1 /* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Authors: Ross Biro 9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 10 * Mark Evans, <evansmp@uhura.aston.ac.uk> 11 * Corey Minyard <wf-rch!minyard@relay.EU.net> 12 * Florian La Roche, <flla@stud.uni-sb.de> 13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 14 * Linus Torvalds, <torvalds@cs.helsinki.fi> 15 * Alan Cox, <gw4pts@gw4pts.ampr.org> 16 * Matthew Dillon, <dillon@apollo.west.oic.com> 17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 18 * Jorge Cwik, <jorge@laser.satlink.net> 19 */ 20 21 /* 22 * Changes: Pedro Roque : Retransmit queue handled by TCP. 23 * : Fragmentation on mtu decrease 24 * : Segment collapse on retransmit 25 * : AF independence 26 * 27 * Linus Torvalds : send_delayed_ack 28 * David S. Miller : Charge memory using the right skb 29 * during syn/ack processing. 30 * David S. Miller : Output engine completely rewritten. 31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 32 * Cacophonix Gaul : draft-minshall-nagle-01 33 * J Hadi Salim : ECN support 34 * 35 */ 36 37 #define pr_fmt(fmt) "TCP: " fmt 38 39 #include <net/tcp.h> 40 41 #include <linux/compiler.h> 42 #include <linux/gfp.h> 43 #include <linux/module.h> 44 45 /* People can turn this off for buggy TCP's found in printers etc. */ 46 int sysctl_tcp_retrans_collapse __read_mostly = 1; 47 48 /* People can turn this on to work with those rare, broken TCPs that 49 * interpret the window field as a signed quantity. 50 */ 51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 52 53 /* Default TSQ limit of two TSO segments */ 54 int sysctl_tcp_limit_output_bytes __read_mostly = 131072; 55 56 /* This limits the percentage of the congestion window which we 57 * will allow a single TSO frame to consume. Building TSO frames 58 * which are too large can cause TCP streams to be bursty. 59 */ 60 int sysctl_tcp_tso_win_divisor __read_mostly = 3; 61 62 int sysctl_tcp_mtu_probing __read_mostly = 0; 63 int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS; 64 65 /* By default, RFC2861 behavior. */ 66 int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 67 68 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 69 int push_one, gfp_t gfp); 70 71 /* Account for new data that has been sent to the network. */ 72 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) 73 { 74 struct inet_connection_sock *icsk = inet_csk(sk); 75 struct tcp_sock *tp = tcp_sk(sk); 76 unsigned int prior_packets = tp->packets_out; 77 78 tcp_advance_send_head(sk, skb); 79 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 80 81 tp->packets_out += tcp_skb_pcount(skb); 82 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 83 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 84 tcp_rearm_rto(sk); 85 } 86 } 87 88 /* SND.NXT, if window was not shrunk. 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 return tp->snd_nxt; 100 else 101 return tcp_wnd_end(tp); 102 } 103 104 /* Calculate mss to advertise in SYN segment. 105 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 106 * 107 * 1. It is independent of path mtu. 108 * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 109 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 110 * attached devices, because some buggy hosts are confused by 111 * large MSS. 112 * 4. We do not make 3, we advertise MSS, calculated from first 113 * hop device mtu, but allow to raise it to ip_rt_min_advmss. 114 * This may be overridden via information stored in routing table. 115 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 116 * probably even Jumbo". 117 */ 118 static __u16 tcp_advertise_mss(struct sock *sk) 119 { 120 struct tcp_sock *tp = tcp_sk(sk); 121 const struct dst_entry *dst = __sk_dst_get(sk); 122 int mss = tp->advmss; 123 124 if (dst) { 125 unsigned int metric = dst_metric_advmss(dst); 126 127 if (metric < mss) { 128 mss = metric; 129 tp->advmss = mss; 130 } 131 } 132 133 return (__u16)mss; 134 } 135 136 /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 137 * This is the first part of cwnd validation mechanism. */ 138 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) 139 { 140 struct tcp_sock *tp = tcp_sk(sk); 141 s32 delta = tcp_time_stamp - tp->lsndtime; 142 u32 restart_cwnd = tcp_init_cwnd(tp, dst); 143 u32 cwnd = tp->snd_cwnd; 144 145 tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 146 147 tp->snd_ssthresh = tcp_current_ssthresh(sk); 148 restart_cwnd = min(restart_cwnd, cwnd); 149 150 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 151 cwnd >>= 1; 152 tp->snd_cwnd = max(cwnd, restart_cwnd); 153 tp->snd_cwnd_stamp = tcp_time_stamp; 154 tp->snd_cwnd_used = 0; 155 } 156 157 /* Congestion state accounting after a packet has been sent. */ 158 static void tcp_event_data_sent(struct tcp_sock *tp, 159 struct sock *sk) 160 { 161 struct inet_connection_sock *icsk = inet_csk(sk); 162 const u32 now = tcp_time_stamp; 163 164 if (sysctl_tcp_slow_start_after_idle && 165 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 166 tcp_cwnd_restart(sk, __sk_dst_get(sk)); 167 168 tp->lsndtime = now; 169 170 /* If it is a reply for ato after last received 171 * packet, enter pingpong mode. 172 */ 173 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 174 icsk->icsk_ack.pingpong = 1; 175 } 176 177 /* Account for an ACK we sent. */ 178 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 179 { 180 tcp_dec_quickack_mode(sk, pkts); 181 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 182 } 183 184 /* Determine a window scaling and initial window to offer. 185 * Based on the assumption that the given amount of space 186 * will be offered. Store the results in the tp structure. 187 * NOTE: for smooth operation initial space offering should 188 * be a multiple of mss if possible. We assume here that mss >= 1. 189 * This MUST be enforced by all callers. 190 */ 191 void tcp_select_initial_window(int __space, __u32 mss, 192 __u32 *rcv_wnd, __u32 *window_clamp, 193 int wscale_ok, __u8 *rcv_wscale, 194 __u32 init_rcv_wnd) 195 { 196 unsigned int space = (__space < 0 ? 0 : __space); 197 198 /* If no clamp set the clamp to the max possible scaled window */ 199 if (*window_clamp == 0) 200 (*window_clamp) = (65535 << 14); 201 space = min(*window_clamp, space); 202 203 /* Quantize space offering to a multiple of mss if possible. */ 204 if (space > mss) 205 space = (space / mss) * mss; 206 207 /* NOTE: offering an initial window larger than 32767 208 * will break some buggy TCP stacks. If the admin tells us 209 * it is likely we could be speaking with such a buggy stack 210 * we will truncate our initial window offering to 32K-1 211 * unless the remote has sent us a window scaling option, 212 * which we interpret as a sign the remote TCP is not 213 * misinterpreting the window field as a signed quantity. 214 */ 215 if (sysctl_tcp_workaround_signed_windows) 216 (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 217 else 218 (*rcv_wnd) = space; 219 220 (*rcv_wscale) = 0; 221 if (wscale_ok) { 222 /* Set window scaling on max possible window 223 * See RFC1323 for an explanation of the limit to 14 224 */ 225 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 226 space = min_t(u32, space, *window_clamp); 227 while (space > 65535 && (*rcv_wscale) < 14) { 228 space >>= 1; 229 (*rcv_wscale)++; 230 } 231 } 232 233 /* Set initial window to a value enough for senders starting with 234 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place 235 * a limit on the initial window when mss is larger than 1460. 236 */ 237 if (mss > (1 << *rcv_wscale)) { 238 int init_cwnd = TCP_DEFAULT_INIT_RCVWND; 239 if (mss > 1460) 240 init_cwnd = 241 max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2); 242 /* when initializing use the value from init_rcv_wnd 243 * rather than the default from above 244 */ 245 if (init_rcv_wnd) 246 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 247 else 248 *rcv_wnd = min(*rcv_wnd, init_cwnd * mss); 249 } 250 251 /* Set the clamp no higher than max representable value */ 252 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 253 } 254 EXPORT_SYMBOL(tcp_select_initial_window); 255 256 /* Chose a new window to advertise, update state in tcp_sock for the 257 * socket, and return result with RFC1323 scaling applied. The return 258 * value can be stuffed directly into th->window for an outgoing 259 * frame. 260 */ 261 static u16 tcp_select_window(struct sock *sk) 262 { 263 struct tcp_sock *tp = tcp_sk(sk); 264 u32 cur_win = tcp_receive_window(tp); 265 u32 new_win = __tcp_select_window(sk); 266 267 /* Never shrink the offered window */ 268 if (new_win < cur_win) { 269 /* Danger Will Robinson! 270 * Don't update rcv_wup/rcv_wnd here or else 271 * we will not be able to advertise a zero 272 * window in time. --DaveM 273 * 274 * Relax Will Robinson. 275 */ 276 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 277 } 278 tp->rcv_wnd = new_win; 279 tp->rcv_wup = tp->rcv_nxt; 280 281 /* Make sure we do not exceed the maximum possible 282 * scaled window. 283 */ 284 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 285 new_win = min(new_win, MAX_TCP_WINDOW); 286 else 287 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 288 289 /* RFC1323 scaling applied */ 290 new_win >>= tp->rx_opt.rcv_wscale; 291 292 /* If we advertise zero window, disable fast path. */ 293 if (new_win == 0) 294 tp->pred_flags = 0; 295 296 return new_win; 297 } 298 299 /* Packet ECN state for a SYN-ACK */ 300 static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) 301 { 302 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 303 if (!(tp->ecn_flags & TCP_ECN_OK)) 304 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 305 } 306 307 /* Packet ECN state for a SYN. */ 308 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 309 { 310 struct tcp_sock *tp = tcp_sk(sk); 311 312 tp->ecn_flags = 0; 313 if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) { 314 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 315 tp->ecn_flags = TCP_ECN_OK; 316 } 317 } 318 319 static __inline__ void 320 TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) 321 { 322 if (inet_rsk(req)->ecn_ok) 323 th->ece = 1; 324 } 325 326 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 327 * be sent. 328 */ 329 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 330 int tcp_header_len) 331 { 332 struct tcp_sock *tp = tcp_sk(sk); 333 334 if (tp->ecn_flags & TCP_ECN_OK) { 335 /* Not-retransmitted data segment: set ECT and inject CWR. */ 336 if (skb->len != tcp_header_len && 337 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 338 INET_ECN_xmit(sk); 339 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 340 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 341 tcp_hdr(skb)->cwr = 1; 342 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 343 } 344 } else { 345 /* ACK or retransmitted segment: clear ECT|CE */ 346 INET_ECN_dontxmit(sk); 347 } 348 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 349 tcp_hdr(skb)->ece = 1; 350 } 351 } 352 353 /* Constructs common control bits of non-data skb. If SYN/FIN is present, 354 * auto increment end seqno. 355 */ 356 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 357 { 358 skb->ip_summed = CHECKSUM_PARTIAL; 359 skb->csum = 0; 360 361 TCP_SKB_CB(skb)->tcp_flags = flags; 362 TCP_SKB_CB(skb)->sacked = 0; 363 364 skb_shinfo(skb)->gso_segs = 1; 365 skb_shinfo(skb)->gso_size = 0; 366 skb_shinfo(skb)->gso_type = 0; 367 368 TCP_SKB_CB(skb)->seq = seq; 369 if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 370 seq++; 371 TCP_SKB_CB(skb)->end_seq = seq; 372 } 373 374 static inline bool tcp_urg_mode(const struct tcp_sock *tp) 375 { 376 return tp->snd_una != tp->snd_up; 377 } 378 379 #define OPTION_SACK_ADVERTISE (1 << 0) 380 #define OPTION_TS (1 << 1) 381 #define OPTION_MD5 (1 << 2) 382 #define OPTION_WSCALE (1 << 3) 383 #define OPTION_FAST_OPEN_COOKIE (1 << 8) 384 385 struct tcp_out_options { 386 u16 options; /* bit field of OPTION_* */ 387 u16 mss; /* 0 to disable */ 388 u8 ws; /* window scale, 0 to disable */ 389 u8 num_sack_blocks; /* number of SACK blocks to include */ 390 u8 hash_size; /* bytes in hash_location */ 391 __u8 *hash_location; /* temporary pointer, overloaded */ 392 __u32 tsval, tsecr; /* need to include OPTION_TS */ 393 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 394 }; 395 396 /* Write previously computed TCP options to the packet. 397 * 398 * Beware: Something in the Internet is very sensitive to the ordering of 399 * TCP options, we learned this through the hard way, so be careful here. 400 * Luckily we can at least blame others for their non-compliance but from 401 * inter-operatibility perspective it seems that we're somewhat stuck with 402 * the ordering which we have been using if we want to keep working with 403 * those broken things (not that it currently hurts anybody as there isn't 404 * particular reason why the ordering would need to be changed). 