calculations in tcp_sendmsg are simpler etc. So account for this 783 * fact here if necessary. If we don't do this correctly, as a 784 * receiver we won't recognize data packets as being full sized when we 785 * should, and thus we won't abide by the delayed ACK rules correctly. 786 * SACKs don't matter, we never delay an ACK when we have any of those 787 * going out. */ 788 opts->mss = tcp_advertise_mss(sk); 789 remaining -= TCPOLEN_MSS_ALIGNED; 790 791 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 792 opts->options |= OPTION_TS; 793 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 794 opts->tsecr = tp->rx_opt.ts_recent; 795 remaining -= TCPOLEN_TSTAMP_ALIGNED; 796 } 797 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 798 opts->ws = tp->rx_opt.rcv_wscale; 799 opts->options |= OPTION_WSCALE; 800 remaining -= TCPOLEN_WSCALE_ALIGNED; 801 } 802 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 803 opts->options |= OPTION_SACK_ADVERTISE; 804 if (unlikely(!(OPTION_TS & opts->options))) 805 remaining -= TCPOLEN_SACKPERM_ALIGNED; 806 } 807 808 if (fastopen && fastopen->cookie.len >= 0) { 809 u32 need = fastopen->cookie.len; 810 811 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 812 TCPOLEN_FASTOPEN_BASE; 813 need = (need + 3) & ~3U; /* Align to 32 bits */ 814 if (remaining >= need) { 815 opts->options |= OPTION_FAST_OPEN_COOKIE; 816 opts->fastopen_cookie = &fastopen->cookie; 817 remaining -= need; 818 tp->syn_fastopen = 1; 819 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 820 } 821 } 822 823 smc_set_option(tp, opts, &remaining); 824 825 if (sk_is_mptcp(sk)) { 826 unsigned int size; 827 828 if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) { 829 opts->options |= OPTION_MPTCP; 830 remaining -= size; 831 } 832 } 833 834 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 835 836 return MAX_TCP_OPTION_SPACE - remaining; 837 } 838 839 /* Set up TCP options for SYN-ACKs. */ 840 static unsigned int tcp_synack_options(const struct sock *sk, 841 struct request_sock *req, 842 unsigned int mss, struct sk_buff *skb, 843 struct tcp_out_options *opts, 844 const struct tcp_md5sig_key *md5, 845 struct tcp_fastopen_cookie *foc, 846 enum tcp_synack_type synack_type, 847 struct sk_buff *syn_skb) 848 { 849 struct inet_request_sock *ireq = inet_rsk(req); 850 unsigned int remaining = MAX_TCP_OPTION_SPACE; 851 852 #ifdef CONFIG_TCP_MD5SIG 853 if (md5) { 854 opts->options |= OPTION_MD5; 855 remaining -= TCPOLEN_MD5SIG_ALIGNED; 856 857 /* We can't fit any SACK blocks in a packet with MD5 + TS 858 * options. There was discussion about disabling SACK 859 * rather than TS in order to fit in better with old, 860 * buggy kernels, but that was deemed to be unnecessary. 861 */ 862 if (synack_type != TCP_SYNACK_COOKIE) 863 ireq->tstamp_ok &= !ireq->sack_ok; 864 } 865 #endif 866 867 /* We always send an MSS option. */ 868 opts->mss = mss; 869 remaining -= TCPOLEN_MSS_ALIGNED; 870 871 if (likely(ireq->wscale_ok)) { 872 opts->ws = ireq->rcv_wscale; 873 opts->options |= OPTION_WSCALE; 874 remaining -= TCPOLEN_WSCALE_ALIGNED; 875 } 876 if (likely(ireq->tstamp_ok)) { 877 opts->options |= OPTION_TS; 878 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 879 opts->tsecr = req->ts_recent; 880 remaining -= TCPOLEN_TSTAMP_ALIGNED; 881 } 882 if (likely(ireq->sack_ok)) { 883 opts->options |= OPTION_SACK_ADVERTISE; 884 if (unlikely(!ireq->tstamp_ok)) 885 remaining -= TCPOLEN_SACKPERM_ALIGNED; 886 } 887 if (foc != NULL && foc->len >= 0) { 888 u32 need = foc->len; 889 890 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 891 TCPOLEN_FASTOPEN_BASE; 892 need = (need + 3) & ~3U; /* Align to 32 bits */ 893 if (remaining >= need) { 894 opts->options |= OPTION_FAST_OPEN_COOKIE; 895 opts->fastopen_cookie = foc; 896 remaining -= need; 897 } 898 } 899 900 mptcp_set_option_cond(req, opts, &remaining); 901 902 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 903 904 bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb, 905 synack_type, opts, &remaining); 906 907 return MAX_TCP_OPTION_SPACE - remaining; 908 } 909 910 /* Compute TCP options for ESTABLISHED sockets. This is not the 911 * final wire format yet. 912 */ 913 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 914 struct tcp_out_options *opts, 915 struct tcp_md5sig_key **md5) 916 { 917 struct tcp_sock *tp = tcp_sk(sk); 918 unsigned int size = 0; 919 unsigned int eff_sacks; 920 921 opts->options = 0; 922 923 *md5 = NULL; 924 #ifdef CONFIG_TCP_MD5SIG 925 if (static_branch_unlikely(&tcp_md5_needed) && 926 rcu_access_pointer(tp->md5sig_info)) { 927 *md5 = tp->af_specific->md5_lookup(sk, sk); 928 if (*md5) { 929 opts->options |= OPTION_MD5; 930 size += TCPOLEN_MD5SIG_ALIGNED; 931 } 932 } 933 #endif 934 935 if (likely(tp->rx_opt.tstamp_ok)) { 936 opts->options |= OPTION_TS; 937 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 938 opts->tsecr = tp->rx_opt.ts_recent; 939 size += TCPOLEN_TSTAMP_ALIGNED; 940 } 941 942 /* MPTCP options have precedence over SACK for the limited TCP 943 * option space because a MPTCP connection would be forced to 944 * fall back to regular TCP if a required multipath option is 945 * missing. SACK still gets a chance to use whatever space is 946 * left. 947 */ 948 if (sk_is_mptcp(sk)) { 949 unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 950 unsigned int opt_size = 0; 951 952 if (mptcp_established_options(sk, skb, &opt_size, remaining, 953 &opts->mptcp)) { 954 opts->options |= OPTION_MPTCP; 955 size += opt_size; 956 } 957 } 958 959 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 960 if (unlikely(eff_sacks)) { 961 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 962 if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED + 963 TCPOLEN_SACK_PERBLOCK)) 964 return size; 965 966 opts->num_sack_blocks = 967 min_t(unsigned int, eff_sacks, 968 (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 969 TCPOLEN_SACK_PERBLOCK); 970 971 size += TCPOLEN_SACK_BASE_ALIGNED + 972 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 973 } 974 975 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp, 976 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) { 977 unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 978 979 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); 980 981 size = MAX_TCP_OPTION_SPACE - remaining; 982 } 983 984 return size; 985 } 986 987 988 /* TCP SMALL QUEUES (TSQ) 989 * 990 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 991 * to reduce RTT and bufferbloat. 992 * We do this using a special skb destructor (tcp_wfree). 993 * 994 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 995 * needs to be reallocated in a driver. 996 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 997 * 998 * Since transmit from skb destructor is forbidden, we use a tasklet 999 * to process all sockets that eventually need to send more skbs. 1000 * We use one tasklet per cpu, with its own queue of sockets. 1001 */ 1002 struct tsq_tasklet { 1003 struct tasklet_struct tasklet; 1004 struct list_head head; /* queue of tcp sockets */ 1005 }; 1006 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 1007 1008 static void tcp_tsq_write(struct sock *sk) 1009 { 1010 if ((1 << sk->sk_state) & 1011 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 1012 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 1013 struct tcp_sock *tp = tcp_sk(sk); 1014 1015 if (tp->lost_out > tp->retrans_out && 1016 tp->snd_cwnd > tcp_packets_in_flight(tp)) { 1017 tcp_mstamp_refresh(tp); 1018 tcp_xmit_retransmit_queue(sk); 1019 } 1020 1021 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 1022 0, GFP_ATOMIC); 1023 } 1024 } 1025 1026 static void tcp_tsq_handler(struct sock *sk) 1027 { 1028 bh_lock_sock(sk); 1029 if (!sock_owned_by_user(sk)) 1030 tcp_tsq_write(sk); 1031 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 1032 sock_hold(sk); 1033 bh_unlock_sock(sk); 1034 } 1035 /* 1036 * One tasklet per cpu tries to send more skbs. 1037 * We run in tasklet context but need to disable irqs when 1038 * transferring tsq->head because tcp_wfree() might 1039 * interrupt us (non NAPI drivers) 1040 */ 1041 static void tcp_tasklet_func(struct tasklet_struct *t) 1042 { 1043 struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet); 1044 LIST_HEAD(list); 1045 unsigned long flags; 1046 struct list_head *q, *n; 1047 struct tcp_sock *tp; 1048 struct sock *sk; 1049 1050 local_irq_save(flags); 1051 list_splice_init(&tsq->head, &list); 1052 local_irq_restore(flags); 1053 1054 list_for_each_safe(q, n, &list) { 1055 tp = list_entry(q, struct tcp_sock, tsq_node); 1056 list_del(&tp->tsq_node); 1057 1058 sk = (struct sock *)tp; 1059 smp_mb__before_atomic(); 1060 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 1061 1062 tcp_tsq_handler(sk); 1063 sk_free(sk); 1064 } 1065 } 1066 1067 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 1068 TCPF_WRITE_TIMER_DEFERRED | \ 1069 TCPF_DELACK_TIMER_DEFERRED | \ 1070 TCPF_MTU_REDUCED_DEFERRED) 1071 /** 1072 * tcp_release_cb - tcp release_sock() callback 1073 * @sk: socket 1074 * 1075 * called from release_sock() to perform protocol dependent 1076 * actions before socket release. 1077 */ 1078 void tcp_release_cb(struct sock *sk) 1079 { 1080 unsigned long flags, nflags; 1081 1082 /* perform an atomic operation only if at least one flag is set */ 1083 do { 1084 flags = sk->sk_tsq_flags; 1085 if (!(flags & TCP_DEFERRED_ALL)) 1086 return; 1087 nflags = flags & ~TCP_DEFERRED_ALL; 1088 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 1089 1090 if (flags & TCPF_TSQ_DEFERRED) { 1091 tcp_tsq_write(sk); 1092 __sock_put(sk); 1093 } 1094 /* Here begins the tricky part : 1095 * We are called from release_sock() with : 1096 * 1) BH disabled 1097 * 2) sk_lock.slock spinlock held 1098 * 3) socket owned by us (sk->sk_lock.owned == 1) 1099 * 1100 * But following code is meant to be called from BH handlers, 1101 * so we should keep BH disabled, but early release socket ownership 1102 */ 1103 sock_release_ownership(sk); 1104 1105 if (flags & TCPF_WRITE_TIMER_DEFERRED) { 1106 tcp_write_timer_handler(sk); 1107 __sock_put(sk); 1108 } 1109 if (flags & TCPF_DELACK_TIMER_DEFERRED) { 1110 tcp_delack_timer_handler(sk); 1111 __sock_put(sk); 1112 } 1113 if (flags & TCPF_MTU_REDUCED_DEFERRED) { 1114 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 1115 __sock_put(sk); 1116 } 1117 } 1118 EXPORT_SYMBOL(tcp_release_cb); 1119 1120 void __init tcp_tasklet_init(void) 1121 { 1122 int i; 1123 1124 for_each_possible_cpu(i) { 1125 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 1126 1127 INIT_LIST_HEAD(&tsq->head); 1128 tasklet_setup(&tsq->tasklet, tcp_tasklet_func); 1129 } 1130 } 1131 1132 /* 1133 * Write buffer destructor automatically called from kfree_skb. 1134 * We can't xmit new skbs from this context, as we might already 1135 * hold qdisc lock. 1136 */ 1137 void tcp_wfree(struct sk_buff *skb) 1138 { 1139 struct sock *sk = skb->sk; 1140 struct tcp_sock *tp = tcp_sk(sk); 1141 unsigned long flags, nval, oval; 1142 1143 /* Keep one reference on sk_wmem_alloc. 1144 * Will be released by sk_free() from here or tcp_tasklet_func() 1145 */ 1146 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 1147 1148 /* If this softirq is serviced by ksoftirqd, we are likely under stress. 1149 * Wait until our queues (qdisc + devices) are drained. 1150 * This gives : 1151 * - less callbacks to tcp_write_xmit(), reducing stress (batches) 1152 * - chance for incoming ACK (processed by another cpu maybe) 1153 * to migrate this flow (skb->ooo_okay will be eventually set) 1154 */ 1155 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 1156 goto out; 1157 1158 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 1159 struct tsq_tasklet *tsq; 1160 bool empty; 1161 1162 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 1163 goto out; 1164 1165 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 1166 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 1167 if (nval != oval) 1168 continue; 1169 1170 /* queue this socket to tasklet queue */ 1171 local_irq_save(flags); 1172 tsq = this_cpu_ptr(&tsq_tasklet); 1173 empty = list_empty(&tsq->head); 1174 list_add(&tp->tsq_node, &tsq->head); 1175 if (empty) 1176 tasklet_schedule(&tsq->tasklet); 1177 local_irq_restore(flags); 1178 return; 1179 } 1180 out: 1181 sk_free(sk); 1182 } 1183 1184 /* Note: Called under soft irq. 1185 * We can call TCP stack right away, unless socket is owned by user. 1186 */ 1187 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 1188 { 1189 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 1190 struct sock *sk = (struct sock *)tp; 1191 1192 tcp_tsq_handler(sk); 1193 sock_put(sk); 1194 1195 return HRTIMER_NORESTART; 1196 } 1197 1198 static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 1199 u64 prior_wstamp) 1200 { 1201 struct tcp_sock *tp = tcp_sk(sk); 1202 1203 if (sk->sk_pacing_status != SK_PACING_NONE) { 1204 unsigned long rate = sk->sk_pacing_rate; 1205 1206 /* Original sch_fq does not pace first 10 MSS 1207 * Note that tp->data_segs_out overflows after 2^32 packets, 1208 * this is a minor annoyance. 