405 * 406 * At least SACK_PERM as the first option is known to lead to a disaster 407 * (but it may well be that other scenarios fail similarly). 408 */ 409 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 410 struct tcp_out_options *opts) 411 { 412 u16 options = opts->options; /* mungable copy */ 413 414 if (unlikely(OPTION_MD5 & options)) { 415 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 416 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 417 /* overload cookie hash location */ 418 opts->hash_location = (__u8 *)ptr; 419 ptr += 4; 420 } 421 422 if (unlikely(opts->mss)) { 423 *ptr++ = htonl((TCPOPT_MSS << 24) | 424 (TCPOLEN_MSS << 16) | 425 opts->mss); 426 } 427 428 if (likely(OPTION_TS & options)) { 429 if (unlikely(OPTION_SACK_ADVERTISE & options)) { 430 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 431 (TCPOLEN_SACK_PERM << 16) | 432 (TCPOPT_TIMESTAMP << 8) | 433 TCPOLEN_TIMESTAMP); 434 options &= ~OPTION_SACK_ADVERTISE; 435 } else { 436 *ptr++ = htonl((TCPOPT_NOP << 24) | 437 (TCPOPT_NOP << 16) | 438 (TCPOPT_TIMESTAMP << 8) | 439 TCPOLEN_TIMESTAMP); 440 } 441 *ptr++ = htonl(opts->tsval); 442 *ptr++ = htonl(opts->tsecr); 443 } 444 445 if (unlikely(OPTION_SACK_ADVERTISE & options)) { 446 *ptr++ = htonl((TCPOPT_NOP << 24) | 447 (TCPOPT_NOP << 16) | 448 (TCPOPT_SACK_PERM << 8) | 449 TCPOLEN_SACK_PERM); 450 } 451 452 if (unlikely(OPTION_WSCALE & options)) { 453 *ptr++ = htonl((TCPOPT_NOP << 24) | 454 (TCPOPT_WINDOW << 16) | 455 (TCPOLEN_WINDOW << 8) | 456 opts->ws); 457 } 458 459 if (unlikely(opts->num_sack_blocks)) { 460 struct tcp_sack_block *sp = tp->rx_opt.dsack ? 461 tp->duplicate_sack : tp->selective_acks; 462 int this_sack; 463 464 *ptr++ = htonl((TCPOPT_NOP << 24) | 465 (TCPOPT_NOP << 16) | 466 (TCPOPT_SACK << 8) | 467 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 468 TCPOLEN_SACK_PERBLOCK))); 469 470 for (this_sack = 0; this_sack < opts->num_sack_blocks; 471 ++this_sack) { 472 *ptr++ = htonl(sp[this_sack].start_seq); 473 *ptr++ = htonl(sp[this_sack].end_seq); 474 } 475 476 tp->rx_opt.dsack = 0; 477 } 478 479 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 480 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 481 482 *ptr++ = htonl((TCPOPT_EXP << 24) | 483 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | 484 TCPOPT_FASTOPEN_MAGIC); 485 486 memcpy(ptr, foc->val, foc->len); 487 if ((foc->len & 3) == 2) { 488 u8 *align = ((u8 *)ptr) + foc->len; 489 align[0] = align[1] = TCPOPT_NOP; 490 } 491 ptr += (foc->len + 3) >> 2; 492 } 493 } 494 495 /* Compute TCP options for SYN packets. This is not the final 496 * network wire format yet. 497 */ 498 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 499 struct tcp_out_options *opts, 500 struct tcp_md5sig_key **md5) 501 { 502 struct tcp_sock *tp = tcp_sk(sk); 503 unsigned int remaining = MAX_TCP_OPTION_SPACE; 504 struct tcp_fastopen_request *fastopen = tp->fastopen_req; 505 506 #ifdef CONFIG_TCP_MD5SIG 507 *md5 = tp->af_specific->md5_lookup(sk, sk); 508 if (*md5) { 509 opts->options |= OPTION_MD5; 510 remaining -= TCPOLEN_MD5SIG_ALIGNED; 511 } 512 #else 513 *md5 = NULL; 514 #endif 515 516 /* We always get an MSS option. The option bytes which will be seen in 517 * normal data packets should timestamps be used, must be in the MSS 518 * advertised. But we subtract them from tp->mss_cache so that 519 * calculations in tcp_sendmsg are simpler etc. So account for this 520 * fact here if necessary. If we don't do this correctly, as a 521 * receiver we won't recognize data packets as being full sized when we 522 * should, and thus we won't abide by the delayed ACK rules correctly. 523 * SACKs don't matter, we never delay an ACK when we have any of those 524 * going out. */ 525 opts->mss = tcp_advertise_mss(sk); 526 remaining -= TCPOLEN_MSS_ALIGNED; 527 528 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 529 opts->options |= OPTION_TS; 530 opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset; 531 opts->tsecr = tp->rx_opt.ts_recent; 532 remaining -= TCPOLEN_TSTAMP_ALIGNED; 533 } 534 if (likely(sysctl_tcp_window_scaling)) { 535 opts->ws = tp->rx_opt.rcv_wscale; 536 opts->options |= OPTION_WSCALE; 537 remaining -= TCPOLEN_WSCALE_ALIGNED; 538 } 539 if (likely(sysctl_tcp_sack)) { 540 opts->options |= OPTION_SACK_ADVERTISE; 541 if (unlikely(!(OPTION_TS & opts->options))) 542 remaining -= TCPOLEN_SACKPERM_ALIGNED; 543 } 544 545 if (fastopen && fastopen->cookie.len >= 0) { 546 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; 547 need = (need + 3) & ~3U; /* Align to 32 bits */ 548 if (remaining >= need) { 549 opts->options |= OPTION_FAST_OPEN_COOKIE; 550 opts->fastopen_cookie = &fastopen->cookie; 551 remaining -= need; 552 tp->syn_fastopen = 1; 553 } 554 } 555 556 return MAX_TCP_OPTION_SPACE - remaining; 557 } 558 559 /* Set up TCP options for SYN-ACKs. */ 560 static unsigned int tcp_synack_options(struct sock *sk, 561 struct request_sock *req, 562 unsigned int mss, struct sk_buff *skb, 563 struct tcp_out_options *opts, 564 struct tcp_md5sig_key **md5, 565 struct tcp_fastopen_cookie *foc) 566 { 567 struct inet_request_sock *ireq = inet_rsk(req); 568 unsigned int remaining = MAX_TCP_OPTION_SPACE; 569 570 #ifdef CONFIG_TCP_MD5SIG 571 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 572 if (*md5) { 573 opts->options |= OPTION_MD5; 574 remaining -= TCPOLEN_MD5SIG_ALIGNED; 575 576 /* We can't fit any SACK blocks in a packet with MD5 + TS 577 * options. There was discussion about disabling SACK 578 * rather than TS in order to fit in better with old, 579 * buggy kernels, but that was deemed to be unnecessary. 580 */ 581 ireq->tstamp_ok &= !ireq->sack_ok; 582 } 583 #else 584 *md5 = NULL; 585 #endif 586 587 /* We always send an MSS option. */ 588 opts->mss = mss; 589 remaining -= TCPOLEN_MSS_ALIGNED; 590 591 if (likely(ireq->wscale_ok)) { 592 opts->ws = ireq->rcv_wscale; 593 opts->options |= OPTION_WSCALE; 594 remaining -= TCPOLEN_WSCALE_ALIGNED; 595 } 596 if (likely(ireq->tstamp_ok)) { 597 opts->options |= OPTION_TS; 598 opts->tsval = TCP_SKB_CB(skb)->when; 599 opts->tsecr = req->ts_recent; 600 remaining -= TCPOLEN_TSTAMP_ALIGNED; 601 } 602 if (likely(ireq->sack_ok)) { 603 opts->options |= OPTION_SACK_ADVERTISE; 604 if (unlikely(!ireq->tstamp_ok)) 605 remaining -= TCPOLEN_SACKPERM_ALIGNED; 606 } 607 if (foc != NULL) { 608 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 609 need = (need + 3) & ~3U; /* Align to 32 bits */ 610 if (remaining >= need) { 611 opts->options |= OPTION_FAST_OPEN_COOKIE; 612 opts->fastopen_cookie = foc; 613 remaining -= need; 614 } 615 } 616 617 return MAX_TCP_OPTION_SPACE - remaining; 618 } 619 620 /* Compute TCP options for ESTABLISHED sockets. This is not the 621 * final wire format yet. 622 */ 623 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 624 struct tcp_out_options *opts, 625 struct tcp_md5sig_key **md5) 626 { 627 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 628 struct tcp_sock *tp = tcp_sk(sk); 629 unsigned int size = 0; 630 unsigned int eff_sacks; 631 632 #ifdef CONFIG_TCP_MD5SIG 633 *md5 = tp->af_specific->md5_lookup(sk, sk); 634 if (unlikely(*md5)) { 635 opts->options |= OPTION_MD5; 636 size += TCPOLEN_MD5SIG_ALIGNED; 637 } 638 #else 639 *md5 = NULL; 640 #endif 641 642 if (likely(tp->rx_opt.tstamp_ok)) { 643 opts->options |= OPTION_TS; 644 opts->tsval = tcb ? tcb->when + tp->tsoffset : 0; 645 opts->tsecr = tp->rx_opt.ts_recent; 646 size += TCPOLEN_TSTAMP_ALIGNED; 647 } 648 649 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 650 if (unlikely(eff_sacks)) { 651 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 652 opts->num_sack_blocks = 653 min_t(unsigned int, eff_sacks, 654 (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 655 TCPOLEN_SACK_PERBLOCK); 656 size += TCPOLEN_SACK_BASE_ALIGNED + 657 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 658 } 659 660 return size; 661 } 662 663 664 /* TCP SMALL QUEUES (TSQ) 665 * 666 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 667 * to reduce RTT and bufferbloat. 668 * We do this using a special skb destructor (tcp_wfree). 669 * 670 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 671 * needs to be reallocated in a driver. 672 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc 673 * 674 * Since transmit from skb destructor is forbidden, we use a tasklet 675 * to process all sockets that eventually need to send more skbs. 676 * We use one tasklet per cpu, with its own queue of sockets. 677 */ 678 struct tsq_tasklet { 679 struct tasklet_struct tasklet; 680 struct list_head head; /* queue of tcp sockets */ 681 }; 682 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 683 684 static void tcp_tsq_handler(struct sock *sk) 685 { 686 if ((1 << sk->sk_state) & 687 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 688 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 689 tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC); 690 } 691 /* 692 * One tasklest per cpu tries to send more skbs. 693 * We run in tasklet context but need to disable irqs when 694 * transfering tsq->head because tcp_wfree() might 695 * interrupt us (non NAPI drivers) 696 */ 697 static void tcp_tasklet_func(unsigned long data) 698 { 699 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 700 LIST_HEAD(list); 701 unsigned long flags; 702 struct list_head *q, *n; 703 struct tcp_sock *tp; 704 struct sock *sk; 705 706 local_irq_save(flags); 707 list_splice_init(&tsq->head, &list); 708 local_irq_restore(flags); 709 710 list_for_each_safe(q, n, &list) { 711 tp = list_entry(q, struct tcp_sock, tsq_node); 712 list_del(&tp->tsq_node); 713 714 sk = (struct sock *)tp; 715 bh_lock_sock(sk); 716 717 if (!sock_owned_by_user(sk)) { 718 tcp_tsq_handler(sk); 719 } else { 720 /* defer the work to tcp_release_cb() */ 721 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 722 } 723 bh_unlock_sock(sk); 724 725 clear_bit(TSQ_QUEUED, &tp->tsq_flags); 726 sk_free(sk); 727 } 728 } 729 730 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 731 (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 732 (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 733 (1UL << TCP_MTU_REDUCED_DEFERRED)) 734 /** 735 * tcp_release_cb - tcp release_sock() callback 736 * @sk: socket 737 * 738 * called from release_sock() to perform protocol dependent 739 * actions before socket release. 740 */ 741 void tcp_release_cb(struct sock *sk) 742 { 743 struct tcp_sock *tp = tcp_sk(sk); 744 unsigned long flags, nflags; 745 746 /* perform an atomic operation only if at least one flag is set */ 747 do { 748 flags = tp->tsq_flags; 749 if (!(flags & TCP_DEFERRED_ALL)) 750 return; 751 nflags = flags & ~TCP_DEFERRED_ALL; 752 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 753 754 if (flags & (1UL << TCP_TSQ_DEFERRED)) 755 tcp_tsq_handler(sk); 756 757 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 758 tcp_write_timer_handler(sk); 759 __sock_put(sk); 760 } 761 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 762 tcp_delack_timer_handler(sk); 763 __sock_put(sk); 764 } 765 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 766 sk->sk_prot->mtu_reduced(sk); 767 __sock_put(sk); 768 } 769 } 770 EXPORT_SYMBOL(tcp_release_cb); 771 772 void __init tcp_tasklet_init(void) 773 { 774 int i; 775 776 for_each_possible_cpu(i) { 777 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 778 779 INIT_LIST_HEAD(&tsq->head); 780 tasklet_init(&tsq->tasklet, 781 tcp_tasklet_func, 782 (unsigned long)tsq); 783 } 784 } 785 786 /* 787 * Write buffer destructor automatically called from kfree_skb. 788 * We cant xmit new skbs from this context, as we might already 789 * hold qdisc lock. 790 */ 791 void tcp_wfree(struct sk_buff *skb) 792 { 793 struct sock *sk = skb->sk; 794 struct tcp_sock *tp = tcp_sk(sk); 795 796 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 797 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 798 unsigned long flags; 799 struct tsq_tasklet *tsq; 800 801 /* Keep a ref on socket. 802 * This last ref will be released in tcp_tasklet_func() 803 */ 804 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); 805 806 /* queue this socket to tasklet queue */ 807 local_irq_save(flags); 808 tsq = &__get_cpu_var(tsq_tasklet); 809 list_add(&tp->tsq_node, &tsq->head); 810 tasklet_schedule(&tsq->tasklet); 811 local_irq_restore(flags); 812 } else { 813 sock_wfree(skb); 814 } 815 } 816 817 /* This routine actually transmits TCP packets queued in by 818 * tcp_do_sendmsg(). This is used by both the initial 819 * transmission and possible later retransmissions. 820 * All SKB's seen here are completely headerless. It is our 821 * job to build the TCP header, and pass the packet down to 822 * IP so it can do the same plus pass the packet off to the 823 * device. 824 * 825 * We are working here with either a clone of the original 826 * SKB, or a fresh unique copy made by the retransmit engine. 827 */ 828 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 829 gfp_t gfp_mask) 830 { 831 const struct inet_connection_sock *icsk = inet_csk(sk); 832 struct inet_sock *inet; 833 struct tcp_sock *tp; 834 struct tcp_skb_cb *tcb; 835 struct tcp_out_options opts; 836 unsigned int tcp_options_size, tcp_header_size; 837 struct tcp_md5sig_key *md5; 838 struct tcphdr *th; 839 int err; 840 841 BUG_ON(!skb || !tcp_skb_pcount(skb)); 842 843 /* If congestion control is doing timestamping, we must 844 * take such a timestamp before we potentially clone/copy. 845 */ 846 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 847 __net_timestamp(skb); 848 849 if (likely(clone_it)) { 850 const struct sk_buff *fclone = skb + 1; 851 852 if (unlikely(skb->fclone == SKB_FCLONE_ORIG && 853 fclone->fclone == SKB_FCLONE_CLONE)) 854 NET_INC_STATS_BH(sock_net(sk), 855 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 856 857 if (unlikely(skb_cloned(skb))) 858 skb = pskb_copy(skb, gfp_mask); 859 else 860 skb = skb_clone(skb, gfp_mask); 861 if (unlikely(!skb)) 862 return -ENOBUFS; 863 } 864 865 inet = inet_sk(sk); 866 tp = tcp_sk(sk); 867 tcb = TCP_SKB_CB(skb); 868 memset(&opts, 0, sizeof(opts)); 869 870 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 871 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 872 else 873 tcp_options_size = tcp_established_options(sk, skb, &opts, 874 &md5); 875 tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 876 877 if (tcp_packets_in_flight(tp) == 0) 878 tcp_ca_event(sk, CA_EVENT_TX_START); 879 880 /* if no packet is in qdisc/device queue, then allow XPS to select 881 * another queue. 882 */ 883 skb->ooo_okay = sk_wmem_alloc_get(sk) == 0; 884 885 skb_push(skb, tcp_header_size); 886 skb_reset_transport_header(skb); 887 888 skb_orphan(skb); 889 skb->sk = sk; 890 skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ? 891 tcp_wfree : sock_wfree; 892 atomic_add(skb->truesize, &sk->sk_wmem_alloc); 893 894 /* Build TCP header and checksum it. */ 895 th = tcp_hdr(skb); 896 th->source = inet->inet_sport; 897 th->dest = inet->inet_dport; 898 th->seq = htonl(tcb->seq); 899 th->ack_seq = htonl(tp->rcv_nxt); 900 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 901 tcb->tcp_flags); 902 903 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 904 /* RFC1323: The window in SYN & SYN/ACK segments 905 * is never scaled. 906 */ 907 th->window = htons(min(tp->rcv_wnd, 65535U)); 908 } else { 909 th->window = htons(tcp_select_window(sk)); 910 } 911 th->check = 0; 912 th->urg_ptr = 0; 913 914 /* The urg_mode check is necessary during a below snd_una win probe */ 915 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 916 if (before(tp->snd_up, tcb->seq + 0x10000)) { 917 th->urg_ptr = htons(tp->snd_up - tcb->seq); 918 th->urg = 1; 919 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 920 th->urg_ptr = htons(0xFFFF); 921 th->urg = 1; 922 } 923 } 924 925 tcp_options_write((__be32 *)(th + 1), tp, &opts); 926 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 927 TCP_ECN_send(sk, skb, tcp_header_size); 928 929 #ifdef CONFIG_TCP_MD5SIG 930 /* Calculate the MD5 hash, as we have all we need now */ 931 if (md5) { 932 sk_nocaps_add(sk, NETIF_F_GSO_MASK); 933 tp->af_specific->calc_md5_hash(opts.hash_location, 934 md5, sk, NULL, skb); 935 } 936 #endif 937 938 icsk->icsk_af_ops->send_check(sk, skb); 939 940 if (likely(tcb->tcp_flags & TCPHDR_ACK)) 941 tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 942 943 if (skb->len != tcp_header_size) 944 tcp_event_data_sent(tp, sk); 945 946 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 947 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 948 tcp_skb_pcount(skb)); 949 950 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); 951 if (likely(err <= 0)) 952 return err; 953 954 tcp_enter_cwr(sk, 1); 955 956 return net_xmit_eval(err); 957 } 958 959 /* This routine just queues the buffer for sending. 960 * 961 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 962 * otherwise socket can stall. 963 */ 964 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 965 { 966 struct tcp_sock *tp = tcp_sk(sk); 967 968 /* Advance write_seq and place onto the write_queue. */ 969 tp->write_seq = TCP_SKB_CB(skb)->end_seq; 970 skb_header_release(skb); 971 tcp_add_write_queue_tail(sk, skb); 972 sk->sk_wmem_queued += skb->truesize; 973 sk_mem_charge(sk, skb->truesize); 974 } 975 976 /* Initialize TSO segments for a packet. */ 977 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 978 unsigned int mss_now) 979 { 980 if (skb->len <= mss_now || !sk_can_gso(sk) || 981 skb->ip_summed == CHECKSUM_NONE) { 982 /* Avoid the costly divide in the normal 983 * non-TSO case. 984 */ 985 skb_shinfo(skb)->gso_segs = 1; 986 skb_shinfo(skb)->gso_size = 0; 987 skb_shinfo(skb)->gso_type = 0; 988 } else { 989 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 990 skb_shinfo(skb)->gso_size = mss_now; 991 skb_shinfo(skb)->gso_type = sk->sk_gso_type; 992 } 993 } 994 995 /* When a modification to fackets out becomes necessary, we need to check 996 * skb is counted to fackets_out or not. 997 */ 998 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 999 int decr) 1000 { 1001 struct tcp_sock *tp = tcp_sk(sk); 1002 1003 if (!tp->sacked_out || tcp_is_reno(tp)) 1004 return; 1005 1006 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 1007 tp->fackets_out -= decr; 1008 } 1009 1010 /* Pcount in the middle of the write queue got changed, we need to do various 1011 * tweaks to fix counters 1012 */ 1013 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1014 { 1015 struct tcp_sock *tp = tcp_sk(sk); 1016 1017 tp->packets_out -= decr; 1018 1019 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1020 tp->sacked_out -= decr; 1021 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1022 tp->retrans_out -= decr; 1023 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1024 tp->lost_out -= decr; 1025 1026 /* Reno case is special. Sigh... */ 1027 if (tcp_is_reno(tp) && decr > 0) 1028 tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1029 1030 tcp_adjust_fackets_out(sk, skb, decr); 1031 1032 if (tp->lost_skb_hint && 1033 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1034 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1035 tp->lost_cnt_hint -= decr; 1036 1037 tcp_verify_left_out(tp); 1038 } 1039 1040 /* Function to create two new TCP segments. Shrinks the given segment 1041 * to the specified size and appends a new segment with the rest of the 1042 * packet to the list. This won't be called frequently, I hope. 1043 * Remember, these are still headerless SKBs at this point. 1044 */ 1045 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 1046 unsigned int mss_now) 1047 { 1048 struct tcp_sock *tp = tcp_sk(sk); 1049 struct sk_buff *buff; 1050 int nsize, old_factor; 1051 int nlen; 1052 u8 flags; 1053 1054 if (WARN_ON(len > skb->len)) 1055 return -EINVAL; 1056 1057 nsize = skb_headlen(skb) - len; 1058 if (nsize < 0) 1059 nsize = 0; 1060 1061 if (skb_cloned(skb) && 1062 skb_is_nonlinear(skb) && 1063 pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 1064 return -ENOMEM; 1065 1066 /* Get a new skb... force flag on. */ 1067 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 1068 if (buff == NULL) 1069 return -ENOMEM; /* We'll just try again later. */ 1070 1071 sk->sk_wmem_queued += buff->truesize; 1072 sk_mem_charge(sk, buff->truesize); 1073 nlen = skb->len - len - nsize; 1074 buff->truesize += nlen; 1075 skb->truesize -= nlen; 1076 1077 /* Correct the sequence numbers. */ 1078 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1079 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1080 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1081 1082 /* PSH and FIN should only be set in the second packet. */ 1083 flags = TCP_SKB_CB(skb)->tcp_flags; 1084 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 1085 TCP_SKB_CB(buff)->tcp_flags = flags; 1086 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1087 1088 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 1089 /* Copy and checksum data tail into the new buffer. */ 1090 buff->csum = csum_partial_copy_nocheck(skb->data + len, 1091 skb_put(buff, nsize), 1092 nsize, 0); 1093 1094 skb_trim(skb, len); 1095 1096 skb->csum = csum_block_sub(skb->csum, buff->csum, len); 1097 } else { 1098 skb->ip_summed = CHECKSUM_PARTIAL; 1099 skb_split(skb, buff, len); 1100 } 1101 1102 buff->ip_summed = skb->ip_summed; 1103 1104 /* Looks stupid, but our code really uses when of 1105 * skbs, which it never sent before. --ANK 1106 */ 1107 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1108 buff->tstamp = skb->tstamp; 1109 1110 old_factor = tcp_skb_pcount(skb); 1111 1112 /* Fix up tso_factor for both original and new SKB. */ 1113 tcp_set_skb_tso_segs(sk, skb, mss_now); 1114 tcp_set_skb_tso_segs(sk, buff, mss_now); 1115 1116 /* If this packet has been sent out already, we must 1117 * adjust the various packet counters. 1118 */ 1119 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 1120 int diff = old_factor - tcp_skb_pcount(skb) - 1121 tcp_skb_pcount(buff); 1122 1123 if (diff) 1124 tcp_adjust_pcount(sk, skb, diff); 1125 } 1126 1127 /* Link BUFF into the send queue. */ 1128 skb_header_release(buff); 1129 tcp_insert_write_queue_after(skb, buff, sk); 1130 1131 return 0; 1132 } 1133 1134 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 1135 * eventually). The difference is that pulled data not copied, but 1136 * immediately discarded. 1137 */ 1138 static void __pskb_trim_head(struct sk_buff *skb, int len) 1139 { 1140 int i, k, eat; 1141 1142 eat = min_t(int, len, skb_headlen(skb)); 1143 if (eat) { 1144 __skb_pull(skb, eat); 1145 len -= eat; 1146 if (!len) 1147 return; 1148 } 1149 eat = len; 1150 k = 0; 1151 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 1152 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); 1153 1154 if (size <= eat) { 1155 skb_frag_unref(skb, i); 1156 eat -= size; 1157 } else { 1158 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 1159 if (eat) { 1160 skb_shinfo(skb)->frags[k].page_offset += eat; 1161 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); 1162 eat = 0; 1163 } 1164 k++; 1165 } 1166 } 1167 skb_shinfo(skb)->nr_frags = k; 1168 1169 skb_reset_tail_pointer(skb); 1170 skb->data_len -= len; 1171 skb->len = skb->data_len; 1172 } 1173 1174 /* Remove acked data from a packet in the transmit queue. */ 1175 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 1176 { 1177 if (skb_unclone(skb, GFP_ATOMIC)) 1178 return -ENOMEM; 1179 1180 __pskb_trim_head(skb, len); 1181 1182 TCP_SKB_CB(skb)->seq += len; 1183 skb->ip_summed = CHECKSUM_PARTIAL; 1184 1185 skb->truesize -= len; 1186 sk->sk_wmem_queued -= len; 1187 sk_mem_uncharge(sk, len); 1188 sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 1189 1190 /* Any change of skb->len requires recalculation of tso factor. */ 1191 if (tcp_skb_pcount(skb) > 1) 1192 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 1193 1194 return 0; 1195 } 1196 1197 /* Calculate MSS not accounting any TCP options. */ 1198 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 1199 { 1200 const struct tcp_sock *tp = tcp_sk(sk); 1201 const struct inet_connection_sock *icsk = inet_csk(sk); 1202 int mss_now; 1203 1204 /* Calculate base mss without TCP options: 1205 It is MMS_S - sizeof(tcphdr) of rfc1122 1206 */ 1207 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 1208 1209 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 1210 if (icsk->icsk_af_ops->net_frag_header_len) { 1211 const struct dst_entry *dst = __sk_dst_get(sk); 1212 1213 if (dst && dst_allfrag(dst)) 1214 mss_now -= icsk->icsk_af_ops->net_frag_header_len; 1215 } 1216 1217 /* Clamp it (mss_clamp does not include tcp options) */ 1218 if (mss_now > tp->rx_opt.mss_clamp) 1219 mss_now = tp->rx_opt.mss_clamp; 1220 1221 /* Now subtract optional transport overhead */ 1222 mss_now -= icsk->icsk_ext_hdr_len; 1223 1224 /* Then reserve room for full set of TCP options and 8 bytes of data */ 1225 if (mss_now < 48) 1226 mss_now = 48; 1227 return mss_now; 1228 } 1229 1230 /* Calculate MSS. Not accounting for SACKs here. */ 1231 int tcp_mtu_to_mss(struct sock *sk, int pmtu) 1232 { 1233 /* Subtract TCP options size, not including SACKs */ 1234 return __tcp_mtu_to_mss(sk, pmtu) - 1235 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 1236 } 1237 1238 /* Inverse of above */ 1239 int tcp_mss_to_mtu(struct sock *sk, int mss) 1240 { 1241 const struct tcp_sock *tp = tcp_sk(sk); 1242 const struct inet_connection_sock *icsk = inet_csk(sk); 1243 int mtu; 1244 1245 mtu = mss + 1246 tp->tcp_header_len + 1247 icsk->icsk_ext_hdr_len + 1248 icsk->icsk_af_ops->net_header_len; 1249 1250 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 1251 if (icsk->icsk_af_ops->net_frag_header_len) { 1252 const struct dst_entry *dst = __sk_dst_get(sk); 1253 1254 if (dst && dst_allfrag(dst)) 1255 mtu += icsk->icsk_af_ops->net_frag_header_len; 1256 } 1257 return mtu; 1258 } 1259 1260 /* MTU probing init per socket */ 1261 void tcp_mtup_init(struct sock *sk) 1262 { 1263 struct tcp_sock *tp = tcp_sk(sk); 1264 struct inet_connection_sock *icsk = inet_csk(sk); 1265 1266 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 1267 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 1268 icsk->icsk_af_ops->net_header_len; 1269 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 1270 icsk->icsk_mtup.probe_size = 0; 1271 } 1272 EXPORT_SYMBOL(tcp_mtup_init); 1273 1274 /* This function synchronize snd mss to current pmtu/exthdr set. 1275 1276 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 1277 for TCP options, but includes only bare TCP header. 