1209 */ 1210 if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 1211 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 1212 u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 1213 1214 /* take into account OS jitter */ 1215 len_ns -= min_t(u64, len_ns / 2, credit); 1216 tp->tcp_wstamp_ns += len_ns; 1217 } 1218 } 1219 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1220 } 1221 1222 INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 1223 INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); 1224 INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)); 1225 1226 /* This routine actually transmits TCP packets queued in by 1227 * tcp_do_sendmsg(). This is used by both the initial 1228 * transmission and possible later retransmissions. 1229 * All SKB's seen here are completely headerless. It is our 1230 * job to build the TCP header, and pass the packet down to 1231 * IP so it can do the same plus pass the packet off to the 1232 * device. 1233 * 1234 * We are working here with either a clone of the original 1235 * SKB, or a fresh unique copy made by the retransmit engine. 1236 */ 1237 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 1238 int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 1239 { 1240 const struct inet_connection_sock *icsk = inet_csk(sk); 1241 struct inet_sock *inet; 1242 struct tcp_sock *tp; 1243 struct tcp_skb_cb *tcb; 1244 struct tcp_out_options opts; 1245 unsigned int tcp_options_size, tcp_header_size; 1246 struct sk_buff *oskb = NULL; 1247 struct tcp_md5sig_key *md5; 1248 struct tcphdr *th; 1249 u64 prior_wstamp; 1250 int err; 1251 1252 BUG_ON(!skb || !tcp_skb_pcount(skb)); 1253 tp = tcp_sk(sk); 1254 prior_wstamp = tp->tcp_wstamp_ns; 1255 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 1256 skb->skb_mstamp_ns = tp->tcp_wstamp_ns; 1257 if (clone_it) { 1258 oskb = skb; 1259 1260 tcp_skb_tsorted_save(oskb) { 1261 if (unlikely(skb_cloned(oskb))) 1262 skb = pskb_copy(oskb, gfp_mask); 1263 else 1264 skb = skb_clone(oskb, gfp_mask); 1265 } tcp_skb_tsorted_restore(oskb); 1266 1267 if (unlikely(!skb)) 1268 return -ENOBUFS; 1269 /* retransmit skbs might have a non zero value in skb->dev 1270 * because skb->dev is aliased with skb->rbnode.rb_left 1271 */ 1272 skb->dev = NULL; 1273 } 1274 1275 inet = inet_sk(sk); 1276 tcb = TCP_SKB_CB(skb); 1277 memset(&opts, 0, sizeof(opts)); 1278 1279 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 1280 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 1281 } else { 1282 tcp_options_size = tcp_established_options(sk, skb, &opts, 1283 &md5); 1284 /* Force a PSH flag on all (GSO) packets to expedite GRO flush 1285 * at receiver : This slightly improve GRO performance. 1286 * Note that we do not force the PSH flag for non GSO packets, 1287 * because they might be sent under high congestion events, 1288 * and in this case it is better to delay the delivery of 1-MSS 1289 * packets and thus the corresponding ACK packet that would 1290 * release the following packet. 1291 */ 1292 if (tcp_skb_pcount(skb) > 1) 1293 tcb->tcp_flags |= TCPHDR_PSH; 1294 } 1295 tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 1296 1297 /* if no packet is in qdisc/device queue, then allow XPS to select 1298 * another queue. We can be called from tcp_tsq_handler() 1299 * which holds one reference to sk. 1300 * 1301 * TODO: Ideally, in-flight pure ACK packets should not matter here. 1302 * One way to get this would be to set skb->truesize = 2 on them. 1303 */ 1304 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 1305 1306 /* If we had to use memory reserve to allocate this skb, 1307 * this might cause drops if packet is looped back : 1308 * Other socket might not have SOCK_MEMALLOC. 1309 * Packets not looped back do not care about pfmemalloc. 1310 */ 1311 skb->pfmemalloc = 0; 1312 1313 skb_push(skb, tcp_header_size); 1314 skb_reset_transport_header(skb); 1315 1316 skb_orphan(skb); 1317 skb->sk = sk; 1318 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1319 refcount_add(skb->truesize, &sk->sk_wmem_alloc); 1320 1321 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1322 1323 /* Build TCP header and checksum it. */ 1324 th = (struct tcphdr *)skb->data; 1325 th->source = inet->inet_sport; 1326 th->dest = inet->inet_dport; 1327 th->seq = htonl(tcb->seq); 1328 th->ack_seq = htonl(rcv_nxt); 1329 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 1330 tcb->tcp_flags); 1331 1332 th->check = 0; 1333 th->urg_ptr = 0; 1334 1335 /* The urg_mode check is necessary during a below snd_una win probe */ 1336 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 1337 if (before(tp->snd_up, tcb->seq + 0x10000)) { 1338 th->urg_ptr = htons(tp->snd_up - tcb->seq); 1339 th->urg = 1; 1340 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 1341 th->urg_ptr = htons(0xFFFF); 1342 th->urg = 1; 1343 } 1344 } 1345 1346 skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1347 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1348 th->window = htons(tcp_select_window(sk)); 1349 tcp_ecn_send(sk, skb, th, tcp_header_size); 1350 } else { 1351 /* RFC1323: The window in SYN & SYN/ACK segments 1352 * is never scaled. 1353 */ 1354 th->window = htons(min(tp->rcv_wnd, 65535U)); 1355 } 1356 1357 tcp_options_write((__be32 *)(th + 1), tp, &opts); 1358 1359 #ifdef CONFIG_TCP_MD5SIG 1360 /* Calculate the MD5 hash, as we have all we need now */ 1361 if (md5) { 1362 sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1363 tp->af_specific->calc_md5_hash(opts.hash_location, 1364 md5, sk, skb); 1365 } 1366 #endif 1367 1368 /* BPF prog is the last one writing header option */ 1369 bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts); 1370 1371 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check, 1372 tcp_v6_send_check, tcp_v4_send_check, 1373 sk, skb); 1374 1375 if (likely(tcb->tcp_flags & TCPHDR_ACK)) 1376 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); 1377 1378 if (skb->len != tcp_header_size) { 1379 tcp_event_data_sent(tp, sk); 1380 tp->data_segs_out += tcp_skb_pcount(skb); 1381 tp->bytes_sent += skb->len - tcp_header_size; 1382 } 1383 1384 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1385 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1386 tcp_skb_pcount(skb)); 1387 1388 tp->segs_out += tcp_skb_pcount(skb); 1389 skb_set_hash_from_sk(skb, sk); 1390 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1391 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1392 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1393 1394 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1395 1396 /* Cleanup our debris for IP stacks */ 1397 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1398 sizeof(struct inet6_skb_parm))); 1399 1400 tcp_add_tx_delay(skb, tp); 1401 1402 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit, 1403 inet6_csk_xmit, ip_queue_xmit, 1404 sk, skb, &inet->cork.fl); 1405 1406 if (unlikely(err > 0)) { 1407 tcp_enter_cwr(sk); 1408 err = net_xmit_eval(err); 1409 } 1410 if (!err && oskb) { 1411 tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1412 tcp_rate_skb_sent(sk, oskb); 1413 } 1414 return err; 1415 } 1416 1417 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 1418 gfp_t gfp_mask) 1419 { 1420 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 1421 tcp_sk(sk)->rcv_nxt); 1422 } 1423 1424 /* This routine just queues the buffer for sending. 1425 * 1426 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 1427 * otherwise socket can stall. 1428 */ 1429 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 1430 { 1431 struct tcp_sock *tp = tcp_sk(sk); 1432 1433 /* Advance write_seq and place onto the write_queue. */ 1434 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq); 1435 __skb_header_release(skb); 1436 tcp_add_write_queue_tail(sk, skb); 1437 sk_wmem_queued_add(sk, skb->truesize); 1438 sk_mem_charge(sk, skb->truesize); 1439 } 1440 1441 /* Initialize TSO segments for a packet. */ 1442 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1443 { 1444 if (skb->len <= mss_now) { 1445 /* Avoid the costly divide in the normal 1446 * non-TSO case. 1447 */ 1448 tcp_skb_pcount_set(skb, 1); 1449 TCP_SKB_CB(skb)->tcp_gso_size = 0; 1450 } else { 1451 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1452 TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 1453 } 1454 } 1455 1456 /* Pcount in the middle of the write queue got changed, we need to do various 1457 * tweaks to fix counters 1458 */ 1459 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1460 { 1461 struct tcp_sock *tp = tcp_sk(sk); 1462 1463 tp->packets_out -= decr; 1464 1465 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1466 tp->sacked_out -= decr; 1467 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1468 tp->retrans_out -= decr; 1469 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1470 tp->lost_out -= decr; 1471 1472 /* Reno case is special. Sigh... */ 1473 if (tcp_is_reno(tp) && decr > 0) 1474 tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1475 1476 if (tp->lost_skb_hint && 1477 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1478 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1479 tp->lost_cnt_hint -= decr; 1480 1481 tcp_verify_left_out(tp); 1482 } 1483 1484 static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 1485 { 1486 return TCP_SKB_CB(skb)->txstamp_ack || 1487 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 1488 } 1489 1490 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1491 { 1492 struct skb_shared_info *shinfo = skb_shinfo(skb); 1493 1494 if (unlikely(tcp_has_tx_tstamp(skb)) && 1495 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1496 struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1497 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1498 1499 shinfo->tx_flags &= ~tsflags; 1500 shinfo2->tx_flags |= tsflags; 1501 swap(shinfo->tskey, shinfo2->tskey); 1502 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1503 TCP_SKB_CB(skb)->txstamp_ack = 0; 1504 } 1505 } 1506 1507 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1508 { 1509 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1510 TCP_SKB_CB(skb)->eor = 0; 1511 } 1512 1513 /* Insert buff after skb on the write or rtx queue of sk. */ 1514 static void tcp_insert_write_queue_after(struct sk_buff *skb, 1515 struct sk_buff *buff, 1516 struct sock *sk, 1517 enum tcp_queue tcp_queue) 1518 { 1519 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 1520 __skb_queue_after(&sk->sk_write_queue, skb, buff); 1521 else 1522 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 1523 } 1524 1525 /* Function to create two new TCP segments. Shrinks the given segment 1526 * to the specified size and appends a new segment with the rest of the 1527 * packet to the list. This won't be called frequently, I hope. 1528 * Remember, these are still headerless SKBs at this point. 1529 */ 1530 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 1531 struct sk_buff *skb, u32 len, 1532 unsigned int mss_now, gfp_t gfp) 1533 { 1534 struct tcp_sock *tp = tcp_sk(sk); 1535 struct sk_buff *buff; 1536 int nsize, old_factor; 1537 long limit; 1538 int nlen; 1539 u8 flags; 1540 1541 if (WARN_ON(len > skb->len)) 1542 return -EINVAL; 1543 1544 nsize = skb_headlen(skb) - len; 1545 if (nsize < 0) 1546 nsize = 0; 1547 1548 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. 1549 * We need some allowance to not penalize applications setting small 1550 * SO_SNDBUF values. 1551 * Also allow first and last skb in retransmit queue to be split. 1552 */ 1553 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE); 1554 if (unlikely((sk->sk_wmem_queued >> 1) > limit && 1555 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && 1556 skb != tcp_rtx_queue_head(sk) && 1557 skb != tcp_rtx_queue_tail(sk))) { 1558 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); 1559 return -ENOMEM; 1560 } 1561 1562 if (skb_unclone(skb, gfp)) 1563 return -ENOMEM; 1564 1565 /* Get a new skb... force flag on. */ 1566 buff = tcp_stream_alloc_skb(sk, nsize, gfp, true); 1567 if (!buff) 1568 return -ENOMEM; /* We'll just try again later. */ 1569 skb_copy_decrypted(buff, skb); 1570 mptcp_skb_ext_copy(buff, skb); 1571 1572 sk_wmem_queued_add(sk, buff->truesize); 1573 sk_mem_charge(sk, buff->truesize); 1574 nlen = skb->len - len - nsize; 1575 buff->truesize += nlen; 1576 skb->truesize -= nlen; 1577 1578 /* Correct the sequence numbers. */ 1579 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1580 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1581 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1582 1583 /* PSH and FIN should only be set in the second packet. */ 1584 flags = TCP_SKB_CB(skb)->tcp_flags; 1585 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 1586 TCP_SKB_CB(buff)->tcp_flags = flags; 1587 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1588 tcp_skb_fragment_eor(skb, buff); 1589 1590 skb_split(skb, buff, len); 1591 1592 buff->tstamp = skb->tstamp; 1593 tcp_fragment_tstamp(skb, buff); 1594 1595 old_factor = tcp_skb_pcount(skb); 1596 1597 /* Fix up tso_factor for both original and new SKB. */ 1598 tcp_set_skb_tso_segs(skb, mss_now); 1599 tcp_set_skb_tso_segs(buff, mss_now); 1600 1601 /* Update delivered info for the new segment */ 1602 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1603 1604 /* If this packet has been sent out already, we must 1605 * adjust the various packet counters. 1606 */ 1607 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 1608 int diff = old_factor - tcp_skb_pcount(skb) - 1609 tcp_skb_pcount(buff); 1610 1611 if (diff) 1612 tcp_adjust_pcount(sk, skb, diff); 1613 } 1614 1615 /* Link BUFF into the send queue. */ 1616 __skb_header_release(buff); 1617 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1618 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1619 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 1620 1621 return 0; 1622 } 1623 1624 /* This is similar to __pskb_pull_tail(). The difference is that pulled 1625 * data is not copied, but immediately discarded. 1626 */ 1627 static int __pskb_trim_head(struct sk_buff *skb, int len) 1628 { 1629 struct skb_shared_info *shinfo; 1630 int i, k, eat; 1631 1632 eat = min_t(int, len, skb_headlen(skb)); 1633 if (eat) { 1634 __skb_pull(skb, eat); 1635 len -= eat; 1636 if (!len) 1637 return 0; 1638 } 1639 eat = len; 1640 k = 0; 1641 shinfo = skb_shinfo(skb); 1642 for (i = 0; i < shinfo->nr_frags; i++) { 1643 int size = skb_frag_size(&shinfo->frags[i]); 1644 1645 if (size <= eat) { 1646 skb_frag_unref(skb, i); 1647 eat -= size; 1648 } else { 1649 shinfo->frags[k] = shinfo->frags[i]; 1650 if (eat) { 1651 skb_frag_off_add(&shinfo->frags[k], eat); 1652 skb_frag_size_sub(&shinfo->frags[k], eat); 1653 eat = 0; 1654 } 1655 k++; 1656 } 1657 } 1658 shinfo->nr_frags = k; 1659 1660 skb->data_len -= len; 1661 skb->len = skb->data_len; 1662 return len; 1663 } 1664 1665 /* Remove acked data from a packet in the transmit queue. */ 1666 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 1667 { 1668 u32 delta_truesize; 1669 1670 if (skb_unclone(skb, GFP_ATOMIC)) 1671 return -ENOMEM; 1672 1673 delta_truesize = __pskb_trim_head(skb, len); 1674 1675 TCP_SKB_CB(skb)->seq += len; 1676 1677 if (delta_truesize) { 1678 skb->truesize -= delta_truesize; 1679 sk_wmem_queued_add(sk, -delta_truesize); 1680 sk_mem_uncharge(sk, delta_truesize); 1681 } 1682 1683 /* Any change of skb->len requires recalculation of tso factor. */ 1684 if (tcp_skb_pcount(skb) > 1) 1685 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 1686 1687 return 0; 1688 } 1689 1690 /* Calculate MSS not accounting any TCP options. */ 1691 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 1692 { 1693 const struct tcp_sock *tp = tcp_sk(sk); 1694 const struct inet_connection_sock *icsk = inet_csk(sk); 1695 int mss_now; 1696 1697 /* Calculate base mss without TCP options: 1698 It is MMS_S - sizeof(tcphdr) of rfc1122 1699 */ 1700 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 1701 1702 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 1703 if (icsk->icsk_af_ops->net_frag_header_len) { 1704 const struct dst_entry *dst = __sk_dst_get(sk); 1705 1706 if (dst && dst_allfrag(dst)) 1707 mss_now -= icsk->icsk_af_ops->net_frag_header_len; 1708 } 1709 1710 /* Clamp it (mss_clamp does not include tcp options) */ 1711 if (mss_now > tp->rx_opt.mss_clamp) 1712 mss_now = tp->rx_opt.mss_clamp; 1713 1714 /* Now subtract optional transport overhead */ 1715 mss_now -= icsk->icsk_ext_hdr_len; 1716 1717 /* Then reserve room for full set of TCP options and 8 bytes of data */ 1718 mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss); 1719 return mss_now; 1720 } 1721 1722 /* Calculate MSS. Not accounting for SACKs here. */ 1723 int tcp_mtu_to_mss(struct sock *sk, int pmtu) 1724 { 1725 /* Subtract TCP options size, not including SACKs */ 1726 return __tcp_mtu_to_mss(sk, pmtu) - 1727 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 1728 } 1729 EXPORT_SYMBOL(tcp_mtu_to_mss); 1730 1731 /* Inverse of above */ 1732 int tcp_mss_to_mtu(struct sock *sk, int mss) 1733 { 1734 const struct tcp_sock *tp = tcp_sk(sk); 1735 const struct inet_connection_sock *icsk = inet_csk(sk); 1736 int mtu; 1737 1738 mtu = mss + 1739 tp->tcp_header_len + 1740 icsk->icsk_ext_hdr_len + 1741 icsk->icsk_af_ops->net_header_len; 1742 1743 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 1744 if (icsk->icsk_af_ops->net_frag_header_len) { 1745 const struct dst_entry *dst = __sk_dst_get(sk); 1746 1747 if (dst && dst_allfrag(dst)) 1748 mtu += icsk->icsk_af_ops->net_frag_header_len; 1749 } 1750 return mtu; 1751 } 1752 EXPORT_SYMBOL(tcp_mss_to_mtu); 1753 1754 /* MTU probing init per socket */ 1755 void tcp_mtup_init(struct sock *sk) 1756 { 1757 struct tcp_sock *tp = tcp_sk(sk); 1758 struct inet_connection_sock *icsk = inet_csk(sk); 1759 struct net *net = sock_net(sk); 1760 1761 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 1762 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 1763 icsk->icsk_af_ops->net_header_len; 1764 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 1765 icsk->icsk_mtup.probe_size = 0; 1766 if (icsk->icsk_mtup.enabled) 1767 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 1768 } 1769 EXPORT_SYMBOL(tcp_mtup_init); 1770 1771 /* This function synchronize snd mss to current pmtu/exthdr set. 1772 1773 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 1774 for TCP options, but includes only bare TCP header. 1775 1776 tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1777 It is minimum of user_mss and mss received with SYN. 1778 It also does not include TCP options. 1779 1780 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 1781 1782 tp->mss_cache is current effective sending mss, including 1783 all tcp options except for SACKs. It is evaluated, 1784 taking into account current pmtu, but never exceeds 1785 tp->rx_opt.mss_clamp. 1786 1787 NOTE1. rfc1122 clearly states that advertised MSS 1788 DOES NOT include either tcp or ip options. 1789 1790 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1791 are READ ONLY outside this function. --ANK (980731) 1792 */ 1793 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 1794 { 1795 struct tcp_sock *tp = tcp_sk(sk); 1796 struct inet_connection_sock *icsk = inet_csk(sk); 1797 int mss_now; 1798 1799 if (icsk->icsk_mtup.search_high > pmtu) 1800 icsk->icsk_mtup.search_high = pmtu; 1801 1802 mss_now = tcp_mtu_to_mss(sk, pmtu); 1803 mss_now = tcp_bound_to_half_wnd(tp, mss_now); 1804 1805 /* And store cached results */ 1806 icsk->icsk_pmtu_cookie = pmtu; 1807 if (icsk->icsk_mtup.enabled) 1808 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1809 tp->mss_cache = mss_now; 1810 1811 return mss_now; 1812 } 1813 EXPORT_SYMBOL(tcp_sync_mss); 1814 1815 /* Compute the current effective MSS, taking SACKs and IP options, 1816 * and even PMTU discovery events into account. 1817 */ 1818 unsigned int tcp_current_mss(struct sock *sk) 1819 { 1820 const struct tcp_sock *tp = tcp_sk(sk); 1821 const struct dst_entry *dst = __sk_dst_get(sk); 1822 u32 mss_now; 1823 unsigned int header_len; 1824 struct tcp_out_options opts; 1825 struct tcp_md5sig_key *md5; 1826 1827 mss_now = tp->mss_cache; 1828 1829 if (dst) { 1830 u32 mtu = dst_mtu(dst); 1831 if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 1832 mss_now = tcp_sync_mss(sk, mtu); 1833 } 1834 1835 header_len = tcp_established_options(sk, NULL, &opts, &md5) + 1836 sizeof(struct tcphdr); 1837 /* The mss_cache is sized based on tp->tcp_header_len, which assumes 1838 * some common options. If this is an odd packet (because we have SACK 1839 * blocks etc) then our calculated header_len will be different, and 1840 * we have to adjust mss_now correspondingly */ 1841 if (header_len != tp->tcp_header_len) { 1842 int delta = (int) header_len - tp->tcp_header_len; 1843 mss_now -= delta; 1844 } 1845 1846 return mss_now; 1847 } 1848 1849 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 1850 * As additional protections, we do not touch cwnd in retransmission phases, 1851 * and if application hit its sndbuf limit recently. 1852 */ 1853 static void tcp_cwnd_application_limited(struct sock *sk) 1854 { 1855 struct tcp_sock *tp = tcp_sk(sk); 1856 1857 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 1858 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 1859 /* Limited by application or receiver window. */ 1860 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1861 u32 win_used = max(tp->snd_cwnd_used, init_win); 1862 if (win_used < tp->snd_cwnd) { 1863 tp->snd_ssthresh = tcp_current_ssthresh(sk); 1864 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 1865 } 1866 tp->snd_cwnd_used = 0; 1867 } 1868 tp->snd_cwnd_stamp = tcp_jiffies32; 1869 } 1870 1871 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1872 { 1873 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1874 struct tcp_sock *tp = tcp_sk(sk); 1875 1876 /* Track the maximum number of outstanding packets in each 1877 * window, and remember whether we were cwnd-limited then. 1878 */ 1879 if (!before(tp->snd_una, tp->max_packets_seq) || 1880 tp->packets_out > tp->max_packets_out || 1881 is_cwnd_limited) { 1882 tp->max_packets_out = tp->packets_out; 1883 tp->max_packets_seq = tp->snd_nxt; 1884 tp->is_cwnd_limited = is_cwnd_limited; 1885 } 1886 1887 if (tcp_is_cwnd_limited(sk)) { 1888 /* Network is feed fully. */ 1889 tp->snd_cwnd_used = 0; 1890 tp->snd_cwnd_stamp = tcp_jiffies32; 1891 } else { 1892 /* Network starves. */ 1893 if (tp->packets_out > tp->snd_cwnd_used) 1894 tp->snd_cwnd_used = tp->packets_out; 1895 1896 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && 1897 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 1898 !ca_ops->cong_control) 1899 tcp_cwnd_application_limited(sk); 1900 1901 /* The following conditions together indicate the starvation 1902 * is caused by insufficient sender buffer: 1903 * 1) just sent some data (see tcp_write_xmit) 1904 * 2) not cwnd limited (this else condition) 1905 * 3) no more data to send (tcp_write_queue_empty()) 1906 * 4) application is hitting buffer limit (SOCK_NOSPACE) 1907 */ 1908 if (tcp_write_queue_empty(sk) && sk->sk_socket && 1909 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1910 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1911 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1912 } 1913 } 1914 1915 /* Minshall's variant of the Nagle send check. */ 1916 static bool tcp_minshall_check(const struct tcp_sock *tp) 1917 { 1918 return after(tp->snd_sml, tp->snd_una) && 1919 !after(tp->snd_sml, tp->snd_nxt); 1920 } 1921 1922 /* Update snd_sml if this skb is under mss 1923 * Note that a TSO packet might end with a sub-mss segment 1924 * The test is really : 1925 * if ((skb->len % mss) != 0) 1926 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1927 * But we can avoid doing the divide again given we already have 1928 * skb_pcount = skb->len / mss_now 1929 */ 1930 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1931 const struct sk_buff *skb) 1932 { 1933 if (skb->len < tcp_skb_pcount(skb) * mss_now) 1934 tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1935 } 1936 1937 /* Return false, if packet can be sent now without violation Nagle's rules: 1938 * 1. It is full sized. (provided by caller in %partial bool) 1939 * 2. Or it contains FIN. (already checked by caller) 1940 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1941 * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1942 * With Minshall's modification: all sent small packets are ACKed. 1943 */ 1944 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1945 int nonagle) 1946 { 1947 return partial && 1948 ((nonagle & TCP_NAGLE_CORK) || 1949 (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1950 } 1951 1952 /* Return how many segs we'd like on a TSO packet, 1953 * to send one TSO packet per ms 1954 */ 1955 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 1956 int min_tso_segs) 1957 { 1958 u32 bytes, segs; 1959 1960 bytes = min_t(unsigned long, 1961 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift), 1962 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1963 1964 /* Goal is to send at least one packet per ms, 1965 * not one big TSO packet every 100 ms. 1966 * This preserves ACK clocking and is consistent 1967 * with tcp_tso_should_defer() heuristic. 1968 */ 1969 segs = max_t(u32, bytes / mss_now, min_tso_segs); 1970 1971 return segs; 1972 } 1973 1974 /* Return the number of segments we want in the skb we are transmitting. 1975 * See if congestion control module wants to decide; otherwise, autosize. 1976 */ 1977 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1978 { 1979 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1980 u32 min_tso, tso_segs; 1981 1982 min_tso = ca_ops->min_tso_segs ? 1983 ca_ops->min_tso_segs(sk) : 1984 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs; 1985 1986 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 1987 return min_t(u32, tso_segs, sk->sk_gso_max_segs); 1988 } 1989 1990 /* Returns the portion of skb which can be sent right away */ 1991 static unsigned int tcp_mss_split_point(const struct sock *sk, 1992 const struct sk_buff *skb, 1993 unsigned int mss_now, 1994 unsigned int max_segs, 1995 int nonagle) 1996 { 1997 const struct tcp_sock *tp = tcp_sk(sk); 1998 u32 partial, needed, window, max_len; 1999 2000 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2001 max_len = mss_now * max_segs; 2002 2003 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 2004 return max_len; 2005 2006 needed = min(skb->len, window); 2007 2008 if (max_len <= needed) 2009 return max_len; 2010 2011 partial = needed % mss_now; 2012 /* If last segment is not a full MSS, check if Nagle rules allow us 2013 * to include this last segment in this skb. 2014 * Otherwise, we'll split the skb at last MSS boundary 2015 */ 2016 if (tcp_nagle_check(partial != 0, tp, nonagle)) 2017 return needed - partial; 2018 2019 return needed; 2020 } 2021 2022 /* Can at least one segment of SKB be sent right now, according to the 2023 * congestion window rules? If so, return how many segments are allowed. 