1278 1279 tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1280 It is minimum of user_mss and mss received with SYN. 1281 It also does not include TCP options. 1282 1283 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 1284 1285 tp->mss_cache is current effective sending mss, including 1286 all tcp options except for SACKs. It is evaluated, 1287 taking into account current pmtu, but never exceeds 1288 tp->rx_opt.mss_clamp. 1289 1290 NOTE1. rfc1122 clearly states that advertised MSS 1291 DOES NOT include either tcp or ip options. 1292 1293 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1294 are READ ONLY outside this function. --ANK (980731) 1295 */ 1296 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 1297 { 1298 struct tcp_sock *tp = tcp_sk(sk); 1299 struct inet_connection_sock *icsk = inet_csk(sk); 1300 int mss_now; 1301 1302 if (icsk->icsk_mtup.search_high > pmtu) 1303 icsk->icsk_mtup.search_high = pmtu; 1304 1305 mss_now = tcp_mtu_to_mss(sk, pmtu); 1306 mss_now = tcp_bound_to_half_wnd(tp, mss_now); 1307 1308 /* And store cached results */ 1309 icsk->icsk_pmtu_cookie = pmtu; 1310 if (icsk->icsk_mtup.enabled) 1311 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1312 tp->mss_cache = mss_now; 1313 1314 return mss_now; 1315 } 1316 EXPORT_SYMBOL(tcp_sync_mss); 1317 1318 /* Compute the current effective MSS, taking SACKs and IP options, 1319 * and even PMTU discovery events into account. 1320 */ 1321 unsigned int tcp_current_mss(struct sock *sk) 1322 { 1323 const struct tcp_sock *tp = tcp_sk(sk); 1324 const struct dst_entry *dst = __sk_dst_get(sk); 1325 u32 mss_now; 1326 unsigned int header_len; 1327 struct tcp_out_options opts; 1328 struct tcp_md5sig_key *md5; 1329 1330 mss_now = tp->mss_cache; 1331 1332 if (dst) { 1333 u32 mtu = dst_mtu(dst); 1334 if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 1335 mss_now = tcp_sync_mss(sk, mtu); 1336 } 1337 1338 header_len = tcp_established_options(sk, NULL, &opts, &md5) + 1339 sizeof(struct tcphdr); 1340 /* The mss_cache is sized based on tp->tcp_header_len, which assumes 1341 * some common options. If this is an odd packet (because we have SACK 1342 * blocks etc) then our calculated header_len will be different, and 1343 * we have to adjust mss_now correspondingly */ 1344 if (header_len != tp->tcp_header_len) { 1345 int delta = (int) header_len - tp->tcp_header_len; 1346 mss_now -= delta; 1347 } 1348 1349 return mss_now; 1350 } 1351 1352 /* Congestion window validation. (RFC2861) */ 1353 static void tcp_cwnd_validate(struct sock *sk) 1354 { 1355 struct tcp_sock *tp = tcp_sk(sk); 1356 1357 if (tp->packets_out >= tp->snd_cwnd) { 1358 /* Network is feed fully. */ 1359 tp->snd_cwnd_used = 0; 1360 tp->snd_cwnd_stamp = tcp_time_stamp; 1361 } else { 1362 /* Network starves. */ 1363 if (tp->packets_out > tp->snd_cwnd_used) 1364 tp->snd_cwnd_used = tp->packets_out; 1365 1366 if (sysctl_tcp_slow_start_after_idle && 1367 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1368 tcp_cwnd_application_limited(sk); 1369 } 1370 } 1371 1372 /* Returns the portion of skb which can be sent right away without 1373 * introducing MSS oddities to segment boundaries. In rare cases where 1374 * mss_now != mss_cache, we will request caller to create a small skb 1375 * per input skb which could be mostly avoided here (if desired). 1376 * 1377 * We explicitly want to create a request for splitting write queue tail 1378 * to a small skb for Nagle purposes while avoiding unnecessary modulos, 1379 * thus all the complexity (cwnd_len is always MSS multiple which we 1380 * return whenever allowed by the other factors). Basically we need the 1381 * modulo only when the receiver window alone is the limiting factor or 1382 * when we would be allowed to send the split-due-to-Nagle skb fully. 1383 */ 1384 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, 1385 unsigned int mss_now, unsigned int max_segs) 1386 { 1387 const struct tcp_sock *tp = tcp_sk(sk); 1388 u32 needed, window, max_len; 1389 1390 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1391 max_len = mss_now * max_segs; 1392 1393 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 1394 return max_len; 1395 1396 needed = min(skb->len, window); 1397 1398 if (max_len <= needed) 1399 return max_len; 1400 1401 return needed - needed % mss_now; 1402 } 1403 1404 /* Can at least one segment of SKB be sent right now, according to the 1405 * congestion window rules? If so, return how many segments are allowed. 1406 */ 1407 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1408 const struct sk_buff *skb) 1409 { 1410 u32 in_flight, cwnd; 1411 1412 /* Don't be strict about the congestion window for the final FIN. */ 1413 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 1414 tcp_skb_pcount(skb) == 1) 1415 return 1; 1416 1417 in_flight = tcp_packets_in_flight(tp); 1418 cwnd = tp->snd_cwnd; 1419 if (in_flight < cwnd) 1420 return (cwnd - in_flight); 1421 1422 return 0; 1423 } 1424 1425 /* Initialize TSO state of a skb. 1426 * This must be invoked the first time we consider transmitting 1427 * SKB onto the wire. 1428 */ 1429 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1430 unsigned int mss_now) 1431 { 1432 int tso_segs = tcp_skb_pcount(skb); 1433 1434 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1435 tcp_set_skb_tso_segs(sk, skb, mss_now); 1436 tso_segs = tcp_skb_pcount(skb); 1437 } 1438 return tso_segs; 1439 } 1440 1441 /* Minshall's variant of the Nagle send check. */ 1442 static inline bool tcp_minshall_check(const struct tcp_sock *tp) 1443 { 1444 return after(tp->snd_sml, tp->snd_una) && 1445 !after(tp->snd_sml, tp->snd_nxt); 1446 } 1447 1448 /* Return false, if packet can be sent now without violation Nagle's rules: 1449 * 1. It is full sized. 1450 * 2. Or it contains FIN. (already checked by caller) 1451 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1452 * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1453 * With Minshall's modification: all sent small packets are ACKed. 1454 */ 1455 static inline bool tcp_nagle_check(const struct tcp_sock *tp, 1456 const struct sk_buff *skb, 1457 unsigned int mss_now, int nonagle) 1458 { 1459 return skb->len < mss_now && 1460 ((nonagle & TCP_NAGLE_CORK) || 1461 (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1462 } 1463 1464 /* Return true if the Nagle test allows this packet to be 1465 * sent now. 1466 */ 1467 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1468 unsigned int cur_mss, int nonagle) 1469 { 1470 /* Nagle rule does not apply to frames, which sit in the middle of the 1471 * write_queue (they have no chances to get new data). 1472 * 1473 * This is implemented in the callers, where they modify the 'nonagle' 1474 * argument based upon the location of SKB in the send queue. 1475 */ 1476 if (nonagle & TCP_NAGLE_PUSH) 1477 return true; 1478 1479 /* Don't use the nagle rule for urgent data (or for the final FIN). */ 1480 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1481 return true; 1482 1483 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1484 return true; 1485 1486 return false; 1487 } 1488 1489 /* Does at least the first segment of SKB fit into the send window? */ 1490 static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1491 const struct sk_buff *skb, 1492 unsigned int cur_mss) 1493 { 1494 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1495 1496 if (skb->len > cur_mss) 1497 end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1498 1499 return !after(end_seq, tcp_wnd_end(tp)); 1500 } 1501 1502 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1503 * should be put on the wire right now. If so, it returns the number of 1504 * packets allowed by the congestion window. 1505 */ 1506 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1507 unsigned int cur_mss, int nonagle) 1508 { 1509 const struct tcp_sock *tp = tcp_sk(sk); 1510 unsigned int cwnd_quota; 1511 1512 tcp_init_tso_segs(sk, skb, cur_mss); 1513 1514 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1515 return 0; 1516 1517 cwnd_quota = tcp_cwnd_test(tp, skb); 1518 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1519 cwnd_quota = 0; 1520 1521 return cwnd_quota; 1522 } 1523 1524 /* Test if sending is allowed right now. */ 1525 bool tcp_may_send_now(struct sock *sk) 1526 { 1527 const struct tcp_sock *tp = tcp_sk(sk); 1528 struct sk_buff *skb = tcp_send_head(sk); 1529 1530 return skb && 1531 tcp_snd_test(sk, skb, tcp_current_mss(sk), 1532 (tcp_skb_is_last(sk, skb) ? 1533 tp->nonagle : TCP_NAGLE_PUSH)); 1534 } 1535 1536 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1537 * which is put after SKB on the list. It is very much like 1538 * tcp_fragment() except that it may make several kinds of assumptions 1539 * in order to speed up the splitting operation. In particular, we 1540 * know that all the data is in scatter-gather pages, and that the 1541 * packet has never been sent out before (and thus is not cloned). 1542 */ 1543 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1544 unsigned int mss_now, gfp_t gfp) 1545 { 1546 struct sk_buff *buff; 1547 int nlen = skb->len - len; 1548 u8 flags; 1549 1550 /* All of a TSO frame must be composed of paged data. */ 1551 if (skb->len != skb->data_len) 1552 return tcp_fragment(sk, skb, len, mss_now); 1553 1554 buff = sk_stream_alloc_skb(sk, 0, gfp); 1555 if (unlikely(buff == NULL)) 1556 return -ENOMEM; 1557 1558 sk->sk_wmem_queued += buff->truesize; 1559 sk_mem_charge(sk, buff->truesize); 1560 buff->truesize += nlen; 1561 skb->truesize -= nlen; 1562 1563 /* Correct the sequence numbers. */ 1564 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1565 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1566 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1567 1568 /* PSH and FIN should only be set in the second packet. */ 1569 flags = TCP_SKB_CB(skb)->tcp_flags; 1570 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 1571 TCP_SKB_CB(buff)->tcp_flags = flags; 1572 1573 /* This packet was never sent out yet, so no SACK bits. */ 1574 TCP_SKB_CB(buff)->sacked = 0; 1575 1576 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1577 skb_split(skb, buff, len); 1578 1579 /* Fix up tso_factor for both original and new SKB. */ 1580 tcp_set_skb_tso_segs(sk, skb, mss_now); 1581 tcp_set_skb_tso_segs(sk, buff, mss_now); 1582 1583 /* Link BUFF into the send queue. */ 1584 skb_header_release(buff); 1585 tcp_insert_write_queue_after(skb, buff, sk); 1586 1587 return 0; 1588 } 1589 1590 /* Try to defer sending, if possible, in order to minimize the amount 1591 * of TSO splitting we do. View it as a kind of TSO Nagle test. 1592 * 1593 * This algorithm is from John Heffner. 