2024 */ 2025 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 2026 const struct sk_buff *skb) 2027 { 2028 u32 in_flight, cwnd, halfcwnd; 2029 2030 /* Don't be strict about the congestion window for the final FIN. */ 2031 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 2032 tcp_skb_pcount(skb) == 1) 2033 return 1; 2034 2035 in_flight = tcp_packets_in_flight(tp); 2036 cwnd = tp->snd_cwnd; 2037 if (in_flight >= cwnd) 2038 return 0; 2039 2040 /* For better scheduling, ensure we have at least 2041 * 2 GSO packets in flight. 2042 */ 2043 halfcwnd = max(cwnd >> 1, 1U); 2044 return min(halfcwnd, cwnd - in_flight); 2045 } 2046 2047 /* Initialize TSO state of a skb. 2048 * This must be invoked the first time we consider transmitting 2049 * SKB onto the wire. 2050 */ 2051 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 2052 { 2053 int tso_segs = tcp_skb_pcount(skb); 2054 2055 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 2056 tcp_set_skb_tso_segs(skb, mss_now); 2057 tso_segs = tcp_skb_pcount(skb); 2058 } 2059 return tso_segs; 2060 } 2061 2062 2063 /* Return true if the Nagle test allows this packet to be 2064 * sent now. 2065 */ 2066 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 2067 unsigned int cur_mss, int nonagle) 2068 { 2069 /* Nagle rule does not apply to frames, which sit in the middle of the 2070 * write_queue (they have no chances to get new data). 2071 * 2072 * This is implemented in the callers, where they modify the 'nonagle' 2073 * argument based upon the location of SKB in the send queue. 2074 */ 2075 if (nonagle & TCP_NAGLE_PUSH) 2076 return true; 2077 2078 /* Don't use the nagle rule for urgent data (or for the final FIN). */ 2079 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 2080 return true; 2081 2082 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 2083 return true; 2084 2085 return false; 2086 } 2087 2088 /* Does at least the first segment of SKB fit into the send window? */ 2089 static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 2090 const struct sk_buff *skb, 2091 unsigned int cur_mss) 2092 { 2093 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 2094 2095 if (skb->len > cur_mss) 2096 end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 2097 2098 return !after(end_seq, tcp_wnd_end(tp)); 2099 } 2100 2101 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 2102 * which is put after SKB on the list. It is very much like 2103 * tcp_fragment() except that it may make several kinds of assumptions 2104 * in order to speed up the splitting operation. In particular, we 2105 * know that all the data is in scatter-gather pages, and that the 2106 * packet has never been sent out before (and thus is not cloned). 2107 */ 2108 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 2109 unsigned int mss_now, gfp_t gfp) 2110 { 2111 int nlen = skb->len - len; 2112 struct sk_buff *buff; 2113 u8 flags; 2114 2115 /* All of a TSO frame must be composed of paged data. */ 2116 if (skb->len != skb->data_len) 2117 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 2118 skb, len, mss_now, gfp); 2119 2120 buff = tcp_stream_alloc_skb(sk, 0, gfp, true); 2121 if (unlikely(!buff)) 2122 return -ENOMEM; 2123 skb_copy_decrypted(buff, skb); 2124 mptcp_skb_ext_copy(buff, skb); 2125 2126 sk_wmem_queued_add(sk, buff->truesize); 2127 sk_mem_charge(sk, buff->truesize); 2128 buff->truesize += nlen; 2129 skb->truesize -= nlen; 2130 2131 /* Correct the sequence numbers. */ 2132 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 2133 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 2134 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 2135 2136 /* PSH and FIN should only be set in the second packet. */ 2137 flags = TCP_SKB_CB(skb)->tcp_flags; 2138 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 2139 TCP_SKB_CB(buff)->tcp_flags = flags; 2140 2141 tcp_skb_fragment_eor(skb, buff); 2142 2143 skb_split(skb, buff, len); 2144 tcp_fragment_tstamp(skb, buff); 2145 2146 /* Fix up tso_factor for both original and new SKB. */ 2147 tcp_set_skb_tso_segs(skb, mss_now); 2148 tcp_set_skb_tso_segs(buff, mss_now); 2149 2150 /* Link BUFF into the send queue. */ 2151 __skb_header_release(buff); 2152 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 2153 2154 return 0; 2155 } 2156 2157 /* Try to defer sending, if possible, in order to minimize the amount 2158 * of TSO splitting we do. View it as a kind of TSO Nagle test. 2159 * 2160 * This algorithm is from John Heffner. 2161 */ 2162 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 2163 bool *is_cwnd_limited, 2164 bool *is_rwnd_limited, 2165 u32 max_segs) 2166 { 2167 const struct inet_connection_sock *icsk = inet_csk(sk); 2168 u32 send_win, cong_win, limit, in_flight; 2169 struct tcp_sock *tp = tcp_sk(sk); 2170 struct sk_buff *head; 2171 int win_divisor; 2172 s64 delta; 2173 2174 if (icsk->icsk_ca_state >= TCP_CA_Recovery) 2175 goto send_now; 2176 2177 /* Avoid bursty behavior by allowing defer 2178 * only if the last write was recent (1 ms). 2179 * Note that tp->tcp_wstamp_ns can be in the future if we have 2180 * packets waiting in a qdisc or device for EDT delivery. 2181 */ 2182 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 2183 if (delta > 0) 2184 goto send_now; 2185 2186 in_flight = tcp_packets_in_flight(tp); 2187 2188 BUG_ON(tcp_skb_pcount(skb) <= 1); 2189 BUG_ON(tp->snd_cwnd <= in_flight); 2190 2191 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2192 2193 /* From in_flight test above, we know that cwnd > in_flight. */ 2194 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 2195 2196 limit = min(send_win, cong_win); 2197 2198 /* If a full-sized TSO skb can be sent, do it. */ 2199 if (limit >= max_segs * tp->mss_cache) 2200 goto send_now; 2201 2202 /* Middle in queue won't get any more data, full sendable already? */ 2203 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 2204 goto send_now; 2205 2206 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 2207 if (win_divisor) { 2208 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 2209 2210 /* If at least some fraction of a window is available, 2211 * just use it. 2212 */ 2213 chunk /= win_divisor; 2214 if (limit >= chunk) 2215 goto send_now; 2216 } else { 2217 /* Different approach, try not to defer past a single 2218 * ACK. Receiver should ACK every other full sized 2219 * frame, so if we have space for more than 3 frames 2220 * then send now. 2221 */ 2222 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 2223 goto send_now; 2224 } 2225 2226 /* TODO : use tsorted_sent_queue ? */ 2227 head = tcp_rtx_queue_head(sk); 2228 if (!head) 2229 goto send_now; 2230 delta = tp->tcp_clock_cache - head->tstamp; 2231 /* If next ACK is likely to come too late (half srtt), do not defer */ 2232 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 2233 goto send_now; 2234 2235 /* Ok, it looks like it is advisable to defer. 2236 * Three cases are tracked : 2237 * 1) We are cwnd-limited 2238 * 2) We are rwnd-limited 2239 * 3) We are application limited. 2240 */ 2241 if (cong_win < send_win) { 2242 if (cong_win <= skb->len) { 2243 *is_cwnd_limited = true; 2244 return true; 2245 } 2246 } else { 2247 if (send_win <= skb->len) { 2248 *is_rwnd_limited = true; 2249 return true; 2250 } 2251 } 2252 2253 /* If this packet won't get more data, do not wait. */ 2254 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 2255 TCP_SKB_CB(skb)->eor) 2256 goto send_now; 2257 2258 return true; 2259 2260 send_now: 2261 return false; 2262 } 2263 2264 static inline void tcp_mtu_check_reprobe(struct sock *sk) 2265 { 2266 struct inet_connection_sock *icsk = inet_csk(sk); 2267 struct tcp_sock *tp = tcp_sk(sk); 2268 struct net *net = sock_net(sk); 2269 u32 interval; 2270 s32 delta; 2271 2272 interval = net->ipv4.sysctl_tcp_probe_interval; 2273 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 2274 if (unlikely(delta >= interval * HZ)) { 2275 int mss = tcp_current_mss(sk); 2276 2277 /* Update current search range */ 2278 icsk->icsk_mtup.probe_size = 0; 2279 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 2280 sizeof(struct tcphdr) + 2281 icsk->icsk_af_ops->net_header_len; 2282 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 2283 2284 /* Update probe time stamp */ 2285 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 2286 } 2287 } 2288 2289 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2290 { 2291 struct sk_buff *skb, *next; 2292 2293 skb = tcp_send_head(sk); 2294 tcp_for_write_queue_from_safe(skb, next, sk) { 2295 if (len <= skb->len) 2296 break; 2297 2298 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb)) 2299 return false; 2300 2301 len -= skb->len; 2302 } 2303 2304 return true; 2305 } 2306 2307 /* Create a new MTU probe if we are ready. 2308 * MTU probe is regularly attempting to increase the path MTU by 2309 * deliberately sending larger packets. This discovers routing 2310 * changes resulting in larger path MTUs. 2311 * 2312 * Returns 0 if we should wait to probe (no cwnd available), 2313 * 1 if a probe was sent, 2314 * -1 otherwise 2315 */ 2316 static int tcp_mtu_probe(struct sock *sk) 2317 { 2318 struct inet_connection_sock *icsk = inet_csk(sk); 2319 struct tcp_sock *tp = tcp_sk(sk); 2320 struct sk_buff *skb, *nskb, *next; 2321 struct net *net = sock_net(sk); 2322 int probe_size; 2323 int size_needed; 2324 int copy, len; 2325 int mss_now; 2326 int interval; 2327 2328 /* Not currently probing/verifying, 2329 * not in recovery, 2330 * have enough cwnd, and 2331 * not SACKing (the variable headers throw things off) 2332 */ 2333 if (likely(!icsk->icsk_mtup.enabled || 2334 icsk->icsk_mtup.probe_size || 2335 inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 2336 tp->snd_cwnd < 11 || 2337 tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 2338 return -1; 2339 2340 /* Use binary search for probe_size between tcp_mss_base, 2341 * and current mss_clamp. if (search_high - search_low) 2342 * smaller than a threshold, backoff from probing. 2343 */ 2344 mss_now = tcp_current_mss(sk); 2345 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 2346 icsk->icsk_mtup.search_low) >> 1); 2347 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 2348 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 2349 /* When misfortune happens, we are reprobing actively, 2350 * and then reprobe timer has expired. We stick with current 2351 * probing process by not resetting search range to its orignal. 2352 */ 2353 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 2354 interval < net->ipv4.sysctl_tcp_probe_threshold) { 2355 /* Check whether enough time has elaplased for 2356 * another round of probing. 2357 */ 2358 tcp_mtu_check_reprobe(sk); 2359 return -1; 2360 } 2361 2362 /* Have enough data in the send queue to probe? */ 2363 if (tp->write_seq - tp->snd_nxt < size_needed) 2364 return -1; 2365 2366 if (tp->snd_wnd < size_needed) 2367 return -1; 2368 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 2369 return 0; 2370 2371 /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2372 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2373 if (!tcp_packets_in_flight(tp)) 2374 return -1; 2375 else 2376 return 0; 2377 } 2378 2379 if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2380 return -1; 2381 2382 /* We're allowed to probe. Build it now. */ 2383 nskb = tcp_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 2384 if (!nskb) 2385 return -1; 2386 sk_wmem_queued_add(sk, nskb->truesize); 2387 sk_mem_charge(sk, nskb->truesize); 2388 2389 skb = tcp_send_head(sk); 2390 skb_copy_decrypted(nskb, skb); 2391 mptcp_skb_ext_copy(nskb, skb); 2392 2393 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 2394 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 2395 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 2396 2397 tcp_insert_write_queue_before(nskb, skb, sk); 2398 tcp_highest_sack_replace(sk, skb, nskb); 2399 2400 len = 0; 2401 tcp_for_write_queue_from_safe(skb, next, sk) { 2402 copy = min_t(int, skb->len, probe_size - len); 2403 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 2404 2405 if (skb->len <= copy) { 2406 /* We've eaten all the data from this skb. 2407 * Throw it away. */ 2408 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2409 /* If this is the last SKB we copy and eor is set 2410 * we need to propagate it to the new skb. 2411 */ 2412 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2413 tcp_skb_collapse_tstamp(nskb, skb); 2414 tcp_unlink_write_queue(skb, sk); 2415 tcp_wmem_free_skb(sk, skb); 2416 } else { 2417 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2418 ~(TCPHDR_FIN|TCPHDR_PSH); 2419 if (!skb_shinfo(skb)->nr_frags) { 2420 skb_pull(skb, copy); 2421 } else { 2422 __pskb_trim_head(skb, copy); 2423 tcp_set_skb_tso_segs(skb, mss_now); 2424 } 2425 TCP_SKB_CB(skb)->seq += copy; 2426 } 2427 2428 len += copy; 2429 2430 if (len >= probe_size) 2431 break; 2432 } 2433 tcp_init_tso_segs(nskb, nskb->len); 2434 2435 /* We're ready to send. If this fails, the probe will 2436 * be resegmented into mss-sized pieces by tcp_write_xmit(). 