1594 */ 1595 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1596 { 1597 struct tcp_sock *tp = tcp_sk(sk); 1598 const struct inet_connection_sock *icsk = inet_csk(sk); 1599 u32 send_win, cong_win, limit, in_flight; 1600 int win_divisor; 1601 1602 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1603 goto send_now; 1604 1605 if (icsk->icsk_ca_state != TCP_CA_Open) 1606 goto send_now; 1607 1608 /* Defer for less than two clock ticks. */ 1609 if (tp->tso_deferred && 1610 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1611 goto send_now; 1612 1613 in_flight = tcp_packets_in_flight(tp); 1614 1615 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1616 1617 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1618 1619 /* From in_flight test above, we know that cwnd > in_flight. */ 1620 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1621 1622 limit = min(send_win, cong_win); 1623 1624 /* If a full-sized TSO skb can be sent, do it. */ 1625 if (limit >= min_t(unsigned int, sk->sk_gso_max_size, 1626 sk->sk_gso_max_segs * tp->mss_cache)) 1627 goto send_now; 1628 1629 /* Middle in queue won't get any more data, full sendable already? */ 1630 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 1631 goto send_now; 1632 1633 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1634 if (win_divisor) { 1635 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1636 1637 /* If at least some fraction of a window is available, 1638 * just use it. 1639 */ 1640 chunk /= win_divisor; 1641 if (limit >= chunk) 1642 goto send_now; 1643 } else { 1644 /* Different approach, try not to defer past a single 1645 * ACK. Receiver should ACK every other full sized 1646 * frame, so if we have space for more than 3 frames 1647 * then send now. 1648 */ 1649 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1650 goto send_now; 1651 } 1652 1653 /* Ok, it looks like it is advisable to defer. 1654 * Do not rearm the timer if already set to not break TCP ACK clocking. 1655 */ 1656 if (!tp->tso_deferred) 1657 tp->tso_deferred = 1 | (jiffies << 1); 1658 1659 return true; 1660 1661 send_now: 1662 tp->tso_deferred = 0; 1663 return false; 1664 } 1665 1666 /* Create a new MTU probe if we are ready. 1667 * MTU probe is regularly attempting to increase the path MTU by 1668 * deliberately sending larger packets. This discovers routing 1669 * changes resulting in larger path MTUs. 1670 * 1671 * Returns 0 if we should wait to probe (no cwnd available), 1672 * 1 if a probe was sent, 1673 * -1 otherwise 1674 */ 1675 static int tcp_mtu_probe(struct sock *sk) 1676 { 1677 struct tcp_sock *tp = tcp_sk(sk); 1678 struct inet_connection_sock *icsk = inet_csk(sk); 1679 struct sk_buff *skb, *nskb, *next; 1680 int len; 1681 int probe_size; 1682 int size_needed; 1683 int copy; 1684 int mss_now; 1685 1686 /* Not currently probing/verifying, 1687 * not in recovery, 1688 * have enough cwnd, and 1689 * not SACKing (the variable headers throw things off) */ 1690 if (!icsk->icsk_mtup.enabled || 1691 icsk->icsk_mtup.probe_size || 1692 inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 1693 tp->snd_cwnd < 11 || 1694 tp->rx_opt.num_sacks || tp->rx_opt.dsack) 1695 return -1; 1696 1697 /* Very simple search strategy: just double the MSS. */ 1698 mss_now = tcp_current_mss(sk); 1699 probe_size = 2 * tp->mss_cache; 1700 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 1701 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 1702 /* TODO: set timer for probe_converge_event */ 1703 return -1; 1704 } 1705 1706 /* Have enough data in the send queue to probe? */ 1707 if (tp->write_seq - tp->snd_nxt < size_needed) 1708 return -1; 1709 1710 if (tp->snd_wnd < size_needed) 1711 return -1; 1712 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 1713 return 0; 1714 1715 /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1716 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1717 if (!tcp_packets_in_flight(tp)) 1718 return -1; 1719 else 1720 return 0; 1721 } 1722 1723 /* We're allowed to probe. Build it now. */ 1724 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 1725 return -1; 1726 sk->sk_wmem_queued += nskb->truesize; 1727 sk_mem_charge(sk, nskb->truesize); 1728 1729 skb = tcp_send_head(sk); 1730 1731 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 1732 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 1733 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 1734 TCP_SKB_CB(nskb)->sacked = 0; 1735 nskb->csum = 0; 1736 nskb->ip_summed = skb->ip_summed; 1737 1738 tcp_insert_write_queue_before(nskb, skb, sk); 1739 1740 len = 0; 1741 tcp_for_write_queue_from_safe(skb, next, sk) { 1742 copy = min_t(int, skb->len, probe_size - len); 1743 if (nskb->ip_summed) 1744 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 1745 else 1746 nskb->csum = skb_copy_and_csum_bits(skb, 0, 1747 skb_put(nskb, copy), 1748 copy, nskb->csum); 1749 1750 if (skb->len <= copy) { 1751 /* We've eaten all the data from this skb. 1752 * Throw it away. */ 1753 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1754 tcp_unlink_write_queue(skb, sk); 1755 sk_wmem_free_skb(sk, skb); 1756 } else { 1757 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1758 ~(TCPHDR_FIN|TCPHDR_PSH); 1759 if (!skb_shinfo(skb)->nr_frags) { 1760 skb_pull(skb, copy); 1761 if (skb->ip_summed != CHECKSUM_PARTIAL) 1762 skb->csum = csum_partial(skb->data, 1763 skb->len, 0); 1764 } else { 1765 __pskb_trim_head(skb, copy); 1766 tcp_set_skb_tso_segs(sk, skb, mss_now); 1767 } 1768 TCP_SKB_CB(skb)->seq += copy; 1769 } 1770 1771 len += copy; 1772 1773 if (len >= probe_size) 1774 break; 1775 } 1776 tcp_init_tso_segs(sk, nskb, nskb->len); 1777 1778 /* We're ready to send. If this fails, the probe will 1779 * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 1780 TCP_SKB_CB(nskb)->when = tcp_time_stamp; 1781 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 1782 /* Decrement cwnd here because we are sending 1783 * effectively two packets. */ 1784 tp->snd_cwnd--; 1785 tcp_event_new_data_sent(sk, nskb); 1786 1787 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 1788 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 1789 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 1790 1791 return 1; 1792 } 1793 1794 return -1; 1795 } 1796 1797 /* This routine writes packets to the network. It advances the 1798 * send_head. This happens as incoming acks open up the remote 1799 * window for us. 1800 * 1801 * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1802 * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1803 * account rare use of URG, this is not a big flaw. 1804 * 1805 * Send at most one packet when push_one > 0. Temporarily ignore 1806 * cwnd limit to force at most one packet out when push_one == 2. 1807 1808 * Returns true, if no segments are in flight and we have queued segments, 1809 * but cannot send anything now because of SWS or another problem. 1810 */ 1811 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1812 int push_one, gfp_t gfp) 1813 { 1814 struct tcp_sock *tp = tcp_sk(sk); 1815 struct sk_buff *skb; 1816 unsigned int tso_segs, sent_pkts; 1817 int cwnd_quota; 1818 int result; 1819 1820 sent_pkts = 0; 1821 1822 if (!push_one) { 1823 /* Do MTU probing. */ 1824 result = tcp_mtu_probe(sk); 1825 if (!result) { 1826 return false; 1827 } else if (result > 0) { 1828 sent_pkts = 1; 1829 } 1830 } 1831 1832 while ((skb = tcp_send_head(sk))) { 1833 unsigned int limit; 1834 1835 1836 tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1837 BUG_ON(!tso_segs); 1838 1839 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) 1840 goto repair; /* Skip network transmission */ 1841 1842 cwnd_quota = tcp_cwnd_test(tp, skb); 1843 if (!cwnd_quota) { 1844 if (push_one == 2) 1845 /* Force out a loss probe pkt. */ 1846 cwnd_quota = 1; 1847 else 1848 break; 1849 } 1850 1851 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1852 break; 1853 1854 if (tso_segs == 1) { 1855 if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1856 (tcp_skb_is_last(sk, skb) ? 1857 nonagle : TCP_NAGLE_PUSH)))) 1858 break; 1859 } else { 1860 if (!push_one && tcp_tso_should_defer(sk, skb)) 1861 break; 1862 } 1863 1864 /* TSQ : sk_wmem_alloc accounts skb truesize, 1865 * including skb overhead. But thats OK. 1866 */ 1867 if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) { 1868 set_bit(TSQ_THROTTLED, &tp->tsq_flags); 1869 break; 1870 } 1871 limit = mss_now; 1872 if (tso_segs > 1 && !tcp_urg_mode(tp)) 1873 limit = tcp_mss_split_point(sk, skb, mss_now, 1874 min_t(unsigned int, 1875 cwnd_quota, 1876 sk->sk_gso_max_segs)); 1877 1878 if (skb->len > limit && 1879 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 1880 break; 1881 1882 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1883 1884 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 1885 break; 1886 1887 repair: 1888 /* Advance the send_head. This one is sent out. 1889 * This call will increment packets_out. 1890 */ 1891 tcp_event_new_data_sent(sk, skb); 1892 1893 tcp_minshall_update(tp, mss_now, skb); 1894 sent_pkts += tcp_skb_pcount(skb); 1895 1896 if (push_one) 1897 break; 1898 } 1899 1900 if (likely(sent_pkts)) { 1901 if (tcp_in_cwnd_reduction(sk)) 1902 tp->prr_out += sent_pkts; 1903 1904 /* Send one loss probe per tail loss episode. */ 1905 if (push_one != 2) 1906 tcp_schedule_loss_probe(sk); 1907 tcp_cwnd_validate(sk); 1908 return false; 1909 } 1910 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); 1911 } 1912 1913 bool tcp_schedule_loss_probe(struct sock *sk) 1914 { 1915 struct inet_connection_sock *icsk = inet_csk(sk); 1916 struct tcp_sock *tp = tcp_sk(sk); 1917 u32 timeout, tlp_time_stamp, rto_time_stamp; 1918 u32 rtt = tp->srtt >> 3; 1919 1920 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 1921 return false; 1922 /* No consecutive loss probes. */ 1923 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 1924 tcp_rearm_rto(sk); 1925 return false; 1926 } 1927 /* Don't do any loss probe on a Fast Open connection before 3WHS 1928 * finishes. 1929 */ 1930 if (sk->sk_state == TCP_SYN_RECV) 1931 return false; 1932 1933 /* TLP is only scheduled when next timer event is RTO. */ 1934 if (icsk->icsk_pending != ICSK_TIME_RETRANS) 1935 return false; 1936 1937 /* Schedule a loss probe in 2*RTT for SACK capable connections 1938 * in Open state, that are either limited by cwnd or application. 1939 */ 1940 if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out || 1941 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 1942 return false; 1943 1944 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 1945 tcp_send_head(sk)) 1946 return false; 1947 1948 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 1949 * for delayed ack when there's one outstanding packet. 1950 */ 1951 timeout = rtt << 1; 1952 if (tp->packets_out == 1) 1953 timeout = max_t(u32, timeout, 1954 (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 1955 timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 1956 1957 /* If RTO is shorter, just schedule TLP in its place. */ 1958 tlp_time_stamp = tcp_time_stamp + timeout; 1959 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 1960 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 1961 s32 delta = rto_time_stamp - tcp_time_stamp; 1962 if (delta > 0) 1963 timeout = delta; 1964 } 1965 1966 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 1967 TCP_RTO_MAX); 1968 return true; 1969 } 1970 1971 /* When probe timeout (PTO) fires, send a new segment if one exists, else 1972 * retransmit the last segment. 1973 */ 1974 void tcp_send_loss_probe(struct sock *sk) 1975 { 1976 struct tcp_sock *tp = tcp_sk(sk); 1977 struct sk_buff *skb; 1978 int pcount; 1979 int mss = tcp_current_mss(sk); 1980 int err = -1; 1981 1982 if (tcp_send_head(sk) != NULL) { 1983 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 1984 goto rearm_timer; 1985 } 1986 1987 /* At most one outstanding TLP retransmission. */ 1988 if (tp->tlp_high_seq) 1989 goto rearm_timer; 1990 1991 /* Retransmit last segment. */ 1992 skb = tcp_write_queue_tail(sk); 1993 if (WARN_ON(!skb)) 1994 goto rearm_timer; 1995 1996 pcount = tcp_skb_pcount(skb); 1997 if (WARN_ON(!pcount)) 1998 goto rearm_timer; 1999 2000 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 2001 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss))) 2002 goto rearm_timer; 2003 skb = tcp_write_queue_tail(sk); 2004 } 2005 2006 if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 2007 goto rearm_timer; 2008 2009 /* Probe with zero data doesn't trigger fast recovery. */ 2010 if (skb->len > 0) 2011 err = __tcp_retransmit_skb(sk, skb); 2012 2013 /* Record snd_nxt for loss detection. */ 2014 if (likely(!err)) 2015 tp->tlp_high_seq = tp->snd_nxt; 2016 2017 rearm_timer: 2018 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2019 inet_csk(sk)->icsk_rto, 2020 TCP_RTO_MAX); 2021 2022 if (likely(!err)) 2023 NET_INC_STATS_BH(sock_net(sk), 2024 LINUX_MIB_TCPLOSSPROBES); 2025 return; 2026 } 2027 2028 /* Push out any pending frames which were held back due to 2029 * TCP_CORK or attempt at coalescing tiny packets. 2030 * The socket must be locked by the caller. 