2437 */ 2438 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 2439 /* Decrement cwnd here because we are sending 2440 * effectively two packets. */ 2441 tp->snd_cwnd--; 2442 tcp_event_new_data_sent(sk, nskb); 2443 2444 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 2445 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 2446 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 2447 2448 return 1; 2449 } 2450 2451 return -1; 2452 } 2453 2454 static bool tcp_pacing_check(struct sock *sk) 2455 { 2456 struct tcp_sock *tp = tcp_sk(sk); 2457 2458 if (!tcp_needs_internal_pacing(sk)) 2459 return false; 2460 2461 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2462 return false; 2463 2464 if (!hrtimer_is_queued(&tp->pacing_timer)) { 2465 hrtimer_start(&tp->pacing_timer, 2466 ns_to_ktime(tp->tcp_wstamp_ns), 2467 HRTIMER_MODE_ABS_PINNED_SOFT); 2468 sock_hold(sk); 2469 } 2470 return true; 2471 } 2472 2473 /* TCP Small Queues : 2474 * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2475 * (These limits are doubled for retransmits) 2476 * This allows for : 2477 * - better RTT estimation and ACK scheduling 2478 * - faster recovery 2479 * - high rates 2480 * Alas, some drivers / subsystems require a fair amount 2481 * of queued bytes to ensure line rate. 2482 * One example is wifi aggregation (802.11 AMPDU) 2483 */ 2484 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2485 unsigned int factor) 2486 { 2487 unsigned long limit; 2488 2489 limit = max_t(unsigned long, 2490 2 * skb->truesize, 2491 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift)); 2492 if (sk->sk_pacing_status == SK_PACING_NONE) 2493 limit = min_t(unsigned long, limit, 2494 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); 2495 limit <<= factor; 2496 2497 if (static_branch_unlikely(&tcp_tx_delay_enabled) && 2498 tcp_sk(sk)->tcp_tx_delay) { 2499 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay; 2500 2501 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we 2502 * approximate our needs assuming an ~100% skb->truesize overhead. 2503 * USEC_PER_SEC is approximated by 2^20. 2504 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. 2505 */ 2506 extra_bytes >>= (20 - 1); 2507 limit += extra_bytes; 2508 } 2509 if (refcount_read(&sk->sk_wmem_alloc) > limit) { 2510 /* Always send skb if rtx queue is empty. 2511 * No need to wait for TX completion to call us back, 2512 * after softirq/tasklet schedule. 2513 * This helps when TX completions are delayed too much. 2514 */ 2515 if (tcp_rtx_queue_empty(sk)) 2516 return false; 2517 2518 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2519 /* It is possible TX completion already happened 2520 * before we set TSQ_THROTTLED, so we must 2521 * test again the condition. 2522 */ 2523 smp_mb__after_atomic(); 2524 if (refcount_read(&sk->sk_wmem_alloc) > limit) 2525 return true; 2526 } 2527 return false; 2528 } 2529 2530 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 2531 { 2532 const u32 now = tcp_jiffies32; 2533 enum tcp_chrono old = tp->chrono_type; 2534 2535 if (old > TCP_CHRONO_UNSPEC) 2536 tp->chrono_stat[old - 1] += now - tp->chrono_start; 2537 tp->chrono_start = now; 2538 tp->chrono_type = new; 2539 } 2540 2541 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 2542 { 2543 struct tcp_sock *tp = tcp_sk(sk); 2544 2545 /* If there are multiple conditions worthy of tracking in a 2546 * chronograph then the highest priority enum takes precedence 2547 * over the other conditions. So that if something "more interesting" 2548 * starts happening, stop the previous chrono and start a new one. 2549 */ 2550 if (type > tp->chrono_type) 2551 tcp_chrono_set(tp, type); 2552 } 2553 2554 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 2555 { 2556 struct tcp_sock *tp = tcp_sk(sk); 2557 2558 2559 /* There are multiple conditions worthy of tracking in a 2560 * chronograph, so that the highest priority enum takes 2561 * precedence over the other conditions (see tcp_chrono_start). 2562 * If a condition stops, we only stop chrono tracking if 2563 * it's the "most interesting" or current chrono we are 2564 * tracking and starts busy chrono if we have pending data. 2565 */ 2566 if (tcp_rtx_and_write_queues_empty(sk)) 2567 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 2568 else if (type == tp->chrono_type) 2569 tcp_chrono_set(tp, TCP_CHRONO_BUSY); 2570 } 2571 2572 /* This routine writes packets to the network. It advances the 2573 * send_head. This happens as incoming acks open up the remote 2574 * window for us. 2575 * 2576 * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2577 * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2578 * account rare use of URG, this is not a big flaw. 2579 * 2580 * Send at most one packet when push_one > 0. Temporarily ignore 2581 * cwnd limit to force at most one packet out when push_one == 2. 2582 2583 * Returns true, if no segments are in flight and we have queued segments, 2584 * but cannot send anything now because of SWS or another problem. 2585 */ 2586 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2587 int push_one, gfp_t gfp) 2588 { 2589 struct tcp_sock *tp = tcp_sk(sk); 2590 struct sk_buff *skb; 2591 unsigned int tso_segs, sent_pkts; 2592 int cwnd_quota; 2593 int result; 2594 bool is_cwnd_limited = false, is_rwnd_limited = false; 2595 u32 max_segs; 2596 2597 sent_pkts = 0; 2598 2599 tcp_mstamp_refresh(tp); 2600 if (!push_one) { 2601 /* Do MTU probing. */ 2602 result = tcp_mtu_probe(sk); 2603 if (!result) { 2604 return false; 2605 } else if (result > 0) { 2606 sent_pkts = 1; 2607 } 2608 } 2609 2610 max_segs = tcp_tso_segs(sk, mss_now); 2611 while ((skb = tcp_send_head(sk))) { 2612 unsigned int limit; 2613 2614 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 2615 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 2616 skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache; 2617 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2618 tcp_init_tso_segs(skb, mss_now); 2619 goto repair; /* Skip network transmission */ 2620 } 2621 2622 if (tcp_pacing_check(sk)) 2623 break; 2624 2625 tso_segs = tcp_init_tso_segs(skb, mss_now); 2626 BUG_ON(!tso_segs); 2627 2628 cwnd_quota = tcp_cwnd_test(tp, skb); 2629 if (!cwnd_quota) { 2630 if (push_one == 2) 2631 /* Force out a loss probe pkt. */ 2632 cwnd_quota = 1; 2633 else 2634 break; 2635 } 2636 2637 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 2638 is_rwnd_limited = true; 2639 break; 2640 } 2641 2642 if (tso_segs == 1) { 2643 if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2644 (tcp_skb_is_last(sk, skb) ? 2645 nonagle : TCP_NAGLE_PUSH)))) 2646 break; 2647 } else { 2648 if (!push_one && 2649 tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2650 &is_rwnd_limited, max_segs)) 2651 break; 2652 } 2653 2654 limit = mss_now; 2655 if (tso_segs > 1 && !tcp_urg_mode(tp)) 2656 limit = tcp_mss_split_point(sk, skb, mss_now, 2657 min_t(unsigned int, 2658 cwnd_quota, 2659 max_segs), 2660 nonagle); 2661 2662 if (skb->len > limit && 2663 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2664 break; 2665 2666 if (tcp_small_queue_check(sk, skb, 0)) 2667 break; 2668 2669 /* Argh, we hit an empty skb(), presumably a thread 2670 * is sleeping in sendmsg()/sk_stream_wait_memory(). 2671 * We do not want to send a pure-ack packet and have 2672 * a strange looking rtx queue with empty packet(s). 2673 */ 2674 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) 2675 break; 2676 2677 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 2678 break; 2679 2680 repair: 2681 /* Advance the send_head. This one is sent out. 2682 * This call will increment packets_out. 2683 */ 2684 tcp_event_new_data_sent(sk, skb); 2685 2686 tcp_minshall_update(tp, mss_now, skb); 2687 sent_pkts += tcp_skb_pcount(skb); 2688 2689 if (push_one) 2690 break; 2691 } 2692 2693 if (is_rwnd_limited) 2694 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 2695 else 2696 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 2697 2698 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2699 if (likely(sent_pkts || is_cwnd_limited)) 2700 tcp_cwnd_validate(sk, is_cwnd_limited); 2701 2702 if (likely(sent_pkts)) { 2703 if (tcp_in_cwnd_reduction(sk)) 2704 tp->prr_out += sent_pkts; 2705 2706 /* Send one loss probe per tail loss episode. */ 2707 if (push_one != 2) 2708 tcp_schedule_loss_probe(sk, false); 2709 return false; 2710 } 2711 return !tp->packets_out && !tcp_write_queue_empty(sk); 2712 } 2713 2714 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 2715 { 2716 struct inet_connection_sock *icsk = inet_csk(sk); 2717 struct tcp_sock *tp = tcp_sk(sk); 2718 u32 timeout, rto_delta_us; 2719 int early_retrans; 2720 2721 /* Don't do any loss probe on a Fast Open connection before 3WHS 2722 * finishes. 2723 */ 2724 if (rcu_access_pointer(tp->fastopen_rsk)) 2725 return false; 2726 2727 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 2728 /* Schedule a loss probe in 2*RTT for SACK capable connections 2729 * not in loss recovery, that are either limited by cwnd or application. 2730 */ 2731 if ((early_retrans != 3 && early_retrans != 4) || 2732 !tp->packets_out || !tcp_is_sack(tp) || 2733 (icsk->icsk_ca_state != TCP_CA_Open && 2734 icsk->icsk_ca_state != TCP_CA_CWR)) 2735 return false; 2736 2737 /* Probe timeout is 2*rtt. Add minimum RTO to account 2738 * for delayed ack when there's one outstanding packet. If no RTT 2739 * sample is available then probe after TCP_TIMEOUT_INIT. 2740 */ 2741 if (tp->srtt_us) { 2742 timeout = usecs_to_jiffies(tp->srtt_us >> 2); 2743 if (tp->packets_out == 1) 2744 timeout += TCP_RTO_MIN; 2745 else 2746 timeout += TCP_TIMEOUT_MIN; 2747 } else { 2748 timeout = TCP_TIMEOUT_INIT; 2749 } 2750 2751 /* If the RTO formula yields an earlier time, then use that time. */ 2752 rto_delta_us = advancing_rto ? 2753 jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2754 tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2755 if (rto_delta_us > 0) 2756 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 2757 2758 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); 2759 return true; 2760 } 2761 2762 /* Thanks to skb fast clones, we can detect if a prior transmit of 2763 * a packet is still in a qdisc or driver queue. 2764 * In this case, there is very little point doing a retransmit ! 2765 */ 2766 static bool skb_still_in_host_queue(struct sock *sk, 2767 const struct sk_buff *skb) 2768 { 2769 if (unlikely(skb_fclone_busy(sk, skb))) { 2770 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2771 smp_mb__after_atomic(); 2772 if (skb_fclone_busy(sk, skb)) { 2773 NET_INC_STATS(sock_net(sk), 2774 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 2775 return true; 2776 } 2777 } 2778 return false; 2779 } 2780 2781 /* When probe timeout (PTO) fires, try send a new segment if possible, else 2782 * retransmit the last segment. 2783 */ 2784 void tcp_send_loss_probe(struct sock *sk) 2785 { 2786 struct tcp_sock *tp = tcp_sk(sk); 2787 struct sk_buff *skb; 2788 int pcount; 2789 int mss = tcp_current_mss(sk); 2790 2791 /* At most one outstanding TLP */ 2792 if (tp->tlp_high_seq) 2793 goto rearm_timer; 2794 2795 tp->tlp_retrans = 0; 2796 skb = tcp_send_head(sk); 2797 if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2798 pcount = tp->packets_out; 2799 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2800 if (tp->packets_out > pcount) 2801 goto probe_sent; 2802 goto rearm_timer; 2803 } 2804 skb = skb_rb_last(&sk->tcp_rtx_queue); 2805 if (unlikely(!skb)) { 2806 WARN_ONCE(tp->packets_out, 2807 "invalid inflight: %u state %u cwnd %u mss %d\n", 2808 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss); 2809 inet_csk(sk)->icsk_pending = 0; 2810 return; 2811 } 2812 2813 if (skb_still_in_host_queue(sk, skb)) 2814 goto rearm_timer; 2815 2816 pcount = tcp_skb_pcount(skb); 2817 if (WARN_ON(!pcount)) 2818 goto rearm_timer; 2819 2820 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 2821 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 2822 (pcount - 1) * mss, mss, 2823 GFP_ATOMIC))) 2824 goto rearm_timer; 2825 skb = skb_rb_next(skb); 2826 } 2827 2828 if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 2829 goto rearm_timer; 2830 2831 if (__tcp_retransmit_skb(sk, skb, 1)) 2832 goto rearm_timer; 2833 2834 tp->tlp_retrans = 1; 2835 2836 probe_sent: 2837 /* Record snd_nxt for loss detection. */ 2838 tp->tlp_high_seq = tp->snd_nxt; 2839 2840 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2841 /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2842 inet_csk(sk)->icsk_pending = 0; 2843 rearm_timer: 2844 tcp_rearm_rto(sk); 2845 } 2846 2847 /* Push out any pending frames which were held back due to 2848 * TCP_CORK or attempt at coalescing tiny packets. 2849 * The socket must be locked by the caller. 2850 */ 2851 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 2852 int nonagle) 2853 { 2854 /* If we are closed, the bytes will have to remain here. 2855 * In time closedown will finish, we empty the write queue and 2856 * all will be happy. 2857 */ 2858 if (unlikely(sk->sk_state == TCP_CLOSE)) 2859 return; 2860 2861 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 2862 sk_gfp_mask(sk, GFP_ATOMIC))) 2863 tcp_check_probe_timer(sk); 2864 } 2865 2866 /* Send _single_ skb sitting at the send head. This function requires 2867 * true push pending frames to setup probe timer etc. 2868 */ 2869 void tcp_push_one(struct sock *sk, unsigned int mss_now) 2870 { 2871 struct sk_buff *skb = tcp_send_head(sk); 2872 2873 BUG_ON(!skb || skb->len < mss_now); 2874 2875 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2876 } 2877 2878 /* This function returns the amount that we can raise the 2879 * usable window based on the following constraints 2880 * 2881 * 1. The window can never be shrunk once it is offered (RFC 793) 2882 * 2. We limit memory per socket 2883 * 2884 * RFC 1122: 2885 * "the suggested [SWS] avoidance algorithm for the receiver is to keep 2886 * RECV.NEXT + RCV.WIN fixed until: 2887 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 2888 * 2889 * i.e. don't raise the right edge of the window until you can raise 2890 * it at least MSS bytes. 