2031 */ 2032 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 2033 int nonagle) 2034 { 2035 /* If we are closed, the bytes will have to remain here. 2036 * In time closedown will finish, we empty the write queue and 2037 * all will be happy. 2038 */ 2039 if (unlikely(sk->sk_state == TCP_CLOSE)) 2040 return; 2041 2042 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 2043 sk_gfp_atomic(sk, GFP_ATOMIC))) 2044 tcp_check_probe_timer(sk); 2045 } 2046 2047 /* Send _single_ skb sitting at the send head. This function requires 2048 * true push pending frames to setup probe timer etc. 2049 */ 2050 void tcp_push_one(struct sock *sk, unsigned int mss_now) 2051 { 2052 struct sk_buff *skb = tcp_send_head(sk); 2053 2054 BUG_ON(!skb || skb->len < mss_now); 2055 2056 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2057 } 2058 2059 /* This function returns the amount that we can raise the 2060 * usable window based on the following constraints 2061 * 2062 * 1. The window can never be shrunk once it is offered (RFC 793) 2063 * 2. We limit memory per socket 2064 * 2065 * RFC 1122: 2066 * "the suggested [SWS] avoidance algorithm for the receiver is to keep 2067 * RECV.NEXT + RCV.WIN fixed until: 2068 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 2069 * 2070 * i.e. don't raise the right edge of the window until you can raise 2071 * it at least MSS bytes. 2072 * 2073 * Unfortunately, the recommended algorithm breaks header prediction, 2074 * since header prediction assumes th->window stays fixed. 2075 * 2076 * Strictly speaking, keeping th->window fixed violates the receiver 2077 * side SWS prevention criteria. The problem is that under this rule 2078 * a stream of single byte packets will cause the right side of the 2079 * window to always advance by a single byte. 2080 * 2081 * Of course, if the sender implements sender side SWS prevention 2082 * then this will not be a problem. 2083 * 2084 * BSD seems to make the following compromise: 2085 * 2086 * If the free space is less than the 1/4 of the maximum 2087 * space available and the free space is less than 1/2 mss, 2088 * then set the window to 0. 2089 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 2090 * Otherwise, just prevent the window from shrinking 2091 * and from being larger than the largest representable value. 2092 * 2093 * This prevents incremental opening of the window in the regime 2094 * where TCP is limited by the speed of the reader side taking 2095 * data out of the TCP receive queue. It does nothing about 2096 * those cases where the window is constrained on the sender side 2097 * because the pipeline is full. 2098 * 2099 * BSD also seems to "accidentally" limit itself to windows that are a 2100 * multiple of MSS, at least until the free space gets quite small. 2101 * This would appear to be a side effect of the mbuf implementation. 2102 * Combining these two algorithms results in the observed behavior 2103 * of having a fixed window size at almost all times. 2104 * 2105 * Below we obtain similar behavior by forcing the offered window to 2106 * a multiple of the mss when it is feasible to do so. 2107 * 2108 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 2109 * Regular options like TIMESTAMP are taken into account. 2110 */ 2111 u32 __tcp_select_window(struct sock *sk) 2112 { 2113 struct inet_connection_sock *icsk = inet_csk(sk); 2114 struct tcp_sock *tp = tcp_sk(sk); 2115 /* MSS for the peer's data. Previous versions used mss_clamp 2116 * here. I don't know if the value based on our guesses 2117 * of peer's MSS is better for the performance. It's more correct 2118 * but may be worse for the performance because of rcv_mss 2119 * fluctuations. --SAW 1998/11/1 2120 */ 2121 int mss = icsk->icsk_ack.rcv_mss; 2122 int free_space = tcp_space(sk); 2123 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 2124 int window; 2125 2126 if (mss > full_space) 2127 mss = full_space; 2128 2129 if (free_space < (full_space >> 1)) { 2130 icsk->icsk_ack.quick = 0; 2131 2132 if (sk_under_memory_pressure(sk)) 2133 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2134 4U * tp->advmss); 2135 2136 if (free_space < mss) 2137 return 0; 2138 } 2139 2140 if (free_space > tp->rcv_ssthresh) 2141 free_space = tp->rcv_ssthresh; 2142 2143 /* Don't do rounding if we are using window scaling, since the 2144 * scaled window will not line up with the MSS boundary anyway. 2145 */ 2146 window = tp->rcv_wnd; 2147 if (tp->rx_opt.rcv_wscale) { 2148 window = free_space; 2149 2150 /* Advertise enough space so that it won't get scaled away. 2151 * Import case: prevent zero window announcement if 2152 * 1<<rcv_wscale > mss. 2153 */ 2154 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 2155 window = (((window >> tp->rx_opt.rcv_wscale) + 1) 2156 << tp->rx_opt.rcv_wscale); 2157 } else { 2158 /* Get the largest window that is a nice multiple of mss. 2159 * Window clamp already applied above. 2160 * If our current window offering is within 1 mss of the 2161 * free space we just keep it. This prevents the divide 2162 * and multiply from happening most of the time. 2163 * We also don't do any window rounding when the free space 2164 * is too small. 2165 */ 2166 if (window <= free_space - mss || window > free_space) 2167 window = (free_space / mss) * mss; 2168 else if (mss == full_space && 2169 free_space > window + (full_space >> 1)) 2170 window = free_space; 2171 } 2172 2173 return window; 2174 } 2175 2176 /* Collapses two adjacent SKB's during retransmission. */ 2177 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 2178 { 2179 struct tcp_sock *tp = tcp_sk(sk); 2180 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2181 int skb_size, next_skb_size; 2182 2183 skb_size = skb->len; 2184 next_skb_size = next_skb->len; 2185 2186 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 2187 2188 tcp_highest_sack_combine(sk, next_skb, skb); 2189 2190 tcp_unlink_write_queue(next_skb, sk); 2191 2192 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 2193 next_skb_size); 2194 2195 if (next_skb->ip_summed == CHECKSUM_PARTIAL) 2196 skb->ip_summed = CHECKSUM_PARTIAL; 2197 2198 if (skb->ip_summed != CHECKSUM_PARTIAL) 2199 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 2200 2201 /* Update sequence range on original skb. */ 2202 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 2203 2204 /* Merge over control information. This moves PSH/FIN etc. over */ 2205 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 2206 2207 /* All done, get rid of second SKB and account for it so 2208 * packet counting does not break. 2209 */ 2210 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2211 2212 /* changed transmit queue under us so clear hints */ 2213 tcp_clear_retrans_hints_partial(tp); 2214 if (next_skb == tp->retransmit_skb_hint) 2215 tp->retransmit_skb_hint = skb; 2216 2217 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2218 2219 sk_wmem_free_skb(sk, next_skb); 2220 } 2221 2222 /* Check if coalescing SKBs is legal. */ 2223 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 2224 { 2225 if (tcp_skb_pcount(skb) > 1) 2226 return false; 2227 /* TODO: SACK collapsing could be used to remove this condition */ 2228 if (skb_shinfo(skb)->nr_frags != 0) 2229 return false; 2230 if (skb_cloned(skb)) 2231 return false; 2232 if (skb == tcp_send_head(sk)) 2233 return false; 2234 /* Some heurestics for collapsing over SACK'd could be invented */ 2235 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2236 return false; 2237 2238 return true; 2239 } 2240 2241 /* Collapse packets in the retransmit queue to make to create 2242 * less packets on the wire. This is only done on retransmission. 2243 */ 2244 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 2245 int space) 2246 { 2247 struct tcp_sock *tp = tcp_sk(sk); 2248 struct sk_buff *skb = to, *tmp; 2249 bool first = true; 2250 2251 if (!sysctl_tcp_retrans_collapse) 2252 return; 2253 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2254 return; 2255 2256 tcp_for_write_queue_from_safe(skb, tmp, sk) { 2257 if (!tcp_can_collapse(sk, skb)) 2258 break; 2259 2260 space -= skb->len; 2261 2262 if (first) { 2263 first = false; 2264 continue; 2265 } 2266 2267 if (space < 0) 2268 break; 2269 /* Punt if not enough space exists in the first SKB for 2270 * the data in the second 2271 */ 2272 if (skb->len > skb_availroom(to)) 2273 break; 2274 2275 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 2276 break; 2277 2278 tcp_collapse_retrans(sk, to); 2279 } 2280 } 2281 2282 /* This retransmits one SKB. Policy decisions and retransmit queue 2283 * state updates are done by the caller. Returns non-zero if an 2284 * error occurred which prevented the send. 2285 */ 2286 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 2287 { 2288 struct tcp_sock *tp = tcp_sk(sk); 2289 struct inet_connection_sock *icsk = inet_csk(sk); 2290 unsigned int cur_mss; 2291 2292 /* Inconslusive MTU probe */ 2293 if (icsk->icsk_mtup.probe_size) { 2294 icsk->icsk_mtup.probe_size = 0; 2295 } 2296 2297 /* Do not sent more than we queued. 1/4 is reserved for possible 2298 * copying overhead: fragmentation, tunneling, mangling etc. 2299 */ 2300 if (atomic_read(&sk->sk_wmem_alloc) > 2301 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 2302 return -EAGAIN; 2303 2304 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 2305 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 2306 BUG(); 2307 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 2308 return -ENOMEM; 2309 } 2310 2311 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 2312 return -EHOSTUNREACH; /* Routing failure or similar. */ 2313 2314 cur_mss = tcp_current_mss(sk); 2315 2316 /* If receiver has shrunk his window, and skb is out of 2317 * new window, do not retransmit it. The exception is the 2318 * case, when window is shrunk to zero. In this case 2319 * our retransmit serves as a zero window probe. 2320 */ 2321 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 2322 TCP_SKB_CB(skb)->seq != tp->snd_una) 2323 return -EAGAIN; 2324 2325 if (skb->len > cur_mss) { 2326 if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 2327 return -ENOMEM; /* We'll try again later. */ 2328 } else { 2329 int oldpcount = tcp_skb_pcount(skb); 2330 2331 if (unlikely(oldpcount > 1)) { 2332 tcp_init_tso_segs(sk, skb, cur_mss); 2333 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 2334 } 2335 } 2336 2337 tcp_retrans_try_collapse(sk, skb, cur_mss); 2338 2339 /* Some Solaris stacks overoptimize and ignore the FIN on a 2340 * retransmit when old data is attached. So strip it off 2341 * since it is cheap to do so and saves bytes on the network. 2342 */ 2343 if (skb->len > 0 && 2344 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 2345 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 2346 if (!pskb_trim(skb, 0)) { 2347 /* Reuse, even though it does some unnecessary work */ 2348 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 2349 TCP_SKB_CB(skb)->tcp_flags); 2350 skb->ip_summed = CHECKSUM_NONE; 2351 } 2352 } 2353 2354 /* Make a copy, if the first transmission SKB clone we made 2355 * is still in somebody's hands, else make a clone. 2356 */ 2357 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2358 2359 /* make sure skb->data is aligned on arches that require it 2360 * and check if ack-trimming & collapsing extended the headroom 2361 * beyond what csum_start can cover. 2362 */ 2363 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 2364 skb_headroom(skb) >= 0xFFFF)) { 2365 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2366 GFP_ATOMIC); 2367 return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2368 -ENOBUFS; 2369 } else { 2370 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2371 } 2372 } 2373 2374 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 2375 { 2376 struct tcp_sock *tp = tcp_sk(sk); 2377 int err = __tcp_retransmit_skb(sk, skb); 2378 2379 if (err == 0) { 2380 /* Update global TCP statistics. */ 2381 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 2382 2383 tp->total_retrans++; 2384 2385 #if FASTRETRANS_DEBUG > 0 2386 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2387 net_dbg_ratelimited("retrans_out leaked\n"); 2388 } 2389 #endif 2390 if (!tp->retrans_out) 2391 tp->lost_retrans_low = tp->snd_nxt; 2392 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 2393 tp->retrans_out += tcp_skb_pcount(skb); 2394 2395 /* Save stamp of the first retransmit. */ 2396 if (!tp->retrans_stamp) 2397 tp->retrans_stamp = TCP_SKB_CB(skb)->when; 2398 2399 tp->undo_retrans += tcp_skb_pcount(skb); 2400 2401 /* snd_nxt is stored to detect loss of retransmitted segment, 2402 * see tcp_input.c tcp_sacktag_write_queue(). 