2891 * 2892 * Unfortunately, the recommended algorithm breaks header prediction, 2893 * since header prediction assumes th->window stays fixed. 2894 * 2895 * Strictly speaking, keeping th->window fixed violates the receiver 2896 * side SWS prevention criteria. The problem is that under this rule 2897 * a stream of single byte packets will cause the right side of the 2898 * window to always advance by a single byte. 2899 * 2900 * Of course, if the sender implements sender side SWS prevention 2901 * then this will not be a problem. 2902 * 2903 * BSD seems to make the following compromise: 2904 * 2905 * If the free space is less than the 1/4 of the maximum 2906 * space available and the free space is less than 1/2 mss, 2907 * then set the window to 0. 2908 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 2909 * Otherwise, just prevent the window from shrinking 2910 * and from being larger than the largest representable value. 2911 * 2912 * This prevents incremental opening of the window in the regime 2913 * where TCP is limited by the speed of the reader side taking 2914 * data out of the TCP receive queue. It does nothing about 2915 * those cases where the window is constrained on the sender side 2916 * because the pipeline is full. 2917 * 2918 * BSD also seems to "accidentally" limit itself to windows that are a 2919 * multiple of MSS, at least until the free space gets quite small. 2920 * This would appear to be a side effect of the mbuf implementation. 2921 * Combining these two algorithms results in the observed behavior 2922 * of having a fixed window size at almost all times. 2923 * 2924 * Below we obtain similar behavior by forcing the offered window to 2925 * a multiple of the mss when it is feasible to do so. 2926 * 2927 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 2928 * Regular options like TIMESTAMP are taken into account. 2929 */ 2930 u32 __tcp_select_window(struct sock *sk) 2931 { 2932 struct inet_connection_sock *icsk = inet_csk(sk); 2933 struct tcp_sock *tp = tcp_sk(sk); 2934 /* MSS for the peer's data. Previous versions used mss_clamp 2935 * here. I don't know if the value based on our guesses 2936 * of peer's MSS is better for the performance. It's more correct 2937 * but may be worse for the performance because of rcv_mss 2938 * fluctuations. --SAW 1998/11/1 2939 */ 2940 int mss = icsk->icsk_ack.rcv_mss; 2941 int free_space = tcp_space(sk); 2942 int allowed_space = tcp_full_space(sk); 2943 int full_space, window; 2944 2945 if (sk_is_mptcp(sk)) 2946 mptcp_space(sk, &free_space, &allowed_space); 2947 2948 full_space = min_t(int, tp->window_clamp, allowed_space); 2949 2950 if (unlikely(mss > full_space)) { 2951 mss = full_space; 2952 if (mss <= 0) 2953 return 0; 2954 } 2955 if (free_space < (full_space >> 1)) { 2956 icsk->icsk_ack.quick = 0; 2957 2958 if (tcp_under_memory_pressure(sk)) 2959 tcp_adjust_rcv_ssthresh(sk); 2960 2961 /* free_space might become our new window, make sure we don't 2962 * increase it due to wscale. 2963 */ 2964 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 2965 2966 /* if free space is less than mss estimate, or is below 1/16th 2967 * of the maximum allowed, try to move to zero-window, else 2968 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 2969 * new incoming data is dropped due to memory limits. 2970 * With large window, mss test triggers way too late in order 2971 * to announce zero window in time before rmem limit kicks in. 2972 */ 2973 if (free_space < (allowed_space >> 4) || free_space < mss) 2974 return 0; 2975 } 2976 2977 if (free_space > tp->rcv_ssthresh) 2978 free_space = tp->rcv_ssthresh; 2979 2980 /* Don't do rounding if we are using window scaling, since the 2981 * scaled window will not line up with the MSS boundary anyway. 2982 */ 2983 if (tp->rx_opt.rcv_wscale) { 2984 window = free_space; 2985 2986 /* Advertise enough space so that it won't get scaled away. 2987 * Import case: prevent zero window announcement if 2988 * 1<<rcv_wscale > mss. 2989 */ 2990 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 2991 } else { 2992 window = tp->rcv_wnd; 2993 /* Get the largest window that is a nice multiple of mss. 2994 * Window clamp already applied above. 2995 * If our current window offering is within 1 mss of the 2996 * free space we just keep it. This prevents the divide 2997 * and multiply from happening most of the time. 2998 * We also don't do any window rounding when the free space 2999 * is too small. 3000 */ 3001 if (window <= free_space - mss || window > free_space) 3002 window = rounddown(free_space, mss); 3003 else if (mss == full_space && 3004 free_space > window + (full_space >> 1)) 3005 window = free_space; 3006 } 3007 3008 return window; 3009 } 3010 3011 void tcp_skb_collapse_tstamp(struct sk_buff *skb, 3012 const struct sk_buff *next_skb) 3013 { 3014 if (unlikely(tcp_has_tx_tstamp(next_skb))) { 3015 const struct skb_shared_info *next_shinfo = 3016 skb_shinfo(next_skb); 3017 struct skb_shared_info *shinfo = skb_shinfo(skb); 3018 3019 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 3020 shinfo->tskey = next_shinfo->tskey; 3021 TCP_SKB_CB(skb)->txstamp_ack |= 3022 TCP_SKB_CB(next_skb)->txstamp_ack; 3023 } 3024 } 3025 3026 /* Collapses two adjacent SKB's during retransmission. */ 3027 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 3028 { 3029 struct tcp_sock *tp = tcp_sk(sk); 3030 struct sk_buff *next_skb = skb_rb_next(skb); 3031 int next_skb_size; 3032 3033 next_skb_size = next_skb->len; 3034 3035 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 3036 3037 if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size)) 3038 return false; 3039 3040 tcp_highest_sack_replace(sk, next_skb, skb); 3041 3042 /* Update sequence range on original skb. */ 3043 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 3044 3045 /* Merge over control information. This moves PSH/FIN etc. over */ 3046 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 3047 3048 /* All done, get rid of second SKB and account for it so 3049 * packet counting does not break. 3050 */ 3051 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 3052 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 3053 3054 /* changed transmit queue under us so clear hints */ 3055 tcp_clear_retrans_hints_partial(tp); 3056 if (next_skb == tp->retransmit_skb_hint) 3057 tp->retransmit_skb_hint = skb; 3058 3059 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 3060 3061 tcp_skb_collapse_tstamp(skb, next_skb); 3062 3063 tcp_rtx_queue_unlink_and_free(next_skb, sk); 3064 return true; 3065 } 3066 3067 /* Check if coalescing SKBs is legal. */ 3068 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 3069 { 3070 if (tcp_skb_pcount(skb) > 1) 3071 return false; 3072 if (skb_cloned(skb)) 3073 return false; 3074 /* Some heuristics for collapsing over SACK'd could be invented */ 3075 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 3076 return false; 3077 3078 return true; 3079 } 3080 3081 /* Collapse packets in the retransmit queue to make to create 3082 * less packets on the wire. This is only done on retransmission. 3083 */ 3084 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 3085 int space) 3086 { 3087 struct tcp_sock *tp = tcp_sk(sk); 3088 struct sk_buff *skb = to, *tmp; 3089 bool first = true; 3090 3091 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) 3092 return; 3093 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 3094 return; 3095 3096 skb_rbtree_walk_from_safe(skb, tmp) { 3097 if (!tcp_can_collapse(sk, skb)) 3098 break; 3099 3100 if (!tcp_skb_can_collapse(to, skb)) 3101 break; 3102 3103 space -= skb->len; 3104 3105 if (first) { 3106 first = false; 3107 continue; 3108 } 3109 3110 if (space < 0) 3111 break; 3112 3113 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 3114 break; 3115 3116 if (!tcp_collapse_retrans(sk, to)) 3117 break; 3118 } 3119 } 3120 3121 /* This retransmits one SKB. Policy decisions and retransmit queue 3122 * state updates are done by the caller. Returns non-zero if an 3123 * error occurred which prevented the send. 3124 */ 3125 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 3126 { 3127 struct inet_connection_sock *icsk = inet_csk(sk); 3128 struct tcp_sock *tp = tcp_sk(sk); 3129 unsigned int cur_mss; 3130 int diff, len, err; 3131 3132 3133 /* Inconclusive MTU probe */ 3134 if (icsk->icsk_mtup.probe_size) 3135 icsk->icsk_mtup.probe_size = 0; 3136 3137 if (skb_still_in_host_queue(sk, skb)) 3138 return -EBUSY; 3139 3140 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 3141 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 3142 WARN_ON_ONCE(1); 3143 return -EINVAL; 3144 } 3145 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 3146 return -ENOMEM; 3147 } 3148 3149 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 3150 return -EHOSTUNREACH; /* Routing failure or similar. */ 3151 3152 cur_mss = tcp_current_mss(sk); 3153 3154 /* If receiver has shrunk his window, and skb is out of 3155 * new window, do not retransmit it. The exception is the 3156 * case, when window is shrunk to zero. In this case 3157 * our retransmit serves as a zero window probe. 3158 */ 3159 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 3160 TCP_SKB_CB(skb)->seq != tp->snd_una) 3161 return -EAGAIN; 3162 3163 len = cur_mss * segs; 3164 if (skb->len > len) { 3165 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 3166 cur_mss, GFP_ATOMIC)) 3167 return -ENOMEM; /* We'll try again later. */ 3168 } else { 3169 if (skb_unclone(skb, GFP_ATOMIC)) 3170 return -ENOMEM; 3171 3172 diff = tcp_skb_pcount(skb); 3173 tcp_set_skb_tso_segs(skb, cur_mss); 3174 diff -= tcp_skb_pcount(skb); 3175 if (diff) 3176 tcp_adjust_pcount(sk, skb, diff); 3177 if (skb->len < cur_mss) 3178 tcp_retrans_try_collapse(sk, skb, cur_mss); 3179 } 3180 3181 /* RFC3168, section 6.1.1.1. ECN fallback */ 3182 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 3183 tcp_ecn_clear_syn(sk, skb); 3184 3185 /* Update global and local TCP statistics. */ 3186 segs = tcp_skb_pcount(skb); 3187 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 3188 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 3189 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 3190 tp->total_retrans += segs; 3191 tp->bytes_retrans += skb->len; 3192 3193 /* make sure skb->data is aligned on arches that require it 3194 * and check if ack-trimming & collapsing extended the headroom 3195 * beyond what csum_start can cover. 3196 */ 3197 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 3198 skb_headroom(skb) >= 0xFFFF)) { 3199 struct sk_buff *nskb; 3200 3201 tcp_skb_tsorted_save(skb) { 3202 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 3203 if (nskb) { 3204 nskb->dev = NULL; 3205 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); 3206 } else { 3207 err = -ENOBUFS; 3208 } 3209 } tcp_skb_tsorted_restore(skb); 3210 3211 if (!err) { 3212 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 3213 tcp_rate_skb_sent(sk, skb); 3214 } 3215 } else { 3216 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3217 } 3218 3219 /* To avoid taking spuriously low RTT samples based on a timestamp 3220 * for a transmit that never happened, always mark EVER_RETRANS 3221 */ 3222 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 3223 3224 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 3225 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 3226 TCP_SKB_CB(skb)->seq, segs, err); 3227 3228 if (likely(!err)) { 3229 trace_tcp_retransmit_skb(sk, skb); 3230 } else if (err != -EBUSY) { 3231 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 3232 } 3233 return err; 3234 } 3235 3236 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 3237 { 3238 struct tcp_sock *tp = tcp_sk(sk); 3239 int err = __tcp_retransmit_skb(sk, skb, segs); 3240 3241 if (err == 0) { 3242 #if FASTRETRANS_DEBUG > 0 3243 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 3244 net_dbg_ratelimited("retrans_out leaked\n"); 3245 } 3246 #endif 3247 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 3248 tp->retrans_out += tcp_skb_pcount(skb); 3249 } 3250 3251 /* Save stamp of the first (attempted) retransmit. */ 3252 if (!tp->retrans_stamp) 3253 tp->retrans_stamp = tcp_skb_timestamp(skb); 3254 3255 if (tp->undo_retrans < 0) 3256 tp->undo_retrans = 0; 3257 tp->undo_retrans += tcp_skb_pcount(skb); 3258 return err; 3259 } 3260 3261 /* This gets called after a retransmit timeout, and the initially 3262 * retransmitted data is acknowledged. It tries to continue 3263 * resending the rest of the retransmit queue, until either 3264 * we've sent it all or the congestion window limit is reached. 3265 */ 3266 void tcp_xmit_retransmit_queue(struct sock *sk) 3267 { 3268 const struct inet_connection_sock *icsk = inet_csk(sk); 3269 struct sk_buff *skb, *rtx_head, *hole = NULL; 3270 struct tcp_sock *tp = tcp_sk(sk); 3271 bool rearm_timer = false; 3272 u32 max_segs; 3273 int mib_idx; 3274 3275 if (!tp->packets_out) 3276 return; 3277 3278 rtx_head = tcp_rtx_queue_head(sk); 3279 skb = tp->retransmit_skb_hint ?: rtx_head; 3280 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 3281 skb_rbtree_walk_from(skb) { 3282 __u8 sacked; 3283 int segs; 3284 3285 if (tcp_pacing_check(sk)) 3286 break; 3287 3288 /* we could do better than to assign each time */ 3289 if (!hole) 3290 tp->retransmit_skb_hint = skb; 3291 3292 segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 3293 if (segs <= 0) 3294 break; 3295 sacked = TCP_SKB_CB(skb)->sacked; 3296 /* In case tcp_shift_skb_data() have aggregated large skbs, 3297 * we need to make sure not sending too bigs TSO packets 3298 */ 3299 segs = min_t(int, segs, max_segs); 3300 3301 if (tp->retrans_out >= tp->lost_out) { 3302 break; 3303 } else if (!(sacked & TCPCB_LOST)) { 3304 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 3305 hole = skb; 3306 continue; 3307 3308 } else { 3309 if (icsk->icsk_ca_state != TCP_CA_Loss) 3310 mib_idx = LINUX_MIB_TCPFASTRETRANS; 3311 else 3312 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 3313 } 3314 3315 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 3316 continue; 3317 3318 if (tcp_small_queue_check(sk, skb, 1)) 3319 break; 3320 3321 if (tcp_retransmit_skb(sk, skb, segs)) 3322 break; 3323 3324 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 3325 3326 if (tcp_in_cwnd_reduction(sk)) 3327 tp->prr_out += tcp_skb_pcount(skb); 3328 3329 if (skb == rtx_head && 3330 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 3331 rearm_timer = true; 3332 3333 } 3334 if (rearm_timer) 3335 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 3336 inet_csk(sk)->icsk_rto, 3337 TCP_RTO_MAX); 3338 } 3339 3340 /* We allow to exceed memory limits for FIN packets to expedite 3341 * connection tear down and (memory) recovery. 3342 * Otherwise tcp_send_fin() could be tempted to either delay FIN 3343 * or even be forced to close flow without any FIN. 3344 * In general, we want to allow one skb per socket to avoid hangs 3345 * with edge trigger epoll() 3346 */ 3347 void sk_forced_mem_schedule(struct sock *sk, int size) 3348 { 3349 int amt; 3350 3351 if (size <= sk->sk_forward_alloc) 3352 return; 3353 amt = sk_mem_pages(size); 3354 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3355 sk_memory_allocated_add(sk, amt); 3356 3357 if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3358 mem_cgroup_charge_skmem(sk->sk_memcg, amt, 3359 gfp_memcg_charge() | __GFP_NOFAIL); 3360 } 3361 3362 /* Send a FIN. The caller locks the socket for us. 3363 * We should try to send a FIN packet really hard, but eventually give up. 3364 */ 3365 void tcp_send_fin(struct sock *sk) 3366 { 3367 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); 3368 struct tcp_sock *tp = tcp_sk(sk); 3369 3370 /* Optimization, tack on the FIN if we have one skb in write queue and 3371 * this skb was not yet sent, or we are under memory pressure. 3372 * Note: in the latter case, FIN packet will be sent after a timeout, 3373 * as TCP stack thinks it has already been transmitted. 3374 */ 3375 tskb = tail; 3376 if (!tskb && tcp_under_memory_pressure(sk)) 3377 tskb = skb_rb_last(&sk->tcp_rtx_queue); 3378 3379 if (tskb) { 3380 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3381 TCP_SKB_CB(tskb)->end_seq++; 3382 tp->write_seq++; 3383 if (!tail) { 3384 /* This means tskb was already sent. 3385 * Pretend we included the FIN on previous transmit. 3386 * We need to set tp->snd_nxt to the value it would have 3387 * if FIN had been sent. This is because retransmit path 3388 * does not change tp->snd_nxt. 3389 */ 3390 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); 3391 return; 3392 } 3393 } else { 3394 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3395 if (unlikely(!skb)) 3396 return; 3397 3398 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3399 skb_reserve(skb, MAX_TCP_HEADER); 3400 sk_forced_mem_schedule(sk, skb->truesize); 3401 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3402 tcp_init_nondata_skb(skb, tp->write_seq, 3403 TCPHDR_ACK | TCPHDR_FIN); 3404 tcp_queue_skb(sk, skb); 3405 } 3406 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 3407 } 3408 3409 /* We get here when a process closes a file descriptor (either due to 3410 * an explicit close() or as a byproduct of exit()'ing) and there 3411 * was unread data in the receive queue. This behavior is recommended 3412 * by RFC 2525, section 2.17. -DaveM 3413 */ 3414 void tcp_send_active_reset(struct sock *sk, gfp_t priority) 3415 { 3416 struct sk_buff *skb; 3417 3418 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 3419 3420 /* NOTE: No TCP options attached and we never retransmit this. */ 3421 skb = alloc_skb(MAX_TCP_HEADER, priority); 3422 if (!skb) { 3423 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3424 return; 3425 } 3426 3427 /* Reserve space for headers and prepare control bits. */ 3428 skb_reserve(skb, MAX_TCP_HEADER); 3429 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3430 TCPHDR_ACK | TCPHDR_RST); 3431 tcp_mstamp_refresh(tcp_sk(sk)); 3432 /* Send it off. */ 3433 if (tcp_transmit_skb(sk, skb, 0, priority)) 3434 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3435 3436 /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3437 * skb here is different to the troublesome skb, so use NULL 3438 */ 3439 trace_tcp_send_reset(sk, NULL); 3440 } 3441 3442 /* Send a crossed SYN-ACK during socket establishment. 3443 * WARNING: This routine must only be called when we have already sent 3444 * a SYN packet that crossed the incoming SYN that caused this routine 3445 * to get called. If this assumption fails then the initial rcv_wnd 3446 * and rcv_wscale values will not be correct. 3447 */ 3448 int tcp_send_synack(struct sock *sk) 3449 { 3450 struct sk_buff *skb; 3451 3452 skb = tcp_rtx_queue_head(sk); 3453 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 3454 pr_err("%s: wrong queue state\n", __func__); 3455 return -EFAULT; 3456 } 3457 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 3458 if (skb_cloned(skb)) { 3459 struct sk_buff *nskb; 3460 3461 tcp_skb_tsorted_save(skb) { 3462 nskb = skb_copy(skb, GFP_ATOMIC); 3463 } tcp_skb_tsorted_restore(skb); 3464 if (!nskb) 3465 return -ENOMEM; 3466 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 3467 tcp_highest_sack_replace(sk, skb, nskb); 3468 tcp_rtx_queue_unlink_and_free(skb, sk); 3469 __skb_header_release(nskb); 3470 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 3471 sk_wmem_queued_add(sk, nskb->truesize); 3472 sk_mem_charge(sk, nskb->truesize); 3473 skb = nskb; 3474 } 3475 3476 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3477 tcp_ecn_send_synack(sk, skb); 3478 } 3479 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3480 } 3481 3482 /** 3483 * tcp_make_synack - Allocate one skb and build a SYNACK packet. 3484 * @sk: listener socket 3485 * @dst: dst entry attached to the SYNACK. It is consumed and caller 3486 * should not use it again. 3487 * @req: request_sock pointer 3488 * @foc: cookie for tcp fast open 3489 * @synack_type: Type of synack to prepare 3490 * @syn_skb: SYN packet just received. It could be NULL for rtx case. 3491 */ 3492 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3493 struct request_sock *req, 3494 struct tcp_fastopen_cookie *foc, 3495 enum tcp_synack_type synack_type, 3496 struct sk_buff *syn_skb) 3497 { 3498 struct inet_request_sock *ireq = inet_rsk(req); 3499 const struct tcp_sock *tp = tcp_sk(sk); 3500 struct tcp_md5sig_key *md5 = NULL; 3501 struct tcp_out_options opts; 3502 struct sk_buff *skb; 3503 int tcp_header_size; 3504 struct tcphdr *th; 3505 int mss; 3506 u64 now; 3507 3508 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 3509 if (unlikely(!skb)) { 3510 dst_release(dst); 3511 return NULL; 3512 } 3513 /* Reserve space for headers. */ 3514 skb_reserve(skb, MAX_TCP_HEADER); 3515 3516 switch (synack_type) { 3517 case TCP_SYNACK_NORMAL: 3518 skb_set_owner_w(skb, req_to_sk(req)); 3519 break; 3520 case TCP_SYNACK_COOKIE: 3521 /* Under synflood, we do not attach skb to a socket, 3522 * to avoid false sharing. 3523 */ 3524 break; 3525 case TCP_SYNACK_FASTOPEN: 3526 /* sk is a const pointer, because we want to express multiple 3527 * cpu might call us concurrently. 3528 * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3529 */ 3530 skb_set_owner_w(skb, (struct sock *)sk); 3531 break; 3532 } 3533 skb_dst_set(skb, dst); 3534 3535 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3536 3537 memset(&opts, 0, sizeof(opts)); 3538 now = tcp_clock_ns(); 3539 #ifdef CONFIG_SYN_COOKIES 3540 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) 3541 skb->skb_mstamp_ns = cookie_init_timestamp(req, now); 3542 else 3543 #endif 3544 { 3545 skb->skb_mstamp_ns = now; 3546 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ 3547 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); 3548 } 3549 3550 #ifdef CONFIG_TCP_MD5SIG 3551 rcu_read_lock(); 3552 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 3553 #endif 3554 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 3555 /* bpf program will be interested in the tcp_flags */ 3556 TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK; 3557 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 3558 foc, synack_type, 3559 syn_skb) + sizeof(*th); 3560 3561 skb_push(skb, tcp_header_size); 3562 skb_reset_transport_header(skb); 3563 3564 th = (struct tcphdr *)skb->data; 3565 memset(th, 0, sizeof(struct tcphdr)); 3566 th->syn = 1; 3567 th->ack = 1; 3568 tcp_ecn_make_synack(req, th); 3569 th->source = htons(ireq->ir_num); 3570 th->dest = ireq->ir_rmt_port; 3571 skb->mark = ireq->ir_mark; 3572 skb->ip_summed = CHECKSUM_PARTIAL; 3573 th->seq = htonl(tcp_rsk(req)->snt_isn); 3574 /* XXX data is queued and acked as is. No buffer/window check */ 3575 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 3576 3577 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3578 th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 3579 tcp_options_write((__be32 *)(th + 1), NULL, &opts); 3580 th->doff = (tcp_header_size >> 2); 3581 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3582 3583 #ifdef CONFIG_TCP_MD5SIG 3584 /* Okay, we have all we need - do the md5 hash if needed */ 3585 if (md5) 3586 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 3587 md5, req_to_sk(req), skb); 3588 rcu_read_unlock(); 3589 #endif 3590 3591 bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb, 3592 synack_type, &opts); 3593 3594 skb->skb_mstamp_ns = now; 3595 tcp_add_tx_delay(skb, tp); 3596 3597 return skb; 3598 } 3599 EXPORT_SYMBOL(tcp_make_synack); 3600 3601 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 3602 { 3603 struct inet_connection_sock *icsk = inet_csk(sk); 3604 const struct tcp_congestion_ops *ca; 3605 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 3606 3607 if (ca_key == TCP_CA_UNSPEC) 3608 return; 3609 3610 rcu_read_lock(); 3611 ca = tcp_ca_find_key(ca_key); 3612 if (likely(ca && bpf_try_module_get(ca, ca->owner))) { 3613 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); 3614 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 3615 icsk->icsk_ca_ops = ca; 3616 } 3617 rcu_read_unlock(); 3618 } 3619 3620 /* Do all connect socket setups that can be done AF independent. */ 3621 static void tcp_connect_init(struct sock *sk) 3622 { 3623 const struct dst_entry *dst = __sk_dst_get(sk); 3624 struct tcp_sock *tp = tcp_sk(sk); 3625 __u8 rcv_wscale; 3626 u32 rcv_wnd; 3627 3628 /* We'll fix this up when we get a response from the other end. 3629 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 3630 */ 3631 tp->tcp_header_len = sizeof(struct tcphdr); 3632 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 3633 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 3634 3635 #ifdef CONFIG_TCP_MD5SIG 3636 if (tp->af_specific->md5_lookup(sk, sk)) 3637 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3638 #endif 3639 3640 /* If user gave his TCP_MAXSEG, record it to clamp */ 3641 if (tp->rx_opt.user_mss) 3642 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 3643 tp->max_window = 0; 3644 tcp_mtup_init(sk); 3645 tcp_sync_mss(sk, dst_mtu(dst)); 3646 3647 tcp_ca_dst_init(sk, dst); 3648 3649 if (!tp->window_clamp) 3650 tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 3651 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3652 3653 tcp_initialize_rcv_mss(sk); 3654 3655 /* limit the window selection if the user enforce a smaller rx buffer */ 3656 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3657 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3658 tp->window_clamp = tcp_full_space(sk); 3659 3660 rcv_wnd = tcp_rwnd_init_bpf(sk); 3661 if (rcv_wnd == 0) 3662 rcv_wnd = dst_metric(dst, RTAX_INITRWND); 3663 3664 tcp_select_initial_window(sk, tcp_full_space(sk), 3665 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 3666 &tp->rcv_wnd, 3667 &tp->window_clamp, 3668 sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 3669 &rcv_wscale, 3670 rcv_wnd); 3671 3672 tp->rx_opt.rcv_wscale = rcv_wscale; 3673 tp->rcv_ssthresh = tp->rcv_wnd; 3674 3675 sk->sk_err = 0; 3676 sock_reset_flag(sk, SOCK_DONE); 3677 tp->snd_wnd = 0; 3678 tcp_init_wl(tp, 0); 3679 tcp_write_queue_purge(sk); 3680 tp->snd_una = tp->write_seq; 3681 tp->snd_sml = tp->write_seq; 3682 tp->snd_up = tp->write_seq; 3683 WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3684 3685 if (likely(!tp->repair)) 3686 tp->rcv_nxt = 0; 3687 else 3688 tp->rcv_tstamp = tcp_jiffies32; 3689 tp->rcv_wup = tp->rcv_nxt; 3690 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); 3691 3692 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3693 inet_csk(sk)->icsk_retransmits = 0; 3694 tcp_clear_retrans(tp); 3695 } 3696 3697 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3698 { 3699 struct tcp_sock *tp = tcp_sk(sk); 3700 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3701 3702 tcb->end_seq += skb->len; 3703 __skb_header_release(skb); 3704 sk_wmem_queued_add(sk, skb->truesize); 3705 sk_mem_charge(sk, skb->truesize); 3706 WRITE_ONCE(tp->write_seq, tcb->end_seq); 3707 tp->packets_out += tcp_skb_pcount(skb); 3708 } 3709 3710 /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3711 * queue a data-only packet after the regular SYN, such that regular SYNs 3712 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3713 * only the SYN sequence, the data are retransmitted in the first ACK. 3714 * If cookie is not cached or other error occurs, falls back to send a 3715 * regular SYN with Fast Open cookie request option. 