2403 */ 2404 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 2405 } 2406 return err; 2407 } 2408 2409 /* Check if we forward retransmits are possible in the current 2410 * window/congestion state. 2411 */ 2412 static bool tcp_can_forward_retransmit(struct sock *sk) 2413 { 2414 const struct inet_connection_sock *icsk = inet_csk(sk); 2415 const struct tcp_sock *tp = tcp_sk(sk); 2416 2417 /* Forward retransmissions are possible only during Recovery. */ 2418 if (icsk->icsk_ca_state != TCP_CA_Recovery) 2419 return false; 2420 2421 /* No forward retransmissions in Reno are possible. */ 2422 if (tcp_is_reno(tp)) 2423 return false; 2424 2425 /* Yeah, we have to make difficult choice between forward transmission 2426 * and retransmission... Both ways have their merits... 2427 * 2428 * For now we do not retransmit anything, while we have some new 2429 * segments to send. In the other cases, follow rule 3 for 2430 * NextSeg() specified in RFC3517. 2431 */ 2432 2433 if (tcp_may_send_now(sk)) 2434 return false; 2435 2436 return true; 2437 } 2438 2439 /* This gets called after a retransmit timeout, and the initially 2440 * retransmitted data is acknowledged. It tries to continue 2441 * resending the rest of the retransmit queue, until either 2442 * we've sent it all or the congestion window limit is reached. 2443 * If doing SACK, the first ACK which comes back for a timeout 2444 * based retransmit packet might feed us FACK information again. 2445 * If so, we use it to avoid unnecessarily retransmissions. 2446 */ 2447 void tcp_xmit_retransmit_queue(struct sock *sk) 2448 { 2449 const struct inet_connection_sock *icsk = inet_csk(sk); 2450 struct tcp_sock *tp = tcp_sk(sk); 2451 struct sk_buff *skb; 2452 struct sk_buff *hole = NULL; 2453 u32 last_lost; 2454 int mib_idx; 2455 int fwd_rexmitting = 0; 2456 2457 if (!tp->packets_out) 2458 return; 2459 2460 if (!tp->lost_out) 2461 tp->retransmit_high = tp->snd_una; 2462 2463 if (tp->retransmit_skb_hint) { 2464 skb = tp->retransmit_skb_hint; 2465 last_lost = TCP_SKB_CB(skb)->end_seq; 2466 if (after(last_lost, tp->retransmit_high)) 2467 last_lost = tp->retransmit_high; 2468 } else { 2469 skb = tcp_write_queue_head(sk); 2470 last_lost = tp->snd_una; 2471 } 2472 2473 tcp_for_write_queue_from(skb, sk) { 2474 __u8 sacked = TCP_SKB_CB(skb)->sacked; 2475 2476 if (skb == tcp_send_head(sk)) 2477 break; 2478 /* we could do better than to assign each time */ 2479 if (hole == NULL) 2480 tp->retransmit_skb_hint = skb; 2481 2482 /* Assume this retransmit will generate 2483 * only one packet for congestion window 2484 * calculation purposes. This works because 2485 * tcp_retransmit_skb() will chop up the 2486 * packet to be MSS sized and all the 2487 * packet counting works out. 2488 */ 2489 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 2490 return; 2491 2492 if (fwd_rexmitting) { 2493 begin_fwd: 2494 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2495 break; 2496 mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 2497 2498 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2499 tp->retransmit_high = last_lost; 2500 if (!tcp_can_forward_retransmit(sk)) 2501 break; 2502 /* Backtrack if necessary to non-L'ed skb */ 2503 if (hole != NULL) { 2504 skb = hole; 2505 hole = NULL; 2506 } 2507 fwd_rexmitting = 1; 2508 goto begin_fwd; 2509 2510 } else if (!(sacked & TCPCB_LOST)) { 2511 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 2512 hole = skb; 2513 continue; 2514 2515 } else { 2516 last_lost = TCP_SKB_CB(skb)->end_seq; 2517 if (icsk->icsk_ca_state != TCP_CA_Loss) 2518 mib_idx = LINUX_MIB_TCPFASTRETRANS; 2519 else 2520 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 2521 } 2522 2523 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 2524 continue; 2525 2526 if (tcp_retransmit_skb(sk, skb)) { 2527 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2528 return; 2529 } 2530 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2531 2532 if (tcp_in_cwnd_reduction(sk)) 2533 tp->prr_out += tcp_skb_pcount(skb); 2534 2535 if (skb == tcp_write_queue_head(sk)) 2536 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2537 inet_csk(sk)->icsk_rto, 2538 TCP_RTO_MAX); 2539 } 2540 } 2541 2542 /* Send a fin. The caller locks the socket for us. This cannot be 2543 * allowed to fail queueing a FIN frame under any circumstances. 2544 */ 2545 void tcp_send_fin(struct sock *sk) 2546 { 2547 struct tcp_sock *tp = tcp_sk(sk); 2548 struct sk_buff *skb = tcp_write_queue_tail(sk); 2549 int mss_now; 2550 2551 /* Optimization, tack on the FIN if we have a queue of 2552 * unsent frames. But be careful about outgoing SACKS 2553 * and IP options. 2554 */ 2555 mss_now = tcp_current_mss(sk); 2556 2557 if (tcp_send_head(sk) != NULL) { 2558 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 2559 TCP_SKB_CB(skb)->end_seq++; 2560 tp->write_seq++; 2561 } else { 2562 /* Socket is locked, keep trying until memory is available. */ 2563 for (;;) { 2564 skb = alloc_skb_fclone(MAX_TCP_HEADER, 2565 sk->sk_allocation); 2566 if (skb) 2567 break; 2568 yield(); 2569 } 2570 2571 /* Reserve space for headers and prepare control bits. */ 2572 skb_reserve(skb, MAX_TCP_HEADER); 2573 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2574 tcp_init_nondata_skb(skb, tp->write_seq, 2575 TCPHDR_ACK | TCPHDR_FIN); 2576 tcp_queue_skb(sk, skb); 2577 } 2578 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 2579 } 2580 2581 /* We get here when a process closes a file descriptor (either due to 2582 * an explicit close() or as a byproduct of exit()'ing) and there 2583 * was unread data in the receive queue. This behavior is recommended 2584 * by RFC 2525, section 2.17. -DaveM 2585 */ 2586 void tcp_send_active_reset(struct sock *sk, gfp_t priority) 2587 { 2588 struct sk_buff *skb; 2589 2590 /* NOTE: No TCP options attached and we never retransmit this. */ 2591 skb = alloc_skb(MAX_TCP_HEADER, priority); 2592 if (!skb) { 2593 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 2594 return; 2595 } 2596 2597 /* Reserve space for headers and prepare control bits. */ 2598 skb_reserve(skb, MAX_TCP_HEADER); 2599 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2600 TCPHDR_ACK | TCPHDR_RST); 2601 /* Send it off. */ 2602 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2603 if (tcp_transmit_skb(sk, skb, 0, priority)) 2604 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 2605 2606 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 2607 } 2608 2609 /* Send a crossed SYN-ACK during socket establishment. 2610 * WARNING: This routine must only be called when we have already sent 2611 * a SYN packet that crossed the incoming SYN that caused this routine 2612 * to get called. If this assumption fails then the initial rcv_wnd 2613 * and rcv_wscale values will not be correct. 2614 */ 2615 int tcp_send_synack(struct sock *sk) 2616 { 2617 struct sk_buff *skb; 2618 2619 skb = tcp_write_queue_head(sk); 2620 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 2621 pr_debug("%s: wrong queue state\n", __func__); 2622 return -EFAULT; 2623 } 2624 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 2625 if (skb_cloned(skb)) { 2626 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 2627 if (nskb == NULL) 2628 return -ENOMEM; 2629 tcp_unlink_write_queue(skb, sk); 2630 skb_header_release(nskb); 2631 __tcp_add_write_queue_head(sk, nskb); 2632 sk_wmem_free_skb(sk, skb); 2633 sk->sk_wmem_queued += nskb->truesize; 2634 sk_mem_charge(sk, nskb->truesize); 2635 skb = nskb; 2636 } 2637 2638 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 2639 TCP_ECN_send_synack(tcp_sk(sk), skb); 2640 } 2641 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2642 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2643 } 2644 2645 /** 2646 * tcp_make_synack - Prepare a SYN-ACK. 2647 * sk: listener socket 2648 * dst: dst entry attached to the SYNACK 2649 * req: request_sock pointer 2650 * 2651 * Allocate one skb and build a SYNACK packet. 2652 * @dst is consumed : Caller should not use it again. 2653 */ 2654 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2655 struct request_sock *req, 2656 struct tcp_fastopen_cookie *foc) 2657 { 2658 struct tcp_out_options opts; 2659 struct inet_request_sock *ireq = inet_rsk(req); 2660 struct tcp_sock *tp = tcp_sk(sk); 2661 struct tcphdr *th; 2662 struct sk_buff *skb; 2663 struct tcp_md5sig_key *md5; 2664 int tcp_header_size; 2665 int mss; 2666 2667 skb = alloc_skb(MAX_TCP_HEADER + 15, sk_gfp_atomic(sk, GFP_ATOMIC)); 2668 if (unlikely(!skb)) { 2669 dst_release(dst); 2670 return NULL; 2671 } 2672 /* Reserve space for headers. */ 2673 skb_reserve(skb, MAX_TCP_HEADER); 2674 2675 skb_dst_set(skb, dst); 2676 security_skb_owned_by(skb, sk); 2677 2678 mss = dst_metric_advmss(dst); 2679 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2680 mss = tp->rx_opt.user_mss; 2681 2682 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 2683 __u8 rcv_wscale; 2684 /* Set this up on the first call only */ 2685 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2686 2687 /* limit the window selection if the user enforce a smaller rx buffer */ 2688 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2689 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2690 req->window_clamp = tcp_full_space(sk); 2691 2692 /* tcp_full_space because it is guaranteed to be the first packet */ 2693 tcp_select_initial_window(tcp_full_space(sk), 2694 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 2695 &req->rcv_wnd, 2696 &req->window_clamp, 2697 ireq->wscale_ok, 2698 &rcv_wscale, 2699 dst_metric(dst, RTAX_INITRWND)); 2700 ireq->rcv_wscale = rcv_wscale; 2701 } 2702 2703 memset(&opts, 0, sizeof(opts)); 2704 #ifdef CONFIG_SYN_COOKIES 2705 if (unlikely(req->cookie_ts)) 2706 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 2707 else 2708 #endif 2709 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2710 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5, 2711 foc) + sizeof(*th); 2712 2713 skb_push(skb, tcp_header_size); 2714 skb_reset_transport_header(skb); 2715 2716 th = tcp_hdr(skb); 2717 memset(th, 0, sizeof(struct tcphdr)); 2718 th->syn = 1; 2719 th->ack = 1; 2720 TCP_ECN_make_synack(req, th); 2721 th->source = ireq->loc_port; 2722 th->dest = ireq->rmt_port; 2723 /* Setting of flags are superfluous here for callers (and ECE is 2724 * not even correctly set) 2725 */ 2726 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2727 TCPHDR_SYN | TCPHDR_ACK); 2728 2729 th->seq = htonl(TCP_SKB_CB(skb)->seq); 2730 /* XXX data is queued and acked as is. No buffer/window check */ 2731 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 2732 2733 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2734 th->window = htons(min(req->rcv_wnd, 65535U)); 2735 tcp_options_write((__be32 *)(th + 1), tp, &opts); 2736 th->doff = (tcp_header_size >> 2); 2737 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); 2738 2739 #ifdef CONFIG_TCP_MD5SIG 2740 /* Okay, we have all we need - do the md5 hash if needed */ 2741 if (md5) { 2742 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 2743 md5, NULL, req, skb); 2744 } 2745 #endif 2746 2747 return skb; 2748 } 2749 EXPORT_SYMBOL(tcp_make_synack); 2750 2751 /* Do all connect socket setups that can be done AF independent. */ 2752 void tcp_connect_init(struct sock *sk) 2753 { 2754 const struct dst_entry *dst = __sk_dst_get(sk); 2755 struct tcp_sock *tp = tcp_sk(sk); 2756 __u8 rcv_wscale; 2757 2758 /* We'll fix this up when we get a response from the other end. 2759 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 2760 */ 2761 tp->tcp_header_len = sizeof(struct tcphdr) + 2762 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 2763 2764 #ifdef CONFIG_TCP_MD5SIG 2765 if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2766 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2767 #endif 2768 2769 /* If user gave his TCP_MAXSEG, record it to clamp */ 2770 if (tp->rx_opt.user_mss) 2771 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 2772 tp->max_window = 0; 2773 tcp_mtup_init(sk); 2774 tcp_sync_mss(sk, dst_mtu(dst)); 2775 2776 if (!tp->window_clamp) 2777 tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 2778 tp->advmss = dst_metric_advmss(dst); 2779 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2780 tp->advmss = tp->rx_opt.user_mss; 2781 2782 tcp_initialize_rcv_mss(sk); 2783 2784 /* limit the window selection if the user enforce a smaller rx buffer */ 2785 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2786 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2787 tp->window_clamp = tcp_full_space(sk); 2788 2789 tcp_select_initial_window(tcp_full_space(sk), 2790 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 2791 &tp->rcv_wnd, 2792 &tp->window_clamp, 2793 sysctl_tcp_window_scaling, 2794 &rcv_wscale, 2795 dst_metric(dst, RTAX_INITRWND)); 2796 2797 tp->rx_opt.rcv_wscale = rcv_wscale; 2798 tp->rcv_ssthresh = tp->rcv_wnd; 2799 2800 sk->sk_err = 0; 2801 sock_reset_flag(sk, SOCK_DONE); 2802 tp->snd_wnd = 0; 2803 tcp_init_wl(tp, 0); 2804 tp->snd_una = tp->write_seq; 2805 tp->snd_sml = tp->write_seq; 2806 tp->snd_up = tp->write_seq; 2807 tp->snd_nxt = tp->write_seq; 2808 2809 if (likely(!tp->repair)) 2810 tp->rcv_nxt = 0; 2811 tp->rcv_wup = tp->rcv_nxt; 2812 tp->copied_seq = tp->rcv_nxt; 2813 2814 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2815 inet_csk(sk)->icsk_retransmits = 0; 2816 tcp_clear_retrans(tp); 2817 } 2818 2819 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 2820 { 2821 struct tcp_sock *tp = tcp_sk(sk); 2822 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 2823 2824 tcb->end_seq += skb->len; 2825 skb_header_release(skb); 2826 __tcp_add_write_queue_tail(sk, skb); 2827 sk->sk_wmem_queued += skb->truesize; 2828 sk_mem_charge(sk, skb->truesize); 2829 tp->write_seq = tcb->end_seq; 2830 tp->packets_out += tcp_skb_pcount(skb); 2831 } 2832 2833 /* Build and send a SYN with data and (cached) Fast Open cookie. However, 2834 * queue a data-only packet after the regular SYN, such that regular SYNs 2835 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 2836 * only the SYN sequence, the data are retransmitted in the first ACK. 2837 * If cookie is not cached or other error occurs, falls back to send a 2838 * regular SYN with Fast Open cookie request option. 