3716 */ 3717 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3718 { 3719 struct tcp_sock *tp = tcp_sk(sk); 3720 struct tcp_fastopen_request *fo = tp->fastopen_req; 3721 int space, err = 0; 3722 struct sk_buff *syn_data; 3723 3724 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3725 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3726 goto fallback; 3727 3728 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3729 * user-MSS. Reserve maximum option space for middleboxes that add 3730 * private TCP options. The cost is reduced data space in SYN :( 3731 */ 3732 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 3733 3734 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3735 MAX_TCP_OPTION_SPACE; 3736 3737 space = min_t(size_t, space, fo->size); 3738 3739 /* limit to order-0 allocations */ 3740 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3741 3742 syn_data = tcp_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3743 if (!syn_data) 3744 goto fallback; 3745 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 3746 if (space) { 3747 int copied = copy_from_iter(skb_put(syn_data, space), space, 3748 &fo->data->msg_iter); 3749 if (unlikely(!copied)) { 3750 tcp_skb_tsorted_anchor_cleanup(syn_data); 3751 kfree_skb(syn_data); 3752 goto fallback; 3753 } 3754 if (copied != space) { 3755 skb_trim(syn_data, copied); 3756 space = copied; 3757 } 3758 skb_zcopy_set(syn_data, fo->uarg, NULL); 3759 } 3760 /* No more data pending in inet_wait_for_connect() */ 3761 if (space == fo->size) 3762 fo->data = NULL; 3763 fo->copied = space; 3764 3765 tcp_connect_queue_skb(sk, syn_data); 3766 if (syn_data->len) 3767 tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3768 3769 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3770 3771 syn->skb_mstamp_ns = syn_data->skb_mstamp_ns; 3772 3773 /* Now full SYN+DATA was cloned and sent (or not), 3774 * remove the SYN from the original skb (syn_data) 3775 * we keep in write queue in case of a retransmit, as we 3776 * also have the SYN packet (with no data) in the same queue. 3777 */ 3778 TCP_SKB_CB(syn_data)->seq++; 3779 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3780 if (!err) { 3781 tp->syn_data = (fo->copied > 0); 3782 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3783 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3784 goto done; 3785 } 3786 3787 /* data was not sent, put it in write_queue */ 3788 __skb_queue_tail(&sk->sk_write_queue, syn_data); 3789 tp->packets_out -= tcp_skb_pcount(syn_data); 3790 3791 fallback: 3792 /* Send a regular SYN with Fast Open cookie request option */ 3793 if (fo->cookie.len > 0) 3794 fo->cookie.len = 0; 3795 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3796 if (err) 3797 tp->syn_fastopen = 0; 3798 done: 3799 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3800 return err; 3801 } 3802 3803 /* Build a SYN and send it off. */ 3804 int tcp_connect(struct sock *sk) 3805 { 3806 struct tcp_sock *tp = tcp_sk(sk); 3807 struct sk_buff *buff; 3808 int err; 3809 3810 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 3811 3812 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 3813 return -EHOSTUNREACH; /* Routing failure or similar. */ 3814 3815 tcp_connect_init(sk); 3816 3817 if (unlikely(tp->repair)) { 3818 tcp_finish_connect(sk, NULL); 3819 return 0; 3820 } 3821 3822 buff = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3823 if (unlikely(!buff)) 3824 return -ENOBUFS; 3825 3826 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 3827 tcp_mstamp_refresh(tp); 3828 tp->retrans_stamp = tcp_time_stamp(tp); 3829 tcp_connect_queue_skb(sk, buff); 3830 tcp_ecn_send_syn(sk, buff); 3831 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 3832 3833 /* Send off SYN; include data in Fast Open. */ 3834 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3835 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3836 if (err == -ECONNREFUSED) 3837 return err; 3838 3839 /* We change tp->snd_nxt after the tcp_transmit_skb() call 3840 * in order to make this packet get counted in tcpOutSegs. 3841 */ 3842 WRITE_ONCE(tp->snd_nxt, tp->write_seq); 3843 tp->pushed_seq = tp->write_seq; 3844 buff = tcp_send_head(sk); 3845 if (unlikely(buff)) { 3846 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); 3847 tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3848 } 3849 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 3850 3851 /* Timer for repeating the SYN until an answer. */ 3852 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 3853 inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 3854 return 0; 3855 } 3856 EXPORT_SYMBOL(tcp_connect); 3857 3858 /* Send out a delayed ack, the caller does the policy checking 3859 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 3860 * for details. 3861 */ 3862 void tcp_send_delayed_ack(struct sock *sk) 3863 { 3864 struct inet_connection_sock *icsk = inet_csk(sk); 3865 int ato = icsk->icsk_ack.ato; 3866 unsigned long timeout; 3867 3868 if (ato > TCP_DELACK_MIN) { 3869 const struct tcp_sock *tp = tcp_sk(sk); 3870 int max_ato = HZ / 2; 3871 3872 if (inet_csk_in_pingpong_mode(sk) || 3873 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 3874 max_ato = TCP_DELACK_MAX; 3875 3876 /* Slow path, intersegment interval is "high". */ 3877 3878 /* If some rtt estimate is known, use it to bound delayed ack. 3879 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 3880 * directly. 3881 */ 3882 if (tp->srtt_us) { 3883 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3884 TCP_DELACK_MIN); 3885 3886 if (rtt < max_ato) 3887 max_ato = rtt; 3888 } 3889 3890 ato = min(ato, max_ato); 3891 } 3892 3893 ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max); 3894 3895 /* Stay within the limit we were given */ 3896 timeout = jiffies + ato; 3897 3898 /* Use new timeout only if there wasn't a older one earlier. */ 3899 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 3900 /* If delack timer is about to expire, send ACK now. */ 3901 if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 3902 tcp_send_ack(sk); 3903 return; 3904 } 3905 3906 if (!time_before(timeout, icsk->icsk_ack.timeout)) 3907 timeout = icsk->icsk_ack.timeout; 3908 } 3909 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3910 icsk->icsk_ack.timeout = timeout; 3911 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 3912 } 3913 3914 /* This routine sends an ack and also updates the window. */ 3915 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 3916 { 3917 struct sk_buff *buff; 3918 3919 /* If we have been reset, we may not send again. */ 3920 if (sk->sk_state == TCP_CLOSE) 3921 return; 3922 3923 /* We are not putting this on the write queue, so 3924 * tcp_transmit_skb() will set the ownership to this 3925 * sock. 3926 */ 3927 buff = alloc_skb(MAX_TCP_HEADER, 3928 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 3929 if (unlikely(!buff)) { 3930 struct inet_connection_sock *icsk = inet_csk(sk); 3931 unsigned long delay; 3932 3933 delay = TCP_DELACK_MAX << icsk->icsk_ack.retry; 3934 if (delay < TCP_RTO_MAX) 3935 icsk->icsk_ack.retry++; 3936 inet_csk_schedule_ack(sk); 3937 icsk->icsk_ack.ato = TCP_ATO_MIN; 3938 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX); 3939 return; 3940 } 3941 3942 /* Reserve space for headers and prepare control bits. */ 3943 skb_reserve(buff, MAX_TCP_HEADER); 3944 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 3945 3946 /* We do not want pure acks influencing TCP Small Queues or fq/pacing 3947 * too much. 3948 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 3949 */ 3950 skb_set_tcp_pure_ack(buff); 3951 3952 /* Send it off, this clears delayed acks for us. */ 3953 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 3954 } 3955 EXPORT_SYMBOL_GPL(__tcp_send_ack); 3956 3957 void tcp_send_ack(struct sock *sk) 3958 { 3959 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 3960 } 3961 3962 /* This routine sends a packet with an out of date sequence 3963 * number. It assumes the other end will try to ack it. 3964 * 3965 * Question: what should we make while urgent mode? 3966 * 4.4BSD forces sending single byte of data. We cannot send 3967 * out of window data, because we have SND.NXT==SND.MAX... 3968 * 3969 * Current solution: to send TWO zero-length segments in urgent mode: 3970 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 3971 * out-of-date with SND.UNA-1 to probe window. 3972 */ 3973 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 3974 { 3975 struct tcp_sock *tp = tcp_sk(sk); 3976 struct sk_buff *skb; 3977 3978 /* We don't queue it, tcp_transmit_skb() sets ownership. */ 3979 skb = alloc_skb(MAX_TCP_HEADER, 3980 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 3981 if (!skb) 3982 return -1; 3983 3984 /* Reserve space for headers and set control bits. */ 3985 skb_reserve(skb, MAX_TCP_HEADER); 3986 /* Use a previous sequence. This should cause the other 3987 * end to send an ack. Don't queue or clone SKB, just 3988 * send it. 3989 */ 3990 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3991 NET_INC_STATS(sock_net(sk), mib); 3992 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 3993 } 3994 3995 /* Called from setsockopt( ... TCP_REPAIR ) */ 3996 void tcp_send_window_probe(struct sock *sk) 3997 { 3998 if (sk->sk_state == TCP_ESTABLISHED) { 3999 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 4000 tcp_mstamp_refresh(tcp_sk(sk)); 4001 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 4002 } 4003 } 4004 4005 /* Initiate keepalive or window probe from timer. */ 4006 int tcp_write_wakeup(struct sock *sk, int mib) 4007 { 4008 struct tcp_sock *tp = tcp_sk(sk); 4009 struct sk_buff *skb; 4010 4011 if (sk->sk_state == TCP_CLOSE) 4012 return -1; 4013 4014 skb = tcp_send_head(sk); 4015 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 4016 int err; 4017 unsigned int mss = tcp_current_mss(sk); 4018 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 4019 4020 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 4021 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 4022 4023 /* We are probing the opening of a window 4024 * but the window size is != 0 4025 * must have been a result SWS avoidance ( sender ) 4026 */ 4027 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 4028 skb->len > mss) { 4029 seg_size = min(seg_size, mss); 4030 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 4031 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 4032 skb, seg_size, mss, GFP_ATOMIC)) 4033 return -1; 4034 } else if (!tcp_skb_pcount(skb)) 4035 tcp_set_skb_tso_segs(skb, mss); 4036 4037 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 4038 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 4039 if (!err) 4040 tcp_event_new_data_sent(sk, skb); 4041 return err; 4042 } else { 4043 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 4044 tcp_xmit_probe_skb(sk, 1, mib); 4045 return tcp_xmit_probe_skb(sk, 0, mib); 4046 } 4047 } 4048 4049 /* A window probe timeout has occurred. If window is not closed send 4050 * a partial packet else a zero probe. 4051 */ 4052 void tcp_send_probe0(struct sock *sk) 4053 { 4054 struct inet_connection_sock *icsk = inet_csk(sk); 4055 struct tcp_sock *tp = tcp_sk(sk); 4056 struct net *net = sock_net(sk); 4057 unsigned long timeout; 4058 int err; 4059 4060 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 4061 4062 if (tp->packets_out || tcp_write_queue_empty(sk)) { 4063 /* Cancel probe timer, if it is not required. */ 4064 icsk->icsk_probes_out = 0; 4065 icsk->icsk_backoff = 0; 4066 icsk->icsk_probes_tstamp = 0; 4067 return; 4068 } 4069 4070 icsk->icsk_probes_out++; 4071 if (err <= 0) { 4072 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 4073 icsk->icsk_backoff++; 4074 timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 4075 } else { 4076 /* If packet was not sent due to local congestion, 4077 * Let senders fight for local resources conservatively. 4078 */ 4079 timeout = TCP_RESOURCE_PROBE_INTERVAL; 4080 } 4081 4082 timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout); 4083 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); 4084 } 4085 4086 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 4087 { 4088 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 4089 struct flowi fl; 4090 int res; 4091 4092 tcp_rsk(req)->txhash = net_tx_rndhash(); 4093 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL, 4094 NULL); 4095 if (!res) { 4096 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 4097 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 4098 if (unlikely(tcp_passive_fastopen(sk))) 4099 tcp_sk(sk)->total_retrans++; 4100 trace_tcp_retransmit_synack(sk, req); 4101 } 4102 return res; 4103 } 4104 EXPORT_SYMBOL(tcp_rtx_synack); 4105