2839 */ 2840 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 2841 { 2842 struct tcp_sock *tp = tcp_sk(sk); 2843 struct tcp_fastopen_request *fo = tp->fastopen_req; 2844 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; 2845 struct sk_buff *syn_data = NULL, *data; 2846 unsigned long last_syn_loss = 0; 2847 2848 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 2849 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 2850 &syn_loss, &last_syn_loss); 2851 /* Recurring FO SYN losses: revert to regular handshake temporarily */ 2852 if (syn_loss > 1 && 2853 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 2854 fo->cookie.len = -1; 2855 goto fallback; 2856 } 2857 2858 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 2859 fo->cookie.len = -1; 2860 else if (fo->cookie.len <= 0) 2861 goto fallback; 2862 2863 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 2864 * user-MSS. Reserve maximum option space for middleboxes that add 2865 * private TCP options. The cost is reduced data space in SYN :( 2866 */ 2867 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 2868 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 2869 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 2870 MAX_TCP_OPTION_SPACE; 2871 2872 syn_data = skb_copy_expand(syn, skb_headroom(syn), space, 2873 sk->sk_allocation); 2874 if (syn_data == NULL) 2875 goto fallback; 2876 2877 for (i = 0; i < iovlen && syn_data->len < space; ++i) { 2878 struct iovec *iov = &fo->data->msg_iov[i]; 2879 unsigned char __user *from = iov->iov_base; 2880 int len = iov->iov_len; 2881 2882 if (syn_data->len + len > space) 2883 len = space - syn_data->len; 2884 else if (i + 1 == iovlen) 2885 /* No more data pending in inet_wait_for_connect() */ 2886 fo->data = NULL; 2887 2888 if (skb_add_data(syn_data, from, len)) 2889 goto fallback; 2890 } 2891 2892 /* Queue a data-only packet after the regular SYN for retransmission */ 2893 data = pskb_copy(syn_data, sk->sk_allocation); 2894 if (data == NULL) 2895 goto fallback; 2896 TCP_SKB_CB(data)->seq++; 2897 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; 2898 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); 2899 tcp_connect_queue_skb(sk, data); 2900 fo->copied = data->len; 2901 2902 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { 2903 tp->syn_data = (fo->copied > 0); 2904 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); 2905 goto done; 2906 } 2907 syn_data = NULL; 2908 2909 fallback: 2910 /* Send a regular SYN with Fast Open cookie request option */ 2911 if (fo->cookie.len > 0) 2912 fo->cookie.len = 0; 2913 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 2914 if (err) 2915 tp->syn_fastopen = 0; 2916 kfree_skb(syn_data); 2917 done: 2918 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 2919 return err; 2920 } 2921 2922 /* Build a SYN and send it off. */ 2923 int tcp_connect(struct sock *sk) 2924 { 2925 struct tcp_sock *tp = tcp_sk(sk); 2926 struct sk_buff *buff; 2927 int err; 2928 2929 tcp_connect_init(sk); 2930 2931 if (unlikely(tp->repair)) { 2932 tcp_finish_connect(sk, NULL); 2933 return 0; 2934 } 2935 2936 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 2937 if (unlikely(buff == NULL)) 2938 return -ENOBUFS; 2939 2940 /* Reserve space for headers. */ 2941 skb_reserve(buff, MAX_TCP_HEADER); 2942 2943 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 2944 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; 2945 tcp_connect_queue_skb(sk, buff); 2946 TCP_ECN_send_syn(sk, buff); 2947 2948 /* Send off SYN; include data in Fast Open. */ 2949 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 2950 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 2951 if (err == -ECONNREFUSED) 2952 return err; 2953 2954 /* We change tp->snd_nxt after the tcp_transmit_skb() call 2955 * in order to make this packet get counted in tcpOutSegs. 2956 */ 2957 tp->snd_nxt = tp->write_seq; 2958 tp->pushed_seq = tp->write_seq; 2959 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 2960 2961 /* Timer for repeating the SYN until an answer. */ 2962 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2963 inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 2964 return 0; 2965 } 2966 EXPORT_SYMBOL(tcp_connect); 2967 2968 /* Send out a delayed ack, the caller does the policy checking 2969 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 2970 * for details. 2971 */ 2972 void tcp_send_delayed_ack(struct sock *sk) 2973 { 2974 struct inet_connection_sock *icsk = inet_csk(sk); 2975 int ato = icsk->icsk_ack.ato; 2976 unsigned long timeout; 2977 2978 if (ato > TCP_DELACK_MIN) { 2979 const struct tcp_sock *tp = tcp_sk(sk); 2980 int max_ato = HZ / 2; 2981 2982 if (icsk->icsk_ack.pingpong || 2983 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 2984 max_ato = TCP_DELACK_MAX; 2985 2986 /* Slow path, intersegment interval is "high". */ 2987 2988 /* If some rtt estimate is known, use it to bound delayed ack. 2989 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 2990 * directly. 2991 */ 2992 if (tp->srtt) { 2993 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 2994 2995 if (rtt < max_ato) 2996 max_ato = rtt; 2997 } 2998 2999 ato = min(ato, max_ato); 3000 } 3001 3002 /* Stay within the limit we were given */ 3003 timeout = jiffies + ato; 3004 3005 /* Use new timeout only if there wasn't a older one earlier. */ 3006 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 3007 /* If delack timer was blocked or is about to expire, 3008 * send ACK now. 3009 */ 3010 if (icsk->icsk_ack.blocked || 3011 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 3012 tcp_send_ack(sk); 3013 return; 3014 } 3015 3016 if (!time_before(timeout, icsk->icsk_ack.timeout)) 3017 timeout = icsk->icsk_ack.timeout; 3018 } 3019 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3020 icsk->icsk_ack.timeout = timeout; 3021 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 3022 } 3023 3024 /* This routine sends an ack and also updates the window. */ 3025 void tcp_send_ack(struct sock *sk) 3026 { 3027 struct sk_buff *buff; 3028 3029 /* If we have been reset, we may not send again. */ 3030 if (sk->sk_state == TCP_CLOSE) 3031 return; 3032 3033 /* We are not putting this on the write queue, so 3034 * tcp_transmit_skb() will set the ownership to this 3035 * sock. 3036 */ 3037 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 3038 if (buff == NULL) { 3039 inet_csk_schedule_ack(sk); 3040 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 3041 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 3042 TCP_DELACK_MAX, TCP_RTO_MAX); 3043 return; 3044 } 3045 3046 /* Reserve space for headers and prepare control bits. */ 3047 skb_reserve(buff, MAX_TCP_HEADER); 3048 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 3049 3050 /* Send it off, this clears delayed acks for us. */ 3051 TCP_SKB_CB(buff)->when = tcp_time_stamp; 3052 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 3053 } 3054 3055 /* This routine sends a packet with an out of date sequence 3056 * number. It assumes the other end will try to ack it. 3057 * 3058 * Question: what should we make while urgent mode? 3059 * 4.4BSD forces sending single byte of data. We cannot send 3060 * out of window data, because we have SND.NXT==SND.MAX... 3061 * 3062 * Current solution: to send TWO zero-length segments in urgent mode: 3063 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 3064 * out-of-date with SND.UNA-1 to probe window. 3065 */ 3066 static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 3067 { 3068 struct tcp_sock *tp = tcp_sk(sk); 3069 struct sk_buff *skb; 3070 3071 /* We don't queue it, tcp_transmit_skb() sets ownership. */ 3072 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 3073 if (skb == NULL) 3074 return -1; 3075 3076 /* Reserve space for headers and set control bits. */ 3077 skb_reserve(skb, MAX_TCP_HEADER); 3078 /* Use a previous sequence. This should cause the other 3079 * end to send an ack. Don't queue or clone SKB, just 3080 * send it. 3081 */ 3082 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3083 TCP_SKB_CB(skb)->when = tcp_time_stamp; 3084 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 3085 } 3086 3087 void tcp_send_window_probe(struct sock *sk) 3088 { 3089 if (sk->sk_state == TCP_ESTABLISHED) { 3090 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3091 tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq; 3092 tcp_xmit_probe_skb(sk, 0); 3093 } 3094 } 3095 3096 /* Initiate keepalive or window probe from timer. */ 3097 int tcp_write_wakeup(struct sock *sk) 3098 { 3099 struct tcp_sock *tp = tcp_sk(sk); 3100 struct sk_buff *skb; 3101 3102 if (sk->sk_state == TCP_CLOSE) 3103 return -1; 3104 3105 if ((skb = tcp_send_head(sk)) != NULL && 3106 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 3107 int err; 3108 unsigned int mss = tcp_current_mss(sk); 3109 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 3110 3111 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 3112 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 3113 3114 /* We are probing the opening of a window 3115 * but the window size is != 0 3116 * must have been a result SWS avoidance ( sender ) 3117 */ 3118 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 3119 skb->len > mss) { 3120 seg_size = min(seg_size, mss); 3121 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3122 if (tcp_fragment(sk, skb, seg_size, mss)) 3123 return -1; 3124 } else if (!tcp_skb_pcount(skb)) 3125 tcp_set_skb_tso_segs(sk, skb, mss); 3126 3127 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3128 TCP_SKB_CB(skb)->when = tcp_time_stamp; 3129 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3130 if (!err) 3131 tcp_event_new_data_sent(sk, skb); 3132 return err; 3133 } else { 3134 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3135 tcp_xmit_probe_skb(sk, 1); 3136 return tcp_xmit_probe_skb(sk, 0); 3137 } 3138 } 3139 3140 /* A window probe timeout has occurred. If window is not closed send 3141 * a partial packet else a zero probe. 3142 */ 3143 void tcp_send_probe0(struct sock *sk) 3144 { 3145 struct inet_connection_sock *icsk = inet_csk(sk); 3146 struct tcp_sock *tp = tcp_sk(sk); 3147 int err; 3148 3149 err = tcp_write_wakeup(sk); 3150 3151 if (tp->packets_out || !tcp_send_head(sk)) { 3152 /* Cancel probe timer, if it is not required. */ 3153 icsk->icsk_probes_out = 0; 3154 icsk->icsk_backoff = 0; 3155 return; 3156 } 3157 3158 if (err <= 0) { 3159 if (icsk->icsk_backoff < sysctl_tcp_retries2) 3160 icsk->icsk_backoff++; 3161 icsk->icsk_probes_out++; 3162 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3163 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 3164 TCP_RTO_MAX); 3165 } else { 3166 /* If packet was not sent due to local congestion, 3167 * do not backoff and do not remember icsk_probes_out. 3168 * Let local senders to fight for local resources. 3169 * 3170 * Use accumulated backoff yet. 3171 */ 3172 if (!icsk->icsk_probes_out) 3173 icsk->icsk_probes_out = 1; 3174 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3175 min(icsk->icsk_rto << icsk->icsk_backoff, 3176 TCP_RESOURCE_PROBE_INTERVAL), 3177 TCP_RTO_MAX); 3178 } 3179 } 3180